Coursang5 PDF

Part V Iconoclasm
383
Chapter 16 Heterodox exponentials

We denitely enter the experimental part of these lectures. From now on, we present ideas whose main quality is their radical novelty ; they had thence not the time to be polished by use : an excess of rigour or details in their presentation would be out of place.
16.1
The quarrell of images
One can contend that ludics solves in principle perfect logic. Only in principle : it provides a setting the analytical theorems where one can perform a not too fabricated synthesis. We are not done, but we can see the end of the tunnel ! The question is to determine what to do with the imperfect part, i.e., with exponentials. The absence of continuity between a perfect world, of very restricted expressivity, but harmonious, and an imperfect one where the growth of functions cannot be controlled is the sign that something goes wrong. Remember the towers of exponentials whose height is a tower of exponentials. . . do we really believe in that ? This is however the necessary consequence of the mental image of the logical world that we bear in mind ; while the experience of perfect logic entitles one to question this badly innite innity, this very perennial perenniality. The question is therefore : should one respect mental images, be iconodule, or should one be iconoclast1 ?
16.1.1
Classical absolutism
The strongest iconodule argument is evidence : the world is classical, because our fundamental intuitions are classical ; the classical is an absolute that one
The quarrel of images ravaged Byzantium during the VIIIth and IXth centuries ; the iconoclast enemies of images emperors, like Constantin Copronymos (sic), destroyed all mosaics but in the places they no longer ruled, Ravenna for instance.
1
384
16.1. THE QUARRELL OF IMAGES
385
cannot surpass. . . This conformism rests upon a long experience, upon an undeniable internal coherence ; at the foundational level, it also rests upon a marked taste for essence, for revealed truth. One must say that even constructivists are of the same opinion : thus, Martin-Lf believes in the set of integers ; should we conclude that his theory of types is only a dancer, a layer of constructive varnish applied on a classical and set-theoretic wall ? Everything you just read is nice, but not that serious : when the recess is over, the children must go back to the classical. What is almost unstoppable. Except that this is circular, that this is the very blind spot ; you may be right, Mr. Iconodule, but admit that one cannot see anything : in a case like this, one should only trust indirect evidence. Fortunately for iconoclasm, there are LLL and ELL, the systems that we shall soon introduce. These experimental systems invert to their prot the revealed aspect of exponentials, by giving of them lightened versions, with a maintenance of a subtle innity, inaccessible to set-theorical methods. The sole existence of these systems is enough to refute the a priori objections, resting upon so-called priority of the classical : while admitting a certain amount of perenniality, they present it in a less absolute, less desperately frozen, version.
16.1.2
Mathematics
Modifying logic since this is the eventual goal of iconoclasm means giving up mathematics, which can be expressed, as one knows, in set-theory. A light logic would thence lose mathematical results : You want to destroy the mosaics ! say the iconodules. This is not that obvious : First, one should not confuse mathematics with mathematics revisited by logicians, which contain an enormous amount of innite combinatorics. Real mathematics makes a more restricted use of the innite, and is thence less sensitive to its precise formulation. If, as in LLL, the function m ; 2m is no longer available, this does not mean that a result mnA[m, n] where the solution n is bounded by 2m is irremediably lost : one could still write it mnA[log m, n]. This is more complicated, but this removes the objection of principle. Take a musical analogy : the equal temperament is extremely convenient ; however, the tempered intervals are slightly out of tune. This does not mean that all music written for the equal temperament but perhaps for certain exaggerations, e.g., dodecaphonism should be dumped. In the same way, classical mathematics is only slightly wrong, as long as one does not enter into logicist exaggerations.
386
CHAPTER 16. HETERODOX EXPONENTIALS
16.1.3
Sophistics
A sophist is the guy that says to his teacher of sophistics : of two things, either you were a bad teacher and I owe no money to you, or you educated me well and I can produce a sophism proving that I owe nothing to you . The same kind of argument proves the impossibility of motion or the dumbness of general relativity, not to speak of quantum mechanics ; if a three-dimensional variety is embeddable in a four-dimensional Euclidian space, sophistics will conclude that the world is eventually Euclidian : this is the theme of hyperspace, familiar to readers of science-ction. The impressive work done around 1900 enables one to code everything in set-theory. Hence even the most delirious iconoclasm can be represented in set-theory : which is thence primal, according to a certain sophistics, as in the case of the aforementioned hyperspace. But nobody takes hyperspace seriously : the Euclidian space is rather seen as a convenient frame for the approximation of true geometry. In the same way, one can contend that set-theory has no real sense, that it is only a convenient reication of a reality dicult to access.
16.1.4
Iconoclast inconsistencies
The iconoclast viewpoint is delicate, since lacking in coherence ; in particular, the extant systems, like LLL and ELL are only experimental. But this is a dynamic position, with its future ahead, and a captivating motivation : complexity theory. The gradual setting of an iconoclast logic should lead to pose questions in a radically dierent way. In particular, to nd nuances, mistakes, in the prevailing foundational paradigm, which mainly rests on an uncouth approach to natural numbers.
16.2
16.2.1
Exponentials
Kronecker
Any foundational iconoclasm sooner or later stumbles on the absoluteness of integers. Thus, the various techniques introduced in these lectures which mainly belong in nite combinatorics , bring us back to natural numbers. And one does not know how to unscrew natural numbers. Said Kronecker : God created the integers, everything else is the deed of man . How can one call into question this absoluteness of integers ? Non standard : Non-standard models of (classical) arithmetic introduce integers after the true ones . This is not very convincing : who has ever seen, who has ever been able to compute, a non-standard integer ?
16.2. EXPONENTIALS
387
Ultranitist : A proposition, not quite serious, by Essenin-Volpin : there would be integers only up to say 19. This want of earnestness is conrmed by his claim that these methods can be used to prove. . . the consistency of set theory ZF ! Dynamic : One no longer poses the question of enlarging or shortening the set N about integers, one ignores it : N is presumably only a reication. When one enunciates a theorem on integers, it is true because it has been proved, which in no way presupposes that {0, 1, 2, 3, . . . } makes any sense. What is important is the process of construction, the dynamics.
16.2.2
A challenge : complexity
The dynamical viewpoint tries to rationalise the intuition that a tower of exponentials whose height is in turn a tower of exponential is a pervert eect of the classical formalism ; that one accepts as a fall-back solution, by no way a reality. One cannot not refuse it for a given value of the argument what would bring us back to the rut of ultranitism , but as a parametric construction. Set-theory does allow the distinction between a parametric construction and the set of its instanciations corresponding to the values n = 0, 1, 2, . . . of the parameter, i.e., ignores dynamics. The question became central in the last part of the XXth because of the emergence of algorithmic complexity. The main achievement of this theory is the individuation of complexity classes, corresponding to the time or the space needed for the computation. Above all one knows P (problems computable in polynomial time) and NP (problems veriable in polynomial time), and the famous 1000000 $ question : P = NP Each complexity class possesses indirect and interchangeable characterisations, none of them being a mathematical denition in the noble sense of the term what would be for instance a preservation property. Which might be perhaps related to the rather modest outcome of the area : in more than thirty years, not a sole serious separation result between complexity classe ! I propose to take complexity seriously, not as a problem of informatics, but as a problem of logic. Although iconoclast, the following thesis is very exciting : Complexity classes do correspond to various sizes of innity. Obviously one is no longer in the Cantorian innity, which classies the settheoretic, static, innite, in terms of cardinals. One is no more in an intensional innity that would classify the innite according to stronger or weaker
?
388
systems, such as the reverse mathematics of section 3.C.4 or various bounded arithmetics , those bleak bureaucracies of complexity. Or, if one prefers, one surpassed this infantile stage : one will try to play on the only formal trace of innity in logic, exponentials.
16.2.3
Exponentials and integers
And one has one trump card, precisely this abrupt, sharp, modal, side of exponentials : this is like that . Which can be read a contrario, since one can modify them as we please : one will thence explore the possibilities of modications. Everything is permitted, but in order to take only the fair side of our freedom, one will impose a constraint : the modied exponentials must of tame growth. Polynomial, exponential, this does not matter much, one will try to tame the growth. This this what will do the lightened logics, LLL and ELL : Light Linear Logic and Elementary Linear Logic rst introduced in [45]. System F enables one to dene integers la Dedekind, remember, section 6.1.6 : nat := X (X ((X X ) X )), that one will write2 nat := X ((X X ) (X X )), what is more legible. This denition translates as nat := X (!(!X X ) (!X X )), with a big stock of exclamation marks. The same eect would have been obtained with the simplied version : nat := X (!(X X ) (X X )) (16.1) which is the type of the functional sending f X X to f . . . f X X . Analysing integers, surpassing Kroneckers absolute, this could thence be unscrewing the ! that struts about denition (16.1). What is not that simple, since the possible denitions of the exponential deal with nite reuses, nite cliques, i.e., presuppose integers. More generally, one sees that all analyses based upon categories eventually run into circles. One will search our happiness with layer 3, with proof-nets. Later, with operator algebras, if possible, see chapter 20.
16.2.4
The lesson of nets
Without much noise, something essential did occur with multiplicative nets. Let us indeed compare the proof of normalisation for natural deduction, section 4.3.5 with the erroneous proof of normalisation for proof-structures of section 11.2.6. In the rst case, a parameter, the degree, measures in its way the logical complexity, thence the algorithmic complexity of normalisation, i.e., the height of the tower of exponentials of section 4.C.4. In the second case,
2
Neglecting the technicalities linked to , see section 7.4.2 !
16.3. RUSSELLS ANTINOMY
389
on sees that normalisation is performed in linear time, since every step shrinks the size i.e., the number of links of the structure. What could happen is that one eventually reaches a vicious circle, but one would then be wise enough to stop. Let us go further and think of system F ; there is no longer any degree, since functions grow too fast, but there is still a logical control, operated by reducibility candidates, see section 6.2.4. On the other hand, if I extend multiplicative logic to second order, no signicant increase in the complexity of normalisation can be observed : the number of steps remains bounded by the number of links. To sum up, in the perfect world, logic does not control complexity while its does in the imperfect world. Could one nd a weakened version of the imperfect world lightened exponentials in which one would observe the same phenomenon, i.e., a priori bounds as to the complexity of normalisation, independent from the logical complexity of formulas, and even from the correction of their proofs ?
16.3
16.3.1
Russells antinomy
A paragon of complexity
What is the paragon of algorithmic or logical complexity ? Imagine a very inexpressive system : for him, there is not much dierence between a tower of towers of exponentials and a diverging normalisation, between a system of high logical complexity and an inconsistent system. In other terms, one can consider metaphorically non-termination, Russells antinomy, as the worse possible complexity. Thence the idea of analysing what, in Russells antinomy provokes non-termination ; one knows that this is due to exponentials. If one can isolate the principles causing non-termination, and if one can redesign viable exponentials excluding these principles, one will get a logical system whose complexity is bounded a priori. What follows is by many standards very unexpected : in one century, thousands of pages have been written on Russells antinomy without any signicant progress. The decomposition perfect/imperfect operated by linear logic enables one to make a breakthrough on this issue.
16.3.2
Dissection of exponentials
Russells antinomy can be decomposed in two steps : (i) The construction of a xed point for negation, say A =! A.
390
(ii) The logical transformation of this xed point into a contradiction. The rst step is without interest : it means that complexity is not bridled. Everything concentrates on the second step and on the ne grain analysis of exponentials. Here follows a list of micro-properties of exponentials : ! (A & B ) ! A ! B ! A ! B ! (A & B ) AB (16.2) (16.3) (16.4) (16.5) (16.6) (16.7) (16.8)
The rst two principles correspond to the isomorphism at the origin of the qualication exponentials : (16.2) expresses contraction and (16.3) expresses weakening. (16.4) expresses the functoriality of ! : it is the solution of a universal problem. (16.5) is a weak form of dereliction. These four principles constitute the base of LLL. (16.6), combined with (16.4), corresponds to multilinear functoriality : from B conclude ! !B ). ELL corresponds to the rst ve principles. (16.7) is dereliction. Burying (16.8) corresponds to the fact that, from ! A on deduces ! !A, and not simply ! ! !A. Both principles are separately faulty : in presence of a xed point A = ! A, one gets the empty sequent which is not quite contradictory, but not cut-free provable anyway in two ways : (i) From (16.2) + (16.4) + (16.7) (ii) From (16.2) + (16.4) + (16.5) + (16.8) Indeed, A, ?A is an identity axiom ; one deduces ?A, ?A, as we please : (i) By a dereliction (16.7) which directly yields ?A, ?A. (ii) (16.4) yields !A, ? ?A, then (16.5) yields ?A, ? ?A and (16.8) removes the extra ? . From ?A, ?A, contraction (16.2) yields ?A, i.e., A, which, by (16.4), entails ! A. On concludes by a cut between ?A and ! A. The principles (16.7) and (16.8) are excluded once for all. The pseudodereliction (16.5), which occurs in the reference version [45], has not been retained here. LLL, with the sole (16.4) would be too weak : what explains the spare modality, A, which behaves like a castrated ! , i.e., a supertype of ! .
!A !B !A ?A ! A ! B ! (A B ) !A A !A !!A
16.4. LLL AND ELL
391
16.3.3
Russell and normalisation
Let us try to understand how cut-elimination works in the two possible translations of Russells antinomy. The cut between ?A and ! A reduces to (two) cuts between ?A, ?A and two copies of ! A. Then one proceeds dierently : (i) If ?A, ?A is obtained through a dereliction from A, ?A, one opens one of the boxes ! A which yields A, which yields in turn two cuts between A, ?A, A and ! A. Since A, ?A is an identity axiom, the cut between A, ?A and A reduces to ?A, which nally gives back a cut between ?A and ! A, the beginning of an innite loop of an eternal golden braid, would say the poet ! (ii) In the second case, the system reduces into two cuts between ?A, ? ?A, ! A and ! ! A, next between !A, ? ?A, ? A and ! ! A. Which simplies into a cut between ? ?A and ! ! A. One loops too, but by entering one notch inside the boxes . One sees that dereliction opens a box (depth 1) to pour it out at depth 0, while burying moves the cut from depth 0 to depth 1. The principles (16.7) and (16.8) which do not respect the depth contrarily to the others are the cause of non-termination and more generally of unlimited complexity.
16.4
16.4.1
LLL and ELL

The systems
The language contains multiplicatives, rst and second order quantiers, as well as the exponentials !, ?, , , with A := A, . . . , are useful only in the case of LLL ; notice the absence of additives. The exponential rules are as follows : (i) Weakening on all negative formulas ( A being declared negative). (ii) Contraction on formulas ?A. (iii) Two promotion rules : B1 , . . . , Bn , A (16.9)
1 B1 , . . . , n Bn , B
Either = ! and for i = 1, . . . , n, i =? ; in the case of LLL, n 1. Or = and for i = 1, . . . , n, i = or i = ?.
392
What is not quite the original system, but rather a simplied variant due to Asperti [6]. The original system [45] is indeed complicated by the presence of additives ; but, in presence of weakening, they can be dened without exponentials by means of : A B := X ((A X ) (B X ) X ) A & B := X ((X A) (X B ) X ) (16.10) (16.11)
which are simplied versions of the translations of section 12.B.2. Weakening is acceptable in a polarised world, see sections 14.C and 15.3.1, which is the case here. In gneral, there are many variants of ELL and especially LLL, without one being able to arbitrate in favour of such or such. Thus, the original version of LLL contained the extra principle (16.5), which was not retained here ; it declared the paragraph as self-dual : = ; it had no weakening and was fussing about a pedantic restriction : exactly one formula in the context of the promotion of ! . All these variants are based upon the respect of the depth of boxes, but one hardly sees anything beyond. So, if one is to navigate by sight, the better is to go to the simplest, i.e., Aspertis variant.
16.4.2
Light nets
One uses underlinings like in section 11.C.2 ; but here the only underlined formulas will be conclusions of boxes, what corresponds to the absence of dereliction. From a net with conclusions , A3 , one can construct a box of conclusions , !A or , A ; in LLL, the boxes , !A are restricted to the case where has at most one element. The link , of conclusion A, has at most one premise A ; this premise cannot be the conclusion of a link ! . The link ?, of conclusion ?A, has n premises A, (n 0). The 0-ary links and ? are particular cases of a more general weakening link, without premise and of negative conclusion. The size is measured depth by depth : s0 , s1 , s2 , . . . . For a certain d, sd = 0, and this degree will not change during normalisation. s0 counts the links at depth 0, with a certain pounderation : each link has a weight, the number of its non underlined conclusions ; thence boxes, multiplicative and quantier links count for 1 ; the same for the links , ? and weakening. But the axiom counts for 2, while cut is not counted.
3
These conclusions are not underlined, what corresponds to the absence of burying.
16.4. LLL AND ELL
393
16.4.3
Bounds for LLL
One knows that everything normalises, so I will content myself with bounds on the size of the normalised net. These bounds are easily converted into bounds on the computation time. (i) One begins by normalising at depth 0 all cuts but exponential ones. One knows that the size shrinks, so that one can keep the same bounds. (ii) A cut between two paragraphs, coming from boxes of respective conclusions , A et A, , B normalises by burying the cut at depth 1, between A and A, and by reforming a box of conclusions , , B ; the size shrinks. Finally, it only remains cuts !/?. These cuts are handled like in the case of the paragraph, with an essential dierence, due to duplication. If I start with ?A0 , cut with ! A0 , there is a multiplicative factor equal to n 1, where n is the arity of the link of conclusion ?A0 . What really causes a problem is the context A1 of the box introducing ! A0 : this context is also multiplied by n 1. One sees that the duplication in A0 is transmitted just so to A1 ; one can proceed in the same way if A1 is premise of a cut contraction, what can lead to A2 . Indeed, if one wants to count the duplications, one must take the paths A0 , A1 , A2 , . . . , Ak which are maximal, i.e., which cannot be extended, neither before 0, nor after k . These paths are nite, i.e., A0 = Ak : it suces to choose the adequate switching. How many of them ? This is very simple, as much as possible choices for A0 , i.e., less than the size s0 of the net at depth 0. Therefore, the size does not increase in depth 0, but it is at most multiplied by a factor s0 in the lower depths. One can redo this at depths 1, then 2, etc. One gets the following bounds : Profondeur 0 Profondeur 1 Profondeur 2 Profondeur d : : : : s0 , s1 , s2 , . . . , sd s0 , s0 s1 , s0 s2 , . . . , s0 sd 2 s0 , s0 s1 , s2 0 s1 s2 , . . . , s0 s1 sd
2 s0 , s0 s1 , s2 0 s1 s2 , . . . , s0
d 1
2 s1
d 2
. . . s2 d2 sd1 sd
One thence sees that, when normalisation is over, the global size has been d changed from s = s0 + s1 + s2 + . . . + sd to at most s2 . For a given d, this a polynomial, thence the polynomial time.
16.4.4
Bounds for ELL
In ELL, exponential boxes are of conclusions , !A, without constraint on . One sees that the cleansing of depth 0 induces a multiplication by an
394
exponential factor of the lower depths, since there are much more sequences A0 , A1 , A2 , . . . , Ak : the number of such sequences is bounded by 2s0 . One sees that the complexity of ELL, is, for a xed d, dominated by a tower of exponentials.
16.5
Expressive power
We know that the complexity of normalisation for a given depth is polynomial (LLL) or elementary, i.e., bounded by a tower of exponentials (ELL). We shall prove the converse. The essential eort will put on LLL.
16.5.1
Coding of polynomials
nat := X (!(X X ) (X X )) (16.12)
In LLL, integers are typed by :
One must have an exponential in output , so as to preserve depth. In ELL, one can take ! ; this does not work with LLL, since one could not type integers 2, 3, . . . which require (16.6). This is why one uses the paragraph and this is by the way the unique reason for its existence. Given A and f A A, one can form !f !(A A). If x nat, then ({x}A)!f (A A). In other terms, one can iterate and the result is of type (A A). One can represent the following functions : Sum : m, n ; m + n of type nat, nat nat : one basically uses (X X ), (X X ) (X X ). Product : m, n ; n m of type nat, !nat nat ; one iterates addition, rewritten of type !nat !(nat nat), which yields nat, !nat (nat nat), hence nat, !nat, nat nat, corresponding to m, !n, p ; (p + n m). Square: One can build an object of type nat (nat!nat) corresponding to n ; (n!n). Combining this with the product, one sees that squaring can receive the type nat nat, what corresponds to n ; n2 . General polynomials : No problem : for instance, n4 can be typed nat nat.
16.5.2
nat2
Coding of polynomial time

:= X (!(X X ) (!(X X ) (X X ))) (16.13)
One must code integers in base 2 :
16.5. EXPRESSIVE POWER
395
One can type polynomials in the length |n| of a binary integer. For this is suces to remark that the identication of the two arguments in a binary integer n (by contraction) yields the integer |n| of nat. The next step is to code a Turing machine in LLL : Tape : The tape is can be represented by a generalisation of nat2 , say natN , where N corresponds to the number of symbols that can written on it. States : A type of the form boolS := X (X . . . X X ), with S elements, can be used for the current state of the machine. Reading head : The type nat can be used for the current position of the reading head. So that the current position of a Turing machine can be given the type Tur := natN boolS nat, and the machine itself the type Tur Tur. Given the input n nat2 , one can type the initial position of the machine pn Tur, as well as the iteration P (|n|) of the machine from the initial position pn . By cooking all this together, one actually gets that every polynomial time algorithm can be typed nat2 . . . nat2 , with a number of paragraphs depending on the degree of the polynomial.
16.5.3
Remark on the paragraph
One would search in vain a proof of (A B ) A B . I give a procedural argument ; the same would also work against !(A B ) !A!B . In presence of this principle (supposedly doing what one thinks), every polynomial time algorithm with a boolean output (hence of type A B ) normally typed nat . . . (A B), could be retyped nat . . . A . . . B. But normalisation is done in such a way that depth 0 is cleansed in linear time. Nothing new will befall later ; now the bit left/right which distinguishes the two values has been taken back to depth 0. No need to proceed with normalisation, one already knows the result !
16.5.4
Case of ELL
nat := X (!(X X ) !(X X )) (16.14)
On denes, without scheming :
and one check that multiplication by 2 can be given the type nat nat. Which enables one to type the iteration of multiplication par 2, a function of type nat !(nat nat), hence nat, !nat !nat, corresponding to n, !p ; !(p.2n ). One thence gets the type nat !nat for the exponential n ; !2n , hence nat !! . . .!!nat for a tower of exponentials.
Chapter 17 Quantum coherent spaces

The experimental gait of this last part is easily regressive. Thus, we relinquished coherent spaces for reason of unfaithfulness, see the discussion in section 12.1 ; this being said, it is a most simple technique which can sometimes directly lead to the essential. This is why we shall meet them gain to establish a link with the quantum world. This link, see [48], is valid for the sole nite dimensional spaces ; to go further, other techniques (Geometry of Interaction ) will be needed, but what we shall see here is worth the dtour.
17.1
17.1.1
Logic and quantic

