You are on page 1of 76

320

V. Boundecl Arithmetic

Fnnctions 1. Witnessing
(a) Xntrocluction
Suppose a sou.nd bheory 7 proves a senLence (Vr)(ly)p(*,y). If ? is weak and tp is simple, then we have some ailditional information on the functions which u'itness the existential quanbifier in Lhis sentence, i.e. functions / for r,vhichwe have (Yx)tp(r,/(")). A classical result says that if ? is ID1 ancL p is botrnded, then there is such an / which is primitiae recursiue) see T'heoremIV.3.6. Parikh fParikh 71] has notecl that / is linear time computable if ? is /Xs and gais bounded. Buss [Buss 86, Bounded Ariih.] has definecl a theory Si2 (a conservative extension of IDo + J?1), fragments S| ancl subsets D! ot bounded formulae in the extended language. His theorems characterize the classesof functions in the polynomial hierarchy using the above phenomenon. In parbicular, if p is a Xf formLrla ancl f is S|, then / is polynomial time computable. The converse of this special case is also true: if / is polynomialtime computable, then there exists El fbrmula which clefines / and for " which it is provable in 5] that / is total. We start this secbion with the defi.nitions of these classesof forrnulae ancl fragrnents of bounded arithmetic. The next step will be positive results about definability of functions of certain complexity in these tragments. It turns out that this is qr-riteimportant inforrnation about the fragments. For instance, knowing ftar T) defines functions of trf*r, for i > 1, it is easy to prove Buss' theorem for Tj. The main part is sr-rbsection (d) where witnessing theorems of Buss and of Krajfcek and Taheuti are proved. In subsection (e) we shall show a reduction of the problem of finite axiomatizability of bounded aribhmetic bo the problem of collapsing the polynomial time hierarchy. Since the fragments of bounded aribhmeLichave been exbensively studied in [Buss 86, Bounded Arith.], we concentrate on results and proofs not includeci in that booh. In particular, as Buss uses proof theory, we shall give proofs of the witnessing theorems based on model theory. We shali refer to Buss's book for basic facts provable in these fragmentsl holvever the reaclers with some training in bounded arithmetic, for example those who have worlced through Sect. 3, shoulcl be able to mal<etheir own proofs r,vithout consulting ihat boolc.

(h) Xfuagrnents of Bo'unclecl Arithmetic


We shall e:ctendthe usual languageof arithmetic .Ls by adcling two unary frrnctionsymbolslrl and trf 2t, anclone binary operation r lf y. The intendecl inter:nrebation of lr.I is

4. lVitnessing Functions

321

r1og2(r+ 1.)t: the length of the binary expansion of ltl r if.n ) 0, other:wise 0, i . e . t h e s a me a s i n S e ct.3 . l l e re r yr denotesthe least integerz ) ,y; u,yL denotes the integer parb of y, bhusthe inberprel,ation c>f. urfL-t is clear. The inbenclecl inberprebation of the smashfunc-tionis x #y: 2ltl*lvl : the po'werof 2 whosebinary expansion has

lengthl"l * lyl -r t .
Observe bhat the seLof the lengths of nurnJrersin a sound arithmeticzr,ltheory conbaining the smash function is closecluncler rnr-Llbiplicabion, hence uncler any polynomial increase (for standard polynomials). This is important, since it enables us to formaiize many standard constructions, in particular polynomial time computations. We shall denote this extension of Lo by L2. In th.is section we shal.l work in the language L2 and in an extension L2 of .L2. Now we recali the basic system of open axioms for the extendecl language. It is callecl simply BAS.IC and it plays a similar role Lo Q in bhe usual language. 4.1"Definition. BASIC is the following theory: (i) v1n--+ylSq

(2) r I Sr; ( 3 )0 ( " ;

( 4 )n l a k x / y : S r S u ; ( 5 ) - 1 0- , 2 r l 6 i (6)" ylrVrly;
(7) rlUky<a--+r-Ui (B)rlUky<z---+x1z;

( e ) l 0 l: 0 ; ( 1 0r) * 6 - l 2 r l: , 9 ( l " l ) & 1 . 9 ( 2:r )5 l '(lrl); : (11) r; lTl

(12r ) 1a -, l"lS lvl; ( t 3 )l r # y l : 5 ' ( l r l * l y l ) ; ( 1 40 ) #y :T; :2(r # : t(1 # (15)r l6 -, t lf (2rc) k't ll (s(2n)) "); ") (16)rlfa:y#r; (17)lrl : lyl -, x:# z : u # z, (18) l r l - l r l * l u l - +r = t ' f y - ( u " = f f a ) , r ( u f t ' y ) ;
( 1 9 ) z< r * y i

(20)r < y kx * a -- S'(2r)<2y k S(2x)#2y; (21)r-la:y-lr; (22,)r-F0-z-; (23)c -FS'y- 5(r -f u);

322

V. Bounded Arithmetic

(r+ v) + z : n 1-(y -r z); (2,1) (25)rl-y<r+z=y1z; (26)x*0-0;


(2,7) r r, (5y) - :ty -l n; (28) ay - y'"c; (29) rc(y+ z) : ry -l-:cz; (30)r ) 1 --r(rA < xz : y S z);

(3r.) r #6 -, l"l : ,9(1,-nl2Ll);

(32)a : ex12' = (2a : r V S(2y): r),

where T and 2 arc numerals as clefinedin Sect. 3. This system is probably not cluite optimal. One can play with it in order: l,o find a shorber or nicer system. For instance, if we redefine

. 1 if r > o, ltl :'rog2(r + 1)-0 otherwise,


then r # y : 2l*l*lul becomes associaLiveancl distributive with respect to multiplication, hence one would obtain a more homogeneous axiomatic system for such functions. We shall not do it here, as we want to concentrate on more interesting questions. We shall use the bounded quantifiers (Vr ( r) and (1, < z); note that r is a term in the language L2, nol just in Lg (not containing r, of course)' lVloreoverwe also need the so-called sharply bounded guantifiers. T'hey have the form

( V r < l r l ) a n d( l r < l r l ) ,
term is the length function. Their function in the bou.nding i.e. the ourtermost relabionto ordinary bounded qttantifiersin bounded arithmetic is similar to to unbounded cluantifiersin Pzl. The the relation of bounded quantifi.ers classesE! and,.A! of bouncledformulae are defined as follows. 4.2 Deffnition. (1) n3: (2) Xl,, ttL only; Xf, consistof formulaewith sharply bounded quantifi.ers are the leasi seis satisfying ancl[I!-., [-rL

g rI!4; zut,n! ancl (") z!,n! e D!+t,

(b) a e E!-4 + (-u 1r)a e E!+r,(Vz( l"l)' e E!+r, a e II!*1=+(Vr 1r)a e n!+r, (lr S l"l)" e II!-4; ( . ) a , g E l n r 4 a k g , a v g e E ! - r - t , a,0 n!.-, + ak g, aV B e n!*1; g e n!*r4-g, g -, * x,l-rr, (.t) ae Dln1, a e II!*r, g e D!-rr=* -,9,9 -r t Qn!-rr.

4. Witnessing F\rncLions

323

( 3 ) cvis in zli with respect to a l,heoryT,i:t cvis eqr-rivalenL Lo a D! formula ( 4 \ D|@i)


h., ,'l

ancl a l/rb formula in ?. is defined as the closutre of E! uncler conn.ecbives aircl sharply bouncledquantification; nobeth.at this classis som.ebimes denobedalso
rtb r-t n'b "i-F.\ttr'ri_l_1.

L,ater on lve shall see. that, for d ) 1, D! (resp. I/ob)formulae clefineiust Llresets in E! (resp. n).'th.e most important formalizations of incLuction for bouncled formulae are the usual schema of inclLrction ancl the foll.owine ones. 4.3 Definition. (1) PIND(a(x)) is the formula

ct(0) & (Vz)(a('-rlz-') -, *(")). -' (Vr)a(z) ; (2) LIND(*(z)) is bheformr-rla

a(0)& (Vy< l"lX*(y) -, a(,S(y))). * *(l"l) ;


(3) LMIN (a(r)) is the formula

(lr)a(r) -* a(0) v (tr)(a(r)sz(Vy < nl2)(-a(z))); (4) PIND? (LIND|, LMINT, respectively) is the schema (or the set) PIND(*(r)) (LIND(a(x)), LMIN(*(z)), respectively) for a(x) e-t'. Here we assumethat a(r) may contain other free variables, which are called paramebers. We have written LIND in the form lvhich contairrsonly one additional quantifier and this cluanbifier is sharply bounded. Thus if a is A!-, th.n LIND(*(r)) is also ab;; it a(z) is in n! u i!, t]nent,IND(a(z)) is

El,Q!)

iVow we are ready to define the hierarchy of subtheories of bouncled arithmetic. 4.4 Definition.

(1) For z > 0, St is BASIC+ PINDE?. (2) Fori > 0,r; is BASIC+ I D!, (3) 52 is l; Sig; Tz is l);T$.
The most importanL fragmenL of .92 is .9;1.IL serves as a basic theory rnoclr-Llo which lesults on tlagrnents of 52 are provecl. Ibs role is similar to

32,1

V. BouncledAribhmetic

Ti is noL quite l<nown. lVe sh.allnot repeat tlr.e fbrm.alization of elementary r.rr.a,themai;ical coi.rceptsin 5'l which has been clone in fBr-rss86, Bounclecl r\ritli.]. One can use some technicluesfrom SecL.3, mainly those r,vhichwele mentioned in connection wibh .tEo + J71. (Those which recluced the size of. are nselessfor .921, since we the cluantified nu.mbersby using more qr-ranLifi.ers

the role of.t E1 in .PA. In parLicularwe shall show in the next subsecLion thtr,t all polynomial time compu.table ftincbionsare clefinableand provab.ly tota,l in 5}. As fbr weakerfragments:,9! is boo weak (it does nob prove the existenceof the predecessor, seefTaheuti,Sharp]y Bouncled]); the strength of - n .

neeclXf definiLions.) Here we briefly rnention a definition of codingin S). Pow(r) (* is a power of 2) by an open formula in L2: tr'irst we L:e-clefine Pow(r):2r:r#L
we shall show in .BASIC lhal Pou satisfies In orcler to see Lhat this woL:ks, condition: the foLlovzingrecr-rrsive

'

