Annals of Global Analysis and Geometry 23: 359372, 2003.

2003 Kluwer Academic Publishers. Printed in the Netherlands.

359

A Bernstein Property of Affine Maximal

Hypersurfaces
AN-MIN LI and FANG JIA
Department of Mathematics, Sichuan University, Chengdu, Sichuan, P. R. China.
e-mail: {math-li,galaxyly2000}@yahoo.com.cn
(Received: 18 March 2002; accepted: 5 December 2002)
Communicated by: D. Ferus (TU Berlin)
Abstract. Let x: M An+1 be a locally strongly convex hypersurface, given as a graph of a
n
locally strongly convex function

xn+1 = f (x1 , . . . , xn ) defined in a domain
A . We introduce
a Riemannian metric G# = ( 2 f /xi xj ) dxi dxj on M. In this paper, we investigate the affine
maximal hypersurfaces which are complete with respect to the metric G# and prove a Bernstein
property for the affine maximal hypersurfaces.
Mathematics Subject Classification (2000): 53A15.
Key words: Bernstein property, affine maximal hypersurface.

Introduction
Affine maximal hypersurfaces are extremals of the interior variation of affinely
invariant volume. The corresponding EulerLagrange equation is a fourth-order
PDE. Originally, these hypersurfaces were called affine minimal hypersurfaces.
Calabi calculated the second variation and proposed calling them affine maximal
(see [1]). For affine maximal surfaces, there are different versions of so called affine
Bernstein conjectures. One is Cherns conjecture (see [6]). Another is a problem
raised by Calabi (see [2]), called Calabis conjecture (see [12]). The two conjectures differ in the assumptions on the completeness of the affine maximal surface
considered. While Chern assumed that the surface is a convex graph over R 2 , which
means that the surface is Euclidean complete, Calabi assumed that the surface is
complete with respect to the Blaschke metric. In [13], the authors present a proof of
Cherns conjecture. Under an additional assumption, the first author gave a partial
answer to Calabis conjecture (see [8]). Calabis conjecture was recently solved,
see [9, 14].
Let x: M An+1 be an affine maximal hypersurface given by a locally strongly
convex function
xn+1 = f (x1 , x2 , . . . , xn )

360

AN-MIN LI AND FANG JIA

defined in a domain
An . In this paper, we say that f is locally strongly
convex which means that the Hessian of the function f is positive definite. Then
x (M) is an affine maximal hypersurface if and only if f satisfies the following
fourth-order PDE:
  2 1/(n+2)
f
 det
= 0,
xi xj
where  denotes the Laplacian with respect to the Blaschke metric G, which is
defined by

  

1
ij
G det(Gkl )
.
=
xi
xj
det(Gkl )
Following Calabi [3] and Pogorelov [11], we consider the Riemannian metric G#
on M, defined by

fij dxi dxj ,
G# =
where fij = 2 f /xi xj . This is a very natural metric for a convex graph. Our
main result can be stated as follows:
THEOREM. Let xn+1 = f (x1 , . . . , xn ) be a locally strongly convex function defined in a domain
An . If
M = {(x1 , . . . , xn , f (x1 , . . . , xn ) | (x1 , . . . , xn )
}
is an affine maximal hypersurface, and if M is complete with respect to the metric
G# , then, in the case where n = 2 or n = 3, M must be an elliptic paraboloid.
For the proof of the theorem, we first show that if the maximal hypersurface M
is complete with respect to the metric G# and if the norm of its Ricci curvature
Ric# G# is bounded, then M must be an elliptic paraboloid. Next, we use Hofers
Lemma to prove that Ric# G# must be bounded.
1. Preliminaries
Let An+1 be the unimodular affine space of dimension n + 1, M be a connected and oriented C manifold of dimension n, and x: M An+1 a locally
strongly convex hypersurface. We choose a local unimodular affine frame field x,
e1 , e2 , . . . , en , en+1 on M such that
e1 , . . . , en Tx M,
det(e1 , . . . , en , en+1 ) = 1,
Gij = ij ,

en+1 = Y,

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

361

where we denote the Blaschke metric and the affine normal vector field by Gij and
Y , respectively. Denote by U , Aij k and Bij the affine conormal vector field, the
FubiniPick tensor and the affine Weingarten tensor with respect to the frame field
x, e1 , . . . , en , and by Rij denote the Ricci curvature. We have the following local
formulas (see [10]):

