A Simplification of Girard’s Paradox Antonius J.C, Hurkens Klaasstoksewog 7, 5443 NS Haps, The Netherlands e-mail: hurkens@sci..un.nl Abstract. In 1972 J.-Y. Girard showed that the Burali-Forti paradox ‘can be formalised in the type system U. In 1991 Th. Coquand formalised another paradox in U~. ‘The corresponding proof terzis (that have no normal form) are large, We present a shorter term of type L in the Pure ‘Type System AU~ and analyse its reduction behaviour. ‘The idea is to construct a universe W and two funetions such that a certain equality holds, Using this equality, we prove and disprove that a certain abject in U is well-founded, 1 Introduction Jean-Yves Girard (1972) derived a contradiction in the type system U by for- malising a paradox inspired by those of Burali-Forti and Russell. By formalising another paradox, Thierry Coquand (1994) showed that the type system U~ is also inconsistent. So there are large proof terms of type in these type systems. In Section 3 we present a relatively short term of type 1 in AU-. This Pure ‘Type System and some notation is described in Section 2. In the last section we show that the -reduction behaviour of the proof term is very simple. In the other sections we will see that the proof has the same ingredients as Burali-Porti’s paradox: a universe U/, a relation < on U, an object 2 in U, and the question whether {2 is well-founded or not. In Section 4 we describe Burali-Forti’s paradox and some simplifications. We analyse the connection between the universe of all ordinals at its power set. In Section 5 we introduce paradorical universes. These are connected to their power set in such a way that we can derive a Burali-Forti like contradiction. This can be formalised in Pure Type Systems. The formalisation can be simplified by considering powerful universes. In Section 6 we see how these universes are connected to the power set of their power set. 2 Pure Type Systems In this section, we describe some Pure Type Systems. For more details, se for example (Barendregt 1992) or (Geuvers 1993). 2.1, The Pure Type Systems AHOL, AU-, and AU ‘The typed A-calculus AHOL (iigher Order Logi) is the Pure Type System (with S-conversion) given by the sorts *, O, and A, the arioms +: and O: A, andzr the rules (+,¢), (0,0), and (0,+). It is a consistent system, even if one adds the rule (A,+), Adding the rule (A,C), one gets the Pure Type System AU. Adding both rules, one gets AU. 2.2. Typing Terms in a Pure Type System Each term A in a Pure Type System is either a variable 2, a sort s, a product Ix: B.C, an abstraction Ax : B.C, or an application (B C). By BiC/z] we denote the result of substituting the term C for the free cocourrences of the variable z in B (renaming bound variables if necessary). By ‘we denote the equivalence relation between terms that is induced by reduction: replacing a subterm of the form (Az : A.B C) by the term B[C/s]. Ifa term does not contain such a subterm, then the term is called normal In a Pure Type System, we can derive formal judgements 2 + Ai,-**42n : An +B: C, expressing that B has type C in the given context, that is, assuming that for i= 1,...,m, variable x, has type As We start in the empty context. If, in some context, A has type # for some sort s, then we are allowed to introduce a new variable z of type A. ‘The context gives the types of some variables. ‘The axioms give the types of some sorts. We use the rules (s',s) to type products as follows: if A has type s' and (under the extra assumption « : A) B has type s, then (in the original context) Tir : A.B also has type s. If liz : A.B has type s and (under the extra assumption z : A) C has type B, then (in the original context) Ar: A.C has type Tz : A. B. If F has type Iz : A. B and C has type A, then (F C) has type B{C/=] Finally, we use B-reduction to change types: if A has type B, B =3 C, and C has type s, then we may conclude that A has type C. Note that if a variable, abstraction or