P o w ( x ) - - ( r : T v ( 1 a < r ) ( r : 2 y k P o u ( y ) ) ). One can easily show -Potr(0). Sttpposec andy > 0.By the defi.nition B A S I C , * - 2 y o r f f _ S ( 2 y ) , f o ry : t r l 2 : h e n c e( b v ( 1 ) , ( 6 ) , ( 7 ) a n d of. P o w a n d ( 1 5 ) ,w e h a v e 2 r - * # T : 2 , ( v # T ) , (30)) we have r: U /lT. SinceU ) 0, we get from (15) that r is even.Hence (B.ASIC provesthat .9(2tt)l2u). we must have the first possibility*:2y rn r.,,-.-," , 4 4 ,l t'T ,i.e. pow( y) . bit(x,i) (the i-th digit,of r) by a Xl formula: No'w we can re-clefine
I IIL!J

< r ) ( 1 u 1< y ) ( P o w ( y k )lal: I bit(r,i): t = z 1I k(1y,u,2 k x : 2 u z A - l z t J+ u 1 ) , i f i > 0 , : 0 otherwise. Since the decompositionaboveis unique (seeTheorem I.1.30), we have aLso
'n'lt

a .L.l; clen.nl[10n:

bit(r,'i):,

= ( V t o< l ) ( V y ,u 2 < r ) ( V t r 1 <v)(Po'w(v)klal:t k x : 2 t t 2 y + u a f u 1 - - +t t s: z ) , i f i > 0 ,

: 0 otherwise. 'f'hus l,vehave Alr defi.nitiono,f bit in S). Similarly we can define lunction " If we rvant to code (see 3.18). The f'Lrncbion bil codes 0-1 secluences. sgn'r,, will be a triple we can proceed as follows. A secluence arbiLrary secluences, ( ( u,u) ,' r l) , Lhusever y num ber is a sec l l enc e) , p : (u ,u ,ru ) (mo re pleciseJ.y

4. lVitnessing Functions

J2b

where tr is the code of the secluence, u is its murimal element and ro is the of lengtlr, th,esequence,t.e. ma.r(p)- u,

lh(p) - u) . The fun.ction the i-tlr, element, of p is definecl by -t-j)), if i <Lu 1 ( p ) ; - n : r / > u k ( V r S l u l ) ( b i t @j , ) -bi,t(tr,,iu ( P) ; : 0 o th e rw i se . Flence all these functions are A!. Finally we observe'Lhat we can bound u by

l z l< l u l . t r .
Thus for some term bound we have (a.r)

p 1 b o u n d ( m a x (, p2 )t h ( n ) 5 .

4.5 Proposition.

( 1 ) F o r i> r , S i = S l +z I N D r r ! = s l + L T N D = ES ! l+trNDn!. (2) Fori > 7,, T4= S+ + rn!.


Proof. We shall show only Si = SB l- LINDE!; the prool's of the remaining equivalences are the same as for fragments of.PA, seeTheorem.I.2.z.L. First we shall derive LINDE,b, in Si.fet rp(z) be X!.fet 1.,(r)be,p(lrl). Tlren by (t0) and (32) of BASIC, the inductive clauseof LIND(,p("))

p(x) - p(5'('))
implies the incluctiveclauseof PIND(rlr@))

-, lhfu). tl.,(_y l2_:)


Thus LINDX| is reduceclto PIND E!. Now we shall derive PIND D! in .g| + LIJVDD!. Ler p(n) be a ):! fbrmtrLa for which we want to prove PIND. Thi.rs we assurne !a(0) ancl (Vx)(pQnl2L) --+ tp(r)), Let msp(r,y) 6e the nr-rmberwirosebinary representationconsis.ts of the first y digits of u (and, say, it is r if a > lrl). T'h.is function has a A! definition in 5'21 (use the function sgm). The icleaof the proof is to show, using LIND, that all end segmentsof r satis,fyv,, hence also z itself sabisfies g. l-ormally, we apply L.IND to gc(ntsp(",y)) wibh y as the inducbiouvariableand c as a parameLer. Thus PIIYD(go(r))is recluLcecl to

326

V. Bounded Alithrnetic

L I N D ( t p ( r r s p ( t j y ) ) ) s i n c e r n s p ( r) 0 ) : 0 , t m s p ( r , S ' ( y ) )l 2 ' - r n s p ( r , y ) a n d

m s 2 t ( r , l c l :) z .

4.6 Theorern.For i ) I, Sif tAbo. ProoJ'. Recaj.lbhemeaningof .tA!:'f.or every a E! and, B e n! (Vr)(a(r) = 0@)) -' /(a(r)) . The icleaof ihe proof is the following.The schemaof induction telis us that we 0 and is closeduncler cannot definea proper subsetof numberswhich conLains The schemaPIND implies that there is no subset of numbers the successor. and the fnnction 2u. We already the successor containing 0 ancl closedr-Lnder we can definea subcut closedunder know (Theorem II.3.5) thai for every cut under 2r.We shall use this construction to reduce addition, hencealso closed induction to PIND. There is a small technical complication,sincewe have to uses unbounclecl use only bounded formulae,while the original consbruction quanLifiers. Let o and B be given. Put ,,h(a,o) : af (Y* < a)(a(r) + B(min(n -l-Utc,))). Thus 1.,is II!. Working in ,9$,assume*(*) : 0@),,a(0) and (Vr)(c(z) -r that a(a) fails for some a. We haveT/(0,a) a(S'(r))). Furthermoreassume tri.rially and r!(y,a) * ,h6@),4), since (Vz)(c(r) --+ a(,9(r))). We shall and a( z) for som ex l a. By r h( a,o) p ro ve,h @,n )-r rh ( 2y,cl) Assumeth( . A,a) a( x * a,a) . w e h a ve a (rn i n (n +y,a) : ) . Thus r *U ( a, since- c( a) . Flence get a(min(n*2y,a)). Applying ,h@,") onceagain(with rf y insteadof :v ) r,ve ---> Hencewe have ,b(A,a) tlt(2y,4).Since :c: 2tr 12: V r : S( Zr xl2L) ,
the formula

n) , h @ @ ) , a 8)c { ( 2 y , a ) ) (Vy)(dfu,o
implies

y f 2-', a) -> t[(y,a)). (Va)?h(

usingPIND(/) *. getlt(a,a). Since*(0), this impliesa(n). We have Hence henceit mtLstbe true. Thus we haveshown deriveda(a) from its negation, n inductionfor a.
Note that, for i 2 2, a weaker version of this theorem can be derived ou"t T'i of S'L1-1 also from Corollary 4.28 belowusing the If*, conservabiviby (Corollary 4.34).

4. lVitnessing F\rnctions

2r'7

4.7 Theorem. c,i+1. f , \ t /) F b r i > 0 . , 9 l c T l C'J2


(2) Sz = T2 and they are conservative e;rtensions of IEo * At.
.ProoJ. (1) Tb prove th.e first inclusion, observe th.at T; trivialllz proves 'Iheorem LIND E! and apply 4,.5(1). The seconcl inclusion is a corollary of Theorem 4.6. '(2) I'he equivalence oL theories,S'2andl'2 follows:[rom (i) of the preceding proposition. In order to show that /Xg * fit is contained tn ,52 we have only to prove axiom A1 in,52, which .followseasily from the boLrnd u2(r) < (* =|t' da . Let us prove the conser:vabivityof.,92over -ftr'g*At. We already know that ;nf2L and lrl are definable in l'Eo+ J71. As we have a X9 definition of the exponenbiabion relation, we car] define the graph of the function x {f y. This funcbion is provably total in /Xs iQt, becauseof the bound r # y < ut2@*U). The axioms of BAS.IC are easy too. To prove indtrctior-r.for bounded formulae in language I'2in IEo l-.f21, recall that by Proposition 1.3 we can eliminate terms as bounds in the formulae used in incluction. Then we use the definiLions of the new operations to translate the fbrrrruJae into the language .tg and

apply /X6. 4.8 Theorem. (a) For i ) 1, and a(u,r) e E|pb,

tr

(1ru)(V" S:$F< lyl)("(",n) : bit(w, r + T) - T) .

(b)For i ) L,sir uwD(r8(rl)).


Note that (a) is a version of the :;$(l.f) schema in Si "otnptehension where $/e use bit(tu,r) as a convenient coding of 0-1 sequences. ProoJ.(a) Consicler a more general schema for a D\@b:

( b . 1 ) 5i F ( l t u ) ( V r t< l y t l ) . . . ( V " , < l y " l ) ( * ( " , x r t . . . , r n ) : bit(u,,T : + ("r ,(*2,.. . , ",r)))


'We

T)

should shour that (b.1) holcls for a e D! ancl thal, tlle set of formulae a satisfying (b.1) is closed under boolean operations and sharply boundecl quanl,ificabions. Suppose a Eo;. Take the formirla B(z,z) defined by

:u t ( l t u ) ( l . r5 l l v r l* l y z l , . ) , . . , r l.y , l k n u , o n ( z ru & ( v r 1 < l v t l ) .. . ( v ' " i l ' v , , 1 ) ( b i t (-z (c r , ( * 2 ,. . , r " ) ) ) ,T


- , a ( u , , x r t ..