A missed encounter
No need to go very far to understand why the relation between logic and quantic has been this complete failure : in the same way Frege had the nerve to make fun of the revolutionary ideas of Riemann, logicians were not afraid to declare that nature makes mistakes and therefore to try to reform, to reformat it. Thence the notorious quantum logic an expression of the style popular democracy , where the role of the adjective is to negate the noun1 . It is necessary to make it clear from the very start : What follows has nothing to do with quantum
logic .
17.1.2
Caracteristics of the quantic
What stroke people in the quantic, it its non-determinism, hardly accepted for essentially ideological reasons, even by the great Einstein : God does not play dice with the world . However, in a strictly deterministic world,
Another example : labeled deductive systems ; labeled meaning that the system is not deductive ; think also of the slogan logic plus control .
1
396
17.1. LOGIC AND QUANTIC
397
the theory of chaos, initiated by Poincar, shows the practical impossibility of predicting the outcome of the national lottery or the position of the solar system in 106 years : this the famous metaphor of the buttery. But the chaos is not too shocking, since it leaves open the possibility of an abstract, inhuman determinism : if one actually knew how to photograph the world at instant t, one could deduce its position at instant t + 10. It is however necessary to remark that this if one knew is just as dubious as the idea that beyond incompleteness there would be a stable, eternal, notion of truth : chaos looks like a nite, eective, version of incompleteness, even in some of its readings. Nothing of the like in the quantum world : one cannot be reassured with the idea that there would be too many parameters, hardly measurable. One must admit that measurement inuences the result in a very strong sense : not only it modies it, but it creates it. Thus, when I measure a spin, I measure it w.r.t. an axis, say, Z , and I nd that there is actually a spin 1/2 w.r.t. this axis, while the electron under measurement had none before. This sort of behaviour entails non-determinism, what shocked more than one ; it induces various hidden variables theories, all of them worn out2 . Deeper than non-determinism is the imbrication between the observer and the system under observation. Contrarily to the usual physical world, the quantic does not accept a dichotomy suject/object.
17.1.3
Quantum logic
Von Neumann himself should be held responsible for the birth of quantum logic ; but not guilty, since, in the beginning of the years 1930, it was natural to make attempts. It was indeed an approach of level 1 : one modies the truth values. One knows that classical logic admits a semantics in terms of the truth values v, f , and, more generally, in terms of Boolean algebras. Von Neumann did propose to replace Boolean algebras with the lattice of closed subspaces of a Hilbert space. . . which yielded strictly nothing. By the way, von Neumann soon took an otherwise more fruitful direction that of what is known to us as von Neumann algebras. Independently from its technical vacuousness, quantum logic was a mistake a priori. Indeed, level 1 rests upon the duality syntax/semantics, i.e., the schizophrenia subjet/object, in opposition to quantum mechanics which rests upon their imbrication. This level of reading supposes a Fregean position : thence the expression the impulsion of M has a denotation, i.e., a value,
But in logic : the fashion of quantum computation induced a come back to old moons, style Gleasons theorem , in a new rear-guard attempt at justifying hidden variables and rehabilitate determinism !
2
398
CHAPTER 17. QUANTUM COHERENT SPACES
which is a real number ; one can make function the swing which separates to relate them better the sense and its denotation. But when I say the spin of e , this has strictly no denotation, no value , in no space, over-ornate or not. Let us quickly conclude : if an idea is bad, one cannot x it by a formalisation. This is nevertheless what quantum logicians did by introducing orthomodular lattices , thence kicking out the only interesting datum, the Hilbert space. Since that time, quantum logic vegetates as a theory of lattices, preferably ill-behaved, without any relation to quantum mechanics. Some quantum logicians dont even know what is a Hilbert space. The attitude of logic w.r.t. quantic can be summarised by a sophism : one can describe quantic with mathematics, and mathematics can be embedded in set-theory, hence in logic. Only remains to code this doohickey . Anything is good, preferably the most ad hoc possible, to mark a very Fregean reprobation in front of the mistake committed by Nature. One thinks of the anecdote reported by Herodotus (VII,35) : a tempest having destroyed the boat of ships he had set over the Hellespont, Xerxes had the sea whipped. In the same way, quantum logic is a whipping of nature, guilty of illogicality .
17.1.4
The question
Remains anyway the problem of the relation between logic and quantic. Here, more than ever, one sees that the question at stake is precisely that of the choice of the question. There is a problem, but which one indeed ? This problem is not that of the logical explanation of quantic, although one necessarily expects clarications. There is a reason, the quantum formalism wave functions, density matrices cannot be replaced with sets, graphs, in two words with what we meet in the logical world. A contrario, observe that it is an essentialist prejudice to believe in a world of ideas a word moreover rather set-theoretic which would rule the universe from above. What if the world of ideas were not what one believes, but were rather quantic ? After all, rather than teaching nature, why not trying to be its pupil ? Instead of trying to interpret quantic into logic, one will interpret logic into quantic3 .
3
This shift of viewpoint took me 30 years of (part time) reection.
17.2. PROBABILISTIC COHERENT SPACES
399
17.1.5
Methodology
One must precise what I mean by quantic . Surely not quantum physics which keeps its own life, far from logic ; indeed, I only mean the processus of quantum measurement, the way in which quantum beings interact. This could go as far as trying to integrate the types of quantum socialisation, e.g., the fermions asocial creatures, like electrons or bosons gregarious particles in a re-reading of the notion of equality. One must perhaps consider a third partner, quantum computation, a fashionable and interesting idea, albeit still at the stage of science-ction. Quantum coherent spaces perhaps bear some relation to quantum computation, and such a relation would be welcome ; the papers [79] and [95] are anyway encouraging. If not, their theoretical interest lies in a break with set-theory, a break that category-theory sought without nding it. This is why the viewpoint taken here is the sole opening of the logical space to non set-theoretic techniques. After reading this chapter, it will no longer be tenable to indulge into sophisms like ideas are language, the language being written with symbols a, b, c. . . ; indeed, what if those symbols were not commuting ?
17.1.6
PCS and QCS
The quantic is based upon superposition, just like linear logic which comes from coherent spaces, and a sort of superposition principle, remember, section 9.1.1 : if a = i ai , then F (a) = i F (ai ) (preservation of disjoint unions). We shall start from that, with the idea of an analogy cliques/wave functions (rather : density operators), to revitalise level 2. One proceeds in two steps, rst a probabilistic, i.e., commutative generalisation, then the general case. As to usual coherent spaces, the Ariadne thread is as follows : points do constitute a distinguished basis and cliques correspond to subspaces which express diagonally in this base, i.e., with coecients 0, 1. The probabilistic version allows real coecients on the diagonal. As to the quantic version, it steps out from the diagonal , by allowing, among others, the introduction of an identity function which cannot be described as a clique, i.e. as a subspace.
17.2
17.2.1
Probabilistic coherent spaces

Desessentialisation
We shall lazily approach the quantic, by rst adding a probabilistic, i.e., nondeterministic, dimension. Here, the main reference is the desessentialisation
400
of coherent spaces of section 9.1.5. Remember that coherent spaces can be dened, given a support |X |, by the duality between subsets of |X | :
| a b:
(a b) 1
(17.1)
and that linear implication corresponds to the adjunction : ((F )a b) = (Sq(F ) a b) (17.2)
Non-determinism essentially concerns the connective : one will randomly choose between A and B in A B . One could thence get expressions a + (1 ) b, with a A, b B . Which suggests replacing cliques with functions taking their values in [0, 1]. The formulation (17.1) as well as the adjunction (17.2) will survive this generalisation : typically (17.1) becomes (17.3) infra. By the way, should I recall it ? We are at level 2, and this has thank you my God ! nothing to do with fuzzy logic !
17.2.2
The bipolar theorem
Definition 78 (Duality) Let |X | be a nite set ; if f, g : |X |R, one denes the scalar product f | g := x|X | f (x) g (x). The positive functions f, g are polar, notation | f g , when : f |g 1 (17.3) The space R+ |X | of all functions from |X | into R+ is thence equipped with a polarity, whose pole is the segment [0, 1], see section 7.1.1. One denes as usual the polar of a set A R+ |X |, and : Definition 79 (Probabilistic coherent spaces) A probabilistic coherent space (PCS) is the pair (|X |, X ) of a nite support |X | and a subset X R+ |X | equal to its bipolar. Theorem 59 (Bipolar) Let X be a PCS ; then (i) X is non empty ; it indeed contains the null function : 0|X | X . (ii) X is a closed convex set. (iii) X is downwards stable : if f g X , then f X . Conversely, every subset of R+ (|X |) enjoying (i)-(iii) is a PCS.
17.2. PROBABILISTIC COHERENT SPACES
401
Proof : Il is immediate that any PCS satises (i)-(iii) ; for instance, if f, g X , if h X and if 0 1, then f + (1 )g | h = f | h + (1 ) g | h 1, hence X is convex. Conversely, if X satises (i)-(iii), and if f X , the Hahn-Banach theorem applied to the real Banach space R(|X |) says that a hyperplane separates X (closed convex) from f . In other terms, one can nd a linear form such that (X ) 1, (f ) > 1. This linear form can be identied with an element h R(|X |), in other terms, (g ) = g | h ; one denes h (x) := sup(h(x), 0). Obviously, f | h f | h > 1. If g X , dene g (x) := g (x) if h(x) 0, g (x) := 0 otherwise. Since g g , (iii) yields g X . But g | h = g | h 1, hence h X . Then f X = X . 2 Additives One adopts here a rather locative viewpoint : one supposes that the supports |X | and |Y | are disjoint. The additive connectives will build PCS with support |X | |Y |. If f R+ (|X |), g R+ (|Y |), one denes f g R+ (|X | |Y |) in the obvious way, by gluing ; one identies f with f 0|Y | , g with 0|X | g . The set : X & Y := {f g ; f X, g Y } (17.4) is the polar of X Y . On the other hand, X Y is not a PCS ; X Y must be dened as (X Y ), without hope of removing the bipolar. But theorem 59 yields : Proposition 31 X Y = {f (1 )g ; f X, g Y, 0 1} (17.5)
Proof : It is enough to remark that the right hand side of (17.5) enjoys conditions (i)-(iii) : it is therefore a PCS, moreover the smallest containing X Y , since it is its convex envelope. 2 Multiplicatives The multiplicative connectives will produce PCS of support |X | |Y |. Definition 80 (Adjunction) If R+ (|X | |Y |), if f R+ (|X |), one denes ()f R+ (|Y |) : (()f )(y ) :=
x|X |
(x, y ) f (x) = (, y ) | f
(17.6)
What makes sense due to the niteness of |X |.
402
Theorem 60 The function ; () is a bijection between R+ (|X | |Y |) and the set of linear applications from the convex cone R+ (|X |) to the convex cone R+ (|Y |). can be recovered from the associated linear function = () by means of : (x, y ) = (x )(y ) with x (x) := 1, x (y ) := 0 for y = x. Proof : A linear function satises (f + g ) = (f )+ (g ) for , 0, and is thence determined by its value on the x , which explains (17.7). Everything is by the way more or less immediate. 2 In the ground case (sets, coherent spaces) this would not work : if and f are sets (characteristic functions), ()f has no reason to be a set. This is why one introduced coherence together with its corollary, the unicity of the witness a in equation (8.13). Definition 81 (Linear implication) If X, Y are PCS, one denes the PCS X Y of support |X | |Y |, as the set of all such that () sends X into Y . Thus the characteristic function |X | of the diagonal belongs to X X ; indeed (|X | )f = f . X Y is the polar of {f g ; f X, g Y }, it is why it is a PCS. Which enables one to introduce X Y := X Y , and dually, X Y = {f g ; f X, g Y }. Proposition 32 is commutative, associative, and distributes over &. Proof : By introducing the obvious notation (), one could as well dene X Y as the set of such that () sends Y into X . The result follows by imitation of theorem 51, see chapter 14. 2 (17.7)
17.3
Quantum coherent spaces
It is only a slight exaggeration to say the quantum version corresponds to the probabilistic one when one has forgotten the base {x ; x |X |}. Anyway, the rst thing to do is to come back to PCS so as to draw some general considerations.
17.3. QUANTUM COHERENT SPACES
403
17.3.1
Methodological backlash
PCS have a vague resemblance to the coherent Banach spaces (CBS) of section 15.A. If one forgets exponentials, one can restrict to nite dimensional real spaces. In such a case, a CBS can be handled by means of an Euclidian (i.e., real Hilbertian) space E , by means of the duality :
| x y
| x|y |1
(17.8)
Indeed, if X denotes the unit ball of a normed space on the same E , it is immediate that the unit ball of its dual can be identied with X , see equation (15.8). Hence, any real nite dimensional CBS can be described as a set equal to its bipolar in a appropriate Euclidian space E . This being said, the denition by means of equation (17.8) is slightly more general than CBS ; indeed, if X = X , x 1 := sup {; x X } does not necessarily denes a norm. It is possible that x = 0 or worse, that x = +. What one usually handles by restricting to points of nite norm , then by quotienting by the points of null norm . This eventually amounts at modifying E , which shows that (17.8) is not quite more general than the denition of CBS. But, if one takes into account the locative viewpoint, the subtyping X Y means that, on the same vector space E , one can have more coherent objects, i.e., that the unit ball increases. In other terms, the norm decreases, Y X . It can thence become null, i.e., become a semi-norm ; dually, it can become innite, and in this case, there is not even a name for what one gets. PCS are not dened on Euclidian spaces, but on positive cones linked to a distinguished base, what is foreign to the spirit of linear algebra. This being said, positivity can be desessentialised : Proposition 33 f R(|X |) is in R+ (|X |) i for all g R+ (|X |) the scalar product f | g is positive. Proof : Immediate. 2
In particular, we shall see that the QCS which have however an intrinsic notion of positivity, positive hermitians call for this variability of positivity. To sum up, the bilinear form x | y induces three dualities, depending one takes as pole [1, +1] (which yields the norm, i.e., coherence), [0, +] (which yields positivity), or the intersection of both, [0, 1], which corresponds to PCS, and what we shall keep just so.
404
17.3.2
The bipolar theorem strikes back
Let us go back to theorem 59, in a more general setting. In what follows, E is a nite dimensional Euclidian space. The duality is dened by :
| x y
0 x|y 1
(17.9)
The problem is the charcaterisation of bipolars. Theorem 61 (Bipolar) A subset C E is its own bipolar i : (i) 0 C . (ii) C is a closed convex set. (iii) If nx C for all n N, then x C . (iv) If x, y C , if , 0 and x + y C , then x C . Proof : We begin with necessity ; (i) et (ii) are immediate. (iii) : if nx C for n N, and z C , then x | z [0, 1/n] for n N, hence x | z = x | z = 0 [0, 1]. (iv) : if z C , then 0 x + y | z 1, 0 x | z , 0 y | z , hence 0 x | z 1. (iv) induces a sort of converse to (iii) : if x, x C , then nx + n(x) = 0 C , hence nx C . Let us now pass to suciency, and assume that C satises (i)-(iv) ; let C + be the cone nN n C (= R+ C ). One can rewrite (iv) : C = C + (C C + ) If b C , then, by (17.10), one must consider two cases : b C + : by Hahn-Banach, there is a d E such that b | d < 0 c | d for all c C . Condition (iii) implies that I = {c ; n N nc C } is a vector space ; | d thence vanishes on I , and one can write C = I C , with C = I C . C is compact : if one embeds E in the projective space, C has a compact closure, whose frontier is made of the lines R a included in C . But there is no such line (they have been removed and put in I ) : the frontier is therefore empty and C compact. It follows that | d is bounded on C , hence on C , and b | d < 0 c | d . By renormalising d one can suppose that = 1, and then d C and b C. (17.10)
17.3. QUANTUM COHERENT SPACES
405
b C C + : the same Hahn-Banach yields d E such that p | d 1 < b | d for all p C C + . Suppose that c | d < 0 for some c C ; then nc C C + for n N and the values nc | d cannot be bounded by 1. One deduces that 0 c | d 1 < b | d for all c C . As above, d C , and b C . 2
17.3.3
Norm and order

nN
With the notations of theorem 61, in particular C + :=
nC :
Definition 82 (Domain) The domain FinC of C is the vector space C + C + generated by C . Proposition 34 FinC = ( C ( C )) . Proof : If c C, d C ( C ), then c | d = 0, which subsists for c FinC , hence FinC ( C ( C )) . Conversely, if c FinC there is a vector d (FinC ) such that c | d = 0. But (FinC ) = C C ( C ), hence c ( C ( C )) . 2
In other terms, the domain of C is the orthogonal of the null space of C , that we shall soon dene and characterise. Definition 83 The domain FinC is equipped with a semi-norm x x
C C
and a preorder
= sup {| x | d | ; d C } y d C x | d y | d
C.
(17.11) (17.12)
Let =C be the equivalence associated with Proposition 35 The kernel 0C of the semi-norm modulo =C . Proof : Obvious.
C
is identical to the equivalence class of 0 2
In particular, FinC /0C is a partially ordered Banach space. Proposition 36 (i) C + is the set of positive elements w.r.t. (ii) 0C = C + (C + ) = C (C ).
C.
406

C
(iii) The unit ball w.r.t.
is (C C + ) (C + C ). 2
Proof : (i) and (iii) are respectively the cases b C + and b C C + of the proof of theorem 61. (ii) is immediate.
A few remarks of a slightly repetitive nature : (i) The partial order C is continuous w.r.t. the . C : if xn C yn and (xn ), (yn ) are Cauchy sequence for C with limits x, y , then x C y . (ii) If 0
C
y , then x
C. C
(iii) If x FinC , then there exist y, z
0 such that x = y z and y x .
What relation between norm and order for C and norm and order for C ? Nothing new in what follows, it is just a compilation : Equivalence x =C y x , y (x = C y x | y = x | y ) (17.13)
The introduction of the domain FinC , i.e., the fact of considering a partial, non reexive, relation, enables this symmetrical formulation. Positivity x C+ y (y ( C )+ x | y 0) (17.14)
The relation C is a preorder on the domain FinC . There is no standard terminology for such a relation, seen as a relation on E ; it enjoys weak reexivity : x yx xy y (17.15)
The next result generalises the decomposition of a hermitians as a dierences u = u+ u of two positive hermitians : Theorem 62 If x E one can nd x+ C + and x ( C )+ such that x = x+ x and x+ | x = 0 ; this decomposition is unique. Proof : Let x+ be the projection of x on the convex C , and let x := x x+ . One knows that x is the unique y such that y | x y y | z for all z C . This condition is easily transformed into y ( C )+ and y | x y = 0. 2
17.3. QUANTUM COHERENT SPACES Semi-norm x