application has type A, then A is of type s for some sort. 2.3 Some Useful Properties of AU ‘Two terms A and B are S-equal if and only if for some C, both A and B reduce to C. If term B has a type D, then this type is unique up to #-equality. Furthermore, if B g-reduces to C, then Cis also a term of type D. We can calculate the level of a term (and its subterms) in a given context 2: Ayn! An a8 follows: The sorts +, 0, and 4 have level 2, 3, and 4, respectively. The level of variable x; is one less then the level of Ax in the context ny: Ajys++yt4-1 : Aina. The level of a product Ilr : B.C or an abstraction Az : B-C is the level of C in the extended context xy : At,---5ty : Anz: B. ‘The level of an application (B C) is the level of B in the original context. ‘One can prove that if B has type C in some context, then the level of B is cone less than that of C. So each term has level 0, 1, 2, 3, or 4. One can also ‘show that no term in AU contains a subterm of lower level (in the corresponding,288 context). This implies that if we use a rule (3,8) to form a product Iz : B.C, then level(z) < level(B) = level(s) ~ 1 = level(C), so the variable x has no free occurrence in C. It turns out that each term A of level 1 is strongly normalising: there is no infinite sequence A+ A’ 1g A" 9 --- of dereduction steps. The terms of higher level are normal, since each abstraction or application has level O or 1 2.4 The Five Levels of Terms in AU ‘We describe the five levels and introduce some notation to distinguish terms of x SM: y.x Va: S.x =r: S.x va: O.ySI¥:0.x Other connectives can be defined as usual. We only need falsehood and negation: ‘There are two kinds of abstractions and applications of level 1. We introduce some new notation only for the ‘polymorphic’ ones AYO. oT) Ax: 0, Here b and c are objects and T is a set.269 ‘There are three kinds of abstractions and applications of level 0: suppose ny. P = An: y.P let 2:5. P= 0:5.P let ¥:0.P =X 0.P Note that for proofs P and Q, the application [P Q] corresponds to modus ponens in a natural deduction. 3 A Term of Type 1 in AU~ We consider the following universe: U=TX:0.((ppX+2)-4pp%) For each term tof type ppl, we define a term of type TESAX:D.AY: (pe FA). Ap: pA. (CAE: U (P(E (CX) NY) For each term s of type U, we define a term of type ppl {s U) Mt: ppl.rt) (Go we do not consider ¢ and 7 as terms.) ‘We define normal terms of type pl and U, respectively: A= dy :U. Np: il. [(ou 9) > (P 700) 12 = the normal form of 7 Ap: pll.¥r :U.[(oz p) > (P 2)) In other words, 2 = AX: DAF : (pp¥A).Ap : pw. + U.[(ox Ay Ur (F (ty X} I) > F lz *} NY), ‘We claim that the following is a term of type 1 in AU: [suppose 0: Vp: pl. (Wz :U.[(o2 p) = (P 2)] = (P 9) (0 A) let x :W. suppose 2: (or A). suppose 3: Vp : ell. [(oz p) > (P t22)] {[(3-4) 2] let p pl (8 Ay :U.(p roy)) I] lt p: pU(0 Ay Up ra))} let p: pll.suppose 1: Vz :U. (ox p) = (p a)}-{(1 2) let 2:4. (1 rez] Note that each subterm (except for the term itself) is normal. One easily verifies that (in the empty context) there is no normal term of type 1 in AUC. At the cend of this article, we analyse the f-reduction behaviour of this proof term. ‘The proof is simple in the sense that it contains just 6 applications corre- sponding to modus ponens. In order to get an idea of the infiuence of abbre- Nations, one can also calculate the length: the total number of applications, abstractions, products, and occurrences of variables and sorts. For example, the terms abbreviated by L, U, A, and 2 have length 3, 15, 241, and 145. The complete proof term has length 2039. ‘In order to explain the idea of this proof, we first describe the paradox of Burali-Forti.270 4 Burali-Forti’s Paradox Cesare Burali-Forti (1897) published a result that lead to the first paradox in naive set theory. He showed that there are different ordinal numbers a and such that neither a < 8 nor § pT as follows: for ‘each subset X of S, f.X = {f2|2 in X}. Using this notation, we see that (U,0,7) is paradoxical if and only if the composition 707 is equal to (roa). Note that if (U,0,7) is paradoxical, then (g2, 04,74) is also paradoxical: 0, 07, = (707), (7 oo)ee = (T+ 04)e- (Here we need extensionality: if two sets have the same elements, then they are equal.)are 5.2 Example of a Paradoxical Universe Let UW be the universe of all triples (A,,a) consisting of a set A, a binary relation < on A, and an element a of A. For each triple (A, <,a), let o(A, <,a) be the collection of all triples of the form (4, ,b), where 6 ~ a. So @ is a function from U to gil. It induces a relation < on pli as follows: For all collections X and ¥ of triples, ¥ < X if and only if ¥ is in 04 X, that is, if ¥ is of the form o(A, <,a) for some triple (A, <,a) in X For each X in pl, let 7X denote the triple (p44, <, X). Now orX = o(pl.<,X) = {(U,2) and defined functions match : Ag > ppAa and intro : ppAo —> Ag. Then he showed that these functions constitute an isomorphism with respect to certain partial equivalence relations. In fact, (Ae;match, intro) can also be used to formulate a Burali-Forti like paradox: itis an example of a powerful universe. 6.1 From a Universe to the Power Set of Its Power Set and Back Let us call a universe U, together with functions « : U > ppl and r : ppl > U, powerful if for each C in ppl, orC = {X|the set {ylroy is in X} is in C} Each function f : S -+ T induces a function f* : pT + pS as follows: for each subset Y of T, f*Y = {x in S|fr is in X}. Using this notation, we see that (U,0,7) is powerful if and only if the composition « © 7 is equal to (r0.0)**. Note that if (U,0,7) is powerful, then (p2l,7*,0*) is also powerful: T*00* = (g0r)* = (roa)*** = (a*or*)**. (Here we do not need extensionality.)278 6.2 Example of a Powerful Universe Let U be the universe of all triples (A,,a) consisting of a set A, a function 5: A ppA, and an element a of A. For each triple (A, 3,0), let o(A,s,a) be the collection of all subsets X of such that {6 in Al(A, s,8) is in X} is in sa, Since o is a function from U/ to ppl, o** is a function from ppl to ppppl. For each C in ppl, let rC denote the triple (ppl, o**,C). In order to verify that (Ud, 0,7) is powerful, let Cin ppl and X in pl. Then the following propositions are equivalent (by definition): X isin oC; X is in o(ppU,0**,C); {bin pplil(ppl,o**,b) is in X} is in oC; o*{b in ppli|rb is in X} is in C; {y in Ulray is in X} is in C. 6.3 Contradiction from the Existence of a Powerful Universe Let (U,a,) be powerful. We will derive a contradiction in a similar way as for paradoxical universes Blements of U will be denoted by 2,y,... and subsets of U by X,Y,.... For each 2, a is in ppl. (U,0,7) is powerful, so oron = {X|the set {ylroy is in X) is in ox} We say that y is a predecessor of x (and we write y <2) if for each X in o, y is in X (in other words, if y is in (oz). One can easily prove that if y < + then tay < raz. We will only do this for the special case y = rox. Note that if X is in oz, then each predecessor of z is in X. X is called inductive if the following holds: for each x in U, if X isin o, then zis in X. We say that 2 is well-founded if z is in each inductive X. (Note that it is not clear whether {2[z is well-founded) is inductive: if one tries to prove this, one would like to use something like: if ¥ is in ox and ¥ C°X, then X is in oz.) Let 2 = r{X|X is inductive). (U,,7) is powerful, so: 19 = {X| the set {ylroy isin X} is inductive} We claim that 12 is well-founded: Let X be inductive. In order to prove that 12 is in X, we only need to show that X is in 0. In other words, we show that the set {ylray is in X} is inductive. So let zx be in U. Since X is inductive, we have the following: if X is in oroz, then raz is in X. In other words, ifthe set {ylroy is in X) is in a, then = is in {ylray is in X). This is exactly what we had to prove, Tin order to show that 2 is not well-founded, we first prove that the set {ulray # y} is inductive: Let x be such that {ylroy £ y} is in ox. Then rox ¢ x. For suppose that rox < 2. In other words, for each X in a2, raz is in X. Applying this to the set {ylroy ¢ y}, which is in oz, we see that rorox ¢ rox. On the other hand,276 ToTox < Tax: Let X be in pl. We have to show the following: if X is in orox, then roraz is in X. In other words, if {ylroy is in X) is in oz, then raz is in {y|roy is in X}. This follows from the assumption that rox (p O2)] > (p 0") Ya .