328

V. Boundecl Aritlimetic

J,etns lvork in 5i. T'hisformula is true for z :0 and fails lbr z : lyrl'r lyzl r. .'.',, l u n l + 1 , si n cemton( r o)( lr .r .'Since a is E!, nlr on is,d! an c l w e c an 1. lrotrndu, say by yt#yz=t'|=. .#yn, Ihe formr-rla B is also X;o.Hence we can z for rvhich B(z,u) fails. Then take r.usr-r.ch we LINDII! to fincl the smallesb t,h a t, B (z 1 ,tu ).If

z - t : l a t l* l y z l* . . . , r l y n l ,

Lhen all biLs of ?, are 1, hence we h.a.re(b.1) lbr this tr.'.Now suppose that

* "',r lynl. z - I < lytl* lazl


If (b.1) were not true for this tu, then.we could add another 1 in the binary r e p r e s e n t a t i o no f u ; w h i c h c o d e sa n i n s t a n c e o f a ( u , r 1 t 1 2 , . . . , f i n ) . B u t t h i s vrould contradict the condition that z - 1 is the last parameter for which. is true. Flenceto codes exactly the set determined by a. 0Q - 1,T.o) The case of propositional connectives and sharply bor-rndedcluantifiers is easy. n (b) follows immediately from (a). In Botrndect Arithmetic there is a counterpart of the collection ariom for Peano Arithmetic, see L2.I. We shall call it the bounded collection a.riom. Sometimes it is also called replacement ariom. 4.9 Definition. (i) BB(a(r,A)) is the following formula boundedcoLlection

z ltl)((.)* ( ska(n, (.),)). ( V " S l t l X l y ( s ) a ( ca , ) . t ( l u . ' ) ( V< BB(a) for a e T. Gl BB f is the schema
(a.1) above, we can bound the variable tu in the bounded collection axiom; thus it has the following form (ecluivalent in 5]): 'Using

y).* (Vr < lrlxfy ( s)a(r,

- ' ( 1 , ( r ( s , t ) ) ( V r< l t l ) ( ( r ) , ( s k a ( x , ( - ) . , ) ) ,
bound'(s,Zltl+t;;.

w h e r e r i s a s u i t a b l et e r m ( r ( s , t ) :

AAfi. For i> I, S'$V 4.10Tireorern. Assume Let a(r,y) be a trf forrnula. Proof,

< s)a(r,y). (V,< ltlX=y


Then we obLain
/a | -t4Il

( r ( s , t ) ) ( V< r l i l X ( . ) , ( sk a ( r , ( - ) , ) )

4. Wibnessing Functions

329

by L.fND((1z( u r ( s , t ) ) ( V r < l z l ) ( ( t u ) .S , " k a(r,(.r),))), where the induction variable is z. The forrnula is Xf , sincecoclingis Al. We
r-rse the elementary facts that the empty seqllence has a code, each.secltrence ca,n be ex1;enclecl bv a,n a,rbiLrarv elem.ent and the estimaLe of the size of:r sequence. LJ Note thab the implication converse to that above is already provable in 5l (using only the fact that (u), is a defintr,blefunction in 5l). T'he bor:rndell forrn of the bouncled colleciion axiom enables u.s to intercb.ange sh"arply bouncled quantifiers with bounded cluantifiers. However we cannot push insicle all sharply bouncl.ecl cluantifiers withor-r.t exbencling the langr-r.age, since the coding function is not clefineclby a term in .L2. Therefore we need to extend the language. 4.1-t Definition. (i) The language -t2 augmented with the.binary operation (r)r, will be clenoteclby L2; the corresponding classesX!,n! eic. will be denoteclbv E!,II!, etc. ii it is of the fbrm (ii) We say that a formula is strict\f (lri

< t 1 ) ( V a 2< t z ) . . . ( Q r ; 1 . t ; ) e ,

where Q is 3 if i is odd and V if i is even, and rp is s]rzr,rplybouncled. We shall not introcluce new notation for theories ob'bained by extencling the language and aclding the definition of the coding relaiion. It will always be clear from the context which theory we mean. 4.12 Corollary. For i formula, provably in S$ (augmentecl with the clefinition of (r)"). Proof. We can shi.ft all sharply boundecl qua,ntifiers beyond bounded (br-tt nob sharply) quantifiels using the bounded collection schema. Successiye cy-tanLifiers of the same lcind can be merged ttsing the coding function. tl 4.13 Corollary. For i ) 1 ancl *(r,y) in E!,

S i F ( 3 t u ) ( V "< l t l ) t ( : y ( s ) a ( r , y ) : ( ( t r ) . , ( s k a ( r , ( . r ) . , ) ) l . (Tlris is sometimes calleds'trong replacement.) Proo!. Let a(r,y) it f! be given.By T'heorem 4.8, we can cocle the sebof r's s u c ht h a b ( ly S s)a (r,y) u si n gsom e,uand fbr m ula bit( u,o) : 1. Now take *'(*,3/) to be t(r,y)V bit(u,n):0.

330

Alithmetic V. Bouncled

Then we have (V" S ltl)(:y ( s)a/(r, y), hencewe can apply bounded coln given by iL sabisfies the clauseof the theorem. r{., lection; the sequence

Cornputations (c) Definability of T'tuingMlachine Arithmetic of Bounalecl i.n .[-'ragrnentr;s


clenol,es bhe set of functions compuLable in polynoRecall ihat for i) I,lf in palticular, trf are just the polynomizrl m.ial time using oracles ior EeO-.rsebs; time computable funcLions, since Xfi oracles are sLrperfluous.In this sr-tbsecthat functions from n{ have suitable clefinibionsin ,9}, an.i[ tion we shall shor,v i > t. ffris means that for functions from Tj,-I ,,i:f. the same holcls for nf a,nd. J! we haxe A!*., definitions in the corresponding theories and that basic properties of the functions are provable in these th.eories.lVe shall only show that the defining formulae have appropriate complexity and that they define total ftinctions. Other properLies follow from the construction and their proofs are omitted. We shall consicler delerminisbic Turing machines, possibly wiih oracles. Let an oracle Turing machine fuI be given, lei e be the code of it, let A be an oracle. We formalize the computabion of NI on input a as two secluences u and g, where ('u); is the instantaneous description of the state of. IVI, the position of the heads and the content of the tapes, and (q), is the query asked in step i.1 The computation is determined by a constant f,rstq (let us set it eqnal to 0) and five functions fi,rstw, nertwg, nentw1,,nertqg, nettql as fbllows: (tu)s : f,rstu(ct), (q)o : f,rstq; (q)l+r - nentql(@);,(q);), if (q); A.

- n e r t q s ( ( . ) , ; , ( q ) ; ) i,f ( q ) ; A ; ( r ) ; - r r - n e u t w s ( ( - ) t , , ( q ) ; )(,q ) ; - r r 4
( . r ) ; + r - n e i l t w r( ( . ) 1 , ( q ) ; ) ,

We can think of (r.o);as a sequencewhich encodesthe current situation on the tapes and encodes the state of IVI. Then the functions above are only local transformations of these sequenceslhence iL is not diffi.cult to write down a! aennitions for them using o.rr Al clefi.nitionof coding. If the oracle A is clefinedby a formula g(r), we thus get a formula C ornPIuI,,p61(ru , Q, cr') explessing Lhat (tu, q) is a computa'tion of IuI on inpttt ct,.'It IUI is a machine withoLrt an oracle, .weshall wriie simply Compp1(ro,c). We shall assume that
1 !V.l.o.g. lve can assulrrethat M asks a query in each compuLation sbep; if it does not need information it asks some defaulb lixed qr-rery.

4. lVitnessing Irunctions

331

(ru); : (.r);-rr i:tr AI stops in the i-th sLep.Thus computaLionso.F lengLh n exist for everlyn, anclwe have a simple conclitionfor testing the termination. FnrLhermorelve shall assumethat th"is(u); 'is tlte ou,tput of I\tI. We shall identify oracles with bheir defining formulael thus we shzr.il. terlk abouL Xf oracLes insteacl o.tEl oracles 4.14 Lemma. (i) Comp ct)is At', in S). 1r7(-u, (ii) If 9@) is D! , i )- I, then Com? q, a) is ,8@!). M,e@)(ru, (iii) There e:<ists a bermcr such that S) | Comp . , q.,a)-* w, Q 1 o(a, Lh(ro)) rur ,,p(x)(* Proof. (i) and (ii). We have (t, a) = (r.r.') Comp 1y1 s : firrtw (a) & (Vi < lh(u) - t X(.)t+ r : nertw( (.r); )), and q,,a ) = (-)o : f,r stw( a) &( Vi < lhQu)- ' I) C o *p M , ,o (c;(tu ) 1 : n e n t'w tqs( ( "u) 0 (( ., ) ;, Q) t) 8, ( q) ;+r - n e$ t, ( q) ;) ) l ( - , p j ) r k (zo ;a t (-);-rr - n e rtw l (( .r ) t, ( q) ;)& ( q) l+r - ner tql ( ( .r ) ;, ( q) ;) ) l . v ( , p (rD& used in these formr,rlaeare A! clefinWe already knor,vthab all Lhefr-rnctions able. (iii) The length of tu and g is polynomial in the number of steps of the n computation and the length of input, .whence we get the bound. Now we prove the relation between the classes of formulae E! and the classes of sebsff which we have promisecl. Note, however,that the relation is not cluite direct tbr z 0.
4.15 Theolem. (i) (ii) Every set definable by a X$ formula is in LogSpace. Every set in P is clefinable by a Al formula. (iii) For i ) I, the sets in E! (resp. III) are just the sets clefinable by Xf (resp. //;b ) forrnulae. ("Definable" means here "deflnable in the standard model".) Proof . (i) A11the fr-rnctionsand predicates of the language of 32 a,re LogSpace preserves LogSpace corrlpr-lcomputable and sharply bouncled cluantifi.caLion tabions.

332

V. Boundecl Arithmetic

(ii) Let ,K be in P. L'et fuI be a Turing machine which always sbops after 0 iff a belongsbo X. T'l:en we can define p(lol) stepson inptLt a anclor:bpubs

xbv
or br

ct) (f tu )(/h (to)S ' p( lol) & Comyt 8z( .r ) plnl) : 0) , x4( r u, (V tu )(/h (tr )5 p( lol)& Com p y( u,,ct) - n ( r ) p1"1) - 0) .

By l,emm.a 4.7rL, the inner parts of the formulaeare A!; by the sameletnma bhe bound lh(u) S p(lol) implies bhat tu can be bouncledby a term in ct, to El ancl //f formulae respectively,thus hence the formtLlaeare ecluivalent they are C!. to a strict Xf formula (iii) By Corollary 4.72 everyX,f furmula is equivalenb in bhelanguageextendedby the codingrelaLion;thus it can be written in the fbrm ( l y r < t t ) ( V a z < t z ) . . . ( Q u ; I t ; ) c I, where @ is sharply bounded in the extendedlanguageand where Q is : if i is odd and V if f is even. Sincethe coding relation also belongsto P, formula O definesa predicate in P. The bounds are just of the order neededfor the definition o,tD! and the relationsaj < t j are in P as well. Thus, by T'heorem by a Xf formtrla is in El. 2.!!, eachset clefinable Thus X is definedby To prove the converse let X be in Epo.