C
407
= inf { ; y ( C )+
| x|y | y
C
C} C
(17.16)
It is however not the case that | x | y | x y Fin C . Proposition 37 If C D , alors :
for all x FinC ,
FinC FinD C D =C =D C D
The last inequality takes its full sense if one modies the notion of semi-norm so as to admit innite values.
17.3.4
Quantum coherent spaces
Let |X | be a nite dimensional (complex) Hilbert space ; one can apply what precedes to E := H(|X |), the space of hermitian operators on |X | ; they are those operators which are self-adjoint, i.e., such that h = h , i.e. : h(x) | y = x | h(y ) (17.17)
equivalently h(x) | x R. Remember that a hermitian is positive when h(x) | x R+ . Among positive hermitians, all the uu ; indeed every positive hermitian is of this very form, with u in turn positive, i.e., u = h. E is a real vector space, whose dimension can easily be computed : if |X | is of (complex) dimension n, then L(|X |) has the complex dimension n2 , hence, as a real vector space, the dimension 2n2 . Now every operator can uniquely be written as 2u = (u + u ) + i(iu iu), i.e. as h + ik , with h, k hermitian, which shows that the dimension of the real space H(|X |) is n2 . This space is equipped with a scalar product (denite positive bilinear form) : h | k := tr(hk ) (17.18)
which makes it Euclidian : tr(hk ) = tr(kh) = tr(hk ), tr(h2 ) > 0 for h = 0. Two hermitians h, k are polar when 0 h | k 1. Definition 84 (Quantum coherent spaces) A quantum coherent space (QCS) of support |X | is a a subset X H(|X |) equal to its bipolar.
408
Theorem 61 characterises QCS. One nds below two canonical examples. In both cases the order relation is the usual ordering of hermitians (h k i k h is positive) ; on the other hand, according to the case, one gets a norm of type sup or sum , consistently with the remarks on polarity made in section 15.A.1. Negative canonical : N is made of the positive hermitians of norm 1. N + therefore corresponds to positive hermitians ; on N + , N correspond to the usual norm . Positive canonical : P is made of the positive hermitians of trace 1. P + therefore corresponds to positive hermitians ; on P + , P coincides with the trace norm u 1 = tr( uu) ; in this case, h 1 = tr(h). Indeed, P = N : one uses |tr( uv )| u v 1 , and, for h, k 0, tr(hk ) = tr( hk h) = tr(( h k )( h k ) ) 0, as well as tr(uxx ) = u(x) | x .
17.4
17.4.1
Additives
A few reminders on quantum
Some basics of quantum mechanics, limited to nite dimension : (i) The state of a system is represented by a wave function, i.e., a vector x of norm 1 in a Hilbert space |X |. (ii) The measurement is represented by a hermitian on |X |. To say that the value of x w.r.t. is means that (x) = x. In other terms, there are practically never any value. Worse, if , do not commute, it is likely that they have no common eigenvector, i.e., x cannot have a value w.r.t. both of and , like in the celebrated uncertainty principle. Thus, the Pauli matrices, see infra, which measure the spin along the axes X, Y , Z , do not commute : if the spin is +1/2 along axis Z , it is completely undetermined along X . (iii) The measurement process is a Procustuss bed4 , which forces the system to have a value . In other terms, once the measurement performed, the wave function x is replaced with an eigenvector x of . This process is non-deterministic : if |X | is decomposed into a direct sum of eigenspaces
According to Plutarch, Procustus was a sort of old-style communist, who rubbed out the dierences between men by equating them to the length of his beds ; he was thence stretching the short and shortening the tall : thence the Procustuss bed .
4
17.4. ADDITIVES
409
of : |X | = |X | , so that x = x , then x is one of the components x , renormalised (multiplied by 1/ x ), and the probability of the transition x ; x / x is x 2 . This process is the reduction of the wave packet, reduction for short. (iv) In this approach, wave functions are dened up to a scalar of norm 1. For instance, the spin x of an electron is replaced with its opposite x during a rotation of angle 2 , without aecting the system. (v) The density matrices (or operators) are due to von Neumann ; they take into account the scalar indetermination of the wave function ; above all, they maintain the probabilistic aspect of the measurement process. A density operator is a positive hermitian of trace 1. Density operators form a compact convex set whose extremal points are of the form xx , where x is a vector of norm 1, i.e., a wave function, transformed into the orthogonal projection on the subspace it generates. During the measurement process, xx is replaced with x x : this density operator is a mixture , a convex combination of the projections x x / x 2 , whose coecients x 2 correspond to the probabilities of the possible transitions. (vi) This formalism is iterative, i.e., one can measure a density operator, not necessarily extremal. This means that, if one writes our density operator h under the matricial form (h ) w.r.t. the decomposition |X | = |X | (h L(|X | , |X | )), then reduction kills the coecients h outside the diagonal : after measurement, h becomes k = (k ), with k = h , k = 0 for = . (vii) The measurement process is irreversible : if u ; v by measurement, then tr(v 2 ) tr(u2). If X is of nite dimension n, tr(u2 ) will therefore vary, trough successive measurements, from the value 1 (extremal point xx ), up to 1/n : 1/n I , the tepid mixture, which conveys no information. Remark the nesse, the creativity, of this interpretation, compared to what logicians may have produced, typically those bleak Kripke models which put side to side arbitrary parallel universes.
17.4.2
Quantum booleans
Booleans form a system with two states, which suggests a two-dimensional space, hence 2 2 matrices. The classical viewpoint will soon be breathless for reasons of triviality, while the quantum viewpoint which really speaks of the same thing, a two-state system will appear of an incredible depth.
410
Commutative booleans There is little to say if one sticks to the
traditional
viewpoint : and 0 0 . 0 1
(i) The booleans true,false are represented by (ii) A diagonal matrix
1 0 0 0
0 , of trace 1, i.e., such that + = 1, and 0 positive, i.e., such that , 0, represents a probabilistic boolean. With a computer-science background, one can even admit + 1, which corresponds to a partial probabilistic boolean, true with probability , false with probability , undened with probability 1 ( + ).
None of this is that exciting. Now, remember that matrices are operators written w.r.t. a certain base ; and let us imagine that for the reason you prefer say, a travel accident that would have distorted the gyroscopes the basis has been lost . The booleans are still there, but one can no longer read them ! Then this bleak bureaucracy starts to live ! One is lead to studying hermitians in dimension 2, which leads to the space R3 , not to speak of space-time ! Space-time Any hermitian writes h = 1/2 t + z x iy , i.e., t.s0 + x.s1 + y.s2 + z.s3 , x + iy t z with t, x, y, z real, where the si are the Pauli matrices : 1/2 1 0 0 1 1/2 0 1 1 0 1/2 0 i i 0 1/2 1 0 0 1 (17.19)
The time t is the trace, t = tr(h). Computation of a determinant yields 4 det(h) = (t2 (x2 + y 2 + z 2 )), the square of the pseudo-metrics. Also remark that tr((t.s0 + x.s1 + y.s2 + z.s3 )(t .s0 + x .s1 + y .s2 + z .s3 )) = tt + xx + yy + zz . For 1 i = j 3, one gets the anti-commutations si .sj + sj .si = 0. In order to characterise positive hermitians, remember that any hermitian is diagonalisable : modulo a change of base, i.e., a unitary transformation u, 0 uhu = , with , R ; h is positive i , 0. In other terms, 0 the condition condition det(h) 0 (vectors in position time ) caracterises hermitians which are either positive or negative. Positivity requires the additional condition : tr(h) 0 : which corresponds to the cone of the future : t x2 + y 2 + z 2 . The most general transformation preserving positivity is of the form h ; uhu , with det(u) = 1, i.e., u SL(2) : this is the positive Lorenz
17.4. ADDITIVES
411
group, which preserves both the pseudo-metrics and the future. While here, the inverse of u SL(2) is given by : a b c d
1
d b c a
(17.20)
Hence, inversion extends into an involutive anti-automorphism of the algebra M2 (C) of 2 2 matrices. In terms of space-time, this anti-automorphism replaces the spacial coordinates with their opposites. The group SO(3) of rotations, which only act on space, corresponds to preservation of time, i.e., of trace. Since tr(uhu ) = tr(u uh), one sees that they are induced by unitary u. In other terms, SO(3) admits a (double) covering by SU(2), the group of unitaries of determinant 1, whose general a b form is , with aa + b b = 1. The rotations of axes X, Y , Z and angle b a are induced by the eisk , respectively : cos /2 i sin /2 i sin /2 cos /2 cos /2 sin /2 sin /2 cos /2 ei/2 0 i/2 0 e (17.21)
It is obviously a heresy to divide an angle by 2, since there are two solutions. This is why the covering is double ; this is also why a rotation of angle 2 acts on the spin seen as a wave function as a multiplication by 1. This corresponds to the possibility of replacing u by u in h ; uhu, and that one cannot continuously choose between both determinations just as one cannot continuously determine a complex square root. Quantum booleans
Classical booleans are the orthogonal projections of two 1-dimensional subspaces distinguished by the matricial representation. A quantum boolean will simply be a 1-dimensional subspace. This approach refuses from the start any distinction between true and false : if E is a boolean, its negation is E , period ! Also remark that, for foreseeable reasons of commutation, it will not be possible to construct convincing binary connectives. It remains to determine the 1-dimensional subspaces, i.e., the matrices of orthogonal projections of rank 1. They are the matrices of trace 1 and determinant 0, i.e., the points of space-time t.s0 + x.s1 + y.s2 + z.s3 , such that t = 1 and x2 + y 2 + z 2 = 1, and are therefore in 1-1 correspondence with the sphere S 2 . What we just called quantum boolean is known in physics as the spin of an electron. The measurement of the spin along the axis say Z is given by the Pauli matrix s3 , whose eigenvalues are 1/2. The value +1/2
412
is obtained on the subspace corresponding to
1 0 (spin up along Z ), the 0 0 0 0 value 1/2 (spin down along Z ) corresponding to . 0 1 Probabilistic quantum booleans
The most general case is that of a convex combination of quantum booleans, what corresponds to a positive hermitian of trace 1, a density matrix . One can diagonalise a hermitian in appropriate orthonormal base ; in which sense is this unique ? (i) 1/2 0 is diagonal in all orthonormal bases ; no form of unicity. 0 1/ 2
(ii) Outside this case, our boolean writes as b + (1 )c, where b, c are orthogonal booleans and 0 < 1/2, and this, in a unique way. Reduction occurs when one measures a boolean, what corresponds to the measurement of a spin. One must specify a base, which diagonalises the measurement operator. One writes our hermitian, quantum boolean, probabilistic or a b . Once the measurement done, it becomes : not, in this base, namely : b c a 0 , i.e., true with probability a, false with probability c = 1 a, along the 0 c axis true/false corresponding to our base. Negation Choosing an orthonormal basis consists in choosing two subspaces of dimension 1, i.e., two quantum booleans et I , whose space-time coordinates will therefore be (1, x, y, z ) et (1, x, y, z ). The vectors A = (x, y, z ) and A correspond to the two possible sense on the same axis (spin up, spin dwon). The symmetry w.r.t. the origin corresponds to the anti-automorphism a b d b ; of the algebra M2 (C). This transformation corresponds c d c a to negation. Since symmetry w.r.t. the origin is of determinant 1, it is not in S0(3), and is not induced by an element of SU(2) ; by the way, the elemnets of SU(2) induce automorphisms, not anti. Binary boolean connectives Negation, truly involutive operation, does not involve reduction. Which is no longer the case for binary connectives :
17.4. ADDITIVES
413
(i) One cannot combine 1-dimensional projections which do not commute so as to produce another projection. (ii) Common sense says that, if one cannot tell the truth from the false, it will be even more dicult to tell a conjunction from a disjunction. Hence binary connectives are probabilistic : they yield a probabilistic boolean even when the arguments are pure . Moreover, they depend upon a base and a b a b a + ca cb an order of evaluation, for instance : . The := b c b c c b cc rst argument is reduced in the base : true with probability a, in which case the 1 0 a b answer is . , false with probability c, in which case the answer is b c 0 0 There is a symmetrical choice, reducing the second argument. And also Jivaro a + ca 0 choice , which reduces both : , which is symmetrical, since 0 cc a + ca = a + ca = a + a aa .
17.4.3
Additives and quantum
Plus and With Definition 85 (Additives) If X, Y are QCS of respective supports |X |, |Y |, one denes X Y and X & Y , QCS of base |X | |Y | : X Y X &Y = {h (1 )k ; h X, k Y, 0 1} = {g ; |X |g |X | X, |Y |g |Y | Y } (17.22) (17.23)
The subspaces |X |, |Y | have been identied with their orthogonal projections. Proposition 38 and & are swapped by negation. Proof : Since h k | h k = h | h + k | k . 2
Neither X Y , nor X &Y are norms. This denition mistreats hermitians that cannot be written h k . W.r.t. a block decomposition, any hermitian on h u |X | |Y | writes : , with h, k hermitian. If u = 0, this takes an innite u k norm in X Y (if one prefers, is not in FinX Y ). A contrario, its norm w.r.t. 0 u X & Y does not depend on u : the kernel 0X &Y contains all the . u 0
414 Dimension 2
If |X | is of dimension 1, then H(|X |) is of dimension 1 (isomorphic to R) ; the two canonical QCS of section 17.3.4 coincide, to yield a QCS 1, corresponding to the segment [0, 1] of R, ordered and normed naturally . In dimension 2, H(X ) is of dimension 4, with several natural choices : Spin : the positive canonical. The elements of Spin are the positive hermitians of trace at most 1, the partial probabilistic quantum booleans . Spin : the negative canonical. The elements of Spin are positive hermitians of norm (in the usual sense) at most 1, the anti-booleans . Bool : the of two copies of 1. The QCS 1 1 is made of all matrices a 0 such that 0 a, c a + c 1, i.e., of partial probabilistic 0 c booleans. It is a subset, a subtype of Spin. Bool : the negation of the previous, i.e., 1 & 1. It is made of the matrices a b such that 0 a, c 1. b c Our construction of Bool depends upon a 1-dimensional subspace corresponding to true , here, the one encoded by Z . Which means that, given a vector A S 2 , there is a QCS of the booleans of axis A , noted BoolA . Proposition 39 Spin = AS 2 BoolA . Proof : Of course, BoolA Spin. Conversely, if h Spin, it can be diagona 0 , with 0 a, c a + c 1, w.r.t. a certain orthonormal base alised as 0 c e, f . If A S 2 corresponds to e, then h BoolA . 2 Corollary 39.1 Spin = AS 2 BoolA . Reduction : a discussion I said that we are seeking a logical explanation of the quantic. This being said, what we did enables one to clarify the question of reduction. In the next section, one deals with multiplicatives, thence with linear implication. In particular, one will be able to transform a boolean h Spin into something else, by making use of an element of a QCS Spin . . ., then
17.5. MULTIPLICATIVES
415
transform the result by means of another implication. . . Some of these transformations will behave like negation, they will be of wave style , other like the binary connectives, will make use of reduction, they will be of particle style . By the way, one will see that logical operations are rather of the particle style, with a noticeable exception, the non-etaspansed identity axiom, rather of style wave . In which measure is reduction subjective ? Let us oer us an impossible hypothesis : let us assume this process of transformation of a boolean completed, i.e., that, in this succession of implications, one was eventually able to close the system . Which means that everything ended with a last implication, with values in 1. If I compose all my implications, I see that this sequence of transformations which eventually closes the system is nothing but an anti-boolean k Spin. The result is objective : h | k = tr(hk ). But the choice of k (the transformations, the measurements made on h) is very subjective. We are not neutral, since on the side of k . Might as well, since we stand on the side of k , let us diagonalise it in an appropriate base e, f . a 0 a b 0 , then , and so h | k = a + c . If h = Then h = ,k = 0 c 0 b c h | k = h | k , i.e., it is as if we had reduced h. It could be the case that we know that f is a boolean in a certain base (e.g., if f comes from the measurement of a spin ). One selects this base, and a 0 h= ,k = , and one can then write h | k = a + c . In 0 c 0 this case, one reduced the observer k into k = in such a way that 0 h|k = h|k . Finally, reduction is a very subjective matter.
17.5
17.5.1
Multiplicatives
Linear functionals
Theorem 63 (Folklore) Let |X |, |Y | nite dimensional Hilbert spaces. Then L(L(|X |), L(|Y |)) L(|X | |Y |). Proof : The complex vector space L(|X |) is generated by rank 1 endomorphisms : xw (y ) := y | w x. If L(L(|X |), L(|Y |)), one denes L(|X | |Y |) by : (x y ) | w z = (xw )(y ) | z (17.24)
416
Conversely, given L(|X | |Y |), if f L(|X |), one denes ()f L(|Y |) by : (()f )(y ) | z = tr( (f yz )) (17.25) hence () L(L(|X |), L(|Y |)). 2 Corollary 39.2 If H(|X ||Y |), if f H(|X |), then ()f H(Y ). The function ; () is a bijection between H(|X ||Y |) and the set of linear functionals from H(|X |) into H(|Y |). Proof : One easily sees that ( )f = (()f ) , hence a hermitian sends hermitians to hermitians. Conversely, if is a linear functional from H(|X |) to H(|Y |), then uniquely extends into a C-linear functional from L(|X |) to L(|Y |) : (u) := 1/2((u + u ) + i(iu iu)). The C-linear functionals thence obtained are hermitian, i.e., enjoy (f ) = (f ) , and are therefore in bijection with the hermitians of H(|X | |Y |). 2 The essential property of () is summarised by the equation : tr((()f ) g ) = tr( (f g )) (17.26) Which is reminiscent of stability, introduced in section 8.2.6. It was a matter of linearity in the absence of true linear operations, since one could only use unions and intersections ; the way to true linearity is marked with the reformulations of section 9.1, and we eventually nd it again under its literal form, with true sums, true coecients. Unfortunately, just like Moses passing away at the gates of Israel, the story of coherent spaces stops here : they will not survive, under this ambitious form, to innite dimension, see section 17.6.1.
17.5.2
A few examples
Example 3 If E H(E E ) is such that (x y ) = y x (the ( )(xw )(y ) | z
twist ), then
= (x y ) | w z = yx|wz = y |w x|z = (xw )(y ) | z
(17.27)
hence ( )(xw ) = xw . By linearity, ( )f = f . Example 4 More generally, let u be a linear map from E to F . Then u u sends E F to F E , and if EF is the twist from F E to E F , then U = (u u ) H(E F ). It goes without saying that (U )f = uf u.
17.5. MULTIPLICATIVES
417
Example 5 Let I = E + F be a decomposition of the identity into orthogonal projections (subspaces). Then R = (E E + F F ) acts as follows : (R)f = Ef E + F f F . R is the typical reduction, which kills the nondiagonal blocks Ef F and F f E of f . One can ask what the identity map of E F is doing. One easily sees that (IE F )u = tr(u) IF . Not very inspired. . . But this will be used in our last example : Example 6 a b c b If E is of dimension 2, then (IE E E ) = , i.e., acts as b c b a the negation. IE E E = 2 , where is the orthogonal projection on the antisymmetric subspace of E E , i.e., the 1-dimensional space made of the x y y x.
17.5.3
Connectives
Definition 86 (Multiplicatives) If X, Y are QCS, one denes the QCS X Y , of support |X | |Y |, as the set of all sending X into Y : X Y = { ; f X ()f Y } (17.28)
X Y might as well be dened by : X Y = { ; g Y g () X } (17.29)
and also as {f g ; f X, g Y }. This last expression shows that X Y Y and is a QCS. From this, one denes X Y = X X Y = {f g ; f X, g Y }. As usual, is commutative, associative and distributes over & (up to isomorphism). Observe that multiplicatives force us to depart from the standard ordering of hermitians. For this, one will suppose that X, Y are positive canonicals, e.g., X = Y = Spin. Then X Y will consider as positive any hermitian sending the (true) positive to the (true) positive. Hence, the twist which behaves as an identity map ; but is a symmetry, a hermitian not at all positive ! Hence X Y is more liberal as to positivity than what one could expect. Dually, X Y is more restrictive, here one is more positive than the King ! The positive cone of X Y is the closure of the set of sums i fi gi , fi , gi 0. Most of (truly) positive hermitians of |X | |Y | are not of this form, typically the zz , when z is not a pure tensor.
418
17.5.4
and reduction
We already met , see, e.g., section 7.4.2. Can one imagine a bleaker, a more bureaucratic topic ? One would like so much to see this principle wrong ; thence one observed that ludic faithfulness stumbled grosso modo on this idiocy, see the discussion in section 14.B.4. The literature on this topic is depressing. One will make an exception for Plotkin who was the rst to refute this doohickey, but in such a fabricated way, that is the case to say what kills me reinforces me : at least he was the rst one to do so. But what followed, my God. . . ! A PhD not long enough, at the time when everything is measured in number of pages, of publications, etc. and thats it : you redo everything with . Dear , the providence of PhD students : 100% transpiration how painful it can sometimes be 0% inspiration. This dear , precisely, one will brush it the wrong way, not necessarily once for all, but in a simple and earnest way. One will, in the case of a Plus C = A B , dierenciate the native identity from its etaspansion. Intuitively the dierence is very simple : the identity is (x) = x, it recopies an object without really caring about ; if one thinks of delocation, it is a genuine wavy operation. Etaspansion of the same is based upon the idea that one is either of type A or of type B : the etaspansed identity asks the object whether it is of type A or B , and whatever the answer, recopies it identically : it is a cop who asks for your identity papers before letting you pass. The commutative, set-theoretic, world cannot separate the real identity from its etaspansion, the inquisitive , essentialist, identity ; it leaves no real room for the potential, reduced to a list of possibilities. In quantum, this is dierent ; etaspansion is a measuring process followed by a trivial reconstruction : it is therefore a plain reduction. Without working too hard, let us assume that A, B have dimension 1 supports, and let us consider : The twist : The generic identity of a space |C | of dimension 2. Which writes : 1 0 = 0 0 0 0 1 0 0 1 0 0 0 0 0 1
(17.30)
in any base e e, e f , f e, f f of |C | |C |. The etaspansed twist : This is the gluing of two identities, that of A and that of B . W.r.t. a well-dened base, corresponding to the decomposi-
17.5. MULTIPLICATIVES tion of |C | in direct sum, this yields : 1 0 0 0 = 0 0 0 0
419
0 0 0 0
0 0 0 1
(17.31)
a b c 0 Equivalently ( ) = : measures the spin along the axis Z , b c 0 a then swaps it. While here, the matrix represents the double of a projection, the one projecting on the EPR state , by the name of a famous paradox imagined by Einstein to refute quantum non-determinism : it is about the space of the x y y x, a space of dimension 1. Two correlated particles, simultaneously measured at a big distance are supposed to yield opposite spins, in apparent contradiction with relativity ; but the EPR paradox has been veried. For us, we shall say that it is a Par , which works as the communicating vessels of gure 10.2.2 : if one destroys (measures) one side, one nds it again reversed on the other side.
c b a b = ; it can is such that ( ) b c b a 0 0 0 0 1 0 = 0 0 1 0 0 0
a b a b These two things are necessarily distinct : ( ) = is the true b c b c a b a 0 identity. On the other hand () = is only an identity of Procusb c 0 c tus . It is the identity for those matrices which are already of the appropriate a 0 . For those which do not t in this setting (in the essence), it form 0 c severs the coecients outside the diagonal. Indeed, is only the reduction of the wave packet, corresponding to the measure of the spin along the axis Z . In logic, only the identity enjoys etaspansion. What is not the case here, since negation can be etaspansed too : 0 0 0 0 0 1 1 0 = (17.32) 0 1 1 0 0 0 0 0 be etaspansed into : 0 0 0 0
(17.33)
420
17.6
17.6.1
Discussion
Innite dimension
This cannot be convincingly extended to innite dimension. (i) The pivot of the construction is the trace, which becomes problematic in innite dimension. One can dene the two-sided ideal of trace-class operators, but this ideal contains no invertible operator, hence nothing like the twist. (ii) Some von Neumann algebras (those of type II1 ) have a trace. One could try to dene QCS whose support would be of type II1 . Restricted to PCS, this would consist in replacing discrete supports with the segment [0, 1], equipped with Lebesgue measure, nite sums becoming integrals. This stumbles on the identity function, which still would be represented by the characteristic function of the diagonal. . . which is of measure zero, hence does not exists . This impossibility extends a fortiori to the non-commutative case. (iii) Technically, the problem is caused by the tensor products of Hibert spaces. One should interpret the identity by something lighter, namely direct sums : this is what geometry of interaction will do. The failure of level 2 must be related to the impossibility of a convincing topological treatment at this level. Although of an innitely higher quality, level 2 reproduces the basic error of Kripke models, i.e., reduces the potential to the set of all potentialities. Thus, a proof of A B which consists in a proof of A and a proof of B , should be represented by a direct sum ; while every concrete use combine both proofs in a rather unpredictable way, which requires a tensor product to represent all possibilities. What does not withstand the innite limit, is thence actualisation , i.e., the reication of possibles, a common attitude to levels 1 and 2.
17.6.2
Operators vs. sets
The most impressive foundational endeavour of the turn of the XXIth century owes nothing to logic : it is the Non Commutative Geometry of Connes, [13]. It is an anti-set-theoretic rereading based upon the familiar result : A commutative operator algebra is a function space. Typically, a commutative C -algebra can be written C(X ), the algebra of continuous functions on the compact X . Connes proposes to consider non
17.A. INITIATION TO C -ALGEBRAS
421
commutative operator algebras as sorts of algebras of functions over. . . non existing sets. An impressive blow against set-theoretic essentialism ! What does this changes for logic, a priori far astray from considerations internal to geometry ? I would say that this changes our ideas of nite set, of point, of graph, etc. The commutative, set-theoretical, world appears as a vector space equipped with a distinguished base. All operations are organised in relation to this base, in particular they can be represented by linear functions whose matrix is diagonal in this base. The non-commutative world forgets the base ; there is still one, but it is subjective, the one where one diagonalises the hermitian operator one uses : his set-theory, so to speak. But, if two hermitians f and g have non commuting set-theories , one sees that f + g has a third set-theory bearing no relation to the previous. Which enables one to come back for the last time to the unfortunate quantum logic. Closed subspaces are perfectly enough to speak of whatever we want in particular of the set-theory relative to a given hermitian. But this does not socialise not to speak of the gigantic mistake of orthomod0 0 ular lattices , where the Hilbert space has disappeared. Indeed, 0 1 1/2 1/2 et have set-theories corresponding to the bases {X, Y } and 1/2 1/2 { 2/2(X + Y ), 2/2(X Y )}, but what is the set-theory of their sum 1/2 1/2 ? The solution does not belong in lattice theory, it involves the 1/2 3/2 solution of an algebraic equation, 2 2 + 1/2 = 0. In other terms, the order structure of subspaces does not socialise with the basic quantum operation, superposition.
17.A
Initiation to C -algebras
What follows addresses to the reader that would have forgotten the bases of functional analysis or to the one that would seek his way through a classic book on the topic, e.g., [59]. This sort of culture est primal and could by no way be replaced with category-theoretic ruminations loosely inspired from linear or operator algebra, e.g., traced monoidal categories.
17.A.1
Hilbert spaces
The most popular Hilbert space is :
422
Definition 87 ( 2 ) 2 is made of square summable sequences of complex numbers, equipped with the form : (an ) | (bn ) := an bn (17.34)
n
This is more or less the sole example : indeed, any Hilbert space admits an orthonormal base (ei )iI . Which makes it isomorphic with the space 2 (I ) of the square summable families indexed by I ; and remember Parsevals formula : x
2 2
=
i
| x | ei |2
(17.35)
is the most frequent case, corresponding to I denumerable ; the other important case is I nite, what corresponds to this very chapter. The use of complex numbers comes from the spectral theory, see infra : the spectrum must be non empty (in nite dimension, the characteristic polynomial must have a root). From that, the form is not quite bilinear, but only x | y . Which is enough sesquilinear (sesqui = one and a half) : x | y = to ensure x | x R (in other terms x | y = y | x ). Finally, the form enjoys x | x > 0 for x = 0. Which yields the celebrated : Theorem 64 (Cauchy-Schwarz) | x | y |2 x | x y | y , equality occurring only in case of colinearity. Further, 2 , equipped with the norm x := x | x 1/2 is a Banach space, i.e., is complete. This being said, x | x 0 is enough in practice : one then quotients by the kernel of the form ; similarly, if the space is not complete, . . . one completes it ! These two operations are known as the process of separation/completion of a pre-Hilbert space. One can sum Hilbert spaces : the algebraic direct sum H K equipped with the form : x y | x y := x | x + y | y (17.36) In particular, x y 2 = x 2 + y 2. The tensor product is obtained by equipping the vector space generated by the formal tensors x y, x H, y K with the sesquilinear form dened on pure tensors by : x y | x y := x | x y | y (17.37)
The Hilbert space H K is dened as the separation/completion of this preHilbert space. Observe that the tensor product does not factorise bilinear maps, but only some of them (styled Hilbert-Schmidt) : the form | cannot be factorised through a linear map from H H into C.
423
17.A.2
The dual space
A Hilbert space is a normed complex vector space, i.e., a Banach space. Its dual is the set of continuous linear forms, equipped with the norm := sup {|(x)|; x 1}. A Banach which is its own bidual is styled reexive, which is an exceptional situation, at least in innite dimension. Hilbert spaces are the main examples of reexive Banach spaces. The main tool on Banach spaces is the Hahn-Banach theorem, which takes various forms, e.g. : Theorem 65 (Separation) If C is a closed convex set in a Banach space and x C , then there is a continuous linear form and a real number r such that ((C )) r < ((x)). This theorem is already very interesting in nite dimension. In this case, all forms are continuous. A Banach space E can be equipped with a weakened topology : xi x i (xi ) (x) for all continuous linear forms on E . This topology has an extraordinary property : the unit ball of E is weakly compact. Indeed, the weakened topology is nothing but convergence coecientwise . W.r.t. weakened topologies, there are many compacts sets. Theorem 66 (Krein-Milman) Any compact convex set is the closed convex envelope of its extremal border, i.e., of the set of its extremal points. Remember that an extremal point of the convex set K is a point x K which cannot be written as the barycenter of two other points of the convex. The extremal frontier of a disk is the frontier circle, the extremal frontier of a convex polygon is its vertices. This theorem applies to density operators, i.e., to positive hermitians of trace 1 (w.r.t the weakened topology when the dimension is innite). It simply says that any hermitian can be approximated by barycenters of extremal hermitians by projections of rank 1. What is implemented in nite in an orthonormal base dimension by diagonalisation 1 0 . . . 0 0 2 . . . 0 (ei ) under the form . . . . . . . . . . . . ; the operator can thence be written 0 0 . . . n as the barycenter of the projections associated with the ei with the respective weights i . Some Euclidian geometry : let us consider the triangle of vertices 0, x, y and the median starting with 0, in other terms the vector (x + y )/2. An immediate computation yields : x
2
+ y
= 2( (x y )/2
+ (x + y )/ 2 2 )
(17.38)
424
This equality enables one to majorise, in certain cases, the norm of the third edge, x y . Thus, if E H is a non-empty closed convex set, (17.38) enables one to project on the convex E : indeed, if (xn E ) is such that xn converges to inf { x ; x E }, then it is a Cauchy sequence. Proposition 40 The minimum inf { x ; x E } is reached in a unique point of E . This point is also the unique e E such that e | e f is negative for all f E . What applies, mutatis mutandis to the projection of an arbitrary point on a closed subspace E ; hence, let the map obtained in this way. One sees that is linear and that 2 = ; the range of is E , its kernel is E , and the projection associated with E is I . These spaces are supplementary, i.e., each vector of H uniquely writes as x = e + e , e E, e E , i.e., H is isomorphic to E E . If e H, x x | e is a continuous linear form : by Cauchy-Schwarz | x | e | e x , equality being reached with x = e, hence e = e . Conversely, any linear form continuous on H is of the form e for a wellchosen e necessarily unique : it suces to consider {x; (x) = 1} and apply projection techniques. Hence the dual H of H is canonically isomorphic to H by means of the map b b ; but, pay attention to the scalar multiplication : ! (b) = b The linear forms b induce the weak topology on H, hence : Proposition 41 The unit ball of H is weakly compact. Weak convergence does not imply norm convergence, but in the case : Proposition 42 If xi x weakly and xi x , then xi x Among the applications of weak compacity : Proposition 43 The image under an operator u B(H, K) (see infra) of the unit ball of H is norm-closed in H. Proof : Indeed, u is weakly continuous, hence the image B of the unit ball is weakly compact and therefore closed. It remains closed in any stronger (with more closed sets) topology. 2
normwise .
425
17.A.3
Spectral theory
If H and K are Hilbert spaces one notes B(H, K) the set of all bounded linear maps from H into K, i.e., such that the norm : u := sup { u(x) ; x H, x 1} (17.39)
is nite. u then induces a linear map u from the dual K of K to the dual H of H. Now K and H are isomorphic to K and H, in other terms u denes a linear map (the adjoint ) u from K to H ; u is indeed dened by the adjunction : u(x) | y = x | u (y ) (17.40) The most important case is that of H = K ; one simply notes B(H) the Banach algebra thence obtained. The adjunction is an involution of B(H), satisfying : , (uv ) = v u , etc., and above all : (.u) = .u uu = u
2
(17.41)
An involutive Banach algebra enjoying (17.41) is called a C -algebra. The most natural example of a C -algebra is B(H). A sub-algebra of B(H) closed under the norm and self-adjoint (closed under adjunction) is the most general form of a C -algebra : this is the contents of the GNS theorem, see infra. There is a more elementary example, the algebra C(X ) of complex continuous functions on a compact X , equipped with multiplication, adjunction being conjugation, all of this dened pointwise, with : f = sup {|f (x)|; x X } (17.42)
The peculiarity of C(X ) is to be commutative ; it is indeed the most general form a commutative C -algebra, see infra. In a Banach algebra C , one can dene the spectrum of an element u : Definition 88 If u C , Sp(u) is the set of C such that u .I is not invertible. If H is nite dimensional and u B(H), Sp(u) is the set of eigenvalues of u. If X if compact and f C (X ), then Sp(f ) is the range of the function f . Some results involve Sp (u) := Sp(u) {0}. For instance : Proposition 44 Sp (uv ) = Sp (vu) The fundamental result is : Theorem 67 Sp(u) is a non empty compact set.
426
The theorem has an eective version : one can compute the size of the spectrum, in this case, give an explicit formula for the spectral radius r (u) := inf {r ; z Sp(u) |z | r } (by compactness, this value is eectively taken) : r (u) = lim un 1/n = inf un 1/n (17.43) In particular, if u is nilpotent, its spectral radius is null ; what is well-know in nite dimension, remember the strictly triangular Jordan matrices and their null eigenvalues. In a C -algebra, the norm of a normal, e.g., hermitian (see infra ), operator equals its spectral radius. Proposition 45 Let P be a complex polynomial ; then Sp(P (u)) = P (Sp(u)). Corollary 45.1 The equation uv vu = I has no solution among bounded operators. Proposition 46 Sp(u ) = Sp(u) ; if u is invertible, then Sp(u1 ) = Sp(u)1 .
17.A.4
Taxonomy of operators
The elements of a C -algebra, hence the operators on a Hilbert space, are mainly classied according to the relation with their own adjoint. Here are the most typical cases : Normal : is said of an operator commuting with its adjoint : uu = u u. Then, the C -algebra generated by u is commutative and u enjoys a sort of diagonalisation . Among normal operators one nds hermitians and unitaries. Unitary : is said of an operator u admitting u as inverse, in other terms, such that uu = u u = I . They correspond to the isometries of H, since u(x) | u(y ) = x | u u(x) = x | y , and therefore form a group. The spectrum of a unitary is included in the circle T = {z ; |z | = 1}. This is obvious since u = 1 (from uu = u 2) hence Sp(u) is included in the unit circle D ; the same is true of Sp(u ) = Sp(u)1 , which shows that Sp(u) D D 1 = T . Hermitian : (or self-adjoint) is said of an operator u equal to its adjoint, in other terms, such that u(x) | x is real for all x. The spectrum of a hermitian is real and the extremal bounds of its spectrum are the reals sup { u(x) | x ; x = 1} et inf { u(x) | x ; x = 1}. The most typical hermitian (indeed, every hermitian is of this form) is a sum u + u .
427
Symmetry : is said of a unitary hermitian, i.e., such that u = u = u1 . The spectrum is included in {1, +1}, and one can indeed diagonalise u as the dierence of the projections (see infra ) (I + u)/2 (eigenspace of +1) and (I u)/2 (eigenspace of 1). Projection : is said of an idempotent hermitian (hence positive) : u = u = u2 . The spectrum is included in {0, +1}, and u corresponds to the orthogonal projection on a closed subspace, the range E of u, what one notes u = IE . Positive hermitian : is said of a hermitian such that u(x) | x 0 for all x. Positive hermitians are of extreme importance, since the order structure of R compensates the deciencies of topology, for instance in questions of convergence of series. Positive hermitians have a spectrum included in R+ . The typical positive hermitian is a product uu ; indeed, indeed, they are all of that form, and one can even suppose u hermitian, and, better u in turn positive : the capital fact is that a hermitian has a square root. The standard analogy is as follows : operators are a non-commutative version of their spectrum, in other terms, hermitians are the non-commutative reals , the unitaries playing the part of the complex arguments ei . By the way, the polar decomposition expresses any operator as the product u = u u of a module (the positive hermitian u u) and an isometry (partial, however). Positive hermitians are so important that we have better coming back to them in details : Definition 89 (Positive hermitians) u is positive when it is of spectrum included in R+ . Hermitians are ordered (see theorem 68) by u v i v u is positive. Theorem 68 Let C + be the set of positive hermitians of C ; (i) C + is a closed convex cone in C ; (ii) If u, u C + , then u = 0. Theorem 69 A hermitian u can be expressed as the dierence u+ u of two positive hermitians such that u+ u = u u+ = 0. This expression is unique and moreover u = sup( u+ , u ). This by no way establishes a lattice structure on hermitians ; indeed, in B(H), two hermitians have a supremum exactly when they are comparable !
428
Theorem 70 If u C , the following properties are equivalent : (i) u C + ; (ii) u = v 2 for a v hermitian ; (iii) u = v v for an arbitrary v . Moreover in (ii), v can in turn be chosen positive, in case the choice is unique. When C is of the form B(H), one can add the following equivalents : (iv) u = v v for an operator v in some B(H, K) ; (v) u(x) | x 0 for all x in H. Corollary 70.1 If u, v C + commute, then uv C + . If u is positive and v is arbitrary, v uv is positive. Proof : Always think of the square root ! Thus, v uv = (v u)( uv ) = ( uv ) ( uv ). 2
17.A.5
Commutative case
If u is hermitian, its spectrum is real ; moreover, if P is a polynomial, one veries that P (u) = sup {|P (x)|; x Sp(u)}. Using Stone-Weierstrass, this equality enables one to dene f (u) for any continuous function f C(Sp(u)) from Sp(u) to C ; for instance, if u is positive, one denes u. One just dened an isometry between C(Sp(u)) and the C -algebra generated by u ; u corresponds to the canonical inclusion from Sp(u) into C. This easily extends to the case where u is normal : one has still an isomorphism between the C -algebra generated by u and C(Sp(u)). This can be seen as a sort of diagonalisation. More generally, any commutative C -algebra is isomorphic to C(X ) for a certain compact set X . If an operator is normal, it is enough to look at his spectrum to know whether it is hermitian, unitary, etc. For instance, one says that u not necessarily normal is a partial isometry when uu is a projection : Proposition 47 If uu is a projection, then uu is a projection too. Hence u is a partial isometry. Proof : Using proposition 44, one sees that Sp(u u) {0, 1} ; since u u is hermitian, one can apply spectral calculus. The canonical inclusion of Sp(u) in C enjoys f 2 = f , hence u2 = u. 2
429
Definition 90 (Domaine, Image) If u is a partial isometry, one styles domain (resp. image) of u the closed subspace associated with the projection u u (resp. uu). In fact u establishes an isometric bijection between its domain and its image. One will similarly toy and prove that, if u is normal and Sp(u) {1, 0, 1}, then u is a partial symmetry, i.e., that u is a hermitian such that u2 is a projection (equivalently : u3 = u).
17.A.6
Finite dimension
If one extracts from what precedes what concerns nite dimension, one gets : Normal, e.g., hermitian, unitary, operators are diagonalisable in an orthonormal base. Which means that a hermitian or unitary matrix writes UDU for well chosen unitary U and diagonal D . A hermitian operator corresponds to the case where D has real entries ; a positive operator corresponds to positive entries. A unitary operator diagonalises through a D whose diagonal entries lay in the unit circle. This is obtained through C(X ) : in nite dimension, the compact X is a nite set (with the discrete topology). In nite dimension, the trace is dened as the sum of all diagonal coecients of any matrix representation. This does not depend on the base (on the matrix), this is cyclicity : tr(uv ) := tr(vu) (17.44) which yields independence from the base : tr(u) = tr(v 1 vu) = tr(vuv 1). While here, a formula : every matrix is triangulable, i.e., let us establish 11 12 . . . 1n 0 22 . . . 2n write M = . . . . . . . . . . . . in a base (orthonormal if one wants it). In 0 0 . . . nn e11 12 . . . 1n 0 e 22 . . . 2n particular, eM = . . . . . . . . . . . . . Hence : 0 0 . . . e nn Proposition 48 det(eM ) = etr(M ) . := tr( uu) denes a norm, the
Still in nite dimension, the quantity u trace-norm. One easily sees that : |tr(u)| u 1 uvw 1 u v
(17.45)
430
Indeed the two norms are put in duality by : u = sup {|tr(uv )|; v 1 1} v 1 = sup {|tr(uv )|; u 1} (17.46)
I 1 = n, the dimension of the space ; it tends to innity when the dimension increases. In innite dimension, one must therefore restrict one to trace-class operators, which form a two-sided ideal, thence do not include invertible operators like the twist . Trace-class operators do not seem to be suited for the interpretation of proofs.
17.A.7
The GNS construction
Short for Gelfand-Neumark-Segal ; it is the result stating that any C -algebra is subalgebra of some B(H). This result mainly rests on the possibility of transforming a C -algebra into a Hilbert space, this what is called the GNS construction. For this one starts with a state , i.e., i.e., a linear functional from the C -algebra C into C satisfying (I ) = 1, (u) 0 for u positive. A state is hermitian : (u ) = (u), hence, if u is hermitian (u) is real. Above all, a state is monotonous : if u v are hermitian, then (u) (v ). Given a state on C , the formula : u | v := (v u) (17.47)
denes a pre-Hilbert space, whose separation/completion is denoted C . Every f C operates (on the left) on the pre-Hilbert space C by : f (u) := f u, and one sees that f (u) 2 = (u f f u) f 2 (u u) (this uses f f f 2 I , etc. and monotonicity). One sees that the left action of f can be extended into an operator of norm at most f on the separation/completion C . One thence obtains a morphism from the C -algebra C into the space B(C ), what one calls a representation. One must sum up many GNS representations so as to eventually get one which is faithful, i.e., does not diminish the norm.
Chapter 18 Yet more proof-nets !

