Complex Geometry - Jean-Pierre Demailly

Complex Analytic and Dierential Geometry
Jean-Pierre Demailly
Universit de Grenoble I e Institut Fourier, UMR 5582 du CNRS 38402 Saint-Martin dH`res, France e
Version of Thursday September 10, 2009
Table of Contents
Chapter I. Complex Dierential Calculus and Pseudoconvexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1. 2. 3. 4. 5. 6. 7. 8. Dierential Calculus on Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Currents on Dierentiable Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Holomorphic Functions and Complex Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 Subharmonic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Plurisubharmonic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Domains of Holomorphy and Stein Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Pseudoconvex Open Sets in Cn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .60
Chapter II. Coherent Sheaves and Analytic Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 1. Presheaves and Sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 2. The Local Ring of Germs of Analytic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3. Coherent Sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 4. Complex Analytic Sets. Local Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 5. Complex Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 6. Analytic Cycles and Meromorphic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 7. Normal Spaces and Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 8. Holomorphic Mappings and Extension Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 9. Complex Analytic Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 10. Bimeromorphic maps, Modications and Blow-ups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 11. Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Chapter III. Positive Currents and Lelong Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 1. Basic Concepts of Positivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 2. Closed Positive Currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 3. Denition of Monge-Amp`re Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 e 4. Case of Unbounded Plurisubharmonic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 5. Generalized Lelong Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 6. The Jensen-Lelong Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 7. Comparison Theorems for Lelong Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 8. Sius Semicontinuity Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 9. Transformation of Lelong Numbers by Direct Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 10. A Schwarz Lemma. Application to Number Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Chapter IV. Sheaf Cohomology and Spectral Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 1. 2. 3. 4. Basic Results of Homological Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 The Simplicial Flabby Resolution of a Sheaf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Cohomology Groups with Values in a Sheaf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Acyclic Sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Table of Contents 5. Cech Cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 6. The De Rham-Weil Isomorphism Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 7. Cohomology with Supports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .216 8. Cup Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 9. Inverse Images and Cartesian Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 10. Spectral Sequence of a Filtered Complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 11. Spectral Sequence of a Double Complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 12. Hypercohomology Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 13. Direct Images and the Leray Spectral Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 14. Alexander-Spanier Cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 15. Knneth Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 u 16. Poincar duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 e
Chapter V. Hermitian Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 1. Denition of Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 2. Linear Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254 3. Curvature Tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 4. Operations on Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 5. Pull-Back of a Vector Bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258 6. Parallel Translation and Flat Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 7. Hermitian Vector Bundles and Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 8. Vector Bundles and Locally Free Sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 9. First Chern Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 10. Connections of Type (1,0) and (0,1) over Complex Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 11. Holomorphic Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 12. Chern Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 13. Lelong-Poincar Equation and First Chern Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270 e 14. Exact Sequences of Hermitian Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 15. Line Bundles (k) over Pn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276 16. Grassmannians and Universal Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Chapter VI. Hodge Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 1. Dierential Operators on Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 2. Formalism of PseudoDierential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 3. Harmonic Forms and Hodge Theory on Riemannian Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 4. Hermitian and Khler Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 a 5. Basic Results of Khler Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 a 6. Commutation Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 7. Groups p,q (X, E) and Serre Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 8. Cohomology of Compact Khler Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 a 9. Jacobian and Albanese Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313 10. Complex Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 11. Hodge-Frlicher Spectral Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 o 12. Eect of a Modication on Hodge Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 Chapter VII. Positive Vector Bundles and Vanishing Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 329 1. 2. 3. 4. 5. 6. Bochner-Kodaira-Nakano Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 Basic a Priori Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332 Kodaira-Akizuki-Nakano Vanishing Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 Girbaus Vanishing Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334 Vanishing Theorem for Partially Positive Line Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336 Positivity Concepts for Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
Table of Contents
7. Nakano Vanishing Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340 8. Relations Between Nakano and Griths Positivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342 9. Applications to Griths Positive Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 10. Cohomology Groups of (k) over Pn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 11. Ample Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 12. Blowing-up along a Submanifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353 13. Equivalence of Positivity and Ampleness for Line Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 14. Kodairas Projectivity Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 Chapter VIII. L2 Estimates on Pseudoconvex Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .363 1. Non Bounded Operators on Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 2. Complete Riemannian Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365 3. L2 Hodge Theory on Complete Riemannian Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 4. General Estimate for d on Hermitian Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370 5. Estimates on Weakly Pseudoconvex Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372 6. Hrmanders Estimates for non Complete Khler Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375 o a 7. Extension of Holomorphic Functions from Subvarieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 8. Applications to Hypersurface Singularities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 9. Skodas L2 Estimates for Surjective Bundle Morphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388 10. Application of Skodas L2 Estimates to Local Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 11. Integrability of Almost Complex Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396 Chapter IX. Finiteness Theorems for q-Convex Spaces and Stein Spaces . . . . . . . . . . . . . . . .403 1. 2. 3. 4. 5. Topological Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 q-Convex Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409 q-Convexity Properties in Top Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414 Andreotti-Grauert Finiteness Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 Grauerts Direct Image Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Chapter I
Complex Dierential Calculus and Pseudoconvexity
This introductive chapter is mainly a review of the basic tools and concepts which will be employed in the rest of the book: dierential forms, currents, holomorphic and plurisubharmonic functions, holomorphic convexity and pseudoconvexity. Our study of holomorphic convexity is principally concentrated here on the case of domains in Cn . The more powerful machinery needed for the study of general complex varieties (sheaves, positive currents, hermitian dierential geometry) will be introduced in Chapters II to V. Although our exposition pretends to be almost self-contained, the reader is assumed to have at least a vague familiarity with a few basic topics, such as dierential calculus, measure theory and distributions, holomorphic functions of one complex variable, . . . . Most of the necessary background can be found in the books of [Rudin 1966] and [Warner 1971]; the basics of distribution theory can be found in Chapter I of [Hrmander 1963]. On the other hand, the reader who has already some knowledge of o complex analysis in several variables should probably bypass this chapter.
1. Dierential Calculus on Manifolds

1.A. Dierentiable Manifolds The notion of manifold is a natural extension of the notion of submanifold dened by a set of equations in Rn . However, as already observed by Riemann during the 19th century, it is important to dene the notion of a manifold in a exible way, without necessarily requiring that the underlying topological space is embedded in an ane space. The precise formal denition was rst introduced by H. Weyl in [Weyl 1913]. Let m N and k N {, }. We denote by k the class of functions which are k-times dierentiable with continuous derivatives if k = , and by C the class of real analytic functions. A dierentiable manifold M of real dimension m and of class k is a topological space (which we shall always assume Hausdor and separable, i.e. possessing a countable basis of the topology), equipped with an atlas of class k with values in Rm . An atlas of class k is a collection of homeomorphisms : U V , I, called dierentiable charts, such that (U )I is an open covering of M and V an open subset of Rm , and such that for all , I the transition map (1.1)
1 = : (U U ) (U U )
is a k dieomorphism from an open subset of V onto an open subset of V (see Fig. 1). Then the components (x) = (x , . . . , x ) are called the local coordinates on U dened 1 m by the chart ; they are related by the transition relation x = (x ).
Chapter I. Complex Dierential Calculus and Pseudoconvexity
V
(U U )
U
U U

(U U )
U M
V Rm Fig. I-1 Charts and transition maps
If M is open and s N {, }, 0 s k, we denote by C s (, R) the set of 1 functions f of class C s on , i.e. such that f is of class C s on (U ) for each ; if is not open, C s (, R) is the set of functions which have a C s extension to some neighborhood of . A tangent vector at a point a M is by denition a dierential operator acting on functions, of the type C 1 (, R) f f = j
1 j m
f (a) xj
in any local coordinate system (x1 , . . . , xm ) on an open set a. We then simply write = j /xj . For every a , the n-tuple (/xj )1 j m is therefore a basis of the tangent space to M at a, which we denote by TM,a . The dierential of a function f at a is the linear form on TM,a dened by dfa () = f = j f /xj (a), TM,a .
In particular dxj () = j and we may consequently write df = (f /xj )dxj . From this, we see that (dx1 , . . . , dxm ) is the dual basis of (/x1 , . . . , /xm ) in the cotangent space TM,a . The disjoint unions TM = xM TM,x and TM = xM TM,x are called the tangent and cotangent bundles of M . If is a vector eld of class C s over , that is, a map x (x) TM,x such that (x) = j (x) /xj has C s coecients, and if is another vector eld of class C s with s 1, the Lie bracket [, ] is the vector eld such that (1.2) [, ] f = ( f ) ( f ). k k j xj xj . xk
In coordinates, it is easy to check that (1.3) [, ] =

1 j,k m
1.B. Dierential Forms A dierential form u of degree p, or briey a p-form over M , is a map u on M with values u(x) p TM,x . In a coordinate open set M , a dierential p-form can be written u(x) = uI (x) dxI ,
|I|=p
where I = (i1 , . . . , ip ) is a multi-index with integer components, i1 < . . . < ip and dxI := dxi1 . . . dxip . The notation |I| stands for the number of components of I, and is read length of I. For all integers p = 0, 1, . . . , m and s N {}, s k, we denote by C s (M, p TM ) the space of dierential p-forms of class C s , i.e. with C s coecients uI . Several natural operations on dierential forms can be dened. 1.B.1. Wedge Product. If v(x) = vJ (x) dxJ is a q-form, the wedge product of u and v is the form of degree (p + q) dened by (1.4) u v(x) = uI (x)vJ (x) dxI dxJ .
|I|=p,|J|=q
1.B.2. Contraction by a tangent vector. A p-form u can be viewed as an antisymmetric p-linear form on TM . If = j /xj is a tangent vector, we dene the contraction u to be the dierential form of degree p 1 such that (1.5) ( u)(1 , . . . , p1 ) = u(, 1 , . . . , p1 ) u is bilinear and we nd easily if j I, / if j = il I. for all tangent vectors j . Then (, u) xj dxI =
0 (1)l1 dxI
{j}
A simple computation based on the above formula shows that contraction by a tangent vector is a derivation, i.e. (1.6) (u v) = ( u) v + (1)deg u u ( v).
1.B.3. Exterior derivative. This is the dierential operator

d : C s (M, p TM ) C s1 (M, p+1 TM )
dened in local coordinates by the formula (1.7) du =

|I|=p, 1 k m
uI dxk dxI . xk
Alternatively, one can dene du by its action on arbitrary vector elds 0 , . . . , p on M . The formula is as follows du(0 , . . . , p ) = (1.7 ) +
0 j p
(1)j j u(0 , . . . , j , . . . , p ) (1)j+k u([j , k ], 0 , . . . , j , . . . , k , . . . , p ).
0 j<k p
10
The reader will easily check that (1.7) actually implies (1.7 ). The advantage of (1.7 ) is that it does not depend on the choice of coordinates, thus du is intrinsically dened. The two basic properties of the exterior derivative (again left to the reader) are: (1.8) (1.9) d(u v) = du v + (1)deg u u dv, d2 = 0. ( Leibnitz rule )
A form u is said to be closed if du = 0 and exact if u can be written u = dv for some form v. 1.B.4. De Rham Cohomology Groups. Recall that a cohomological complex K = p pZ is a collection of modules K over some ring, equipped with dierentials, i.e., linear maps dp : K p K p+1 such that dp+1 dp = 0. The cocycle, coboundary and cohomology modules Z p (K ), B p (K ) and H p (K ) are dened respectively by Z p (K ) K p , Z p (K ) = Ker dp : K p K p+1 , p p1 p1 p (1.10) B (K ) = Im d :K K , B p (K ) Z p (K ) K p , p H (K ) = Z p (K )/B p (K ).
Now, let M be a dierentiable manifold, say of class for simplicity. The De Rham complex of M is dened to be the complex K p = (M, p TM ) of smooth dierential p forms, together with the exterior derivative d = d as dierential, and K p = {0}, dp = 0 for p < 0. We denote by Z p (M, R) the cocycles (closed p-forms) and by B p (M, R) the coboundaries (exact p-forms). By convention B 0 (M, R) = {0}. The De Rham cohomology group of M in degree p is (1.11)
p HDR (M, R) = Z p (M, R)/B p(M, R).
When no confusion with other types of cohomology groups may occur, we sometimes denote these groups simply by H p (M, R). The symbol R is used here to stress that we are p considering real valued p-forms; of course one can introduce a similar group HDR (M, C) p for complex valued forms, i.e. forms with values in C p TM . Then HDR (M, C) = p C HDR (M, R) is the complexication of the real De Rham cohomology group. It is 0 clear that HDR (M, R) can be identied with the space of locally constant functions on M , thus 0 HDR (M, R) = R0 (X) , where 0 (X) denotes the set of connected components of M . Similarly, we introduce the De Rham cohomology groups with compact support (1.12)
p p p HDR,c (M, R) = Zc (M, R)/Bc (M, R), p c (M, TM )
associated with the De Rham complex K p = with compact support.
of smooth dierential forms
1.B.5. Pull-Back. If F : M M is a dierentiable map to another manifold M , dimR M = m , and if v(y) = vJ (y) dyJ is a dierential p-form on M , the pull-back F v is the dierential p-form on M obtained after making the substitution y = F (x) in v, i.e. (1.13) F v(x) = vI F (x) dFi1 . . . dFip .
11
If we have a second map G : M M and if w is a dierential form on M , then F (G w) is obtained by means of the substitutions z = G(y), y = F (x), thus (1.14) F (G w) = (G F ) w.
Moreover, we always have d(F v) = F (dv). It follows that the pull-back F is closed if v is closed and exact if v is exact. Therefore F induces a morphism on the quotient spaces (1.15)
p p F : HDR (M , R) HDR (M, R).
1.C. Integration of Dierential Forms A manifold M is orientable if and only if there exists an atlas ( ) such that all transition maps preserve the orientation, i.e. have positive jacobian determinants. Suppose that M is oriented, that is, equipped with such an atlas. If u(x) = f (x1 , . . . , xm ) dx1 . . . dxm is a continuous form of maximum degree m = dimR M , with compact support in a coordinate open set , we set (1.16)
M
u=
Rm
f (x1 , . . . , xm ) dx1 . . . dxm .
By the change of variable formula, the result is independent of the choice of coordinates, provided we consider only coordinates corresponding to the given orientation. When u is an arbitrary form with compact support, the denition of M u is easily extended by means of a partition of unity with respect to coordinate open sets covering Supp u. Let F : M M be a dieomorphism between oriented manifolds and v a volume form on M . The change of variable formula yields (1.17)
M
F v =
v
M
according whether F preserves orientation or not. We now state Stokes formula, which is basic in many contexts. Let K be a compact subset of M with piecewise C 1 boundary. By this, we mean that for each point a K there are coordinates (x1 , . . . , xm ) on a neighborhood V of a, centered at a, such that K V = x V ; x1 for some index l C 1 boundaries: 0, . . . , xl 0
1. Then K V is a union of smooth hypersurfaces with piecewise x V ; x1 0, . . . , xj = 0, . . . , xl 0 .
K V =
1 j l
At points of K where xj = 0, then (x1 , . . . , xj , , . . . , xm ) dene coordinates on K. We take the orientation of K given by these coordinates or the opposite one, according to the sign (1)j1 . For any dierential form u of class C 1 and degree m 1 on M , we then have
12
(1.18) Stokes formula.

K
u=
K
du.
The formula is easily checked by an explicit computation when u has compact support in V : indeed if u = 1 j n uj dx1 . . . dxj . . . dxm and j K V is the part of K V where xj = 0, a partial integration with respect to xj yields uj dx1 . . . dxj . . . dxm = u=
KV
j KV
uj dx1 . . . dxm , xj du.

V
(1)j1
1 j m j KV
uj dx1 . . . dxj . . . dxm =
The general case follows by a partition of unity. In particular, if u has compact support in M , we nd M du = 0 by choosing K Supp u. 1.D. Homotopy Formula and Poincar Lemma e Let u be a dierential form on [0, 1] M . For (t, x) [0, 1] M , we write u(t, x) =
|I|=p
uI (t, x) dxI +
|J|=p1
uJ (t, x) dt dxJ .
We dene an operator
K : C s ([0, 1] M, p T[0,1]M ) C s (M, p1 TM ) 1
(1.19)
Ku(x) =
|J|=p1 0
uJ (t, x) dt dxJ
and say that Ku is the form obtained by integrating u along [0, 1]. A computation of the operator dK + Kd shows that all terms involving partial derivatives uJ /xk cancel, hence
1
Kdu + dKu = (1.20) Kdu + dKu =

|I|=p i u i u, 1 0 0
uI (t, x) dt dxI = t
|I|=p
uI (1, x) uI (0, x) dxI ,
where it : M [0, 1] M is the injection x (t, x). (1.20) Corollary. Let F, G : M M be maps. Suppose that F, G are smoothly homotopic, i.e. that there exists a map H : [0, 1] M M such that H(0, x) = F (x) and H(1, x) = G(x). Then
p p F = G : HDR (M , R) HDR (M, R).
Proof. If v is a p-form on M , then G v F v = (H i1 ) v (H i0 ) v = i (H v) i (H v) 1 0 = d(KH v) + KH (dv)
2. Currents on Dierentiable Manifolds
13
by (1.20) applied to u = H v. If v is closed, then F v and G v dier by an exact form, p so they dene the same class in HDR (M, R). (1.21) Corollary. If the manifold M is contractible, i.e. if there is a smooth homotopy 0 H : [0, 1] M M from a constant map F : M {x0 } to G = IdX , then HDR (M, R) = p R and HDR (M, R) = 0 for p 1.
0 Proof. F is clearly zero in degree p 1, while F : HDR (M, R) R is induced by the evaluation map u u(x0 ). The conclusion then follows from the equality F = G = Id on cohomology groups.
(1.22) Poincar lemma. Let Rm be a starshaped open set. If a form v = e vI dxI C s (, p T ), p 1, satises dv = 0, there exists a form u C s (, p1 T ) such that du = v. Proof. Let H(t, x) = tx be the homotopy between the identity map and the constant map {0}. By the above formula d(KH v) = G v F v = v v(0) if p = 0, v if p 1.
Hence u = KH v is the (p 1)-form we are looking for. An explicit computation based on (1.19) easily gives
1
(1.23)
u(x) =
|I|=p 1 k p 0
tp1 vI (tx) dt (1)k1 xik dxi1 . . . dxik . . . dxip .

2.A. Denition and Examples Let M be a dierentiable manifold, m = dimR M . All the manifolds considered in Sect. 2 will be assumed to be oriented. We rst introduce a topology on the space of dierential forms C s (M, p TM ). Let M be a coordinate open set and u a p-form on M , written u(x) = uI (x) dxI on . To every compact subset L and every integer s N, we associate a seminorm (2.1) ps (u) = sup L
xL |I|=p,|| s
max
|D uI (x)|,
where = (1 , . . . , m ) runs over Nm and D = || /x1 1 . . . xm is a derivation of m order || = 1 + + m . This type of multi-index, which will always be denoted by Greek letters, should not be confused with multi-indices of the type I = (i1 , . . . , ip ) introduced in Sect. 1.
(2.2) Denition. We introduce as follows spaces of p-forms on manifolds.
14
a) We denote by p (M ) resp. s p (M ) the space (M, p TM ) resp. the space C s (M, p TM ) , equipped with the topology dened by all seminorms ps when s, L, L vary (resp. when L, vary).
b) If K M is a compact subset, p (K) will denote the subspace of elements u p (M ) with support contained in K, together with the induced topology; p (M ) will stand for the set of all elements with compact support, i.e. p (M ) := K p (K). c) The spaces of C s -forms
s p
(K) and
(M ) are dened similarly.
Since our manifolds are assumed to be separable, the topology of p (M ) can be dened by means of a countable set of seminorms ps , hence p (M ) (and likewise s p (M )) L is a Frchet space. The topology of s p (K) is induced by any nite set of seminorms ps j e K such that the compact sets Kj cover K ; hence s p (K) is a Banach space. It should be e observed however that p (M ) is not a Frchet space; in fact p (M ) is dense in p (M ) and thus non complete for the induced topology. According to [De Rham 1955] spaces of currents are dened as the topological duals of the above spaces, in analogy with the usual denition of distributions. (2.3) Denition. The space of currents of dimension p (or degree m p) on M is the space (M ) of linear forms T on p (M ) such that the restriction of T to all subspaces p p (K), K M , is continuous. The degree is indicated by raising the index, hence we set mp p (M ) = (M ) := topological dual (M ) . p The space s (M ) = s mp (M ) := p space of currents of order s on M .
s p
(M )
is dened similarly and is called the
In the sequel, we let T, u be the pairing between a current T and a test form u p (M ). It is clear that s (M ) can be identied with the subspace of currents p T (M ) which are continuous for the seminorm ps on p (K) for every compact set p K K contained in a coordinate patch . The support of T , denoted Supp T , is the smallest closed subset A M such that the restriction of T to p (M A) is zero. The topological dual (M ) can be identied with the set of currents of (M ) with compact support: p p indeed, let T be a linear form on p (M ) such that | T, u | C max{ps j (u)} K
for some s N, C 0 and a nite number of compact sets Kj ; it follows that Supp T Kj . Conversely let T (M ) with support in a compact set K. Let Kj be compact p patches such that K is contained in the interior of Kj and (M ) equal to 1 on K with Supp Kj . For u p (M ), we dene T, u = T, u ; this is independent of and the resulting T is clearly continuous on p (M ). The terminology used for the dimension and degree of a current is justied by the following two examples. (2.4) Example. Let Z M be a closed oriented submanifold of M of dimension p and class C 1 ; Z may have a boundary Z. The current of integration over Z, denoted [Z], is dened by [Z], u =
Z
u,
u0
(M ).
15
It is clear that [Z] is a current of order 0 on M and that Supp[Z] = Z. Its dimension is p = dim Z. (2.5) Example. If f is a dierential form of degree q on M with L1 coecients, we loc can associate to f the current of dimension m q : Tf , u =
M
f u,
u0
mq
(M ).
Tf is of degree q and of order 0. The correspondence f Tf is injective. In the same way L1 functions on Rm are identied to distributions, we will identify f with its image loc Tf 0 q (M ) = 0 mq (M ). 2.B. Exterior Derivative and Wedge Product 2.B.1. Exterior Derivative. Many of the operations available for dierential forms can be extended to currents by simple duality arguments. Let T s q (M ) = s mp (M ). The exterior derivative dT s+1 is dened by (2.6) dT, u = (1)q+1 T, du , u s+1
mq1 q+1
(M ) = s+1
mq1
(M ).
The continuity of the linear form dT on s+1 mq1 (M ) follows from the continuity of the map d : s+1 mq1 (K) s mq (K). For all forms f 1 q (M ) and u mq1 (M ), Stokes formula implies 0=
M
d(f u) =
df u + (1)q f du,
thus in example (2.5) one actually has dTf = Tdf as it should be. In example (2.4), another application of Stokes formula yields Z du = Z u, therefore [Z], du = [Z], u and (2.7) d[Z] = (1)mp+1 [Z].
s q
2.B.2. Wedge Product. For T T g s q+r (M ) is dened by (2.8)
(M ) and g us
(M ), the wedge product
T g, u = T, g u ,
mqr
(M ).
This denition is licit because u g u is continuous in the C s -topology. The relation d(T g) = dT g + (1)deg T T dg is easily veried from the denitions.
16
(2.9) Proposition. Let (x1 , . . . , xm ) be a coordinate system on an open subset M . Every current T s q (M ) of degree q can be written in a unique way T =
|I|=q
TI dxI
on ,
where TI are distributions of order s on , considered as currents of degree 0. Proof. If the result is true, for all f s
0
() we must have
T, f dxI = TI , dxI f dxI = (I, I) TI , f dx1 . . . dxm , where (I, I) is the signature of the permutation (1, . . . , m) (I, I). Conversely, this can be taken as a denition of the coecient TI : (2.10) TI (f ) = TI , f dx1 . . . dxm := (I, I) T, f dxI , f s
0
().
Then TI is a distribution of order s and it is easy to check that T =
TI dxI .
In particular, currents of order 0 on M can be considered as dierential forms with measure coecients. In order to unify the notations concerning forms and currents, we set T, u =
M
T u
whenever T s (M ) = s mp (M ) and u s p (M ) are such that Supp T Supp u p is compact. This convention is made so that the notation becomes compatible with the identication of a form f to the current Tf . 2.C. Direct and Inverse Images 2.C.1. Direct Images. Assume now that M1 , M2 are oriented dierentiable manifolds of respective dimensions m1 , m2 , and that (2.11) is a (2.12)
F : M1 M2 map. The pull-back morphism

s p
(M2 ) s
(M1 ),
u F u
is continuous in the C s topology and we have Supp F u F 1 (Supp u), but in general Supp F u is not compact. If T s (M1 ) is such that the restriction of F to Supp T p is proper, i.e. if Supp T F 1 (K) is compact for every compact subset K M2 , then the linear form u T, F u is well dened and continuous on s p (M2 ). There exists therefore a unique current denoted F T s (M2 ), called the direct image of T by F , p such that (2.13) F T, u = T, F u , u s
p
(M2 ).
We leave the straightforward proof of the following properties to the reader.
17
(2.14) Theorem. For every T s F T s (M2 ) is such that p a) Supp F T F (Supp T ) ; b) d(F T ) = F (dT ) ; c) F (T F g) = (F T ) g, d) If G : M2 M3 is a
p (M1 )
such that FSupp T is proper, the direct image
g s
(M2 , R) ;
map such that (G F )Supp T is proper, then G (F T ) = (G F ) T.
(2.15) Special case. Assume that F is a submersion, i.e. that F is surjective and that for every x M1 the dierential map dx F : TM1 ,x TM2 ,F (x) is surjective. Let g be a dierential form of degree q on M1 , with L1 coecients, such that FSupp g is proper. loc We claim that F g 0 1 q (M2 ) is the form of degree q (m1 m2 ) obtained from g m by integration along the bers of F , also denoted F g(y) =
zF 1 (y)
g(z).
Supp g
z=(x,y)
M1
M2
Fig. I-2 Local description of a submersion as a projection. In fact, this assertion is equivalent to the following generalized form of Fubinis theorem: g F u = g(z) u(y), u 0
m1 q
(M2 ).
M1
yM2
zF 1 (y)
By using a partition of unity on M1 and the constant rank theorem, the verication of this formula is easily reduced to the case where M1 = A M2 and F = pr2 , cf. Fig. 2. The bers F 1 (y) A have to be oriented in such a way that the orientation of M1 is the product of the orientation of A and M2 . Let us write r = dim A = m1 m2 and let z = (x, y) A M2 be any point of M1 . The above formula becomes
AM2
g(x, y) u(y) =
yM2
xA
g(x, y) u(y),
18
where the direct image of g is computed from g = by the formula (2.16) F g(y) =
xA
gI,J (x, y) dxI dyJ , |I| + |J| = q,
g(x, y) g(1,...,r),J (x, y) dx1 . . . dxr dyJ .
=
|J|=qr xA
In this situation, we see that F g has L1 coecients on M2 if g is L1 on M1 , and that loc loc the map g F g is continuous in the C s topology. (2.17) Remark. If F : M1 M2 is a dieomorphism, then we have F g = (F 1 ) g according whether F preserves the orientation or not. In fact formula (1.17) gives F g, u =
M1
g F u =
M2
(F 1 ) (g F u) =
M2
(F 1 ) g u.
2.C.2. Inverse Images. Assume that F : M1 M2 is a submersion. As a consequence of the continuity statement after (2.16), one can always dene the inverse image F T s q (M1 ) of a current T s q (M2 ) by F T, u = T, F u , us
q+m1 m2
(M1 ).
Then dim F T = dim T + m1 m2 and Th. 2.14 yields the formulas: (2.18) d(F T ) = F (dT ), F (T g) = F T F g, g s
(M2 ).
Take in particular T = [Z], where Z is an oriented C 1 -submanifold of M2 . Then F 1 (Z) is a submanifold of M1 and has a natural orientation given by the isomorphism TM1 ,x /TF 1 (Z),x TM2 ,F (x) /TZ,F (x) , induced by dx F at every point x Z. We claim that (2.19) F [Z] = [F 1 (Z)].
Indeed, we have to check that Z F u = F 1 (Z) u for every u s (M1 ). By using a partition of unity on M1 , we may again assume M1 = A M2 and F = pr2 . The above equality can be written F u(y) =
yZ (x,y)AZ
u(x, y).
This follows precisely from (2.16) and Fubinis theorem.
19
2.C.3. Weak Topology. The weak topology on (M ) is the topology dened by the p collection of seminorms T | T, f | for all f p (M ). With respect to the weak topology, all the operations (2.20) T dT, T T g, T F T, T F T
dened above are continuous. A set B (M ) is bounded for the weak topology p (weakly bounded for short) if and only if T, f is bounded when T runs over B, for p (M ). The standard Banach-Alaoglu theorem implies that every every xed f weakly bounded closed subset B (M ) is weakly compact. p 2.D. Tensor Products, Homotopies and Poincar Lemma e 2.D.1. Tensor Products. If S, T are currents on manifolds M , M there exists a unique current on M M , denoted S T and dened in a way analogous to the tensor product of distributions, such that for all u (M ) and v (M ) (2.21) S T, pr u pr v = (1)deg T deg u S, u T, v . 1 2 One veries easily that d(S T ) = dS T + (1)deg S S dT . 2.D.2. Homotopy Formula. Assume that H : [0, 1] M1 M2 is a homotopy from F (x) = H(0, x) to G(x) = H(1, x) and that T (M1 ) is a current such that H[0,1]Supp T is proper. If [0, 1] is considered as the current of degree 0 on R associated to its characteristic function, we nd d[0, 1] = 0 1 , thus d H ([0, 1] T ) = H (0 T 1 T + [0, 1] dT ) = F T G T + H ([0, 1] dT ). Therefore we obtain the homotopy formula (2.22) F T G T = d H ([0, 1] T ) H ([0, 1] dT ).
When T is closed, i.e. dT = 0, we see that F T and G T are cohomologous on M2 , i.e. they dier by an exact current dS. 2.D.3. Regularization of Currents. Let (Rm ) be a function with support in B(0, 1), such that (x) depends only on |x| = ( |xi |2 )1/2 , 0 and Rm (x) dx = 1. We associate to the family of functions ( ) such that (2.23) (x) = 1 x , m Supp B(0, ), (x) dx = 1.
Rm
We shall refer to this construction by saying that ( ) is a family of smoothing kernels. For every current T = TI dxI on an open subset Rm , the family of smooth forms T =
I
(TI ) dxI ,
dened on = {x Rm ; d(x, ) > }, converges weakly to T as tends to 0. Indeed, T , f = T, f and f converges to f in p () with respect to all seminorms ps . K
20
2.D.4. Poincar Lemma for Currents. Let T s q () be a closed current on an e open set Rm . We rst show that T is cohomologous to a smooth form. In fact, let (Rm ) be a cut-o function such that Supp , 0 < 1 and |d| 1 on . For any vector v B(0, 1) we set Fv (x) = x + (x)v. Since x (x)v is a contraction, Fv is a dieomorphism of Rm which leaves invariant pointwise, so Fv () = . This dieomorphism is homotopic to the identity through the homotopy Hv (t, x) = Ftv (x) : [0, 1] which is proper for every v. Formula (2.22) implies (Fv ) T T = d (Hv ) ([0, 1] T ) . After averaging with a smoothing kernel (v) we get T = dS where =
B(0,)
(Fv ) T (v) dv,
S=
B(0,)
(Hv ) ([0, 1] T ) (v) dv.

p
Then S is a current of the same order s as T and is smooth. Indeed, for u we have , u = T, u where u (x) = Fv u(x) (v) dv ;
B(0,)
()
we can make a change of variable z = Fv (x) v = (x)1 (z x) in the last integral and perform derivatives on to see that each seminorm pt (u ) is controlled by the sup K norm of u. Thus and all its derivatives are currents of order 0, so is smooth. Now we have d = 0 and by the usual Poincar lemma (1.22) applied to we obtain e (2.24) Theorem. Let Rm be a starshaped open subset and T s q () a current of degree q 1 and order s such that dT = 0. There exists a current S s q1 () of degree q 1 and order s such that dS = T on .
3. Holomorphic Functions and Complex Manifolds

3.A. Cauchy Formula in One Variable We start by recalling a few elementary facts in one complex variable theory. Let C be an open set and let z = x + iy be the complex variable, where x, y R. If f is a function of class C 1 on , we have df = with the usual notations (3.1) 1 , = i z 2 x y 1 . = +i z 2 x y f f f f dx + dy = dz + dz x y z z
The function f is holomorphic on if df is C-linear, that is, f /z = 0.
21
(3.2) Cauchy formula. Let K C be a compact set with piecewise C 1 boundary K. Then for every f C 1 (K, C) f (w) = 1 2i
K
f (z) dz zw
1 f d(z), (z w) z
w K
i where d(z) = 2 dz dz = dx dy is the Lebesgue measure on C.
Proof. Assume for simplicity w = 0. As the function z 1/z is locally integrable at z = 0, we get 1 f d(z) = lim 0 z z 1 f i dz dz K D(0,) z z 2 dz 1 f (z) = lim d 0 K D(0,) 2i z dz 1 dz 1 f (z) f (z) lim = 0 2i D(0,) 2i K z z
1 2 2 0
by Stokes formula. The last integral is equal to as tends to 0.
f (ei ) d and converges to f (0)
When f is holomorphic on , we get the usual Cauchy formula (3.3) f (w) = 1 2i f (z) dz, zw w K ,
from which many basic properties of holomorphic functions can be derived: power and Laurent series expansions, Cauchy residue formula, . . . Another interesting consequence is: (3.4) Corollary. The L1 function E(z) = 1/z is a fundamental solution of the loc operator /z on C, i.e. E/z = 0 (Dirac measure at 0). As a consequence, if v is a distribution with compact support in C, then the convolution u = (1/z) v is a solution of the equation u/z = v. Proof. Apply (3.2) with w = 0, f (C) and K Supp f , so that f = 0 on the boundary K and f (0) = 1/z, f /z . (3.5) Remark. It should be observed that this formula cannot be used to solve the equation u/z = v when Supp v is not compact; moreover, if Supp v is compact, a solution u with compact support need not always exist. Indeed, we have a necessary condition v, z n = u, z n /z = 0 for all integers n 0. Conversely, when the necessary condition v, z n = 0 is satised, the canonical solution u = (1/z) v has compact support: this is easily seen by means of the power series expansion (w z)1 = z n wn1 , if we suppose that Supp v is contained in the disk |z| < R and that |w| > R.
22
3.B. Holomorphic Functions of Several Variables Let Cn be an open set. A function f : C is said to be holomorphic if f is continuous and separately holomorphic with respect to each variable, i.e. zj f (. . . , zj , . . .) is holomorphic when z1 , . . . , zj1 , zj+1 , . . . , zn are xed. The set of holomorphic functions on is a ring and will be denoted (). We rst extend the Cauchy formula to the case of polydisks. The open polydisk D(z0 , R) of center (z0,1 , . . . , z0,n ) and (multi)radius R = (R1 , , . . . , Rn ) is dened as the product of the disks of center z0,j and radius Rj > 0 in each factor C : (3.6) D(z0 , R) = D(z0,1 , R1 ) . . . D(z0,n , Rn ) Cn .
The distinguished boundary of D(z0 , R) is by denition the product of the boundary circles (3.7) (z0 , R) = (z0,1 , R1 ) . . . (z0,n , Rn ).
It is important to observe that the distinguished boundary is smaller than the topological boundary D(z0 , R) = j {z D(z0 , R) ; |zj z0,j | = Rj } when n 2. By induction on n, we easily get the (3.8) Cauchy formula on polydisks. If D(z0 , R) is a closed polydisk contained in and f (), then for all w D(z0 , R) we have f (w) = 1 (2i)n
(z0 ,R)
f (z1 , . . . , zn ) dz1 . . . dzn . (z1 w1 ) . . . (zn wn )
The expansion (zj wj )1 = (wj z0,j )j (zj z0,j )j 1 , j N, 1 j n, shows that f can be expanded as a convergent power series f (w) = Nn a (w z0 ) over the polydisk D(z0 , R), with the standard notations z = z1 1 . . . zn n , ! = 1 ! . . . n ! and with (3.9) a = 1 (2i)n f (z1 , . . . , zn ) dz1 . . . dzn f () (z0 ) = . (z1 z0,1 )1 +1 . . . (zn z0,n )n +1 !
(z0 ,R)
As a consequence, f is holomorphic over if and only if f is C-analytic. Arguments similar to the one variable case easily yield the (3.10) Analytic continuation theorem. If is connected and if there exists a point z0 such that f () (z0 ) = 0 for all Nn , then f = 0 on . Another consequence of (3.9) is the Cauchy inequality (3.11) |f () (z0 )| ! sup |f |, R (z0 ,R) D(z0 , R) ,
From this, it follows that every bounded holomorphic function on Cn is constant (Liouvilles theorem), and more generally, every holomorphic function F on Cn such that
23
We endow () with the topology of uniform convergence on compact sets K , that is, the topology induced by C 0 (, C). Then () is closed in C 0 (, C). The Cauchy inequalities (3.11) show that all derivations D are continuous operators on () and that any sequence fj () that is uniformly bounded on all compact sets K is locally equicontinuous. By Ascolis theorem, we obtain (3.12) Montels theorem. Every locally uniformly bounded sequence (fj ) in a convergent subsequence (fj() ).
|F (z)| degree
A(1 + |z|)B with suitable constants A, B B.
0 is in fact a polynomial of total
() has
In other words, bounded subsets of the Frchet space () are relatively compact (a e Frchet space possessing this property is called a Montel space). e 3.C. Dierential Calculus on Complex Analytic Manifolds A complex analytic manifold X of dimension dimC X = n is a dierentiable manifold equipped with a holomorphic atlas ( ) with values in Cn ; this means by denition that the transition maps are holomorphic. The tangent spaces TX,x then have a natural complex vector space structure, given by the coordinate isomorphisms d (x) : TX,x Cn , U x ;
the induced complex structure on TX,x is indeed independent of since the dierentials R d are C-linear isomorphisms. We denote by TX the underlying real tangent space R and by End(TX ) the almost complex structure, i.e. the operator of multiplication J by i = 1. If (z1 , . . . , zn ) are complex analytic coordinates on an open subset X R and zk = xk + iyk , then (x1 , y1 , . . . , xn , yn ) dene real coordinates on , and TX admits (/x1 , /y1 , . . ., /xn , /yn) as a basis ; the almost complex structure is given by J(/xk ) = /yk , J(/yk ) = /xk . The complexied tangent space R R R C TX = C R TX = TX iTX splits into conjugate complex subspaces which are the eigenspaces of the complexied endomorphism Id J associated to the eigenvalues i and i. These subspaces have respective bases (3.13) 1 , = i zk 2 xk yk 1 , = +i z k 2 xk yk 1 k n
and are denoted T 1,0 X (holomorphic vectors or vectors of type (1, 0)) and T 0,1 X (antiholomorphic vectors or vectors of type (0, 1)). The subspaces T 1,0 X and T 0,1 X are canonically isomorphic to the complex tangent space TX (with complex structure J) and its conjugate TX (with conjugate complex structure J), via the C-linear embeddings
1,0 0,1 TX TX C TX , TX TX C TX 1 1 ( + iJ). 2 ( iJ), 2 0,1 1,0 We thus have a canonical decomposition CTX = TX TX TX TX , and by duality a decomposition R HomR (TX ; C) HomC (C TX ; C) TX TX
24
where TX is the space of C-linear forms and TX the space of conjugate C-linear forms. With these notations, (dxk , dyk ) is a basis of HomR (TR X, C), (dzj ) a basis of TX , (dz j ) a basis of TX , and the dierential of a function f C 1 (, C) can be written n
(3.14)
df =
k=1
f f dxk + dyk = xk yk
k=1
f f dzk + dz k . zk z k
The function f is holomorphic on if and only if df is C-linear, i.e. if and only if f satises the Cauchy-Riemann equations f /zk = 0 on , 1 k n. We still denote here by (X) the algebra of holomorphic functions on X. Now, we study the basic rules of complex dierential calculus. The complexied R exterior algebra C R (TX ) = (C TX ) is given by R C k (C TX ) = k TX TX
=
p+q=k
p,q TX ,
2n
where the exterior products are taken over C, and where the components p,q TX are dened by
(3.15)
p,q TX = p TX q TX .
A complex dierential form u on X is said to be of bidegree or type (p, q) if its value at every point lies in the component p,q TX ; we shall denote by C s (, p,q TX ) the space of dierential forms of bidegree (p, q) and class C s on any open subset of X. If is a coordinate open set, such a form can be written u(z) =
|I|=p,|J|=q
uI,J (z) dzI dz J ,
uI,J C s (, C).
This writing is usually much more convenient than the expression in terms of the real basis (dxI dyJ )|I|+|J|=k which is not compatible with the splitting of k TC X in its (p, q) components. Formula (3.14) shows that the exterior derivative d splits into d = d + d , where d : d : (3.16 ) (3.16 ) d u = d u =
I,J 1 k n. (X, p,q TX ) (X, p+1,q TX ), (X, p,q TX ) (X, p,q+1 TX ), uI,J dzk dzI dz J , zk
I,J 1 k n
uI,J dz k dzI dz J . z k
The identity d2 = (d + d )2 = 0 is equivalent to (3.17) d2 = 0, d d + d d = 0, d2 = 0,
since these three operators send (p, q)-forms in (p+2, q), (p+1, q +1) and (p, q +2)-forms, respectively. In particular, the operator d denes for each p = 0, 1, . . . , n a complex, called the Dolbeault complex
(X, p,0 TX ) d (X, p,q TX ) d (X, p,q+1 TX )
25
and corresponding Dolbeault cohomology groups (3.18) H p,q (X, C) = Ker d p,q , Im d p,q1
with the convention that the image of d is zero for q = 0. The cohomology group H p,0 (X, C) consists of (p, 0)-forms u = |I|=p uI (z) dzI such that uI /z k = 0 for all I, k, i.e. such that all coecients uI are holomorphic. Such a form is called a holomorphic p-form on X. Let F : X1 X2 be a holomorphic map between complex manifolds. The pullback F u of a (p, q)-form u of bidegree (p, q) on X2 is again homogeneous of bidegree (p, q), because the components Fk of F in any coordinate chart are holomorphic, hence F dzk = dFk is C-linear. In particular, the equality dF u = F du implies (3.19) d F u = F d u, d F u = F d u.
Note that these commutation relations are no longer true for a non holomorphic change of variable. As in the case of the De Rham cohomology groups, we get a pull-back morphism F : H p,q (X2 , C) H p,q (X1 , C). The rules of complex dierential calculus can be easily extended to currents. We use the following notation. (3.20) Denition. There are decompositions
k
(X, C) =
p+q=k
p,q
(X, C),
k (X, C)
=
p+q=k
p,q (X, C).
The space (X, C) is called the space of currents of bidimension (p, q) and bidegree p,q (n p, n q) on X, and is also denoted np,nq (X, C). 3.D. Newton and Bochner-Martinelli Kernels The Newton kernel is the elementary solution of the usual Laplace operator = 2 /x2 in Rm . We rst recall a construction of the Newton kernel. j Let d = dx1 . . . dxm be the Lebesgue measure on Rm . We denote by B(a, r) the euclidean open ball of center a and radius r in Rm and by S(a, r) = B(a, r) the corresponding sphere. Finally, we set m = Vol B(0, 1) and m1 = mm so that (3.21) Vol B(a, r) = m r m , Area S(a, r) = m1 r m1 .
The second equality follows from the rst by derivation. An explicit computation of 2 the integral Rm e|x| d(x) in polar coordinates shows that m = m/2 /(m/2)! where x! = (x + 1) is the Euler Gamma function. The Newton kernel is then given by: N (x) = 1 log |x| 2 1 N (x) = |x|2m (m 2)m1 if m = 2, if m = 2.
(3.22)
26
The function N (x) is locally integrable on Rm and satises N = 0 . When m = 2, this follows from Cor. 3.4 and the fact that = 4 2 /zz. When m = 2, this can be checked by computing the weak limit
0
lim (|x|2 + 2 )1m/2 = lim m(2 m)2 (|x|2 + 2 )1m/2

0
= m(2 m) Im 0 with Im = Rm (|x|2 + 1)1m/2 d(x). The last equality is easily seen by performing the change of variable y = x in the integral 2 (|x|2 + 2 )1m/2 f (x) d(x) =
Rm Rm
(|y|2 + 1)1m/2 f (y) d(y),
where f is an arbitrary test function. Using polar coordinates, we nd that Im = m1 /m and our formula follows. The Bochner-Martinelli kernel is the (n, n 1)-dierential form on Cn with L1 loc coecients dened by (3.23) kBM (z) = cn
1 j n
(1)j
z j dz1 . . . dzn dz 1 . . . dz j . . . dz n , |z|2n
cn = (1)n(n1)/2
(n 1)! . (2i)n
(3.24) Lemma. d kBM = 0 on Cn . Proof. Since the Lebesgue measure on Cn is d(z) =

1 j n
i i dzj dz j = 2 2
(1)
n(n1) 2 dz1
. . . dzn dz 1 . . . dz n ,
we nd d kBM = = (n 1)! n zj d(z) z j |z|2n 1 2 d(z) 2n2 zj z j |z|
1 j n
1 n(n 1)2n
1 j n
= N (z)d(z) = 0 . We let KBM (z, ) be the pull-back of kBM by the map : Cn Cn Cn , (z, ) z . Then Formula (2.19) implies (3.25) d KBM = 0 = [],
where [] denotes the current of integration on the diagonal Cn Cn .
27
(3.26) Koppelman formula. Let Cn be a bounded open set with piecewise C 1 boundary. Then for every (p, q)-form v of class C 1 on we have v(z) =
p,q KBM (z, ) v()
+ d z
p,q1 KBM (z, ) v() +
p,q KBM (z, ) d v()
p,q on , where KBM (z, ) denotes the component of KBM (z, ) of type (p, q) in z and (n p, n q 1) in .
Proof. Given w
np,nq
(), we consider the integral KBM (z, ) v() w(z).
It is well dened since KBM has no singularities on Supp v . Since w(z) vanishes on the integral can be extended as well to (). As KBM (z, )v()w(z) is of total bidegree (2n, 2n 1), its dierential d vanishes. Hence Stokes formula yields
KBM (z, ) v() w(z) = =
d KBM (z, ) v() w(z)
p,q d KBM (z, ) v() w(z) KBM (z, ) d v() w(z) p,q1 KBM (z, ) v() d w(z).
(1)p+q By (3.25) we have
d KBM (z, ) v() w(z) =
[] v() w(z) =
v(z) w(z)
Denoting , the pairing between currents and test forms on , the above equality is thus equivalent to KBM (z, ) v(), w(z) = v(z) (1)p+q
p,q KBM (z, ) d v(), w(z) p,q1 KBM (z, ) v(), d w(z) ,
which is itself equivalent to the Koppelman formula by integrating d v by parts. (3.27) Corollary. Let v s p,q (Cn ) be a form of class C s with compact support such that d v = 0, q 1. Then the (p, q 1)-form u(z) =
Cn p,q1 KBM (z, ) v()
is a C s solution of the equation d u = v. Moreover, if (p, q) = (0, 1) and n 2 then u has 0,1 compact support, thus the Dolbeault cohomology group with compact support Hc (Cn , C) vanishes for n 2.
28
Proof. Apply the Koppelman formula on a suciently large ball = B(0, R) containing Supp v. Then the formula immediately gives d u = v. Observe that the coecients of KBM (z, ) are O(|z |(2n1) ), hence |u(z)| = O(|z|(2n1) ) at innity. If q = 1, then u is holomorphic on Cn B(0, R). Now, this complement is a union of complex lines when n 2, hence u = 0 on Cn B(0, R) by Liouvilles theorem. (3.28) Hartogs extension theorem. Let be an open set in Cn , n 2, and let K be a compact subset such that K is connected. Then every holomorphic function f ( K) extends into a function f (). Proof. Let () be a cut-o function equal to 1 on a neighborhood of K. Set f0 = (1 )f (), dened as 0 on K. Then v = d f0 = f d can be extended by 0 outside , and can thus be seen as a smooth (0, 1)-form with compact support in Cn , such that d v = 0. By Cor. 3.27, there is a smooth function u with compact support in Cn such that d u = v. Then f = f0 u (). Now u is holomorphic outside Supp , so u vanishes on the unbounded component G of Cn Supp . The boundary G is contained in Supp K, so f = (1 )f u coincides with f on the non empty open set G K. Therefore f = f on the connected open set K. A rened version of the Hartogs extension theorem due to Bochner will be given in Exercise 8.13. It shows that f need only be given as a C 1 function on , satisfying the tangential Cauchy-Riemann equations (a so-called CR-function). Then f extends as a holomorphic function f () C 0 (), provided that is connected. 3.E. The Dolbeault-Grothendieck Lemma We are now in a position to prove the Dolbeault-Grothendieck lemma [Dolbeault 1953], which is the analogue for d of the Poincar lemma. The proof given below makes e use of the Bochner-Martinelli kernel. Many other proofs can be given, e.g. by using a reduction to the one dimensional case in combination with the Cauchy formula (3.2), see Exercise 8.5 or [Hrmander 1966]. o (3.29) Dolbeault-Grothendieck lemma. Let be a neighborhood of 0 in Cn and v s p,q (, C), [resp. v s p,q (, C)], such that d v = 0, where 1 s . a) If q = 0, then v(z) = |I|=p vI (z) dzI is a holomorphic p-form, i.e. a form whose coecients are holomorphic functions. b) If q 1, there exists a neighborhood of 0 and a form u in a current u s p,q1 (, C)] such that d u = v on .
s p,q1
(, C) [resp.
Proof. We assume that is a ball B(0, r) Cn and take for simplicity r > 1 (possibly after a dilation of coordinates). We then set = B(0, 1). Let () be a cut-o function equal to 1 on . The Koppelman formula (3.26) applied to the form v on gives (z)v(z) = d z
p,q1 KBM (z, ) ()v() + p,q KBM (z, ) d () v().
This formula is valid even when v is a current, because we may regularize v as v and take the limit. We introduce on Cn Cn Cn the kernel n (1)j (wj j ) (dzk dk ) (dwk d k ). K(z, w, ) = cn ((z ) (w ))n k j=1 k=j
4. Subharmonic Functions
29
By construction, KBM (z, ) is the result of the substitution w = z in K(z, w, ), i.e. KBM = h K where h(z, ) = (z, z, ). We denote by K p,q the component of K of p,q bidegree (p, 0) in z, (q, 0) in w and (n p, n q 1) in . Then KBM = h K p,q and we nd v = d u0 + g v1 on , where g(z) = (z, z) and u0 (z) =
p,q1 KBM (z, ) ()v(),
v1 (z, w) =
K p,q (z, w, ) d () v().
By denition of K p,q (z, w, ), v1 is holomorphic on the open set U = (z, w) ; , Re(z ) (w ) > 0 , / which contains the conjugate-diagonal points (z, z) as well as the points (z, 0) and (0, w) in . Moreover U clearly has convex slices ({z} Cn ) U and (Cn {w}) U . In particular U is starshaped with respect to w, i.e. (z, w) U = (z, tw) U, t [0, 1].
As u1 is of type (p, 0) in z and (q, 0) in w, we get d (g v1 ) = g dw v1 = 0, hence dw v1 = 0. z p,q1 For q = 0 we have KBM = 0, thus u0 = 0, and v1 does not depend on w, thus v is holomorphic on . For q 1, we can use the homotopy formula (1.23) with respect to w (considering z as a parameter) to get a holomorphic form u1 (z, w) of type (p, 0) in z and (q 1, 0) in w, such that dw u1 (z, w) = v1 (z, w). Then we get d g u1 = g dw u1 = g v1 , hence v = d (u0 + g u1 ) on . Finally, the coecients of u0 are obtained as linear combinations of convolutions of the coecients of v with L1 functions of the form j ||2n . Hence u0 is of class C s (resp. loc is a current of order s), if v is. (3.30) Corollary. The operator d is hypoelliptic in bidegree (p, 0), i.e. if a current f p,0 (X, C) satises d f p,1 (X, C), then f p,0 (X, C). Proof. The result is local, so we may assume that X = is a neighborhood of 0 in Cn . The (p, 1)-form v = d f p,1 (X, C) satises d v = 0, hence there exists u p,0 (, C) such that d u = d f . Then f u is holomorphic and f = (f u) + u p,0 (, C).
A harmonic (resp. subharmonic) function on an open subset of Rm is essentially a function (or distribution) u such that u = 0 (resp. u 0). A fundamental example of subharmonic function is given by the Newton kernel N , which is actually harmonic on Rm {0}. Subharmonic functions are an essential tool of harmonic analysis and potential theory. Before giving their precise denition and properties, we derive a basic integral formula involving the Green kernel of the Laplace operator on the ball.
30
4.A. Construction of the Green Kernel The Green kernel G (x, y) of a smoothly bounded domain Rm is the solution of the following Dirichlet boundary problem for the Laplace operator on : (4.1) Denition. The Green kernel of a smoothly bounded domain Rm is a function G (x, y) : [, 0] with the following properties: a) G (x, y) is
on
Diag
(Diag = diagonal ) ;
b) G (x, y) = G (y, x) ; c) G (x, y) < 0 on and G (x, y) = 0 on ; d) x G (x, y) = y on for every xed y . It can be shown that G always exists and is unique. The uniqueness is an easy consequence of the maximum principle (see Th. 4.14 below). In the case where = B(0, r) is a ball (the only case we are going to deal with), the existence can be shown through explicit calculations. In fact the Green kernel Gr (x, y) of B(0, r) is (4.2) Gr (x, y) = N (x y) N r2 |y| x 2 y r |y| , x, y B(0, r).
A substitution of the explicit value of N (x) yields: |x y|2 1 log 2 if m = 2, otherwise 4 r 2 x, y + r12 |x|2 |y|2 1 1 1m/2 |x y|2m r 2 2 x, y + 2 |x|2 |y|2 . Gr (x, y) = (m 2)m1 r Gr (x, y) = (4.3) Theorem. The above dened function Gr satises all four properties (4.1 ad) on = B(0, r), thus Gr is the Green kernel of B(0, r). Proof. The rst three properties are immediately veried on the formulas, because r 2 2 x, y + 1 2 2 1 |x| |y| = |x y|2 + 2 r 2 |x|2 r 2 |y|2 . 2 r r
For property d), observe that r 2 y/|y|2 B(0, r) whenever y B(0, r) {0}. The second / Newton kernel in the right hand side of (4.1) is thus harmonic in x on B(0, r), and x Gr (x, y) = x N (x y) = y on B(0, r).
4.B. Green-Riesz Representation Formula and Dirichlet Problem 4.B.1. Green-Riesz Formula. For all smooth functions u, v on a smoothly bounded domain Rm , we have (4.4)
(u v v u) d =
v u d v
31
where / is the derivative along the outward normal unit vector of and d the euclidean area measure. Indeed (1)j1 dx1 . . . dxj . . . dxm = j d, for the wedge product of , dx with the left hand side is j d. Therefore v d =
m j=1
v j d = xj
(1)j1
j=1
v dx1 . . . dxj . . . dxm . xj
Formula (4.4) is then an easy consequence of Stokes theorem. Observe that (4.4) is still valid if v is a distribution with singular support relatively compact in . For = B(0, r), u C 2 B(0, r), R and v(y) = Gr (x, y), we get the Green-Riesz representation formula: (4.5) u(x) =
B(0,r)
u(y) Gr (x, y) d(y) +

S(0,r)
u(y) Pr (x, y) d(y)
where Pr (x, y) = Gr (x, y)/(y), (x, y) B(0, r) S(0, r). The function Pr (x, y) is called the Poisson kernel. It is smooth and satises x Pr (x, y) = 0 on B(0, r) by (4.1 d). A simple computation left to the reader yields: (4.6) Pr (x, y) = r 2 |x|2 . m1 r |x y|m 1
Formula (4.5) for u 1 shows that S(0,r) Pr (x, y) d(y) = 1. When x in B(0, r) tends to x0 S(0, r), we see that Pr (x, y) converges uniformly to 0 on every compact subset of S(0, r) {x0 } ; it follows that the measure Pr (x, y) d(y) converges weakly to x0 on S(0, r). 4.B.2. Solution of the Dirichlet Problem. For any bounded measurable function v on S(a, r) we dene (4.7) Pa,r [v](x) =
S(a,r)
v(y) Pr (x a, y a) d(y),
x B(a, r).
If u C 0 B(a, r), R C 2 B(a, r), R is harmonic, i.e. u = 0 on B(a, r), then (4.5) gives u = Pa,r [u] on B(a, r), i.e. the Poisson kernel reproduces harmonic functions. Suppose now that v C 0 S(a, r), R is given. Then Pr (x a, y a) d(y) converges weakly to x0 when x tends to x0 S(a, r), so Pa,r [v](x) converges to v(x0 ). It follows that the function u dened by u = Pa,r [v] on B(a, r), u=v on S(a, r) is continuous on B(a, r) and harmonic on B(a, r) ; thus u is the solution of the Dirichlet problem with boundary values v. 4.C. Denition and Basic Properties of Subharmonic Functions
32
4.C.1. Denition. Mean Value Inequalities. If u is a Borel function on B(a, r) which is bounded above or below, we consider the mean values of u over the ball or sphere: (4.8) (4.8 ) B (u ; a, r) = 1 u(x) d(x), m r m B(a,r) 1 u(x) d(x). S (u ; a, r) = m1 r m1 S(a,r)
As d = dr d these mean values are related by (4.9) B (u ; a, r) = 1 m r m

1 0 r
m1 tm1 S (u ; a, t) dt
0
=m
tm1 S (u ; a, rt) dt.
Now, apply formula (4.5) with x = 0. We get Pr (0, y) = 1/m1 r m1 and Gr (0, y) = r (|y|2m r 2m )/(2 m)m1 = (1/m1 ) |y| t1m dt, thus u(y) Gr (0, y) d(y) = 1 m1
r 0 0 r
dt tm1
|y|<t
u(y) d(y)
B(0,r)
1 = m
B (u ; 0, t) t dt
thanks to the Fubini formula. By translating S(0, r) to S(a, r), (4.5) implies the Gauss formula (4.10) S (u ; a, r) = u(a) + 1 m
r
B (u ; a, t) t dt.
0
Let be an open subset of Rm and u C 2 (, R). If a and u(a) > 0 (resp. u(a) < 0), Formula (4.10) shows that S (u ; a, r) > u(a) (resp. S (u ; a, r) < u(a)) for r small enough. In particular, u is harmonic (i.e. u = 0) if and only if u satises the mean value equality S (u ; a, r) = u(a), B(a, r) . Now, observe that if ( ) is a family of radially symmetric smoothing kernels associated with (x) = (|x|) and if u is a Borel locally bounded function, an easy computation yields u (a) =
B(0,1) 1
u(a + x) (x) d S (u ; a, t) (t) tm1 dt.

0
(4.11)
= m1
Thus, if u is a Borel locally bounded function satisfying the mean value equality on , (4.11) shows that u = u on , in particular u must be smooth. Similarly, if we replace the mean value equality by an inequality, the relevant regularity property to be required for u is just semicontinuity.
33
(4.12) Theorem and denition. Let u : [, +[ be an upper semicontinuous function. The following various forms of mean value inequalities are equivalent: a) u(x) b) u(a) c) u(a) Pa,r [u](x), S (u ; a, r), B (u ; a, r), B(a, r) , B(a, r) ; B(a, r) ; x B(a, r) ;
d) for every a , there exists a sequence (r ) decreasing to 0 such that u(a) B (u ; a, r ) ;
e) for every a , there exists a sequence (r ) decreasing to 0 such that u(a) S (u ; a, r ) .
A function u satisfying one of the above properties is said to be subharmonic on . The set of subharmonic functions will be denoted by Sh(). By (4.10) we see that a function u C 2 (, R) is subharmonic if and only if u 0 : in fact S (u ; a, r) < u(a) for r small if u(a) < 0. It is also clear on the denitions that every (locally) convex function on is subharmonic. Proof. We have obvious implications a) = b) = c) = d) = e), the second and last ones by (4.10) and the fact that B (u ; a, r ) S (u ; a, t) for at least one t ]0, r [. In order to prove e) = a), we rst need a suitable version of the maximum principle. (4.13) Lemma. Let u : [, +[ be an upper semicontinuous function satisfying property 4.12 e). If u attains its supremum at a point x0 , then u is constant on the connected component of x0 in . Proof. We may assume that is connected. Let W = {x ; u(x) < u(x0 )}. W is open by the upper semicontinuity, and distinct from since x0 W . We want to / show that W = . Otherwise W has a non empty connected component W0 , and W0 has a boundary point a . We have a W , thus u(a) = u(x0 ). By assumption 4.12 e), we get u(a) S (u ; a, r ) for some sequence r 0. For r small enough, W0 intersects B(a, r ) and B(a, r ) ; as W0 is connected, we also have S(a, r ) W0 = . Since u u(x0 ) on the sphere S(a, r ) and u < u(x0 ) on its open subset S(a, r ) W0 , we get u(a) S (u ; a, r) < u(x0 ), a contradiction.
34
(4.14) Maximum principle. If u is subharmonic in (in the sense that u satises the weakest property 4.12 e)), then sup u =
lim sup
z{}
u(z),
and supK u = supK u(z) for every compact subset K . Proof. We have of course lim supz{} u(z) sup u. If the inequality is strict, this means that the supremum is achieved on some compact subset L . Thus, by the upper semicontinuity, there is x0 L such that sup u = supL u = u(x0 ). Lemma 4.13 shows that u is constant on the connected component 0 of x0 in , hence sup u = u(x0 ) =
lim sup
0 z0 {}
u(z)
lim sup
z{}
u(z),
contradiction. The statement involving a compact subset K is obtained by applying the rst statement to = K . Proof of (4.12) e) = a). Let u be an upper semicontinuous function satisfying 4.12 e) and B(a, r) an arbitrary closed ball. One can nd a decreasing sequence of continuous functions vk C 0 S(a, r), R such that lim vk = u. Set hk = Pa,r [vk ] C 0 B(a, r), R . As hk is harmonic on B(a, r), the function u hk satises 4.12 e) on B(a, r). Furthermore lim supxS(a,r) u(x) hk (x) u() vk () 0, so u hk 0 on B(a, r) by Th. 4.14. By monotone convergence, we nd u Pa,r [u] on B(a, r) when k tends to +. 4.C.2. Basic Properties. Here is a short list of the most basic properties. (4.15) Theorem. For any decreasing sequence (uk ) of subharmonic functions, the limit u = lim uk is subharmonic. Proof. A decreasing limit of upper semicontinuous functions is again upper semicontinuous, and the mean value inequalities 4.12 remain valid for u by Lebesgues monotone convergence theorem. (4.16) Theorem. Let u1 , . . . , up Sh() and : Rp R be a convex function such that (t1 , . . . , tp ) is non decreasing in each tj . If is extended by continuity into a function [, +[p [, +[, then (u1 , . . . , up ) Sh(). In particular u1 + + up , max{u1 , . . . , up }, log(eu1 + + eup ) Sh(). Proof. Every convex function is continuous, hence (u1 , . . . , up ) is upper semicontinuous. One can write (t) = sup Ai (t)
iI
where Ai (t) = a1 t1 + + ap tp + b is the family of ane functions that dene supporting hyperplanes of the graph of . As (t1 , . . . , tp ) is non-decreasing in each tj , we have aj 0, thus aj uj (x) + b
1 j p
aj uj + b ; x, r
B (u1 , . . . , up ) ; x, r
35
for every ball B(x, r) . If one takes the supremum of this inequality over all the Ai s , it follows that (u1 , . . . , up ) satises the mean value inequality 4.12 c). In the last example, the function (t1 , . . . , tp ) = log(et1 + + etp ) is convex because 2 j k = e tj tk
2 j etj e2
j etj
1 j,k p
and
j etj
2 j etj e by the Cauchy-Schwarz inequality.
(4.17) Theorem. If is connected and u Sh(), then either u or u L1 (). loc Proof. Note that a subharmonic function is always locally bounded above. Let W be the set of points x such that u is integrable in a neighborhood of x. Then W is open by denition and u > almost everywhere on W . If x W , one can choose a W such that |a x| < r = 1 d(x, ) and u(a) > . Then B(a, r) is a neighborhood of 2 x, B(a, r) and B (u ; a, r) u(a) > . Therefore x W , W is also closed. We must have W = or W = ; in the last case u by the mean value inequality. (4.18) Theorem. Let u Sh() be such that u on each connected component of . Then a) r S (u ; a, r), r B (u ; a, r) are non decreasing functions in the interval ]0, d(a, )[ , and B (u ; a, r) S (u ; a, r). b) For any family ( ) of smoothing kernels, u Sh( ) (u ) is non decreasing in and lim0 u = u.
( , R), the family
Proof. We rst verify statements a) and b) when u C 2 (, R). Then u 0 and S (u ; a, r) is non decreasing in virtue of (4.10). By (4.9), we nd that B (u ; a, r) is also non decreasing and that B (u ; a, r) S (u ; a, r). Furthermore, Formula (4.11) shows that u (a) is non decreasing (provided that is radially symmetric). u u r on r , r > 0,
In the general case, we rst observe that property 4.12 c) is equivalent to the inequality
where r is the probability measure of uniform density on B(0, r). This inequality implies u u r on (r ) = r+ , thus u ( , R) is subharmonic on . It follows that u is non decreasing in ; by symmetry, it is also non decreasing in , and so is u = lim0 u . We have u u by (4.19) and lim sup0 u u by the upper semicontinuity. Hence lim0 u = u. Property a) for u follows now from its validity for u and from the monotone convergence theorem. (4.19) Corollary. If u Sh() is such that u on each connected component of , then u computed in the sense of distribution theory is a positive measure. Indeed (u ) 0 as a function, and (u ) converges weakly to u in (). Corollary 4.19 has a converse, but the correct statement is slightly more involved than for the direct property:
36
(4.20) Theorem. If v () is such that v is a positive measure, there exists a unique function u Sh() locally integrable such that v is the distribution associated to u. We must point out that u need not coincide everywhere with v, even when v is a locally integrable upper semicontinuous function: for example, if v is the characteristic function of a compact subset K of measure 0, the subharmonic representant of v is u = 0. 0, thus v Sh( ). Proof. Set v = v ( , R). Then v = (v) Arguments similar to those in the proof of Th. 4.18 show that (v ) is non decreasing in . Then u := lim0 v Sh() by Th. 4.15. Since v converges weakly to v, the monotone convergence theorem shows that v, f = lim v f d =

u f d,
(),
0,
which concludes the existence part. The uniqueness of u is clear from the fact that u must satisfy u = lim u = lim v . The most natural topology on the space Sh() of subharmonic functions is the topology induced by the vector space topology of L1 () (Frchet topology of convergence e loc 1 in L norm on every compact subset of ). (4.21) Proposition. The convex cone Sh() L1 () is closed in L1 (), and it has loc loc the property that every bounded subset is relatively compact. Proof. Let (uj ) be a sequence in Sh() L1 (). If uj u in L1 () then uj u loc loc in the weak topology of distributions, hence u 0 and u can be represented by a subharmonic function thanks to Th. 4.20. Now, suppose that uj L1 (K) is uniformly bounded for every compact subset K of . Let j = uj 0. If () is a test function equal to 1 on a neighborhood of K and such that 0 1 on , we nd j (K)
uj d =
uj d
C uj
L1 (K ) ,
where K = Supp , hence the sequence of measures (j ) is uniformly bounded in mass on every compact subset of . By weak compactness, there is a subsequence (j ) which converges weakly to a positive measure on . We claim that f (j ) converges to f () in L1 (Rm ) for every function f L1 (Rm ). In fact, this is clear if f (Rm ), loc loc and in general we use an approximation of f by a smooth function g together with the estimate (f g) (j )
L1 (A)
(f g)
L1 (A+K ) j (K
),
A Rm
to get the conclusion. We apply this when f = N is the Newton kernel. Then hj = uj N (j ) is harmonic on and bounded in L1 (). As hj = hj for any smoothing kernel , we see that all derivatives D hj = hj (D ) are in fact uniformly locally bounded in . Hence, after extracting a new subsequence, we may suppose that
37
hj converges uniformly to a limit h on . Then uj = hj + N (j ) converges to u = h + N () in L1 (), as desired. loc We conclude this subsection by stating a generalized version of the Green-Riesz formula. (4.22) Proposition. Let u Sh() L1 () and B(0, r) . loc a) The Green-Riesz formula still holds true for such an u, namely, for every x B(0, r) u(x) =
B(0,r)
u(y) Gr (x, y) d(y) +

S(0,r)
u(y) Pr (x, y) d(y).
b) (Harnack inequality) If u 0 on B(0, r), then for all x B(0, r) 0 If u u(x)

S(0,r)
u(y) Pr (x, y) d(y)
r m2 (r + |x|) S (u ; 0, r). (r |x|)m1
0 on B(0, r), then for all x B(0, r) u(x)

S(0,r)
u(y) Pr (x, y) d(y)
r m2 (r |x|) S (u ; 0, r) (r + |x|)m1
0.
Proof. We know that a) holds true if u is of class C 2 . In general, we replace u by u and take the limit. We only have to check that (y) Gr (x, y) d(y) = lim
B(0,r) 0
(y) Gr (x, y) d(y)

B(0,r)
for the positive measure = u. Let us denote by Gx (y) the function such that Gx (y) = Then (y) Gr (x, y) d(y) =
B(0,r) Rm
Gr (x, y) if x B(0, r) 0 if x B(0, r). / (y) Gx (y) d(y) (y) Gx (y) d(y).
Rm
However Gx is continuous on Rm {x} and subharmonic in a neighborhood of x, hence Gx converges uniformly to Gx on every compact subset of Rm {x}, and converges pointwise monotonically in a neighborhood of x. The desired equality follows by the monotone convergence theorem. Finally, b) is a consequence of a), for the integral involving u is nonpositive and 1 r m2 (r |x|) m1 r m1 (r + |x|)m1 Pr (x, y) 1 r m2 (r + |x|) m1 r m1 (r |x|)m1 |x y|m (r + |x|)m .
by (4.6) combined with the obvious inequality (r |x|)m
38
4.C.3. Upper Envelopes and Choquets Lemma. Let Rn and let (u )I be a family of upper semicontinuous functions [, +[. We assume that (u ) is locally uniformly bounded above. Then the upper envelope u = sup u need not be upper semicontinuous, so we consider its upper semicontinuous regularization: u (z) = lim sup u u(z).
0 B(z,)
It is easy to check that u is the smallest upper semicontinuous function which is u. Our goal is to show that u can be computed with a countable subfamily of (u ). Let B(zj , j ) be a countable basis of the topology of . For each j, let (zjk ) be a sequence of points in B(zj , j ) such that sup u(zjk ) =
k
sup u,
B(zj ,j )
and for each pair (j, k), let (j, k, l) be a sequence of indices I such that u(zjk ) = supl u(j,k,l) (zjk ). Set v = sup u(j,k,l) .
j,k,l
Then v
u and v sup v
B(zj ,j )
u . On the other hand sup v(zjk )

k
sup u(j,k,l) (zjk ) = sup u(zjk ) =

k,l k
sup u.
B(zj ,j )
As every ball B(z, ) is a union of balls B(zj , j ), we easily conclude that v v = u . Therefore:
u , hence
(4.23) Choquets lemma. Every family (u ) has a countable subfamily (vj ) = (u(j) ) such that its upper envelope v satises v u u = v . (4.24) Proposition. If all u are subharmonic, the upper regularization u is subharmonic and equal almost everywhere to u. Proof. By Choquets lemma we may assume that (u ) is countable. Then u = sup u is a Borel function. As each u satises the mean value inequality on every ball B(z, r) , we get u(z) = sup u (z) sup B (u ; z, r) B (u ; z, r). The right-hand side is a continuous function of z, so we infer u (z) B (u ; z, r) B (u ; z, r)
and u is subharmonic. By the upper semicontinuity of u and the above inequality we nd u (z) = limr0 B (u ; z, r), thus u = u almost everywhere by Lebesgues lemma.
5. Plurisubharmonic Functions
39
5.A. Denition and Basic Properties Plurisubharmonic functions have been introduced independently by [Lelong 1942] and [Oka 1942] for the study of holomorphic convexity. They are the complex counterparts of subharmonic functions. (5.1) Denition. A function u : [, +[ dened on an open subset Cn is said to be plurisubharmonic if a) u is upper semicontinuous ; b) for every complex line L Cn , uL is subharmonic on L. The set of plurisubharmonic functions on is denoted by Psh(). An equivalent way of stating property b) is: for all a , Cn , || < d(a, ), then (5.2) u(a) 1 2
2
u(a + ei ) d.
0
An integration of (5.2) over S(0, r) yields u(a) (5.3) Psh() Sh().
S (u ; a, r), therefore
The following results have already been proved for subharmonic functions and are easy to extend to the case of plurisubharmonic functions: (5.4) Theorem. For any decreasing sequence of plurisubharmonic functions uk Psh(), the limit u = lim uk is plurisubharmonic on . (5.5) Theorem. Let u Psh() be such that u on every connected component of . If ( ) is a family of smoothing kernels, then u is and plurisubharmonic on , the family (u ) is non decreasing in and lim0 u = u. (5.6) Theorem. Let u1 , . . . , up Psh() and : Rp R be a convex function such that (t1 , . . . , tp ) is non decreasing in each tj . Then (u1 , . . . , up ) is plurisubharmonic on . In particular u1 + +up , max{u1 , . . . , up }, log(eu1 + +eup ) are plurisubharmonic on . (5.7) Theorem. Let {u } Psh() be locally uniformly bounded from above and u = sup u . Then the regularized upper envelope u is plurisubharmonic and is equal to u almost everywhere. Proof. By Choquets lemma, we may assume that (u ) is countable. Then u is a Borel function which clearly satises (5.2), and thus u also satises (5.2). Hence u is plurisubharmonic. By Proposition 4.24, u = u almost everywhere and u is subharmonic, so u = lim u = lim u
40
is plurisubharmonic. If u C 2 (, R), the subharmonicity of restrictions of u to complex lines, C w u(a + w), a , Cn , is equivalent to 2 u(a + w) = ww 2u (a + w) j k zj z k 0.
1 j,k n
Therefore, u is plurisubharmonic on if and only if 2 u/zj z k (a) j k is a semipositive hermitian form at every point a . This equivalence is still true for arbitrary plurisubharmonic functions, under the following form: (5.8) Theorem. If u Psh(), u on every connected component of , then for all Cn 2u Hu() := j k () zj z k
1 j,k n
is a positive measure. Conversely, if v () is such that Hv() is a positive measure for every Cn , there exists a unique function u Psh() locally integrable on such that v is the distribution associated to u. Proof. If u Psh(), then Hu() = weak lim H(u )() 0. Conversely, Hv 0 implies H(v ) = (Hv) 0, thus v Psh(), and also v 0, hence (v ) is non decreasing in and u = lim0 v Psh() by Th. 5.4. (5.9) Proposition. The convex cone Psh() L1 () is closed in L1 (), and it has loc loc the property that every bounded subset is relatively compact. 5.B. Relations with Holomorphic Functions In order to get a better geometric insight, we assume more generally that u is a C 2 function on a complex n-dimensional manifold X. The complex Hessian of u at a point a X is the hermitian form on TX dened by (5.10) Hua =
1 j,k n
2u (a) dzj dz k . zj z k
If F : X Y is a holomorphic mapping and if v C 2 (Y, R), we have d d (v F ) = F d d v. In equivalent notations, a direct calculation gives for all TX,a H(v F )a () = 2 v F (a) Fl a) Fm a) j k = HvF (a) F (a). . zl z m zj zk
j,k,l,m
In particular Hua does not depend on the choice of coordinates (z1 , . . . , zn ) on X, and Hva 0 on Y implies H(v F )a 0 on X. Therefore, the notion of plurisubharmonic function makes sense on any complex manifold. (5.11) Theorem. If F : X Y is a holomorphic map and v Psh(Y ), then v F Psh(X).
41
Proof. It is enough to prove the result when X = 1 Cn and X = 2 Cp are open subsets . The conclusion is already known when v is of class C 2 , and it can be extended to an arbitrary upper semicontinuous function v by using Th. 5.4 and the fact that v = lim v . (5.12) Example. By (3.22) we see that log |z| is subharmonic on C, thus log |f | Psh(X) for every holomorphic function f (X). More generally for every fj (X) and j log |f1 |1 + + |fq |q Psh(X) 0 (apply Th. 5.6 with uj = j log |fj | ).
5.C. Convexity Properties The close analogy of plurisubharmonicity with the concept of convexity strongly suggests that there are deeper connections between these notions. We describe here a few elementary facts illustrating this philosophy. Another interesting connection between plurisubharmonicity and convexity will be seen in 7.B (Kiselmans minimum principle). (5.13) Theorem. If = + i where , are open subsets of Rn , and if u(z) is a plurisubharmonic function on that depends only on x = Re z, then x u(x) is convex.
1 Proof. This is clear when u C 2 (, R), for 2 u/zj z k = 4 2 u/xj xk . In the general case, write u = lim u and observe that u (z) depends only on x.
(5.14) Corollary. If u is a plurisubharmonic function in the open polydisk D(a, R) = D(aj , Rj ) Cn , then
2 1 u(a1 + r1 ei1 , . . . , an + rn ein ) d1 . . . dn , (2)n 0 m(u ; r1 , . . . , rn ) = sup u(z1 , . . . , zn ), rj < R j
(u ; r1 , . . . , rn ) =
zD(a,r)
are convex functions of (log r1 , . . . , log rn ) that are non decreasing in each variable. Proof. That is non decreasing follows from the subharmonicity of u along every coordinate axis. Now, it is easy to verify that the functions
2 1 u(a1 + ez1 ei1 , . . . , an + ezn ein ) d1 . . . dn , n (2) 0 m(z1 , . . . , zn ) = sup u(a1 + ez1 w1 , . . . , an + ezn wn )
(z1 , . . . , zn ) =
|wj | 1
are upper semicontinuous, satisfy the mean value inequality, and depend only on Re zj ]0, log Rj [. Therefore and M are convex. Cor. 5.14 follows from the equalities (u ; r1 , . . . , rn ) = (log r1 , . . . , log rn ), m(u ; r1 , . . . , rn ) = m(log r1 , . . . , log rn ).
42
5.D. Pluriharmonic Functions Pluriharmonic functions are the counterpart of harmonic functions in the case of functions of complex variables: (5.15) Denition. A function u is said to be pluriharmonic if u and u are plurisubharmonic. A pluriharmonic function is harmonic (in particular smooth) in any C-analytic coordinate system, and is characterized by the condition Hu = 0, i.e. d d u = 0 or 2 u/zj z k = 0 for all j, k.
If f (X), it follows that the functions Re f, Im f are pluriharmonic. Conversely:

1 (5.16) Theorem. If the De Rham cohomology group HDR (X, R) is zero, every pluriharmonic function u on X can be written u = Re f where f is a holomorphic function on X. 1 Proof. By hypothesis HDR (X, R) = 0, u (X) and d(d u) = d d u = 0, hence there (X) such that dg = d u. Then dg is of type (1, 0), i.e. g (X) and exists g
d(u 2 Re g) = d(u g g) = (d u dg) + (d u dg) = 0. Therefore u = Re(2g + C), where C is a locally constant function. 5.E. Global Regularization of Plurisubharmonic Functions We now study a very ecient regularization and patching procedure for continuous plurisubharmonic functions, essentially due to [Richberg 1968]. The main idea is contained in the following lemma: (5.17) Lemma. Let u Psh( ) where X is a locally nite open covering of X. Assume that for every index lim sup u () < max {u (z)}
z z
at all points z . Then the function u(z) = max u (z)

z
is plurisubharmonic on X. Proof. Fix z0 X. Then the indices such that z0 or z0 do not contribute / to the maximum in a neighborhood of z0 . Hence there is a a nite set I of indices such that z0 and a neighborhood V I on which u(z) = maxI u (z). Therefore u is plurisubharmonic on V . The above patching procedure produces functions which are in general only continuous. When smooth functions are needed, one has to use a regularized max function. Let
43
and
(R, R) be a nonnegative function with support in [1, 1] such that h(h) dh = 0. R
(h) dh = 1
(5.18) Lemma. For arbitrary = (1 , . . . , p ) ]0, +[p, the function M (t1 , . . . , tp ) =

Rn
max{t1 + h1 , . . . , tp + hp }
(hj /j ) dh1 . . . dhp

1 j n
possesses the following properties: a) M (t1 , . . . , tp ) is non decreasing in all variables, smooth and convex on Rn ; b) max{t1 , . . . , tp } M (t1 , . . . , tp ) max{t1 + 1 , . . . , tp + p } ;
c) M (t1 , . . . , tp ) = M(1 ,...,j ,...,p ) (t1 , . . . , tj , , . . . , tp ) if tj + j maxk=j {tk k } ; d) M (t1 + a, . . . , tp + a) = M (t1 , . . . , tp ) + a, a R ; e) if u1 , . . . , up are plurisubharmonic and satisfy H(uj )z () z () where z z is a continuous hermitian form on TX , then u = M (u1 , . . . , up ) is plurisubharmonic and satises Huz () z (). Proof. The change of variables hj hj tj shows that M is smooth. All properties are immediate consequences of the denition, except perhaps e). That M (u1 , . . . , up ) is plurisubharmonic follows from a) and Th. 5.6. Fix a point z0 and > 0. All functions u (z) = uj (z) z0 (z z0 ) + |z z0 |2 are plurisubharmonic near z0 . It follows that j M (u , . . . , u ) = u z0 (z z0 ) + |z z0 |2 1 p is also plurisubharmonic near z0 . Since > 0 was arbitrary, e) follows. (5.19) Corollary. Let u ( ) Psh( ) where X is a locally nite open covering of X. Assume that u (z) < max{u (z)} at every point z , when runs over the indices such that z. Choose a family ( ) of positive numbers so small that u (z) + max z {u (z) } for all and z . Then the function dened by u(z) = M( ) u (z) is smooth and plurisubharmonic on X. (5.20) Denition. A function u Psh(X) is said to be strictly plurisubharmonic if u L1 (X) and if for every point x0 X there exists a neighborhood of x0 and c > 0 loc such that u(z) c|z|2 is plurisubharmonic on , i.e. ( 2 u/zj z k )j k c||2 (as n distributions on ) for all C . (5.21) Theorem ([Richberg 1968]). Let u Psh(X) be a continuous function which is strictly plurisubharmonic on an open subset X, with Hu for some continuous positive hermitian form on . For any continuous function C 0 (), > 0, there exists a plurisubharmonic function u in C 0 (X) () such that u u u + on for such that z
44
and u = u on X , which is strictly plurisubharmonic on and satises H u (1). In particular, u can be chosen strictly plurisubharmonic on X if u has the same property. Proof. Let ( ) be a locally nite open covering of by relatively compact open balls contained in coordinate patches of X. Choose concentric balls of respec tive radii r < r < r and center z = 0 in the given coordinates z = (z1 , . . . , zn ) near , such that still cover . We set
2 u (z) = u (z) + (r |z|2 )
on . (1 )
For < ,0 and < ,0 small enough, we have u u + /2 and Hu on . Set 2 2 2 2 = min{r r , (r r )/2}.
Choose rst < ,0 such that < min /2, and then < ,0 so small that u u on , we have u u + on , so that the condition required in Corollary 5.19 is satised. We dene u= u M( ) (u ) on X on . ,
By construction, u is smooth on and satises u u u + , Hu (1 ) thanks to 5.18 (b,e). In order to see that u is plurisubharmonic on X, observe that u is the uniform limit of u with u = max u , M(1 ... ) (u1 . . . u ) and u = u on the complement. 5.F. Polar and Pluripolar Sets. Polar and pluripolar sets are sets of poles of subharmonic and plurisubharmonic functions. Although these functions possess a large amount of exibility, pluripolar sets have some properties which remind their loose relationship with holomorphic functions. (5.22) Denition. A set A Rm (resp. A X, dimC X = n) is said to be polar (resp. pluripolar) if for every point x there exist a connected neighborhood W of x and u Sh(W ) (resp. u Psh(W )), u , such that A W {x W ; u(x) = }. Theorem 4.17 implies that a polar or pluripolar set is of zero Lebesgue measure. Now, we prove a simple extension theorem. (5.23) Theorem. Let A be a closed polar set and v Sh( A) such that v is bounded above in a neighborhood of every point of A. Then v has a unique extension v Sh(). Proof. The uniqueness is clear because A has zero Lebesgue measure. On the other hand, every point of A has a neighborhood W such that A W {x W ; u(x) = }, u Sh(W ), u . on
1
6. Domains of Holomorphy and Stein Manifolds
45
After shrinking W and subtracting a constant to u, we may assume u 0. Then for every > 0 the function v = v + u Sh(W A) can be extended as an upper semicontinuous on W by setting v = on A W . Moreover, v satises the mean value inequality v (a) S (v ; a, r) if a W A, r < d(a, A W ), and also clearly if a A, r < d(a, W ). Therefore v Sh(W ) and v = (sup v ) Sh(W ). Clearly v coincides with v on W A. A similar proof gives: (5.24) Theorem. Let A be a closed pluripolar set in a complex analytic manifold X. Then every function v Psh(X A) that is locally bounded above near A extends uniquely into a function v Psh(X). (5.25) Corollary. Let A X be a closed pluripolar set. Every holomorphic function f (X A) that is locally bounded near A extends to a holomorphic function f (X). Proof. Apply Th. 5.24 to Re f and Im f . It follows that Re f and Im f have pluriharmonic extensions to X, in particular f extends to f (X). By density of X A, d f = 0 on X. (5.26) Corollary. Let A (resp. A X) be a closed (pluri)polar set. If (resp. X) is connected, then A (resp. X A) is connected. Proof. If A (resp. X A) is a disjoint union 1 2 of non empty open subsets, the function dened by f 0 on 1 , f 1 on 2 would have a harmonic (resp. holomorphic) extension through A, a contradiction.

6.A. Domains of Holomorphy in Cn . Examples Loosely speaking, a domain of holomorphy is an open subset in Cn such that there is no part of across which all functions f () can be extended. More precisely: (6.1) Denition. Let Cn be an open subset. is said to be a domain of holomorphy if for every connected open set U Cn which meets and every connected component V of U there exists f () such that fV has no holomorphic extension to U . Under the hypotheses made on U , we have = V U . In order to show that is a domain of holomorphy, it is thus sucient to nd for every z0 a function f () which is unbounded near z0 . (6.2) Examples. Every open subset C is a domain of holomorphy (for any z0 , f (z) = (zz0 )1 cannot be extended at z0 ). In Cn , every convex open subset is a domain of holomorphy: if Re z z0 , 0 = 0 is a supporting hyperplane of at z0 , the function f (z) = ( z z0 , 0 )1 is holomorphic on but cannot be extended at z0 . (6.3) Hartogs gure. Assume that n 2. Let Cn1 be a connected open set and an open subset. Consider the open sets in Cn : = (D(R) D(r)) D(R) = D(R) (Hartogs gure), (lled Hartogs gure).
46
where 0
r < R and D(r) C denotes the open disk of center 0 and radius r in C. Cn1 z (1 , z ) 0 R z1 r C z
Fig. I-3 Hartogs gure Then every function f () can be extended to = D(R) by means of the Cauchy formula: f (z1 , z ) = 1 2i f (1 , z ) d1 , 1 z1 z , max{|z1 |, r} < < R.
|1 |=
In fact f (D(R) ) and f = f on D(R) , so we must have f = f on since is connected. It follows that is not a domain of holomorphy. Let us quote two interesting consequences of this example. (6.4) Corollary (Riemanns extension theorem). Let X be a complex analytic manifold, and S a closed submanifold of codimension 2. Then every f (X S) extends holomorphically to X. Proof. This is a local result. We may choose coordinates (z1 , . . . , zn ) and a polydisk D(R)n in the corresponding chart such that S D(R)n is given by equations z1 = . . . = zp = 0, p = codim S 2. Then, denoting = D(R)n1 and = {z2 = . . . = zp = 0}, the complement D(R)n S can be written as the Hartogs gure D(R)n S = (D(R) {0}) D(R) .
It follows that f can be extended to = D(R)n . 6.B. Holomorphic Convexity and Pseudoconvexity Let X be a complex manifold. We rst introduce the notion of holomorphic hull of a compact set K X. This can be seen somehow as the complex analogue of the notion of (ane) convex hull for a compact set in a real vector space. It is shown that domains of holomorphy in Cn are characterized a property of holomorphic convexity. Finally, we prove that holomorphic convexity implies pseudoconvexity a complex analogue of the geometric notion of convexity.
47
(6.5) Denition. Let X be a complex manifold and let K be a compact subset of X. Then the holomorphic hull of K in X is dened to be K = K(X) = z X ; |f (z)| (6.6) Elementary properties. a) K is a closed subset of X containing K. Moreover we have sup |f | = sup |f |,
K K
sup |f |, f (X) .
K
f (X),
hence K = K. b) If h : X Y is a holomorphic map and K X is a compact set, then h(K(X) ) h(K)(Y ) . In particular, if X Y , then K(X) K(Y ) X. This is immediate from the denition.
c) K contains the union of K with all relatively compact connected components of X K (thus K lls the holes of K). In fact, for every connected component U of X K we have U K, hence if U is compact the maximum principle yields sup |f | = sup |f |
U U
sup |f |,
K
for all f (X).
d) More generally, suppose that there is a holomorphic map h : U X dened on a relatively compact open set U in a complex manifold S, such that h extends as a continuous map h : U X and h(U ) K. Then h(U ) K. Indeed, for f (X), the maximum principle again yields sup |f h| = sup |f h|
U U
sup |f |.
K
This is especially useful when U is the unit disk in C. e) Suppose that X = Cn is an open set. By taking f (z) = exp(A(z)) where A is an arbitrary ane function, we see that K() is contained in the intersection of all ane half-spaces containing K. Hence K() is contained in the ane convex hull Ka . As a consequence K() is always bounded and K(Cn ) is a compact set. However, when is arbitrary, K() is not always compact; for example, in case = Cn {0}, n 2, then () = (Cn ) and the holomorphic hull of K = S(0, 1) is the non compact set K = B(0, 1) {0}. (6.7) Denition. A complex manifold X is said to be holomorphically convex if the holomorphic hull K(X) of every compact set K X is compact. (6.8) Remark. A complex manifold X is holomorphically convex if and only if there is an exhausting sequence of holomorphically compact subsets K X, i.e. compact sets such that X= K , K = K , K K1 .
48
Indeed, if X is holomorphically convex, we may dene K inductively by K0 = and K+1 = (K L )(X) , where K is a neighborhood of K and L a sequence of compact sets of X such that X = L . The converse is obvious: if such a sequence (K ) exists, then every compact subset K X is contained in some K , hence K K = K is compact. We now concentrate on domains of holomorphy in Cn . We denote by d and B(z, r) the distance and the open balls associated to an arbitrary norm on Cn , and we set for simplicity B = B(0, 1). (6.9) Proposition. If is a domain of holomorphy and K is a compact subset, then d(K, ) = d(K, ) and K is compact. Proof. Let f (). Given r < d(K, ), we denote by M the supremum of |f | on the compact subset K + rB . Then for every z K and B, the function
+
(6.10)
C t f (z + t) =
k=0
1 k D f (z)()k tk k!
is analytic in the disk |t| < r and bounded by M . The Cauchy inequalities imply |Dk f (z)()k | M k! r k , z K, B.
As the left hand side is an analytic fuction of z in , the inequality must also hold for z K, B. Every f () can thus be extended to any ball B(z, r), z K, by means of the power series (6.10). Hence B(z, r) must be contained in , and this shows that d(K, ) r. As r < d(K, ) was arbitrary, we get d(K, ) d(K, ) and the converse inequality is clear, so d(K, ) = d(K, ). As K is bounded and closed in , this shows that K is compact. (6.11) Theorem. Let be an open subset of Cn . The following properties are equivalent: a) is a domain of holomorphy; b) is holomorphically convex; c) For every countable subset {zj }jN without accumulation points in and every sequence of complex numbers (aj ), there exists an interpolation function F () such that F (zj ) = aj . d) There exists a function F point of .
() which is unbounded on any neighborhood of any
Proof. d) = a) is obvious and a) = b) is a consequence of Prop. 6.9. c) = d). If = Cn there is nothing to prove. Otherwise, select a dense sequence (j ) in and take zj such that d(zj , j ) < 2j . Then the interpolation function F () such that F (zj ) = j satises d).
49
b) = c). Let K be an exhausting sequence of holomorphically convex compact sets as in Remark 6.8. Let (j) be the unique index such that zj K(j)+1 K(j) . By the denition of a holomorphic hull, we can nd a function gj () such that
K(j)
sup |gj | < |gj (zj )|.
After multiplying gj by a constant, we may assume that gj (zj ) = 1. Let Pj C[z1 , . . . , zn ] be a polynomial equal to 1 at zj and to 0 at z0 , z1 , . . . , zj1 . We set
+
F =
j=0
j Pj gj j ,
where j C and mj N are chosen inductively such that j = aj

m |j Pj gj j |
k Pk (zj )gk (zj )mk ,

0 k<j
2j
on K(j) ;
once j has been chosen, the second condition holds as soon as mj is large enough. Since {zj } has no accumulation point in , the sequence (j) tends to +, hence the series converges uniformly on compact sets. We now show that a holomorphically convex manifold must satisfy some more geometric convexity condition, known as pseudoconvexity, which is most easily described in terms of the existence of plurisubharmonic exhaustion functions. (6.12) Denition. A function : X [, +[ on a topological space X is said to be an exhaustion if all sublevel sets Xc := {z X ; (z) < c}, c R, are relatively compact. Equivalently, is an exhaustion if and only if tends to + relatively to the lter of complements X K of compact subsets of X. A function on an open set Rn is thus an exhaustion if and only if (x) + as x or x . It is easy to check, cf. Exercise 8.8, that a connected open set Rn is convex if and only if has a locally convex exhaustion function. Since plurisubharmonic functions appear as the natural generalization of convex functions in complex analysis, we are led to the following denition. (6.13) Denition. Let X be a complex n-dimensional manifold. Then X is said to be a) weakly pseudoconvex if there exists a smooth plurisubharmonic exhaustion function Psh(X) (X) ; b) strongly pseudoconvex if there exists a smooth strictly plurisubharmonic exhaustion function Psh(X) (X), i.e. H is positive denite at every point. (6.14) Theorem. Every holomorphically convex manifold X is weakly pseudoconvex. Proof. Let (K ) be an exhausting sequence of holomorphically convex compact sets as in Remark 6.8. For every point a L := K+2 K+1 , one can select g,a () such
50
that supK |g,a | < 1 and |g,a (a)| > 1. Then |g,a (z)| > 1 in a neighborhood of a ; by the Borel-Lebesgue lemma, one can nd nitely many functions (g,a )aI such that max |g,a (z)| > 1 for z L ,
aI
max |g,a (z)| < 1 for z K .

aI
For a suciently large exponent p() we get |g,a |2p() on L ,

aI
aI
|g,a |2p()
2 on K .
It follows that the series (z) =

N aI
|g,a (z)|2p()
converges uniformly to a real analytic function Psh(X) (see Exercise 8.11). By construction (z) for z L , hence is an exhaustion. (6.15) Example. The converse to Theorem 6.14 does not hold. In fact let X = C2 / be the quotient of C2 by the free abelian group of rank 2 generated by the ane automorphisms g1 (z, w) = (z + 1, ei1 w), g2 (z, w) = (z + i, ei2 w), 1 , 2 R.
Since acts properly discontinuously on C2 , the quotient has a structure of a complex (non compact) 2-dimensional manifold. The function w |w|2 is -invariant, hence it induces a function ((z, w) ) = |w|2 on X which is in fact a plurisubharmonic exhaustion function. Therefore X is weakly pseudoconvex. On the other hand, any holomorphic function f (X) corresponds to a -invariant holomorphic function f (z, w) on C2 . Then z f (z, w) is bounded for w xed, because f (z, w) lies in the image of the compact set K D(0, |w|), K = unit square in C. By Liouvilles theorem, f (z, w) does not depend on z. Hence functions f (X) are in one-to-one correspondence with holomorphic functions f (w) on C such that f (eij w) = f (w). By looking at the Taylor expansion at the origin, we conclude that f must be a constant if 1 Q or 1 Q (if / / 1 , 2 Q and m is the least common denominator of 1 , 2 , then f is a power series of the form k wmk ). From this, it follows easily that X is holomorphically convex if and only if 1 , 2 Q. 6.C. Stein Manifolds The class of holomorphically convex manifolds contains two types of manifolds of a rather dierent nature: domains of holomorphy X = Cn ; In the rst case we have a lot of holomorphic functions, in fact the functions in () separate any pair of points of . On the other hand, if X is compact and connected, the sets Psh(X) and (X) consist of constant functions merely (by the maximum principle). It is therefore desirable to introduce a clear distinction between these two subclasses. For compact complex manifolds.
51
this purpose, [Stein 1951] introduced the class of manifolds which are now called Stein manifolds. (6.16) Denition. A complex manifold X is said to be a Stein manifold if a) X is holomorphically convex ; b)
(X) locally separates points in X, i.e. every point x X has a neighborhood V that for any y V {x} there exists f (X) with f (y) = f (x).
such
The second condition is automatic if X = is an open subset of Cn . Hence an open set Cn is Stein if and only if is a domain of holomorphy. (6.17) Lemma. If a complex manifold X satises the axiom (6.16 b) of local separation, there exists a smooth nonnegative strictly plurisubharmonic function u Psh(X). Proof. Fix x0 X. We rst show that there exists a smooth nonnegative function u0 Psh(X) which is strictly plurisubharmonic on a neighborhood of x0 . Let (z1 , . . . , zn ) be local analytic coordinates centered at x0 , and if necessary, replace zj by zj so that the closed unit ball B = { |zj |2 1} is contained in the neighborhood V x0 on which (6.16 b) holds. Then, for every point y B, there exists a holomorphic function f (X) such that f (y) = f (x0). Replacing f with (f f (x0)), we can achieve f (x0) = 0 and |f (y)| > 1. By compactness of B, we nd nitely many functions f1 , . . . , fN (X) such that v0 = |fj |2 satises v0 (x0 ) = 0, while v0 1 on B. Now, we set u0 (z) = v0 (z) on X 2 M {v0 (z), (|z| + 1)/3} on B. B,
where M are the regularized max functions dened in 5.18. Then u0 is smooth and plurisubharmonic, coincides with v0 near B and with (|z|2 + 1)/3 on a neighborhood of x0 . We can cover X by countably many neighborhoods (Vj )j 1 , for which we have a smooth plurisubharmonic functions uj Psh(X) such that uj is strictly plurisubharmonic on Vj . Then select a sequence j > 0 converging to 0 so fast that u = j uj (X). The function u is nonnegative and strictly plurisubharmonic everywhere on X. (6.18) Theorem. Every Stein manifold is strongly pseudoconvex. Proof. By Th. 6.14, there is a smooth exhaustion function Psh(X). If u 0 is strictly plurisubharmonic, then = + u is a strictly plurisubharmonic exhaustion. The converse problem to know whether every strongly pseudoconvex manifold is actually a Stein manifold is known as the Levi problem, and was raised by [Levi 1910] in the case of domains Cn . In that case, the problem has been solved in the armative independently by [Oka 1953], [Norguet 1954] and [Bremermann 1954]. The general solution of the Levi problem has been obtained by [Grauert 1958]. Our proof will rely on the theory of L2 estimates for d , which will be available only in Chapter VIII.
52
|z2 | X
C2
z1 C Fig. I-4 Hartogs gure with excrescence (6.19) Remark. It will be shown later that Stein manifolds always have enough holomorphic functions to separate nitely many points, and one can even interpolate given values of a function and its derivatives of some xed order at any discrete set of points. In particular, we might have replaced condition (6.16 b) by the stronger requirement that (X) separates any pair of points. On the other hand, there are examples of manifolds satisfying the local separation condition (6.16 b), but not global separation. A simple example is obtained by attaching an excrescence inside a Hartogs gure, in such a way that the resulting map : X D = D(0, 1)2 is not one-to-one (see Figure I-4 above); then (X) coincides with (D). 6.D. Heredity Properties Holomorphic convexity and pseudoconvexity are preserved under quite a number of natural constructions. The main heredity properties can be summarized in the following Proposition. (6.20) Proposition. Let denote the class of holomorphically convex (resp. of Stein, or weakly pseudoconvex, strongly pseudoconvex manifolds). a) If X, Y b) If X , then X Y . .
and S is a closed complex submanifold of X, then S
c) If (Sj )1 j N is a collection of (not necessarily closed) submanifolds of a complex manifold X such that S = Sj is a submanifold of X, and if Sj for all j, then S . d) If F : X Y is a holomorphic map and S X, S Y are (not necessarily closed) submanifolds in the class , then S F 1 (S ) is in , as long as it is a submanifold of X. e) If X is a weakly (resp. strongly) pseudoconvex manifold and u is a smooth plurisubharmonic function on X, then the open set = u1 (], c[ is weakly (resp. strongly) pseudoconvex. In particular the sublevel sets Xc = 1 (] , c[)
7. Pseudoconvex Open Sets in Cn
53
of a (strictly) plurisubharmonic exhaustion function are weakly (resp. strongly) pseudoconvex. Proof. All properties are more or less immediate to check, so we only give the main facts. a) For K X, L Y compact, we have (K L)(XY ) = K(X) K(Y ) , and if , are plurisubharmonic exhaustions of X, Y , then (x) + (y) is a plurisubharmonic exhaustion of X Y . b) For a compact set K S, we have K(S) K(X) S, and if Psh(X) is an exhaustion, then S Psh(S) is an exhaustion (since S is closed). c) Sj is a closed submanifold in X N ). Sj (equal to its intersection with the diagonal of
d) For a compact set K S F 1 (S ), we have K(SF 1 (S )) K(S) F 1 (F (K)(S ) ), and if , are plurisubharmonic exhaustions of S, S , then +F is a plurisubharmonic exhaustion of S F 1 (S ). e) (z) := (z) + 1/(c u(z)) is a (strictly) plurisubharmonic exhaustion function on .

7.A. Geometric Characterizations of Pseudoconvex Open Sets We rst discuss some characterizations of pseudoconvex open sets in Cn . We will need the following elementary criterion for plurisubharmonicity. (7.1) Criterion. Let v : [, +[ be an upper semicontinuous function. Then v is plurisubharmonic if and only if for every closed disk = z0 + D(1) and every polynomial P C[t] such that v(z0 + t) Re P (t) for |t| = 1, then v(z0 ) Re P (0). Proof. The condition is necessary because t v(z0 + t) Re P (t) is subharmonic in a neighborhood of D(1), so it satises the maximum principle on D(1) by Th. 4.14. Let us prove now the suciency. The upper semicontinuity of v implies v = lim+ v on where (v ) is a strictly decreasing sequence of continuous functions on . As trigonometric polynomials are dense in C 0 (S 1 , R), we may assume v (z0 + ei ) = Re P (ei ), P C[t]. Then v(z0 + t) Re P (t) for |t| = 1, and the hypothesis implies v(z0 ) 1 Re P (0) = 2
2 0
1 Re P (e ) d = 2
i
v (z0 + ei ) d.
0
Taking the limit when tends to + shows that v satises the mean value inequality (5.2). For any z and Cn , we denote by (z, ) = sup r > 0 ; z + D(r)
54
the distance from z to in the complex direction . (7.2) Theorem. Let Cn be an open subset. The following properties are equivalent: a) is strongly pseudoconvex (according to Def. 6.13 b); b) is weakly pseudoconvex ; c) has a plurisubharmonic exhaustion function . d) log (z, ) is plurisubharmonic on Cn ; e) log d(z, ) is plurisubharmonic on . If one of these properties hold, is said to be a pseudoconvex open set. Proof. The implications a) = b) = c) are obvious. For the implication c) = d), we use Criterion 7.1. Consider a disk = (z0 , 0 ) + D(1) (, ) in Cn and a polynomial P C[t] such that log (z0 + t, 0 + t) Re P (t) for |t| = 1.
We have to verify that the inequality also holds when |t| < 1. Consider the holomorphic mapping h : C2 Cn dened by h(t, w) = z0 + t + weP (t) (0 + t). By hypothesis h D(1) {0} = pr1 () ,
h D(1) D(1) (since |eP |
on ),
and the desired conclusion is that h D(1) D(1) . Let J be the set of radii r 0 such that h D(1) D(r) . Then J is an open interval [0, R[, R > 0. If R < 1, we get a contradiction as follows. Let Psh() be an exhaustion function and K = h D(1) D(R) , c = sup .
K
As h is plurisubharmonic on a neighborhood of D(1) D(R), the maximum principle applied with respect to t implies h(t, w) c on D(1) D(R),
hence h D(1) D(R) c and h D(1) D(R + ) for some > 0, a contradiction. d) = e). The function log d(z, ) is continuous on and satises the mean value inequality because log d(z, ) = sup log (z, ) .
B
55
e) = a). It is clear that u(z) = |z|2 + max{log d(z, )1 , 0} is a continuous strictly plurisubharmonic exhaustion function. Richbergs theorem 5.21 implies that there exists () strictly plurisubharmonic such that u u + 1. Then is the required exhaustion function. (7.3) Proposition. a) Let Cn and Cp be pseudoconvex. Then is pseudoconvex. For every holomorphic map F : Cp the inverse image F 1 ( ) is pseudoconvex. b) If ( )I is a family of pseudoconvex open subsets of Cn , the interior of the inter section = is pseudoconvex. I c) If (j )jN is a non decreasing sequence of pseudoconvex open subsets of Cn , then = jN j is pseudoconvex. Proof. a) Let , be smooth plurisubharmonic exhaustions of , . Then (z, w) (z) + (w) is an exhaustion of and z (z) + (F (z)) is an exhaustion of F 1 ( ). b) We have log d(z, ) = supI log d(z, ), so this function is plurisubharmonic. c) The limit log d(z, ) = lim j+ log d(z, j ) is plurisubharmonic, hence is pseudoconvex. This result cannot be generalized to strongly pseudoconvex manifolds: J.E. Fornaess in [Fornaess 1977] has constructed an increasing sequence of 2-dimensional Stein (even ane algebraic) manifolds X whose union is not Stein; see Exercise 8.16. (7.4) Examples. a) An analytic polyhedron in Cn is an open subset of the form P = {z Cn ; |fj (z)| < 1, 1 j N}
where (fj )1 j N is a family of analytic functions on Cn . By 7.3 a), every analytic polyhedron is pseudoconvex. b) Let Cn1 be pseudoconvex and let u : [, +[ be an upper semicontinuous function. Then the Hartogs domain = (z1 , z ) C ; log |z1 | + u(z ) < 0 is pseudoconvex if and and only if u is plurisubharmonic. To see that the plurisubharmonicity of u is necessary, observe that u(z ) = log (0, z ), (1, 0) . Conversely, assume that u is plurisubharmonic and continuous. If is a plurisubharmonic exhaustion of , then 1 (z ) + log |z1 | + u(z )
56
is an exhaustion of . This is no longer true if u is not continuous, but in this case we may apply Property 7.3 c) to conclude that = (z1 , z ) ; d(z , ) > , log |z1 | + u (z ) < 0 , are pseudoconvex. c) An open set Cn is called a tube of base if = + iRn for some open subset Rn . Then of course log d(z, ) = log(x, ) depends only on the real part x = Re z. By Th. 5.13, this function is plurisubharmonic if and only if it is locally convex in x. Therefore if pseudoconvex if and only if every connected component of is convex. d) An open set Cn is called a Reinhardt domain if (ei1 z1 , . . . , ein zn ) is in for every z = (z1 , . . . , zn ) and 1 , . . . , n Rn . For such a domain, we consider the logarithmic indicatrix = R n with = { Cn ; (e1 , . . . , en ) }. =
It is clear that is a tube of base . Therefore every connected component of must be convex if is pseudoconvex. The converse is not true: = Cn {0} is not pseudoconvex for n 2 although = Rn is convex. However, the Reinhardt open set = (z1 , . . . , zn ) (C {0})n ; (log |z1 |, . . . , log |zn |)
is easily seen to be pseudoconvex if is convex: if is a convex exhaustion of , then (z) = (log |z1 |, . . . , log |zn |) is a plurisubharmonic exhaustion of . Similarly, if is convex and such that x = y for yj xj , we can take increasing in all variables and tending to + on , hence the set = (z1 , . . . , zn ) Cn ; |zj | exj for some x
is a pseudoconvex Reinhardt open set containing 0. 7.B. Kiselmans Minimum Principle We already know that a maximum of plurisubharmonic functions is plurisubharmonic. However, if v is a plurisubharmonic function on X Cn , the partial minimum function on X dened by u() = inf z v(, z) need not be plurisubharmonic. A simple counterexample in C C is given by v(, z) = |z|2 + 2 Re(z) = |z + |2 ||2 , u() = ||2 .
It follows that the image F () of a pseudoconvex open set by a holomorphic map F need not be pseudoconvex. In fact, if = {(t, , z) C3 ; log |t| + v(, z) < 0} and if C2 is the image of by the projection map (t, , z) (t, ), then = {(t, ) C2 ; log |t| + u() < 0} is not pseudoconvex. However, the minimum property holds true when v(, z) depends only on Re z :
57
is a convex tube + iRn , Rn . For every plurisubharmonic function v(, z) on that does not depend on Im z, the function u() = inf v(, z)
z
(7.5) Theorem ([Kiselman 1978]). Let Cp Cn be a pseudoconvex open set such that each slice = {z Cn ; (, z) }, Cp ,
is plurisubharmonic or locally on = prCn (). Proof. The hypothesis implies that v(, z) is convex in x = Re z. In addition, we rst assume that v is smooth, plurisubharmonic in (, z), strictly convex in x and limx{} v(, x) = + for every . Then x v(, x) has a unique minimum point x = g(), solution of the equations v/xj (x, ) = 0. As the matrix ( 2 v/xj xk ) is positive denite, the implicit function theorem shows that g is smooth. Now, if C w 0 + wa, a Cn , |w| 1 is a complex disk contained in , there exists a holomorphic function f on the unit disk, smooth up to the boundary, whose real part solves the Dirichlet problem Re f (ei ) = g(0 + ei a). Since v(0 + wa, f (w)) is subharmonic in w, we get the mean value inequality v(0 , f (0)) 1 2
2
v 0 + ei a, f (ei ) d =
0
1 2
v(, g())d.
The last equality holds because Re f = g on and v(, z) = v(, Re z) by hypothesis. As u(0 ) v(0 , f (0)) and u() = v(, g()), we see that u satises the mean value inequality, thus u is plurisubharmonic. Now, this result can be extended to arbitrary functions v as follows: let (, z) 0 be a continuous plurisubharmonic function on which is independent of Im z and is an exhaustion of (Cp Rn ), e.g. (, z) = max{||2 + | Re z|2 , log (, z)}. There is slowly increasing sequence Cj + such that each function j = (Cj 1/j )1 is an exhaustion of a pseudoconvex open set j whose slices are convex tubes and such that d(j , ) > 2/j. Then 1 vj (, z) = v 1/j (, z) + | Re z|2 + j (, z) j is a decreasing sequence of plurisubharmonic functions on j satisfying our previous conditions. As v = lim vj , we see that u = lim uj is plurisubharmonic. (7.6) Corollary. Let Cp Cn be a pseudoconvex open set such that all slices , Cp , are convex tubes in Cn . Then the projection of on Cp is pseudoconvex. Proof. Take v Psh() equal to the function dened in the proof of Th. 7.5. Then u is a plurisubharmonic exhaustion of .
58
7.C. Levi Form of the Boundary For an arbitrary domain in Cn , we rst show that pseudoconvexity is a local property of the boundary. (7.7) Theorem. Let Cn be an open subset such that every point z0 has a neighborhood V such that V is pseudoconvex. Then is pseudoconvex. Proof. As d(z, ) coincides with d z, ( V ) in a neighborhood of z0 , we see that there exists a neighborhood U of such that log d(z, ) is plurisubharmonic on U . Choose a convex increasing function such that (r) > Then the function (z) = max (|z|), log d(z, ) coincides with (|z|) in a neighborhood of an exhaustion. U . Therefore Psh(), and is clearly sup
( U)B(0,r)
log d(z, ),
0.
Now, we give a geometric characterization of the pseudoconvexity property when is of class C 2 . Let C 2 () be a dening function of , i.e. a function such that (7.9) < 0 on , = 0 and d = 0 on .
The holomorphic tangent space to is by denition the largest complex subspace which is contained in the tangent space T to the boundary: (7.9)
h
T = T JT .
It is easy to see that h T,z is the complex hyperplane of vectors Cn such that d (z) = j = 0. zj
1 j n
The Levi form on h T is dened at every point z by (7.10) L,z () = 1 |(z)| 2 j k , zj z k h T,z .
j,k
The Levi form does not depend on the particular choice of , as can be seen from the following intrinsic computation of L (we still denote by L the associated sesquilinear form). (7.11) Lemma. Let , be C 1 vector elds on with values in h T . Then [, ], J = 4 Im L (, ) where is the outward normal unit vector to , [ , ] the Lie bracket of vector elds and , the hermitian inner product.
59
Proof. Extend rst , as vector elds in a neighborhood of and set = j 1 = ( iJ), zj 2 = k 1 = ( + iJ). z k 2
As , J, , J are tangent to , we get on : 0 = .( .) + .( .) =

1 j,k n
j 2 j k + j k + k . zj z k zj z k z k zj
Since [, ] is also tangent to , we have Re [, ], = 0, hence J[, ], is real and [, ], J = J[, ], = 1 2 J[, ]. = Re J[ , ]. || ||
because J[ , ] = i[ , ] and its conjugate J[ , ] are tangent to . We nd now J[ , ] = i Re J[ , ]. = Im [, ], J = j j k j + k , zj z k z k z j k j + k = 2 Im zj z k z k zj 2 j k = 4 Im L (, ). zj z k 2 j k , zj z k
4 Im ||
(7.12) Theorem. An open subset Cn with C 2 boundary is pseudoconvex if and only if the Levi form L is semipositive at every point of . Proof. Set (z) = d(z, ), z . Then = is C 2 near and satises (7.9). If is pseudoconvex, the plurisubharmonicity of log() means that for all z near and all Cn one has
1 j,k n
1 1 2 + 2 j k || zj z k zj z k
0.
Hence ( 2 /zj z k )j k 0 if this is also true at the limit on .
(/zj )j = 0, and an easy argument shows that
Conversely, if is not pseudoconvex, Th. 7.2 and 7.7 show that log is not plurisubharmonic in any neighborhood of . Hence there exists Cn such that c= 2 log (z + t) tt
|t=0
>0
for some z in the neighborhood of where C 2 . By Taylors formula, we have log (z + t) = log (z) + Re(at + bt2 ) + c|t|2 + o(|t|2 ) with a, b C. Now, choose z0 such that (z) = |z z0 | and set h(t) = z + t + eat+bt (z0 z),
2
t C.
60
Then we get h(0) = z0 and (h(t)) (z + t) (z) eat+bt (z) eat+bt