<; p . i ), ( 1 v t , l s l< l p1(lrl)) l ; ( l " l ) ) p ( t , u r , .,.u . .( Q v t . , l v


P by a 4l formula. Similarly to above wlrere P is in P. By (ii) we can clefi.ne we can replace the bouncls lyil Spi(l"l) bV bouncls of the forrn lli ( 'ti tor tr stiiLableterrns tr,,...,f;. t'hus we obtain a Xf formula defining -,(. 4.L6 Theorern. (1) Let M be a Turing machine. Then 5l prov"s that for every input a and every b, there exists a unique computaLion of .&/ which has lbl steps. Formally:

ri)
(i i )

a) k lh( zu): l6l); S j F (l tu ) (Com p x,1( u, S) f Comptt(w, a,)k Comp , cr.) 1r1(.ut :lr' k l h ( r u ) , l h ( w t ) : l b l - - +, J )

(2) Let i > I,Iet fuI be an oracle'furing machineanclleb tp(u) be in Eb;. T'lren Tj provesthat fbr every input a and every b, there exisbsa unique computaLionof IrtI with oracle p(r) which has lbl steps.Formally: (i) f)f l, k) lh(-u): 1 n 1 , . p ( r ;c (t uc, L Qw,q)(Comp l b l );

qt c1, a) k Comp (i i ) f) | C o mp ,r , a) tur ,,p( t:) ( ot tu ,,p( r u) ( zu, ' cq : k l h ( w ) , l h ( z u t ) : l b l + , , : , . u8

4. WiLnessing f,\rnct,ions

JJJ

Proof. LetAI anclp be given,,p in D!. L,et1,,be clefinecl by l h ( . , Q , a , b ) : L h ( r u-) l b l- F t & ( t u ) 6- . f i " r s t u ( o , ) & ( q) of , r s t q . 3 z ( V <7l b l ) : n e x t q o ( (jr,), ( q ) i ) ) i, q ) i )& ( q ) r - r r [ ( ( , r ) ir r : n e r t u ) 0 ( ( r ) ( rt (,p(kt)i)& (tu)i11,- nexL,ur((.r)i, (q)i) k ( q ) i - r t * n e x t q(l( , r ) i ,( q ) i ) ) l.
The meaning of bhe formula is that u, q is like a computahionin which we fbllow th.e aclviceof bhe oracle if its answer is negaLive, but not always do we follow it if its answer is positi've.This forrnula is A! it i - 0, and z! it i > 0. llence we can prove, using respectively LINDE9, or LINDE! on b, that for.givena,,bthere exists at leastone pair u,g such that r!(n,Q,a,b). Using IE! we now minimize Lhenumber of sLeps *h"r" tu ancl q.lo t;t foilow the positive advice. Let V be the E! formuLadefineclby i [ t ( t t , a , b-)- ( h r , q a o ( a , l b l ) ) ( / ( t oQ , , d , b ) U ( V j< l b l ) j,G) ) [(.r)i -rr : n e rtw y((*) i,kt) i) k( q) i+t : ner tqo( lu) -> bit(tr',l6 -l i ) : 1 ) l,

-,

where a is the term frorn Lemrna 4.14(iii). For a and b there exists at ieast one u satisfyingtIt(tr",a., {), sio.e for tu,r7satisfying{,Qo,Q,a,b)the existence of u follows by LINDAl. By IE! we have the least u for which V(u,,a,b) holds (this is the place where we usethe full strength of.T).Let u,g be the wibnesses of.V(zt,a, 6) for this u. We shall shor,v that for this r-u, g also the implication complementaryto the one in t/ holds for all f < lbl, 7 r , ( q) i) & ( q) r _r r- nextql ( ( .r ) i ,Q) ) . w ( Q ) i ) -, (r)j -rr : n e n 'tw((r) Suppose not, then we have p((q)J.)and ( . ) i - r r : n e r t w y ( ( r u ) iQ , )i ) & ( q ) i + r : nentcto(@)i,(q)i),

ibr some j < l b l . H e n c eb i t ( u , , l b l -j ) : 1 . T a k e u r l a n d q / o f l e n g t h l b l s r L c h that ( r r ' ) n : ( . r ) nk ( q ' ) n : ( g ) r , fbr /c : 0,


t J t

( - ' ) j + t - - n e n t u l ( ( . r ) i , ( q ) i ) k ( q ' ) i t - t - n e x t , q(l ( - ) i , ( i l ) and

(,r')t+r : nex.t0 ' a( ( . r ' ) 1 , ( q ' h ) & ( q ' ) f t + r- n , e x t q s ( ( r ' ) t(,g ' ) r ) ,

qq,l

V. Bounded Arithmetic

fo r /c : j -l 1 ,. ..,l ql - 2. T' he exisLence of such tol ancl g/ is ag ai n c as i l y provable by L.INDEl. Let d be cletermined by

- -j - ' I , b i t ' ( i l , k ) : t , f o rk : ! , . . . , 1 6 1 b i t ( u.t, l b-l j ) : o , l c : l [ ,b i t ( i l , k ): b i t ( u , , h ) ,f . o r ] j - l \ , . . . , | b |.


., tb t r l r r t < u , w h i c h i s Tlren ''are h a v e V ( u ' , o , b ) , s i n c e u t c o n e s p o n d sL o t o t , c 1 a cont.radiction. Thus we have shown

c, i , aA ) l h Q u ) : l b l. Compr,v(,;(u apply LINDA! ro To proveuniqueness'ure

( V i < l a l ) ( ( r ) i: ( t u t ) i8 r ( i l i : k t ' ) ) .

4.1? Theorem" tr{ functions are definableby formulae which arc Ab, in,9}; for i > 0, trf+, functions are definable by formulae which are A!*, in T$. F\rrthermore, for each such formula the correspondingtheory proves that the formula defines a total function. Proof.Let, i ) 1. Let f he a nf*r function defined by an oracle machine -42l 1>y and a El oracleA. By Theorem4.I5, A can be clefined a E! formula ,p(r). Then (sirnilarly as in Theol'em Let p(r) be a polynomial time bound for /rz.I. by 4.15) f (n) : b can be definecl q,, , c t ) k (.r)e11,: ;; b), ( 1 u , q 1 o ( c t , , p ( l r l ) )C Xo m p p l , * ( r ; ( t u or by ) ,- , ( ' u ) p 1 . 1:) b ) , ( Y r u , c1 1 o ( c t , p ( l " l ) ) ) ( C o m p p 1 , , p ( rq ), (a tu
lvhere o is the term from Lemma 2I.14.Using ihe same lemma to estimate Lhe complexity of. Comp,y1,,p(r) o.d the existence and unicluenessof the computation (Theorem 4.16) we see that it is a Ab;*l definition inT{,. For i:0 the proof is similar. The fact ihat the formulae define total functions in the f corresponcling tb.eoriesfollows {rom Theorem 4.16. In order to be able to use the clefinitions of functions for other consbrucLions in'I, we neecl more propelties of forrnulae defining i;he functions. In particular, we shail use bhe fbllowing properties:

4. Witnessing Fr-rnctions

335

arrd* is the clefi.nition (1) if l(x) is a beL:rn of the corresponcling n{ frLnction, then l(x) : g is ectuivalent to r/(x, y) in S:); (2) nf functions are provably closedr-rnder composibion, i.e. iI tp(x,.g) c\efines ./(x) and t!;(x,y) defineg;(x), then for some defrnitionX(x, y) of / ( s / r ( ") ,. . . , e n (x)) T ; (re sp .Sl, lf i : 1) pr oves , t ) & . . . k r h n ( * ,r n ) k g ( 2 t , . . . t z n , y ) ) ; X G , y ) : ( 1 r t , . . . , r r ) ( r l t t ( xz
e Lr; runccions are closeduncler definiLions by casesdei;erminedby E!-, formulae (it is clear what is the corresponding formula); (4) nf functions are closedunder Jrouncled recursion.

(r/

/n\

t-D

The last conclition can be formalizedasfollows.Let /(x) ,9(z,x), ['(y, x) be functions in nf ; then the functionh(y,x) definedby the followingrecursion must also belong to nf : h ( 0 , x )- / ( x ) , h ( y , * ) - m i n ( e ( h ( ' y f2 ' , x ) , x ) , k ( y , * ) ) . We recltrirethat this schemabe provable in Tj (resp. S$, it. i : 1) for the clefining formulae of f ,g and h. To prove these conditions one has only to checlcthat the proofs in the standard model can be carried out in the fragments of Bounded Arithmetic, We omit the proofs since they are not difiEcult and contain no essentiallynew ideas. So fzr,rwe are able to tallc about a single function form some class nf . Later we shall need formalization of Turing machine computations such that we can talh about Turing machinesand oraclesin the theory, t.e.we want a formula defining the computation which has parameters also for Turing machines and oracles. In the model theoretical language it means that we ',vant to consider also nonstandard Turing machines and oracles. A simple solution is to take a universalTuring machine,which is explicibly defined and thus has a formalization in Srt by Theorem 4.77. Then the code of an arbitrary Turing machine will be the word which mr-rst be written on tlr.e input of the universal one in order to simulateit. A natural universal'Iuring machine simulaLeseach machine IVI in such a way that the running time is at most polynomially longer than the running time of IuL lvlore precisely, the simulation time is boundedby a polynomial in the original rttrrning time and the size of IvI'. In particular, if. lvI runs in polynomial time so does the simulation of iVI. This was abouL Turing machines.Oracles can be presentedin a simjlar tashion. Using a universai Turing machine we can representevery predicate for somefixed 4t fbrmul* po. Here e is P(r) in P in the form l.Lo(e,,r,t"(x)) the code (the incle:r)of P and l" is a suitableterm determinedby the runnirrg time of a Turing machine fbr P. This is still not cluite what we need, since the dependence of f. on e is noL explicit. Thus we shall make an adclitional in time n". thab the predicatelvith code e is deciclable natural assumpLion

336

V. Bottucled Alithmetic

in P will be of the form of preclicabes Then our representation

p o ( e*,, 2 1 * f ).
of Zltl" is provable in 52t. Fbr the the exisbence N'ote thab foL:e sl,anclarcl ALto l,hei:e clui-r,nlifier:s. i:r Xf we ltave only Loadcl. of preclicales representabion is no need to use t.woindices,one for the machineand one for the oracle. We as a theorem. our conclusions shall p.resen.t to,: e Al, LLi E!, ru e l$(ff) 4.18 Theorem. There are formulae F0.,u0 '1, such that i > in S'2, (i) each preclicatewh.ichis in P, provably in 5l ) carrbe representecl r,zl*lur,for somenumeralE; as plg(a, (ii) for i ) 1 andfor every V@) e X;b,there exists an e such that S| V ,p( *1= t) i( e,r ,2l*1",' (iii) for every function / which.is polynomial time cornpulable,provably in. in 52[by the relation f ("): ?Jcanbe represenLed .9r1, ug( E,x,a,2l*1"r , for somenumer ala; (iv) for i >ancl for every function / which is in trf*r, provably in Ti,, the in Tit:l canbe represented relation f (*):a u;(e,x.,y,,21'f ), for somenumerale . 'I

A preclicateP in P is determined by a 0-1 Let us staLe(i) more precisely. tunction /. For this function we have a defining polynomial time cornputable some basic conditions in.921.Thus ior-:.,Ia V@,il which is A! and satisfi.es .Ihe precise statement of (i) is that for sorne P(r) is defined by p(a,O). number e : p o ( a , r , 2 l * 1 " S) l,p(r,0) 7. in the sameway. The statement (iii) should be und.erstood In order to simplifv notabionwe shali clefine { " } ; ( t ) : ? lz d f u ; ( e , x , y , T W)f. We should keep in mind thai iL is nob a bounded formula. However, if we a A!r-r, prove the existenceof 2l*l' , then we can worlc with ii as if it r,vere formula.

cl.