Aij k U,k Bij U,
(1.1)
U,ij =
U = nL1 U,

Aiik = 0,
Rij =

Amli Amlj +

(1.2)
(1.3)
n2
n
Bij + L1 ij ,
2
2

(1.4)

where L1 denotes the affine mean curvature, and , denotes the covariant differentiation with respect to the Blaschke metric (note that with respect to the orthonormal
j
frame field x, e1 , . . . , en , we have Aij k = Aik ). A locally strongly convex
hypersurface is called an affine maximal hypersurface if L1 = 0 everywhere.
Let x: M An+1 be given by a locally strongly convex function
xn+1 = f (x1 , . . . , xn ).
We choose the following unimodular affine frame field:


f
,
e1 = 1, 0, . . . , 0,
x1


f
,
e2 = 0, 1, 0, . . . , 0,
x2
.........


f
,
en = 0, 0, . . . , 1,
xn
en+1 = (0, 0, . . . , 0, 1).
Then the Blaschke metric is given by (see [10])
  2 1/(n+2)  2
f
f
dxi dxj .
G = det
xi xj
xi xj
The affine conormal vector field U can be identified with

  2 1/(n+2) 
f
f
f
,...,
,1 .

U = det
xi xj
x1
xn

362

AN-MIN LI AND FANG JIA

The formula U = nL1 U implies that x(M) is an affine maximal hypersurface

if and only if f satisfies the following PDE:
  2 1/(n+2)
f
 det
= 0,
(1.5)
xi xj
where  denotes the Laplacian with respect to the Blaschke metric, which is
defined by

  

1
ij
G det(Gkl )
.
=
xi
xj
det(Gkl )
Denote
  2 1/(n+2)
f
,
= det
xi xj
and
=

2G
.

(1.6)

Then (1.5) gives

 = 0.

(1.7)

Now we calculate . Let p M. We choose an orthonormal frame field

around p M. Then

2
,j
,
=

,i ,j2
2 ,j ,j i

,
,i =

2






2 ,j2 i + 2 ,j ,j ii
,j ,i ,j i
4G
4
+
2
.
 =

2
3
In the case where (p) = 0, it is easy to get, at p,

2
2 ,ij
.


(1.8)

Now we assume that (p)  = 0 and choose an orthonormal frame field such that,
at p M, ,1 = , ,i = 0, i > 1. Then




2
4
2 ,j2 i + 2 ,j ,j ii
,1
,1
,11
4
+2 3.
(1.9)
 =

363

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

The Schwarz inequality gives




2
2
2
2
,11
+
,ii
+2
,1i
,ij
i>1

i>1


1
2
,11
+
,ii
n 1 i>1
=
Similarly,

A2ml1

n
2 + 2
n 1 ,11

2
+2

2
,1i

i>1

2
,1i
.

(1.10)

i>1

n
A2 .
n 1 111

(1.11)

Taking the (n + 1)-th component of U,ij =

Aij k U,k Bij U , we have

,11 = A111 ,1 B11 .

(1.12)

Using the formula (1.4) and (1.12), we get



2
2
2
=2
A2ml1 ,1
+ (n 2)B11 ,1
2
,j ,j ii = 2R11 ,1
= 2

2
A2ml1 ,1
(n 2)

(n 2)

2
,11 ,1

2
,11 ,1

(n 2)A111

4
(n 2)2 (n 1) ,1
.
8n
2

Substituting (1.10) and (1.13) into (1.9), we obtain

2
2
2n ,11
i>1 ,1i
+4


n1

 4

2
,11
,1
(n 2)2 (n 1) ,1
+ 2
.
(n + 2)
2
8n
3
Note that



2
2
2
4
2
,11
,1
,1
,11
,i
i>1 ,1i
=4
+4
+ 3 4
.