2 2
ec|t|
/2
(z) c|t|2 /3
when |t| is suciently small. Since (h(0)) = (z0 ) = 0, we obtain at t = 0 : (h(t)) = t 2 (h(t)) = tt (z0 ) h (0) = 0, j zj 2 (z0 ) h (0)h (0) > 0, j k zj z k
hence h (0) h T,z0 and L,z0 (h (0)) < 0. (7.13) Denition. The boundary is said to be weakly (resp. strongly) pseudoconvex if L is semipositive (resp. positive denite) on . The boundary is said to be Levi at if L 0. (7.14) Remark. Lemma 7.11 shows that is Levi at if and only if the subbundle h T T is integrable (i.e. stable under the Lie bracket). Assume that is of class k , k 2. Then h T is of class C k1 . By Frobenius theorem, the integrability condition implies that h T is the tangent bundle to a k foliation of whose leaves have real dimension 2n 2. But the leaves themselves must be complex analytic since h T is a complex vector space (cf. Lemma 7.15 below). Therefore is Levi at if and only if it is foliated by complex analytic hypersurfaces. (7.15) Lemma. Let Y be a C 1 -submanifold of a complex analytic manifold X. If the tangent space TY,x is a complex subspace of TX,x at every point x Y , then Y is complex analytic. Proof. Let x0 Y . Select holomorphic coordinates (z1 , . . . , zn ) on X centered at x0 such that TY,x0 is spanned by /z1 , . . . , /zp. Then there exists a neighborhood U = U U of x0 such that Y U is a graph z = h(z ), z = (z1 , . . . , zp ) U , z = (zp+1 , . . . , zn ) with h C 1 (U ) and dh(0) = 0. The dierential of h at z is the composite of the projection of Cp {0} on TY,(z ,h(z )) along {0} Cnp and of the second projection Cn Cnp . Hence dh(z ) is C-linear at every point and h is holomorphic.
8. Exercises
8.1.
Let Cn be an open set such that z , C, || 1 = z .
Show that is a union of polydisks of center 0 (with arbitrary linear changes of coordinates) and infer that the space of polynomials C[z1 , . . . , zn ] is dense in () for the topology of uniform convergence on compact subsets and in () C 0 () for the topology of uniform convergence on . Hint: consider the Taylor expansion of a function f () at the origin, writing it as a series of homogeneous polynomials. To deal with the case of , rst apply a dilation to f .
8. Exercises
61
8.2.
Let B Cn be the unit euclidean ball, S = B and f (B) C 0 (B). Our goal is to check the following Cauchy formula: f (w) = 1 2n1 Z f (z) d(z). (1 w, z )n
a) By means of a unitary transformation and Exercise 8.1, reduce the question to the case when w = (w1 , 0, . . . , 0) and f (z) is a monomial z . R b) Show that the integral B z z k d(z) vanishes unless = (k, 0, . . . , 0). Compute the value of the 1 R remaining integral by the Fubini theorem, as well as the integrals S z z k d(z). 1 c) Prove the formula by a suitable power series expansion.
(M) p
8.3.
A current T coecients.
is said to be normal if both T and dT are of order zero, i.e. have measure
a) If T is normal and has support contained in a C 1 submanifold Y M, show that there exists a normal current on Y such that T = j , where j : Y M is the inclusion. Hint: if x1 = . . . = xq = 0 are equations of Y in a coordinate system (x1 , . . . , xn ), observe that xj T = xj dT = 0 for 1 j q and infer that dx1 . . . dxq can be factorized in all terms of T . b) What happens if p > dim Y ? c) Are a) and b) valid when the normality assumption is dropped ? P 8.4. Let T = 1 j n Tj dzj be a closed current of bidegree (0, 1) with compact support in Cn such that d T = 0. a) Show that the partial convolution S = (1/z1 ) 1 T1 is a solution of the equation d S = T . e b) Let K = Supp T . If n 2, show that S has support in the compact set K equal to the union of K and of all bounded components of Cn K. Hint: observe that S is holomorphic on Cn K and that S vanishes for |z2 | + . . . + |zn | large. P 1, 8.5. Alternative proof of the Dolbeault-Grothendieck lemma. Let v = |J |=qvJ dzJ , q be a smooth form of bidegree (0, q) on a polydisk = D(0, R) Cn , such that d v = 0, and let = D(0, r) . Let k be the smallest integer such that the monomials dz J appearing in v only involve dz 1 , . . ., dz k . Prove by induction on k that the equation d u = v can be solved on . Hint: set v = f dz k + g where f , g only involve dz 1 , . . ., dz k1 . Then consider v d F where F = X FJ dz J , FJ (z) = ((zk )fJ ) k 1 zk ,
|J |=q1
where k denotes theP partial convolution with respect to zk , (zk ) is a cut-o function equal to 1 on D(0, rk + ) and f = |J |=q1 fJ dz J .
8.6.
8.7.
Construct locally bounded non continuous subharmonic functions on C. P Hint: consider eu where u(z) = j 1 2j log |z 1/j|. Let be an open subset of Rn , n
2, and u a subharmonic function which is not locally .
a) For every open set , show that there is a positive measure with support in and a harmonic function h on such that u = N + h on . b) Use this representation to prove the following properties: u Lp for all p < n/(n 2) and loc u/xj Lp for all p < n/(n 1). loc Show that a connected open set Rn is convex if and only if has a locally convex exhaustion function . Hint: to show the suciency, take a path : [0, 1] joining two arbitrary points a, b and consider the restriction of to [a, (t0 )] where t0 is the supremum of all t such that [a, (u)] for u [0, t].
8.8.
62
Chapter I. Complex Dierential Calculus and Pseudoconvexity Let r1 , r2 ]1, +[. Consider the compact set K = {|z1 | r1 , |z2 | 1} {|z1 | 1 , |z2 | r2 } C2 .
8.9.
Show that the holomorphic hull of K in C2 is b K = {|z1 | r1 , |z2 | r2 , |z1 |1/ log r1 |z2 |1/ log r2 e}.
b Hint: to show that K is contained in this set, consider all holomorphic monomials f (z1 , z2 ) = z1 1 z2 2 . To show the converse inclusion, apply the maximum principle to the domain |z1 | r1 , |z2 | r2 on suitably chosen Riemann surfaces z1 1 z2 2 = .
Compute the rank of the Levi form of the ellipsoid |z1 |2 + |z3 |4 + |z3 |6 < 1 at every point of the boundary. P 8.11. Let X be a complex manifold and let u(z) = jN |fj |2 , fj (X), be a series converging uniformly on every compact subset of X. Prove that the limit is real analytic and that the series remains uniformly convergent by taking derivatives term by term. Hint: since the problem is local, take X = B(0, r), a ball in Cn . Let gj (z) = gj (z) be the conjugate P function of fj and let U(z, w) = jN fj (z)gj (w) on X X. Using the Cauchy-Schwarz inequality, show that this series of holomorphic functions is uniformly convergent on every compact subset of X X.
8.10.
8.12.
Let Cn be a bounded open set with C 2 boundary.
a) Let a be a given point. Let en be the outward normal vector to T,a , (e1 , . . . , en1 ) an orthonormal basis of h Ta () in which the Levi form is diagonal and (z1 , . . . , zn ) the associated linear coordinates centered at a. Show that there is a neighborhood V of a such that V is the graph Re zn = (z1 , . . . , zn1 , Im zn ) of a function such that (z) = O(|z|2 ) and the matrix 2 /zj z k (0), 1 j, k n 1 is diagonal. P b) Show that there exist local analytic coordinates w1 = z1 , . . . , wn1 = zn1 , wn = zn + cjk zj zk on a neighborhood V of a = 0 such that V = V {Re wn + X j |wj |2 + o(|w|2 ) < 0},
n
j R
1 j
c) Prove that is strongly pseudoconvex at a if and only if there is a neighborhood U of a and a biholomorphism of U onto some open set of Cn such that ( U) is strongly convex.
and that n can be assigned to any given value by a suitable choice of the coordinates. Hint: Consider the Taylor expansion of order 2 of the dening function (z) = (Re zn + (z))(1 + P Re cj zj ) where cj C are chosen in a suitable way.
d) Assume that the Levi form of is not semipositive. Show that all holomorphic functions f () extend to some (xed) neighborhood of a. Hint: assume for example 1 < 0. For > 0 small, show that contains the Hartogs gure {/2 < |w1 | < } {|wj | < 2 }1<j<n {|wn | < 3/2 , Re wn < 3 } {|w1 | < } {|wj | < 2 }1<j<n {|wn | < 3/2 , Re wn < 2 }.
8.13.
Let Cn be a bounded open set with C 2 boundary and C 2 (, R) such that < 0 on , = 0 and d = 0 on . Let f C 1 (, C) be a function satisfying the tangential Cauchy-Riemann equations 1 f = 0, h T , = ( + iJ). 2 a) Let f0 be a C 1 extension of f to . Show that d f0 d = 0 on and infer that v = 1 d f0 is l a d -closed current on Cn . b) Show that the solution u of d u = v provided by Cor. 3.27 is continuous and that f admits an e extension f () C 0 () if is connected.
8. Exercises
63
8.14. Let Cn be a bounded pseudoconvex domain with C 2 boundary and let (z) = d(z, ) be the euclidean distance to the boundary.
a) Use the plurisubharmonicity of log to prove the following fact: for every > 0 there is a constant C > 0 such that Hz () |d z .|2 + + C ||2 0 (z) |(z)|2 for Cn and z near . b) Set (z) = log (z) + K|z|2 . Show that for K large and small the function 2 (z) = exp (z) = eK|z| (z) is plurisubharmonic. c) Prove the existence of a plurisubharmonic exhaustion function u : [1, 0[ of class C 2 such that |u(z)| has the same order of magnitude as (z) when z tends to . Hint: consult [Diederich-Fornaess 1976].
8.15.
Let = + iRn be a connected tube in Cn of base .
a) Assume rst that n = 2. Let T R2 be the triangle x1 0, x2 0, x1 + x2 1, and assume that the two edges [0, 1] {0} and {0} [0, 1] are contained in . Show that every holomorphic function f () extends to a neighborhood of T + iR2 . Hint: let : C2 R2 be the projection on the real part and M the intersection of 1 ((1 + )T ) 2 2 with the Riemann surface z1 + z2 2 (z1 + z2 ) = 1 (a non degenerate ane conic). Show that M is compact and that (M ) ([0, 1 + ] {0}) ({0} [0, 1 + ]) , ([0, 1] M ) T for small. Use the Cauchy formula along M (in some parametrization of the conic) to obtain an extension of f to [0, 1] M + iRn . b b) In general, show that every f () extends to the convex hull . Hint: given a, b , consider a polygonal line joining a and b and apply a) inductively to obtain an extension along [a, b] + iRn .
8.16.
For each integer 1, consider the algebraic variety n o X = (z, w, t) C3 ; wt = p (z) , p (z) =
1 k
(z 1/k),
and the map j : X X+1 such that j (z, w, t) = z, w, t z 1 +1 .
a) Show that X is a Stein manifold, and that j is an embedding of X onto an open subset of X+1 . b) Dene X = lim(X , j ), and let : X C2 be the projection to the rst two coordinates. Since +1 j = , there exists a holomorphic map : X C2 , = lim . Show that n o C2 (X) = (z, 0) C2 ; z = 1/, N, 1 , and especially, that (0, 0) (X). / c) Consider the compact set K = 1 {(z, w) C2 ; |z| 1, |w| = 1} .
b By looking at points of the forms (1/, w, 0), |w| = 1, show that 1 (1/, 1/) K(X) . Conclude from this that X is not holomorphically convex (this example is due to [Fornaess 1977]).
64
8.17.
e Let X be a complex manifold, and let : X X be a holomorphic unramied covering of X e are assumed to be connected). (X and X
b) Let (U ) be a locally nite covering of X by open balls contained in coordinate open sets, such that all intersections U U are dieomorphic to convex open sets (see Lemma IV-6.9). Let be a partition of unity subordinate to the covering (U ), and let be the convolution of 0 with a regularizing kernel on each piece of 1 (U ) which is mapped biholomorphically onto U . P Finally, set = ( ) . Show that if ( ) is a collection of suciently small positive numbers, e then is a smooth exhaustion function on X.
e a) Let g be a complete riemannian metric on X, and let e be the geodesic distance on X associated d to e = g (see VIII-2.3 for denitions). Show that e is complete and that 0 (x) := e x0 ) is a g g d(x, e for any given point x0 X. e continuous exhaustion function on X,
c) Using the fact that 0 is 1-Lipschitz with respect to e show that derivatives || (x)/x of a given d, order with respect to coordinates in U are uniformly bounded in all components of 1 (U ), at least when x lies in the compact subset Supp . Conclude from this that there exists a positive e hermitian form with continuous coecients on X such that H on X. e d) If X is strongly pseudoconvex, show that X is also strongly pseudoconvex. Hint: let be a smooth strictly plurisubharmonic exhaustion function on X. Show that there exists a smooth convex increasing function : R R such that + is strictly plurisubharmonic.
65
Chapter II
Coherent Sheaves and Analytic Spaces
The chapter starts with rather general and abstract concepts concerning sheaves and ringed spaces. Introduced in the decade 1950-1960 by Leray, Cartan, Serre and Grothendieck, sheaves and ringed spaces have since been recognized as the adequate tools to handle algebraic varieties and analytic spaces in a unied framework. We then concentrate ourselves on the theory of complex analytic functions. The second section is devoted to a proof of the Weierstrass preparation theorem, which is nothing but a division algorithm for holomorphic functions. It is used to derive algebraic properties of the ring n of germs of holomorphic functions in Cn . Coherent analytic sheaves are then introduced and the fundamental coherence theorem of Oka is proved. Basic properties of analytic sets are investigated in detail: local parametrization theorem, Hilberts Nullstellensatz, coherence of the ideal sheaf of an analytic set, analyticity of the singular set. The formalism of complex spaces is then developed and gives a natural setting for the proof of more global properties (decomposition into global irreducible components, maximum principle). After a few denitions concerning cycles, divisors and meromorphic functions, we investigate the important notion of normal space and establish the Oka normalization theorem. Next, the Remmert-Stein extension theorem and the Remmert proper mapping theorem on images of analytic sets are proved by means of semi-continuity results on the rank of morphisms. As an application, we give a proof of Chows theorem asserting that every analytic subset of Pn is algebraic. Finally, the concept of analytic scheme with nilpotent elements is introduced as a generalization of complex spaces, and we discuss the concepts of bimeromorphic maps, modications and blowing-up.
1. Presheaves and Sheaves

1.A. Main Denitions. Sheaves have become a very important tool in analytic or algebraic geometry as well as in algebraic topology. They are especially useful when one wants to relate global properties of an object to its local properties (the latter being usually easier to establish). We rst introduce the axioms of presheaves and sheaves in full generality and give some basic examples. (1.1) Denition. Let X be a topological space. A presheaf following data: a) a collection of non empty sets b) a collection of maps U,V : the transitivity property c) U,V V,W = U,W The set on X consists of the
(U ) associated with every open set U X, (U ) dened whenever U V and satisfying U,U = IdU for every U . over U .
(V )
for U V W,
(U ) is called the set of sections of the presheaf
66
Chapter II. Coherent Sheaves and Analytic Spaces
Most often, the presheaf For instance:
is supposed to carry an additional algebraic structure.
(1.2) Denition. A presheaf is said to be a presheaf of abelian groups (resp. rings, Rmodules, algebras) if all sets (U ) are abelian groups (resp. rings, R-modules, algebras) and if the maps U,V are morphisms of these algebraic structures. In this case, we always assume that () = {0}. (1.3) Example. If we assign to each open set U X the set (U ) of all real valued continuous functions on U and let U,V be the obvious restriction morphism (V ) (U ), then is a presheaf of rings on X. Similarly if X is dierentiable (resp. complex analytic) manifold, there are well dened presheaves of rings k of functions of class k (resp. ) of holomorphic functions) on X. Because of these examples, the maps U,V in Def. 1.1 are often viewed intuitively as restriction homomorphisms, although the sets (U ) are not necessarily sets of functions dened over U . For the simplicity of notation we often just write U,V (f ) = fU whenever f (V ), V U . For the above presheaves , k , , the properties of functions under consideration are purely local. As a consequence, these presheaves satisfy the following additional gluing axioms, where (U ) and U = U are arbitrary open subsets of X : (1.4 ) If F (U ) are such that U U ,U (F ) = U U ,U (F ) (U ) such that U ,U (F ) = F ;
for all , , there exists F If F, G
(1.4 )
(U ) and U ,U (F ) = U ,U (G) for all , then F = G ;
in other words, local sections over the sets U can be glued together if they coincide in the intersections and the resulting section on U is uniquely dened. Not all presheaves satisfy (1.4 ) and (1.4 ): (1.5) Example. Let E be an arbitrary set with a distinguished element 0 (e.g. an abelian group, a R-module, . . .). The constant presheaf EX on X is dened to be EX (U ) = E for all = U X and EX () = {0}, with restriction maps U,V = IdE if = U V and U,V = 0 if U = . Then axiom (1.4 ) is not satised if U is the union of two disjoint open sets U1 , U2 and E contains a non zero element. (1.6) Denition. A presheaf (1.4 ) and (1.4 ). is said to be a sheaf if it satises the gluing axioms
are presheaves of abelian groups (or of some other algebraic structure) on If , the same space X, a presheaf morphism : is a collection of morphisms U : (U ) (U ) commuting with the restriction morphisms, i.e. such that for each pair U V there is a commutative diagram V (V ) (V ) U,V U,V U (U ) (U ).
67
We say that is a subpresheaf of in the case where U : (U ) (U ) is the inclusion morphism; the commutation property then means that U,V ( (V )) (U ) for all U , V , and that U,V coincides with U,V on (V ). If is a subpresheaf of a presheaf of abelian groups, there is a presheaf quotient = / dened by (U ) = (U )/ (U ). In a similar way, one denes the presheaf kernel (resp. presheaf image, presheaf cokernel) to be the presheaves of a presheaf morphism : U Ker U , U Im U , U CokerU . The direct sum of presheaves of abelian groups , is the presheaf U (U ) (U ), the tensor product of presheaves of R-modules is U (U ) R (U ), etc . . . (1.7) Remark. The reader should take care of the fact that the presheaf quotient of a sheaf by a subsheaf is not necessarily a sheaf. To give a specic example, let X = S 1 be the unit circle in R2 , let be the sheaf of continuous complex valued functions and the subsheaf of integral valued continuous functions (i.e. locally constant functions to Z). The exponential map = exp(2i) : to the sheaf of invertible continuous functions, and the kernel is a morphism from of is precisely . However U is surjective for all U = X but maps (X) onto the multiplicative subgroup of continuous functions of (X) of degree 0. Therefore the quotient presheaf / is not isomorphic with , although their groups of sections are the same for all U = X. Since is a sheaf, we see that / does not satisfy property (1.4 ).
In order to overcome the diculty appearing in Example 1.7, it is necessary to introduce a suitable process by which we can produce a sheaf from a presheaf. For this, it is convenient to introduce a slightly modied viewpoint for sheaves. is a presheaf, we dene the set (1.8) Denition. If x X to be the abstract inductive limit
x x
of germs of
at a point
= lim
(U ), U,V .
Ux
More explicitely, x is the set of equivalence classes of elements in the disjoint union (U1 ), Ux (U ) taken over all open neighborhoods U of x, with two elements F1 F2 (U2 ) being equivalent, F1 F2 , if and only if there is a neighborhood V U1 , U2 such that F1V = F2V , i.e., V U1 (F1 ) = V U2 (F2 ). The germ of an element F (U ) at a point x U will be denoted by Fx . Let be an arbitrary presheaf. The disjoint union = xX with a natural topology as follows: for every F (U ), we set F,U = Fx ; x U and choose the F,U to be a basis of the topology of ; note that this family is stable by intersection: F,U G,V = H,W where W is the (open) set of points x U V at
x
can be equipped
68
which Fx = Gx and H = W,U (F ). The obvious projection map : X which sends x to {x} is then a local homeomorphism (it is actually a homeomorphism from F,U onto U ). This leads in a natural way to the following denition: (1.9) Denition. Let X and be topological spaces (not necessarily Hausdor), and let : X be a mapping such that a) maps b) is a local homeomorphism, that is, every point in has an open neighborhood which is mapped homeomorphically by onto an open subset of X. Then is called a sheaf-space on X and is called the projection of then x = 1 (x) is called the stalk of at x.
onto X ;
on X. If x X,
If Y is a subset of X, we denote by (Y, ) the set of sections of on Y , i.e. the set of continuous functions F : Y such that F = IdY . It is clear that the presheaf dened by the collection of sets (U ) := (U, ) for all open sets U X together with the restriction maps U,V satises axioms (1.4 ) and (1.4 ), hence is a sheaf. The set of germs of at x is in one-to-one correspondence with the stalk x = 1 (x), thanks to the local homeomorphism assumption 1.9 b). This shows that one can associate in a natural way a sheaf to every sheaf-space , and that the sheaf-space ( ) can be considered to be identical to the original sheaf-space . Since the assignment from sheaf-spaces to sheaves is an equivalence of categories, we will usually omit the prime sign in the notation of and thus use the same symbols for a sheaf-space and its associated sheaf of sections; in a corresponding way, we write (U, ) = (U ) when U is an open set. Conversely, given a presheaf obvious presheaf morphism (1.10) (U )
on X, we have an associated sheaf-space F F = (U x Fx ).
and an
(U ) = (U, ),
satises axiom (1.4 ), and it is not This morphism is clearly injective if and only if dicult to see that (1.4 ) and (1.4 ) together imply surjectivity. Therefore is is a sheaf. According to the equivalence of categories an isomorphism if and only if between sheaves and sheaf-spaces mentioned above, we will use from now on the same symbol for the sheaf-space and its associated sheaf ; one says that is the sheaf associated with the presheaf . If itself is a sheaf, we will again identify and , but we will of course keep the notational dierence for a presheaf which is not a sheaf. (1.11) Example. The sheaf associated to the constant presheaf of stalk E over X is the sheaf of locally constant functions X E. This sheaf will be denoted merely by EX or E if there is no risk of confusion with the corresponding presheaf. In Example 1.7, we have = ZX and the sheaf ( /ZX ) associated with the quotient presheaf /ZX is isomorphic to via the exponential map. In the sequel, we usually work in the category of sheaves rather than in the category by a subsheaf of presheaves themselves. For instance, the quotient / of a sheaf generally refers to the sheaf associated with the quotient presheaf: its stalks are equal
69
to x / x , but a section G of / over an open set U need not necessarily come from a global section of (U ) ; what can be only said is that there is a covering (U ) of U and local sections F (U ) representing GU such that (F F )U U belongs to (U U ). A sheaf morphism : is said to be injective (resp. surjective) if the germ morphism x : x x is injective (resp. surjective) for every x X. Let us note again that a surjective sheaf morphism does not necessarily give rise to surjective morphisms U : (U ) (U ). 1.B. Direct and Inverse Images of Sheaves Let X, Y be topological spaces and let f : X Y be a continuous map. If presheaf on X, the direct image f is the presheaf on Y dened by (1.12) f (U ) = f 1 (U ) also satises axioms is a
is a sheaf, it is clear that f for all open sets U Y . When (1.4 ) and (1.4 ), thus f is a sheaf. Its stalks are given by (1.13) (f )y = lim
V y
f 1 (V )
where V runs over all open neighborhoods of y Y .
Now, let be a sheaf on Y , viewed as a sheaf-space with projection map : We dene the inverse image f 1 by f 1 = Y X = (s, x) X ; (s) = f (x)
Y.
(1.14)
X. It is then easy to see with the topology induced by the product topology on that the projection = pr2 : f 1 X is a local homeomorphism, therefore f 1 is a sheaf on X. By construction, the stalks of f 1 are (1.15) (f 1 )x =
f (x) ,
and the sections f 1 (U ) can be considered as continuous mappings s : U such that = f . In particular, any section s (V ) on an open set V Y has a pull-back (1.16) f s = s f f 1 f 1 (V ) .
There are always natural sheaf morphisms (1.17) f 1 f , f f 1
dened as follows. A germ in (f 1 f )x = (f )f (x) is dened by a local section s (f )(V ) = (f 1 (V )) for some neighborhood V of f (x) ; this section can be mapped to the germ sx x . In the opposite direction, the pull-back f s of a section s (V ) can be seen by (1.16) as a section of f f 1 (V ). It is not dicult to see that these natural morphisms are not isomorphisms in general. For instance, if f is a nite = EX , = EY to be constant sheaves, covering map with q sheets and if we take q q q then f EX EY and f 1 EY = EX , thus f 1 f EX EX and f f 1 EY EY .
70
1.C. Ringed Spaces Many natural geometric structures considered in analytic or algebraic geometry can be described in a convenient way as topological spaces equipped with a suitable structure sheaf which, most often, is a sheaf of commutative rings. For instance, a lot of properties of k dierentiable (resp. real analytic, complex analytic) manifolds can be described in terms of their sheaf of rings k of dierentiable functions (resp. of X X real analytic functions, X of holomorphic functions). We rst recall a few standard denitions concerning rings, referring to textbooks on algebra for more details (see e.g. [Lang 1965]). (1.18) Some denitions and conventions about rings. All our rings R are supposed implicitly to have a unit element 1R (if R = {0}, we agree that 1R = 0R ), and a ring morphism R R is supposed to map 1R to 1R . In the subsequent denitions, we assume that all rings under consideration are commutative. a) An ideal I R is said to be prime if xy I implies x I or y I, i.e., if the quotient ring R/I is entire. b) An ideal I R is said to be maximal if I = R and there are no ideals J such that I J R (equivalently, if the quotient ring R/I is a eld). c) The ring R is said to be a local ring if R has a unique maximal ideal m (equivalently, if R has an ideal m such that all elements of R m are invertible). Its residual eld is dened to be the quotient eld R/m. d) The ring R is said to be Noetherian if every ideal I R is nitely generated (equivalently, if every increasing sequence of ideals I1 I2 . . . is stationary). the e) The radical I of an ideal I is set of all elements x R such that some power xm , m N , lies in in I. Then I is again an ideal of R. f) The nilradical N (R) = {0} is the ideal of nilpotent elements of R. The ring R is said to be reduced if N (R) = {0}. Otherwise, its reduction is dened to be the reduced ring R/N (R). We now introduce the general notion of a ringed space. (1.19) Denition. A ringed space is a pair (X, X ) consisting of a topological space X and of a sheaf of rings X on X, called the structure sheaf. A morphism of ringed spaces is a pair (f, F ) where f : X Y is a continuous map and a homomorphism of sheaves of rings on X, called the comorphism of F . If F : (X, X ) (Y, Y ) and G : (Y, Y ) (Z, Z ) are morphisms of ringed spaces, the composite G F is the pair consisting of the map g f : X Z and of the comorphism (G F ) = F f 1 G : (1.20) X , F f 1 G : f 1 g 1 Z f 1 Y X,x . Fx Gf (x) : Z,gf (x) Y,f (x)
f 1 G F
F : (X, X ) (Y, Y )
Fx :
F : f 1 Y X ,
Y,f (x) X,x
71
We say of course that F is an isomorphism of ringed spaces if there exists G such that G F = IdX and F G = IdY . If (X, X ) is a ringed space, the nilradical of X denes an ideal subsheaf X of X , and the identity map IdX : X X together with the ring homomorphism X X /X denes a ringed space morphism (1.21) (X, X /X ) (X, X )
called the reduction morphism. Quite often, the letter X by itself is used to denote the ringed space (X, X ) ; we then denote by Xred = (X, X /X ) its reduction. The ringed space X is said to be reduced if X = 0, in which case the reduction morphism Xred X is an isomorphism. In all examples considered later on in this book, the structure sheaf X will be a sheaf of local rings over some eld k. The relevant denition is as follows. (1.22) Denition. a) A local ringed space is a ringed space (X, X ) such that all stalks X,x are local rings. The maximal ideal of X,x will be denoted by mX,x . A morphism F = (f, F ) : (X, X ) (Y, Y ) of local ringed spaces is a morphism of ringed spaces such that Fx (mY,f (x) ) mX,x at any point x X (i.e., Fx is a local homomorphism of rings). b) A local ringed space over a eld k is a local ringed space (X, X ) such that all rings X,x are local k-algebras with residual eld X,x /mX,x k. A morphism F between such spaces is supposed to have its comorphism dened by local k-homomorphisms Fx : Y,f (x) X,x . If (X, X ) is a local ringed space over k, we can associate to each section s X (U ) a function s : U k, x s(x) k = X,x /mX,x ,
and we get a sheaf morphism X X onto a subsheaf of rings functions from X to k. We clearly have a factorization
of the sheaf of
X X /X X ,
and thus a corresponding factorization of ringed space morphisms (with IdX as the underlying set theoretic map) Xst-red Xred X where Xst-red = (X, RX ) is called the strong reduction of (X, X ). It is easy to see that Xst-red is actually a reduced local ringed space over k. We say that X is strongly reduced if X X is an isomorphism, that is, if X can be identied with a subsheaf of the sheaf of functions X k (in our applications to the theory of algebraic or analytic schemes, the concepts of reduction and strong reduction will actually be the same ; in general, these notions dier, see Exercise ??.??). It is important to observe that reduction (resp. strong reduction) is a fonctorial process:
72
if F = (f, F ) : (X, X ) (Y, Y ) is a morphism of ringed spaces (resp. of local ringed spaces over k), there are natural reductions
Fred = (f, Fred ) : Xred Yred , Fred : Y,f (x) /Y,f (x) X,x /X,x ,
where f is the usual pull-back comorphism associated with f . Therefore, if (X, X ) and (Y, Y ) are strongly reduced, the morphism F is completely determined by the underlying set-theoretic map f . Our rst basic examples of (strongly reduced) ringed spaces are the various types of manifolds already dened in Chapter I. The language of ringed spaces provides an equivalent but more elegant and more intrinsic denition. (1.23) Denition. Let X be a Hausdor separable topological space. One can dene the category of k , k N {, }, dierentiable manifolds (resp. complex analytic manifolds) to be the category of reduced local ringed spaces (X, X ) over R (resp. over C), such that every point x X has a neighborhood U on which the restriction (U, XU ) is isomorphic to a ringed space (, k ) where Rn is an open set and k is the sheaf of k dierentiable functions (resp. (, ), where Cn is an open subset, and is the sheaf of holomorphic functions on ). We say that the ringed spaces (, k ) and (, ) are the models of the category of dierentiable (resp. complex analytic) manifolds, and that a general object (X, X ) in the category is locally isomorphic to one of the given model spaces. It is easy to see that the corresponding ringed spaces morphisms are nothing but the usual concepts of dierentiable and holomorphic maps. 1.D. Algebraic Varieties over a Field As a second illustration of the notion of ringed space, we present here a brief introduction to the formalism of algebraic varieties, referring to [Hartshorne 1977] or [EGA 1967] for a much more detailed exposition. Our hope is that the reader who already has some background of analytic or algebraic geometry will nd some hints of the strong interconnections between both theories. Beginners are invited to skip this section and proceed directly to the theory of complex analytic sheaves in ,2. All rings or algebras occurring in this section are supposed to be commutative rings with unit. 1.D.1. Ane Algebraic Sets. Let k be an algebraically closed eld of any characteristic. An ane algebraic set is a subset X k N of the ane space k N dened by an arbitrary collection S k[T1 , . . . , TN ] of polynomials, that is, X = V (S) = (z1 , . . . , zN ) k N ; P (z1 , . . . , zN ) = 0, P S . Of course, if J k[T1 , . . . , TN ] is the ideal generated by S, then V (S) = V (J). As k[T1 , . . . , TN ] is Noetherian, J is generated by nitely many elements (P1 , . . . , Pm ), thus X = V ({P1 , . . . , Pm }) is always dened by nitely many equations. Conversely, for any subset Y k N , we consider the ideal I(Y ) of k[T1 , . . . , TN ], dened by I(Y ) = P k[T1 , . . . , TN ] ; P (z) = 0, z Y . Of course, if Y k N is an algebraic set, we have V (I(Y )) = Y . In the opposite direction, we have the following fundamental result.
Fst-red = (f, f ) : Xst-red Yst-red ,
f : Y,f (x) X,x ,
ssf
73
(1.24) Hilberts Nullstellensatz (see [Lang 1965]). If J k[T1 , . . . , TN ] is an ideal, then I(V (J)) = J. If X = V (J) k N is an ane algebraic set, we dene the (reduced) ring (X) of algebraic functions on X to be the set of all functions X k which are restrictions of polynomials, i.e., (X) = k[T1 , . . . , TN ]/I(X) = k[T1 , . . . , TN ]/ J. (1.25) This is clearly a reduced k-algebra. An (algebraic) morphism of ane algebraic sets X = V (J) k N , Y = V (J ) k N is a map f : Y X which is the restriction of a polynomial map k N tok N . We then get a k-algebra homomomorphism f : (X) (Y ), (1.26) X (X), s s f, f f
called the comorphism of f . In this way, we have dened a contravariant fonctor
from the category of ane algebraic sets to the category of nitely generated reduced k-algebras. We are going to show the existence of a natural fonctor going in the opposite direction. In fact, let us start with an arbitrary nitely generated algebra A (not necessarily reduced at this moment). For any choice of generators (g1 , . . . , gN ) of A we get a surjective morphism of the polynomial ring k[T1 , . . . , TN ] onto A, k[T1 , . . . , TN ] A, Tj gj ,
and thus A k[T1 , . . . , TN ]/J with the ideal J being the kernel of this morphism. It is well-known that every maximal ideal m of A has codimension 1 in A (see [Lang 1965]), so that m gives rise to a k-algebra homomorphism A A/m = k. We thus get a bijection Homalg (A, k) Spm(A), u Ker u between the set of k-algebra homomorphisms and the set Spm(A) of maximal ideals of A. In fact, if A = k[T1 , . . . , TN ]/J, an element Homalg (A, k) is completely determined by the values zj = (Tj mod J), and the corresponding algebra homomorphism k[T1 , . . . , TN ] k, P P (z1 , . . . , zN ) can be factorized mod J if and only if z = (z1 , . . . , zN ) k N satises the equations P (z1 , . . . , zN ) = 0, We infer from this that Spm(A) V (J) = (z1 , . . . , zN ) k N ; P (z1 , . . . , zN ) = 0, P J can be identied with the ane algebraic set V (J) k N . If we are given an algebra homomorphism : A B of nitely generated k-algebras we get a corresponding map Spm() : Spm(B) Spm(A) described either as Spm(B) Spm(A), Homalg (B, k) Homalg (A, k), m 1 (m) or v v . P J.
74
If B = k[T1 , . . . , TN ]/J and Spm(B) = V (J ) k N , it is easy to see that Spm() : Spm(B) Spm(A) is the restriction of the polynomial map
f : kN kN ,
w f (w) = (P1 (w), . . . , PN (w)),
where Pj k[T1 , . . . , TN ] are polynomials such that Pj = (Tj ) mod J in B. We have in this way dened a contravariant fonctor
(1.27)
A Spm(A),
Spm()
from the category of nitely generated k-algebras to the category of ane algebraic sets. Since A = k[T1 , . . . , TN ]/J and its reduction A/N (A) = k[T1 , . . . , TN ]/ J give rise to the same algebraic set V (J) = Spm(A) = Spm(A/N (A)) = V ( J), we see that the category of ane algebraic sets is actually equivalent to the subcategory of reduced nitely generated k-algebras. (1.28) Example. The simplest example of an ane algebraic set is the ane space k N = Spm(k[T1 , . . . , TN ]), in particular Spm(k) = k 0 is just one point. We agree that Spm({0}) = (observe that V (J) = when J is the unit ideal in k[T1 , . . . , TN ]). 1.D.2. Zariski Topology and Ane Algebraic Schemes. Let A be a nitely generated algebra and X = Spm(A). To each ideal a A we associate the zero variety V (a) X which consists of all elements m X = Spm(A) such that m a ; if A k[T1 , . . . , TN ]/J and X V (J) k N ,
then V (a) can be identied with the zero variety V (Ja ) X of the inverse image Ja of a in k[T1 , . . . , TN ]. For any family (a ) of ideals in A we have V( a ) = V (a ), V (a1 ) V (a2 ) = V (a1 a2 ),
hence there exists a unique topology on X such that the closed sets consist precisely of all algebraic subsets (V (a))aA of X. This topology is called the Zariski topology. The Zariski topology is almost never Hausdor (for example, if X = k is the ane line, the open sets are and complements of nite sets, thus any two nonempty open sets have nonempty intersection). However, X is a Noetherian space, that is, a topological space in which every decreasing sequence of closed sets is stationary; an equivalent denition is to require that every open set is quasi-compact (from any open covering of an open set, one can extract a nite covering). We now come to the concept of ane open subsets. For s A, the open set D(s) = X V (s) can be given the structure of an ane algebraic variety. In fact, if A = k[T1 , . . . , TN ]/J and s is represented by a polynomial in k[T1 , . . . , TN ], the localized ring
75
A[1/s] can be written as A[1/s] = k[T1 , . . . , TN , TN+1 ]/Js where Js = J[TN+1 ]+(sTN+1 1), thus V (Js ) = {(z, w) V (J) k ; s(z) w = 1} V (I) s1 (0) and D(s) can be identied with Spm(A[1/s]). We have D(s1 ) D(s2 ) = D(s1 s2 ), and the sets (D(s))sA are easily seen to be a basis of the Zariski topology on X. The open sets D(s) are called ane open sets. Since the open sets D(s) containing a given point x X form a basis of neighborhoods, one can dene a sheaf space X such that the ring of germs X,x is the inductive limit
X,x =
D(s)x
lim A[1/s] = {fractions p/q ; p, q A, q(x) = 0}.
This is a local ring with maximal ideal mX,x = {p/q ; p, q A, p(x) = 0, q(x) = 0},
and residual eld X,x /mX,x = k. In this way, we get a ringed space (X, X ) over k. It is easy to see that (X, X ) coincides with the nitely generated k-algebra A. In fact, from the denition of X , a global section is obtained by gluing together local sections pj /sj on ane open sets D(sj ) with D(sj ) = X, 1 j m. This means that the ideal a = (s1 , . . . , sm ) A has an empty zero variety V (a), thus a = A and there are elements uj A with uj sj = 1. The compatibility condition pj /sj = pk /sk implies that these elements are induced by uj pj / uj s j = uj pj A,
as desired. More generally, since the open sets D(s) are ane, we get (D(s), X ) = A[1/s]. It is easy to see that the ringed space (X, X ) is reduced if and only if A itself is reduced; in this case, X is even strongly reduced as Hilberts Nullstellensatz shows. Otherwise, the reduction Xred can obtained from the reduced algebra Ared = A/N (A).
Ringed spaces (X, X ) as above are called ane algebraic schemes over k (although substantially dierent from the usual denition, our denition can be shown to be equivalent in this special situation; compare with [Hartshorne 1977]); see also Exercise ??.??). The category of ane algebraic schemes is equivalent to the category of nitely generated k-algebras (with the arrows reversed). 1.D.3. Algebraic Schemes. Algebraic schemes over k are dened to be ringed spaces over k which are locally isomorphic to ane algebraic schemes, modulo an ad hoc separation condition. (1.29) Denition. An algebraic scheme over k is a local ringed space (X, X ) over k such that a) X has a nite covering by open sets U such that (U , XU ) is isomorphic as a ringed space to an ane algebraic scheme (Spm(A ), Spm(A ) ).
76
b) X satises the algebraic separation axiom, namely the diagonal X of X X is closed for the Zariski topology. A morphism of algebraic schemes is just a morphism of the underlying local ringed spaces. An (abstract) algebraic variety is the same as a reduced algebraic scheme. In the above denition, some words of explanation are needed for b), since the product X Y of algebraic schemes over k is not the ringed space theoretic product, i.e., the product topological space equipped with the structure sheaf pr X k pr Y . Instead, 2 1 we dene the product of two ane algebraic schemes X = Spm(A) and Y = Spm(B) to be X Y = Spm(A k B), equipped with the Zariski topology and the structural sheaf associated with A k B. Notice that the Zariski topology on X Y is not the product topology of the Zariski topologies on X, Y , as the example k 2 = k k shows; also, the rational function 1/(1 z1 z2 ) k2 ,(0,0) is not in k,0 k k,0 . In general, if X, Y are written as X = U and Y = V with ane open sets U , V , we dene X Y to be the union of all open ane charts U V with their associated structure sheaves of ane algebraic varieties, the open sets of X Y being all unions of open sets in the various charts U V . The separation axiom b) is introduced for the sake of excluding pathological examples such as an ane line k {0 } with the origin changed into a double point. 1.D.4. Subschemes. If (X, X ) is an ane algebraic scheme and A = (X, X ) is the associated algebra, we say that (Y, Y ) is a subscheme of (X, X ) if there is an ideal a of A such that Y X is the morphism dened by the algebra morphism A A/a as its comorphism. As Spm(A/a) Spm(A) has for image the set V (a) of maximal ideals m of A containing a, we see that Y = V (a) as a set; let us introduce the ideal subsheaf = aX X . Since the structural sheaf Y is obtained by taken localizations A/a[1/s], it is easy to see that Y coincides with the quotient sheaf X / restricted to Y . Since a has nitely many generators, the ideal sheaf is locally nitely generated (see 2 below). This leads to the following denition. (1.30) Denition. If (X, X ) is an algebraic scheme, a (closed) subscheme is an algebraic scheme (Y, Y ) such that Y is a Zariski closed subset of X, and there is a locally nitely generated ideal subsheaf X such that Y = V () and Y = (X /)Y . If (Y, Y ), (Z, Z ) are subschemes of (X, X ) dened by ideal subsheaves , X , there are corresponding subschemes Y Z and Y Z dened as ringed spaces (Y Z, X /( + )), (Y Z, X / ).
1.D.5. Projective Algebraic Varieties. A very important subcategory of the category of algebraic varieties is provided by projective algebraic varieties. Let PN be the k projective N -space, that is, the set k N+1 {0}/k of equivalence classes of (N + 1)tuples (z0 , . . . , zN ) k N+1 {0} under the equivalence relation given by (z0 , . . . , zN ) (z0 , . . . , zN ), k . The corresponding element of PN will be denoted [z0 : z1 : . . . : k zN ]. It is clear that Pk can be covered by the (N + 1) ane charts U , 0 N , such N that U = [z0 : z1 : . . . : zN ] PN z = 0 . k
77
The set U can be identied with the ane N -space k N by the map U k N , [z0 : z1 : . . . : zN ] z1 z+1 zN z0 z1 . , ,..., , ,..., z z z z z
With this identication, (U ) is the algebra of homogeneous rational functions of degree 0 in z0 , . . . , zN which have just a power of z in their denominator. It is easy to see that the structure sheaves U and U coincide in the intersections U U ; they can be glued together to dene an algebraic variety structure (PN , PN ), such that PN ,[z] k consists of all homogeneous rational functions p/q of degree 0 (i.e., deg p = deg q), such that q(z) = 0. (1.30) Denition. An algebraic scheme or variety (X, X ) is said to be projective if it is isomorphic to a closed subscheme of some projective space (PN , PN ). k We now indicate a standard way of constructing projective schemes. Let S be a collection of homogeneous polynomials Pj k[z0 , . . . , zN ], of degree dj N. We dene an associated projective algebraic set V (S) = [z0 : . . . : zN ] PN ; P (z) = 0, P S . k Let J be the homogeneous ideal of k[z0 , . . . , zN ] generated by S (recall that an ideal J is said to be homogeneous if J = Jm is the direct sum of its homogeneous components, or equivalently, if J is generated by homogeneous elements). We have an associated graded algebra B = k[z0 , . . . , zN ]/J = Bm , Bm = k[z0 , . . . , zN ]m /Jm
such that B is generated by B1 and Bm is a nite dimensional vector space over k for each k. This is enough to construct the desired scheme structure on V (J) := V (Jm ), as we see in the next subsection. 1.D.6. Projective Scheme Associated with a Graded Algebra. Let us start with a reduced graded k-algebra B= Bm
mN
such that B is generated by B0 and B1 as an algebra, and B0 , B1 are nite dimensional vector spaces over k (it then follows that B is nitely generated and that all Bm are nite dimensional vector spaces). Given s Bm , m > 0, we dene a k-algebra As to be the ring of all fractions of homogeneous degree 0 with a power of s as their denominator, i.e., (1.31) As = p/sd ; p Bdm , d N .
m Since As is generated by 1 B1 over B0 , As is a nitely generated algebra. We dene s Us = Spm(As ) to be the associated ane algebraic variety. For s Bm and s Bm , we clearly have algebra homomorphisms
As Ass ,
As Ass ,
78
since Ass is the algebra of all 0-homogeneous fractions with powers of s and s in the denominator. As Ass is the same as the localized ring As [sm /sm ], we see that Uss can be identied with an ane open set in Us , and we thus get canonical injections Uss Us , Uss Us .
(1.32) Denition. If B = mN Bm is a reduced graded algebra generated by its nite dimensional vector subspaces B0 and B1 , we associate an algebraic scheme (X, X ) = Proj(B) as follows. To each nitely generated algebra As = p/sd ; p Bdm , d N we associate an ane algebraic variety Us = Spm(As ). We let X be the union of all open charts Us with the identications Us Us = Uss ; then the collection (Us ) is a basis of the topology of X, and X is the unique sheaf of local k-algebras such that (Us , X ) = As for each Us . The following proposition shows that only nitely many open charts are actually needed to describe X (as required in Def. 1.29 a)). (1.33) Lemma. If s0 , . . . , sN is a basis of B1 , then Proj(B) =
0 j N
Usj .
Proof. In fact, if x X is contained in a chart Us for some s Bm , then Us = m Spm(As ) = , and therefore As = {0}. As As is generated by 1 B1 , we can nd a s fraction f = sj1 . . . sjm /s representing an element f (Us ) such that f (x) = 0. Then x Us f 1 (0), and Us f 1 (0) = Spm(As [1/f ]) = Us Usj1 . . . Usjm . In particular x Usj1 . (1.34) Example. One can consider the projective space PN to be the algebraic scheme k PN = Proj(k[T0 , . . . , TN ]). k
The Proj construction is fonctorial in the following sense: if we have a graded ho momorphism : B B (i.e. an algebra homomorphism such that (Bm ) Bm , then there are corresponding morphisms As A , U(s) Us , and we thus nd a scheme (s) morphism F : Proj(B ) Proj(B). Also, since p/sd = psl /sd+l , the algebras As depend only on components Bm of large degree, and we have As = Asl . It follows easily that there is a canonical isomorphism Proj(B) Proj Blm .
m
Similarly, we may if we wish change a nite number of components Bm without affecting Proj(B). In particular, we may alway assume that B0 = k 1B . By selecting nitely many generators g0 , . . . , gN in B1 , we then nd a surjective graded homomorphism k[T0 , . . . , TN ] B, thus B k[T0 , . . . , TN ]/J for some graded ideal J B. The algebra homomorphism k[T0 , . . . , TN ] B therefore yields a scheme embedding Proj(B) PN onto V (J).
2. The Local Ring of Germs of Analytic Functions
79
We will not pursue further the study of algebraic varieties from this point of view ; in fact we are mostly interested in the case k = C, and algebraic varieties over C are a special case of the more general concept of complex analytic space.

2.A. The Weierstrass Preparation Theorem Our rst goal is to establish a basic factorization and division theorem for analytic functions of several variables, which is essentially due to Weierstrass. We follow here a simple proof given by C.L. Siegel, based on a clever use of the Cauchy formula. Let g be a holomorphic function dened on a neighborhood of 0 in Cn , g 0. There exists a dense set of vectors v Cn {0} such that the function C t g(tv) is not identically zero. In fact the Taylor series of g at the origin can be written
+
g(tv) =
k=0
1 k (k) t g (v) k!
where g (k) is a homogeneous polynomial of degree k on Cn and g (k0 ) 0 for some index k0 . Thus it suces to select v such that g (k0 ) (v) = 0. After a change of coordinates, we may assume that v = (0, . . . , 0, 1). Let s be the vanishing order of zn g(0, . . . , 0, zn ) at zn = 0. There exists rn > 0 such that g(0, . . . , 0, zn ) = 0 when 0 < |zn | rn . By continuity of g and compactness of the circle |zn | = rn , there exists r > 0 and > 0 such that g(z , zn ) = 0 for z Cn1 , |z | r , rn |zn | rn + .
For every integer k N, let us consider the integral Sk (z ) = 1 2i 1

|zn |=rn
g(z , zn )
g k (z , zn ) zn dzn . zn
Then Sk is holomorphic in a neighborhood of |z | r . Rouchs theorem shows that e S0 (z ) is the number of roots zn of g(z , zn ) = 0 in the disk |zn | < rn , thus by continuity S0 (z ) must be a constant s. Let us denote by w1 (z ), . . . , ws (z ) these roots, counted with multiplicity. By denition of rn , we have w1 (0) = . . . = ws (0) = 0, and by the choice of r , we have |wj (z )| < rn for |z | r . The Cauchy residue formula yields
s
Sk (z ) =
j=1
wj (z )k .
Newtons formula shows that the elementary symmetric function ck (z ) of degree k in w1 (z ), . . . , ws (z ) is a polynomial in S1 (z ), . . . , Sk (z ). Hence ck (z ) is holomorphic in a neighborhood of |z | r . Let us set
s
P (z , zn ) =
s zn
c1 (z
s1 )zn
+ + (1) cs (z ) =
j=1
zn wj (z ) .
80
For |z | r , the quotient f = g/P (resp. f = P/g) is holomorphic in zn on the disk |zn | < rn + , because g and P have the same zeros with the same multiplicities, and f (z , zn ) is holomorphic in z for rn |zn | rn + . Therefore f (z , zn ) = 1 2i
|wn |=rn +
f (z , wn ) dwn wn zn r }
is holomorphic in z on a neighborhood of the closed polydisk (r , rn ) = {|z | {|zn | rn }. Thus g/P is invertible and we obtain:
(2.1) Weierstrass preparation theorem. Let g be holomorphic on a neighborhood of s 0 in Cn , such that g(0, zn)/zn has a not zero nite limit at zn = 0. With the above choice of r and rn , one can write g(z) = u(z)P (z , zn ) where u is an invertible holomorphic function in a neighborhood of the polydisk (r , rn ), and P is a Weierstrass polynomial in zn , that is, a polynomial of the form
s s1 P (z , zn ) = zn + a1 (z )zn + + as (z ),
ak (0) = 0, r in Cn1 .
with holomorphic coecients ak (z ) on a neighborhood of |z |
(2.2) Remark. If g vanishes at order m at 0 and v Cn {0} is selected such that g (m) (v) = 0, then s = m and P must also vanish at order m at 0. In that case, the coecients ak (z ) are such that ak (z ) = O(|z |k ), 1 k s. (2.3) Weierstrass division theorem. Every bounded holomorphic function f on = (r , rn ) can be represented in the form (2.4) f (z) = g(z)q(z) + R(z , zn ), s 1 in zn , and
where q and R are analytic in , R(z , zn ) is a polynomial of degree (2.5) for some constant C sup |q|
C sup |f |,
sup |R|
C sup |f |
0 independent of f . The representation (2.4) is unique.
Proof (Siegel). It is sucient to prove the result when g(z) = P (z , zn ) is a Weierstrass polynomial. Let us rst prove the uniqueness. If f = P q1 + R1 = P q2 + R2 , then P (q2 q1 ) + (R2 R1 ) = 0. It follows that the s roots zn of P (z , ) = 0 are zeros of R2 R1 . Since degzn (R2 R1 ) s 1, we must have R2 R1 0, thus q2 q1 0. In order to prove the existence of (q, R), we set q(z , zn ) = lim 1 2i f (z , wn ) dwn , P (z , wn )(wn zn )
0+
|wn |=rn
z;
81
observe that the integral does not depend on when < rn |zn | is small enough. Then q is holomorphic on . The function R = f P q is also holomorphic on and R(z) = lim
0+
1 2i
|wn |=rn
f (z , wn ) P (z , wn ) P (z , zn ) dwn . P (z , wn ) (wn zn )
The expression in brackets has the form

s s (wn
s zn )
+
j=1
sj sj aj (z )(wn zn ) /(wn zn )
hence is a polynomial in zn of degree s 1 with coecients that are holomorphic functions of z . Thus we have the asserted decomposition f = P q + R and sup |R|
C1 sup |f |
where C1 depends on bounds for the aj (z ) and on = min |P (z , zn )| on the compact set {|z | r } {|zn | = rn }. By the maximum principle applied to q = (f R)/P on each disk {z } {|zn | < rn }, we easily get sup |q|
1 (1 + C1 ) sup |f |.
2.B. Algebraic Properties of the Ring
We give here important applications of the Weierstrass preparation theorem to the study of the ring of germs of holomorphic functions in Cn . (2.6) Notation. We let n be the ring of germs of holomorphic functions on Cn at 0. Alternatively, n can be identied with the ring C{z1 , . . . , zn } of convergent power series in z1 , . . . , zn . Proof. By induction on n. For n = 1, n is principal: every ideal = {0} is generated by z s , where s is the minimum of the vanishing orders at 0 of the non zero elements of . Let n 2 and n , = {0}. After a change of variables, we may assume that contains a Weierstrass polynomial P (z , zn ). For every f , the Weierstrass division theorem yields
s1
(2.7) Theorem. The ring n is Noetherian, i.e. every ideal of n is nitely generated.
f (z) = P (z , zn )q(z) + R(z , zn ),
R(z , zn ) =
k=0
k ck (z ) zn ,
and we have R . Let us consider the set of coecients (c0 , . . . , cs1 ) in s n1 corresponding to the polynomials R(z , zn ) which belong to . Then is a n1 submodule of s . By the induction hypothesis n1 is Noetherian; furthermore, every n1 submodule of a nitely generated module over a Noetherian ring is nitely generated ([Lang 1965], Chapter VI). Therefore is nitely generated, and is generated by P and by polynomials R1 , . . . , RN associated with a nite set of generators of .
82
Before going further, we need two lemmas which relate the algebraic properties of n to those of the polynomial ring n1 [zn ]. (2.8) Lemma. Let P, F n1 [zn ] where P is a Weierstrass polynomial. If P divides F in n , then P divides F in n1 [zn ]. Proof. Assume that F (z , zn ) = P (z , zn )h(z), h n . The standard division algorithm of F by P in n1 [zn ] yields F = P Q + R, Q, R n1 [zn ], deg R < deg P.
The uniqueness part of Th. 2.3 implies h(z) = Q(z , zn ) and R 0. (2.9) Lemma. Let P (z , zn ) be a Weierstrass polynomial. a) If P = P1 . . . PN with Pj n1 [zn ], then, up to invertible elements of are Weierstrass polynomials. b) P (z , zn ) is irreducible in
n1 , all Pj
Proof. a) Assume that P degrees sj , 1 j N sj = s. The product of the leading coecients of P1 , . . . , PN in n1 is equal to 1; after normalizing these polynomials, we may assume that P1 , . . . , PN are unitary and sj > 0 for all j. Then
s P (0, zn ) = zn = P1 (0, zn ) . . . PN (0, zn ),
n if and only if it is irreducible in n1 [zn ]. = P1 . . . PN with polynomials Pj n1 [zn ] of respective
hence Pj (0, zn ) = znj and therefore Pj is a Weierstrass polynomial.

s b) Set s = deg P and P (0, zn ) = zn . Assume that P is reducible in n , with P (z , zn ) = g1 (z)g2 (z) for non invertible elements g1 , g2 n . Then g1 (0, zn ) and g2 (0, zn ) have vanishing orders s1 , s2 > 0 with s1 + s2 = s, and
g j = uj P j ,
deg Pj = sj ,
j = 1, 2,
where Pj is a Weierstrass polynomial and uj n is invertible. Therefore P1 P2 = uP for an invertible germ u n . Lemma 2.8 shows that P divides P1 P2 in n1 [zn ] ; since P1 , P2 are unitary and s = s1 + s2 , we get P = P1 P2 , hence P is reducible in n1 [zn ]. The converse implication is obvious from a). (2.10) Theorem. a) every non elements ;
n is a factorial ring, i.e. n is entire and: zero germ f n admits a factorization f =
f1 . . . fN in irreducible
b) the factorization is unique up to invertible elements. Proof. The existence part a) follows from Lemma 2.9 if we take f to be a Weierstrass polynomial and f = f1 . . . fN be a decomposition of maximal length N into polynomials of positive degree. In order to prove the uniqueness, it is sucient to verify the following statement:
3. Coherent Sheaves
83
bIf g is an irreducible element that divides a product f1 f2 , then g divides either f1 or ) f2 .
By Th. 2.1, we may assume that f1 , f2 , g are Weierstrass polynomials in zn . Then g is irreducible and divides f1 f2 in n1 [zn ] thanks to Lemmas 2.8 and 2.9 b). By induction on n, we may assume that n1 is factorial. The standard Gauss lemma ([Lang 1965], Chapter V) says that the polynomial ring A[T ] is factorial if the ring A is factorial. Hence n1 [zn ] is factorial by induction and thus g must divide f1 or f2 in n1 [zn ]. (2.11) Theorem. If f, g n are relatively prime, then the germs fz , gz at every point z Cn near 0 are again relatively prime. Proof. One may assume that f = P, g = Q are Weierstrass polynomials. Let us recall that unitary polynomials P, Q [X] ( = a factorial ring) are relatively prime if and only if their resultant R is non zero. Then the resultant R(z ) n1 of P (z , zn ) and Q(z , zn ) has a non zero germ at 0. Therefore the germ Rz at points z Cn1 near 0 is also non zero.
3. Coherent Sheaves
3.1. Locally Free Sheaves and Vector Bundles Section 9 will greatly develope this philosophy. Before introducing the more general notion of a coherent sheaf, we discuss the notion of locally free sheaves over a sheaf a ring. All rings occurring in the sequel are supposed to be commutative with unit (the non commutative case is also of considerable interest, e.g. in view of the theory of -modules, but this subject is beyond the scope of the present book). (3.1) Denition. Let be a sheaf of rings on a topological space X and let a sheaf (or briey a -module). Then is said to be locally free of rank r of modules over over , if is locally isomorphic to r on a neighborhood of every point, i.e. for every x0 X one can nd a neighborhood and sections F1 , . . . , Fr () such that the sheaf homomorphism F :
r
r x
(w1 , . . . , wr )
1 j r
wj Fj,x x
is an isomorphism. By denition, if is locally free, there is a covering (U )I by open sets on which 1 r admits free generators F , . . . , F (U). Because the generators can be uniquely expressed in terms of any other system of independent generators, there is for each pair (, ) a r r matrix G = (Gjk )1 such that
k F = 1 j r j,k r ,
Gjk
(U U ),
j F Gjk
on U U .
84
In other words, we have a commutative diagram

r U U
F U U U U F
r U U
1 It follows easily from the equality G = F F that the transition matrices G are invertible matrices satisfying the transition relation
(3.2)
G = G G
on U U U
The notion of locally free sheaf is closely related to another essential notion of differential geometry, namely the notion of vector bundle (resp. topological, dierentiable, holomorphic . . ., vector bundle). To describe the relation between these notions, we asis a subsheaf of the sheaf K of continous functions on sume that the sheaf of rings X with values in the eld K = R or K = C, containing the sheaf of locally constant functions X K. Then, for each x X, there is an evaluation map
x
Conversely, if we are given a system of invertible r r matrices G with coecients (U U ) satisfying the transition relation (3.2), we can dene a locally free sheaf of rank r over by taking r over each U , the identication over U U being given by the isomorphism G . A section H of over an open set X can 1 r just be seen as a collection of sections H = (H , . . . , H ) of r ( U ) satisfying the transition relations H = G H over U U . in
for all indices , , I. In particular G = Id on U and G1 = G on U U .
whose kernel is a maximal ideal mx of x , and x /mx = K. Let be a locally free sheaf of rank r over . To each x X, we can associate a K-vector space Ex = x /mx x : since x r , we have Ex ( x /mx )r = Kr . The set E = xX Ex is equipped x with a natural projection : E X, Ex () := x,
K,
w w(x)
and the bers Ex = 1 (x) have a structure of r-dimensional K-vector space: such a structure E is called a K-vector bundle of rank r over X. Every section s (U ) gives rise to a section of E over U by setting s(x) = sx mod mx . We obtain a function (still denoted by the same symbol) s : U E such that s(x) Ex for every x U , i.e. s = IdU . It is clear that (U ) can be considered as a (U )-submodule of the K-vector space of functions U E mapping a point x U to an element in the ber Ex . Thus we get a subsheaf of the sheaf of E-valued sections, which is in a natural way a -module isomorphic to . This subsheaf will be denoted by (E) and will be called the sheaf of -sections of E. If we are given a K-vector bundle E over X and a subsheaf = (E) of the sheaf of all sections of E which is in a natural way a locally free -module of rank r, we say that E (or more precisely the pair (E, (E))) is a -vector bundle of rank r over X.
3. Coherent Sheaves
85
= X,K is the sheaf of all K-valued continuous functions (3.3) Example. In case on X, we say that E is a topological vector bundle over X. When X is a manifold and = p , we say that E is a C p -dierentiable vector bundle; nally, when X is complex X,K analytic and = X , we say that E is a holomorphic vector bundle. Let us introduce still a little more notation. Since (E) is a locally free sheaf of rank 1 r r over any open set U in a suitable covering of X, a choice of generators (F , . . . , F ) for (E)U yields corresponding generators (e1 (x), . . . , er (x)) of the bers Ex over K. Such a system of generators is called a -admissible frame of E over U . There is a corresponding isomorphism (3.4) : EU := 1 (U ) U Kr
1 r which to each Ex associates the pair (x, ( , . . . , )) U Kr composed of x j and of the components ( )1 j r of in the basis (e1 (x), . . . , er (x)) of Ex . The bundle E is said to be trivial if it is of the form X Kr , which is the same as saying that (E) = r . For this reason, the isomorphisms are called trivializations of E over U . The corresponding transition automorphisms are 1 := : (U U ) Kr (U U ) Kr ,
(x, ) = (x, g (x) ),
(x, ) (U U ) Kr ,
where (g ) GLr ( )(U U ) are the transition matrices already described (except that they are just seen as matrices with coecients in K rather than with coecients in jk a sheaf). Conversely, if we are given a collection of matrices g = (g ) GLr ( )(U U ) satisfying the transition relation g = g g we can dene a -vector bundle E=
I
on U U U ,
U Kr /
by gluing the charts U Kr via the identication (x , ) (x , ) if and only if x = x = x U U and = g (x) . (3.5) Example. When X is a real dierentiable manifold, an interesting example of real vector bundle is the tangent bundle TX ; if : U Rn is a collection of coordinate charts on X, then = d : TXU U Rm dene trivializations of TX and the 1 transition matrices are given by g (x) = d (x ) where = and x = (x). The dual TX of TX is called the cotangent bundle of X. If X is complex analytic, then TX has the structure of a holomorphic vector bundle. We now briey discuss the concept of sheaf and bundle morphisms. If and are sheaves of -modules over a topological space X, then by a morphism : we just mean a -linear sheaf morphism. If = (E) and = (E ) are locally free sheaves, this is the same as a -linear bundle morphism, that is, a ber preserving
86
K-linear morphism (x) : Ex Ex such that the matrix representing in any local -admissible frames of E and E has coecients in .
(3.6) Proposition. Suppose that is a sheaf of local rings, i.e. that a section of is invertible in if and only if it never takes the zero value in K. Let : be a -morphism of locally free -modules of rank r, r . If the rank of the r r matrix (x) Mr r (K) is constant for all x X, then Ker and Im are locally free subsheaves of , respectively, and Coker = / Im is locally free. Proof. This is just a consequence of elementary linear algebra, once we know that non zero determinants with coecients in can be inverted. Note that all three sheaves X,K , p , X are sheaves of local rings, so Prop. 3.6 X,K applies to these cases. However, even if we work in the holomorphic category ( = X ), a diculty immediately appears that the kernel or cokernel of an arbitrary morphism of locally free sheaves is in general not locally free. (3.7) Examples. a) Take X = C, let = = be the trivial sheaf, and let : be the morphism u(z) z u(z). It is immediately seen that is injective as a sheaf morphism ( being an entire ring), and that Coker is the skyscraper sheaf C0 of stalk C at z = 0, having zero stalks at all other points z = 0. Thus Coker is not a locally free sheaf, although is everywhere injective (note however that the corresponding morphism : E E , (z, ) (z, z) of trivial rank 1 vector bundles E = E = C C is not injective on the zero ber E0 ). b) Take X = C3 ,
= 3, = and
: 3 , (u1 , u2 , u3 ) zj uj (z1 , z2 , z3 ).
1 j 3
Since yields a surjective bundle morphism on C3 {0}, one easily sees that Ker is locally free of rank 2 over C3 {0}. However, by looking at the Taylor expansion of the uj s at 0, it is not dicult to check that Ker is the -submodule of 3 generated by the three sections (z2 , z1 , 0), (z3 , 0, z1 ) and (0, z3 , z2 ), and that any two of these three sections cannot generate the 0-stalk (Ker )0 . Hence Ker is not locally free. Since the category of locally free -modules is not stable by taking kernels or cokernels, one is led to introduce a more general category which will be stable under these operations. This leads to the notion of coherent sheaves. 3.2. Notion of Coherence The notion of coherence again deals with sheaves of modules over a sheaf of rings. It is a semi-local property which says roughly that the sheaf of modules locally has a nite presentation in terms of generators and relations. We describe here some general properties of this notion, before concentrating ourselves on the case of coherent X modules.
3. Coherent Sheaves
87
be a sheaf of rings on a topological space X and a sheaf of (3.8) Denition. Let (or briey a -module). Then is said to be locally nitely generated modules over if for every point x0 X one can nd a neighborhood and sections F1 , . . . , Fq () such that for every x the stalk x is generated by the germs F1,x , . . . , Fq,x as an x -module. (3.9) Lemma. Let be a locally nitely generated sheaf of -modules on X and G1 , . . . , GN sections in (U ) such that G1,x0 , . . . , GN,x0 generate x0 at x0 U . Then G1,x , . . . , GN,x generate x for x near x0 . Proof. Take F1 , . . . , Fq as in Def. 3.8. As G1 , . . . , GN generate x0 , one can nd a Hjk Gk on . Thus neighborhood of x0 and Hjk ( ) such that Fj = G1,x , . . . , GN,x generate x for all x . 3.2.1. Denition of Coherent Sheaves. If U is an open subset of X, we denote by U the restriction of to U , i.e. the union of all stalks x for x U . If F1 , . . . , Fq (U ), the kernel of the sheaf homomorphism F : q U dened by U (3.10)
q x
(g 1 , . . . , g q )
q U ,
1 j q
g j Fj,x x ,
xU
is a subsheaf
(F1, . . . , Fq ) of
called the sheaf of relations between F1 , . . . , Fq . -modules on X is said to be coherent if:
(3.11) Denition. A sheaf a)
of
is locally nitely generated ; (U ),

the sheaf of relations
b) for any open subset U of X and any F1 , . . . , Fq (F1 , . . . , Fq ) is locally nitely generated.
Assumption a) means that every point x X has a neighborhood such that there is a surjective sheaf morphism F : q , and assumption b) implies that the kernel of F is locally nitely generated. Thus, after shrinking , we see that admits over a nite presentation under the form of an exact sequence (3.12)
p
0,
F
where G is given by a q p matrix (Gjk ) of sections of (Gjp ) are generators of (F1 , . . . , Fq ).
() whose columns (Gj1 ), . . . ,
It is clear that every locally nitely generated subsheaf of a coherent sheaf is coherent. From this we easily infer: (3.13) Theorem. Let : and ker are coherent. be a -morphism of coherent sheaves. Then Im
Proof. Clearly Im is a locally nitely generated subsheaf of , so it is coherent. Let x0 X, let F1 , . . . , Fq () be generators of on a neighborhood of x0 , and
88
G1 , . . . , Gr ( )q be generators of (F1 ), . . . , (Fq ) on a neighborhood of x0 . Then ker is generated over by the sections
q
Hj =
k=1
Gk Fk j
( ),
r.
are coherent, then = ker( ) is coherent by Th. 3.13. If Proof. If and and are coherent, then is locally nitely generated; to prove the coherence, let G1 , . . . , Gq (U ) and x0 U . Then there is a neighborhood of x0 and sections G1 , . . . , Gq () which are mapped to G1 , . . . , Gq on . After shrinking , we may assume also that is generated by sections F1 , . . . , Fp (). Then (G1 , . . . , Gq ) is the projection on the last q-components of (F1 , . . . , Fp , G1 , . . . , Gq ) p+q , which is nitely generated near x0 by the coherence of . Hence (G1 , . . . , Gq ) is nitely generated near x0 and is coherent. Finally, assume that and are coherent. Let x0 X be any point, let F1 , . . . , Fp () and G1 , . . . , Gq () be generators of , on a neighborhood of x0 . There is a neighborhood of x0 such that G1 , . . . , Gq admit liftings G1 , . . . , Gq ( ). Then (F1 , . . . , Fq , G1 , . . . , Gq ) generate , so is locally nitely generated. Now, let S1 , . . . , Sq be arbitrary sections in (U ) and S 1 , . . . , S q their images in (U ). For any x0 U , the sheaf of relations (S 1 , . . . , S q ) is generated by sections P1 , . . . , Ps ()q on a small neighborhood of x0 . Set Pj = (Pjk )1 k q . Then Hj = Pj1 S1 + . . . + Pjq Sq , 1 j s, are mapped to 0 in so they can be seen as sections of . The coherence of shows that (H1 , . . . , Hs ) has generators Q1 , . . . , Qt ( )s on a small neighborhood of x0 . Then (S1 , . . . , Sq ) is generated over by Rj = Qk Pk ( ), j 1 j t, and is coherent. (3.15) Corollary. If and are coherent subsheaves of a coherent analytic sheaf the intersection is a coherent sheaf. Proof. Indeed, the intersection sheaf / , and / is coherent.
(3.14) Theorem. Let 0 0 be an exact sequence of If two of the sheaves , , are coherent, then all three are coherent.
-modules.
is the kernel of the composite morphism
3.2.2. Coherent Sheaf of Rings. A sheaf of rings is said to be coherent if it is coherent as a module over itself. By Def. 3.11, this means that for any open set U X and any sections Fj (U ), the sheaf of relations (F1 , . . . , Fq ) is nitely generated. The above results then imply that all free modules p are coherent. As a consequence: (3.16) Theorem. If is a coherent sheaf of rings, any locally nitely generated subsheaf of p is coherent. In particular, if is a coherent -module and F1 , . . . , Fq (U ), the sheaf of relations (F1 , . . . , Fq ) q is also coherent. Let be a coherent sheaf of modules over a coherent sheaf of ring . By an iteration of construction (3.12), we see that for every integer m 0 and every point x X there is a neighborhood of x on which there is an exact sequence of sheaves (3.17)
pm
m
pm1
p1
p0
0 0,
3. Coherent Sheaves
89
where Fj is given by a pj1 pj matrix of sections in 3.3. Analytic Sheaves and the Oka Theorem
().
Many properties of holomorphic functions which will be considered in this book can be expressed in terms of sheaves. Among them, analytic sheaves play a central role. The Oka theorem [Oka 1950] asserting the coherence of the sheaf of holomorphic functions can be seen as a far-reaching deepening of the noetherian property seen in Sect. 1. The theory of analytic sheaves could not be presented without it. (3.18) Denition. Let M be a n-dimensional complex analytic manifold and let M be the sheaf of germs of analytic functions on M . An analytic sheaf over M is by denition a sheaf of modules over M . (3.19) Coherence theorem of Oka. The sheaf of rings M is coherent for any complex manifold M . Let F1 , . . . , Fq (U ). Since M,x is Noetherian, we already know that every stalk (F1, . . . , Fq )x q is nitely generated, but the important new fact expressed by M,x the theorem is that the sheaf of relations is locally nitely generated, namely that the same generators can be chosen to generate each stalk in a neighborhood of a given point.
Proof. By induction on n = dimC M . For n = 0, the stalks M,x are equal to C and the result is trivial. Assume now that n 1 and that the result has already been proved in dimension n 1. Let U be an open set of M and F1 , . . . , Fq M (U ). To show that (F1, . . . , Fq ) is locally nitely generated, we may assume that U = = n is a polydisk in Cn centered at x0 = 0 ; after a change of coordinates and multiplication of F1 , . . . , Fq by invertible functions, we may also suppose that F1 , . . . , Fq are Weierstrass polynomials in zn with coecients in ( ). We need a lemma. (3.20) Lemma. If x = (x , xn ) , the ,x -module (F1 , . . . , Fq )x is generated by those of its elements whose components are germs of analytic polynomials in ,x [zn ] with a degree in zn at most equal to , the maximum of the degrees of F1 , . . . , Fq . Proof. Assume for example that Fq is of the maximum degree . By the Weierstrass preparation Th. 1.1 and Lemma 1.9 applied at x, we can write Fq,x = f f where f , f ,x [zn ], f is a Weierstrass polynomial in zn xn and f (x) = 0. Let and denote the degrees of f and f with respect to zn , so + = . Now, take (g 1 , . . . , g q ) (F1 , . . . , Fq )x . The Weierstrass division theorem gives where t rq = g q + (3.21)
j
,x and r
1 j q1 t j
g j = Fq,xtj + r j ,

Fj,x . We can write

1 j q
,x [zn ] is a polynomial of degree < .
j = 1, . . . , q 1,
For j = q, dene
(g 1 , . . . , g q ) =
tj (0, . . . , Fq , . . . , 0, Fj )x + (r 1 , . . . , rq )
where Fq is in the j-th position in the q-tuples of the summation. Since these q-tuples are in (F1 , . . . , Fq )x , we have (r 1 , . . . , rq ) (F1 , . . . , Fq )x , thus Fj,x r j + f f r q = 0.
1 j q1
90
As the sum is a polynomial in zn of degree < + , it follows from Lemma 1.9 that f r q is a polynomial in zn of degree < . Now we have (r 1 , . . . , rq ) = 1/f (f r 1 , . . . , f r q ) where f r j is of degree < + = . In combination with (3.21) this proves the lemma. Proof of Theorem 3.19 (end).. If g = (g 1 , . . . , g q ) is one of the polynomials of (F1 , . . . , Fq )x described in Lemma 3.20, we can write gj =
0 k k ujk zn ,
ujk ,x .
The condition for (g 1 , . . . , g q ) to belong to (F1 , . . . , Fq )x therefore consists of 2 + 1 linear conditions for the germ u = (ujk ) with coecients in ( ). By the induction hypothesis, is coherent and Th. 3.16 shows that the corresponding modules of relations are generated over ,x , for x in a neighborhood of 0, by nitely many (q )-tuples U1 , . . . , UN ( )q . By Lemma 3.20, (F1 , . . . , Fq )x is generated at every point x = n by the germs of the corresponding polynomials Gl (z) =
0 k k Uljk (z )zn 1 j q
z ,
N.
(3.22) Strong Noetherian property. Let be a coherent analytic sheaf on a complex manifold M and let 1 2 . . . be an increasing sequence of coherent subsheaves of . Then the sequence ( k ) is stationary on every compact subset of M . Proof. Since is locally a quotient of a free module q , we can pull back the sequence M = M (by easy reductions similar to those in the proof of to q and thus suppose M Th. 3.14). Suppose M connected and k0 = {0} for some index k0 (otherwise, there is nothing to prove). By the analytic continuation theorem, we easily see that k0 ,x = {0} for every x M . We can thus nd a non zero Weierstrass polynomial P k0 (V ), degzn P (z , zn ) = , in a coordinate neighborhood V = n of any point x M . A division by P shows that for k k0 and x V , all stalks k,x are generated by Px and by polynomials of degree < in zn with coecients in ,x . Therefore, we can apply induction on n to the coherent -module
Q [zn ] ; deg Q
k
and its increasing sequence of coherent subsheaves
4. Complex Analytic Sets. Local Properties

4.1. Denition. Irreducible Components A complex analytic set is a set which can be dened locally by nitely many holomorphic equations; such a set has in general singular points, because no assumption is
91
made on the dierentials of the equations. We are interested both in the description of the singularities and in the study of algebraic properties of holomorphic functions on analytic sets. For a more detailed study than ours, we refer to H. Cartans seminar [Cartan 1950], to the books of [Gunning-Rossi 1965], [Narasimhan 1966] or the recent book by [Grauert-Remmert 1984]. (4.1) Denition. Let M be a complex analytic manifold. A subset A M is said to be an analytic subset of M if A is closed and if for every point x0 A there exist a neighborhood U of x0 and holomorphic functions g1 , . . . , gn in (U ) such that A U = {z U ; g1 (z) = . . . = gN (z) = 0}. Then g1 , . . . , gN are said to be (local) equations of A in U .
It is easy to see that a nite union or intersection of analytic sets is analytic: if (gj ), are equations of A , A in the open set U , then the family of all products (gj gk ) and the family (gj ) (gk ) dene equations of A A and A A respectively. (gk )
(4.2) Remark. Assume that M is connected. The analytic continuation theorem shows that either A = M or A has no interior point. In the latter case, each piece AU = g 1 (0) is the set of points where the function log |g|2 = log(|g1 |2 + + |gN |2 ) Psh(U ) takes the value , hence A is pluripolar. In particular M A is connected and every function f (M A) that is locally bounded near A can be extended to a function f (M ). We focus now our attention on local properties of analytic sets. By denition, a germ of set at a point x M is an equivalence class of elements in the power set (M ), with A B if there is an open neighborhood V of x such that A V = B V . The germ of a subset A M at x will be denoted by (A, x). We most often consider the case when A M is a analytic set in a neighborhood U of x, and in this case we denote by A,x the ideal of germs f M,x which vanish on (A, x). Conversely, if = (g1, . . . , gN ) is an ideal of M,x , we denote by V (), x the germ at x of the zero variety V () = {z U ; g1 (z) = . . . = gN (z) = 0}, where U is a neighborhood of x such that gj (U ). It is easy to check that the germ (V (cJ), x) does not depend on the choice of generators. Moreover, it is clear that (4.3 ) (4.3 ) for every ideal
in the ring M,x ,
for every germ of analytic set (A, x),
V (),x , V (A,x ), x = (A, x). A,x is a prime ideal of
(4.4) Denition. A germ (A, x) is said to be irreducible if it has no decomposition (A, x) = (A1 , x) (A2 , x) with analytic sets (Aj , x) = (A, x), j = 1, 2. (4.5) Proposition. A germ (A, x) is irreducible if and only if the ring M,x .
Proof. Let us recall that an ideal is said to be prime if f g implies f or g . Assume that (A, x) is irreducible and that f g A,x . As we can write (A, x) = (A1 , x) (A2 , x) with A1 = A f 1 (0) and A2 = A g 1 (0), we must have
92
for example (A1 , x) = (A, x) ; thus f A,x and A,x is prime. Conversely, if (A, x) = (A1 , x) (A2 , x) with (Aj , x) = (A, x), there exist f A1 ,x , g A2 ,x such that f, g A,x . However f g A,x , thus A,x is not prime. / (4.6) Theorem. Every decreasing sequence of germs of analytic sets (Ak , x) is stationary. Proof. In fact, the corresponding sequence of ideals k = A ,x is increasing, thus k = k for k k0 large enough by the Noetherian property of M,x . Hence (Ak , x) = V (k ), x is constant for k k0 . This result has the following straightforward consek 0
quence:
(4.7) Theorem. Every analytic germ (A, x) has a nite decomposition (A, x) =
1 k N
(Ak , x)
where the germs (Aj , x) are irreducible and (Aj , x) (Ak , x) for j = k. The decomposition is unique apart from the ordering. Proof. If (A, x) can be split in several components, we split repeatedly each component as long as one of them is reducible. The process must stop by Th. 4.6, whence the existence. For the uniqueness, assume that (A, x) = (A , x), 1 l N , is another l decomposition. Since (Ak , x) = l (Ak A , x), we must have (Ak , x) = (Ak A , x) for l l some l = l(k), i.e. (Ak , x) (A , x), and likewise (A , x) (Aj , x) for some j. Hence l(k) l(k) j = k and (A , x) = (Ak , x). l(k) 4.2. Local Structure of a Germ of Analytic Set We are going to describe the local structure of a germ, both from the holomorphic and topological points of view. By the above decomposition theorem, we may restrict ourselves to the case of irreducible germs Let be a prime ideal of n = Cn ,0 and let A = V () be its zero variety. We set k = C{z1 , . . . , zk } for each k = 0, 1, . . . , n. (4.8) Proposition. There exist an integer d, a basis (e1 , . . . , en ) of Cn and associated coordinates (z1 , . . . , zn ) with the following properties: d = {0} and for every integer k = d + 1, . . . , n there is a Weierstrass polynomial Pk k of the form (4.9)
s Pk (z , zk ) = zkk + 1 j sk s aj,k (z ) zkk j ,
aj,k (z ) k1 ,
where aj,k (z ) = O(|z |j ). Moreover, the basis (e1 , . . . , en ) can be chosen arbitrarily close to any preassigned basis (e0 , . . . , e0 ). 1 n Proof. By induction on n. If = n = {0}, then d = n and there is nothing to prove. Otherwise, select a non zero element gn and a vector en such that C w gn (wen ) has minimum vanishing order sn . This choice excludes at most the algebraic (s ) set gn n (v) = 0, so en can be taken arbitrarily close to e0 . Let (1 , . . . , zn1 , zn ) be z n the coordinates associated to the basis (e0 , . . . , e0 , en ). After multiplication by an 1 n1 invertible element, we may assume that gn is a Weierstrass polynomial
s Pn (, zn ) = znn + z 1 j sn
aj,n () znn j , z s
aj,n n1 ,
93
z and aj,n () = O(||j ) by Remark 2.2. If n1 = C{} = {0} then d = n 1 and z z the construction is nished. Otherwise we apply the induction hypothesis to the ideal n1 n1 in order to nd a new basis (e1 , . . . , en1 ) of Vect(e0, . . . , e0 ), associated 1 n1 coordinates (z1 , . . . , zn1 ) and Weierstrass polynomials Pk k , d + 1 k n 1, as stated in the lemma. (4.10) Lemma. If w C is a root of wd + a1 wd1 + + ad = 0, aj C, then |w| 2 max |aj |1/j . Proof. Otherwise |w| > 2|aj |1/j for all j = 1, . . . , d and the given equation 1 = a1 /w + + ad /wd implies 1 21 + + 2d , a contradiction. (4.11) Corollary. Set z = (z1 , . . . , zd ), z = (zd+1 , . . . , zn ), and let in Cd , in Cnd be polydisks of center 0 and radii r , r > 0. Then the germ (A, 0) is contained in a cone |z | C|z |, C = constant, and the restriction of the projection map Cn Cd , (z , z ) z : : A ( ) is proper if r is small enough and r r /C. Proof. The polynomials Pk (z1 , . . . , zk1 ; zk ) vanish on (A, 0). By Lemma 4.10 and (4.9), every point z A suciently close to 0 satises |zk | thus |z | Since (4.12) (4.13) Proposition. Ck (|z1 | + + |zk1 |), d+1 k n,
d = {0}, we have an injective ring morphism
C|z | and the Corollary follows.
Proof. Let f n . A division by Pn yields f = Pn qn + Rn with a remainder Rn n1 [zn ], degzn Rn < sn . Further divisions of the coecients of Rn by Pn1 , Pn2 etc . . . yield Rk+1 = Pk qk + Rk , Rk k [zk+1 , . . . , zn ], where degzj Rk < sj for j > k. Hence (4.14) f = Rd +
d+1 k n
n / is a nite integral extension of d.
d = C{z1 , . . . , zd} n /.
Pk qk = Rd
mod (Pd+1 , . . . , Pn )
d+1 and n / is nitely generated as an d -module by the family of monomials zd+1 . . . zn n with j < sj .
As is prime, n / is an entire ring. We denote by f the class of any germ f n in n /, by A and d the quotient elds of n / and d respectively. Then A = d [d+1 , . . . , zn ] is a nite algebraic extension of d . Let q = [A :d ] be its degree and z let 1 , . . . , q be the embeddings of A over d in an algebraic closure A . Let us recall that a factorial ring is integrally closed in its quotient eld ([Lang 1965], Chapter IX).
94
Hence every element of d which is integral over d lies in fact in d . By the primitive element theorem, there exists a linear form u(z ) = cd+1 zd+1 + + cn zn , ck C, such u that A = d [] ; in fact, u is of degree q if and only if 1 u, . . . , q u are all distinct, and this excludes at most a nite number of vector subspaces in the space Cnd of coecients (cd+1 , . . . , cn ). As u n / is integral over the integrally closed ring d , the unitary irreducible polynomial Wu of u over d has coecients in d : Wu (z ; T ) = T q +
1 j q
aj (z1 , . . . , zd ) T qj ,
aj d .
Wu must be a Weierstrass polynomial, otherwise there would exist a factorization Wu = W Q in d [T ] with a Weierstrass polynomial W of degree deg W < q = deg u and Q(0) = 0, hence W () = 0, a contradiction. In the same way, we see that zd+1 , . . . , zn u have irreducible equations Wk (z ; zk ) = 0 where Wk d [T ] is a Weierstrass polynomial of degree = deg zk q, d + 1 k n. (4.15) Lemma. Let (z ) d be the discriminant of Wu (z ; T ). For every element g of A which is integral over d (or equivalently over n /) we have g d []. u Proof. We have (z ) = j<k (k u g=
0 j q1
u j u)2 0 , and g A = d [] can be written b j uj , bj d ,
where b0 , . . . , bd1 are the solutions of the linear system k g = bj (k u)j ; the deter 1/2 minant (of Van der Monde type) is . It follows that bj d are polynomials in k g and k u, thus bj is integral over d . As d is integrally closed, we must have bj d , hence g d []. u In particular, there exist unique polynomials Bd+1 , . . ., Bn q 1, such that (4.16) Then we have (4.17) (z )q Wk z ; Bk (z ; T )/(z ) ideal Wu (z ; T ) d [T ] ; (z )zk = Bk (z ; u(z )) (mod
d[T ] with deg Bk
).
indeed, the left-hand side is a polynomial in d [T ] and admits T = u as a root in since Bk (z ; u)/(z ) = zk and Wk (z ; zk ) = 0. (4.18) Lemma. Consider the ideal = Wu (z ; u(z )) , (z )zk Bk (z ; u(z )) and set m = max{q, (n d)(q 1)}. For every germ f polynomial R d [T ], degT R q 1, such that (z )m f (z) = R(z ; u(z )) (mod
n /
n ,
there exists a unique
).
95
Moreover f implies R = 0, hence m
. . The analogue
Proof. By (4.17) and a substitution of zk , we nd (z )q Wk (z ; zk ) of formula (4.14) with Wk in place of Pk yields f = Rd +

d+1 k n
Wk qk ,
Rd d [zd+1 , . . . , zn ],
with degzk Rd < deg Wk q, thus m f = m Rd mod . We may therefore replace f by Rd and assume that f d [zd+1 , . . . , zn ] is a polynomial of total degree (nd)(q1) m. A substitution of zk by Bk (z ; u(z ))/(z ) yields G d [T ] such that Finally, a division G = Wu Q + R gives the required polynomial R d [T ]. The last statement is clear: if f satises m (z )f (z) = R(z ; u(z )) mod for degT R < q, then R(z ; u) = 0, and as u n / is of degree q, we must have R = 0. The uniqueness of R is proved similarly. (4.19) Local parametrization theorem. Let V (). Assume that the coordinates (z )m f (z) = G(z ; u(z )) mod (z )zk Bk (z ; u(z )) .
be a prime ideal of n and let A =
are chosen as above. Then the ring n / is a nite integral extension of d ; let q be the degree of the extension and let (z ) d be the discriminant of the irreducible polynomial of a primitive element u(z ) = k>d ck zk . If , are polydisks of suciently small radii r , r and if r r /C with C large, the projection map : A ( ) is a ramied covering with q sheets, whose ramication locus is contained in S = {z ; (z ) = 0}. This means that: a) the open subset AS = A ( in A ( ) ; S) is a smooth d-dimensional manifold, dense
(z ; z ) = (z1 , . . . , zd ; zd+1 , . . . , zn )
b) : AS
S is a covering ; S, and A ( ) is contained in a
Moreover, AS is a connected covering of cone |z | C|z | (see Fig. 1). z Cnp
c) the bers 1 (z ) have exactly q elements if z S and at most q if z S. /
0 A
z Cp
Fig. II-1 Ramied covering from A to Cp .
96
Proof. After a linear change in the coordinates zd+1 , . . . , zn , we may assume u(z ) = zd+1 , so Wu = Wd+1 and Bd+1 (z ; T ) = (z )T . By Lemma 4.18, we have = Wd+1 (z , zd+1 ) , (z )zk Bk (z , zd+1 )
k d+2
We can thus nd a polydisk = of suciently small radii r , r such that V () V ( ) V ( m ) in . As V () = A and V () = S , this implies A V ( ) (A ) (S ). In particular, the set AS = A ( S) lying above S coincides with V ( ) ( S) , which is the set of points z parametrized by the equations (4.20) (z ) = 0, Wd+1 (z , zd+1 ) = 0, zk = Bk (z , zd+1 )/(z ), d + 2 k n.
As (z ) is the resultant of Wd+1 and Wd+1 /T , we have Wd+1 /T (z , zd+1 ) = 0 on AS .
The implicit function theorem shows that zd+1 is locally a holomorphic function of z on AS , and the same is true for zk = Bk (z , zd+1 )/(z ), k d + 2. Hence AS is a smooth manifold, and for r r /C small, the projection map : AS S is a proper local dieomorphism; by (4.20) the bers 1 (z ) have at most q points corresponding to some of the q roots w of Wd+1 (z ; w) = 0. Since S is connected (Remark 4.2), either AS = or the map is a covering of constant sheet number q q. However, if w is a root of Wd+1 (z , w) = 0 with z S and if we set zd+1 = w, zk = Bk (z , w)/(z ), k d + 2, relation (4.17) shows that Wk (z , zk ) = 0, in particular |zk | = O(|z |1/q ) by Lemma 4.10. For z small enough, the q points z = (z , z ) dened in this way lie in , thus q = q. Property b) and the rst parts of a) and c) follow. Now, we need the following lemma.
n is prime and A = V (), then A,0 = . I.t is obvious that A,0 . Now, for any f A,0 , Prop. 4.13 implies that f satises in n / an irreducible equation f r + b1 (z ) f r1 + + br (z ) = 0 (mod ).
(4.21) Lemma. If Then br (z ) vanishes on (A, 0) and the rst part of c) gives br = 0 on S. Hence r = 0 and the irreducibility of the equation of f implies r = 1, so f , as desired. b Proof of Theorem 4.19 (end).. It only remains to prove that AS is connected and dense in A and that the bers 1 (z ), z S, have at most q elements. Let AS,1 , . . . , AS,N be the connected components of AS . Then : AS,j S is a covering with qj sheets, q1 + + qN = q. For every point S, there exists a neighborhood U of such that AS,j 1 (U ) is a disjoint union of graphs z = gj,k (z ) of analytic k qj . If (z ) is an arbitrary linear form in zd+1 , . . . , zn functions gj,k (U ), 1 and z S, we set P,j (z ; T ) =
{z ; (z ,z )AS,j }
T (z ) =
1 k kj
T gj,k (z ) .
97
This denes a polynomial in T with bounded analytic coecients on S. These coecients have analytic extensions to (Remark 4.2), thus P,j ( )[T ]. By construction, P,j z ; (z ) vanishes identically on AS,j . Set P =
1 j N
P,j ,
f (z) = (z ) P z ; (z ) ;
f vanishes on AS,1 . . . AS,N (S ) A . Lemma 4.21 shows that A,0 is prime; as A,0 , we get P,j z ; (z ) A,0 for some j. This is a contradiction if / N 2 and if is chosen in such a way that separates the q points z in each ber 1 (z ), for a sequence z 0 in S. Hence N = 1, AS is connected, and for every nd (C ) we have P z , (z ) (A,0) . By construction P z , (z ) vanishes on AS , so it vanishes on AS ; hence, for every z S, the ber AS 1 (z ) has at most q elements, otherwise selecting which separates q + 1 of these points would yield q + 1 roots (z ) of P (z ; T ), a contradiction. Assume now that AS is not dense in A for arbitrarily small polydisks . Then there exists a sequence A z = (z , z ) 0 such that z S and z is not in F := pr AS 1 (z ) . The continuity of the roots of the polynomial P (z ; T ) as S z z implies that the set of roots of P (z ; T ) is (F ). Select such that (z ) (F ) for all . Then P z ; (z ) = 0 for every / and P z ; (z ) A,0 , a contradiction. / At this point, it should be observed that many of the above statements completely fail in the case of real analytic sets. In R2 , for example, the prime ideal = (x5 + y 4 ) denes an irreducible germ of curve (A, 0) and there is an injective integral extension of rings R{x} R{x, y}/ of degree 4; however, the projection of (A, 0) on the rst factor, (x, y) x, has not a constant sheet number near 0, and this number is not related to the degree of the extension. Also, the prime ideal = (x2 + y 2 ) has an associated variety V () reduced to {0}, hence A,0 = (x, y) is strictly larger than , in contrast with Lemma 4.21. Let us return to the complex situation, which is much better behaved. The result obtained in Lemma 4.21 can then be extended to non prime ideals and we get the following important result: (4.22) Hilberts Nullstellensatz. For every ideal
V (),0 = ,
Proof. Set B = V (). If f k , then f k vanishes on (B, 0) and f B,0 . Thus B,0 . Conversely, it is well known that is the intersection of all prime ideals ([Lang 1965], Chapter VI). For such an ideal (B, 0) = V (), 0) V), 0 , ( thus B,0 V (),0 = in view of Lemma 4.21. Therefore B,0 = and the Theorem is proved. In other words, if a germ (B, 0) is dened by an arbitrary ideal n and if f n vanishes on (B, 0), then some power f k lies in . where in .
is the radical of , i.e. the set of germs f n such that some power f k lies
98
4.3. Regular and Singular Points. Dimension The above powerful results enable us to investigate the structure of singularities of an analytic set. We rst give a few denitions. (4.23) Denition. Let A M be an analytic set and x A. We say that x A is a regular point of A if A is a C-analytic submanifold of for some neighborhood of x. Otherwise x is said to be singular. The corresponding subsets of A will be denoted respectively Areg and Asing . It is clear from the denition that Areg is an open subset of A (thus Asing is closed), and that the connected components of Areg are C-analytic submanifolds of M (non necessarily closed). (4.24) Proposition. If (A, x) is irreducible, there exist arbitrarily small neighborhoods of x such that Areg is dense and connected in A . Proof. Take = as in Th. 4.19. Then AS Areg A , where AS is connected and dense in A ; hence Areg has the same properties. (4.25) Denition. The dimension of an irreducible germ of analytic set (A, x) is dened by dim(A, x) = dim(Areg , x). If (A, x) has several irreducible components (Al , x), we set dim(A, x) = max{dim(Al , x)}, codim(A, x) = n dim(A, x).
(4.26) Proposition. Let (B, x) (A, x) be germs of analytic sets. If (A, x) is irreducible and (B, x) = (A, x), then dim(B, x) < dim(A, x) and B has empty interior in A for all suciently small neighborhoods of x. Proof. We may assume x = 0, (A, 0) (Cn , 0) and (B, 0) irreducible. Then A,0 B,0 are prime ideals. When we choose suitable coordinates for the ramied coverings, we may at each step select vectors en , en1 , . . . that work simultaneously for A and B. If dim B = dim A, the process stops for both at the same time, i.e. we get ramied coverings : A ( ) , : B ( )
with ramication loci SA , SB . Then B ( (SA SB )) is an open subset of the manifold AS = A ( SA ) , therefore B AS is an analytic subset of AS with non empty interior. The same conclusion would hold if B had non empty interior in A . As AS is connected, we get B AS = AS , and as B is closed in we infer B AS = A , hence (B, 0) = (A, 0), in contradiction with the hypothesis. (4.27) Example: parametrization of curves. Suppose that (A, 0) is an irreducible germ of curve (dim(A, 0) = 1). If the disk C is chosen so small that S = {0}, then AS is a connected covering of {0} with q sheets. Hence, there exists a covering isomorphism between and the standard covering C (r) {0} (r q ) {0}, t tq , r q = radius of ,
99
i.e. a map : (r) {0} AS such that (t) = tq . This map extends into a bijective holomorphic map : (r) A with (0) = 0. This means that every irreducible germ of curve can be parametrized by a bijective holomorphic map dened on a disk in C (see also Exercise 10.8). 4.4. Coherence of Ideal Sheaves
Let A be an analytic set in a complex manifold M . The sheaf of ideals A is the subsheaf of M consisting of germs of holomorphic functions on M which vanish on A. Its stalks are the ideals A,x already considered; note that A,x = M,x if x A. If / x A, we let A,x be the ring of germs of functions on (A, x) which can be extended as germs of holomorphic functions on (M, x). By denition, there is a surjective morphism M,x A,x whose kernel is A,x , thus (4.28) i.e. A = (M /A )A . on M A.
A,x = M,x /A,x , x A, Since A,x = M,x for x A, the quotient sheaf M /A is zero / A
(4.29) Theorem ([Cartan 1950]). For any analytic set A M , the sheaf of ideals is a coherent analytic sheaf.
Proof. It is sucient to prove the result when A is an analytic subset in a neighborhood of 0 in Cn . If (A, 0) is not irreducible, there exists a neighborhood such that A = A1 . . . AN where Ak are analytic sets in and (Ak , 0) is irreducible. We have A = Ak , so by Cor. 3.15 we may assume that (A, 0) is irreducible. Then we can choose coordinates z , z , polydisks , and a primitive element u(z ) = cd+1 zd+1 + + cn zn such that Th. 4.19 is valid. Since (z ) = j<k (k u j u)2 , we see that (z ) is in fact a polynomial in the cj s with coecients in d . The same is true for the coecients of the polynomials Wu (z ; T ) and Bk (z ; T ) which can be expressed in terms of the elementary symmetric functions of the k u s. We suppose that is chosen small enough in order that all coecients of these d [cd+1 , . . . , cn ] polynomials are in ( ). Let ( ) be some non zero coecient appearing in m = c . Also, let G1 , . . . , GN ( )[z ] be the coecients of all monomials c appearing in the expansion of the functions Wu (z ; u(z )) or (z )zk Bk (z ; u(z )). Clearly, G1 , . . . , GN vanish on A . We contend that
A,x = f M,x ; f (G1,x , . . . , GN,x) . This implies that the sheaf A is the projection on the rst factor of the sheaf of relations ( , G1, . . . , GN ) N+1 , which is coherent by the Oka theorem; Theorem 4.29 then
(4.30) follows.
We rst prove that the inclusion A,x {. . .} holds in (4.30). In fact, if f (G1,x , . . . , GN,x ), then f vanishes on A { = 0} in some neighborhood of x. Since (A ) { = 0} is dense in A , we conclude that f A,x .
To prove the other inclusion A,x {. . .}, we repeat the proof of Lemma 4.18 with a few modications. Let x be a given point. At x, the irreducible polynomials Wu (z ; T ) and Wk (z ; T ) of u and zk in M,0 /A,0 split into Wu (z ; T ) = Wu,x z ; T u(x ) Qu,x z ; T u(x ) , Wk (z ; T ) = Wk,x (z ; T xk ) Qk,x (z ; T xk ),
100
where Wu,x (z ; T ) and Wk,x (z ; T ) are Weierstrass polynomials in T and Qu,x (x , 0) = 0, Qk,x (x , 0) = 0. For all z , the roots of Wu (z ; T ) are the values u(z ) at all points z A 1 (z ). As A is closed, any point z A 1 (z ) with z near x has to be in a small neighborhood of one of the points y A 1 (x ). Choose cd+1 , . . . , cn such that the linear form u(z ) separates all points in the ber A 1 (x ). Then, for a root u(z ) of Wu,x z ; T u(x ) , the point z must be in a neighborhood of y = x, otherwise u(z ) would be near u(y ) = u(x ) and the Weierstrass polynomial Wu,x (z ; T ) would have a root away from 0, in contradiction with (4.10). Conversely, if z A 1 (z ) is near x, then Qu,x z ; u(z ) u(x ) = 0 and u(z ) is a root of Wu,x z ; T u(x ) . From this, we infer that every polynomial P (z ; T ) ,x [T ] such that P z ; u(z ) = 0 on (A, x) is a multiple of Wu,x z ; T u(x ) , because the roots of the latter polynomial are simple for z in the dense set ( S, x). In particular deg P < deg Wu,x implies P = 0 and (z )q Wk,x z ; Bk (z ; u(z ))/(z ) xk is a multiple of Wu,x z ; T u(x ) . If we replace Wu , Wk by Wu,x , Wk,x respectively, the proof of Lemma 4.18 shows that for every f M,x there is a polynomial R ,x [T ] of degree deg R < deg Wu,x such that (z )m f (z) = R z ; u(z ) modulo the ideal , Wu,x z ; u(z ) u(x ) , (z )zk Bk z ; u(z )
and f A,x implies R = 0. Since Wu,x diers from Wu only by an invertible element in M,x , we conclude that c
A,x = mA,x (G1,x , . . . , GN,x ).

for all .
This is true for a dense open set of coecients cd+1 , . . . , cn , therefore by expressing the coecients through interpolation of c at suitable points c we infer A,x (G1,x , . . . , GN,x ) (4.31) Theorem. Asing is an analytic subset of A. Proof. The statement is local. Assume rst that (A, 0) is an irreducible germ in Cn . Let g1 , . . . , gN be generators of the sheaf A on a neighborhood of 0. Set d = dim A. In a neighborhood of every point x Areg , A can be dened by holomorphic equations u1 (z) = . . . = und (z) = 0 such that du1 , . . . , dund are linearly independant. As u1 , . . . , und are generated by g1 , . . . , gN , one can extract a subfamily gj1 , . . . , gjnd that has at least one non zero Jacobian determinant of rank n d at x. Therefore Asing is dened by the equations det gj zk
jJ kK
= 0,
J {1, . . . , N }, K {1, . . . , n}, |J| = |K| = n d.
Assume now that (A, 0) = (Al , 0) with (Al , 0) irreducible. The germ of an analytic set at a regular point is irreducible, thus every point which belongs simultaneously to at least two components is singular. Hence (Asing , 0) = (Al,sing , 0) (Ak Al , 0),
k=l
5. Complex Spaces
101
and Asing is analytic. Now, we give a characterization of regular points in terms of a simple algebraic property of the ring A,x . (4.32) Proposition. Let (A, x) be a germ of analytic set of dimension d and let mA,x A,x be the maximal ideal of functions that vanish at x. Then mA,x cannot have less than d generators and mA,x has d generators if and only if x is a regular point. Proof. If A Cn is a d-dimensional submanifold in a neighborhood of x, there are local coordinates centered at x such that A is given by the equations zd+1 = . . . = zn near z = 0. Then A,x d and mA,x is generated by z1 , . . . , zd . Conversely, assume that mA,x has s generators g1 (z), . . . , gs (z) in A,x = Cn ,x /A,x . Letting x = 0 for simplicity, we can write zj =
1 k s
ujk (z)gk (z) + fj (z),
ujk n , fj A,0 , 1
n.
Then we nd dzj = cjk (0)dgk (0)+dfj (0), so that the rank of the system of dierentials dfj (0) 1 j n is at least equal to n s. Assume for example that df1 (0), . . . , dfns (0) are linearly independent. By the implicit function theorem, the equations f1 (z) = . . . = fns (z) = 0 dene a germ of submanifold of dimension s containing (A, 0), thus s d and (A, 0) equals this submanifold if s = d. (4.33) Corollary. Let A M be an analytic set of pure dimension d and let B A be an analytic subset of codimension p in A. Then, as an A,x -module, the ideal B,x cannot be generated by less than p generators at any point x B, and by less than p + 1 generators at any point x Breg Asing .
Proof. Suppose that B,x admits s-generators (g1 , . . . , gs ) at x. By coherence of B these germs also generate B in a neighborhood of x, so we may assume that x is a regular point of B. Then there are local coordinates (z1 , . . . , zn ) on M centered at x such that (B, x) is dened by zk+1 = . . . = zn = 0, where k = dim(B, x). Then the maximal ideal mB,x = mA,x /B,x is generated by z1 , . . . , zk , so that mA,x is generated by (z1 , . . . , zk , g1 , . . . , gs ). By Prop. 4.32, we get k + s d, thus s d k p, and we have strict inequalities when x Asing .
5. Complex Spaces
Much in the same way a manifold is constructed by piecing together open patches isomorphic to open sets in a vector space, a complex space is obtained by gluing together open patches isomorphic to analytic subsets. The general concept of analytic morphism (or holomorphic map between analytic sets) is rst needed. 5.1. Morphisms and Comorphisms Let A Cn and B Cp be analytic sets. A morphism from A to B is by denition a map F : A B such that for every x A there is a neighborhood U of x and a holomorphic map F : U Cp such that FAU = FAU . Equivalently, such a
102
morphism can be dened as a continuous map F : A B such that for all x A and g B,F (x) we have g F A,x . The induced ring morphism (5.1)
Fx :
B,F (x) g g F A,x
is called the comorphism of F at point x. 5.1. Denition of Complex Spaces (5.2) Denition. A complex space X is a locally compact Hausdor space, countable at innity, together with a sheaf X of continuous functions on X, such that there exists an open covering (U ) of X and for each a homeomorphism F : U A onto an analytic set A Cn such that the comorphism F : A X U is an isomorphism of sheaves of rings. X is called the structure sheaf of X. By denition a complex space X is locally isomorphic to an analytic set, so the concepts of holomorphic function on X, of analytic subset, of analytic morphism, etc . . . are meaningful. If X is a complex space, Th. 4.31 implies that Xsing is an analytic subset of X. (5.3) Theorem and denition. For every complex space X, the set Xreg is a dense open subset of X, and consists of a disjoint union of connected complex manifolds X . Let X be the closure of X in X. Then (X ) is a locally nite family of analytic subsets of X, and X = X . The sets X are called the global irreducible components of X.
A1 (1, 0) (0, 0) A2
Fig. II-2 The irreducible curve y 2 = x2 (1 + x) in C2 . Observe that the germ at a given point of a global irreducible component can be reducible, as shows the example of the cubic curve : y 2 = x2 (1 + x) ; the germ (, 0) has two analytic branches y = x 1 + x, however {0} is easily seen to be a connected smooth Riemann surface (the real points of corresponding to 1 x 0 form a path connecting the two branches). This example shows that the notion of global irreducible component is quite dierent from the notion of local irreducible component introduced in (4.4).
Proof. By denition of Xreg , the connected components X are (disjoint) complex man ifolds. Let us show that the germ of X = X at any point x X is analytic. We may
5. Complex Spaces
103
assume that (X, x) is a germ of analytic set A in an open subset of Cn . Let (Al , x), 1 l N , be the irreducible components of this germ and U a neighborhood of x such that X U = Al U . Let l U be a neighborhood of x such that Al,reg l is connected and dense in Al l (Prop. 4.24). Then A := Xreg Al l equals l (Al,reg l ) Al,reg l Ak . However, Al,reg l Ak is an analytic subset k=l of Al,reg l , distinct from Al,reg l , otherwise Al,reg l would be contained in Ak , thus (Al , x) (Ak , x) by density. Remark 4.2 implies that A is connected and dense l in Al,reg l , hence in Al l . Set = l and let (X )J be the family of global components which meet (i.e. such that X = ). As Xreg = A , l each X , J, meets at least one set A , and as A Xreg is connected, we have in l l fact A X . It follows that there exists a partition (L )J of {1, . . . , N } such that l X = lL A , J. Hence J is nite, card J N , and l X = X = is analytic for all J. (5.4) Corollary. If A, B are analytic subsets in a complex space X, then the closure A B is an analytic subset, consisting of the union of all global irreducible components A of A which are not contained in B. Proof. Let C = A be the union of these components. Since (A ) is locally nite, C is analytic. Clearly A B = C B = A B. The regular part A of each A is a connected manifold and A B is a proper analytic subset (otherwise A B would imply A B). Thus A (A B) is dense in A which is dense in A , so A B = A = C. (5.5) Theorem. For any family (A ) of analytic sets in a complex space X, the intersection A = A is an analytic subset of X. Moreover, the intersection is stationary on every compact subset of X. Proof. It is sucient to prove the last statement, namely that every point x X has a neighborhood such that A is already obtained as a nite intersection. However, since X,x is Noetherian, the family of germs of nite intersections has a minimum element (B, x), B = Aj , 1 j N . Let B be the union of the global irreducible components B of B which contain the point x ; clearly (B, x) = (B, x). For any set A in the family, the minimality of B implies (B, x) (A , x). Let B be the regular part of any global irreducible component B of B. Then B A is a closed analytic subset of B containing a non empty open subset (the intersection of B with some neighborhood of x), so we must have B A = B . Hence B = B A for all B B and all A , thus B A = A . We infer (B, x) = (B, x) (A, x) (B, x), and the proof is complete. As a consequence of these general results, it is not dicult to show that a complex space always admits a (locally nite) stratication such that the strata are smooth manifolds.
lL
Al =
lL
Al
104
(5.6) Proposition. Let X be a complex space. Then there is a locally stationary increasing sequence of analytic subsets Yk X, k N, such that Y0 is a discrete set and such that Yk Yk1 is a smooth k-dimensional complex manifold for k 1. Such a sequence is called a stratication of X, and the sets Yk Yk1 are called the strata (the strata may of course be empty for some indices k < dim X). Proof. Let be the family of irreducible analytic subsets Z X which can be obtained through a nite sequence of steps of the following types: a) Z is an irreducible component of X ;
b) Z is an irreducible component of the singular set Zsing of some member Z
c) Z is an irreducible component of some nite intersection of sets Zj
Since X has locally nite dimension and since steps b) or c) decrease the dimension of our sets Z, it is clear that is a locally nite family of analytic sets in X. Let Yk be the of dimension k. It is easily seen that Yk = X and that the union of all sets Z irreducible components of (Yk )sing are contained in Yk1 (these components are either intersections of components Zj Yk or parts of the singular set of some component Z Yk , so there are in either case obtained by step b) or c) above). Hence Yk Yk1 is a smooth manifold and it is of course k-dimensional, because the components of Yk of dimension < k are also contained in Yk1 by denition. (5.7) Theorem. Let X be an irreducible complex space. Then every non constant holomorphic function f on X denes an open map f : X C. Proof. We show that the image f () of any neighborhood of x X contains a neighborhood of f (x). Let (Xl , x) be an irreducible component of the germ (X, x) (embedded in Cn ) and = a polydisk such that the projection : Xl is a ramied covering. The function f is non constant on the dense open manifold Xreg , so we may select a complex line L through 0, not contained in the ramication locus of , such that f is non constant on the one dimensional germ 1 (L). Therefore we can nd a germ of curve (C, 0) t (t) (X, x) such that f is non constant. This implies that the image of every neighborhood of 0 C by f already contains a neighborhood of f (x). (5.8) Corollary. If X is a compact irreducible analytic space, then every holomorphic function f (X) is constant. In fact, if f (X) was non constant, f (X) would be compact and also open in C by Th. 5.7, a contradiction. This result implies immediately the following consequence. (5.9) Theorem. Let X be a complex space such that the global holomorphic functions in (X) separate the points of X. Then every compact analytic subset A of X is nite. Proof. A has a nite number of irreducible components A which are also compact. Every function f (X) is constant on A , so A must be reduced to a single point.
6. Analytic Cycles and Meromorphic Functions
105
5.2. Coherent Sheaves over Complex Spaces Let X be a complex space and X its structure sheaf. Locally, X can be identied to an analytic set A in an open set Cn , and we have X = /A . Thus X is coherent over the sheaf of rings . It follows immediately that X is coherent over itself. Let be a X -module. If denotes the extension of A to obtained by setting x = 0 for x A, then is a -module, and it is easily seen that A is coherent over XA if and only if is coherent over . If Y is an analytic subset of X, then Y is locally given by an analytic subset B of A and the sheaf of ideals of Y in X is the quotient Y = B /A ; hence Y is coherent. Let us mention the following important property of supports. (5.10) Theorem. If is a coherent X -module, the support of , dened as Supp {x X ; x = 0} is an analytic subset of X.
Proof. The result is local, thus after extending by 0, we may as well assume that X is an open subset Cn . By (3.12), there is an exact sequence of sheaves
p U
q U 0 U
G F
in a neighborhood U of any point. If G : p q is surjective it is clear that the x x linear map G(x) : Cp Cq must be surjective; conversely, if G(x) is surjective, there is a q-dimensional subspace E Cp on which the restriction of G(x) is a bijection onto Cq ; then GE : U C E q is bijective near x and G is surjective. The support U of U is thus equal to the set of points x U such that all minors of G(x) of order q vanish.