'Witnessins

F\rnctions

337

Fmnc{;iorm (d) Wi il;nessing


In this subsecLion we shall clevelop the moclei th.eory o:[fragrnenl,s ol 52 ancl prove som.ebh.eorems of the Lype menbioneclaL the beginning of Lhis section. We shall use bhe fact provecl above thaL 'Iuring rn.achinecom.pul,al,ionsare clefinable in f::agments of 52. Firsb we shall show Lhat a substructure of a model IVI o,t ?l. that is cl.osecl nncler the funcl,ions of nl,r,is I! elementary in IVl. ancl is a mo.lel of 1-j. The same holds for substructures of moclels of ,5) ancl the functions ol: nf . Tlris is appiiecl to prove witnessing Theor:ems 4.27 and 4.29 'tor fragmenl,s Zj. Then we prove a witnessing Th.eorem 4.32 for:flragments 5l-F1. The theorem has a sbronger conciusion than in Lhe original form used by Bussl it sbates tlrat (Vr)p(*,f (*)) is provable in the wealcertheory T'j,if. ? > 0. T'his allows us to prove imtnediabely th.e well-lcnown conservation result for sr-rchpairs of bheories. As the moclel-theoretical proof of Theorem 4.27 is rather diffi,cult, we first prove an aurxiliary i['heorem 4.31 abou.t extensions of moclels of 5l-l-1. 4.19 Definition" Let IVI be a structui:e for the language of. 52, let A g IVI, i > 0. Assume that each {"};(") defines a function in .Att. The n! cLosu,re of A is the set

{{"h(o)lcretr&ee1/}.
We say thab A is n!, closedi:f. A is its own closure. We want bo use the fact that polynomi,al time prerl'icatesare a,bsolu't,e in eaerAJ! closedsubstructu,re of NI, a model of .9|. As'we clo nob have symbols for all polynomial time predicabes in our language, we have to formalize this statement. Before th.e next definition let us note that every formula can be transforrned to an ecluivalent formula which has only &, V and as connecbives, and ail negations occur only at atomic formulae. What is important is that this transformabion preserves the quantifier complexity. We shall say bhab such a formula is in negat,ion normal form. 4.20 Definition. Let i > 0, leL f be Slif i : 0 and itj otherwise. For a formula rp in negation normal form and a modei IVI ? 7,, we define in.ch-rci;ive1y i;hat g has Jf,_r, Skolem futtc'tions in XI: (1) every open formula has nf_r, Skolem functions in M; (2) if g ancl {, huve nl;*l SlcolemfuncLions i.n lVI,l,hen Vzp, v k rh and gY {
T ) f l have nf_r, Skolemfunctions in IUI; (3) ii p(x,y) has Jli-rt Skolernfnncbionsin IuI, bhen (ly) ,p1r,y) hzr,s nf-r, Slc.olem functions in NI, if for some I in nf*, t

-, p ( " , / ( * ) ) ) . .v/F (Vx)((3y)v@,a)


F l e r e/ ( x ) i s l e p r e s e n b e c bly { a } ; ( ( * t , ( * 2 , . ( " r , . ' . , r r r ): : c .

",rr)

'. .)),

e ly' and "vhere

338

V. Bouncled Ariihmetic

ancl 4,2L Lernma. Let 7 be as above, Iet IVI F T' ,Iet I( C fuI be trf*, closed, in &1. Then lbr a e I i , tet p(x) have nf-,-, Skolemfr-Lnctions

IUIF e@) + K F e ( a ) .
tp. Proof. By inducbion on the complexit'y o'f.

4.22 Cerollary. If p(x) and -?(x) have A'o*t Skolem functions in M F T, :[' closed substructures of IUI, i.e. il A as above, then p(x) is absolule for !f*, is a trf*., closed substructure of IVI and a is a string of elements of .A then

IvItr e@) iff AF e@).


have n{ The naburaldefinitionsof polynomial time computablepredicabes Skolernfunctionsin M.In particular we shall need that ihe formula * - (a), and its negationhave n{ Skolemfunctionsin IuL Now iet P(r) be an arbitrary by polynomial time computable predicate.Then it can be clefined (d.1)

- 11 ); '1 (:.Xl-l Sp(l"l)kCornp(w,*)k(T-u)01

It is clear thab r.oand the existentially quantified numbers in Comp can be computed in polynomial time, bhus (d.1) has nf Skolem functions in IuI. The same is true for the negation of (d.t), hence (d.1) is absolute. Thus in our formaljzation, polynomial time predicates are absolute in the situation consicleredabove. The same argument worlts, of course, lbr polynomiaL time computable funcLions. This enables us to work with polynomial time computable predicates as if they were present in the language. 4.23 Lernma. Let IuI be a model of S$,leI K be a trf closed substructure. Then

K tr LlND-striclib, + K tr Si . 'for stric'tEf forrnulae that in /{ every . First we shall show using LIND Proof will bave LIND E! ^in X '.. Xnb formula is equivalenl,Io a stri,cti'f one. lthus we^ of quanl,ifiers on the number of quantifiers.Several by indu.ction This ir proved. relation. A sharply the samekincl are replacedby a singleone using lhe cocling wilh the next existentialbounded bolndecl universalquantifier is exchalgecl which LIND-stnct\! is sufEcient quantifier using an lnstance of BBi!,lor (seethe proof tf Th.or.m 4.10). We musb checkthat the propertiesof the coclingfunction that we are using are provablealreadyin LII{D-stridtbr. We need the following properties:

4. WiLnessing fiuncbions

339

(1) (2) (3)

: r o a ' " u( a) i : *i) , fbr ever Y ( V " 0 .. . * i ) G y ) ( ( y ) o i ; ( : s ) ( / h ( s) 0); ( V r ,s ) ( l l ) ( t h ( t ) : / h ( s )- FT & ( l ) / A ( , ) : 3 '


k ( V z < / h ( s ) )( t ) z : ( ' ) , ) ) ;

((i)rnf"i is the lasb elementof the sec;-Lence f). Again it is clear that these
lormuLlaeh.avetr1i Slcolemfunctions in IuI. Since they ai:e true in Lrue also rn Ii by Lem.ma <1.22. the proof of
^ f t n D Atf rv4 ll^o, urrLJ ono cLf, rr'

Similarly vre shall show tha.. LIN.D D!, implies PIND Z! in lC. $/e repeaL

z! , sl + LII{D z! r zuwn

only Lhe follorving properbies (Proposition 4.5 (1)) where we needeclfrom ,5r1 of tlre frnction msp:

: 0; msq t(a ,,O) < l " l * L r n s p ( a , S ( b ) ) l 2 L 'r: n s p @ , b ) ; S'(b) - o. msp(a,lol) These formulae are true in IVI and have n{ Skolem functions, hence they are n trtre in K loo. Let i > 0, Iet T be ,9| i:f.i - 0 and T'ij otherwise. Then each 4,2,1'Letmrna. EBQ\ formula has nf-r, Sholem functions in each IVI F T. Proof. We shall use inducLion on the depth of tl.e forrnula. For open forrrr-ulae the lemrna is true by definition. Consider a X,f formula of the form (l t < f ( a) ) qa( a, o) ,
where we assume that ,p(a,x) has nf*, searclt using -I;o oracle
-A

fu.nctions. We apply binary Slcolem

( 1 " ) ( p( r (

qke@,,r))

to find an r satislyingr <t(a)8zp(r,a), fbr a given a. At the beginningwe r). If iL is not set p :: 0 and q :: f(a). tr'irstwe ask whether (1" < il(a)),p(a, true, then the outpr-rtis, say, 0. Otherwise we coirLinueas follo.ws.In each subsecli-rent step we set either

p :: p,

q :: IIf 2(p1-,i-.', if (:r)(p ( ., ( ,-Ilz('p-f q)-'8c a(rr,z))

340 or

V. BoundedArilhmetic

p :: tIf 2(p -F lt).t, Q i: (1, obherwise. By Theorem 4..|-6the computation of length lt(")l always exisbs. In o::cler to prove its correctness we neecl only to shor,vthaL in l,he 7-Lh step of bhe ^"';- ^. f.-;n ^^rclitions arc sabiificd: ^^.-.-",+^ r.;,,- ri. ^ {^]1 Ul-l.(JIL.r.r.I!,JwIIJ.S t vvu uull UUIlfPLtt,d,l,lUII q-

p s t(a)l2r ;
\

(:'Xp \ ; { ,

q k tp6, r ) )

by induction on 1. Oncethe computation is given, bheconciition This is pi:oved we neecl only .L.[NDX!.tt the formu].alras is Xf , h.ence that we ale pL:o.zing bhefbrm ( l r < l t ( a ) l ) p ( ar,) , r,vhere g i.stI!, we proceedsimilarly, but we use thorough searchinstead of binary search.The caseof the universal bounded cluantifierand connectives lf &, v is trivial. 4.25 Corollary. Let T be as above, Iet IVI F ?, ancl let K be a nr?*r closecl substru.ctureof IuI. Then I( is Es(Df ) elernentary. and Lem.ma4.24. 4.2.1' .Proof.By l-,ernrna tl

Since LINDT! is a XN(Xf) formula, we get that /( is a moclelot,Si. BuL 'we can prove more. ancl T) oiherwise.Let l'{ be a 4.26 Tlreorern.Let i > 0, Iet T be,l',) if.i:0 of somemoclelNI oI ?. Then K is a moclelof 1')' sr-LbstrucLure al*, closecl ProoJ.Let i ) 0, let IVI F T, tet .I( g IvI be nf*r closed.We shall show that for for a(r, b)) to t(a(r,b)) (inctuction b) in D!, a formulaecluivalenb every cv(2, this is not (lcnownto be equivalentto) a, thouLgh has nf_,_,Skolemir-rnctions, Then, by Lemma tL22,.t(a(r,,b)) must hold in If . We shall t8@!) forrnurla. tormula the fbllowing eqLrivalent consicler - a ( 0 , b ) V ( 3 r < a ) ( a ( r , b ) 8 z- a ( r + f , b ) ) t za ( c , b ) ' Again we shail use binary search.\rVestart with p :- 0 ancl q :- a, alr-d repeat: q)12', I t: Qt if a('(p -t ct)l2L,lr) yt :: '(p -t, p t: p, q :-- ,(p -l q)12,t, othelrn'ise.