2
Then (1.14) and (1.15) together give us

2
,i
n2 n 2 
,i
n

,i
+

2(n 1) 
2(n 1)

 4

n2 2 ,1
(n 2)2 (n 1)

.
+ 2
8n
2(n 1) 3

3
,1

(1.13)

(1.14)

(1.15)

(1.16)

364

AN-MIN LI AND FANG JIA

For n = 2, by (1.16) we have

2
,i
1

+ 2 .


(1.17)

For n = 3, by (1.16) we have

2
,i
2 
,i
3
+ 16

,i .
 4


(1.18)

Denote by # and Ric# the Laplacian and the Ricci curvature with respect to
the metric G# , resp. By definition of Laplacian and a direct calculation, we get
 2G#

= 2G ,

# r = r

(1.19)

n 2 < , r >G#
,
2

(1.20)

where r is the geodesic distance function with respect to the metric G# on M.

Denote fij k = 3 f /xi xj xk . We have the following formula (see [11, p. 38]):

(1.21)
f j l f hm (fhil fmj k fhik fmj l ),
Ric#ik = 14
where (f ij ) denotes the inverse matrix of (fij ).
2. Proof of the Theorem
We first prove the following lemma:
LEMMA 1. Let x: M An+1 be a locally strongly convex affine maximal
hypersurface, which is given by a locally strongly convex function:
xn+1 = f (x1 , . . . , xn ).
If M is complete with respect to the metric G# , and if there is a constant N > 0
such that Ric# 2G# N everywhere, then in the case where n = 2 or n = 3, M
must be an elliptic paraboloid.
Proof. Let p0 M. By adding a linear function and taking a parameter
transformation we may assume that p0 has coordinates (0, . . . , 0) and
f (0) = 0,

fi (0) = 0,

fij (0) = ij .

Denote by r(p0 , p) the geodesic distance function from p0 with respect to the
metric G# . For any positive number a, let Ba (p0 ) = {p M | r(p0 , p) a}.
Consider the function
F = (a 2 r 2 )2 

365

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

defined on Ba (p0 ). Obviously, F attains its supremum at some interior point p .

We may assume that r 2

is a C 2 -function in a neighborhood of p , and  > 0 at

p . Then at p , F,i = 0, F,ii 0, where , denotes the covariant differentiation

with respect to the Blaschke metric. We calculate both expression explicitly
2(r 2 ),i
,i
2
= 0,

a r2

2
,i
2r 2 2G
2r 2


0.

2
(a 2 r 2 )2 a 2 r 2

(2.1)
(2.2)

Inserting (2.1) into (2.2) and noting that

r 2 2G = 4r 2 r2G ,
r 2 = 2r2G + 2rr,
we get
6r 2 2G
2r 2


+

(a 2 r 2 )2 a 2 r 2
=

24r 2 r2G
4r2G
4rr
+
+ 2
.
2
2
2
2
2
(a r )
a r
a r2

(2.3)

Since
2(n 2)r < , r >G#
4r # r
4rr
+ 2
=
2
2
2
2
a r
a r
a r2
2

4r
# r
+

a2 r 2

1
24

2G#
3

24(n 2)2 r 2 r2G#

,
(a 2 r 2 )2

and
2G#
= 2G ,

r2G#

= r2G ,

r2G# = 1,
we have
24r 2 r2G#
4r2G#
4r# r


+
+
+

(a 2 r 2 )2
(a 2 r 2 )
(a 2 r 2 )
+
=

1
24

 24(n 2)2 r 2
+ 2
r2G#

(a r 2 )2

4
4r# r
24(1 + (n 2)2 )r 2
+
+
+
(a 2 r 2 )2
(a 2 r 2 )
(a 2 r 2 )

1
24


.

(2.4)

366

AN-MIN LI AND FANG JIA

Recall that (M, G# ) is a complete Riemannian manifold with the Ricci curvature bounded from below by a constant K, K > 0. We have

r# r (n 1)(1 + Kr).