6.1. Complete Intersections Our goal is to study in more details the dimension of a subspace given by a set of equations. The following proposition is our starting point. (6.1) Proposition. Let X be a complex space of pure dimension p and A an analytic subset of X with codimX A 2. Then every function f (X A) is locally bounded near A. Proof. The statement is local on X, so we may assume that X is an irreducible germ of analytic set in (Cn , 0). Let (Ak , 0) be the irreducible components of (A, 0). By a reasoning analogous to that of Prop. 4.26, we can choose coordinates (z1 , . . . , zn ) on Cn such that all projections : z (z1 , . . . , zp ), p = dim X, k : z (z1 , . . . , zpk ), pk = dim Ak , dene ramied coverings : X , k : Ak . By construction k (Ak ) is contained in the set Bk dened by some Weierstrass polynomials in the
106
variables zpk +1 , . . . , zp and codim Bk = p pk 2. Let S be the ramication locus of and B = Bk . We have (A ) B. For z (S B), we let
k (z ) = elementary symmetric function of degree k in f (z , z ), where (z , z ) are the q points of X projecting on z . Then k is holomorphic on (S B) and locally bounded near every point of S B, thus k extends holomorphically to B by Remark 4.2. Since codim B 2, k extends to by Cor. 1.4.5. Now, f satises f q 1 f q1 + . . . + (1)q q = 0, thus f is locally bounded on X .
(6.2) Theorem. Let X be an irreducible complex space and f f 1 (0) is empty or of pure dimension dim X 1.
(X), f
0. Then
Proof. Let A = f 1 (0). By Prop. 4.26, we know that dim A dim X 1. If A had an irreducible branch Aj of dimension dim X 2, then in virtue of Prop. 6.1 the function 1/f would be bounded in a neighborhood of Aj k=j Ak , a contradiction. (6.3) Corollary. If f1 , . . . , fp are holomorphic functions on an irreducible complex space 1 1 X, then all irreducible components of f1 (0) . . . fp (0) have codimension p. (6.4) Denition. Let X be a complex space of pure dimension n and A an analytic subset of X of pure dimension. Then A is said to be a local (set theoretic) complete intersection in X if every point of A has a neighborhood such that with exactly p = codim A functions fj (). A = {x ; f1 (x) = . . . = fp (x) = 0}
(6.5) Remark. As a converse to Th. 6.2, one may ask whether every hypersurface A in X is locally dened by a single equation f = 0. In general the answer is negative. A simple counterexample for dim X = 3 is obtained with the singular quadric X = {z1 z2 + z3 z4 = 0} C4 and the plane A = {z1 = z3 = 0} X. Then A cannot be dened by a single equation f = 0 near the origin, otherwise the plane B = {z2 = z4 = 0} would be such that f 1 (0) B = A B = {0}, in contradiction with Th. 6.2 (also, by Exercise 10.11, we would get the inequality codimX A B 2). However, the answer is positive when X is a manifold: (6.6) Theorem. Let M be a complex manifold with dimC M = n, let (A, x) be an analytic germ of pure dimension n 1 and let Aj , 1 j N , be its irreducible components. a) The ideal of (A, x) is a principal ideal germs gj such that Aj ,x = (gj ).
A,x = (g) where g is a product of irreducible
b) For every f M,x such that f 1 (0) (A, x), there is a unique decomposition m m f = ug1 1 . . . gN N where u is an invertible germ and mj is the order of vanishing of f at any point z Aj,reg k=j Ak . Proof. a) In a suitable local coordinate system centered at x, the projection : Cn Cn1 realizes all Aj as ramied coverings : Aj Cn1 , ramication locus = Sj .
107
The function gj (z , zn ) =
wAj 1 (z )
(zn wn ),
Sj
extends into a holomorphic function in [zn ] and is irreducible at x. Set g = gj A,x . For any f A,x , the Weierstrass division theorem yields f = gQ + R with R n1 [zn ] and deg R < deg g. As R(z , zn ) vanishes when zn is equal to wn for each point w A 1 (z ), R has exactly deg g roots when z Sj (Aj Ak ) , so R = 0. Hence A,x = (g) and similarly Aj ,x = (gj ). Since Aj is coherent, gj is also a generator of Aj ,z for z near x and we infer that gj has order 1 at any regular point z Aj,reg .
m m b) As M,x is factorial, any f M,x can be written f = u g1 1 . . . gN N where u is either invertible or a product of irreducible elements distinct from the gj s. In the latter case the hypersurface u1 (0) cannot be contained in (A, x), otherwise it would be a union of some of the components Aj and u would be divisible by some gj . This proves b).
(6.7) Denition. Let X be an complex space of pure dimension n. a) An analytic q-cycle Z on X is a formal linear combination j Aj where (Aj ) is a locally nite family of irreducible analytic sets of dimension q in X and j Z. The support of Z is |Z| = j =0 Aj . The group of all q-cycles on X is denoted Cyclq (X). Eective q-cycles are elements of the subset Cyclq (X) of cycles such that + all coecients j are 0 ; rational, real cycles are cycles with coecients j Q, R. b) An analytic (n 1)-cycle is called a (Weil ) divisor, and we set Div(X) = Cycln1 (X). c) Assume that dim Xsing n 2. If f (X) does not vanish identically on any irreducible component of X, we associate to f a divisor div(f ) = mj Aj Div+ (X)
in the following way: the components Aj are the irreducible components of f 1 (0) and the coecient mj is the vanishing order of f at every regular point in Xreg Aj,reg k=j Ak . It is clear that we have div(f g) = div(f ) + div(g). d) A Cartier divisor is a divisor D = j Aj that is equal locally to a Z-linear combination of divisors of the form div(f ). It is easy to check that the collection of abelian groups Cyclq (U ) over all open sets U X, together with the obvious restriction morphisms, satises axioms (1.4) of sheaves; observe however that the restriction of an irreducible component Aj to a smaller open set may subdivide in several components. Hence we obtain sheaves of abelian groups Cyclq and Div = Cycln1 on X. The stalk Cyclq is the free abelian group generated by x the set of irreducible germs of q-dimensional analytic sets at the point x. These sheaves
108
carry a natural partial ordering determined by the subsheaf of positive elements Cyclq . + We dene the sup and inf of two analytic cycles Z = j Aj , Z = j Aj by (6.8) sup{Z, Z } = sup{j , j } Aj , inf{Z, Z } = inf{j , j } Aj ;
it is clear that these operations are compatible with restrictions, i.e. they are dened as sheaf operations. (6.9) Remark. When X is a manifold, Th. 6.6 shows that every eective Z-divisor is locally the divisor of a holomorphic function; thus, for manifolds, the concepts of Weil and Cartier divisors coincide. This is not always the case in general: in Example 6.5, one can show that A is not a Cartier divisor (exercise 10.?). 6.2. Divisors and Meromorphic Functions
Let X be a complex space. For x X, let X,x be the ring of quotients of X,x , i.e. the set of formal quotients g/h, g, h X,x , where h is not a zero divisor in X,x , with the identication g/h = g /h if gh = g h. We consider the disjoint union
(6.10)
X =
xX
X,x
A point x X is called a pole of a meromorphic function f on X if fx X,x . / Clearly, the set Pf of poles of f is a closed subset of X with empty interior: if f = g/h on U , then h 0 on any irreducible component and Pf U h1 (0). For x Pf , / one can speak of the value f (x). If the restriction of f to Xreg Pf does not vanish identically on any irreducible component of (X, x), then 1/f is a meromorphic function in a neighborhood of x ; the set of poles of 1/f will be denoted Zf and called the zero set of f . If f vanishes on some connected open subset of Xreg Pf , then f vanishes identically (outside Pf ) on the global irreducible component X containing this set; we agree that these components X are contained in Zf . For every point x in the complement of Zf Pf , we have either fx X,x or (1/f )x X,x , thus f denes a holomorphic map X (Zf Pf ) C {} = P1 with values in the projective line. In general, no value (nite or innite) can be assigned to f at a point x Zf Pf , as shows the example of the function f (z) = z2 /z1 in C2 . The set Zf Pf is called the indeterminacy set of f . (6.11) Theorem. For every meromorphic function f on X, the sets Pf , Zf and the indeterminacy set Zf Pf are analytic subsets.
with the topology in which the open sets open sets are unions of sets of the type {Gx /Hx ; x V } X where V is open in X and G, H X (V ). Then X is a sheaf over X, and the sections of X over an open set U are called meromorphic functions on U . By denition, these sections can be represented locally as quotients of holomorphic functions, but there need not exist such a global representation on U .
Proof. Let x be the ideal of germs u X,x such that ufx X,x . Let us write f = g/h on a small open set U . Then U appears as the projection on the rst factor of the sheaf of relations (g, h) U U , so is a coherent sheaf of ideals. Now Pf = x X ;
x = X,x
= Supp
X /,
109
thus Pf is analytic by Th. 5.10. Similarly, the projection of denes a sheaf of ideals such that Zf = Supp X / .
(g, h) on the second factor
When X has pure dimension n and dim Xsing n 2, Def. 6.7 c) can be extended to meromorphic functions: if f = g/h locally, we set (6.12) div(f ) = div(g) div(h).
By 6.7 c), we immediately see that this denition does not depend on the choice of the local representant g/h. Furthermore, Cartier divisors are precisely those divisors which are associated locally to meromorphic functions. Assume from now on that M is a connected n-dimensional complex manifold. Then, for every point x M , the ring M,x n is factorial. This property makes the study of meromorphic functions much easier. (6.13) Theorem. Let f be a non zero meromorphic function on a manifold M , with dimC M = n. Then the sets Zf , Pf are purely (n1)-dimensional, and the indeterminacy set Zf Pf is purely (n 2)-dimensional.
Proof. For every point a M , the germ fa can be written ga /ha where ga , ha M,a are relatively prime holomorphic germs. By Th. 1.12, the germs gx , hx are still relatively prime for x in a neighborhood U of a. Thus the ideal associated to f coincides with (h) on U , and we have Pf U = Supp
U /(h) = h1 (0),
Zf U = g 1 (0).
Th. 6.2 implies our contentions: if g and h are the irreducible components of g, h, then 1 Zf Pf = g (0) h1 (0) is (n 2)-dimensional. As we will see in the next section, Th. 6.13 does not hold on an arbitrary complex space. Let (Aj ), resp. (Bj ), be the global irreducible components of Zf , resp. Pf . In a neighborhood Vj of the (n 1)-dimensional analytic set A = Aj j Pf Ak )
k=j
f is holomorphic and V f 1 (0) = A . As A j j,reg is connected, we must have div(fVj ) = mj A for some constant multiplicity mj equal to the vanishing order of f along A . j j,reg Similarly, 1/f is holomorphic in a neighborhood Wj of
Bj = Bj
Zf
Bk )
k=j
and we have div(fV ) = pj Bj where pj is the vanishing order of 1/f along Bj,reg . At a point x M the germs Aj,x and Bj,x may subdivide in irreducible local components Aj,,x and Bj,,x . If gj, and hj, are local generators of the corresponding ideals, we may a priori write
fx = u g/h where
g=
j, gj, ,
h=
j, hj,
110
and where u is invertible. Then necessarily mj, = mj and pj, = pj for all , and we see that the global divisor of f on M is (6.14) div(f ) = mj Aj pj Bj .
Let us denote by the multiplicative sheaf of germs of non zero meromorphic functions, and by the sheaf of germs of invertible holomorphic functions. Then we have an exact sequence of sheaves (6.15) 1 Div 0.
div
Indeed, the surjectivity of div is a consequence of Th. 6.6. Moreover, any meromorphic function that has a positive divisor must be holomorphic by the fact that n is factorial. Hence a meromorphic function f with div(f ) = 0 is an invertible holomorphic function.
7. Normal Spaces and Normalization

7.1. Weakly Holomorphic Functions The goal of this section is to show that the singularities of X can be studied by en larging the structure sheaf X into a sheaf X of so-called weakly holomorphic functions. (7.1) Denition. Let X be a complex space. A weakly holomorphic function f on X is a holomorphic function on Xreg such that every point of Xsing has a neighborhood V for which f is bounded on Xreg V . We denote by X,x the ring of germs of weakly holomorphic functions over neighborhoods of x and X the associated sheaf. Clearly, X,x is a ring containing X,x . If (Xj , x) are the irreducible components of (X, x), there is a fundamental system of neighborhoods V of x such that Xreg V is a disjoint union of connected open sets Xj,reg V Xk Xj,reg V
k=j
which are dense in Xj,reg V . Therefore any bounded holomorphic function on Xreg V extends to each component Xj,reg V and we see that X,x = X ,x .
j
The rst important fact is that weakly holomorphic functions are always meromorphic and possess universal denominators. (7.2) Theorem. For every point x X, there is a neighborhood V of x and h X (V ) such that h1 (0) is nowhere dense in V and hy X,y X,y for all y V ; such a function h is called a universal denominator on V . In particular X is contained in the ring X of meromorphic functions. Proof. First assume that (X, x) is irreducible and that we have a ramied covering : X with ramication locus S. We claim that the discriminant (z ) of a
111
primitive element u(z ) = cd+1 zd+1 + + cn zn is a universal denominator on X . To see this, we imitate the proof of Lemma 4.15. Let f X,y , y X . Then we solve the equation f (z) = bj (z )u(z )j
0 j q in a neighborhood of y. For z S, let us denote by (z , z ), 1 q, the points in the ber X 1 (z ). Among these, only q are close to y, where q is the sum of the sheet numbers of the irreducible components of (X, y) by the projection . The other points (z , z ), say q < q, are in neighborhoods of the points of 1 (y ) {y}. We take bj (z ) to be the solution of the linear system bj (z )u(z )j = 0 j q f (z , z ) for 1 0 for q <
q , n.
The solutions bj (z ) are holomorphic on S near y . Since the determinant is (z )1/2 , we see that bj is bounded, thus bj ,y and y f X,y .
Now, assume that (X, x) (Cn , 0) has irreducible components (Xj , x). We can nd for each j a neighborhood j of 0 in Cn and a function j n (j ) which is a universal denominator on Xj j . After adding to j a function which is identically zero on 1 (Xj , x) and non zero on (Xk , x), k = j, we may assume that j (0) Xk is nowhere dense in Xk for all j and k and some small j . Then = j is a universal denominator on each component Xj . For some possibly smaller , select a function 1 vj n () such that vj vanishes identically on k=j Xk and vj (0) is nowhere dense in Xj , and set h = vk . For any germ f X,y , y X , there is a germ gj ,y with f = gj on (Xj , y). We have h = vj on Xj , so h1 (0) is nowhere dense in X and hf = vj f = vj gj = Since vk gk ,y , we get hX,y X,y . vk gk on each (Xj , y).
(7.3) Theorem. If (X, x) is irreducible, X,x is the integral closure of quotient eld X,x . Moreover, every germ f X,x admits a limit
Xreg zx
X,x
in its
lim
f (z).
Observe that X,x is an entire ring, so the ring of quotients X,x is actually a eld. A simple illustration of the theorem is obtained with the irreducible germ of curve 3 2 X : z1 = z2 in (C2 , 0). Then X can be parametrized by z1 = t2 , z2 = t3 , t C, and 3 2 X,0 = C{z1 , z2}/(z1 z2 ) = C{t2 , t3} consists of all convergent series antn with a1 = 0. The function z2 /z1 = t is weakly holomorphic on X and satises the integral equation t2 z1 = 0. Here we have X,0 = C{t}. Proof. a) Let f = g/h be an element in
X,x satisfying an integral equation f m + a1 f m1 + . . . + am = 0, ak X,x .
112
b) Let f X,x and let : X be a ramied covering in a neighborhood of x, with ramication locus S. As in the proof of Th. 6.1, f satises an equation f q 1 f q1 + + (1)q q = 0, k ( ) ;
Set A = h1 (0). Then f is holomorphic on X A near x, and Lemma 4.10 shows that f is bounded on a neighborhood of x. By Remark 4.2, f can be extended as a holomorphic function on Xreg in a neighborhood of x, thus f X,x .
indeed the elementary symmetric functions k (z ) are holomorphic on S and boun ded, so they extend holomorphically to . Hence X,x is integral over X,x and we already know that X,x X,x . c) Finally, the cluster set V x f (Xreg V ) is connected, because there is a fundamental system of neighborhoods V of x such that Xreg V is connected, and any intersection of a decreasing sequence of compact connected sets is connected. However the limit set is contained in the nite set of roots of equation b) at point x , so it must be reduced to one element. 7.2. Normal Spaces Normal spaces are spaces for which all weakly holomorphic functions are actually holomorphic. These spaces will be seen later to have simpler singularities than general analytic spaces. (7.4) Denition. A complex space X is said to be normal at a point x if (X, x) is irreducible and X,x = X,x , that is, X,x is integrally closed in its eld of quotients. The set of normal (resp. non-normal) points will be denoted Xnorm (resp. Xn-n ). The space X itself is said to be normal if X is normal at every point. Observe that any regular point x is normal: in fact X,x n is then factorial, hence integrally closed. Therefore Xn-n Xsing . (7.5) Theorem. The non-normal set Xn-n is an analytic subset of X. In particular, Xnorm is open in X. Proof. We give here a beautifully simple proof due to [Grauert and Remmert 1984]. Let h be a universal denominator on a neighborhood V of a given point and let = hX be the sheaf of ideals of h1 (0) by Hilberts Nullstellensatz. Finally, let = hom (, ) be the sheaf of X -endomorphisms of . Since is coherent, so is (cf. Exercise 10.?). over V . We claim that there Clearly, the homotheties of give an injection X is a natural injection X . In fact, any endomorphism of yields by restriction a homomorphism hX X , and by X -linearity such a homomorphism is obtained by multiplication by an element in h1 X . Thus h1 X X . Since each stalk x is a nite X,x -module containing non-zero divisors, it follows that that any meromorphic germ f such that f x x is integral over X,x ([Lang 1965], Chapter IX, 1), hence X . Now, we assert that x X,x . Thus we have inclusions X Xn-n V = {x V ;
x
= X,x } =
/ X .
113
This will imply the theorem by 5.10. To prove the equality, we rst observe that x = X,x implies X,x = X,x , thus x Xn-n . Conversely, assume that x is non normal, that is, X,x = X,x . Let k be the smallest integer such that k X,x X,x ; such x an integer exists since l X,x hX,x X,x for l large. Then there is an element x w k1 X,x such that w X,x . We have wx X,x ; moreover, as w is locally / x bounded near Xsing , any germ wg in wx satises lim w(z)g(z) = 0 when z Xreg tends to a point of the zero variety h1 (0) of x . Hence wx x , i.e. w x , but w X,x , / so x = X,x . (7.6) Theorem. If x X is a normal point, then (Xsing , x) has codimension at least 2 in (X, x). Proof. We suppose that = Xsing has codimension 1 in a neighborhood of x and try to get a contradiction. By restriction to a smaller neighborhood, we may assume that X itself is normal and irreducible (since Xnorm is open), dim X = n, that has pure dimension n 1 and that the ideal sheaf has global generators (g1 , . . . , gk ). Then 1 gj (0) ; both sets have pure dimension n 1 and thus singular sets of dimension 1 n 2. Hence there is a point a that is regular on and on gj (0), in particular 1 1 there is a neighborhood V of a such that g1 (0) V = . . . = gk (0) V = V is a smooth (n 1)-dimensional manifold. Since codimX = 1 and a is a singular point of X, ,a cannot have less than 2 generators in X,a by Cor. 4.33. Take (g1 , . . . , gl ), l 2, to be a minimal subset of generators. Then f = g2 /g1 cannot belong to X,a , but f is holomorphic on V . We may assume that there is a sequence a V converging to a such that f (a ) remains bounded (otherwise reverse g1 and g2 and pass 1 to a subsequence). Since g1 (0) V = V , Hilberts Nullstellensatz gives an integer m m such that ,a g1 X,a , hence fa m X,a . We take m to be the smallest ,a integer such that the latter inclusion holds. Then there is a product g = g1 1 . . . gl l with || = m 1 such that f g X,a but f g gj X,a for each j. Since the sequence / f (a ) is bounded we conclude that f g gj vanishes at a. The zero set of this function 1 has dimension n 1 and is contained in gk (0) V = V so it must contain the germ (, a). Hence f g gj ,a and f g ,a ,a . As ,a is a nitely generated X,a -module, this implies f g X,a = X,a , a contradiction. (7.7) Corollary. A complex curve is normal if and only if it is regular. (7.8) Corollary. Let X be a normal complex space and Y an analytic subset of X such that dim(Y, x) dim(X, x) 2 for any x X. Then any holomorphic function on X Y can be extended to a holomorphic function on X. Proof. By Cor. 1.4.5, every holomorphic function f on Xreg Y extends to Xreg . Since codim Xsing 2, Th. 6.1 shows that f is locally bounded near Xsing . Therefore f extends to X by denition of a normal space. 7.3. The Oka Normalization Theorem The important normalization theorem of [Oka 1950] shows that X can be used to dene the structure sheaf of a new analytic space X which is normal and is obtained by simplifying the singular set of X. More precisely:
114
(7.9) Denition. Let X be a complex space. A normalization (Y, ) of X is a normal complex space Y together with a holomorphic map : Y X such that the following conditions are satised. a) : Y X is proper and has nite bers; b) if is the set of singular points of X and A = 1 (), then Y and : Y A X = Xreg is an analytic isomorphism. A is dense in Y
It follows from b) that Y A Yreg . Thus Y is obtained from X by a suitable modication of its singular points. Observe that Yreg may be larger than Y A, as is the case in the following two examples. (7.10) Examples. a) Let X = C{0}{0}C be the complex curve z1 z2 = 0 in C2 . Then the normalization of X is the disjoint union Y = C {1, 2} of two copies of C, with the map (t1 ) = (t1 , 0), (t2 ) = (0, t2 ). The set A = 1 (0, 0) consists of exactly two points.
3 2 b) The cubic curve X : z1 = z2 is normalized by the map : C X, t (t2 , t3 ). Here is a homeomorphism but 1 is not analytic at (0, 0).
We rst show that the normalization is essentially unique up to isomorphism and postpone the proof of its existence for a while. (7.11) Lemma. If (Y1 , 1 ) and (Y2 , 2 ) are normalizations of X, there is a unique analytic isomorphism : Y1 Y2 such that 1 = 2 .
1 Proof. Let be the set of singular points of X and Aj = j (), j = 1, 2. Let 1 : Y1 A1 Y2 A2 be the analytic isomorphism 2 1 . We assert that can be extended into a map : Y1 Y2 . In fact, let a A1 and s = 1 (a) . Then 1 2 (s) consists of a nite set of points yj Y2 . Take disjoint neighborhoods Uj of yj such that Uj is an analytic subset in an open set j CN . Since 2 is proper, there 1 is a neighborhood V of s in X such that 2 (V ) Uj and by continuity of 1 a 1 neighborhood W of a such that 1 (W ) V . Then = 2 1 maps W A1 into Uj and can be seen as a bounded holomorphic map into CN through the embeddings Uj j CN . Since Y1 is normal, extends to W , and the extension takes values in U j which is contained in Y2 (shrink Uj if necessary). Thus extends into a map : Y1 Y2 and similarly 1 extends into a map : Y2 Y1 . By density of Yj Aj , we have = IdY1 , = IdY2 .
(7.12) Oka normalization theorem. Let X be any complex space. Then X has a normalization (Y, ). Proof. Because of the previous lemma, it suces to prove that any point x X has a neighborhood U such that U admits a normalization; all these local normalizations will then glue together. Hence we may suppose that X is an analytic set in an open set of Cn . Moreover, if (X, x) splits into irreducible components (Xj , x) and if (Yj , j ) is a normalization of Xj U , then the disjoint union Y = Yj with = j is easily seen to be a normalization of X U . We may therefore assume that (X, x) is irreducible.
115
Let h be a universal denominator in a neighborhood of x. Then X,x is isomorphic to its image hX,x X,x , so it is a nitely generated X,x -module. Let (f1 , . . . , fm ) be a nite set of generators of X,x . After shrinking X again, we may assume the following two points: fj is holomorphic in Xreg , can be written fj = gj /h on X with gj , h in n () and satises an integral equation Pj (z ; fj (z)) = 0 where Pj (z ; T ) is a unitary polynomial with holomorphic coecients on X. Set X = X h1 (0). Consider the holomorphic map F : Xreg Cm , z z, f1 (z), . . . , fm (z) X is an analytic set in an open set Cn , (X, x) is irreducible and Xreg is connected;
and the image Y = F (X ). We claim that the closure Y of Y in Cm is an analytic set. In fact, the set Z = (z, w) Cm ; z X , h(z)wj = gj (z) is analytic and Y = Z {h(z) = 0}, so we may apply Cor. 5.4. Observe that Y is contained in the set dened by Pj (z ; wj ) = 0, thus so is its closure Y . The rst projection Cm gives a holomorphic map : Y X such that F = Id on X , hence also on Xreg . If = Xsing and A = 1 (), the restriction : Y A X = Xreg is thus an analytic isomorphism and F is its inverse. Since (X, x) is irreducible, each fj has a limit j at x by Th. 7.3 and the ber 1 (x) is reduced to the single point y = (x, ). The other bers 1 (z) are nite because they are contained in the nite set of roots of the equations Pj (z ; wj ) = 0. The same argument easily shows that is proper (use Lemma 4.10). Next, we show that Y is normal at the point y = 1 (x). In fact, for any bounded holomorphic function u on (Yreg , y) the function u F is bounded and holomorphic on (Xreg , x). Hence u F X,x = X,x [f1 , . . . , fm ] and we can write u F (z) = Q(z ; f1 (z), . . . , fm (z)) = Q F (z) where Q(z ; w) = a (z)w is a polynomial in w with coecients in X,x . Thus u coincides with Q on (Yreg , y), and as Q is holomorphic on (X, x) Cm (Y, y), we conclude that u Y,y . Therefore Y,y = Y,y . Finally, by Th. 7.5, there is a neighborhood V Y of y such that every point of V is normal. As is proper, we can nd a neighborhood U of x with 1 (U ) V . Then : 1 (U ) U is the required normalization in a neighborhood of x.
The proof of Th. 7.12 shows that the ber 1 (x) has exactly one point yj for each irreducible component (Xj , x) of (X, x). As a one-to-one proper map is a homeomorphism, we get in particular: (7.13) Corollary. If X is a locally irreducible complex space, the normalization : Y X is a homeomorphism. (7.14) Remark. In general, for any open set U X, we have an isomorphism (7.15)
: X (U ) Y 1 (U ) ,
116
whose inverse is given by the comorphism of 1 : Xreg Y ; note that Y (U ) = Y (U ) since Y is normal. Taking the direct limit over all neighborhoods U of a given point x X, we get an isomorphism (7.15 ) : X,x
Y,y .
j
yj 1 (x)
In other words, X is isomorphic to the direct image sheaf Y , see (1.12). We will prove later on the deep fact that the direct image of a coherent sheaf by a proper holomorphic map is always coherent ([Grauert 1960], see 9.?.1). Hence X = Y is a coherent sheaf over X .
8. Holomorphic Mappings and Extension Theorems

8.1. Rank of a Holomorphic Mapping Our goal here is to introduce the general concept of the rank of a holomorphic map and to relate the rank to the dimension of the bers. As in the smooth case, the rank is shown to satisfy semi-continuity properties. (8.1) Lemma. Let F : X Y be a holomorphic map from a complex space X to a complex space Y . a) If F is nite, i.e. proper with nite bers, then dim X b) If F is nite and surjective, then dim X = dim Y . Proof. a) Let x X, (Xj , x) an irreducible component and m = dim(Xj , x). If (Yk , y) are the irreducible components of Y at y = F (x), then (Xj , x) is contained in F 1 (Yk ), hence (Xj , x) is contained in one of the sets F 1 (Yk ). If p = dim(Yk , y), there is a ramied covering from some neighborhood of y in Yk onto a polydisk in Cp . Replacing X by some neighborhood of x in Xj and F by the nite map FXj : Xj , we may suppose that Y = and that X is irreducible, dim X = m. Let r = rank dFx0 be the maximum of the rank of the dierential of F on Xreg . Then r min{m, p} and the rank of dF is constant equal to r on a neighborhood U of x0 . The constant rank theorem implies that the bers F 1 (y) U are (m r)-dimensional submanifolds, hence m r = 0 and m = r p. b) We only have to show that dim X dim Y . Fix a regular point y Y of maximal dimension. By taking the restriction F : F 1 (U ) U to a small neighborhood U of y, we may assume that Y is an open subset of Cp . If dim X < dim Y , then X is a union of analytic manifolds of dimension < dim Y and Sards theorem implies that F (X) has zero Lebesgue measure in Y , a contradiction. (8.2) Proposition. For any holomorphic map F : X Y , the ber dimension dim F 1 (F (x)), x is an upper semi-continuous function of x. Proof. Without loss of generality, we may suppose that X is an analytic set in Cn , that F (X) is contained in a small neighborhood of F (x) in Y which is embedded in CN , dim Y .
117
and that x = 0, F (x) = 0. Set A = F 1 (0) and s = dim(A, 0). We can nd a linear form 1 1 on Cn such that dim(A 1 (0), 0) = s 1 ; in fact we need only select a point xj = 0 on each irreducible component (Aj , 0) of (A, 0) and take 1 (xj ) = 0. By induction, we can nd linearly independent forms 1 , . . . , s on Cn such that
1 1 dim A 1 (0) . . . j (0), 0 = s j
for all j = 1, . . . , s ; in particular 0 is an isolated point in the intersection when j = s. After a change of coordinates, we may suppose that j (z) = zj . Fix r so small that the ball B Cns of center 0 and radius r satises A ({0} B ) = {0}. Then A is disjoint from the compact set {0} B , so there exists a small ball B Cs of center 0 such that A (B B ) = , i.e. F does not vanish on the compact set K = X (B B ). Set = minK |F |. Then for |y| < the ber F 1 (y) does not intersect B B . This implies that the projection map : F 1 (y) (B B ) B is proper. The bers of are then compact analytic subsets of B , so they are nite by 5.9. Lemma 8.1 a) implies dim F 1 (y) (B B ) dim B = s = dim(A, 0) = dim(F 1 (0), 0).
Let X be a pure dimensional complex space and F : X Y a holomorphic map. For any point x X, we dene the rank of F at x by (8.3) F (x) = dim(X, x) dim F 1 (F (x)), x .
By the above proposition, F is a lower semi-continuous function on X. In particular, if F is maximum at some point x0 , it must be constant in a neighborhood of x0 . The maximum (F ) = maxX F is thus attained on Xreg or on any dense open subset X Xreg . If X is not pure dimensional, we dene (F ) = max (FX ) where (X ) are the irreducible components of X. For a map F : X CN , the constant rank theorem implies that (F ) is equal to the maximum of the rank of the jacobian matrix dF at points of Xreg (or of X ). (8.4) Proposition. If F : X Y is a holomorphic map and Z an analytic subset of X, then (FZ ) (F ). Proof. Since each irreducible component of Z is contained in an irreducible component of X, we may assume X irreducible. Let : X X be the normalization of X and 1 Z = (Z). Since is nite and surjective, the ber of F at point x has the same dimension than the ber of F at (x) by Lemma 8.1 b). Therefore (F ) = (F ) and (F Z ) = (FZ ), so we may assume X normal. By induction on dim X, we may also suppose that Z has pure codimension 1 in X (every point of Z has a neighborhood V X such that Z V is contained in a pure one codimensional analytic subset of V ). But then Zreg Xreg is dense in Zreg because codim Xsing 2. Thus we are reduced to the case when X is a manifold and Z a submanifold, and this case is clear if we consider the rank of the jacobian matrix. (8.5) Theorem. Let F : X Y be a holomorphic map. If Y is pure dimensional and (F ) < dim Y , then F (X) has empty interior in Y .
118
Proof. Taking the restriction of F to F 1 (Yreg ), we may assume that Y is a manifold. Since X is a countable union of compact sets, so is F (X), and Baires theorem shows that the result is local for X. By Prop. 8.4 and an induction on dim X, F (Xsing ) has empty interior in Y . The set Z Xreg of points where the jacobian matrix of F has rank < (F ) is an analytic subset hence, by induction again, F (Z) has empty interior. The constant rank theorem nally shows that every point x Xreg Z has a neighborhood V such that F (V ) is a submanifold of dimension (F ) in Y , thus F (V ) has empty interior and Baires theorem completes the proof. (8.6) Corollary. Let F : X Y be a surjective holomorphic map. Then dim Y = (F ). Proof. By the remark before Prop. 8.4, there is a regular point x0 X such that the jacobian matrix of F has rank (F ). Hence, by the constant rank theorem dim Y (F ). Conversely, let Y be an irreducible component of Y of dimension equal to dim Y , and Z = F 1 (Y ) X. Then F (Z) = Y and Th. 8.5 implies (F ) (FZ ) dim Y . 8.2. Remmert and Remmert-Stein Theorems We are now ready to prove two important results: the extension theorem for analytic subsets due to [Remmert and Stein 1953] and the theorem of [Remmert 1956, 1957] which asserts that the image of a complex space under a proper holomorphic map is an analytic set. These will be obtained by a simultaneous induction on the dimension. (8.7) Remmert-Stein theorem. Let X be a complex space, A an analytic subset of X and Z an analytic subset of X A. Suppose that there is an integer p 0 such that dim A p, while dim(Z, x) > p for all x Z. Then the closure Z of Z in X is an analytic subset. (8.8) Remmerts proper mapping theorem. Let F : X Y be a proper holomorphic map. Then F (X) is an analytic subset of Y . Proof. We let (8.7m ) denote statement (8.7) for dim Z m and (8.8m ) denote statement (8.8) for dim X m. We proceed by induction on m in two steps: Step 1. (8.7m ) and (8.8m1 ) imply (8.8m ). Step 2. (8.8m1 ) implies (8.7m ). As (8.8m ) is obvious for m = 0, our statements will then be valid for all m, i.e. for all complex spaces of bounded dimension. However, Th. 8.7 is local on X and Th. 8.8 is local on Y , so the general case is immediately reduced to the nite dimensional case. Proof of step 1. The analyticity of F (X) is a local question in Y . Since F : F 1 (U ) U is proper for any open set U Y and F 1 (U ) X if U Y , we may suppose that Y is embedded in an open set Cn and that X only has nitely many irreducible components X . Then we have F (X) = F (X ) and we are reduced to the case when X is irreducible, dim X = m and Y = . First assume that X is a manifold and that the rank of dF is constant. The constant rank theorem implies that every point in X has a neighborhood V such that F (V ) is a closed submanifold in a neighborhood W of F (x) in Y . For any point y Y , the
119
ber F 1 (y) can be covered by nitely many neighborhoods Vj of points xj F 1 (y) such that F (Vj ) is a closed submanifold in a neighborhood Wj of y. Then there is a neighborhood of y W Wj such that F 1 (W ) Vj , so F (X) W = F (Vj ) W is a nite union of closed submanifolds in W and F (X) is analytic in Y . Now suppose that X is a manifold, set r = (F ) and let Z X be the analytic subset of points x where the rank of dFx is < r. Since dim Z < m = dim X, the hypothesis (8.8m1 ) shows that F (Z) is analytic. We have dim F (Z) = (FZ ) < r. If F (Z) = F (X), then F (X) is analytic. Otherwise A = F 1 F (Z) is a proper analytic subset of X, dF has constant rank on X A X Z and the morphism F : X A Y F (Z) is proper. Hence the image F (X A) is analytic in Y F (Z). Since dim F (X A) = r m and dim F (Z) < r, hypothesis (8.7m ) implies that F (X) = F (X A) is analytic in Y . When X is not a manifold, we apply the same reasoning with Z = Xsing in order to be reduced to the case of F : X A Y F (Z) where X A is a manifold. Proof of step 2. Since Th. 8.7 is local on X, we may suppose that X is an open set Cn . Then we use induction on p to reduce the situation to the case when A is a p-dimensional submanifold (if this case is taken for granted, the closure of Z in Asing is analytic and we conclude by the induction hypothesis). By a local analytic change of coordinates, we may assume that 0 A and that A = L where L is a vector subspace of Cn of dimension p. By writing Z = p<s m Zs where Zs is an analytic subset of Y of pure dimension s, we may suppose that Z has pure dimension s, p < s m. We are going to show that Z is analytic in a neighborhood of 0. Let 1 be a linear form on Cn which is not identically zero on L nor on any irreducible component of Z (just pick a point x on each component and take 1 (x ) = 0 for all ). 1 1 Then dim L 1 (0) = p 1 and the analytic set Z 1 (0) has pure dimension s 1. By induction, there exist linearly independent forms 1 , . . . , s such that
(8.9)
1 1 dim Z 1 (0) . . . j (0) = s j,
1 1 dim L 1 (0) . . . j (0) = p j,
1 1
j j
p, s.
By adding a suitable linear combination of 1 , . . . , p to each j , p < j s, we may take jL = 0 for p < j s. After a linear change of coordinates, we may suppose that j (z) = zj , L = Cp {0} and A = (Cp {0}). Let = (1 , . . . , s ) : Cn Cs be the projection onto the rst s variables. As Z is closed in A, Z A is closed in . Moreover, our construction gives (Z A) 1 (0) = Z 1 (0) {0} and the case j = s of (8.9) shows that Z 1 (0) is a locally nite sequence in ({0} Cns ) {0}. Therefore, we can nd a small ball B of center 0 in Cns such that Z ({0} B ) = . As {0} B is compact and disjoint from the closed set Z A, there is a small ball B of center 0 in Cs such that (Z A) (B B ) = . This implies that the projection : (Z A) (B B ) B is proper. Set A = B (Cp {0}).
120
Z A B
B B B
A
S1 S2
Fig. II-3 Projection : Z ((B Then the restriction = : Z (B B )
A ) B ) B
A .
(A B ) B
is proper, and Z (B B ) is analytic in (B B ) A, so has nite bers by Th. 5.9. By denition of the rank we have () = s. Let S1 = Zsing 1 (B A ) and S1 = (S1 ) ; further, let S2 be the set of points x Z 1 B (A S1 ) Zreg such that dx has rank < s and S2 = (S2 ). We have dim Sj s 1 m 1. Hypothesis (8.8)m1 implies that S1 is analytic in B A and that S2 is analytic in B (A S1 ). By Remark 4.2, B (A S1 S2 ) is connected and every bounded holomorphic function on this set extends to B . As is a (non ramied) covering over B (A S1 S2 ), the sheet number is a constant q. Let (z) = j>s j zj be a linear form on Cn in the coordinates of index j > s. For z B (A S1 S2 ), we let j (z ) be the elementary symmetric functions in the q complex numbers (z) corresponding to z 1 (z ). Then these functions can be extended as bounded holomorphic functions on B and we get a polynomial P (z ; T ) such that P z ; (z ) vanishes identically on Z 1 (A S1 S2 ). Since is nite, Z 1 (A S1 S2 ) is a union of three (non necessarily closed) analytic subsets of dimension s 1, thus has empty interior in Z. It follows that the closure Z (B B ) is contained in the analytic set W B B equal to the common zero set of all functions P z ; (z ) . Moreover, by construction, Z
1 (A S1 S2 ) = W 1 (A S1 S2 ).
9. Complex Analytic Schemes
121
As in the proof of Cor. 5.4, we easily conclude that Z (B B ) is equal to the union of all irreducible components of W that are not contained in 1 (A S1 S2 ). Hence Z is analytic. Finally, we give two interesting applications of the Remmert-Stein theorem. We assume here that the reader knows what is the complex projective space Pn . For more details, see Sect. 5.15. (8.10) Chows theorem ([Chow 1949]). Let A be an analytic subset of the complex projective space Pn . Then A is algebraic, i.e. A is the common zero set of nitely many homogeneous polynomials Pj (z0 , . . . , zn ), 1 j N . Proof. Let : Cn+1 {0} Pn be the natural projection and Z = 1 (A). Then Z is an analytic subset of Cn+1 {0} which is invariant by homotheties and dim Z = dim A + 1 1. The Remmert-Stein theorem implies that Z = Z {0} is an analytic subset of Cn+1 . Let f1 , . . . , fN be holomorphic functions on a small polydisk Cn+1 1 of center 0 such that Z = fj (0). The Taylor series at 0 gives an expansion + fj = k=0 Pj,k where Pj,k is a homogeneous polynomial of degree k. We claim that Z coincides with the common zero W set of the polynomials Pj,k . In fact, we clearly have 1 W fj (0) = Z . Conversely, for z Z , the invariance of Z by homotheties shows that fj (tz) = Pj,k (z)tk vanishes for every complex number t of modulus < 1, so all coecients Pj,k (z) vanish and z W . By homogeneity Z = W ; since C[z0 , . . . , zn ] is Noetherian, W can be dened by nitely many polynomial equations. (8.11) E.E. Levis continuation theorem. Let X be a normal complex space and A an analytic subset such that dim(A, x) dim(X, x) 2 for all x A. Then every meromorphic function on X A has a meromorphic extension to X. Proof. We may suppose X irreducible, dim X = n. Let f be a meromorphic function on X A. By Th. 6.13, the pole set Pf has pure dimension (n 1), so the Remmert-Stein theorem implies that P f is analytic in X. Fix a point x A. There is a connected neighborhood V of x and a non zero holomorphic function h X (V ) such that P f V has nitely many irreducible components P f,j and P f V h1 (0). Select a point xj in P f,j (Xsing (P f )sing A). As xj is a regular point on X and on P f , there is a mj local coordinate z1,j at xj dening an equation of P f,j , such that z1,j f X,xj for some integer mj . Since h vanishes along Pf , we have hmj f X,x . Thus, for m = max{mj }, the pole set Pg of g = hm f in V A does not contain xj . As Pg is (n 1)-dimensional and contained in Pf V , it is a union of irreducible components P f,j A. Hence Pg must be empty and g is holomorphic on V A. By Cor. 7.8, g has an extension to a m holomorphic function g on V . Then g /h is the required meromorphic extension of f on V .

Our goal is to introduce a generalization of the notion of complex space given in Def. 5.2. A complex space is a space locally isomorphic to an analytic set A in an open subset Cn , together with the sheaf of rings A = ( /A )A . Our desire is to enrich the structure sheaf A by replacing A with a possibly smaller ideal dening the same
122
zero variety V () = A. In this way holomorphic functions are described not merely by their values on A, but also possibly by some transversal derivatives along A. 9.1. Ringed Spaces We start by an abstract notion of ringed space on an arbitrary topological space. (9.1) Denition. A ringed space is a pair (X, X ) consisting of a topological space X and of a sheaf of rings X on X, called the structure sheaf. A morphism F : (X, X ) (Y, Y ) of ringed spaces is a pair (f, F ) where f : X Y is a continuous map and F : f 1 Y X ,
Fx : (Y )f (x) (X )x
a homomorphism of sheaves of rings on X, called the comorphism of F . If F : (X, X ) (Y, Y ) and G : (Y, Y ) (Z, Z ) are morphisms of ringed spaces, the composite G F is the pair consisting of the map g f : X Z and of the comorphism (G F ) = F f 1 G : (9.2) X , F f 1 G : f 1 g 1 Z f 1 Y (X )x . Fx G (x) : (Z )gf (x) (Y )f (x) f
f 1 G F
9.2. Denition of Complex Analytic Schemes We begin by a description of what will be the local model of an analytic scheme. Let Cn be an open subset, a coherent sheaf of ideals and A = V () the analytic set in dened locally as the zero set of a system of generators of . By Hilberts Nullstellensatz 4.22 we have A = , but A diers in general from . The sheaf of rings / is supported on A, i.e. ( /)x = 0 if x A. Ringed spaces of the type / (A, /) will be used as the local models of analytic schemes. (9.3) Denition. A morphism F = (f, F ) : (A, /A ) (A , / ) A is said to be analytic if for every point x A there exists a neighborhood Wx of x in and a holomorphic function : Wx such that fAWx = AWx and such that the comorphism Fx : ( / )f (x) ( /)x
2 (9.4) Example. Take = Cn and = (zn ). Then A is the hyperplane Cn1 {0}, and the sheaf Cn / can be identied with the sheaf of rings of functions u + zn u , 2 u, u Cn1 , with the relation zn = 0. In particular, zn is a nilpotent element of Cn /. A morphism F of (A, Cn /) into itself is induced (at least locally) by a holomorphic
is induced by : ,f (x) u u ,x with .
123
map = (, n ) dened on a neighborhood of A in Cn with values in Cn , such that (A) A, i.e. nA = 0. We see that F is completely determined by the data f (z1 , . . . , zn1 )= (z1 , . . . , zn1 , 0), (z1 , . . . , zn1 , 0), f (z1 , . . . , zn1 )= zn which can be chosen arbitrarily. (9.5) Denition. A complex analytic scheme is a ringed space (X, X ) over a separable Hausdor topological space X, satisfying the following property: there exist an open covering (U ) of X and isomorphisms of ringed spaces G : (U , XU ) (A , / A ) f : Cn1 Cn , f : Cn1 Cn1 ,
where A is the zero set of a coherent sheaf of ideals on an open subset CN , such that every transition morphism G G1 is a holomorphic isomorphism from g (U U ) A onto g (U U ) A , equipped with the respective structure sheaves / A , / A . We shall often consider the maps G as identications and write simply U = A . A morphism F : (X, X ) (Y, Y ) of analytic schemes obtained by gluing patches (A , / A ) and (A , / A ), respectively, is a morphism F of ringed spaces such that for each pair (, ), the restriction of F from A f 1 (A ) X to A Y is holomorphic in the sense of Def. 9.3. Let (X, X ) be an analytic scheme. The set of nilpotent elements is the sheaf of ideals X dened by (9.6) X = {u X ; uk = 0 for some k N}. Locally, we have XA = ( / )A , thus of

9.3. Nilpotent Elements and Reduced Schemes
(9.7) (9.8)
ringed space The scheme (X, X /X ) is reduced by construction; it is called the reduced scheme of (X, X ). We shall often denote the original scheme by the letter X merely, the associated reduced scheme by Xred , and let X,red = X /X . There is a canonical morphism Xred X whose comorphism is the reduction morphism (9.9)
XA = ( / )A , (X /X )A ( / )A = ( /A )A = A . (X, X ) is said to be reduced if X = 0. The associated

X (U ) X,red (U ) = (X /X )(U ),
U open set in X.
By (9.8), the notion of reduced scheme is equivalent to the notion of complex space introduced in Def. 5.2. It is easy to see that a morphism F of reduced schemes X, Y is completely determined by the set-theoretic map f : X Y .
124
If (X, X ) is an analytic scheme, a sheaf of X -modules is said to be coherent if it satises the same properties as those already stated when X is a manifold: (9.10) is locally nitely generated over X ; (9.10 ) for any open set U X and any sections G1 , . . . , Gq (U ), the sheaf (G1 , . . . , Gq ) q is locally nitely generated. XU relation
9.4. Coherent Sheaves on Analytic Schemes
Locally, we have XA = / , so if i : A is the injection, the direct image = (i ) (A ) is a module over such that . = 0. It is clear that is coherent if and only if is coherent as a module over . It follows immediately that the Oka theorem and its consequences 3.1620 are still valid over analytic schemes. 9.5. Subschemes Let X be an analytic scheme and a coherent sheaf of ideals in X . The image of in X,red is a coherent sheaf of ideals, and its zero set Y is clearly an analytic subset of Xred . We can make Y into a scheme by introducing the structure sheaf (9.11)
= (X / )Y ,
and we have a scheme morphism F : (Y, Y ) (X, X ) such that f is the inclusion and F : f 1 X Y the obvious map of XY onto its quotient Y . The scheme (Y, Y ) will be denoted V ( ). When the analytic set Y is given, the structure sheaf of V ( ) depends of course on the choice of the equations of Y in the ideal ; in general Y has nilpotent elements. 9.6. Inverse Images of Coherent Sheaves Let F : (X, X ) (Y, Y ) be a morphism of analytic schemes and a coherent sheaf over Y . The sheaf theoretic inverse image f 1 , whose stalks are (f 1 )x = f (x) , is a sheaf of modules over f 1 Y . We dene the analytic inverse image F by (9.12) F = X f 1 Y f 1 , (F )x = X,x Y,f (x)
f (x) .
where A is a (q p)-matrix with coecients in Y (U ), our denition shows that F is a coherent sheaf over X , given over f 1 (U ) by the local presentation (9.13)
Here the tensor product is taken with respect to the comorphism F : f 1 Y X , which yields a ring morphism Y,f (x) X,x . If is given over U Y by a local presentation A p q U 0 Y U Y U
p Xf
1 (U)
q 1 (U) F f 1 (U) 0. Xf
F A
9.7. Products of Analytic Schemes
Let (X, X ) and (Y, Y ) be analytic schemes, and let (A , / ), (B , / ) be local models of X, Y , respectively. The product scheme (X Y, XY ) is obtained by gluing the open patches (9.14) A B ,
pr1 + pr1 2 1
125
In other words, if A , B are the subschemes of , dened by the equations g,j (x) = 0, g,k (y) = 0, where (g,j ) and (g,k ) are generators of and respectively, then A B is equipped with the structure sheaf g,j (x), g,k (y) . Now, let be a coherent sheaf over X and let be a coherent sheaf over (analytic) external tensor product is dened to be
Y . The
(9.15)
= pr 1
XY
pr . 2
If we go back to the denition of the inverse image, we see that the stalks of given by (9.15 ) ( )(x,y) = XY,(x,y) X,x Y,y (x C ) , y
are
in particular ( )(x,y) does not coincide with the sheaf theoretic tensor product x y which is merely a module over X,x Y,y . If and are given by local presentations
A p q U 0, p U XU XU Y then is the coherent sheaf given by

q U 0, U Y
B
pq qp XY UU

qq UU ( )UU 0. XY
(A(x)Id,Id B(y))
If x is a point of an analytic scheme (X, X ), the Zariski embedding dimension of the germ (X, x) is the smallest integer N such that (X, x) can be embedded in CN , i.e. such that there exists a patch of X near x isomorphic to (A, /) where is an open subset of CN . This dimension is denoted (9.16) embdim(X, x) = smallest such N.
9.8. Zariski Embedding Dimension
Consider the maximal ideal mX,x X,x of functions which vanish at point x. If (X, x) is embedded in (, x) = (CN , 0), then mX,x /m2 is generated by z1 , . . . , zN , so d = X,x N . Let s1 , . . . , sd be germs in m,x which yield a basis of mX,x /m2 dim mX,x /m2 X,x X,x m,x /(m2 + x ). We can write ,x zj =
1 k d
cjk sk + uj + fj ,
cjk C, uj m2 , fj x , 1 ,x
n.
Then we nd dzj = cjk dsk (x) + dfj (x), so that the rank of the system of dierentials dfj (x) is at least N d. Assume for example that df1 (x), . . . , dfNd (x) are linearly independant . By the implicit function theorem, the equations f1 = . . . = fNd = 0 dene a germ of smooth subvariety (Z, x) (, x) of dimension d which contains (X, x). We have Z = /(f1 , . . . , fNd ) in a neighborhood of x, thus
X = / Z /
(9.17)
where
= /(f1, . . . , fNd ).
This shows that (X, x) can be imbedded in Cd , and we get embdim(X, x) = dim mX,x /m2 . X,x
126
(9.18) Remark. For a given dimension n = dim(X, x), the embedding dimension d can be arbitrarily large. Consider for example the curve CN parametrized by C t (tN , tN+1 , . . . , t2N1 ). Then ,0 is the ring of convergent series in C{t} which have no terms t, t2 , . . . , tN1 , and m,0 /m2 admits precisely z1 = tN , . . . , zN = t2N1 ,0 as a basis. Therefore n = 1 but d = N can be as large as we want.
10. Bimeromorphic maps, Modications and Blow-ups

It is a very frequent situation in analytic or algebraic geometry that two complex spaces have isomorphic dense open subsets but are nevertheless dierent along some analytic subset. These ideas are made precise by the notions of modication and bimeromorphic map. This will also lead us to generalize the notion of meromorphic function to maps between analytic schemes. If (X, X ) is an analytic scheme, X denotes the sheaf of meromorphic functions on X, dened at the beginning of 6.2. (10.1) Denition. Let (X, X ), (Y, Y ) be analytic schemes. An analytic morphism F : X Y is said to be a modication if F is proper and if there exists a nowhere dense closed analytic subset B Y such that the restriction F : X F 1 (B) Y B is an isomorphism. (10.2) Denition. If F : X Y is a modication, then the comorphism F : f Y X induces an isomorphism F : f Y X for the sheaves of meromorphic functions on X and Y . Proof. Let v = g/h be a section of Y on a small open set where u is actually given as a quotient of functions g, h Y (). Then F u = (g F )/(h F ) is a section of X on F 1 (), for h F cannot vanish identically on any open subset W of F 1 () (otherwise h would vanish on the open subset F (W F 1 (B)) of B). Thus the extension of the comorphism to sheaves of meromorphic functions is well dened. Our claim is that this is an isomorphism. The injectivity of F is clear: F u = 0 implies g F = 0, which implies g = 0 on B and thus g = 0 on because B is nowhere dense. In order to prove surjectivity, we need only show that every section u X (F 1 ()) is in the image of Y () by F . For this, we may shrink into a relatively compact subset and thus assume that u is bounded (here we use the properness of F through the fact that F 1 ( ) is relatively compact in F 1 ()). Then v = u F 1 denes a bounded holomorphic function on B. By Th. 7.2, it follows that v is weakly holomorphic for the reduced structure of Y . Our claim now follows from the following Lemma. (10.3) Lemma. If (X, X ) is an analytic scheme, then every holomorphic function v in the complement of a nowhere dense analytic subset B Y which is weakly holomorphic on Xred is meromorphic on X. Proof. It is enough to argue with the germ of v at any point x Y , and thus we may suppose that (Y, Y ) = (A, /) is embedded in CN . Because v is weakly holomorphic, we can write v = g/h in Yred , for some germs of holomorphic functions g, h. Let g and h be extensions of g, h to ,x . Then there is a neighborhood U of x such that g v h is a nilpotent section of cO (U B) which is in on
11. Exercises
127
(10.4) Denition. A meromorphic map F : X - - Y is a scheme morphism F : X A Y dened in the complement of a nowhere dense analytic subset A X, such that the closure of the graph of F in X Y is an analytic subset (for the reduced complex space structure of X Y ).
11. Exercises
be a sheaf on a topological space X. If the sheaf space e is Hausdor, show that Let satises the following unique continuation principle: any two sections s, s (U) on a connected open set U which coincide on some non empty open subset V U must coincide identically on U. Show that the converse holds if X is Hausdor and locally connected.
11.1.
11.2.
Let be a sheaf of abelian groups on X and let s (X). The support of s, denoted Supp s, is dened to be {x X ; s(x) = 0}. Show that Supp s is a closed subset of X. The support of is dened to be Supp = {x X ; x = 0}. Show that Supp is not necessarily closed: if is an open set in X, consider the sheaf such that (U) is the set of continuous functions f (U) which vanish on a neighborhood of U (X ).
11.3. Let be a sheaf of rings on a topological space X and let , be sheaves of -modules. We dene a presheaf = om ( , ) such that (U) is the module of all sheaf-homomorphisms -linear. U U which are
a) Show that om ( , ) is a sheaf and that there exists a canonical homomorphism x : om ( , )x hom x ( x , x ) for every x X. is locally nitely generated, then x is injective, and if b) If (3.12), then x is bijective. has local nite presentations as in
c) Suppose that is a coherent sheaf of rings and that , are coherent modules over . Then om ( , ) is a coherent -module. Hint: observe that the result is true if = p and use a local presentation of to get the conclusion.
11.4.
Let f : X Y be a continuous map of topological spaces. Given sheaves of abelian groups on X and on Y , show that there is a natural isomorphism homX (f 1 , ) = homY ( , f ).
Hint: use the natural morphisms (2.17). Show that the sheaf of polynomials over Cn is a coherent sheaf of rings (with either the ordinary topology or the Zariski topology on Cn ). Extend this result to the case of regular algebraic functions on an algebraic variety. Hint: check that the proof of the Oka theorem still applies. Let P be a non zero polynomial on Cn . If P is irreducible in C[z1 , . . . , zn ], show that the hypersurface H = P 1 (0) is globally irreducible as an analytic set. In general, show that the irreducible components of H are in a one-to-one correspondence with the irreducible factors of P . Hint: for the rst part, take coordinates such that P (0, . . . , 0, zn ) has degree equal to P ; if H splits in two components H1 , H2 , then P can be written as a product P1 P2 where the roots of Pj (z , zn ) correspond to points in Hj .
11.5.
11.6.
11.7.
Prove the following facts:
a) For every algebraic variety A of pure dimension p in Cn , there are coordinates z = (z1 , . . . , zp ), z = (zp+1 , . . . , zn ) such that : A Cp , z z is proper with nite bers, and such that A is entirely contained in a cone |z | C(|z | + 1).
128
Chapter II. Coherent Sheaves and Analytic Spaces Hint: imitate the proof of Cor. 4.11.
b) Conversely if an analytic set A of pure dimension p in Cn is contained in a cone |z | C(|z | + 1), then A is algebraic. Hint: rst apply (5.9) to conclude that the projection : A Cp is nite. Then repeat the arguments used in the nal part of the proof of Th. 4.19. c) Deduce from a), b) that an algebraic set in Cn is irreducible if and only if it is irreducible as an analytic set. Let : f (x, y) = 0 be a germ of analytic curve in C2 through (0, 0) and let (j , 0) be the irreducible components of (, 0). Suppose that f (0, y) 0. Show that the roots y of f (x, y) = 0 corresponding to points of near 0 are given by Puiseux expansions of the form y = gj (x1/qj ), where gj C,0 and where qj is the sheet number of the projection j C, (x, y) x. Hint: special case of the parametrization obtained in (4.27).
11.8.
11.9. The goal of this exercise is to prove the existence and the analyticity of the tangent cone to an arbitrary analytic germ (A, 0) in Cn . Suppose that A is dened by holomorphic equations f1 = . . . = fN = 0 in a ball = B(0, r). Then the (set theoretic) tangent cone to A at 0 is the set C(A, 0) of all limits of sequences t1 z with z A and C t converging to 0.
a) Let E be the set of points (z, t) C such that z t1 A. Show that the closure E in C is analytic. Hint: observe that E = A ( {0}) where A = {fj (tz) = 0} and apply Cor. 5.4. b) Show that C(A, 0) is a conic set and that E ( {0}) = C(A, 0) {0} and conclude. Infer from this that C(A, 0) is an algebraic subset of Cn .
11.10.
Give a new proof of Theorem 5.5 based on the coherence of ideal sheaves and on the strong noetherian property.
11.11.
Let X be an analytic space and let A, B be analytic subsets of pure dimensions. Show that codimX (A B) codimX A + codimX B if A or B is a local complete intersection, but that the equality does not necessarily hold in general. Hint: see Remark (6.5).
Let be the curve in C3 parametrized by C t (x, y, z) = (t3 , t4 , t5 ). Show that the ideal sheaf is generated by the polynomials xz y 2 , x3 yz and x2 y z 2 , and that the germ (, 0) is not a (sheaf theoretic) complete intersection. P Hint: is smooth except at the origin. Let f (x, y, z) = a x y z be a convergent power series near 0. Show that f ,0 if and only if all weighted homogeneous components fk = X a x y z
11.12.
3+4+5=k
are in ,0 . By means of suitable substitutions, reduce the proof to the case when f = fk is homogeneous with all non zero monomials satisfying 2, 1, 1; then check that there is at most one such monomial in each weighted degree 15 the product of a power of x by a homogeneous polynomial of weighted degree 8.
129
Chapter III
Positive Currents and Lelong Numbers
In 1957, P. Lelong introduced natural positivity concepts for currents of pure bidimension (p, p) on complex manifolds. With every analytic subset is associated a current of integration over its set of regular points and all such currents are positive and closed. The important closedness property is proved here via the Skoda-El Mir extension theorem. Positive currents have become an important tool for the study of global geometric problems as well as for questions related to local algebra and intersection theory. We develope here a dierential geometric approach to intersection theory through a detailed study of wedge products of closed positive currents (Monge-Amp`re operators). The Lelong-Poincar equation e e and the Jensen-Lelong formula are basic in this context, providing a useful tool for studying the location and multiplicities of zeroes of entire functions on Cn or on a manifold, in relation with the growth at innity. Lelong numbers of closed positive currents are then introduced; these numbers can be seen as a generalization to currents of the notion of multiplicity of a germ of analytic set at a singular point. We prove various properties of Lelong numbers (e.g. comparison theorems, semi-continuity theorem of Siu, transformation under holomorphic maps). As an application to Number Theory, we prove a general Schwarz lemma in Cn and derive from it Bombieris theorem on algebraic values of meromorphic maps and the famous theorems of Gelfond-Schneider and Baker on the transcendence of exponentials and logarithms of algebraic numbers.
1. Basic Concepts of Positivity

1.A. Positive and Strongly Positive Forms Let V be a complex vector space of dimension n and (z1 , . . . , zn ) coordinates on V . We denote by (/z1 , . . . , /zn) the corresponding basis of V , by (dz1 , . . . , dzn ) its dual basis in V and consider the exterior algebra
VC =
p,q V ,
p,q V = p V q V .
We are of course especially interested in the case where V = Tx X is the tangent space to a complex manifold X, but we want to emphasize here that our considerations only involve linear algebra. Let us rst observe that V has a canonical orientation, given by the (n, n)-form (z) = idz1 dz 1 . . . idzn dz n = 2n dx1 dy1 . . . dxn dyn where zj = xj + iyj . In fact, if (w1 , . . . , wn ) are other coordinates, we nd dw1 . . . dwn = det(wj /zk ) dz1 . . . dzn , (w) = det(wj /zk ) (z).
2
130
Chapter III. Positive Currents and Lelong Numbers
In particular, a complex manifold always has a canonical orientation. More generally, natural positivity concepts for (p, p)-forms can be dened. (1.1) Denition. A (p, p)-form u p,p V is said to be positive if for all j V , 1 j q = n p, then u i1 1 . . . iq q
is a positive (n, n)-form. A (q, q)-form v q,q V is said to be strongly positive if v is a convex combination v= where s,j V and s 0.
2
s is,1 s,1 . . . is,q s,q
(1.2) Example. Since ip (1)p(p1)/2 = ip , we have the commutation rules i1 1 . . . ip p = ip ,

2 2 2 2
ip im = i(p+m) ,
= 1 . . . p p,0 V ,
p,0 V , m,0 V .
Take m = q to be the complementary degree of p. Then = dz1 . . . dzn for some 2 2 C and in = ||2 (z). If we set = 1 . . . q , we nd that ip is a positive (p, p)-form for every p,0 V ; in particular, strongly positive forms are positive. The sets of positive and strongly positive forms are closed convex cones, i.e. closed and stable under convex combinations. By denition, the positive cone is dual to the strongly positive cone via the pairing (1.3) p,p V q,q V C (u,v) u v/,
that is, u p,p V is positive if and only if u v 0 for all strongly positive forms q,q v V . Since the bidual of an arbitrary convex cone is equal to its closure , we also obtain that v is strongly positive if and only if v u = u v is 0 for all positive forms u. Later on, we will need the following elementary lemma. (1.4) Lemma. Let (z1 , . . . , zn ) be arbitrary coordinates on V . Then p,p V admits a basis consisting of strongly positive forms s = is,1 s,1 . . . is,p s,p , where each s,l is of the type dzj dzk or dzj idzk , 1 1 s j, k n p n.
2
Proof. Since one can always extract a basis from a set of generators, it is sucient to see that the family of forms of the above type generates p,p V . This follows from the identities 4dzj dz k = (dzj + dzk ) (dzj + dzk ) (dzj dzk ) (dzj dzk ) +i(dzj + idzk )(dzj + idzk )i(dzj idzk )(dzj idzk ),
131
dzj1 . . . dzjp dz k1 . . . dz kp =
1 s p
dzjs dz ks .
(1.5) Corollary. All positive forms u are real, i.e. satisfy u = u. In terms of coordinates, 2 if u = ip |I|=|J|=p uI,J dzI dz J , then the coecients satisfy the hermitian symmetry relation uI,J = uJ, I . Proof. Clearly, every strongly positive (q, q)-form is real. By Lemma 1.4, these forms generate over R the real elements of q,q V , so we conclude by duality that positive (p, p)-forms are also real. (1.6) Criterion. A form u p,p V is positive if and only if its restriction uS to every p-dimensional subspace S V is a positive volume form on S. Proof. If S is an arbitrary p-dimensional subspace of V we can nd linear coordinates (z1 , . . . , zn ) on V such that S = {zp+1 = . . . = zn = 0}. Then uS = S idz1 dz 1 . . . idzp dz p where S is given by u idzp+1 dz p+1 . . . idzn dz n = S (z). If u is positive then S 0 so uS is positive for every S. The converse is true because the (n p, n p)-forms j>p idzj dz j generate all strongly positive forms when S runs over all p-dimensional subspaces. (1.7) Corollary. A form u = i j,k ujk dzj dz k of bidegree (1, 1) is positive if and only if ujk j k is a semi-positive hermitian form on Cn . Proof. If S is the complex line generated by and t t the parametrization of S, then uS = ujk j k idt dt. Observe that there is a canonical one-to-one correspondence between hermitian forms and real (1, 1)-forms on V . The correspondence is given by (1.8) h=
1 j,k n
hjk (z) dzj dz k u = i
1 j,k n
hjk (z) dzj dz k
and does not depend on the choice of coordinates: indeed, as hjk = hkj , one nds immediately u(, ) = i hjk (z)(j k j k ) = 2 Im h(, ), , T X.
Moreover, h is 0 as a hermitian form if and only if u 0 as a (1, 1)-form. A diagonalization of h shows that every positive (1, 1)-form u 1,1 V can be written u=
1 j r
ij j ,
V , r = rank of u,
132
in particular, every positive (1, 1)-form is strongly positive. By duality, this is also true for (n 1, n 1)-forms. (1.9) Corollary. The notions of positive and strongly positive (p, p)-forms coincide for p = 0, 1, n 1, n. (1.10) Remark. It is not dicult to see, however, that positivity and strong positivity 2 dier in all bidegrees (p, p) such that 2 p n 2. Indeed, a positive form ip with p,0 V is strongly positive if and only if is decomposable as a product 1 . . . p . To see this, suppose that 2 2 ip = ip j j
1 j N
where all j p,0 V are decomposable. Take N minimal. The equality can be also written as an equality of hermitian forms ||2 = |j |2 if , j are seen as linear forms p 2 on V . The hermitian form || has rank one, so we must have N = 1 and = j , as desired. Note that there are many non decomposable p-forms in all degrees p such that 2 p n 2, e.g. (dz1 dz2 + dz3 dz4 ) . . . dzp+2 : if a p-form is decomposable, p1 the vector space of its contractions by elements of V is a p-dimensional subspace of V ; in the above example the dimension is p + 2. (1.11) Proposition. If u1 , . . . , us are positive forms, all of them strongly positive (resp. all except perhaps one), then u1 . . . us is strongly positive (resp. positive). Proof. Immediate consequence of Def. 1.1. Observe however that the wedge product of two positive forms is not positive in general (otherwise we would infer that positivity coincides with strong positivity). (1.12) Proposition. If : W V is a complex linear map and u p,p V is (strongly) positive, then u p,p W is (strongly) positive. Proof. This is clear for strong positivity, since (i1 1 . . . ip p ) = i1 1 . . . ip p with j = j W , for all j V . For u positive, we may apply Criterion 1.6: if S is a p-dimensional subspace of W , then u(S) and ( u)S = ( S) u(S) are positive when S : S (S) is an isomorphism; otherwise we get ( u)S = 0. 1.B. Positive Currents The duality between the positive and strongly positive cones of forms can be used to dene corresponding positivity notions for currents. (1.13) Denition. A current T (X) is said to be positive (resp. strongly positive) p,p if T, u 0 for all test forms u p,p (X) that are strongly positive (resp. positive) at each point. The set of positive (resp. strongly positive) currents of bidimension (p, p) will be denoted + resp. p,p (X), p,p (X).
133
It is clear that (strong) positivity is a local property and that the sets (X) p,p + (X) are closed convex cones with respect to the weak topology. Another way of p,p stating Def. 1.13 is: T is positive (strongly positive) if and only if T u all strongly positive (positive) forms u (X). p,p
0,0 (X)
is a positive measure for 0 for every non-negative
This is so because a distribution S (X) such that S(f ) function f (X) is a positive measure.
2
(1.14) Proposition. Every positive current T = i(np) TI,J dzI dz J in + (X) is p,p real and of order 0, i.e. its coecients TI,J are complex measures and satisfy TI,J = TJ, I for all multi-indices |I| = |J| = n p. Moreover TI,I 0, and the absolute values |TI,J | of the measures TI,J satisfy the inequality I J |TI,J | where k 2p
M
2 TM,M , M
I J M I J
kI
0 are arbitrary coecients and I =
k .
Proof. Since positive forms are real, positive currents have to be real by duality. Let us denote by K = I and L = J the ordered complementary multi-indices of I, J in {1, 2, . . . , n}. The distribution TI,I is a positive measure because TI,I = T ip dzK dz K On the other hand, the proof of Lemma 1.4 yields TI,J = T ip dzK dz L = a =
1 s p
2 2
0.
a(Z/4Z)p
a T a
where a = 1, i.
i (dzks + ias dzls ) (dzks + ias dzls ), 4
Now, each T a is a positive measure, hence TI,J is a complex measure and |TI,J |
a
T a = T
a
a
=T =T
1 s p
as Z/4Z
i (dzks + ias dzls ) (dzks + ias dzls ) 4
1 s p
idzks dz ks + idzls dz ls .
The last wedge product is a sum of at most 2p terms, each of which is of the type 2 2 ip dzM dz M with |M | = p and M K L. Since T ip dzM dz M = TM,M and M K L = I J, we nd |TI,J | 2p
M IJ
TM,M .
134
Now, consider a change of coordinates (z1 , . . . , zn ) = w = (1 w1 , . . . , n wn ) with 1 , . . . , n > 0. In the new coordinates, the current T becomes T and its coecients become I J TI,J (w). Hence, the above inequality implies I J |TI,J | 2p
M IJ
2 TM,M . M
This inequality is still true for k 0 by passing to the limit. The inequality of Prop. 1.14 follows when all coecients k , k I J, are replaced by 0, so that M = 0 for M I J. / (1.15) Remark. If T is of order 0, we dene the mass measure of T by T = |TI,J | (of course T depends on the choice of coordinates). By the Radon-Nikodym theorem, we can write TI,J = fI,J T with a Borel function fI,J such that |fI,J | = 1. Hence T = T f, where f = i(np)
2
fI,J dzI dz J .
Then T is (strongly) positive if and only if the form f (x) np,np Tx X is (strongly) positive at T -almost all points x X. Indeed, this condition is clearly sucient. On the other hand, if T is (strongly) positive and uj is a sequence of forms with constant coecients in p,p T X which is dense in the set of strongly positive (positive) forms, then T uj = ||T || f uj , so f (x) uj has to be a positive (n, n)-form except perhaps for x in a set N (uj ) of T -measure 0. By a simple density argument, we see that f (x) is (strongly) positive outside the T -negligible set N = N (uj ).
As a consequence of this proof, T is positive (strongly positive) if and only if T u is a positive measure for all strongly positive (positive) forms u of bidegree (p, p) with constant coecients in the given coordinates (z1 , . . . , zn ). It follows that if T is (strongly) positive in a coordinate patch , then the convolution T is (strongly) positive in = {x ; d(x, ) > }.
0 (1.16) Corollary. If T (X) and v Cs,s (X) are positive, one of them (resp. p,p both of them) strongly positive, then the wedge product T v is a positive (resp. strongly positive) current.
This follows immediately from Remark 1.15 and Prop. 1.11 for forms. Similarly, Prop. 1.12 on pull-backs of positive forms easily shows that positivity properties of currents are preserved under direct or inverse images by holomorphic maps. (1.17) Proposition. Let : X Y be a holomorphic map between complex analytic manifolds. a) If T b) If T
+ p,p (X)
and Supp T is proper, then T
+ p,p (Y
).
+ p,p (Y ) + p+m,p+m (X).
and if is a submersion with m-dimensional bers, then T
Similar properties hold for strongly positive currents.
135
1.C. Basic Examples of Positive Currents We present here two fundamental examples which will be of interest in many circumstances. 1.18. Current Associated to a Plurisubharmonic Function. Let X be a complex manifold and u Psh(X) L1 (X) a plurisubharmonic function. Then loc T = id d u = i
1 j,k n
2u dzj dz k zj z k
is a positive current of bidegree (1, 1). Moreover T is closed (we always mean here dclosed, that is, dT = 0). Assume conversely that is a closed real (1, 1)-current on X. Poincars lemma implies that every point x0 X has a neighborhood 0 such e that = dS with S (0 , R). Write S = S 1,0 + S 0,1 , where S 0,1 = S 1,0 . Then 1 d S = 0,2 = 0, and the Dolbeault-Grothendieck lemma shows that S 0,1 = d v on some neighborhood 0 , with v (, C). Thus S = d v + d v = d v + d v, = dS = d d (v v) = id d u, where u = 2 Re v (, R). If (X), the hypoellipticity of d in bidegree (p, 0) 1,1 shows that d u is of class , so u (). When is positive, the distribution u is a plurisubharmonic function (Th. I.3.31). We have thus proved: (1.19) Proposition. If + n1,n1 (X) is a closed positive current of bidegree (1, 1), then for every point x0 X there exists a neighborhood of x0 and u Psh() such that = id d u. 1.20. Current of Integration on a Complex Submanifold. Let Z X be a closed p-dimensional complex submanifold with its canonical orientation and T = [Z]. Then T (X). Indeed, every (r, s)-form of total degree r + s = 2p has zero restriction p,p to Z unless (r, s) = (p, p), therefore we have [Z] (X). Now, if u p,p (X) is a p,p positive test form, then uZ is a positive volume form on Z by Criterion 1.6, therefore [Z], u =
Z
uZ
0.
In this example the current [Z] is also closed, because d[Z] = [Z] = 0 by Stokes theorem. 1.D. Trace Measure and Wirtingers Inequality We discuss now some questions related to the concept of area on complex submanifolds. Assume that X is equipped with a hermitian metric h, i.e. a positive denite hermitian form h = hjk dzj dz k of class ; we denote by = i hjk dzj dz k (X) 1,1 the associated positive (1, 1)-form. (1.21) Denition. For every T the positive measure
+ p,p (X),
the trace measure of T with respect to is T p.
T =
1 2p p!
136
If (1 , . . . , n ) is an orthonormal frame of T X with respect to h on an open subset U X, we may write =i

1 j n
j j ,
p = ip p!
|K|=p
K K ,
T = i(np)
|I|=|J|=np
TI,J I J ,
TI,J
(U ),
where I = i1 . . . inp . An easy computation yields (1.22) T = 2p

|I|=np
TI,I i1 1 . . . in n . dzj dz j , we get in particular
For X = Cn with the standard hermitian metric h = (1.22 ) T = 2p

|I|=np
TI,I idz1 dz 1 . . . idzn dz n .
Proposition 1.14 shows that the mass measure ||T || = |TI,J | of a positive current T is always dominated by CT where C > 0 is a constant. It follows easily that the weak topology of (X) and of 0 (X) coincide on + (X), which is moreover a metrizable p p p subspace: its weak topology is in fact dened by the collection of semi-norms T | T, f | where (f ) is an arbitrary dense sequence in p (X). By the Banach-Alaoglu theorem, the unit ball in the dual of a Banach space is weakly compact, thus: (1.23) Proposition. Let be a positive continuous function on X. Then the set of currents T + (X) such that X T p 1 is weakly compact. p Proof. Note that our set is weakly closed, since a weak limit of positive currents is positive and X T p = sup T, p when runs over all elements of (X) such that 0 1. Now, let Z be a p-dimensional complex analytic submanifold of X. We claim that (1.24) [Z] = 1 2p p! [Z] p = Riemannian volume measure on Z.
This result is in fact a special case of the following important inequality. (1.25) Wirtingers inequality. Let Y be an oriented real submanifold of class C 1 and dimension 2p in X, and let dVY be the Riemannian volume form on Y associated with the metric hY . Set 1 p = dVY , C 0 (Y ). p p! Y 2 Then || 1 and the equality holds if and only if Y is a complex analytic submanifold of X. In that case = 1 if the orientation of Y is the canonical one, = 1 otherwise.
2. Closed Positive Currents
137
Proof. The restriction Y is a real antisymmetric 2-form on T Y . At any point z Y , we can thus nd an oriented orthonormal R-basis (e1 , e2 , . . . , e2p ) of Tz Y such that 1 = 2 k =
k e 2k1 e2k
on Tz Y,
where
1 k p
1 (e2k1 , e2k ) = Im h(e2k1 , e2k ). 2
We have dVY = e . . . e by denition of the Riemannian volume form. By taking 1 2p the p-th power of , we get 1 2p p!
p Tz Y = 1 . . . p e . . . e = 1 . . . p dVY . 1 2p
Since (ek ) is an orthonormal R-basis, we have Re h(e2k1 , e2k ) = 0, thus h(e2k1 , e2k ) = ik . As |e2k1 | = |e2k | = 1, we get 0 Therefore |k | 1, || = |1 . . . p | 1, with equality if and only if k = 1 for all k, i.e. e2k = Je2k1 . In this case Tz Y Tz X is a complex vector subspace at every point z Y , thus Y is complex analytic by Lemma I.4.23. Conversely, assume that Y is a C-analytic submanifold and that (e1 , e3 , . . . , e2p1 ) is an orthonormal complex basis of Tz Y . If e2k := Je2k1 , then (e1 , . . . , e2p ) is an orthonormal R-basis corresponding to the canonical orientation and 1 Y = 2
e 2k1 e2k ,
|e2k Je2k1 |2 = 2 1 Re h(Je2k1 , e2k ) = 2(1 k ).
1 2p p!
1 k p
p Y = e . . . e = dVY . 1 2p
Note that in the case of the standard hermitian metric on X = Cn , the form = i dzj dz j = d i zj dz j is globally exact. Under this hypothesis, we are going to see that C-analytic submanifolds are always minimal surfaces for the Plateau problem, which consists in nding a compact subvariety Y of minimal area with prescribed boundary Y . (1.26) Theorem. Assume that the (1, 1)-form is exact, say = d with 1 (X, R), and let Y, Z X be (2p)-dimensional oriented compact real submanifolds of class C 1 with boundary. If Y = Z and Z is complex analytic, then Vol(Y ) Vol(Z).
Proof. Write = d. Wirtingers inequality and Stokes theorem imply Vol(Y ) 1 2p p!

Y
p =
1 2p p!
Y
d( p1 ) =
1 2p p!
Y Y
p1 ,
1 Vol(Z) = p 2 p!
1 p = p 2 p! Z
1 p1 = p 2 p! Z
p1 .
138

2.A. The Skoda-El Mir Extension Theorem We rst prove the Skoda-El Mir extension theorem ([Skoda 1982], [El Mir 1984]), which shows in particular that a closed positive current dened in the complement of an analytic set E can be extended through E if (and only if) the mass of the current is locally nite near E. El Mir simplied Skodas argument and showed that it is enough to assume E complete pluripolar. We follow here the exposition of Sibonys survey article [Sibony 1985]. (2.1) Denition. A subset E X is said to be complete pluripolar in X if for every point x0 X there exist a neighborhood of x0 and a function u Psh() L1 () loc such that E = {z ; u(z) = }. Note that any closed analytic subset A X is complete pluripolar: if g1 = . . . = gN = 0 are holomorphic equations of A on an open set X, we can take u = log(|g1 |2 + . . . + |gN |2 ). (2.2) Lemma. Let E X be a closed complete pluripolar set. If x0 X and is a suciently small neighborhood of x0 , there exists: a) a function v Psh()
E) such that v = on E ;
b) an increasing sequence vk Psh() (), 0 vk 1, converging uniformly to 1 on every compact subset of E, such that vk = 0 on a neighborhood of E . Proof. Assume that 0 X is a coordinate patch of X containing x0 and that E 0 = {z 0 ; u(z) = }, u Psh(0 ). In addition, we can achieve u 0 by shrinking 0 and subtracting a constant to u. Select a convex increasing function ([0, 1], R) such that (t) = 0 on [0, 1/2] and (1) = 1. We set uk = exp(u/k) . Then 0 uk 1, uk is plurisubharmonic on 0 , uk = 0 in a neighborhood k of E 0 and lim uk = 1 on 0 E. Let 0 be a neighborhood of x0 , let 0 = d(, 0 ) and k ]0, 0 [ be such that k < d(E , k ). Then wk := max{uj j } Psh() C 0 (),
j k
0 wk 1, wk = 0 on a neighborhood of E and wk is an increasing sequence converging to 1 on E (note that wk uk ). Hence, the convergence is uniform on every compact subset of E by Dinis lemma. We may therefore choose a subsequence wks such that wks (z) 1 2s on an increasing sequence of open sets with = E. s s Then
+ s=0
w(z) := |z|2 +
(wks (z) 1)
is a strictly plurisubharmonic function on that is continuous on E, and w = on E . Richbergs theorem I.3.40 applied on E produces v Psh( E) ( E)
139
such that w v w+1. If we set v = on E , then v is plurisubharmonic on and has the properties required in a). After subtraction of a constant, we may assume v 0 on . Then the sequence (vk ) of statement b) is obtained by letting vk = exp(v/k) . (2.3) Theorem (El Mir). Let E X be a closed complete pluripolar set and T + E) a closed positive current. Assume that T has nite mass in a neighborhood p,p (X of every point of E. Then the trivial extension T + (X) obtained by extending the p,p measures TI,J by 0 on E is closed on X. Proof. The statement is local on X, so we may work on a small open set such that there exists a sequence vk Psh() () as in 2.2 b). Let ([0, 1]) be a function such that = 0 on [0, 1/3], = 1 on [2/3, 1] and 0 1. Then vk = 0 near E and vk = 1 for k large on every xed compact subset of E. Therefore T = limk+ ( vk )T and d T = lim T d ( vk )
k+
in the weak topology of currents. It is therefore sucient to check that T d ( vk ) converges weakly to 0 in p1,p () (note that d T is conjugate to d T , thus d T will also vanish). Assume rst that p = 1. Then T d ( vk )
0,1 (),
and we have to show that

1,0 (), 1,0 ().
T d ( vk ), = T, (vk )d vk 0, As T, i is a non-negative hermitian form on inequality yields T, i Let (), 0

2
the Cauchy-Schwarz
T, i
T, i ,
1,0 ().
1, be equal to 1 in a neighborhood of Supp . We nd

2
T, (vk )d vk
T, id vk d vk
T, (vk )2 i .
By hypothesis E T i < + and (vk ) converges everywhere to 0 on , thus T, (vk )2 i converges to 0 by Lebesgues dominated convergence theorem. On the other hand 2 id d vk = 2vk id d vk + 2id vk d vk 2id vk d vk , 2 T, id vk d vk
2 T, id d vk .
As yields
(), vk = 0 near E and d T = d T = 0 on

2 2 T, id d vk = T, vk id d
E, an integration by parts T < +
C
E
where C is a bound for the coecients of . Thus T, id vk d vk is bounded, and the proof is complete when p = 1. In the general case, let s = is,1 s,1 . . . is,p1 s,p1 be a basis of forms of bidegree (p1, p1) with constant coecients (Lemma 1.4). Then T s + ( E) 1,1
140
has nite mass near E and is closed on for all s, and we conclude that dT = 0.
E. Therefore d(T s ) = (dT ) s = 0 on
(2.4) Corollary. If T + (X) is closed, if E X is a closed complete pluripolar set p,p and 1 E is its characteristic function, then 1 E T and 1 X E T are closed (and, of course, l l l positive). Proof. If we set = TX and 1 E T = T . l
E,
then has nite mass near E and we have 1 X l
ET
2.B. Current of Integration over an Analytic Set Let A be a pure p-dimensional analytic subset of a complex manifold X. We would like to generalize Example 1.20 and to dene a current of integration [A] by letting (2.5) [A], v =
Areg
v,
p,p (X).
One diculty is of course to verify that the integral converges near Asing . This follows from the following lemma, due to [Lelong 1957]. (2.6) Lemma. The current [Areg ] of every point z0 Asing .
+ p,p (X
Asing ) has nite mass in a neighborhood
Proof. Set T = [Areg ] and let z0 be a coordinate open set. If we write the monomials dzK dz L in terms of an arbitrary basis of p,p T consisting of decomposable forms s = is,1 s,1 . . . s,p s,p (cf. Lemma 1.4), we see that the measures TI,J . are linear combinations of the positive measures T s . It is thus sucient to prove that all T s have nite mass near Asing . Without loss of generality, we may assume that (A, z0 ) is irreducible. Take new coordinates w = (w1 , . . . , wn ) such that wj = s,j (z z0 ), 1 j p. After a slight perturbation of the s,j , we may assume that each projection s : A ( ), w w = (w1 , . . . , wp )
denes a ramied covering of A (cf. Prop. II.3.8 and Th. II.3.19), and that (s ) remains a basis of p,p T . Let S be the ramication locus of s and AS = A ( S) Areg . The restriction of s : AS S is then a covering with nite sheet number qs and we nd

[Areg ] s =
Areg ( )
idw1 dw1 . . . idwp dwp

S
=
AS
idw1 dw1 . . . dwp = qs
idw1 dw1 . . . dwp < +.
The second equality holds because AS is the complement in Areg ( ) of an analytic subset (such a set is of zero Lebesgue measure in Areg ). (2.7) Theorem ([Lelong 1957]). For every pure p-dimensional analytic subset A X, the current of integration [A] + (X) is a closed positive current on X. p,p Proof. Indeed, [Areg ] has nite mass near Asing and [A] is the trivial extension of [Areg ] to X through the complete pluripolar set E = Asing . Theorem 2.7 is then a consequence of El Mirs theorem.
141
2.C. Support Theorems and Lelong-Poincar Equation e Let M X be a closed C 1 real submanifold of X. The holomorphic tangent space at a point x M is (2.8)
h
Tx M = Tx M JTx M,
that is, the largest complex subspace of Tx X contained in Tx M . We dene the CauchyRiemann dimension of M at x by CRdimx M = dimC h Tx M and say that M is a CR submanifold of X if CRdimx M is a constant. In general, we set (2.9) CRdim M = max CRdimx M = max dimC h Tx M.
xM xM
A current is said to be normal if and d are currents of order 0. For instance, every closed positive current is normal. We are going to prove two important theorems describing the structure of normal currents with support in CR submanifolds. (2.10) First theorem of support. Let (X) be a normal current. If Supp p,p is contained in a real submanifold M of CR dimension < p, then = 0. Proof. Let x0 M and let g1 , . . . , gm be real C 1 functions in a neighborhood of x0 such that M = {z ; g1 (z) = . . . = gm (z) = 0} and dg1 . . . dgm = 0 on . Then
h
T M = T M JT M =
1 k m
ker dgk ker(dgk J) =
ker d gk
1 k m
because d gk = 1 dgk i(dgk ) J . As dimC h T M < p, the rank of the system of (1, 0)2 forms (d gk ) must be > n p at every point of M . After a change of the ordering, we may assume for example that 1 = d g1 , 2 = d g2 , . . ., np+1 = d gnp+1 are linearly independent on (shrink if necessary). Complete (1 , . . . , np+1 ) into a continuous frame (1 , . . . , n ) of T X and set =
|I|=|J|=np
I,J I J
on .
As and d have measure coecients supported on M and gk = 0 on M , we get gk = gk d = 0, thus d gk = d (gk ) gk d = 0, 1 k m,
in particular k = 0 for all 1 k n p + 1. When |I| = n p, the multi-index I contains at least one of the elements 1, . . . , n p + 1, hence I J = 0 and I,J = 0. (2.11) Corollary. Let (X) be a normal current. If Supp is contained in an p,p analytic subset A of dimension < p, then = 0. Proof. As Areg is a submanifold of CRdim < p in X Asing , Theorem 2.9 shows that Supp Asing and we conclude by induction on dim A.
142
Now, assume that M X is a CR submanifold of class C 1 with CRdim M = p and that h T M is an integrable subbundle of T M ; this means that the Lie bracket of two vector elds in h T M is in h T M . The Frobenius integrability theorem then shows that M is locally bered by complex analytic p-dimensional submanifolds. More precisely, in a neighborhood of every point of M , there is a submersion : M Y onto a real C 1 manifold Y such that the tangent space to each ber Ft = 1 (t), t Y , is the holomorphic tangent space h T M ; moreover, the bers Ft are necessarily complex analytic in view of Lemma 1.7.18. Under these assumptions, with any complex measure on Y we associate a current with support in M by (2.12) =
tY
[Ft ] d(t),
i.e.
, u =
tY Ft
u d(t)
for all u (X). Then clearly (X) is a closed current of order 0, for all p,p p,p bers [Ft ] have the same properties. When the bers Ft are connected, the following converse statement holds: (2.13) Second theorem of support. Let M X be a CR submanifold of CR dimension p such that there is a submersion : M Y of class C 1 whose bers Ft = 1 (t) are connected and are the integral manifolds of the holomorphic tangent space h T M . Then any closed current (X) of order 0 with support in M can be written p,p = Y [Ft ] d(t) with a unique complex measure on Y . Moreover is (strongly) positive if and only if the measure is positive. Proof. Fix a compact set K Y and a C 1 retraction from a neighborhood V of M onto M . By means of a partition of unity, it is easy to construct a positive form 0 (V ) p,p such that Ft = 1 for each ber Ft with t K. Then the uniqueness and positivity statements for follow from the obvious formula f (t) d(t) = , (f ) , f C 0 (Y ), Supp f K.
Now, let us prove the existence of . Let x0 M . There is a small neighborhood of x0 and real coordinates (x1 , y1 , . . . , xp , yp , t1 , . . . , tq , g1 , . . . , gm ) such that zj = xj +iyj , 1 j on all bers Ft . p, are holomorphic functions on that dene complex coordinates
t1 , . . . , tq restricted to M are pull-backs by : M Y of local coordinates on an open set U Y such that : M U is a trivial ber space. g1 = . . . = gm = 0 are equations of M in . Then T Ft = {dtj = dgk = 0} equals h T M = {d gk = 0} and the rank of (d g1 , . . . , d gm ) is equal to n p at every point of M . After a change of the ordering we may suppose that 1 = d g1 , . . ., np = d gnp are linearly independent on . As in Prop. 2.10, we k n p and infer that I J = 0 unless get k = k = 0 for 1 I = J = L where L = {1, 2, . . . , n p}. Hence = L,L 1 . . . np 1 . . . np on .
143
Now 1 . . . np is proportional to dt1 . . . dtq dg1 . . . dgm because both induce a volume form on the quotient space T XM /h T M . Therefore, there is a complex measure supported on M such that = dt1 . . . dtq dg1 . . . dgm on .
As is supposed to be closed, we have /xj = /yj = 0. Hence is a measure depending only on (t, g), with support in g = 0. We may write = dU (t) 0 (g) where U is a measure on U = (M ) and 0 is the Dirac measure at 0. If j : M X is the injection, this means precisely that = j U on , i.e. =
tU
[Ft ] dU (t)
on .
The uniqueness statement shows that for two open sets 1 , 2 as above, the associated measures U1 and U2 coincide on (M 1 2 ). Since the bers Ft are connected, there is a unique measure which coincides with all measures U . (2.14) Corollary. Let A be an analytic subset of X with global irreducible components Aj of pure dimension p. Then any closed current (X) of order 0 with support p,p in A is of the form = j [Aj ] where j C. Moreover, is (strongly) positive if and only if all coecients j are 0. Proof. The regular part M = Areg is a complex submanifold of X Asing and its connected components are Aj Areg . Thus, we may apply Th. 2.13 in the case where Y is discrete to see that = j [Aj ] on X Asing . Now dim Asing < p and the dierence j [Aj ] p,p (X) is a closed current of order 0 with support in Asing , so this current must vanish by Cor. 2.11. (2.15) Lelong-Poincar equation. Let f (X) be a meromorphic function which e does not vanish identically on any connected component of X and let mj Zj be the divisor of f . Then the function log |f | is locally integrable on X and satises the equation i d d log |f | = in the space
n1,n1 (X)
mj [Zj ]
of currents of bidimension (n 1, n 1).
We refer to Sect. 2.6 for the denition of divisors, and especially to (2.6.14). Observe that if f is holomorphic, then log |f | Psh(X), the coecients mj are positive integers and the right hand side is a positive current in + n1,n1 (X). Proof. Let Z = Zj be the support of div(f ). Observe that the sum in the right hand side is locally nite and that d d log |f | is supported on Z, since d log |f |2 = d log(f f ) = f df df = f ff on X Z.
In a neighborhood of a point a Zj Zreg , we can nd local coordinates (w1 , . . . , wn ) such that Zj is given by the equation w1 = 0. Then Th. 2.6.6 shows that f can
144
m
be written f (w) = u(w)w1 j with an invertible holomorphic function u on a smaller neighborhood . Then we have id d log |f | = id d log |u| + mj log |w1 | = mj id d log |w1 |. For z C, Cor. I.3.4 implies id d log |z|2 = id dz z = i0 dz dz = 2 [0].
If : Cn C is the projection z z1 and H Cn the hyperplane {z1 = 0}, formula (1.2.19) shows that id d log |z1 | = id d log |(z)| = (id d log |z|) = ([0]) = [H],
i because is a submersion. We get therefore d d log |f | = mj [Zj ] in . This implies that the Lelong-Poincar equation is valid at least on X Zsing . As dim Zsing < n 1, e Cor. 2.11 shows that the equation holds everywhere on X.
3. Denition of Monge-Amp`re Operators e

Let X be a n-dimensional complex manifold. We denote by d = d + d the usual decomposition of the exterior derivative in terms of its (1, 0) and (0, 1) parts, and we set dc = 1 (d d ). 2i
i It follows in particular that dc is a real operator, i.e. dc u = dc u, and that ddc = d d . Although not quite standard, the 1/2i normalization is very convenient for many purposes, since we may then forget the factor 2 almost everywhere (e.g. in the Lelong-Poincar e equation (2.15)). In this context, we have the following integration by part formula.
(3.1) Formula. Let X be a smoothly bounded open set in X and let f, g be forms of class C 2 on of pure bidegrees (p, p) and (q, q) with p + q = n 1. Then
f ddc g ddc f g =
f dc g dc f g.
Proof. By Stokes theorem the right hand side is the integral over of d(f dc g dc f g) = f ddc g ddc f g + (df dc g + dc f dg). As all forms of total degree 2n and bidegree = (n, n) are zero, we get df dc g = 1 (d f d g d f d g) = dc f dg. 2i
Let u be a plurisubharmonic function on X and let T be a closed positive current of bidimension (p, p), i.e. of bidegree (n p, n p). Our desire is to dene the wedge
145
product ddc u T even when neither u nor T are smooth. A priori, this product does not make sense because ddc u and T have measure coecients and measures cannot be multiplied; see [Kiselman 1983] for interesting counterexamples. Assume however that u is a locally bounded plurisubharmonic function. Then the current uT is well dened since u is a locally bounded Borel function and T has measure coecients. According to [Bedford-Taylor 1982] we dene ddc u T = ddc (uT ) where ddc ( ) is taken in the sense of distribution (or current) theory. (3.2) Proposition. The wedge product ddc u T is again a closed positive current. Proof. The result is local. In an open set Cn , we can use convolution with a family of regularizing kernels to nd a decreasing sequence of smooth plurisubharmonic functions uk = u 1/k converging pointwise to u. Then u uk u1 and Lebesgues dominated convergence theorem shows that uk T converges weakly to uT ; thus ddc (uk T ) converges weakly to ddc (uT ) by the weak continuity of dierentiations. However, since uk is smooth, ddc (uk T ) coincides with the product ddc uk T in its usual sense. As T 0 and as ddc uk is a positive (1, 1)-form, we have ddc uk T 0, hence the weak limit ddc u T is 0 (and obviously closed). Given locally bounded plurisubharmonic functions u1 , . . . , uq , we dene inductively ddc u1 ddc u2 . . . ddc uq T = ddc (u1 ddc u2 . . . ddc uq T ). By (3.2) the product is a closed positive current. In particular, when u is a locally bounded plurisubharmonic function, the bidegree (n, n) current (ddc u)n is well dened and is a positive measure. If u is of class C 2 , a computation in local coordinates gives (ddc u)n = det 2u zj z k n! idz1 dz 1 . . . idzn dz n . n
The expression Monge-Amp`re operator classically refers to the non-linear partial dife ferential operator u det( 2 u/zj z k ). By extension, all operators (ddc )q dened above are also called Monge-Amp`re operators. e Now, let be a current of order 0. When K X is an arbitrary compact subset, we dene a mass semi-norm ||||K = |I,J |
Kj
I,J
by taking a partition K = Kj where each K j is contained in a coordinate patch and where I,J are the corresponding measure coecients. Up to constants, the semi-norm ||||K does not depend on the choice of the coordinate systems involved. When K itself is contained in a coordinate patch, we set = ddc |z|2 over K ; then, if 0, there are constants C1 , C2 > 0 such that C1 ||||K p C2 ||||K .
146
We denote by L1 (K), resp. by L (K), the space of integrable (resp. bounded measurable) functions on K with respect to any smooth positive density on X. (3.3) Chern-Levine-Nirenberg inequalities (1969). For all compact subsets K, L of X with L K , there exists a constant CK,L 0 such that ||ddc u1 . . . ddc uq T ||L CK,L ||u1 ||L (K) . . . ||uq ||L (K) ||T ||K .
Proof. By induction, it is sucient to prove the result for q = 1 and u1 = u. There is a covering of L by a family of balls Bj Bj K contained in coordinate patches of X. Let (Bj ) be equal to 1 on B j . Then ||ddc u T ||LB
j
Bj
ddc u T p1
C
Bj
ddc u T p1 .
As T and are closed, an integration by parts yields ||ddc u T ||LB C

Bj
u T ddc p1
C ||u||L (K) ||T ||K
where C is equal to C multiplied by a bound for the coecients of the smooth form ddc p1 . (3.4) Remark. With the same notations as above, any plurisubharmonic function V on X satises inequalities of the type a) b) ||ddc V ||L sup V+
L
CK,L ||V ||L1 (K) . CK,L ||V ||L1 (K) .
In fact the inequality ddc V n1 ddc V n1 = V ddc n1
LB j
Bj
Bj
implies a), and b) follows from the mean value inequality. (3.5) Remark. Products of the form = 1 . . . q T with mixed (1, 1)-forms j = ddc uj or j = dvj dc wj + dwj dc vj are also well dened whenever uj , vj , wj are locally bounded plurisubharmonic functions. Moreover, for L K , we have ||||L CK,L ||T ||K ||uj ||L (K) ||vj ||L (K) ||wj ||L (K) .
To check this, we may suppose vj , wj 0 and ||vj || = ||wj || = 1 in L (K). Then the inequality follows from (3.3) by the polarization identity
2 2 2(dvj dc wj + dwj dc vj ) = ddc (vj + wj )2 ddc vj ddc wj vj ddc wj wj ddc vj
in which all ddc operators act on plurisubharmonic functions.
147
(3.6) Corollary. Let u1 , . . . , uq be continuous (nite) plurisubharmonic functions and let uk , . . . , uk be sequences of plurisubharmonic functions converging locally uniformly to 1 q u1 , . . . , uq . If Tk is a sequence of closed positive currents converging weakly to T , then a) uk ddc uk . . . ddc uk Tk u1 ddc u2 . . . ddc uq T q 2 1 b) ddc uk . . . ddc uk Tk ddc u1 . . . ddc uq T 1 q weakly.
weakly.
Proof. We observe that b) is an immediate consequence of a) by the weak continuity of ddc . By using induction on q, it is enough to prove result a) when q = 1. If (uk ) converges locally uniformly to a nite continuous plurisubharmonic function u, we introduce local regularizations u = u and write uk Tk uT = (uk u)Tk + (u u )Tk + u (Tk T ). As the sequence Tk is weakly convergent, it is locally uniformly bounded in mass, thus ||(uk u)Tk ||K ||uk u||L (K) ||Tk ||K converges to 0 on every compact set K. The same argument shows that ||(u u )Tk ||K can be made arbitrarily small by choosing small enough. Finally u is smooth, so u (Tk T ) converges weakly to 0. Now, we prove a deeper monotone continuity theorem due to [Bedford-Taylor 1982] according to which the continuity and uniform convergence assumptions can be dropped if each sequence (uk ) is decreasing and Tk is a constant sequence. j (3.7) Theorem. Let u1 , . . . , uq be locally bounded plurisubharmonic functions and let uk , . . . , uk be decreasing sequences of plurisubharmonic functions converging pointwise to 1 q u1 , . . . , uq . Then a) uk ddc uk . . . ddc uk T u1 ddc u2 . . . ddc uq T 1 2 q b) ddc uk . . . ddc uk T ddc u1 . . . ddc uq T 1 q weakly.
weakly.
Proof. Again by induction, observing that a) = b) and that a) is obvious for q = 1 thanks to Lebesgues bounded convergence theorem. To proceed with the induction step, we rst have to make some slight modications of our functions uj and uk . j As the sequence (uk ) is decreasing and as uj is locally bounded, the family (uk )kN j j is locally uniformly bounded. The results are local, so we can work on a Stein open set X with strongly pseudoconvex boundary. We use the following notations: (3.8) let be a strongly plurisubharmonic function of class on and = 0, d = 0 on ; we set = {z ; (z) < } for all > 0.
near with < 0
(3.8 )
148
0 1 M R
A uk j
k III-1 Construction of vj
After addition of a constant we can assume that M uk 1 near . Let us denote j by (uk, ), ]0, 0 ], an increasing family of regularizations converging to uk as 0 j j 1 on . Set A = M/ with > 0 small and replace uk and such that M uk, j j k, k by vj = max{A, uk } and uk, by vj = max {A, uk,} where max = max is a j j j regularized max function.
k k Then vj coincides with uk on since A < A = M on , and vj is equal to j A on the corona \ /M . Without loss of generality, we can therefore assume that all uk (and similarly all uk, ) coincide with A on a xed neighborhood of . We need a j j lemma.
(3.9) Lemma. Let fk be a decreasing sequence of upper semi-continuous functions converging to f on some separable locally compact space X and k a sequence of positive measures converging weakly to on X. Then every weak limit of fk k satises f . Indeed if (gp ) is a decreasing sequence of continuous functions converging to fk0 for some k0 , then fk k fk0 k gp k for k k0 , thus gp as k +. The monotone convergence theorem then gives fk0 as p + and f as k0 +. Proof of Theorem 3.7 (end).. Assume that a) has been proved for q 1. Then S k = ddc uk . . . ddc uk T S = ddc u2 . . . ddc uq T. 2 q By 3.3 the sequence (uk S k ) has locally bounded mass, hence is relatively compact for 1 the weak topology. In order to prove a), we only have to show that every weak limit of uk S k is equal to u1 S. Let (m, m) be the bidimension of S and let be an arbitrary 1 smooth and strongly positive form of bidegree (m, m). Then the positive measures S k converge weakly to S and Lemma 3.9 shows that u1 S , hence u1 S. m , i.e. To get the equality, we set = ddc > 0 and show that u1 S m u1 ddc u2 . . . ddc uq T m lim inf
k+
uk ddc uk . . . ddc uk T m . 1 2 q
149
As u1
uk 1
uk,1 for every 1 > 0, we get 1 u1 ddc u2 . . . ddc uq T m
uk,1 ddc u2 . . . ddc uq T m 1 ddc uk,1 u2 ddc u3 . . . ddc uq T m 1
after an integration by parts (there is no boundary term because uk,1 and u2 both vanish 1 on ). Repeating this argument with u2 , . . . , uq , we obtain u1 ddc u2 . . . ddc uq T m
q1 ddc uk,1 . . . ddc uq1 uq T m 1
k,
uk,1 ddc uk,2 . . . ddc uk,q T m . q 1 2
Now let q 0, . . . , 1 0 in this order. We have weak convergence at each step and uk,1 = 0 on the boundary; therefore the integral in the last line converges and we get 1 the desired inequality u1 ddc u2 . . . ddc uq T m uk ddc uk . . . ddc uk T m . 1 2 q
(3.10) Corollary. The product ddc u1 . . . ddc uq T is symmetric with respect to u1 , . . . , uq . Proof. Observe that the denition was unsymmetric. The result is true when u1 , . . . , uq are smooth and follows in general from Th. 3.7 applied to the sequences uk = uj 1/k , j 1 j q. (3.11) Proposition. Let K, L be compact subsets of X such that L K . For any plurisubharmonic functions V, u1 , . . . , uq on X such that u1 , . . . , uq are locally bounded, there is an inequality ||V ddc u1 . . . ddc uq ||L CK,L ||V ||L1 (K) ||u1 ||L (K) . . . ||uq ||L (K) .
Proof. We may assume that L is contained in a strongly pseudoconvex open set = { < 0} K (otherwise we cover L by small balls contained in K). A suitable normalization gives 2 uj 1 on K ; then we can modify uj on \ L so that uj = A on \ with a xed constant A and > 0 such that L . Let 0 be a smooth function equal to on with compact support in . If we take ||V ||L1 (K) = 1, we see that V+ is uniformly bounded on by 3.4 b); after subtraction of a xed constant we can
150
assume V
0 on . First suppose q
n 1. As uj = A on \ , we nd
V ddc u1 . . . ddc uq nq
= =
V ddc u1 . . . ddc uq nq1 ddc Aq ddc V ddc u1 . . . ddc uq nq1 Aq
V n1 ddc V n1 ddc .
The rst integral of the last line is uniformly bounded thanks to 3.3 and 3.4 a), and the second one is bounded by ||V ||L1 () constant. Inequality 3.11 follows for q n 1. If q = n, we can work instead on X C and consider V, u1 , . . . , uq as functions on X C independent of the extra variable in C.
4. Case of Unbounded Plurisubharmonic Functions

We would like to dene ddc u1 . . . ddc uq T also in some cases when u1 , . . . , uq are not bounded below everywhere, especially when the uj have logarithmic poles. Consider rst the case q = 1 and let u be a plurisubharmonic function on X. The pole set of u is by denition P (u) = u1 (). We dene the unbounded locus L(u) to be the set of points x X such that u is unbounded in every neighborhood of x. Clearly L(u) is closed and we have L(u) P (u) but in general these sets are dierent: in fact, 3 u(z) = k 2 log(|z 1/k| + ek ) is everywhere nite in C but L(u) = {0}. (4.1) Proposition. We make two additional assumptions: a) T has non zero bidimension (p, p) (i.e. degree of T < 2n). b) X is covered by a family of Stein open sets X whose boundaries do not intersect L(u) Supp T . Then the current uT has locally nite mass in X. For any current T , hypothesis 4.1 b) is clearly satised when u has a discrete unbounded locus L(u); an interesting example is u = log |F | where F = (F1 , . . . , FN ) are holomorphic functions having a discrete set of common zeros. Observe that the current uT need not have locally nite mass when T has degree 2n (i.e. T is a measure); example: T = 0 and u(z) = log |z| in Cn . The result also fails when the sets are not assumed to be Stein; example: X = blow-up of Cn at 0, T = [E] = current of integration on the exceptional divisor and u(z) = log |z| (see 7.12 for the denition of blow-ups). Proof. By shrinking slightly, we may assume that has a smooth strongly pseudoconvex boundary. Let be a dening function of as in (3.8). By subtracting a constant to u, we may assume u on . We x so small that does not intersect L(u) Supp T and we select a neighborhood of ( ) Supp T such that L(u) = . Then we dene us (z) = max{u(z), A(z)} on , max{u(z), s} on = { < }.
151
By construction u M on for some constant M > 0. We x A s M , so max{u(z), A(z)} = max{u(z), s} = u(z) on
M/ and take
and our denition of us is coherent. Observe that us is dened on , which is a neighborhood of Supp T . Now, us = A on ( /A ), hence Stokes theorem implies ddc us T (ddc )p1 =
Addc T (ddc )p1 ddc (us A)T (ddc )p1 = 0
because the current [. . .] has a compact support contained in /A . Since us and both vanish on , an integration by parts gives us T (ddc )p = ddc us T (ddc )p1
||||L ()
T ddc us (ddc )p1
= ||||L () A
T (ddc )p . M on . We
Finally, take A = M/, let s tend to and use the inequality u obtain u T (ddc )p M T (ddc )p + lim
s
us T (ddc )p
M + ||||L () M/ The last integral is nite. This concludes the proof.
T (ddc )p .
(4.2) Remark. If is smooth and strongly pseudoconvex, the above proof shows in fact that C ||uT || ||u||L (( )Supp T ) ||T || when L(u) Supp T . In fact, if u is continuous and if is chosen suciently small, the constant M can be taken arbitrarily close to ||u||L (( )Supp T ) . Moreover, the maximum principle implies ||u+ ||L (Supp T ) = ||u+ ||L (Supp T ) , so we can achieve that u < on a neighborhood of Supp T by subtracting ||u||L (( )Supp T ) + 2 [Proof of maximum principle: if u(z0 ) > 0 at z0 Supp T and u 0 near Supp T , then
u+ T (ddc )p =
ddc u+ T (ddc )p1
0,
152
a contradiction]. (4.3) Corollary. Let u1 , . . . , uq be plurisubharmonic functions on X such that X is covered by Stein open sets with L(uj ) Supp T = . We use again induction to dene ddc u1 ddc u2 . . . ddc uq T = ddc (u1 ddc u2 . . . ddc uq T ).
Then, if uk , . . . , uk are decreasing sequences of plurisubharmonic functions converging 1 q pointwise to u1 , . . . , uq , q p, properties (3.7 a, b) hold.
0 1 M R Supp T
A uk j
L(uj )
k III-2 Modied construction of vj
Proof. Same proof as for Th. 3.7, with the following minor modication: the max k procedure vj := max{uk , A} is applied only on a neighborhood of Supp T ( ) j with > 0 small, and uk is left unchanged near Supp T . Observe that the integration j by part process requires the functions uk and uk, to be dened only near Supp T . j j (4.4) Proposition. Let X be a Stein open subset. If V is a plurisubharmonic function on X and u1 , . . . , uq , 1 q n 1, are plurisubharmonic functions such that L(uj ) = , then V ddc u1 . . . ddc uq has locally nite mass in . Proof. Same proof as for 3.11, when > 0 is taken so small that L(uj ) for all 1 j q. Finally, we show that Monge-Amp`re operators can also be dened in the case of e plurisubharmonic functions with non compact pole sets, provided that the mutual intersections of the pole sets are of suciently small Hausdor dimension with respect to the dimension p of T . (4.5) Theorem. Let u1 , . . . , uq be plurisubharmonic functions on X. The currents u1 ddc u2 . . . ddc uq T and ddc u1 . . . ddc uq T are well dened and have locally nite mass in X as soon as q p and
2p2m+1
L(uj1 ) . . . L(ujm ) Supp T = 0
for all choices of indices j1 < . . . < jm in {1, . . . , q}.
153
The proof is an easy induction on q, thanks to the following improved version of the Chern-Levine-Nirenberg inequalities. (4.6) Proposition. Let A1 , . . . , Aq X be closed sets such that
2p2m+1
Aj1 . . . Ajm Supp T = 0
for all choices of j1 < . . . < jm in {1, . . . , q}. Then for all compact sets K, L of X with L K , there exist neighborhoods Vj of K Aj and a constant C = C(K, L, Aj ) such that the conditions uj 0 on K and L(uj ) Aj imply a) ||u1 ddc u2 . . . ddc uq T ||L b) ||ddc u1 . . . ddc uq T ||L C||u1 ||L (K C||u1 ||L (K
V1 )
. . . ||uq ||L (K
Vq ) ||T ||K
V1 ) . . . ||uq ||L (K Vq ) ||T ||K .
Proof. We need only show that every point x0 K has a neighborhood L such that a), b) hold. Hence it is enough to work in a coordinate open set. We may thus assume that X Cn is open, and after a regularization process uj = lim uj for j = q, q 1, . . . , 1 in this order, that u1 , . . . , uq are smooth. We proceed by induction on q in two steps: Step 1. (bq1 ) = (bq ), Step 2. (aq1 ) and (bq ) = (aq ), where (b0 ) is the trivial statement ||T ||L (aq ) = (a ) and (bq ) = (b ) for q need the following elementary fact. ||T ||K and (a0 ) is void. Observe that we have p by taking u+1 (z) = . . . = uq (z) = |z|2 . We
(4.7) Lemma. Let F Cn be a closed set such that 2s+1 (F ) = 0 for some integer 0 s < n. Then for almost all choices of unitary coordinates (z1 , . . . , zn ) = (z , z ) with z = (z1 , . . . , zs ), z = (zs+1 , . . . , zn ) and almost all radii of balls B = B(0, r ) Cns , the set {0} B does not intersect F . Proof. The unitary group U (n) has real dimension n2 . There is a proper submersion : U (n) Cns {0} Cn {0}, (g, z ) g(0, z ),
whose bers have real dimension N = n2 2s. It follows that the inverse image 1 (F ) has zero Hausdor measure N+2s+1 = n2 +1 . The set of pairs (g, r ) U (n) R + such that g({0} B ) intersects F is precisely the image of 1 (F ) in U (n) R by + the Lipschitz map (g, z ) (g, |z |). Hence this set has zero n2 +1 -measure. Proof of step 1.. Take x0 = 0 K . Suppose rst 0 A1 . . . Aq and set F = A1 . . . Aq Supp T . Since 2p2q+1 (F ) = 0, Lemma 4.7 implies that there are coordinates z = (z1 , . . . , zs ), z = (zs+1 , . . . , zn ) with s = p q and a ball B such that F {0} B = and {0} B K . By compactness of K, we can nd neighborhoods Wj of K Aj and a ball B = B(0, r ) Cs such that B B K and (4.8) W 1 . . . W q Supp T B B

(1 )B
for > 0 small. If 0 Aj for some j, we choose instead Wj to be a small neighborhood / of 0 such that W j (B (1 )B ) Aj ; property (4.8) is then automatically satised.
154
Let j 0 be a function with compact support in Wj , equal to 1 near K Aj if Aj 0 (resp. equal to 1 near 0 if Aj 0) and let (z ) 0 be a function equal to 1 on 1/2 B with compact support in B . Then ddc (1 u1 ) . . . ddc (q uq ) T (z ) (ddc |z |2 )s = 0
B B
because the integrand is ddc exact and has compact support in B B thanks to (4.8). If we expand all factors ddc (j uj ), we nd a term 1 . . . q (z )ddc u1 . . . ddc uq T 0
which coincides with ddc u1 . . .ddc uq T on a small neighborhood of 0 where j = = 1. The other terms involve dj dc uj + duj dc j + uj ddc j for at least one index j. However dj and ddc j vanish on some neighborhood Vj of K Aj and therefore uj is bounded on B B Vj . We then apply the induction hypothesis (bq1 ) to the current = ddc u1 . . . ddc uj . . . ddc uq T and the usual Chern-Levine-Nirenberg inequality to the product of with the mixed term dj dc uj + duj dc j . Remark 3.5 can be applied because j is smooth and is (1) (2) therefore a dierence j j of locally bounded plurisubharmonic functions in Cn . Let K be a compact neighborhood of B B with K K , and let Vj be a neighborhood of K Aj with V j Vj . Then with L := (B B ) Vj (K Vj ) we obtain ||(dj dc uj + duj dc j ) ||B B = ||(dj dc uj + duj dc j ) ||L C1 ||uj ||L (K
Vj ) ||||K Vj ,
||||K
Vj
||||K
C2 ||u1 ||L (K
V1 ) . . . ||uj || . . . ||uq ||L (K Vq ) ||T ||K .
Now, we may slightly move the unitary basis in Cn and get coordinate systems z m = m m (z1 , . . . , zn ) with the same properties as above, such that the forms (ddc |z m |2 )s = dene a basis of
s,s
s! m m i dz1 dz m . . . i dzs dz m , 1 s s
(Cn ) . It follows that all measures
m m ddc u1 . . . ddc uq T i dz1 dz m . . . i dzs dz m 1 s
satisfy estimate (bq ) on a small neighborhood L of 0. Proof of Step 2.. We argue in a similar way with the integrals 1 u1 ddc (2 u2 ) . . . ddc (q uq ) T (z )(ddc |z |2 )s ddc |zs+1 |2 =
B B
B B
|zs+1 |2 ddc (1 u1 ) . . . ddc (q uq ) T (z )(ddc |z |2 )s .
155
We already know by (bq ) and Remark 3.5 that all terms in the right hand integral admit the desired bound. For q = 1, this shows that (b1 ) = (a1 ). Except for 1 . . . q (z ) u1 ddc u2 . . . ddc uq T , all terms in the left hand integral involve derivatives of j . By construction, the support of these derivatives is disjoint from Aj , thus we only have to obtain a bound for u1 ddc u2 . . . ddc uq T
when L = B(x0 , r) is disjoint from Aj for some j 2, say L A2 = , and is a constant positive form of type (p q, p q). Then B(x0 , r + ) K V 2 for some > 0 and some neighborhood V2 of K A2 . By the max construction used e.g. in Prop. 4.1, we can replace u2 by a plurisubharmonic function u2 equal to u2 in L and to A(|z x0 |2 r 2 )M in B(x0 , r + ) B(x0 , r + /2), with M = ||u2 ||L (K V2 ) and A = M/r. Let 0 be a smooth function equal to 1 on B(x0 , r + /2) with support in B(x0 , r). Then u1 ddc (u2 ) ddc u3 . . . ddc uq T =
B(x0 ,r+)
B(x0 ,r+)
u2 ddc u1 ddc u3 . . . ddc uq T

V1 )
O(1) ||u1 ||L (K
. . . ||uq ||L (K
Vq ) ||T ||K
where the last estimate is obtained by the induction hypothesis (bq1 ) applied to ddc u1 ddc u3 . . . ddc uq T . By construction ddc (u2 ) = ddc u2 + (smooth terms involving d) coincides with ddc u2 in L, and (aq1 ) implies the required estimate for the other terms in the left hand integral. (4.9) Proposition. With the assumptions of Th. 4.5, the analogue of the monotone convergence Theorem 3.7 (a,b) holds. Proof. By the arguments already used in the proof of Th. 3.7 (e.g. by Lemma 3.9), it is enough to show that 1 . . . q u1 ddc u2 . . . ddc uq T lim inf
k+ B B
B B
1 . . . q uk ddc uk . . . ddc uk T 1 2 q
where = (z )(ddc |z |2 )s is closed. Here the functions j , are chosen as in the proof of Step 1 in 4.7, especially their product has compact support in B B and j = = 1 in a neighborhood of the given point x0 . We argue by induction on q and also on the number m of functions (uj )j 1 which are unbounded near x0 . If uj is bounded near x0 , we take Wj Wj Wj to be small balls of center x0 on which uj is bounded and we modify the sequence uk on the corona Wj Wj so as to make it constant and equal j to a smooth function A|z x0 |2 + B on the smaller corona Wj Wj . In that case, we
156
take j equal to 1 near W j and Supp j Wj . For every = 1, . . . , q, we are going to check that lim inf
k+ B B
1 uk ddc (2 uk ) . . . 1 2
ddc (1 uk ) ddc ( u ) ddc (+1 u+1 ) . . . ddc (q uq ) T 1 lim inf

k+
ddc (1 uk ) ddc ( uk ) ddc (+1 u+1 ) . . . ddc (q uq ) T . 1 In order to do this, we integrate by parts 1 uk ddc ( u ) into u ddc (1 uk ) for 2, 1 1 and we use the inequality u uk . Of course, the derivatives dj , dc j , ddc j produce terms which are no longer positive and we have to take care of these. However, Supp dj is disjoint from the unbounded locus of uj when uj is unbounded, and contained in W j when uj is bounded. The number m of unbounded functions is therefore Wj replaced by m 1 in the rst case, whereas in the second case uk = uj is constant and j smooth on Supp dj , so q can be replaced by q 1. By induction on q + m (and thanks to the polarization technique 3.5), the limit of the terms involving derivatives of j is equal on both sides to the corresponding terms obtained by suppressing all indices k. Hence these terms do not give any contribution in the inequalities. We nally quote the following simple consequences of Th. 4.5 when T is arbitrary and q = 1, resp. when T = 1 has bidegree (0, 0) and q is arbitrary. (4.10) Corollary. Let T be a closed positive current of bidimension (p, p) and let u be a plurisubharmonic function on X such that L(u) Supp T is contained in an analytic set of dimension at most p 1. Then uT and ddc u T are well dened and have locally nite mass in X. (4.11) Corollary. Let u1 , . . . , uq be plurisubharmonic functions on X such that L(uj ) is contained in an analytic set Aj X for every j. Then ddc u1 . . . ddc uq is well dened as soon as Aj1 . . . Ajm has codimension at least m for all choices of indices j1 < . . . < jm in {1, . . . , q}. In the particular case when uj = log |fj | for some non zero holomorphic function fj on X, we see that the intersection product of the associated zero divisors [Zj ] = ddc uj is well dened as soon as the supports |Zj | satisfy codim |Zj1 | . . . |Zjm | = m for every m. Similarly, when T = [A] is an analytic p-cycle, Cor. 4.10 shows that [Z] [A] is well dened for every divisor Z such that dim |Z| |A| = p 1. These observations easily imply the following (4.12) Proposition. Suppose that the divisors Zj satisfy the above codimension condition and let (Ck )k 1 be the irreducible components of the point set intersection |Z1 | . . . |Zq |. Then there exist integers mk > 0 such that [Z1 ] . . . [Zq ] = mk [Ck ].
B B
1 uk ddc (2 uk ) . . . 1 2
The integer mk is called the multiplicity of intersection of Z1 , . . . , Zq along the component Ck .
5. Generalized Lelong Numbers
157
Proof. The wedge product has bidegree (q, q) and support in C = Ck where codim C = q, so it must be a sum as above with mk R+ . We check by induction on q that mk is a positive integer. If we denote by A some irreducible component of |Z1 | . . . |Zq1 |, we need only check that [A] [Zq ] is an integral analytic cycle of codimension q with positive coecients on each component Ck of the intersection. However [A] [Zq ] = 1 ddc (log |fq | [A]). First suppose that no component of A fq (0) is contained in the singular part Asing . Then the Lelong-Poincar equation applied on Areg shows that e c dd (log |fq | [A]) = mk [Ck ] on X Asing , where mk is the vanishing order of fq along Ck in Areg . Since C Asing has codimension q + 1 at least, the equality must hold on X. In general, we replace fq by fq so that the divisor of fq has no component contained in Asing . Then ddc (log |fq | [A]) is an integral codimension q cycle with 1 positive multiplicities on each component of A fq () and we conclude by letting tend to zero.

The concepts we are going to study mostly concern the behaviour of currents or plurisubharmonic functions in a neighborhood of a point at which we have for instance a logarithmic pole. Since the interesting applications are local, we assume from now on (unless otherwise stated) that X is a Stein manifold, i.e. that X has a strictly plurisubharmonic exhaustion function. Let : X [, +[ be a continuous plurisubharmonic function (in general may have poles, our continuity assumption means that e is continuous). The sets (5.1) (5.1 ) (5.1 ) S(r) = {x X ; (x) = r}, B(r) = {x X ; (x) < r}, B(r) = {x X ; (x) r}
will be called pseudo-spheres and pseudo-balls associated with . Note that B(r) is not necessarily equal to the closure of B(r), but this is often true in concrete situations. The most simple example we have in mind is the case of the function (z) = log |z a| on an open subset X Cn ; in this case B(r) is the euclidean ball of center a and radius er ; moreover, the forms (5.2) i 2 1 c 2 dd e = d d |z| , 2 2 ddc = i d d log |z a|
can be interpreted respectively as the at hermitian metric on Cn and as the pull-back over Cn of the Fubini-Study metric of Pn1 , translated by a. (5.3) Denition. We say that is semi-exhaustive if there exists a real number R such that B(R) X. Similarly, is said to be semi-exhaustive on a closed subset A X if there exists R such that A B(R) X. We are interested especially in the set of poles S() = { = } and in the behaviour of near S(). Let T be a closed positive current of bidimension (p, p) on X. Assume that is semi-exhaustive on Supp T and that B(R) Supp T X. Then
158
P = S() SuppT is compact and the results of 2 show that the measure T (ddc )p is well dened. Following [Demailly 1982b, 1987a], we introduce: (5.4) Denition. If is semi-exhaustive on Supp T and if R is such that B(R) Supp T X, we set for all r ] , R[ (T, , r) =
B(r)
T (ddc )p , T (ddc )p = lim (T, , r).

r
(T, ) =
S()
The number (T, ) will be called the (generalized) Lelong number of T with respect to the weight . If we had not required T (ddc )p to be dened pointwise on 1 (), the assumption that X is Stein could have been dropped: in fact, the integral over B(r) always makes sense if we dene (T, , r) =
B(r)
T ddc max{, s}
with s < r.
Stokes formula shows that the right hand integral is actually independent of s. The example given after (4.1) shows however that T (ddc )p need not exist on 1 () if 1 () contains an exceptional compact analytic subset. We leave the reader consider by himself this more general situation and extend our statements by the max{, s} technique. Observe that r (T, , r) is always an increasing function of r. Before giving examples, we need a formula. (5.5) Formula. For any convex increasing function : R R we have
B(r)
T (ddc )p = (r 0)p (T, , r)
where (r 0) denotes the left derivative of at r. Proof. Let be the convex function equal to on [r , +[ and to a linear function of slope (r 0) on ] , r ]. We get ddc ( ) = (r 0)ddc on B(r ) and Stokes theorem implies T (ddc )p = T (ddc )p T (ddc )p
B(r)
B(r)
= (r 0)p (T, , r ). Similarly, taking equal to on ] , r ] and linear on [r , r], we obtain

B(r)
B(r)
T (ddc )p
B(r)
T (ddc )p = (r 0)p (T, , r).
159
The expected formula follows when tends to 0. We get in particular (5.6)

B(r)
T (ddc e2 )p = (2e2r )p (T, , r), whence the formula T 1 c 2 dd e 2

p
(T, , r) = e2pr
B(r)
Now, assume that X is an open subset of Cn and that (z) = log |z a| for some a X. Formula (5.6) gives (T, , log r) = r 2p
|za|<r
i 2 d d |z| 2
i 1 The positive measure T = p! T ( 2 d d |z|2 )p = 2p the trace measure of T . We get
TI,I . in dz1 . . . dz n is called
(5.7)
(T, , log r) =
T B(a, r) p r 2p /p!
and (T, ) is the limit of this ratio as r 0. This limit is called the (ordinary) Lelong number of T at point a and is denoted (T, a). This was precisely the original denition of Lelong, see [Lelong 1968]. Let us mention a simple but important consequence. (5.8) Consequence. The ratio T B(a, r) /r 2p is an increasing function of r. Moreover, for every compact subset K X and every r0 < d(K, X) we have T B(a, r)
2p where C = T K + B(0, r0 ) /r0 .
Cr 2p
for a K and r
r0 ,
All these results are particularly interesting when T = [A] is the current of integration over an analytic subset A X of pure dimension p. Then T B(a, r) is the euclidean area of A B(a, r), while p r 2p /p! is the area of a ball of radius r in a p-dimensional subspace of Cn . Thus (T, , log r) is the ratio of these areas and the Lelong number (T, a) is the limit ratio. (5.9) Remark. It is immediate to check that ([A], x) = 0 for x A, / 1 when x A is a regular point.
We will see later that ([A], x) is always an integer (Thies theorem 8.7). (5.10) Remark. When X = Cn , (z) = log |z a| and A = X (i.e. T = 1), we obtain in particular B(a,r) (ddc log |z a|)n = 1 for all r. This implies (ddc log |z a|)n = a .
160
This fundamental formula can be viewed as a higher dimensional analogue of the usual formula log |z a| = 2a in C. We next prove a result which shows in particular that the Lelong numbers of a closed positive current are zero except on a very small set. (5.11) Proposition. If T is a closed positive current of bidimension (p, p), then for each c > 0 the set Ec = {x X ; (T, x) c} is a closed set of locally nite 2p Hausdor measure in X. Proof. By (5.7), we infer (T, a) = limr0 T B(a, r) p!/ p r 2p . The function a T B(a, r) is clearly upper semicontinuous. Hence the decreasing limit (T, a) as r decreases to 0 is also upper semicontinuous in a. This implies that Ec is closed. Now, let K be a compact subset in X and let {aj }1 j N , N = N (), be a maximal collection of points in Ec K such that |aj ak | 2 for j = k. The balls B(aj , 2) cover Ec K, whereas the balls B(aj , ) are disjoint. If Kc, is the set of points which are at distance of Ec K, we get T (Kc, ) since (T, aj ) T B(aj , ) N () c p 2p /p!,
c. By the denition of Hausdor measure, we infer
2p (Ec K)
lim inf
0 0
diam B(aj , 2)
2p
lim inf N ()(4)2p
p!42p T (Ec K). c p
Finally, we conclude this section by proving two simple semi-continuity results for Lelong numbers. (5.12) Proposition. Let Tk be a sequence of closed positive currents of bidimension (p, p) converging weakly to a limit T . Suppose that there is a closed set A such that Supp Tk A for all k and such that is semi-exhaustive on A with A B(R) X. Then for all r < R we have T (ddc )p lim inf
k+ B(r)
B(r)
Tk (ddc )p Tk (ddc )p
B(r)
lim sup
k+ B(r)
T (ddc )p .
When r tends to , we nd in particular lim sup (Tk , )

k+
(T, ).
Proof. Let us prove for instance the third inequality. Let be a sequence of smooth plurisubharmonic approximations of with < + 1/ on {r r + }. We set on B(r), = max{, (1 + )( 1/) r} on X B(r).
161
This denition is coherent since = near S(r), and we have = (1 + )( 1/) r near S(r + /2)
as soon as is large enough, i.e. (1 + )/ 2 /2. Let be a cut-o function equal to 1 in B(r + /2) with support in B(r + ). Then Tk (ddc )p Tk (ddc )p
B(r+/2)
B(r)
B(r+/2)
= (1 + )p (1 + )p
B(r+)
Tk (ddc )p
Tk (ddc )p .
As (ddc )p is smooth with compact support and as Tk converges weakly to T , we infer T (ddc )p . Tk (ddc )p (1 + )p lim sup
k+ B(r) B(r+)
We then let tend to + and tend to 0 to get the desired inequality. The rst inequality is obtained in a similar way, we dene so that = on X B(r) and = max{(1 )( 1/) + r} on B(r), and we take = 1 on B(r ) with Supp B(r /2). Then for large
B(r)
Tk (ddc )p
B(r/2)
Tk (ddc )p
B(r/2)
(1 )p
Tk (ddc )p .
(5.13) Proposition. Let k be a (non necessarily monotone) sequence of continuous plurisubharmonic functions such that ek converges uniformly to e on every compact subset of X. Suppose that { < R} Supp T X. Then for r < R we have lim sup
k+ {k r}{<R}
T (ddc k )p (T, ).
{ r}
T (ddc )p .
In particular lim supk+ (T, k )
When we take k (z) = log |z ak | with ak a, Prop. 5.13 implies the upper semicontinuity of a (T, a) which was already noticed in the proof of Prop. 5.11. Proof. Our assumption is equivalent to saying that max{k , t} converges locally uniformly to max{, t} for every t. Then Cor. 3.6 shows that T (ddc max{k , t})p converges weakly to T (ddc max{, t})p . If is a cut-o function equal to 1 on { r + /2} with support in { < r + }, we get
k+
lim
T (ddc max{k , t})p =
T (ddc max{, t})p .
162
For k large, we have {k infer lim sup

k+ {k r}{<R}
r} { < R} { < r + /2}, thus when tends to 0 we T (ddc max{k , t})p T (ddc max{, t})p .
{ r}
When we choose t < r, this is equivalent to the rst inequality in statement (5.13).
6. The Jensen-Lelong Formula

We assume in this section that X is Stein, that is semi-exhaustive on X and that B(R) X. We set for simplicity r = max{, r}. For every r ] , R[, the measures ddc ( r )n are well dened. By Cor. 3.6, the map r (ddc r )n is continuous on ] , R[ with respect to the weak topology. As (ddc r )n = (ddc )n on X \ B(r) and as r r, (ddc r )n = 0 on B(r), the left continuity implies (ddc r )n 1 X\B(r) (ddc )n . Here 1 A denotes the characteristic function of any subset A X. l l According to the denition introduced in [Demailly 1985a], the collection of MongeAmp`re measures associated with is the family of positive measures r such that e (6.1) r = (ddc r )n 1 X\B(r) (ddc )n , l r ] , R[.
The measure r is supported on S(r) and r r is weakly continuous on the left by the bounded convergence theorem. Stokes formula shows that B(s) (ddc r )n (ddc )n = 0 for s > r, hence the total mass r (S(r)) = r (B(s)) is equal to the dierence between the masses of (ddc )n and 1 X\B(r) (ddc )n over B(s), i.e. l (6.2) r S(r) =
B(r)
(ddc )n .
(6.3) Example. When (ddc )n = 0 on X \ 1 (), formula (6.1) can be simplied into r = (ddc r )n . This is so for (z) = log |z|. In this case, the invariance of under unitary transformations implies that r is also invariant. As the total mass of r is equal to 1 by 5.10 and (6.2), we see that r is the invariant measure of mass 1 on the euclidean sphere of radius er . (6.4) Proposition. Assume that is smooth near S(r) and that d = 0 on S(r), i.e. r is a non critical value. Then S(r) = B(r) is a smooth oriented real hypersurface and the measure r is given by the (2n 1)-volume form (ddc )n1 dc S(r) . Proof. Write max{t, r} = limk+ k (t) where is a decreasing sequence of smooth convex functions with k (t) = r for t r 1/k, k (t) = t for t r + 1/k. Theorem 3.6 shows that (ddc k )n converges weakly to (ddc r )n . Let h be a smooth function h with compact support near S(r). Let us apply Stokes theorem with S(r) considered as the boundary of X \ B(r) : h(ddc r )n = lim
X k+ X
h(ddc k )n dh (ddc k )n1 dc (k )
= lim
k+
163
= lim =
k+
()n dh (ddc )n1 dc k
X\B(r)
dh (ddc )n1 dc
X\B(r)
=
S(r)
h (ddc )n1 dc +
h (ddc )n1 ddc .
Near S(r) we thus have an equality of measures (ddc r )n = (ddc )n1 dc S(r) + 1 X\B(r) (ddc )n . l (6.5) Jensen-Lelong formula. Let V be any plurisubharmonic function on X. Then V is r -integrable for every r ] , R[ and
r
r (V )
V (ddc )n =
B(r)
(ddc V, , t) dt.
Proof. Proposition 3.11 shows that V is integrable with respect to the measure (ddc r )n , hence V is r -integrable. By denition (ddc V, , t) =
(z)<t
ddc V (ddc )n1
and the Fubini theorem gives

r
(ddc V, , t) dt =
(z)<t<r
ddc V (z) (ddc (z))n1 dt
(6.6)
=
B(r)
(r )ddc V (ddc )n1 .
We rst show that Formula 6.5 is true when and V are smooth. As both members of the formula are left continuous with respect to r and as almost all values of are non critical by Sards theorem, we may assume r non critical. Formula 3.1 applied with f = (r )(ddc )n1 and g = V shows that integral (6.6) is equal to
S(r)
V (ddc )n1 dc
B(r)
V (ddc )n = r (V )
V (ddc )n .
B(r)
Formula 6.5 is thus proved when and V are smooth. If V is smooth and merely continuous and nite, one can write = lim k where k is a decreasing sequence of smooth plurisubharmonic functions (because X is Stein). Then ddc V (ddc k )n1 converges weakly to ddc V (ddc )n1 and (6.6) converges, since 1 B(r) (r) is continuous l with compact support on X. The left hand side of Formula 6.5 also converges because the denition of r implies k,r (V ) V (ddc k )n =
k <r X
V (ddc k, r )n (ddc k )n
164
and we can apply again weak convergence on a neighborhood of B(r). If takes values, replace by k where k +. Then r (V ) is unchanged, B(r) V (ddc k )n converges to B(r) V (ddc )n and the right hand side of Formula 6.5 is replaced by r (ddc V, , t) dt. Finally, for V arbitrary, write V = lim Vk with a sequence of k smooth functions Vk . Then ddc Vk (ddc )n1 converges weakly to ddc V (ddc )n1 by Prop. 4.4, so the integral (6.6) converges to the expected limit and the same is true for the left hand side of 6.5 by the monotone convergence theorem. For r < r0 < R, the Jensen-Lelong formula implies (6.7) r (V ) r0 (V ) + V (ddc )n =
B(r0 )\B(r) r r0
(ddc V, , t) dt.
(6.8) Corollary. Assume that (ddc )n = 0 on X \S(). Then r r (V ) is a convex increasing function of r and the lelong number (ddc V, ) is given by (ddc V, ) = lim r (V ) . r r
Proof. By (6.7) we have

r
r (V ) = r0 (V ) +
r0
(ddc V, , t) dt.
As (ddc V, , t) is increasing and nonnegative, it follows that r r (V ) is convex and increasing. The formula for (ddc V, ) = limt (ddc V, , t) is then obvious. (6.9) Example. Let X be an open subset of Cn equipped with the semi-exhaustive function (z) = log |z a|, a X. Then (ddc )n = a and the Jensen-Lelong formula becomes r r (V ) = V (a) +
(ddc V, , t) dt.
As r is the mean value measure on the sphere S(a, er ), we make the change of variables r log r, t log t and obtain the more familiar formula
r
(6.9 a)
(V, S(a, r)) = V (a) +

0
(ddc V, a, t)
dt t
where (ddc V, a, t) = (ddc V, , log t) is given by (5.7): (6.9 b) (ddc V, a, t) = 1 n1 t2n2 /(n 1)!
B(a,t)
1 V. 2
In this setting, Cor. 6.8 implies (6.9 c) (ddc V, a) = lim supS(a,r) V V, S(a, r) = lim . r0 r0 log r log r
165
To prove the last equality, we may assume V 0 after subtraction of a constant. Inequality follows from the obvious estimate (V, S(a, r)) supS(a,r) V , while inequality follows from the standard Harnack estimate (6.9 d) sup V
S(a,r)
1 V, S(a, r) (1 + )2n1
when is small (this estimate follows easily from the Green-Riesz representation formula 1.4.6 and 1.4.7). As supS(a,r) V = supB(a,r) V , Formula (6.9 c) can also be rewritten (ddc V, a) = lim inf za V (z)/ log |z a|. Since supS(a,r) V is a convex (increasing) function of log r, we infer that (6.9 e) V (z) log |z a| + O(1)
with = (ddc V, a), and (ddc V, a) is the largest constant which satises this inequality. Thus (ddc V, a) = is equivalent to V having a logarithmic pole of coecient . 6.10. Special case. Take in particular V = log |f | where f is a holomorphic function on X. The Lelong-Poincar formula shows that ddc log |f | is equal to the zero divisor e [Zf ] = mj [Hj ], where Hj are the irreducible components of f 1 (0) and mj is the 1 multiplicity of f on Hj . The trace 2 log |f | is then the euclidean area measure of Zf (with corresponding multiplicities mj ). By Formula (6.9 c), we see that the Lelong number ([Zf ], a) is equal to the vanishing order orda (f ), that is, the smallest integer m such that D f (a) = 0 for some multiindex with || = m. In dimension n = 1, we have 1 mj aj . Then (6.9 a) is the usual Jensen formula 2 log f =
r
log |f |, S(0, r) log |f (0)| =
(t)
0
dt = t
mj log
r |aj |
where (t) is the number of zeros aj in the disk D(0, t), counted with multiplicities mj . (6.11) Example. Take (z) = log max |zj |j where j > 0. Then B(r) is the polydisk of radii (er/1 , . . . , er/n ). If some coordinate zj is non zero, say z1 , we can write (z) as / 1 log |z1 | plus some function depending only on the (n 1) variables zj /z1 1 j . Hence (ddc )n = 0 on Cn \ {0}. It will be shown later that (6.11 a) (ddc )n = 1 . . . n 0 . We now determine the measures r . At any point z where not all terms |zj |j are equal, the smallest one can be omitted without changing in a neighborhood of z. Thus depends only on (n 1)-variables and (ddc r )n = 0, r = 0 near z. It follows that r is supported by the distinguished boundary |zj | = er/j of the polydisk B(r). As is invariant by all rotations zj eij zj , the measure r is also invariant and we see that r is a constant multiple of d1 . . . dn . By formula (6.2) and (6.11 a) we get (6.11 b) r = 1 . . . n (2)n d1 . . . dn .
In particular, the Lelong number (ddc V, ) is given by (ddc V, ) = lim 1 . . . n r r V (er/1 +i1 , . . . , er/n +in )
j [0,2]
d1 . . . dn . (2)n
166
These numbers have been introduced and studied by [Kiselman 1986]. We call them directional Lelong numbers with coecients (1 , . . . , n ). For an arbitrary current T , we dene (6.11 c) (T, x, ) = T, log max |zj xj |j .
The above formula for (ddc V, ) combined with the analogue of Harnacks inequality (6.9 d) for polydisks gives (ddc V, x, ) = lim (6.11 d) 1 . . . n r0 log r 1 . . . n = lim r0 log r V (r 1/1 ei1 , . . . , r1/n ein ) d1 . . . dn (2)n
sup V (r 1/1 ei1 , . . . , r1/n ein ).