We shall revisit proof-nets, especially the correctness criterion. What will eventually lead us to Geometry of Interaction (GoI. What follows essentially belongs in the early times of linear logic, and was expounded in a paper impossible to nd, [35].
18.1
Duality and correctness
We begin with a simplied version of [35], adapted to the Danos-Regnier version of correctness [18].
18.1.1
Switchings and counterproofs
Remember the neologism paraproof , see section 12.5.3, to speak of a slightly generalised notion of proof. A paraproof of A is a test for A, seen as sort of proof. We shall be interested is a specic sort of tests, the switchings of proof-nets. One will take the most simplistic hypotheses : etaspansed multiplicative logic without cut. In other terms, we have three links, , and an axiom link between atoms. What is thence a proof of A by means of a net ? One has not too much freedom : each formula is conclusion of a well-dened link, corresponding to its rst symbol. When the formula is a literal p or p, it is conclusion of an axiom link of which one a priori ignores the other conclusion. Thus, the formula A := p p p p, which has four atoms p, p, p, p has only two proofs, corresponding to the two ways of displaying axiom links 431
432 between the atoms :
CHAPTER 18. YET MORE PROOF-NETS !
p p
pp
p p
pp
p p p p
p p p p (18.1)
One could rewrite this more simply, by only retaining the four atoms, labeled 1, 2, 3, 4 : it is a matter of two graphs, one with the edges 13, 24, the other with 14, 23. To seek a proof of A is to seek a graph between the vertices 1, 2, 3, 4 subject to certain constraints : (i) The edges must relate matching atoms, p with p, q with q . We shall ignore this constraint, which only concerns proofs, not paraproofs. Indeed, one will allow any choice of axioms under the form of boxes with atomic conclusions : daimons, so to speak. (ii) The correction criterion must be veried ; it is where intervenes the formula A, which links together the atoms by means of connectives : depending on the choice of A, one does not get the same switchings. Let us try to solve the problem in the case (18.1) : by erasing the axiom links, one nds oneself with the graphs induced by the various switchings of A ; each of the four choices yields three connected components. They dier, but, in the four cases, they dispatch the atoms in the same way : {1}, {2}, {3, 4}. The solution of my problem consists in all partitions of the set {1, 2, 3, 4} enjoying the correctness criterion. There are of course the two we already know, {1, 3}, {2, 4} et {1, 4}, {2, 3}, but also {1, 2, 3}, {4} and {1, 2, 4}, {3} which correspond to paraproofs of A. Which suggests the following denition : Definition 91 (Incidence graph) Let I be a non-empty set. Two partitions P and Q of I induce a bipartite graph, styled incidence graph : Vertices : On one hand, the elements of P , on the other hand, those of Q. Edges : The edges between a P and b Q are the elements of a b.
| P and Q are polar, notation P Q, when their incidence graph is a tree, i.e., est connected and acyclic.
18.1. DUALITY AND CORRECTNESS
433
| There are therefore (I ) edges ; if P Q, a P , b Q can intersect in at most one point. One will consider, as usual, sets (here of partitions on a given set I ) equal to their own bipolar ; the question is then that of multiplicative connectives ; suppose that the supports I, J are disjoint, and that A, B are sets of partitions equal to their bipolars, whose polars are non-empty :
Theorem 71 (Bipolar) The set A B := {P Q; P A, Q B } is equal to its bipolar. Proof : Let R be a partition of I J such that R (A B ). Choose partitions S , T in respectively A, B . If i I, j J , one can dene S ij T as made of the elements of S and T , up to one exception : the component of i in S and that of j in T have been glued to form a single element of the partition S ij T . One successively veries that : S ij T (A B ). Indeed, the gluing of the component of i with that of j enables glues two disjoint trees into a tree.
| S ij T . In particular, one sees that i, j are not in the same Hence R component of R.
Therefore R = P Q, the union of a partition of I and a partition of J . | | | R S ij T implies P S et Q T. 2 Remark the use of the locative product (14.11). This result should be put side to side with the internal completeness of multiplicatives, see section 14.2.4.
18.1.2
The criterion
Theorem 71 is close to the correction criterion. If one interprets a formula by a set of partitions equal to its bipolar, the sequentialisation theorem precisely says that this interpretation is logically correct. This is indeed because the switchings of A provide us with a prepolar for A. This works as follows : Remark that an appropriate switching can always link the conclusion A the ground with an atom given in advance. The switch puts together two partitions of the atoms without mixing them, on one hand what is aboveA, on the other hand what is above B . It therefore behaves like a tensor product, which is natural, since on is in the dual world. This being said, this switch puts into contact with the ground , as we please, either a class of A, or a class of B .
434
The tensor is without switching. This being said, it glues through the ground A B the class linked with A and the class linked with B . It thence exactly reproduces the construction S ij T of theorem 71. Of course, this prepolar is not everything : thence, if one forms a between A (located in {1, . . . , N }) and A (delocated in {N + 1, . . . , 2N }), one nds the partition {{1, N + 1}, {2, N + 2}, . . . , {N, 2N }}, which is not in the prepolar of A A. A switching literally induces a proof of the negation. Beginning with sequent calculus, one writes the conclusion, namely A. One recursively proceeds from the conclusion : if one has written a sequent , one tries to make it appear as the conclusion of a rule : : If = C D, , one writes a rule

C, D,
D,
: If = C
D, , one writes a rule D or
: C
C,
C D,
C D, gives
D, the full
depending on the switch. One sees that the switch context to the subformula it selects.
Atoms : If only contains atoms, one admits it as an axiom, it is therefore a sort of daimon. All of this can eventually be rewritten by forgetting sequents, as a multiplicative nets with boxes (daimons ) of atomic conclusions. One could by the way treat what precedes, i.e., the duality between partitions associated to switchings, by means of cuts between the associated nets, which reduce to cuts between daimons solved by gluing of boxes. Due the connectedness/acyclicity of the incidence graph, it eventually remains nothing. One could also investigate subtyping. For instance, p (q r ) admits as prepolar the set formed of the partitions {{p, q }, {r }} and {{p, r }, {q }} ; it is therefore reduced to {{p}, {q, r }}. But (p q ) r has a smaller prepolar, reduced to {{p, q }, {r }}, and is therefore formed of {{p}, {q, r }} and {{q }, {p, r }}. And this is why the implication (p q ) r p (q r ) is incorrect, see (11.8). The incorrectness of the net written there corresponds to the incidence graph between {{q }, {p, r }} and {{q }, {p, r }}.
18.2. THE ORIGINAL CRITERION
435
18.2
The original criterion
The original criterion, the one of the paper [33], analysed in [35], was not formulated in terms of partitions, but in terms of trips, i.e., in terms of permutations. This criterion was relegated to the middleground for pedagogical reasons. This being said, it oers a larger latitude, and above all, by replacing partitions which run into no mathematical concept with permutations, it opens the door of operator algebras, since a partition is a unitary operator on a nite-dimensional Hilbert space.
18.2.1
Trips
The idea is easily understood from the version graph . One will go through the net, by following the edges ; since it is a tree, one will be forced to make round trips. Hence each formula of the net will be visited twice, once upwards , notation1 A, a second time downwards , notation A. The travel instructions are as follows2 : Conclusion : From A, go to A. Cut link : From A, go to A, and from A, go to A. Axiom link : From A, go to A, and from A, go to A.
link : One must set a switch left/right : Left : From A B .
B go to A ; from A, go to A
B , from B , go to
Right : From A to A.
B go to B ; from B , go to A
B , from A, go
link : One must also set a switch left/right : Left : From A B go to A ; from A, go to B , from B , go to A B .
Right : From A B go to B ; from B , go to A, from A, go to A B .
Let us proceed quickly, since the essential has been gathered in the simplied version :
1 2
Without any relation to the ludic signposting of section 13.9. One styles left the switch going up from the conclusion to the left premise.
436
The correction criterion, it the absence of short trip, i.e., the fact that this travel is done in a single long trip . For us, it will be enough to bring back the criterion to its graph version. For this, one considers a link ; if one dug it well, the switch of the link does not speak of the link, but precisely of what occurs outside the link. One tries to know, in the cyclic course constituted by a long trip, the order of passage out of the link : one sees that, after A, one cannot directly come back trough B (right switching), nor through A B (left switching) ; one thence comes back trough A. Similarly, after B one comes back trough B . Finally, after A B , one comes back through A B . In other terms, one always comes back through whence one departed. It is thence a matter of course in a connected and acyclic graph. The two switchings of the exclude congurations A . . . B . . . A . . . B . . ., which are not tree courses . A contrario, the restriction to only one of the two switchings will enable considerations inaccessible to the criterion graph . This is what underlies non commutative logic, see section 18.B.
18.2.2
Duality
Definition 92 (Polar permutations) Let I be a non empty set. Two permutations , of I are polar, notation | , when the product is cyclic.
| If I = {1, . . . , N }, means that (1), ( )2 (1), . . . , ( )N (1) are pairwise | | distinct. It is immediate that . This condition translates the correctness criterion and leads to developments of which we already saw a simplied version in section 18.1. The passage from the original to the simplied version is easy to understand : if is a permutation of I , then it splits as a sum of cyclic permutations. In other terms, one can write I = I1 . . . Ik , with the Ik disjoint and non empty, and such that the Ii are cyclic. On replaces with the partition P = {I1 , . . . , Ik }. To prove the analogue of theorem 71, one needs a way of gluing two permutations , : one denes = ij as the disjoint union of both permutations, with a small modication, (i) = (j ) and (j ) = (i), which glues the cycles of i and j . A permutation can easily by self-polar, for instance a circular permutation of three elements, whose square remains circular, hence cyclic. One avoids this sort of things with partitions since one implicitly considers all permutations corresponding to a given partition.
18.2. THE ORIGINAL CRITERION
437
18.2.3
Execution
Let us now consider the elimination of a cut between of support || |A| (corresponding to a proof of , A) and of support |A| || (corresponding to a proof of A, ). The sets ||, |A|, || are pairwise disjoint, moreover one assumes that || || = . The nal output of cut-elimination will be a permutation of || ||. Coming back to the sources, i.e., to cut-elimination in proof-nets, section 11.2.5, one sees that the reduction of the cuts / does nothing but propagating them upwards. Once this propagation is completed, one is left with cuts between axiom links, which can be solved by shortening the links, i.e., by composition of the permutations. The computation of is done as follows : one takes an argument i || ||, and either , or applies, this in an exclusive way. Say that i ||, hence (i) || |A|. If (i) ||, one denes (i) := (i). Otherwise one entered the combat zone |A|, and one thence forms (i) |A| || ; if (i) ||, one denes (i) := (i). Otherwise one resumes with (i) || |A|, etc. One alternates and until one exits. One necessarily exits, without any hypothesis3 : indeed, if the iterations stay forever in |A|, it because of a loop ( )n (a) = a, a |A| ; but since the functions are injective, one cannot access the loop from outside, in particular from i. The rst value obtained outside |A| is by denition (i). Moreover, observe that the logical hypotheses forbid de facto the existence of an internal loop. Indeed, if , are respectively polar to and , | then , which excludes any cycle internal to |A| which would logically constitute a vicious circle, see (11.1). The Tortoise principle, section 3.A.2, enables one to slightly simplify the structure. One will suppose that || = , which by the way corresponds to the original cut, the Modus Ponens. One can write : := ||( . . .)|| (18.2)
where || stands for the (partial) identity permutation of ||. Partial permutations that one composes, union of functions, this does not look very nice : let us change horses.
18.2.4
The execution formula
Let us consider the Hilbert space C|A||| . Total or partial, any function f from |A| || in itself induces an operator : (
i|A|||
3
i ei ) =
i
i ef (i)
(18.3)
See section 19.A.1 for an operator-theoretic explanation.
438
Moreover, when the function is injective, the operator is indeed a partial isometry, see proposition 47, and is thence of norm 1, but when f is completely undened in which case = 0. In particular, , , || induce operators of C|A||| ; , become unitaries, || becomes a projection. Our formula (18.2) rewrites almost identically as : := ||( + + + . . .)|| (18.4)
At least in the logical case whence there is no cycle internal to |A| : the central sum is nite, one can write it, as we please, .(1 )1 or (1 )1 : this is because is nilpotent. The inversion suggest the solution of a linear equation ; this is the feedback equation, which writes, w.r.t. the direct sum decomposition C| A | C| | : (y x) = y x (y ) = y (18.5)
Finally, (x) = x , i.e., only retains the emerging part of the iceberg. The introspective part y (or y ) corresponds to the computation process. This equation underlies the Geometry of Interaction which will occupy the last two chapters of this book.
18.A
Trips and coherent spaces
Trips in proof nets are performed in two times, A question and A answer . One will try to make this machinery work to establish the coherent interpretation of a net, without transiting through sequentialisation. This is a matter of trying to understand the correctness criterion by using it, by no way of establishing a new result. Thus, let R be a multiplicative etaspansed cut-free proof-net of conclusion A, and p, q, r, . . ., be the variables of A. One substitutes for these variables coherent spaces X, Y, Z, . . ., of non-empty supports. The interpretation of R is made of sequences, (x1 , . . . , xn ), indexed by the literals of A, and enjoying : If the literal of index i corresponds to a variable say p or its negation, then xi is in the associated coherent space say xi |X |. If an axiom link relates the literal of index i to the literal of index j , then xi = xj . One will prove the following result : Theorem 72 The set of sequences associated to R is a clique of the coherent space associated to A.
18.B. NON COMMUTATIVE LOGIC
439
One supposes that (x1 , . . . , xn ) and (x1 , . . . , xn ) are two such sequences ; one will show their coherence. For this, one will remark that any formula B of the net induces two subsequences (xi , . . . , xj ) and (xi , . . . , xj ) of the coherent space associated with B : these two sequences are simply denoted B and B . The proof works as follows : one supposes that (x1 , . . . , xn ) (x1 , . . . , xn ), and one eventually concludes that (x1 , . . . , xn ) (x1 , . . . , xn ). This is done by means of a trip : one starts with A with the hypothesis A A , and one comes back to A, with the conclusion A A . The trip is done by means of switchings, chosen in such a way that, when in B , then B B , when in B , then B B. Atom : If one is in a and if a a , and if one then goes to b through an axiom link, hence in b , one get b b , because of linear negation. Link ,upwards : If one enters for the rst time in B C from below, then B C . Hence B B,C C . One sets the switch as one B C pleases, say on left . Then one moves to B , where B B ; later, one will arrive in C , so C C , hence C = C , then one proceeds to come down through B , which yields B B ; one has thence B = B , and, nally, since B C = B C , one can complete the third passage by going to B C . Link , downwards : If one enters for the rst time in B C from above, say, through B , one has thence B C , one B . If B C B can switch to left so as to go down : when coming back, one has coherence, thence eventually equality. Otherwise, it is that B = B and C C , exactly what is needed to follow a right . Later, reaching C , one gets C C , a contradiction. B C , and one can assume strict Link , upwards : One has B C incoherence, in case one side is strictly incoherent, say C C . One switches to right . When back through C , one gets C C, a contradiction. Link , downwards : If one enters for the rst time in B C from above, say, through B , one has thence B B . If B C B C , one can switch to right and go down : when going up one gets equality, . . . Otherwise, this is because we have C C and one switches to left ; after C , one comes back to C , hence C C , a contradiction.
18.B
Non commutative logic
The non commutative logic of Ruet and Abrusci [3] is a very interesting system, albeit slightly experimental and isolated. As to terminology : the expression non-commutative logic is far better than non-commutative linear logic . Indeed, layer 1 considerations tell us that contraction is idempotency which
440
implies in practice commutativity. In general, the accumulation of adjectives conveys an impression of marginality ; this must therefore be avoided when it is like here a pleonasm. Non-commutative unfortunately bears little relation with the non commutative geometry of Connes [13] ; one can however remark that its correctness criterion is stated in terms of trips and not in terms of graphs : it therefore stands slightly closer to functional analysis.
18.B.1
Contexts
Order varieties
What makes the superiority of non commutative over the cyclic logic of which I spoke in section 11.1.3, is that non commutative does not exclude commutativity : besides the commutative tensor A B , there is non commutative one A B , both combining harmoniously, in particular : AA= A B A B (18.6)
One must therefore allow a certain amount of commutativity, and consider that sequents are partially ordered. In the purely non commutative case, one acknowledged the interest of cyclicity, i.e., of orders, up to a cyclic permutation. The notion of an order variety will be to partial orders what the oriented circle is to the segment. Taking as reference the relation circle/segment, one eventually reaches the following covenant : Presentation : every partial order denes (presents) a unique order variety. Existence : every order variety can be presented by a partial order. Unicity : under certain hypotheses, this presentation is unique. Main results Let us state the results ; it is possible to recover the denition of order varieties from the synthesis that follows. One will use , to speak of nite partial orders and , to speak of their domains. One will use < or to speak of their series or parallel sum (assuming the domains disjoint), and + to speak of any of the three sums < , > , . Presentation : to each partial order is associated a variety v () with the same domain. One has v ( < ) = v ( ) which corresponds to cyclicity hence non-unicity of the presentation.
441
Unicity : if one splits the domain of the variety V into , non empty, then there is unicity of , such that V = v ( + ). By what precedes, the sum + is irrelevant, <, >, . Existence : there is existence when ( ) = 1 : if I choose a point A in the domain, I can write V = v (+ A) for a certain partial order (necessarily unique) on the complementary of A. In other terms, an order variety is a structure that can be seen as a partial order every time I remove a point (i.e., I choose a viewpoint). One can thence always focalise on a point A in an order variety, what induces a unique ordered context . Observe that A is in no position (series, parallel) w.r.t. its context, which enables one to have a single cut rule (and a single negation). In the rules of multiplicative disjunctions, we must place two formulas A, B w.r.t. a context ; one denes a partial order between A and B depending whether the disjunction is commutative or not. Then one must be able to present the sequent by + ; this is not always possible, but, when it works, remains unique. The rules Order varieties constitute an abstract, synthetic, notion. Practice is better at ease with analytical notions : one thence presents a sequent by means of a partial series/parallel order. The most general structural rule enables one to replace with when v ( ) v (). Which reduces to two cases : (i) Replacing in the series/parallel decomposition of a series with a parallel . Thus (A < B ) C can be replaced with A B C . B C with
(ii) Replace an external parallel with a series, thence A A < (B C ).
In particular the inclusion of presentations is neither necessary nor sucient : A < (B C ), which is ner than (A < B ) C denes a strictly smaller variety. The logical rules are easy to recover : Pars : They suppose a decomposition + , with = {A, B } ; depending whether is ordered (A < B ) or not (A B ), one concludes + A B or + A B . Tensors : From + A and B +, one concludes in the commutative case ( ) + A B , while in the non commutative case one concludes ( > ) + A B . Identity : Axiom and cut are indierent to all order nuances : the axiom is A+ A ; as to the cut, from + A and A + , it yields + .
442 The solution
The solution can be recovered from the constraint : v (a < b < c) = v (c < a < b) = v ((a < b) c) An order variety is a nite domain equipped with a ternary relation a < b < c ( b between a et c ) enjoying : Irreexivity : a < b < c implies a = c. Cyclicity : a < b < c implies b < c < a. Transitivity : a < b < c and a < c < d imply a < b < d. Repartition : If a < b < c and d = a, b, c, then d < b < c or a < d < c or a < b < d. v () is dened as the set of a < b < c with a < b and either b < c modulo or c incomparable to a, b modulo ; of course v () must be closed under circular permutations. One veries the conditions without diculty, but this is sometimes tedious (repartition condition). Our covenant is respected : Presentation : one immediately sees that v ( < ) = v ( ). Unicity : the unicity of the writing V = v ( < ) = v ( ) is obtained as follows : if I select d , then a < b (modulo ) i d < a < b V .
Existence : the existence of the writing V = v ( + d) comes from the remark that d < a < b V always denes an order relation a < b on , namely . But one must still show the equality V = v ( + d), which reduces to :
The inclusion v ( + d) V , easy to prove : for instance if a < b and c a, b in , one has d < a < b V ; repartition yields d < c < b V or d < a < c V (impossible cases) or a < b < c V. The inclusion v ( + d) V is also proved by repartition : if a < b < c V , a, b, c = d, then say that d < b < c V ; if b < a modulo , then d < b < a V and by transitivity d < c < a V hence b < c < a modulo , which yields a < b < c v ( + d). One reaches the same conclusion from a < c modulo . Finally, the only remaining possibility is that a is not comparable to b, c modulo , but this again yields a < b < c v ( + d). Various The topsy-turvying of orders corresponds to topsy-turvying of varieties, i.e., to the replacement of cycles a < b < c with c < b < a. The repartition condition says that one can put d in place of a or b or c in the cycle a < b < c. Of course, this does not forbid sharper repartitions, e.g. a < d < b (which implies a < d < c and d < b < c).
443
18.B.2
The criterion
Keeping in mind that A B is a subtype proofs of the negation, it appears that :
of A
B , that switchings are the
: The non commutative Tensor retains only one switching, i.e.,
right .
: The non commutative Par has one more, slightly weird : from A, go to A B ; from A B go to B . The additional switching is partial : indeed, once in B , the cycle breaks down ; similarly, there no way of reaching A. Whatsoever, a proof-structure whose conclusion is a single formula A will be correct when there is a sole uninterrupted cycle, the one passing through the conclusion. The relation with order varieties is as follows : in the presence of several conclusions, partially ordered, if the variety contains B < C < D, then any cycle will follow the order . . . B, . . . C, . . . D . . .. In other terms, it is a matter of constraint on the course. An example : say that the unique conclusion A is indeed (B C ) D ). Let us look at the subnet of conclusions B, C, D : By switching ordinarily , it is easily seen that the unique long trip actually passes trough B, C, D ; but in which order ? Let us imagine a passage BDC . I switch on right , on the third position . Then B goes to C , D goes to A, A goes to D ; B and C are lost . By hypothesis, the order of passage is BDC , in other terms, after B comes D , after D comes C , after C comes B . Let us look at what happens with our switching : A D . . . C , lost ! At the other end : B . . . D A, lost too. There is still a long trip, but internal, C . . . B C . The criterion is sequentialisable in the sole cut-free case. But since is preserved by normalisation, this by no means aect the theory : as long as nets do normalise, one is truly interested in the sequentialisation of cut-free ones. Even if non commutative logic which is more a poetical idea that an urging need were bound to disappear, one would at least remember the order varieties, and, as to our methodological reection, this idea of speaking of a cycle passing through the conclusion. This means that the conclusion plays a distinguished role : this is a possible reading of polarisation, we shall see this again in chapter 20.
Chapter 19 Geometry of Interaction