(2.5)
Consequently, from (2.4) it follows that

24(1 + (n 2)2 )r 2
4n
4(n 1) K r


+ 2
+
+

(a 2 r 2 )2
(a r 2 )
(a 2 r 2 )

1
24


. (2.6)

For n = 3, we have by (1.18)



3
4

 2,i
2

3
4

1
6

  ,i i

  2,i
3
+
2
36r 2
+
(a 2 r 2 )2

1
12

1
12


,

(2.7)

where we used (2.1). Inserting (2.7) into (2.6) we get

288
192 K r
2016r 2
+
+
.
 2
(a r 2 )2 a 2 r 2
a2 r 2

(2.8)

Multiply both sides of (2.8) by (a 2 r 2 )2 . We obtain, at p ,

(a 2 r 2 )2  2304a 2 + 192 Ka 3
= b1 a 2 + b2 a 3 ,
(2.9)

where b1 = 2304 and b2 = 192 K. It is obvious from (1.17) that (2.9) also holds
for n = 2. Hence, at any interior point of Ba (p0 ), we have
 b1

a2 1

r2
a2

1
2 + b2

a 1


r2 2
a2

Let a , then
 0.
It follows that
 2 
f
= 1,
det
xi xj
and
G# = G.

(2.10)

367

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

This means that M is an affine complete parabolic affine hypersphere. From a

result of Calabi (see [10]) we conclude that M must be an elliptic paraboloid.

We now want to show that there is a constant N > 0 such that ||Ric# ||2G# N
everywhere. To this end, we need the following lemma (see [7, p. 635, lemma 26]),
which was applied several times in symplectic geometry.
LEMMA 2 (Hofer [7]). Let (X, d) be a complete metric space with metric d, and
Ba (p) = {x|d(p, x) a} be a ball with center p and radius a. Let H be a positive
continuous function defined on B2a (p). Then there is a point q Ba (p) and a
positive number - a/2 such that
H (x) 2H (q)

for all x B- (q) and -H (q)

a
H (p).
2

Now we assume that ||Ric# ||2G# is not bounded above. Then there is a sequence of
points p. M such that ||Ric# ||2G# (p. ) . Let B1 (p. ) be the geodesic ball with
center p. and radius 1. Consider a family of functions /(.): B2 (p. ) R, . N,
defined by
/(.) = ||Ric# ||G# +  + L,

(2.11)

where  is defined by (1.6) and


L=
f il f j m f kn fij k flmn .
Using Hofers Lemma with H = / 1/2 , we find a sequence of points q. and positive
numbers -. such that
/ 1/2 (x) 2/ 1/2 (q. ),

x B-. (q. ),

-. / 1/2 (q. ) 12 / 1/2 (p. ) .

(2.12)
(2.13)

The restriction of the hypersurface x to the balls B-. (q. ) defines a family M(.)
of maximal hypersurfaces. For every ., we normalize M(.) as follows:
Step 1. By adding a linear function we may assume that, at ql , (x1 , . . . , xn ) =
(0, . . . , 0) and
f (0) = 0,

fi (0) = 0.

We take a parameter transformation:

 j
ai (.)xj (.),

xi (.) =
j

(2.14)
j

where ai (.) are constants. Choosing ai (.) appropriately and using an obvious
 , we may assume that, for every ., we have fij (0) = ij . Note that,
notation f, /

368

AN-MIN LI AND FANG JIA

 is invariant.
under the parameter transformation (2.14), /
Step 2. We take an affine transformation by
xi (.),

xi (.) = a(.)

1 i n,

xn+1 (.),

xn+1 (.) = (.)

where (.) and a(.) are constants. It is easy to verify that each M(.)
again is
a locally strongly convex maximal hypersurface. We now choose (.) = a(.)2 ,
, one can see that
 (q. ). Using again an obvious notation f , /
(.) = /
f ij (.) = fij (.),

(.) =
/

1
 (.).
/
(.)