1 ,...,n
7. Comparison Theorems for Lelong Numbers

Let T be a closed positive current of bidimension (p, p) on a Stein manifold X equipped with a semi-exhaustive plurisubharmonic weight . We rst show that the Lelong numbers (T, ) only depend on the asymptotic behaviour of near the polar set S(). In a precise way: (7.1) First comparison theorem. Let , : X [, +[ be continuous plurisubharmonic functions. We assume that , are semi-exhaustive on Supp T and that := lim sup Then (T, ) (x) < + (x) as x Supp T and (x) .
p (T, ), and the equality holds if = lim /.
Proof. Denition 6.4 shows immediately that (T, ) = p (T, ) for every scalar > 0. It is thus sucient to verify the inequality (T, ) (T, ) under the hypothesis lim sup / < 1. For all c > 0, consider the plurisubharmonic function uc = max( c, ). Let R and R be such that B (R ) Supp T and B (R ) Supp T be relatively compact in X. Let r < R and a < r be xed. For c > 0 large enough, we have uc = on 1 ([a, r]) and Stokes formula gives (T, , r) = (T, uc , r) (T, uc ).
The hypothesis lim sup / < 1 implies on the other hand that there exists t0 < 0 such that uc = c on {uc < t0 } Supp T . We infer (T, uc ) = (T, c) = (T, ), hence (T, ) (T, ). The equality case is obtained by reversing the roles of and and observing that lim / = 1/l.
167
k k Assume in particular that z k = (z1 , . . . , zn ), k = 1, 2, are coordinate systems centered at a point x X and let k k k (z) = log |z k | = log |z1 |2 + . . . + |zn |2 1/2
We have limzx 2 (z)/1 (z) = 1, hence (T, 1 ) = (T, 2 ) by Th. 7.1. (7.2) Corollary. The usual Lelong numbers (T, x) are independent of the choice of local coordinates. This result had been originally proved by [Siu 1974] with a much more delicate proof. Another interesting consequence is: (7.3) Corollary. On an open subset of Cn , the Lelong numbers and Kiselman numbers are related by (T, x) = T, x, (1, . . . , 1) . Proof. By denition, the Lelong number (T, x) is associated with the weight (z) = log |z x| and the Kiselman number T, x, (1, . . . , 1) to the weight (z) = log max |zj xj |. It is clear that limzx (z)/(z) = 1, whence the conclusion. Another consequence of Th. 7.1 is that (T, x, ) is an increasing function of each variable j . Moreover, if 1 . . . n , we get the inequalities p (T, x) 1 (T, x, ) p (T, x). n
These inequalities will be improved in section 7 (see Cor. 9.16). For the moment, we just prove the following special case. (7.4) Corollary. For all 1 , . . . , n > 0 we have ddc log max |zj |j
1 j n n
= ddc log
1 j n
|zj |j
= 1 . . . n 0 .
Proof. In fact, our measures vanish on Cn {0} by the arguments explained in example 6.11. Hence they are equal to c 0 for some constant c 0 which is simply the Lelong number of the bidimension (n, n)-current T = [X] = 1 with the corresponding weight. The comparison theorem shows that the rst equality holds and that ddc log
1 j n
|zj |j
= n ddc log
1 j n
|zj |j
for all > 0. By taking large and approximating j with 2[j /2], we may assume that j = 2sj is an even integer. Then formula (5.6) gives ddc log |zj |2sj
|wj |2 <r2 n
|zj |2sj <r2
= r 2n
= s1 . . . sn r 2n
2n
i 2 d d |w| 2
|zj |2sj <r2 n
ddc
|zj |2sj
= 1 . . . n
168

s
by using the s1 . . . sn -sheeted change of variables wj = zj j . Now, we assume that T = [A] is the current of integration over an analytic set A X of pure dimension p. The above comparison theorem will enable us to give a simple proof of P. Thies main result [Thie 1967]: the Lelong number ([A], x) can be interpreted as the multiplicity of the analytic set A at point x. Our starting point is the following consequence of Th. II.3.19 applied simultaneously to all irreducible components of (A, x). (7.5) Lemma. For a generic choice of local coordinates z = (z1 , . . . , zp ) and z = (zp+1 , . . . , zn ) on (X, x), the germ (A, x) is contained in a cone |z | C|z |. If B Cp is a ball of center 0 and radius r small, and B Cnp is the ball of center 0 and radius r = Cr , then the projection pr : A (B B ) B is a ramied covering with nite sheet number m. We use these properties to compute the Lelong number of [A] at point x. When z A tends to x, the functions (z) = log |z| = log(|z |2 + |z |2 )1/2 , (z) = log |z |.
are equivalent. As , are semi-exhaustive on A, Th. 7.1 implies ([A], x) = ([A], ) = ([A], ). Let us apply formula (5.6) to : for every t < r we get ([A], , log t) = t2p
{<log t}
[A]
1 c 2 dd e 2
1 c 2 pr dd |z | A{|z |<t} 2 1 c 2 dd |z | = m t2p Cp {|z |<t} 2 = t2p
= m,
hence ([A], ) = m. Here, we have used the fact that pr is an tale covering with m e sheets over the complement of the ramication locus S B , and the fact that S is of zero Lebesgue measure in B . We have thus obtained simultaneously the following two results: (7.6) Theorem and Denition. Let A be an analytic set of dimension p in a complex manifold X of dimension n. For a generic choice of local coordinates z = (z1 , . . . , zp ), z = (zp+1 , . . . , zn ) near a point x A such that the germ (A, x) is contained in a cone |z | C|z |, the sheet number m of the projection (A, x) (Cp , 0) onto the rst p coordinates is independent of the choice of z , z . This number m is called the multiplicity of A at x. (7.7) Theorem ([Thie 1967]). One has ([A], x) = m.
169
There is another interesting version of the comparison theorem which compares the Lelong numbers of two currents obtained as intersection products (in that case, we take the same weight for both). (7.8) Second comparison theorem. Let u1 , . . . , uq and v1 , . . . , vq be plurisubharmonic functions such that each q-tuple satises the hypotheses of Th. 4.5 with respect to T . Suppose moreover that uj = on Supp T 1 () and that j := lim sup Then (ddc v1 . . . ddc vq T, ) 1 . . . q (ddc u1 . . . ddc uq T, ). vj (z) < + uj (z) when z Supp T u1 (), (z) . j
Proof. By homogeneity in each factor vj , it is enough to prove the inequality with constants j = 1 under the hypothesis lim sup vj /uj < 1. We set wj,c = max{vj c, uj }. Our assumption implies that wj,c coincides with vj c on a neighborhood Supp T { < r0 } of Supp T { < }, thus (ddc v1 . . . ddc vq T, ) = (ddc w1,c . . . ddc wq,c T, ) for every c. Now, x r < R . Proposition 4.9 shows that the current ddc w1,c . . . ddc wq,c T converges weakly to ddc u1 . . .ddc uq T when c tends to +. By Prop. 5.12 we get lim sup (ddc w1,c . . . ddc wq,c T, )
c+
(ddc u1 . . . ddc uq T, ).
(7.9) Corollary. If ddc u1 . . . ddc uq T is well dened, then at every point x X we have ddc u1 . . . ddc uq T, x (ddc u1 , x) . . . (ddc uq , x) (T, x).
Proof. Apply (7.8) with (z) = v1 (z) = . . . = vq (z) = log |z x| and observe that j := lim sup vj /uj = 1/(ddc uj , x) (there is nothing to prove if (ddc uj , x) = 0). Finally, we present an interesting stability property of Lelong numbers due to [Siu 1974]: almost all slices of a closed positive current T along linear subspaces passing through a given point have the same Lelong number as T . Before giving a proof of this, we need a useful formula known as Croftons formula. (7.10) Lemma. Let be a closed positive (p, p)-form on Cn under the unitary group U (n). Then has the form = ddc (log |z|)
p
{0} which is invariant
170
where is a convex increasing function. Moreover is invariant by homotheties if and only if is an ane function, i.e. = (ddc log |z|)p . Proof. A radial convolution (z) = R (t/) (et z) dt produces a smooth form with the same properties as and lim0 = . Hence we can suppose that is smooth p,p on Cn {0}. At a point z = (0, . . . , 0, zn ), the (p, p)-form (z) (Cn ) must be invariant by U (n1) acting on the rst (n1) coordinates. We claim that the subspace of p,p U (n 1)-invariants in (Cn ) is generated by (ddc |z|2 )p and (ddc |z|2 )p1 idzn dz n . In fact, a form = I,J dzI dz J is invariant by U (1)n1 U (n 1) if and only if I,J = 0 for I = J, and invariant by the permutation group n1 U (n 1) if and only if all coecients I,I (resp. Jn,Jn ) with I, J {1, . . . , n 1} are equal. Hence =
|I|=p
dzI dz I +
|J|=p1
dzJ dz J dzn dz n .
i This proves our claim. As d|z|2 dc |z|2 = |zn |2 dzn dz n at (0, . . . , 0, zn ), we conclude that (z) = f (z)(ddc |z|2 )p + g(z)(ddc |z|2 )p1 d|z|2 dc |z|2
for some smooth functions f, g on Cn {0}. The U (n)-invariance of shows that f and g are radial functions. We may rewrite the last formula as (z) = u(log |z|)(ddc log |z|)p + v(log |z|)(ddc log |z|)p1 d log |z| dc log |z|. Here (ddc log |z|)p is a positive (p, p)-form coming from Pn1 , hence it has zero contraction in the radial direction, while the contraction of the form (ddc log |z|)p1 d log |z| dc log |z| by the radial vector eld is (ddc log |z|)p1 . This shows easily that (z) 0 if and only if u, v 0. Next, the closedness condition d = 0 gives u v = 0. Thus u is increasing and we dene a convex increasing function by = u1/p . Then v = u = pp1 and p (z) = ddc (log |z|) . If is invariant by homotheties, the functions u and v must be constant, thus v = 0 and = (ddc log |z|)p . (7.11) Corollary (Croftons formula). Let dv be the unique U (n)-invariant measure of mass 1 on the Grassmannian G(p, n) of p-dimensional subspaces in Cn . Then [S] dv(S) = (ddc log |z|)np .
SG(p,n)
Proof. The left hand integral is a closed positive bidegree (n p, n p) current which is invariant by U (n) and by homotheties. By Lemma 7.10, this current must coincide with the form (ddc log |z|)np for some 0. The coecient is the Lelong number at 0. As ([S], 0) = 1 for every S, we get = G(p,n) dv(S) = 1. We now recall a few basic facts of slicing theory; see [Federer 1969] for details. Let : M M be a submersion of smooth dierentiable manifolds and let be a locally at current on M , that is a current which can be written locally as = U + dV where
171
U , V have locally integrable coecients. It can be shown that every current such that both and d have measure coecients is locally at; in particular, closed positive currents are locally ats. Then, for almost every x M , there is a well dened slice x , which is the current on the ber 1 (x ) dened by x = U1 (x ) + dV1 (x ) . The restrictions of U , V to the bers exist for almost all x by the Fubini theorem. It is easy to show by a regularization = that the slices of a closed positive current are again closed and positive: in fact U,x and V,x converge to Ux and Vx in L1 , thus loc ,x converges weakly to x for almost every x . This kind of slicing can be referred to as parallel slicing (if we think of as being a projection map). The kind of slicing we need (where the slices are taken over linear subspaces passing through a given point) is of a slightly dierent nature and is called concurrent slicing. The possibility of concurrent slicing is proved as follows. Let T be a closed positive current of bidimension (p, p) in the ball B(0, R) Cn . Let Y = (x, S) Cn G(q, n) ; x S be the total space of the tautological rank q vector bundle over the Grassmannian G(q, n), equipped with the obvious projections : Y G(q, n), : Y Cn .
We set YR = 1 (B(0, R)) and YR = 1 (B(0, R) {0}). The restriction 0 of to YR is a submersion onto B(0, R) {0}, so we have a well dened pull-back 0 T over YR . We would like to extend it as a pull-back T over YR , so as to dene slices TS = ( T )1 (S) ; of course, these slices can be non zero only if the dimension of S is at least equal to the degree of T , i.e. if q n p. We rst claim that 0 T has locally nite mass near the zero section 1 (0) of . In fact let G be a unitary invariant Khler a metric over G(q, n) and let = ddc |z|2 in Cn . Then we get a Khler metric on Y dened a by Y = G + . If N = (q 1)(n q) is the dimension of the bers of , the projection formula (u v) = ( u) v gives N+p = Y 1 k p
N +p k N+pk ). ( G k
N+pk ) is a unitary and homothety invariant (pk, pk) closed positive form Here ( G N+pk n ) is proportional to (ddc log |z|)pk . With some constants on C {0}, so ( G k > 0, we thus get N+p = 0 T Y
k
0 k p B(0,r) {0}
Yr
T k (ddc log |z|)pk T p < +.
=
0 k p
k 2(pk) r 2(pk)
B(0,r) {0}
The Skoda-El Mir theorem 2.3 shows that the trivial extension 0 T of 0 T is a closed positive current on YR . Of course, the zero section 1 (0) might also carry some extra
172
mass of the desired current T . Since 1 (0) has codimension q, this extra mass cannot exist when q > n p = codim T and we simply set T = 0 T . On the other hand, if q = n p, we set (7.12)
T := 0 T + (T, 0) [ 1 (0)].
We can now apply parallel slicing with respect to : YR G(q, n), which is a submersion: for almost all S G(q, n), there is a well dened slice TS = ( T )1 (S) . These slices coincide with the usual restrictions of T to S if T is smooth. (7.13) Theorem ([Siu 1974]). For almost all S G(q, n) with q satises (TS , 0) = (T, 0). n p, the slice TS
Proof. If q = n p, the slice TS consists of some positive measure with support in S {0} plus a Dirac measure (T, 0) 0 coming from the slice of (T, 0) [ 1 (0)]. The equality (TS , 0) = (T, 0) thus follows directly from (7.12). In the general case q > n p, it is clearly sucient to prove the following two properties: a) (T, 0, r) =
SG(q,n)
(TS , 0, r) dv(S) for all r ]0, R[ ;
b)
(TS , 0)
(T, 0) for almost all S.
In fact, a) implies that (T, 0) is the average of all Lelong numbers (TS , 0) and the conjunction with b) implies that these numbers must be equal to (T, 0) for almost all S. In order to prove a) and b), we can suppose without loss of generality that T is smooth on B(0, R) {0}. Otherwise, we perform a small convolution with respect to the action of Gln (C) on Cn : T =
gGln (C)
(g) g T dv(g)
where ( ) is a regularizing family with support in an -neighborhood of the unit element of Gln (C). Then T is smooth in B(0, (1 )R) {0} and converges weakly to T . Moreover, we have (T , 0) = (T, 0) by (7.2) and (TS , 0) lim sup0 (T,S , 0) by (5.12), thus a), b) are preserved in the limit. If T is smooth on B(0, R) {0}, the slice TS is dened for all S and is simply the restriction of T to S {0} (carrying no mass at the origin). a) Here we may even assume that T is smooth at 0 by performing an ordinary convolution. As TS has bidegree (n p, n p), we have (TS , 0, r) =
SB(0,r)
T S
q(np)
=
B(0,r)
p+qn T [S] S
where S = ddc log |w| and w = (w1 , . . . , wq ) are orthonormal coordinates on S. We simply have to check that
p+qn dv(S) = (ddc log |z|)p . [S] S
SG(q,n)
8. Sius Semicontinuity Theorem
173
However, both sides are unitary and homothety invariant (p, p)-forms with Lelong number 1 at the origin, so they must coincide by Lemma 7.11. b) We prove the inequality when S = Cq {0}. By the comparison theorem 7.1, for every r > 0 and > 0 we have (7.14)
B(0,r) p T
(T, 0)
where
1 c dd log(|z1 |2 + . . . + |zq |2 + |zq+1 |2 + . . . + |zn |2 ). 2
p We claim that the current converges weakly to p+qn = [S] [S] S
1 c dd log(|z1 |2 + . . . + |zq |2 ) 2
p+qn
p as tends to 0. In fact, the Lelong number of at 0 is 1, hence by homogeneity p (ddc |z|2 )np = (2r 2 )np
B(0,r)
p for all , r > 0. Therefore the family ( ) is relatively compact in the weak topology. Since 0 = lim is smooth on Cn S and depends only on n q variables (n q p), p p p we have lim = 0 = 0 on Cn S. This shows that every weak limit of ( ) has support in S. Each of these is the direct image by inclusion of a unitary and homothety invariant (p + q n, p + q n)-form on S with Lelong number equal to 1 at 0. Therefore we must have p+qn p+qn p lim = (iS ) (S ) = [S] S , 0
and our claim is proved (of course, this can also be checked by direct elementary calculations). By taking the limsup in (7.14) we obtain (TS , 0, r + 0) =
B(0,r) p+qn T [S] S
(T, 0)
(the singularity of T at 0 does not create any diculty because we can modify T by a ddc -exact form near 0 to make it smooth everywhere). Property b) follows when r tends to 0.

Let X, Y be complex manifolds of dimension n, m such that X is Stein. Let : X Y [, +[ be a continuous plurisubharmonic function. We assume that is semi-exhaustive with respect to Supp T , i.e. that for every compact subset L Y there exists R = R(L) < 0 such that (8.1) {(x, y) Supp T L ; (x, y) R} X Y.
Let T be a closed positive current of bidimension (p, p) on X. For every point y Y , the function y (x) := (x, y) is semi-exhaustive on Supp T ; one can therefore associate
174
with y a generalized Lelong number (T, y ). Proposition 5.13 implies that the map y (T, y ) is upper semi-continuous, hence the upperlevel sets (8.2) Ec = Ec (T, ) = {y Y ; (T, y ) c} , c > 0
are closed. Under mild additional hypotheses, we are going to show that the sets Ec are in fact analytic subsets of Y , following [Demailly 1987a]. (8.3) Denition. We say that a function f (x, y) is locally Hlder continuous with o respect to y on X Y if every point of X Y has a neighborhood on which |f (x, y1 ) f (x, y2 )| M |y1 y2 |
for all (x, y1 ) , (x, y2 ) , with some constants M > 0, ]0, 1], and suitable coordinates on Y . (8.4) Theorem ([Demailly 1987a]). Let T be a closed positive current on X and let : X Y [, +[ be a continuous plurisubharmonic function. Assume that is semi-exhaustive on Supp T and that e(x,y) is locally Hlder continuous with respect to y on X Y . Then the o upperlevel sets Ec (T, ) = {y Y ; (T, y ) c} are analytic subsets of Y . This theorem can be rephrased by saying that y (T, y ) is upper semi-continuous with respect to the analytic Zariski topology. As a special case, we get the following important result of [Siu 1974]: (8.5) Corollary. If T is a closed positive current of bidimension (p, p) on a complex manifold X, the upperlevel sets Ec (T ) = {x X ; (T, x) c} of the usual Lelong numbers are analytic subsets of dimension p. Proof. The result is local, so we may assume that X Cn is an open subset. Theorem 8.4 with Y = X and (x, y) = log |x y| shows that Ec (T ) is analytic. Moreover, Prop. 5.11 implies dim Ec (T ) p. (8.6) Generalization. Theorem 8.4 can be applied more generally to weight functions of the type (x, y) = max log |Fj,k (x, y)|j,k
j k
where Fj,k are holomorphic functions on X Y and where j,k are positive real constants; in this case e is Hlder continuous of exponent = min{j,k , 1} and is semi-exhaustive o 1 with respect to the whole space X as soon as the projection pr2 : Fj,k (0) Y is proper and nite. For example, when (x, y) = log max |xj yj | on an open subset X of Cn , we see j that the upperlevel sets for Kiselmans numbers (T, x, ) are analytic in X (a result rst
175
proved in [Kiselman 1986]. More generally, set (z) = log max |zj |j and (x, y, g) = g(xy) where x, y Cn and g Gl(Cn ). Then (T, y,g ) is the Kiselman number of T at y when the coordinates have been rotated by g. It is clear that is plurisubharmonic in (x, y, g) and semi-exhaustive with respect to x, and that e is locally Hlder continuous o with respect to (y, g). Thus the upperlevel sets Ec = {(y, g) X Gl(Cn ) ; (T, y,g ) c}
are analytic in X Gl(Cn ). However this result is not meaningful on a manifold, because it is not invariant under coordinate changes. One can obtain an invariant version as follows. Let X be a manifold and let J k X be the bundle of k-jets of holomorphic functions on X. We consider the bundle Sk over X whose ber Sk,y is the set of n-tuples of k-jets u = (u1 , . . . , un ) (J k X,y )n such that uj (y) = 0 and du1 . . . dun (y) = 0. Let (zj ) be local coordinates on an open set X. Modulo O(|z y|k+1 ), we can write uj (z) =
1 || k
aj, (z y)
with det(aj,(0,...,1k ,...,0) ) = 0. The numbers ((yj ), (aj,)) dene a coordinate system on the total space of Sk . For (x, (y, u)) X Sk , we introduce the function (x, y, u) = log max |uj (x)|j = log max aj, (x y)
j
1 || k
which has all properties required by Th. 8.4 on a neighborhood of the diagonal x = y, i.e. a neighborhood of X X Sk in X Sk . For k large, we claim that Kiselmans directional Lelong numbers (T, y, u, ) := (T, y,u ) with respect to the coordinate system (uj ) at y do not depend on the selection of the jet representives and are therefore canonically dened on Sk . In fact, a change of uj by O(|z y|k+1 ) adds O(|z y|(k+1)j ) to e , and we have e O(|z y|max j ). Hence by (7.1) it is enough to take k + 1 max j / min j . Theorem 8.4 then shows that the upperlevel sets Ec (T, ) are analytic in Sk . Proof of the Semicontinuity Theorem 8.4. As the result is local on Y , we may assume without loss of generality that Y is a ball in Cm . After addition of a constant to , we may also assume that there exists a compact subset K X such that {(x, y) X Y ; (x, y) 0} K Y.
By Th. 7.1, the Lelong numbers depend only on the asymptotic behaviour of near the (compact) polar set 1 () (SuppT Y ). We can add a smooth strictly plurisubharmonic function on X Y to make strictly plurisuharmonic. Then Richbergs approximation theorem for continuous plurisubharmonic functions shows that there exists a smooth plurisubharmonic function such that + 1. We may therefore assume that is smooth on (X Y ) \ 1 (). First step: construction of a local plurisubharmonic potential.
176
Our goal is to generalize the usual construction of plurisubharmonic potentials associated with a closed positive current ([Lelong 1967], [Skoda 1972a]). We replace here the usual kernel |z |2p arising from the hermitian metric of Cn by a kernel depending on the weight . Let (R, R) be an increasing function such that (t) = t for t 1 and (t) = 0 for t 0. We consider the half-plane H = {z C ; Rez < 1} and associate with T the potential function V on Y H dened by
0
(8.7)
V (y, z) =
(T, y , t) (t) dt.

Rez
For every t > Re z, Stokes formula gives (T, y , t) =

(x,y)<t
T (x) (ddc (x, y, z))p x
with (x, y, z) := max{(x, y), Rez}. The Fubini theorem applied to (8.7) gives V (y, z) = =
xX xX,(x,y)<t Re z<t<0
T (x) (ddc (x, y, z))p (t)dt x
T (x) ((x, y, z)) (ddc (x, y, z))p. x
For all (n 1, n 1)-form h of class C with compact support in Y H, we get ddc V, h = V, ddc h =
XY H
T (x) ((x, y, z))(ddc(x, y, z))p ddc h(y, z).
Observe that the replacement of ddc by the total dierentiation ddc = ddc x x,y,z does not modify the integrand, because the terms in dx, dx must have total bidegree (n, n). The current T (x) ((x, y, z))h(y, z) has compact support in X Y H. An integration by parts can thus be performed to obtain ddc V, h =
XY H
T (x) ddc ( (x, y, z)) (ddc (x, y, z))p h(y, z).
On the corona {1 (x, y) 0} we have (x, y, z) = (x, y), whereas for (x, y) < 1 we get < 1 and = . As is plurisubharmonic, we see that ddc V (y, z) is the sum of the positive (1, 1)-form (y, z) T (x) (ddc (x, y, z))p+1 x,y,z
{xX;(x,y)<1}
and of the (1, 1)-form independent of z y T ddc ( ) (ddc )p . x,y x,y
{xX;1 (x,y) 0}
As is smooth outside 1 (), this last form has locally bounded coecients. Hence ddc V (y, z) is 0 except perhaps for locally bounded terms. In addition, V is continuous
177
on Y H because T (ddc y,z )p is weakly continuous in the variables (y, z) by Th. 3.5. We therefore obtain the following result. (8.8) Proposition. There exists a positive plurisubharmonic function in that (y) + V (y, z) is plurisubharmonic on Y H. If we let Re z tend to , we see that the function
0
(Y ) such
U0 (y) = (y) + V (y, ) = (y)
(T, y , t) (t)dt
is locally plurisubharmonic or on Y . Furthermore, it is clear that U0 (y) = at every point y such that (T, y ) > 0. If Y is connected and U0 , we already conclude that the density set c>0 Ec is pluripolar in Y . Second step: application of Kiselmans minimum principle. Let a 0 be arbitrary. The function Y H (y, z) (y) + V (y, z) aRez is plurisubharmonic and independent of Im z. By Kiselmans theorem 1.7.8, the Legendre transform Ua (y) = inf (y) + V (y, r) ar
r<1
is locally plurisubharmonic or on Y . (8.9) Lemma. Let y0 Y be a given point. a) If a > (T, y0 ), then Ua is bounded below on a neighborhood of y0 . b) If a < (T, y0 ), then Ua (y0 ) = . Proof. By denition of V (cf. (8.7)) we have
0
(8.10)
V (y, r)
(T, y , r)
(t)dt = r(T, y , r)
r
r(T, y ).
Then clearly Ua (y0 ) = if a < (T, y0 ). On the other hand, if (T, y0 ) < a, there exists t0 < 0 such that (T, y0 , t0 ) < a. Fix r0 < t0 . The semi-continuity property (5.13) shows that there exists a neighborhood of y0 such that supy (T, y , r0 ) < a. For all y , we get
r0
V (y, r) and this implies Ua (y)
C a
(t)dt = C + a(r r0 ),
C ar0 .
(8.11) Theorem. If Y is connected and if Ec = Y , then Ec is a closed complete pluripolar subset of Y , i.e. there exists a continuous plurisubharmonic function w : Y [, +[ such that Ec = w1 ().
178
Proof. We rst observe that the family (Ua ) is increasing in a, that Ua = on Ec for all a < c and that supa<c Ua (y) > if y Y \ Ec (apply Lemma 8.9). For any integer k 1, let wk (Y ) be a plurisubharmonic regularization of Uc1/k such that wk Uc1/k on Y and wk 2k on Ec Yk where Yk = {y Y ; d(y, Y ) 1/k}. Then Lemma 8.9 a) shows that the family (wk ) is uniformly bounded below on every compact subset of Y \ Ec . We can also choose wk uniformly bounded above on every compact subset of Y because Uc1/k Uc . The function
+
w=
k=1
2k wk
saties our requirements. Third step: estimation of the singularities of the potentials Ua . (8.12) Lemma. Let y0 Y be a given point, L a compact neighborhood of y0 , K X a compact subset and r0 a real number < 1 such that {(x, y) X L; (x, y) r0 } K L.
Assume that e(x,y) is locally Hlder continuous in y and that o |f (x, y1 ) f (x, y2 )| M |y1 y2 |
for all (x, y1 , y2 ) K L L. Then, for all ]0, 1[, there exists a real number () > 0 such that all y Y with |y y0 | < () satisfy Ua (y) (y) + (1 )p (T, y0 ) a log |y y0 | + log 2eM .
Proof. First, we try to estimate (T, y , r) when y L is near y0 . Set (x) = (1 )y0 (x) + r /2 (x) = max y (x), (1 )y0 (x) + r /2 (x) = y (x) |ey (x) ey0 (x) | This inequality implies y (x) y (x) y0 (x) + log 1 + M |y y0 | ey0 (x) if if r 1 if r y0 (x) y0 (x) y0 (x) r1 r r0
and verify that this denition is coherent when |y y0 | is small enough. By hypothesis M |y y0 | .
y0 (x) + log 1 M |y y0 | ey0 (x) .
In particular, for y0 (x) = r, we have (1 )y0 (x) + r /2 = r /2, thus y (x) r + log(1 M |y y0 | er ).
179
Similarly, for y0 (x) = r 1, we have (1 )y0 (x) + r /2 = r 1 + /2, thus y (x) r 1 + log(1 + M |y y0 | e1r ). /2 , i.e. 2eM The denition of is thus coherent as soon as M |y y0 | e1r log |y y0 | + log r.
In this case coincides with y on a neighborhood of { = r} , and with (1 )y0 (x) + r /2 on a neighborhood of the polar set 1 (). By Stokes formula applied to (T, , r), we infer (T, y , r) = (T, , r) (T, ) = (1 )p (T, y0 ). From (8.10) we get V (y, r) Ua (y) (8.13) Ua (y) r(T, y , r), hence (y) + r (T, y , r) a , (y) + V (y, r) ar
(y) + r (1 )p (T, y0 ) a .
Suppose log |y y0 | + log(2eM/) r0 , i.e. |y y0 | (/2eM )1/ er0 / ; one can then choose r = log |y y0 | + log(2eM/), and by (8.13) this yields the inequality asserted in Th. 8.12. Fourth step: application of the Hrmander-Bombieri-Skoda theorem. o
The end of the proof relies on the following crucial result, which is a consequence of the Hrmander-Bombieri-Skoda theorem ([Bombieri 1970], [Skoda 1972a], [Skoda 1976]); o it will be proved in Chapter 8, see Cor. 8.?.?. (8.14) Proposition. Let u be a plurisubharmonic function on a complex manifold Y . The set of points in a neighborhood of which eu is not integrable is an analytic subset of Y . Proof of Theorem 8.4 (end).. The main idea in what follows is due to [Kiselman 1979]. For a, b > 0, we let Za,b be the set of points in a neighborhood of which exp(Ua /b) is not integrable. Then Za,b is analytic, and as the family (Ua ) is increasing in a, we have Za ,b Za ,b if a a , b b . Let y0 Y be a given point. If y0 Ec , then (T, y0 ) < c by denition of Ec . / Choose a such that (T, y0 ) < a < c. Lemma 8.9 a) implies that Ua is bounded below in a neighborhood of y0 , thus exp(Ua /b) is integrable and y0 Za,b for all b > 0. / On the other hand, if y0 Ec and if a < c, then Lemma 8.12 implies for all > 0 that Ua (y) (1 )(c a) log |y y0 | + C()
on a neighborhood of y0 . Hence exp(Ua /b) is non integrable at y0 as soon as b < (c a)/2m, where m = dim Y . We obtain therefore Ec =
a<c b<(ca)/2m
Za,b .
180
This proves that Ec is an analytic subset of Y . Finally, we use Cor. 8.5 to derive an important decomposition formula for currents, which is again due to [Siu 1974]. We rst begin by two simple observations. (8.15) Lemma. If T is a closed positive current of bidimension (p, p) and A is an irreducible analytic set in X, we set mA = inf{(T, x) ; x A}. Then (T, x) = mA for all x A A , where (A ) is a countable family of proper j j analytic subsets of A. We say that mA is the generic Lelong number of T along A. Proof. By denition of mA and Ec (T ), we have (T, x) (T, x) = mA on A
cQ, c>mA
mA for every x A and A Ec (T ).
However, for c > mA , the intersection A Ec (T ) is a proper analytic subset of A. (8.16) Proposition. Let T be a closed positive current of bidimension (p, p) and let A be an irreducible p-dimensional analytic subset of X. Then 1 A T = mA [A], in particular l T mA [A] is positive. Proof. As the question is local and as a closed positive current of bidimension (p, p) cannot carry any mass on a (p 1)-dimensional analytic subset, it is enough to work in a neighborhood of a regular point x0 A. Hence, by choosing suitable coordinates, we can suppose that X is an open set in Cn and that A is the intersection of X with a p-dimensional linear subspace. Then, for every point a A, the inequality (T, a) mA implies T B(a, r) mA p r 2p /p! = mA [A] B(a, r) for all r such that B(a, r) X. Now, set = T mA [A] and = ddc |z|2 . Our inequality says that 1 B(a,r) p l 0. If we integrate this with respect to some positive continuous function f with compact support in A, we get X gr p 0 where gr (z) =
A
1 B(a,r) (z) f (a) d(a) = l

aAB(z,r)
f (a) d(a).
Here gr is continuous on Cn , and as r tends to 0 the function gr (z)/( p r 2p /p!) converges to f on A and to 0 on X A, with a global uniform bound. Hence we obtain 1 A f p l 0. Since this inequality is true for all continuous functions f 0 with compact support in A, we conclude that the measure 1 A p is positive. By a linear change of coordinates, l we see that n 0 1 A ddc l j |uj |2
1 j n
for every basis (u1 , . . . , un ) of linear forms and for all coecients j > 0. Take 1 = . . . = p = 1 and let the other j tend to 0. Then we get 1 A idu1 du1 . . . dup dup l 0.
181
This implies 1 A 0, or equivalently 1 A T mA [A]. By Cor. 2.4 we know that 1 A T is l l l a closed positive current, thus 1 A T = [A] with 0. We have just seen that mA . l On the other hand, T 1 A T = [A] clearly implies mA . l (8.16) Sius decomposition formula. If T is a closed positive current of bidimension (p, p), there is a unique decomposition of T as a (possibly nite) weakly convergent series T = j [Aj ] + R, j > 0,
j 1
where [Aj ] is the current of integration over an irreducible p-dimensional analytic set Aj X and where R is a closed positive current with the property that dim Ec (R) 0. Proof of uniqueness.. If T has such a decomposition, the p-dimensional components of Ec (T ) are (Aj )j c , for (T, x) = j ([Aj ], x) + (R, x) is non zero only on Aj Ec (R), and is equal to j generically on Aj more precisely, (T, x) = j at every regular point of Aj which does not belong to any intersection Aj Ak , k = j or to Ec (R) . In particular Aj and j are unique. Proof of existence.. Let (Aj )j 1 be the countable collection of p-dimensional components occurring in one of the sets Ec (T ), c Q , and let j > 0 be the generic Lelong number + of T along Aj . Then Prop. 8.16 shows by induction on N that RN = T 1 j N j [Aj ] is positive. As RN is a decreasing sequence, there must be a limit R = limN+ RN in the weak topology. Thus we have the asserted decomposition. By construction, R has zero generic Lelong number along Aj , so dim Ec (R) 0. It is very important to note that some components of lower dimension can actually occur in Ec (R), but they cannot be subtracted because R has bidimension (p, p). A typical case is the case of a bidimension (n 1, n 1) current T = ddc u with u = log(|Fj |1 +. . . |FN |N ) and Fj (X). In general Ec (T ) = Fj1 (0) has dimension < n 1. In that case, an important formula due to King plays the role of (8.17). We state it in a somewhat more general form than its original version [King 1970]. (8.18) Kings formula. Let F1 , . . . , FN be holomorphic functions on a complex manp, and set u = ifold X, such that the zero variety Z = Fj1 (0) has codimension log |Fj |j with arbitrary coecients j > 0. Let (Zk )k 1 be the irreducible components of Z of codimension p exactly. Then there exist multiplicities k > 0 such that (ddc u)p =
k 1
k [Zk ] + R,
where R is a closed positive current such that 1 Z R = 0 and codim Ec (R) > p for every l c > 0. Moreover the multiplicities k are integers if 1 , . . . , N are integers, and k = 1 . . . p if 1 . . . N and some partial Jacobian determinant of (F1 , . . . , Fp ) of order p does not vanish identically along Zk . Proof. Observe that (ddc u)p is well dened thanks to Cor. 4.11. The comparison theorem 7.8 applied with (z) = log |z x|, v1 = . . . = vp = u, u1 = . . . = up = and T = 1 shows that the Lelong number of (ddc u)p is equal to 0 at every point of X Z. Hence Ec ((ddc u)p ) is contained in Z and its (np)-dimensional components are members of the
182
family (Zk ). The asserted decomposition follows from Sius formula 8.16. We must have 1 Zk R = 0 for all irreducible components of Z: when codim Zk > p this is automatically l true, and when codim Zk = p this follows from (8.16) and the fact that codim Ec (R) > p. If det(Fj /zk )1 j,k p = 0 at some point x0 Zk , then (Z, x0 ) = (Zk , x0 ) is a smooth germ dened by the equations F1 = . . . = Fp = 0. If we denote v = log j p |Fj |j with 1 . . . N , then u v near Zk and Th. 7.8 implies ((ddc u)p , x) = ((ddc v)p , x) for all x Zk near x0 . On the other hand, if G := (F1 , . . . , Fp ) : X Cp , Cor. 7.4 gives (ddc v)p = G ddc log
1 j p
|zj |j
= 1 . . . p G 0 = 1 . . . p [Zk ]
near x0 . This implies that the generic Lelong number of (ddc u)p along Zk is k = 1 . . . p . The integrality of k when 1 , . . . , N are integers will be proved in the next section.
9. Transformation of Lelong Numbers by Direct Images

Let F : X Y be a holomorphic map between complex manifolds of respective dimensions dim X = n, dim Y = m, and let T be a closed positive current of bidimension (p, p) on X. If FSupp T is proper, the direct image F T is dened by (9.1) F T, = T, F
for every test form of bidegree (p, p) on Y . This makes sense because Supp T F 1 (Supp ) is compact. It is easily seen that F T is a closed positive current of bidimension (p, p) on Y . (9.2) Example. Let T = [A] where A is a p-dimensional irreducible analytic set in X such that FA is proper. We know by Remmerts theorem 2.7.8 that F (A) is an analytic set in Y . Two cases may occur. Either FA is generically nite and F induces an tale e 1 covering A F (Z) F (A) Z for some nowhere dense analytic subset Z F (A), or FA has generic bers of positive dimension and dim F (A) < dim A. In the rst case, let s < + be the covering degree. Then for every test form of bidegree (p, p) on Y we get F [A], =
A
F =
A F 1 (Z)
F = s
F (A) Z
= s [F (A)],
because Z and F 1 (Z) are negligible sets. Hence F [A] = s[F (A)]. On the other hand, if dim F (A) < dim A = p, the restriction of to F (A)reg is zero, and therefore so is this the restriction of F to Areg . Hence F [A] = 0. Now, let be a continuous plurisubharmonic function on Y which is semi-exhaustive on F (Supp T ) (this set certainly contains Supp F T ). Since FSupp T is proper, it follows that F is semi-exhaustive on Supp T , for Supp T { F < R} = F 1 F (Supp T ) { < R} . (9.3) Proposition. If F (Supp T ) { < R} Y , we have (F T, , r) = (T, F, r) for all r < R,
183
in particular (F T, ) = (T, F ). Here, we do not necessarily assume that X or Y are Stein; we thus replace with s = max{, s}, s < r, in the denition of (F T, , r) and (T, F, r).
Proof. The rst equality can be written F T 1 {<r} (ddc s )p = l T (1 {<r} F )(ddc l
s
F )p .
This follows almost immediately from the adjunction formula (9.1) when is smooth and when we write 1 {<R} = lim gk for some sequence of smooth functions gk . In l general, we write s as a decreasing limit of smooth plurisubharmonic functions and we apply our monotone continuity theorems (if Y is not Stein, Richbergs theorem shows that we can obtain a decreasing sequence of almost plurisubharmonic approximations such that the negative part of ddc converges uniformly to 0 ; this is good enough to apply the monotone continuity theorem; note that the integration is made on compact subsets, thanks to the semi-exhaustivity assumption on ). It follows from this that understanding the transformation of Lelong numbers under direct images is equivalent to understanding the eect of F on the weight. We are mostly interested in computing the ordinary Lelong numbers (F T, y) associated with the weight (w) = log |w y| in some local coordinates (w1 , . . . , wm ) on Y near y. Then Prop. 9.3 gives (9.4) (F T, y) = (T, log |F y|) with 1 |Fj (z) yj |2 , Fj = wj F. log |F (z) y| = log 2
We are going to show that (T, log |F y|) is bounded below by a linear combination of the Lelong numbers of T at points x in the ber F 1 (y), with suitable multiplicities attached to F at these points. These multiplicities can be seen as generalizations of the notion of multiplicity of an analytic map introduced by [Stoll 1966]. (9.5) Denition. Let x X and y = F (x). Suppose that the codimension of the ber F 1 (y) at x is p. Then we set p (F, x) = (ddc log |F y|)p , x . Observe that (ddc log |F y|)p is well dened thanks to Cor. 4.10. The second comparison theorem 7.8 immediately shows that p (F, x) is independent of the choice of local coordinates on Y (and also on X, since Lelong nombers do not depend on coordinates). By denition, p (F, x) is the mass carried by {x} of the measure (ddc log |F (z) y|)p (ddc log |z x|)np . We are going to give a more geometric interpretation of this multiplicity, from which it will follow that p (F, x) is always a positive integer (in particular, the proof of (8.18) will be complete).
184
(9.6) Example. For p = n = dim X, the assumption codimx F 1 (y) p means that the germ of map F : (X, x) (Y, y) is nite. Let Ux be a neighborhood of x such that U x F 1 (y) = {x}, let Wy be a neighborhood of y disjoint from F (Ux ) and let Vx = Ux F 1 (Wy ). Then F : Vx Wy is proper and nite, and we have F [Vx ] = s [F (Vx )] where s is the local covering degree of F : Vx F (Vx ) at x. Therefore n (F, x) =
{x}
ddc log |F y|
= [Vx ], log |F y| = F [Vx ], y
= s F (Vx ), y . In the particular case when dim Y = dim X, we have (F (Vx ), y) = (Y, y), so n (F, x) = s. In general, it is a well known fact that the ideal generated by (F1 y1 , . . . , Fm ym ) in X,x has the same integral closure as the ideal generated by n generic linear combinations of the generators, that is, for a generic choice of coordinates w = (w1 , . . . , wn ), w = (wn+1 , . . . , wm ) on (Y, y), we have |F (z) y| C|w F (z)| (this is a simple consequence of Lemma 7.5 applied to A = F (Vx )). Hence for p = n, the comparison theorem 7.1 gives n (F, x) = n (w F, x) = local covering degree of w F at x, for a generic choice of coordinates (w , w ) on (Y, y). (9.7) Geometric interpretation of p (F, x). An application of Croftons formula 7.11 shows, after a translation, that there is a small ball B(x, r0 ) on which (ddc log |F (z) y|)p (ddc log |z x|)np = (9.7 a)
SG(p,n)
(ddc log |F (z) y|)p [x + S] dv(S).
For a rigorous proof of (9.7 a), we replace log |F (z) y| by the regularized function 1 2 2 2 log(|F (z) y| + ) and let tend to 0 on both sides. By (4.3) (resp. by (4.10)), the c wedge product (dd log |F (z) y|)p [x + S] is well dened on a small ball B(x, r0 ) as soon as x + S does not intersect F 1 (y) B(x, r0 ) (resp. intersects F 1 (y) B(x, r0 ) at nitely many points); thanks to the assumption codim(F 1 (y), x) p, Sards theorem shows that this is the case for all S outside a negligible closed subset E in G(p, n) (resp. by Bertini, an analytic subset A in G(p, n) with A E). Fatous lemma then implies that the inequality holds in (9.7 a). To get equality, we observe that we have bounded convergence on all complements G(p, n) V (E) of neighborhoods V (E) of E. However the mass of V (E) [x + S] dv(S) in B(x, r0 ) is proportional to v(V (E)) and therefore tends to 0 when V (E) is small; this is sucient to complete the proof, since Prop. 4.6 b) gives ddc log(|F (z) y|2 + 2 )
p
zB(x,r0 )
[x + S] dv(S)
SV (E)
C v(V (E))
with a constant C independent of . By evaluating (9.7 a) on {x}, we get (9.7 b) p (F, x) =

SG(p,n) A
(ddc log |Fx+S z|)p , x dv(S).
185
Let us choose a linear parametrization gS : Cp S depending analytically on local coordinates of S in G(p, n). Then Theorem 8.4 with T = [Cp ] and (z, S) = log |F gS (z) y| shows that (ddc log |Fx+S z|)p , x = [Cp ], log |F gS (z) y| is Zariski upper semicontinuous in S on G(p, n) A. However, (9.6) shows that these numbers are integers, so S (ddc log |Fx+S z|)p , x must be constant on a Zariski open subset in G(p, n). By (9.7 b), we obtain (9.7 c) p (F, x) = p (Fx+S , x) = local degree of w Fx+S at x
for generic subspaces S G(p, n) and generic coordinates w = (w1 , . . . , wp ), w = (wp+1 , . . . , wm ) on (Y, y). (9.8) Example. Let F : Cn Cn be dened by
s s F (z1 , . . . , zn ) = (z11 , . . . , znn ),
s1
...
sn .
We claim that p (F, 0) = s1 . . . sp . In fact, for a generic p-dimensional subspace S Cn such that z1 , . . . , zp are coordinates on S and zp+1 , . . . , zn are linear forms in z1 , . . . , zp , and for generic coordinates w = (w1 , . . . , wp ), w = (wp+1 , . . . , wn ) on Cn , we can rears s range w by linear combinations so that wj FS is a linear combination of (zj j , . . . , znn ) sj and has non zero coecient in zj as a polynomial in (zj , . . . , zp ). It is then an exercise to show that w FS has covering degree s1 . . . sp at 0 [ compute inductively the roots zn , zn1 , . . . , zj of wj FS (z) = aj and use Lemma II.3.10 to show that the sj values of zj lie near 0 when (a1 , . . . , ap ) are small ]. We are now ready to prove the main result of this section, which describes the behaviour of Lelong numbers under proper morphisms. A similar weaker result was already proved in [Demailly 1982b] with some other non optimal multiplicities p (F, x). (9.9) Theorem. Let T be a closed positive current of bidimension (p, p) on X and let F : X Y be an analytic map such that the restriction FSupp T is proper. Let I(y) be the set of points x Supp T F 1 (y) such that x is equal to its connected component in Supp T F 1 (y) and codim(F 1 (y), x) p. Then we have (F T, y)
xI(y)
p (F, x) (T, x).
In particular, we have (F T, y) xI(y) (T, x). This inequality no longer holds if the summation is extended to all points x Supp T F 1 (Y ) and if this set contains positive dimensional connected components: for example, if F : X Y contracts some exceptional subspace E in X to a point y0 (e.g. if F is a blow-up map, see 7.12), then T = [E] has direct image F [E] = 0 thanks to (9.2). Proof. We proceed in three steps.
186
Step 1. Reduction to the case of a single point x in the ber. It is sucient to prove the inequality when the summation is taken over an arbitrary nite subset {x1 , . . . , xN } of I(y). As xj is equal to its connected component in Supp T F 1 (y), it has a fondamental system of relative open-closed neighborhoods, hence there are disjoint neighborhoods Uj of xj such that Uj does not intersect Supp T F 1 (y). Then the image F (Uj Supp T ) is a closed set which does not contain y. Let W be a neighborhood of y disjoint from all sets F (Uj Supp T ), and let Vj = Uj F 1 (W ). It is clear that Vj is a neighborhood of xj and that FVj : Vj W has a proper restriction to Supp T Vj . Moreover, we obviously have F T j (FVj ) T on W . Therefore, it is enough to check the inequality (F T, y) p (F, x) (T, x) for a single point x I(y), in the case when X Cn , Y Cm are open subsets and x = y = 0. Step 2. Reduction to the case when F is nite. By (9.4), we have (F T, 0) = inf T (ddc log |F |)p
V
V 0
= inf lim
V 0 0
T ddc log(|F | + |z|N ) ,
and the integrals are well dened as soon as V does not intersect the set Supp T F 1 (0) (may be after replacing log |F | by max{log |F |, s} with s 0). For every V and , the last integral is larger than (G T, 0) where G is the nite morphism dened by G : X Y Cn ,
N N (z1 , . . . , zn ) (F1 (z), . . . , Fm (z), z1 , . . . , zn ).
We claim that for N large enough we have p (F, 0) = p (G, 0). In fact, x I(y) implies by denition codim(F 1 (0), 0) p. Hence, if S = {u1 = . . . = unp = 0} is a generic p-dimensional subspace of Cn , the germ of variety F 1 (0) S dened by (F1 , . . . , Fm , u1 , . . . , unp ) is {0}. Hilberts Nullstellensatz implies that some powers of z1 , . . . , zn are in the ideal (Fj , uk ). Therefore |F (z)| + |u(z)| C|z|a near 0 for some integer a independent of S (to see this, take coecients of the uk s as additional variables); in particular |F (z)| C|z|a for z S near 0. The comparison theorem 7.1 then shows that p (F, 0) = p (G, 0) for N a. If we are able to prove that (G T, 0) p (G, 0)(T, 0) in case G is nite, the obvious inequality (F T, 0) (G T, 0) concludes the proof. Step 3. Proof of the inequality (F T, y) p (F, x) (T, x) when F is nite and F 1 (y) = x. Then (z) = log |F (z) y| has a single isolated pole at x and we have p (F, x) = ((ddc )p , x). It is therefore sucient to apply to following Proposition. (9.10) Proposition. Let be a semi-exhaustive continuous plurisubharmonic function on X with a single isolated pole at x. Then (T, ) (T, x) ((ddc )p , x).
Proof. Since the question is local, we can suppose that X is the ball B(0, r0 ) in Cn and x = 0. Set X = B(0, r1 ) with r1 < r0 and (z, g) = g(z) for g Gln (C). Then there is a small neighborhood of the unitary group U (n) Gln (C) such that is plurisubharmonic on X and semi-exhaustive with respect to X . Theorem 8.4
187
implies that the map g (T, g) is Zariski upper semi-continuous on . In particular, we must have (T, g) (T, ) for all g A in the complement of a complex analytic set A. Since Gln (C) is the complexication of U (n), the intersection U (n) A must be a nowhere dense real analytic subset of U (n). Therefore, if dv is the Haar measure of mass 1 on U (n), we have (T, )
gU(n)
(T, g) dv(g) dv(g)

B(0,r)
(9.11) Since
= lim
r0
gU(n)
T (ddc g)p .
gU(n)
(ddc g)p dv(g) is a unitary invariant (p, p)-form on B, Lemma 7.10 implies (ddc g)p dv(g) = ddc (log |z|)
p
gU(n)
where is a convex increasing function. The Lelong number at 0 of the left hand side is equal to ((ddc )p , 0), and must be equal to the Lelong number of the right hand side, which is limt (t)p (to see this, use either Formula (5.5) or Th. 7.8). Thanks to the last equality, Formulas (9.11) and (5.5) imply (T, ) lim T ddc (log |z|)
p
r0 r0
= lim (log r 0)p (T, 0, r)
B(0,r)
((ddc )p , 0) (T, 0).
Another interesting question is to know whether it is possible to get inequalities in the opposite direction, i.e. to nd upper bounds for (F T, y) in terms of the Lelong numbers (T, x). The example T = [] with the curve : t (ta , ta+1 , t) in C3 and F : C3 C2 , (z1 , z2 , z3 ) (z1 , z2 ), for which (T, 0) = 1 and (F T, 0) = a, shows that this may be possible only when F is nite. In this case, we have: (9.12) Theorem. Let F : X Y be a proper and nite analytic map and let T be a closed positive current of bidimension (p, p) on X. Then (a) (F T, y)
xSupp T F 1 (y)
p (F, x) (T, x)
where p (F, x) is the multiplicity dened as follows: if H : (X, x) (Cn , 0) is a germ of nite map, we set (b) (c) (H, x) = inf > 0 ; C > 0, |H(z)| p (F, x) = inf
G
(G F, x)p , p (G, 0)
C|z x| near x ,
where G runs over all germs of maps (Y, y) (Cn , 0) such that G F is nite. Proof. If F 1 (y) = {x1 , . . . , xN }, there is a neighborhood W of y and disjoint neighborhoods Vj of xj such that F 1 (W ) = Vj . Then F T = (FVj ) T on W , so it
188
is enough to consider the case when F 1 (y) consists of a single point x. Therefore, we assume that F : V W is proper and nite, where V , W are neighborhoods of 0 in Cn , Cm and F 1 (0) = {0}. Let G : (Cm , 0) (Cn , 0) be a germ of map such that G F is nite. Hilberts Nullstellensatz shows that there exists > 0 and C > 0 such that |G F (z)| C|z| near 0. Then the comparison theorem 7.1 implies (G F T, 0) = (T, log |G F |) p (T, log |z|) = p (T, 0).
On the other hand, Th. 9.9 applied to = F T on W gives (G F T, 0) Therefore (F T, 0) p (G, 0) (F T, 0).
p (T, 0). p (G, 0)
The inmum of all possible values of is by denition (G F, 0), thus by taking the inmum over G we obtain (F T, 0) p (F, 0) (T, 0).
s s (9.13) Example. Let F (z1 , . . . , zn ) = (z11 , . . . , znn ), s1 . . . sn as in 9.8. Then we have p (F, 0) = s1 . . . sp , p (F, 0) = snp+1 . . . sn .
To see this, let s be the lowest common multiple of s1 , . . . , sn and let G(z1 , . . . , zn ) = s/s s/s (z1 1 , . . . , zn n ). Clearly p (G, 0) = (s/snp+1 ) . . . (s/sn ) and (G F, 0) = s, so we get by denition p (F, 0) snp+1 . . . sn . Finally, if T = [A] is the current of integration over the p-dimensional subspace A = {z1 = . . . = znp = 0}, then F [A] = snp+1 . . . sn [A] because FA has covering degree snp+1 . . . sn . Theorem 9.12 shows that we must have snp+1 . . . sn p (F, 0), QED. If 1 . . . n are positive real numbers and sj is taken to be the integer part of kj as k tends to +, Theorems 9.9 and 9.12 imply in the limit the following: (9.14) Corollary. For 0 < 1 the inequalities 1 . . . p (T, x) ... n , Kiselmans directional Lelong numbers satisfy
(T, x, )
np+1 . . . n (T, x).
(9.15) Remark. It would be interesting to have a direct geometric interpretation of p (F, x). In fact, we do not even know whether p (F, x) is always an integer.
10. A Schwarz Lemma. Application to Number Theory
189

In this section, we show how Jensens formula and Lelong numbers can be used to prove a fairly general Schwarz lemma relating growth and zeros of entire functions in Cn . In order to simplify notations, we denote by |F |r the supremum of the modulus of a function F on the ball of center 0 and radius r. Then, following [Demailly 1982a], we present some applications with a more arithmetical avour. (10.1) Schwarz lemma. Let P1 , . . . , PN C[z1 , . . . , zn ] be polynomials of degree , such that their homogeneous parts of degree do not vanish simultaneously except at 0. Then there is a constant C 2 such that for all entire functions F (Cn ) and all R r 1 we have log |F |r log |F |R 1n ([ZF ], log |P |) log R Cr
where ZF is the zero divisor of F and P = (P1 , . . . , PN ) : Cn CN . Moreover ([ZF ], log |P |) ord(F, w) n1 (P, w)
wP 1 (0)
where ord(F, w) denotes the vanishing order of F at w and n1 (P, w) is the (n 1)multiplicity of P at w, as dened in (9.5) and (9.7). Proof. Our assumptions imply that P is a proper and nite map. The last inequality is then just a formal consequence of formula (9.4) and Th. 9.9 applied to T = [ZF ]. Let Qj be the homogeneous part of degree in Pj . For z0 B(0, r), we introduce the weight functions (z) = log |P (z)|, (z) = log |Q(z z0 )|. Since Q1 (0) = {0} by hypothesis, the homogeneity of Q shows that there are constants C1 , C2 > 0 such that (10.2) C1 |z| |Q(z)| C2 |z| on Cn .
The homogeneity also implies (ddc )n = n z0 . We apply the Lelong Jensen formula 6.5 to the measures ,s associated with and to V = log |F |. This gives
s
(10.3)
,s (log |F |) n log |F (z0 )| =
dt
{<t}
[ZF ] (ddc )n1 .
By (6.2), ,s has total mass n and has support in {(z) = s} = {Q(z z0 ) = es } B 0, r + (es /C1 )1/ . Note that the inequality in the Schwarz lemma is obvious if R R Cr 2r. We take s = log(R/2) + log C1 ; then {(z) = s} B(0, r + R/2) B(0, R). Cr, so we can assume
190
In particular, we get ,s (log |F |) (10.4) log |F |R log |F (z0 )|
n log |F |R and formula (10.3) gives

s
n
s0
dt
{<t}
[ZF ] (ddc )n1
for any real number s0 < s. The proof will be complete if we are able to compare the integral in (10.4) to the corresponding integral with in place of . The argument for this is quite similar to the proof of the comparison theorem, if we observe that at innity. We introduce the auxiliary function w= max{, (1 ) + t } on { (1 ) + t on { t 2}, t 2},
with a constant to be determined later, such that (1 ) + t > near { = t 2} and (1 ) + t < near { = t}. Then Stokes theorem implies
{<t}
[ZF ] (ddc )n1 = (1 )n1
{<t}
[ZF ] (ddc w)n1 (1 )n1 ([ZF ], log |P |).
(10.5)
{<t2}
[ZF ] (ddc )n1
By (10.2) and our hypothesis |z0 | < r, the condition (z) = t implies |P (z) Q(z z0 )| P (z) 1 Q(z z0 ) |Q(z z0 )| = et = et/ /C1
1/
C3 (1 + |z0 |)(1 + |z| + |z0 |)1 C4 ret/ (ret/ + 1)1 s0
|z z0 |
et/ /C2 , C4 r(r + et/ )1 , 21 C4 ret/ ,
1/
provided that t
log r. Hence for (z) = t |(z) (z)| = log
log(2 C4 r), we get C5 ret/ .
|P (z)| |Q(z z0 )|
Now, we have (1 ) + t = (1 )( ) + (t 1 ), so this dierence is < C5 ret/ on { = t} and > C5 re(2t)/ + on { = t 2}. Hence it is sucient to take = C5 re(2t)/ . This number has to be < 1, so we take t s0 2 + log(C5 r). Moreover, (10.5) actually holds only if P 1 (0) { < t 2}, so by (10.2) it is enough to take t s0 2 + log(C2 (r + C6 ) ) where C6 is such that 1 P (0) B(0, C6 ). Finally, we see that we can choose s = log R C7 , and inequalities (10.4), (10.5) together imply
s
s0 = log r + C8 ,
log |F |R log |F (z0 )|
n
s0
(1 C5 re(2t)/ )n1 dt ([ZF ], log |P |).
191
The integral is bounded below by

log(R/r)C7 C8
(1 C9 et/ ) dt
log(R/Cr).
This concludes the proof, by taking the inmum when z0 runs over B(0, r). (10.6) Corollary. Let S be a nite subset of Cn and let be the minimal degree of algebraic hypersurfaces containing S. Then there is a constant C 2 such that for all n F (C ) and all R r 1 we have log |F |r log |F |R ord(F, S) + n(n 1)/2 R log n! Cr
where ord(F, S) = minwS ord(F, w). Proof. In view of Th. 10.1, we only have to select suitable polynomials P1 , . . . , PN . The vector space C[z1 , . . . , zn ]< of polynomials of degree < in Cn has dimension m() = +n1 n = ( + 1) . . . ( + n 1) . n!
By denition of , the linear forms C[z1 , . . . , zn ]< C, P P (w), w S
vanish simultaneously only when P = 0. Hence we can nd m = m() points w1 , . . . , wm S such that the linear forms P P (wj ) dene a basis of C[z1 , . . . , zn ] . This < means that there is a unique polynomial P C[z1 , . . . , zn ]< which takes given values P (wj ) for 1 j m. In particular, for every multiindex , || = , there is a unique polynomial R C[z1 , . . . , zn ]< such that R (wj ) = wj . Then the polynomials P (z) = z R (z) have degree , vanish at all points wj and their homogeneous parts of maximum degree Q (z) = z do not vanish simultaneously except at 0. We simply use the fact that n1 (P, wj ) 1 to get ([ZF ], log |P |) ord(F, w)
wP 1 (0)
m() ord(F, S).
Theorem 10.1 then gives the desired inequality, because m() is a polynomial with positive coecients and with leading terms 1 n + n(n 1)/2 n1 + . . . . n! Let S be a nite subset of Cn . According to [Waldschmidt 1976], we introduce for every integer t > 0 a number t (S) equal to the minimal degree of polynomials P C[z1 , . . . , zn ] which vanish at order t at every point of S. The obvious subadditivity property t1 +t2 (S) t1 (S) + t2 (S)
192
easily shows that (S) := inf t (S) t (S) = lim . t+ t>0 t t
We call 1 (S) the degree of S (minimal degree of algebraic hypersurfaces containing S) and (S) the singular degree of S. If we apply Cor. 10.6 to a polynomial F vanishing at order t on S and x r = 1, we get log |F |R with = 1 (S), in particular deg F t 1 (S) + n(n 1)/2 . n! t + n(n 1)/2 R log + log |F |1 n! C
The minimum of deg F over all such F is by denition t (S). If we divide by t and take the inmum over t, we get the interesting inequality (10.7) t (S) t (S) 1 (S) + n(n 1)/2 . n!
(10.8) Remark. The constant 1 (S)+n(n1)/2 in (10.6) and (10.7) is optimal for n = 1, 2 n! but not for n 3. It can be shown by means of Hrmanders L2 estimates [Waldschmidt o 1978] that for every > 0 the Schwarz lemma (10.6) holds with coecient (S) : log |F |r log |F |R ord(F, S)((S) ) log R , C r
and that (S) (u (S) + 1)/(u + n 1) for every u 1 ; this last inequality is due to [Esnault-Viehweg 1983], who used deep tools of algebraic geometry; [Azhari 1990] reproved it recently by means of Hrmanders L2 estimates. Rather simple examples o [Demailly 1982a] lead to the conjecture (S) u (S) + n 1 u+n1 for every u 1.
The special case u = 1 of the conjecture was rst stated by [Chudnovsky 1979]. Finally, let us mention that Cor. 10.6 contains Bombieris theorem on algebraic values of meromorphic maps satisfying algebraic dierential equations [Bombieri 1970]. Recall that an entire function F (Cn ) is said to be of order if for every > 0 there is a constant C such that |F (z)| C exp(|z|+ ). A meromorphic function is said to be of order if it can be written G/H where G, H are entire functions of order . (10.9) Theorem ([Bombieri 1970]). Let F1 , . . . , FN be meromorphic functions on Cn , such that F1 , . . . , Fd , n < d N , are algebraically independent over Q and have nite orders 1 , . . . , d . Let K be a number eld of degree [K : Q]. Suppose that the ring K[f1 , . . . , fN ] is stable under all derivations d/dz1 , . . . , d/dzn . Then the set S of
193
points z Cn , distinct from the poles of the Fj s, such that (F1 (z), . . . , FN (z)) K N is contained in an algebraic hypersurface whose degree satises + n(n 1)/2 n! 1 + . . . + d [K : Q]. dn
Proof. If the set S is not contained in any algebraic hypersurface of degree < , the linear algebra argument used in the proof of Cor. 10.6 shows that we can nd m = m() points w1 , . . . , wm S which are not located on any algebraic hypersurface of degree < . Let H1 , . . . , Hd be the denominators of F1 , . . . , Fd . The standard arithmetical methods of transcendental number theory allow us to construct a sequence of entire functions in the following way: we set G = P (F1 , . . . , Fd )(H1 . . . Hd )s where P is a polynomial of degree s in each variable with integer coecients. The polynomials P are chosen so that G vanishes at a very high order at each point wj . This amounts to solving a linear system whose unknowns are the coecients of P and whose coecients are polynomials in the derivatives of the Fj s (hence lying in the number eld K). Careful estimates of size and denominators and a use of the Dirichlet-Siegel box principle lead to the following lemma, see e.g. [Waldschmidt 1978]. (10.10) Lemma. For every > 0, there exist constants C1 , C2 > 0, r 1 and an innite sequence Gt of entire functions, t T N (depending on m and on the choice of the points wj ), such that a) Gt vanishes at order b) |Gt |r c) |Gt |R(t) (C1 t)t [K:Q] ;
t C2
t at all points w1 , . . . , wm ;
where R(t) = (tdn / log t)1/(1 +...+d +) .
An application of Cor. 10.6 to F = Gt and R = R(t) gives the desired bound for the degree as t tends to + and tends to 0. If 0 is the largest integer which satises the inequality of Th. 10.9, we get a contradiction if we take = 0 + 1. This shows that S must be contained in an algebraic hypersurface of degree 0 .
195
Chapter IV
Sheaf Cohomology and Spectral Sequences
One of the main topics of this book is the computation of various cohomology groups arising in algebraic geometry. The theory of sheaves provides a general framework in which many cohomology theories can be treated in a unied way. The cohomology theory of sheaves will be constructed here by means of Godements simplicial abby resolution. However, we have emphasized the analogy with Alexander-Spanier cochains in order to give a simple denition of the cup product. In this way, all the basic properties of cohomology groups (long exact sequences, Mayer Vietoris exact sequence, Lerays theorem, relations with Cech cohomology, De Rham-Weil isomorphism theorem) can be derived in a very elementary way from the denitions. Spectral sequences and hypercohomology groups are then introduced, with two principal examples in view: the Leray spectral sequence and the Hodge-Frlicher o spectral sequence. The basic results concerning cohomology groups with constant or locally constant coecients (invariance by homotopy, Poincar duality, Leray-Hirsch theorem) are also included, in order e to present a self-contained approach of algebraic topology.
1. Basic Results of Homological Algebra

Let us rst recall briey some standard notations and results of homological algebra that will be used systematically in the sequel. Let R be a commutative ring with unit. A dierential module (K, d) is a R-module K together with an endomorphism d : K K, called the dierential, such that d d = 0. The modules of cycles and of boundaries of K are dened respectively by (1.1) Z(K) = ker d, B(K) = Im d.
Our hypothesis dd = 0 implies B(K) Z(K). The homology group of K is by denition the quotient module (1.2) H(K) = Z(K)/B(K).
A morphism of dierential modules : K L is a R-homomorphism : K L such that d = d ; here we denote by the same symbol d the dierentials of K and L. It is then clear that Z(K) Z(L) and B(K) B(L). Therefore, we get an induced morphism on homology groups, denoted (1.3) H() : H(K) H(L).
It is easily seen that H is a functor, i.e. H( ) = H() H(). We say that two morphisms , : K L are homotopic if there exists a R-linear map h : K L such that (1.4) d h + h d = .
196
Chapter IV. Sheaf Cohomology and Spectral Sequences
Then h is said to be a homotopy between and . For every cocycle z Z(K), we infer (z) (z) = dh(z), hence the maps H() and H() coincide. The module K itself is said to be homotopic to 0 if IdK is homotopic to 0 ; then H(K) = 0. (1.5) Snake lemma. Let 0 K L M 0 be a short exact sequence of morphisms of dierential modules. Then there exists a homomorphism : H(M ) H(K), called the connecting homomorphism, and a homology exact sequence H(K) H(L) H(M ) / Moreover, to any commutative diagram of short exact sequences 0 K L M 0
H() H()
0 K L M 0
is associated a commutative diagram of homology exact sequences H(K) H(L) H(M ) H(K)

H(K) H(L) H(M ) H(K) . Proof. We rst dene the connecting homomorphism : let m Z(M ) represent a given cohomology class {m} in H(M ). Then {m} = {k} H(K) is the class of any element k 1 d 1 (m), as obtained through the following construction: l L m M d d k K dl L 0 M. The element l is chosen to be a preimage of m by the surjective map ; as (dl) = d(m) = 0, there exists a unique element k K such that (k) = dl. The element k is actually a cocycle in Z(K) because is injective and (dk) = d(k) = d(dl) = 0 = dk = 0. The map will be well dened if we show that the cohomology class {k} depends only on {m} and not on the choices made for the representatives m and l. Consider another
2. The Simplicial Flabby Resolution of a Sheaf
197
representative m = m + dm1 . Let l1 L such that (l1 ) = m1 . Then l has to be replaced by an element l L such that (l ) = m + dm1 = (l + dl1 ). It follows that l = l + dl1 + (k1 ) for some k1 K, hence dl = dl + d(k1 ) = (k) + (dk1 ) = (k ), therefore k = k + dk1 and k has the same cohomology class as k. Now, let us show that ker = Im H(). If {m} is in the image of H(), we can take m = (l) with dl = 0, thus {m} = 0. Conversely, if {m} = {k} = 0, we have k = dk1 for some k1 K, hence dl = (k) = d(k1 ), z := l (k1 ) Z(L) and m = (l) = (z) is in Im H(). We leave the verication of the other equalities Im H() = ker H(), Im = ker H() and of the commutation statement to the reader. In most applications, the dierential modules come with a natural Z-grading. A homological complex is a graded dierential module K = qZ Kq together with a dierential d of degree 1, i.e. d = dq with dq : Kq Kq1 and dq1 dq = q 0. Similarly, a cohomological complex is a graded dierential module K = qZ K with dierentials dq : K q K q+1 such that dq+1 dq = 0 (superscripts are always used instead of subscripts in that case). The corresponding (co)cycle, (co)boundary and (co)homology modules inherit a natural Z-grading. In the case of cohomology, say, these modules will be denoted Z (K ) = Z q (K ), B (K ) = B q (K ), H (K ) = H q (K ).
Unless otherwise stated, morphisms of complexes are assumed to be of degree 0, i.e. of the form = q with q : K q Lq . Any short exact sequence 0 K L M 0 gives rise to a corresponding long exact sequence of cohomology groups (1.6) H q (K ) H q (L ) H q (M ) H q+1 (K )
H q ( ) H q ( ) q H q+1 ( )
and there is a similar homology long exact sequence with a connecting homomorphism q of degree 1. When dealing with commutative diagrams of such sequences, the following simple lemma is often useful; the proof consists in a straightforward diagram chasing. (1.7) Five lemma. Consider a commutative diagram of R-modules A1 A2 A3 A4 A5 5 4 3 2 1 B1 B2 B3 B4 B5 where the rows are exact sequences. If 2 and 4 are injective and 1 surjective, then 3 is injective. If 2 and 4 is surjective and 5 injective, then 3 is surjective. In particular, 3 is an isomorphism as soon as 1 , 2 , 4 , 5 are isomorphisms.
198
2. The Simplicial Flabby Resolution of a Sheaf

Let X be a topological space and let be a sheaf of abelian groups on X (see II-2 for the denition). All the sheaves appearing in the sequel are assumed implicitly to be sheaves of abelian groups, unless otherwise stated. The rst useful notion is that of resolution. (2.1) Denition. A (cohomological) resolution of is a dierential complex of sheaves ( , d) with q = 0, dq = 0 for q < 0, such that there is an exact sequence 0 0 1 q q+1 .
j d0 dq
If : is a morphism of sheaves and ( , d) a resolution of resolutions : is a commutative diagram 0 0

j j
, a morphism of
0 0

d
0
d0
1 1
q q

dq
dq
q+1
q+1 q+1
(2.2) Example. Let X be a dierentiable manifold and q the sheaf of germs of dierential forms of degree q with real values. The exterior derivative d denes a resolution ( , d) of the sheaf R of locally constant functions with real values. In fact Poincars lemma asserts that d is locally exact in degree q 1, and it is clear that the e sections of ker d0 on connected open sets are constants. In the sequel, we will be interested by special resolutions in which the sheaves q have no local rigidity. For that purpose, we introduce abby sheaves, which have become a standard tool in sheaf theory since the publication of Godements book [Godement 1957]. is called abby if for every open subset U of X, the (2.3) Denition. A sheaf restriction map (X) (U ) is onto, i.e. if every section of on U can be extended to X. Let : X be a sheaf on X. We denote by [0] the sheaf of germs of sections X which are not necessarily continuous. In other words, [0] (U ) is the set of all maps f : U such that f (x) x for all x U , or equivalently [0] (U ) = xU x . It is clear that [0] is abby and there is a canonical injection j:
[0]
dened as follows: to any s x we associate the germ s x equal to the continuous section y s(y) near x such that s(x) = s. In the sequel we merely denote s : y s(y) for simplicity. The sheaf [0] is called the canonical abby sheaf associated to . We dene inductively [q] = ( [q1] )[0] .
[0]
3. Cohomology Groups with Values in a Sheaf

[q]
199
The stalk x can be considered as the set of equivalence classes of maps f : X q+1 such that f (x0 , . . . , xq ) xq , with two such maps identied if they coincide on a set of the form (2.4) where V is an open neighborhood of x and V (x0 , . . . , xj ) an open neighborhood of xj , depending on x0 , . . . , xj . This is easily seen by induction on q, if we identify a to the map X [q1], x0 fx0 such that fx0 (x1 , . . . , xq ) = map f : X q+1 f (x0 , x1 , . . . , xq ). Similarly, [q] (U ) is the set of equivalence classes of functions X q+1 (x0 , . . . , xq ) f (x0 , . . . , xq ) xq , with two such functions identied if they coincide on a set of the form (2.4 ) Here, we may of course suppose V (x0 , . . . , xq1 ) . . . V (x0 , x1 ) V (x0 ) U . We dene a dierential dq : [q] [q+1] by (dq f )(x0 , . . . , xq+1 ) =
0 j q
x0 V,
x1 V (x0 ), . . . ,
xq V (x0 , . . . , xq1 ),
x0 U,
x1 V (x0 ), . . . ,
xq V (x0 , . . . , xq1 ).
(2.5)
(1)j f (x0 , . . . , xj , . . . , xq+1 ) + (1)q+1 f (x0 , . . . , xq )(xq+1 ). The meaning of the last term is to be understood as follows: the element s = f (x0 , . . . , xq ) is a germ in xq , therefore s denes a continuous section xq+1 s(xq+1 ) of in a neighborhood V (x0 , . . . , xq ) of xq . In low degrees, we have the formulas (js)(x0 ) = s(x0 ), (2.6)
0
(d f )(x0 , x1 ) = f (x1 ) f (x0 )(x1 ), f [0] , x 1 (d f )(x0 , x1 , x2 ) = f (x1 , x2 ) f (x0 , x2 ) + f (x0 , x1 )(x2 ),
[]
x, [1] x .
(2.7) Theorem ([Godement 1957]). The complex ( , called the simplicial abby resolution of .
, d) is a resolution of the sheaf
Proof. For s x , the associated continuous germ obviously satises s(x0 )(x1 ) = s(x1 ) for x0 V , x1 V (x0 ) small enough. The reader will easily infer from this that d0 j = 0 and dq+1 dq = 0. In order to verify that ( [] , d) is a resolution of , we show that the complex is homotopic to zero for every point x X. Set h0 : hq :
[0] x [q] x
[0] x
d0
[q] x
[1]
dq
,d
[q+1] x 1
= j and
x, [q1] , x
h0 (f ) = f (x) x , hq (f )(x0 , . . . , xq1 ) = f (x, x0 , . . . , xq1 ).
A straightforward computation shows that (hq+1 dq + dq1 hq )(f ) = f for all q Z [q] and f x . If : (2.8) is a sheaf morphism, it is clear that induces a morphism of resolutions [] :
[]
For every short exact sequence exact sequence of sheaf complexes (2.9)
[]
[]
of sheaves, we get a corresponding short

[]
[]
200
3. Cohomology Groups with Values in a Sheaf

3.A. Denition and Functorial Properties If : X is a sheaf of abelian groups, the cohomology groups of on X are (in a vague sense) algebraic invariants which describe the rigidity properties of the global sections of . (3.1) Denition. For every q Z, the q-th cohomology group of X with values in H q (X, ) = H q = ker d : with the convention
[q] q []
is
(X) = (X)
[q+1]
[q]
(X) / Im(dq1 :
[q1]
(X)
[q]
(X)
= 0, dq = 0, H q (X, ) = 0 when q < 0.
For any subset S X, we denote by S the restriction of to S, i.e. the sheaf 1 (S) equipped with the projection S onto S. Then we write H q (S, S ) = S = q H (S, ) for simplicity. When U is open, we see that ( [q] )U coincides with ( U )[q] , thus we have H q (U, ) = H q [] (U ) . It is easy to show that every exact sequence of sheaves 0 0 1 induces an exact sequence (3.2) If we apply this to (3.3) Let : (3.4)
q =
0
[q]
(X) 0 (X) 1 (X). H 0 (X, ) = (X).
, q = 0, 1, we conclude that
be a sheaf morphism; (2.8) shows that there is an induced morphism H q () : H q (X, ) H q (X, )
on cohomology groups. Let 0 Then we have an exact sequence of groups 0 because groups
[0] [0] x. [q]
0 be an exact sequence of sheaves.

[0]
(X)
[0]
(X)
(X) 0
(X) =
xX
Similarly, (2.9) yields for every q an exact sequence of (X)

[q]
(X)
[q]
(X) 0. 0 is associated
If we take (3.3) into account, the snake lemma implies: (3.5) Theorem. To any exact sequence of sheaves 0 a long exact sequence of cohomology groups 0 (X) (X) (X)
H q (X, ) H q (X,
) H q (X,
) H q+1 (X, ) . be its kernel. If
H 1 (X, )
be a surjective sheaf morphism and let (3.6) Corollary. Let 1 H (X, ) = 0, then (X) (X) is surjective.
4. Acyclic Sheaves
201
3.B. Exact Sequence Associated to a Closed Subset Let S be a closed subset of X and U = X (S ), S. For any sheaf X open on X, the presheaf
with the obvious restriction maps satises axioms (II-2.4 ) and (II-2.4 ), so it denes a sheaf on X which we denote by S . This sheaf should not be confused with the restriction sheaf S , which is a sheaf on S. We easily nd (3.7) (
S
)x =
if x S,
)x = 0
if x U.
Observe that these relations would completely fail if S were not closed. The restriction S . We let U be its morphism f fS induces a surjective sheaf morphism kernel, so that we have the relations (3.8) (
U )x
=0
if x S,
U )x
if x U.
From the denition, we obtain in particular (3.9)

S
(X) =
(S),
U (X)
= {sections of
U
(X) vanishing on S}.

S
Theorem 3.5 applied to the exact sequence 0 exact sequence (3.9)
0 on X gives a long
U)
(X) (S) 0 U (X) q q q H (X, U ) H (X, ) H (X,
H 1 (X, S ) H q+1 (X,
U)
3.C. Mayer-Vietoris Exact Sequence Let U1 , U2 be open subsets of X and U = U1 U2 , V = U1 U2 . For any sheaf X and any q we have an exact sequence 0
[q]
on
(U )
[q]
(U1 )
[q]
(U2 )
[q]
(V ) 0
where the injection is given by f (fU1 , fU2 ) and the surjection by (g1 , g2 ) g2V g1V ; the surjectivity of this map follows immediately from the fact that [q] is abby. An application of the snake lemma yields: (3.11) Theorem. For any sheaf on X and any open sets U1 , U2 X, set U = U1 U2 , V = U1 U2 . Then there is an exact sequence H q (U, ) H q (U1 , ) H q (U2 , ) H q (V, ) H q+1 (U, )
202
4. Acyclic Sheaves
Given a sheaf on X, it is usually very important to decide whether the cohomology groups H q (U, ) vanish for q 1, and if this is the case, for which type of open sets U . Note that one cannot expect to have H 0 (U, ) = 0 in general, since a sheaf always has local sections. (4.1) Denition. A sheaf for q 1. is said to be acyclic on an open subset U if H q (U, ) = 0
4.A. Case of Flabby Sheaves We are going to show that abby sheaves are acyclic. First we need the following simple result. (4.2) Proposition. Let be a sheaf with the following property: for every section f of on an open subset U X and every point x X, there exists a neighborhood of x and a section h () such that h = f on U . Then is abby. A consequence of this proposition is that abbiness is a local property: a sheaf abby on X if and only if it is abby on a neighborhood of every point of X. is
Proof. Let f (U ) be given. Consider the set of pairs (v, V ) where v in (V ) is an extension of f on an open subset V U . This set is inductively ordered, so there exists a maximal extension (v, V ) by Zorns lemma. The assumption shows that V must be equal to X. (4.3) Proposition. Let 0 If is abby, the sequence of groups 0 (U ) and
j j
0 be an exact sequence of sheaves.

p
(U )
(U ) 0 is abby.
is exact for every open set U . If
are abby, then
Proof. Let g (U ) be given. Consider the set E of pairs (v, V ) where V is an open subset of U and v (V ) is such that p(v) = g on V . It is clear that E is inductively ordered, so E has a maximal element (v, V ), and we will prove that V = U . Otherwise, in a neighborhood of x such that p(hx ) = gx . let x U V and let h be a section of Then p(h) = g on a neighborhood of x, thus p(v h) = 0 on V and v h = j(u) is abby, u has an extension u (X) and we can dene a with u (V ). If section w (V ) such that p(w) = g by w = v on V, w = h + j(u) on ,
contradicting the maximality of (v, V ). Therefore V = U , v (U ) and p(v) = g on is also abby, v has an extension v (X) and U . The rst statement is proved. If g = p(v) (X) is an extension of g. Hence is abby. (4.4) Theorem. A abby sheaf is acyclic on all open sets U X.
4. Acyclic Sheaves
203
Proof. Let of sheaves
= ker dq :
[q]
[q+1] q
. Then
[q]
=
q+1
and we have an exact sequence 0
dq
because Im dq = ker dq+1 = q+1 . Proposition 4.3 implies by induction on q that all sheaves q are abby, and yields exact sequences 0 For q 1, we nd therefore ker dq :
[q] q
(U )
[q]
(U )
dq
q+1
(U ) 0.
(U )
[]
[q+1]
(U ) =
(U )
[q1]
= Im dq1 :
(U )
[q]
(U ) ,
that is, H q (U, ) = H q
(U ) = 0.
4.B. Soft Sheaves over Paracompact Spaces We now discuss another general situation which produces acyclic sheaves. Recall that a topological space X is said to be paracompact if X is Hausdor and if every open covering of X has a locally nite renement. For instance, it is well known that every metric space is paracompact. A paracompact space X is always normal; in particular, for any locally nite open covering (U ) of X there exists an open covering (V ) such that V U . We will also need another closely related concept. (4.5) Denition. We say that a subspace S is strongly paracompact in X if S is Hausdor and if the following property is satised: for every covering (U ) of S by open sets in X, there exists another such covering (V ) and a neighborhood W of S such that each set W V is contained in some U , and such that every point of S has a neighborhood intersecting only nitely many sets V . It is clear that a strongly paracompact subspace S is itself paracompact. Conversely, the following result is easy to check: (4.6) Lemma. A subspace S is strongly paracompact in X as soon as one of the following situations occurs: a) X is paracompact and S is closed; b) S has a fundamental family of paracompact neighborhoods in X ; c) S is paracompact and has a neighborhood homeomorphic to some product S T , in which S is embedded as a slice S {t0 }. be a sheaf on X and S a strongly paracompact subspace of (4.7) Theorem. Let X. Then every section f of on S can be extended to a section of on some open neighborhood of . Proof. Let f (S). For every point z S there exists an open neighborhood Uz and a section f z (Uz ) such that f z (z) = f (z). After shrinking Uz , we may assume that f z
204
and f coincide on S Uz . Let (V ) be an open covering of S that is locally nite near S and W a neighborhood of S such that W V Uz() (Def. 4.5). We let = xW V ; f z() (x) = f z() (x), , with x V V .
Then ( V ) is an open covering of and all pairs of sections f z() coincide in pairwise intersections. Thus there exists a section F of on which is equal to f z() on V . It remains only to show that is a neighborhood of S. Let z0 S. There exists a neighborhood U of z0 which meets only nitely many sets V1 , . . . , Vp . After shrinking U , we may keep only those Vl such that z0 V l . The sections f z(l ) coincide at z0 , so they coincide on some neighborhood U of this point. Hence W U , so is a neighborhood of S. (4.8) Corollary. If X is paracompact, every section f S extends to a neighborhood of S. (S) dened on a closed set
(4.9) Denition. A sheaf on X is said to be soft if every section f of on a closed set S can be extended to X, i.e. if the restriction map (X) (S) is onto for every closed set S. (4.10) Example. On a paracompact space, every abby sheaf consequence of Cor. 4.8. is soft: this is a
(4.11) Example. On a paracompact space, the Tietze-Urysohn extension theorem shows that the sheaf X of germs of continuous functions on X is a soft sheaf of rings. However, observe that X is not abby as soon as X is not discrete. (4.12) Example. If X is a paracompact dierentiable manifold, the sheaf of functions on X is a soft sheaf of rings.
X
of germs
Until the end of this section, we assume that X is a paracompact topological space. We rst show that softness is a local property. (4.13) Proposition. A sheaf of every point x X. is soft on X if and only if it is soft in a neighborhood
Proof. If is soft on X, it is soft on any closed neighborhood of a given point. Conversely, let (U )I be a locally nite open covering of X which rene some covering by neighborhoods on which is soft. Let (V ) be a ner covering such that V U , and on a closed subset S of X. We consider the set E of pairs f (S) be a section of (g, J), where J I and where g is a section over FJ := S J V , such that g = f on S. As the family (V ) is locally nite, a section of over FJ is continuous as soon it is continuous on S and on each V . Then (f, ) E and E is inductively ordered by the relation (g , J ) (g , J ) if J J and g = g on FJ No element (g, J), J = I, can be maximal: the assumption shows that gFJ V has an / extension to V , thus such a g has an extension to FJ{} for any J. Hence E has a maximal element (g, I) dened on FI = X.
5. Cech Cohomology
205
0 be an exact sequence of sheaves. If (4.14) Proposition. Let 0 is soft, the map (S) (S) is onto for any closed subset S of X. If and are soft, then is soft. By the above inductive method, this result can be proved in a way similar to its analogue for abby sheaves. We therefore obtain: (4.15) Theorem. On a paracompact space, a soft sheaf is acyclic on all closed subsets. (4.16) Denition. The support of a section f (X) is dened by
Supp f = x X ; f (x) = 0 .
X is a local homeomorphism, the equality Supp f is always a closed set: as f (x) = 0 implies f = 0 in a neighborhood of x. is soft and f (X), (4.17) Theorem. Let (U )I be an open covering of X. If there exists a partition of f subordinate to (U ), i.e. a family of sections f (X) such that (Supp f ) is locally nite, Supp f U and f = f on X. Proof. Assume rst that (U ) is locally nite. There exists an open covering (V ) such that V U . Let (f )J , J I, be a maximal family of sections f (X) such that Supp f U and J f = f on S = J V . If J = I and I J, there exists a section f (X) such that f = 0 on X U and f = f f
J
on S V U ) S. This is a
because (X U ) S V is closed and f contradiction unless J = I.
f = 0 on (X
In general, let (Vj ) be a locally nite renement of (U ), such that Vj U(j) , and let (fj ) be a partition of f subordinate to (Vj ). Then f = j1 () fj is the required partition of f . Finally, we discuss a special situation which occurs very often in practice. Let be a sheaf of commutative rings on X ; the rings x are supposed to have a unit element. Assume that is a sheaf of modules over . It is clear that [0] is a [0] -module, and thus also a -module. Therefore all sheaves [q] are -modules and the cohomology groups H q (U, ) have a natural structure of (U )-module. (4.18) Lemma. If
is soft, every sheaf
of
-modules is soft.
Proof. Every section f (S) dened on a closed set S has an extension to some open neighborhood . Let (X) be such that = 1 on S and = 0 on X . Then f , dened as 0 on X , is an extension of f to X. (4.19) Corollary. Let be a sheaf of manifold X. Then H q (X, ) = 0 for all q
X -modules
on a paracompact dierentiable
1.
206
5. Cech Cohomology
5.A. Denitions can In many important circumstances, cohomology groups with values in a sheaf be computed by means of the complex of Cech cochains, which is directly related to the spaces of sections of on suciently ne coverings of X. This more concrete approach was historically the rst one used to dene sheaf cohomology ([Leray 1950], [Cartan 1950]); however Cech cohomology does not always coincide with the good cohomology on non paracompact spaces. Let = (U )I be an open covering of X. For the sake of simplicity, we denote U0 1 ...q = U0 U1 . . . Uq .
The group C q (, ) of Cech q-cochains is the set of families c = (c0 1 ...q )
(U0 1 ...q ).
(0 ,...,q )I q+1
The group structure on C q (, ) is the obvious one deduced from the addition law on sections of . The Cech dierential q : C q (, ) C q+1 (, ) is dened by the formula (5.1) ( q c)0 ...q+1 = (1)j c0 ...j ...q+1 U
0 j q+1
0 ...q+1
and we set C q (, ) = 0, q = 0 for q < 0. In degrees 0 and 1, we get for example (5.2) (5.2 ) q = 0, q = 1, c = (c ), c = (c ), ( 0 c) = c c U , ( 1 c) = c c + c U .
Easy verications left to the reader show that q+1 q = 0. We get therefore a cochain complex C (, ), , called the complex of Cech cochains relative to the covering . (5.3) Denition. The Cech cohomology group of relative to
is
H q (, ) = H q C (, ) . Formula (5.2) shows that the set of Cech 0-cocycles is the set of families (c ) (U ) such that c = c on U U . Such a family denes in a unique way a global section f (X) with fU = c . Hence (5.4) H 0 (, ) = (X). Now, let = (V )J be another open covering of X that is ner than ; this means that there exists a map : J I such that V U() for every J. Then we can dene a morphism : C (, ) C (, ) by (5.5) (q c)0 ...q = c(0 )...(q ) V0 ...q ;
5. Cech Cohomology
207
the commutation property = is immediate. If : J I is another renement map such that V U () for all , the morphisms , are homotopic. To see this, we dene a map hq : C q (, ) C q1 (, ) by (hq c)0 ...q1 = (1)j c(0 )...(j ) (j )... (q1 ) V0 ...q1 .
0 j q1
The homotopy identity q1 hq + hq+1 q = q q is easy to verify. Hence and induce a map depending only on , : (5.6) H q ( ) = H q ( ) : H q (, ) H q (, ).
Now, we want to dene a direct limit H q (X, ) of the groups H q (, ) by means of the renement mappings (5.6). In order to avoid set theoretic diculties, the coverings used in this denition will be considered as subsets of the power set (X), so that the collection of all coverings becomes actually a set. (5.7) Denition. The Cech cohomology group H q (X, ) is the direct limit H q (X, ) = lim H q (, )
when runs over the collection of all open coverings of X. Explicitly, this means that the elements of H q (X, ) are the equivalence classes in the disjoint union of the groups q (, ), with an element in H q (, ) and another in H q (, ) identied if their H q images in H (, ) coincide for some renement of the coverings and . If : is a sheaf morphism, we have an obvious induced morphism : C (, ) C (, ), and therefore we nd a morphism
H q ( ) : H q (, ) H q (,
).
Let 0 of groups (5.8)
0 be an exact sequence of sheaves. We have an exact sequence 0 C q (, ) C q (, ) C q (, ),
but in general the last map is not surjective, because every section in (U0 ,...,q ) need not have a lifting in (U0 ,...,q ). The image of C (, ) in C (, ) will be denoted C (, ) and called the complex of liftable cochains of in . By construction, the sequence (5.9) 0 C q (, ) C q (, H q (, ) H q (, ) C q (, ) 0
is exact, thus we get a corresponding long exact sequence of cohomology (5.10) ) H q (, ) H q+1 (, ) . ) = C q ( ,
If is abby, Prop. 4.3 shows that we have C q (, H q (, ).
), hence H q (,
)=
208
(5.11) Proposition. Let a) is abby, or :
be a sheaf on X. Assume that either
b) X is paracompact and
is a sheaf of modules over a soft sheaf of rings 1 and every open covering
on X.
We
= (U)I of X. Proof. b) Let ( )I be a partition of unity in subordinate to (Prop. 4.17). dene a map hq : C q (, ) C q1 (, ) by
(5.12) (hq c)0 ...q1 =
I
Then H q (, ) = 0 for every q
c0 ...q1
where c0 ...q1 is extended by 0 on U0 ...q1 U . It is clear that ( q1 hq c)0 ...q =

I
c0 ...q ( q c)0 ...q , 1.
i.e. q1 hq + hq+1 q = Id. Hence q c = 0 implies q1 hq c = c if q
a) First we show that the result is true for the sheaf [0] . One can nd a family of sets L U such that (L ) is a partition of X. If is the characteristic function of L , Formula (5.12) makes sense for any cochain c C q (, [0] ) because [0] is a module over the ring Z[0] of germs of arbitrary functions X Z. Hence H q (, [0]) = 0 for q 1. We shall prove this property for all abby sheaves by induction on q. Consider the exact sequence 0 [0] 0 where =
[0]
/ . By the remark after (5.10), we have exact sequences

[0]
H q (,
(X)
(X) H 1 (, ) H 1 (, [0] ) = 0, ) H q+1 (, ) H q+1 (, [0]) = 0.
Then [0] (X) (X) is surjective by Prop. 4.3, thus H 1 (, ) = 0. By 4.3 again, q (, ) = 0 implies that H q+1 (, ) = 0. is abby; the induction hypothesis H 5.B. Lerays Theorem for Acyclic Coverings We rst show the existence of a natural morphism from Cech cohomology to ordinary cohomology. Let = (U )I be a covering of X. Select a map : X I such that x U(x) for every x X. To every cochain c C q (, ) we associate the section q c = f [q] (X) such that (5.13) f (x0 , . . . , xq ) = c(x0 )...(xq ) (xq )
xq
note that the right hand side is well dened as soon as x0 X, x1 U(x0 ) , . . . , xq U(x0 )...(xq1 ) .
5. Cech Cohomology
209
[q+1]
A comparison of (2.5) and (5.13) immediately shows that the section of ciated to q c is (1)j c
0 j q+1 (x0 )...(xj )...(xq+1 )
(X) asso-
(xq+1 ) = (dq f )(x0 , . . . , xq+1 ).

[]
In this way we get a morphism of complexes : C (, ) corresponding morphism (5.14) H q ( ) : H q (, ) H q (X, ).
(X). There is a
If = (V )J is a renement of get a commutative diagram
such that V U() and x V(x) for all x, , we

q
H ( ) H q (, ) H q (, ) H q ( ) H q ( )
H q (X, )
with = . In particular, (5.6) shows that the map H q ( ) in (5.14) does not depend on the choice of : if is another choice, then H q ( ) and H q ( ) can be both factorized through the group H q (, ) with Vx = U(x) U (x) and = IdX . By the universal property of direct limits, we get an induced morphism (5.15) Let 0 diagram H q (X, ) H q (X, ). 0 C (, ) C (, 0
[]
0 be an exact sequence of sheaves. There is a commutative ) C (,

[]
) 0 0
(X)
[]
(X)
(X)
where the vertical arrows are given by the morphisms . We obtain therefore a commutative diagram H q (, ) H q (, ) H q (, ) H q (X, ) H q+1 (, ) H q+1 (, ) ).
(5.16)
H q (X, ) H q (X,
) H q+1 (X, ) H q+1 (X,
(5.17) Theorem (Leray). Assume that H s (U0 ...t , ) = 0 for all indices 0 , . . . , t and s H q (X, ). 1. Then (5.14) gives an isomorphism H q (, )
We say that the covering is acyclic (with respect to ) if the hypothesis of Th. 5.17 is satised. Lerays theorem asserts that the cohomology groups of on X can be
210
computed by means of an arbitrary acyclic covering (if such a covering exists), without using the direct limit procedure. Proof. By induction on q, the result being obvious for q = 0. Consider the exact sequence 0 0 with = [0] and = [0] / . As is acyclic, the hypothesis and the long exact sequence of cohomology imply H s (U0 ...t , ) = 0 for s 1, on t 0. Moreover C (, ) = C (, ) thanks to Cor. 3.6. The induction hypothesis in degree q and diagram (5.16) give H q (, ) H q (, H q (X, ) H q+1 (, ) 0
(5.18) Remark. The morphism H 1 ( ) : H 1 (, ) H 1 (X, ) is always injective. Indeed, we have a commutative diagram H 0 (, ) H 0 (, = ) H 0 (X, ) H 1 (, ) 0
hence H q+1 (, ) H q+1 (X, ) is also an isomorphism.
) H q (X,
) H q+1 (X, ) 0,
H 0 (X,
where H 0 (, ) is the subspace of (X) = H 0 (X, ) consisting of sections which can over each U . As a consequence, the renement mappings be lifted in H 1 ( ) : H 1 (, ) H 1 (, ) are also injective. 5.C. Cech Cohomology on Paracompact Spaces We will prove here that Cech cohomology theory coincides with the ordinary one on paracompact spaces. (5.19) Proposition. Assume that X is paracompact. If 0 H q (X, ) H q (X, 0 ) H q+1 (X, )
) H 1 (X, ) 0,
is an exact sequence of sheaves, there is an exact sequence ) H q (X,
which is the direct limit of the exact sequences (5.10) over all coverings Proof. We have to show that the natural map lim H q (, ) lim H q (, )
is an isomorphism. This follows easily from the following lemma, which says essentially that every cochain in becomes liftable in after a renement of the covering.
5. Cech Cohomology
211
(5.20) Lifting lemma. Let = (U )I be an open covering of X and c C q (, ). If X is paracompact, there exists a ner covering = (V )J and a renement map : J I such that q c C q (, ). Proof. Since admits a locally nite renement, we may assume that itself is locally nite. There exists an open covering = (W )I of X such that W U . For every point x X, we can select an open neighborhood Vx of x with the following properties:
a) if x W , then Vx W ; b) if x U or if Vx W = , then Vx U ; c) if x U0 ...q , then c0 ...q C q (U0 ...q , ) admits a lifting in (Vx ). Indeed, a) (resp. c)) can be achieved because x belongs to only nitely many sets W have to be lifted in . b) can be (resp. U ), and so only nitely many sections of achieved because x has a neighborhood Vx that meets only nitely many sets U ; then we take Vx Vx U (Vx W ).
U x U x
= (Vx )xX is ner than , and denes a renement map. If Vx ...x

Vx0 W(xj ) Vx0 Vxj = for 0 j q,
Choose : X I such that x W(x) for every x. Then a) implies Vx W(x) , so = , we have 0 q
thus Vx0 U(x0 )...(xq ) by b). Now, c) implies that the section c(x0 )...(xq ) admits a lifting in (Vx0 ), and in particular in (Vx0 ...xq ). Therefore q c is liftable in . (5.21) Theorem. If X is a paracompact topological space, the canonical morphism H q (X, ) H q (X, ) is an isomorphism.
Proof. Argue by induction on q as in Lerays theorem, with the Cech cohomology exact sequence over replaced by its direct limit in (5.16).
In the next chapters, we will be concerned only by paracompact spaces, and most often in fact by manifolds that are either compact or countable at innity. In these cases, we will not distinguish H q (X, ) and H q (X, ). 5.D. Alternate Cech Cochains For explicit calculations, it is sometimes useful to consider a slightly modied Cech complex which has the advantage of producing much smaller cochain groups. If is a q q sheaf and = (U )I an open covering of X, we let AC (, ) C (, ) be the subgroup of alternate Cech cochains, consisting of Cech cochains c = (c0 ...q ) such that (5.22) c0 ...q = 0 c(0) ...(q) = () c0 ...q if i = j , i = j,
for any permutation of {1, . . . , q} of signature (). Then the Cech dierential (5.1) of an alternate cochain is still alternate, so AC (, ) is a subcomplex of C (, ). We are going to show that the inclusion induces an isomorphism in cohomology: (5.23) H q AC (, ) H q C (, ) = H q (, ).
212
Select a total ordering on the index set I. For each such ordering, we can dene a projection q : C q (, ) AC q (, ) C q (, ) by c alternate c such that c0 ...q = c0 ...q whenever 0 < . . . < q . As is a morphism of complexes, it is enough to verify that is homotopic to the identity on C (, ). For a given multi-index = (0 , . . . , q ), which may contain repeated indices, there is a unique permutation m(0), . . . , m(q) of (0, . . . , q) such that m(0) For p ... m(q) and m(l) < m(l + 1) whenever m(l) = m(l+1) .
q, we let (, p) be the sign of the permutation (0, . . . , q) m(0), . . . , m(p 1), 0, 1, . . . , m(0), . . . , m(p 1), . . . , q
if the elements m(0) , . . . , m(p) are all distinct, and (, p) = 0 otherwise. Finally, we set hq = 0 for q 0 and (hq c)0 ...q1 = (1)p (, p) c
0 p q1
m(0) ...m(p) 0 1 ...m(0) ...m(p1) ...q1
for q 1 ; observe that the index m(p) is repeated twice in the right hand side. A rather tedious calculation left to the reader shows that ( q1 hq c + hq+1 q c)0 ...q = c0 ...q (, q) cm(0) ...m(q) = (c q c)0 ...q . An interesting consequence of the isomorphism (5.23) is the following: (5.24) Proposition. Let be a sheaf on a paracompact space X. If X has arbitrarily ne open coverings or at least one acyclic open covering = (U ) such that more than n + 1 distinct sets U0 , . . . , Un have empty intersection, then H q (X, ) = 0 for q > n. Proof. In fact, we have AC q (, ) = 0 for q > n.
6. The De Rham-Weil Isomorphism Theorem

In 3 we dened cohomology groups by means of the simplicial abby resolution. We show here that any resolution by acyclic sheaves could have been used instead. Let ( , d) be a resolution of a sheaf . We assume in addition that all q are acyclic on X, i.e. H s (X, q ) = 0 for all q 0 and s 1. Set q = ker dq . Then 0 = and for every q 1 we get a short exact sequence 0
q1
q1
dq1
0.
Theorem 3.5 yields an exact sequence (6.1) H s (X, q1 ) H s (X,

dq1 q
) H s+1 (X,
s,q
q1
) H s+1 (X, q1 ) = 0.
213
If s
1, the rst group is also zero and we get an isomorphism s,q : H s (X,
q
For s = 0 we have H 0 (X, q1) = q1 (X) and H 0 (X, q ) = q (X) is the q-cocycle group of (X), so the connecting map 0,q gives an isomorphism Hq
) H s+1 (X,
q1
).
(X)
(X)/dq1 q1 (x) H 1 (X,

0,q
q1
).
The composite map q1,1 1,q1 0,q therefore denes an isomorphism (6.2) Hq
(X)
0,q
H 1 (X,
1,q1 q1
) H q (X,
q1,1
)=H q (X, ).
This isomorphism behaves functorially with respect to morphisms of resolutions. Our and every morphism of assertion means that for every sheaf morphism : resolutions : , there is a commutative diagram Hs (6.3) Hs If q = ker(dq : q commutative diagrams 0
q1
H s (X, ) s s H ( ) H ()
(X) (X)
q+1),
H s (X,
).
the functoriality comes from the fact that we have H s ( X, q ) H s+1 (X, q1 ) s q s+1 q1 H ( ) H ( ) H s (X, q ) H s+1 (X, q1 ).
s,q s,q
(6.4) De Rham-Weil isomorphism theorem. If ( , d) is a resolution of sheaves q which are acyclic on X, there is a functorial isomorphism Hq
0 q1 q1 q 0 ,
q1 q 0 , q1 q1 q
by
(X)
H q (X, ).
(6.5) Example: De Rham cohomology. Let X be a n-dimensional paracompact dierential manifold. Consider the resolution 0R
0
q+1
given by the exterior derivative d acting on germs of dierential q-forms (cf. Example 2.2). The De Rham cohomology groups of X are precisely (6.6) All sheaves isomorphism (6.7)
q q HDR (X, R) = H q
(X) .
are
X -modules,
so
is acyclic by Cor. 4.19. Therefore, we get an
q HDR (X, R) H q (X, R)
214
from the De Rham cohomology onto the cohomology with values in the constant sheaf R. Instead of using dierential forms, one can consider the resolution of R given by the exterior derivative d acting on currents: 0R
n
n1
nq
nq1
0.
The sheaves are also X -modules, hence acyclic. Thanks to (6.3), the inclusion q q nq induces an isomorphism (6.8) Hq
(X) H q
n (X)
both groups being isomorphic to H q (X, R). The isomorphism between cohomology of dierential forms and singular cohomology (another topological invariant) was rst established by [De Rham 1931]. The above proof follows essentially the method given by [Weil 1952], in a more abstract setting. As we will see, the isomorphism (6.7) can be put under a very explicit form in terms of Cech cohomology. We need a simple lemma. (6.9) Lemma. Let X be a paracompact dierentiable manifold. There are arbitrarily ne open coverings = (U ) such that all intersections U0 ...q are dieomorphic to convex sets. Proof. Select locally nite coverings j of X by open sets dieomorphic to j concentric euclidean balls in Rn . Let us denote by jk the transition dieomorphism from the coordinates in k to those in j . For any point a , the function x |x a|2 j computed in terms of the coordinates of j becomes |jk (x) jk (a)|2 on any patch k a. It is clear that these functions are strictly convex at a, thus there is a euclidean ball B(a, ) such that all functions are strictly convex on B(a, ) k (only j k a nite number of indices k is involved). Now, choose to be a (locally nite) covering of X by such balls U = B(a , ) with U . Then the intersection U0 ...q is () dened in k , k = (0 ), by the equations |jk (x) jk (am )|2 < 2 m where j = (m ), 0 chart (0 ) . m q. Hence the intersection is convex in the open coordinate
Let be an open subset of Rn which is starshaped with respect to the origin. Then the e De Rham complex R () is acyclic: indeed, Poincars lemma yields a homotopy q q q1 operator k : () () such that
1
k q fx (1 , . . . , q1 ) =
0 0
tq1 ftx (x, 1 , . . . , q1 ) dt,

0
x , j Rn ,
k f = f (0) R for f
().
q Hence HDR (, R) = 0 for q 1. Now, consider the resolution of the constant sheaf R on X, and apply the proof of the De Rham-Weil isomorphism theorem to Cech cohomology groups over a covering chosen as in Lemma 6.9. Since the intersections U0 ...s are convex, all Cech cochains in C s (, q ) are liftable in q1 by means of k q . Hence
215
qs
for all s = 1, . . . , q we have isomorphisms s,qs : H s (, for s 1 and we get a resulting isomorphism
) H s+1 (,
qs1
q q1,1 1,q1 0,q : HDR (X, R) H q (, R)
Let c in C s (, qs ) such that s c = 0. As c0 ...s is d-closed, we can write c = d(k c) where the cochain k qs c C s (, qs1) is dened as the family of sections k qs c0 ...s qs1 (U0 ...s ). Then d( s k qs c) = s (dk qs c) = s c = 0 and
qs
We are going to compute the connecting homomorphisms s,qs and their inverses explicitly.
s,qs {c} = { s k qs c} H s+1 (,
qs1
).
q The isomorphism HDR (X, R) H q (, R) is thus dened as follows: to the cohomology class {f } of a closed q-form f q (X), we associate the cocycle (c0 ) = (fU ) C 0 (, q ), then the cocycle
and by induction cocycles (cs 0 ...s ) C s (, (6.10) cs+1 s+1 = 0 ...