Geometry of Interaction (GoI) is the natural continuation of the interpretation of chapter 18 : the permutations of a nite set are naturally relieved by the bounded operators of a Hilbert space.
19.1
19.1.1
From syllogistics to feedback

Chiasmi
seems to deserve a category-theoretic interpretation. Indeed, let us associate to the formulas A, B, C Hilbert spaces E, F, G and let us assume that the proofs of AB and B C do correspond to operators u B(E, F), v B(F, G). The syllogism is thence interpreted as composition : the conclusion corresponds to w := v u B(E, G). Moreover, the axiom A A is rightfully interpreted as the identity of B(E). Which enables one to indulge in syllogistics and nothing more, by the way ! Let U B(E F), V B(F G), W B(E G) be the block-matrices 0 u 0 v 0 w (see section 19.C.2) : . These matrices are u 0 v 0 w 0 indeed chiasmi, i.e., operators of the form (a a ), with a crossing of inputs/outputs operated by (x y := y . W.r.t. example 4, section 17.5.2, the tensor product has been replaced with a direct sum. It is easy to see that W is the solution of a linear equation. Given x F, z H, one denes W by W (x z ) := x z , where x , z are dened 444
Let us forget for a while the preparatory work done in chapter 18, and let us try an interpretation inside Hilbert spaces. The syllogism Barbara, section 11.B.2, that one can rewrite as : A B B C (19.1) A C
19.1. FROM SYLLOGISTICS TO FEEDBACK by : U (x y ) = x y V (y z ) = y z
445
(19.2)
which admits the unique solution : y = u(x), z = v (u(x)), y = v (z ), x = u (v (z )). It is a feedback equation, since the outputs y , y are reinjected here, one more chiasmus as inputs y, y . While here, observe that category-theoretic compositionality i.e., associativity consists in equating the various solutions of the feedback equation between three chiasmi (here, w B(G, K)) : U (x y ) = x y V (y z ) = y z W (z t) = z t (19.3)
One can either solve the equation between U, V , then solve the equation between this solution and W , or solve the equation in a single step : due to the existence and unicity of the solution, the result is the same. Among the immediate neighbours of the syllogism, the Modus Ponens : B B C C (19.4)
appears as the particular case of (19.1) with E = 0. If this rule is that universal remember that Hilbert-style systems essentially rely on it , it is surely that it is not limited to chiasmi (with E = 0, one would get U = 0 !). Which shows that the general form of a proof of A B is a 2 2 matrix. It is reasonable to a b , with a = a, c = c, and also of assume it to be hermitian, i.e., U = b c norm at most 1. Equations of the type (19.2) correspond to the generalisation of category-theoretic composition, but by no way reduce to it : while (19.2) is trivial in the case of chiasmi1 , its general solution is at least problematic. We shall rewrite equation (19.2) under the equivalent form of cut-systems. Let us give an example : if U B(E F), V B(F G), one introduces H := E F F G, as well as h := U V . It goes without saying that solving equation (19.2) is quite the same as solving : h(x y y z ) = x y y z Which is indeed an equation of Modus Ponens : A A (A A) (B C )
1
(19.5)
B C
(19.6)
It is handled by nilpotency, see infra.
446
CHAPTER 19. GEOMETRY OF INTERACTION
0 I of the space E E (a true matI 0 rix). This type of transformation, which complicates a problem to simplify its structure has been styled the principle of the Tortoise, see section 3.A.2. The chiasmus , extended into a partial symmetry of the whole space, will become a feedback. between h and the chiasmus :=
19.1.2
Cut-systems
Definition 93 (Partial symmetries) Let H be a Hilbert space. A partial symmetry of H is a hermitian operator such that 3 = . The spectral calculus puts in correspondence with the inclusion between Sp( ) and C which thence veries 3 = . Which shows that the spectrum is discrete : Sp( ) {1, 0, +1}. The eigenspaces 1 , 0 , +1 of the values 1, 0, +1 correspond to the projections : 1/2( 2 ), I 2 , 1/2( 2 + ). One will note S the sum of the subspaces 1 and +1 , i.e., the space corresponding to the projection 2 ; R will stand for 0 . Hence H = R + S, the direct sum of the kernel and the support of . Definition 94 (Cut-systems) A cut-system is a three-tuple (H, h, ) such that : H is a complex Hilbert space. h is a hermitian of H, of norm at most 1. (the cut, the loop, the feedback) is a partial symmetry, i.e., = = 3. The system (H, h, ) induces a feedback equation : given x R, nd x R, y S, such that : h(x + y ) = x + (y ) (19.7) It is the visible part of the equation, the result x , which interests us ; the introspective part y will correspond to the computation process. It is the existence of this hidden component which makes that one is really at layer 3. This equation is indeed a variant of (19.2) reformulated by means of (19.5).
19.1.3
Normal forms
The normal form corresponds to the visible, extraspective, part of the equation :
19.1. FROM SYLLOGISTICS TO FEEDBACK
447
Definition 95 (Norma form) Under the proviso of existence and unicity of the solution of equation (19.7), the normal form of the system is the operator h (x) := x . h 1 : indeed, x 2 + (y ) 2 x 2 + y 2 et comme (y ) = y , x 2 x 2 . In the same way, one can see that h is hermitian. Therefore, the normal form can be seen as the trivial cut-system : (R, h , 0), a cutfree system, whose feedback is null. Which poses the principle of iteratibility of the normal form : if is a feedback of R, then one can form under the proviso of existence and unicity the normal form h . At the price of a small abuse of notation, one can assume that is dened on the full space H, in which case + is a feedback too, i.e., a partial symmetry. Definition 96 (Independence) Two feedbacks , are independent when = 0 (= ). In other terms when their supports S, T are orthogonal. The solution of equation (19.3) in two steps writes as h or h , and its solution in one step writes as ( + ) h . Under reasonable hypotheses, these three expressions will be equal, this is associativity. Associativity is all what remains of layer 2 ; one can use it in a regressive way to build categories based upon GoI, traced if one insists on wasting paper. But the associativity of GoI includes a case inaccessible to ruminants : one can write = + + , i.e., as the sum of a positive feedback a projection and a negative feedback the opposite of a projection. This decomposition makes sense neither in set theory (graphs, etc.) nor in category theory (morphisms, etc.), since it cuts out objects in a forbidden way. Thus, in dimension 3, the splits two exchange of indices 1, 2 as the sum of 0 1 0 1 / 2 1/ 2 0 1/2 1/2 0 independent feedbacks : 1 0 0 = 1/2 1/2 0 + 1/2 1/2 0. 0 0 0 0 0 0 0 0 0
19.1.4
Variants
Let us discuss some variants of the notion of cut-system. Positive case One could try to restrict equation (19.2) to operators U, V which are positive. Since positive operators enjoy exceptional properties linked to the square root, this simplies a lot of things. I acknowledge to have been tempted by the replacement of identity chiasmi with their positive parts ; syntactical normalisation thence translates as an
448
inequality h h , what is far from being unpleasant. But I found myself too far from the base camp and I turned back. Anyway, there is something to dig n/n + 1 1/n + 1 here : for instance, if, for n 0 Pn := , the equation (19.2) 1/n + 1 n/n + 1 between U := Pm and V := Pn admits the solution W := Pm+n . P0 is not positive, it is the identity chiasmus ; on the other hand, the Pn are positive for n 1, thus P1 is the positive part of P0 . N equations involving P1 yield the result PN , thence keeping track of the computation process. Non hermitian systems The hypothesis h hermitian corresponds to the usual conditions of use of GoI ; moreover, this restriction is most useful in the general case. But it is of no use stricto sensu, this is what the Tortoise tells us. Let us start indeed with an equation (19.2) with U, V arbitrary of norms 1, and let us add duplicates E1 , F1 , G1 of the spaces E, F, G ; the equation : U (x y ) = x1 y1 U (x1 y1 ) = x y V (y1 z1 ) = y z V (y z ) = y1 z1
(19.8)
is of the same type on twice bigger spaces ; but it relates the hermitians 0 U 0 V and , still of norms less than 1. Its solution immediately U 0 V 0 induces a solution of the original equation : W (x z1 ) := x1 z . Arbitrary norm The removal of the restriction h 1 opens the Pandora box. There is no general result in that case ; in the invertible case, see infra, one has a few partial results.
19.2
19.2.1
Invertible systems
Normal form
Let (H, h, ) be a cut-system ; one decomposes the space as the direct sum H = R + S between the kernel and the support of ; this induces the block 0 0 a b ,= ; s is a symmetry of S. writings : h = b c 0 s
19.2. INVERTIBLE SYSTEMS
449
Definition 97 (Invertibility) The system (H, u, ) is invertible when s c is invertible as an endomorphism of S. Theorem 73 (Normal form) If (H, h, ) is invertible, then the feedback equation admits the normal form h := a + b (s c)1 b as its unique solution. Proof : One wants to solve the equation :
x a b x = . Since s c is 0 b cs y invertible, it is immediate that x = (a + b (s c)1 b)(x), y = ((s c)1 b)(x) is solution of the equation. It is indeed the solution : if x 0 a b , then (c s)(y ) = 0, hence y = 0 and x = 0. Hence = 0 b cs y the visible part of the solution is given by h := a + b (s c)1 b. 2 Remark 2 The shape of the solution renders explicit the hermitian character of the solution. The dependency h ; h is norm-continuous ; we shall soon see that it is also order-continuous. The logical interpretation is happy with invertibility ; indeed the termination of computations translates as the nilpotency of sc, see infra, which implies that (s c)1 = s + scs + scscs + . . . + s(cs)n , a nite sum.
19.2.2
Associativity
In presence of independent feedbacks , , there are several ways to get a normal form for (H, h, + ). One can either form at once ( + ) h , or rst normalise (H, h, ), which yields h , then normalise (R, h , ), which yields h . One can work in the opposite order, which leads to h . The associativity theorem relieves us, all of this amounts at the same thing : Theorem 74 (Associativity) Assume , independent and write H = R + S + T. Then (H, h, + ) is invertible i (H, h, ) and (R + S, h , ) are invertible. Moreover : ( + ) h = h Proof : The theorem essentially reduces to the following lemma : Lemma 74.1 If the hermitian f = c d d e (w.r.t. a decomposition S + T) is of norm at (19.9)
most 1, is s, t are symmetries of S, T, then g :=
s c d is invertible d t e i t e are s (c + d (t e)1 d) invertible in T and S respectively.
450
Proof : Hermitian (more generally, normal) operators admit approximate eigenvectors, see [59], p. 183. Hence, if t e is not invertible, there is a sequence (yn ) T of norm 1 vectors such that (t e)(yn ) 0 ; since t(yn ) = 1, one sees that e(yn ) 1, thence, from f 1, d(yn ) 0. One deduces that g (yn) 0, hence g is not invertible. Dening c := c + d (t e)1 d and assuming now s c not invertible, there is a sequence (xn ) S of norm 1 vectors such that (s c )(xn ) 0. But c is a normal form, which means that one can nd yn T (unique) such that g (xn + yn ) = (s c )(xn ) ; then g (xn + yn ) 0. One just proved that the condition is necessary ; the suciency is almost obvious, due to the explicit formula : (s c ) 1 (s c )1 d (t e)1 . 2 g 1 = (t e)1 d(s c )1 (t e)1 + (t e)1 d(s c )1 d (t e)1 2 Remark 3 The theorem persists under a weaker form systems of arbitrary norms : if (H, h, ) and (R + S, h , ) are invertible, then (H, h, + ) is invertible too, and its normal form is given by (19.9).
19.2.3
Semi-invertible systems
Since any hermitian can uniquely be written as the dierence of two positive hermitians whose product is null, the feedback writes as the sum = + + of two feedbacks, one positive, one negative. Definition 98 (Semi-invertibility) The system (H, h, ) is lower semi-invertible (l.s.i.) i (H, h, ) is invertible. Proposition 49 (i) If (H, h, ) is l.s.i. and h k , then (H, k, ) is l.s.i. (ii) If (H, k, ) is l.s.i. and h k , if 0 < < 1, then (H, (1 )h + k, ) is l.s.i. Proof : (i) By restricting to the space 1 , this simply means that, if c c I and if I + c is invertible, then I + c est inversible. (ii) If I + c is invertible and c c I , then I + c + (1 )c I + c (1 )I = (I + c) is invertible. One symmetrically denes upper semi-invertible (u.s.i.) systems.
19.2. INVERTIBLE SYSTEMS Proposition 50 The system (H, h, ) is invertible ssi it is both u.s.i. and l.s.i.
451
Proof : Theorem 74 shows the necessity. Conversely, s c writes, w.r.t. the u v , with u, w positive decomposition S = +1 + 1 under the form : v w u 0 and invertible. Such a matrix is necessarily invertible : let m := , 0 w I +1 t then m1 (s c)m1 is of the form n := for which there is the t I1 (I+1 + t t)1 (I+1 + t t)1 t explicit inverse : n1 := . Thence t(I1 + t t)1 (I1 + tt )1 (s c)1 = mn1 m. 2 Corollary 50.1 If h k are such that (H, h, ) is u.s.i. and (H, k, ) is l.s.i., then (H, 1/2(h + k ), ) is invertible.
19.2.4
Order continuity
In the invertible case, the normal form commutes to directed sups and infs. Theorem 75 (Monotonicity of the normal form) The function h ; h is monotonous and continuous w.r.t. the order. I sketch here the main steps of the proof, see [50] : (i) The principle of the Tortoise enables one to let vary the sole parameter c : in R R H, there are three copies of R, and let (resp. ) be the partial symmetry exchanging copies 1, 2 (resp. 2, 3). Then (R R H, + h, + ) is an invertible system with the same normal 0 0 form as (H, h, ), but its coecient c writes . 0 h (ii) Let p be a positive hermitian ; a power series expansion shows that the function ; a + b (c + p)1 b is dened and monotonous in a open neighbourhood of 0 in R. (iii) Let (H, h , ) be another invertible system, where h diers from h only through its coecient c c. Let k := (1 )h + h be a convex combination, i.e., with 0 1. h invertible is l.s.i. and the same is true of k h ; symmetrically, one sees that k is u.s.i., hence by proposition 50, it is invertible. If p = c c, the function ; a + b (c + p)1 b is dened and locally monotonous. One deduces that a + b c1 b a + b c 1 b.
452
(iv) A pointwise supremum is dened by a weak limit : h(x) | x = supi hi (x) | x . But proposition 55 shows that this is indeed a strong limit. Since the product is strongly continuous on balls, the same will be true of the inverse. This is why the normal form commutes to sups (and to infs).
19.3
19.3.1
Old style GoI