4 . W i l , n e s s i n gF u n c t i o L r s

341

--c,(cr, b), then this algorithm finds in polynornially rrany If a(0, b) and sLeps(in the length of numbers a, b) some c such that c 1 ct, a(c,b), and -a(c-FT, b). Ib uses Iif oracle *("). The existenceof the compr-rLation follor,vs from Theorem 4.'17. We must show the correctness of the algorithm, which wibnessesthe exisheni,ial means bhat bhe value obtzr,inedby the cornputabion 'We 'wanl, to show tha.t bhe in.stance prouably i,n T$. in the induction cluantifier following conclitions are satisfiecl in tl-le j-th step of computaLion:

(1) (2) (3)

q-p<2,-r; *(P,b) ;
- - a ( q ,b ) .

We cannot prove by incluction th.at (2) k (3) holds in each step, since this (1) anct(2) using LINDX| ancl formula has too large complexity,So we prov,e ! tlren we prove separately(3) using LINDni. Note thab we are not able to prove the theorem wibh 5l instead of T), since we are not able to formalize [Jf,*, functions in S$. As a corollary we cor:ollary get a version of the theorem of Buss for fragments Tj.It will be ar, of a stronger result which we shall prove later, but we have to prove iL now, the proof since we shall neeclit for the proof of the stronger one. iVloreover is rnuch easierthen the proof of Buss's Lheorem. 4.27 Theorem. Let i > 0, IeLg(n,y) b" II! "t strictDls and suppose

T; F ( V r ) ( l e ) , p ( " , y ) .
Then for some f in J!*,

,n/F (Vr)p(*, J @D.


a n d ? l j o t h e r w i s e ,b h e n T | ( V r ) p ( r , / ( r ) ) , ili:0 Moreover,i:lT is S21 more precisely, for some e /V, ?'F (Vr)(1y)({e};+r(t) : y kp(r,a)) . i.e.

a tttncbion, Proof. Since for every e, T provesthat {eh-r-r(r) : y defi-nes rivecan use the fr-rnctional notabion. First supposethab cp is nf. By way of for any e. not prove (Vr)g(x,{eh+r{r}) that 7 cloes contraclicbion, suppose Then also, for any k, 7' does not prove

t )' )" v , p ( t ,{ t u } ; + r ( t ),) , p ( x{ ,6 } ; - r , r ( r Y ) )t p ( x , { T } ; - r - r (v

342

V. BouncledArithmetic

sinceotherwise,using W@,y) as an oracle,we could combine these Turring bhereis a machinesinto one prodrlcingy for a given r.'By compactness, model M hor

-r '' ' , T { -<p(c, + -V(c,"[T]l,r-r(.)) {0};-rr(.))


Let K be the closrtreof {c} uncler Jlr-, functions wh.e.r:e c is a new consbanl,. in. /'ul,i.e K : {a e NI l(le N)(III F {eh1-r(.): ")}. -,tp(c, 'tr.25., ,[( is E! elementary a). By Corollary Then, for every a .K, fu[ l= hence/( F (V9)-Wk,A). By Theorern4.26,I( tr T;, thus 7j doesnol, in M-[, p ro ve (V rxl s)g @,a ). blo'uvsuppose thrat ,p(r,u) is strictD!-rr. Then lbr some II! formule-r. someterm l, V@,A) is (321 < t)rh@,y, zr). Hence 1.,(r,zs, z1) an,d (vr)(=y)tp( r ,y) : ( Vz) ( 32) r h@, Q) a, Q) ) k ( q < ,) . by solrre{e};+r(r), b.ence 9 By the first part of the proof z canbe wibnessecl fl can be witnessedbv ({ah+r("))0, which is a nf*r function. 4.28 corollary. Let, i ) induction fbr O(r). 1-, let @(r) be stricta!.-, in T'j. T'hen T| proves

Proof. The assutnptionof Lhetheorem meansthat f(r)

is striclL'f*, and for

@and V by V(x) sl;ridilnr,TiF Q(*):'-w(x). We canrepresent some


Q(*):- (1y)tp(x,y) ,,

V(r) = (1a)rh@,y) ,
where tp,$ arei" nb;. Thus we have

. v ,p@,aD r$ts$r)(1y)(,p(x,y)
By Theorern 4.27 we get some e such that

) )t l : @{,z } ; - r r ( " ),) T i t s t p ( r , , { z } ; + r ( 'v


hence

= V ( r , { e } ; - r r (.' ) ) | r $ la @ )

V/e have nob gained any reduction of cornplexiby, but this was not our aim, funcbion. the incltiction formula for @(z) using our aim is bo.wiLness "Jf,+t Suppose 'we have @(0) and -{r(cL), and ure wanb to consbruct sorne t 1 ct sr-rch that 4I(z) & -'rtr(r * 1). We shall tLsebinary search in a sirnilarway as in

4. lViLrressirrgJ_,\rnctions

3,13

'we Tlreorem4.26. use ,p(r,y) as a II! oracreancl {a};11(r) as a subroutine in order to ask cluestions of the form vr(r,{e}*F1(")).Crr."."-p'Lation goes as follows: s l e p0 : j-l-1: s'tep 12:: A, cl::. ct,) tr :_ {e};a1(0), u:: r :: {a };a r(l (p + i l 1 2 _,) ; i t p Q(p * cl l 2 L , r) b l ren p i: L( p * o)l2r , q ': r-(p-f (Dl2-,, u :.: r. e.[se { e } ; . r r ( a ); r L:: r

Tlre co:rrecbness ir; proved by L.\ND.tr.f showing ihat in th.e7-t6 step we h.ave (1) (2) q-p<z-ja; eQt,t) & gkt, u) . fi

1n! formula consisbs of a prefix of existential clu.antifiers followeclby a .A Zobtbrmula. 4.29 Theorern.LeL > 0, rer cp(x,y,z)he a=II! lbrmula anclsuppose T) ts (Vr)(zy)(V z)tp(*,y ., . ") Then for some fo,. . . , f,, in nl*, ( d . 1 ) , n / F ( V z ,z y t . . . , z n ) ( r p ( r , f o ( * ) , 2 0 V ,f t ( * , " 0 ) , " r )y . . . V) @
. .V V(*, fn(*, zot. . . ,.?n-I),z")) .

Moreover,this is provablein S| , if i :0,

anclin 1',j, if i > 0.

Proof. L e L? b e 5 ] , i f . i - 0 , a n d r j , i t i ) 0 , r e t p , (n,y.,z)bealII! formula. strppose7' does nob prove Lheformula in (d.1) for any choice ib ,. . . , fr. "f Take some entimeration of functions in nr?_pr stich that (1) b.he n-ih lunction /,, depencls on ( n arg.Llmenrs; (2) each / trr?-,-r occurs in the enumeralioninfinitely many times. (1 , ' o i rn s t a n ce ,{e };-r-r(((ri,rz),. . . ,n") ) ,, wher e ( - , - ) is the pair ing t' unction, has bhis properby, assr-rrning natr-rral coding of Turing maclr.ines.) By compacLness,

'f' -p-,9(c, fo(r),do)-F-'v(c,f {r,do),d) -i-' . .


wlrere c, cLg, dr, . .. are ner,vcon.stants, is a consistent theory. Let lVl' be a rnocleL of this Lheory and let

I ( : { J o ( " )f, t ( c , c L sJ .. )z , ( " , c l g , c L.1 )} ,.

34+

V. Bouncled.Alithmetic

Since all pr:ojections occr,rrin the enumeration ancl each function occtLrs in the enr:.merabion infinitely many tirnes, we have (3) c, clg, cl1,cl2,. . . I( , ('L) /( is nf ,., closecl.
,a_ t

By (3), ll'e have

ya .K 1d e I( IUIF _,pG,o,, d) . Since'-tpis VE!,,,usingCorollary4.25we get,


/f F ( Ve) ( 12) - ,e( c,y, z)

Now we woulcl like to strengthen Theorem 21.27 by taliing the weaker assumption Si-rr F (Vr)(:y)v@,,y) insteacl of Tj F (Vr)Qyjrp(r, s). The proof above d.oesnot work for Sl+r, since we are not able to show that a nfn, closecL substructure is a model of SiL-FI. It is worthwhile to reali ze lhe reason, then we shall better understand the forthcoming proofs. Let (!y St)rh@,y) be a E!*r formula, where t! is A!.In a nf*, closedsubstructure the cluantifier lg can be witnessed by different elements of the form {e}l+r(a). Since e runs over standard numbers, we can obtain a cut (for instance contained in some segment [0, b]) for which (ly < t)$(*,y) holds, hence the induction fails. Therefore we shall talce closures under some functions wil]n nonstandard indi,ces. The basic idea of the proof is the same as above: if (=y)gQ, y) is not witnessed by a nf*, function, we construct a subrnodel in which it does not hold. Thus we must be careful when adding {"};+r(c) for e nonstandard. We shall Lrse overspill. If (ly)tp(c,y) is not witiressed by any {"}l+r(c), for e sbaniLard,then this must be also true for all e up to some small nonstandard 11. Tlrtrs we ensure the failure of (1y)tp(c, y) by taking the closure ouly uncler functions with such small indices. However it is not so easy to ensttre the induction. Now we sketch the idea of the model-theoretic proof of this strengthening. IL is essentially the proof of Wilkie with some changes. Let i ) 0, let T be ? does noL S$ it i : 0 and Ti oth.erwise. Suppose that for every e .Ay', prove tp(n,{e};-rr(r)).Take a model fuI tr 7 with solne c }uI, such liraL M t r - - r 1 c , { E } ; + t ( . ) ) . B y o v e r s p i l lt h i s i s t r u e a l s o f o r a l l e - - 1 1 . W e s h a l l consisting of sorne elements of the forrn {eh-rt(c), for-' construct sr-Lbstructures e 111, thus we shallhavg -(Vr)(1a)V@,y) i"the sttbstructures.So we have shall do it inco steps (we use only countable only to ensure LINDEh..We moclels) adding LIND for one formula and one string of parameters at zrtime. Let us talce the formula V(*) : (:y S t)rh@,y) consideredabove, ancl leL a be alreacly in our substrucl,ure K.'We want to extend /f so that it satisfi.es tlre following insian ce of LIIttD D!-rr:

(d 2)

-u7(0 ): j S l " l ) ( v ( j ) k - v ,+ v( ( jf ) ) v r ( l a l )

4. lViinessincFunclions

q,ltr d'tt,

lVe take a suibable12 ( 11) and bry to find a witness gs for 131 in tZ(0) of we try to find erwitness g.1f.or the tbr:n {"hr-r("), r 1 rz. If we succeed, ( 1 ) o f t h e f o r m [ t h + r ( ( . , ( 0 , 9 0 ) ) ), 1 r 2 , a n d s o o n . T h e b o u n d r 2 i s so small tirabwe never consLruct chosen tbr {rh+r(.) wibh e ) rt.Sr-rppose, g1, bLrt bhereis no witness example, Lhat for some j < l"l, we have fbr-rnci -f T). Then we take the |lf*, tj+t o,f the form {t}l-rr((.,(r, gi))') for V(j gi)) (c, (j of and -I( in N|' (ci.2). clostrre We repeat this extension , , anclwe have processfor other formulae and other parameters.In order Lo preserve(d.2) aswe usedfor -(Vr)(=y)rp(r,A).In our par:ticular we use a sirnilarargumenb V(j) is preservecl by Xobelementary extensions,since ii is 3I{, while case, -V(j -FT) will be preservedbecausewe shall add only elementsof the form {r};+1(c), for e .--12. The following key lernma {brmalizesone step of the above construction. if. i :0 and Tl otherwise.Let IuI be a 4.30 Lernrna"Let i ) 0, let ? be .92r o,f. IVI,,a,b /(, let Q(r,y)be stricttf-r, ancllet moclelof ?, K sr-tbstructure f(y) be strictll!4. that Suppose (1) .K is a n'rn, closureof one element of IVI; (2) K is not cofinal in IVI;

(3) /( F r(b).
such that.I{ g J('' g IVI, (1)-(3) holds Then Lhereexistsa substructure.K'k for y'(* and

b). b)& -,a(i+ T,b))v o(lal, b) v (:j < lolXo(j, (4) K'' tr -@(0,
Proof. Leb the assumptionsbe satisfied.Let J( be generaiedby an element in IvI, there is some d e IUI such bhat c. Since K is nob cofi.nal

x K + l r l < 1 .. /l
Since errery r e I( is computed in polynomial time from c, and by overspill, we have

(.t.3)

l"l" 5 ld, l
*, U,21'l'7 , {t};-rr(") : It = u;-Ft(",

fbr some rs nonstandard. Recall ihat

formula n Abo+, we canwrite {t};+r(") - lJ z,s By (c1.3), wlrere u;11is Abo*r.


wibh an adclitional parameter d, since : ( V t < d ) ( , - 2 l x l "- > ' u i - r r ( " , r , ' y , z ) ) fuIF u;a1(",r,y,2ltl"; - 2l'"1 : (1, 3 rL)(, k' r ; - t - t ( " , r , y , 2 ) ) .

J.,to

V. Bouncled Arithrnetic

Let

@ ( r ,b ) : ( 1 , < t ( ' ) ) q r ( r ,z ) , :(Yris)t|,,(z), !Z(h)

',,2i:.ere g(rc,:t) is llb, andrl.,(z) is x;b; we shall. ornibbhepararnebers fr.[rom now on.. Consiclerthe follo''ring :fbrmula

(.1.4)

( V t< l t l X { r } ; - r - r ( < .) ) ) c )s - , , 1 , ( [ r ] ; - r - r ( "

Sin-ce .K ? V.,I( is L'! eLementaryinlVI,ancle.reryelementof /{ is of the fbr.m r\lso {"}*vt(c), for e standard,this fbrmula is true in h'|, for every r sl,anclard. this fbrmrr.la is ecluival.ent to u Ao;+t formr-rla in AI. Since,S:]. I r A1, (Theorem 4.7) and for i ) I, T; l- I A!-r, (Corollary tL.2B), we can use overspillto cl.eclu.ce tlrat (cl.zi) holclsin I'tl for some r nonstanclard. L,et (.1 .5 )

r r : m i n ( l " olll,r l l , l . l ).

Now r,ve construct a Turing machine with an oracle which searches for the r,vitnessof the existential. cluantifier in condition (<t) of the lemma. The machine works on input c as follows:

1: 2:

j :: --I; g :: 0; find tlt,e f,rst e I 11 snch that

(i , g)))< t(j + T)& v(j + r, {'};-rr ((r,(i,g)))) {'};+r((., ;


3:
-J.

( - r , 0 ;: 9 .

iJ strrch an e doesnot eri,st or j - l.l - 1, th,enprinl. (j,g); j : : i + T , g : : { e } ; + r ( ( c , ( f , g ) ) ) ; g o t o 2 w h e r ew e d e f " n e otherwise

Lel, e6 be bhecodeo:[this Tr-rring rnachine.We shall esLimateee. Recall tha,t it is also the ercponent in the time bouncl of bhis Turing rnachine and this malcesthis estimatenonl,rivial.To rtln some {.};-fr r,vith e ( 11 on input c we neeclat most lrlt < lcl"' sbeps. In 2: we cloi{, 11 times. The time neecleil to check tl.reII! conclitiono:[ 2: is a sbanclard. polynornial in lcl. (NoLe that we lrtr,ve to evaluatealso the implicit paramebers b, for i,vhichvyerreedl,imea sfandarclpolynomialin lcl.) Finally the cyclegiver:by 4: is repeatedlal -f 1 times. Thus the total rurrring tirne is boundedby

l 1 ), r - r ,p r ( l r l )' r . ( l a + l . l ' ' '' , .


where p(z) is somestandardpo1.ynornial. We have11 < lcl and lal-f 1 is aiso polynomial in lcl, since i'u is an elemenl,of.I:i. botLnc[ed by some stanclarcl (,ve Tl-r.uls cz-rii. bounil the abo've expression by lcl2t''. Sincethe program of bhe T'lLringmach.ine has stanclalillength) trve can bounclits cocieby
eg<.2ry

4. lVil,nessirtg FLrncbions

3'17

prinLs(j,g) on iLs or-rLpLr.t. lVe define.I(* as fbllo'',vs. Supposebhemachin.e : -cl(O) of r7(c), then,I'f I{'t', is t.henl*, closuLre F otherwise -/('* V /( If c'o: (c, (j , tJ)): (", { h_rr(")) . "o assrllrre lhal, K l= Norv lve hzrve lo checLicorrdibions(2)-(,+). lVe s.Lrall -Q(ct), in. MI. henceIVIF 0(6), as./( is E! ele:mentary 4,(d)82
1y" [aire

j l"l2''' , 1."'l
hence the len.gth of each elemeut r. o'f..K"' ie boun.ded by'

Since "t

, /c iV . l:tl < 1c1t'r'2" 'bounded ( Irol, th.e by is; last ex1:ression

< l d l, l"l"o
.I("' cannobbe cofinal in IVI. LhrLs 'Io prove (3), sr-rppose that it fails 'f.orI("', which lneans thaL for some e 1 I V a n d h : { e r } ; + r ( . * ) , Jr*Fhlsk-nb@). By CorolLary4.25, "I("'is Xf elementaryin jlr.I,thus bhisformula holds alsoin from c using a machiuewith code eg 1211. /rzl.blow c"' has been constructed runs fi.-steg anitthen e1 has a the Turing rnachine.,vhich Sincee1 is standarcl., by (d.5), e2 I lr I. But this is in conbradicbion codee2 polynomialin 11,hence, wittr (c1.4). Ib remains to prove condition (.1)for .I(*. We shall show that -FT) , y'f* F r < lal ,AA(j) 8z-<D(j
'I'he first inecluality.is clear, of the machine es. where (j,g) is bhe ou.tpr-Lb since the.n:achine stops before reach.inglcrl.AIso O(j) is easy)because,by the program lines 2: ancl ,l:, ? tl < t(j) k,p(i , s) , ML' of contraclicbion suppose thai; hence also in 1f"' by elemeni,ariness.By r^rz'r,y

e [f , tor sorne bhat, usingelem.entari:.ress) (again fi ? rD(j-fT). Thisrneans . T,{r};+r(r'*)) tlvt tr {e};1-r(c*) < t(j -FI)& v(j -tIn pzrrtrcuLlar e ( 11. Recarllthal, c,' :(c,(j,tJD, stop al, 7, ',.2[ichis a conLrzrcticl,ion. shoulcl nol, thtts the lrrerchine n

348

V. Bounded Alithmetic

4.3L Theorem. Let i ) 0,Iet T be S'21 if i : 0 anclT$ otherwise. Let M be zt, cottn'tab/e moclel o,f T, K a substi:ticttLre of AI, h Ii, ancl let f(y) be ..:.4 strtctlli_,.1.Su.ppose bhat (1) K is a Jl,_r, closureof one elementof fuI; (2) K is not cofinal in M;

(3) r{ trv(h).

Then there e:cisl,s a substructure J{'F sr,rch that K C I('' C AI and

(4) K* F r(b); (5) I{'' F S,i*'.