The first equation is trivial. We calculate the second one. From (1.6), (1.21) we
can easily get
,
= 1


||G # =
||Ric
#

1 #
Ric G# ,

.
= 1L
L

Then the second equality follows.

a (q. ) = {x M(.)
|
r(.)(q. , x) a}, where
r(.) is the geodesic
We denote B

# on M(.).
(.) is defined on
Then /
distance function with respect to the metric G
d(.)(q. ) with
the geodesic ball B
d(.) = -. / 1/2 (q. ) 12 / 1/2 (p. ) .
From (2.12) we have
(q. ) = 1,
/
(x) 4,
/

(2.15)
d(.)(q. ).
x B

(2.16)

We may identify the parametrization as (1 , . . . , n ) for any index .. Then f (.)

is a sequence of functions defined in a domain
(.) with 0
(.). Thus we have

a sequence M(.)
of maximal hypersurfaces given by f (.). We have
f (.)(0) = 0,

f (.)
(0) = 0,
i

2 f (.)
(0) = ij ,
i j

(2.17)

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

(.) = 1,
/

369
(2.18)

(.)(x) 4,
/
d(.) ,

d(.)(0).
x B

(2.19)

as . .

(2.20)

We need the following lemma:

LEMMA 3. Let M be an affine maximal hypersurface defined in a neighborhood
of 0 Rn . Suppose that, with the notations from above,
(i)

fij (0) = ij ,

(ii)

Ric# G# +  +

f il f j m f km fij k flmn 4.

2
i 1/4n2 }. Then there is constant C1 > 0 such
Denote D := {(1 , . . . , n )|
that, for (1 , . . . , n ) D, the following estimates hold:


(1)
fii 4n,
(2) 1/C1 det(fij ) C1 ,


(3) Define do by do2 = 1/7n2 (4n)n1 C1 then
do { i2 < 1/7n2 } D, where

do is the geodesic ball with center 0 and radius do with respect to the metric
G# .
Proof. (1) Consider an arbitrary curve




ai2 = 1, s 0 .
4 = 1 = a1 s, . . . , n = an s 
By assumption, we have

f il f j m f kn fij k flmn 4,

fii (0) = n.

Since f il f j m f kn fij k flmn is independent of the choice of coordinates 1 , . . . , n ,
for any point (s) we may assume that fij = i ij . Then


il

f f

jm

f fij k flmn =
kn

2
fij k
1
2
fij k

3 .
i j k
fii

It follows that

2

1
f

iik 3

fij2 k 4
( fii )
( fii )3

370

AN-MIN LI AND FANG JIA

and, hence,


1
fii (x(s)))
=

fiik (x(s))ak
ds
( fii (x(s)))3/2

2
1/2  
1/2

f (x(s))
2
n
iik
a
k
( fii (x(s)))3

2 n.


Solving this differential inequality with fii (0) = n, we get
1
d(


3/2
( fii (x(s)))

1
1
.
s n

( fii (x(s)))1/2
n
From the assumption we have s 1/2n , then (1) follows.
(2) Again consider an arbitrary curve




ai2 = 1, s 0 .
4 = 1 = a1 s, . . . , n = an s 
By assumption, we have

ij
f i j
4.
2
It follows that

2

i 2 4.
( fii )
By (1) we get

1 d(x(s))
4 n.

ds
Solving this differential inequality with (0) = 1, we obtain

4 ns ln(x(s)) 4 ns.
Recall that s 1/2n, then (2) follows.
(3) Denote by min , max the minimal and maximal eigenvalues of (fij ). Then,
from (1) and (2), we have max 4n and
1
n1
det(fij ) min n1
min .
max (4n)
C1

(2.21)

Hence, by (1) and (2.21)

4n

xi2 2


1
xi2 ,
n1
C1 (4n)

(2.22)

A BERNSTEIN PROPERTY OF AFFINE MAXIMAL HYPERSURFACES

371

and (3) follows.

We continue with the proof of the theorem. Since d(.) , we have D

(.)
for . big enough. In fact, by (2.22), the geodesic distance from
0 to the boundary

# on M(.)
is less than 1/ n. By Lemma 3 and
of D with respect to the metric G
k
bootstrapping, we may

get2 a C - estimate, independent of ., for any k. It follows
that there is a ball { i C2 } and a subsequence (still indexed by .) such
that f (.) converges to f on the ball and correspondingly all derivatives, where
C2 < 1/4n2 is very close to 1/4n2 . Thus, as a limit, we get a maximal hypersurface
defined on the ball, which contains a geodesic ball
d0 . We now extend the
M,
as follows: For every boundary point p = (10, . . . , n0 ) of the
hypersurface M
geodesic ball
d0 , we first make a parameter transformation
 j


ai j
i =
we
such that, at p, (
1 , . . . ,
n ) = (0, . . . , 0) and for the limit hypersurface M,
have f ij (0) = ij . We have
(i ) f ij (.)(p) f ij (0) = ij ,

as . .