0 j s+1
c1 = k q c0 k q c0 C 1 (,
qs
q1
),
) given by on U0 ...s+1 .
(1)j k qs cs
0 ...j ...s+1
Conversely, let ( ) be a partition of unity subordinate to . Any Cech cocycle c C s+1 (, qs1 ) can be written c = s with C s (, qs1 ) given by 0 ...s =
I
The image of {f } in H q (, R) is the class of the q-cocycle (cq 0 ...q ) in C q (, R).
c0 ...s ,
(cf. Prop. 5.11 b)), thus {c } = ( s,qs)1 {c} can be represented by the cochain c = d C s (, qs ) such that c 0 ...s =
I
d c0 ...s = (1)qs1
c0 ...s d .
For a reason that will become apparent later, we shall in fact modify the sign of our isomorphism s,qs by the factor (1)qs1 . Starting from a class {c} H q (, R), we q 0 ) such that obtain inductively {b} H (, (6.11) b 0 =
0 ,...,q1
c0 ...q1 0 d0 . . . dq1
on U0 ,
q corresponding to {f } HDR (X, R) given by the explicit formula
(6.12)
f=
q
q b q =
0 ,...,q
c0 ...q q d0 . . . dq1 .
216
The choice of sign corresponds to (6.2) multiplied by (1)q(q1)/2. (6.13) Example: Dolbeault cohomology groups. Let X be a C-analytic manifold of dimension n, and let p,q be the sheaf of germs of dierential forms of type (p, q) with complex values. For every p = 0, 1, . . . , n, the Dolbeault-Grothendieck Lemma I-2.9 shows that ( p, , d ) is a resolution of the sheaf p of germs of holomorphic forms of X degree p on X. The Dolbeault cohomology groups of X already considered in chapter 1 can be dened by (6.14) H p,q (X, C) = H q
p,
(X) .
The sheaves p,q are acyclic, so we get the Dolbeault isomorphism theorem, originally proved in [Dolbeault 1953], which relates d -cohomology and sheaf cohomology: (6.15) H p,q (X, C) H q (X, p ). X H 0,q (X, C) H q (X, X ).
The case p = 0 is especially interesting: (6.16)
As in the case of De Rham cohomology, there is an inclusion p,q np,nq and the p complex of currents ( np,n , d ) denes also a resolution of X . Hence there is an isomorphism: (6.17) H p,q (X, C) = H q
p,
(X) H q
np,n (X)
7. Cohomology with Supports

As its name indicates, cohomology with supports deals with sections of sheaves having supports in prescribed closed sets. We rst introduce what is an admissible family of supports. (7.1) Denition. A family of supports on a topological space X is a collection of closed subsets of X with the following two properties: a) If F , F , then F F ; b) If F and F F is closed, then F . (7.2) Example. Let S be an arbitrary subset of X. Then the family of all closed subsets of X contained in S is a family of supports. (7.3) Example. The collection of all compact (non necessarily Hausdor) subsets of X is a family of supports, which will be denoted simply c in the sequel. and any family of supports on X, (7.4) Denition. For any sheaf denote the set of all sections f (X) such that Supp f .
(X)
will
7. Cohomology with Supports
217
It is clear that (X) is a subgroup of groups with arbitrary supports. (7.5) Denition. The cohomology groups of
(X). We can now introduce cohomology with supports in are

[] (X)
q H (X, ) = H q
q The cohomology groups with compact supports will be denoted Hc (X, ) and the cohoq mology groups with supports in a subset S will be denoted HS (X, ). 0 In particular H (X, ) = (X). If 0 there are corresponding exact sequences
[q]
0 is an exact sequence,
(7.6) 0 When (7.7)
[q] (X) q H (X,
(X) (X) q q q+1 ) H (X, ) H (X, ) H (X, ) .
[q]
is abby, there is an exact sequence 0

(X)
(X)
(X)
and every g (X) can be lifted to v (X) without enlarging the support: apply the proof of Prop. 4.3 to a maximal lifting which extends w = 0 on W = (Supp g). It q is -acyclic, i.e. H (X, ) = 0 for all q 1. Similarly, follows that a abby sheaf assume that X is paracompact and that is soft, and suppose that has the following additional property: every set F has a neighborhood G . An adaptation of the proofs of Prop. 4.3 and 4.13 shows that (7.7) is again exact. Therefore every soft sheaf is also -acyclic in that case. As a consequence of (7.6), any resolution canonical De Rham-Weil isomorphism (7.8) Hq
of
by -acyclic sheaves provides a
(X)
q H (X, ).
(7.9) Example: De Rham cohomology with compact support. In the special case of the De Rham resolution R on a paracompact manifold, we get an isomorphism (7.10)
q HDR,c (X, R) := H q ( q (X) Hc (X, R),
where q (X) is the space of smooth dierential q-forms with compact support in X. These groups are called the De Rham cohomology groups of X with compact support. When X is oriented, dim X = n, we can also consider the complex of compactly supported currents: 0
n (X)
n1 (X)
nq (X)
nq1 (X)
Note that (X) and (X) are respectively the subgroups of compactly supported n sections in and , both of which are acyclic resolutions of R. Therefore the n inclusion (X) (X) induces an isomorphism n (7.11) Hq
(X) H q
n (X)
218
q both groups being isomorphic to Hc (X, R).
Now, we concentrate our attention on cohomology groups with compact support. We assume until the end of this section that X is a locally compact space. (7.12) Proposition. There is an isomorphism
q Hc (X, ) = lim
H q (U ,
U)
U X
where
is the sheaf of sections of
vanishing on X
U (cf. 3).
[]
Proof. By denition
q Hc (X, ) = H q [] c (X)
= lim
Hq (
)U (U )
U X
since sections of ( [] )U (U ) can be extended by 0 on X U . However, ( [] )U is a resolution of U and ( [q] )U is a Z[q] -module, so it is acyclic on U . The De Rham-Weil isomorphism theorem implies Hq (
[]
)U (U) = H q (U ,
U) [q]
and the proposition follows. The reader should take care of the fact that ( not coincide in general with ( U )[q] .
)U does
The cohomology groups with compact support can also be dened by means of Cech cohomology. (7.13) Denition. Assume that X is a separable locally compact space. If = (U ) is q a locally nite covering of X by relatively compact open subsets, we let Cc (, ) be the subgroups of cochains such that only nitely many coecients c0 ...q are non zero. The Cech cohomology groups with compact support are dened by
q Hc (, ) = H q Cc (, ) q Hc (X, ) = lim H q Cc (, )
For such coverings (7.14)
, Formula (5.13) yields canonical morphisms q q H q ( ) : Hc (, ) Hc (X, ).

q q Hc (X, ) Hc (X, ).
Now, the lifting Lemma 5.20 is valid for cochains with compact supports, and the same proof as the one given in 5 gives an isomorphism (7.15)
8. Cup Product
219
8. Cup Product
sheaves of Let be a sheaf of commutative rings and , X. We denote by the sheaf on X dened by (8.1) ( )x =
x
-modules on a space
x,
with the weakest topology such that the range of any section given by is open in for any open set U X. Given f [p+q] is dened by f g ( )x (8.2) f
[p] x
and g
[q] x ,
(U ) (U)
(U )
the cup product
g(x0 , . . . , xp+q ) = f (x0 , . . . , xp )(xp+q ) g(xp , . . . , xp+q ).
A simple computation shows that (8.3) In particular, f (f + dp1 f ) dp+q (f g) = (dp f ) g + (1)p f (dq g).
g is a cocycle if f, g are cocycles, and we have (g + dq1 g ) = f g + dp+q1 f g + (1)p f g + f dg .
Consequently, there is a well dened (8.4)
(X)-bilinear morphism
) H p+q (X, ) g}.
H p (X, ) H q (X,
0 be an exact sequence of sheaves. Assume that the Let 0 sequence obtained after taking the tensor product by is also exact: 0
which maps a pair ({f }, {g}) to {f
0.
Then we obtain connecting homomorphisms q : H q (X, q : H q (X,
) H q+1 (X,

), ).
) H q+1 (X,
For every H p (X, ), H q (X, (8.5) (8.5 )
) we have
p+q ( ) = (1)p ( q ), p+q ( ) = ( q ) ,
where the second line corresponds to the tensor product of the exact sequence by on the right side. These formulas are deduced from (8.3) applied to a representant f [p] (X) of and to a lifting g [q] (X) of a representative g of (note that dp f = 0). (8.6) Associativity and anticommutativity. Let i : q p canonical isomorphism s t t s. For all H (X, ), H (X, = (1)pq i( ). be the ) we have
220
If
is another sheaf of
-modules and H r (X,

( ) = (
), then ).
Proof. The associativity property is easily seen to hold already for all cochains (f g) h=f (g h), f
[p] x ,
[q] x ,
[r] x .
The commutation property is obvious for p = q = 0, and can be proved in general by induction on p + q. Assume for example q 1. Consider the exact sequence 0
where = [0] and = [0] / . This exact sequence splits on each stalk (but [0] not globally, nor even locally): a left inverse x x of the inclusion is given by g g(x). Hence the sequence remains exact after taking the tensor product with . Now, as is acyclic, the connecting homomorphism H q1 (X, ) H q (X, ) is onto, so there is H q1 (X, ) such that = q1 . Using (8.5 ), (8.5) and the induction hypothesis, we nd = p+q1 ( ) = p+q1 (1)p(q1) i( ) ).
= (1)p(q1) i p+q1 (
) = (1)p(q1) (1)p i(
Theorem 8.6 shows in particular that H (X, ) is a graded associative and supercommutative algebra, i.e. = (1)pq for all classes H p (X, ), H q (X, ). If is a -module, then H (X, ) is a graded H (X, )-module. (8.7) Remark. The cup product can also be dened for Cech cochains. Given c q p+q p (, ) is dened by C (, ) and c C (, ), the cochain c c C (c c )0 ...p+q = c0 ...p c p ...p+q p+q (c c ) = ( p c) on U0 ...p+q .
Straightforward calculations show that c + (1)p c ( q c )
and that there is a commutative diagram H p (, )H q (, ) H p+q ( , ) H p+q (X, ) ),
H p (X, )H q (X,
where the vertical arrows are the canonical morphisms H s ( ) of (5.14). Note that the commutativity already holds in fact on cochains. (8.8) Remark. Let and be families of supports on X. Then is again a family of supports, and Formula (8.2) denes a bilinear map (8.9)
p q H (X, ) H (X, p+q ) H (X,
9. Inverse Images and Cartesian Products
221
on cohomology groups with supports. Supp(f g) Supp f Supp g.
This follows immediately from the fact that
(8.10) Remark. Assume that X is a dierentiable manifold. Then the cohomology complex HDR (X, R) has a natural structure of supercommutative algebra given by the wedge product of dierential forms. We shall prove the following compatibility statement:
q Let H q (X, R) HDR (X, R) be the De Rham-Weil isomorphism given by Formula (6.12). Then the cup product c c is mapped on the wedge product f f of the corresponding De Rham cohomology classes.
By remark 8.7, we may suppose that c , c are Cech cohomology classes of respective degrees p, q. Formulas (6.11) and (6.12) give
fUp = 0 ,...,p1
c 0 ...p1 p d0 . . . dp1 , cp ...p+q p+q dp . . . dp+q1 .
f =
p ,...,p+q
We get therefore f f = c 0 ...p cp ...p+q p+q d0 . . . p+q1 , c in the De Rham cohomology.
0 ,...,p+q
which is precisely the image of c
9. Inverse Images and Cartesian Products

9.A. Inverse Image of a Sheaf Let F : X Y be a continuous map between topological spaces X, Y , and let : Y be a sheaf of abelian groups. Recall that inverse image F 1 is dened as the sheaf-space F 1 = Y X = (s, x) ; (s) = F (x) with projection = pr2 : F 1 (9.1) X. The stalks of F 1 (F 1 )x =
F (x) ,
are given by
and the sections F 1 (U ) can be considered as continuous mappings : U such that = F . In particular, any section s (U ) has a pull-back (9.2) For any v (9.3)
[q] y ,
F s = s F F 1 we dene F v (F 1 )x by
[q]
F 1 (U ) .
F v(x0 , . . . , xq ) = v F (x0 ), . . . , F (xq ) (F 1 )xq =
F (xq )
222
for x0 V (x), x1 V (x0 ), . . . , xq V (x0 , . . . , xq1 ). We get in this way a morphism of complexes F : [] (Y ) (F 1 )[] (X). On cohomology groups, we thus have an induced morphism (9.4) F : H q (Y, ) H q (X, F 1 ).
Let 0 0 be an exact sequence of sheaves on X. Thanks to property (9.1), there is an exact sequence 0 F 1 F 1 F 1 0.
It is clear on the denitions that the morphism F in (9.4) commutes with the associated cohomology exact sequences. Also, F preserves the cup product, i.e. F ( ) = F F whenever , are cohomology classes with values in sheaves , on X. Furthermore, if G : Y Z is a continuous map, we have (9.5) (9.6) Remark. Similar denitions can be given for Cech cohomology. If = (U )I 1 1 is an open covering of Y , then F = F (U ) I is an open covering of X. For c C q (, ), we set (F c)0 ...q = c0 ...q F C q (F 1 , F 1 ). This denition is obviously compatible with the morphism from Cech cohomology to ordinary cohomology. (9.7) Remark. Let be a family of supports on Y . We dene F 1 to be the family of closed sets K X such that F (K) is contained in some set L . Then (9.4) can be generalized in the form (9.8)
q q F : H (Y, ) HF 1 (X, F 1 ).
(G F ) = F G .
(9.9) Remark. Assume that X and Y are paracompact dierentiable manifolds and that F : X Y is a map. If ( )I is a partition of unity subordinate to , then ( F )I is a partition of unity on X subordinate to F 1 . Let c C q (, R). The dierential form associated to F c in the De Rham cohomology is g=
0 ,...,q
c0 ...q (q F )d(0 F ) . . . d(q1 F ) c0 ...q q d0 . . . dq1 .
= F
0 ,...,q
Hence we have a commutative diagram

q HDR (Y, R) H q (Y, R) H q (Y, R) F F F q q q HDR (X, R) H (X, R) H (X, R).
10. Spectral Sequence of a Filtered Complex
223
9.B. Cohomology Groups of a Subspace Let be a sheaf on a topological space X, let S be a subspace of X and iS : S X 1 1 q q is the restriction of to S, so that H (S, ) = H (S, iS ) by the inclusion. Then iS denition. For any two subspaces S S, the inclusion of S in S induces a restriction morphism H q (S, ) H q (S , ). (9.10) Theorem. Let be a sheaf on X and S a strongly paracompact subspace in X. When ranges over open neighborhoods of S, we have H q (S, ) = lim H q (, ).
S
Proof. When q = 0, the property is equivalent to Prop. 4.7. The general case follows by induction on q if we use the long cohomology exact sequences associated to the short exact sequence 0 [0] [0] / 0 on S and on its neighborhoods (note that the restriction of a abby sheaf to S is soft by Prop. 4.7 and the fact that every closed subspace of a strongly paracompact subspace is strongly paracompact). 9.C. Cartesian Product We use here the formalism of inverse images to deduce the cartesian product from X, Y sheaves of the cup product. Let R be a xed commutative ring and R-modules. We dene the external tensor product by (9.11) R = pr1 1 R pr1 2 R
where pr1 , pr2 are the projections of X Y onto X, Y respectively. The sheaf is thus the sheaf on X Y whose stalks are (9.12) For all cohomology classes H p (X, ), H q (Y, H p+q (X Y, R ) is dened by (9.13) = (pr ) 1 ( R )(x,y) =
x
y.
) the cartesian product
(pr ). 2
More generally, let and be families of supports in X and Y respectively. If denotes the family of all closed subsets of X Y contained in products K L of elements K , L , the cartesian product denes a R-bilinear map (9.14) If X, Y are sheaves of abelian groups and if , are cohomology classes of degree p , q with values in , , one gets easily (9.15) Furthermore, if F : X X and G : Y Y are continuous maps, then (9.16) (F G) ( ) = (F ) (G ). ( ) ( ) = (1)qp (
p q H (X, ) H (Y,
p+q ) H (X Y,
).
) (
).
224

10.A. Construction of the Spectral Sequence The theory of spectral sequences consists essentially in computing the homology groups of a dierential module (K, d) by successive approximations, once a ltration Fp (K) is given in K and the cohomology groups of the graded modules Gp (K) are known. Let us rst recall some standard denitions and notations concerning ltrations. (10.1) Denition. Let R be a commutative ring. A ltration of a R-module M is a sequence of submodules Mp M , p Z, also denoted Mp = Fp (M ), such that Mp+1 Mp for all p Z, Mp = M and Mp = {0}. The associated graded module is G(M ) =
pZ
Gp (M ),
Gp (M ) = Mp /Mp+1 .
Let (K, d) be a dierential module equipped with a ltration (Kp ) by dierential submodules (i.e. dKp Kp for every p). For any number r N {}, we dene p p Zr , Br Gp (K) = Kp /Kp+1 by (10.2) (10.2 )
p Zr = Kp d1 Kp+r mod Kp+1 , p Br = Kp dKpr+1 mod Kp+1 , p Z = Kp d1 {0} mod Kp+1 , p B = Kp dK mod Kp+1 .
(10.3) Lemma. For every p and r, there are inclusions

p p p p p p . . . Br Br+1 . . . B Z . . . Zr+1 Zr . . .
and the dierential d induces an isomorphism

p p+r p p+r d : Zr /Zr+1 Br+1 /Br . p p Proof. It is clear that (Zr ) decreases with r, that (Br ) increases with r, and that p p B Z . By denition p Br = (Kp dKpr+1 )/(Kp+1 dKpr+1 ). p Zr = (Kp d1 Kp+r )/(Kp+1 d1 Kp+r ),
The dierential d induces a morphism

p Zr (dKp Kp+r )/(dKp+1 Kp+r ) p whose kernel is (Kp d1 {0})/(Kp+1 d1 {0}) = Z , whence isomorphisms p p d : Zr /Z (Kp+r dKp )/(Kp+r dKp+1 ),
p p d : Zr /Zr+1 (Kp+r dKp )/(Kp+r dKp+1 + Kp+r+1 dKp ).
p+r p+r The right hand side of the last arrow can be identied to Br+1 /Br , for p+r Br = (Kp+r dKp+1 )/(Kp+r+1 dKp+1 ), p+r Br+1 = (Kp+r dKp )/(Kp+r+1 dKp ).
225
p Now, for each r N, we dene a complex Er = pZ Er with a dierential dr : p p+r p p p Er Er of degree r as follows: we set Er = Zr /Br and take
(10.4)
d p+r p+r p+r p+r p p dr : Zr /Br Zr /Zr+1 Br+1 /Br Zr /Br p p
where the rst arrow is the obvious projection and the third arrow the obvious inclusion. Since dr is induced by d, we actually have dr dr = 0 ; this can also be seen directly by p+r p+r the fact that Br+1 Zr+1 .
(10.5) Theorem and denition. There is a canonical isomorphism Er+1 H (Er ). The sequence of dierential complexes (Er , d ) is called the spectral sequence of the ltered r dierential module (K, d).
Proof. Since d is an isomorphism in (10.4), we have

p p ker dr = Zr+1 /Br , p+r p+r Im dr = Br+1 /Br .
p pr p p Hence the image of dr : Er Er is Br+1 /Br and p p p p p p p H p (Er ) = (Zr+1 /Br )/(Br+1 /Br ) Zr+1 /Br+1 = Er+1 .
(10.6) Theorem. Consider the ltration of the homology module H(K) dened by Fp H(K) = Im H(Kp ) H(K) . Then there is a canonical isomorphism
p E = Gp H(K) .
Proof. Clearly Fp H(K) = (Kp d1 {0})/(Kp dK), whereas

p p p E = Z /B = (Kp d1 {0})/(Kp+1 d1 {0} + Kp dK). p It follows that E Fp H(K) /Fp+1 H(K) . p p Z = (Kp d1 {0})/(Kp+1 d1 {0}), B = (Kp dK)/(Kp+1 dK),
In most applications, the dierential module K has a natural grading compatible with the ltration. Let us consider for example the case of a cohomology complex K = l lZ K . The ltration Kp = Fp (K ) is said to be compatible with the dierential complex structure if each Kp is a subcomplex of K , i.e.
Kp = lZ l p,q p,q p,q where (Kp ) is a ltration of K l . Then we dene Zr , Br , Er to be the sets of elements p p p of Zr , Br , Er of total degree p + q. Therefore l Kp
226

p+q+1 p,q p+q Zr = Kp d1 Kp+r p+q mod Kp+1 , p Zr = p Br = p Er = p,q Zr , p,q Br , p,q Er ,
(10.7) (10.7 ) (10.7 )
p+q1 p,q p+q Br = Kp dKpr+1 p,q p,q p,q Er = Zr /Br ,
p+q mod Kp+1 ,
and the dierential dr has bidegree (r, r + 1), i.e. (10.8)

p,q p+r dr : Er Er , qr+1 .
For an element of pure bidegree (p, q), p is called the ltering degree, q the complementary degree and p + q the total degree.
(10.9) Denition. A ltration (Kp ) of a complex K is said to be regular if for each l l degree l there are indices (l) N (l) such that Kp = K l for p < (l) and Kp = 0 for p > N (l).
If the ltration is regular, then (10.7) and (10.7 ) show that

p,q p,q p,q Zr = Zr+1 = . . . = Z p,q Br+1
p,q Br
= ... =
p,q B
for r > p + 1 (p + q 1),
for r > N (p + q + 1) p,
p,q p,q therefore Er = E for r r0 (p, q). We say that the spectral sequence converges to its limit term, and we write symbolically
(10.10)
p,q Er = H p+q (K )
to express the following facts: there is a spectral sequence whose terms of the r-th p,q p,q p,lp generation are Er , the sequence converges to a limit term E , and E is the term l l Gp H (K ) in the graded module associated to some ltration of H (K ).
p,q (10.11) Denition. The spectral sequence is said to collapse in Er if all terms Zk , p,q p,q Bk , Ek are constant for k r, or equivalently if dk = 0 in all bidegrees for k r.
(10.12) Special case. Assume that there exists an integer r 2 and an index q0 such p,q that Er = 0 for q = q0 . Then this property remains true for larger values of r, and we must have dr = 0. It follows that the spectral sequence collapses in Er and that
lq H l (K ) = Er 0 ,q0 . p,q Similarly, if Er = 0 for p = p0 and some r
1 then
p H l (K ) = Er 0 ,lp0 .
10.B. Computation of the First Terms

p p Consider an arbitrary spectral sequence. For r = 0, we have Z0 = Kp /Kp+1 , B0 = {0}, thus
(10.13)
p E0 = Kp /Kp+1 = Gp (K).
11. Spectral Sequence of a Double Complex
227
The dierential d0 is induced by d on the quotients, and (10.14)

p E1 = H Gp (K) .
Now, there is a short exact sequence of dierential modules 0 Gp+1 (K) Kp /Kp+2 Gp (K) 0. We get therefore a connecting homomorphism (10.15)
p E1 = H Gp (K) p+1 H Gp+1 (K) = E1 .
We claim that coincides with the dierential d1 : indeed, both are induced by d. When K is a ltered cohomology complex, d1 is the connecting homomorphism (10.16)
p,q E1 = H p+q Gp (K ) p+1,q H p+q+1 Gp+1 (K ) = E1 .
11. Spectral Sequence of a Double Complex

A double complex is a bigraded module K , = d = d + d such that (11.1) d : K p,q K p+1,q , K p,q together with a dierential
d : K p,q+1 K p,q+1 ,
and d d = 0. In particular, d and d satisfy the relations (11.2) d2 = d2 = 0, d d + d d = 0.
The simple complex associated to K , is dened by Kl =

p+q=l
K p,q
together with the dierential d. We will suppose here that both graduations of K , are positive, i.e. K p,q = 0 for p < 0 or q < 0. The rst ltration of K is dened by (11.3)
l Kp = i+j=l, i p
K i,j =
p i l
K i,li .
The graded module associated to this ltration is of course Gp (K l ) = K p,lp , and the l+1 l dierential induced by d on the quotient coincides with d because d takes Kp to Kp+1 . Thus we have a spectral sequence beginning by (11.4)
p,q E0 = K p,q ,
d0 = d ,
p,q q E1 = Hd (K p, ).
By (10.16), d1 is the connecting homomorphism associated to the short exact sequence 0 K p+1, K p, K p+1, K p, 0
228
where the dierential is given by (d mod K p+2, ) for the central term and by d for the two others. The denition of the connecting homomorphism in the proof of Th. 1.5 shows that q q d1 = : Hd (K p, ) Hd (K p+1, ) is induced by d . Consequently, we nd (11.5)
p,q p ,q p q E2 = Hd (E1 ) = Hd Hd (K , ) .
For such a spectral sequence, several interesting additional features can be pointed out. For all r and l, there is an injective homomorphism
0,l Er+1 Er 0,l 0,l whose image can be identied with the set of dr -cocycles in Er ; the coboundary group p,q l,0 is zero because Er = 0 for q < 0. Similarly, Er is equal to its cocycle submodule, and there is a surjective homomorphism l,0 l,0 lr,r1 Er Er+1 Er /dr Er l,0 .
Furthermore, the ltration on H l (K ) begins at p = 0 and stops at p = l, i.e. (11.6) F0 H l (K ) = H l (K ), Fp H l (K ) = 0 for p > l.
Therefore, there are canonical maps (11.7)

0,l 0,l H l (K ) G0 H l (K ) = E Er ,
l,0 Er E = Gl H l (K ) H l (K ). l,0
These maps are called the edge homomorphisms of the spectral sequence. (11.8) Theorem. There is an exact sequence
2,0 1,0 0,1 2 0 E2 H 1 (K ) E2 E2 H 2 (K ) d
where the non indicated arrows are edge homomorphisms. Proof. By 11.6, the graded module associated to H 1 (K ) has only two components, and we have an exact sequence
1,0 0,1 0 E H 1 (K ) E 0. 1,0 1,0 1,0 1,0 However E = E2 because all dierentials dr starting from Er or abuting to Er 0,1 2,0 0,1 2,0 must be zero for r 2. Similarly, E = E3 and E = E3 , thus there is an exact sequence 0,1 d2 2,0 0,1 2,0 0 E E2 E2 E 0.
A combination of the two above exact sequences yields

2 2,0 1,0 0,1 2,0 0 E2 H 1 (K ) E2 E2 E 0.
12. Hypercohomology Groups
229
2,0 Taking into account the injection E H 2 (K ) in (11.7), we get the required exact sequence.
(11.9) Example. Let X be a complex manifold of dimension n. Consider the double complex K p,q = (X, C) together with the exterior derivative d = d + d . Then there p,q is a spectral sequence which starts from the Dolbeault cohomology groups
p,q E1 = H p,q (X, C)
and which converges to the graded module associated to a ltration of the De Rham cohomology groups: p+q p,q Er = HDR (X, C). This spectral sequence is called the Hodge-Frlicher spectral sequence [Frlicher 1955]. o o We will study it in much more detail in chapter 6 when X is compact. Frequently, the spectral sequence under consideration can be obtained from two distinct double complexes and one needs to compare the nal cohomology groups. The following lemma can often be applied. (11.10) Lemma. Let K p,q Lp,q be a morphism of double complexes (i.e. a double sequence of maps commuting with d and d ). Then there are induced morphisms
, K Er , L Er ,
of the associated spectral sequences. If one of these morphisms is an isomorphism for some r, then H l (K ) H l (L ) is an isomorphism. Proof. If the r-terms are isomorphic, they have the same cohomology groups, thus , , , , K Er+1 L Er+1 and K E L E in the limit. The lemma follows from the fact that if a morphism of graded modules : M M induces an isomorphism G (M ) G (M ), then is an isomorphism.
12. Hypercohomology Groups

Let ( , ) be a complex of sheaves
0
on a topological space X. We denote by q = q ( ) the q-th sheaf of cohomology of this complex; thus q is a sheaf of abelian groups over X. Our goal is to dene generalized cohomology groups attached to on X, in such a way that these groups only depend on the cohomology sheaves q . For this, we associate to the double complex of groups (12.1)
p,q K = (q )[p] (X)
0 0 1 q
q
with dierential d = dp given by (2.5), and with d = (1)p ( q )[p] . As ( q )[] : (q )[] (q+1 )[] is a morphism of complexes, we get the expected relation d d + d d = 0.
230
(12.2) Denition. The groups H q (K ) are called the hypercohomology groups of and are denoted Hq (X, ).
Clearly H0 (X, ) = 0 (X) where 0 = ker 0 is the rst cohomology sheaf of . If : is a morphism of sheaf complexes, there is an associated morphism of , , double complexes , : K K , hence a natural morphism (12.3) Hq ( ) : Hq (X, ) Hq (X, ). We rst list a few immediate properties of hypercohomology groups, whose proofs are left to the reader. (12.4) Proposition. The following properties hold: a) If b) c)
q = 0 for q = 0, then Hq (X, ) = H q (X, 0). If [s] denotes the complex shifted of s indices to the right, i.e. [s]q = qs , then Hq (X, [s]) = Hqs (X, ). If 0 0 is an exact sequence of sheaf complexes, there is a
long exact sequence Hq (X, ) Hq (X, ) Hq (X, ) Hq+1 (X, ) .
If is a sheaf complex, the spectral sequence associated to the rst ltration of K is given by p,q q p, E1 = Hd (K ) = H q ( )[p] (X) .
However by (2.9) the functor get (12.5) (12.5 )
p,q E1 =
[p]
(X) preserves exact sequences. Therefore, we
p,q p ,q since E2 = Hd (E1 ). If 11.10 to the E2 -term of the associated rst spectral sequences of K and K yields:
q () [p] (X), p,q E2 = H p X, q ( ) , : is a morphism, an application of Lemma , ,
(12.6) Corollary. If : is a quasi-isomorphism this means that induces an isomorphism ( ) ( ) , then Hl ( ) : Hl (X, ) Hl (X, ) is an isomorphism. Now, we may reverse the roles of the indices p, q and of the dierentials d , d . The lj,j l second ltration Fp (K ) = j p K is associated to a spectral sequence such that q p,q q ,p p [] E 1 = Hd (K ) = Hd ( ) (X) , hence (12.7) (12.7 )
p,q E1 = H q (X, p ), p,q p E2 = H H q (X, ) .
13. Direct Images and the Leray Spectral Sequence
231
These two spectral sequences converge to limit terms which are the graded modules associated to ltrations of H (X, ) ; these ltrations are in general dierent. Let us mention two interesting special cases. Assume rst that the complex q = 0 for q 1. Then we nd
is a resolution of a sheaf
p,q E2 = 0 for q
, so that 1.
0 =
and
p,0 E2 = H p (X, ),
It follows that the rst spectral sequence collapses in E2 , and 10.12 implies
(12.8) Now, assume that the sheaves (12.9)
by acyclic sheaves, If both conditions hold, i.e. if is a resolution of a sheaf then (12.8) and (12.9) can be combined to obtain a new proof of the De Rham-Weil isomorphism H l (X, ) H l (X) .
q are acyclic. The second spectral sequence gives p,0 p,q E2 = H p (X) , E2 = 0 for q 1, Hl (X, ) H l (X) .
Hl (X, ) H l (X, ).

13.A. Direct Images of a Sheaf Let X, Y be topological spaces, F : X Y a continuous map and a sheaf of abelian groups on X. Recall that the direct image F is the presheaf on Y dened for any open set U Y by (13.1) (F )(U ) = F 1 (U ) . is in fact a sheaf. The Axioms (II-2.4 and (II-2.4 ) are clearly satised, thus F following result is obvious: (13.2) Every sheaf morphism : is abby = F is abby.
induces a corresponding morphism F : F F ,
so F is a functor on the category of sheaves on X to the category of sheaves on Y . This functor is exact on the left: indeed, to every exact sequence 0 is associated an exact sequence 0 F F F ,
but F F need not be onto if is. All this follows immediately from the considerations of 3. In particular, the simplicial abby resolution ( [] , d) yields a complex of sheaves (13.3) 0 F
[0]
[1]
q [q] F d
[q+1]
232
(13.4) Denition. The q-th direct image of sheaf complex (13.3). It is denoted Rq F
by F is the q-th cohomology sheaf of the

[]
= q (F
).
As F is exact on the left, the sequence 0 F (13.5) R0 F = F .
[0]
[1]
is exact, thus
We now compute the stalks of Rq F . As the kernel or cokernel of a sheaf morphism is obtained stalk by stalk, we have (Rq F )y = H q (F
[]
)y = lim H q F
Uy
[]
(U ) .
The very denition of F and of sheaf cohomology groups implies H q F hence we nd (13.6) (Rq F )y = lim H q F 1 (U ),
Uy []
(U ) = H q
[]
(F 1 (U )) = H q F 1 (U ),
i.e. Rq F is the sheaf associated to the presheaf U H q F 1 (U ), . We must stress need not be true in general. here that the stronger relation Rq F (U ) = H q F 1 (U ), If the ber F 1 (y) is strongly paracompact in X and if the family of open sets F 1 (U ) is a fundamental family of neighborhoods of F 1 (y) (this situation occurs for example if X and Y are locally compact spaces and F a proper map, or if F = pr1 : X = Y S Y where S is compact), Th. 9.10 implies the more natural relation (13.6 ) (Rq F )y = H q F 1 (y), .
0 be an exact sequence of sheaves on X. Apply the long Let 0 exact sequence of cohomology on every open set F 1 (U ) and take the direct limit over U . We get an exact sequence of sheaves: (13.7) 0 F Rq F F Rq F F Rq F R1 F Rq+1 F .
13.B. Leray Spectral Sequence For any continuous map F : X Y , the Leray spectral sequence relates the cohomology groups of a sheaf on X and those of its direct images Rq F on Y . Consider the two spectral sequences Er , Er associated with the complex of sheaves = F [] on Y , as in 12. By denition we have q ( ) = Rq F . By (12.5 ) the second term of the rst spectral sequence is
p,q E2 = H p (Y, Rq F ),
233
and this spectral sequence converges to the graded module associated to a ltration of Hl (Y, F [] ). On the other hand, (13.2) implies that F [q] is abby. Hence, the second special case (12.9) can be applied: Hl (Y, F
[]
) H l F
[]
(Y ) = H l
[]
(X) = H l (X, ).
We may conclude this discussion by the following (13.8) Theorem. For any continuous map F : X Y and any sheaf groups on X, there exists a spectral sequence whose E2 term is
p,q E2 = H p (Y, Rq F ), p,lp which converges to a limit term E equal to the graded module associated with a l ltration of H (X, ). The edge homomorphism
of abelian
l,0 H l (Y, F ) E H l (X, ) coincides with the composite morphism F # : H l (Y, F ) H l (X, F 1 F ) H l (X, ) where F : F 1 F is the canonical sheaf morphism.
F H l (F )
Proof. Only the last statement remains to be proved. The morphism F is dened as follows: every element s (F 1 F )x = (F )F (x) is the germ of a section s F (V ) = F 1 (V ) on a neighborhood V of F (x). Then F 1 (V ) is a neighborhood of x and we let F s be the germ of s at x. Now, we observe that to any commutative diagram of topological spaces and continuous maps is associated a commutative diagram involving the direct image sheaves and their cohomology groups: X Y
F
X Y
H l (X , G ) H l (Y , F G ).
H l (X, ) H l (Y, F ) H # G#
F #
F#
There is a similar commutative diagram in which F # and F # are replaced by the edge homomorphisms of the spectral sequences of F and F : indeed there is a natural on X , so we get a morphism of sheaf morphism H 1 F F G1 for any sheaf complexes
H 1 F (G )[] F G1 (G )[] F (G1 G )[] F []
hence also a morphism of the spectral sequences associated to both ends. The special case X = Y = Y , G = F , F = H = IdY then shows that our statement is true for F if it is true for F . Hence we may assume that F is the identity map; in this case, we need only show that the edge homomorphism of the spectral sequence of
234
F [] = [] is the identity map. This is an immediate consequence of the fact that we have a quasi-isomorphism ( 0 (13.9) Corollary. If Rq F H l (X, ) induced by F # . 0 )
[]
= 0 for q
1, there is an isomorphism H l (Y, F )
p,q Proof. We are in the special case 10.12 with E2 = 0 for q = 0, so l,0 H l (Y, F ) = E2 H l (X, ).
(13.10) Corollary. Let F : X Y be a proper nite-to-one map. For any sheaf on X, we have Rq F = 0 for q 1 and there is an isomorphism H l (Y, F ) H l (X, ). Proof. By denition of higher direct images, we have (Rq F )y = lim H q
Uy []
F 1 (U ) .
If F 1 (y) = {x1 , . . . , xm }, the assumptions imply that F 1 (U ) is a fundamental system of neighborhoods of {x1 , . . . , xm }. Therefore (Rq F )y =
1 j m
Hq
[] xj
xj
for q = 0, for q 1,
and we conclude by Cor. 13.9. Corollary 13.10 can be applied in particular to the inclusion j : S X of a closed subspace S. Then j coincides with the sheaf S dened in 3 and we get H q (S, ) = H q (X, S ). It is very important to observe that the property Rq j = 0 for q 1 need not be true if S is not closed. 13.C. Topological Dimension As a rst application of the Leray spectral sequence, we are going to derive some properties related to the concept of topological dimension. (13.11) Denition. A non empty space X is said to be of topological dimension n if on X. We let topdim X be the smallest H q (X, ) = 0 for any q > n and any sheaf such integer n if it exists, and + otherwise. (13.12) Criterion. For a paracompact space X, the following conditions are equivalent: a) topdim X b) the sheaf c) every sheaf sheaves. n;
n
= ker(
[n]
[n+1]
) is soft for every sheaf
; 0 of length n by soft
admits a resolution 0
14. Alexander-Spanier Cohomology
235
n
Proof. b) = c). Take
q =
[q]
for q < n and
n =
c) = a). For every sheaf , the De Rham-Weil isomorphism implies H q (X, ) = H q (X) = 0 when q > n. a) = b). Let S be a closed set and U = X S. As in Prop. 7.12, ( [] )U is an acyclic resolution of U . Clearly ker ( [n] )U ( [n+1] )U = n , so the isomorphisms (6.2) U obtained in the proof of the De Rham-Weil theorem imply H 1 (X, By (3.10), the restriction map
n n U)
(X)
H n+1 (X,
n
U)
= 0.
n
(S) is onto, so
is soft.
(13.13) Theorem. The following properties hold: a) If X is paracompact and if every point of X has a neighborhood of topological dimension n, then topdim X n. b) If S X, then topdim S topdim X provided that S is closed or X metrizable.
c) If X, Y are metrizable spaces, one of them locally compact, then topdim (X Y ) topdim X + topdim Y. n. d) If X is metrizable and locally homeomorphic to a subspace of Rn , then topdim X
Proof. a) Apply criterion 13.12 b) and the fact that softness is a local property (Prop. 4.12). b) When S is closed in X, the property follows from Cor. 13.10. When X is metrizable, a sheaf any subset S is strongly paracompact. Let j : S X be the injection and on S. As = (j )S , we have H q (S, ) = H q (S, j ) = lim H q (, j )
S
by Th. 9.10. We may therefore assume that S is open in X. Then every point of S has a neighborhood which is closed in X, so we conclude by a) and the rst case of b). c) Thanks to a) and b), we may assume for example that X is compact. Let be a sheaf on X Y and : X Y Y the second projection. Set nX = topdim X, nY = topdim Y . In virtue of (13.6 ), we have Rq = 0 for q > nX . In the Leray spectral sequence, we obtain therefore
p,q E2 = H p (Y, Rq ) = 0
for p > nY or q > nX , 1, because [0, 1] admits

0 k
p,lp thus E = 0 when l > nX + nY and we infer H l (X Y, ) = 0.
d) The unit interval [0, 1] R is of topological dimension arbitrarily ne coverings (13.14)
k =
[0, 1] ]( 1)/k, ( + 1)/k[
for which only consecutive open sets U , U+1 intersect; we may therefore apply Prop. 5.24. Hence Rn ]0, 1[n [0, 1]n is of topological dimension n by b) and c). Property d) follows.
236

14.A. Invariance by Homotopy Alexander-Spaniers theory can be viewed as the special case of sheaf cohomology theory with constant coecients, i.e. with values in constant sheaves. (14.1) Denition. Let X be a topological space, R a commutative ring and M a Rmodule. The constant sheaf X M is denoted M for simplicity. The Alexander-Spanier q-th cohomology group with values in M is the sheaf cohomology group H q (X, M ). In particular H 0 (X, M ) is the set of locally constant functions X M , therefore H (X, M ) M E , where E is the set of connected components of X. We will not repeat here the properties of Alexander-Spanier cohomology groups that are formal consequences of those of general sheaf theory, but we focus our attention instead on new features, such as invariance by homotopy.
0
(14.2) Lemma. Let I denote the interval [0, 1] of real numbers. Then H 0 (I, M ) = M and H q (I, M ) = 0 for q = 0. Proof. Let us employ alternate Cech cochains for the coverings As I is paracompact, we have H q (I, M ) = lim H q (n , M ).
n dened in (13.14).
However, the alternate Cech complex has only two non zero components and one non zero dierential: AC 0 (n , M ) = (c )0 n = M n+1 , AC 1 (n , M ) = (c +1 )0
n1
= M n,
We see that d0 is surjective, and that ker d0 = (m, m, . . . , m) = M . For any continuous map f : X Y , the inverse image of the constant sheaf M on Y is f 1 M = M . We get therefore a morphism (14.3) f : H q (Y, M ) H q (X, M ),
d0 : (c ) (c +1 ) = (c+1 c ).
which, as already mentioned in 9, is compatible with cup product. (14.4) Proposition. For any space X, the projection : X I X and the injections it : X X I, x (x, t) induce inverse isomorphisms H q (X, M ) H q (X I, M ).
it
In particular, i does not depend on t. t Proof. As it = Id, we have i = Id, so it is sucient to check that is an t isomorphism. However (Rq M )x = H q (I, M ) in virtue of (13.6 ), so we get R0 M = M, Rq M = 0 for q = 0
237
and conclude by Cor. 13.9. (14.4) Theorem. If f, g : X Y are homotopic maps, then f = g : H q (Y, M ) H q (X, M ). Proof. Let H : X I Y be a homotopy between f and g, with f = H i0 and g = H i1 . Proposition 14.3 implies f = i H = i H = g . 1 0 We denote f g the homotopy equivalence relation. Two spaces X, Y are said to be homotopically equivalent (X Y ) if there exist continuous maps u : X Y , v : Y X such that v u IdX and u v IdY . Then H q (X, M ) H q (Y, M ) and u , v are inverse isomorphisms. (14.5) Example. A subspace S X is said to be a (strong) deformation retract of X if there exists a retraction of X onto S, i.e. a map r : X S such that r j = IdS (j = inclusion of S in X), which is a deformation of IdX , i.e. there exists a homotopy H : X I X relative to S between IdX and j r : H(x, 0) = x, H(x, 1) = r(x) on X, H(x, t) = x on S I.
Then X and S are homotopically equivalent. In particular X is said to be contractible if X has a deformation retraction onto a point x0 . In this case H q (X, M ) = H q ({x0 }, M ) = M 0 for q = 0 for q = 0.
(14.6) Corollary. If X is a compact dierentiable manifold, the cohomology groups H q (X, R) are nitely generated over R. Proof. Lemma 6.9 shows that X has a nite covering such that the intersections U0 ...q are contractible. Hence is acyclic, H q (X, R) = H q C (, R) and each Cech cochain space is a nitely generated (free) module. (14.7) Example: Cohomology Groups of Spheres. Set S n = x Rn+1 ; x2 + x2 + . . . + x2 = 1 , 0 1 n We will prove by induction on n that (14.8) H q (S n , M ) = M 0 for q = 0 or q = n otherwise. n 1.
As S n is connected, we have H 0 (S n , M ) = M . In order to compute the higher cohomology groups, we apply the Mayer-Vietoris exact sequence 3.11 to the covering (U1 , U2 ) with U1 = S n {(1, 0, . . . , 0)}, U2 = S n {(1, 0, . . . , 0)}.
238
Then U1 , U2 Rn are contractible, and U1 U2 can be retracted by deformation on the equator S n {x0 = 0} S n1 . Omitting M in the notations of cohomology groups, we get exact sequences (14.9 ) (14.9 ) H 0 (U1 ) H 0 (U2 ) H 0 (U1 U2 ) H 1 (S n ) 0, 0 H q1 (U1 U2 ) H q (S n ) 0, q 2.
For n = 1, U1 U2 consists of two open arcs, so we have H 0 (U1 ) H 0 (U2 ) = H 0 (U1 U2 ) = M M, and the rst arrow in (14.9 ) is (m1 , m2 ) (m2 m1 , m2 m1 ). We infer easily that H 1 (S 1 ) = M and that H q (S 1 ) = H q1 (U1 U2 ) = 0 for q 2.
For n 2, U1 U2 is connected, so the rst arrow in (14.9 ) is onto and H 1 (S n ) = 0. The second sequence (14.9 ) yields H q (S n ) H q1 (S n1 ). An induction concludes the proof. 14.B. Relative Cohomology Groups and Excision Theorem Let X be a topological space and S a subspace. We denote by M [q](X, S) the subgroup of sections u M [q] (X) such that u(x0 , . . . , xq ) = 0 when (x0 , . . . , xq ) S q , x1 V (x0 ), . . . , xq V (x0 , . . . , xq1 ).
Then M [] (X, S) is a subcomplex of M [] (X) and we dene the relative cohomology group of the pair (X, S) with values in M as (14.10) H q (X, S ; M ) = H q M [] (X, S) .
By denition of M [q](X, S), there is an exact sequence (14.11) 0 M [q] (X, S) M [q](X) (MS )[q](S) 0.
The reader should take care of the fact that (MS )[q](S) does not coincide with the module of sections M [q] (S) of the sheaf M [q] on X, except if S is open. The snake lemma shows that there is an exact sequence of the pair: (14.12) H q (X, S ; M ) H q (X, M ) H q (S, M ) H q+1 (X, S ; M ) . We have in particular H 0 (X, S ; M ) = M E , where E is the set of connected components of X which do not meet S. More generally, for a triple (X, S, T ) with X S T , there is an exact sequence of the triple: (14.12 ) 0 M [q](X, S) M [q](X, T ) M [q] (S, T ) 0,
H q (X, S ; M ) H q (X, T ; M ) H q (S, T ; M ) H q+1 (X, S ; M ).
239
The denition of the cup product in (8.2) shows that vanishes on S S if , vanish on S, S respectively. Therefore, we obtain a bilinear map (14.13) H p (X, S ; M ) H q (X, S ; M ) H p+q (X, S S ; M M ).
If f : (X, S) (Y, T ) is a morphism of pairs, i.e. a continuous map X Y such that f (S) T , there is an induced pull-back morphism (14.14) f : H q (Y, T ; M ) H q (X, S ; M )
which is compatible with the cup product. Two morphisms of pairs f, g are said to be homotopic when there is a pair homotopy H : (X I, S I) (Y, T ). An application of the exact sequence of the pair shows that : H q (X, S ; M ) H q (X I, S I ; M ) is an isomorphism. It follows as above that f = g as soon as f, g are homotopic. (14.15) Excision theorem. For subspaces T S of X, the restriction morphism H q (X, S ; M ) H q (X T, S T ; M ) is an isomorphism. Proof. Under our assumption, it is not dicult to check that the surjective restriction map M [q] (X, S) M [q] (X T, S T ) is also injective, because the kernel consists of sections u M [q] (X) such that u(x0 , . . . , xq ) = 0 on (X T )q+1 S q+1 , and this set is a neighborhood of the diagonal of X q+1 . (14.16) Proposition. If S is open or strongly paracompact in X, the relative cohomology groups can be written in terms of cohomology groups with supports in X S :
q H q (X, S ; M ) HX S (X, M ).
In particular, if X
S is relatively compact in X, we have

q H q (X, S ; M ) Hc (X
S, M ).
Proof. We have an exact sequence (14.17) where MX

[] S (X)
0 MX
[]
[]
S (X)
M [] (X) M [] (S) 0 S. If S is open, then M [] (S) =
denotes sections with support in X
(MS )[] (S), hence MX S (X) = M [] (X, S) and the result follows. If S is strongly paracompact, Prop. 4.7 and Th. 9.10 show that H q M [] (S) = H q lim M [] () = lim H q (, M ) = H q (S, MS ).
S S
If we consider the diagram 0 MX

[]
0 M [] (X, S) M [] (X) (MS )[] (S) 0
S (X)
M [] (X) M [] (S) S Id
240
we see that the last two vertical arrows induce isomorphisms in cohomology. Therefore, the rst one also does. (14.18) Corollary. Let X, Y be locally compact spaces and f, g : X Y proper maps. We say that f, g are properly homotopic if they are homotopic through a proper homotopy H : X I Y . Then
q q f = g : Hc (Y, M ) Hc (X, M ).
Proof. Let X = X {}, Y = Y {} be the Alexandrov compactications of X, Y . Then f, g, H can be extended as continuous maps f , g : X Y , H : X I Y
with f () = g() = H(, t) = , so that f , g are homotopic as maps (X, ) q (Y , ). Proposition 14.16 implies Hc (X, M ) = H q (X, ; M ) and the result follows.
15. Knneth Formula u

15.A. Flat Modules and Tor Functors The goal of this section is to investigate homological properties related to tensor products. We work in the category of modules over a commutative ring R with unit. All tensor products appearing here are tensor products over R. The starting point is the observation that tensor product with a given module is a right exact functor: if 0 A B C 0 is an exact sequence and M a R-module, then A M B M C M 0 is exact, but the map A M B M need not be injective. A counterexample is given by the sequence 0 Z Z Z/2Z 0
2
over R = Z
tensorized by M = Z/2Z. However, the injectivity holds if M is a free R-module. More generally, one says that M is a at R-module if the tensor product by M preserves injectivity, or equivalently, if M is a left exact functor. A at resolution C of a R-module A is a homology exact sequence Cq Cq1 C1 C0 A 0 where Cq are at R-modules and Cq = 0 for q < 0. Such a resolution always exists because every module A is a quotient of a free module C0 . Inductively, we take Cq+1 to be a free module such that ker(Cq Cq1 ) is a quotient of Cq+1 . In terms of homology groups, we have H0 (C ) = A and Hq (C ) = 0 for q = 0. Given R-modules A, B and free resolutions d : C A, d : D B, we consider the double homology complex Kp,q = Cp Dq , d = d Id +(1)p Id d
241
and the associated rst and second spectral sequences. Since Cp is free, we have
1 Ep,q = Hq (Cp D ) =
Cp B 0
for q = 0, for q = 0.
Similarly, the second spectral sequence also collapses and we have Hl (K ) = Hl (C B) = Hl (A D ). This implies in particular that the homology groups Hl (K ) do not depend on the choice of the resolutions C or D . (15.1) Denition. The q-th torsion module of A and B is Torq (A, B) = Hq (K ) = Hq (C B) = Hq (A D ). Since the denition of K is symmetric with respect to A, B, we have Torq (A, B) Torq (B, A). By the right-exactness of B, we nd in particular Tor0 (A, B) = A B. Moreover, if B is at, B is also left exact, thus Torq (A, B) = 0 for all q 1 and all modules A. If 0 A A A 0 is an exact sequence, there is a corresponding exact sequence of homology complexes 0 A D A D A D 0, thus a long exact sequence (15.2) Torq (A, B) Torq (A , B) Torq (A , B) Torq1 (A, B) A B A B A B 0.
It follows that B is at if and only if Tor1 (A, B) = 0 for every R-module A. Suppose now that R is a principal ring. Then every module A has a free resolution 0 C1 C0 A 0 because the kernel of any surjective map C0 A is free (every submodule of a free module is free). It follows that one always has Torq (A, B) = 0 for q 2. In this case, we denote Tor1 (A, B) = A B and call it the torsion product of A and B. The above exact sequence (15.2) reduces to (15.3) 0 A B A B A B A B A B A B 0.
In order to compute A B, we may restrict ourselves to nitely generated modules, because every module is a direct limit of such modules and the product commutes with direct limits. Over a principal ring R, every nitely generated module is a direct sum of a free module and of cyclic modules R/aR. It is thus sucient to compute R/aR R/bR. a The obvious free resolution R R of R/aR shows that R/aR R/bR is the kernel of a the map R/bR R/bR. Hence (15.4) R/aR R/bR R/(a b)R
where a b denotes the greatest common divisor of a and b. It follows that a module B is at if and only if it is torsion free. If R is a eld, every R-module is free, thus A B = 0 for all A and B.
242
15.B. K nneth and Universal Coecient Formulas u The algebraic Knneth formula describes the cohomology groups of the tensor product u of two dierential complexes. (15.5) Algebraic K nneth formula. Let (K , d ), (L , d ) be complexes of R-modules u and (K L) the simple complex associated to the double complex (K L)p,q = K p Lq . If K or L is torsion free, there is a split exact sequence 0 H p (K ) H q (L ) H l (K L)
H p (K ) H q (L )
p+q=l+1
p+q=l
where the map is dened by ({k p } {lq }) = {k p lq } for all cocycles {k p } H p (K ), {lq } H q (L ). (15.6) Corollary. If R is a eld, or if one of the graded modules H (K ), H (L ) is torsion free, then H l (K L) H p (K ) H q (L ).
p+q=l
Proof. Assume for example that K is torsion free. There is a short exact sequence of complexes 0 Z K B +1 0 where Z , B K are respectively the graded modules of cocycles and coboundaries in K , considered as subcomplexes with zero dierential. As B +1 is torsion free, the tensor product of the above sequence with L is still exact. The corresponding long exact sequence for the associated simple complexes yields: H l (B L) H l (Z L) H l (K L) H l+1 (Z L) . H l+1 (B L)
d d
(15.7)
The rst and last arrows are connecting homomorphisms; in this situation, they are easily seen to be induced by the inclusion B Z . Since the dierential of Z is zero, the p p simple complex (Z L) is isomorphic to the direct sum , where Z p is pZ L torsion free. Similar properties hold for (B L) , hence H l (Z L) = The exact sequence 0 B p Z p H p (K ) 0 tensorized by H q (L ) yields an exact sequence of the type (15.3): 0 H p (K ) H q (L ) B p H q (L ) Z p H q (L )
p+q=l
Z p H q (L ),
H l (B L) =
p+q=l
B p H q (L ).
H p (K ) H q (L ) 0.
243
By the above equalities, we get 0

p+q=l
H p (K ) H q (L ) H l (B L) H l (Z L) H p (K ) H q (L ) 0.
p+q=l
In our initial long exact sequence (15.7), the cokernel of the rst arrow is thus H p (K ) H q (L )
p+q=l
p q and the kernel of the last arrow is the torsion sum p+q=l+1 H (K ) H (L ). This gives the exact sequence of the lemma. We leave the computation of the map as an exercise for the reader. The splitting assertion can be obtained by observing that there always exists a torsion free complex K that splits (i.e. Z K splits), and a morphism K K inducing an isomorphism in cohomology; then the projection K Z yields a projection
H l (K L) H l (Z L)
p+q=l
Z p H q (L ) H p (K ) H q (L ).
p+q=l
To construct K , let Z Z be a surjective map with Z free, B the inverse image of B in and K = Z B +1 , where the dierential K K +1 is given by Z 0 and B +1 Z +1 0 ; as B is free, the map B +1 B +1 can be lifted to a map B +1 K , and this lifting combined with the composite Z Z K yields the required complex morphism K = Z B +1 K . (15.8) Universal coecient formula. Let K be a complex of R-modules and M a R-module such that either K or M is torsion free. Then there is a split exact sequence 0 H p (K ) M H p (K M ) H p+1 (K ) M 0. Indeed, this is a special case of Formula 15.5 when the complex L is reduced to one term L0 = M . In general, it is interesting to observe that the spectral sequence of K L collapses in E2 if K is torsion free: H k (K L) is in fact the direct sum of p,q the terms E2 = H p K H q (L ) thanks to (15.8). 15.C. K nneth Formula for Sheaf Cohomology u Here we apply the general algebraic machinery to compute cohomology groups over a product space X Y . The main argument is a combination of the Leray spectral sequence with the universal coecient formula for sheaf cohomology. (15.9) Theorem. Let be a sheaf of R-modules over a topological space X and M a R-module. Assume that either or M is torsion free and that either X is compact or M is nitely generated. Then there is a split exact sequence 0 H p (X, ) M H p (X, M ) H p+1 (X, ) M 0.
244
Proof. If M is nitely generated, we get ( M )[] (X) = [] (X) M easily, so the above exact sequence is a consequence of Formula 15.8. If X is compact, we may consider Cech cochains C q (, M ) over nite coverings. There is an obvious morphism C q (, ) M C q (, M)
but this morphism need not be surjective nor injective. However, since ( M )x =
x
M = lim
V x
(V ) M,
the following properties are easy to verify: a) If c C q (, M ), there is a renement of and : c C q (, M ) is in the image of C q (, ) M .
such that
b) If a tensor t C q (, ) M is mapped to 0 in C q (, of such that t C q (, ) M equals 0. From a) and b) it follows that H q (X, M ) = lim H q C (,
M ), there is a renement
M ) = lim H q C (, ) M
and the desired exact sequence is the direct limit of the exact sequences of Formula 15.8 obtained for K = C (, ). (15.10) Theorem (Knneth). Let u and be sheaves of R-modules over topological is torsion free, that Y is compact and that either X is spaces X and Y . Assume that compact or the cohomology groups H q (Y, ) are nitely generated R-modules. There is a split exact sequence 0 H p (X, ) H q (Y, ) H l (X Y, H p (X, ) H q (Y,
p+q=l+1
p+q=l
) ) 0 p q .
where is the map given by the cartesian product
p q
Proof. We compute H l (X, ) by means of the Leray spectral sequence of the projection : X Y X. This means that we are considering the dierential sheaf q = ( )[q] and the double complex K p,q = (q )[p] (X).
p,q By (12.5 ) we have K E2 = H p X, q ( ) . As Y is compact, the cohomology sheaves q () = Rq ( ) are given by
R q (
)x = H q ({x} Y,
{x}Y
) = H q (Y,
)=
H q (Y,
16. Poincar duality e
245
thanks to the compact case of Th. 15.9 where M = x is torsion free. We obtain therefore Rq ( ) = H q (Y, ),
p,q K E2
= H p X,
Theorem 15.9 shows that the E2 -term is actually given by the desired exact sequence, but it is not a priori clear that the spectral sequence collapses in E2 . In order to check this, we consider the double complex C p,q =
[p]
H q (Y,
) .
(X)
[q]
(Y )
and construct a natural morphism C , K , . We may consider the set K p,q = [p] ( )[q] (X) as the set of equivalence classes of functions h 0 , . . . , p ) ( or more precisely
[q] )p = lim (
)[q] 1 V (p )
(xj , yj ) V 0 , . . . , p ; (x0 , y0 ), . . . , (xj1 , yj1 ) , Then f g C p,q is mapped to h K p,q by the formula
0 X, j V (0 , . . . , j1 ), (x0 , y0 ) V (0 , . . . , p ) Y,
h 0 , . . . , p ; (x0 , y0 ), . . . , (xq , yq ) 1
xq
yq
with
p, 1 j q.
h 0 , . . . , p ; (x0 , y0 ), . . . , (xq , yq ) = f (0 , . . . , p )(xq ) g(y0 , . . . , yq ). As

[p]
(X) is torsion free, we nd

p,q C E1
[p]
(X) H q (Y, H q (Y,
).
Since either X is compact or H q (Y,

p,q C E2
) nitely generated, Th. 15.9 yields

p,q ) K E2
= H p X,
hence H l (K ) H l (C ) and the algebraic Knneth formula 15.5 concludes the proof. u (15.11) Remark. The exact sequences of Th. 15.9 and of Knneths theorem also u hold for cohomology groups with compact support, provided that X and Y are locally compact and (or ) is torsion free. This is an immediate consequence of Prop. 7.12 on direct limits of cohomology groups over compact subsets. are torsion free constant sheaves, e.g. = =Z (15.12) Corollary. When and or R, the Knneth formula holds as soon as X or Y has the same homotopy type as a u nite cell complex. Proof. If Y satises the assumption, we may suppose in fact that Y is a nite cell complex by the homotopy invariance. Then Y is compact and H (Y, ) is nitely generated, so Th. 15.10 can be applied.
246

16.A. Injective Modules and Ext Functors The study of duality requires some algebraic preliminaries on the Hom functor and its derived functors Extq . Let R be a commutative ring with unit, M a R-module and 0 A B C 0 an exact sequence of R-modules. Then we have exact sequences 0 HomR (M, A) HomR (M, B) HomR (M, C), HomR (A, M ) HomR (B, M ) HomR (C, M ) 0, i.e. Hom(M, ) is a covariant left exact functor and Hom(, M ) a contravariant right exact functor. The module M is said to be projective if Hom(M, ) is also right exact, and injective if Hom(, M ) is also left exact. Every free R-module is projective. Conversely, if M is projective, any surjective morphism F M from a free module F onto M must split IdM has a preimage in Hom(M, F ) ; if R is a principal ring, projective is therefore equivalent to free. (16.1) Proposition. Over a principal ring R, a module M is injective if and only if it is divisible, i.e. if for every x M and R {0}, there exists y M such that y = x. Proof. If M is injective, the exact sequence 0 R R R/R 0 shows that M = Hom(R, M ) Hom(R, M ) = M must be surjective, thus M is divisible. Conversely, assume that R is divisible. Let f : A M be a morphism and B A. Zorns lemma implies that there is a maximal extension f : A M of f where A A B. If A = B, select x B A and consider the ideal I of elements R such that x A. As R is principal we have I = 0 R for some 0 . If 0 = 0, select y M such that 0 y = f (0 x) and if 0 = 0 take y arbitrary. Then f can be extended to A + Rx by letting f (x) = y. This is a contradiction, so we must have A = B. (16.2) Proposition. Every module M can be embedded in an injective module M . Proof. Assume rst R = Z. Then set M = HomZ (M, Q/Z), M = HomZ (M , Q/Z) Q/ZM .

Since Q/Z is divisible, Q/Z and Q/ZM are injective. It is therefore sucient to show that the canonical morphism M M is injective. In fact, for any x M {0}, the subgroup Zx is cyclic (nite or innite), so there is a non trivial morphism Zx Q/Z, and we can extend this morphism into a morphism u : M Q/Z. Then u M and u(x) = 0, so M M is injective.
Now, for an arbitrary ring R, we set M = HomZ R, Q/ZM . There are R-linear embeddings M = HomR (R, M ) HomZ (R, M ) HomZ R, Q/ZM
=M
247
and since HomR (, M ) HomZ , Q/ZM , it is clear that M is injective over the ring R. As a consequence, any module has a (cohomological) resolution by injective modules. Let A, B be given R-modules, let d : B D be an injective resolution of B and let d : C A be a free (or projective) resolution of A. We consider the cohomology double complex K p,q = Hom(Cq , Dp ), d = d + (1)p (d ) ( means transposition) and the associated rst and second spectral sequences. Since Hom(, Dp ) and Hom(Cq , ) are exact, we get
p,0 E1 = Hom(A, Dp ), p,q p,q E1 = E1 = 0 p,0 E1 = Hom(Cp , B),
for q = 0.
Therefore, both spectral sequences collapse in E1 and we get H l (K ) = H l Hom(A, D ) = H l Hom(C , B) ; in particular, the cohomology groups H l (K ) do not depend on the choice of the resolutions C or D . (16.3) Denition. The q-th extension module of A, B is Extq (A, B) = H q (K ) = H q Hom(A, D ) = H q Hom(C , B) . R By the left exactness of Hom(A, ), we get in particular Ext0 (A, B) = Hom(A, B). If A is projective or B injective, then clearly Extq (A, B) = 0 for all q 1. Any exact sequence 0 A A A 0 is converted into an exact sequence by Hom(, D ), thus we get a long exact sequence 0 Hom(A , B) Hom(A , B) Hom(A, B) Ext1 (A , B) Similarly, any exact sequence 0 B B B 0 yields 0 Hom(A, B) Hom(A, B ) Hom(A, B ) Ext1 (A, B) Extq (A, B) Extq (A, B ) Extq (A, B ) Extq+1 (A, B)
Extq (A , B) Extq (A , B) Extq (A, B) Extq+1 (A , B)
Suppose now that R is a principal ring. Then the resolutions C or D can be taken of length 1 (any quotient of a divisible module is divisible), thus Extq (A, B) is always 0 for q 2. In this case, we simply denote Ext1 (A, B) = Ext(A, B). When A is nitely generated, the computation of Ext(A, B) can be reduced to the cyclic case, since a Ext(A, B) = 0 when A is free. For A = R/aR, the obvious free resolution R R gives (16.4) ExtR (R/aR, B) = B/aB.
248
(16.5) Lemma. Let K be a homology complex and let M M be an injective resolution of a R-module M . Let L be the simple complex associated to Lp,q = HomR (Kq , M p). There is a split exact sequence 0 Ext Hq1 (K ), M H q (L ) Hom Hq (K ), M 0. Proof. As the functor HomR (, M p) is exact, we get
p,q L E1 p,q L E2
The spectral sequence collapses in E2 , therefore we get G0 H q (L ) = Hom Hq (K ), M , G1 H q (L ) = Ext Hq1 (K ), M and the expected exact sequence follows. By the same arguments as at the end of the proof of Formula 15.5, we may assume that K is split, so that there is a projection Kq Zq . Then the composite morphism Hom Hq (K ), M = Hom(Zq /Bq , M ) Hom(Kq /Bq , M ) denes a splitting of the exact sequence. 16.B. Poincar Duality for Sheaves e Let be a sheaf of abelian groups on a locally compact topological space X of nite admits a soft resolution of topological dimension n = topdim X. By 13.12 c), length n. If M M 0 M 1 0 is an injective resolution of a R-module M , we introduce the double complex of presheaves p,q dened by ,M (16.6)
p,q ,M (U )
= Hom Hq (K ), M p , Hom Hq (K ), M = Ext Hq (K ), M 0
for p = 0, for p = 1, for p 2.
Z q (L ) H q (L )
= HomR
nq (U ), M p c
where the restriction map p,q (U ) p,q (V ) is the adjoint of the inclusion map ,M ,M nq (V ) nq (U ) when V U . As nq is soft, any f nq (U ) can be written c c c as f = f with (f ) subordinate to any open covering (U ) of U ; it follows easily that p,q satisfy axioms (II-2.4) of sheaves. The injectivity of M p implies that p,q ,M ,M is a abby sheaf. By Lemma 16.5, we get a split exact sequence
nq+1 0 Ext Hc (X, ), M H q ,M (X)
(16.7)
nq Hom Hc (X, ), M 0.
This can be seen as an abstract Poincar duality formula, relating the cohomology groups e of a dierential sheaf to the dual of the cohomology with compact ,M dual of support of . In concrete applications, it still remains to compute H q ,M (X) . This can be done easily when X is a manifold and is a constant or locally constant sheaf.
249
16.C. Poincar Duality on Topological Manifolds e Here, X denotes a paracompact topological manifold of dimension n. (16.8) Denition. Let L be a R-module. A locally constant sheaf of stalk L on X is such that every point has a neighborhood on which is R-isomorphic to a sheaf the constant sheaf L. Thus, a locally constant sheaf can be seen as a discrete ber bundle over X whose bers are R-modules and whose transition automorphisms are R-linear. If X is locally contractible, a locally constant sheaf of stalk L is given, up to isomorphism, by a representation : 1 (X) AutR (L) of the fundamental group of X, up to conjugation; denoting by X the universal covering of X, the sheaf associated to can be viewed as the quotient of X L by the diagonal action of 1 (X). We leave the reader check himself the details of these assertions: in fact similar arguments will be explained in full details in V-6 when properties of at vector bundles are discussed.
p,q ,M
be a locally constant sheaf of stalk L, let be a soft resolution of and Let the associated abby sheaves. For an arbitrary open set U X, Formula (16.7) gives a (non canonical) isomorphism Hq
,M (U ) nq nq+1 Hom Hc (U, ), M Ext Hc (U, ), M
and in the special case q = 0 a canonical isomorphism (16.9) H0

,M (U ) n = Hom Hc (U, ), M .
For an open subset homeomorphic to Rn , we have the exact sequence of the pair yield
q Hc (, L) H q (S n , {} ; L) =
L. Proposition 14.16 and
L for q = n, 0 for q = n.
If Rn , we nd H0
,M ()
Hom(L, M ),
H1
,M ()
Ext(L, M )
and H q ,M () = 0 for q = 0, 1. Consider open sets V where V is a deformation retract of . Then the restriction map H q ,M () H q ,M (V ) is an isomorphism. Taking the direct limit over all such neighborhoods V of a given point x , we see that 0 ( ,M ) and 1 ( ,M ) are locally constant sheaves of stalks Hom(L, M ) and Ext(L, M ), and that q ( ,M ) = 0 for q = 0, 1. When Ext(L, M ) = 0, the complex 0 0 ,M is thus a abby resolution of = ( ,M ) and we get isomorphisms (16.10) (16.11) Hq
0(U ) = H
,M (X) 0
= H q (X, 0 ),
,M (U ) n = Hom Hc (U, ), M . Z,Z )
(16.12) Denition. The locally constant sheaf X = 0 (

n X (U ) = HomZ Hc (U, Z), Z
of stalk Z dened by
250
is called the orientation sheaf (or dualizing sheaf ) of X. This sheaf is given by a homomorphism 1 (X) {1, 1} ; it is not dicult to check that X coincides with the trivial sheaf Z if and only if X is orientable (cf. exercice 18.?). n n In general, Hc (U, ) = Hc (U, Z) Z (U ) for any small open set U on which is trivial, thus 0 ( ,M ) = X Z Hom( , M ). A combination of (16.7) and (16.10) then gives: (16.13) Poincar duality theorem. Let X be a topological manifold, let be a locally e constant sheaf over X of stalk L and let M be a R-module such that Ext(L, M ) = 0. There is a split exact sequence
nq+1 0 Ext Hc (X, ), M H q X, X Hom( , M )
nq Hom Hc (X, ), M 0.
In particular, if either X is orientable or R has characteristic 2, then

nq+1 nq 0 Ext Hc (X, R), R H q (X, R) Hom Hc (X, R), R
0.
(16.14) Corollary. Let X be a connected topological manifold, n = dim X. Then for any R-module L
n a) Hc (X, X L) L ; n b) Hc (X, L) L/2L if X is not orientable.
Proof. First assume that L is free. For q = 0 and formula gives an isomorphism
= X L, the Poincar duality e
n Hom Hc (X, X L), M Hom(L, M )
and the isomorphism is functorial with respect to morphisms M M . Taking M = L n or M = Hc (X, X L), we easily obtain the existence of inverse morphisms
n n Hc (X, X L) L and L Hc (X, X L),
hence equality a). Similarly, for
= L we get
n Hom Hc (X, L), M H 0 X, X Hom(L, M ) .
If X is non orientable, then X is non trivial and the global sections of the sheaf X Hom(L, M ) consist of 2-torsion elements of Hom(L, M ), that is
n Hom Hc (X, L), M Hom(L/2L, M ).
Formula b) follows. If L is not free, the result can be extended by using a free resolution 0 L1 L0 L 0 and the associated long exact sequence.
251
(16.15) Remark. If X is a connected non compact n-dimensional manifold, it can be proved (exercise 18.?) that H n (X, ) = 0 for every locally constant sheaf on X. Assume from now on that X is oriented. Replacing M by LM and using the obvious morphism M Hom(L, L M ), the Poincar duality theorem yields a morphism e (16.16)
nq H q (X, M ) Hom Hc (X, L), L M ,
in other words, a bilinear pairing (16.16 )

nq Hc (X, L) H q (X, M ) L M.
(16.17) Proposition. Up to the sign, the above pairing is given by the cup product, n modulo the identication Hc (X, L M ) L M . Proof. By functoriality in L, we may assume L = R. Then we make the following special choices of resolutions:
q = R[q]
for q < n,
[n] [n1] M 0 = an injective module containing Mc (X)/dn1 Mc (X).
n = ker(R[q] R[q+1]),
We embed M in M 0 by u Z where u Z[n] (X) is a representative of a generator n of Hc (X, Z), and we set M 1 = M 0 /M . The projection map M 0 M 1 can be seen as an extension of
[n] [n1] [n] dn : Mc (X)/dn1 Mc (X) dn Mc (X), n since Ker dn Hc (X, M ) = M . The inclusion dn Mc (X) M 1 can be extended into a [n+1] map : Mc (X) M 1 . The cup product bilinear map [nq] [n] M [q](U ) Rc (U ) Mc (X) M 0 [n]
gives rise to a morphism M [q] (U ) (16.18)
q R,M (U )
dened by
M [q] (U ) Hom nq (U ), M 0 Hom nq+1 (U ), M 1 c c f (g f g) h (f h) .
This morphism is easily seen to give a morphism of dierential sheaves M [] , R,M when M [] is truncated in degree n, i.e. when M [n] is replaced by Ker dn . The induced morphism M = 0 (M [] ) 0 ( ) R,M is then the identity map, hence the cup product morphism (16.18) actually induces the Poincar duality map (16.16). e (16.19) Remark. If X is an oriented dierentiable manifold, the natural isomorphism n Hc (X, R) R given by 16.14 a) corresponds in De Rham cohomology to the integration morphism f X f , f n (X). Indeed, by a partition of unity, we may assume that Supp f Rn . The proof is thus reduced to the case X = Rn , which itself reduces
252
to X = R since the cup product is compatible with the wedge product of forms. Let us consider the covering = (]k 1, k + 1[)kZ and a partition of unity (k ) subordinate to . The Cech dierential AC 0 (, Z) AC 1 (, Z) (ck ) (ck k+1 ) = (ck+1 ck )
1 shows immediately that the generators of Hc (R, Z) are the 1-cocycles c such that c01 = 1 and ck k+1 = 0 for k = 0. By Formula (6.12), the associated closed dierential form is f = c01 1 d0 + c10 0 d1 ,
hence f = 1[0,1] d0 and f does satisfy (16.20) Corollary. If X is an oriented
f = 1.
manifold, the bilinear map ({f }, {g}) f g
nq Hc (X, R) H q (X, R) R,
nq is well dened and identies H q (X, R) to the dual of Hc (X, R).
253
Chapter V
Hermitian Vector Bundles
This chapter introduces the basic denitions concerning vector bundles and connections. In the rst sections, the concepts of connection, curvature form, rst Chern class are considered in the framework of dierentiable manifolds. Although we are mainly interested in complex manifolds, the ideas which will be developed in the next chapters also involve real analysis and real geometry as essential tools. In the second part, the special features of connections over complex manifolds are investigated in detail: Chern connections, rst Chern class of type (1, 1), induced curvature forms on sub- and quotient bundles, . . . . These general concepts are then illustrated by the example of universal vector bundles over Pn and over Grassmannians.
1. Denition of Vector Bundles

Let M be a dierentiable manifold of dimension m and let K = R or K = C be the scalar eld. A (real, complex) vector bundle of rank r over M is a manifold E together with i) a
ii) a K-vector space structure of dimension r on each ber Ex = 1 (x) such that the vector space structure is locally trivial. This means that there exists an open covering (V )I of M and dieomorphisms called trivializations : EV V Kr , such that for every x V the map
Ex {x} Kr Kr
map : E M called the projection,
where EV = 1 (V ),
is a linear isomorphism. For each , I, the map

1 = : (V V ) Kr (V V ) Kr
acts as a linear automorphism on each ber {x} Kr . It can thus be written (x, ) = (x, g (x) ), (x, ) (V V ) Kr
where (g )(,)II is a collection of invertible matrices with coecients in V , K), satisfying the cocycle relation (1.1) g g = g on V V V .
(V
254
Chapter V. Hermitian Vector Bundles
The collection (g ) is called a system of transition matrices. Conversely, any collection of invertible matrices satisfying (1.1) denes a vector bundle E, obtained by gluing the charts V Kr via the identications . (1.2) Example. The product manifold E = M Kr is a vector bundle over M , and is called the trivial vector bundle of rank r over M . We shall often simply denote it Kr for brevity. (1.3) Example. A much more interesting example of real vector bundle is the tangent bundle T M ; if : V Rn is a collection of coordinate charts on M , then = d : T MV V Rm dene trivializations of T M and the transition matrices 1 are given by g (x) = d (x ) where = and x = (x). The dual T M of T M is called the cotangent bundle and the p-th exterior power p T M is called the bundle of dierential forms of degree p on M . (1.4) Denition. If M is an open subset and k a positive integer or +, we let k (, E) denote the space of k sections of E , i.e. the space of k maps s : E such that s(x) Ex for all x (that is s = Id ). Let : EV V Kr be a trivialization of E. To , we associate the (e1 , . . . , er ) of EV dened by e (x) = 1 (x, ), x V,
k
frame
where ( ) is the standard basis of Kr . A section s in terms of its components (s) = = (1 , . . . , r ) by s=

1 r
(V, E) can then be represented
on V,
(V, K).
Let ( ) be a family of trivializations relative to a covering (V ) of M . Given a global section s k (M, E), the components (s) = = (1 , . . . , r ) satisfy the transition relations (1.5) = g on V V .
Conversely, any collection of vector valued functions : V Kr satisfying the transition relations denes a global section s of E. More generally, we shall also consider dierential forms on M with values in E. Such forms are nothing else than sections of the tensor product bundle p T M R E. We shall write (1.6) (1.7)
k p (, E) k (, E)
= =
(, p T M R E)
k p (, E).
0 p m
3. Curvature Tensor
255
2. Linear Connections
A (linear) connection D on the bundle E is a linear dierential operator of order 1 acting on (M, E) and satisfying the following properties: (2.1) (2.1 ) D : (M, E) (M, E), q q+1 D(f s) = df s + (1)p f Ds
q (M, E),
for any f (M, K) and s p derivative of f .
where df stands for the usual exterior
Assume that : E Kr is a trivialization of E , and let (e1 , . . . , er ) be the corresponding frame of E . Then any s (, E) can be written in a unique way q s=
1 r
e ,
q (, K).
By axiom (2.1 ) we get Ds =

1 r
d e + (1)p De .
1 (, K),
If we write De =
1 r
a e where a d +

we thus have
Ds =
a e .
Identify E with Kr via and denote by d the trivial connection d = (d ) on Kr . Then the operator D can be written (2.2) Ds d + A where A = (a ) (, Hom(Kr , Kr )). Conversely, it is clear that any operator D 1 dened in such a way is a connection on E . The matrix 1-form A will be called the connection form of D associated to the trivialization . If : E Kr is another trivialization and if we set g = 1
(, Gl(Kr ))
then the new components = ( ) are related to the old ones by = g. Let A be the connection form of D with respect to . Then Ds d + A =
Ds g 1 (d + A ) = g 1 (d(g) + A g) d + (g 1 Ag + g 1 dg) . A = g 1 Ag + g 1 dg. Therefore we obtain the gauge transformation law : (2.3)
256
3. Curvature Tensor
Let us compute D2 : (M, E) q : E Kr . We obtain
q+2 (M, E)
with respect to the trivialization
D2 s d(d + A ) + A (d + A )
= d2 + (dA A d) + (A d + A A ) = (dA + A A) .
2 (M, Hom(E, E))
It follows that there exists a global 2-form (D) vature tensor of D, such that D2 s = (D) s, given with respect to any trivialization by (3.1)
called the cur-
(D) dA + A A.
(3.2) Remark. If E is of rank r = 1, then A (M, K) and Hom(E, E) is canonically 1 isomorphic to the trivial bundle M K, because the endomorphisms of each ber Ex are homotheties. With the identication Hom(E, E) = K, the curvature tensor (D) can be considered as a closed 2-form with values in K: (3.3) (D) = dA.
In this case, the gauge transformation law can be written (3.4) A = A + g 1 dg, g = 1
(, K ).
It is then immediately clear that dA = dA, and this equality shows again that dA does not depend on . Now, we show that the curvature tensor is closely related to commutation properties of covariant derivatives. (3.5) Denition. If is a vector eld with values in T M , the covariant derivative of a section s (M, E) in the direction is the section D s (M, E) dened by D s = Ds . (3.6) Proposition. For all sections s (M, T M ), we have
(M, E) and all vector elds ,
D (D s) D (D s) = [, ]D s + (D)(, ) s where [, ]
(M, T M ) is the Lie bracket of , .
Proof. Let (x1 , . . . , xm ) be local coordinates on an open set M . Let : E Kr be a trivialization of E and let A be the corresponding connection form. If = j /xj and A = Aj dxj , we nd (3.7) D s (d + A) = j
j
+ Aj . xj
4. Operations on Vector Bundles
257
Now, we compute the above commutator [D , D ] at a given point z0 . Without loss of generality, we may assume A(z0 ) = 0 ; in fact, one can always nd a gauge transformation g near z0 such that g(z0 ) = Id and dg(z0 ) = A(z0 ) ; then (2.3) yields A(z0 ) = 0. If = k /xk , we nd D s k /xk at z0 , hence D (D s) k
k
xk
j
j
+ Aj , xj Aj (j k j k ) xk
D (D s) D (D s) j j k k + xk xk xj
j,k
j,k
= d([, ]) + dA(, ) , whereas (D) dA and [, ]D s d([, ]) at point z0 .
4. Operations on Vector Bundles

Let E, F be vector bundles of rank r1 , r2 over M . Given any functorial operation on vector spaces, a corresponding operation can be dened on bundles by applying the operation on each ber. For example E , E F , Hom(E, F ) are dened by (E )x = (Ex ) , (E F )x = Ex Fx , Hom(E, F )x = Hom(Ex , Fx ).
The bundles E and F can be trivialized over the same covering V of M (otherwise take a common renement). If (g ) and ( ) are the transition matrices of E and F , then for example E F , k E, E are the bundles dened by the transition matrices g , k g , (g )1 where denotes transposition. Suppose now that E, F are equipped with connections DE , DF . Then natural connections can be associated to all derived bundles. Let us mention a few cases. First, we let (4.1) It follows immediately that (4.1 ) (DEF ) = (DE ) (DF ). DEF = DE DF .
DEF will be dened in such a way that the usual formula for the dierentiation of a product remains valid. For every s (M, E), t (M, F ), the wedge product s t can be combined with the bilinear map E F E F in order to obtain a section s t (M, E F ) of degree deg s + deg t. Then there exists a unique connection DEF such that (4.2) DEF (s t) = DE s t + (1)deg s s DF t.
As the products s t generate (M, E F ), the uniqueness is clear. If E, F are trivial on an open set M and if AE , AF , are their connection 1-forms, the induced
258
connection DEF is given by the connection form AE IdF + IdE AF . The existence 2 2 2 follows. An easy computation shows that DEF (s t) = DE s t + s DF t, thus (4.2 ) (DEF ) = (DE ) IdF + IdE (DF ).
Similarly, there are unique connections DE and DHom(E,F ) such that (4.3) (4.4) whenever s (DE u) s = d(u s) (1)deg u
(M, E ), u
(DHom(E,F ) v) s = DF (v s) (1)
(M, E),
deg v
u DE s,
v DE s
Hom(E, F ) . It follows that
0 = d2 (u s) = (DE ) u s + u (DE ) s . If denotes the transposition operator Hom(E, E) Hom(E , E ), we thus get (4.3 ) (DE ) = (DE ) .
With the identication Hom(E, F ) = E F , Formula (4.2 ) implies (4.4 ) (DHom(E,F ) ) = IdE (DF ) (DE ) IdF .
(M, E)
Finally, k E carries a natural connection Dk E . For every s1 , . . . , sk in respective degrees p1 , . . . , pk , this connection satises (4.5) (4.5 ) Dk E (s1 . . . sk ) = (1)p1 +...+pj1 s1 . . . DE sj . . . sk ,
of
1 j k
(Dk E ) (s1 . . . sk ) =
1 j k
s1 . . . (DE ) sj . . . sk .
In particular, the determinant bundle, dened by det E = r E where r is the rank of E, has a curvature form given by (4.6) (Ddet E ) = E (DE )
where E : Hom(E, E) K is the trace operator. As a conclusion of this paragraph, we mention the following simple identity. (4.7) Bianchi identity. DHom(E,E) (DE ) = 0. Proof. By denition of DHom(E,E) , we nd for any s
3 3 = DE s DE s = 0.
(M, E)
DHom(E,E) (DE ) s = DE (DE ) s (DE ) (DE s)
6. Parallel Translation and Flat Vector Bundles
259
5. Pull-Back of a Vector Bundle

Let M , M be manifolds and : M M a smooth map. If E is a vector bundle on M , one can dene in a natural way a vector bundle : E M and a linear morphism : E E such that the diagram E M E
commutes and such that : Ex E(x) is an isomorphism for every x M . The bundle E can be dened by (5.1) E = {(x, ) M E ; (x) = ()}
and the maps and are then the restrictions to E of the projections of M E on M and E respectively. If : EV V Kr are trivializations of E, the maps = : E 1 (V ) 1 (V ) Kr dene trivializations of E with respect to the covering V = 1 (V ) of M . The corresponding system of transition matrices is given by (5.2) g = g on V V .
(5.3) Denition. E is termed the pull-back of E under the map and is denoted E = E. Let D be a connection on E. If (A ) is the collection of connection forms of D with respect to the s, one can dene a connection D on E by the collection of connection forms A = A V , Hom(Kr , Kr ) , i.e. for every s (V , E) 1 p Ds d + A .
Given any section s (M, E), one denes a pull back s which is a section in p (M , E) : for s = f u, f (M, K), u (M, E), set s = f (u ). p p Then we have the formula (5.4) D( s) = (Ds).
Using (5.4), a simple computation yields (5.5) (D) = ((D)).
260
6. Parallel Translation and Flat Vector Bundles

Let : [0, 1] M be a smooth curve and s : [0, 1] E a section of E along , i.e. a map s such that s(t) E(t) for all t [0, 1]. Then s can be viewed as a section of E = E over [0, 1]. The covariant derivative of s is the section of E along dened by (6.1) d Ds = Ds(t) E(t) , dt dt
where D is the induced connection on E. If A is a connection form of D with respect to a trivialization : E Kr , we have Ds d + A , i.e. (6.2) Ds d + A((t)) (t) (t) for (t) . dt dt
For v E(0) given, the Cauchy uniqueness and existence theorem for ordinary linear dierential equations implies that there exists a unique section s of E such that s(0) = v and Ds/dt = 0. (6.3) Denition. The linear map T : E(0) E(1) , is called parallel translation along . If = 2 1 is the composite of two paths 1 , 2 such that 2 (0) = 1 (1), it is clear 1 that T = T2 T1 , and the inverse path 1 : t (1 t) is such that T 1 = T . It follows that T is a linear isomorphism from E(0) onto E(1) . More generally, if h : W M is a map from a domain W Rp into M and if s is a section of h E, we dene covariant derivatives Ds/tj , 1 j p, by D = h D and (6.4) Ds = Ds . tj tj v = s(0) s(1)
Since /tj , /tk commute and since (D) = h (D), Prop. 3.6 implies (6.5) D Ds D Ds = (D) , tj tk tk tj tj tk s = (D)h(t) h h , tj tk s(t).
(6.6) Denition. The connection D is said to be at if (D) = 0. Assume from now on that D is at. We then show that T only depends on the homotopy class of . Let h : [0, 1] [0, 1] M be a smooth homotopy h(t, u) = u (t) from 0 to 1 with xed end points a = u (0), b = u (1). Let v Ea be given and let s(t, u) be such that s(0, u) = v and Ds/t = 0 for all u [0, 1]. Then s is in both variables (t, u) by standard theorems on the dependence of parameters. Moreover (6.5) implies that the covariant derivatives D/t, D/u commute. Therefore, if we set s = Ds/u, we nd Ds /t = 0 with initial condition s (0, u) = 0 (recall that s(0, u)
7. Hermitian Vector Bundles and Connections
261
is a constant). The uniqueness of solutions of dierential equations implies that s is identically zero on [0, 1] [0, 1], in particular Tu (v) = s(1, u) must be constant, as desired. (6.7) Proposition. Assume that D is at. If is a simply connected open subset of M , then E admits a parallel frame (e1 , . . . , er ), in the sense that De = 0 on , 1 r. For any two simply connected open subsets , the transition automorphism between the corresponding parallel frames (e ) and (e ) is locally constant. The converse statement E has parallel frames near every point implies that (D) = 0 can be immediately veried from the equality (D) = D2 . Proof. Choose a base point a and dene a linear isomorphism : Ea E by sending (x, v) on T (v) Ex , where is any path from a to x in (two such paths are always homotopic by hypothesis). Now, for any path from a to x, we have by construction (D/dt)((t), v) = 0. Set ev (x) = (x, v). As may reach any point x with an arbitrary tangent vector = (1) Tx M , we get Dev (x) = (D/dt)((t), v)t=1 = 0. Hence Dev is parallel for any xed vector v Ea ; Prop. 6.7 follows. Assume that M is connected. Let a be a base point and M M the universal covering of M . The manifold M can be considered as the set of pairs (x, []), where [] is a homotopy class of paths from a to x. Let 1 (M ) be the fundamental group of M with base point a, acting on M on the left by [] (x, []) = (x, [1 ]). If D is at, 1 (M ) acts also on Ea by ([], v) T (v), [] 1 (M ), v Ea , and we have a well dened map : M Ea E, (x, []) = T (v). Then is invariant under the left action of 1 (M ) on M Ea dened by [] (x, []), v = (x, [1 ]), T (v) , Conversely, let S be a K-vector space of dimension r together with a left action of 1 (M ). The quotient E = (M S)/1 (M ) is a vector bundle over M with locally constant transition automorphisms (g ) relatively to any covering (V ) of M by simply connected open sets. The relation = g implies d = g d on V V . We may therefore dene a connection D on E by letting Ds d on each V . Then clearly (D) = 0. therefore we have an isomorphism E (M Ea )/1 (M ).
7. Hermitian Vector Bundles and Connections

A complex vector bundle E is said to be hermitian if a positive denite hermitian form | |2 is given on each ber Ex in such a way that the map E R+ , ||2 is smooth. The notion of a euclidean (real) vector bundle is similar, so we leave the reader adapt our notations to that case. Let : E Cr be a trivialization and let (e1 , . . . , er ) be the corresponding frame of E . The associated inner product of E is given by a positive denite hermitian
262
matrix (h ) with
coecients on , such that e (x), e (x) = h (x), x .
When E is hermitian, one can dene a natural sesquilinear map

p (M, E) q (M, E)
(7.1)
(s, t) {s, t} e ,
p+q (M, C)
combining the wedge product of forms with the hermitian metric on E ; if s = t= e , we let {s, t} = e , e .
1 , r
A connection D is said to be compatible with the hermitian structure of E, or briey hermitian, if for every s (M, E), t (M, E) we have p q (7.2) d{s, t} = {Ds, t} + (1)p {s, Dt}.
Let (e1 , . . . , er ) be an orthonormal frame of E . Denote (s) = = ( ) and (t) = = ( ). Then {s, t} = {, } = ,
1 r
d{s, t} = {d, } + (1)p {, d }. Therefore D is hermitian if and only if its connection form A satises {A, } + (1)p {, A } = {(A + A ) , } = 0 for all , , i.e. (7.3) A = A or (a ) = (a ).
This means that iA is a 1-form with values in the space Herm(Cr , Cr ) of hermitian matrices. The identity d2 {s, t} = 0 implies {D2 s, t} + {s, D2 t} = 0, i.e. {(D) s, t} + {s, (D) t} = 0. Therefore (D) = (D) and the curvature tensor (D) is such that i (D) (M, Herm(E, E)). 2 (7.4) Special case. If E is a hermitian line bundle (r = 1), D is a hermitian connection if and only if its connection form A associated to any given orthonormal frame of E is a 1-form with purely imaginary values. If , : E are two such trivializations on a simply connected open subset M , then g = 1 = ei for some real phase function (, R). The gauge transformation law can be written A = A + i d. In this case, we see that i (D)
2 (M, R).
8. Vector Bundles and Locally Free Sheaves
263
(7.5) Remark. If s, s (M, E) are two sections of E along a smooth curve : [0, 1] M , one can easily verify the formula d Ds Ds s(t), s (t) = , s + s, . dt dt dt In particular, if (e1 , . . . , er ) is a parallel frame of E along such that e1 (0), . . . , er (0) is orthonormal, then e1 (t), . . . , er (t) is orthonormal for all t. All parallel translation operators T dened in 6 are thus isometries of the bers. It follows that E has a at hermitian connection D if and only if E can be dened by means of locally constant unitary transition automorphisms g , or equivalently if E is isomorphic to the hermitian bundle (M S)/1 (M ) dened by a unitary representation of 1 (M ) in a hermitian vector space S. Such a bundle E is said to be hermitian at.
8. Vector Bundles and Locally Free Sheaves

We denote here by the sheaf of germs of complex functions on M . Let F be the sheaf of germs of M be a complex vector bundle of rank r. We let sections of F , i.e. the sheaf whose space of sections on an open subset U M is (U ) = (U, F ). It is clear that is a -module. Furthermore, if F Cr is trivial, the sheaf is isomorphic to r as a -module. (8.1) Denition. A sheaf of modules over a sheaf of rings is said to be locally free of rank k if every point in the base has a neighborhood such that is -isomorphic to k . Suppose that is a locally free of M and sheaf isomorphisms -module of rank r. There exists a covering (V )I
r V .
: V
1 Then we have transition isomorphisms g = : r r dened on V V , and such an isomorphism is the multiplication by an invertible matrix with coecients on V V . The concepts of vector bundle and of locally free -module are thus completely equivalent.
Assume now that F M is a line bundle (r = 1). Then every collection of transition automorphisms g = (g ) denes a Cech 1-cocycle with values in the multiplicative sheaf of invertible functions on M . In fact the denition of the Cech dierential (cf. 1 (IV-5.1)) gives (g) = g g g , and we have g = 1 in view of (1.1). Let be another family of trivializations and (g ) the associated cocycle (it is no loss of generality to assume that both are dened on the same covering since we may otherwise take a renement). Then we have
1 : V C V C,
(x, ) (x, u (x)),
(V ).
It follows that g = g u1 u , i.e. that the Cech 1-cocycles g, g dier only by the Cech 1-coboundary u. Therefore, there is a well dened map which associates to every line
264
bundle F over M the Cech cohomology class {g} H 1 (M, ) of its cocycle of transition automorphisms. It is easy to verify that the cohomology classes associated to two line bundles F, F are equal if and only if these bundles are isomorphic: if g = g u, then the collection of maps
FV V C V C FV
(x, ) (x, u (x))
denes a global isomorphism F F . It is clear that the multiplicative group structure on H 1 (M, ) corresponds to the tensor product of line bundles (the inverse of a line bundle being given by its dual). We may summarize this discussion by the following: (8.2) Theorem. The group of isomorphism classes of complex line bundles is in one-to-one correspondence with the Cech cohomology group H 1 (M, ).
9. First Chern Class

Throughout this section, we assume that E is a complex line bundle (that is, rk E = r = 1). Let D be a connection on E. By (3.3), (D) is a closed 2-form on M . Moreover, if D is another connection on E, then (2.2) shows that D = D + where 1 (M, C). By (3.3), we get (9.1) (D ) = (D) + d.
2 This formula shows that the De Rham class {(D)} HDR (M, C) does not depend on the particular choice of D. If D is chosen to be hermitian with respect to a given hermitian metric on E (such a connection can always be constructed by means of a 2 partition of unity) then i (D) is a real 2-form, thus {i (D)} HDR (M, R). Consider now the one-to-one correspondence given by Th. 8.2:
class {E} dened by the cocycle (g ) class of (g ). Using the exponential exact sequence of sheaves 0 Z
{isomorphism classes of line bundles} H 1 (M,
f e
2if
and the fact that H 1 (M, ) = H 2 (M, ) = 0, we obtain: (9.2) Theorem and Denition. The coboundary morphism H 1 (M,
) H 2 (M, Z)
is an isomorphism. The rst Chern class of a line bundle E is the image c1 (E) in H 2 (M, Z) of the Cech cohomology class of the 1-cocycle (g ) associated to E : (9.3) c1 (E) = {(g )}.
9. First Chern Class
265
Consider the natural morphism (9.4)

2 H 2 (M, Z) H 2 (M, R) HDR (M, R)
where the isomorphism is that given by the De Rham-Weil isomorphism theorem and the sign convention of Formula (IV-6.11).
2 (9.5) Theorem. The image of c1 (E) in HDR (M, R) under (9.4) coincides with the De i Rham cohomology class { 2 (D)} associated to any (hermitian) connection D on E.
Proof. Choose an open covering (V )I of M such that E is trivial on each V , and such that all intersections V V are simply connected (as in IV-6, choose the V to be small balls relative to a given locally nite covering of M by coordinate patches). Denote by A the connection forms of D with respect to a family of isometric trivializations : EV V Cr . Let g (V V ) be the corresponding transition automorphisms. Then |g | = 1, and as V V is simply connected, we may choose real functions u (V V ) such that g = exp(2i u ). By denition, the rst Chern class c1 (E) is the Cech 2-cocycle c1 (E) ={(g )} = {(u) )} H 2 (M, Z) (u) :=u u + u . where
Now, if q (resp. q ) denotes the sheaf of real (resp. real d-closed) q-forms on M , the short exact sequences d 0 1 1 2 0 d 0 R 0 1 0 yield isomorphisms (with the sign convention of (IV-6.11)) (9.6) (9.7)
2 HDR (M, R) := H 0 (M, 2
)/dH 0 (M,
) H 1 (M,
),
H 1 (M,
) H 2 (M, R).
1 i Formula 3.4 gives A = A +g dg . Since (D) = dA on V , the image of { 2 (D)} under (9.6) is the Cech 1-cocycle with values in 1
i (A A ) = 2
1 1 g dg 2i
= {du }
and the image of this cocycle under (9.7) is the Cech 2-cocycle {u} in H 2 (M, R). But {u} is precisely the image of c1 (E) H 2 (M, Z) in H 2 (M, R). Let us assume now that M is oriented and that s (M, E) is transverse to the zero section of E, i.e. that Ds Hom(T M, E) is surjective at every point of the zero set Z := s1 (0). Then Z is an oriented 2-codimensional submanifold of M (the orientation of
266
Z is uniquely dened by those of M and E). We denote by [Z] the current of integration 2 over Z and by {[Z]} HDR (M, R) its cohomology class. (9.8) Theorem. We have {[Z]} = c1 (E)R . Proof. Consider the dierential 1-form u = s1 Ds
1 (M
Z, C).
Relatively to any trivialization of E , one has D d + A , thus u = d +A where = (s).
It follows that u has locally integrable coecients on M . If d/ is considered as a current on , then d d = d dz z = d dz z = (2i0 ) = 2i[Z]
because of the Cauchy residue formula (cf. Lemma I-2.10) and because is a submersion in a neighborhood of Z (cf. (I-1.19)). Now, we have dA = (D) and Th. 9.8 follows from the resulting equality: (9.9) du = 2i [Z] + (E).
10. Connections of Type (1,0) and (0,1) over Complex Manifolds

Let X be a complex manifold, dimC X = n and E a vector bundle of rank r over X ; here, E is not assumed to be holomorphic. We denote by (X, E) the space of p,q sections of the bundle p,q T X E. We have therefore a direct sum decomposition
l (X, E)
=
p+q=l
p,q (X, E).
Connections of type (1, 0) or (0, 1) are operators acting on vector valued forms, which imitate the usual operators d , d acting on (X, C). More precisely, a connection of p,q type (1,0) on E is a dierential operator D of order 1 acting on (X, E) and satisfying , the following two properties: (10.1) (10.1 ) D :
p,q (X, E)
D (f s) = d f s + (1)
p+1,q (X, E), deg f
f D s
for any f ,q1 (X, C), s ,q2 (X, E). The denition of a connection D of type p1 p2 (0,1) is similar. If : E Cr is a trivialization of E and if = ( ) = (s), then all such connections D and D can be written (10.2 ) (10.2 ) D s d + A D s d + A ,
11. Holomorphic Vector Bundles
267
where A , Hom(Cr , Cr ) , A 1,0 matrix coecients.
0,1
, Hom(Cr , Cr ) are arbitrary forms with
It is clear that D = D + D is then a connection in the sense of 2 ; conversely any connection D admits a unique decomposition D = D +D in terms of a (1,0)-connection and a (0,1)-connection. Assume now that E has a hermitian structure and that is an isometry. The connection D is hermitian if and only if the connection form A = A +A satises A = A, and this condition is equivalent to A = (A ) . From this observation, we get immediately:
(10.3) Proposition. Let D0 be a given (0, 1)-connection on a hermitian bundle : E X. Then there exists a unique hermitian connection D = D + D such that D = D0 .
11. Holomorphic Vector Bundles

From now on, the vector bundles E in which we are interested are supposed to have a holomorphic structure: (11.1) Denition. A vector bundle : E X is said to be holomorphic if E is a complex manifold, if the projection map is holomorphic and if there exists a covering (V )I of X and a family of holomorphic trivializations : EV V Cr . It follows that the transition matrices g are holomorphic on V V . In complete analogy with the discussion of 8, we see that the concept of holomorphic vector bundle is equivalent to the concept of locally free sheaf of modules over the ring of germs of holomorphic functions on X. We shall denote by (E) the associated sheaf of germs of holomorphic sections of E. In the case r = 1, there is a one-to-one correspondence between the isomorphism classes of holomorphic line bundles and the Cech cohomology 1 group H (X, ). (11.2) Denition. The group H 1 (X, ) of isomorphism classes of holomorphic line bundles is called the Picard group of X. If s
p,q (X, E), the components = ( )1 r
= (s) of s under are related
by
= g Since d g = 0, it follows that d = g d
on V V . on V V .
The collection of forms (d ) therefore corresponds to a unique global (p, q+1)-form d s such that (d s) = d , and the operator d dened in this way is a (0, 1)-connection on E. (11.3) Denition. The operator d is called the canonical (0, 1)-connection of the holomorphic bundle E.
268
It is clear that d2 = 0. Therefore, for any integer p = 0, 1, . . . , n, we get a complex

p,0 (X, E)
p,q (X, E)
p,q+1 (X, E)
known as the Dolbeault complex of (p, )-forms with values in E. (11.4) Notation. The q-th cohomology group of the Dolbeault complex is denoted H p,q (X, E) and is called the (p, q) Dolbeault cohomology group with values in E. The Dolbeault-Grothendieck lemma I-2.11 shows that the complex of sheaves d : 0, (X, E) is a soft resolution of the sheaf (E). By the De Rham-Weil isomorphism theorem IV-6.4, we get: (11.5) Proposition. H 0,q (X, E) H q X, (E) . Most often, we will identify the locally free sheaf (E) and the bundle E itself ; the above sheaf cohomology group will therefore be simply denoted H q (X, E). Another standard notation in analytic or algebraic geometry is: (11.6) Notation. If X is a complex manifold, p denotes the vector bundle p T X or X its sheaf of sections. It is clear that the complex (X, E) is identical to the complex p, therefore we obtain a canonical isomorphism: (11.7) Dolbeault isomorphism. H p,q (X, E) H q (X, p E). X p X In particular, H p,0 (X, E) is the space of global holomorphic sections of the bundle E.
p 0, (X, X
E),
12. Chern Connection

Let : E X be a hermitian holomorphic vector bundle of rank r. By Prop. 10.3, there exists a unique hermitian connection D such that D = d . (12.1) Denition. The unique hermitian connection D such that D = d is called the Chern connection of E. The curvature tensor of this connection will be denoted by (E) and is called the Chern curvature tensor of E. Let us compute D with respect to an arbitrary holomorphic trivialization : E Cr . Let H = (h )1 , r denote the hermitian matrix with coecients representing the metric along the bers of E . For any s, t (X, E) and = (s), = , (t) one can write {s, t} =
,
h = H ,
12. Chern Connection
269
where is the transposed matrix of . It follows that {Ds, t}+(1)deg s {s, Dt} = d{s, t} = d + H
1
= (d) H + (1)deg (dH + Hd ) dH
H + (1)deg (d + H
1 dH
using the fact that dH = d H + d H and H = H. Therefore the Chern connection D coincides with the hermitian connection dened by (12.2) (12.3) Ds d + H D d + H
1
d H ,
d H =H
d (H),
D = d .
It is clear from this relations that D2 = D2 = 0. Consequently D2 is given by to D2 = D D +D D , and the curvature tensor (E) is of type (1, 1). Since d d +d d = 0, we get (D D + D D )s H
1
d H d + d (H
d H ) = d (H
d H) .
(12.4) Theorem. The Chern curvature tensor is such that i (E)

1,1 (X, Herm(E, E)).
If : E Cr is a holomorphic trivialization and if H is the hermitian matrix representing the metric along the bers of E , then i (E) = i d (H
1
d H)
on
Let (e1 , . . . , er ) be a orthonormal frame of E over a coordinate patch X with complex coordinates (z1 , . . . , zn ). On the Chern curvature tensor can be written (12.5) i(E) = i
1 j,k n, 1 , r
cjk dzj dz k e e
for some coecients cjk C. The hermitian property of i(E) means that cjk = ckj . (12.6) Special case. When r = rank E = 1, the hermitian matrix H is a positive function which we write H = e , (, R). By the above formulas we get (12.7) D d d = e d (e ), i(E) = id d on .
(12.8)
Especially, we see that i (E) is a closed real (1,1)-form on X.
270
(12.9) Remark. In general, it is not possible to nd local frames (e1 , . . . , er ) of E that are simultaneously holomorphic and orthonormal. In fact, we have in this case H = ( ), so a necessary condition for the existence of such a frame is that (E) = 0 on . Conversely, if (E) = 0, Prop. 6.7 and Rem. 7.5 show that E possesses local orthonormal parallel frames (e ) ; we have in particular D e = 0, so (e ) is holomorphic; such a bundle E arising from a unitary representation of 1 (X) is said to be hermitian at. The next proposition shows in a more local way that the Chern curvature tensor is the obstruction to the existence of orthonormal holomorphic frames: a holomorphic frame can be made almost orthonormal only up to curvature terms of order 2 in a neighborhood of any point. (12.10) Proposition. For every point x0 X and every coordinate system (zj )1 j at x0 , there exists a holomorphic frame (e )1 r in a neighborhood of x0 such that e (z), e (z) = cjk zj z k + O(|z|3 )
1 j,k n n
where (cjk ) are the coecients of the Chern curvature tensor (E)x0 . Such a frame (e ) is called a normal coordinate frame at x0 . Proof. Let (h ) be a holomorphic frame of E. After replacing (h ) by suitable linear combinations with constant coecients, we may assume that h (x0 ) is an orthonormal basis of Ex0 . Then the inner products h , h have an expansion h (z), h (z) = +
j
(aj zj + a z j ) + O(|z|2 ) j such that a = aj . Set rst j aj zj h (z).

j,
for some complex coecients
aj , a j
g (z) = h (z)
Then there are coecients ajk , a jk , ajk such that
g (z), g (z) = + O(|z|2 ) = +

j,k 3 ajk zj z k + a jk zj zk + ajk z j z k + O(|z| ).
The holomorphic frame (e ) we are looking for is e (z) = g (z)

Since a jk = ajk , we easily nd
a jk zj zk g (z).
j,k,
e (z), e (z) = +
j,k
ajk zj z k + O(|z|3 ), ajk z k dzj + O(|z|2 ),

j,k
d e , e = {D e , e } = (E) e = D (D e ) = therefore cjk = ajk .
j,k,
ajk dz k dzj e + O(|z|),
13. Lelong-Poincar Equation and First Chern Class e
271
13. Lelong-Poincar Equation and First Chern Class e

Our goal here is to extend the Lelong-Poincar equation III-2.15 to any meromorphic e section of a holomorphic line bundle. (13.1) Denition. A meromorphic section of a bundle E X is a section s dened on an open dense subset of X, such that for every trivialization : EV V Cr the components of = (s) are meromorphic functions on V . Let E be a hermitian line bundle, s a meromorphic section which does not vanish on any component of X and = (s) the corresponding meromorphic function in a trivialization : E C. The divisor of s is the current on X dened by div s = div for all trivializing open sets . One can write div s = mj Zj , where the sets Zj are the irreducible components of the sets of zeroes and poles of s (cf. II-5). The Lelong-Poincar equation (II-5.32) gives e i d d log || = mj [Zj ],
and from the equalities |s|2 = ||2 e and d d = (E) we get (13.2) id d log |s|2 = 2 mj [Zj ] i (E).
This equality can be viewed as a complex analogue of (9.9) (except that here the hypersurfaces Zj are not necessarily smooth). In particular, if s is a non vanishing holomorphic section of E , we have (13.3) i (E) = id d log |s|2 on .
(13.4) Theorem. Let E X be a line bundle and let s be a meromorphic section of E which does not vanish identically on any component of X. If mj Zj is the divisor of s, then c1 (E)R = mj [Zj ] H 2 (X, R). Proof. Apply Formula (13.2) and Th. 9.5, and observe that the bidimension (1, 1)-current id d log |s|2 = d id log |s|2 has zero cohomology class. (13.5) Example. If = mj Zj is an arbitrary divisor on X, we associate to the sheaf () of germs of meromorphic functions f such that div(f ) + 0. Let (V ) be a covering of X and u a meromorphic function on V such that div(u ) = on V . Then ()V = u1 , thus () is a locally free -module of rank 1. This sheaf can be identied to the line bundle E over X dened by the cocycle g := u /u (V V ). In fact, there is a sheaf isomorphism () (E) dened by
()() f s (E)()
with (s) = f u on V .
The constant meromorphic function f = 1 induces a meromorphic section s of E such that div s = div u = ; in the special case when 0, the section s is holomorphic and its zero set s1 (0) is the support of . By Th. 13.4, we have (13.6) c1
() R = {[]}.
272
Let us consider the exact sequence 1 Div 0 already described in (II-5.36). There is a corresponding cohomology exact sequence (13.7)
(X) Div(X)
H 1 (X, ).
0
The connecting homomorphism 0 is equal to the map isomorphism class of
()
dened above. The kernel of this map consists of divisors which are divisors of global meromorphic functions in (X). In particular, two divisors 1 and 2 give rise to isomorphic line bundles (1 ) (2 ) if and only if 2 1 = div(f ) for some global meromorphic function f (X) ; such divisors are called linearly equivalent. The image of 0 consists of classes of line bundles E such that E has a global meromorphic section which does not vanish on any component of X. Indeed, if s is such a section and = div s, there is an isomorphism (13.8)
()
(E),
f f s.
The last result of this section is a characterization of 2-forms on X which can be written as the curvature form of a hermitian holomorphic line bundle. (13.9) Theorem. Let X be an arbitrary complex manifold.
i a) For any hermitian line bundle E over M , the Chern curvature form 2 (E) is a closed real (1, 1)-form whose De Rham cohomology class is the image of an integral class. 2 b) Conversely, let be a closed real (1, 1)-form such that {} HDR (X, R) is the image of an integral class. Then there exists a hermitian line bundle E X such i that 2 (E) = .
Proof. a) is an immediate consequence of Formula (12.9) and Th. 9.5, so we have only to prove the converse part b). By Prop. III-1.20, there exist an open covering (V ) of X and i functions (V , R) such that 2 d d = on V . It follows that the function is pluriharmonic on V V . If (V ) is chosen such that the intersections V V are simply connected, then Th. I-3.35 yields holomorphic functions f on V V such that 2 Re f = on V V . Now, our aim is to prove (roughly speaking) that exp(f ) is a cocycle in that denes the line bundle E we are looking for. The Cech dierential (f ) = f f + f takes values in the constant sheaf iR because 2 Re (f ) = ( ) ( ) + ( ) = 0. Consider the real 1-forms A = f ) = df , we get
i 4 (d
d ). As d ( ) is equal to d (f +
(A) = A A =
1 i d(f f ) = d Im f . 4 2
14. Exact Sequences of Hermitian Vector Bundles
273
1 Since = dA , it follows by (9.6) and (9.7) that the Cech cohomology class {( 2 Im f )} is equal to {} H 2 (X, R), which is by hypothesis the image of a 2-cocycle (n ) H 2 (X, Z). Thus we can write
1 Im f = (n ) + (c ) 2
for some 1-chain (c ) with values in R. If we replace f by f 2ic , then we can achieve c = 0, so (f ) 2iZ and g := exp(f ) will be a cocycle with values in . Since = 2 Re f = log |g |2 , the line bundle E associated to this cocycle admits a global hermitian metric dened in every trivialization by the matrix H = (exp( )) and therefore i i (E) = d d = 2 2 on V .

Let us consider an exact sequence of holomorphic vector bundles over X : (14.1) 0 S E Q 0.
j g
Then E is said to be an extension of S by Q. A (holomorphic, resp. ) splitting of the exact sequence is a (holomorphic, resp. ) homomorphism h : Q E which is a right inverse of the projection E Q, i.e. such that g h = IdQ . Assume that a hermitian metric on E is given. Then S and Q can be endowed with the induced and quotient metrics respectively. Let us denote by DE , DS , DQ the corresponding Chern connections. The adjoint homomorphisms j : E S, g : Q E
are and can be described respectively as the orthogonal projection of E onto S and as the orthogonal splitting of the exact sequence (14.1). They yield a (in general non analytic) isomorphism (14.2) j g : E S Q.

(14.3) Theorem. According to the
isomorphism j g, DE can be written DS DQ
DE =
where X, Hom(S, Q) is called the second fundamental of S in E and where 1,0 X, Hom(Q, S) is the adjoint of . Furthermore, the following identities hold: 0,1
274

DHom(S,E) j = g , DHom(E,Q) g = j , DHom(E,S) j = 0, DHom(Q,E) g = 0,
a) b) c) d)
d j = 0 ; d g = 0 ; d j = g ; d g = j .
Proof. If we dene E DS DQ via (14.2), then E is a hermitian connection on E. By (7.3), we have therefore DE = E + , where (X, Hom(E, E)) and 1 = . Let us write = , = , = , = , DS + DQ +
(14.4) For any section u
DE =
, (X, E)
we have
DE u = DE (jj u+g gu) = jDS (j u)+g DQ (gu)+(DHom(S,E) j)j u+(DHom(E,Q) g )gu. A comparison with (14.4) yields DHom(S,E) j = j + g ,
DHom(E,Q) g = j + g ,
Since j is holomorphic, we have d j = j 0,1 + g 0,1 = 0, thus 0,1 = 0,1 = 0. But = , hence = 0 and (Hom(S, Q)) ; identity a) follows. Similarly, we get 1,0 DS (j u) = j DE u + (DHom(E,S) j ) u, DQ (gu) = gDE u + (DHom(E,Q) g) u, and comparison with (14.4) yields DHom(E,S) j = j g = g, DHom(E,Q) g = j g. Since d g = 0, we get 0,1 = 0, hence = 0. Identities b), c), d) follow from the above computations. (14.5) Theorem. We have d ( ) = 0, and the Chern curvature of E is (E) = (S) d
DHom(Q,S) (Q)
2 Proof. A computation of DE yields 2 DE = 2 DS DS + DQ
(DS + DQ ) 2 DQ
275
Formula (13.4) implies DHom(S,Q) = DS + DQ ,
DHom(Q,S) = DS + DQ .
2 Since DE is of type (1,1), it follows that d = DHom(Q,S) = 0. The proof is achieved.
A consequence of Th. 14.5 is that (S) and (Q) are given in terms of (E) by the following formulas, where (E)S , (E)Q denote the blocks in the matrix of (E) corresponding to Hom(S, S) and Hom(Q, Q): (14.6) (14.7) (S) = (E)S + ,
(Q) = (E)Q + .
By 14.3 c) the second fundamental form vanishes identically if and only if the orthogonal splitting E S Q is holomorphic ; then we have (E) = (S) (Q). Next, we show that the d -cohomology class { }H 0,1 X, Hom(Q, S) characterizes the isomorphism class of E among all extensions of S by Q. Two extensions E and F are said to be isomorphic if there is a commutative diagram of holomorphic maps 0 S E Q 0
(14.8)
0 S F Q 0
in which the rows are exact sequences. The central vertical arrow is then necessarily an isomorphism. It is easily seen that 0 S E Q 0 has a holomorphic splitting if and only if E is isomorphic to the trivial extension S Q. (14.9) Proposition. The correspondence {E} { } induces a bijection from the set of isomorphism classes of extensions of S by Q onto the cohomology group H 1 X, Hom(Q, S) . In particular { } vanishes if and only if the exact sequence 0 S E Q 0 splits holomorphically. Proof. a) The map is well dened, i.e. { } does not depend on the choice of the hermitian metric on E. Indeed, a new hermitian metric produces a new splitting g and a new form such that d g = j . Then gg = gg = IdQ , thus g g = j v for some section v X, Hom(Q, S) . It follows that = d v. Moreover, it is clear that an isomorphic extension F has the same associated form if F is endowed with the image of the hermitian metric of E.
j g
276
b) The map is injective. Let E and F be extensions of S by Q. Select splittings E, F S Q. We endow S, Q with arbitrary hermitian metrics and E, F with the direct sum metric. Then we have corresponding (0, 1)-connections
DE = DS 0
DQ
DF =
DS 0
DQ
Assume that = + d v for some v E F of class dened by the matrix IdS 0
X, Hom(Q, S) . The isomorphism :
v IdQ
is then holomorphic, because the relation DS v v DQ = d v = implies DHom(E,F ) = DF DE
= =
DS 0
DQ
IdS 0
v IdQ
IdS 0
v IdQ = 0.
DS 0
DQ
0 0
+ + (DS v v DQ ) 0
Hence the extensions E and F are isomorphic. c) The map is surjective. Let be an arbitrary d -closed (0, 1)-form on X with values in Hom(Q, S). We dene E as the hermitian vector bundle S Q endowed with the (0, 1)-connection DS . DE = 0 DQ We only have to show that this connection is induced by a holomorphic structure on E ; then we will have = . However, the Dolbeault-Grothendieck lemma implies that there is a covering of X by open sets U on which = d v for some v U , Hom(Q, S) . Part b) above shows that the matrix IdS 0 v IdQ
denes an isomorphism from EU onto the trivial extension (S Q)U such that DHom(E,SQ) = 0. The required holomorphic structure on EU is the inverse image of the holomorphic structure of (S Q)U by ; it is independent of because v v 1 and are holomorphic on U U . (14.10) Remark. If E and F are extensions of S by Q such that the corresponding forms and = u v 1 dier by u H 0 X, Aut(S) , v H 0 X, Aut(Q) , it is easy to see that the bundles E and F are isomorphic. To see this, we need only replace the vertical arrows representing the identity maps of S and Q in (14.8) by u and v respectively. Thus, if we want to classify isomorphism classes of bundles E which are extensions of S by Q rather than the extensions themselves, the
15. Line Bundles
(k) over Pn
277
set of classes is the quotient of H 1 X, Hom(Q, S) by the action of H 0 X, Aut(S) H 0 X, Aut(Q) . In particular, if S, Q are line bundles and if X is compact connected, then H 0 X, Aut(S) , H 0 X, Aut(Q) are equal to C and the set of classes is the projective space P H 1 (X, Hom(Q, S)) .
15. Line Bundles (k) over Pn

15.A. Algebraic properties of
(k)
Let V be a complex vector space of dimension n + 1, n 1. The quotient topological space P (V ) = (V {0})/C is called the projective space of V , and can be considered as the set of lines in V if {0} is added to each class C x. Let : V {0} P (V ) x [x] = C x
be the canonical projection. When V = Cn+1 , we simply denote P (V ) = Pn . The space Pn is the quotient S 2n+1 /S 1 of the unit sphere S 2n+1 Cn+1 by the multiplicative action of the unit circle S 1 C, so Pn is compact. Let (e0 , . . . , en ) be a basis of V , and let (x0 , . . . , xn ) be the coordinates of a vector x V {0}. Then (x0 , . . . , xn ) are called the homogeneous coordinates of [x] P (V ). The space P (V ) can be covered by the open sets j dened by j = {[x] P (V ) ; xj = 0} and there are homeomorphisms j : j Cn [x] (z0 , . . . , zj , . . . , zn ), zl = xl /xj for l = j.
The collection (j ) denes a holomorphic atlas on P (V ), thus P (V ) = Pn is a compact n-dimensional complex analytic manifold.
V Let be the trivial bundle P (V ) V . We denote by (1) the tautological line V subbundle (15.1)
(1) =
([x], ) P (V ) V ; C x {0}. Then
such that (1)[x] = C x V , x V holomorphic section
(1)
admits a non vanishing
[x] j ([x]) = x/xj = z0 e0 + . . . + ej + zj+1 ej+1 + . . . + zn en , and this shows in particular that (1) is a holomorphic line bundle. (15.2) Denition. For every k Z, the line bundle
(k) is dened by
k 1,
(1) = (1), (0) = P (V ) C, (k) = (1)k = (1) (1) for (k) = (1)k for k 1
278
We also introduce the quotient vector bundle H = /(1) of rank n. Therefore we V have canonical exact sequences of vector bundles over P (V ) : (15.3) 0 (1) H 0, V 0 H (1) 0. V
The total manifold of the line bundle transformation, or Hopf -process:
(1)
gives rise to the so called monoidal
(15.4) Lemma. The holomorphic map : (1) V dened by V : (1) = P (V ) V V

pr2
sends the zero section P (V ){0} of (1) to the point {0} and induces a biholomorphism of (1) P (V ) {0} onto V {0}. Proof. The inverse map 1 : V {0} (1) is clearly dened by 1 : x [x], x . The space H 0 (Pn , (k)) of global holomorphic sections of (k) can be easily computed by means of the above map . (15.5) Theorem. H 0 P (V ), (k) = 0 for k < 0, and there is a canonical isomorphism H 0 P (V ), (k) S k V , where S k V denotes the k-th symmetric power of V . (15.6) Corollary. We have dim H 0 Pn , (k) = for k < 0. Proof. Assume rst that k
n+k n
0,
for k
0, and this group is 0
0. There exists a canonical morphism : S k V H 0 P (V ), (k) ;
indeed, any element a S k V denes a homogeneous polynomial of degree k on V V and thus by restriction to (1) a section (a) = a of ((1) )k = (k) ; in other words is induced by the k-th symmetric power S k (k) of the canonical V morphism (1) in (15.3). V Assume now that k Z is arbitrary and that s is a holomorphic section of (k). For every x V {0} we have s([x]) (k)[x] and 1 (x) (1)[x] . We can therefore associate to s a holomorphic function on V {0} dened by f (x) = s([x]) 1 (x)k , xV {0}.
Since dim V = n + 1 2, f can be extended to a holomorphic function on V and f is clearly homogeneous of degree k ( and 1 are homogeneous of degree 1). It follows
15. Line Bundles
(k) over Pn
279
that f = 0, s = 0 if k < 0 and that f is a homogeneous polynomial of degree k on V if k 0. Thus, there exists a unique element a S k V such that f (x) = a xk = a([x]) 1 (x)k . Therefore is an isomorphism. The tangent bundle on Pn is closely related to the bundles H and the following proposition. (15.7) Proposition. There is a canonical isomorphism of bundles T P (V ) H (1). Proof. The dierential dx of the projection : V {0} P (V ) may be considered as a map dx : V T[x] P (V ). As dx (x) = 0, dx can be factorized through V /C x = V /(1)[x] = H[x] . Hence we get an isomorphism dx : H[x] T[x] P (V ),
(1) as shown by
depends only on [x]. Therefore H T P (V ) (1).
but this isomorphism depends on x and not only on the base point [x] in P (V ). The formula (x + ) = (x + 1 ), C , V , shows that dx = 1 dx , hence the map dx 1 (x) : H[x] T P (V ) (1) [x]
15.B. Curvature of the Tautological Line Bundle Assume now that V is a hermitian vector space. Then (15.3) yields exact sequences of hermitian vector bundles. We shall compute the curvature of (1) and H. Let a P (V ) be xed. Choose an orthonormal basis (e0 , e1 , . . . , en ) of V such that a = [e0 ]. Consider the embedding Cn P (V ), 0 a
which sends z = (z1 , . . . , zn ) to [e0 + z1 e1 + + zn en ]. Then (z) = e0 + z1 e1 + + zn en denes a non-zero holomorphic section of (1) implies (15.8) (15.8 )
(1)C
and Formula (13.3) for
(1)
(1) = d d log |(z)|2 = d d log(1 + |z|2 ) (1) a = dzj dz j .

1 j n
on Cn ,
280
On the other hand, Th. 14.3 and (14.7) imply d g = j , (H) = ,
V V where j : (1) is the inclusion, g : H the orthogonal splitting and V 0,1 P (V ), Hom(H, (1)) . The images (e1 , . . . , en ) of e1 , . . . , en in H = /(1) n and we have dene a holomorphic frame of HC g ej = ej
a = 1 j n
zj ej , = ej , 2 || 1 + |z|2 dz j e , j a =
d ga ej = dz j ,
1 j n
dzj ej ,
(15.9)
(H)a =
1 j,k n
dzj dz k e ej . k
(15.10) Theorem. The cohomology algebra H (Pn , Z) is isomorphic to the quotient ring Z[h]/(hn+1 ) where the generator h is given by h = c1 ((1)) in H 2 (Pn , Z). Proof. Consider the inclusion Pn1 = P (Cn {0}) Pn . Topologically, Pn is obtained from Pn1 by attaching a 2n-cell B2n to Pn1 , via the map f : B2n Pn
z [z, 1 |z|2 ],
z Cn , |z|
which sends S 2n1 = {|z| = 1} onto Pn1 . That is, Pn is homeomorphic to the quotient space of B2n Pn1 , where every point z S 2n1 is identied with its image f (z) Pn1 . We shall prove by induction on n that (15.11) H 2k (Pn , Z) = Z, 0 k n, otherwise H l (Pn , Z) = 0.
The result is clear for P0 , which is reduced to a single point. For n 1, consider n the covering (U1 , U2 ) of P such that U1 is the image by f of the open ball B2n and n U2 = P {f (0)}. Then U1 B2n is contractible, whereas U2 = (B2n {0})S 2n1 Pn1 . Moreover U1 U2 B2n {0} can be retracted on the (2n 1)-sphere of radius 1/2. For q 2, the Mayer-Vietoris exact sequence IV-3.11 yields H q1 (Pn1 , Z) H q1 (S 2n1 , Z)
H q (Pn , Z) H q (Pn1 , Z) H q (S 2n1 , Z) . For q = 1, the rst term has to be replaced by H 0 (Pn1 , Z) Z, so that the rst arrow is onto. Formula (15.11) follows easily by induction, thanks to our computation of the cohomology groups of spheres in IV-14.6.
We know that h = c1 ((1)) H 2 (Pn , Z). It will follow necessarily that hk is a generator of H 2k (Pn , Z) if we can prove that hn is the fundamental class in H 2n (P, Z), or equivalently that (15.12) c1
(1) n = R
Pn
i ((1)) 2
= 1.
15. Line Bundles
(k) over Pn
281
This equality can be veried directly by means of (15.8), but we will avoid this com putation. Observe that the element e Cn+1 denes a section e of H 0 (Pn , (1)) n n i transverse to 0, whose zero set is the hyperplane Pn1 . As { 2 ((1))} = {[Pn1 ]} by Th. 13.4, we get c1 ((1)) = c1 ((1))n = [P0 ] = 1
P1
for n = 1 and i ((1)) 2

n1
Pn
[Pn1 ]
=
Pn1
i ((1)) 2
n1
in general. Since (1)Pn1 can be identied with the tautological line subbundle Pn1 (1) over Pn1 , we have ((1))Pn1 = (Pn1 (1)) and the proof is achieved by induction on n. 15.C. Tautological Line Bundle Associated to a Vector Bundle Let E be a holomorphic vector bundle of rank r over a complex manifold X. The projectivized bundle P (E) is the bundle with Pr1 bers over X dened by P (E)x = P (Ex ) for all x X. The points of P (E) can thus be identied with the lines in the bers of E. For any trivialization : EU U Cr of E we have a corresponding trivialization : P (E)U U Pr1 , and it is clear that the transition automorphisms are the projectivizations g H 0 U U , P GL(r, C) of the transition automorphisms g of E. Similarly, we have a dual projectivized bundle P (E ) whose points can be identied with the hyperplanes of E (every hyperplane F in Ex corresponds bijectively to the line of linear forms in Ex which vanish on F ); note that P (E) and P (E ) coincide only when r = rk E = 2. If : P (E ) X is the natural projection, there is a tautological hyperplane subbundle S of E over P (E ) such that S[] = 1 (0) Ex for all Ex {0}. exercise: check that S is actually locally trivial over P (E ) . (15.13) Denition. The quotient line bundle E/S is denoted E (1) and is called the tautological line bundle associated to E. Hence there is an exact sequence 0 S E E (1) 0 of vector bundles over P (E ).
Note that (13.3) applied with V = Ex implies that the restriction of E (1) to each r1 ber P (Ex ) P coincides with the line bundle (1) introduced in Def. 15.2. Theorem 15.5 can then be extended to the present situation and yields:
Proof. For k 0, the k-th symmetric power of the morphism E E (1) gives a morphism S k E E (k). This morphism together with the pull-back morphism yield canonical arrows U : H 0 (U, S k E) H 0 1 (U ), S k E H 0 1 (U ), E (k)
(15.14) Theorem. For every k Z, the direct image sheaf E (k) on X vanishes for k < 0 and is isomorphic to (S k E) for k 0.
282
for any open set U X. The right hand side is by denition the space of sections of E (k) over U , hence we get a canonical sheaf morphism : (S k E) E (k). It is easy to check that this coincides with the map introduced in the proof of Cor. 15.6 when X is reduced to a point. In order to check that is an isomorphism, we may suppose that U is chosen so small that EU is trivial, say EU = U V with dim V = r. Then P (E ) = U P (V ) and E (1) = p (1) where (1) is the tautological line bundle over P (V ) and p : P (E ) P (V ) is the second projection. Hence we get H 0 1 (U ), E (k) = H 0 U P (V ), p (1) = X (U ) H 0 P (V ), (1)
as desired; the reason for the second equality is that p (1) coincides with (1) on each ber {x} P (V ) of p, thus any section of p (1) over U P (V ) yields a family of sections H 0 {x} P (V ), (k) depending holomorphically in x. When k < 0 there are no non zero such sections, thus E (k) = 0. Finally, suppose that E is equipped with a hermitian metric. Then the morphism E E (1) endows E (1) with a quotient metric. We are going to compute the associated curvature form E (1) .
Fix a point x0 X and a P (Ex0 ). Then Prop. 12.10 implies the existence of a normal coordinate frame (e )1 r ) of E at x0 such that a is the hyperplane e2 , . . . , er = (e )1 (0) at x0 . Let (z1 , . . . , zn ) be local coordinates on X near x0 and 1 let (1 , . . . , r ) be coordinates on E with respect to the dual frame (e , . . . , e ). If we 1 r assign 1 = 1, then (z1 , . . . , zn , 2 , . . . , r ) dene local coordinates on P (E ) near a, and we have a local section of E (1) := E (1) E dened by
= X (U ) S k V = H 0 (U, S k E),
(z, ) = e (z) + 1
2 r
e (z).
The hermitian matrix ( e , e ) is just the congugate of the inverse of matrix ( e , e ) = Id cjk zj z k + O(|z|3 ), hence we get e (z), e (z) = +
1 j,k n
cjk zj z k + O(|z|3 ),
where (cjk ) are the curvature coecients of (E) ; accordingly we have (E ) = (E) . We infer from this |(z, )|2 = 1 + Since cjk11 zj z k +
1 j,k n 2 r
| |2 + O(|z|3 ).
E (1)
= d d log |(z, )|2 , we get
E (1) a =
1 j,k n
cjk11 dzj dz k +
2 r
d d .
16. Grassmannians and Universal Vector Bundles
283
Note that the rst summation is simply (E )a, a /|a|2 = curvature of E in the direction a. A unitary change of variables then gives the slightly more general formula: (15.15) Formula. Let (e ) be a normal coordinate frame of E at x0 X and let (E)x0 = cjk dzj dz k e e . At any point a P (E ) represented by a vector a e Ex0 of norm 1, the curvature of E (1) is
E (1) a =
1 j,k n, 1 , r
cjk a a dzj dz k +
1 r1
d d ,
where ( ) are coordinates near a on P (E ), induced by unitary coordinates on the hy perplane a Ex0 .