The programme
Geometry of Interaction originates in the paper [35] (section 18.2), conceptualised in a programme [39]. This programme nds its rst realisation in the paper [37] : the GoI of system F, indeed second order linear propositional logic minus the additives. The paper [40] extends GoI to pure -calculus. As to [43], it tries to integrate with ups and downs the additive connectives. The quality of the programme is above all its originality, since it was not obvious that such a thing was possible. Its obvious limitation, is the aspect ad hoc, with the forceps of the general results style associativity ; but this has since been xed in [50], see sections 19.2, 19.A and 19.B. Remains another criticism, much deeper : these interpretations are done in B(H), i.e., in a trivial von Neumann algebra (of type I ), see section 20.1.3. They thence have no hope of running into an iconoclast rereading of logic, i.e., of exponentials, see chapter 16. A second GoI, based upon type II1 algebras is currently under development, see chapter 20.
19.3.2
Generalities
To each proof is associated a cut-system ; more precisely let us introduce the notation , [] to speak of the sequent proven by means of cuts on . The idea of the rst GoI, is to dispose of one Hilbert space by formula of , , always the same, H. One will suppose H of innite dimension and separable ; it is thence isomorphic to 2 , the isomorphism depending upon a Hilbertian base (en ) : ( n en ) := (n ) (19.10) see section 17.A.1. A proof of , [], i.e., of with cuts on the , will be interpreted by a cut-system (H, , h, ), where : H, is the sum of several copies of H. h is a bounded operator on the previous space ; what can be written as a matrix of type (, ) (, ), with coecients in B(H).
19.3. OLD STYLE GOI
453
The elements of being the cut formulas, they go by pairs. is thence the matrix whose coecients are null but those corresponding to two indices A, A paired by a cut : mAA = I . In [37] the following is (more or less) proven : Theorem 76 (Normal form) If t is a closed term of system F of normal form u, if (H, h, ) and (H, k, 0) are the respective interpretations of t and u, then : (i) h is nilpotent. (ii) h = k Proof : One must basically understand that the sequence of reductions leading to the normal form constitutes a solution of the feedback equation by successive substitutions, what translates as the nilpotency of h, a phenomenon already observed in section 18.2.4. In terms of proof-nets, the reduction of cuts translates as follows : The logical cuts, e.g., / , are reduced by replacing the system with an isomorphic one. Beware, one is in innite dimension, and this replacement is everything but innocent : thus it augments the size of matrices, thence that of , i.e., the number of cuts. The cuts between identity axioms ; they are chains of chiasmi of which (19.2) or (19.3) are examples for the lengths N = 2, 3. If one translates (19.2) or (19.3) as cut-systems, one sees that h is nilpotent : h = 0 or (h)2 = 0. More generally, for a chain of N chiasmi, one will get (h)N 1 = 0. 2 Remark 4 The statement is approximative, since h is not quite k , but something close, see section 19.3.6.
19.3.3
Identity group
0 I . Similarly, I 0 the cut between cut-free proofs M of , A and N of A, consists in the juxtaposition of both matrices, so as to get a matrix of type M 0 (, A, A, ) (, A, A, ), i.e., , to which one must add the 0 N feedback AA : The identity axiom A, A is interpreted by the matrix :
454
0 . . . 0 0 . . . 0
... 0 0 ... 0 . . . . . . . . . . . . . . . ... 0 I ... 0 ... I 0 ... 0 . . . . . . . . . . . . . . . ... 0 0 ... 0
corresponding to the swapping of A et A. If M, N already have cuts , , one denes := + + A A . In what follows, one will content oneself, for reasons of legibility, of the cut-free case.
19.3.4
Multiplicatives
If one were following, instead of a matricial technique, a block technique, there would not be the slightest problem. , written in block-style : from a Take for instance the case of a proof of say A, B, C written as : uAA uAB uAC uBA uBB uBC (19.11) uCA uCB uCC one would pass to A
Similarly for the
: from proofs of D, A and B, E written as : et uBB uBE uEB uEE (19.13)
B, C by uAA uBA uCA
means of : uAB uBB uCB
uAC uBC uCC
(19.12)
uDD uDA uAD uAA
In both cases, one groups the A and the B in the same block.
one would pass to D, A B, E : uDD uDA 0 uAD uAA 0 0 0 uBB 0 0 uEB
0 0 uBE uE
(19.14)
19.3. OLD STYLE GOI
455
But these blocks only have a subjective value, in other terms, a cut between (19.12) and (19.14) is the same thing as a cut between (19.11) and : uDD uDA 0 0 uAD uAA 0 0 (19.15) 0 0 uBB uBE 0 0 uEB uEE i.e., tow cuts between (19.11) and the two matrices of (19.13). It is indeed what we already did with permutations, see section 18.2.3. This is not tenable for system F : polymorphism, i.e., subtyping, asks for a space independent from logic. One would like, for instance, to replace A B with a variable X which has forgotten that it stands for A B . Which is only possible if all coecients become homogeneous, and this is why one replaces block-matrices with true matrices. The Hilbert hotel yields a solution, under the form of an isometry of H H into H. Taking inspiration from section 14.1.2 and the delocations , of (14.8), one can dene the partial isometries p, q of the space 2 : p( q( of respective adjoints : p ( q( n en ) := n en ) := 3n en 3n+1 en (19.17) n en ) := n en ) := n e3n n e3n+1 (19.16)
One sees that the following equations are satised : p p = q q = I p q = q p = 0 (19.18) (19.19)
These equations are those of the isometry of HH into H : xy ; p(x)+q (x). An isometry onto H e.g., 2n, 2n + 1 would correspond to : pp + qq = I (19.20)
of which (19.19) is a weak version. Let us conclude with multiplicatives : p, q enable one to embed say matrices of type (A, B, C ) (A, B, C ) in the matrices of type (A B, C ) (A B, C ), by contracting two indices, thence :
456
becomes :
uAA uAB uAC uBA uBB uBC uCA uCB uCC puAA p + puAB q + uBA p + uBB q puAC + quBC uCA p + uCB q uCC
The previous transformation which works as well on the Tensor side is an isomorphism of C -algebras, which preserves everything, but the identity (equation 19.20 would be required). Concretely, this means that two cuts between , A and B, and A, B, , represented by a matrix of type (, A, B, , A, B, ) (, A, B, , A, B, ) and a feedback swapping A, A and B, B are replaced,using the isometry, with : a matrix of type (, A B, , A B, ) (, A B, , A B, ) and a feedback exchanging A B and A B . The isometricity ensures that the normal form of both systems is the same, which corresponds to the reduction of a cut / . Two remarks : The refusal of (19.20) is a refutation of etaspansion, but not really honest : nobody forbids me from taking a surjective isometry, of the kind (2n, 2n + 1). On the other hand, in section 17.5.4, there was no other way ; moreover, without the help of one-eyed hotels with innitely many dimensions : by remaining in dimension 2. One can see p, q as the maintenance of an operational stack. p, q respectively push 0, 1, while p , q pull them. This means that, if the top of the stack is a 0 , then p removes it, otherwise, it is not dened. Which justies the equations (19.18) and (19.19), but by no way (19.20) which would for instance require a stack of innite depth.
19.3.5
Exponentials
This was the chunk of [37], and it took more than one year to nd the solution, the problem being with contraction, i.e., duplication. One tensorises with a message space isomorphic to H, which makes that u is promoted into !u := I u. One will take care of duplications by means of the p I, q I, . . ., which commute to the I u. But one is not in H H, one is in H. The previous operations are therefore brought back in H by means of an isomorphism between H H and H. In particular, there is an internal tensor product : u v := (u v ) .
19.3. OLD STYLE GOI
457
Which enables one to interpret rules of constant depth , see section 16.4.1. To get dereliction, one must be able to recover, from !u, a copy of u, this what is done by the operator d, which satises d (!u)d = u. One must also interpret the principle !A !!A, which consists in exchanging I (I u) with (I I ) u = I u. What is ensured by the unitary operator t : t(u (v w ))t = (u v ) w. Remark that t belongs to the same kind as the associativity functors of monoidal categories, see section 9.A.2. Let us x a Hilbertian base (en ) de H ; any partial injective function from N into N induces a partial isometry of H, u( n en ) := n e(n) , what we already used for p, q, p, q . The isomorphism between H H and H is induced by a bijection (m, n) ; m, n between N N and N. If u B(H), then I u B(H H), what can be brought back into !u := (I u) B(H). If u is induced by a partial injection from N into itself, then !u is induced by !( m, n ) := m, (n) . A possible choice for d is the operator induced by the function n ; 1, n . t corresponds to the bijection m, n, p ; m, n , p . The rules are interpreted as follows (the matrix u corresponds to the premise, the matrix v corresponds to the conclusion) : Weakening : The passage from to , ?A : vij := uij if i, j =?A, vij := 0 else. Contraction : The passage from , ?A, ?A to , ?A : one works as if the conclusion were , ?A ?A ; but, instead of p, q, . . . , one uses (p I ) , (q I ) , etc. Dereliction : The passage from , A to , ?A : let D be the diagonal matrix DCC := I for c , DAA := d. Then v := DuD . Promotion : The passage from , A to ?, !A is easily interpreted : vij := (I uij ) . But this is not enough, one also needs : Burying : The passage from , ??A to , ?A : let T be the diagonal matrix with coecients TCC := I when c , TAA := t. Then v := T uT . One easily checks that normalisation rules are correctly interpreted, provided one restricts to exponential cuts whose premise ?, !A is without context, i.e., with = . In the general case, there is a slight dierence between GoI and syntactic normalisation.
458
19.3.6
Mistakes
of GoI
W.r.t. exponentials, GoI makes mistakes. Thus, it forgets erasing. Take for 0 0 instance a prof of ?B, A, obtained by weakening, thence of the form , 0 u a b that one cuts with a proof of A, ! B . The solution of the feedback c d a 0 equation yields the result : , while, if one were following syntax, the 0 u coecient a should disappear too. In general, GoI makes small mistakes w.r.t. syntax. But this is not that bad, indeed : If the conclusion is a 1 formula which does not use ? , there is no mistake, even in presence of exponential cuts. Hence, by associativity, GoI does not contradict syntactical normalisation. Moreover, GoI corresponds to optimal reduction, see infra. In other terms, who makes a mistake ? Not the one you may think. . .
19.3.7
Weak nilpotency
The extension of GoI to pure -calculus is a rather interesting endeavour. W.r.t. the work on system F, it only requires pretending to make typing mistakes. Not quite, since the results of this rst round of GoI were relying on logical correctness. It was not possible to release operators in the wilderness and say please normalise . It was necessary to develop a specic method for solving the feedback equation in this case. Instead of proving that h is nilpotent, one shows that it is weakly nilpotent, i.e., that the (h)n converge to 0 in the weak topology : h)n (x) | x 0. -calculus is interpreted by weakly nilpotent cut-systems, what is enough to get a normal form together enjoying the right properties. This result had an unexpected consequence : Lamping had devised a rather hermetic algorithm of optimal reduction for -calculus. Gonthier [52] was then able to translate the algorithm in GoI, thence proving the mathematical correctness of Lampings ideas. I had no second look at this weak nilpotency. This should be revisited under the light of the general results of this chapter. Whatsoever, one presumably needs nothing to show that the equations of -calculus are satised by the normal form, in the general sense of section 19.B : the characterisation of section 19.B.4 is indeed very amenable.
19.3. OLD STYLE GOI
459
19.3.8
Digression : the dictionary
We saw that normalisation translates in terms of nilpotency. The idea of nilpotency is very natural in logic : let us use the metaphor of the dictionary. This dictionary is perfect in its kind : from a set of basic knowledge, it enables one to dene a whole vocabulary, say, technical. Which means that, taking a word , the dictionary denes it clearly, with perhaps references to other words ; what sends back to these other words, and, recursively, to other references. To say that one has a good dictionary means that one is not endlessly walked from reference to reference. What one can express, as we please, in two ways : The reexive/transitive closure of the relation refers to the word is an order relation.
Dene the la matrix (a ) : a = 1 when the word refers to , a = 0 else ; then (a ) is nilpotent.
19.3.9
Additives
The extension to additives done in [43] is part of those works that cannot be justied. This was done in time when additives didnt nd their real place at layer 3. One would by the way be surprised to see how many interesting works forget the additives2 . Technically speaking, the reason for this diculty was simple : GoI interprets the tensor product by a direct sum of spaces, thence producing a logarithmic simplication . But where to nd a logarithm for the sum ? In extremis, one must reintroduce what is missing, under the form of a dialect, i.e., a message space. La solution is heavy, unsubtle, in contrast to the necessary elegance of an ultimate explanation of logic. On can consider [43] as the rst draft of chapter 20, where multiplicatives and additives are eventually reconciled. A bungled draft, since the solution was not found by pushing the idea of dialects, but by transiting through quantic intuitions.
19.3.10
Limitations of the rst GoI
Fundamentally, the rst GoI makes used of the language of operator algebras without really respecting their spirit. Thus, one never steps out of partial
In the same way that circulate sequent calculi without Hauptsatz, one witnessed extensions of GoI to additives, but where the analogue of theorem 76 is restricted to the fragment without additives. By the way, the new generation of portable phones can also be used as mustard pots. . . as long as one puts no mustard in them !
2
460
isometries, what indeed amounts at restricting to the partial permutations of a distinguished base. One hardly gets more than a preposterous dressing of a theory of the partial permutations of N. Practical jokers had even the bad taste of rewriting GoI inside Scott domains, what is vexing, albeit perhaps merited. This being said, the idea is excellent, one must only make the equation live in another milieu than the one induced by logical proofs. Our brief incursion in the quantum freed us from this limited standpoint. The general results of this chapter are genuine operator algebra results, inaccessible to kindergarten topology. The next chapter will develop the new GoI, worthy, I hope, of our ambitions.
19.A
Termination
One takes an interest in the solution stricto sensu of the feedback equation (19.7). The section on deadlocks deals with the unicity of the introspective component y : one essentially can, i.e., remove the deadlocks. The section on the computational size introduces an execution operator et proves an inequality related to associativity.
19.A.1
Deadlocks
a b ,= b c 0 0 . 0 s
With the usual notations H = R + S, h =
Definition 99 (Injectivity) The system (H, h, ) is injective if s c is injectif, as an endomorphism of S. Invertible systems are injective, but the converse is wrong : in innite dimension, injectivity does not entail surjectivity. Which is by the way what dierenciates spectrum and eigenvalues : an operator which is injective but not surjective has 0 in its spectrum, while there is no eigenvector for 0. Definition 100 (Deadlock) The shortcut of (H, h, ) is the subspace Z S dened Z := ker(s c) ; its (non-zero) elements are called deadlocks. Proposition 51 IZ commute to , s, c, h. Proof : If z Z, then c(z ) = s(z ), hence c(z ) = z ; one deduces c2 (z ) | z = c(z ) | c(z ) = z | z ; since c 1, one is in the equality case of Cauchy-Schwarz (theorem 64), which forces c2 (z ) = z , and since c2 is positive,
19.A. TERMINATION
461
c2 (z ) = z . One immediately deduces that s(c(z )) = c(c(z )), i.e., that c(z ) Z. Thence c, s send Z into Z, which means that they have block-matrices of the u v form w.r.t. the decomposition H = Z + Z ; since these operators 0 w are hermitian, v = 0. Which shows that Iz commutes with s, c, and therefore with , h. 2 Proposition 51 says that the deadlock Z is rather nice : one can always remove it by replacing h with (IZ )h without changing anything to the feedback equation but for the fact that the introspective component y becomes unique. The deadlocks exactly correspond to the short trips of the proof-net criterion, see section 18.2.1, which come in turn from congurations of the type (11.1). One already observed their autistic character, their non-communication with the outside : this is what expresses proposition 51. In nite dimension, injectivity equals invertibility, and one can thence solve the equation by removing the shortcut .
19.A.2
Computational size
Let (H, h, ) be a cut-system and let Z be its shortcut : two solutions of the feedback equation (19.7) x + y , x + y for the same datum x are such that x = x : from h(0 + (y y )) = (x x ) + (y y ) (19.21) and h 1, (y y ) = y y , one gets x x = 0, thence x = x . But the two answers x + y , x , y can still dier by their introspective components y, y , in case y y is a deadlock. Proposition 51 yields a sort of unicity : (i) If x + y is solution for the input x, x + Z(y ) is also solution. (ii) x + Z(y ) is amongst all solutions x + y for the input x that of smallest norm. (iii) This smallest solution consists in removing the shortcut , i.e., in replacing (H, h, ) with (H, IZ h, ). Definition 101 (Termination) The cut-system (H, h, ) terminates when equation (19.7) has a solution x + y for any input x R. In case one denes the execution operator ex(h, ) from R into H : it associates to x R the vector3 x + Z(y ).
3
This is not a typo, I didnt mean x + Z(y ).
462
The typical case for termination is invertibility ; in case ex(h, ) = IR + ( c)1 b. But there are many other cases, e.g., b = 0. Definition 102 (Size) When (H, h, ) terminates, its computational size is dened by : size(h, ) := ex(h, ) (19.22)
Termination is a rather accidental feature ; this is why the next theorem is less powerful than its model, theorem 74. In what follows, h refers to the literal solution, and does not necessarily agrees with the general solution that we shall dene infra, section 19.B. Theorem 77 (Associativity of size) If , are independent, and if (H, h, ) and (R + S, h , ) terminate, then (H, h, + ) terminates too ; moreover size(h, + ) size( h , ) size(h, ) (19.23)
Our denition of size can be grasped from the naive solution of (19.7) by means of the series expansion : h = IR (h + hh + hhh + . . .)IR a correct formula when h, i.e., sc, is nilpotent : ex(h, ) = IR + b sb + b scsb + b scscsb + . . . + b s(cs)n2 b (19.25) (19.24)
where n is the smallest integer such that (h)n = 0 (the order of nilpotency of sc). Then size(h, ) n ; in practice, e.g., for a converging normalisation in say system F, size(h, ) n. One should perhaps replace the size with its logarithm : Danos observed in his PhD (see, e.g., [19]) that, in -calculus, the order of nilpotency of the cut-system associated with t is exponential in the normalisation length of t. In this case, theorem 77 involves a sum in place of a product.
19.B
19.B.1
General case
Digresssion : the Lebesgue integral
The invertible case, possibly improved with minor extensions, see section 19.A, cannot proceed further : this is because the feedback equation (19.7) is an impossible problem. We shall take our inspiration from Lebesgue integral, the celebrated solution of another impossible problem : the integration of discontinuous functions.
19.B. GENERAL CASE
463
(i) The Riemann integral is a continuous linear form on the Banach space R([0, 1]) of real-valued continuous functions on [0, 1]. Our normal form is slightly of the same kind : the function h is dened on certain operators h (invertible systems) of norm 1. One could even suppose R of dimension 1, what would make that our result lies in R. An essential dierence, the dependency is not linear ; on the other hand, the result remains bounded : h 1. (ii) The next step consists in remarking that the Riemann integral is monotonous, thence in extending it, by passage to suprema, to lower semicontinuous (l.s.c.) functions, which are the sups of continuous functions. We shall do the same, extend our normal form to sups of invertible systems, lower semi-invertible (l.s.i.) systems. (iii) Of course, only a monotonous function can thus be extended : both the Riemann integral and the normal form are monotonous. Moreover the extension must be compatible with what has already been dened. In the case of the integral, one uses Dinis theorem : if fn is a monotonous sequence of continuous functions with a continuous sup f , then the convergence fn f is uniform, i.e., normwise. The same is true of the normal form which commutes to invertible sups of invertible systems. The idea is the same as in the previous case : it is a matter of topological bonication due to the order structure. At Dinis, from simple to uniform, here from weak to strong. (iv) A last step must be performed : after the sups, the infs (of sups). Thats all : at Lebesgues, any measurable function is equivalent to an inf of l.s.c. functions. For us too, since any system is an inf of l.s.i. systems. (v) But the answers are dierent : the Lebesgue integral can proceed with the second extension because the extension to l.s.c. already commutes to infs. On the other hand, the extension of the normal form to l.s.i. does not commute to infs : this is because we are in a non commutative world. Something else must be found.
19.B.2
Semi-invertible case
We shall extend the normal form to l.s.i. systems ; since it is easy to mistake l.s.i. and u.s.i., let us say that the l.s.i. are rather large , that they are maintained by directed sups, and that they socialise with positive feedbacks : indeed, if is positive, = 0 hence any system (H, h, ) is l.s.i. The anity between l.s.i. and directed sups comes from :
464
Proposition 52 Any l.s.i. system can be written as a directed sup of invertible system. Proof : (H, I, ) is u.s.i., hence, for 0 < < 1, (H, ( 1)I + h, ) is invertible by corollary 50.1. One denes h := (1)I +h : it is a monotonous family of supremum h. 2 Corollary 52.1 Any system can be written as a directed sup of u.s.i. systems. Theorem 78 (Semi-invertible case) The normal form extends to l.s.i. systems so as to preserve directed sups. This extension is monotonous and associative. Proof : By theorem 75, the normal form is monotonous and order-continuous for invertible systems. It thence extends by order continuity to directed sups of invertible systems, i.e. to l.s.i. systems. Associativity extends without problem to this case. 2 Remark 5 The extension to l.s.i. systems does not commute to directed infs. In [50], section 6.6., one gives the example of a decreasing family (H, hn , ), with positive thence a l.s.i. family such that inf n hn < infn hn . A symmetric extension to u.s.i. systems, by passage to infs is possible. These two extensions are compatible ; indeed, if h k are such that (H, h, ) is u.s.i. and (H, k, ) is l.s.i., then (H, 1/2(h + k ), ) is invertible, thence h 1/2(h + k ) k . This being said, this extension is not associative a priori : the normal form is well dened for all lopsided feedbacks, either positive or negative, since in that case any system is semi-invertible. One easily gets the inequality : + h + h (19.26)
But this all what commutation to sups or infs can oer us. One must use other techniques to get associativity.
19.B.3
The resolvant
In what follows one supposes which costs little, but simplies the exposition that our system is injective. Remember (section 19.A.1) that one can transform any system into an injective one by removing the deadlocks . We shall write our feedback as the sum of two independent biased feedbacks, = + + , what one will note = + ; , are thence projections ; one poses := I 2 , thence I = + , which corresponds to a decomposition
19.B. GENERAL CASE
465
a b d H = R + P + N w.r.t. which one can write h = b c e . What follows d e f requires a certain familiarity with unbounded operators, see section 19.C.3. Observe that c and f are positive ; they thence admit positive square roots, which are injective by hypothesis : these square roots admit inverses , , which are partial operators, dened on dense subspaces (the ranges of c, f ) and with closed graphs (the inverse has the same graph). Theorem 79 (Resolvant) The coecients of the block-matrix : a b d b 0 e are densely dened and closable. d e 0
Proof : From h I , one gets 2 ( x | b (y ) ) (a)(x) | x + ( c)(y ) | y , which forces the discriminant | x | b (y ) |2 ( a )(x) | x + ( c)(y ) | y 1/2 c)(y ) ; if to be negative. This yields b ( y ) a ( 1/2 y := ( c)(z ), one gets b (z ) a z . Since b is total and has a dense domain, it has a dense domain too ; thence b extends to an of norm by a 1/2 , its closure. By continuity, bounded operator c = b , hence c = b : b is indeed the bounded operator . By taking into account the system (H, h, ) (which has dierent blocks), one concludes that e is total and bounded, and that e closes into a bounded operator. Symmetrically, one sees that d, e are bounded operators and that d , e have bounded operators as closures. Consider e ; writing it (e ) , one sees that it is densely dened. On the other hand, it is included in (e ), which is of closed graph : it is indeed its closure. Idem for the coecient e. 2 Definition 103 (Resolvant) The resolvant of the system (H, h, ) is the block-matrix : a b d a B D res(h, ) := b 0 e = B 0 C . D C 0 d e 0
All coecients are bounded operators, but perhaps C, C . Remember that P + C C and N + CC are invertible, with positive inverses of norms bounded by 1 ; also, C (P + C C )1 and C (N + CC )1 are bounded operators. Although far from ocial denitions, the system (H, res(h, ), ) is invertible. P C One already had the occasion to invert a block-matrix n := , see C N the proof of proposition 50. What suggests the following denition :
466
Definition 104 (Normal form) One calls normal form of (H, h, ) the operator h := a + B (P+C C )B B C (N+CC )1 D D C (P+C C )1 B D (N+CC )1 D . Proposition 53 In the case of a positive feedback, the normal form of denition 104 coincides with that of theorem 78. Proof : In the invertible case, h = a + b ( c)1 b = a + (b ( c)1/2 )(( c)1/2 b). In other terms, h (x) | x = a(x) | x + ( c)1/2 b(x)
2
(19.27)
In the general case, dene for 0 < < 1, c := c + (1 ) , so that c is the sup of the c with the c invertible. Lemma 79.1 If y is in the range of c, then ( c)1/2 (y ) = sup ( c )1/2 (y ) . Proof : u := c ( c )1 c is an increasing family of hermitians bounded by . It thence admits a supremum u which is, by bonication (see section 19.C.1) a strong limit. We are indeed in the commutative (von Neumann) algebra generated by c, hence ( c ) c( c )1 c = c( c )( c )1 c = c admits as a limit, by the strong product, ( c)u, continuity of the bounded 1/2 which proves u = . If y = c ( x ) , then sup ( c ) (y ) = 1/2 1/2 c(x) | ( c ) c(x) = sup u(x) | x = x 2 . sup ( c ) 2 The lemma enables one to pass to suprema in equation (19.27), the left hand side corresponding to theorem 78, the right hand side to denition 104. 2 Theorem 80 (General associativity) + h = + h (19.28)
Proof : h = res(h, ) ; since (H, res(h, ), ) is l.s.i., res(h, ) = res(h, ) = res(res(h, ), ) = res(h, ) . Since res(h, ) is not of norm 1, one must be careful with these equalities : res(h, ) = res(h, ) (corollary 52.1). : Case u.s.i., then passage to the supremum
res(h, ) = res(res(h, ), ) : Proposition 53 still works in this case. res(res(h, ), ) = res(h, ) : A property of associativity in an invertible, albeit unbounded, case. In the same way, h = res(h, ) . 2
19.C. COMPLEMENTS ON OPERATORS
467
19.B.4
The normal form theorem
Theorem 81 (Normal form) The normal form of denition 104 enjoys the following properties which dene it uniquely : (i) It solves the equation in the invertible case. (ii) It commutes to directed sups of l.s.i., to directed infs of u.s.i. (iii) It is associative, ( + ) h = h , for all , independent. Proof : We begin with unicity : the normal form is uniquely dened in the invertible case, than by sup or inf in the semi-invertible case, in particular for lopsided feedbacks. Finally, associativity denes h := + h = + h , a combination of semi-invertible cases. Let us look at the properties : (i) Immediate. (ii) The normal form commutes to directed sups of l.s.i., to directed infs of s.i.s. Indeed it splits into to steps (19.28) which both commute. (iii) In fact, the normal form is associative on the l.s.i., hence between positive feedbacks ; the same between negative feedbacks. Finally, the equation (19.28) shows associativity between a positive and a negative feedback. One concludes by writing + = + + + + + . 2 These conditions are quite amenable ; thence one can show that : Proposition 54 ( ) h = h (19.29)
by considering rst the invertible case, then the semi-invertible case by ordercontinuity, nally the general case using associativity.
19.C
19.C.1
Complements on operators
Topologies
On B(H), there are several topologies, in particular : Norm : The norm u = sup { u (x) ; x 1} makes product and adjunction continuous.
468
Strong : (ui) converges strongly to u when for all x H ui (x) u(x) converges to 0. The strong topology is slightly weird : indeed, adjunction is not strongly continuous. On the other hand u, v ; uv is strongly continuous, provided the argument u remains bounded in norm. Weak : (ui ) converges weakly to u when for all x H ui(x) | x converges to u(x) | x . Adjunction is weakly continuous, but the product is only separately continuous, what is of no interest. As a compensation, the unit ball B1 (H) := {u; u 1} is weakly compact. The inequality : | u(x) | x | u(x) x u x 2 explains the relative strengths of these topologies. Remember that a stronger topology has more open sets, thence makes convergence more dicult. A directed and bounded family of hermitians admits a (weak) supremum : h(x) | x := sup hi (x) | x
iI
(19.30)
There a sort of Dinis theorem , see [59], p. 307. : Proposition 55 If h = supiI hi , then hi h in the strong topology. In mathematics, function spaces always bear several topologies (four, ve on B(H) for instance), and vehicles to make them communicate. Which contrast with the arrogant navete of certain logicians which contend that they can reduce everything to a single topology, not even Hausdor ! A single topology, this is the all terrain vehicle which does everything thence, does nothing. The topology la Scott on R, it is the one making continuous the l.s.c. functions : its opens are the ]a, +[, a [, +] ; since its is not Hausdor, it oers little interest. Mathematicians (Dini, proposition 55) are perfectly aware of the interest of order-continuity, but prefer to use it as a bonication between good topologies rather than introducing a mediocre topology devoted to the sole l.s.c. (and not adapted to u.s.c. functions, for instance). But all of this is a matter of good taste ; and I can hear the sophism : How do you dene good taste ?
19.C.2
Matrices and blocks
On uses two styles of matricial writing : Block matrices : In the case of a direct sum decomposition H = n 1 Hi , if i is the orthogonal projection of Hi , one can write any operator f of H as the sum ij i f j . For n = 2 :
19.C. COMPLEMENTS ON OPERATORS
469
f=
1 f 1 1 f 2 . These block-matrices compose in the usual way, 2 f 1 2 f 2