It is easy to see that under the conditions (i ) and (ii) in Lemma 3, the estimates
(1), (2) and (3) in Lemma 3 still hold. By the same argument as above, we conclude
that there is a ball around p and a subsequence .k , such that f(.k ) converges to
f on the ball and, correspondingly, all derivatives. As a limit, we get a maximal
hypersurface M  , which contains a geodesic ball of radius do around p. Then we
return to the original parameters. Note that the geodesic distance is independent of
the choice of the parameters. It is obvious that M and M  agree on the common
part. We repeat this procedure to extend M to be defined on
2do , etc. In this way,
we may extend M to be a maximal hypersurface defined in a domain
Rn ,
# . Using (2.18) and (2.19), we get
which is complete with respect to the metric G
# ||G # 4,
||Ric

/(0)
= 1.

By Lemma 1, M must be an elliptic paraboloid. For a paraboloid, we have

= 0. Thus, we get a contradiction. So ||Ric# ||2 # must be bounded
identically /
G
above on M. By Lemma 1, M is an elliptic paraboloid. This completes the proof
of the theorem.


Before concluding this paper, we wish to state the following problem for higher
dimensions:
PROBLEM. Let x: M An+1 be a locally strongly convex hypersurface, given as
a graph of a locally strongly convex function
xn+1 = f (x1 , . . . , xn )

372

AN-MIN LI AND FANG JIA

defined in a domain
An . If x(M) is an affine maximal hypersurface and if
x(M) is complete with respect to the metric G# , is it an elliptic paraboloid?
Acknowledgements
The first author would like to thank Professors U. Simon and L. Vrancken for many
valuable discussions.
Both authors are partially supported by 973 project, NSFC 10271083 grant and
a Chinese-German exchange project of NSFC and DFG.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.

Calabi, E.: Hypersurfaces with maximal affinely invariant area, Amer. J. Math. 104 (1982),
91126.
Calabi, E.: Convex affine maximal surfaces, Results in Math. 13 (1988), 209223.
Calabi, E.: Improper affine hyperspheres of convex type and a generalization of a theorem by
K. Jogens, Michigan Math. J. 5 (1958), 105126.
Calabi, E.: Affine differential geometry and holomorphic curves, in: Lecture Notes in
Math. 1422, Springer, New York, 1990, pp. 1521.
Caffareli, L. A. and Guitierrez, C. E.: Properties of the solutions of the linearrized Monge
Ampere equations, Amer. J. Math. 119 (1997), 423465.
Chern, S. S.: Affine minimal hypersurfaces, in: Proc. Japanese-U.S. Seminar, 1977, Tokyo,
1978, pp. 1730.
Hofer, H.: Pseudoholomorphic curves in symplectizations with applications to the Weinstein
conjecture in dimension three, Invent. Math. 114 (1993), 515563.
Li, A.-M.: Some theorems in affine differential geometry, Acta Math. Sinica (NS) 5, 345354.
Li, A.-M. and Jia, F.: The Calabi conjecture on affine maximal surfaces, Results in Math. 40
(2001), 256272.
Li, A.-M., Simon, U. and Zhao G.: Global Affine Differential Geometry of Hypersurfaces, De
Gruyter, Berlin, 1993.
Pogorelov, A. V.: The Minkowski Multidimensional Problem, Wiley, New York, 1978.
Simon, U.: Affine differential geometry, in: Handbook of Differential Geometry, NorthHolland, Amsterdam, 2000, pp. 905961.
Trudinger, N. and Wang, X. J.: The Bernstein problem for affine maximal hypersurfaces, Invent.
Math. 140 (2000), 399422.
Trudinger, N. and Wang, X. J.: Affine complete locally convex hypersurfaces, Invent. Math.
150 (2002), 4560.
Yau, S. T. and Schoen, R.: Differential Geometry, Science Press, Beijing, 1988 [in Chinese].