16.A. Universal Subbundles and Quotient Vector Bundles If V is a complex vector space of dimension d, we denote by Gr (V ) the set of all r-codimensional vector subspaces of V . Let a Gr (V ) and W V be xed such that V = a W, dimC W = r. Then any subspace x Gr (V ) in the open subset W = {x Gr (V ) ; x W = V } can be represented in a unique way as the graph of a linear map u in Hom(a, W ). This gives rise to a covering of Gr (V ) by ane coordinate charts W Hom(a, W ) Cr(dr) . Indeed, let (e1 , . . . , er ) and (er+1 , . . . , en ) be respective bases of W and a. Every point x W is the graph of a linear map (16.1) u : a W, u(ek ) =
1 j r
zjk ej ,
r+1
d,
i.e. x = Vect ek + 1 j r zjk ej r+1 k d . We choose (zjk ) as complex coordinates on W . These coordinates are centered at a = Vect(er+1 , . . . , ed ). (16.2) Proposition. Gr (V ) is a compact complex analytic manifold of dimension n = r(d r). Proof. It is immediate to verify that the coordinate change between two ane charts of Gr (V ) is holomorphic. Fix an arbitrary hermitian metric on V . Then the unitary group U (V ) is compact and acts transitively on Gr (V ). The isotropy subgroup of a point a Gr (V ) is U (a) U (a ), hence Gr (V ) is dieomorphic to the compact quotient space U (V )/U (a) U (a ). Next, we consider the tautological subbundle S := Gr (V ) V dened by Sx = x V for all x Gr (V ), and the quotient bundle Q = /S of rank r : V (16.3) 0 S Q 0. V
284
An interesting special case is r = d 1, Gd1 (V ) = P (V ), S = (1), Q = H. The case r = 1 is dual, we have the identication G1 (V ) = P (V ) because every hyperplane x V corresponds bijectively to the line in V of linear forms V that vanish on x. V Then the bundles (1) and H on P (V ) are given by
(1)[] = C. (V /x) = Q , x
H[] = V /C. x = Sx ,
therefore S = H , Q = (1). This special case will allow us to compute H 0 (Gr (V ), Q) in general. (16.4) Proposition. There is an isomorphism V = H 0 Gr (V ), V H 0 Gr (V ), Q .
Proof. Let V = W W be an arbitrary direct sum decomposition of V with codim W = r 1. Consider the projective space P (W ) = G1 (W ) Gr (V ), its tautological hyperplane subbundle H = P (W ) W and the exact sequence W 0 H (1) 0. Then SP (W ) coincides with H and W QP (W ) = ( )/H = ( W W W/H ) = (1) . W W Theorem 15.5 implies H 0 (P (W ), (1)) = W , therefore the space H 0 (P (W ), QP (W ) ) = W W is generated by the images of the constant sections of . Since W is arbitrary, Prop. 16.4 V follows immediately. Let us compute the tangent space T Gr (V ). The linear group Gl(V ) acts transitively on Gr (V ), and the tangent space to the isotropy subgroup of a point x Gr (V ) is the set of elements u Hom(V, V ) in the Lie algebra such that u(x) x. We get therefore Tx Gr (V ) Hom(V, V )/{u ; u(x) x}
Hom(V, V /x)/ u ; u(x) = {0} Hom(x, V /x) = Hom(Sx , Qx ).
(16.5) Corollary. T Gr (V ) = Hom(S, Q) = S Q.
285
16.B. Pl cker Embedding u There is a natural map, called the Plcker embedding, u (16.6) jr : Gr (V ) P (r V )
constructed as follows. If x Gr (V ) is dened by r independent linear forms 1 , . . . , r V , we set jr (x) = [1 r ]. Then x is the subspace of vectors v V such that v (1 r ) = 0, so jr is injective. Since the linear group Gl(V ) acts transitively on Gr (V ), the rank of the dierential djr is a constant. As jr is injective, the constant rank theorem implies: (16.7) Proposition. The map jr is a holomorphic embedding. Now, we dene a commutative diagram r Q Gr (V ) Jr jr
(16.8)
(1)
P (r V )
1 1 as follows: for x = 1 (0) r (0) Gr (V ) and v = v1 vr r Qx where vk Qx = V /x is the image of vk V in the quotient, we let Jr (v) (1)jr (x) be the linear form on (1)jr (x) = C.1 . . . r such that
Jr (v), 1 . . . r = det j (vk ) ,
C.
Then Jr is an isomorphism on the bers, so r Q can be identied with the pull-back of (1) by jr . 16.C. Curvature of the Universal Vector Bundles Assume now that V is a hermitian vector space. We shall generalize our curvature computations of 15.C to the present situation. Let a Gr (V ) be a given point. We take W to be the orthogonal complement of a in V and select an orthonormal basis (e1 , . . . , ed ) of V such that W = Vect(e1 , . . . , er ), a = Vect(er+1 , . . . , ed ). For any point x Gr (V ) in W with coordinates (zjk ), we set k (x) = ek +
1 j r
zjk ej ,
r+1
k 1
d, j r.
ej (x) = image of ej in Qx = V /x,
Then (e1 , . . . , er ) and (r+1 , . . . , d ) are holomorphic frames of Q and S respectively. If g : Q is the orthogonal splitting of g : Q, then V V g ej = ej + jk k
r+1 k d
286
for some jk C. After an easy computation we nd 0 = ej , gk = g ej , k = jk + z jk +

l,m
jm zlm z lk ,
so that jk = z jk + O(|z|2 ). Formula (13.3) yields

d ga ej = a = j,k r+1 k d
dz jk k , a =
j,k
dz jk e k , j
dzjk ej , k
(16.9) (16.10)
(Q)a = ( )a =
j,k,l
dzjk dz lk e ej , l dzjk dz jl l . k
(S)a = ( )a =
j,k,l
287
Chapter VI
Hodge Theory
The goal of this chapter is to prove a number of basic facts in the Hodge theory of real or complex manifolds. The theory rests essentially on the fact that the De Rham (or Dolbeault) cohomology groups of a compact manifold can be represented by means of spaces of harmonic forms, once a Riemannian metric has been chosen. At this point, some knowledge of basic results about elliptic dierential operators is required. The special properties of compact Khler manifolds are then investigated in detail: Hodge a decomposition theorem, hard Lefschetz theorem, Jacobian and Albanese variety, . . . ; the example of curves is treated in detail. Finally, the Hodge-Frlicher spectral sequence is applied to get some results o on general compact complex manifolds, and it is shown that Hodge decomposition still holds for manifolds in the Fujiki class ( ).
1. Dierential Operators on Vector Bundles

We rst describe some basic concepts concerning dierential operators (symbol, composition, adjunction, ellipticity), in the general setting of vector bundles. Let M be a dierentiable manifold, dimR M = m, and let E, F be K-vector bundles over M , with K = R or K = C, rank E = r, rank F = r . (1.1) Denition. A (linear) dierential operator of degree from E to F is a K-linear operator P : (M, E) (M, F ), u P u of the form P u(x) =
||
a (x)D u(x),
where E Kr , F Kr are trivialized locally on some open chart M equipped with local coordinates (x1 , . . . , xm ), and where a (x) = a (x) 1 r , 1 r are r r-matrices with coecients on . Here D = (/x1 )1 . . . (/xm )m as usual, and u = (u )1 r , D u = (D u )1 r are viewed as column matrices. If t K is a parameter, a simple calculation shows that etu(x) P (etu(x) ) is a polynomial of degree in t, of the form etu(x) P (etu(x) ) = t P (x, du(x)) + lower order terms cj (x)tj , j < ,
where P is the polynomial map from TM Hom(E, F ) dened by
(1.2)
TM,x P (x, ) Hom(Ex , Fx ),
P (x, ) =
||=
a (x) .
288
Chapter VI. Hodge Theory
The formula involving etu P (etu ) shows that P (x, ) actually does not depend on the choice of coordinates nor on the trivializations used for E, F . It is clear that P (x, ) is smooth on TM as a function of (x, ), and is a homogeneous polynomial of degree in . We say that P is the principal symbol of P . Now, if E, F , G are vector bundles and P :
(M, E)
(M, F ),
Q:
(M, F )
(M, G)
are dierential operators of respective degrees P , Q , it is easy to check that Q P : (M, E) (M, G) is a dierential operator of degree P + Q and that (1.3) QP (x, ) = Q (x, )P (x, ). Here the product of symbols is computed as a product of matrices. Now, assume that M is oriented and is equipped with a smooth volume form dV (x) = (x)dx1 . . . dxm , where (x) > 0 is a smooth density. If E is a euclidean or hermitian vector bundle, we have a Hilbert space L2 (M, E) of global sections u of E with measurable coecients, satisfying the L2 estimate (1.4) We denote by (1.4 ) u, v =
M
=
M
|u(x)|2 dV (x) < +.
u(x), v(x) dV (x),
u, v L2 (M, E)
the corresponding L2 inner product. (1.5) Denition. If P : (M, E) (M, F ) is a dierential operator and both E, F are euclidean or hermitian, there exists a unique dierential operator P :
(M, F )
(M, E),
called the formal adjoint of P , such that for all sections u (M, F ) there is an identity P u, v = u, P v ,
(M, E) and v
whenever Supp u Supp v M .
Proof. The uniqueness is easy, using the density of the set of elements u (M, E) with compact support in L2 (M, E). Since uniqueness is clear, it is enough, by a partition of unity argument, to show the existence of P locally. Now, let P u(x) = || a (x)D u(x) be the expansion of P with respect to trivializations of E, F given by orthonormal frames over some coordinate open set M . Assuming Supp u Supp v , an integration by parts yields P u, v =
|| ,,
a D u (x)v (x) (x) dx1, . . . , dxm (1)|| u (x)D ((x) a v (x) dx1 , . . . , dxm
|| ,,
= =
u,
||
(1)|| (x)1 D (x) ta v(x)
dV (x).
2. Formalism of PseudoDierential Operators
289
Hence we see that P exists and is uniquely dened by (1.6) P v(x) =

||
(1)|| (x)1 D (x) t a v(x) .
It follows immediately from (1.6) that the principal symbol of P is (1.7) P (x, ) = (1)
||= t
a = (1) P (x, ) .
(1.8) Denition. A dierential operator P is said to be elliptic if P (x, ) Hom(Ex , Fx )

is injective for every x M and TM,x
{0}.
2. Formalism of PseudoDierential Operators

We assume throughout this section that (M, g) is a compact Riemannian manifold. For any positive integer k and any hermitian bundle F M , we denote by W k (M, F ) the Sobolev space of sections s : M F whose derivatives up to order k are in L2 . k Let k be the norm of the Hilbert space W (M, F ). Let P be an elliptic dierential operator of order d acting on (M, F ). We need the following basic facts of elliptic P DE theory, see e.g. [Hrmander 1963]. o (2.1) Sobolev lemma. For k > l +
m 2,
W k (M, F ) C l (M, F ).
(2.2) Rellich lemma. For every integer k, the inclusion W k+1 (M, F ) W k (M, F ) is a compact linear operator. (2.3) G ardings inequality. Let P be the extension of P to sections with distribution coecients. For any u W 0 (M, F ) such that P u W k (M, F ), then u W k+d (M, F ) and u k+d Ck ( P u k + u 0 ), where Ck is a positive constant depending only on k. (2.4) Corollary. The operator P : ties: i) ker P is nite dimensional. ii) P (M, F ) is closed and of nite codimension; furthermore, if P is the formal adjoint of P , there is a decomposition

(M, F )
(M, F ) has the following proper-
(M, F ) = P
(M, F ) ker P
290
as an orthogonal direct sum in W 0 (M, F ) = L2 (M, F ). Proof. (i) G ardings inequality shows that u k+d Ck u 0 for any u in ker P . Thanks to the Sobolev lemma, this implies that ker P is closed in W 0 (M, F ). Moreover, the unit closed 0 -ball of ker P is contained in the d -ball of radius C0 , thus compact by the Rellich lemma. Riesz theorem implies that dim ker P < +. (ii) We rst show that the extension P : W k+d (M, F ) W k (M, F ) has a closed range for any k. For every > 0, there exists a nite number of elements v1 , . . . , vN W k+d (M, F ), N = N (), such that
N
(2.5) indeed the set
k+d
+
j=1
| u, vj
N
0|
K(vj ) = u W
k+d
(M, F ) ; u
k+d
+
j=1
| u, vj
0|
is relatively compact in W 0 (M, F ) and (vj ) K (vj ) = {0}. It follows that there exist elements (vj ) such that K (vj ) is contained in the unit ball of W 0 (M, F ), QED. Substitute ||u||0 by the upper bound (2.5) in G ardings inequality; we get
N
(1 Ck ) u Dene G = u W
k+d
k+d
Ck
Pu j
+
j=1
| u, vj
0|
(M, F ) ; u vj , 1 u
k+d
n} and choose = 1/2Ck . We obtain
2Ck P u k , u G.
This implies that P (G) is closed. Therefore P W k+d (M, F ) = P (G) + Vect P (v1 ), . . . , P (vN ) is closed in W k (M, F ). Take in particular k = 0. Since we see that in W 0 (M, F ) P W d (M, F ) We have proved that (2.6) W 0 (M, F ) = P W d (M, F ) ker P .

(M, F ) is dense in W d (M, F ), = ker P .
= P
(M, F )
Since P is also elliptic, it follows that ker P is nite dimensional and that ker P = ker P is contained in (M, F ). Thanks to G ardings inequality, the decomposition formula (2.6) yields (2.7) (2.8) W k (M, F ) = P W k+d (M, F ) ker P ,
(M, F ) = P
(M, F ) ker P .
3. Harmonic Forms and Hodge Theory on Riemannian Manifolds
291

3.1. Euclidean Structure of the Exterior Algebra Let (M, g) be an oriented Riemannian -manifold, dimR M = m, and E M a hermitian vector bundle of rank r over M . We denote respectively by (1 , . . . , m ) and (e1 , . . . , er ) orthonormal frames of TM and E over an open subset M , and by (1 , . . . , m ), (e , . . . , e ) the corresponding dual frames of TM , E . Let dV stand for the 1 r Riemannian volume form on M . The exterior algebra TM has a natural inner product , such that (3.1) u1 . . . up , v1 . . . vp = det( uj , vk )1
j,k p , uj , vk TM
for all p, with TM = p TM as an orthogonal sum. Then the covectors I = i1 ip , i1 < i2 < < ip , provide an orthonormal basis of TM . We also denote by , the corresponding inner product on TM E.
(3.2) Hodge Star Operator. The Hodge-Poincar-De Rham operator is the collece tion of linear maps dened by
: p TM mp TM ,
u v = u, v dV,
u, v p TM .
The existence and uniqueness of this operator is easily seen by using the duality pairing
p TM mp TM R
(3.3)
(u, v) u v/dV =
(I, I) uI vI ,
where u = |I|=p uI I , v = |J|=mp vJ J , where I stands for the (ordered) complementary multi-index of I and (I, I) for the signature of the permutation (1, 2, . . . , m) (I, I). From this, we nd
(3.4)
v =
|I|=p
(I, I)vI I .
More generally, the sesquilinear pairing {, } dened in (V-7.1) yields an operator on vector valued forms, such that
(3.3 ) : p TM E mp TM E,
(3.4 ) t =
|I|=p,
(I, I) tI, I e
{s, t} = s, t dV,
s, t p TM E,
for t = diately (3.5)
tI, I e . Since (I, I)(I, I) = (1)p(mp) = (1)p(m1) , we get imme-
t = (1)p(m1) t
on p TM E.
292
It is clear that is an isometry of TM E.
We shall also need a variant of the operator, namely the conjugate-linear operator
# : p TM E mp TM E
dened by s # t = s, t dV, where the wedge product is combined with the canonical pairing E E C. We have (3.6) #t =
|I|=p, (I, I) tI, I e .
(3.7) Contraction by a Vector Field.. Given a tangent vector TM and a form u p TM , the contraction u p1 TM is dened by u (1 , . . . , p1 ) = u(, 1 , . . . , p1 ), j TM .
In terms of the basis (j ), l

(i1 . . . ip ) =
is the bilinear operation characterized by 0 (1)k1 i1 . . . ik . . . ip if l {i1 , . . . , ip }, / if l = ik .
This formula is in fact valid even when (j ) is non orthonormal. A rather easy computation shows that is a derivation of the exterior algebra, i.e. that (u v) = ( u) v + (1)deg u u ( v).
Moreover, if = , TM , the operator
is the adjoint map of , that is,

u, v TM .
(3.8)
u, v = u, v ,
Indeed, this property is immediately checked when = l , u = I , v = J .
3.2. Laplace-Beltrami Operators

Let us consider the Hilbert space L2 (M, p TM ) of p-forms u on M with measurable coecients such that u 2= |u|2 dV < +. M
We denote by , the global inner product on L2 -forms. One can dene similarly the Hilbert space L2 (M, p TM E). (3.9) Theorem. The operator d = (1)mp+1 d is the formal adjoint of the exterior derivative d acting on (M, p TM E). Proof. If u
(M, p TM ), v (M, p+1 TM ) are compactly supported we get
du, v =
M
du, v dV =
M
du v
M
=
M
d(u v) (1)p u d v = (1)p
udv
293
by Stokes formula. Therefore (3.4) implies du, v = (1)p (1)p(m1) u d v = (1)mp+1 u, d v .
(3.10) Remark. If m is even, the formula reduces to d = d . (3.11) Denition. The operator = dd + d d is called the Laplace-Beltrami operator of M . Since d is the adjoint of d, the Laplace operator is formally self-adjoint, i.e. u, v = u, v when the forms u, v are of class and compactly supported. (3.12) Example. Let M be an open subset of Rm and g = i=1 dx2 . In that case we i get uI dxj dxI , u= uI dxI , du = xj
|I|=p |I|=p,j m
u, v =
M
u, v dV =
M I
uI vI dV
One can write dv = dxj (v/xj ) where v/xj denotes the form v in which all coecients vI are dierentiated as vI /xj . An integration by parts combined with contraction gives d u, v = u, dv =
M
u,
j
dxj
v dV xj xj u , v dV, xj
=
M j
xj xj
u,
v dV = xj
d u = We get therefore
u = xj
I,j
uI xj xj
dxI .
dd u = d du = Since xj
I,j,k
2 uI dxk xj xk xj 2 uI xj xk xj
dxI ,
I,j,k
(dxk dxI ). xj
(dxk dxI ) =
xj
dxk dxI dxk 2 uI dxI . x2 j
dxI ,
we obtain u =
I j
294
In the case of an arbitrary riemannian manifold (M, g) we have u= du =

I,j uI I , (j uI ) j I + uI dI , I I,K uI K , I,K
d u =
I,j
(j uI ) j
I +
for some coecients I,K , |I| = p, |K| = p 1. It follows that the principal part of is the same as that of the second order operator u As a consequence, is elliptic. Assume now that DE is a hermitian connection on E. The formal adjoint operator of DE acting on (M, p TM E) is (3.13) Indeed, if s get
DE = (1)mp+1 DE . (M, p TM E), t (M, p+1 TM E) have compact support, we 2 j uI I .
DE s, t =
M
DE s, t dV =
M
{DE s, t}
=
M
d{s, t} (1)p {s, DE t} = (1)mp+1 s, DE t .
(3.14) Denition. The Laplace-Beltrami operator associated to DE is the second order operator E = DE DE + DE DE . E is a self-adjoint elliptic operator with principal part s
2 j sI, I e .
I,
3.3. Harmonic Forms and Hodge Isomorphism Let E be a hermitian vector bundle over a compact Riemannian manifold (M, g). We assume that E possesses a at hermitian connection DE (this means that (DE ) = 2 DE = 0, or equivalently, that E is given by a representation 1 (M ) U (r), cf. V-6). A fundamental example is of course the trivial bundle E = M C with the connection DE = d. Thanks to our atness assumption, DE denes a generalized De Rham complex DE :
(M, p TM E) (M, p+1 TM E).
p The cohomology groups of this complex will be denoted by HDR (M, E).
295
The space of harmonic forms of degree p with respect to the Laplace-Beltrami operator E = DE DE + DE DE is dened by (3.15)
p (M, E) =
(M, p TM E) ; E s = 0 .
Since E s, s = ||DE s||2 + ||DE s||2 , we see that s DE s = 0.
p (M, E) if and only if DE s =
(3.16) Theorem. For any p, there exists an orthogonal decomposition

(M, p TM E) = p (M, E) Im DE Im DE , (M, p1 TM E) , (M, p+1 TM E) .
Im DE = DE
Im DE = DE
Proof. It is immediate that p (M, E) is orthogonal to both subspaces Im DE and Im DE . 2 The orthogonality of these two subspaces is also clear, thanks to the assumption DE = 0 : 2 DE s, DE t = DE s, t = 0.
We apply now Cor. 2.4 to the elliptic operator E = acting on p-forms, i.e. on the E bundle F = p TM E. We get
(M, p TM E) = p (M, E) E (M, p TM E) , Im E = Im(DE DE + DE DE ) Im DE + Im DE .
p (3.17) Hodge isomorphism theorem. The De Rham cohomology group HDR (M, E) p is nite dimensional and HDR (M, E) p (M, E).
Proof. According to decomposition 3.16, we get

p BDR (M, E) = DE p ZDR (M, E) = ker DE = (Im DE ) = p (M, E) Im DE . (M, p1 TM E) ,
This shows that every De Rham cohomology class contains a unique harmonic representative. (3.18) Poincar duality. The bilinear pairing e
mp p HDR (M, E) HDR (M, E ) C,
(s, t)
st
is a non degenerate duality. Proof. First note that there exists a naturally dened at connection DE such that for any s1 (M, E), s2 (M, E ) we have (3.19) d(s1 s2 ) = (DE s1 ) s2 + (1)deg s1 s1 DE s2 .
296
It is then a consequence of Stokes formula that the map (s, t) M st can be factorized through cohomology groups. Let s (M, p TM E). We leave to the reader the proof of the following formulas (use (3.19) in analogy with the proof of Th. 3.9):
DE (# s) = (1)p # DE s, E (# s) = (1)p+1 # DE s,
(3.20)
E (# s) = # E s, Consequently #s mp (M, E ) if and only if s p (M, E). Since

M
s #s =
|s|2 dV = s 2 ,
we see that the Poincar pairing has zero kernel in the left hand factor p (M, E) e p HDR (M, E). By symmetry, it has also zero kernel on the right. The proof is achieved.
4. Hermitian and Khler Manifolds a

Let X be a complex n-dimensional manifold. A hermitian metric on X is a positive denite hermitian form of class on TX ; in a coordinate system (z1 , . . . , zn ), such a form can be written h(z) = 1 j,k n hjk (z) dzj dz k , where (hjk ) is a positive hermitian matrix with coecients. According to (III-1.8), the fundamental (1, 1)-form associated to h is the positive form of type (1, 1) = Im h = (4.1) Denition. a) A hermitian manifold is a pair (X, ) where is a on X. b) The metric is said to be khler if d = 0. a c) X is said to be a Khler manifold if X carries at least one Khler metric. a a Since is real, the conditions d = 0, d = 0, d = 0 are all equivalent. In local coordinates we see that d = 0 if and only if hjk hlk = zl zj A simple computation gives n = det(hjk ) n! i dzj dz j = det(hjk ) dx1 dy1 dxn dyn , 2 , 1 j, k, l n.
i 2
hjk dzj dz k ,
j, k
n.
positive denite (1, 1)-form
1 j n
where zn = xn + iyn . Therefore the (n, n)-form (4.2) dV = 1 n n!
4. Hermitian and Khler Manifolds a
297
is positive and coincides with the hermitian volume element of X. If X is compact, a then X n = n! Vol (X) > 0. This simple remark already implies that compact Khler manifolds must satisfy some restrictive topological conditions: (4.3) Consequence. a) If (X, ) is compact Khler and if {} denotes the cohomology class of in H 2 (X, R), a n then {} = 0. b) If X is compact Khler, then H 2k (X, R) = 0 for 0 a zero class in H 2k (X, R). k n. In fact, {}k is a non
(4.4) Example. The complex projective space Pn is Khler. A natural Khler metric a a n on P , called the Fubini-Study metric, is dened by p = i d d log |0 |2 + |1 |2 + + |n |2 2
where 0 , 1 , . . . , n are coordinates of Cn+1 and where p : Cn+1 \ {0} Pn is the projection. Let z = (1 /0 , . . . , n /0 ) be non homogeneous coordinates on Cn Pn . Then (V-15.8) and (V-15.12) show that = i i d d log(1 + |z|2 ) = c 2 2
(1) ,
n = 1.
Pn
Furthermore {} H 2 (Pn , Z) is a generator of the cohomology algebra H (Pn , Z) in virtue of Th. V-15.10. (4.5) Example. A complex torus is a quotient X = Cn / by a lattice of rank 2n. Then X is a compact complex manifold. Any positive denite hermitian form = i hjk dzj dz k with constant coecients denes a Khler metric on X. a (4.6) Example. Every (complex) submanifold Y of a Khler manifold (X, ) is Khler a a n with metric Y . Especially, all submanifolds of P are Khler. a (4.7) Example. Consider the complex surface X = (C2 \ {0})/ where = {n ; n Z}, < 1, acts as a group of homotheties. Since C2 \ {0} is dieomorphic to R S 3 , we have X S 1 S 3 . Therefore H 2 (X, R) = 0 by Knneths u + formula IV-15.10, and property 4.3 b) shows that X is not Khler. More generally, one a can obtaintake to be an innite cyclic group generated by a holomorphic contraction of C2 , of the form z1 z2 1 z1 2 z2 , resp. z1 z2 z1 p z2 + z1 ,
where , 1 , 2 are complex numbers such that 0 < |1 | |2 | < 1, 0 < || < 1, and p a positive integer. These non Khler surfaces are called Hopf surfaces. a
298
The following Theorem shows that a hermitian metric on X is Khler if and only a if the metric is tangent at order 2 to a hermitian metric with constant coecients at every point of X. (4.8) Theorem. Let be a positive denite (1, 1)-form on X. In order that be Khler, it is necessary and sucient that to every point x0 X corresponds a a holomorphic coordinate system (z1 , . . . , zn ) centered at x0 such that (4.9) =i
1 l,m n
lm dzl dz m ,
j n
lm = lm + O(|z|2 ).
If is Khler, the coordinates (zj )1 a (4.10) lm =
can be chosen such that cjklm zj z k + O(|z|3 ),

1 j,k n
, = lm zl zm
where (cjklm ) are the coecients of the Chern curvature tensor (4.11) (TX )x0 =
j,k,l,m
cjklm dzj dz k
zl
zm
associated to (TX , ) at x0 . Such a system (zj ) will be called a geodesic coordinate system at x0 . Proof. It is clear that (4.9) implies dx0 = 0, so the condition is sucient. Assume now that is Khler. Then one can choose local coordinates (x1 , . . . , xn ) such that a (dx1 , . . . , dxn ) is an -orthonormal basis of Tx0 X. Therefore =i
1 l,m n
lm dxl dxm ,
where (ajlm xj + a xj ) + O(|x|2 ). jlm
(4.12)
lm = lm + O(|x|) = lm +
1 j n
Since is real, we have a = ajml ; on the other hand the Khler condition lm /xj = a jlm jm /xl at x0 implies ajlm = aljm . Set now zm = xm + 1 2 ajlm xj xl ,
j,l
n.
Then (zm ) is a coordinate system at x0 , and dzm = dxm +

j,l
ajlm xj dxl , ajlm xj dxl dxm ajlm xj dxm dxl + O(|x|2 )
i
m
dzm dz m = i
dxm dxm + i +i
j,l,m
j,l,m
=i
l,m
lm dxl dxm + O(|x|2 ) = + O(|z|2 ).
5. Basic Results of Khler Geometry a
299
Condition (4.9) is proved. Suppose the coordinates (xm ) chosen from the beginning so that (4.9) holds with respect to (xm ). Then the Taylor expansion (4.12) can be rened into (4.13) lm = lm + O(|x|2 ) = lm +
j,k 3 ajklm xj xk + a jklm xj xk + ajklm xj xk + O(|x| ).
These new coecients satisfy the relations

a jklm = akjlm , a jklm = ajkml ,
ajklm = akjml .
The Khler condition lm /xj = jm /xl at x = 0 gives the equality a a jklm = alkjm ; in particular a jklm is invariant under all permutations of j, k, l. If we set
zm = xm + then by (4.13) we nd dzm = dxm +

j,k,l
1 3
a jklm xj xk xl ,
j,k,l
n,
a jklm xj xk dxl , dzm dz m + i dzm dz m + i
n,
=i
1 m n
j,k,l,m
ajklm xj xk dxl dxm + O(|x|3 ), ajklm zj z k dzl dz m + O(|z|3 ).
(4.14)
=i
1 m n
j,k,l,m
It is now easy to compute the Chern curvature tensor (TX )x0 in terms of the coecients ajklm . Indeed , = lm + zl zm d ajklm zj z k + O(|z|3 ),
j,k
, = D , zl zm zl zm = D D zl zl
=
j,k
ajklm z k dzj + O(|z|2 ), ajklm dzj dz k + O(|z|), zm
(TX )
j,k,m
therefore cjklm = ajklm and the expansion (4.10) follows from (4.14). (4.15) Remark. As a by-product of our computations, we nd that on a Khler mania fold the coecients of (TX ) satisfy the symmetry relations cjklm = ckjml , cjklm = clkjm = cjmlk = clmjk .
300

5.1. Operators of Hermitian Geometry Let (X, ) be a hermitian manifold and let zj = xj + iyj , 1 j n, be analytic coordinates at a point x X such that (x) = i dzj dz j is diagonalized at this point. The associated hermitian form is the h(x) = 2 dzj dz j and its real part is the euclidean metric 2 (dxj )2 + (dyj )2 . It follows from this that |dxj | = |dyj | = 1/ 2, |dzj | = |dz j | = 1, and that (/z1 , . . . , /zn) is an orthonormal basis of (Tx X, ). Formula (3.1) with uj , vk in the orthogonal sum (C TX ) = TX TX denes a natural inner product on the exterior algebra (C TX ) . The norm of a form u=
I,J
uI,J dzI dz J (C TX )
at the given point x is then equal to (5.1) |u(x)|2 = |uI,J (x)|2 .
I,J
The Hodge operator (3.2) can be extended to C-valued forms by the formula (5.2) u v = u, v dV.
It follows that is a C-linear isometry

: p,q TX nq,np TX .
The usual operators of hermitian geometry are the operators d, = d , = d + d already dened, and their complex counterparts d = d + d , (5.3) = d + d , d = (d ) = d , d = (d ) = d , = d d + d d , = d d + d d . Another important operator is the operator L of type (1,1) dened by Lu = u (5.4) and its adjoint = 1 L : (5.5) 5.2. Commutation Identities If A, B are endomorphisms of the algebra graded Lie bracket) is dened by (5.6)
, (X, C),
u, v = Lu, v .
their graded commutator (or
[A, B] = AB (1)ab BA
301
where a, b are the degrees of A and B respectively. If C is another endomorphism of degree c, the following Jacobi identity is easy to check: (5.7) (1)ca A, [B, C] + (1)ab B, [C, A] + (1)bc C, [A, B] = 0.
For any p,q TX , we still denote by the endomorphism of type (p, q) on , TX dened by u u. Consider further a real (1,1)-form
There exists an -orthogonal basis (1 , 2 , . . . , n ) in TX which diagonalizes both forms and : j j j , j R. =i j j , = i

1 j n 1 j n
1,1 TX .
(5.8) Proposition. For every form u = [, ]u =

J,K jJ
uJ,K J K , one has
j +
jK
1 j n
j uJ,K J K .
Proof. If u is of type (p, q), a brute-force computation yields u = i(1)p

J,K,l
uJ,K (l
J ) ( l
K ),
1 1
l m
n, n,
u = i(1)p [, ]u =
J,K,l,m
J,K,m
m uJ,K m J m K ,
m uJ,K
l (m m
J ) l ( m (l J ) m J ) K
K) ( l K )

=
J,K,m
m uJ,K m (m
+ J m ( m
K ) J K
=
J,K mJ
m +
mK
1 m n
m uJ,K J K .
(5.9) Corollary. For every u p,q TX , we have
[L, ]u = (p + q n)u.
Proof. Indeed, if = , we have 1 = = n = 1. (5.10)
This result can be generalized as follows: for every u k (C TX ) , we have [Lr , ]u = r(k n + r 1) Lr1 u.
In fact, it is clear that [Lr , ]u = =

0 m r1
Lr1m [L, ]Lm u (2m + k n)Lr1m Lm u = r(r 1) + r(k n) Lr1 u.
0 m r1
302
5.3. Primitive Elements and Hard Lefschetz Theorem In this subsection, we prove a fundamental decomposition theorem for the represen tation of the unitary group U (TX ) U (n) acting on the spaces p,q TX of (p, q)-forms. It turns out that the representation is never irreducible if 0 < p, q < n. (5.11) Denition. A homogeneous element u k (C TX ) is called primitive if u = 0. The space of primitive elements of total degree k will be denoted
Primk TX = p+q=k Primp,q TX .
Let u Primk TX . Then
s Lr u = s1 (Lr Lr )u = r(n k r + 1)s1 Lr1 u. By induction, we get for r (5.12) s
s Lr u = r(r 1) (r s + 1) (n k r + 1) (n k r + s)Lrs u.
Apply (5.12) for r = n + 1. Then Ln+1 u is of degree > 2n and therefore we have Ln+1 u = 0. This gives (n + 1) n + 1 (s 1) (k)(k + 1) (k + s 1)Ln+1s u = 0. The integral coecient is = 0 if s k, hence: (n + 1 k)+ .
(5.13) Corollary. If u Primk TX , then Ls u = 0 for s (5.14) Corollary. Primk TX = 0 for n + 1
2n.
Proof. Apply Corollary 5.13 with s = 0. (5.15) Primitive decomposition formula. For every u k (C TX ) , there is a unique decomposition u=
r (kn)+
L r ur ,
ur Primk2r TX .
Furthermore ur = k,r (L, )u where k,r is a non commutative polynomial in L, with rational coecients. As a consequence, there are direct sum decompositions of U (n)representations k (C TX ) = Lr Primk2r TX ,
r (kn)+ p,q TX = r (p+qn)+ Lr Primpr,qr TX .
303
Proof of the uniqueness of the decomposition. Assume that u = 0 and that ur = 0 for some r. Let s be the largest integer such that us = 0. Then s u = 0 =
(kn)+ r s
s Lr ur =
(kn)+ r s
sr r Lr ur .
But formula (5.12) shows that r Lr ur = ck,r ur for some non zero integral coecient ck,r = r!(n k + r + 1) (n k + 2r). Since ur is primitive we get s Lr ur = 0 when r < s, hence us = 0, a contradiction. Proof of the existence of the decomposition. We prove by induction on s s u = 0 implies (5.16) u=
(kn)+ r<s
(k n)+ that
L r ur ,
ur = k,r,s (L, )u Primk2r TX .
The Theorem will follow from the step s = n + 1.

Assume that the result is true for s and that s+1 u = 0. Then s u is in Primk2s TX . Since s (k n)+ we have ck,s = 0 and we set
us =
1 ck,s
s u Primk2s TX ,
u = u L s us = 1 By formula (5.12), we get
1 ck,s
Ls s u.
s u = s u s Ls us = s u ck,s us = 0. The induction hypothesis implies u =

(kn)+ r<s Lr u , u = k,r,s (L, )u Primk2r TX , r r
hence u =
(kn)+ r s
Lr ur with
It remains to prove the validity of the decomposition 5.16) for the initial step s = (kn)+ , i.e. that s u = 0 implies u = 0. If k n, then s = 0 and there is nothing to prove. We are left with the case k > n, kn u = 0. Then v = u 2nk (C TX ) and 2n k < n. Since the decomposition exists in degree n by what we have just proved, we get v = u= Lr vr , vr Prim2nk2r TX ,
r 0
ur = u = k,r,s (L, ) 1 1 Ls s u, r < s, r ck,s 1 s u. us = ck,s
0 = kn u = Lkn u =
r 0
Lr+kn vr ,
304
with degree (Lr+kn vr ) = 2n k + 2(k n) = k. The uniqueness part shows that vr = 0 for all r , hence u = 0. The Theorem is proved. (5.17) Corollary. The linear operators Lnk : k (C TX ) 2nk (C TX ) ,
Lnpq : p,q TX nq,np TX ,
are isomorphisms for all integers k
n, p + q
n.
r 0
Proof. For every u k TX , the primitive decomposition u = C bijectively onto that of Lnk u :
Lr ur is mapped
Lnk u =
r 0
Lr+nk ur .
6. Commutation Relations
6.1. Commutation Relations on a Khler Manifold a Assume rst that X = Cn is an open subset and that is the standard Khler a metric =i dzj dz j .
1 j n
For any form u (6.1 ) (6.1 )
(, p,q TX ) we have
d u =
I,J,k
uI,J dzk dzI dz J , zk uI,J dz k dzI dz J . z k
d u =
I,J,k
Since the global L2 inner product is given by u, v =

I,J
uI,J v I,J dV,
easy computations analogous to those of Example 3.12 show that (6.2 ) (6.2 ) d u = d u = uI,J z k zk uI,J zk z k (dzI dz J ), (dzI dz J ).
I,J,k
I,J,k
We rst prove a lemma due to [Akizuki and Nakano 1954].
305
(6.3) Lemma. In Cn , we have [d , L] = id . Proof. Formula (6.2 ) can be written more briey d u = Then we get [d , L]u = z k ( u) + zk z k u . zk z k u . zk
Since has constant coecients, we have [d , L] u = = Clearly z k = idzk , so [d , L] u = i

k
u ( u) = and therefore zk zk u zk u . zk z k u zk
z k z k
dzk
u = id u. zk
We are now ready to derive the basic commutation relations in the case of an arbitrary Khler manifold (X, ). a (6.4) Theorem. If (X, ) is Khler, then a [d , L]= id , [, d ] = id , [d , L]= id , [, d ] = id .
Proof. It is sucient to verify the rst relation, because the second one is the conjugate of the rst, and the relations of the second line are the adjoint of those of the rst line. If (zj ) is a geodesic coordinate system at a point x0 X, then for any (p, q)-forms u, v with compact support in a neighborhood of x0 , (4.9) implies u, v =
M I,J
uIJ v IJ +
I,J,K,L
aIJKL uIJ v KL dV,
with aIJKL (z) = O(|z|2 ) at x0 . An integration by parts as in (3.12) and (6.2 ) yields d u = uI,J zk z k (dzI dz J ) + bIJKL uIJ dzK dz L ,
I,J,k
I,J,K,L
306
where the coecients bIJKL are obtained by derivation of the aIJKL s. Therefore bIJKL = O(|z|). Since /zk = O(|z|), the proof of Lemma 6.3 implies here [d , L]u = id u + O(|z|), in particular both terms coincide at every given point x0 X. (6.5) Corollary. If (X, ) is Khler, the complex Laplace-Beltrami operators satisfy a = = 1 . 2
Proof. It will be rst shown that = . We have = [d , d ] = i d , [, d ] . Since [d , d ] = 0, Jacobis identity (5.7) implies d , [, d ] + d , [d , ] = 0, hence = d , i[d , ] = [d , d ] = . On the other hand = [d + d , d + d ] = + + [d , d ] + [d , d ]. Thus, it is enough to prove: (6.6) Lemma. [d , d ] = 0, [d , d ] = 0. Proof. We have [d , d ] = i d , [, d ] and (5.7) implies d , [, d ] + , [d , d ] + d , [d , ] = 0, hence 2 d , [, d ] = 0 and [d , d ] = 0. The second relation [d , d ] = 0 is the adjoint of the rst. (6.7) Theorem. commutes with all operators , d , d , d , d , L, . Proof. The identities [d , ] = [d , ] = 0, [d , ] = [d , ] = 0 and [, ] = 0 are immediate. Furthermore, the equality [d , L] = d = 0 together with the Jacobi identity implies [L, ] = L, [d , d ] = d , [d , L] = i[d , d ] = 0. By adjunction, we also get [ , ] = 0. 6.2. Commutation Relations on Hermitian Manifolds We are going to extend the commutation relations of 6.1 to an arbitrary hermitian manifold (X, ). In that case is no longer tangent to a constant metric, and the commutation relations involve extra terms arising from the torsion of . Theorem 6.8 below is taken from [Demailly 1984], but the idea was already contained in [Griths 1966]. (6.8) Theorem. Let be the operator of type (1, 0) and order 0 dened by = [, d ]. Then
307
a) b) c) d)
[d , L]= i(d + ), [d , L] = i(d + ), [, d ] = i(d + ), [, d ] = i(d + ) ;
d will be called the torsion form of , and the torsion operator. Proof. b) follows from a) by conjugation, whereas c), d) follow from a), b) by adjunction. It is therefore enough to prove relation a). Let (zj )1 j n be complex coordinates centered at a point x0 X, such that the ntuple (/z1 , . . . , /zn ) is an orthonormal basis of Tx0 X for the metric (x0 ). Consider the metric with constant coecients 0 = i
1 j n
dzj dz j .
The metric can then be written = 0 + with = O(|z|). Denote by , 0 , L0 , 0 , d , d the inner product and the operators associated to 0 0 n the constant metric 0 , and let dV0 = 0 /2n n!. The proof of relation a) is based on a Taylor expansion of L, , d , d in terms of the operators with constant coecients L0 , 0 , d , d . 0 0 (6.9) Lemma. Let u, v
(X, p,q TX ). Then in a neighborhood of x0 0 dV0
u, v dV = u [, 0 ]u, v Proof. In a neighborhood of x0 , let =i

1 j n j j j ,
+ O(|z|2 ).
n ,
j n
be a diagonalization of the (1,1)-form (z) with respect to an orthonormal basis (j )1 of Tz X for 0 (z). We thus have = 0 + = i with j = 1 + j and j = O(|z|). Set now J = {j1 , . . . , jp }, u=
J = j1 jp , j j j
J = j1 jp ,
vJ,K J K
uJ,K J K ,
v=
where summations are extended to increasing multi-indices J, K such that |J| = p, |K| = q. With respect to we have j , j = 1 , hence j u, v dV =
J,K
1 1 uJ,K v J,K 1 n dV0 K J 1 j j +

jK 1 j n
=
J,K
j uJ,K v J,K dV0 + O(|z|2 ).
jJ
308
Lemma 6.9 follows if we take Prop. 5.8 into account. (6.10) Lemma. d = d + 0 , [d , ] at point x0 , i.e. at this point both operators 0 0 have the same formal expansion. Proof. Since d is an operator of order 1, Lemma 6.9 shows that d coincides at point x0 with the formal adjoint of d for the metric u, v
1
=
X
u [, 0 ]u, v
0 dV0 . (X, p,q1 TX ) we get by
For any compactly supported u denition u, d v

1
(X, p,q TX ), v
=
X
u [, 0 ]u, d v
dV0 =
X
d u d [, 0 ]u, v 0 0
0 dV0 .
Since and 0 coincide at point x0 and since (x0 ) = 0 we obtain at this point d u = d u d [, 0 ]u = d u d , [, 0] u ; 0 0 0 0 d = d d , [, 0 ] . 0 0 We have [0 , d ] = [d , L0 ] = 0 since d 0 = 0. The Jacobi identity (5.7) implies 0 d , [, 0] + 0 , [d , ] = 0, 0 0 and Lemma 6.10 follows. Proof Proof of formula 6.8 a). The equality L = L0 + and Lemma 6.10 yield (6.11) [L, d ] = [L0 , d ] + L0 , 0 , [d , ] 0 0 + [, d ] 0
at point x0 , because the triple bracket involving twice vanishes at x0 . From the Jacobi identity applied to C = [d , ], we get 0 L0 , [0 , C] = [0 , [C, L0] C, [L0 , 0 ] , (6.12) [C, L0 ] = L0 , [d , ] = , [L0, d ] (since [, L0] = 0). 0 0 Lemma 6.3 yields [L0 , d ] = id , hence 0 (6.13) [C, L0 ] = [, id ] = id = id .
On the other hand, C is of type (1, 0) and Cor. 5.9 gives (6.14) C, [L0 , 0 ] = C = [d , ]. 0
From (6.12), (6.13), (6.14) we get L0 , 0 , [d , ] 0 = [0 , id ] + [d , ]. 0
7. Groups
p,q (X, E) and Serre Duality
309
This last equality combined with (6.11) implies [L, d ] = [L0 , d ] [0 , id ] = i(d + ) 0 at point x0 . Formula 6.8 a) is proved. (6.15) Corollary. The complex Laplace-Beltrami operators satisfy = + [d , ] [d , ],
[d , d ] = [d , ], [d , d ] = [d , ], = + [d , ] [d , ].
1 Therefore , and 2 no longer coincide, but they dier by linear dierential operators of order 1 only.
Proof. As in the Khler case (Cor. 6.5 and Lemma 6.6), we nd a = [d , d ] = d , i[, d ] ] [d , d + ] = i d , [, d ] = 0,
= d , i[d , ] [d , = + [d , ] [d , ],
and the rst two lines are proved. The third one is an immediate consequence of the second.
7. Groups
p,q (X, E) and Serre Duality
Let (X, ) be a compact hermitian manifold and E a holomorphic hermitian vector bundle of rank r over X. We denote by DE the Chern connection of E, by DE = DE the formal adjoint of DE , and by DE , DE the components of DE of type (1, 0) and (0, 1).
Corollary 6.8 implies that the principal part of the operator = D DE + DE D E is one half that of E . The operator acting on each space (X, p,q TX E) is E 2 thus a self-adjoint elliptic operator. Since D = 0, the following results can be obtained in a way similar to those of 3.3.
(7.1) Theorem. For every bidegree (p, q), there exists an orthogonal decomposition
(X, p,q TX E) = p,q (X, E) Im DE Im DE (X, p,q TX E). (X, p,q TX
where
p,q (X, E) is the space of -harmonic forms in E
The above decomposition shows that the subspace of d -cocycles in E) is p,q (X, E) Im DE . From this, we infer
(7.2) Hodge isomorphism theorem. The Dolbeault cohomology group H p,q (X, E) is nite dimensional, and there is an isomorphism H p,q (X, E) p,q (X, E).
310
(7.3) Serre duality theorem. The bilinear pairing H p,q (X, E) H np,nq (X, E ) C, is a non degenerate duality.
Proof. Let s1 (X, p,q TX E), s2 of bidegree (n, n 1), we have (X, np,nq1 TX E). Since s1 s2 is
(s, t)
st
(7.4)
d(s1 s2 ) = d (s1 s2 ) = d s1 s2 + (1)p+q s1 d s2 .
Stokes formula implies that the above bilinear pairing can be factorized through Dolbeault cohomology groups. The # operator dened in 3.1 is such that #:
(X, p,q TX E) (X, np,nq TX E ).
Furthermore, (3.20) implies

d (# s) = (1)deg s # DE s, DE (# s) = (1)deg s+1 # DE s,
where DE is the Chern connection of E . Consequently, s p,q (X, E) if and only if # s np,nq (X, E ). Theorem 7.3 is then a consequence of the fact that the integral s 2 = X s # s does not vanish unless s = 0.
(# s) = # s, E E
8. Cohomology of Compact Khler Manifolds a

8.1. Bott-Chern Cohomology Groups Let X be for the moment an arbitrary complex manifold. The following cohomology groups are helpful to describe Hodge theory on compact complex manifolds which are not necessarily Khler. a (8.1) Denition. We dene the Bott-Chern cohomology groups of X to be
p,q HBC (X, C) = (X, p,q TX ) ker d /d d (X, p1,q1TX ).
, Then HBC (X, C) has the structure of a bigraded algebra, which we call the Bott-Chern cohomology algebra of X. As the group d d (X, p1,q1 TX ) is contained in both coboundary groups d (X, p,q1 TX ) or d (X, p+q1 (C TX ) ), there are canonical morphisms
(8.2) (8.3)
p,q HBC (X, C) H p,q (X, C),
p+q p,q HBC (X, C) HDR (X, C),
of the Bott-Chern cohomology to the Dolbeault or De Rham cohomology. These morphisms are homomorphisms of C-algebras. It is also clear from the denition that we q,p p,q have the symmetry property HBC (X, C) = HBC (X, C). It can be shown from the Hodgep,q Frlicher spectral sequence (see 11 and Exercise 13.??) that HBC (X, C) is always nite o dimensional if X is compact.
8. Cohomology of Compact Khler Manifolds a
311
8.2. Hodge Decomposition Theorem We suppose from now on that (X, ) is a compact Khler manifold. The equality a = 2 shows that is homogeneous with respect to bidegree and that there is an orthogonal decomposition (8.4)
k (X, C) =
p+q=k
p,q (X, C).
As = = , we also have q,p (X, C) = p,q (X, C). Using the Hodge isomorphism theorems for the De Rham and Dolbeault cohomology, we get: (8.5) Hodge decomposition theorem. On a compact Khler manifold, there are a canonical isomorphisms H k (X, C) H
q,p
H p,q (X, C)
(Hodge decomposition), (Hodge symmetry).
(X, C)
p+q=k H p,q (X, C)
The only point which is not a priori completely clear is that this decomposition is independent of the Khler metric. In order to show that this is the case, one can use a the following Lemma, which allows us to compare all three types of cohomology groups considered in 8.1. (8.6) Lemma. Let u be a d-closed (p, q)-form. The following properties are equivalent: a) u is d-exact ;
b ) u is d -exact ; b ) u is d -exact ; c) d) u is d d -exact, i.e. u can be written u = d d v. u is orthogonal to
p,q (X, C).
Proof. It is obvious that c) implies a), b ), b ) and that a) or b ) or b ) implies d). It is thus sucient to prove that d) implies c). As du = 0, we have d u = d u = 0, and as u is supposed to be orthogonal to p,q (X, C), Th. 7.1 implies u = d s, s (X, p,q1TX ). By the analogue of Th. 7.1 for d , we have s = h + d v + d w, with h p,q1 (X, C), v (X, p1,q1 TX ) and w (X, p+1,q1 TX ). Therefore u = d d v + d d w = d d v d d w in view of Lemma 6.6. As d u = 0, the component d d w orthogonal to ker d must be zero. From Lemma 8.6 we infer the following Corollary, which in turn implies that the Hodge decomposition does not depend on the Khler metric. a
312
(8.7) Corollary. Let X be a compact Khler manifold. Then the natural morphisms a
p,q HBC (X, C) H p,q (X, C), p,q k HBC (X, C) HDR (X, C)
p+q=k
are isomorphisms.
p,q Proof. The surjectivity of HBC (X, C) H p,q (X, C) comes from the fact that every class in H p,q (X, C) can be represented by a harmonic (p, q)-form, thus by a d-closed (p, q)form; the injectivity means nothing more than the equivalence (8.5 b ) (8.5 c). Hence p,q p,q HBC (X, C) H p,q (X, C) p,q (X, C), and the isomorphism p+q=k HBC (X, C) k HDR (X, C) follows from (8.4).
Let us quote now two simple applications of Hodge theory. The rst of these is 2p a computation of the Dolbeault cohomology groups of Pn . As HDR (Pn , C) = C and H p,p (Pn , C) { p } = 0, the Hodge decomposition formula implies: (8.8) Application. The Dolbeault cohomology groups of Pn are H p,p (Pn , C) = C for 0 p n, H p,q (Pn , C) = 0 for p = q.
(8.9) Proposition. Every holomorphic p-form on a compact Khler manifold X is a d-closed. Proof. If u is a holomorphic form of type (p, 0) then d u = 0. Furthermore d u is of type (p, 1), hence d u = 0. Therefore u = 2 u = 0, which implies du = 0. (8.10) Example. Consider the Heisenberg group G Gl3 (C), dened as the subgroup of matrices 1 x z 1 y , (x, y, z) C3 . M = 0 0 0 1
Let be the discrete subgroup of matrices with entries x, y, z Z[i] (or more generally in the ring of integers of an imaginary quadratic eld). Then X = G/ is a compact complex 3-fold, known as the Iwasawa manifold. The equality 0 dx dz xdy M 1 dM = 0 0 dy 0 0 0 shows that dx, dy, dz xdy are left invariant holomorphic 1-forms on G. These forms induce holomorphic 1-forms on the quotient X = G/. Since dz xdy is not d-closed, we see that X cannot be Khler. a 8.3. Primitive Decomposition and Hard Lefschetz Theorem We rst introduce some standard notation. The Betti numbers and Hodge numbers of X are by denition (8.11) bk = dimC H k (X, C), hp,q = dimC H p,q (X, C).
9. Jacobian and Albanese Varieties
313
Thanks to Hodge decomposition, these numbers satisfy the relations (8.12) bk =

p+q=k
hp,q ,
hq,p = hp,q .
As a consequence, the Betti numbers b2k+1 of a compact Khler manifold are even. a Note that the Serre duality theorem gives the additional relation hp,q = hnp,nq , which holds as soon as X is compact. The existence of primitive decomposition implies other interesting specic features of the cohomology algebra of compact Khler manifolds. a (8.13) Lemma. If u = r (kn)+ Lr ur is the primitive decomposition of a harmonic k-form u, then all components ur are harmonic. Proof. Since [, L] = 0, we get 0 = u =
r
Lr ur , hence ur = 0 by uniqueness.
Let us denote by Prim k (X, C) = p+q=k Prim p,q (X, C) the spaces of primitive harmonic k-forms and let bk,prim , hp,q be their respective dimensions. Lemma 8.13 prim yields (8.14) (8.15)
p,q (X, C) =
r (p+qn)+
Lr Prim pr,qr (X, C), hpr,qr . prim

r (p+qn)+
hp,q =
Formula (8.15) can be rewritten
(8.15 )
if p + q
(8.16) Corollary. The Hodge and Betti numbers satisfy the inequalities a) if k = p + q b) if k = p + q n, then hp,q n, then hp,q hp1,q1 , bk hp+1,q+1 , bk bk2 , bk+2 .
if p + q
p1,q1 + n, hp,q = hp,q + hprim prim
n, hp,q = hnq,np + hnq1,np1 + . prim prim
Another important result of Hodge theory (which is in fact a direct consequence of Cor. 5.17) is the (8.17) Hard Lefschetz theorem. The mappings Lnk : H k (X, C) H 2nk (X, C), Lnpq : H p,q (X, C) H nq,np (X, C), are isomorphisms. k n, p + q n,
314
9.1. Description of the Picard Group An important application of Hodge theory is a description of the Picard group H 1 (X, ) of a compact Khler manifold. We assume here that X is connected. The a exponential exact sequence 0 Z 1 gives (9.1) 0 H 1 (X, Z) H 1 (X, ) H 1 (X, )
1 H 2 (X, Z) H 2 (X, )
because the map exp(2i) : H 0 (X, ) = C H 0 (X, ) = C is onto. We have H 1 (X, ) H 0,1 (X, C) by (V-11.6). The dimension of this group is called the irregularity of X and is usually denoted (9.2) Therefore we have b1 = 2q and (9.3) H 1 (X, ) Cq , H 0 (X, 1 ) = H 1,0 (X, C) Cq . X q = q(X) = h0,1 = h1,0 .
(9.4) Lemma. The image of H 1 (X, Z) in H 1 (X, ) is a lattice. Proof. Consider the morphisms
induced by the inclusions Z R C . Since the Cech cohomology groups with values in Z, R can be computed by nite acyclic coverings, we see that H 1 (X, Z) is a nitely generated Z-module and that the image of H 1 (X, Z) in H 1 (X, R) is a lattice. It is enough to check that the map H 1 (X, R) H 1 (X, ) is an isomorphism. However, the commutative diagram 0C 0 1 2 0
0,0 d
H 1 (X, Z) H 1 (X, R) H 1 (X, C) H 1 (X, )
shows that the map H 1 (X, R) H 1 (X, ) corresponds in De Rham and Dolbeault cohomologies to the composite mapping
1 1 HDR (X, R) HDR (X, C) H 0,1 (X, C). 1 Since H 1,0 (X, C) and H 0,1 (X, C) are complex conjugate subspaces in HDR (X, C), we 1 0,1 conclude that HDR (X, R) H (X, C) is an isomorphism.
0,1 d
0,2
As a consequence of this lemma, H 1 (X, Z) Z2q . The q-dimensional complex torus (9.5) Jac(X) = H 1 (X, )/H 1 (X, Z) is called the Jacobian variety of X and is isomorphic to the subgroup of H 1 (X, ) corresponding to line bundles of zero rst Chern class. On the other hand, the kernel of H 2 (X, Z) H 2 (X, ) = H 0,2 (X, C)
315
which consists of integral cohomology classes of type (1, 1), is equal to the image of c1 in H 2 (X, Z). This subgroup is called the Neron-Severi group of X, and is denoted N S(X). The exact sequence (9.1) yields (9.6)
1 0 Jac(X) H 1 (X, ) N S(X) 0.
The Picard group H 1 (X, ) is thus an extension of the complex torus Jac(X) by the nitely generated Z-module N S(X). (9.7) Corollary. The Picard group of Pn is H 1 (Pn , ) Z, and every line bundle over Pn is isomorphic to one of the line bundles (k), k Z.
Proof. We have H k (Pn , ) = H 0,k (Pn , C) = 0 for k 1 by Appl. 8.8, thus Jac(Pn ) = 0 and N S(Pn ) = H 2 (Pn , Z) Z. Moreover, c1 (1) is a generator of H 2 (Pn , Z) in virtue of Th. V-15.10. 9.2. Albanese Variety A proof similar to that of Lemma 9.4 shows that the image of H 2n1 (X, Z) in H n1,n (X, C) via the composite map (9.8) H 2n1 (X, Z) H 2n1 (X, R) H 2n1 (X, C) H n1,n (X, C)
is a lattice. The q-dimensional complex torus (9.9) Alb(X) = H n1,n (X, C)/ Im H 2n1 (X, Z)
is called the Albanese variety of X. We rst give a slightly dierent description of Alb(X), based on the Serre duality isomorphism H n1,n (X, C) H 1,0 (X, C)
H 0 (X, 1 ) . X
(9.10) Lemma. For any compact oriented dierentiable manifold M with dimR M = m, there is a natural isomorphism H1 (M, Z) H m1 (M, Z) where H1 (M, Z) is the rst homology group of M , that is, the abelianization of 1 (M ). Proof. This is a well known consequence of Poincar duality, see e.g. [Spanier 1966]. e We will content ourselves with a description of the morphism. Fix a base point a M . Every homotopy class [] 1 (M, a) can be represented by as a composition of closed loops dieomorphic to S 1 . Let be such a loop. As every oriented vector bundle over S 1 is trivial, the normal bundle to is trivial. Hence (S 1 ) has a neighborhood U dieomorphic to S 1 Rm1 , and there is a dieomorphism : S 1 Rm1 U with m1 S 1 {0} = . Let {0 } Hc (Rm1 , Z) be the fundamental class represented by the Dirac measure 0 (Rm1 ) in De Rham cohomology. Then the cartesian product 1 0 m1 {0 } Hc (S 1 Rm1 , Z) is represented by the current [S 1 ] {0 } (S 1 Rm1 ) 1 and the current of integration over is precisely the direct image current I := ([S 1 ] 0 ) = (1 ) ([S 1 ] 0 ).
316
m1 Its cohomology class {I } Hc (U, R) is thus the image of the class m1 (1 ) 1 {0 } Hc (U, Z).
Therefore, we have obtained a well dened morphism

m1 1 (M, a) Hc (U, Z) H m1 (M, Z),
[] (1 ) 1 {0 }
and the image of [] in H m1 (M, R) is the De Rham cohomology class of the integration current I . Thanks to Lemma 9.10, we can reformulate the denition of the Albanese variety as (9.11) Alb(X) = H 0 (X, 1 ) / Im H1 (X, Z) X

where H1 (X, Z) is mapped to H 0 (X, 1 ) X
by u .
[] I = u
Observe that the integral only depends on the homotopy class [] because all holomorphic 1-forms u on X are closed by Prop. 8.9. We are going to show that there exists a canonical holomorphic map : X Alb(X). Let a be a base point in X. For any x X, we select a path from a to x and associate to x the linear form in H 0 (X, 1 ) dened by I . By construction the class of this linear X form mod Im H1 (X, Z) does not depend on , since I 1 is in the image of H1 (X, Z) for any other path . It is thus legitimate to dene the Albanese map as
x
(9.12)
: X Alb(X),
x u
u
a
mod Im H1 (X, Z).
Of course, if we x a basis (u1 , . . . , uq ) of H 0 (X, 1 ), the Albanese map can be seen in X coordinates as the map (9.13) : X Cq /,
x x
u1 , . . . ,
a a
uq
mod ,
where Cq is the group of periods of (u1 , . . . , uq ) : (9.13 ) =
u1 , . . . ,
uq
; [] 1 (X, a) .
It is then clear that is a holomorphic map. With the original denition (9.9) of the Albanese variety, it is not dicult to see that is the map given by (9.14) : X Alb(X),
n1,n } mod H 2n1 (X, Z), x {I
n1,n where {I } H n1,n (X, C) denotes the (n 1, n)-component of the De Rham cohomology class {I }.
10. Complex Curves
317
10. Complex Curves

We show here how Hodge theory can be used to derive quickly the basic properties of compact manifolds of complex dimension 1 (also called complex curves or Riemann surfaces). Let X be such a curve. We shall always assume in this section that X is compact and connected. Since every positive (1, 1)-form on a curve denes a Khler a metric, the results of 8 and 9 can be applied. 10.1. Riemann-Roch Formula Denoting g = h1 (X, ), we nd H 0 (X, Z) = Z, H 1 (X, ) Cg ,
(10.1) (10.2)
H 1 (X, Z) = Z2g ,
H 0 (X, 1 ) Cg , X
H 2 (X, Z) = Z.
The classication of oriented topological surfaces shows that X is homeomorphic to a sphere with g handles ( = torus with g holes), but this property will not be used in the sequel. The number g is called the genus of X. Any divisor on X can be written = mj aj where (aj ) is a nite sequence of points and mj Z. Let E be a line bundle over X. We shall identify E and the associated locally free sheaf (E). According to V-13.2, we denote by E() the sheaf of germs of meromorphic sections f of E such that div f + 0, i.e. which have a pole of order mj at aj if mj > 0, and which have a zero of order |mj | at aj if mj < 0. Clearly (10.3) E() = E (),
( + ) = () ().
For any point a X and any integer m > 0, there is an exact sequence where = E(m[a])/E is a sheaf with only one non zero stalk a isomorphic to Cm . Indeed, if z is a holomorphic coordinate near a, the stalk a corresponds to the polar parts m k<0 ck z k in the power series expansions of germs of meromorphic sections at point a. We get an exact sequence H 0 X, E(m[a]) Cm H 1 (X, E). When m is chosen larger than dim H 1 (X, E), we see that E(m[a]) has a non zero section and conclude: (10.4) Theorem. Let a be a given point on a curve. Then every line bundle E has non zero meromorphic sections f with a pole at a and no other poles. If is the divisor of a meromorphic section f of E, we have E (), so the map Div(X) H 1 (X, ), () 0 E E(m[a]) 0
is onto (cf. (V-13.8)). On the other hand, Div is clearly a soft sheaf, thus H 1 (X, Div) = 0. The long cohomology sequence associated to the exact sequence 1 Div 0 implies:
318
(10.5) Corollary. On any complex curve, one has H 1 (X, ) = 0 and there is an exact sequence 0 C (X) Div(X) H 1 (X, ) 0. The rst Chern class c1 (E) H 2 (X, Z) can be interpreted as an integer. This integer mj aj , formula V-13.6 shows that the is called the degree of E. If E () with = 2 image of c1 (E) in H (X, R) is the De Rham cohomology class of the associated current [] = mj aj , hence (10.6) c1 (E) =
X
[] =
mj . mj = 0 because the
If mj aj is the divisor of a meromorphic function, we have associated bundle E = ( mj aj ) is trivial.
(10.7) Theorem. Let E be a line bundle on a complex curve X. Then a) H 0 (X, E) = 0 if c1 (E) < 0 or if c1 (E) = 0 and E is non trivial ; b) For every positive (1, 1)-form on X with X = 1, E has a hermitian metric such i that 2 (E) = c1 (E) . In particular, E has a metric of positive (resp. negative) curvature if and only if c1 (E) > 0 (resp. if and only if c1 (E) < 0). Proof. a) If E has a non zero holomorphic section f , then its degree is c1 (E) = X div f 0. In fact, we even have c1 (E) > 0 unless f does not vanish, in which case E is trivial.
i b) Select an arbitrary hermitian metric h on E. Then c1 (E) 2 h (E) is a real (1, 1)-form cohomologous to zero (the integral over X is zero), so Lemma 8.6 c) implies
c1 (E)
i h (E) = id d 2
for some real function (X, R). If we replace the initial metric of E by h = h e , we get a metric of constant curvature c1 (E) . (10.8) Riemann-Roch formula. Let E be a holomorphic line bundle and let hq (E) = dim H q (X, E). Then h0 (E) h1 (E) = c1 (E) g + 1. Moreover h1 (E) = h0 (K E ), where K = 1 is the canonical line bundle of X. X Proof. We claim that for every line bundle F and every divisor we have the equality (10.9) h0 F () h1 F () = h0 (F ) h1 (F ) + [].
X
If we write E = () and apply the above equality with F = formula results from (10.6), (10.9) and from the equalities h0 () = dim H 0 (X, ) = 1,
, the Riemann-Roch
h1 () = dim H 1 (X, ) = g.
10. Complex Curves
319
However, (10.9) need only be proved when 0 : otherwise we are reduced to this case by writing = 1 2 with 1 , 2 0 and by applying the result to the pairs (F, 1 ) and F (), 2 . If = mj aj 0, there is an exact sequence 0 F F () 0
where aj Cmj and the other stalks are zero. Let m = mj = X []. The sheaf is acyclic, because its support {aj } is of dimension 0. Hence there is an exact sequence 0 H 0 (F ) H 0 F () Cm H 1 (F ) H 1 F () 0 and (10.9) follows. The equality h1 (E) = h0 (K E ) is a consequence of the Serre duality theorem H 0,1 (X, E)
H 1,0 (X, E ),
i.e.
H 1 (X, E)
H 0 (X, K E ).
(10.10) Corollary (Hurwitz formula). c1 (K) = 2g 2. Proof. Apply Riemann-Roch to E = K and observe that (10.11) h0 (K) = dim H 0 (X, 1 ) = g X h1 (K) = dim H 1 (X, 1 ) = h1,1 = b2 = 1 X
(10.12) Corollary. For every a X and every m Z h0 K(m[a]) = h1
(m[a])
= h0
(m[a])
m + g 1.
10.2. Jacobian of a Curve By the Neron-Severi sequence (9.6), there is an exact sequence (10.13)
1 0 Jac(X) H 1 (X, ) Z 0,
where the Jacobian Jac(X) is a g-dimensional torus. Choose a base point a X. For every point x X, the line bundle ([x] [a]) has zero rst Chern class, so we have a well-dened map (10.14) a : X Jac(X), x ([x] [a]).
Observe that the Jacobian Jac(X) of a curve coincides by denition with the Albanese variety Alb(X). (10.15) Lemma. The above map a coincides with the Albanese map : X Alb(X) dened in (9.12). Proof. By holomorphic continuation, it is enough to prove that a (x) = (x) when x is near a. Let z be a complex coordinate and let D D be open disks centered at a. Relatively to the covering U1 = D, U2 = X \ D ,
320
the line bundle ([x] [a]) is dened by the Cech cocycle c C 1 (, ) such that c12 (z) = zx za on U12 = D \ D .
On the other hand, we compute (x) by Formula (9.14). The path integral current I[a,x] 1 (X) is equal to 0 on U2 . Lemma I-2.10 implies d (dz/2iz) = 0 dz dz/2i = 0 according to the usual identication of distributions and currents of degree 0, thus
0,1 I[a,x] = d
dz 0,1 I[a,x] 2iz
on U1 .
0,1 [ Therefore {I[a,x] } H 0,1 (X, C) is equal to the Cech cohomology class } in H 1 (X, ) represented by the cocycle
c (z) = 12
1 dw 0,1 I[a,x] = 2iw 2i
x a
and we have c = exp(2ic ) in H 1 (X, ).
dw 1 zx = log wz 2i za
on U12
The nature of a depends on the value of the genus g. A careful examination of a will enable us to determine all curves of genus 0 and 1. (10.16) Theorem. The following properties are equivalent: a) g = 0 ; b) X has a meromorphic function f having only one simple pole p ; c) X is biholomorphic to P1 . Proof. c) = a) is clear. a) = b). Since g = 0, we have Jac(X) = 0. If p, p X are distinct points, the bundle ([p ] [p]) has zero rst Chern class, therefore it is trivial and there exists a meromorphic function f with div f = [p ] [p]. In particular p is the only pole of f , and this pole is simple. b) = c). We may consider f as a map X P1 = C {}. For every value w C, the function f w must have exactly one simple zero x X because X div(f w) = 0 and p is a simple pole. Therefore f : X P1 is bijective and X is biholomorphic to P1 . (10.17) Theorem. The map a is always injective for g 1.
a) If g = 1, a is a biholomorphism. In particular every curve of genus 1 is biholomorphic to a complex torus C/. b) If g 2, a is an embedding. Proof. If a is not injective, there exist points x1 = x2 such that ([x1 ] [x2 ]) is trivial; then there is a meromorphic function f such that div f = [x1 ][x2 ] and Th. 10.16 implies that g = 0.
10. Complex Curves
321
When g = 1, a is an injective map X Jac(X) C/, thus a is open. It follows that a (X) is a compact open subset of C/, so a (X) = C/ and a is a biholomorphism of X onto C/. In order to prove that a is an embedding when g 2, it is sucient to show that the holomorphic 1-forms u1 , . . . , ug do not all vanish at a given point x X. In fact, X has no non constant meromorphic function having only a simple pole at x, thus h0 ([x]) = 1 and Cor. 10.12 implies h0 K([x]) = g 1 < h0 (K) = g. Hence K has a section u which does not vanish at x. 10.3. Weierstrass Points of a Curve We want to study how many meromorphic functions have a unique pole of multiplicity m at a given point a X, i.e. we want to compute h0 (m[a]) . As we shall see soon, these numbers may depend on a only if m is small. We have c1 K(m[a]) = 2g 2 m, so the degree is < 0 and h0 K(m[a]) = 0 for m 2g 1 by 10.7 a). Cor. 10.12 implies (10.18) h0
(m[a])
=mg+1
for m
2g 1.
It remains to compute h0 K(m[a]) for 0 m 2g 2 and g 1. Let u1 , . . . , ug be a basis of H 0 (X, K) and let z be a complex coordinate centered at a. Any germ u (K)a 1 can be written u = U (z) dz with U (z) = mN m! U (m) (a)z m dz. The unique non zero stalk of the quotient sheaf K(m[a]) / K((m + 1)[a]) is canonically isomorphic m+1 to Ka via the map u U (m) (a)(dz)m+1 , which is independant of the choice of z. g 1+2+...+g (K)/(K g[a]) Ka and the Wronskian Hence U1 (z) U1 (z) . . . U1
(g1)
... ...
(10.19)
W (u1 , . . . , ug ) =
Ug (z) Ug (z) . . .
(g1)
dz 1+2+...+g
(z)
. . . Ug
(z)
denes a global section W (u1 , . . . , ug ) H 0 (X, K g(g+1)/2). At the given point a, we can nd linear combinations u1 , . . . , ug of u1 , . . . , ug such that uj (z) = z sj 1 + O(z sj ) dz, s 1 < . . . < sg .
We know that not all sections of K vanish at a and that c1 (K) = 2g 2, thus s1 = 1 and sg 2g 1. We have W (u1 , . . . , ug ) W (z s1 1 dz, . . . , z sg 1 dz) at point a, and an easy induction on sj combined with dierentiation in z yields W (z s1 1 dz, . . . , z sg 1 dz) = C z s1 +...+sg g(g+1)/2 dz g(g+1)/2 for some positive integer constant C. In particular, W (u1 , . . . , ug ) is not identically zero and vanishes at a with multiplicity (10.20) a = s1 + . . . + sg g(g + 1)/2 > 0
322
unless s1 = 1, s2 = 2, . . ., sg = g. Now, we have h0 K(m[a]) = card{j ; sj > m} = g card{j ; sj and Cor. 10.12 gives (10.21) h0 m}
(m[a])
= m + 1 card{j ; sj
m}.
If a is not a zero of W (u1 , . . . , ug ), we nd (10.22) h0 h0
(m[a]) (m[a])
=1 =m+1g
for m g, for m > g.
The zeroes of W (u1 , . . . , ug ) are called the Weierstrass points of X, and the associated Weierstrass sequence is the sequence wm = h0 (m[a]) , m N. We have wm1 wm wm1 + 1 and s1 < . . . < sg are precisely the integers m 1 such that wm = wm1 . The numbers sj {1, 2, . . . , 2g 1} are called the gaps and a the weight of the Weierstrass point a. Since W (u1 , . . . , ug ) is a section of K g(g+1)/2 , Hurwitz formula implies (10.23)
aX
a = c1 (K g(g+1)/2 ) = g(g + 1)(g 1).
In particular, a curve of genus g has at most g(g + 1)(g 1) Weierstrass points.
11. Hodge-Frlicher Spectral Sequence o

Let X be a compact complex n-dimensional manifold. We consider the double complex K p,q = (X, p,q TX ), d = d + d . The Hodge-Frlicher spectral sequence is by o denition the spectral sequence associated to this double complex (cf. IV-11.9). It starts with (11.1)
p,q E1 = H p,q (X, C)
p,q and the limit term E is the graded module associated to a ltration of the De Rham cohomology group H k (X, C), k = p + q. In particular, if the numbers bk and hp,q are still dened as in (8.11), we have
(11.2)
bk =
p+q=k
p,q dim E
p,q dim E1 = p+q=k p+q=k
hp,q .
The equality is equivalent to the degeneration of the spectral sequence at E1 . As a consequence, the Hodge-Frlicher spectral sequence of a compact Khler manifold degenerates o a in E1 .
(11.3) Theorem and Denition. The existence of an isomorphism

k HDR (X, C)
H p,q (X, C)
p+q=k
12. Eect of a Modication on Hodge Decomposition
323
is equivalent to the degeneration of the Hodge-Frlicher spectral sequence at E1 . In this o case, the isomorphism is canonically dened and we say that X admits a Hodge decomposition. In general, interesting informations can be deduced from the spectral sequence. Theorem IV-11.8 shows in particular that (11.4) b1
1,0 However, E2 is the central cohomology group in the sequence 0,0 1,0 2,0 d1 = d : E1 E1 E1 , 1,0 0,1 2,0 dim E2 + (dim E2 dim E2 )+ .
0,0 and as E1 is the space of holomorphic functions on X, the rst map d1 is zero (by the maximum principle, holomorphic functions are constant on each connected component 1,0 0,1 0,1 1,1 of X ). Hence dim E2 h1,0 h2,0 . Similarly, E2 is the kernel of a map E1 E1 , 0,1 thus dim E2 h0,1 h1,1 . By (11.4) we obtain
(11.5)
b1
Another interesting relation concerns the topological Euler-Poincar characteristic e We need the following simple lemma. top (X) = b0 b1 + . . . b2n1 + b2n .
(h1,0 h2,0 )+ + (h0,1 h1,1 h2,0 )+ .
(11.6) Lemma. Let (C , d) a bounded complex of nite dimensional vector spaces over some eld. Then, the Euler characteristic (C ) = (1)q dim C q
is equal to the Euler characteristic H (C ) of the cohomology module. Proof. Set cq = dim C q , Then cq = zq + bq+1 , Therefore we nd (1)q cq = (1)q zq hq = zq bq . (1)q bq = (1)q hq . zq = dim Z q (C ), bq = dim B q (C ), hq = dim H q (C ).
In particular, if the term Er of the spectral sequence of a ltered complex K is a bounded and nite dimensional complex, we have (Er ) = (Er+1 ) = . . . = (E ) = H (K )
l because Er+1 = H (Er ) and dim E = dim H l (K ). In the Hodge-Frlicher spectral o l p,q sequence, we have dim E1 = p+q=l h , hence:
(11.7) Theorem. For any compact complex manifold X, one has top (X) = (1)k bk =
0 k 2n
(1)p+q hp,q .
0 p,q n
324

In this section, we show that the existence of a Hodge decomposition on a compact complex manifold X is guaranteed as soon as there exists such a decomposition on a modication X of X (see II-??.?? for the Denition). This leads us to extend Hodge theory to a class of manifolds which are non necessarily Khler, the so called Fujiki class a a ( ) of manifolds bimeromorphic to Khler manifolds.
p,q 12.1. Sheaf Cohomology Reinterpretation of HBC (X, C) p,q We rst give a description of HBC (X, C) in terms of the hypercohomology of a suitable complex of sheaves. This interpretation, combined with the analogue of the Hodgep,q Frlicher spectral sequence, will imply in particular that HBC (X, C) is always nite o dimensional when X is compact. Let us denote by p,q the sheaf of germs of forms of bidegree (p, q), and by p the sheaf of germs of holomorphic p-forms on X. For a xed bidegree (p0 , q0 ), we let k0 = p0 + q0 and we introduce a complex of sheaves (0 ,q0 , ), p also denoted for simplicity, such that
k = k1 =
p+q=k,p<p0 ,q<q0 p+q=k,p p0 ,q q0
p,q
for k for k
k0 2, k0 .
p,q
The dierential k on k is chosen equal to the exterior derivative d for k = k0 2 (in the case k k0 3, we neglect the components which fall outside k+1 ), and we set k0 2 = d d : k0 2 =
p0 1,q0 1
p0 ,q We nd in particular HBC 0 (X, C) = H k0 1 plexes ( , d ) and ( , d ) dened by
(X) . We observe that has subcom k = 0 k = 0

otherwise, otherwise.
k0 1 =
p0 ,q0
k = k X k = k X
for 0 for 0
k k
If p0 = 0 or q0 = 0 we set instead 0 = C or 0 = C, and take the other components to be zero. Finally, we let = + (the sum is direct except for 0 ); we denote by the sheaf complex dened in the same way as , except that the sheaves p,q are replaced by the sheaves of currents np,nq . (12.1) Lemma. The inclusions
p0 1, q0 1,
induce isomorphisms k ( ) k () k ( ),
Proof. We will prove the result only for the inclusion , the other case is identical. Let us denote by p,q the sheaf of d -closed dierential forms of bidegree (p, q). We consider the ltration Fp (k ) = k
r, r p
and these cohomology sheaves vanish for k = 0, p0 1, q0 1.
325
and the induced ltration on following terms E0 and E1 : if p < p0 if p p0 if p < p0 if p p0

p, E0 :
In the case of
d
, the rst spectral sequence has the

d p,q0 1 p,q
d p, E0 : 0 p,q0 p,q0 +1 p,q p,q p,0 E1 = p , E1 0 1 p,q0 , E1 = 0 for X p,q p,q E1 0 1 = p,q0 , E1 = 0 for q = q0 1.
p,0
p,1
, q = 0, q0 1,
0,
d
The isomorphism in the third line is given by

p,q0 1
/d
p,q0 2
p,q0 1
p,q0
p p The map d1 : E1 0 1,q0 1 E1 0 ,q0 1 is induced by d d acting on p0 1,q0 1 , but thanks to the previous identication, this map becomes d acting on p0 1,q0 . Hence E1 consists of two sequences ,0 E1 : 0 0 1 p0 1 0, X X X ,q E1 0 1 : 0 0,q0 d d
1,q0
p,q0
if these sequences overlap (q0 = 1), only the second one has to be considered. The term E1 in the spectral sequence of has the same rst line, but the second is reduced to 0,q E1 0 1 = dq0 2 (resp. = C for q0 = 1). Thanks to Lemma 12.2 below, we see that the X two spectral sequences coincide in E2 , with at most three non zero terms:
0,0 E2 = C, p E2 0 1,0 = dp0 2 for p0 X
2,
0,q E2 0 1 = dq0 2 for q0 X
2.
Hence
k ( ) k () and these sheaves vanish for k = 0, p0 1, q0 1.

0
,q0 0,q0
(12.2) Lemma. The complex of sheaves

d 1,q0
p,q0
is a resolution of dq0 1 for q0 X Proof. Embed
1, resp. of C for q0 = 0.
in the double complex

p,q
K p,q =
for q < q0 ,
K p,q = 0
for q
q0 .
For the rst tration of K , we nd

p,q E1 0 1 = p,q0
p,q E1 = 0
for q = q0 1
p,q The second tration gives E1 = 0 for q p,0 E1 = H 0 (K ,p ) =
1 and
p, )
H 0( 0
= p X
for p for p
q0 1 q0 ,
p<q0 .
thus the cohomology of
,q0
coincides with that of (p , d)0 X
326
Lemma IV-11.10 and formula (IV-12.9) imply (12.3) Hk (X, ) Hk (X, ) Hk (X, ) H k (X) H k (X)
p,q because the sheaves k and k are soft. In particular, the group HBC (X, C) can be computed either by means of dierential forms or by means of currents. This property also holds for the De Rham or Dolbeault groups H k (X, C), H p,q (X, C), as was already remarked in IV-6. Another important consequence of (12.3) is: p,q (12.4) Theorem. If X is compact, then dim HBC (X, C) < +.
Proof. We show more generally that the hypercohomology groups Hk (X, ) are nite dimensional. As there is an exact sequence 0 C 0 and a corresponding long exact sequence for hypercohomology groups, it is enough to show that the groups Hk (X, ) are nite dimensional. This property is proved for = p by induction on p0 . For p0 = 0 or 1, the complex is reduced to its term 0 0, thus H k (X, C) for p0 = 0 Hk (X, ) = H k (X, 0 ) = H k (X, ) for p0 = 1 and this groups are nite dimensional. In general, we have an exact sequence 0 p0 0 +1 0 0 p p X where p0 denotes the subcomplex of X
0
+1 reduced to one term in degree p0 . As p
Hk (X, p0 ) = H kp0 (X, p0 ) = H p0 ,kp0 (X, C) X X is nite dimensional, the Theorem follows. (12.5) Denition. We say that a compact manifold admits a strong Hodge decomposition if the natural maps
p,q HBC (X, C) H p,q (X, C), p,q HBC (X, C) H k (X, C)
p+q=k
are isomorphisms. This implies of course that there are natural isomorphisms H k (X, C) H p,q (X, C),
p+q=k
H q,p(X, C) H p,q (X, C)
and that the Hodge-Frlicher spectral sequence degenerates in E1 . It follows from 8 o that all Khler manifolds admit a strong Hodge decomposition. a
327
12.2. Direct and Inverse Image Morphisms Let F : X Y be a holomorphic map between complex analytic manifolds of respective dimensions n, m, and r = n m. We have pull-back morphisms (12.6) F : H k (Y, C) H k (X, C), F : H p,q (Y, C) H p,q (X, C), p,q p,q F : HBC (Y, C) HBC (X, C),
commuting with the natural morphisms (8.2), (8.3). Assume now that F is proper. Theorem I-1.14 shows that one can dene direct image morphisms F :
k (X)
k (Y
),
F :
p,q (X)
p,q (Y
),
commuting with d , d . To F therefore correspond cohomology morphisms (12.7) F : H k (X, C) H k2r (Y, C), F : H p,q (X, C) H pr,qr (Y, C), pr,qr p,q (Y, C), F : HBC (X, C) HBC
which commute also with (8.2), (8.3). In addition, I-1.14 c) implies the adjunction formula (12.8) F ( F ) = (F )
whenever is a cohomology class (of any of the three above types) on X, and a cohomology class (of the same type) on Y . 12.3. Modications and the Fujiki Class ( ) Recall that a modication of a compact manifold X is a holomorphic map : X X such that i) X is a compact complex manifold of the same dimension as X ; ii) there exists an analytic subset S X of codimension X \ S is a biholomorphism. 1 such that : X \1 (S)
(12.9) Theorem. If X admits a strong Hodge decomposition, and if : X X is a modication, then X also admits a strong Hodge decomposition. Proof. We rst observe that f = f for every smooth form f on Y . In fact, this property is equivalent to the equality ( f ) g = (f g) = f g
for every smooth form g on Y , and this equality is clear because is a biholomorphism outside sets of Lebesgue measure 0. Consequently, the induced cohomology morphism
328
is surjective and is injective (but these maps need not be isomorphisms). Now, we have commutative diagrams
p,q HBC (X, C) H p,q (X, C), p,q HBC (X, C) H k (X, C) p+q=k p+q=k p,q HBC (X, C) H k (X, C)
p,q HBC (X, C) H p,q (X, C),
with either upward or downward vertical arrows. Hence the surjectivity or injectivity of the top horizontal arrows implies that of the bottom horizontal arrows. (12.10) Denition. A manifold X is said to be in the Fujiki class ( ) if X admits a Khler modication X. a By Th. 12.9, Hodge decomposition still holds for a manifold in the class ( ). We will see later that there exist non-Khler manifolds in ( ), for example all non projective a Moiezon manifolds (cf. ?.?). The class ( ) has been rst introduced in [Fujiki 1978]. s
329
Chapter VII
Positive Vector Bundles and Vanishing Theorems
In this chapter, we prove a few vanishing theorems for hermitian vector bundles over compact complex manifolds. All these theorems are based on an a priori inequality for (p, q)-forms with values in a vector bundle, known as the Bochner-Kodaira-Nakano inequality. This inequality naturally leads to several positivity notions for the curvature of a vector bundle ([Kodaira 1953, 1954], [Griths 1969] and [Nakano 1955, 1973]). The corresponding algebraic notion of ampleness introduced by Grothendieck [EGA 196067] and developped further by [Nakai 1963] and [Hartshorne 1966] is also discussed. The dierential geometric techniques yield optimal vanishing results in the case of line bundles (Kodaira-Akizuki-Nakano and Girbau vanishing theorems) and also some partial results in the case of vector bundles (Nakano vanishing theorem). As an illustration, we compute the cohomology groups H p,q (Pn , (k)) ; much ner results will be obtained in chapters 811. Finally, the Kodaira vanishing theorem is combined with a blowing-up technique in order to establish the projective embedding theorem for manifolds admitting a Hodge metric.
1. Bochner-Kodaira-Nakano Identity
Let (X, ) be a hermitian manifold, dimC X = n, and let E be a hermitian holomorphic vector bundle of rank r over X. We denote by D = D +D its Chern connection (or DE if we want to specify the bundle), and by = + the formal adjoint operator of D. The operators L, of chapter 6 are extended to vector valued forms in p,q T X E by taking their tensor product with IdE . The following result extends the commutation relations of chapter 6 to the case of bundle valued operators. (1.1) Theorem. If is the operator of type (1, 0) dened by = [, d ] on then a) b) c) d)
[E , L] = i(DE + ), [E , L] = i(DE + ), [, DE ]= i(E + ), [, DE ]= i(E + ). , (X, E),
Proof. Fix a point x0 in X and a coordinate system z = (z1 , . . . , zn ) centered at x0 . Then Prop. V-12.?? shows the existence of a normal coordinate frame (e ) at x0 . Given any section s = e (X, E), it is easy to check that the operators DE , E , . . . p,q have Taylor expansions of the type DE s =
d e + O(|z|),
E s =
e + O(|z|), . . .
330
Chapter VII. Positive Vector Bundles and Vanishing Theorems
in terms of the scalar valued operators d, , . . .. Here the terms O(|z|) depend on the curvature coecients of E. The proof of Th. 1.1 is then reduced to the case of scalar valued operators, which is granted by Th. VI-10.1. The Bochner-Kodaira-Nakano identity expresses the antiholomorphic Laplace operator = D + D acting on (X, E) in terms of its conjugate operator , = D + D , plus some extra terms involving the curvature of E and the torsion of the metric (in case is not Khler). Such identities appear frequently in riemannian a geometry (Weitzenbck formula). o (1.2) Theorem. = + [i(E), ] + [D , ] [D , ]. Proof. Equality 1.1 d) yields = i[, D ] , hence = [D , ] = i[D , , D ] [D , ]. The Jacobi identity VI-10.2 and relation 1.1 c) imply D , [, D ] = , [D , D ]] + D , [D , ] = [, (E)] + i[D , + ], taking into account that [D , D ] = D2 = (E). Theorem 1.2 follows. (1.3) Corollary ([Akizuki-Nakano 1955]). If is Khler, then a = + [i(E), ].
In the latter case, is therefore an operator of order 0 closely related to the curvature of E. When is not Khler, Formula 1.2 is not really satisfactory, because it a involves the rst order operators [D , ] and [D , ]. In fact, these operators can be combined with in order to yield a new positive self-adjoint operator . (1.4) Theorem ([Demailly 1985]). The operator = [D + , + ] is a positive and formally self-adjoint operator with the same principal part as the Laplace operator . Moreover = + [i(E), ] + T , where T is an operator of order 0 depending only on the torsion of the hermitian metric : i T = , , d d d , (d ) . 2 Proof. The rst assertion is clear, because the equality (D + ) = + implies the self-adjointness of and u, u = D u + u for any compactly supported form u need two lemmas. (1.5) Lemma. a) [L, ] = 3d ,
2
+ u + u
p,q (X, E).
In order to prove the formula, we
b)
[, ] = 2i .
1. Bochner-Kodaira-Nakano Identity
331
Proof. a) Since [L, d ] = 0, the Jacobi identity implies [L, ] = L, [, d ] = d , [L, ] = 3d , taking into account Cor. VI-10.4 and the fact that d is of degree 3. b) By 1.1 a) we have = i[ , L] D , hence [, ] = i , [ , L] [, D ] = i , [ , L] + + . Using again VI-10.4 and the Jacobi identity, we get , [ , L] = L, [, ] , [L, ] = [d , L],
= [d , ] = .
A substitution in the previous equality gives [, ] = 2i . (1.6) Lemma. The following identities hold: a) [D , ] = [D , ] = [, ], b) [D , ] = [, + ] + T . Proof. a) The Jacobi identity implies D , [, D ] + D , [D , ] + , [D , D ] = 0, hence 2 D , [, D ] = 0 and likewise , [ , L] = 0. Assertion a) is now a consequence of 1.1 a) and d).
i b) In order to verify b), we start from the equality = 2 [, ] provided by Lemma 1.5 b). It follows that
(1.7)
[D , ] =
i D , [, ] . 2
The Jacobi identity will now be used several times. One obtains (1.8) (1.9) D , [, ] = , [, D ] + , [D , ] ; [, D ] = [D , ] = D , [, d ] = , [d , D ] + d , [D , ] = [, d d ] + [d , A]
with A = [D , ] = i( + ). From (1.9) we deduce (1.10) , [, D ] = , [, d d ] + , [d , A] .
Let us compute now the second Lie bracket in the right hand side of (1.10: (1.11) (1.12) , [d , A] = A, [, d ] d , [A, ] = [, A] + d , [, A] ; [, A] = i[, + ] = i[D + , L] .
332
Lemma 1.5 b) provides [, L] = 3d , and it is clear that [D , L] = d . Equalities (1.12) and (1.11) yield therefore [, A] = 2i(d ) ,
(1.13)
, [d , A] = , [D , ] 2i[d , (d ) ].
Substituting (1.10) and (1.13) in (1.8) we get (1.14) D , [, ] = , [, d d ] + 2 , [D , ] 2i d , (d ) = 2i T + [, + ] .
Formula b) is a consequence of (1.7) and (1.14). Theorem 1.4 follows now from Th. 1.2 if Formula 1.6 b) is rewritten + [D , ] [D , ] = [D + , + ] + T . When is Khler, then = T = 0 and Lemma 1.6 a) shows that [D , ] = 0. a Together with the adjoint relation [D , ] = 0, this equality implies (1.15) = + .
When is not Khler, Lemma 1.6 a) can be written [D + , ] = 0 and we obtain more a generally [D + , + ] = (D + ) + D , ( + ) + = + . (1.16) Proposition. Set = [D + , + ]. Then = + .
2. Basic a Priori Inequality

Let (X, ) be a compact hermitian manifold, dimC X = n, and E a hermitian holomorphic vector bundle over X. For any section u (X, E) we have u, u = p,q 0. Theorem 1.4 D u 2 + u 2 and the similar formula for gives u, u implies therefore (2.1) D u
2
+ u
2 X
[i(E), ]u, u + T u, u dV.
This inequality is known as the Bochner-Kodaira-Nakano inequality. When u is harmonic, we get in particular (2.2)
X
[i(E), ]u, u + T u, u dV
0.
These basic a priori estimates are the starting point of all vanishing theorems. Observe that [i(E), ] + T is a hermitian operator acting pointwise on p,q T X E (the
3. Kodaira-Akizuki-Nakano Vanishing Theorem
333
hermitian property can be seen from the fact that this operator coincides with on smooth sections). Using Hodge theory (Cor. VI-11.2), we get: (2.3) Corollary. If the hermitian operator [i(E), ] + T is positive denite on p,q T X E, then H p,q (X, E) = 0. In some circumstances, one can improve Cor. 2.3 thanks to the following analytic continuation lemma due to [Aronszajn 1957]: (2.4) Lemma. Let M be a connected -manifold, F a vector bundle over M , and P a second order elliptic dierential operator acting on (M, F ). Then any section ker P vanishing on a non-empty open subset of M vanishes identically on M . (2.5) Corollary. Assume that X is compact and connected. If [i(E), ] + T Herm p,q T X E is semi-positive on X and positive denite in at least one point x0 X, then we have H p,q (X, E) = 0. Proof. By (2.2) every -harmonic (p, q)-form u must vanish in the neighborhood of x0 where [i(E), ] + T > 0, thus u 0. Hodge theory implies H p,q (X, E) = 0.
3. Kodaira-Akizuki-Nakano Vanishing Theorem

The main goal of vanishing theorems is to nd natural geometric or algebraic conditions on a bundle E that will ensure that some cohomology groups with values in E vanish. In the next three sections, we prove various vanishing theorems for cohomology groups of a hermitian line bundle E over a compact complex manifold X. (3.1) Denition. A hermitian holomorphic line bundle E on X is said to be positive (resp. negative) if the hermitian matrix cjk (z) of its Chern curvature form i(E) = i
1 j,k n
cjk (z) dzj dz k
is positive (resp. negative) denite at every point z X. Assume that X has a Khler metric . Let a 1 (x) ... n (x)
be the eigenvalues of i(E)x with respect to x at each point x X, and let i(E)x = i
1 j n
j (x) j j ,
j Tx X
334
be a diagonalization of i(E)x . By Prop. VI-8.3 we have [i(E), ]u, u =

J,K jJ
j +
jK
1 j n
j |uJ,K |2
(3.2) for any form u =

J,K
(1 + . . . + q p+1 . . . n )|u|2 uJ,K J K p,q T X.
(3.3) Akizuki-Nakano vanishing theorem (1954). Let E be a holomorphic line bundle on X. a) If E is positive, then H p,q (X, E) = 0 b) If E is negative, then H p,q (X, E) = 0 for p + q for p + q n + 1. n 1.
Proof. In case a), choose = i(E) as a Khler metric on X. Then we have j (x) = 1 a for all j and x, so that [i(E), ]u, u (p + q n)||u||2
for any u p,q T X E. Assertion a) follows now from Corollary 2.3. Property b) is proved similarly, by taking = i(E). One can also derive b) from a) by Serre duality (Theorem VI-11.3). When p = 0 or p = n, Th. 3.3 can be generalized to the case where i(E) degenerates at some points. We use here the standard notations (3.4) p = p T X, X KX = n T X, n = dimC X ;
KX is called the canonical line bundle of X. (3.5) Theorem ([Grauert-Riemenschneider 1970]). Let (X, ) be a compact and connected Khler manifold and E a line bundle on X. a a) If i(E) b) If i(E) 0 on X and i(E) > 0 in at least one point x0 X, then H q (X, KX E) = 0 for q 1. 0 on X and i(E) < 0 in at least one point x0 X, then H q (X, E) = 0 for q n 1.
It will be proved in Volume II, by means of holomorphic Morse inequalities, that the Khler assumption is in fact unnecessary. This improvement is a deep result rst proved a by [Siu 1984] with a dierent ad hoc method. Proof. For p = n, formula (3.2) gives (3.6) [i(E), ]u, u (1 + . . . + q )|u|2
and a) follows from Cor. 2.5. Now b) is a consequence of a) by Serre duality.
4. Girbaus Vanishing Theorem
335
4. Girbaus Vanishing Theorem

Let E be a line bundle over a compact connected Khler manifold (X, ). Girbaus a theorem deals with the (possibly everywhere) degenerate semi-positive case. We rst state the corresponding generalization of Th. 4.5. (4.1) Theorem. If i(E) is semi-positive and has at least n s + 1 positive eigenvalues at a point x0 X for some integer s {1, . . . , n}, then H q (X, KX E) = 0 for q s. s.
Proof. Apply 2.5 and inequality (3.6), and observe that q (x0 ) > 0 for all q
(4.2) Theorem ([Girbau 1976]). If i(E) is semi-positive and has at least n s + 1 positive eigenvalues at every point x X, then H p,q (X, E) = 0 for p + q n + s.
Proof. Let us consider on X the new Khler metric a = + i(E), > 0,
and let i(E) = i j j j be a diagonalization of i(E) with respect to and with 1 . . . n . Then = i ( + j ) j j . The eigenvalues of i(E) with respect to are given therefore by (4.3) j, = j /( + j ) [0, 1[, 1 j n. s we have
On the other hand, the hypothesis is equivalent to s > 0 on X. For j j s , thus (4.4) j, = 1 1 + /j 1 1 + /s 1 /s , s j n.
Let us denote the operators and inner products associated to with as an index. Then inequality (3.2) combined with (4.4) implies [i(E), ]u, u
q s + 1) (1 /s ) (n p) |u|2 = p + q n s + 1 (q s + 1)/s |u|2 .
Theorem 4.2 follows now from Cor. 2.3 if we choose < p+qns+1 min s (x). xX qs+1
(4.5) Remark. The following example due to [Ramanujam 1972, 1974] shows that Girbaus result is no longer true for p < n when i(E) is only assumed to have n s + 1 positive eigenvalues on a dense open set.
336
Let V be a hermitian vector space of dimension n + 1 and X the manifold obtained from P (V ) Pn by blowing-up one point a. The manifold X may be described as follows: if P (V /Ca) is the projective space of lines containing a, then X = (x, ) P (V ) P (V /Ca) ; x . We have two natural projections 1 : X P (V ) Pn ,
2 : X Y = P (V /Ca) Pn1 .
1 It is clear that the preimage 1 (x) is the single point x, = (ax) if x = a and that 1 1 (a) = {a} Y Pn1 , therefore
1 : X \ ({a} Y ) P (V ) \ {a} is an isomorphism. On the other hand, 2 is a locally trivial ber bundle over Y with 1 ber 2 () = P1 , in particular X is smooth and n-dimensional. Consider now the line bundle E = 1 (1) over X, with the hermitian metric induced by that of (1). Then E is semi-positive and i(E) has n positive eigenvalues at every point of X \ ({a} Y ), hence the assumption of Th. 4.2 is satised on X \ ({a} Y ). However, we will see that H p,p (X, E) = 0, 0 p n 1,
in contradiction with the expected generalization of (4.2) when 2p n + 1. Let j : Y {a} Y X be the inclusion. Then 1 j : Y {a} and 2 j = IdY ; in particular j E = (1 j) (1) is the trivial bundle Y (1)a . Consider now the composite morphism H p,p (Y, C)H 0 P (V ), (1) H p,p (X, E) H p,p (Y, C) (1)a us 2 u 1 s,
j
given by us (2 j) u(1 j) s = us(a) ; it is surjective and H p,p (Y, C) = 0 for 0 p n 1, so we have H p,p (X, E) = 0.
5. Vanishing Theorem for Partially Positive Line Bundles

Even in the case when the curvature form i(E) is not semi-positive, some cohomology groups of high tensor powers E k still vanish under suitable assumptions. The prototype of such results is the following assertion, which can be seen as a consequence of the Andreotti-Grauert theorem [Andreotti-Grauert 1962], see IX-?.?; the special case where E is > 0 (that is, s = 1) is due to [Kodaira 1953] and [Serre 1956]. (5.1) Theorem. Let F be a holomorphic vector bundle over a compact complex manifold X, s a positive integer and E a hermitian line bundle such that i(E) has at least ns+1 positive eigenvalues at every point x X. Then there exists an integer k0 0 such that H q (X, E k F ) = 0 for q s and k k0 .
5. Vanishing Theorem for Partially Positive Line Bundles
337
Proof. The main idea is to construct a hermitian metric on X in such a way that all negative eigenvalues of i(E) with respect to will be of small absolute value. Let denote a xed hermitian metric on X and let 1 ... n be the corresponding eigenvalues of i(E). (5.2) Lemma. Let (R, R). If A is a hermitian n n matrix with eigenvalues 1 . . . n and corresponding eigenvectors v1 , . . . , vn , we dene [A] as the hermitian matrix with eigenvalues (j ) and eigenvectors vj , 1 j n. Then the map A [A] is on Herm(Cn ). Proof. Although the result is very well known, we give here a short proof. Without loss of generality, we may assume that is compactly supported. Then we have [A] = 1 2
+
(t)eitA dt
t (t u)p uq 0
where is the rapidly decreasing Fourier transform of . The equality p! q!/(p + q + 1)! and obvious power series developments yield
t
du =
DA (e
itA
)B =i
ei(tu)A B eiuA du.

0
Since eiuA is unitary, we get DA (eitA ) |t|. A dierentiation under the integral k sign and Leibniz formula imply by induction on k the bound DA (eitA ) |t|k . Hence A [A] is smooth. Let us consider now the positive numbers t0 = inf s > 0,
X
M = sup max |j | > 0.

X j
We select a function (t) = t for t
(R, R) such that t for 0 t t0 , (t) = M/ for t 0.
t0 , (t)
By Lemma 5.2, := [i(E)] is a smooth hermitian metric on X. Let us write i(E) = i

1 j n
j j j ,
= i
1 j n
(j ) j j
in an orthonormal basis (1 , . . . , n ) of T X for . The eigenvalues of i(E) with respect to are given by j, = j / (j ) and the construction of shows that j, 1, 1 j n, and j, = 1 for s j n. Now, we have H q (X, E k F ) H n,q (X, E k G)
where G = F KX . Let e, (g )1 G and (T X, ) respectively. For r
and (j )1
j n
denote orthonormal frames of E,
u=
|J|=q,
uJ, 1 . . . n J ek g n,q T X E k G,
338
inequality (3.2) yields [i(E), ]u, u
=
J, jJ
j, |uJ, |2
q s + 1 (s 1) |u|2 .
Choosing = 1/s and q s, the right hand side becomes (1/s)|u|2 . Since (E k G) = k(E) IdG +(G), there exists an integer k0 such that i(E k G), + T is positive denite for q s and k acting on n,q T X E k G
k0 . The proof is complete.
6. Positivity Concepts for Vector Bundles

Let E be a hermitian holomorphic vector bundle of rank r over X, where dimC X = n. Denote by (e1 , . . . , er ) an orthonormal frame of E over a coordinate patch X with complex coordinates (z1 , . . . , zn ), and (6.1) i(E) = i
1 j,k n, 1 , r
cjk dzj dz k e e ,
cjk = ckj
the Chern curvature tensor. To i(E) corresponds a natural hermitian form E on T X E dened by E =
j,k,,
cjk (dzj e ) (dzk e ),
and such that E (u, u) =

j,k,,
cjk (x) uj uk ,
u Tx X Ex .(6.2)
(6.3) Denition ([Nakano 1955]). E is said to be Nakano positive (resp. Nakano seminegative) if E is positive denite (resp. semi-negative) as a hermitian form on T X E, i.e. if for every u T X E, u = 0, we have E (u, u) > 0 We write >Nak (resp.
Nak )
(resp.
0).
for Nakano positivity (resp. semi-negativity).
(6.4) Denition ([Griths 1969]). E is said to be Griths positive (resp. Griths semi-negative) if for all Tx X, = 0 and s Ex , s = 0 we have E ( s, s) > 0 We write >Grif (resp.
Grif )
(resp.
0).
for Griths positivity (resp. semi-negativity).
It is clear that Nakano positivity implies Griths positivity and that both concepts coincide if r = 1 (in the case of a line bundle, E is merely said to be positive). One
6. Positivity Concepts for Vector Bundles
339
can generalize further by introducing additional concepts of positivity which interpolate between Griths positivity and Nakano positivity. (6.5) Denition. Let T and E be complex vector spaces of dimensions n, r respectively, and let be a hermitian form on T E. a) A tensor u T E is said to be of rank m if m is the smallest u can be written
m
0 integer such that
u=
j=1
j sj ,
j T, sj E. 0)
b) is said to be m-positive (resp. m-semi-negative) if (u, u) > 0 (resp. (u, u) for every tensor u T E of rank m, u = 0. In this case, we write >m 0 (resp.
m
0).
We say that the bundle E is m-positive if E >m 0. Griths positivity corresponds to m = 1 and Nakano positivity to m min(n, r). (6.6) Proposition. A bundle E is Griths positive if and only if E is Griths negative. Proof. By (V-4.3 ) we get i(E ) = i(E) , hence E (1 s2 , 2 s1 ) = E (1 s1 , 2 s2 ), 1 , 2 T X, s1 , s2 E,
where sj = , sj E . Proposition 6.6 follows immediately. It should be observed that the corresponding duality property for Nakano positive bundles is not true. In fact, using (6.1) we get i(E ) = i cjk dzj dz k e e ,
j,k,,
(6.7)
E (v, v) =
cjk vj v k ,
j,k,,
for any v = vj (/zj ) e T X E . The following example shows that Nakano positivity or negativity of E and E are unrelated. (6.8) Example. Let H be the rank n bundle over Pn dened in V-15. For any u= uj (/zj ) e T X H, v = vj (/zj ) e T X H , 1 j, n, formula (V-15.9) implies H (u, u) =
H (v, v)
(6.9)
uj uj vjj v = vjj .
2
340

Grif
It is then clear that H
0 and H
Nak
0 , but H is neither
Nak
0 nor
Nak
0.
(6.10) Proposition. Let 0 S E Q 0 be an exact sequence of hermitian vector bundles. Then a) E b) E c) E

Grif Grif Nak
0 = Q 0 = S 0 = S
Grif Grif Nak
0, 0, 0,
and analogous implications hold true for strict positivity. Proof. If is written yield dzj j , j hom(S, Q), then formulas (V-14.6) and (V-14.7) i(S) = i(E)S i(Q) = i(E)Q + Since ( s) = j j s and ( s) =
dzj dz k k j , dzj dz k j k . k k s we get
S ( s, s ) = E ( s, s )
j,k
j k j s, k s ,
S (u, u) = E (u, u) | u|2 , Q ( s, s ) = E ( s, s ) +

j,k
j k k s, j s ,
Q ( s, s) = E ( s, s) + | ( s)|2 . Since H is a quotient bundle of the trivial bundle , Example 6.8 shows that E V does not imply Q Nak 0.
Nak
7. Nakano Vanishing Theorem

Let (X, ) be a compact Khler manifold, dimC X = n, and E X a hermitian a vector bundle of rank r. We are going to compute explicitly the hermitian operator [i(E), ] acting on p,q T X E. Let x0 X and (z1 , . . . , zn ) be local coordinates such that (/z1 , . . . , /zn ) is an orthonormal basis of (T X, ) at x0 . One can write x0 = i
1 j n
dzj dz j , cjk dzj dz k e e
i(E)x0 = i
j,k,,
where (e1 , . . . , er ) is an orthonormal basis of Ex0 . Let u=

|J|=p, |K|=q,
uJ,K, dzJ dz K e p,q T X E
x0
7. Nakano Vanishing Theorem
341
A simple computation as in the proof of Prop. VI-8.3 gives u = i(1)p

J,K,,s
uJ,K,
zs
dzJ
z s
dz K e ,
i(E) u = i(1)p [i(E), ]u =
j,k,,,J,K
cjk uJ,K, dzj dzJ dz k dz K e , zk dzJ dz K e z j dz K e
j,k,,,J,K
cjk uJ,K, dzj
+
j,k,,,J,K
cjk uJ,K, dzJ dz k
j,,,J,K
cjj uJ,K, dzJ dz K e .
We extend the denition of uJ,K, to non increasing multi-indices J = (js ), K = (ks ) by deciding that uJ,K, = 0 if J or K contains identical components repeated and that uJ,K, is alternate in the indices (js ), (ks ). Then the above equality can be written [i(E), ]u, u = + cjk uJ,jS, uJ,kS, cjk ukR,K, ujR,K, cjj uJ,K, uJ,K, ,
extended over all indices j, k, , , J, K, R, S with |R| = p 1, |S| = q 1. This hermitian form appears rather dicult to handle for general (p, q) because of sign compensation. Two interesting cases are p = n and q = n. For u = (7.1) uK, dz1 . . . dzn dz K e of type (n, q), we get [i(E), ]u, u =
|S|=q1 j,k,,
cjk ujS, ukS, ,
because of the equality of the second and third summations in the general formula. Since ujS, = 0 for j S, the rank of the tensor (ujS, )j, Cn Cr is in fact min{n q + 1, r}. We obtain therefore: (7.2) Lemma. Assume that E >m 0 in the sense of Def. 6.5. Then the hermitian operator [i(E), ] is positive denite on n,q T X E for q 1 and m min{n q + 1, r}. (7.3) Theorem. Let X be a compact connected Khler manifold of dimension n and a E a hermitian vector bundle of rank r. If E m 0 on X and E >m 0 in at least one point, then H n,q (X, E) = H q (X, KX E) = 0 Similarly, for u = for q 1 and m min{n q + 1, r}.
uJ, dzJ dz 1 . . . dz n e of type (p, n), we get [i(E), ]u, u =

|R|=p1 j,k,,
cjk ukR, ujR, ,
342
because of the equality of the rst and third summations in the general formula. The indices j, k are twisted, thus [i(E), ] denes a positive hermitian form under the assumption i(E) >m 0, i.e. i(E ) <m 0, with m min{n p + 1, r}. Serre duality H p,0 (X, E) = H np,n (X, E ) gives: (7.4) Theorem. Let X and E be as above. If E one point, then H p,0 (X, E) = H 0 (X, p E) = 0 X The special case m = r yields: (7.5) Corollary. For X and E as above: a) Nakano vanishing theorem (1955): E b) E
Nak Nak m
0 on X and E <m 0 in at least min{p + 1, r}.
for p < n and m
0, 0,
strictly in one point strictly in one point
= =
H n,q (X, E) = 0 H p,0 (X, E) = 0
for q
1.
for p < n.
8. Relations Between Nakano and Griths Positivity

It is clear that Nakano positivity implies Griths positivity. The main result of 8 is the following converse to this property [Demailly-Skoda 1979]. (8.1) Theorem. For any hermitian vector bundle E, E >Grif 0 = E det E >Nak 0. To prove this result, we rst use (V-4.2 ) and (V-4.6). If End(E det E) is identied to hom(E, E), one can write (E det E) = (E) + TrE ((E)) IdE , Edet E = E + TrE E h, where h denotes the hermitian metric on E and where TrE E is the hermitian form on T X dened by TrE E (, ) = E ( e , e ), T X,
1 r
for any orthonormal frame (e1 , . . . , er ) of E. Theorem 8.1 is now a consequence of the following simple property of hermitian forms on a tensor product of complex vector spaces. (8.2) Proposition. Let T, E be complex vector spaces of respective dimensions n, r, and h a hermitian metric on E. Then for every hermitian form on T E >Grif 0 = + TrE h >Nak 0.
8. Relations Between Nakano and Griths Positivity
343
We rst need a lemma analogous to Fourier inversion formula for discrete Fourier transforms. (8.3) Lemma. Let q be an integer 3, and x , y , 1 , r, be complex numbers. r Let describe the set Uq of r-tuples of q-th roots of unity and put x =
1 r
x , ,
y = 1 r
y ,
r Uq .
Then for every pair (, ), 1
r, the following identity holds: x y if = , x y if = .
r
r Uq
x y
1 r
Proof. The coecient of x y in the summation q r q r

r Uq
r Uq
x y is given by
This coecient equals 1 when the pairs {, } and {, } are equal (in which case r = 1 for any one of the q r elements of Uq ). Hence, it is sucient to prove that = 0
r Uq
If {, } = {, }, then one of the elements of one of the pairs does not belong to the other pair. As the four indices , , , play the same role, we may suppose for example that {, }. Let us apply to the substitution , where is dened by / = e2i/q , = We get =

when the pairs {, } and {, } are distinct.
for
= .
e2i/q e4i/q
if = ,
if = ,
Since q
3 by hypothesis, it follows that
= 0.
Proof of Proposition 8.2.. Let (tj )1 of E and = j j tj T , u =
be a basis of T , (e )1 r an orthonormal basis j, uj tj e T E. The coecients cjk of

j n
344
with respect to the basis tj e satisfy the symmetry relation cjk = ckj , and we have the formulas (u, u) =
j,k,,
cjk uj uk , cjk j k ,
j,k,
TrE (, ) = ( + TrE h)(u, u) =

r For every Uq (cf. Lemma 8.3), put
cjk uj uk + cjk uj uk .
j,k,,
u = j
1 r
uj C, , e =
u =
j
u t j T j
e E.
Lemma 8.3 implies q r

r Uq
(u e , u e ) = q r =
cjk u u j k
r Uq
cjk uj uk +
j,k,= j,k,,
cjk uj uk . 0, hence
The Griths positivity assumption shows that the left hand side is ( + TrE h)(u, u) with strict positivity if >Grif 0 and u = 0. (8.4) Example. Take E = H over Pn = P (V ). The exact sequence V 0 (1) H 0 cjk uj uk
j,k,
implies det = det H (1). Since det is a trivial bundle, we get (non canonical) V V isomorphisms det H (1), T Pn = H (1) H det H. We already know that H (6.9) shows that
Grif
0, hence T Pn
Nak
0. A direct computation based on
T Pn (u, u) = (H + TrH H h)(u, u) =

1 j,k n
ujk ukj + ujk ujk =
1 2
1 j,k n
|ujk + ukj |2 .
9. Applications to Griths Positive Bundles
345
In addition, we have T Pn >Grif 0. However, the Serre duality theorem gives H q (Pn , KPn T Pn ) H nq (Pn , T Pn ) = H 1,nq (Pn , C) =
C 0
if q = n 1, if q = n 1.
For n 2, Th. 7.3 implies that T Pn has no hermitian metric such that T Pn 2 0 on Pn and T Pn >2 0 in one point. This shows that the notion of 2-positivity is actually stronger than 1-positivity (i.e. Griths positivity). (8.5) Remark. Since TrH H = (1) is positive and T Pn is not >Nak 0 when n 2, we see that Prop. 8.2 is best possible in the sense that there cannot exist any constant c < 1 such that >Grif 0 = + c TrE h Nak 0.
9. Applications to Griths Positive Bundles

We rst need a preliminary result. (9.1) Proposition. Let T be a complex vector space and (E, h) a hermitian vector space of respective dimensions n, r with r 2. Then for any hermitian form on T E and any integer m 1 >Grif 0 = m TrE h >m 0.
Proof. Let us distinguish two cases. a) m = 1. Let u T E be a tensor of rank 1. Then u can be written u = 1 e1 with 1 T, 1 = 0, and e1 E, |e1 | = 1. Complete e1 into an orthonormal basis (e1 , . . . , er ) of E. One gets immediately (TrE h)(u, u) = TrE (1 , 1 ) = (1 e , 1 e )
1 r
> (1 e1 , 1 e1 ) = (u, u). b) m 2. Every tensor u T E of rank u=

1 q
m can be written , T,
with q = min(m, r) and (e )1 r an orthonormal basis of E. Let F be the vector subspace of E generated by (e1 , . . . , eq ) and F the restriction of to T F . The rst part shows that := TrF F h F >Grif 0.
346
Proposition 9.2 applied to on T F yields + TrF h = q TrF F h F >q 0. Since u T F is of rank q m, we get (for u = 0)
(u, u) = F (u, u) < q(TrF F h)(u, u) =q

1 j, q
(j e , j e )
m TrE h(u, u).
Proposition 9.1 is of course also true in the semi-positive case. From these facts, we deduce (9.2) Theorem. Let E be a Griths (semi-)positive bundle of rank r any integer m 1 E (det E)m >m 0 (resp. m 0). Proof. Apply Prop. 8.1 to = E >Grif 0 and observe that det E = det E = TrE . (9.3) Theorem. Let 0 S E Q 0 be an exact sequence of hermitian vector bundles. Then for any m 1 E >m 0 = S (det Q)m >m 0. 2. Then for
Proof. Formulas (V-14.6) and (V-14.7) imply i(S) >m i , i(Q) >m i ,
i(det Q) = TrQ (i(Q)) > TrQ (i ). If we write = dzj j as in the proof of Prop. 6.10, then
idzj dz k TrQ (j k ) idzj dz k TrS (k j ) = TrS (i ). Nak
TrQ (i ) = =
Furthermore, it has been already proved that i the corresponding hermitian form on T X S, we get m TrS (i ) IdS +i and Th. 9.3 follows.
0. By Prop. 8.1 applied to
0,
(9.4) Corollary. Let X be a compact n-dimensional complex manifold, E a vector bundle of rank r 2 and m 1 an integer. Then
10. Cohomology Groups of
(k) over Pn
for q 1; 1 for q
347
b) E >Grif 0 = H n,q X, E (det E)m = 0 and m min{n q + 1, r} ;
a) E >Grif 0 = H n,q (X, E det E) = 0
c) Let 0 S E Q 0 be an exact sequence of vector bundles and m = min{n q + 1, rk S}, q 1. If E >m 0 and if L is a line bundle such that L (det Q)m 0, then H n,q (X, S L) = 0. Proof. Immediate consequence of Theorems 7.3, 8.1, 9.2 and 9.3. Note that under our hypotheses = i TrE (E) = i(r E) is a Khler metric on a X. Corollary 2.5 shows that it is enough in a), b), c) to assume semi-positivity and strict positivity in one point (this is true a priori only if X is supposed in addition to be Khler, a but this hypothesis can be removed by means of Sius result mentioned after (4.5). a) is in fact a special case of a result of [Griths 1969], which we will prove in full generality in volume II (see the chapter on vanishing theorems for ample vector bundles); property b) will be also considerably strengthened there. Property c) is due to [Skoda 1978] for q = 0 and to [Demailly 1982c] in general. Let us take the tensor product of the exact sequence in c) with (det Q)l . The corresponding long cohomology exact sequence implies that the natural morphism H n,q X, E (det Q)l H n,q X, Q (det Q)l is surjective for q 0 and l, m min{n q + 1, rk S}. min{n q, rk S}, bijective for q 1 and l, m
10. Cohomology Groups of (k) over Pn

As an illustration of the above results, we compute now the cohomology groups of all line bundles (k) Pn . This precise evaluation will be needed in the proof of a general vanishing theorem for vector bundles, due to Le Potier (see volume II). As in V-15, we consider a complex vector space V of dimension n + 1 and the exact sequence (10.1) V 0 (1) H 0
of vector bundles over Pn = P (V ). We thus have det = det H (1), and as V T P (V ) = H (1) by Th. V-15.7, we nd (10.2) KP (V ) = det T P (V ) = det H (n) = det (n 1) V
where det is a trivial line bundle. V Before going further, we need some notations. For every integer k N, we consider the homological complex C ,k (V ) with dierential such that p,k p kp V , 0 p k, C (V ) = V S =0 otherwise, (10.3) : p V S kp V p1 V S kp+1 V ,
348
where is the linear map obtained by contraction with the Euler vector eld IdV V V , through the obvious maps V p V p1 V and V S kp V S kp+1 V . If (z0 , . . . , zn ) are coordinates on V , the module C p,k (V ) can be identied with the space of p-forms (z) = I (z) dzI
|I|=p
where the I s are homogeneous polynomials of degree k p. The dierential is given by contraction with the Euler vector eld = 0 j n zj /zj . Let us denote by Z p,k (V ) the space of p-cycles of C ,k (V ), i.e. the space of forms C p,k (V ) such that = 0. The exterior derivative d also acts on C ,k (V ) ; we have d : C p,k (V ) C p+1,k (V ), and a trivial computation shows that d + d = k IdC ,k (V ) . (10.4) Theorem. For k = 0, C ,k (V ) is exact and there exist canonical isomorphisms C ,k (V ) = p V S kp V Z p,k (V ) Z p1,k (V ). Proof. The identity d + d = k Id implies the exactness. The isomorphism is given by 1 1 k d and its inverse by 1 + k d 2 . Let us consider now the canonical mappings : V \ {0} P (V ), : V \ {0} (1)
dened in V-15. As T[z] P (V ) V /C(z) for all z V \ {0}, every form Z p,k (V ) denes a holomorphic section of p T P (V ) , (z) being homogeneous of degree k with respect to z. Hence (z) (z)k is a holomorphic section of p T P (V ) (k) , and since its homogeneity degree is 0, it induces a holomorphic section of p T P (V ) (k). We thus have an injective morphism (10.5) Z p,k (V ) H p,0 P (V ), (k) .
(10.6) Theorem. The groups H p,0 P (V ), (k) are given by b) H p,0 P (V ), (k) = 0 a) H p,0 P (V ), (k) Z p,k (V ) for k p 0,
for k
p and (k, p) = (0, 0).
Proof. Let s be a holomorphic section of p T P (V ) (k). Set (z) = (dz ) s([z]) (z)k , z V \ {0}.
Then is a holomorphic p-form on V \ {0} such that = 0, and the coecients of are homogeneous of degree k p on V \ {0} (recall that dz = 1 dz ). It follows that = 0 if k < p and that Z p,k (V ) if k p. The injective morphism (10.5) is therefore also surjective. Finally, Z p,p (V ) = 0 for p = k = 0, because of the exactness of C ,k (V ) when k = 0. The proof is complete.
11. Ample Vector Bundles
349
(10.7) Theorem. The cohomology groups H p,q P (V ), (k) vanish in the following cases: a) q = 0, n, p ; b) q = 0, k c) q = n, k p and (k, p) = (0, 0) ; n + p and (k, p) = (0, n) ;
d) q = p = 0, n, k = 0. The remaining non vanishing groups are: b) H p,0 P (V ), (k) Z p,k (V ) d) H p,p P (V ), C = C, 0 p for k > p ; for k < n + p ; c) H p,n P (V ), (k) Z np,k (V ) n.
Proof. d) is already known, and so is a) when k = 0 (Th. VI-13.3). b) and b) follow from Th. 10.6, and c), c) are equivalent to b), b) via Serre duality: H p,q P (V ), (k)
= H np,nq P (V ), (k) ,
thanks to the canonical isomorphism Z p,k (V )
= Z p,k (V ).
Let us prove now property a) when k = 0 and property d). By Serre duality, we may assume k > 0. Then p T P (V ) KP (V ) np T P (V ). It is very easy to verify that E T P (V ) Nak 0, we get therefore
Nak
0 implies s E
Nak
0 for every integer s. Since
F = np T P (V ) (k) >Nak 0 and the Nakano vanishing theorem implies
for k > 0,
H p,q P (V ), (k) = H q P (V ), p T P (V ) (k) = H q P (V ), KP (V ) F = 0,
1.