j
i.e., hik =
fij gjk . But they are not true
matrices.
Matrices : They correspond to an isomorphism B(H) B(K) Mn (C). f f In other terms, in 11 12 (remark the dierent graphical style), the f21 f22 coecients belong to B(K), and not to B(H). One can relate these notions in a particular case : if one is given partial isometries ij , such that ii = i , ji = ij , ik = jk ij . Then f can be written as the true matrix (with coecients in B(H1 )), fij = 1i f j 1, e.g., f 11 f 21 . f = 11 11 12 f 11 12 f 21
19.C.3
Unbounded operators
(u) = (u) (19.31)
If u B(H, K), the graphs (u), (u ) H K are linked by the relation : If u is a partial operator from (a subspace of) H into K, the adjoint (dened by (19.31) makes appear two dual hypotheses :
Density : the domain of u est dense. In this case, u is a partial operator of closed graph. Closure : the graph of u is closed : in this case, u is densely dened, but need not be a graph ; which obviously subsists when u is closable, i.e., when the (u) is a graph. u is of closed graph when : xn x and u(xn ) y imply u(x) dened and u(x) = y . u is closable when : xn 0 and u(xn ) y imply y = 0. An unbounded operator is a partial operator of dense domain and closed graph. From what precedes, an unbounded operator has an unbounded adjoint. The typical unbounded operator is the inverse of an injective operator of B(H). A total unbounded operator is bounded (closed graph theorem). Unbounded operators poorly compose, however : ub is closed when u is closed and b is bounded. If u is densely dened, ub need not be densely dened. bu is densely dened when u is densely dened and b is bounded. If u is closed, bu need not even be closable. When u is unbounded, u u is dened on a dense subspace of the domain of u. Moreover it is hermitian (thence closed) and u u + I is injective, with image H, and inverse an operator of norm at most 1. See [59], section 2.7.
Chapter 20 Finite and hypernite GoI

In this chapter, I shall present the two ideas of niteness at work in con Neumann algebras. Grosso modo, it is a matter of spaces which are innite from the outside, nite from the inside ; last but not least, these spaces are not non-standard . By the way, since non-standard refers to a preexisting standard, what could be non-standard foundations ? These ideas are more than fty years old ; but it occurred to nobody, and above all not to the foundational Jurassic Park, that this niteness, this hyperniteness could be of some relevance to foundations. These people are currently discussing the hottest matters, say the latest developments of Hilbertian nitism : witness the list http://www.cs.nyu.edu/mailman/listinfo/fom. After all, in 1452 in Byzantium, the big question was that of the sex of angels, not that of the Turks who were camping beyond the ramparts.
20.1
Von Neumann algebras
The general solution of the feedback equation (section 19.B) involves inversions, norm limits, square roots, which all are C -algebras operations ; and also directed sups which are not C -algebraic. A C -algebra with directed sups, thats what we need : this is (more or less) the denition of a von Neumann algebra, for instance B(H). Among such algebras, some are styled nite, some are styled hypernite ; here are two notions which owe nothing to logic, and do not match Hilbertian nitism. So, what about taking von Neumann seriously by interpreting logic in an algebra both nite and hypernite ?
20.1.1
Introduction : commutative case
Theorem 82 (Extremely disconnected spaces) Let X be a compact space ; the following conditions are equivalent : 470
20.1. VON NEUMANN ALGEBRAS (i) Any bounded family of C(X ) has a supremum in C(X ). (ii) The closure of an open set of X is still open. In case on says that X is extremely disconnected (e.d.).
471
Beware : the supremum in C(X ) need not be pointwise. For instance, in C([0, 1]) the sup of the fn (x) := x1/n equals f (x) := 1, while the pointwise sup is discontinuous. Just like Scott domains, e.d. spaces are eccentric topologies : while the former are too coarse, the latter are too discrete. While the former were lattices in disguise, the latter are hiding measure spaces. Indeed, the clopen sets of an e.d. topology form a complete Boolean algebra : it is enough to dene i Oi := i Oi . If (M, ) is a measured space, then L (M, ) is a C -algebra admitting bounded sups ; by the way, those are pointwise sups. As a C -algebra, L (M, ) thence writes C(X ) for an e..d. space X . X is unique, while two equivalent measures (M, ), (M, ) i.e., with the same negligible sets will dene the same algebra. Since e.d. spaces are primarily complete boolean algebras, one can by the way pull back any space C(X ) to the form L (M, ) ; although this expression is not unique, it has the immense advantage of involving natural notions. Typically, compare , a completely standard algebra, to C( N)), the continuous functions over a bizarre space, the Stone-Cech compactication of N. Summing up : The von Neumann spirit diers from the C spirit : continuity is not measurability. To disguise a measurable function as a continuous function on a warped space is possible, but dishonest. One will thence try to complete to the sups a C -algebra ; the question has no intrinsic sense. Indeed, to pass from C(X ) to L (X, ) depends on the choice of a measure ; but, on the sole space [0, 1], there are non equivalent diuse measures. What resembles most to measure in the non-commutative case, is a state, see section 17.A.7 ; among states, those enjoying (uv ) = (vu), the traces, are sorts of non commutative measures . What we are actually interested in a C -algebra with bounded directed sups. But one is merely able to work on B(H) ; this is why one restricts to those algebras (the W -algebras) which are isomorphic as C -algebras to a subalgebra of some B(H), the isomorphism respecting directed sups. Since W -algebras have no intrinsic denition, one falls back on the concrete subalgebras of some B(H) ; even if they are taken up to a -isomorphism.
472
CHAPTER 20. FINITE AND HYPERFINITE GOI
20.1.2
Basic denitions
Definition 105 (Von Neumann algebras) A von Neumann algebra is a sub-C -algebra A B(H) closed under bounded directed sups. There are many characterisations of von Neumann algebras, e.g. : (i) The closure in the strong topology. (ii) The closure in the weak topology. (iii) The equality to the bicommutant. This last characterisation is the most popular. If A B(H) is closed under adjunction, the commutant Ac := {u B(H); a A au = ua} is a von Neumann algebra ; indeed Ac = Accc. Example 7 L (M, ) operates by multiplication over L2 (M, ) ; one easily checks that it is equal to its commutant, hence to its bicommutant. Definition 106 (Factors) A factor is a connected von Neumann algebra, i.e., whose center is reduced the scalar multiples of the identity. The theory of von Neumann algebras reduces to the study of factors. Since the center of a von Neumann algebra A is of the form L (M, ), A can be written as a sum of factors a sort of integral indexed by its center. Typically, when A is commutative, these factors are all isomorphic to C.
20.1.3
Finite algebras
Definition 107 (Finiteness) Between the projections of a von Neumann algebra A, one denes the preorder relation , with associated equivalence : u (u u = and uu = ) ( = and ) (20.1) (20.2)
A is nite when I stands alone in its equivalence class : uu = I uu = I . Theorem 83 (Type) If A is a factor, the preorder is total.
The preorder induces a classication of the factors over a separable Hilbert space :
20.1. VON NEUMANN ALGEBRAS In : Order type {0, . . . , n}. I : Order type N {+}. II1 : Order type [0, 1]. II : Order type [0, +]. III : Order type {0, +}.
473
Beware ! + is used in the case of an innite algebra to speak of the class of the identity. Let us discuss the types I and II : Type I matches the algebras which are -isomorphic to B(H), with H of Hilbertian dimension n or 0 . Those algebras are of little interest, in the sense that the theory was not intended for them. One sees that the order corresponds to the set-theoretic cardinal except that it stays denumerable. The type II is the most interesting for us. II resolves as a tensor product II1 I , which makes of II1 the real novelty. They are nite factors (the other being the In ), for which there is a notion of dimension. In a type II1 factor, a closed subspace is of dimension 1/2 when its orthogonal projection is such that I , etc. One must see the type II1 as another cardinality, in a sort of pointless setting : since every projection can be halved, one never reaches the atoms ! Dimension extends by linearity to linear combinations of projections, then to the full algebra by strong continuity : this is the trace : nite algebras are algebras with a trace.
20.1.4
Algebra of a discrete group
If G is a discrete (denumerable) group, the Hilbert space 2 (G) of square summable sequences indexed by G is equipped with a convolution : ((xg ) (yh ))k := xg yh
gh=k
(20.3)
(xg ) (yh ) (G) ; one denes A[G] as the set of all sequences (xg ) such that, for all (yh ) 2 (G), (xg ) (yh ) 2 (G). This is indeed a von Neumann algebra : (xg ) := ( xg1 ) yields the adjoint ; on the other hand, A[G] is the commutant of the set [G]A, made of the operators acting on the right : (yh ) ; (yh ) (xg ). The identity of A[G] is given by I := (1g ). A[G] admits a trace : tr((xg )) := x1 (20.4) i.e., a form enjoying tr(uv) = tr(vu), such that tr(u ) = tr(u), tr(u) = tr(u), and which is weakly1 continuous.
1
The ocial denition mentions another topology, styled ultraweak.
474 Proposition 56 A[G] is a nite algebra.
Proof : If uu = I , then the projection (xg ) := u u is of trace 1 ; but tr((xg )) = x1 = |xg |2 , hence x1 = 1, xg = 0 for g = 1.
Definition 108 (Condition i.c.c.) G is with innite conjugacy classes (i.c.c.) i, for all g G, g = 1, the {h1 gh; h G} of the conjugates of g is innite. Proposition 57 The algebra A[G] of an i.c.c. group is a type II1 factor. Proof : If (xg ) is in the center of A[G] and g = 1, then xg = xh1 gh is constant on its conjugacy class ; since (xg ) 2 (G), xg = 0. 2
20.1.5
Hypernite algebras
Definition 109 (Local finiteness) G is locally nite i any nite subset of G generates a nite group. In other terms, G is the union of a increasing sequence of nite groups. Usual groups such as Z are not locally nite. On the other hand, the group of bijections of a denumerable set which move only a nite set of points, is locally nite ; by the way, this group is also i.c.c. Definition 110 (Hyperfiniteness) A is hypernite when it is the closure (strong, weak, or the bicommutant) of the union of an increasing sequence of nite dimensional algebras. Hyperniteness is a property independent from niteness ; typically, an algebra of type III, hence badly innite, can perfectly be hypernite. Since nite dimensional algebras are (sums of) matrix algebras, hyperniteness means that one can approximate the elements of A by matrices in the ordinary sense of matrices with complex entries whose size is however not specied. Proposition 58 If G is locally nite, then A[G] is hypernite. Proof : If H G, then A[G] A[G]. Hence A[G] contains the directed unions of the A[G], where H runs among the nite subgroups of G. On the other hand, any u A[G] is weak limit of its restrictions uH : uH(g ) := ug (g H), uH(g ) := 0 (g H). 2 Theorem 84 (Murray-von Neumann) Up to isomorphism, there is only one hypernite factor of type II1 .
20.2. LALGBRE DE CLIFFORD
475
In particular, all i.c.c. locally nite groups induces the same algebra. This classic result shows that there are much less von Neumann algebras than C algebras ; indeed, the passage to weak limits erases the dierences due to the norm.
20.2
Lalgbre de Cliord
An excellent version for what follows is [83].
20.2.1
The three versions
The Cliord algebra of an Euclidian (i.e., real Hilbertian) space E exists in three versions : algebraic C (E), C -algebraic C [E] and von Neumann C {E}. In what follows, E is separable, i.e., admits a nite or denumerable orthonormal base (ei ) ; in nite dimension, the three versions coincide. Algebraic version It corresponds to the following universal problem : extend E into a complex algebra, i.e., : E C (E) in such a way that, for x E : (x)2 = x | x .I hence, for x, y E :2 (x)(y ) + (y )(x) = 2 x | y .I (20.6) (20.5)
In particular, (x)(y ) = (y )(x) when x, y are orthogonal. C (E) is thence the set of linear combinations of formal expressions (ea ), where a runs among nite sets of indices. The product is dened by : (ea ) (eb ) := (1)n (eab ) (20.7)
where is the symmetric dierence and n := {(i, j ); i a, j b, j < i}, which supposes to order the indices. Obviously, ((e{i} )2 = (e ) = I . C version
a (I + a= a (ea )) := I a= (1) a (ea ) is an involution of C (E). C (E) admits a unique C -norm i.e., enjoying uu = u 2 ; and can thence be completed into a C -algebra C [E].
2
This is why E is assumed to be Euclidian (real) and not complex.
476
If a est odd, (ea ) is a symmetry ; if a is even and a = , then i (ea ) is a symmetry. The universality of the solution has the following consequence : if : E F is an isometry, then there is a unique -morphism C [] from C [E] into C [F] extending . This what one calls a Bogoliubov morphism. In particular, if one writes E = lim (En , n ) as the direct limit of an increasing sequence of nite dimensional spaces, C [E] = lim (C [En ], C [n ]). vN version The linear form of C (E) : tr( a a (ea )) := extends into a state of C [E], hence the GNS construction (section 17.A.7) applies and yields a representation of C (E) ; this representation is faithful, since C (E) admits only one C -semi-norm3 . One denes C {E} as the bicommutant of this representation, in other terms, as the sets of weak limits of operators of C [E] or even of C (E). Since C (E) is the union of an increasing sequence of nite dimensional algebras, C {E} is hypernite. By the way, it is a nite algebra ; it is also a factor when E is of innite (type II1 ) or even (type I2n ) dimension. Remember that there is only one hypernite factor of type II1 (theorem 84, section 20.1.5).
20.2.2
Fock space
The antisymmetric Fock space is the Hilbertian version of the exterior algebra. If H is a Hilbert space and n 2, one notes H[n] the tensor product of n copies of H (with the notation instead of ). Definition 111 n (H) is dened by completion separation of the pre-Hilbert space H[n] , equipped with the unique sesquilinear form | 1 such that : x1 . . . xn | y1 . . . yn For instance, for n = 2, x y | x y
1 1
= det( xi | yj ) y|y x|y
(20.8) y|x .
= x|x
Definition 112 The Fock space (H) is dened as the direct sum of the n (H) (with 0 (H) := C, 1(H) := H). One denes an associative product ; beware, the product is only bounded (dened) as a bilinear map from p (H)q (H) into p+q (H) (bound :
3
p ). p+q
It is a simple algebra.
20.2. LALGBRE DE CLIFFORD
477
Remark 6 One can also see n (H) as the antisymmetric subspace of H[n], provided one denes x1 . . .xn := n!. (x1 . . .xn ). For instance, for n = 2, x1 x2 is not the antisymmetrisation 1/2.(x1 x2 x2 x1 ) of x1 x2 , but its renormalisation 2/2.(x1 x2 x2 x1 ).
20.2.3
Creators et annihilators
A simple calculation of dimension explains the relation between Cliord and Fock : if H is of dimension n, it can be seen as an Euclidian space HR of dimension 2n, which therefore induces a Cliord algebra of dimension 22n = 2n 2n , which is in turn the dimension of the algebra of endomorphisms of a space of dimension 2n , typically of the Fock space (H). The Fock representation thence interprets C (HR ) in B((H)). Definition 113 (Creators, annihilators) If a H, the creator (a) is dened as the operator : xax of (H). The annihilator (a) est dened as its adjoint, indeed :
n
(20.9)
(a)(x0 . . . xn ) =
j =0
(1)j xj | a x0 . . . x j . . . xn
(20.10)
Theorem 85 (CAR relations) Creators and annihilators enjoy the relations CAR ( canonical anticommutation relations) : (a) (b) + (b) (a) = a | b I (a)(b) + (b)(a) = 0 (a) (b) + (b) (a) = 0 (20.11) (20.12) (20.13)
To a H, one associates the operator (a) + (a) ; this operation is only (a). And one actually gets ((a) + (a))2 = a | a I , R-linear, since (a) = which provides a representation of the Cliord algebra C (HR ) (or C [HR ]) ; this representation depends upon the complex structure of H, which is by no means determined by the Euclidian space HR . If X indices an orthonormal base of H, then the set n (X ) of nite subsets of X indices an orthonormal base of (H). Up to the sign, the creator of ex corresponds to the partial function a ; a {x} dened when x a, and the annihilator to its inverse. One must thence understand (a) + (a) as the symmetry exchanging the half-space without a and the half-space with a .
478
20.3
20.3.1
A new nitism ?
Jurassic nitism
The ancient, Hilbertien, nitism, relies on an impossible hypothesis. Indeed, by acknowledging the sole natural numbers, if puts in front the innite set N of integers. And there is no way to avoid this basic misunderstanding ; this is by the way the deep meaning of the incompleteness theorem : nitism is not nite . There is no great divide nite/innite that would separate nitistic , predicativistic , etc., methods from the others ; only remain more or less religious, sectarian postures : one pretends.
20.3.2
Finite algebras
A type II1 factor is a space which is innite from outside, nite from inside. That it be innite from outside goes without saying : one can split the identity as a sum of n orthogonal projections of dimension 1/n, thence nd a unitary un such that un i = i+1 un (modulo n), which is a good approximation of the integer n. But, in an internal way, how can we, given un , build something like un+1 ? It would be necessary to make room for a new projection, by eventually cutting o those already there. A familiar method relies on the Hilbert hotels p, q of section 19.3.4 : these objects makes us increase or shrink as we please, and the pi p , qj q provide what is needed to duplicate our decomposition. Unfortunately, p, q exist no more than Alices mushrooms ; in any case they do not dwell in a nite algebra : p p = I , while pp = I .
20.3.3
Hypernite algebras
Section 19.3.5 enable one to build a discrete group G, an innite subgroup H G and g, h G such that g 1Hg, h1 Hh H commute. Take for G the group generated by the bijections m, n, p ; m, n , p and m, n , p ; n, m , p of N ; H will thence be the subgroup of bijections of the form m, n ; (m), n ; one will easily nd an internal automorphism sending H on the subgroup of bijections m, n , p ; (m), n , p (resp. m, n , p ; m, (n) , p ). Hence, if our decomposition (i ) is done inside A[G], it is possible to send it by an internal automorphism on (i ) or on (i ), hence to duplicate it into (i j ). But this group G is not locally nite.
20.3.4
A logic of LOGSPACE ?
The two possibilities of internal expansion have been barred :
20.3. A NEW FINITISM ?
479
One cannot create new locations in a nite algebra. Indeed, the trace takes care of the bookkeeping of space. Hyperniteness prevents us from creating new dialects , which would duplicate what exists, provided one restricts to twice smaller rooms. Finite and hypernite algebras thence provide us with a rst iconoclast idea on the nite : one can create nothing : even the function n ; n + 1 is problematic. On the other hand, there are arbitrary large integers. This is far from being absurd from the standpoint of complexity theory : the class LOGSPACE corresponds to algorithms without possibility of writing. What can be done in that case is to wander across the datum, without memorisation ; one can simultaneously read, say, 5 compartments, and according to what one reads there, modify the current state of the machine among, say, 17 possible states, etc. So, we should be ready for a logic of logarithmic space rather than of polynomial time !
20.3.5
Syntax and semantics
To speak of syntax is to speak of a subject, hence of a certain arbitrariness, and also of niteness. And, by the way, this niteness can be taken in a more nite sense than what is usually meant : it is indeed possible that the language is nite in an internal way. Assume that we are given a hypernite i.e., an i.c.c. locally nite group G. The subject will be able to use certain elements of the group, of which he will do as he pleases, basically linear combinations and products, i.e., convolutions. But he will never be able to construct an element which is not part of the nite subgroup thence generated. In this way we get internal niteness. One also sees that, if one succeeds in writing GoI in this case, it is an iconoclast system (chapter 16) that we shall deal with. Indeed, the computation time is bounded by the degree of nilpotency, which is in turn a priori bounded by the dimension of the space. Hence we get a termination a priori, the bound being the cardinality of the subgroup generated. Theorem 84 says that the choice of the group G is irrelevant. This is the case if one forgets the subject, if one works sub specie ternatitis. For instance, when one studying the asymptotics : weakly completeness becomes essential, and this unicity of the hypernite type II1 factor sounds like good news. As for myself, the dierence between semantics and syntax is the dierence between the full algebra A[G] and the set of nite combinations of points of G ; if I were working with the Cliord algebra, the dierence between A{E} and C (E). One thence sees that the choice of the hypernite group i.e., of the approximation is crucial. But this is no more than acknowledging the subjective aspect of syntax.
480
In the foundational vulgate, the language is innite ; indeed, one must be able to create an innity of copies, of occurrences , of the same symbol. But is this reasonable ? Should we try to do it concretely, physically, one would need a delocating machine, which would deposit copies of the symbol at distances of 1 km, 2 km,. . . . One sees that, as long as thought is concerned, one is ready to admit that the machine thus launched will never come back to its starting point. However, in the physical world, it is what happens ! And I remember my shock, when, almost a child, I heard about the impossibility of getting farther and farther from Earth. Hyperniteness is simply the niteness of space translated in the realm of thought. And no sophist will convince me that there should be internally speaking innitely many distinct symbols : let him show them to me. Might as well, the innity of galaxies supposedly created by God !
20.4
20.4.1
Finite GoI
Determinant vs. trace
One already observed (section 17.6.1) that quantum coherent spaces do not step out type In : the type I admits no trace, and, as to type II1 , it does not even allow the interpretation of the identity axiom ! However, it is possible to make use of type II1 , provided one replaces the trace with the determinant : instead of tr(uv ), one will use det(I uv ). In nite dimension, the trace appears as an explicitation of the determinant. Indeed, assuming a triangular w.r.t. an orthonormal base e1 , . . . , en of H, with diagonal entries 1 , . . . , n , one sees that det(I + a) = (1 + 1 ) . . . (1 + n ). In general, det((aij )) = (1) a1(1) . . . an(n) sums up all exhaustive travels of the graph represented by a ; and det(I + a) sums up all travels, exhaustive or not. The Fock space H (section 20.2.2) admits a base indexed by ({1, n}), the ( iI ei ). In particular, tr((a)) = 1+ 1in i + 1i<j n i j + . . . + 1 . . . n = det(I + a). Hence, det(I uv ) = tr((uv )) = tr((iu)(iv )). In other terms, the determinant reduces to the trace, modulo the replacement u ; iu. But this Fock space is indeed the space of all travels, exhaustive or not, that one can perform in H. In other terms, replacing the determinant with the trace amounts at replacing the potential with the list of possibilities. With classical logical tools, one can always replace the potential with a catalogue, in the (old-)fashion of those debilitating Kripke models. For the rst time, immanent justice strikes back : this replacement diverges, for deep reasons linked to the passage at the limit in von Neumann algebra. The equation tr((a)) = det(I + a) can only subsist in a type I algebra, where precisely
20.4. FINITE GOI
481
the type no longer exists ; indeed, the factors of types I , II are semi-nite : one can dene a trace for the sole trace class operators ; unfortunately, the operators used in GoI are not of trace class !
20.4.2
The scalar invariant
In a factor of type II1 , one cannot bluntly dene det(I u). However, if the spectral radius r (u) is < 1, then (see equation (17.43), section 17.A.3) the power series log(I u) := u u2 /2 u3 /3 . . . converges. In this case, one can, according to proposition 48, dene the determinant by : det(I u) := etr(log(I u)) (20.14)
Under the proviso of convergence, one can prove the usual properties of the determinant, for instance : Proposition 59 det(I uv ) = det(I vu) Proof : r (uv ) = r (vu) shows that both sides are simultaneously dened. From tr((uv )n /n) = tr((vu)n /n), one concludes tr(log(I uv )) = tr(log(I vu)) thence det(I uv ) = det(I vu). 2 Theorem 86 (Scalar invariant) W.r.t. a decomposition I = S + T, let u = a b c d such that r (a) < 1,
r (d + c(S a)1 b) < 1, and r (u) < 1 ; then det(I u) = det(S a) det(T (d + c(S a)1 b)). Proof : det S a b S a b + (S a)(S a)1 b = det c T d c T d c(S a)1 b 1 = det(S a) det(T (d + c(S a) b)). This is established at the level of power series, since everything converges. 2
The theorem yields a scalar invariant for the normal form of GoI (chapter 19) : indeed, if I = + , with 2 = R, 2 = S, and if (H, v, + ) is a cutsystem, one sets u := v ( + ), and one sees that, provided invertibility holds, v = d + c(S a)1 b. By the way, in nite dimension : det( + v ) = det( Sv S) det( v ) (20.15) In the logical case, S a is nilpotent, thence det(S a) = 1, which shows that we have indeed a sort of scalar invariant, which generalises the invariant of quantum coherent spaces, see equation (17.26), section 17.5.1. By the way, remember that this invariant came in turn from stability and coherent spaces,
482
see equation (17.2) of section 17.2.1. One thence replaced the cardinality with the trace, next with the determinant ; which took twenty-odd years ! It will perhaps not take twenty-odd years to reconstruct to whole logic on the basis of factors of type II1 , but this will anyway take some time. This is why what follows is very sketchy.
20.4.3
Duality
In what follows, A is a factor of type II1 . Proposition 60 Let R be a projection of A distinct from 0, I ; then AR := {.(I R) + RuR; C, u A} is a nite algebra. Proof : It is the commutant of {v + Rw ; v B(I R), w Ac }. It the trace of A, which makes it a nite algebra. 2 Ar is not a factor : its center consists in the (I R) + R and is thence of dimension 2. In fact Ar writes C B, where B is a factor of type II1 . The elements of Ar will thence be written u. Definition 114 (Duality) | v , when : u, v AR are polar, notation u r (uv ) < 1 det(R uv ) R+ \ {1} A behaviour of base R is a subset G AR s.t. G = G. Proposition 61 A behaviour is never empty, it must contain 0 0. If u G et || 1, then 0 u G. Proof : The restriction to the 0 u permits to ignore (20.17). 2 Definition 115 (Sequents of behaviours) If S, T = 0 are disjoint and R = S + T = I , if G, H are of respective bases S, T, one can dene G, H of base R, by : G, H := { (u + v ) ; u G, v H} In what follows, one supposes that F writes F = (20.18) (20.16) (20.17)
F11 F12 w.r.t. the F21 F22 decomposition R = S + T, and [F ]u := F22 + F21 u(S F11 u)1F12 .
20.4. FINITE GOI Theorem 87 (Implication) F G H i for all u G : r (F11 u) det(S F11 u) [F ]u < 1 H
483
(20.19) (20.20)
Proof : First assume that F G, H and u G ; since 0 0 H, it follows that 0 u ( G, H), thence r (F11 u) = r (F u) < 1. If v H, then, for || 1, 0 v H, hence r (F (u + v )) < 1, in particular R F (u + v ) is invertible, which amounts at saying that T [F ]u v is invertible. One concludes that r ([F ]u v ) < 1 ; theorem 87 yields det(S F11 u) det([F ]u v ) = det(F (u + v ) = 1. Hence det(S F11 u) [F ]u H. The converse is proved in the same way. 2 In particular, one gets associativity : Proposition 62 (G, H), K = G, (H, K). Proof : This is proved as theorem 51. 2
20.4.4
Polarisation
In ludics the objects (designs) had a polarity of their own, see section 13.1 ; here the notion is desessentialised : it is only a hypothesis on the behaviour, i.e., on socialisation. Definition 116 (Polarisation) A behaviour is negative when its elements are of the form 0 u or 1 u, and has at least one element of the form 1 u. A behaviour is positive when its negation is a negative behaviour ; which means that it contains all the 0 with R+ {1}, and does not contain 1 0. Polarisation forces sequents of behaviours to contain at most one negative behaviour : only in this way they still will have a polarity, positive when then all are positive, negative when one them is negative. Polarisation enables one to understand what happens with the invariant in the additive case. The additives will combine two behaviours of the same base R and same polarity in G H (positive case) or G& H (negative case). For this one will need a fresh projection , i.e., commuting to everybody, and of dimension 1/2. If u G, then u G H. If v G, w H, then v + (I )w G & H.
484
What is compatible with the duality : it is enough to consider the case = 1, and then det(R (u)(v + (I )w ) = det(R uv ). But, if is fresh, det(R a) = det(R (I )a) and det(R a) det(R (I )a) = det(R a), which nally yields det(R (u)(v + (I )w ) = det(R uv ) = 1.
20.5
Ludics in C {E}
It is natural to interpret perfect logic by means of creators and annihilators. Indeed, in chapter 10, I insisted upon the perfective aspect of logic, i.e., upon non-reusability. What is perfectly caught by creators, since equation (20.11) yields (a)2 = 0 : one cannot twice create the same thing.
20.5.1
Creators et annihilators
One can reproduce the CAR relations inside a Cliord algebra, here supposed to be of innite dimension. Let (en ) be an orthonormal basis of E ; one can make it into a Hilbert space such that HR = E : ie2n := e2n+1 , ie2n+1 := e2n , what suggests the denitions : (n;) (; n) := := 1/2((e2n ) + i(e2n+1 )) 1/2((e2n ) i(e2n+1 )) = = = = (; n) 0 I 0 (n = m) (20.21) (20.22)
It is immediate that the CAR relations are satised : (n;) (n;)(m;) + (m;)(n;) (n;)(; n) + (; n)(; n) (n;)(; m) + (m;)(n;) (20.23) (20.24) (20.25) (20.26)
Corollary 87.1 |n;| := (; n)(n;) et |; n| := (n;)(; n) are projections. Proof : ((; n)(n;))2 = (; n)(n;)(I (n;)(; n)) = (n;)(; n) (; n)(n;)2 (; n). But (n;)2 = 0 by (20.24), hence ((; n)(n;))2 = (; n)(n;). 2 Indeed (n;) is a partial isometry with domain |; n| and image |n;|, see denition 90, section 12.6.4. Since their sum is I , these two equivalent projections have dimension, i.e., trace, 1/2. Proposition 63 The projections |m;| et |; n| (m, n N) generate a commutative subalgebra of C {E}. In particular, the product of a nite number of (m;) and (; n) is still a partial isometry.
20.5. LUDICS IN C {E}
485
Definition 117 If a and b are disjoint nite subsets of N, one denes (a; b) as the product of the (m;) for m a and the (; n) for n b, in the order induced by N. One denes |a; b| as the image of the partial isometry (a; b) : it is the product of the |m;| for m a and the |; n| for n b ; its dimension is 2{ ab} . Proposition 64 If (a b) (a b ) = , then (a a ; b b ) = (a; b)(a ; b ). The proposition illustrates the fact that, whatever order we choose to perform the product of the (m;) for m a and the (; n) for n b the result is the same, up to the sign. The case a = b = is special : (; ) = I is an isometry whose domain and image are |;| = I , i.e., the full space. This is the only case where the initial and nal projection are not orthogonal.
20.5.2
A notational conict
We already discussed in section 9.2.3 the importance of notations, of the ideological challenge hiding behind illegible notations : grosso modo, certain people would like to reduce mathematics to a bureaucratic compulsion : that of a lunatic who tears o wings from ies, puts them in two distinct heaps, and later checks that there are as many on the left and on the right. Here, we meet a serious notational conict between two intuitions : if a a , b b , then |a ; b | |a; b|. In other terms, the projection |a; b| is contravariant w.r.t. the inclusion of images. In which direction should we order, according to the projections, or according to the sets encoding them ? I rst followed the option of the associated projections, and discovered that it was illegible, including by myself. Simply because at this level, one essentially works with formal expressions in the style (a; b), and that the combinatorics of nite sets is more intuitive than that of operators. Hence, I choose to order my objects topsy-turvy , which will enable a combinatoric legibility. This is done to the detriment of the operator intuition ; but, after all, when one eventually comes back to operators, the order relation no longer plays any role.
20.5.3
Loci
Definition 118 (Loci) A locus is the pair = (a; b) of two nite subsets of N such that a b = . The space L of loci is naturally ordered : (a; b) (a ; b ) i a a et b b . Which corresponds to the inclusion of kernels, |a; b| |a ; b | . L is equipped with an involution : (a; b) := (b; a).
486
20.5.4
Relations between loci
We shall dene a certain number of symmetric binary relations between loci, hence capable of giving rise to various coherent spaces, see section 8.2. All these relations are preserved by the involution. Coherence Any non-empty family of loci, nite or not, admits an inmum, ai ; bi ) ; in particular there is a smallest locus, := (;). i (ai ; bi ) = ( On the other hand, the supremum (a; b) (a ; b ) := (a b, a b ) exists only in case a b = a b = , which is by denition the coherence (a; b) (a ; b ). It is immediate that a nite clique admits the supremum bi ). i (ai , bi ) := ( ai , Loci are coherent when their domains intersect, i.e., if their intersection is not reduced to {0}. Since (a a b b ) (a b)+ (a b ), the dimension of the intersection of the domains is greater than the product of the dimensions. Extraneousness Two loci are foreign, notation (a; b) (a ; b ), when they are coherent, of inmum , which can be expressed by (a a ) (b b ) = . Extraneousness ensures that the dimension of the intersection of the domains is equal to the . and product of their dimensions. Remark that i If , one can nd a unique locus such that = : if = (a; b), = (a ; b ), then = (a \ a, b \ b). The locus is noted \ , hence = ( \ ). Extraneousness is the assumption of proposition 64. Incoherence Incoherence is the negation of coherence ; two strictly incoherent loci thence have orthogonal domains and images. (a; b) is strictly incoherent to (b; a), but if a = b = . More generally, if c splits in two ways as c = a b = a b , then (a; b) (a ; b ). Antagonism Two loci , , are antagonistic, notation , when the four loci , , , are strictly incoherent. The simplest example is given by (0; 1) and (0, 1;). Observe that (a; b) is not antagonistic to (b; a). Multiplicatives are handled by extraneousness, additives are handled by antagonism.
487
20.5.5
Biases
Pitchforks
A bias is a non-trivial (i.e., distinct from ) locus. Biases form a coherent space, when equipped with the extraneousness relation . Ramications A ramication is a non-empty nite clique of biases. The space of ramications is equipped with a coherence relation : I J i there exist I, J such that . Since a ramication corresponds to a n-ary multiplicative, this is consistent with the idea of handling multiplicatives with extraneousness. The ramication is non empty, because ludics taught us that there is no multiplicative neutral. The coherence between ramications comes from the primary additive case : if two loci , are antagonistic, then { } and { } may belong to the same directory, i.e., are coherent. Which, modulo the distributivity multiplicative/additive, yields our denition I J . Directories A directory, it is a nite clique R of R. Observe the restriction nite . Indeed, one is no longer interested in innite etaspansion ; only in real nite logic, thence with a nite number of cases for each rule, etc., to sum up, in nite directories. Carrier of a clique Let X be a nite clique of loci (i.e., w.r.t. coherence), and let := { ; X }. Any element X induces a unique decomposition = ( \ ) of as a supremum of foreign loci. If one denes X := ( \ ) , it is immediate that the X are pairwise (strictly) incoherent : in other terms, their initial projections are pairwise orthogonal. Definition 119 (Carrier) as the sum If X is a clique of loci, one denes X X X . This is a partial isometry, whose image |X | := X |X | is called the carrier of X . Examples takes a rather simple shape. If X is made of pairwise foreign loci, then X For instance {, , } = .
488
Moreover, if X , the shape is peculiar : {, , , } = . The most general case is that of coherence. Take for instance will be made of three dierent splitX := {(01; 24), (03; 56), (1; 46)} ; X tings of c := {0, 1, 2, 3, 4, 5, 6}. More precisely, = (0156; 234) + (0234; 156) + (125; 0346). X In this denition we perceive an essential novelty, the role of context. , once put into context , splits according to the context. Thence, in a context , which is foreign, becomes ; one sees that is used as a context from through . In other terms, the dichotomy / separates the direct use (as ) from the contextual use (as ), what the paper [43] tried to do in an awkward way by introducing dialects to speak of the contextual use. In particular, if is separately put into context with , , which produces and , one sees that depending whether , will be in multiplicative (foreign) or additive (antagonistic) position, one will get coherence or incoherence, thence, either the possibility of working together , or blunt exclusion, the possibility of superposing . Pitchforks Definition 120 (Pitchforks) A pitchfork is a pointed clique of loci ; moreover, if the minimum locus occurs in the pitchfork, he is the point . A positive pitchfork is thence a pitchfork with point . A pitchfork is denoted , where the handle is the point . To each pitchfork one associates a carrier | |, which is by denition its carrier as a clique.
20.5.6
Examples
To each nite design one associates a hermitian in C (E). If the design is of base , the operator will be of domain/image included in | |. What I will do by contenting myself with representative examples. Daimon and Faith On a positive base say (0;), (; 1) , of carrier thence equal to |1; 0| + |01;| + |; 01|, the daimon is interpreted by the projection |1; 0|, whereas the faith (i.e., absence) is simply rendered by 0.
489
Tensor
rule
Take a Tensor : the typical example consists in cutting o a positive base (0;), (1;), (2;) along two foreign biases say (3;), (4;) , with a change of polarity, which yields : (03;) (1;) and (04;) (2;). Assume that we are given hermitians h, k of respective bases included in |03; 1| + |1; 03| and |04; 2| + |2; 04|. Here we need a freshness hypothesis, i.e., that 3 and 4 have not been used to build h, k . By multiplying h by the projection (; 24), one get h of image included in |03; 124| + |1; 0234| ; one produces in the same way k of range included in |04; 123| + |2; 0134|. One must still represent the rst action : remark that the carrier of (0;), (1;), (2;) is |; 012| + |0; 12| + |1; 02| + |2; 01|. The operator (partial symmetry) of image |34; 012| + |0; 1234| dened as g := ((0; 34)+(34; 0))|; 12| is of image orthogonal to those of h , k . The desired interpretation is thence h + k + g .
With
rule
Take a With : here one tries an additive cutting o of a negative base (0;) (1;) according two antagonistic biases say (23; 4) and (2; 34) which will yield the bases (023; 4), (1;) and (02; 34), (1;). The focus is put on the shared context (1;). Assume that one is given h, k of images respectively included in |4; 0123| + |023; 14| + |14; 023| and |34; 012| + |02; 134| + |134; 02| which are orthogonal. Our elegant handling of context enables one to sum up h, k , but remains a small problem : one would like the sum to be of image included in |0; 1| + |1; 0|. One can, using the partial symmetry ((0;) + (; 0))(|4; 123| + |34; 12|), replace h, k with h , k of images respectively included in |04; 123| + |023; 14| + |14; 023| and |034; 12| + |02; 134| + |134; 02|. The sum h + k is the desired interpretation.
20.5.7
A question of freshness
Although we didnt dene it rigourously, one sees that the interpretation of perfect logic supposes the systematic use of fresh biases, i.e., new. In particular, two occurrences of the same A, say A and A will use foreign parts of C {E}, for instance A (resp. A ) will only make use of the (a; b) such that the elements of a, b are even (resp. odd). In those circumstances, one sees that second order quantication is called into question for real reasons geometrical ones, not dubious quibble about predicativity . Indeed, if I want to interpret X X , where X and X only dier by their biases, even for X , odd for X , I am penniless, since C {E} being hypernite, it contains no internal operation enabling to permute even and odd. I can perhaps do it when X assumes the modest dimension of a
490
single formula A, with a nite stock of instructions ; on the other hand I am confused when X denotes the potentiality of such an A. A solution would be to act like in ludics and admit innite directories. This poses no problem, but one neglects by the way the nite character of the language in favour of a vaguer nitary character. In other terms : one neglects hypernitism.
20.5.8
Bogoliubov automorphisms
The idea is to enrich our algebra with Bogoliubov automorphisms (section 20.2.1), for instance the one induced by the exchange 2n/2n + 1 : One can always add automorphisms to a factor. This is called a crossed product. A good hypothesis is that of a discrete group of external automorphisms, see [60] : thence the crossed product remains a factor. If, moreover the factor is hypernite and the group is locally nite, then the result is hypernite. Bogoliubov automorphisms will handle quantication, and more generally boxes. Since one can construct matrioshka boxes, one must iterate the crossed products, what becomes more or less illegible. This is why we shall directly construct a hypernite group of boxes . Its algebra contains C {E} as well as enough global operations.
20.A
20.A.1
The algebra of boxes