[(ox O24) > -¥p : pll.[(o2 Orp) > (p O"*2)]) So gn expresses that "0 is in each subset X of U for which {yl6y is in X} is inductive. Note that yo is “is well-founded” and yn. is the normal formarr of "O™412 < ON”. The proposition yn expresses that {ylO"ty ¢ O"y} is inductive. We also define five proof terms: let p: pll. suppose 1: Vx :U.[(ox Op) = (p O"2)] (1) let 2 :4.(1 r02)) Mp = let 2 :U.suppose 2: (ox 03.) suppose 3: Vp: pli.|(ox Op) = (p 0™*12)] [[(3 A) 2] let p: pid. (3 Ay =U. (p ray))] suppose 4 : vn. suppose 0: yn: ({(0 A) 4] let p: pl. (0 Ay =U. (p r04))] Ln = suppose 0: Pn. [(0 A) My] let p: pld.(0 Ay :U(p rou))] Qn = suppose 4: Yn. [(4 2) let 2 Ul. (4 r02)] Then Ry proves Yn, Mn proves Yn, Pa Proves [Wn = Pn) Ln proves —y>n, and Qn proves [Un = ns) Note that [Lo Ro] is the proof term that we presented in Section 3, For each natural number n, [Px My] Rp] reduces in one step to [Ln Rn]- (Variable 4 disappears.) This reduces in a big step to [[Qn Mn] Raa). (Variable 0 disappears and some occurrences of 1 are renamed as 4.) This reduces to (Past Musi] Rosa}: (Variable 4 disappears and some occurrences of 2 and 3 are renamed as 4 and 0.) So these proof terms of type .L in AU~ reduce in three big steps to a similar proof term: only the types of the proof variables change a little bit. References Barendregt, H.P.: Typed lambda calculi, in: Handbook of Logic in Computer Science (Vol. 3), 8. Abramsky et al. (editors), Clarendon Press, Oxford (1992) Burali-Forti, C.: Una questione sui numeri transfiniti, Rendiconti del Circolo Matem- atico di Palermo 11 (1897) 154-164 Cantor, G.: Beitrige zur Begriindung der transfiniten Mengenlehre, Il, Mathematische Annalen 49 (1807) 207-246 (Coquand, Th.: An analysis of Girard’s paradox, in: Proceedings Symposium on Logic in Computer Science: Cambridge, Massachusetts, June 16-18, 1986, BEE Computer Society Press, Washington, D.C. (1986) 227-236 Coquand, Th.: A New Paradox in Type Theory, in: Logic and philosophy of science ‘in Uppsala: papers from the 9th international congress of logic, methodology and philosophy of science, D. Prawitz, D. Westerstaahl (editors), KIuwer Academic Publishers, Dordrecht (1994) ?~? Coquand, Th., Herbelin, H.: A-translation and looping combinators in pure type sys- tems, Journal of Functional Programming 4 (1994) 77-88 Geuvers, JL: Logics and Type Systems, Procfschrift, Katholicke Universiteit Nijmegen (1993) Geuvers, H., Werner, B.: On the Church-Rosser property for Expressive Type Systems and its Consequences for their Metatheoretic Study, in: Proceedings of the Ninth Annual Symposium on Logic in Computer Science, Paris, France, IEEE Computer Society Press, Washington, D.C. (1994) 320-329,278 Girard, 1-Y.: Interpréationfonctionnell et élimination des coupures de Varthmétique ordre supérieur, These de Doctorat d'Etat, Université Paris VII (1972) Howe, D.J: The Computational Behaviour of Girard’s Paradox, in: Proceedings Sym: posium on Logic in Computer Science: Hhaka, New York, June 22-25, 1987, IEEE Computer Society Press, Washington, D.C. (1987) 205-214 Mirimanoff, D. Les antinomies de Russell t de Burali-Forti et le probleme fondamental de la théorie des ensembles, L'Bnscignement Mathématique 19 (1917) 37-52 Reynolds, J.C.: Polymorphism is not Set-Theoretic, in: Semantics of Data Types, G. Kahn et al, (editors), Lecture Notes in Computer Science 178, Springer-Verlag, Berlin Heidelberg (1984) 145-156 Russell, B.: The Principles of Mathematics, Cambridge University Press, Cambridge, GB. (1903)