11.A. Globally Generated Vector Bundles All denitions concerning ampleness are purely algebraic and do not involve dierential geometry. We shall see however that ampleness is intimately connected with the dierential geometric notion of positivity. For a general discussion of properties of ample vector bundles in arbitrary characteristic, we refer to [Hartshorne 1966]. (11.1) Denition. Let E X be a holomorphic vector bundle over an arbitrary complex manifold X.
350
a) E is said to be globally generated if for every x X the evaluation map H 0 (X, E) Ex is onto. b) E is said to be semi-ample if there exists an integer k0 such that S k E is globally generated for k k0 . Any quotient of a trivial vector bundle is globally generated, for example the tautological quotient vector bundle Q over the Grassmannian Gr (V ) is globally generated. Conversely, every globally generated vector bundle E of rank r is isomorphic to the quotient of a trivial vector bundle of rank n + r, as shown by the following result. (11.2) Proposition. If a vector bundle E of rank r is globally generated, then there exists a nite dimensional subspace V H 0 (X, E), dim V n+r, such that V generates all bers Ex , x X. = Proof. Put an arbitrary hermitian metric on E and consider the Frchet space e 0 n+r H (X, E) of (n + r)-tuples of holomorphic sections of E, endowed with the topology of uniform convergence on compact subsets of X. For every compact set K X, we set A(K) = {(s1 , . . . , sn+r ) which do not generate E on K}. : indeed, Baires theorem will It is enough to prove that A(K) is of rst category in imply that A(X) = A(K ) is also of rst category, if (K ) is an exhaustive sequence of compact subsets of X. It is clear that A(K) is closed, because A(K) is characterized by the closed condition |si1 sir | = 0. min
K i1 <<ir
It is therefore sucient to prove that A(K) has no interior point. By hypothesis, each ber Ex , x K, is generated by r global sections s , . . . , s . We have in fact s s = r 1 r 1 0 in a neighborhood Ux of x. By compactness of K, there exist nitely many sections s , . . . , s which generate E in a neighborhood of the set K. 1 N If T is a complex vector space of dimension r, dene Rk (T p ) as the set of p-tuples (x1 , . . . , xp ) T p of rank k. Given a Rk (T p ), we can reorder the p-tuple in such a way that a1 ak = 0. Complete these k vectors into a basis (a1 , . . . , ak , b1 , . . . , brk ) of T . For every point x T p in a neighborhood of a, then (x1 , . . . , xk , b1 , . . . , brk ) is again a basis of T . Therefore, we will have x Rk (T p ) if and only if the coordinates of xl , k + 1 l N , relative to b1 , . . . , brk vanish. It follows that Rk (T p ) is a (non closed) submanifold of T p of codimension (r k)(p k). Now, we have a surjective ane bundle-morphism : CN(n+r) E n+r (x, ) sj (x) +
1 k N
jk s (x) k
1 j n+r
Therefore 1 (Rk (E n+r )) is a locally trivial dierentiable bundle over , and the codimension of its bers in CN(n+r) is (r k)(n + r k) n + 1 if k < r ; it follows that the dimension of the total space 1 (Rk (E n+r )) is N (n + r) 1. By Sards theorem
1 Rk (E n+r )
k<r
351
is of zero measure in CN(n+r) . This means that for almost every value of the parameter the vectors sj (x) + k jk s (x) Ex , 1 j n + r, are of maximum rank r at each k point x . Therefore A(K) has no interior point. Assume now that V H 0 (X, E) generates E on X. Then there is an exact sequence (11.3) 0 S E 0 V
of vector bundles over X, where Sx = {s V ; s(x) = 0}, codimV Sx = r. One obtains therefore a commutative diagram E X V Q
(11.4)
V Gr (V )
where V , V are the holomorphic maps dened by V (x) = Sx , x X,
V (u) = {s V ; s(x) = u} V /Sx , u Ex . In particular, we see that every globally generated vector bundle E of rank r is the pullback of the tautological quotient vector bundle Q of rank r over the Grassmannian by means of some holomorphic map X Gr (V ). In the special case when rk E = r = 1, the above diagram becomes E X V
(11.4 )
V P (V )
(1)
(11.5) Corollary. If E is globally generated, then E possesses a hermitian metric such that E Grif 0 (and also E Nak 0). Proof. Apply Prop. 6.11 to the exact sequence (11.3), where is endowed with an V arbitrary hermitian metric. When E is of rank r = 1, then S k E = E k and any hermitian metric of E k yields a metric on E after extracting k-th roots. Thus: (11.6) Corollary. If E is a semi-ample line bundle, then E 0.
In the case of vector bundles (r 2) the answer is unknown, mainly because there is no known procedure to get a Griths semipositive metric on E from one on S k E. 11.B. Ampleness We are now turning ourselves to the denition of ampleness. If E X is a holomorphic vector bundle, we dene the bundle J k E of k-jets of sections of E by (J k E)x = x (E)/ k+1 x (E) for every x X, where x is the maximal ideal x of x . Let (e1 , . . . , er ) be a holomorphic frame of E and (z1 , . . . , zn ) analytic coordinates
352
on an open subset X. The ber (J k E)x can be identied with the set of Taylor developments of order k : c, (z x) e (z),
1 r,|| k
and the coecients c, dene coordinates along the bers of J k E. It is clear that the choice of another holomorphic frame (e ) would yield a linear change of coordinates (c, ) with holomorphic coecients in x. Hence J k E is a holomorphic vector bundle of rank r n+k . n (11.7) Denition. a) E is said to be very ample if all evaluation maps H 0 (X, E) (J 1 E)x , are surjective. b) E is said to be ample if there exists an integer k0 such that S k E is very ample for k k0 . (11.8) Example. (1) Pn is a very ample line bundle (immediate verication). Since the pull-back of a (very) ample vector bundle by an embedding is clearly also (very) ample, diagram (V-16.8) shows that r Q Gr (V ) is very ample. However, Q itself cannot be very ample if r 2, because dim H 0 (Gr (V ), Q) = dim V = d, whereas rank(J 1 Q) = (rank Q) 1 + dim Gr (V ) = r 1 + r(d r) > d if r 2. H 0 (X, E) Ex Ey , x, y X, x = y,
(11.9) Proposition. If E is very ample of rank r, there exists a subspace V of H 0 (X, E), dim V max nr + n + r, 2(n + r) , such that all the evaluation maps V Ex Ey , x = y, and V (J 1 E)x , x X, are surjective. Proof. The arguments are exactly the same as in the proof of Prop. 11.4, if we consider instead the bundles J 1 E X and E E X X \ X of respective ranks r(n + 1) and 2r, and sections s , . . . , s H 0 (X, E) generating these bundles. 1 N (11.10) Proposition. Let E X be a holomorphic vector bundle. a) If V H 0 (X, E) generates J 1 E X and E E X X \ X , then V is an embedding. b) Conversely, if rank E = 1 and if there exists V H 0 (X, E) generating E such that V is an embedding, then E is very ample.
Proof. b) is immediate, because E = V ((1)) and (1) is very ample. Note that the result is false for r 2 as shown by the example E = Q over X = Gr (V ).
a) Under the assumption of a), it is clear since Sx = {s V ; s(x) = 0} that Sx = Sy implies x = y, hence V is injective. Therefore, it is enough to prove that the map x Sx
353
has an injective dierential. Let x X and W V such that Sx W = V . Choose a coordinate system in a neighborhood of x in X and a small tangent vector h Tx X. The element Sx+h Gr (V ) is the graph of a small linear map u = O(|h|) : Sx W . Thus we have Sx+h = {s = s + t V ; s Sx , t = u(s) W, s (x + h) = 0}. Since s(x) = 0 and |t| = O(|h|), we nd s (x + h) = s (x) + dx s h + O(|s | |h|2 ) = t(x) + dx s h + O(|s| |h|2 ), thus s (x + h) = 0 if and only if t(x) = dx s h + O(|s| |h|2 ). Thanks to the ber isomorphism V : Ex V /Sx W , t(x) t mod Sx , we get: u(s) = t = V (t(x)) = V dx s h + O(|s| |h|2 ) . Recall that Ty Gr (V ) = hom(Sy , Qy ) = hom(y, V /y) (see V-16.5) and use these identications at y = Sx . It follows that (11.11) (dx V ) h = u = Sx V /Sx , s V (dx s h) ,
Now hypothesis a) implies that Sx s dx s hom(Tx X, Ex ) is onto, hence dx V is injective. (11.12) Corollary. If E is an ample line bundle, then E > 0. Proof. If E is very ample, diagram (11.4 ) shows that E is the pull-back of (1) by the embedding V , hence i(E) = V i((1)) > 0 with the induced metric. The ample case follows by extracting roots. (11.13) Corollary. If E is a very ample vector bundle, then E carries a hermitian metric such that E <Nak 0, in particular E >Grif 0. Proof. Choose V as in Prop. 11.9 and select an arbitrary hermitian metric on V . Then diagram 11.4 yields E = V Q, hence E = Q . By formula (V-16.9) we have for every V T Gr (V ) = hom(S, Q) and t Q : Q ( t, t) = jk lk tl tj =
j,k,l k j
tj jk
= | , t |2 .
Let h Tx X, t Ex . Thanks to formula (11.11), we get E (h t, h t) = Q (dx V h) V (t), (dx V h) V (t) = , V (t) (dx V h)
2
= Sx s dx s h, t
= Sx s V (dx s h), V (t) 0.
As Sx s dx s T X E is surjective, it follows that E (h t, h t) = 0 when h = 0, t = 0. Now, dx s denes a linear form on T X E and the above formula for the curvature of E clearly yields E (u, u) = |Sx s dx s u|2 < 0 if u = 0.
354
(11.14) Problem ([Griths 1969]). If E is an ample vector bundle over a compact manifold X, then is E >Grif 0 ?
Griths problem has been solved in the armative when X is a curve (see [Umemura 1973] and also [Campana-Flenner 1990]), but the general case is still unclear and seems very deep. The next sections will be concerned with the important result of Kodaira asserting the equivalence between positivity and ampleness for line bundles.
12. Blowing-up along a Submanifold

Here we generalize the blowing-up process already considered in Remark 4.5 to arbitrary manifolds. Let X be a complex n-dimensional manifold and Y a closed submanifold with codimX Y = s.
(12.1) Notations. The normal bundle of Y in X is the vector bundle over Y dened as the quotient N Y = (T X)Y /T Y . The bers of N Y are thus given by Ny Y = Ty X/Ty Y at every point y Y . We also consider the projectivized normal bundle P (N Y ) Y whose bers are the projective spaces P (Ny Y ) associated to the bers of N Y .
The blow-up of X with center Y (to be constructed later) is a complex n-dimensional manifold X together with a holomorphic map : X X such that:
i) E := 1 (Y ) is a smooth hypersurface in X, and the restriction : E Y is a holomorphic ber bundle isomorphic to the projectivized normal bundle P (N Y ) Y .
ii) : X \ E X \ Y is a biholomorphism.
In order to construct X and , we rst dene the set-theoretic underlying objects as the disjoint sums
X= (X \ Y ) E, = IdX\Y ,
where E := P (N Y ), where : E Y.
355
VII-1 Blow-up of one point in X. This means intuitively that we have replaced each point y Y by the projective space of all directions normal to Y . When Y is reduced to a single point, the geometric picture is given by Fig. 1 below. In general, the picture is obtained by slicing X transversally to Y near each point and by blowing-up each slice at the intersection point with Y . It remains to construct the manifold structure on X and in particular to describe what are the holomorphic functions near a point of E. Let f, g be holomorphic functions on an open set U X such that f = g = 0 on Y U . Then df and dg vanish on T YY U , hence df and dg induce linear forms on N YY U . The holomorphic function h(z) = f (z)/g(z) on the open set Ug := z U ; g(z) = 0 U \ Y can be extended in a natural way to a function h on the set Ug = Ug (z, []) P (N Y )Y U ; dgz () = 0 X by letting h(z, []) = dfz () , dgz () (z, []) P (N Y )Y U .
Using this observation, we now dene the manifold structure on X by giving explicitly an atlas. Every coordinate chart of X\Y is taken to be also a coordinate chart of X. Furthermore, for every point y0 Y , there exists a neighborhood U of y0 in X and a coordinate chart (z) = (z1 , . . . , zn ) : U Cn centered at y0 such that (U ) = B B for some balls B Cs , B Cns , and such that Y U = 1 ({0} B ) = {z1 = . . . =zs =0}.
356
It follows that (zs+1 , . . . , zn ) are local coordinates on Y U and that the vector elds (/z1 , . . . , /zs) yield a holomorphic frame of N YY U . Let us denote by (1 , . . . , s ) the corresponding coordinates along the bers of N Y . Then (1 , . . . , s , zs+1 , . . . , zn ) are coordinates on the total space N Y . For every j = 1, . . . , s, we set Uj = Uzj = z U \ Y ; zj = 0 (z, []) P (N Y )Y U ; j = 0 . Then (Uj )1 j s is a covering of U = 1 (U ) and for each j we dene a coordinate chart j = (w1 , . . . , wn ) : Uj Cn by wk := zk zj
for 1
s, k = j ;
wk := zk
for k > s or k = j.
For z U \ Y , resp. (z, []) P (N Y )Y U , we get z1 zj1 zj+1 zs ,..., , zj , , . . . , , zs+1 , . . . , zn , zj zj zj zj j1 j+1 s 1 ,..., , 0, , . . . , , s+1 , . . . , n . j (z, []) = (w1 , . . . , wn ) = j j j j j (z) = (w1 , . . . , wn ) = With respect to the coordinates (wk ) on Uj and (zk ) on U , the map is given by (12.2) Uj U w (w1 wj , . . . , wj1 wj ; wj ; wj+1 wj , . . . , ws wj ; ws+1 , . . . , wn )
1 where j = j , thus is holomorphic. The range of the coordinate chart j is 1 j (Uj ) = j (U ) , so it is actually open in Cn . Furthermore E Uj is dened by the single equation wj = 0, thus E is a smooth hypersurface in X. It remains only to verify that the coordinate changes w w associated to any coordinate change z z on X are holomorphic. For that purpose, it is sucient to verify that (f /g) is holomorphic in (w1 , . . . , wn ) on Uj Ug . As g vanishes on Y U , we can write g(z) = 1 k s zk Ak (z) for some holomorphic functions Ak on U . Therefore j
g(z) = Aj (j (w)) + zj
wk Ak (j (w))
k=j
has an extension (g/zj ) to Uj which is a holomorphic function of the variables (w1 , . . . , wn ). Since (g/zj ) (z, []) = dgz ()/j on E Uj , it is clear that Uj Ug = w Uj ; (g/zj ) (w) = 0 . Hence Uj Ug is open in Ug and (f /g) = (f /zj ) /(g/zj ) is holomorphic on Uj Ug . (12.3) Denition. The map : X X is called the blow-up of X with center Y and E = 1 (Y ) P (N Y ) is called the exceptional divisor of X.
357
According to (V-13.5), we denote by (E) the line bundle on X associated to the divisor E and by h H 0 (X, (E)) the canonical section such that div(h) = [E]. On the other hand, we denote by P (NY ) (1) (N Y ) the tautological line subbundle over E = P (N Y ) such that the ber above the point (z, []) is C Nz Y . (12.4) Proposition. a)
(E) enjoys the following properties: (E)E is isomorphic to P (NY ) (1).
b) Assume that X is compact. For every positive line bundle L over X, the line bundle (E) (Lk ) over X is positive for k > 0 large enough. Proof. a) The canonical section h H 0 (X, (E)) vanishes at order 1 along E, hence the kernel of its dierential dh : (T X)E (E)E
is T E. We get therefore an isomorphism N E (E)E . Now, the map : X X satises (E) Y , so its dierential d : T X (T X) is such that d(T E) (T Y ). Therefore d induces a morphism (12.5) N E (N Y ) = (N Y )
of vector bundles over E. The vector eld /wj yields a non vanishing section of N E on Uj , and (12.2) implies dj wj = + zj wk
1 k s,k=j
zk
//
1 k s
zk
P (NY ) (1) (N Y ), hence

(12.6)
at every point (z, []) E. This shows that (12.5) is an isomorphism of N E onto
(E)E N E P (NY ) (1).
b) Select an arbitrary hermitian metric on T X and consider the induced metrics on N Y and on P (NY ) (1) E = P (N Y ). The restriction of P (NY ) (1) to each ber P (Nz Y ) is the standard line bundle (1) over Ps1 ; thus by (V-15.10) this restriction has a positive denite curvature form. Extend now the metric of P (NY ) (1) on E to a metric of (E) on X in an arbitrary way. If F = (E) (Lk ), then (F ) = ((E)) + k (L), thus for every t T X we have F (t, t) = (E) (t, t) + k L d(t), d(t) . By the compactness of the unitary tangent bundle to X and the positivity of L , it is sucient to verify that (E) (t, t) > 0 for every unit vector t Tz X such that d(t) = 0. However, from the computations of a), this can only happen when z E and t T E, and in that case d(t) = d(t) = 0, so t is tangent to the ber P (Nz Y ). Therefore (E) (t, t) = P (N Y ) (1) (t, t) > 0.
358
(12.7) Proposition. The canonical line bundle of X is given by KX = (s 1)E KX , where s = codimX Y.
Proof. KX is generated on U by the holomorphic n-form dz1 . . . dzn . Using (12.2), we see that KX is generated on Uj by
s1 (dz1 . . . dzn ) = wj dw1 . . . dwn .
Since the divisor of the section h H 0 (X, (E)) is the hypersurface E dened by the equation wj = 0 in Uj , we have a well dened line bundle isomorphism KX (1 s)E KX , h1s ().
13. Equivalence of Positivity and Ampleness for Line Bundles

We have seen in section 11 that every ample line bundle carries a hermitian metric of positive curvature. The converse will be a consequence of the following result. (13.1) Theorem. Let L X be a positive line bundle and Lk the k-th tensor power of L. For every N -tuple (x1 , . . . , xN ) of distinct points of X, there exists a constant C > 0 such that the evaluation maps H 0 (X, Lk ) (J m Lk )x1 (J m Lk )xN are surjective for all integers m 0, k C(m + 1).
(13.2) Lemma. Let : X X be the blow-up of X with center the nite set Y = {x1 , . . . , xN }, and let (E) be the line bundle associated to the exceptional divisor E. Then H 1 (X, (mE) Lk ) = 0 for m 1, k Cm and C 0 large enough. Proof. By Prop. 12.7 we get KX = (n 1)E KX and H 1 X, (mE) Lk = H n,1 X, K 1 (mE) Lk = H n,1 X, F
X 1 where F = (m + n 1)E (KX Lk ), so the conclusion will follow from the Kodaira-Nakano vanishing theorem if we can show that F > 0 when k is large enough. Fix an arbitrary hermitian metric on KX . Then
(F ) = (m + n 1)((E)) + k(L) (KX ) . There is k0 0 such that i k0 (L) (KX ) > 0 on X, and Prop. 12.4 implies the existence of C0 > 0 such that i ((E)) + C0 (L) > 0 on X. Thus i(F ) > 0 for m 2 n and k k0 + C0 (m + n 1).
14. Kodairas Projectivity Criterion
359
Proof of Theorem 13.1.. Let vj H 0 (j , Lk ) be a holomorphic section of Lk in a neighborhood j of xj having a prescribed m-jet at xj . Set v(x) =
j
j (x)vj (x)
where j = 1 in a neighborhood of xj and j has compact support in j . Then d v = d j vj vanishes in a neighborhood of x1 , . . . , xN . Let h be the canonical section of (E)1 such that div(h) = [E]. The (0, 1)-form dv vanishes in a neighborhood of E = h1 (0), hence w = h(m+1) d v
0,1
X, ((m + 1)E) Lk .
and w is a d -closed form. By Lemma 13.2 there exists a smooth section u

0,0
X, ((m + 1)E) Lk
such that d u = w = h(m+1) d v. This implies v hm+1 u H 0 (X, Lk ), and since L is trivial near E, there exists a section g H 0 (X, Lk ) such that g = v hm+1 u. As h vanishes at order 1 along E, the m-jet of g at xj must be equal to that of v (or vj ). (13.3) Corollary. For any holomorphic line bundle L X, the following conditions are equivalent: a) L is ample; b) L > 0, i.e. L possesses a hermitian metric such that i(L) > 0. Proof. a) = b) is given by Cor. 11.12, whereas b) = a) is a consequence of Th. 13.1 for m = 1.

The following fundamental projectivity criterion is due to [Kodaira 1954]. (14.1) Theorem. Let X be a compact complex manifold, dimC X = n. The following conditions are equivalent. a) X is projective algebraic, i.e. X can be embedded as an algebraic submanifold of the complex projective space PN for N large. b) X carries a positive line bundle L. c) X carries a Hodge metric, i.e. a Khler metric with rational cohomology class a {} H 2 (X, Q).
Proof. a) = b). Take L = (1)X .
360

i 2 (L) ;
b) = c). Take =
then {} is the image of c1 (L) H 2 (X, Z).
c) = b). We can multiply {} by a common denominator of its coecients and suppose that {} is in the image of H 2 (X, Z). Then Th. V-13.9 b) shows that there exists a i hermitian line bundle L such that 2 (L) = > 0. b) = a). Corollary 13.3 shows that F = Lk is very ample for some integer k > 0. Then Prop. 11.9 enables us to nd a subspace V of H 0 (X, F ), dim V 2n + 2, such that V : X G1 (V ) = P (V ) is an embedding. Thus X can be embedded in P2n+1 and Chows theorem II-7.10 shows that the image is an algebraic set in P2n+1 . (14.2) Remark. The above proof shows in particular that every n-dimensional projective manifold X can be embedded in P2n+1 . This can be shown directly by using generic projections PN P2n+1 and Whitney type arguments as in 11.2. (14.3) Corollary. Every compact Riemann surface X is isomorphic to an algebraic curve in P3 . Proof. Any positive smooth form of type (1, 1) is Khler, and is in fact a Hodge a metric if we normalize its volume so that X = 1. This example can be somewhat generalized as follows. (14.4) Corollary. Every Khler manifold (X, ) such that H 2 (X, ) = 0 is projective. a Proof. By hypothesis H 0,2 (X, C) = 0 = H 2,0 (X, C), hence H 2 (X, C) = H 1,1 (X, C) admits a basis {1 }, . . . , {N } H 2 (X, Q) where 1 , . . . , N are harmonic real (1, 1)forms. Since {} is real, we have {} = 1 {1 } + . . . + N {N }, j R, thus = 1 1 + . . . + N N because itself is harmonic. If 1 , . . . , N are rational numbers suciently close to 1 , . . . , N , then := 1 1 + N N is close to , so is a positive denite d-closed (1, 1)-form, and {} H 2 (X, Q). We obtain now as a consequence the celebrated Riemann criterion characterizing abelian varieties ( = projective algebraic complex tori). (14.5) Corollary. A complex torus X = Cn / ( a lattice of Cn ) is an abelian variety if and only if there exists a positive denite hermitian form h on Cn such that Im h(1 , 2 ) Z for all 1 , 2 .
Proof (Suciency of the condition).. Set = Im h. Then denes a constant Khler a n n metric on C , hence also on X = C /. Let (a1 , . . . , a2n ) be an integral basis of the lattice . We denote by Tj , Tjk the real 1- and 2-tori Tj = (R/Z)aj , 1 j n, Tjk = Tj Tk , 1 j<k 2n.
361
Topologically we have X T1 . . . T2n , so the Knneth formula IV-15.7 yields u H (X, Z) H 2 (X, Z)
1 j 2n
H 0 (Tj , Z) H 1 (Tj , Z) , H 1 (Tj , Z) H 1 (Tk , Z) H 2 (Tjk , Z)

1 j<k 2n
1 j<k 2n
where the projection H 2 (X, Z) H 2 (Tjk , Z) is induced by the injection Tjk X. In the identication H 2 (Tjk , R) R, we get (14.6) {} Tjk =
Tjk
= (aj , ak ) = Im h(aj , ak ).
The assumption on h implies {} Tjk H 2 (Tjk , Z) for all j, k, therefore {} H 2 (X, Z) and X is projective by Th. (14.1). Proof (Necessity of the condition).. If X is projective, then X admits a Khler metric a 2 such that {} is in the image of H (X, Z). In general, is not invariant under the translations x (y) = y x of X. Therefore, we replace by its mean value: = 1 Vol(X)
xX (x ) dx,
which has the same cohomology class as (x is homotopic to the identity). Now is the imaginary part of a constant positive denite hermitian form h on Cn , and formula (14.6) shows that Im h(aj , ak ) Z. (14.7) Example. Let X be a projective manifold. We shall prove that the Jacobian Jac(X) and the Albanese variety Alb(X) (cf. VI-13 for denitions) are abelian varieties. In fact, let be a Khler metric on X such that {} is in the image of H 2 (X, Z) and a let h be the hermitian metric on H 1 (X, ) H 0,1 (X, C) dened by h(u, v) =
X
2i u v n1
for all closed (0, 1)-forms u, v. As 2i u v n1 = 2 2 n |u| , n
we see that h is a positive denite hermitian form on H 0,1 (X, C). Consider elements j H 1 (X, Z), j = 1, 2. If we write j = j + j in the decomposition H 1 (X, C) = H 1,0 (X, C) H 0,1 (X, C), we get
h(1 , 2 ) = Im h(1 , 2 ) = X 2i 1 2 n1 , (1 2 + 1 2 ) n1 =
1 2 n1 Z.
362
Therefore Jac(X) is an abelian variety. Now, we observe that H n1,n (X, C) is the anti-dual of H 0,1 (X, C) by Serre duality. We select on H n1,n (X, C) the dual hermitian metric h . Since the Poincar bilinear e pairing yields a unimodular bilinear map H 1 (X, Z) H 2n1 (X, Z) Z,
we easily conclude that Im h (1 , 2 ) Q for all 1 , 2 H 2n1 (X, Z). Therefore Alb(X) is also an abelian variety.
363
Chapter VIII
L2 Estimates on Pseudoconvex Manifolds
The main goal of this chapter is to show that the dierential geometric technique that has been used in order to prove vanishing theorems also yields very precise L2 estimates for the solutions of equations d u = v on pseudoconvex manifolds. The central idea, due to [Hrmander 1965], is to introduce weights o of the type e where is a function satisfying suitable convexity conditions. This method leads to generalizations of many standard vanishing theorems to weakly pseudoconvex manifolds. As a special case, we obtain the original Hrmander estimates for pseudoconvex domains of Cn , and give some applications o to algebraic geometry (Hrmander-Bombieri-Skoda theorem, properties of zero sets of polynomials in o Cn ). We also derive the Ohsawa-Takegoshi extension theorem for L2 holomorphic functions and Skodas L2 estimates for surjective bundle morphisms [Skoda 1972a, 1978], [Demailly 1982c]. Skodas estimates can be used to obtain a quick solution of the Levi problem, and have important applications to local algebra and Nullstellensatz theorems. Finally, L2 estimates are used to prove the Newlander-Nirenberg theorem on the analyticity of almost complex structures. We apply it to establish Kuranishis theorem on deformation theory of compact complex manifolds.
1. Non Bounded Operators on Hilbert Spaces
A few preliminaries of functional analysis will be needed here. Let 1 , 2 be complex Hilbert spaces. We consider a linear operator T dened on a subspace Dom T 1 (called the domain of T ) into 2 . The operator T is said to be densely dened if Dom T is dense in 1 , and closed if its graph
is closed in
is bounded in 1 -norm. Since Dom T is dense, there exists for every y in Dom T a unique element T y 1 such that T x, y 2 = x, T y 1 for all x Dom T . It is immediate to verify that Gr T = Gr(T ) in 1 2 . It follows that T is closed and that every pair (u, v) 1 2 can be written (u, v) = (x, T x) + (T y, y), Take in particular u = 0. Then x + T y = 0, v = y T x = y + T T y, v, y
2
Assume now that T is closed and densely dened. The adjoint T of T (in Von Neumanns sense) is constructed as follows: Dom T is the set of y 2 such that the linear form Dom T x T x, y 2
1 2.
Gr T = (x, T x) ; x Dom T
x Dom T, y Dom T .
2 2
= y
+ T y 2. 1
364
Chapter VIII. L2 Estimates on Pseudoconvex Manifolds
If v (Dom T ) we get v, y 2 = 0, thus y = 0 and v = 0. Therefore T is densely dened and our discussion implies: (1.1) Theorem [Von Neumann 1929]). If T : 1 2 is a closed and densely dened operator, then its adjoint T is also closed and densely dened and (T ) = T . Furthermore, we have the relation Ker T = (Im T ) and its dual (Ker T ) = Im T . Consider now two closed and densely dened operators T , S :
2 3
T S
such that S T = 0. By this, we mean that the range T (Dom T ) is contained in Ker S Dom S, in such a way that there is no problem for dening the composition S T . The starting point of all L2 estimates is the following abstract existence theorem. (1.2) Theorem. There are orthogonal decompositions
Ker S = (Ker S Ker T ) Im T . In order that Im T = Ker S, it suces that (1.3) T x

2 1
2 = (Ker S Ker T ) Im T Im S ,
+ Sx
2 3
C x 2, 2
x Dom S Dom T
for some constant C > 0. In that case, for every v u 1 such that T u = v and 1 u 2 v 2. 1 2 C In particular Im T = Im T = Ker S,
2 such that Sv = 0, there exists
Im S = Im S = Ker T .
Proof. Since S is closed, the kernel Ker S is closed in 2 . The relation (Ker S) = Im S implies (1.4)
and similarly 2 = Ker T Im T . However, the assumption S T = 0 shows that Im T Ker S, therefore (1.5) Ker S = (Ker S Ker T ) Im T .
2 = Ker S Im S
The rst two equalities in Th. 1.2 are then equivalent to the conjunction of (1.4) and (1.5). Now, under assumption (1.3), we are going to show that the equation T u = v is always solvable if Sv = 0. Let x Dom T . One can write x = x + x where x Ker S and x (Ker S) (Im T ) = Ker T .
2. Complete Riemannian Manifolds
365
Since x, x Dom T , we have also x Dom T . We get v, x

2
= v, x
+ v, x
= v, x
because v Ker S and x (Ker S) . As Sx = 0 and T x = 0, the Cauchy-Schwarz inequality combined with (1.3) implies | v, x 2 |2 v
2 2
2 2
1 v C
2 2
T x
2 1
1 v C
2 2
T x 2. 1
This shows that the linear form TX x x, v 2 is continuous on Im T 1 with norm C 1/2 v 2 . By the Hahn-Banach theorem, this form can be extended to a continuous linear form on 1 of norm C 1/2 v 2 , i.e. we can nd u 1 such that u 1 C 1/2 v 2 and
x, v
= T x, u 1 ,
x Dom T .
This means that u Dom (T ) = Dom T and v = T u. We have thus shown that Im T = Ker S, in particular Im T is closed. The dual equality Im S = Ker T follows by considering the dual pair (S , T ).
2. Complete Riemannian Manifolds

Let (M, g) be a riemannian manifold of dimension m, with metric g(x) = gjk (x) dxj dxk , 1 j, k m.
The length of a path : [a, b] M is by denition

b b
() =
a
| (t)|g dt =
gjk (t) j (t)k (t) j,k
1/2
dt.
The geodesic distance of two points x, y M is (x, y) = inf ()
over paths with (a) = x, (b) = y,
if x, y are in the same connected component of M , (x, y) = + otherwise. It is easy to check that satises the usual axioms of distances: for the separation axiom, use the fact that if y is outside some closed coordinate ball B of radius r centered at x and if g c|dx|2 on B, then (x, y) c1/2 r. In addition, satises the axiom: (2.1) for every x, y M , inf max{(x, z), (y, z)} = 1 (x, y). 2
zM
In fact for every > 0 there is a path such that (a) = x, (b) = y, () < (x, y) + and we can take z to be at mid-distance between x and y along . A metric space E with a distance satisfying the additional axiom (2.1) will be called a geodesic metric space. It is then easy to see by dichotomy that any two points x, y E can be joined by a chain of points x = x0 , x1 , . . . , xN = y such that (xj , xj+1 ) < and (xj , xj+1 ) < (x, y) + .
366
(2.2) Lemma (Hopf-Rinow). Let (E, ) be a geodesic metric space. Then the following properties are equivalent: a) E is locally compact and complete ; b) all closed geodesic balls B(x0 , r) are compact. Proof. Since any Cauchy sequence is bounded, it is immediate that b) implies a). We now check that a) = b). Fix x0 and dene R to be the supremum of all r > 0 such that B(x0 , r) is compact. Since E is locally compact, we have R > 0. Suppose that R < +. Then B(x0 , r) is compact for every r < R. Let y be a sequence of points in B(x0 , R). Fix an integer p. As (x0 , y ) R, axiom (2.1) shows that we can nd points z M such that (x0 , z ) (1 2p )R and (z , y ) 21p R. Since B(x0 , (1 2p )R) is compact, there is a subsequence (z(p,q) )qN converging to a limit point wp with (z(p,q) , wp ) 2q . We proceed by induction on p and take (p + 1, q) to be a subsequence of (p, q). Then (y(p,q) , wp ) (y(p,q), z(p,q)) + (z(p,q), wp ) 21p R + 2q . Since (y(p+1,q)) is a subsequence of (y(p,q)), we infer that (wp , wp+1 ) 3 2p R by letting q tend to +. By the completeness hypothesis, the Cauchy sequence (wp ) converges to a limit point w M , and the above inequalities show that (y(p,p) ) converges to w B(x0 , R). Therefore B(x0 , R) is compact. Now, each point y B(x0 , R) can be covered by a compact ball B(y, y ), and the compact set B(x0 , R) admits a nite covering by concentric balls B(yj , yj /2). Set = min yj . Every point z B(x0 , R + /2) is at distance /2 of some point y B(x0 , R), hence at distance /2 + yj /2 of some point yj , in particular B(x0 , R + /2) B(yj , yj ) is compact. This is a contradiction, so R = +. The following standard denitions and properties will be useful in order to deal with the completeness of the metric. (2.3) Denitions. a) A riemannian manifold (M, g) is said to be complete if (M, ) is complete as a metric space. b) A continuous function : M R is said to be exhaustive if for every c R the sublevel set Mc = {x M ; (x) < c} is relatively compact in M . c) A sequence (K )N of compact subsets of M is said to be exhaustive if M = K and if K is contained in the interior of K+1 for all (so that every compact subset of M is contained in some K ). (2.4) Lemma. The following properties are equivalent: a) (M, g) is complete; b) there exists an exhaustive function
(M, R) such that |d|g
1;
c) there exists an exhaustive sequence (K )N of compact subsets of M and functions (M, R) such that = 1 0 in a neighborhood of K , 1 and |d |g 2 .
Supp K+1 ,
3. L2 Hodge Theory on Complete Riemannian Manifolds
367
Proof. a) = b). Without loss of generality, we may assume that M is connected. Select 1 a point x0 M and set 0 (x) = 2 (x0 , x). Then 0 is a Lipschitz function with constant 1 1 2 , thus 0 is dierentiable almost everywhere on M and |d0 |g 2 . We can nd a smoothing of 0 such that |d|g 1 and | 0 | 1. Then is an exhaustion function of M . b) = c). Choose as in a) and a function (R, R) such that = 1 on ] , 1.1], = 0 on [1.9, +[ and 0 2 on [1, 2]. Then K = {x M ; (x) satisfy our requirements. c) = b). Set = 2 (1 ). (x, y) for all x, y M , so 2+1 }, (x) = 21 (x)
b) = a). The inequality |d|g 1 implies |(x) (y)| all -balls must be relatively compact in M .

Let (M, g) be a riemannian manifold and let F1 , F2 be hermitian vector bundles over M . If P : (M, F1 ) (M, F2 ) is a dierential operator with smooth coecients, then P induces a non bounded operator P : L2 (M, F1 ) L2 (M, F2 ) as follows: if u L2 (M, F1 ), we compute P u in the sense of distribution theory and we say that u Dom P if P u L2 (M, F2 ). It follows that P is densely dened, since Dom P contains the set (M, F1 ) of compactly supported sections of (M, F1 ), which is dense in L2 (M, F1 ). Furthermore Gr P is closed: if u u in L2 (M, F1 ) and P u v in L2 (M, F2 ) then P u P u in the weak topology of distributions, thus we must have P u = v and (u, v) Gr P . By the general results of 1, we see that P has a closed and densely dened Von Neumann adjoint P . We want to stress, however, that P does not always coincide with the extension (P ) of the formal adjoint P : (M, F2 ) (M, F1 ), computed in the sense of distribution theory. In fact u Dom (P ) , resp. u Dom (P ) , if and only if there is an element v L2 (M, F1 ) such that u, P f = v, f for all f Dom P , resp. for all f (M, F1 ). Therefore we always have Dom (P ) Dom (P ) and the inclusion may be strict because the integration by parts to perform may involve boundary integrals for (P ) . (3.1) Example. Consider P = d : L2 ]0, 1[ L2 ]0, 1[ dx
where the L2 space is taken with respect to the Lebesgue measure dx. Then Dom P consists of all L2 functions with L2 derivatives on ]0, 1[. Such functions have a continuous extension to the interval [0, 1]. An integration by parts shows that
1
u
0
df dx = dx
1 0
du f dx dx
368
for all f (]0, 1[), thus P = d/dx = P . However for f Dom P the integration by parts involves the extra term u(1)f (1) u(0)f (0) in the right hand side, which is thus continuous in f with respect to the L2 topology if and only if du/dx L2 and u(0) = u(1) = 0. Therefore Dom (P ) consists of all u Dom (P ) = Dom P satisfying the additional boundary condition u(0) = u(1) = 0. Let E M be a dierentiable hermitian bundle. In what follows, we still denote by D, , the dierential operators of VI-2 extended in the sense of distribution theory (as explained above). These operators are thus closed and densely dened operators 2 on L2 (M, E) = p (M, E) of compactly p Lp (M, E). We also introduce the space supported forms in p (M, E). The theory relies heavily on the following important result. (3.2) Theorem. Assume that (M, g) is complete. Then a) norms
(M, E)
is dense in Dom D, Dom and Dom D Dom respectively for the graph u u + u , u u + Du + u .
u u + Du ,
b) D = , = D as adjoint operators in Von Neumanns sense. c) One has u, u = Du

2
+ u
for every u Dom . In particular Ker = Ker D Ker ,
Dom Dom D Dom , and is self-adjoint.
d) If D2 = 0, there are orthogonal decompositions L2 (M, E) = (M, E) Im D Im , where (M, E) = harmonic forms. Ker D = (M, E) Im D, u L2 (M, E) ; u = 0
(M, E)
is the space of L2
Proof. a) We show that every element u Dom D can be approximated in the graph norm of D by smooth and compactly supported forms. By hypothesis, u and Du belong to L2 (M, E). Let ( ) be a sequence of functions as in Lemma 2.4 c). Then u u in L2 (M, E) and D( u) = Du + d u where |d u| |d | |u| 2 |u|.
Therefore d u 0 and D( u) Du. After replacing u by u, we may assume that u has compact support, and by using a nite partition of unity on a neighborhood of Supp u we may also assume that Supp u is contained in a coordinate chart of M on which E is trivial. Let A be the connection form of D on this chart and ( ) a family of smoothing kernels. Then u (M, E) converges to u in L2 (M, E) and D(u ) (Du) = A (u ) (A u)
369
because d commutes with convolution (as any dierential operator with constant coecients). Moreover (Du) converges to Du in L2 (M, E) and A(u ), (Au) both converge to A u since A acts continuously on L2 . Thus D(u ) converges to Du and the density of (M, E) in Dom D follows. The proof for Dom and Dom D Dom is similar, except that the principal part of no longer has constant coecients in general. The convolution technique requires in this case the following lemma due to K.O. Friedrichs (see e.g. [Hrmander 1963]). o (3.3) Lemma. Let P f = ak f /xk + bf be a dierential operator of order 1 on an open set Rn , with coecients ak C 1 (), b C 0 (). Then for any v L2 (Rn ) with compact support in we have
0
lim ||P (v ) (P v) ||L2 = 0.
Proof. It is enough to consider the case when P = a/xk . As the result is obvious if v C 1 , we only have to show that ||P (v ) (P v) ||L2 C||v||L2
and to use a density argument. A computation of w = P (v ) (P v) by means of an integration by parts gives w (x) =
Rn
a(x)
=
Rn
v v (x y)(y) a(x y) (x y)(y) dy xk xk 1 a(x) a(x y) v(x y) k (y) + k a(x y)v(x y)(y) dy.
If C is a bound for |da| in a neighborhood of Supp v, we get |w (x)| C

Rn
|v(x y)| |y| |k (y)| + |(y)| dy, ||f ||L1 ||g||Lp gives |y| |k (y)| + |(y)| dy ||v||L2 .
so Minkowskis inequality ||f g||Lp ||w ||L2 C

Rn
Proof (end).. b) is equivalent to the fact that Du, v = u, v , By a), we can nd u , v u u, v v,

(M, E)
u Dom D, v Dom . such that and v v in L2 (M, E),
Du Du
and the required equality is the limit of the equalities Du , v = u , v .
370
2 c) Let u Dom . As is an elliptic operator of order 2, u must be in W (M, E, loc) by G ardings inequality. In particular Du, u L2 (M, E, loc) and we can perform all integrations by parts that we want if the forms are multiplied by compactly supported functions . Let us compute
Du 2 + u 2 = 2 2 = Du, Du + u, D( u)
2 2 = D( u), Du + u, Du 2 d u, Du + 2 u, d u 2 = u, u 2 d u, Du + 2 u, d ( u) 2 u, u + 2 2 Du
u + 2 u
2
2 u, u
+2
Du
+ u
+2 u
We get therefore Du
2
+ u
1 1 2
2
2 u, u + 21 u 2
By letting tend to +, we obtain Du in L2 (M, E). This implies
+ u
u, u , in particular Du, u are u, v Dom ,
u, v = Du, Dv + u, v ,
because the equality holds for u and v, and because we have u u, D( u) Du and ( u) u in L2 . Therefore is self-adjoint. d) is an immediate consequence of b), c) and Th. 1.2. On a complete hermitian manifold (X, ), there are of course similar results for the operators D , D , , , , attached to a hermitian vector bundle E.
4. General Estimate for d on Hermitian Manifolds

Let (X, ) be a complete hermitian manifold and E a hermitian holomorphic vector bundle of rank r over X. Assume that the hermitian operator (4.1) AE, = [i(E), ] + T
is semi-positive on p,q TX E. Then for every form u Dom D Dom of bidegree (p, q) we have
(4.2)
D u
+ u
2 X
AE, u, u dV.
In fact (4.2) is true for all u p,q (X, E) in view of the Bochner-Kodaira-Nakano identity VII-2.3, and this result is easily extended to every u in Dom D Dom by density of p,q (X, E) (Th. 3.2 a)). Assume now that a form g L2 (X, E) is given such that p,q (4.3) D g = 0,
4. General Estimate for d on Hermitian Manifolds
371
and that for almost every x X there exists [0, +[ such that | g(x), u |2 AE, u, u
for every u (p,q TX E)x . If the operator AE, is invertible, the minimal such number 1/2 is |AE, g(x)|2 = A1 g(x), g(x) , so we shall always denote it in this way even when E, AE, is no longer invertible. Assume furthermore that
(4.4)
X
A1 g, g dV < +. E,
The basic result of L2 theory can be stated as follows. (4.5) Theorem. If (X, ) is complete and AE, 0 in bidegree (p, q), then for any g L2 (X, E) satisfying (4.4) such that D g = 0 there exists f L2 p,q p,q1 (X, E) such that D f = g and f
2 X
A1 g, g dV. E,
Proof. For every u Dom D Dom we have u, g

2
=
X
u, g dV
2 X
AE, u, u
1/2
A1 g, g E,
1/2
dV
A1 g, g dV E,
AE, u, u dV
X
by means of the Cauchy-Schwarz inequality. The a priori estimate (4.2) implies u, g

2
D u
+ u
u Dom D Dom
where C is the integral (4.4). Now we just have to repeat the proof of the existence part of Th. 1.2. For any u Dom , let us write u = u1 + u2 , u1 Ker D , u2 (Ker D ) = Im .
Then D u1 = 0 and u2 = 0. Since g Ker D , we get u, g

2
u1 , g
C u1
= C u 2 .
The Hahn-Banach theorem shows that the continuous linear form

L2 p,q1 (X, E) u u, g
can be extended to a linear form v v, f , f L2 p,q1 (X, E), of norm f This means that u, g = u, f , u Dom , i.e. that D f = g. The theorem is proved.
C 1/2 .
372
(4.6) Remark. One can always nd a solution f (Ker D ) : otherwise replace f by its orthogonal projection on (Ker D ) . This solution is clearly unique and is precisely the solution of minimal L2 norm of the equation D f = g. We have f Im , thus f saties the additional equation (4.7) f = 0.
p,q (X, E),
Consequently f = D f = g. If g f (X, E). p,q1
the ellipticity of shows that
(4.8) Remark. If AE, is positive denite, let (x) > 0 be the smallest eigenvalue of this operator at x X. Then is continuous on X and we have
X
A1 g, g dV E,
(x)1 |g(x)|2 dV.
The above situation occurs for example if is complete Khler, E >m 0 and p = n, a q 1, m min{n q + 1, r} (apply Lemma VII-7.2).
5. Estimates on Weakly Pseudoconvex Manifolds

We rst introduce a large class of complex manifolds on which the L2 estimates will be easily tractable. (5.1) Denition. A complex manifold X is said to be weakly pseudoconvex if there exists an exhaustion function (X, R) such that id d 0 on X, i.e. is plurisubharmonic. For domains Cn , the above weak pseudoconvexity notion is equivalent to pseudoconvexity (cf. Th. I-4.14). Note that every compact manifold is also weakly pseudoconvex (take 0). Other examples that will appear later are Stein manifolds, or the total space of a Griths semi-negative vector bundle over a compact manifold (cf. Prop. IX?.?). (5.2) Theorem. Every weakly pseudoconvex Khler manifold (X, ) carries a complete a Khler metric . a Proof. Let (X, R) be an exhaustive plurisubharmonic function on X. After addition of a constant to , we can assume 0. Then = + id d ( 2 ) is Khler a and = + 2id d + 2id d + 2id d . Since d = d + d , we get |d| = 2|d | 1 and Lemma 2.4 shows that is complete. Observe that we could have set more generally = + id d ( ) where is a convex increasing function. Then = + i( )d d + i( )d d + id ( ) d ( )
(5.3)
5. Estimates on Weakly Pseudoconvex Manifolds
373
where (t) = 0 (u) du. We thus have |d ( )| soon as limt+ (t) = +, i.e.
+
1 and will be complete as
(5.4)
0
(u) du = +. 1.
One can take for example (t) = t log(t) for t
It follows from the above considerations that almost all vanishing theorems for positive vector bundles over compact manifolds are also valid on weakly pseudoconvex manifolds. Let us mention here the analogues of some results proved in Chapter 7.
(5.5) Theorem. For any m-positive vector bundle of rank r over a weakly pseudoconvex manifold X, we have H n,q (X, E) = 0 for all q 1 and m min{n q + 1, r}. Proof. The curvature form i(det E) is a Khler metric on X, hence X possesses a coma plete Khler metric . Let a (X, R) be an exhaustive plurisubharmonic function. For any convex increasing function (R, R), we denote by E the holomorphic vector bundle E together with the modied metric |u|2 = |u|2 exp (x) , u Ex . We get i(E ) = i(E) + id d ( ) IdE m i(E), thus AE , AE, > 0 in bidegree (n, q). Let g be a given form of bidegree (n, q) with L2 coecients, such that D g = 0. The integrals loc A1 , g, g E
dV X
A1 g, g e dV, E,
|g|2 e dV
become convergent if grows fast enough. We can thus apply Th. 4.5 to (X, E , ) and nd a (n, q 1) form f such that D f = g. If g is smooth, Remark 4.6 shows that f can also be chosen smooth. (5.6) Theorem. If E is a positive line bundle over a weakly pseudoconvex manifold X, then H p,q (X, E) = 0 for p + q n + 1. Proof. The proof is similar to that of Th. 5.5, except that we use here the Khler metric a = i(E ) = + id d ( ), = i(E),
which depends on . By (5.4) is complete as soon as is a convex increasing function that grows fast enough. Apply now Th. 4.5 to (X, E , ) and observe that AE , = [i(E ), ] = (p + q n) Id in bidegree (p, q) in virtue of Cor. VI-8.4 It remains to show that for every form g (X, E) there exists a choice of such that g p,q L2 (X, E , ). By (5.3) the norm of a scalar form with respect to is less than its p,q norm with respect to , hence |g|2 |g|2 exp( ). On the other hand dV C 1 + +
n
dV
where C is a positive continuous function on X. The following lemma implies that we can always choose in order that the integral of |g|2 dV converges on X.
374
(5.7) Lemma. For any positive function [0, +[, R , there exists a smooth convex function [0, +[, R such that , , and (1 + + )n e 1/. Proof. We shall construct such that and /2 C for some constant C. Then satisties the conclusion of the lemma after addition of a constant. Without loss of generality, we may assume that is increasing and 1. We dene as a power series
+
(t) =
k=0
a 0 a 1 . . . a k tk ,
where ak > 0 is a decreasing sequence converging to 0 very slowly. Then is real analytic on R and the inequalities are realized if we choose ak 1/k, k 1. 1 Select a strictly increasing sequence of integers (Np )p 1 so large that p (p + 1)1/Np [1/p, 1/(p 1)[. We set a0 = . . . = aN1 1 = e (2), 1 ak = (p + 1)1/Np e1/ k , p
Np
k < Np+1 . (t) and for t [1, +[ the (p + 1) (t).
Then (ak ) is decreasing. For t [0, 1] we have (t) a0 choice k = Np where p = [t] is the integer part of t gives (t) Furthermore, we have (t)2
k 0
(p)
(a0 a1 . . . ak )pk
(ak p)k
(a0 a1 . . . ak )2 t2k , (k + 1)(k + 2) a0 a1 . . . ak+2 tk ,

k 0
(t) = thus we will get (t)
C(t)2 if we can prove that C (a0 a1 . . . am )2 , m 0.
m2 a0 a1 . . . a2m
1 However, as p (p + 1)1/Np is decreasing, we nd
am+1 . . . a2m a0 a1 . . . a2m = 2 (a0 a1 . . . am ) a0 a 1 . . . am 1 1 1 1 + O(1) exp ++ m m+1 2m 1 2 exp 2 2m 4 m + O(1) Cm . As a last application, we generalize the Girbau vanishing theorem in the case of weakly pseudoconvex manifolds. This result is due to [Abdelkader 1980] and [Ohsawa 1981]. We present here a simplied proof which appeared in [Demailly 1985].
6. Hrmanders Estimates for non Complete Khler Metrics o a
375
(5.8) Theorem. Let (X, ) be a weakly pseudoconvex Khler manifold. If E is a semia positive line bundle such that i(E) has at least n s + 1 positive eigenvalues at every point, then H p,q (X, E) = 0 for p + q n + s. Proof. Let , (R, R) be convex increasing functions to be specied later. We use here the hermitian metric = i(E ) + exp( ) = i(E) + id d ( ) + exp( ) .
, , Although is Khler, the metric is not so. Denote by j (resp. j ), 1 j a n, the eigenvalues of i(E ) with respect to (resp. ), rearranged in increasing order. , 0, 0, 0, The minimax principle implies j j , and the hypothesis yields 0 < s s+1 0, . . . n on X. By means of a diagonalization of i(E ) with respect to , we nd , j = , j , j + exp( ) 0, j 0, j + exp( )
Let > 0 be small. Select such that exp( (x)) for j s we get 0, j 1 , j 0, = 1 + 0, j + j and Th. VI-8.3 implies i(E ), u, u
0, s (x) at every point. Then
1 ,
, , , , 1 + + p q+1 . . . n |u|2
1 (p s + 1) |u|2 .
(p s + 1)(1 ) (n q) |u|2
It remains however to control the torsion term T . As is Khler, trivial computations a yield d d = exp( ) Since i( d d + d d ) + exp( ), |d | | exp( )| ( )2 + + . ( ) 2
1
d = exp( ) d ,
( )2 d d d d .
we get the upper bounds |d | |d d |
It is then clear that we can choose growing suciently fast in order that |T | . If 1 is chosen suciently small, we get AE , 2 Id, and the conclusion is obtained in the same way as for Th. 5.6.
376

Our aim here is to derive also estimates for a non complete Khler metric, for example a the standard metric of Cn on a bounded domain Cn . A result of this type can be obtained in the situation described at the end of Remark 4.8. The underlying idea is due to [Hrmander 1966], although we do not apply his so called three weights technique, o but use instead an approximation of the given metric by complete Khler metrics. a (6.1) Theorem. Let (X, ) be a complete Khler manifold, another Khler metric, a a possibly non complete, and E X a m-semi-positive vector bundle. Let g L2 (X, E) n,q be such that D g = 0 and
X
A1 g, g dV < + q 1,
with respect to , where Aq stands for the operator i(E) in bidegree (n, q) and q m min{n q + 1, r}. Then there exists f L2 n,q1 (X, E) such that D f = g and f
2 X
A1 g, g dV. q
Proof. For every > 0, the Khler metric a = + is complete. The idea of the proof is to apply the L2 estimates to and to let tend to zero. Let us put an index to all objects depending on . It follows from Lemma 6.3 below that (6.2) |u|2 dV |u|2 dV, A1 u, u dV q, A1 u, u dV q
for every u n,q TX E. If these estimates are taken for granted, Th. 4.5 applied to yields a section f L2 n,q1 (X, E) such that D f = g and
|f |2 dV
A1 g, g dV q,
A1 g, g dV. q
This implies that the family (f ) is bounded in L2 norm on every compact subset of X. We can thus nd a weakly convergent subsequence (f ) in L2 . The weak limit f is the loc solution we are looking for. (6.3) Lemma. Let , be hermitian metrics on X such that u n,q TX E, q 1, we have |u|2 dV |u|2 dV, A1 u, u q,
. For every
dV
A1 u, u dV q
where an index means that the corresponding term is computed in terms of instead of . Proof. Let x0 X be a given point and (z1 , . . . , zn ) coordinates such that =i
1 j n
dzj dz j ,
=i
1 j n
j dzj dz j
at x0 ,
377
where 1 . . . n are the eigenvalues of with respect to (thus j 1). We have 1 1 |dzj |2 = j and |dzK |2 = K for any multi-index K, with the notation K = jK j . For every u = uK, dz1 . . . dzn dz K e , |K| = q, 1 r, the computations of VII-7 yield |u|2 = |u|2 dV =
K, K, 1 (1 . . . n )1 K |uK, |2 , 1 K |uK, |2 dV
dV = 1 . . . n dV,
|u|2 dV,
u =
|I|=q1 j,
1 i(1)n+j1 j ujI, (dzj ) dz I e ,
where (dzj ) means dz1 . . . dzj . . . dzn , Aq, u =

|I|=q1 j,k,, 1 j cjk ujI, dz1 . . . dzn dz kI e , 1 I |I|=q1 j,k,, 2 I |I|=q1 j,k,, 1 1 j k cjk ujI, ukI, 1 1 j k cjk ujI, ukI,
Aq, u, u
= (1 . . . n )1 (1 . . . n )1
= 1 . . . n Aq S u, S u where S is the operator dened by S u =

K
(1 . . . n K )1 uK, dz1 . . . dzn dz K e .
We get therefore | u, v |2 = | u, S v |2 and the choice v = A1 u implies q, A1 u, u q,
A1 u, u Aq S v, S v q (1 . . . n )1 A1 u, u Aq, v, v , q
(1 . . . n )1 A1 u, u ; q
this relation is equivalent to the last one in the lemma. An important special case is that of a semi-positive line bundle E. If we let 0 1 (x) ... n (x) be the eigenvalues of i(E)x with respect to x for all x X, formula VI-8.3 implies Aq u, u (6.4)
X
A1 g, g dV q
(1 + + q )|u|2 , 1 |g|2 dV. 1 + + q X
378
A typical situation where these estimates can be applied is the case when E is the trivial line bundle X C with metric given by a weight e . One can assume for example that is plurisubharmonic and that id d has at least n q + 1 positive eigenvalues at every point, i.e. q > 0 on X. This situation leads to very important L2 estimates, which are precisely those given by [Hrmander 1965, 1966]. We state here a slightly more general o result. (6.5) Theorem. Let (X, ) be a weakly pseudoconvex Khler manifold, E a hermitian a (X, R) a weight function such that the eigenvalues 1 . . . line bundle on X, n of i(E) + id d are 0. Then for every form g of type (n, q), q 1, with L2 loc (resp. ) coecients such that D g = 0 and 1 |g|2 e dV < +, 1 + + q ) form f of type (n, q 1) such that D f = g and
X
we can nd a L2 (resp. loc
|f |2 e dV
1 |g|2 e dV. 1 + + q
Proof. Apply the general estimates to the bundle E deduced from E by multiplication of the metric by e ; we have i(E ) = i(E) + id d . It is not necessary here to assume in addition that g L2 (X, E ). In fact, g is in L2 and we can exhaust X by n,q loc the relatively compact weakly pseudoconvex domains Xc = x X ; (x) < c where (X, R) is a plurisubharmonic exhaustion function (note that log(c ) is also such a function on Xc ). We get therefore solutions fc on Xc with uniform L2 bounds; any weak limit f gives the desired solution. If estimates for (p, q)-forms instead of (n, q)-forms are needed, one can invoke the isomorphism p TX np TX n TX (obtained through contraction of n-forms by (n p)-vectors) to get
p,q TX E n,q TX F,
F = E np TX .
Let us look more carefully to the case p = 0. The (1, 1)-curvature form of n TX with respect to a hermitian metric on TX is called the Ricci curvature of . We denote: (6.6) Denition. Ricci() = i(n TX ) = i Tr (TX ). For any local coordinate system (z1 , . . . , zn ), the holomorphic n-form dz1 . . . dzn is a section of n TX , hence Formula V-13.3 implies (6.7) Ricci() = id d log |dz1 . . . dzn |2 = id d log det(jk ). ... n
The estimates of Th. 6.5 can therefore be applied to any (0, q)-form g, but 1 must be replaced by the eigenvalues of the (1, 1)-form (6.8) i(E) + Ricci() + id d (supposed 0).
7. Extension of Holomorphic Functions from Subvarieties
379
We consider now domains Cn equipped with the euclidean metric of Cn , and the trivial bundle E = C. The following result is especially convenient because it requires only weak plurisubharmonicity and avoids to compute the curvature eigenvalues. (6.9) Theorem. Let Cn be a weakly pseudoconvex open subset and an upper semi-continuous plurisubharmonic function on . For every ]0, 1] and every g L2 (, loc) such that d g = 0 and p,q
1 + |z|2 |g|2 e dV < +,
we can nd a
form f of type (p, q 1) such that d f = g and 4 1 + |z|2 |g|2 e dV < +. 1 + |z|2 |f |2 e dV q2 Moreover f can be chosen smooth if g and are smooth.
L2 loc
Proof. Since p T is a trivial bundle with trivial metric, the proof is immediately reduced to the case p = 0 (or equivalently p = n). Let us rst suppose that is smooth. We replace by = + where (z) = log 1 + (1 + |z|2 ) . (6.10) Lemma. The smallest eigenvalue 1 (z) of id d (z) satises 1 (z) 2 . 2(1 + |z|2 ) 1 + (1 + |z|2 )
In fact a brute force computation of the complex hessian Hz () and the CauchySchwarz inequality yield Hz () = (1+|z| ) || ( 1)(1+|z|2 )2 | , z |2 2 (1+|z|2 )22 | , z |2 = + 2 1 + (1+|z|2 ) 1 + (1+|z|2 ) 1 + (1+|z|2 )
2 1 2
(1 )(1 + |z|2 )2 |z|2 (1 + |z|2 )22 |z|2 (1 + |z|2 )1 ||2 2 2 ) 2 ) 1 + (1 + |z| 1 + (1 + |z| 1 + (1 + |z|2 ) (1 + |z|2 )2 1 + (1 + |z|2 ) 1 + |z|2 + (1 + |z|2 )
2
||2
2 ||2 . 2 ) 1 + (1 + |z|2 ) 2(1 + |z| The Lemma implies e /1
(1 + |z|2 )1 1 + (1 + |z|2 )
2 ||2
2(1 + |z|2 )/2 , thus Cor. 6.5 provides an f such that 2 1 1 + (1 + |z|2 ) |f |2 e dV 1 + |z|2 |g|2 e dV < +, 2 q and the required estimate follows. If is not smooth, apply the result to a sequence of regularized weights on an increasing sequence of domains c , and extract a weakly convergent subsequence of solutions.
380

The existence theorems for solutions of the d operator easily lead to an extension theorem for sections of a holomorphic line bundle dened in a neighborhood of an analytic subset. The following result [Demailly 1982] is an improvement and a generalization of Jennanes extension theorem [Jennane 1976]. (7.1) Theorem. Let (X, ) be a weakly pseudoconvex Khler manifold, L a hermitian a line bundle and E a hermitian vector bundle over X. Let Y be an analytic subset of X such that Y = 1 (0) for some section of E, and p the maximal codimension of the irreducible components of Y . Let f be a holomorphic section of KX L dened in the open set U Y of points x X such that |(x)| < 1. If U |f |2 dV < + and if the curvature form of L satises i(L) p + {i(E), } 2 || 1 + ||2
for some > 0, there is a section F H 0 (X, KX L) such that FY = fY and |F |2 dV (1 + ||2 )p+ 1+ (p + 1)2 |f |2 dV.
The proof will involve a weight with logarithmic singularities along Y . We must therefore apply the existence theorem over X Y . This requires to know whether X Y has a complete Khler metric. a (7.2) Lemma. Let (X, ) be a Khler manifold, and Y = 1 (0) an analytic subset a dened by a section of a hermitian vector bundle E X. If X is weakly pseudoconvex and exhausted by Xc = {x X ; (x) < c}, then Xc Y has a complete Khler metric a for all c R. The same conclusion holds for X Y if (X, ) is complete and if for some constant C 0 we have E Grif C , E on X. Proof. Set = log ||2 . Then d = {D , }/||2 and D D = D2 = (E), thus id d = i {D , D } {D , } {, D } {i(E), } i . ||2 ||4 ||2
For every TX , we nd therefore H () = ||2 |D |2 | D , |2 E ( , ) ||4 ||2 E ( , ) ||2
by the Cauchy-Schwarz inequality. If C is a bound for the coecients of E on the compact subset X c , we get id d C on Xc . Let (R, R) be a convex increasing function. We set = + id d ( ).
381
Formula 5.3 shows that is positive denite if Y = 1 () as soon as

0
1/2C and that is complete near
(t) dt = +.
1 One can choose for example such that (t) = 5C (t log |t|) for t 1. In order to obtain a complete Khler metric on Xc Y , we need also that the metric be complete a near Xc . Such a metric is given by
= + id d log(c )
id log(c )1 d log(c )1 ; is complete on Xc
id d id d =+ + c (c )2
because log(c )1 tends to + on Xc .
Proof of Theorem 7.1.. When we replace by (1 + ) for some small > 0 and let tend to 0, we see that we can assume f dened in a neighborhood of U . Let h be the continuous section of L such that h = (1 ||p+1 )f on U = {|| < 1} and h = 0 on X U . We have hY = fY and d h = p + 1 p1 || {, D } f 2 on U, d h = 0 on X U.
We consider g = d h as a (n, 1)-form with values in the hermitian line bundle L = L, endowed with the weight e given by = p log ||2 + log(1 + ||2 ). Notice that is singular along Y . The Cauchy-Schwarz inequality implies i{D , } {, D } i{D , D } as in Lemma 7.2, and we nd id d log(1 + ||2 ) = (1 + ||2 )i{D , D } i{D , } {, D } (1 + ||2 )2 {i(E), } i{D , D } {i(E), } . 1 + ||2 (1 + ||2 )2 1 + ||2
The inequality id d log ||2 one imply
{i(E), }/||2 obtained in Lemma 7.2 and the above
i(L ) = i(L) + p id d log ||2 + id d log(1 + ||2 ) i(L)
p i{D , D } + {i(E), } + ||2 1 + ||2 (1 + ||2 )2 i {D , } {, D } , ||2 (1 + ||2 )2
thanks to the hypothesis on the curvature of L and the Cauchy-Schwarz inequality. Set = (p + 1)/2 ||p1 {D , } = j dzj in an -orthonormal basis /zj , and let
382
= j /z j be the dual (0, 1)-vector eld. For every (n, 1)-form v with values in L , we nd |f | | v|, d h, v = f, v = f, v v=
2
ij dzj v = i v, |f |2 | v|2 = |f |2 i v,
2
| d h, v |2
= |f | i v, v = |f | [i , ]v, v
(p + 1)2 2p || (1 + ||2 )2 |f |2 [i(L ), ]v, v . 4
Thus, in the notations of Th. 6.1, the form g = d h satises A1 g, g 1 (p + 1)2 2p || (1 + ||2 )2 |f |2 4 (p + 1)2 2 |f | e ,
where the last equality results from the fact that (1 + ||2 )2 4 on the support of g. Lemma 7.2 shows that the existence theorem 6.1 can be applied on each set Xc Y . Letting c tend to innity, we infer the existence of a (n, 0)-form u with values in L such that d u = g on X Y and |u|2 e dV A1 g, g e, 1 |f |2 dV. thus
X Y
X Y
X Y
|u|2 dV ||2p(1 + ||2 )
(p + 1)2
This estimate implies in particular that u is locally L2 near Y . As g is continuous over X, Lemma 7.3 below shows that the equality d u = g = d h extends to X, thus F = h u is holomorphic everywhere. Hence u = h F is continuous on X. As |(x)| C d(x, Y ) in a neighborhood of every point of Y , we see that ||2p is non integrable at every point x0 Yreg , because codim Y p. It follows that u = 0 on Y , so FY = hY = fY . The nal L2 -estimate of Th. 7.1 follows from the inequality |F |2 = |h u|2 which implies (1 + ||2p ) |u|2 + (1 + ||2p ) |f |2 |u|2 + |f |2 . ||2p
|F |2 (1 + ||2 )p
(7.3) Lemma. Let be an open subset of Cn and Y an analytic subset of . Assume that v is a (p, q 1)-form with L2 coecients and w a (p, q)-form with L1 coecients loc loc such that d v = w on Y (in the sense of distribution theory). Then d v = w on . Proof. An induction on the dimension of Y shows that it is sucient to prove the result in a neighborhood of a regular point a Y . By using a local analytic isomorphism, the proof is reduced to the case where Y is contained in the hyperplane z1 = 0, with a = 0.
383
Let (R, R) be a function such that (t) = 0 for t must show that (7.4)
1 2
and (t) = 1 for t
1. We
w = (1)p+q
v d
for all Then
np,nq ().
Set (z) = (|z1 |/) and replace in the integral by . Y ) and the hypotheses imply np,nq ( v d ( ) = (1)p+q v (d + d ).
w = (1)p+q
As w and v have L1 coecients on , the integrals of w and v d converge loc respectively to the integrals of w and v d as tends to 0. The remaining term can be estimated by means of the Cauchy-Schwarz inequality: v d
2 |z1 |
|v |2 dV.
Supp
|d |2 dV ;
as v L2 (), the integral loc
|z1 |
|v |2 dV converges to 0 with , whereas C Vol Supp {|z1 | 2 } C .
Supp
|d |2 dV
Equality (7.4) follows when tends to 0. (7.5) Corollary. Let Cn be a weakly pseudoconvex domain and let , be plurisubharmonic functions on , where is supposed to be nite and continuous. Let = (1 , . . . , r ) be a family of holomorphic functions on , let Y = 1 (0), p = maximal codimension of Y and set a) U = {z ; |(z)|2 < e(z) }, b) U = {z ; |(z)|2 < e(z) }. For every > 0 and every holomorphic function f on U , there exists a holomorphic function F on such that FY = fY and a) b) |F |2 e+p dV 2 p+ (1 + || e ) |F |2 e dV 2 p+ (e + || ) (p + 1)2 (p + 1)2 1+ 1+ |f |2 e+p dV, resp. resp.
|f |2 e(p+) dV.
Proof. After taking convolutions with smooth kernels on pseudoconvex subdomains c , we may assume , smooth. In either case a) or b), apply Th. 7.1 to a) E = Cr with the weight e , L = C with the weight e+p , and U = {||2 e < 1}. Then i(E) = id d IdE 0, i(L) = id d p id d p i(E).
384
b) E = Cr with the weight e , L = C with the weight e(p+) , and U = {||2 e < 1}. Then i(E) = id d IdE 0, i(L) = id d + (p + ) id d (p + ) i(E).
The condition on (L) is satised in both cases and K is trivial. (7.6) Hrmander-Bombieri-Skoda theorem. Let Cn be a weakly pseudoconvex o domain and a plurisubharmonic function on . For every > 0 and every point z0 such that e is integrable in a neighborhood of z0 , there exists a holomorphic function F on such that F (z0 ) = 1 and |F (z)|2 e(z) dV < +. (1 + |z|2 )n+
[Bombieri 1970] originally stated the theorem with the exponent 3n instead of n + ; the improved exponent n + is due to [Skoda 1975]. The example = Cn , (z) = 0 shows that one cannot replace by 0. Proof. Apply Cor. 7.5 b) to f 1, (z) = zz0 , p = n and log r 2 where U = B(z0 , r) is a ball such that U e dV < +. (7.7) Corollary. Let be a plurisubharmonic function on a complex manifold X. Let A be the set of points z X such that e is not locally integrable in a neighborhood of z. Then A is an analytic subset of X. Proof. Let X be an open coordinate patch isomorphic to a ball of Cn , with coordinates (z1 , . . . , zn ). Dene E H 0 (, ) to be the Hilbert space of holomorphic functions f on such that
1
|f (z)|2 e(z) dV (z) < +.
Then A = f E f (0). In fact, every f in E must obviously vanish on A ; conversely, if z0 A, Th. 7.6 shows that there exists f E such that f (z0 ) = 0. By Th. II-5.5, we / conclude that A is analytic.
8. Applications to Hypersurface Singularities

We rst give some basic denitions and results concerning multiplicities of divisors on a complex manifold. (8.1) Proposition. Let X be a complex manifold and = j [Zj ] a divisor on X with real coecients j 0. Let x X and fj = 0, 1 j N , irreducible equations of Zj on a neighborhood U of x. a) The multiplicity of at x is dened by (, x) = j ordx fj .
385
b) is said to have normal crossings at a point x Supp if all hypersurfaces Zj containing x are smooth at x and intersect transversally, i.e. if the linear forms dfj dening the corresponding tangent spaces Tx Zj are linearly independent at x. The set nnc() of non normal crossing points is an analytic subset of X. c) The non-integrability locus nil() is dened as the set of points x X such that |fj |2j is non integrable near x. Then nil() is an analytic subset of X and there are inclusions {x X ; (, x) n} nil() {x X ; (, x) 1}.
Moreover nil() nnc() if all coecients of satisfy j < 1. Proof. b) The set nnc() U is the union of the analytic sets fj1 = . . . = fjp = 0, dfj1 . . . dfjp = 0,
for each subset {j1 , . . . , jp } of the index set {1, . . . , N }. Thus nnc() is analytic. c) The analyticity of nil() follows from Cor. 7.7 applied to the plurisubharmonic function = 2j log |fj |. Assume rst that j < 1 and that has normal crossings at x. Let fj1 (x) = . . . = fjs (x) = 0 and fj (x) = 0 for j = jl . Then, we can choose local coordinates (w1 , . . . , wn ) on U such that w1 = fj1 (z), . . ., ws = fjs (z), and we have
U
It follows that nil() nnc(). Let us prove now the statement relating nil() with multiplicity sets. Near any point x, we have |fj (z)| Cj |z x|mj with mj = ordx fj , thus |fj |2j C |z x|2(,x) . It follows that x nil() as soon as (, x) n. On the other hand, we are going to prove that (, x) < 1 implies x nil(), i.e. / |fj |2j integrable near x. We may assume j rational; otherwise replace each j by a slightly larger rational number in such k a way that (, x) < 1 is still true. Set f = fj j where k is a common denominator. The result is then a consequence of the following lemma. (8.2) Lemma. If f X,x is not identically 0, there exists a neighborhood U of x such that U |f |2 dV converges for all < 1/m, m = ordx f . Proof. One can assume that f is a Weierstrass polynomial
m m1 f (z) = zn + a1 (z )zn + + am (z ),
d(z) |fj (z)|2j
C d(w) < +. |w1 . . . |ws |2s |21
with respect to some coordinates (z1 , . . . , zn ) centered at x. Let vj (z ), 1 j m, denote the roots zn of f (z) = 0. On a small neighborhood U of x we have |vj (z )| 1. The inequality between arithmetic and geometric mean implies
{|zn | 1}
aj (z ) n1 ,
aj (0) = 0,
|f (z)|2 dxn dyn =
{|zn | 1} 1 j m
|zn vj (z )|2 dxn dyn |zn vj (z )|2m dxn dyn
1 m
{|zn | 1} 1 j m
{|zn | 2}
dxn dyn , |zn |2m
386
so the Lemma follows from the Fubini theorem. Another interesting application concerns the study of multiplicities of singular points for algebraic hypersurfaces in Pn . Following [Waldschmidt 1975], we introduce the following denition. (8.3) Denition. Let S be a nite subset of Pn . For any integer t 1, we dene t (S) as the minimum of the degrees of non zero homogeneous polynomials P C[z0 , . . . , zn ] which vanish at order t at every point of S, i.e. D P (w) = 0 for every w S and every multi-index = (0 , . . . , n ) of length || < t. It is clear that t t (S) is a non-decreasing and subadditive function, i.e. for all integers t1 , t2 1 we have t1 +t2 (S) t1 (S) + t2 (S). One denes (8.4) For all integers t, t t (S) We nd therefore (8.5) (S) = lim t (S) . t+ t (S) = inf
t
t (S) . 1 t
1, the monotonicity and subadditivity of t (S) show that hence (S) t (S) t 1 1 t (S). + t t
([t/t ] + 1) t (S),
Our goal is to nd a lower bound of (S) in terms of t (S). For n = 1, it is obvious that (S) = t (S)/t = card S for all t. From now on, we assume that n 2. (8.6) Theorem. Let t1 , t2 1 be integers, let P be a homogeneous polynomial of k k degree t2 (S) vanishing at order t2 at every point of S. If P = P1 1 . . . PNN is the 1 decomposition of P in irreducible factors and Zj = Pj (0), we set = t1 + n 1 , t2 = (kj [kj ]) [Zj ], a = dim nil() .
Then we have the inequality t1 (S) + n a 1 t1 + n 1 t2 (S) . t2
Let us rst make a few comments before giving the proof. If we let t2 tend to innity and observe that nil() nnc() by Prop. 8.1 c), we get a 2 and (8.7) t1 (S) + 1 t1 + n 1 (S) t2 (S) . t2
Such a result was rst obtained by [Waldschmidt 1975, 1979] with the lower bound t1 (S)/(t1 + n 1), as a consequence of the Hrmander-Bombieri-Skoda theorem. The o
387
above improved inequalities were then found by (Esnault-Viehweg 1983), who used rather deep tools of algebraic geometry. Our proof will consist in a renement of the BombieriWaldschmidt method due to [Azhari 1990]. It has been conjectured by [Chudnovsky 1979] that (S) (1 (S) + n 1)/n. Chudnovskys conjecture is true for n = 2 (as shown by (8.7)); this case was rst veried independently by [Chudnovsky 1979] and [Demailly 1982]. The conjecture can also be veried in case S is a complete polytope, and the lower bound of the conjecture is then optimal (see [Demailly 1982a] and ??.?.?). More generally, it is natural to ask whether the inequality (8.8) t1 (S) + n 1 t1 + n 1 (S) t2 (S) t2
always holds; this is the case if there are innitely many t2 for which P can be chosen in such a way that nil() has dimension a = 0. (8.9) Bertinis lemma. If E Pn is an analytic subset of dimension a, there exists a dense subset in the grassmannian of k-codimensional linear subspaces Y of Pn such that dim(E Y ) a k (when k > a this means that E Y = ). Proof. By induction on n, it suces to show that dim(E H) a 1 for a generic hyperplane H Pn . Let Ej be the (nite) family of irreducible components of E, and wj Ej an arbitrary point. Then E H = Ej H and we have dim Ej H < dim Ej a as soon as H avoids all points wj . Proof of Theorem 8.6.. By Bertinis lemma, there exists a linear subspace Y Pn of codimension a +1 such that nil() Y = . We consider P as a section of the line bundle (D) over Pn , where D = deg P (cf. Th. V-15.5). There are sections 1 , . . . , a+1 of (1) such that Y = 1 (0). We shall apply Th. 7.1 to E = (1) with its standard hermitian metric, and to L = (k) equipped with the additional weight = log |P |2 . We may assume that the open set U = {|| < 1} is such that nil() U = , otherwise it [k ] suces to multiply by a large constant. This implies that the polynomial Q = Pj j satises |Q|2 e dV = |Pj |2(kj [kj ]) dV < +.
U U
e Set = ic (1) . We have id d log |P |2 ic (D) = D by the Lelong-Poincar equation, thus i(L ) (k D). The desired curvature inequality i(L ) (a + 1 + )i(E) is satised if k D (a + 1 + ). We thus take k = [D] + a + 2. The section f H 0 (U, KPn L) = H 0 U, (k n 1) is taken to be a multiple of Q by some polynomial. This is possible provided that kn1 deg Q D + a + 2 n 1 [kj ] deg Pj ,
or equivalently, as D = (8.10)
kj deg Pj , (kj [kj ]) deg Pj n a 1.
388
Then we get f H 0 (U, KPn L) such that U |f |2 e dV < +. Theorem 7.1 implies the existence of F H 0 (Pn , KPn L), i.e. of a polynomial F of degree k n 1, such that |F |2 dV < + ; |F |2 e dV = 2 Pn |P | Pn observe that || is bounded, for we are on a compact manifold. Near any w S, we have |P (z)| C|z w|t2 , thus |P (z)|2 C|z w|2(t1 +n1) . This implies that the above integral can converge only if F vanishes at order t1 at each point w S. Therefore t1 (S) deg F = k n 1 = [D] + a + 1 n t2 (S) + a + 1 n,
which is the desired inequality. However, the above proof only works under the additional assumption (8.10). Assume on the contrary that = (kj [kj ]) deg Pj < n a 1.
Then the polynomial Q has degree [kj ] deg Pj = deg P = D , and Q vanishes at every point w S with order ordw Q [kj ] ordw Pj = ordw P kj ordw Pj (kj [kj ]) ordw Pj 0, we
t2 = t1 ( n + 1).
This implies ordw Q t1 [ n + 1]. As [ n + 1] < n a 1 n + 1 = a can take a derivative of order [ n + 1] of Q to get a polynomial F with deg F = D + [ n + 1] D n + 1,
which vanishes at order t1 on S. In this case, we obtain therefore t1 (S) D n + 1 = t1 + n 1 t2 (S) n + 1 t2
and the proof of Th. 8.6 is complete.
9. Skodas L2 Estimates for Surjective Bundle Morphisms

Let (X, ) be a Khler manifold, dim X = n, and g : E Q a holomorphic a morphism of hermitian vector bundles over X. Assume in the rst instance that g is surjective. We are interested in conditions insuring for example that the induced morphism g : H k (X, KX E) H k (X, KX Q) is also surjective. For that purpose, it is natural to consider the subbundle S = Ker g E and the exact sequence (9.1) 0 S E Q 0.
g
389
Assume for the moment that S and Q are endowed with the metrics induced by that of E. Let L be a line bundle over X. We consider the tensor product of sequence (9.1) by L : (9.2) 0 S L E L Q L 0. k n. Set r = rk E, q = rkQ,
g
(9.3) Theorem. Let k be an integer such that 0 s = rk S = r q and
m = min{n k, s} = min{n k, r q}. Assume that (X, ) possesses also a complete Khler metric , that E a L X is a hermitian line bundle such that i(L) (m + )i(det Q) 0
m
0, and that
for some > 0. Then for every D -closed form f of type (n, k) with values in Q L such that f < +, there exists a D -closed form h of type (n, k) with values in E L such that f = g h and h 2 (1 + m/) f 2 .
The idea of the proof is essentially due to [Skoda 1978], who actually proved the special case k = 0. The general case appeared in [Demailly 1982c]. Proof. Let j : S E be the inclusion morphism, g : Q E and j : E S the adjoints of g, j, and DE = DS DQ ,
1,0
X, hom(S, Q) ,
0,1
X, hom(Q, S) ,
the matrix of DE with respect to the orthogonal splitting E S Q (cf. V-14). Then g f is a lifting of f in E L. We shall try to nd h under the form h = g f + ju, u L2 (X, S L). n,k
As the images of S and Q in E are orthogonal, we have |h|2 = |f |2 + |u|2 at every point of X. On the other hand DQL f = 0 by hypothesis and D g = j by V-14.3 d), hence DEL h = j( f ) + j DSL = j(DSL f ). We are thus led to solve the equation (9.4)
DSL u = f,
and for that, we apply Th. 4.5 to the (n, k + 1)-form f . One observes now that the curvature of S L can be expressed in terms of . This remark will be used to prove: (9.5) Lemma. A1 ( f ), ( f ) k (m/) |f |2 .
390
If the Lemma is taken for granted, Th. 4.5 yields a solution u of (9.4) in L2 (X, S L) n,q such that u 2 (m/) f 2 . As h 2 = f 2 + u 2 , the proof of Th. 9.3 is complete. Proof of Lemma 9.5.. Exactly as in the proof of Th. VII-10.3, formulas (V-14.6) and (V-14.7) yield i(S) Since
m
i ,
i(det Q)
Grif
TrQ (i ) = TrS (i ). 0, Prop. VII-10.1 implies

m
1,1 (X, Herm S)
:= i
m TrS (i ) IdS +i From the hypothesis on the curvature of L we get i(S L)

m m m
0.
i(S) IdL +(m + ) i(det Q) IdSL (/m) (i ) IdS IdL .
i + (m + ) TrS (i ) IdS IdL
For any v n,k+1 TX S L, Lemma VII-7.2 implies
(9.6)
Ak,SL v, v
(/m) i v, v ,
because rk(S L) = s and m = min{n k, s}. Let (dz1 , . . . , dzn ) be an orthonormal basis of TX at a given point x0 X and set =
1 j n
dzj j ,
j hom(S, Q). dened by
The adjoint of the operator = v= z j
dz j j is the contraction
(j v) =
idzj (j v) = i v. v|2 and (9.6) implies v|2 (m/) Ak,SL v, v |f |2.
We get consequently i v, v = | | f, v |2 = | f, v |2 |f |2 |
If X has a plurisubharmonic exhaustion function , we can select a convex increasing function (R, R) and multiply the metric of L by the weight exp( ) in order to make the L2 norm of f converge. Theorem 9.3 implies therefore: (9.7) Corollary. Let (X, ) be a weakly pseudoconvex Khler manifold, let g : E Q a be a surjective bundle morphism with r = rk E, q = rk Q, let m = min{n k, r q} and let L X be a hermitian line bundle. Suppose that E m 0 and i(L) (m + ) i(det Q) 0
391
for some > 0. Then g induces a surjective map H k (X, KX E L) H k (X, KX Q L). The most remarkable feature of this result is that it does not require any strict positivity assumption on the curvature (for instance E can be a at bundle). A careful examination of the proof shows that it amounts to verify that the image of the coboundary morphism : H k (X, KX Q L) H k+1 (X, KX S L) vanishes; however the cohomology group H k+1 (X, KX S L) itself does not vanish in general as it would do under a strict positivity assumption (cf. Th. VII-9.4). We want now to get also estimates when Q is endowed with a metric given a priori, that can be distinct from the quotient metric of E by g. Then the map g (gg )1 : Q E is the lifting of Q orthogonal to S = Ker g. The quotient metric | | on Q is therefore dened in terms of the original metric | | by |v|2 = |g (gg )1 v|2 = (gg )1 v, v = det(gg )1 gg v, v where gg End(Q) denotes the endomorphism of Q whose matrix is the transposed of the comatrix of gg . For every w det Q, we nd |w|2 = det(gg )1 |w|2 . If Q denotes the bundle Q with the quotient metric, we get i(det Q ) = i(det Q) + id d log det(gg ). In order that the hypotheses of Th. 9.3 be satised, we are led to dene a new metric m . Then | | on L by |u|2 = |u|2 det(gg ) i(L ) = i(L) + (m + ) id d log det(gg ) (m + ) i(det Q ). Theorem 9.3 applied to (E, Q , L ) can now be reformulated: (9.8) Theorem. Let X be a complete Khler manifold equipped with a Khler metric a a on X, let E Q be a surjective morphism of hermitian vector bundles and let L X be a hermitian line bundle. Set r = rk E, q = rk Q and m = min{n k, r q} and suppose E m 0, i(L) (m + )i(det Q) 0 for some > 0. Then for every D -closed form f of type (n, k) with values in Q L such that I= gg f, f (det gg )m1 dV < +,
X
there exists a D -closed form h of type (n, k) with values in E L such that f = g h and |h|2 (det gg )m dV (1 + m/) I.
X
392
Our next goal is to extend Th. 9.8 in the case when g : E Q is only generically surjective; this means that the analytic set Y = {x X ; gx : Ex Qx is not surjective } dened by the equation q g = 0 is nowhere dense in X. Here q g is a section of the bundle hom(q E, det Q). (9.9) Theorem. The existence statement and the estimates of Th. 9.8 remain true for a generically surjective morphism g : E Q provided that X is weakly pseudoconvex. Proof. Apply Th. 9.8 to each relatively compact domain Xc Y (these domains are complete Khler by Lemma 7.2). From a sequence of solutions on Xc Y we can extract a a subsequence converging weakly on X Y as c tends to +. One gets a form h satisfying the estimates, such that D h = 0 on X Y and f = g h. In order to see that D h = 0 on X, it suces to apply Lemma 7.3 and to observe that h has L2 coecients on X by loc our estimates. A very special but interesting case is obtained for the trivial bundles E = Cr , Q = C over a pseudoconvex open set Cn . Then the morphism g is given by a r-tuple (g1 , . . . , gr ) of holomorphic functions on . Let us take k = 0 and L = C with the metric given by a weight e . If we observe that gg = Id when rk Q = 1, Th. 9.8 applied on X = g 1 (0) and Lemmas 7.2, 7.3 give: (9.10) Theorem ([Skoda 1978]). Let be a complete Khler open subset of Cn and a a plurisubharmonic function on . Set m = min{n, r 1}. Then for every holomorphic function f on such that I=
Z
|f |2 |g|2(m+1+)e dV < +,
where Z = g 1 (0), there exist holomorphic functions (h1 , . . . , hr ) on such that f = gj hj and
Y
|h|2 |g|2(m+)e dV
(1 + m/)I.
This last theorem can be used in order to obtain a quick solution of the Levi problem mentioned in I-4. It can be used also to prove a result of [Diederich-Pug 1981], relating the pseudoconvexity property and the existence of complete Khler metrics for domains a n of C . (9.11) Theorem. Let Cn be an open subset. Then: a) is a domain of holomorphy if and only if is pseudoconvex ; a b) If () = and if has a complete Khler metric , then is pseudoconvex. Note that statement b) can be false if the assumption () = is omitted: in fact Cn {0} is complete Khler by Lemma 7.2, but it is not pseudoconvex if n 2. a
10. Application of Skodas L2 Estimates to Local Algebra
393
Proof. b) By Th. I-4.12, it is enough to verify that is a domain of holomorphy, i.e. that for every connected open subset U such that U = and every connected component W of U there exists a holomorphic function h on such that hW cannot be continued to U . Since () = , the set U is not empty. We select a U . Then the integral
|z a|2(n+) dV (z)
converges. By Th. 9.10 applied to f (z) = 1, gj (z) = zj aj and = 0, there exist holomorphic functions hj on such that (zj aj ) hj (z) = 1. This shows that at least one of the functions hj cannot be analytically continued at a U . a) Assume that is pseudoconvex. Given any open connected set U such that U = , choose a U . By Th. I-4.14 c) the function (z) = (n + )(log(1 + |z|2 ) 2 log d(z, ) is plurisubharmonic on . Then the integral |z a|2(n+) e(z) dV (z) (1 + |z|2 )n dV (z)
converges, and we conclude as for b).

We apply here Th. 9.10 to the study of ideals in the ring n = C{z1 , . . . , zn } of germs of holomorphic functions on (Cn , 0). Let = (g1 , . . . , gr ) = (0) be an ideal of n . (10.1) Denition. Let k R+ . We associate to a) the ideal of germs u n such that |u| |g|2 = |g1 |2 + + |gr |2 .
(k)
the following ideals:

C|g|k for some constant C 0, where
b) the ideal
(k) of germs u n such that
|u|2 |g|2(k+) dV < +
on a small ball centered at 0, if > 0 is small enough. (10.2) Proposition. For all k, l R+ we have a) b) c) d)
(k) (k) ; k (k) if k N ; (k) .(l) (k+l) ; (k) .(l) (k+l) .
394
All properties are immediate from the denitions except a) which is a consequence of Lemma 8.2. Before stating the main result, we need a simple lemma. (10.3) Lemma. If = (g1 , . . . , gr ) and r > n, we can nd elements g1 , . . . , gn such that C 1 |g| |g| C|g| on a neighborhood of 0. Each gj can be taken to be a linear combination g j = aj . g = ajk gk , aj Cr {0}
1 k r
where the coecients ([a1 ], . . . , [an ]) are chosen in the complement of a proper analytic subset of (Pr1 )n . It follows from the Lemma that the ideal and (k) = (k) for all k.
= (g1, . . . , gn) satises (k) = (k)
Proof. Assume that g ()r . Consider the analytic subsets in (Pr1 )n dened by A = (z, [w1 ], . . . , [wn ]) ; wj . g(z) = 0 , A = irreducible components of A not contained in g 1 (0) (Pr1 )n .
For z g 1 (0) the ber Az = {([w1 ], . . . , [wn ]) ; wj . g(z) = 0} = A is a product of / z n hyperplanes in Pr1 , hence A ( g 1 (0)) (Pr1 )n is a ber bundle with base g 1 (0) and ber (Pr2 )n . As A is the closure of this set in (Pr1 )n , we have dim A = n + n(r 2) = n(r 1) = dim(Pr1 )n . It follows that the zero ber A = A {0} (Pr1 )n 0 is a proper subset of {0} (Pr1 )n . Choose (a1 , . . . , an ) (Cr {0})n such that (0, [a1 ], . . . , [an]) is not in A . By an easy compactness argument the set 0 r1 n A B(0, ) (P ) is disjoint from the neighborhood B(0, ) [B(aj , )]
of (0, [a1 ], . . . , [an ]) for small enough. For z B(0, ) we have |aj . g(z)| |g(z)| for some j, otherwise the inequality |aj . g(z)| < |g(z)| would imply the existence of hj Cr with |hj | < and aj . g(z) = hj . g(z). Since g(z) = 0, we would have (z, [a1 h1 ], . . . , [an hn ]) A B(0, ) (Pr1 )n , a contradiction. We obtain therefore |g(z)| max |aj . g(z)| (max |aj |) |g(z)| on B(0, ).
(10.4) Theorem ([Brianon-Skoda 1974]). Set p = min{n 1, r 1}. Then c a)
(k+1) = (k) = (k)
for k
p.
395
(k+p) (k+p) k for all k N. Proof. a) The inclusions (k) (k) (k+1) are obvious thanks to Prop. 10.2, so we only have to prove that (k+1) (k) . Assume rst that r n. Let f (k+1) be
b) such that
|f |2 |g|2(k+1+) dV < +.
For k p 1, we can apply Th. 9.10 with m = r 1 and with the weight = (k m) log |g|2 . Hence f can be written f = gj hj with
|h|2 |g|2(k+) dV < +,
thus hj (k) and f (k) . When r > n, Lemma 10.3 shows that there is an ideal with n generators such that (k) = (k) . We nd
(k+1) = (k+1) (k) (k)
for k
n 1.
b) Property a) implies inductively (k+p) = k (p) for all k N. This gives in particular (k+p) k . (10.5) Corollary. a) The ideal is the integral closure of which satisfy an equation
, i.e. by denition the set of germs u n

as s , 1 s d.
ud + a1 ud1 + + ad = 0, b) Similarly,
(k) is the set of germs u n which satisfy an equation 1 s d, ud + a1 ud1 + + ad = 0, as ]ks[ ,

t.
(k)
where ]t[ denotes the smallest integer As the ideal rational number l
Conversely, assume that u . The ring n is Noetherian, so the ideal (p) has a nite number of generators v1 , . . . , vN . For every j we have uvj (p) = (p) , hence there exist elements bjk such that uvj =
1 k N
Proof. a) If u n satises a polynomial equation with coecients as s , then clearly |as | Cs |g|s and Lemma II-4.10 implies |u| C |g|.
is nitely generated, property b) shows that there always exists a k such that (l) = (k) .
bjk vk .
The matrix (ujk bjk ) has the non zero vector (vj ) in its kernel, thus u satises the equation det(ujk bjk ) = 0, which is of the required type.
396
If k Q, we can nd m such that d(km + /2, Z) < /2, thus km + N for some / ]0, [. If k Q, we take m such that km N and = 0. Then um vj (N+p) = N (p) with N = km + N,
b) Observe that v1 , . . . , vN satisfy simultaneously some integrability condition |v |2(p+) < +, thus (p) = (p+) for [0, [. Let u (k) . For every in j teger m N we have (km) (p+) um vj (km++p) .
and the reasoning made in a) gives det(um jk bjk ) = 0 for some bjk N . This is an equation of the type described in b), where the coecients as vanish when s is not a multiple of m and ams Ns ]kms[ . Let us mention that Brianon and Skodas result 10.4 b) is optimal for k = 1. Take c r j r, and f (z) = z1 . . . zr . Then for example = (g1 , . . . , gr ) with gj (z) = zj , 1 |f | C|g| and 10.4 b) yields f r ; however, it is easy to verify that f r1 . The / theorem also gives an answer to the following conjecture made by J. Mather. (10.6) Corollary. Let f n and f = (z1 f /z1 , . . . , zn f /zn ). Then f f , and for every integer k 0, f k+n1 k . f The Corollary is also optimal for k = 1 : for example, one can verify that the function 3n1 3n1 f (z) = (z1 . . . zn )3 + z1 + . . . + zn is such that f n1 f . / Proof. Set gj (z) = zj f /zj , 1 j n. By 10.4 b), it suces to show that |f | C|g|. For every germ of analytic curve C t (t), 0, the vanishing order of f (t) at t = 0 is the same as that of d(f ) f t (t) . = t j (t) dt zj
1 j n
We thus obtain |f (t)| C1 |t| d(f ) dt C2

1 j n |t j (t)|
f (t) zj
C3 |g (t)|
and conclude by the following elementary lemma. (10.7) Lemma. Let f, g1 , . . . , gr n be germs of holomorphic functions vanishing at 0. Then we have |f | C|g| for some constant C if and only if for every germ of analytic curve through 0 there exists a constant C such that |f | C |g |. Proof. If the inequality |f | C|g| does not hold on any neighborhood of 0, the germ of analytic set (A, 0) (Cn+r , 0) dened by gj (z) = f (z)zn+j , 1 j r, contains a sequence of points z , gj (z )/f (z ) converging to 0 as tends to +, with f (z ) = 0. Hence (A, 0) contains an irreducible component on which f 0 and there is a germ of curve = (, n+j ) : (C, 0) (Cn+r , 0) contained in (A, 0) such that f 0. We get gj = (f )n+j , hence |g (t)| C|t| |f (t)| and the inequality |f | C |g | does not hold.
11. Integrability of Almost Complex Structures
397

Let M be a manifold of real dimension m = 2n. An almost complex structure on M is by denition an endomorphism J End(T M ) of class such that J 2 = Id. Then T M becomes a complex vector bundle for which the scalar multiplication by i is given by J. The pair (M, J) is said to be an almost complex manifold. For such a manifold, the complexied tangent space TC M = C R T M splits into conjugate complex subspaces (11.1) TC M = T 1,0 M T 0,1 M, dimC T 1,0 M = dimC T 0,1 M = n,
where T 1,0 M , T 0,1 M TC M are the eigenspaces of Id J corresponding to the eigen values i and i. The complexied exterior algebra C R T M = TC M has a corresponding splitting (11.2) where we denote by denition (11.3)
p,q TC M = p (T 1,0 M ) C q (T 0,1 M ) . k TC M = p+q=k p,q TC M
s As for complex manifolds, we let Cp,q (M, E) be the space of dierential forms of class C s and bidegree (p, q) on M with values in a complex vector bundle E. There is a natural antisymmetric bilinear map
(M, T 1,0 M )
(M, T 1,0 M )
(M, T 0,1 M )
which associates to a pair (, ) of (1, 0)-vector elds the (0, 1)-component of the Lie bracket [, ]. Since [, f ] = f [, ] + (.f ) , f
(M, C)
we see that (, f ) = f (, ). It follows that is in fact a (2, 0)-form on M with values in T 0,1 M . If M is a complex analytic manifold and J its natural almost complex structure, we have in fact = 0, because [/zj , /zk ] = 0, 1 j, k n, for any holomorphic local coordinate system (z1 , . . . , zn ). (11.4) Denition. The form (M, T 0,1 M ) is called the torsion form of J. The 2,0 almost complex structure J is said to be integrable if = 0. (11.5) Example. If M is of real dimension m = 2, every almost complex structure is integrable, because n = 1 and alternate (2, 0)-forms must be zero. Assume that M is a smooth oriented surface. To any Riemannian metric g we can associate the endomorphism J End(T M ) equal to the rotation of +/2. A change of orientation changes J into the conjugate structure J. Conversely, if J is given, T M is a complex line bundle, so M is oriented, and a Riemannian metric g is associated to J if and only if g is J-hermitian. As a consequence, there is a one-to-one correspondence between conformal
398
classes of Riemannian metrics on M and almost complex structures corresponding to a given orientation. If (M, J) is an almost complex manifold and u (M, C), we let d u, d u be the p,q components of type (p + 1, q) and (p, q + 1) in the exterior derivative du. Let (1 , . . . , n ) be a frame of T 1,0 M . The torsion form can be written =
1 j n
j j ,
2,0 (, C).
Then yields conjugate operators , on TC M such that
(11.6)
u =
1 j n
j ( j
u),
u =
1 j n
j (j
u).
If u is of bidegree (p, q), then u and u are of bidegree (p + 2, q 1) and (p 1, q + 2). It is clear that , are derivations, i.e. (u v) = ( u) v + (1)deg u u ( v) for all smooth forms u, v, and similarly for . (11.7) Proposition. We have d = d + d . Proof. Since all operators occuring in the formula are derivations, it is sucient to check the formula for forms of degree 0 or 1. If u is of degree 0, the result is obvious because u = u = 0 and du can only have components of types (1, 0) or (0, 1). If u is a 1-form and , are complex vector elds, we have du(, ) = .u() .du() u([, ]). When u is of type (0, 1) and , of type (1, 0), we nd (du)2,0 (, ) = u (, ) thus (du)2,0 = u, and of course (du)1,1 = d u, (du)0,2 = d u, u = 0 by denition. The case of a (1, 0)-form u follows by conjugation. Proposition 11.7 shows that J is integrable if and only if d = d + d . In this case, we infer immediately d2 = 0, d d + d d = 0, d2 = 0. For an integrable almost complex structure, we thus have the same formalism as for a complex analytic structure, and indeed we shall prove: (11.8) Newlander-Nirenberg theorem (1957). Every integrable almost complex structure J on M is dened by a unique analytic structure. The proof we shall give follows rather closely that of [Hrmander 1966], which was o itself based on previous ideas of [Kohn 1963, 1964]. A function f C 1 (, C), M , is
399
said to be J-holomorphic if d f = 0. Let f1 , . . . , fp C 1 (, C) and let h be a function of class C 1 on an open subset of Cp containing the range of f = (f1 , . . . , fp ). An easy computation gives (11.9) d (h f ) = h h f d fj + f d fj , zj z j
1 j p
in particular h f is J-holomorphic as soon as f1 , . . . , fp are J-holomorphic and h holomorphic in the usual sense. Constructing a complex analytic structure on M amounts to show the existence of Jholomorphic complex coordinates (z1 , . . . , zn ) on a neighborhood of every point a M . Formula (11.9) then shows that all coordinate changes h : (zk ) (wk ) are holomorphic in the usual sense, so that M is furnished with a complex analytic atlas. The uniqueness of the analytic structure associated to J is clear, since the holomorphic functions are characterized by the condition d f = 0. In order to show the existence, we need a lemma. (11.10) Lemma. For every point a M and every integer s complex coordinates (z1 , . . . , zn ) centered at a such that d zj = O(|z|s ), 1 j n. 1, there exist
Proof. By induction on s. Let (1 , . . . , n ) be a basis of 1,0 TC M . One can nd complex functions zj such that dzj (a) = j , i.e. d zj (a) = j ,
d zj (a) = 0.
Then (z1 , . . . , zn ) satisfy the conclusions of the Lemma for s = 1. If (z1 , . . . , zn ) are already constructed for the integer s, we have a Taylor expansion d zj =
1 k n
Pjk (z, z) d zk + O(|z|s+1 )
where Pjk (z, w) is a homogeneous polynomial in (z, w) Cn Cn of total degree s. As J is integrable, we have 0 = d2 zj =
1 k,l n
Pjk Pjk d zl d zk + d zl d zk + O(|z|s ) zl z l Pjl Pjk d zl d zk + O(|z|s ) z l z k
=
1 k<l n
because Pjk /zl is of degree s 1 and d zl = O(|z|s ). Since the polynomial between brackets is of degree s 1, we must have Pjl Pjk = 0, z l z k j, k, l.
400
We dene polynomials Qj of degree s + 1

1
Qj (z, z) =
0 1 l n
z l Pjl (z, tz) dt.
Trivial computations show that Qj = z k =

1
Pjk +
0 1 1 l n
zl
Pjl (z, tz) dt z k
d t Pjk (z, tz) dt = Pjk (z, z), 0 dt Qj Qj d zj Qj (z, z) = d zj d zk d zk z k zk

1 k n 1 k n
1 k n
Qj d zk + O(|z|s+1 ) = O(|z|s+1 ) zk
because Qj /zk is of degree s and d zl = O(|z|). The new coordinates zj = zj Qj (z, z), fulll the Lemma at step s + 1. All usual notions dened on complex analytic manifolds can be extended to integrable almost complex manifolds. For example, a smooth function is said to be strictly plurisubharmonic if id d is a positive denite (1, 1)-form. Then = id d is a Khler a metric on (M, J). In this context, all L2 estimates proved in the previous paragraphs still apply to an integrable almost complex manifold; remember that the proof of the Bochner-KodairaNakano identity used only Taylor developments of order 2, and the coordinates given by Lemma 11.10 work perfectly well for that purpose. In particular, Th. 6.5 is still valid. (11.11) Lemma. Let (z1 , . . . , zn ) be coordinates centered at a point a M with d zj = O(|z|s ), s 3. Then the functions (z) = |z|2 , Proof. We have id d = i
1 j n
(z) = |z|2 + log(|z|2 + 2 ),
]0, 1]
are strictly plurisubharmonic on a small ball |z| < r0 . d zj d zj + d z j d zj + zj d d z j + z j d d zj .
The last three terms are O(|z|s ) and the rst one is positive denite at z = 0, so the result is clear for . Moreover id d = id d + i d zj d zj z j d zj (|z|2 + 2 )2 O(|z|s+2 ) O(|z|s ) + . + 2 |z| + 2 (|z|2 + 2 )2 (|z|2 + 2 ) z j d zj
401
We observe that the rst two terms are positive denite, whereas the remainder is O(|z|) uniformly in . Proof of theorem 11.8.. With the notations of the previous lemmas, consider the pseudoconvex open set = {|z| < r} = {(z) r 2 < 0}, r < r0 ,
endowed with the Khler metric = id d . Let h () be a cut-o function with a 0 h 1 and h = 1 on a neighborhood of z = 0. We apply Th. 6.5 to the (0, 1)-forms gj = d zj h(z) for the weight (z) = A|z|2 + (n + 1) log |z|2 = lim A|z|2 + (n + 1) log(|z|2 + 2 ).
0 0,1 (, C)
Lemma 11.11 shows that is plurisubharmonic for A n + 1, and for A large enough we obtain id d + Ricci() on . By Remark (6.8) we get a function fj such that d fj = gj and |fj |2 e dV |gj |2 e dV.
As gj = d zj = O(|z|s ) and e = O(|z|2n2 ) near z = 0, the integral of gj converges provided that s 2. Then |fj (z)|2 |z|2n2 dV converges also at z = 0. Since the solution fj is smooth, we must have fj (0) = dfj (0) = 0. We set zj = zj h(z) fj , 1 j n.
Then zj is J-holomorphic and dzj (0) = dzj (0), so (z1 , . . . , zn ) is a J-holomorphic coordinate system at z = 0. In particular, any Riemannian metric on an oriented 2-dimensional real manifold denes a unique analytic structure. This fact will be used in order to obtain a simple proof of the well-known: (11.12) Uniformization theorem. Every simply connected Riemann surface X is biholomorphic either to P1 , C or the unit disk . Proof. We will merely use the fact that H 1 (X, R) = 0. If X is compact, then X is a complex curve of genus 0, so X P1 by Th. VI-14.16. On the other hand, the elementary Riemann mapping theorem says that an open set C with H 1 (, R) = 0 is either equal to C or biholomorphic to the unit disk. Thus, all we have to show is that a non compact Riemann surface X with H 1 (X, R) = 0 can be embedded in the complex plane C. Let be an exhausting sequence of relatively compact connected open sets with smooth boundary in X. We may assume that X has no relatively compact connected components, otherwise we ll the holes of by taking the union with all such
402
components. We let Y be the double of the manifold with boundary ( , ), i.e. the union of two copies of with opposite orientations and the boundaries identied. Then Y is a compact oriented surface without boundary. (11.13) Lemma. We have H 1 (Y , R) = 0.
1 Proof. Let us rst compute Hc ( , R). Let u be a closed 1-form with compact support 1 in . By Poincar duality Hc (X, R) = 0, so u = df for some function f (X). As e df = 0 on a neighborhood of X and as all connected components of this set are non compact, f must be equal to the constant zero near X . Hence u = df is the zero 1 1 class in Hc ( , R) and we get Hc ( , R) = H 1 ( , R) = 0. The exact sequence of the pair ( , ) yields 1 R = H 0 ( , R) H 0 ( , R) H 1 ( , ; R) Hc ( , R) = 0,
thus H 0 ( , R) = R. Finally, the Mayer-Vietoris sequence applied to small neighborhoods of the two copies of in Y gives an exact sequence H 0 ( , R)2 H 0 ( , R) H 1 (Y , R) H 1 ( , R)2 = 0 where the rst map is onto. Hence H 1 (Y , R) = 0. Proof End of the proof of the uniformization theorem.. Extend the almost complex structure of in an arbitrary way to Y , e.g. by an extension of a Riemannian metric. Then Y becomes a compact Riemann surface of genus 0, thus Y P1 and we obtain in particular a holomorphic embedding : C. Fix a point a 0 and a non zero linear form Ta X. We can take the composition of with an ane linear map C C so that (a) = 0 and d (a) = . By the well-known properties of injective holomorphic maps, ( ) is then uniformly bounded on every small disk centered at a, thus also on every compact subset of X by a connectedness argument. Hence ( ) has a subsequence converging towards an injective holomorphic map : X C.
403
Chapter IX
Finiteness Theorems for q-Convex Spaces and Stein Spaces
1. Topological Preliminaries
1.A. Krull Topology of
n-Modules
We shall use in an essential way dierent kind of topological results. The rst of these concern the topology of modules over a local ring and depend on the Artin-Rees and Krull lemmas. Let R be a noetherian local ring; local means that R has a unique maximal ideal m, or equivalently, that R has an ideal m such that every element R m is invertible. (1.1) Nakayama lemma. Let E be a nitely generated R-module such that mE = E. Then E = {0}. Proof. By induction on the number of generators of E : if E is generated by x1 , . . . , xp , the hypothesis E = mE shows that xp = 1 x1 + +p xp with j m ; as 1p R m is invertible, we see that xp can be expressed in terms of x1 , . . . , xp1 if p > 1 and that x1 = 0 if p = 1. (1.2) Artin-Rees lemma. Let F be a nitely generated R-module and let E be a submodule. There exists an integer s such that E mk F = mks (E ms F ) for k s.
Proof. Let Rt be the polynomial ring R[mt] = R + mt + + mk tk + where t is an indeterminate. If g1 , . . . , gp is a set of generators of the ideal m over R, we see that the ring Rt is generated by g1 t, . . . , gp t over R, hence Rt is also noetherian. Now, we consider the Rt -modules Et = E tk , Ft = (mk F ) tk . Then Ft is generated over Rt by the generators of F over R, hence the submodule Et Ft is nitely generated. Let s be the highest exponent of t in a set of generators P1 (t), . . . , PN (t) of Et Ft . If we identify the components of tk in the extreme terms of the equality mk tk Pj (t), E mk F tk = Et Ft =
j k
404
Chapter IX. Finiteness Theorems for q-Convex Spaces and Stein Spaces
we get E mk F
l s
mkl (E ml F ) mks (E ms F ).
The opposite inclusion is clear. (1.3) Krull lemma. Let F be a nitely generated R-module and let E be a submodule. Then a) b)
k 0
mk F = {0}. + mk F ) = E.
k 0 (E
Proof. a) Put G = k 0 mk F F . By the Artin-Rees lemma, there exists s N such that G mk F = mks (G ms F ). Taking k = s + 1, we nd G mG, hence mG = G and G = {0} by the Nakayama lemma. b) By applying a) to the quotient module F/E we get follows. mk (F/E) = {0}. Property b)
Now assume that R = n = C{z1 , . . . , zn } and m = (z1 , . . . , zn ). Then n /mk is a nite dimensional vector space generated by the monomials z , || < k. It follows that E/mk E is a nite dimensional vector space for any nitely generated n -module E. As mk E = {0} by 1.3 a), there is an injection (1.4) E E/mk E.
kN
We endow E with the Hausdor topology induced by the product, i.e. with the weakest topology that makes all projections E E/mk E continuous for the complex vector space topology on E/mk E. This topology is called the Krull topology (or rather, the analytic Krull topology; the algebraic Krull topology would be obtained by taking the discrete topology on E/mk E). For E = n , this is the topology of simple convergence on coecients, dened by the collection of semi-norms c z |c |. Observe that this topology is not complete: the completion of n can be identied with the ring of formal power series C[[z1 , . . . , zn ]]. In general, the completion is the inverse limit E = lim E/mk E. Every n -homomorphism E F is continuous, because the induced nite dimensional linear maps E/mk E F/mk F are continuous. (1.5) Theorem. Let E F be nitely generated
n -modules. Then:
a) The map F G = F/E is open, i.e. the Krull topology of G is the quotient of the Krull topology of F ; b) E is closed in F and the topology induced by F on E coincides with the Krull topology of E. Proof. a) is an immediate consequence of the fact that the surjective nite dimensional linear maps F/mk F G/mk G are open.
405
b) Let E be the closure of E in F . The image of E in F/mk F is mapped into the closure of the image of E. As every subspace of a nite dimensional space is closed, the images of E and E must coincide, i.e. E + mk F = E + mk F . Therefore EE (E + mk F ) = E
thanks to 1.3 b). The topology induced by F on E is the weakest that makes all projections E E/E mk F continuous (via the injections E/E mk F F/mk F ). However, the Artin-Rees lemma gives mk E E mk F = mks (E ms F ) mks E so the topology induced by F coincides with that induced by 1.B. Compact Pertubations of Linear Operators We now recall some basic results in the perturbation theory of linear operators. These results will be needed in order to obtain a niteness criterion for cohomology groups. (1.6) Denition. Let E, F be Hausdor locally convex topological vector spaces and g : E F a continuous linear operator. a) g is said to be compact if there exists a neighborhood U of 0 in E such that the image g(U ) is compact in F . b) g is said to be a monomorphism if g is a topological isomorphism of E onto a closed subspace of F , and a quasi-monomorphism if ker g is nite dimensional and g : E/ ker g F a monomorphism. c) g is said to be an epimorphism if g is surjective and open, and a quasi-epimorphism if g is an epimorphism of E onto a closed nite codimensional subspace F F . d) g is said to be a quasi-isomorphism if g is simultaneously a quasi-monomorphism and a quasi-epimorphism. (1.7) Lemma. Assume that E, F are Frchet spaces. Then e a) g is a (quasi-) monomorphism if and only if g(E) is closed in F and g is injective (resp. and ker g is nite dimensional ). b) g is a (quasi-) epimorphism if and only if g is surjective (resp. g(E) is nite codimensional ). Proof. a) If g(E) is closed, the map g : E/ ker g g(E) is a continuous bijective linear map between Frchet spaces, so g is a topological isomorphism by Banachs theorem. e b) If g is surjective, Banachs theorem implies that g is open, thus g is an epimorphism. If g(E) is nite codimensional, let S be a supplementary subspace of g(E) in F , dim S < +. Then the map G : (E/ ker g) S F, x y g(x) + y for k s,
E/mk E.
406
is a bijective linear map between Frchet spaces, so it is a topological isomorphism. In e particular g(E) = G (E/ ker g) {0} is closed as an image of a closed subspace. Hence g(E) is also a Frchet space and g : E g(E) is an epimorphism. e (1.8) Theorem. Let h : E F be a compact linear operator. a) If g : E F is a quasi-monomorphism, then g + h is a quasi-monomorphism. b) If E, F are Frchet spaces and if g : E F is a quasi-epimorphism, then g + h is e a quasi-epimorphism. Proof. Set f = g + h and let U be an open convex symmetric neighborhood of 0 in E such that K = h(U ) is compact. a) It is sucient to show that there is a nite dimensional subspace E E such that fE is a monomorphism. If we take E equal to a supplementary subspace of ker g, we see that we may assume g injective. Then g is a monomorphism, so we may assume in fact that E is a subspace of F and that g is the inclusion. Let V be an open convex symmetric neighborhood of 0 in F such that U = V E. There exists a closed nite codimensional subspace F F such that K F 21 V because F K F = {0}. If we replace E by E = h1 (F ) and U by U = U E , we get K := h(U ) K F 21 V. Hence, we may assume without loss of generality that K 21 V . Then we show that f = g + h is actually a monomorphism. If is an arbitrary open neighborhood of 0 in E, we have to check that there exists a neighborhood W of 0 in F such that f (x) W = x . There is an integer N such that 2N K E . We choose W convex and so small that (W + 2N K) E and 2N W + K 21 V.
Let x E be such that f (x) W . Then x 2n U for n large enough and we infer x = f (x) h(x) W + 2n K 2n1 V provided that n N.
Thus x 2n1 V E = 2n1 U . By induction we nally get x 2N U , so x (W + 2N K) E . b) By Lemma 1.7 b), we only have to show that there is a nite dimensional subspace S F such that the induced map f : E F F/S is surjective. If we take S equal to a supplementary subspace of g(E) and replace g, h by the induced maps g, h : E F/S, we may assume that g itself is surjective. Then g is open, so V = g(U ) is a convex open neighborhood of 0 in F . As K is compact, there exists a nite set of elements b1 , . . . , bN K such that K (bj + 21 V ). If we take now S = Vect(b1 , . . . , bN ), we obtain K 21 V where K is the closure of h(U ) and
407
V = g(U ), so we may assume in addition that K 21 V . Then we show that f = g + h is actually surjective. Let y0 V . There exists x0 U such that g(x0 ) = y0 , thus y1 = y0 f (x0 ) = h(x0 ) K 21 V. By induction, we construct xn 2n U such that g(xn ) = yn and yn+1 = yn f (xn ) = h(xn ) 2n K 2n1 V. Hence yn+1 = y0 f (x0 + +xn ) tends to 0 in F , but we still have to make sure that the series xn converges in E. Let Up be a fundamental system of convex neighborhoods of 0 in E such that Up+1 21 Up . For each p, K is contained in the union of the open sets g(2n Up 21 U ) when n N, equal to g(21 U ) = 21 V . There exists an increasing sequence N (p) such that K g(2N(p) Up 21 U ), thus 21n K g(2N(p)+1n Up 2n U ). As yn 21n K, we see that we can choose xn 2N(p)+1n Up 2n U for N (p) < n N (p + 1) ; then xN(p)+1 + + xN(p+1) (1 + 21 + ) Up 2 Up . As E is complete, the series x = y0 , and f is surjective. xn converges towards an element x such that f (x) =
The following important niteness theorem due to L. Schwartz can be easily deduced from this. (1.9) Theorem. Let (E , d) and (F , ) be complexes of Frchet spaces with continuous e dierentials, and : E F a continuous complex morphism. If q is compact and H q ( ) : H q (E ) H q (F ) surjective, then H q (F ) is a Hausdor nite dimensional space. Proof. Consider the operators g, h : Z q (E ) F q1 Z q (F ),
g(x y) = q (x) + q1 (y),
h(x y) = q (x).
As Z q (E ) E q , Z q (F ) F q are closed, all our spaces are Frchet spaces. Moreover e the hypotheses imply that h is compact and g is surjective since H q ( ) is surjective. Hence g is an epimorphism and f = g + h = 0 q1 is a quasi-epimorphism by 1.8 b). Therefore B q (F ) is closed and nite codimensional in Z q (F ), thus H q (F ) is Hausdor and nite dimensional. (1.10) Remark. If : E F is a continuous morphism of Frchet complexes and e q q if H ( ) is surjective, then H ( ) is in fact open, because the above map g is open. If H q ( ) is bijective, it follows that H q ( ) is necessarily a topological isomorphism (however H q (E ) and H q (F ) need not be Hausdor).
408
1.C. Abstract Mittag-Leer Theorem We will also need the following abstract Mittag-Leer theorem, which is a very ecient tool in order to deal with cohomology groups of inverse limits.
(1.11) Proposition. Let (E , )N be a sequence of Frchet complexes together with e morphisms E+1 E . We assume that the image of E+1 in E is dense and we let E = lim E be the inverse limit complex.
a) If all maps H q (E+1 ) H q (E ), N, are surjective, then the limit q q H (E ) H (E0 ) is surjective.
b) If all maps H q (E+1 ) H q (E ), q q H (E ) H (E0 ) has a dense range.
N, have a dense range, then
c) If all maps H q1 (E+1 ) H q1 (E ) have a dense range and all maps q q H (E+1 ) H (E ) are injective, N, then H q (E ) H q (E0 ) is injective.
d) Let : F E be a morphism of Frchet complexes that has a dense range. If e q q every map H (F ) H (E ) has a dense range, then H q (F ) H q (E ) has a dense range.
Proof. If x is an element of E or of E , , we denote by x its canonical image in E . Let d be a translation invariant distance that denes the topology of E . After replacement of d (x, y) by
d (x, y) = max d (x , y ) ,

x, y E ,
we may assume that all maps E+1 E are Lipschitz continuous with coecient 1. a) Let x0 Z q (E0 ) represent a given cohomology class x0 H q (E0 ). We construct by induction a convergent sequence x Z q (E ) such that x is mapped onto x0 . If x is already chosen, we can nd by assumption x+1 Z q (E+1 ) such that x +1 = x , q1 i.e. x+1 = x + y for some y E . If we replace x+1 by x+1 y+1 where q1 y+1 E+1 yields an approximation y+1 of y , we may assume that
max{d (y , 0), d (y , 0)}
2 .
0 y .
Then (x ) converges to a limit Z q (E ) and we have 0 = x0 +

q1 Z q (E+1 ) E Z q (E ),
b) The density assumption for cohomology groups implies that the map (x+1 , y ) x + y +1
q1 has a dense range. If we approximate y by elements coming from E+1 , we see that the map Z q (E+1 ) Z q (E ) has also a dense range. If x0 Z q (E0 ), we can nd inductively a sequence x Z q (E ) such that d (x , x ) 21 for all , thus (x ) +1 q converges to an element Z (E ) such that d0 ( 0 , x0 ) and Z q (E ) Z q (E0 ) has a dense range. c) Let x Z q (E ) be such that x0 H q (E0 ) is zero. By assumption, the image q1 of x in H q (E ) must be also zero, so we can write x = dy , y E . We have
2. q-Convex Spaces
409
z = y+1 y Z q1 (E ). Let z+1 Z q1 (E+1 ) be such that z+1 approximates z . If we replace y+1 by y+1 z+1 , we still have x+1 = dy+1 and we may assume in addition that d (y+1 , y ) 2 . Then (y ) converges towards an element y E q1 such that x = dy, thus x = 0 and H q (E ) H q (E0 ) is injective.
d) For every class y H q (E ), the hypothesis implies the existence of a sequence x Z q (F ) such that q (x ) converges to y , that is, d (y , q (x ) + z ) tends to 0 for q1 some sequence z E . Approximate z by q1 (w ) for some w F q1 and replace x by x = x + w . Then q (x ) converges to y in Z q (E ).
2. q-Convex Spaces
2.A. q-Convex Functions The concept of q-convexity, rst introduced in [Rothstein 1955], and further developed by [Andreotti-Grauert 1962], generalizes the concepts of pseudoconvexity already considered in chapters 1 and 8. Let M be a complex manifold, dimC M = n. A function v C 2 (M, R) is said to be strongly (resp. weakly) q-convex at a point x M if id d v(x) has at least (n q + 1) strictly positive (resp. nonnegative) eigenvalues, or equivalently if there exists a (n q + 1)-dimensional subspace F Tx M on which the complex Hessian Hx v is positive denite (resp. semi-positive). Weak 1-convexity is thus equivalent to plurisubharmonicity. Some authors use dierent conventions for the number of positive eigenvalues in q-convexity. The reason why we introduce the number n q + 1 instead of q is mainly due to the following result: (2.1) Proposition. If v C 2 (M, R) is strongly (weakly) q-convex and if Y is a submanifold of M , then vY is strongly (weakly) q-convex. Proof. Let d = dim Y . For every x Y , there exists F Tx M with dim F = nq+1 such that Hv is (semi-) positive on F . Then G = F Tx Y has dimension (nq+1)(nd) = d q + 1, and H(vY ) is (semi-) positive on G Tx Y . Hence vY is strongly (weakly) q-convex at x. (2.2) Proposition. Let vj C 2 (M, R) be a weakly (strongly) qj -convex function, 1 j s, and C 2 (Rs , R) a convex function that is increasing (strictly increasing) in all variables. Then v = (v1 , . . . , vs ) is weakly (strongly) q-convex with q 1 = (qj 1). In particular v1 + + vs is weakly (strongly) q-convex. Proof. A simple computation gives (2.3) Hv =
j
(v1 , . . . , vs ) Hvj + tj
j,k
2 (v1 , . . . , vs ) d vj d vk , tj tk
and the second sum denes a semi-positive hermitian form. In every tangent space Tx M there exists a subspace Fj of codimension qj 1 on which Hvj is semi-positive (positive denite). Then F = Fj has codimension q 1 and Hv is semi-positive (positive denite) on F . The above result cannot be improved, as shown by the trivial example v1 (z) = 2|z1 |2 + |z2 |2 + |z3 |2 , v2 (z) = |z1 |2 2|z2 |2 + |z3 |2 on C3 ,
410
in which case q1 = q2 = 2 but v1 + v2 is only 3-convex. However, formula (2.3) implies the following result. (2.4) Proposition. Let vj C 2 (M, R), 1 j s, be such that every convex linear combination j vj , j 0, j = 1, is weakly (strongly) q-convex. If C 2 (Rs , R) is a convex function that is increasing (strictly increasing) in all variables, then (v1 , . . . , vs ) is weakly (strongly) q-convex. The invariance property of Prop. 2.1 immediately suggests the denition of q-convexity on complex spaces or analytic schemes: (2.5) Denition. Let (X, X ) be an analytic scheme. A function v on X is said to be strongly (resp. weakly) q-convex of class k on X if X can be covered by patches G : U A, A CN such that for each patch there exists a function v on with vA G = vU , which is strongly (resp. weakly) q-convex of class k . The notion of q-convexity on a patch U does not depend on the way U is embedded in CN , as shown by the following lemma. (2.6) Lemma. Let G : U A CN and G : U A CN be two patches of X. Let v be a strongly (weakly) q-convex function on and v = vA G. For every x U U there exists a strongly (weakly) q-convex function v on a neighborhood W of G (x) such that vA W G coincides with v on G1 (W ). Proof. The isomorphisms G G1 : A G(U U ) G (U U ) A G G1 : A G (U U ) G(U U ) A are restrictions of holomorphic maps H : W , H : W dened on neighborhoods W G(x), W G (x) ; we can shrink W so that H (W ) W . If we compose the automorphism (z, z ) (z, z H(z)) of W CN with the function v(z) + |z |2 we see that the function (z, z ) = v(z) + |z H(z)|2 is strongly (weakly) q-convex on W . Now, W can be embedded in W via the map z H (z ), z , so that the composite function v (z ) = H (z ), z = v H (z ) + |z H H (z )|2 is strongly (weakly) q-convex on W by Prop. 2.1. Since H G = G and H G = G on G1 (W ), we have v G = v G = v on G1 (W ) and the lemma follows. A consequence of this lemma is that Prop. 2.2 is still valid for an analytic scheme X (all the extensions vj near a given point x X can be obtained with respect to the same local embedding). (2.7) Denition. An analytic scheme (X, X ) is said to be strongly (resp. weakly) qconvex if X has a exhaustion function which is strongly (resp. weakly) q-convex outside an exceptional compact set K X. We say that X is strongly q-complete if
2. q-Convex Spaces
411
can be chosen so that K = . By convention, a compact scheme X is said to be strongly 0-complete, with exceptional compact set K = X. We consider the sublevel sets (2.8) Xc = {x X ; (x) < c}, c R.
If K Xc , we may select a convex increasing function such that = 0 on ] , c] and > 0 on ]c, +[. Then = 0 on Xc , so that is weakly q-convex everywhere in virtue of (2.3). In the weakly q-convex case, we may therefore always assume K = . The following properties are almost immediate consequences of the denition: (2.9) Theorem. a) A scheme X is strongly (weakly) q-convex if and only if the reduced space Xred is strongly (weakly) q-convex. b) If X is strongly (weakly) q-convex, every closed analytic subset Y of Xred is strongly (weakly) q-convex. c) If X is strongly (weakly) q-convex, every sublevel set Xc containing the exceptional compact set K is strongly (weakly) q-convex. d) If Uj is a weakly qj -convex open subset of X, 1 j s, the intersection U = U1 . . . Us is weakly q-convex with q 1 = (qj 1) ; U is strongly q-convex (resp. q-complete) as soon as one of the sets Uj is strongly qj -convex (resp. qj -complete). Proof. a) is clear, since Def. 2.5 does not involve the structure sheaf X . In cases b) and c), let be an exhaustion of the required type on X. Then Y and 1/(c ) are exhaustions on Y and Xc respectively (this is so only if Y is closed). Moreover, these functions are strongly (weakly) q-convex on Y (K Y ) and Xc K, thanks to Prop. 2.1 and 2.2. For property d), note that a sum = 1 + + s of exhaustion functions on the sets Uj is an exhaustion on U , choose the j s weakly qj -convex everywhere, and apply Prop. 2.2. (2.10) Corollary. Any nite intersection U = U1 . . . Us of weakly 1-convex open subsets is weakly 1-convex. The set U is strongly 1-convex (resp. 1-complete) as soon as one of the sets Uj is strongly 1-convex (resp. 1-complete). 2.B. Neighborhoods of q-complete subspaces We prove now a rather useful result asserting the existence of q-complete neighborhoods for q-complete subvarieties. The case q = 1 goes back to [Siu 1976], who used a much more complicated method. The rst step is an approximation-extension theorem for strongly q-convex functions. (2.11) Proposition. Let Y be an analytic set in a complex space X and a strongly q-convex function on Y . For every continuous function > 0 on Y , there exists a strongly q-convex function on a neighborhood V of Y such that Y < + .
412
Proof. Let Zk be a stratication of Y as given by Prop. II.5.6, i.e. Zk is an increasing sequence of analytic subsets of Y such that Y = Zk and Zk Zk1 is a smooth kdimensional manifold (possibly empty for some ks). We shall prove by induction on k the following statement: There exists a function k on X which is strongly q-convex along Y and on a closed neighborhood V k of Zk in X, such that kY < + . We rst observe that any smooth extension 1 of to X satises the requirements with Z1 = V1 = . Assume that Vk1 and k1 have been constructed. Then Zk Vk1 Zk Zk1 is contained in Zk,reg . The closed set Zk Vk1 has a locally nite covering (A ) in X by open coordinate patches A CN in which Zk is given by equations z = (z,k+1 , . . . , z,N ) = 0. Let be functions with compact support in A such that 0 1 and = 1 on Zk Vk1 . We set k (x) = k1 (x) +
(x) 3 log(1 + 4 |z |2 )
on X.
For > 0 small enough, we will have k1Y kY < + . Now, we check that k is still strongly q-convex along Y and near any x0 V k1 , and that k becomes strongly q-convex near any x0 Zk Vk1 . We may assume that x0 Supp for some , otherwise k coincides with k1 in a neighborhood of x0 . Select and a small neighborhood W of x0 such that a) if x0 Zk Vk1 , then (x0 ) > 0 and A W { > 0} ; b) if x0 A for some (there is only a nite set I of such s), then A W A and zA W has a holomorphic extension z to W ; c) if x0 V k1 , then k1A W has a strongly q-convex extension k1 to W ; d) if x0 Y V k1 , then k1Y W has a strongly q-convex extension k1 to W . Otherwise take an arbitrary smooth extension k1 of k1A W to W and let be an extension of A W to W . Then k = k1 +
3 log(1 + 4 |z |2 )
is an extension of kA W to W , resp. of kY W to W in case d). As the function log(1 + 4 |z |2 ) is plurisubharmonic and as its rst derivative z , dz (4 + |z |2 )1 is 2 bounded by O( ), we see that id d k id d k1 O( ).
Therefore, for small enough, k remains q-convex on W in cases c) and d). Since all functions z vanish along Zk W , we have id d k id d k1 +
I 1 id d |z |2 id d k1 + 1 id d |z |2
at every point of Zk W . Moreover id d k1 has at most (q 1)-negative eigenvalues on T Zk since Zk Y , whereas id d |z |2 is positive denite in the normal directions
2. q-Convex Spaces
413
to Zk in . In case a), we thus nd that k is strongly q-convex on W for small enough; we also observe that only nitely many conditions are required on each if we choose a locally nite covering of Supp by neighborhoods W as above. Therefore, for small enough, k is strongly q-convex on a neighborhood V k of Zk Vk1 . The function k and the set Vk = Vk Vk1 satisfy the requirements at order k. It is clear that we can choose the sequence k stationary on every compact subset of X ; the limit and the open set V = Vk fulll the proposition. The second step is the existence of almost plurisubharmonic functions having poles along a prescribed analytic set. By an almost plurisubharmonic function on a manifold, we mean a function that is locally equal to the sum of a plurisubharmonic function and of a smooth function, or equivalently, a function whose complex Hessian has bounded negative part. On a complex space, we require that our function can be locally extended as an almost plurisubharmonic function in the ambient space of an embedding. (2.12) Lemma. Let Y be an analytic subvariety in a complex space X. There is an almost plurisubharmonic function v on X such that v = on Y with logarithmic poles and v (X Y ). Proof. Since Y X is a coherent subsheaf, there is a locally nite covering of X by patches A isomorphic to analytic sets in balls B(0, r ) CN , such that Y admits a system of generators g = (g,j ) on a neighborhood of each set A . We set 1 v (z) = log |g (z)|2 2 on A , r |z z |2 v(z) = M(1,...,1) . . . , v (z), . . . where M is the regularized max function dened in I-3.37. As the generators (g,j ) can be expressed in terms of one another on a neighborhood of A A , we see that the quotient |g |/|g | remains bounded on this set. Therefore none of the values v (z) for 2 A z and z near A contributes to the value of v, since 1/(r |z z |2 ) tends to + on A . It follows that v is smooth on X Y ; as each v is almost plurisubharmonic on A , we also see that v is almost plurisubharmonic on X. (2.13) Theorem. Let X be a complex space and Y a strongly q-complete analytic subset. Then Y has a fundamental family of strongly q-complete neighborhoods V in X. Proof. By Prop. 2.11 applied to a strongly q-convex exhaustion of Y and = 1, there exists a strongly q-convex function on a neighborhood W0 of Y such that Y is an exhaustion. Let W1 be a neighborhood of Y such that W 1 W0 and such that W 1 is an exhaustion. We are going to show that every neighborhood W W1 of Y contains a strongly q-complete neighborhood V . If v is the function given by Lemma 2.12, we set where : R R is a smooth convex increasing function. If grows fast enough, we get v > 0 on W and the (q 1)-codimensional subspace on which id d is positive denite (in some ambient space) is also positive denite for id d v provided that be large enough to compensate the bounded negative part of id d v. Then v is strongly q-convex. Let be a smooth convex increasing function on ] , 0[ such that (t) = 0 for t < 3 and (t) = 1/t on ] 1, 0[. The open set V = {z W ; v(z) < 0} is a neighborhood of Y and = + v is a strongly q-convex exhaustion of V . v =v+ on W for such that A z,
414
2.C. Runge Open Subsets In order to extend the classical Runge theorem into an approximation result for sheaf cohomology groups, we need the concept of a q-Runge open subset. (2.14) Denition. An (resp. q-Runge complete) exhaustion function on strongly q-convex on X open subset U of a complex space X is said to be q-Runge in X if for every compact subset L U there exists a smooth X and a sublevel set Xb of such that L Xb U and is X b (resp. on the whole space X).
(2.15) Example. If X is strongly q-complete and if is a strongly q-convex exhaustion function of X, then every sublevel set Xc of is q-Runge complete in X : every compact set L Xc satises L Xb Xc for some b < c. More generally, if X is strongly q-convex and if is strongly q-convex on X K, every sublevel set Xc containing K is q-Runge in X. Later on, we shall need the following technical result. (2.16) Proposition. Let Y be an analytic subset of a complex space X. If U is a q-Runge complete open subset of Y and L a compact subset, there exist a neighborhood V of Y in X and a strongly q-convex exhaustion on V such that U = Y V and L Y Vb U for some sublevel set Vb of . Proof. Let be a strongly q-convex exhaustion on Y with L { < b} U as in Def. 2.14. Then L { 0 and Lemma 2.11 gives a strongly q-convex function on a neighborhood W0 of Y so that Y < + . The neighborhood V and the function = + v constructed in the proof of Th. 2.13 are the desired ones: we have Y = Y < + , thus L Y Vb { < b} U.
3. q-Convexity Properties in Top Degrees