The group B0
If a1 , . . . , an V , then {a1 , . . . , an } V .
The hereditarily nite sets are constructed as follows :
There is a bijection between V and N, hence V is totally ordered : ({a1 , . . . , an }) := 2(a1 ) + . . . + 2(an ) One denes B0 as the set of all expressions = a, with a V and Definition 121 (Sets in B0 ) If 1 = 1 a1 , . . . , n = n an B0 , then [1 , . . . , n ] := b, with : b is the set of those ai occurring an odd number of times. :=
1
(20.27) = 1.
...
n,
where := (1) {i<j ;ai >aj } .
20.A. THE ALGEBRA OF BOXES The denition is better understood from particular cases : [] [ a] [, ] si a < b, alors [ a, b] et [ b, a] = + = {a} = + = {a, b} = {a, b}
491
(20.28)
Definition 122 (Symmetric difference) One denes the composition law on B0 by : [1 , . . . , n ] [1 , . . . , m ] := [1 , . . . , n , 1 , . . . , m ]. B0 is a (non-commutative) group, with unit +. Observe that ( a)2 = (1) (a)+1 when a = , thence the elements of the group are either involutive, 2 = +, or of square equal to ; in any case, 4 = +. With | a| := a, let us compute the conjugacy classes : are central, in particular, {+, } is a distinguished subgroup. The quotient |B0 | := B0 /{+, } is isomorphic to V equipped with the usual symmetric dierence. If || = , then the conjugacy class of is {, }.
20.A.2
The von Neumann algebra A[B0] is not a factor. The calculation of conjugacy classes shows that it is the direct sum of two algebras : The sequences (xg ) such that xg = xg , which form a commutative algebra isomorphic to L ([0, 1]). The sequences (xg ) such that xg = xg . This algebra is exactly C {E}.
B0 vs. C {E}
20.A.3
The group B
Definition 123 (Odd bijections) A bijection of B0 (on itself) is odd when () = () for all B0 ; if is odd, one denes another odd bijection of B0 by : ([1 , . . . , n ]) := [(1 ), . . . , (n )] (20.29)
0 One denes, for n 0 and B0 , odd bijections n de B :
Proposition 65 is an automorphism of the group B0 . If is the identity, then = , moreover ( ) = . 0 ( ) +1 n := := n
(20.30)
492 Proposition 66
+1 with := n ( ).
+1 0 n+1 n 0 =
(20.31)
Corollary 66.1
n n
= =
n
n+p n
n+p n
(p > 0 ; := p ( ))
(20.32)
Definition 124 (Group of boxes) n for n Z, One denes the group of boxes B , by means of the generators B0 , and the relations : n = = = 1 n n n+p (p > 0 ; := p ( )) np n1 , with ... p 1 (20.33)
n n+p
n n
Any element of B can be uniquely written as n1 > . . . > np and 1 , . . . , p = +. Proposition 67 B is a i.c.c. group. Proposition 68 The group B is locally nite.
Indeed, if X B is nite, it generates a subgroup whose cardinal is bounded by a tower of exponentials 2k ( (X )) (see equation (4.1), section 4.C.4), where k is the amplitude of the upper indices ni occurring in the expression of the elements of X . I therefore plan to interpret logic in A ; as a von Neumann algebra, it is isomorphic to C {E}. But it has not at all the same syntactical operations. 1 In particular, the must be seen as the elements of a box manipulated 0 . by the

Coursang5 PDF

Загружено:

Сведения о документе

Исходное описание:

Оригинальное название

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

Coursang5 PDF

Загружено:

Авторское право:

Доступные форматы

Part V Iconoclasm

Chapter 16 Heterodox exponentials

The quarrell of images

16.1. THE QUARRELL OF IMAGES

CHAPTER 16. HETERODOX EXPONENTIALS

CHAPTER 16. HETERODOX EXPONENTIALS

Exponentials and integers

The lesson of nets

Neglecting the technicalities linked to , see section 7.4.2 !

16.3. RUSSELLS ANTINOMY

CHAPTER 16. HETERODOX EXPONENTIALS

16.4. LLL AND ELL

Russell and normalisation

LLL and ELL

Either = ! and for i = 1, . . . , n, i =? ; in the case of LLL, n 1. Or = and for i = 1, . . . , n, i = or i = ?.

CHAPTER 16. HETERODOX EXPONENTIALS

16.4. LLL AND ELL

Bounds for LLL

Bounds for ELL

CHAPTER 16. HETERODOX EXPONENTIALS

In LLL, integers are typed by :

Coding of polynomial time

One must code integers in base 2 :

16.5. EXPRESSIVE POWER

Remark on the paragraph

On denes, without scheming :

Chapter 17 Quantum coherent spaces

Logic and quantic

Caracteristics of the quantic

17.1. LOGIC AND QUANTIC

CHAPTER 17. QUANTUM COHERENT SPACES

This shift of viewpoint took me 30 years of (part time) reection.

17.2. PROBABILISTIC COHERENT SPACES

PCS and QCS

Probabilistic coherent spaces

CHAPTER 17. QUANTUM COHERENT SPACES

The bipolar theorem

17.2. PROBABILISTIC COHERENT SPACES

What makes sense due to the niteness of |X |.

CHAPTER 17. QUANTUM COHERENT SPACES

Quantum coherent spaces

17.3. QUANTUM COHERENT SPACES

CHAPTER 17. QUANTUM COHERENT SPACES

The bipolar theorem strikes back

17.3. QUANTUM COHERENT SPACES

Norm and order

With the notations of theorem 61, in particular C + :=

is identical to the equivalence class of 0 2

CHAPTER 17. QUANTUM COHERENT SPACES

(iii) The unit ball w.r.t.

(iii) If x FinC , then there exist y, z

0 such that x = y z and y x .

17.3. QUANTUM COHERENT SPACES Semi-norm x

It is however not the case that | x | y | x y Fin C . Proposition 37 If C D , alors :

for all x FinC ,

Quantum coherent spaces

CHAPTER 17. QUANTUM COHERENT SPACES

CHAPTER 17. QUANTUM COHERENT SPACES

Commutative booleans There is little to say if one sticks to the

(i) The booleans true,false are represented by (ii) A diagonal matrix

CHAPTER 17. QUANTUM COHERENT SPACES

is obtained on the subspace corresponding to