It is obvious by denition that a n-dimensional complex manifold M is strongly qcomplete for q n + 1 (an arbitrary smooth function is then strongly q-convex !). If M is connected and non compact, [Greene and Wu 1975] have shown that M is strongly n-complete, i.e. there is a smooth exhaustion function on M such that id d has at least one positive eigenvalue everywhere. We need the following lemmas. (3.1) Lemma. Let be a strongly q-convex function on M and > 0 a given number. There exists a hermitian metric on M such that the eigenvalues 1 . . . n of the Hessian form id d with respect to satisfy 1 and q = . . . = n = 1. Proof. Let 0 be a xed hermitian metric, A0 (End T M ) the hermitian endomor0 0 phism associated to the hermitian form id d with respect to 0 , and 1 ... n the eigenvalues of A0 (or id d ). We can choose a function (M, R) such that 0 0 < (x) q (x) at each point x M . Select a positive function (R, R) such that (t) |t|/ for t 0, (t) t for t 0, (t) = t for t 1.
415
We let be the hermitian metric dened by the hermitian endomorphism A(x) = (x) [((x))1A0 (x)] where [ 1 A0 ] (End T M ) is dened as in Lemma VII-6.2. By construction, the 0 eigenvalues of A(x) are j (x) = (x) j (x)/(x) > 0 and we have
0 j (x) |j (x)|/ 0 j (x) j (x) 0 j (x) = j (x) 0 for j (x) 0 for j (x) for j q
0, 0, 0 then j (x)
(x) .
0 The eigenvalues of id d with respect to are j (x) = j (x)/j (x) and they have the required properties.
On a hermitian manifold (M, ), we consider the Laplace operator dened by (3.2) v = Trace (id d v) =
1 j,k n
jk (z)
2v zj z k
where ( jk ) is the conjugate of the inverse matrix of (jk ). Note that may dier from the usual Laplace-Beltrami operator if is not Khler. We say that v is strongly a -subharmonic if v > 0. This property implies clearly that v is strongly n-convex; however, as (v1 , . . . , vs ) =
j
(v1 , . . . , vs ) vj tj +
j,k
2 (v1 , . . . , vs ) d vj , d vk tj tk
subharmonicity has the advantage of being preserved by all convex increasing transformations. Conversely, if is strongly n-convex and chosen as in Lemma 3.1 with small enough, we get 1 (n 1) > 0, thus is strongly subharmonic for a suitable metric . (3.3) Lemma. Let U, W M be open sets such that for every connected component Us of U there is a connected component Wt(s) of W such that Wt(s) Us = and 0, with support Wt(s) U s = . Then there exists a function v (M, R), v contained in U W , such that v is strongly -subharmonic and > 0 on U . Proof. We rst prove that the result is true when U, W are small cylinders with the same radius and axis. Let a0 M be a given point and z1 , . . . , zn holomorphic coordinates centered at a0 . We set Re zj = x2j1 , Im zj = x2j , x = (x2 , . . . , x2n ) and = jk (x)dxj dxk . Let U be the cylinder |x1 | < r, |x | < r, and W the cylinder r < x1 < r + , |x | < r. There are constants c, C > 0 such that jk (x)j k c||2 and | jk (x)| C on U .
416
Let (R, R) be a nonnegative function equal to 0 on ] , r] [r + , +[ and strictly convex on ] r, r]. We take explicitly (x1 ) = (x1 + r) exp(1/(x1 + r)2 on ] r, r] and v(x) = (x1 ) exp 1/(|x |2 r 2 ) We have v
on U W,
v = 0 on M
(U W ).
(M, R), v > 0 on U , and a simple computation gives
v(x) = 11 (x) 4(x1 + r)5 2(x1 + r)3 v(x) +

j>1
1j (x) 1 + 2(x1 + r)2 (2xj )(r 2 |x |2 )2 jk (x) xj xk 4 8(r 2 |x |2 ) 2(r 2 |x |2 )2 jk (r 2 |x |2 )4 .
+
j,k>1
For r small, we get v(x) v(x) 2c(x1 + r)5 C1 (x1 + r)2 |x |(r 2 |x |2 )2 + (2c|x |2 C2 r 4 )(r 2 |x |2 )4 with constants C1 , C2 independent of r. The negative term is bounded by C3 (x1 + r)4 + c|x |2 (r 2 |x |2 )4 , hence v/v(x) c(x1 + r)5 + (c|x |2 C2 r 4 )(r 2 |x |2 )4 .
The last term is negative only when |x | < C4 r 2 , in which case it is bounded by C5 r 4 < c(x1 + r)5 . Hence v is strongly -subharmonic on U . Next, assume that U and W are connected. Then U W is connected. Fix a point a W U . If z0 U is given, we choose a path U W from z0 to a which is piecewise linear with respect to holomorphic coordinate patches. Then we can nd a nite sequence of cylinders (Uj , Wj ) of the type described above, 1 j N , whose axes are segments contained in , such that Uj Wj U W, W j Uj+1 and z0 U0 , a WN W U.
0, For each such pair, we have a function vj (M ) with support in U j W j , vj strongly -subharmonic and > 0 on Uj . By induction, we can nd constants Cj > 0 such that v0 + C1 v1 + + Cj vj is strongly -subharmonic on U0 . . . Uj and -subharmonic on M W j . Then wz0 = v0 + C1 v1 + . . . + CN vN 0 is -subharmonic on U and strongly -subharmonic > 0 on a neighborhood 0 of the given point z0 . Select a denumerable covering of U by such neighborhoods p and set v(z) = p wzp (z) where p is a sequence converging suciently fast to 0 so that v (M, R). Then v has the required properties.
In the general case, we nd for each pair (Us , Wt(s) ) a function vs with support in U s W t(s) , strongly -subharmonic and > 0 on Us . Any convergent series v = s vs yields a function with the desired properties.
417
(3.4) Lemma. Let X be a connected, locally connected and locally compact topological space. If U is a relatively compact open subset of X, we let U be the union of U with all compact connected components of X U . Then U is open and relatively compact in X, and X U has only nitely many connected components, all non compact. Proof. A rather easy exercise of general topology. Intuitively, U is obtained by lling the holes of U in X. (3.5) Theorem ([Greene-Wu 1975]). Every n-dimensional connected non compact complex manifold M has a strongly subharmonic exhaustion function with respect to any hermitian metric . In particular, M is strongly n-complete. Proof. Let (M, R) be an arbitrary exhaustion function. There exists a sequence . of connected smoothly bounded open sets M such that +1 and M = Let = be the relatively compact open set given by Lemma 3.4. Then +1 , M = and M has no compact connected component. We set U1 = 2 , U = +1 2 for 2.
Then U = +1 2 ; any connected component U,s of U has its boundary U,s 2 , otherwise U ,s would be open and closed in M 2 , hence U ,s would be a compact component of M 2 . Therefore U,s intersects +1 U+1 . If U+1,t(s) is a connected component of U+1 containing a point of U,s , then U+1,t(s) U,s = and U+1,t(s) U ,s = . Lemma 7 implies that there is a nonnegative function v (M, R) with support in U U+1 , which is strongly -subharmonic on U . An induction yields constants C such that = + C1 v1 + + C v is strongly -subharmonic on U0 . . . U , thus = + -subharmonic exhaustion function on M . C v is a strongly
By an induction on the dimension, the above result can be generalized to an arbitrary complex space (or analytic scheme), as was rst shown by T. Ohsawa. (3.6) Theorem ([Ohsawa 1984]). Let X be a complex space of maximal dimension n. a) X is always strongly (n + 1)-complete. b) If X has no compact irreducible component of dimension n, then X is strongly ncomplete. c) If X has only nitely many irreducible components of dimension n, then X is strongly n-convex. Proof. We prove a) and b) by induction on n = dim X. For n = 0, property b) is void and a) is obvious (any function can then be considered as strongly 1-convex). Assume that a) has been proved in dimension n 1. Let X be the union of Xsing and of the irreducible components of X of dimension at most n 1, and M = X X the n-dimensional part of Xreg . As dim X n 1, the induction hypothesis shows that X is strongly n-complete. By Th. 2.13, there exists a strongly n-convex exhaustion function
418
on a neighborhood V of X . Take a closed neighborhood V V and an arbitrary exhaustion on X that extends . Since every function on a n-dimensional manifold V is strongly (n + 1)-convex, we conclude that X is at worst (n + 1)-complete, as stated in a). In case b), the hypothesis means that the connected components Mj of M = X X have non compact closure M j in X. On the other hand, Lemma 3.1 shows that there exists a hermitian metric on M such that M V is strongly -subharmonic. Consider the open sets Uj, Mj provided by Lemma 3.7 below. By the arguments already used in Th. 3.5, we can nd a strongly -subharmonic exhaustion = + j, Cj, vj, on X, with vj, strongly -subharmonic on Uj, , Supp vj, Uj, Uj,+1 and Cj, large. Then is strongly n-convex on X. (3.7) Lemma. For each j, there exists a sequence of open sets Uj, Mj , N, such that a) Mj V
Uj, and (Uj, ) is locally nite in M j ;
b) for every connected component Uj,,s of Uj, there is a connected component Uj,+1,t(s) of Uj,+1 such that Uj,+1,t(s) Uj,,s = and Uj,+1,t(s) U j,,s = . Proof. By Lemma 3.4 applied to the space M j , there exists a sequence of relatively compact connected open sets j, in M j such that M j j, has no compact connected component, j, j,+1 and M j = j, . We dene a compact set Kj, Mj and an open set Wj, M j containing Kj, by Kj, = (j, j,1 ) V , Wj, = j,+1 j,2 .
By induction on , we construct an open set Uj, Wj, X Mj and a nite set Fj, Uj, j, . We let Fj,1 = . If these sets are already constructed for 1, the compact set Kj, Fj,1 is contained in the open set Wj, , thus contained in a nite union of connected components Wj,,s . We can write Kj, Fj,1 = Lj,,s where Lj,,s is contained in Wj,,s X Mj . The open set Wj,,s X is connected and non contained in j, Lj,,s , otherwise its closure W j,,s would have no boundary point j,+1 , thus would be open and compact in M j j,2 , contradiction. We select a point as (Wj,,s X ) (j, Lj,,s ) and a smoothly bounded connected open set Uj,,s Wj,,s X containing Lj,,s with as Uj,,s . Finally, we set Uj, = s Uj,,s and let Fj, be the set of all points as . By construction, we have Uj, Kj, Fj,1 , thus Uj, Kj, = Mj V , and Uj,,s as with as Fj, Uj,+1 . Property b) follows. Proof of Theorem 3.6 c) (end). Let Y X be the union of Xsing with all irreducible components of X that are non compact or of dimension < n. Then dim Y n 1, so Y (X, R) such is n-convex and Th. 2.13 implies that there is an exhaustion function that is strongly n-convex on a neighborhood V of Y . Then the complement K = X V is compact and is strongly n-convex on X K. (3.8) Proposition. Let M be a connected non compact n-dimensional complex manifold and U an open subset of M . Then U is n-Runge complete in M if and only if M U has no compact connected component.
4. Andreotti-Grauert Finiteness Theorems
419
On the other hand, assume that M U has no compact connected component and let L be a compact subset of U . Let be any hermitian metric on M and a strongly -subharmonic exhaustion function on M . Set b = 1 + supL and P = {x M As M y in U ; (x) b}.
Proof. First observe that a strongly n-convex function cannot have any local maximum, so it satises the maximum principle. If M U has a compact connected component T , then T has a compact neighborhood L in M such that L U . We have maxL = maxL for every strongly n-convex function, thus L Mb implies L Mb ; thus we cannot nd a sublevel set Mb such that L Mb U , and U is not n-Runge in M .
U has no compact connected component, all its components T contain a point W = {x X ; (x) > b + 1}.
For every point x P with x T , there exists a connected open set Vx M L containing x such that Vx y (M L is a neighborhood of M U and we can consider a tubular neighborhood of a path from x to y in M L). The compact set P can be covered by a nite number of open sets Vxj . Then Lemma 3.3 yields functions vj with support in V xj W which are strongly -subharmonic on Vxj . Let be a convex increasing function such that (t) = 0 on ] , b] and (t) > 0 on ]b, +[. Consider the function =+ Cj vj + . First, choose Cj large enough so that b on P . Then choose increasing fast enough so that is strongly -subharmonic on W . Then is a strongly n-convex exhaustion function on M , and as on M and = on L, we see that L {x M ; (x) < b} U. This proves that U is n-Runge complete in M .

4.A. Case of Vector Bundles over Manifolds The crucial point in the proof of the Andreotti-Grauert theorems is the following special case, which is easily obtained by the methods of chapter 8. (4.1) Proposition. Let M be a strongly q-complete manifold with q holomorphic vector bundle over M . Then: a) H k M, (E) = 0 for k q. 1, and E a
b) Let U be a q-Runge complete open subset of M . Every d -closed form h 0,q1 (U, E) can be approximated uniformly with all derivatives on every compact subset of U by a sequence of global d -closed forms h (M, E). 0,q1 Proof. We replace E by E = n T M E ; then we can work with forms of bidegree (n, k) instead of (0, k). Let be a strongly q-convex exhaustion function on M and
420
the metric given by Lemma 3.1. Select a function (M, R) which increases rapidly at innity so that the hermitian metric = e is complete on M . Denote by E the bundle E endowed with the hermitian metric obtained by multiplication of a xed metric of E by the weight exp( ) where (R, R) is a convex increasing function. We apply Th. VIII-4.5 for the bundle E over the complete hermitian manifold (M, ). Then ic(E ) = ic(E) + id d ( ) IdE
Nak
ic(E) + id d IdE .
The eigenvalues of id d with respect to are e j , so Lemma VII-7.2 and Prop. VI-8.3 yield [ic(E ), ] + T [ic(E), ] + T + [id d , ] IdE [ic(E), ] + T + e (1 + + k ) IdE q, we have if 1/q. when this curvature tensor acts on (n, k)-forms. For k 1 + + k 1 (q 1) > 0
We choose 0 increasing fast enough so that all the eigenvalues of the above curvature tensor are 1 when = 0 . Then for every g (M, E) with D g = 0 the equation n,k D f = g can be solved with an estimate |f |2 e dV |g|2 e dV,
where = 0 + 1 and where 1 is a convex increasing function chosen so that the integral of g converges. This gives a). In order to prove b), let h (U, E) be such n,q1 that D h = 0 and let L be an arbitrary compact subset of U . Thanks to Def. 2.14, we can choose such that there is a sublevel set Mb with L Mb U . Select b0 < b so that L Mb0 , and let (R, R) be a convex increasing function such that = 0 on ] , b0 [ and 1 on ]b, +[. Let (U ) be a cut-o function such that = 1 on Mb . We solve the equation D f = g for g = D (h) with the weight = 0 + and let tend to innity. As g has compact support in U Mb and 0 + on this set, we nd a solution f such that |f |2 e0 dV |f |2 e dV |g|2 e dV Ce ,
Mb 0
U Mb
thus f converges to 0 in L2 (Mb0 ) and h = h f (M, E) is a D -closed n,q1 form converging to h in L2 (Mb0 ). However, if we choose the minimal solution such that f = 0 as in Rem. VIII-4.6, we get f = g on M and in particular 0 f = 0 on Mb0 . G ardings inequality VI-3.3 applied to the elliptic operator 0 shows that f converges to 0 with all derivatives on L, hence h converges to h on L. Now, replace L by an exhaustion L of U by compact sets; some diagonal subsequence h converges to h in (U, E). n,q1
421
4.B. A Local Vanishing Result for Sheaves
For a given point x X, we choose a patch (A, /) of X containing x, where A is an analytic subset of CN and a sheaf of ideals with zero set A. Let iA : A be the inclusion. Then (iA ) is a coherent -module supported on A. In particular there is a neighborhood W0 of x and a surjective sheaf morphism
Let (X, X ) be an analytic scheme and a coherent sheaf of X -modules. We wish to extend Prop. 4.1 to the cohomology groups H k (X, ). The rst step is to show that the result holds on small open sets, and this is done by means of local resolutions of .
(iA )
on W0 ,
(u1 , . . . , up0 )
uj Gj
1 j p0
where G1 , . . . , Gp0 (A W0 ) are generators of (iA ) on W0 . If we repeat the procedure inductively for the kernel of the above surjective morphism, we get a homological free resolution of (iA ) : (4.3)
p1 p0 (iA ) 0
on Wl
(4.4) Lemma. Let A be a -distinguished patch of X and U a strongly q-convex open subset of A. Then H k (U, ) = 0 for k q. Proof. Theorem 2.13 shows that there exists a strongly q-convex open set V such 1 and that U = A V . Let us denote by l the kernel of pl pl1 for l 0 = ker p0 (iA ) . There are exact sequences 0 0
0 l
of arbitrary large length l, on neighborhoods Wl Wl1 . . . W0 . In particular, after replacing by W2N and A by A W2N , we may assume that (iA ) has a resolution of length 2N on . In this case, we shall say that A is a -distinguished patch of X.
For k
q, Prop. 4.1 a) gives H k (V, pl ) = 0, therefore we get H k (U, ) H k V, (iA ) H k+1 (V,
0
pl
p0 (iA ) 0,
l1
0,
2N.
) . . . H k+2N+1 (V,
2N
),
and the last group vanishes because topdim V
dimR V = 2N .
4.C. Topological Structure on Spaces of Sections and on Cohomology Groups Let V be a strongly 1-complete open set relatively to a A and let U = A V . By the proof of Lemma 4.4, we have H 1 (V, hence we get an exact sequence (4.5) 0
0 0
-distinguished patch
) H 2N+1 (V,
2N
) = 0,
(V ) p0 (V ) (U ) 0.
422
We are going to show that the Frchet space structure on p0 (V ) induces a natural e Frchet space structure on the groups of sections of over any open subset. We rst e note that 0 (V ) is closed in p0 (V ). Indeed, let f 0 (V ) be a sequence converging to a limit f p0 (V ) uniformly on compact subsets of V . For every x V , the germs (f )x converge to fx with respect to the topology dened by (1.4) on p0 . As 0 is x closed in p0 in view of Th. 1.5 b), we get fx 0 for all x V , thus f 0 (V ). x x (4.6) Proposition. The quotient topology on above.
(U ) is independent of the choices made
Proof. For a smaller set U = A V where V is a strongly 1-convex open subset of V , the restriction map p0 (V ) p0 (V ) is continuous, thus (U ) (U ) is continuous. If (V ) is a countable covering of V by such sets and U = A V , we get an injection of (U ) onto the closed subspace of the product (U ) consisting of families which are compatible in the intersections. Therefore, the Frchet topology induced by e the product coincides with the original topology of (U ). If we choose other generators H1 , . . . , Hq0 for (iA ) , the germs Hj,x can be expressed in terms of the Gj,x s, thus we get a commutative diagram G 0 p(V ) (U ) 0
H q (V ) (U ) 0
0
provided that U and V are small enough. If we express the generators Gj in terms of the Hj s, we nd a similar diagram with opposite vertical arrows and we conclude easily that the topology obtained in both cases is the same. Finally, it remains to show that the topology of (U ) is independent of the embedding A near a given point x X. We compare the given embedding with the Zariski embedding (A, x) of minimal dimension d. After shrinking A and changing coordinates, we may assume = CNd and that the embedding iA : A is the composite of i : A A and of the inclusion j : {0} . For V sucient small and U = AV , we have a surjective map G : p0 (V ) (U ) obtained by choosing generators G of j (i ) on a neighborhood of x in . Then we consider the open set V = V CNd A and the surjective map onto (U ) equal to the composite
p (V )
0
This map corresponds to a choice of generators Gj (iA ) (V ) equal to the functions G , considered as functions independent of the last variables zd+1 , . . . , zN . Since j is j open, it is obvious that the quotient topology on (U ) is the same for both embeddings. Now, there is a natural topology on the cohomology groups H k (X, ). In fact, let (U ) be a countable covering of X by strongly 1-complete open sets, such that each U is contained in a -distinguished patch. Since the intersections U0 ...k are again strongly 1-complete, the covering is acyclic by Lemma 4.4 and Lerays theorem shows that H k (X, ) is isomorphic to H q (, ). We consider the product topology on the spaces of Cech cochains C k (, ) = (U0 ...k ) and the quotient topology on H k (, ). It is 0 clear that H (, ) is a Frchet space; however the higher cohomology groups H k (, ) e need not be Hausdor because the coboundary groups may be non closed in the cocycle
p0 (V ) (U ).
j G
423
Finally, observe that when is the locally free sheaf associated to a holomorphic vector bundle E on a smooth manifold X, the topology on H k X, (E) is the same as the topology associated to the Frchet space structure on the Dolbeault complex e : by the analogue of formula (IV-6.11) we have a bijective continuous 0, (X, E), d map H k , (E) H k (X, E) 0, {(c0 ...k )} f (z) =
0 ,...,q
groups. The resulting topology on H k (X, ) is independent of the choice of the covering: in fact we only have to check that the bijective continuous map H k (, ) H k ( , ) is a topological isomorphism if is a renement of , and this follows from Rem. 1.10 applied to the morphism of Cech complexes C (, ) C ( , ).
c0 ...q (z) q d 0 . . . d q1
where ( ) is a partition of unity subordinate to . As in Rem. 1.10, the continuity of the inverse follows by the open mapping theorem applied to the surjective map Z k C (, (E))
0,k1 (X, E)
Z k
0, (X, E)
We shall need a few simple additional results. (4.7) Proposition. The following properties hold: a) For every x X, the map x dened by (1.4).
(X) x is continuous with respect to the topology of , the space of global sections (X) is closed
b) If is a coherent analytic subsheaf of in (X).
c) If U U are open in X, the restriction maps H k (U, ) H k (U , ) are continuous. e) Let be a morphism of coherent sheaves over X. Then the induced maps H k (X, ) H k (X, ) are continuous.
d) If U is relatively compact in U , the restriction operator
(U ) (U ) is compact.
Proof. a) Let V be a strongly 1-convex open neighborhood of x relatively to a -distinguished patch A . The map p0 (V ) p0 is continuous, and the same is x true for p0 x by 1. Therefore the composite p0 (V ) x and its factorization x (U ) x are continuous. b) is a consequence of the above property a) and of the fact that each stalk in x (cf. 1.5 b)).
x is closed
c) The restriction map (U ) (U ) is continuous, and the case of higher cohomology groups follows immediately. d) Assume rst that U = AV and U = AV , where A is a -distinguished patch and V V are strongly 1-convex open subsets of . The operator p0 (V ) p0 (V ) is compact by Montels theorem, thus (U ) (U ) is also compact. In the general
424
case, select a nite family of strongly 1-convex sets U U U such that (U ) covers U and U is contained in some distinguished patch. There is a commutative diagram
(U )
(U ) (U ) (U U)
(U )
e) It is enough to check that (U ) (U ) is continuous, and for this we may assume that U = A V where V is a small neighborhood of a given point x. Let G1 , . . . , Gp0 be generators of x , G , . . . , G 0 their images in . Complete these elements in order x 1 p to obtain a system of generators (G , . . . , G 0 ) of . For V small enough, the map 1 q x (U ) (U ) is induced by the inclusion p0 (V ) p0 (V ) {0} q0 (V ), hence continuous. 4.D. Cartan-Serre Finiteness Theorem The above results enable us to prove a niteness theorem for cohomology groups over compact analytic schemes.
where the right vertical arrow is a monomorphism and where the rst arrow in the bottom line is compact. Thus (U ) (U ) is compact.
(4.8) Theorem (Cartan-Serre). Let be a coherent analytic sheaf over an analytic scheme (X, X ). If X is compact, all cohomology groups H k (X, ) are nite dimensional (and Hausdor ).
Proof. There exist nitely many strongly 1-complete open sets U U such that each U is contained in some -distinguished patch and such that U = X. By Prop. 4.7 d), the restriction map on Cech cochains
C (, ) C ( , ) denes a compact morphism of complexes of Frchet spaces. As the coverings e and = (U ) are acyclic by 4.4, the induced map H k (, ) H k ( , )
= (U )
is an isomorphism, both spaces being isomorphic to H k (X, ). We conclude by Schwartz theorem 1.9. 4.E. Local Approximation Theorem We show that a local analogue of the approximation result 4.1 b) holds for a sheaf over an analytic scheme (X, X ).
(4.9) Lemma. Let A be a -distinguished patch of X, and U U A open subsets such that U is q-Runge complete in U . Then the restriction map H q1 (U, ) H q1 (U , ) has a dense range.
425
Proof. Let L be an arbitrary compact subset of U . Proposition 2.16 applied with Y = U embedded in some neighborhood in shows that there is a neighborhood V of U in such that A V = U and a strongly q-convex function on V such that L Ub U for some Ub = A Vb . The proof of Lemma 4.4 gives H q (V, 0 ) = H q (Vb , 0 ) = 0 and the cohomology exact sequences of 0 0 p0 i 0 over V and Vb yield a A commutative diagram of continuous maps H q1 V, p0 H q1 V, i = H q1 U, ) A
H q1 Vb , p0 H q1 Vb , i = H q1 Ub , ) A
where the horizontal arrows are surjective. Since Vb is q-Runge complete in V , the left vertical arrow has a dense range by Prop. 4.1 b). As U is the union of an increasing sequence of sets Ub , we only have to show that the range remains dense in the inverse limit H q1 (U , ). For that, we apply Property 1.11 d) on a suitable covering of U . Let be a countable basis of the topology of U , consisting of strongly 1-convex open subsets contained in -distinguished patches. We let (resp. ) be the subfamily of W such that W U (resp. W Ub ). Then , , are acyclic coverings of U, U , Ub and each restriction map C (, ) C ( , ) is surjective. Property 1.11 d) can thus be applied and the lemma follows. 4.F. Statement and Proof of the Andreotti-Grauert Theorem (4.10) Theorem ([Andreotti-Grauert 1962]). Let strongly q-convex analytic scheme (X, X ). Then
be a coherent analytic sheaf over a

q.
a) H k (X, ) is Hausdor and nite dimensional for k
Moreover, let U be a q-Runge open subset of X, q
1. Then q;
b) the restriction map H k (X, ) H k (U, ) is an isomorphism for k c) the restriction map H q1 (X, ) H q1 (U, ) has a dense range.
The compact case q = 0 of 4.10 a) is precisely the Cartan-Serre niteness theorem. For q 1, the special case when X is strongly q-complete and U = yields the following very important consequence. (4.11) Corollary. If X is strongly q-complete, then H k (X, ) = 0 for k q.
Assume that q 1 and let be a smooth exhaustion on X that is strongly q-convex on X K. We rst consider sublevel sets Xd Xc K, d > c, and verify assertions 4.10 b), c) for all restriction maps H k (Xd , ) H k (Xc , ), k q 1.
426
The basic idea, already contained in [Andreotti-Grauert 1962], is to deform Xc into Xd through a sequence of strongly q-convex open sets (Gj ) such that Gj+1 is obtained from Gj by making a small bump. (4.12) Lemma. There exist a sequence of strongly q-convex open sets G0 . . . Gs and a sequence of strongly q-complete open sets U0 , . . . , Us1 in X such that a) G0 = Xc , Gs = Xd , Gj+1 = Gj Uj for 0 j s 1;
b) Gj = {x X ; j (x) < cj } where j is an exhaustion function on X that is strongly q-convex on X K ; c) Uj is contained in a
-distinguished patch Aj j of X ;
Proof. There exists a nite covering of the compact set X d Xc by -distinguished patches Aj j , 0 j < s, where j CNj is a euclidean ball and K Aj = . Let j (X) be a family of functions such that Supp j Aj , j 0, j 1 and j = 1 on a neighborhood of X d Xc . We can nd 0 > 0 so small that j = is still strongly q-convex on X s = on X d Xc , thus K for 0 k
0 k<j
d) Gj Uj is strongly q-complete and q-Runge complete in Uj .
s and 0 j
0 . We have 0 = and s
Gj = {x X ; j (x) < c},
is an increasing sequence of strongly q-convex open sets such that G0 = Xc , Gs = Xc+ . Moreover, as j+1 j = j has support in Aj , we have Gj+1 = Gj Uj where Uj = Gj+1 Aj .
It follows that conditions a), b), c) are satised with c + instead of d. Finally, the functions
2 j = 1/(c j+1 ) + 1/(rj |z zj |2 ), 2 j = 1/(c j ) + 1/(rj |z zj |2 )
are strongly q-convex exhaustions on Uj and Gj Uj = Gj Aj . Let L be an arbitrary compact subset of Gj Uj and a = supL j < c. Select b ]a, c[ and set j, = j + j on Uj , > 0.
Then j, is an exhaustion of Uj . As j is bounded below, we have L {j, < b} {j < c} Uj = Gj Uj for small enough. Moreover (1 )j + j+1 = k j
0 k<j
427
is strongly q-convex for all [0, 1] and 0 small enough, so Prop. 2.4 implies that j, is strongly q-convex. By denition, Gj Uj is thus q-Runge complete in Uj , and Lemma 4.12 is proved with Xc+ instead of Xd . In order to achieve the proof, we consider an increasing sequence c = c0 < c1 < . . . < cN = d with ck+1 ck 0 and perform the same construction for each pair Xck Xck+1 , with c replaced by ck and = ck+1 ck . (4.13) Proposition. For every sublevel set Xc K, the group H k (Xc , ) is Hausdor and nite dimensional when k q. Moreover, for d > c, the restriction map H k (Xd , ) H k (Xc , ) is an isomorphism when k q and has a dense range when k = q 1. H k (Gj+1 , ) H k (Gj , ).
Proof. Thanks to Lemma 4.12, we are led to consider the restriction maps (4.14)
Let us apply the Mayer-Vietoris exact sequence IV-3.11 to Gj+1 = Gj Uj . For k q we have H k (Uj , ) = H k (Gj Uj , ) = 0 by Lemma 4.4. Hence we get an exact sequence H q1 (Gj+1 , ) H q1 (Gj , ) H q1 (Uj , ) H q1 (Gj Uj , ) H k (Gj+1 , ) H k (Gj , ) 0 , k q.
In this sequence, all the arrows are induced by restriction maps, so they dene continuous linear operators. We already infer that the map (4.14) is bijective for k > q and surjective for k = q. There exist a -acyclic covering = (V ) of Xd and a nite family = (V1 , . . . , Vp ) of open sets such that Vj Vj and Vj X c . Let be a locally nite -acyclic covering of Xc which renes Xc = (Vj Xc ). The renement map C (, ) C ( Xc , ) C (, ) is compact because the rst arrow is, and it induces a surjective map H k (Xd , ) H k (Xc , ) for k q.
By Schwartz theorem 1.9, we conclude that H k (Xc , ) is Hausdor and nite dimensional for k q. This is equally true for H q (Gj , ) because Gj is also a global sublevel set {x X ; j (x) < cj } containing K. Now, the Mayer-Vietoris exact sequence implies that the composite H q1 (Uj , ) H q1 (Gj Uj , ) H q (Gj+1 , )
is equal to zero. However, the rst arrow has a dense range by Lemma 4.9. As the target space is Hausdor, the second arrow must be zero; we obtain therefore the injectivity of H q (Gj+1 , ) H q (Gj , ) and an exact sequence H q1 (Gj+1 , ) H q1 (Gj , )H q1 (Uj , ) H q1 (Gj Uj , ) 0 gu uGj Uj gGj Uj .
428
The argument used in Rem. 1.10 shows that the surjective arrow is open. Let g H q1 (Gj , ) be given. By Lemma 4.9, we can approximate gGj Uj by a sequence uGj Uj , u H q1 (Uj , ). Then w = uGj Uj gGj Uj tends to zero. As the second map in the exact sequence is open, we can nd a sequence
g u H q1 (Gj , ) H q1 (Uj , ) converging to zero which is mapped on w . Then (g g ) (u u ) is mapped on zero, and there exists a sequence f H q1 (Gj+1 , ) which coincides with g g on Gj and with u u on Uj . In particular fGj converges to g and we have shown that
H q1 (Gj+1 , ) H q1 (Gj , )
has a dense range. Proof of Andreotti-Grauerts Theorem 4.10.. Let be a countable basis of the topology of X consisting of strongly 1-convex open sets W contained in -distinguished patches of X. Let L U be an arbitrary compact subset. Select a smooth exhaustion function on X such that is strongly q-convex on X X b and L Xb U for some sublevel set Xb of ; choose c > b such that Xc U . For every d R, we denote by d the collection of sets W such that W Xd . Then d is a -acyclic covering of Xd . We consider the sequence of Cech complexes
E = C (c+ , ),
together with the surjective projection maps E+1 E , and their inverse limit k k E = C (, ). Then we have H (E ) = H (X, ) and H k (E ) = H k (Xc+ , ). Propositions 1.11 (a,b,c) and 4.13 imply that H k (X, ) H k (Xc , ) is bijective for k q and has a dense range for k = q 1. It already follows that H k (X, ) is Hausdor for k q. Now, take an increasing sequence of open sets Xc equal to sublevel sets of a sequence of exhaustions , such that U = Xc . Then all groups H k (Xc , ) are in bijection with H k (X, ) for k q, and the image of H q1 (Xc+1 , ) in H q1 (Xc , ) is dense because it contains the image of H q1 (X, ). Proposition 1.11 (a,b,c) again shows that H k (U, ) H k (Xc0 , ) is bijective for k q, and d) shows that H q1 (X, ) H q1 (U, ) has a dense range. The theorem follows.
A combination of Andreotti-Grauerts theorem with Th. 3.6 yields the following important consequence. (4.15) Corollary. Let dim X n.
be a
coherent sheaf over an analytic scheme (X, X ) with n+ 1;
a) We have H k (X, ) = 0 for all k
b) If X has no n-dimensional compact irreducible component, then we have H n (X, ) = 0. c) If X has only nitely many n-dimensional compact irreducible components, then H n (X, ) is nite dimensional. The special case of 4.15 b) when X is smooth and locally free has been rst proved by [Malgrange 1955], and the general case is due to [Siu 1969]. Another consequence is
5. Grauerts Direct Image Theorem
429
the following approximation theorem for coherent sheaves over manifolds, which results from Prop. 3.8. (4.16) Proposition. Let be a coherent sheaf over a non compact connected complex manifold M with dim M = n. Let U M be an open subset such that the complement M U has no compact connected component. Then the restriction map H n1 (M, ) H n1 (U, ) has a dense range.

The goal of this section is to prove the following fundamental result on direct images of coherent analytic sheaves, due to [Grauert 1960]. (5.1) Direct image theorem. Let X, Y be complex analytic schemes and let F : X Y be a proper analytic morphism. If is a coherent X -module, the direct images Rq F are coherent Y -modules. We give below a beautiful proof due to [Kiehl-Verdier 1971], which is much simpler than Grauerts original proof; this proof rests on rather deep properties of nuclear modules over nuclear Frchet algebras. We rst introduce the basic concept of topologe ical tensor product. Our presentation owes much to the seminar lectures by [DouadyVerdier 1973]. 5.A. Topological Tensor Products and Nuclear Spaces The algebra of holomorphic functions on a product space X Y is a completion We are going to describe the construction and the basic properties of the required topological tensor products .
(X) (Y ) of the algebraic tensor product (X) (Y ).
Let E, F be (real or complex) vector spaces equipped with semi-norms p and q, respectively. Then E F can be equipped with any one of the two natural semi-norms p q, p q dened by p q(t) = inf p q(t) = p(xj ) q(yj ) ; t =
1 j N 1 j N
xj yj , xj E , yj F
sup
||||p 1, ||||q 1
(t) ,
E , F ; p(x) and |(y)| p(xj ) q(yj ). q(y) for
the inequalities in the last line mean that , satisfy |(x)| all x E, y F . Then clearly p q p q, for p q xj y j p q(xj yj )
Given x E, y F , the Hahn-Banach theorem implies that there exist , such that ||||p = ||||q = 1 with (x) = p(x) and (y) = q(y), hence p q(x y) p(x) q(y). On the other hand p q(x y) p(x) q(y), thus p q(x y) = p q(x y) = p(x) q(y).
430
(5.2) Denition. Let E, F be locally convex topological vector spaces. The topological tensor product E F (resp. E F ) is the Hausdor completion of E F , equipped with the family of semi-norms p q (resp. p q) associated to fundamental families of semi-norms on E and F . Since we may also write p q(t) = inf |j | ; t = j xj yj , p(xj ) 1 , q(yj ) 1
where the j s are scalars, we see that the closed unit ball B(p q) in E F is the closed convex hull of B(p) B(q). From this, we easily infer that the topological dual space (E F ) is isomorphic to the space of continuous bilinear forms on E F . Another simple consequence of this interpretation of B(p q) is example a) below. (5.3) Examples. a) For all discrete spaces I, J, there is an isometry 1 (I) 1 (J) 1 (I J). b) For Banach spaces (E, p), (F, q), the closed unit ball in E F is dual to the unit ball B(p q ) of E F through the natural pairing extending the algebraic pairing of E F and E F . If c0 (I) denotes the space of bounded sequences on I converging to zero at innity, we have c0 (I) = 1 (I), hence by duality c0 (I) c0 (J) is isometric to c0 (I J). c) If X, Y are compact topological spaces and if C(X), C(Y ) are their algebras of continuous functions with the sup norm, then C(X) C(Y ) C(X Y ). Indeed, C(X) is the space of nite Borel measures equipped with the mass norm. Thus for f C(X) C(Y ), the -seminorm is given by ||f || = sup
|||| 1, |||| 1
(f ) = sup |f | ;
XY
the last equality is obtained by taking Dirac measures x , y for , (the inequality is obvious). Now C(X) C(Y ) is dense in C(X Y ) by the Stone-Weierstrass theorem, hence its completion is C(X Y ), as desired. Let f : E1 E2 and g : F1 F2 be continuous morphisms. For all semi-norms p2 , q2 on E2 , F2 , there exist semi-norms p1 , q1 on E1 , F1 and constants ||f || = ||f ||p1 ,p2 , ||g|| = ||g||q1,q2 such that p2 f ||f || p1 and q2 g ||g|| q1. Then we nd (p2 q2 ) (f g) ||f || ||g|| p1 q1
and a similar formula with pj qj . It follows that there are well dened continuous maps
431
(5.4 ) (5.4 )
f g : E1 F1 E2 F2 , f g : E1 F1 E2 F2 .
Another simple fact is that preserves open morphisms: (5.5) Proposition. If f : E1 E2 and g : F1 F2 are epimorphisms, then f g : E1 F1 E2 F2 is an epimorphism. Proof. Recall that when E is locally convex complete and F Hausdor, a morphism u : E F is open if and only if u(V ) is a neighborhood of 0 for every neighborhood of 0 (this can be checked essentially by the same proof as 1.8 b)). Here, for any seminorms p, q on E1 , F1 the closure of f g B(p q) contains the closed convex hull of f B(p) g B(q) in which f B(p) and g B(q) are neighborhoods of 0, so it is a neighborhood of 0 in E F . If E1 E2 is a closed subspace, every continuous semi-norm p1 on E1 is the restriction of a continuous semi-norm on E2 , and every linear form 1 E1 such that ||1 ||p1 1 can be extended to a linear form 2 E2 such that ||2 ||p2 = ||1 ||p1 (Hahn-Banach theorem); similar properties hold for a closed subspace F1 F2 . We infer that (p2 q2 )E1 F1 = p1 q1 , thus E1 F1 is a closed subspace of E2 F2 . In other words: (5.6) Proposition. If f : E1 E2 and g : F1 F2 are monomorphisms, then f g : E1 F1 E2 F2 is a monomorphism. Unfortunately, 5.5 fails for and 5.6 fails for , even with Frchet or Banach spaces. e It follows that neither nor are exact functors in the category of Frchet spaces. e In order to circumvent this diculty, it is necessary to work in a suitable subcategory. (5.7) Denition. A morphism f : E F of complete locally convex spaces is said to be nuclear if f can be written as f (x) = j j (x) yj
where (j ) is a sequence of scalars with |j | < +, j E an equicontinuous sequence of linear forms and yj F a bounded sequence. When E and F are Banach spaces, the space of nuclear morphisms is isomorphic to E F and the nuclear norm ||f || is dened to be the norm in this space, namely (5.8) ||f || = inf |j | ; f = j j yj , ||j || 1, ||yj || 1 .
For general spaces E, F , the equicontinuity of (j ) means that there is a semi-norm p on E and a constant C such that |j (x)| C p(x) for all j. Then the denition shows that f : E F is nuclear if and only if f can be factorized as E E1 F1 F where E1 F1 is a nuclear morphism of Banach spaces: indeed we need only take E1 be equal to the Hausdor completion Ep of (E, p) and let F1 be the subspace of F generated by
432
the closed balanced convex hull of {yj } (= unit ball in F1 ) ; moreover, if u : S E and v : F T are continuous, the nuclearity of f implies the nuclearity of v f u ; its nuclear decomposition is then v f u = j (j u) v(yj ). (5.9) Remark. Every nuclear morphism is compact: indeed, we may assume in Def. 5.7 that (yj ) converges to 0 and |j | 1, otherwise we replace yj by j yj with j converging to zero such that |j /j | 1 ; then, if U F is a neighborhood of 0 such that |j (U )| 1 for all j, the image f (U ) is contained in the closed convex hull of the compact set {yj } {0}, which is compact. (5.10) Proposition. If E, F, G are Banach spaces and if f : E F is nuclear, there is a continuous morphism f IdG : E G F G extending f IdG , such that ||f IdG || Proof. If f = ||f || .
j j yj as in (5.8), then for any t E G we have (f IdG )(t) = j j IdG (t) yj
where (j IdG )(t) G has norm ||(j IdG )(t)|| = Therefore ||f IdG (t)|| yields Proposition 5.10 follows. If E is a Frchet space and (pj ) an increasing sequence of semi-norms on E dening e the topology of E, we have E = lim Epj , where Epj is the Hausdor completion of (E, pj ) and Epj+1 Epj the canonical morphism. Here Epj is a Banach space for the induced norm pj . (5.11) Denition. A Frchet space E is said to be nuclear if the topology of E can be e dened by an increasing sequence of semi-norms pj such that each canonical morphism Epj+1 Epj of Banach spaces is nuclear. If E, F are arbitrary locally convex spaces, we always have a continuous morphism E F E F , because p q p q. If E, say, is nuclear, this morphism yields in fact an isomorphism E F E F : indeed, by Prop. 5.10, we have pj q Cj pj+1 q where Cj is the nuclear norm of Epj+1 Epj . Hence, when E sup
G , |||| 1
j IdG (t)
= sup j (t)
||t|| .
|j | ||t|| , and the inmum over all decompositions of f ||f IdG (t)|| ||f || ||t|| .
433
or F is nuclear, we will identify E F and E F and omit or in the notation E F. (5.12) Example. Let D = (D) with the semi-norm D(0, Rj ) be a polydisk in Cn . For any t ]0, 1[, we equip pt (f ) = sup |f |.
tD
The completion of (D), pt is the Banach space Et of holomorphic functions on tD which are continuous up to the boundary. We claim that for t < t < 1 the restriction map t,t : Et Et is nuclear. In fact, for f (D), we have f (z) = a z where a = a (f ) satises the Cauchy inequalities |a (f )| pt (f )/(t R) for all Nn . The formula f = a (f ) e with e (z) = z shows that ||t,t || ||a ||pt ||e ||pt (t R) (tR) = (1 t/t )n < +.
e We infer that (D) is a nuclear Frchet space. It is also in a natural way a fully nuclear Frchet algebra (see Def. 5.39 below). e (5.13) Proposition. Let E be a nuclear space. A morphism f : E F is nuclear if and only if f admits a factorization E M F through a Banach space M . Proof. By denition, a nuclear map f : E F always has a factorization through a Banach space (even if E is not nuclear). Conversely, if E is nuclear, any continuous linear map E M into a Banach space M is continuous for some semi-norm pj on E, so this map has a factorization E Epj+1 Epj M in which the second arrow is nuclear. Hence any map E M F is nuclear. (5.14) Proposition. a) If E, F are nuclear spaces, then E F is nuclear. b) Any closed subspace or quotient space of a nuclear space is nuclear. c) Any countable product of nuclear spaces is nuclear. d) Any countable inverse limit of nuclear spaces is nuclear. Proof. a) If f : E1 F1 and g : E2 F2 are nuclear morphisms of Banach spaces, it is easy to check that f g and f g are nuclear with ||f ? g|| ||f || ||g|| in both cases. Property a) follows by applying this to the canonical morphisms Epj+1 Epj and Fqj+1 Fqj . c) Let Ek , k N, be nuclear spaces and F = Ek . If (pk ) is an increasing family of j semi-norms on Ek as in Def. 5.11, then the topology of F is dened by the family of semi-norms qj (x) = max pk (xk ), x = (xk ) F. j
0 k j
434
0 k j
Then Fqj =
Ek,pk and j Ek,pk

0 k j
Fqj+1 Fqj = is easily seen to be nuclear.
j+1
Ek,pk Ej+1,pj+1 {0}

j j+1
Fpj+1 Fpj is the restriction of Epj+1 Epj and we have E/F pj Epj /Fpj . It is not true in general that the restriction or quotient of a nuclear morphism is nuclear, but this is true for a binuclear = (nuclear nuclear) morphism, as shown by Lemma 5.15 b) below. Hence Fp2j+2 Ep2j and E/F p2j+2 E/F p2j are nuclear, so (p2j ) is a fundamental family of semi-norms on F or E/F , as required in Def. 5.11. d) follows immediately from b) and c), since lim Ek is a closed subspace of (5.15) Lemma. Let E, F , G be Banach spaces. a) If f : E F is nuclear, then f can be factorized through a Hilbert space H as a morphism E H F . b) Let g : F G be another nuclear morphism. If Im(g f ) is contained in a closed subspace T of G, then g f : E T is nuclear. If ker(g f ) contains a closed subspace S of E, the induced map (g f ) : E/S G is nuclear. Proof. a) Write f = jI j yj E F with ||j || ||yj || < +. Without loss of generality, we may suppose ||j || = ||yj ||. Then f is the composition E 2 (I) F, x j (x) , (j ) j yj . Ek .
b) If F E is closed, then Fpj can be identied to a closed subspace of Epj , the map
b) Decompose g into g = v u as in a) and write g f as the composition E F H G where H is a Hilbert space. If Im(g f ) T and if T G is closed, then H1 = v 1 (T ) is a closed subspace of H containing Im(u f ). Therefore g f : E T is the composition
1 E F H H1 T
pr
vH
where f is nuclear and g f : E T is nuclear. Similar proof for (g f ) : E/S G by using decompositions f = v u : E H F and
1 (g f ) : E/S H/H1 H1 F G
vH
where H1 = u(S) satises H1 ker(g v) H. (5.16) Corollary. Let E be a nuclear space and let E F be a nuclear morphism.
435
a) If f (E) is contained in a closed subspace T of F , then the morphism f1 : E T induced by f is nuclear. b) If ker f contains a closed subspace S of E, then f : E/S F is nuclear. Proof. Let E M F be a factorization of f through a Banach space M . In case a), resp. b), M1 = v 1 (T ) is a closed subspace of M , resp. M/u(S) is a Banach space, and we have factorizations
1 1 f1 : E M1 T,
f : E/S M/u(S) F
where u1 , u are induced by u and v1 , v by v. Hence f1 and f are nuclear. (5.17) Proposition. Let 0 E1 E2 E3 0 be an exact sequence of Frchet e spaces and let F be a Frchet space. If E2 or F is nuclear, there is an exact sequence e 0 E1 F E2 F E3 F 0. Proof. If E2 is nuclear, then so are E1 and E3 by Prop. 5.14 b). Hence E1 F E2 F is a monomorphism and E2 F E3 F an epimorphism by Prop. 5.6 and 5.5. It only remains to show that Im E1 F E2 F = ker E2 F E3 F and for this, we need only show that the left hand side is dense in the right hand side (we already know it is closed). Let (E2 F ) be a linear form, viewed as a continuous bilinear form on E2 F . If vanishes on the image of E1 F , then induces a continuous bilinear form on E3 F by passing to the quotient. Hence must vanish on the kernel of E2 F E3 F , and our density statement follows by the Hahn-Banach theorem. 5.B. K nneth Formula for Coherent Sheaves u As an application of the above general concepts, we now show how topological tensor products can be used to compute holomorphic functions and cohomology of coherent sheaves on product spaces. be a coherent analytic sheaf on a complex analytic scheme (5.18) Proposition. Let (X, X ). Then (X) is a nuclear space. Proof. Let A CN be an open patch of X such that the image sheaf (iA ) has a resolution p1 p0 (iA ) A 0 and let D be a polydisk. As D is Stein, we get an exact sequence (5.19) Hence (A D) is a quotient of the nuclear space p0 (D) and so (A D) is nuclear by (5.14 b). Let (U ) be a countable covering of X by open sets of the form A D. Then (X) is a closed subspace of (U ), thus (X) is nuclear by (5.14 b,c).
A
on
p (D) p (D)
1 0
(A D) 0.
436
(5.20) Proposition. Let , be coherent sheaves on complex analytic schemes X, Y respectively. Then there is a canonical isomorphism (X Y ) (X) (Y ).
Proof. We show the proposition in several steps of increasing generality. a) X = D Cn , Y = D Cp are polydisks, = X , = Y .
Let pt (f ) = suptD |f |, p (f ) = suptD |f | and qt (f ) = supt(DD ) |f | be the semit norms dening the topology of (D), (D ) and (D D ), respectively. Then Ept is a closed subspace of the space C(tD) of continuous functions on tD with the sup norm, and we have pt p = qt by example (5.3 c). Now, (D) (D ) is dense in (D D ), t hence its completion with respect to the family (qt ) is (D) (D ) = (D D ). b) X is embedded in a polydisk D Cn , X = A D D, p1 p0 i is the cokernel of a morphism D D , Y = D Cp is a polydisk and = Y . By taking the external tensor product with (5.21)
i
Y , we get an exact sequence

Y 0. (X) (Y ) 0
p DY
1 1
p0 DY i
Then we nd a commutative diagram
p (D) (Y ) p (D) (Y )
0
p (DY )
1
in which the rst line is exact as the image of (5.19) by the exact functor (Y ), and the second line is exact because the exact sequence of sheaves (5.21) gives an exact sequence of spaces of sections on the Stein space D Y ; note that i Y (D Y ) = Y (X Y ). As the rst two vertical arrows are isomorphisms by a), the third one is also an isomorphism. c) X, are as in b), j Y is embedded in a polydisk D Cp , Y = A D D and j is the cokernel of q1 q0 . D D Taking the external tensor product with q1 D (X) q1 (D ) , we get an exact sequence j 0 q0 D
p (DY )
0
Y (X Y ) 0
and with the same arguments as above we obtain a commutative diagram (X) q0 (D ) (X) (Y ) 0
q1 (X D ) D
q0 (X D ) D
(X Y ) 0.
437
d) X,
are as in b),c) and Y ,
are arbitrary.
Then Y can be covered by open sets U = A D embedded in polydisks D , on which the image of admits a two-step resolution. We have (X U ) (X) (U ) by c), and the same is true over the intersections X U because U = U U can be embedded by the cross product embedding j j : U D D . We have an exact sequence 0 (Y ) (U ) (U )
,
where the last arrow is (c ) (c c ), and a commutative diagram with exact lines 0 0 (X) (Y ) (X) (U ) (X U ) (X) (U ) (X U ).
(X Y )
Therefore the rst vertical arrow is an isomorphism. e) X, , Y, are arbitrary. , and
This case is treated exactly in the same way as d) by reversing the roles of by using d) to get the isomorphism in the last two vertical arrows.
be coherent sheaves over complex analytic schemes X, Y (5.22) Corollary. Let , and let : X Y X be the projection. Suppose that H (Y, ) is Hausdor. a) If X is Stein, then H q (X Y, ) (X) H q (Y, ). b) In general, for every open set U X, Rq ( ) (U ) = c) If H q (Y, ) is nite dimensional, then R q ( ) = H q (Y, ). (U ) H q (Y, ).
Proof. a) Let = (V ) be a countable Stein covering of Y . By the Leray theorem, H (Y, ) is equal to the cohomology of the Cech complex C (, ). Similarly X = (X V ) is a Stein covering of X Y and we have H q (X Y, ) = H q C (X , ) .
However, Prop. 5.20 shows that C (X, ) = (X) C (, ). Our assumption that C (, ) has Hausdor cohomology implies that the cocycle and coboundary groups are (nuclear) Frchet spaces, and that each cohomology group can be computed by means e of short exact sequences in this category. By Prop. 5.17, we thus get the desired equality H q C (X , ) = (X) H q C (, ) .
438
b) The presheaf U (U ) H q (Y, ) is in fact a sheaf, because the tensor product with the nuclear space H q (Y, ) preserves the exactness of all sequences 0 (U ) (U ) (U )
associated to arbitrary coverings (U ) of U . Property b) thus follows from a) and from the fact that Rq ( ) is the sheaf associated to the presheaf U H q (U Y, ). c) is an immediate consequence of b), since the nite dimensionality of H q (Y, ) implies that this space is Hausdor. (5.23) K nneth formula. Let , be coherent sheaves over complex analytic schemes u X, Y and suppose that the cohomology spaces H (X, ) and H (Y, ) are Hausdor. Then there is an isomorphism H p (X,
p+q=k
) H q (Y, ) H k (X Y, p q p q .
Proof. Consider the Leray spectral sequence associated to the coherent sheaf = and to the projection : X Y X. By Cor. 5.22 b) and a use of Cech cohomology, we nd p,q E2 = H p (X, Rq ) = H p (X, ) H q (Y, ). It remains to show that the Leray spectral sequence degenerates in E2 . For this, we argue as in the proof of Th. IV-15.9. In that proof, we dened a morphism of the double complex C p,q = [p] (X) [q] (Y ) into the double complex that denes the Leray spectral sequence (in IV-15.9, we only considered the sheaf theoretic external tensor product , but there is an obvious morphism of that one into the analytic tensor product). We get a morphism of spectral sequences which induces at the E2 -level the obvious morphism H p (X, ) H q (Y, ) H p (X, ) H q (Y, ).
p,q It follows that the Leray spectral sequence Er is obtained for r 2 by taking the , completion of the spectral sequence of C . Since this spectral sequence degenerates in E2 by the algebraic Knneth theorem, the Leray spectral sequence also satises dr = 0 u for r 2.
(5.24) Remark. If X or Y is compact, the Knneth formula holds with instead of u , and the assumption that both cohomology spaces are Hausdor is unnecessary. The proof is exactly the same, except that we use (5.22 c) instead of (5.22 b). 5.C. Modules over Nuclear Frchet Algebras e Throughout this subsection, we work in the category of nuclear Frchet spaces. Recall e that a topological algebra (commutative, with unit element 1) is an algebra A together with a topological vector space structure such that the multiplication A A A is
439
continuous. A is said to be a Frchet (resp. nuclear) algebra if it is Frchet (resp. nuclear) e e as a topological vector space. (5.25) Denition. A (Frchet, resp. nuclear) A-module E is a (Frchet, resp. nuclear) e e space E with a A-module structure such that the multiplication A E E is continuous. The module E is said to be nuclearly free if E is of the form A V where V is a nuclear Frchet space. e Assume that A is nuclear and let E be a nuclear A-module. A nuclearly free resolution L of E is an exact sequence of A-modules and continuous A-linear morphisms (5.26) Lq Lq1 L0 E 0
dq
in which each Lq is a nuclearly free A-module. Such a resolution is said to be direct if each map dq is direct, i.e. if Im dq has a topological supplementary space in Lq1 (as a vector space over R or C, not necessarily as a A-module). (5.27) Proposition. Every nuclear A-module E admits a direct nuclearly free resolution. Proof. We dene the standard resolution of E to be Lq = A . . . A E where A is repeated (q + 1) times; the A-module structure of Lq is chosen to be the one given by the rst factor and we set d0 (a0 x) = a0 x, dq (a0 . . . aq x) =
0 i<q
(1)i a0 . . . ai ai+1 . . . aq x
+ (1)q a0 . . . aq1 aq x. Then there is a homotopy operator hq : Lq Lq+1 given by hq (t) = 1 t for all q (hq , however, is not A-linear). This implies easily that L is a direct nuclearly free resolution. If E and F are two nuclear A-modules, we dene E A F to be (5.28) E A F = coker E A F E F d(x a y) = ax y x ay.
d
where
Then E A F is a A-module which it is not necessarily Hausdor. If E A F is Hausdor, it is in fact a nuclear A-module by Prop. 5.14. If E is nuclearly free, say E = A V V A, we have E A F = V F (which is thus Hausdor): indeed, there is an exact sequence V A A F V A F V F 0, v a0 a1 x v a0 a1 x v a0 a1 x,
v a x v ax,
440
obtained by tensoring the standard resolution of F with V ; observe that the tensor product with a nuclear space preserves exact sequences thanks to Prop. 5.17. We further dene TrA (E, F ) to be oq (5.29) TrA (E, F ) = Hq (E A L ), oq
where L is the standard resolution of F . There is in fact an isomorphism x A (a0 a1 . . . aq y) a0 x a1 . . . aq y where A is repeated q times in the target space. In this isomorphism, the dierential becomes dq (x a1 . . . aq y) = a1 x a2 . . . aq y +
1 i<q
E A L E A A F
(1)i x a1 . . . ai ai+1 . . . aq y
+ (1)q x a1 . . . aq1 aq y. In particular, we get TrA (E, F ) = E A F . Moreover, if we exchange the roles of o0 E and F , we obtain a complex which is isomorphic to the above one up to the sign of dq , hence TrA (E, F ) TrA (F, E). If E = A V is nuclearly free, the complex oq oq E A L = V L is exact, thus E or F nuclearly free = TrA (E, F ) = 0 for q oq 1.
(5.30) Proposition. For any exact sequence 0 E1 E2 E3 0 of nuclear A-modules and any nuclear A-module F , there is an (algebraic) exact sequence TrA (E1 , F ) TrA (E2 , F ) TrA (E3 , F ) TrA (E1 , F ) oq oq oq o q1 E1 A F E2 A F E3 A F 0. Proof. As the standard resolution L F is nuclearly free, Lq = A Vq say, then Ej A L = Ej V for j = 1, 2, 3, so we have a short exact sequence of complexes 0 E1 A L E2 A L E3 A L 0. (5.31) Corollary. For any nuclearly free (possibly non direct) resolution L of F , there is a canonical isomorphism TrA (E, F ) Hq (E A L ). oq Proof. Set Bq = Im(Lq+1 Lq ) for all q 0 and B1 = F . Then apply (5.30) to the short exact sequences 0 Bq Lq Bq1 0 and the fact that Lq is nuclearly free to get TrA (E, Bq ) o k1 for k > 1, TrA (E, Bq1) ok ker(E A Bq E A Lq ) for k = 1.
441
Hence we obtain inductively TrA (E, F ) = TrA (E, B1 ) . . . TrA (E, Bq2) oq oq o1 ker(E A Bq1 E A Lq1 ) and a commutative diagram E A Lq+1 E A Lq E A Bq1 0 E A Bq
in which the horizontal line is exact (thanks to the surjectivity of the left oblique arrow and the exactness of the sequence with E A Bq as rst term). Therefore ker(E A Bq1 E A Lq1 ) can be interpreted as the kernel of E A Lq E A Lq1 modulo the image of E A Lq+1 E A Lq , and this is is precisely the denition of Hq (E A L ). Now, we are ready to introduce the crucial concept of transversality. (5.32) Denition. We say that two nuclear A-modules E, F are transverse if E A F is Hausdor and if TrA (E, F ) = 0 for q 1. oq For example, a nuclearly free A-module E = A V is transverse to any nuclear A-module F . Before proving further general properties, we give a fundamental example. (5.33) Proposition. Let X, Y be Stein spaces and let U U X, V Y be is a coherent sheaf over X Y , then (U ) and (U V ) are Stein open subsets. If transverse over (U ). Moreover
(U ) (U)
(U V ) =
(U V ).
be a free resolution of over U V ; such a resolution exists by Proof. Let Cartans theorem A. Then (U V ) is a resolution of (U V ) which is nuclearly free over (U ), for (U V ) = (U ) (V ) ; in particular, we get
But
(U V ) is a resolution of
Tr(U) oq
(U ) (U) (U V ) = (U ) (V ) = (U V ), (U ) (U) (U V ) = (U V ).
(U V ), so (U V ) = 0 (U V ) for q = 0, for q 1.
(U ),
(5.34) Properties. a) If 0 E1 E2 E3 0 is an exact sequence of nuclear A-modules and if E2 , E3 are transverse to F , then E1 is transverse to F .
442
b) Let A A1 A2 be homomorphisms of nuclear algebras and let E be a nuclear Amodule. if A1 and A2 are transverse to E over A, then A2 is tranverse to A1 A E over A1 . c) Let E be a complex of nuclear A-modules, bounded on the right side, and let M be a nuclear A-module which is transverse to all E n . If E is acyclic in degrees k, then M A E is also acyclic in degrees k.
d) Let E , F be complexes of nuclear A-modules, bounded on the right side. Let f : E F be a A-linear morphism and let M be a nuclear A-module which is transverse to all E q and F q . If f induces an isomorphism H q (f ) : H q (E ) H q (F ) in degrees q k and an epimorphism in degree q = k 1, then IdM A f : M A E M A F has the same property. Proof. a) is an immediate consequence of the Tr exact sequence. o To prove b), we need only check that if A1 is transverse to E over A, then TrA1 (A2 , A1 A E) = TrA (A2 , E), oq oq n 0.
Indeed, if L = A V is a nuclearly free resolution of E over A, then A1 A L = A1 V is a nuclearly free resolution of A1 A E over A1 , since Hq (A1 A L ) = TrA (A1 , E) = 0 for q 1. Hence oq TrA1 (A2 , A1 A E) = Hq A2 A1 (A1 A L ) = Hq A2 A1 (A1 V ) oq = Hq (A2 V ) = Hq (A2 A L ) = TrA (A2 , E). oq
dq
c) The short exact sequences 0 Z q (E ) E q Z q+1 (E ) 0 show by backward q induction on q that M is transverse to Z (E ) for q k 1. Hence for q k 1 we obtain an exact sequence 0 M A Z q (E ) M A E q M A Z q+1 (E ) 0, which gives in particular Z q (M A E ) = B q (M A E ) = M A Z q (E ) for q desired. k, as
dq
d) is obtained by applying c) to the mapping cylinder C(f ), as dened in the following lemma (the proof is straightforward and left to the reader). (5.35) Lemma. If f : E F is a morphism of complexes, the mapping cylinder C = C(f ) is the complex dened by C q = E q F q1 with dierential dq E f q 0 dq1 F : E q F q1 E q+1 F q .
Then there is a short exact sequence 0 F 1 C E 0 and the associated connecting homomorphism q : H q (E ) H q (F ) is equal to H q (f ) ; in particular, C is acyclic in degree q if and only if H q (f ) is injective and H q1 (f ) is surjective.
443
5.D. A-Subnuclear Morphisms and Perturbations We now introduce a notion of nuclearity relatively to an algebra A. This notion is needed for example to describe the properties of the (S)-linear restriction map (S U ) (S U ) when U U . (5.36) Denition. Let E and F be Frchet A-modules over a Frchet algebra A and let e e f : E F be a A-linear map. We say that a) f is A-nuclear if there exist a scalar sequence (j ) with |j | < +, an equicontinuous family of A-linears maps j : E A and a bounded sequence yj in F such that for all x E f (x) = j j (x)yj .
b) f is A-subnuclear if there exists a Frchet A-module M and an epimorphism p : M e E such that f p is A-nuclear; if E is nuclear, we also require M to be nuclear. If f : E F is A-nuclear and if u : S E and v : F T are continuous A-linear maps then v f u is A-nuclear; the same is true for A-subnuclear maps. If V and W are nuclear spaces and if u : V W is C-nuclear, then IdA u : A V A W is A-nuclear. From this we infer: (5.37) Proposition. Let S, Z be Stein spaces and let U U Z be Stein open subsets. Then the restriction : (SU ) (SU ) is (S)-nuclear. If is a coherent sheaf over Y Z with Y Stein and S Y , then the restriction map : (S U ) (S U ) is (S)-subnuclear. Proof. As (S U ) = (S) (U ) and (U ) (U ) is C-nuclear, only the second statement needs a proof. By Cartans theorem A, there exists a free resolution over S U . Then there is a commutative diagram
0(SU )
0(SU )
(SU ) (SU )
in which the top horizontal arrow is an (S)-epimorphism and the left vertical arrow is an (S)-nuclear map; its composition with the bottom horizontal arrow is thus also (S)-nuclear. Let f : E F be a A-linear morphism of Frchet A-modules. Suppose that f (E) e F1 where F1 is a closed A-submodule of F and let f1 : E F1 be the map induced by f . If f is A-nuclear, it is not true in general that f1 is A-nuclear or A-subnuclear, even if A, E, F are nuclear. However: (5.38) Proposition. With the above notations, suppose A, E, F nuclear. Let B be a nuclear Frchet algebra and let : A B be a C-nuclear homomorphism. Suppose e that B is transverse to E, F and F/F1 over A. If f : E F is A-subnuclear, then IdB A f1 : B A E B A F1 is B-subnuclear.
444
Proof. We rst show that A f1 : E = A A E B A F1 is C-nuclear. Since a quotient of a C-nuclear map is C-nuclear by Cor. 5.16 b), we may suppose for this that f is A-nuclear. Write f (x) = (t) = j j (x)yj , k k (t)bk , j : E A, k : A C, |j | < +, yj F, |k | < +, bk B
as in the denition of (A-)nuclearity. Then A f : E B A F is C-nuclear: for any x E, we have (j (x)) = j (x)(1) in the A-module structure of B, hence A f (x) = f (1 x) = = j (j (x)) A yj j k (k j )(x) bk A yj .
By our transversality assumptions, B A F1 is a closed subspace of B A F . As Im( A f ) B A F1 , the induced map A f1 : E B A F1 is C-nuclear by Cor. 5.16 a). Finally, there is a commutative diagram B E B (B A F1 ) B A E A
IdB f1 IdB (A f1 )
B A F1
in which the vertical arrows are B-linear epimorphisms. The top horizontal arrow is B-nuclear by the C-nuclearity of A f1 , hence IdB A f1 is B-subnuclear. Example 5.12 suggests the following denition (which is somewhat less general than some other in current use, but sucient for our purposes). (5.39) Denition. We say that a Frchet algebra A is fully nuclear if the topology of A is e dened by an increasing family (pt )t]0,1[ of multiplicative semi-norms that is, pt (xy) pt (x) pt (y) , such that the Banach algebra homomorphism Apt Apt is nuclear for all t < t < 1. If A is fully nuclear and t ]0, 1], we dene At to be the completion of A equipped with the family of semi-norms pt , ]0, 1[. Then At is again a fully nuclear algebra, and for all t < t < 1 the canonical map At At is nuclear: indeed, for t u < u < t , there is a factorization At Apu Apu At . If E is a nuclear A-module, we say that E is fully A-transverse if E is transverse to all At over A. Then by 5.34 b), each nuclear space (5.40) E t = At A E
is a fully At -transverse At -module. If f : E F is a morphism of fully A-transverse nuclear modules, there is an induced map (5.40 ) ft = IdAt A f : Et Ft , t ]0, 1].
445
(5.41) Example. Let X be a closed analytic subscheme of an open set CN , D = D(a, R) a polydisk and U = D X. We have an epimorphism (D) (U ). Denote by pt the quotient semi-norm of pt (f ) = supD(a,tR) |f | on (U ). Then (U ) equipped with (pt )t]0,1[ is a fully nuclear algebra, and (U )t = D(a, tR) X . Now, let Y be a Stein space, V Y a Stein open subset and a coherent sheaf over X Y . Then Prop. 5.33 shows that (U V ) is a fully transverse nuclear (U )-module.
(5.42) Subnuclear perturbation theorem. Let A be a fully nuclear algebra, let E and F be two fully A-transverse nuclear A-modules and let f, u : E F be A-linear maps. Suppose that u is A-subnuclear and that f is an epimorphism. Then for every t < 1, the cokernel of ft ut : Et Ft is a nitely generated At -module (as an algebraic module; we do not assert that the cokernel is Hausdor). Proof. We argue in several steps. The rst step is the following special case. (5.43) Lemma. Let B be a Banach algebra, S a Frchet B-module and v : S S a e B-nuclear morphism. Then Coker(IdS v) is a nitely generated B-module. Proof. Let v(x) = tion j j (x)yj be a B-nuclear decomposition of v. We have a factorizav = : S 1 (B) S

where (x) = j j (x) and (tj ) = tj yj . Set w = : 1 (B) 1 (B). As is B-nuclear, so is w, and , induce isomorphisms Coker(IdS v) Coker Id1 (B) w .

We are thus reduced to the case when S is a Banach module. Then we write v = v + v with v (x) = j j (x)yj , v (x) = j j (x)yj .
1 j N j>N
For N large enough, we have ||v || < 1, hence IdS v is an automorphism and Coker(IdS v v ) is generated by the classes of y1 , . . . , yN . Proof of Theorem 5.42. a) We may suppose that E is nuclearly free and that u is p A-nuclear, otherwise we replace f , u by their composition with A M M E, where M is nuclear and p : M E is an epimorphism such that u p is A-nuclear. b) As in (5.9), there is a A-nuclear decomposition u(x) = j j (x)yj where (yj ) converges to 0 in F . Since f is an epimorphism, we can nd a sequence (xj ) converging to 0 in E such that f (xj ) = yj . Hence we have u = f v where v(x) = j j (x)xj is a A-nuclear endomorphism of E, and the cokernel of f u is the image by f of the cokernel of IdE v. c) By a), b) we may suppose that F = E = A M , f = IdE and that u is A-nuclear. Let B be the Banach algebra B = Apt . Then B A E = B M is a Frchet B-module and e
446
IdB A u is B-nuclear. By Lemma 5.42, IdB A IdE IdB A u has a nitely generated cokernel over B. Now, there is an obvious morphism B At , hence by taking the tensor product with At B we get At B (B A E) = At B (B M ) = At M = At A E = Et and we see that IdEt ut = IdAt A IdE IdAt A u has a nitely generated cokernel over At . 5.E. Proof of the Direct Image Theorem We rst prove a functional analytic version of the result, which appears as a relative version of Schwartz theorem 1.9. (5.44) Theorem. Let A be a fully nuclear algebra, E and F complexes of fully Atransverse nuclear A-modules. Let f : E F be a morphism of complexes such that each f q is A-subnuclear. Suppose that E and F are bounded on the right and that H q (f ) is an isomorphism for each q. Then for every t < 1, there is a complex L of nitely generated free At -modules and a complex morphism h : L Et which induces an isomorphism on cohomology. Proof. a) We rst show the following statement:
Suppose that Et and Ft are acyclic in degrees > q. Then for every t < t, the cohomology space H q (Et ) H q (Ft ) is a nitely generated At -module. k Indeed, the exact sequences 0 Z k (Et ) Et Z k+1 (Et ) 0 show by backward induction on k that Z k (Et ) is fully At -transverse for k q. The same is true for Z k (Ft ). q Then ft is a At -subnuclear map from Z q (Et ) into Ftq , and its image is contained in the closed subspace Z q (Ft ). By Prop. 5.38, for all t < t, the map ftq = IdAt At ftq is a At -subnuclear map Z q (Et ) Z q (Ft ). By Prop. 5.34 d), H (ft ) is an isomorphism in all degrees, hence dq ftq : Ftq1 Z q (Et ) Z q (Ft ) t
is surjective. By the subnuclear perturbation theorem, the map dq 0 = IdAt At (dq ftq ) (0 ftq ) t t has a nitely generated At -cokernel for t < t < t, as desired. b) Let N be an index such that E k = F k = 0 for k > N . Fix a sequence t < . . . < tq < tq+1 < . . . < tN < 1. To prove the theorem, we construct by backward induction on q a q nitely generated free module Lq over Atq and morphisms dq : Lq Lq+1 , hq : Lq Etq tq such that i) L q, tq : 0 Lq Lq+1 LN 0 is a complex and tq tq h q, tq : L q, tq Etq is a complex morphism.
ii) The mapping cylinder Mq = C(h q, tq ) dened by k1 k Mq = kZ Lk q, tq Etq is acyclic in degrees k > q.
447
Suppose that Lk has been constructed for k C(ft h q, tq ) and the complex morphism q
Mq Nq ,
q. Consider the mapping cylinder Nq =
k1 Lk q, tq Etq Lk q, tq Ftk1 q
given by Id ftk1 . This morphism is Atq -subnuclear in each degree and induces an q isomorphism in cohomology (compare the cohomology of the short exact sequences associated to each mapping cylinder, with the obvious morphism between them). Moreover, Mq and Nq are acyclic in degrees k > q. By step a), the cohomology space H q (Mq, tq1 ) is a nitely generated Atq1 -module. Therefore, we can nd a nitely generated free Atq1 -module Lq1 and a morphism
q q1 dq1 hq1 : Lq1 Mq, tq1 = Lqq1 Etq1 t such that the image is contained in Z q (Mq, tq1 ) and generates the cohomology space q1 q2 H q (Mq, tq1 ). As Mq, tq1 = Etq1 , this means that Mq1 is also acyclic in degree q. Thus Lq1 , together with the maps (dq1 , hq1 ) satises the induction hypotheses for q 1, and L together with the induced map h : L Et is the required morphism of t t t complexes.
Proof of Theorem 5.1. Let X, Y be complex analytic schemes, let F : X Y be a proper analytic morphism and let be a coherent sheaf over X. Fix a point y0 Y , a neighborhood of y0 which is isomorphic to a closed analytic subscheme of a Stein open set 0 W Cn and a polydisk D0 = D(y0 , R0 ) W . The compact set K = F 1 (D Y ) can 0 be covered by nitely many open subsets U X which possess embeddings as closed analytic subschemes of Stein open sets 0 CN . Let 0 be Stein open 0 0 0 subsets such that U = U and U = U still cover K. Let i : U 0 and 0 0 j : Y D D be the embeddings and = i (j F ) the image sheaf of on 0 D0 . Let D D0 be a concentric polydisk. Then U F 1 (D) = ( D) is a fully transverse (D)-module by Ex. 5.41, and so is U F 1 (D) = ( D). Moreover, the restriction map is (D)-subnuclear by Prop. 5.37 applied to Prop. 5.33 shows that
U F 1(D)
U F 1 (D)
For every Stein open set V D,
(V ) (D) U F 1 (D)
= U F 1 (V ) .
Denote by F 1 (D) the collection U F 1 (D) and use a similar notation with = (U ). As F 1 (D), F 1 (D) are Stein coverings of F 1(D), the Leray theorem applied to the alternate Cech complex of over F 1 (D) and F 1 (D) gives an isomorphism By the above discussion, AC ( F 1 (D), ) and AC ( F 1 (D), ) are nite complexes of fully transverse nuclear (D)-modules, the restriction map AC ( F 1 (D), ) AC ( F 1 (D), ) H AC ( F 1 (D), ) = H AC ( F 1 (D), ) = H F 1 (D), .
448
is (D)-subnuclear and induces an isomorphism on cohomology groups. Set D = D(y0 , R) and Dt = D(y0 , tR). Theorem 5.44 shows that for every t < 1 there is a complex of nitely generated free -modules and a (Dt )-linear morphism of complexes
(Dt ) AC ( F 1 (Dt ), )
which induces an isomorphism on cohomology. Let V Dt be an arbitrary Stein open set. By Prop. 5.34 d) applied with M = (V ), we conclude that (V ) AC ( F 1 (V ), ) induces an isomorphism on cohomology. If we take the direct limit as V runs over all Stein neighborhoods of a point y Y Dt , we see that q ( ) Rq F over Y Dt , hence Rq F is Y -coherent near y0 .
449
References
Abdelkader, O. [1980] Annulation de la cohomologie dune varit khlrienne faiblement 1ee a e compl`te ` valeurs dans un br vectoriel holomorphe semi-positif, C. R. Acad. Sci. Paris 290 e a e (1980) 7578 ; Th`se de 3e cycle Universit Paris VI (1980). e e Akizuki, Y., Nakano, S. [1954] Note on Kodaira-Spencers proof of Lefschetz theorems, Proc. Jap. Acad. 30 (1954) 266272. Andreotti, A., Grauert, H. [1962] Thor`mes de nitude pour la cohomologie des espaces e e complexes, Bull. Soc. Math. France 90 (1962) 193259. Andreotti, A., Vesentini, E. [1965] Carleman estimates for the Laplace-Beltrami equation in complex manifolds, Publ. Math. I.H.E.S. 25 (1965) 81130. Aronszajn, N. [1957] A unique continuation theorem for solutions of elliptic partial dierential equations or inequalities of second order, J. Math. Pures Appl. 36 (1957) 235249. Atiyah, M.F., Bott, R., Patodi, V.K. [1973] On the heat equation and the index theorem, Invent. Math. 19 (1973) 279330. Atiyah, M.F., Singer, I.M. [1968] The index of elliptic operators III, IV, Ann. of Math. 87 (1968) 546604 ; 93 (1971) 119138. Avron, J., Herbst, I., Simon, B. [1978] Schrdinger operators with magnetic elds I, Duke o Math. J. 45 (1978) 847883. Azhari, A. [1990] Sur la conjecture de Chudnovsky-Demailly et les singularits des hypersurfaces e algbriques, Ann. Inst. Fourier (Grenoble) 40 (1990) 106117. e Bedford, E., Taylor, B.A. [1976] The Dirichlet problem for a complex Monge-Amp`re equation, e Invent. Math. 37 (1976) 144. Bedford, E., Taylor, B.A. [1982] A new capacity for plurisubharmonic functions, Acta Math. 149 (1982) 141. Bismut, J.M. [1987] Demaillys asymptotic inequalities: a heat equation proof, J. Funct. Anal. 72 (1987) 263278. Bochner, S. [1948] Curvature and Betti numbers (I ) and (II ), Ann. of Math. 49 (1948) 379390 ; 50 (1949) 7793. Bombieri, E. [1970] Algebraic values of meromorphic maps, Invent. Math. 10 (1970) 267287 ; Addendum, Invent. Math. 11 (1970) 163166. Borel, A., Serre, J.P. [1958] Le thor`me de Riemann-Roch, Bull. Soc. Math. France 86 (1958) e e 97136. Borel, A., Weil, A. [1954] Reprsentations linaires et espaces homog`nes khlriens des groupes e e e a e de Lie compacts, Sminaire Bourbaki expos no 100 par J.-P. Serre, (mai 1954), 8p. e e Bott, R. [1957] Homogeneous vector bundles, Ann. of Math. 66 (1957) 203248. Bott, R., Chern S.S. [1965] Hermitian vector bundles and the equidistribution of the zeroes of their holomorphic sections, Acta Math. 114 (1965) 71112. Bremermann, H. [1954] Uber die Aquivalenz der pseudokonvexen Gebiete und der Holomorphiegebiete im Raum von n komplexen Vernderlichen, Math. Ann. 128 (1954) 6391. a Brianon, J., Skoda, H. [1974] Sur la clture intgrale dun idal de germes de fonctions c o e e holomorphes en un point de Cn , C. R. Acad. Sci. srie A 278 (1974) 949951. e Cartan, H. [1954] Sminaire de lEcole Normale Suprieure 1950/54, W.A. Benjamin, New York e e (1967).
450
References
Cartan, H., Serre, J.P. [1953] Un thor`me de nitude concernant les varits analytiques e e ee compactes, C. R. Acad. Sci. Paris 237 (1953) 128130. Chern, S.S. [1946] Characteristic classes of hermitian manifolds, Ann. of Math. 71 (1946) 85121. Chern, S.S., Levine, H.I., Nirenberg, L. [1969] Intrinsic norms on a complex manifold, Global Analysis (papers in honor of K. Kodaira), Univ. of Tokyo Press, Tokyo (1969) 119139. Chow, W.L. [1949] On compact complex analytic varieties, Amer. J. Math. 71 (1949) 893914. Chudnovsky, G.V. [1979] Singular points on complex hypersurfaces and multidimensional Schwarz lemma, Sm. Delange-Pisot-Poitou, anne 1979/80, Progress in Math. no 12, Birkhuser (1981) e e a 2969. Deligne, P., Griths, P.A., Morgan, J., Sullivan, D. [1975] Real homotopy theory of Khler a manifolds, Invent. Math. 29 (1975) 245274. Demailly, J.P. [1982a] Formules de Jensen en plusieurs variables et applications arithmtiques, e Bull. Soc. Math. France 110 (1982) 75102. Demailly, J.P. [1982b] Sur les nombres de Lelong associs ` limage directe dun courant positif e a ferm, Ann. Inst. Fourier (Grenoble) 32 (1982) 3766. e Demailly, J.P. [1982c] Estimations L2 pour loprateur dun br vectoriel holomorphe semie e positif au dessus dune varit khlrienne compl`te, Ann. Sci. Ecole Norm. Sup. 15 (1982) 457511. ee a e e Demailly, J.P. [1984] Sur lidentit de Bochner-Kodaira-Nakano en gomtrie hermitienne, Smie e e e naire P. Lelong - P. Dolbeault - H. Skoda (Analyse) 1983/84, Lecture Notes in Math. no 1198, Springer-Verlag (1985) 8897. Demailly, J.P. [1985a] Mesures de Monge-Amp`re et caractrisation gomtrique des varits e e e e ee algbriques anes, Mm. Soc. Math. France (N.S.) 19 (1985) 1124. e e Demailly, J.P. [1985b] Sur les thor`mes dannulation et de nitude de T. Ohsawa et O. Abe e delkader, Sminaire P. Lelong - P. Dolbeault - H. Skoda (Analyse) 1985/86, Lecture Notes in Math. e no 1295, Springer-Verlag (1987) 4858. Demailly, J.P. [1985c] Une preuve simple de la conjecture de Grauert-Riemenschneider, Sminaire e P. Lelong - P. Dolbeault - H. Skoda (Analyse) 1985/86, Lecture Notes in Math. no 1295, SpringerVerlag (1987) 2447. Demailly, J.P. [1985d] Champs magntiques et ingalits de Morse pour la d -cohomologie, Ann. e e e Inst. Fourier (Grenoble) 35 (1985) 189229. Demailly, J.P. [1987a] Nombres de Lelong gnraliss, thor`mes dintgralit et danalyticit, e e e e e e e e Acta Math. 159 (1987) 153169. Demailly, J.P. [1987b] Mesures de Monge-Amp`re et mesures pluriharmoniques, Math. Zeit. 194 e (1987) 519564. Demailly, J.P. [1988a] Vanishing theorems for tensor powers of a positive vector bundle, Proceedings of the Conference Geometry and Analysis on Manifolds held in Katata-Kyoto (august 1987), edited by T. Sunada, Lecture Notes in Math. no 1339, Springer-Verlag (1988). Demailly, J.P. [1988b] Vanishing theorems for tensor powers of an ample vector bundle, Invent. Math. 91 (1988) 203220. Demailly, J.P., Skoda, H. [1979] Relations entre les notions de positivit de P.A. Griths et de e S. Nakano, Sminaire P. Lelong-H. Skoda (Analyse), anne 1978/79, Lecture notes in Math. no 822, e e Springer-Verlag, Berlin (1980) 304309. Demazure, M. [1976] A very simple proof of Botts theorem, Invent. Math. 33 (1976) 271272. De Rham, G. [1931] Sur lAnalysis situs des varits ` n dimensions (Th`se), J. Math. Pures ee a e Appl. 10 (1931) 115200. De Rham, G. [1952] Sur la rductibilit dun espace de Riemann, Comment. Math. Helv. 26 e e (1952) 328344. De Rham, G. [1955] Varits direntiables, Hermann, Paris (1955). ee e Diederich, K., Fornaess, J.E. [1977] Pseudoconvex domains: bounded strictly plurisubharmonic functions, Invent. Math. 39 (1977) 129141. Diederich, K., Pug, P. [1981] Uber Gebiete mit vollstndiger Khlermetrik, Math. Ann. 257 a a (1981) 191198. Dolbeault, P. [1953] Sur la cohomologie des varits analytiques complexes, C. R. Acad. Sc. Paris ee 236 (1953) 175177. Draper, R.N. [1969] Intersection theory in analytic geometry, Math. Ann. 180 (1969) 175204.
References
451
EGA, Grothendieck, G., Dieudonn, J.A. [1960-67] Elements de gomtrie algbrique I, II, e e e e III, IV, Publ. Math. I.H.E.S. no 4, 8, 11, 17, 20, 24, 28, 32 (19601967). El Mir, H. [1984] Sur le prolongement des courants positifs ferms, Acta Math. 153 (1984) 145. e Esnault, H., Viehweg, E. [1983] Sur une minoration du degr dhypersurfaces sannulant en e certains points, Math. Ann. 263 (1983) 7586. Farkas, H.M., Kra, I. [1980] Riemann surfaces, Graduate Texts in Math. 71, Springer-Verlag, Berlin (1980). Federer, H. [1969] Geometric measure theory, Springer Verlag, Grundlehren der math. Wissenschaften, Band 153, Berlin (1969). Frlicher, A. [1955] Relations between the cohomology groups of Dolbeault and topological invario ants, Proc. Nat. Acad. Sci. U.S.A. 41 (1955) 641644. Fujiki, A. [1978] Closedness of the Douady spaces of compact Khler spaces, Publ. R.I.M.S., Kyoto a Univ. 14 (1978) 152. Girbau, J. [1976] Sur le thor`me de Le Potier dannulation de la cohomologie, C. R. Acad. Sci. e e Paris, srie A 283 (1976) 355358. e Godement, R. [1958] Thorie des faisceaux, Hermann, Paris (1958). e Grauert, H. [1958] On Levis problem and the embedding of real-analytic manifolds, Ann. Math. 68 (1958) 460472. Grauert, H. [1960] Ein Theorem der analytischen Garbentheorie und die Modulrume komplexer a Strukturen, Publ. Math. I.H.E.S. 5 (1960) 233292. Grauert, H. [1962] Uber Modikationen und exzeptionelle analytische Mengen, Math. Ann. 146 (1962) 331368. Grauert, H., Remmert, R. [1984] Coherent analytic sheaves, Grundlehren der math. Wissenschaften 265, Springer-Verlag, Berlin (1984). Grauert, H., Riemenschneider, O. [1970] Khlersche Mannigfaltigkeiten mit hyper-q-konvexem a Rand, Problems in Analysis: a Symposium in honor of Salomon Bochner, Princeton, N.J., Princeton University Press 1970. Grauert, H., Riemenschneider, O. [1971] Verschwindungsstze fr analytische Kohomoloa u giegruppen auf komplexen Rumen, Invent. Math. 11 (1970) 263292. a Griths, P.A. [1966] The extension problem in complex analysis II; embeddings with positive normal bundle, Amer. J. Math. 88 (1966) 366446. Griths, P.A. [1969] Hermitian dierential geometry, Chern classes and positive vector bundles, Global Analysis, papers in honor of K. Kodaira, Princeton Univ. Press, Princeton (1969) 181251. Griths, P.A., Harris, J. [1978] Principles of algebraic geometry, Wiley, New York (1978). Grothendieck, A. [1957] Sur quelques points dalg`bre homologique, Thoku Math. J. 9 (1957) e o 119221. Grothendieck, A. [1958] La thorie des classes de Chern, Bull. Soc. Math. France 86 (1958) e 137154. Grothendieck, A. [1966] On the De Rham cohomology of algebraic varieties, Publ. Math. I.H.E.S. 29 (1966) 351359. Grothendieck, A. [1968] Standard conjectures on algebraic cycles, Bombay Colloquium on Algebraic Geometry, 1968, Oxford Univ. Press (1969). Greene, R.E., Wu, H.H. [1975] Embedding of open riemannian manifolds by harmonic functions, Ann. Inst. Fourier (Grenoble) 25 (1975) 215235. Gunning, R.C., Rossi, H. [1965] Analytic functions of several complex variables, Prentice-Hall, Englewood Clis, N.J. (1965). Hartshorne, R. [1966] Ample vector bundles, Publ. Math. I.H.E.S. 29 (1966) 319350. Hartshorne, R. [1970] Ample subvarieties of algebraic varieties, Lecture Notes in Math. no 156, Springer-Verlag, Berlin (1970). Hartshorne, R. [1977] Algebraic geometry, Springer-Verlag, Berlin (1977). Harvey, R. [1975] Holomorphic chains and their boundaries, Proc. of Symposia in Pure Math., Several Complex Variables, held in Williamstown in 1975, vol. XXX, part I, ed. by R.O. Wells (1977) 309382. Hironaka, H. [1964] Resolution of singularities of an algebraic variety over a eld of characteristic zero, Ann. of Math. 79 (1964) 109326.
452
References
Hirzebruch, F. [1966] Topological methods in algebraic geometry, Springer-Verlag, Berlin (1966). Hodge, W.V.D. [1941] The theory and applications of harmonic integrals, Cambridge University Press (1941). Hrmander, L. [1963] Linear partial dierential operators Grundlehren der math. Wissenschaften, o Band 116, Springer-Verlag, Berlin (1963). Hrmander, L. [1965] L2 estimates and existence theorems for the operator, Acta Math. 113 o (1965) 89152. Hrmander, L. [1966] An introduction to Complex Analysis in several variables, 1st edition, o Elsevier Science Pub., New York, 1966, 3rd revised edition, North-Holland Math. library, Vol 7, Amsterdam (1990). Hopf, H., Rinow, W. [1931] Uber den Begri der vollstndigen dierentialgeometrischen Flche, a a Comment. Math. Helv. 3 (1931) 209255. Jennane, B. [1976] Extension dune fonction holomorphe dnie sur une sous-varit avec contrle e ee o de la croissance, Sminaire P. Lelong - H. Skoda (Analyse) 1976/77, Lecture Notes in Math. no 694, e Springer-Verlag, Berlin (1978) 126133. Kawamata, Y. [1981] Characterization of abelian varieties, Compositio Math. 43 (1981) 253276. Kawamata, Y. [1982] A generalization of Kodaira-Ramanujams vanishing theorem, Math. Ann. 261 (1982) 4346. Kiehl, R., Verdier, J.L. [1971] Ein einfacher Beweis des Kohrenzsatzes von Grauert, Math. a Ann. 195 (1971) 2450. King, J.R. [1970] A residue formula for complex subvarieties, Proc. Carolina Conf. on Holomorphic mappings and minimal surfaces, University of North Carolina, Chapel Hill (1970) 4356. Kiselman, C.O. [1978] The partial Legendre transformation for plurisubharmonic functions, Invent. Math. 49 (1978) 137148. Kiselman, C.O. [1979] Densit des fonctions plurisousharmoniques, Bull. Soc. Math. France 107 e (1979) 295304. Kiselman, C.O. [1983] Sur la dnition de loprateur de Monge-Amp`re complexe, Analyse e e e Complexe, Proceedings of Journes Fermat (SMF), Toulouse, May 1983, Lecture Notes in Math., e vol. 1094, Springer-Verlag, Berlin (1984) 139150. Kiselman, C.O. [1986] Un nombre de Lelong ran, Preprint Uppsala 1986, Sm. dAnalyse e e Complexe et Gomtrie 198587, Facult des Sciences de Monastir, Tunisie (1987). e e e Kleiman, S. [1966] Toward a numerical theory of ampleness, Ann. of Math. 84 (1966) 293344. Kobayashi, S., Nomizu, K. [1963] Foundations of dierential geometry, Vol. I, II, Interscience Publishers, John Wiley & Sons, New York (1963). Kodaira, K. [1953a] On cohomology groups of compact analytic varieties with coecients in some analytic faisceaux, Proc. Nat. Acad. Sci. U.S.A. 39 (1953) 868872. Kodaira, K. [1953b] On a dierential geometric method in the theory of analytic stacks, Proc. Nat. Acad. Sci. U.S.A. 39 (1953) 12681273. Kodaira, K. [1954] On Khler varieties of restricted type, Ann. of Math. 60 (1954) 2848. a Kodaira, K. [1975] Collected works, Vol. I, II, III, Iwanami Shoten Publishers and Princeton University Press (1975). Kohn, J.J. [1963] Harmonic integrals on strongly pseudo-convex manifolds I, Ann. Math. 78 (1963) 206213. Kohn, J.J. [1964] Harmonic integrals on strongly pseudo-convex manifolds II, Ann. Math. 79 (1964) 450472. Lang S. [1965] Algebra, Addison-Wesley (1965). Lelong, P. [1942] Dnition des fonctions plurisousharmoniques, C. R. Acad. Sci. Paris 215 (1942) e p. 398 and p. 454. Lelong, P. [1957] Intgration sur un ensemble analytique complexe, Bull. Soc. Math. France 85 e (1957) 239262. Lelong P. [1967] Fonctionnelles analytiques et fonctions enti`res (n variables), Sm. de Math. e e Suprieures, 6e session, t 1967, Presses Univ. Montral (1968). e e e e Lelong, P. [1968] Fonctions plurisousharmoniques et formes direntielles positives, Dunod, Paris, e Gordon & Breach, New York (1968).
References
453
Le Potier, J. [1975] Annulation de la cohomologie ` valeurs dans un br vectoriel holomorphe a e de rang quelconque, Math. Ann. 218 (1975) 3553. Le Potier, J. [1977] Cohomologie de la grassmannienne ` valeurs dans les puissances extrieures a e et symtriques du br universel, Math. Ann. 226 (1977) 257270. e e Leray, J. [1950] Lanneau spectral et lanneau ltr dhomologie dun espace localement compact e et dune application continue, J. Math. Pures Appl. 29 (1950) 1139. Levi, E.E. [1910] Studii sui punti singolari essenziali delli funzioni analitiche di due o pi` variabili u complesse, Ann. Mat. Pura Appl. (3), 17 (1910) 6187. Malgrange, B. [1955] Existence et approximation des solutions des quations aux drives pare e e tielles et des quations de convolution, Ann. Inst. Fourier (Grenoble) 6 (19551956) 271355. e Moishezon, B.G. [1967] On n-dimensional compact varieties with n algebraically independent meromorphic functions, Amer. Math. Soc. Translations 63 (1967) 51177. Nakai, Y. [1963] A criterion of an ample sheaf on a projective scheme, Amer. J. Math. 85 (1963) 1426. Nakano, S. [1955] On complex analytic vector bundles, J. Math. Soc. Japan 7 (1955) 112. Nakano, S. [1973] Vanishing theorems for weakly 1-complete manifolds, Number Theory, Algebraic Geometry and Commutative Algebra, in honor of Y. Akizuki, Kinokuniya, Tokyo (1973) 169179. Nakano, S. [1974] Vanishing theorems for weakly 1-complete manifolds II, Publ. R.I.M.S., Kyoto Univ. 10 (1974) 101110. Narasimhan, R. [1966] Introduction to the theory of analytic spaces, Lecture Notes in Math. no 25, Springer-Verlag (1966). Narasimhan, R. [1971] Analysis on real and complex manifolds, Masson, New York (1971). Newlander, A., Nirenberg, N. [1957] Complex analytic coordinates in almost complex manifolds, Ann. Math. (2), 65 (1957) 391404. Norguet, F. [1954] Sur les domaines dholomorphie des fonctions uniformes de plusieurs variables complexes, Bull. Soc. Math. France 82 (1954) 137159. Ohsawa, T. [1979] Finiteness theorems on weakly 1-complete manifolds, Publ. R.I.M.S., Kyoto Univ. 15 (1979) 853870. Ohsawa, T. [1981] On H p,q (X, B) of weakly 1-complete manifolds, Publ. R.I.M.S., Kyoto Univ. 17 (1981) 113126. Ohsawa, T. [1982] A reduction theorem for cohomology groups of very strongly q-convex Khler a manifolds, Invent. Math. 63 (1981) 335354 ; 66 (1982) 391393. Ohsawa, T. [1984] Completeness of noncompact analytic spaces, Publ. R.I.M.S., Kyoto Univ. 20 (1984) 683692. Oka, K. [1942] Sur les fonctions de plusieurs variables VI, Domaines pseudoconvexes, Thoku o Math. J. 49 (1942) 1552. Oka, K. [1950] Sur les fonctions de plusieurs variables VII, Sur quelques notions arithmtiques, e Bull. Soc. Math. France 78 (1950) 127. Oka, K. [1953] Sur les fonctions de plusieurs variables IX, Domaines nis sans point critique intrieur, Jap. J. Math. 23 (1953) 97155. e Peternell, Th., Le Potier, J., Schneider, M. [1987a] Vanishing theorems, linear and quadratic normality, Invent. Math. 87 (1987) 573586. Peternell, Th., Le Potier, J., Schneider, M. [1987b] Direct images of sheaves of dierentials and the Atiyah class, Math. Zeit. 196 (1987) 7585. Ramanujam, C.P. [1974] Remarks on the Kodaira vanishing theorem, J. Indian Math. Soc. 36 (1972) 4150 ; 38 (1974) 121124. Remmert, R. [1956] Projectionen analytischer Mengen, Math. Ann. 130 (1956) 410441. Remmert, R. [1957] Holomorphe und meromorphe Abbildungen komplexer Rume, Math. Ann. a 133 (1957) 328370. Remmert, R., Stein, K. [1953] Uber die wesentlichen Singularitten analytischer Mengen, Math. a Ann. 126 (1953) 263306. Richberg, R. [1968] Stetige streng pseudokonvexe Funktionen, Math. Ann. 175 (1968) 257286. Rothstein, W. [1955] Zur Theorie der Analytischen Mannigfaltigkeiten im Raume von n komplexen Vernderlichen, Math. Ann. 129 (1955) 96138. a
454
References
Rudin, W. [1966] Real and complex analysis, McGraw-Hill, New York (1966). Schneider, M. [1974] Ein einfacher Beweis des Verschwindungssatzes fr positive holomorphe u Vektorraumbndel, Manuscripta Math. 11 (1974) 95101. u Schwartz, L. [1953] Homomorphismes et applications compl`tement continues, C. R. Acad. Sci e 236 (1953) 24722473. Serre, J.P. [1954] Fonctions automorphes: quelques majorations dans le cas o` X/G est compact, u Sm. Cartan (195354) 2-1 a 2-9. e ` Serre, J.P. [1955a] Un thor`me de dualit, Comment. Math. Helv. 29 (1955) 926. e e e Serre, J.P. [1955b] Faisceaux algbriques cohrents, Ann. of Math. 61 (1955) 197278. e e Serre, J.P. [1956] Gomtrie algbrique et gomtrie analytique, Ann. Inst. Fourier (Grenoble) 6 e e e e e (1956) 142. Shiman, B., Sommese, A.J. [1985] Vanishing theorems on complex manifolds, Progress in Math. no 56, Birkhuser (1985). a Sibony, N. [1985] Quelques probl`mes de prolongement de courants en analyse complexe, Duke e Math. J. 52 (1985) 157197. Siegel, C.L. [1955] Meromorphic Funktionen auf kompakten Mannigfaltigkeiten, Nachrichten der Akademie der Wissenschaften in Gttingen, Math.-Phys. Klasse 4 (1955) 7177. o Siu, Y.T. [1969] Analytic sheaf cohomology groups of dimension n of n-dimensional complex spaces, Trans. Amer. Math. Soc. 143 (1969) 7794. Siu, Y.T. [1974] Analyticity of sets associated to Lelong numbers and the extension of closed positive currents, Invent. Math. 27 (1974) 53156. Siu, Y.T. [1976] Every Stein subvariety has a Stein neighborhood, Invent. Math. 38 (1976) 89100. Siu, Y.T. [1984] A vanishing theorem for semi-positive line bundles over non-Khler manifolds, a J. Dierential Geom. 19 (1984) 431452. Siu, Y.T. [1985] Some recent results in complex manifold theory related to vanishing theorems for the semi-positive case, Proceedings of the Bonn Arbeitstagung 1984, Lecture Notes in Math. no 1111, Springer (1985) 169192. Siu, Y.T. [1986] Asymptotic Morse inequalities for analytic sheaf cohomology, Sm. Bourbaki, e expos no 666, juin 1986, 16 p. e Siu, Y.T. [1990] Calculus inequalities derived from holomorphic Morse inequalities, Math. Ann. 286 (1990) 549558. Skoda, H. [1972a] Sous-ensembles analytiques dordre ni ou inni dans Cn , Bull. Soc. Math. France 100 (1972) 353408. Skoda, H. [1972b] Application des techniques L2 ` ltude des idaux dune alg`bre de fonctions a e e e holomorphes avec poids, Ann. Sci. Ecole Norm. Sup. 5 (1972) 545579. Skoda, H. [1976] Estimations L2 pour loprateur et applications arithmtiques, Sminaire e e e P. Lelong (Analyse), anne 1975/76, Lecture Notes in Math. no 538, Springer-Verlag, Berlin (1977) e 314323. Skoda, H. [1978] Morphismes surjectifs de brs vectoriels semi-positifs, Ann. Sci. Ecole Norm. e Sup. 11 (1978) 577611. Skoda, H. [1982] Prolongement des courants positifs ferms de masse nie, Invent. Math. 66 e (1982) 361376. Sommese, A.J. [1978] Submanifolds of abelian varieties, Math. Ann. 233 (1978) 229256. Spanier, E.H. [1966] Algebraic topology, McGraw-Hill, New York (1966). Stein, K. [1951] Analytische Funktionen mehrerer komplexer Vernderlichen und das zweite a Cousinsche Problem, Math. Ann. 123 (1951) 201222. Stoll, W. [1966] The multiplicity of a holomorphic map, Invent. Math. 2 (1966) 1558. Thie, P. [1967] The Lelong number of a point of a complex analytic set, Math. Ann. 172 (1967) 269312. Umemura, H. [1973] Some results in the theory of vector bundles, Nagoya Math. J. 52 (1973) 97128. Viehweg, E. [1982] Vanishing theorems, J. Reine Angew. Math. 335 (1982) 18. Von Neumann, J. [1929] Zur Algebra der Funktionaloperationen und Theorie der normalen Operatoren, Math. Ann. 102 (1929) 370-427.
References
455
Waldschmidt, M. [1976] Proprits arithmtiques des fonctions de plusieurs variables II, Smiee e e naire P. Lelong (Analyse), anne 1975/76, Lecture Notes in Math. no 538, Springer-Verlag (1977) e 108135. Waldschmidt, M. [1978] Nombres transcendants et groupes algbriques, Astrisque no 69-70 e e (1979). Warner, F. [1971] Foundations of dierentiable manifolds and Lie groups, Academic Press, New York (1971). Weil, A. [1952] Sur les thor`mes de De Rham, Comment. Math. Helv. 26 (1952) 119145. e e Weil, A. [1957] Varits khlriennes, Hermann, Paris (1957). ee a e Wells, R.O. [1980] Dierential analysis on complex manifolds, Graduate Texts in Math. 65, 2nd edition, Springer-Verlag, Berlin (1980). Weyl, H. [1913] Die Idee der Riemannschen Flche, Teubner, Leipzig, (1913), 3. Auage, 1955. a Witten, E. [1982] Supersymmetry and Morse theory, J. Dierential Geom. 17 (1982) 661692. Witten, E. [1984] Holomorphic Morse inequalities, Taubner-Texte zur Math. 70, Algebraic and Dierential Topology, Ed. G.M. Rassias (1984) 318333.

Complex Geometry - Jean-Pierre Demailly

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Complex Geometry - Jean-Pierre Demailly

Загружено:

Авторское право:

Доступные форматы

Complex Analytic and Dierential Geometry

Version of Thursday September 10, 2009

1. Dierential Calculus on Manifolds

Chapter I. Complex Dierential Calculus and Pseudoconvexity

V Rm Fig. I-1 Charts and transition maps

In coordinates, it is easy to check that (1.3) [, ] =

1. Dierential Calculus on Manifolds

1.B.3. Exterior derivative. This is the dierential operator

dened in local coordinates by the formula (1.7) du =

(1)j j u(0 , . . . , j , . . . , p ) (1)j+k u([j , k ], 0 , . . . , j , . . . , k , . . . , p ).

Chapter I. Complex Dierential Calculus and Pseudoconvexity

associated with the De Rham complex K p = with compact support.

of smooth dierential forms

1. Dierential Calculus on Manifolds

f (x1 , . . . , xm ) dx1 . . . dxm .

1. Then K V is a union of smooth hypersurfaces with piecewise x V ; x1 0, . . . , xj = 0, . . . , xl 0 .

Chapter I. Complex Dierential Calculus and Pseudoconvexity

(1.18) Stokes formula.

uj dx1 . . . dxm , xj du.

uj dx1 . . . dxj . . . dxm =

Kdu + dKu = (1.20) Kdu + dKu =

uI (1, x) uI (0, x) dxI ,

Proof. If v is a p-form on M , then G v F v = (H i1 ) v (H i0 ) v = i (H v) i (H v) 1 0 = d(KH v) + KH (dv)

2. Currents on Dierentiable Manifolds

tp1 vI (tx) dt (1)k1 xik dxi1 . . . dxik . . . dxip .

2. Currents on Dierentiable Manifolds

(2.2) Denition. We introduce as follows spaces of p-forms on manifolds.

Chapter I. Complex Dierential Calculus and Pseudoconvexity

(M ) are dened similarly.

is dened similarly and is called the

2. Currents on Dierentiable Manifolds

2.B.2. Wedge Product. For T T g s q+r (M ) is dened by (2.8)

(M ), the wedge product

Chapter I. Complex Dierential Calculus and Pseudoconvexity

Then TI is a distribution of order s and it is easy to check that T =

F : M1 M2 map. The pull-back morphism

We leave the straightforward proof of the following properties to the reader.

2. Currents on Dierentiable Manifolds

such that FSupp T is proper, the direct image

map such that (G F )Supp T is proper, then G (F T ) = (G F ) T.

Chapter I. Complex Dierential Calculus and Pseudoconvexity

gI,J (x, y) dxI dyJ , |I| + |J| = q,

g(x, y) g(1,...,r),J (x, y) dx1 . . . dxr dyJ .

This follows precisely from (2.16) and Fubinis theorem.

2. Currents on Dierentiable Manifolds

Chapter I. Complex Dierential Calculus and Pseudoconvexity

(Fv ) T (v) dv,

(Hv ) ([0, 1] T ) (v) dv.

3. Holomorphic Functions and Complex Manifolds

The function f is holomorphic on if df is C-linear, that is, f /z = 0.

3. Holomorphic Functions and Complex Manifolds

i where d(z) = 2 dz dz = dx dy is the Lebesgue measure on C.

by Stokes formula. The last integral is equal to as tends to 0.

f (ei ) d and converges to f (0)

Chapter I. Complex Dierential Calculus and Pseudoconvexity

f (z1 , . . . , zn ) dz1 . . . dzn . (z1 w1 ) . . . (zn wn )

3. Holomorphic Functions and Complex Manifolds

A(1 + |z|)B with suitable constants A, B B.

0 is in fact a polynomial of total

Chapter I. Complex Dierential Calculus and Pseudoconvexity

uI,J (z) dzI dz J ,

The identity d2 = (d + d )2 = 0 is equivalent to (3.17) d2 = 0, d d + d d = 0, d2 = 0,

3. Holomorphic Functions and Complex Manifolds

p,q (X, C).