Proof. lVe shall construcb a countable chain of substructures

K:KoeIheI{zc...cIvI,
and s'trict, fi !*, formulae itr:Vo,iLt,Vz,.... Take some enumerationof stric'tD!*. formulae with parametersfrom IVI. In the j-blr step,7 ) 0, we ensureLINb-in Ki ror the first formula d(a,b) from tlris enumeration for which LIND fails in K j-t, and whose parameters b are in Ki-t. Suppose we consider

--e(n - @ ( 0b , )v (!c < lal)(@(r,& - FT ,b ) )v o | a l , b ) lr)


in tlre j-th step. We apply Lemma 4.30 to .Ki, Q andt?i. Thus we obtain I(ia1. Then we define Vj+t by Viry = [,j k"-$(0),, i{ I{ia1F -rD(0); -l-T), Vj_rt = tl/j &,--,O(k i f K i a l F l c1 l a l k 0 ( k ) k - $ ( k + tIi'4 = Vi, otherwise. 1), f o r s o m ek e K i a l ;

EIerc Vial is not written in ihe strictll!-r., lbrm; however to show thab they are ecluivalent to such formulae, we need only properties of the coding relaiion that are expressible by equalitiesand hencepreserved frorrr MI Lon?, closedsubsbructures. Sinc. Vj+-t will be true in all Krr, f.or nt. > j + 1, also LIND@(a, b)) will hold in these models. Considel for instance the second case, wlrereI{ia1F k < lalkaft) k-'Q(lc r- 1). Then (tr(k) wlll be preserved by extensions, sinceall |;hesLructures zr,re nf*, c1osec1, hence E! dement,ary, while -Q(lc *T) will be.preserved by the clefinitionof the extensions.Let Ii* be the ttnion of this If elernentarychain. Then all formuiae Vi are true in /'('*, henceib sabisfi.es conclition(a) and K* F LIND-sirictEb;a1 . By Lemma 4.23 this implies that -I("' is zrmocleloL S'-"rt

ft

4. Witnessing Functions

349

Now we are ready to prove.Buss'switnessingtheorem. The origilal theorern was proved wiih ? Si-lt; the next theorem is a strengtirening (since = T) is a subLheory of S|+1) which is also ch_re to Buss. 4.32 Theorern.Let > 0, Iet tp(r,y) be t!*r. Suppose thaL

s:i-Ft t (vr)(=fie@,y) .
/

tnen lor some7 rn Ui_p1,

T F (Vr)e@, f (*)) ,
wlrere T is S) if. i :0 and 2j otherwise.

Proof. First observe that we can bound y by some term f:

Si+t t- (Vr)(ts < t(r))tp(x:,a) .


This follows from Theorem 1.4. By corollary 4.'12 we can assume thab g is strictE!-r1 ' Supposethat the conclusionis false. By compactness,there exists a countable model IVI of 7 such that

M F -'p(c,{rh+r(r)) for everye standardand moreover,2'."1" , fr -l[, is not cofinalrn ]vI. Let K
be the nr?-,-,closure of the element c in M. Then ,I( is not cofinal in lz1 and

/( F (Vy)-v(c,y), since,by Corollary 4.25, it is xf elementaryin IuI. Let,v be (Vy 1 t)--tp(c,y), then it is eqr-rivalent to a strictfrf*, for*.rla alreaclyin preclicatelogic. Thus we can apply Theorem 4.31 to get some

K* F SL*+ t (Vv5 t)-p(c,y),


which is a contradiction.
-Ilr.e most interesting case is with i: O, lve sbate ib explicitly.

fl

4.33 Corolla-ry. LeI g(x,y) b" a Xf fbrmula. Suppose that

s'r1 F 1Vu . )(1y)w@,,a)


Then fbr some / polynomial time computable function

s'lt-1vr;v@, f (")).

350

V. Bounded Alithmetic

4.34 Corollary.For i 2 t, Strl-t is a i'f-,-, conservaLive extension of Tj, i.e. for e@) ir i'f-p1, s;-t-L | (Vr)p(z) + r; F (Vc),p(c) .

Proof. Acld a dr-Lmmye;<istentialclu.antifierto ,p(r) ancl apply Theorem. 4.32.

n
Recall that A! sets are setscompuLable in polynomial time using oracles from Ef -r. Lebus noi;ethat it is an open problem ',rhetherE!*rnil'f-r, : Al. Thus Af definablesets neeclnot be in Aeo, ',vith respect to S$. Then 4.35 Corollaly. Let i ) t, let tp(r) be a A! forrnr-rla subsetof N. 9(r) definesa APo Proof.For p(r) we can find someformulaeo(r,y),9@,y) in nbr-, anclterms t, s such that

S $t se @ )= ( 1 a< t ) a ( r , y )= - , ( 1 y3 ' ) g @ , , i l.
Hence

< t ) a ( x , y ) v( 1 y< 4 g ( * , i l . S ' $ r1 = y


into i! formula; thus we can The last formula can be easily transformecl " apply T'heorem4.32 to obtain somef inf:! such thai F ( V r ) ( ( f( " ) l t k a ( r , f ( r ) ) ) v ( / ( " ) S s & 0 @ ,f ( " ) ) ) ) . "nf Hence we can decide p(r) bv cornputing/(t) and using " ilf t oracle. ll

Again the most interestingparticttlar caseof the last corollary is the one with i : 0; it can be stated as follows

x e P. o c o N+ P Sl,F slr x e tuP

A"ritlarne{,ic of Eorended (e) Oruthe FirLite Axiornafizabiiity


of ihe finibea,'riomatizability the quesl,ion we shall consicler In bhissubsection Lhe fragmenl,s S:i ancl T$ bhat of BouncleclArithmetic. First we shali sho,,v are finitely a,riomatizablefor i ) .[. Thus the finite axiomatizability of 52 is eclr-r.ivalenb to the statement bhat the hierarchy of these theoriesis infinite.

.1.. Witnessing Functions

351

shallrecluce the finiLe axiornatizability of. S'2 Then r,ve to th.enon.-collapsecrness of Polynomial Flierarchy. 4.36 T'heor:ern.For i > 1.,each of Sl2ancL Tj. is finitely axiomatizable. Proof-sketch. T'he basic iclea of the proof is the sarre as tor fragrnents I Dn, seeTlreorem I.2.52. However, since this theorem is important, we give at leasb a skeLch of the proof. In Theorem 4.IB we s.howedthat there ate L'! formulae p; which are i1 a sense universal; more precisely, for every tp(r) in x.f there exists a1 e such thaL ,5'|t ,p(r): Ltt(E,r,Zlrlt! Furtherm ore, zl'lu p.rovably e:cistsin sl; this fbllows from the bound 2lal,-vrS 2l*lu_gr . For induction and PIND we need formulae which have an incluction variable and parameters. One parameber is suffi.cient since the pairing function and the clecoding functions are A! definable. For the same reason, we can extend the result above to formulae with two free variables:

s ) F p @ , y )= p i ( z(,* , a ) , 2 1 * l.a ;
Another easymodificationof the result aboverei'rcls as follows: n F z > 2 l @ , a )S') 1 "(rp(*,,y ) pt;(e,(n,il,4. = )
Now we only neecl to show that there exists a finite subtheory ? of Srl stich thal for all tp(r,y) in x! the above equivalence is provable in ?. once we have strch a T, we can finitely axiomatize Si by T pLus .B,A.g-IC plus

-. 1,,;(", (V.,U,z) PIN D (z > 2l@'a)1" (*, y), r)) .,


wlrere r is the induction variable (thus boLrnded in PIND), and e)a)z are parameLers of incluction. Similarly we can axiomatize Tj. To find such a T we have to anaryzeformulae pr; undu; of.T'heorem4.1B more closely. The concept that we wa,nt to use is just Tarski's conclitions for the definition of th.e satisfaction of E! formulae. Tarslci's conclitions d.escribe the satisfaction of forr.nulaeof a given class via relating the truth of a fbrm.ula wiih the truth of its components. Fbr eaclr.closure condition of Lhe class of formulae there is one Tarshi conciition. The class es E! were clefinecl in Definition 4,2; Larthermore tve have to use the inductive concliLions fol the value of terrns.

352

V, Bounded Arilhn-retic

hlow we shall describe the conditions more explicitly. (1) Fbr each function symbolof L2,,ve shall have an integer consbant, e5, - 1 , e * , e b c . ,s u c h t h a t

S ')| u s(es t r t a.,2lxlzs ) = y :,5( r )

S ) r - u s ( e +(," r , 1 2 ) , u , 2 1 * , ' ' " ' ) l u ) -- t J

rI t T2;

e tc. (2) For relation symbols : and ( we consLruct binary functions e- ancl e< such that , ,2lal"= ( "' ' "t, ,9 1ts . p o (e :(e 1 ,e2)r

*, yz,2l'1" rt,,U \ , 2 l x l ")' k r o ( " 2 , = ( 1 y t ,y z ) ( u o ( e ; )'k n : a z )


and similarly for (. The functions ur" Alt (3) For each of the conditions of the definition of the class Xf we have a function and a condition. Consider for instance the condition

ae D!+(Vz<ltl)'E!.
Then we have a binary A\ definablefunction elvl such that

e1 z,)r,, 2 l ' l " l o t , | pt;(elvl(e S') = (Vyz , ) ( u s ( e 1 , r , U , 2 1 'k1 z " ;l l y l * ' 7"1 t t ; ( e z , (" r) ,,21G'')1

does the following. T'he flurnctioncompuied by the machine with the code e1y1 If e1 is the code of a machine tor a term t and ez is thei code of a machine fbr a formula p(r), then elvl(er,ez) is the code of a machine which decides tlre formula (Vz < ltl);o(x, z). needed Let ? consist of the above condibions plus finitely many senbences d e f i nition). , nicluenest,A?. ... (existenceu e-, b of o r m a l i z e f u . n c t i o n s "(,elvl, Then we can easily show, using induction on the depth ot tp X;0, that there exists an e such that T t s < p ( r )- - L t ; ( E , , r , 2 1 * 1 " 1 . It remains to prove the above Tarski conditions in 5'21.A formal proof woulcl, perhaps, recluire a lvhole section. BLrt in fact it is not hard to see condition, that they must be provable. Let us look at the last co.nsiclered which concerns the sharply bouncled universal cluan"ti'fier.Formula pl; has been clefineclas being true if ihe'Iuring machine wibh cocle e accepts inpub r. The conclilion describeswhat the machine.,vith code e of the fbrm e1v1Gt,ez) .r,vill machine e1 on r and obtain y (the value of the sinrr-rlai,e wiil do: first it

4. Witnessing fi-rnctions

J53

ierm), Lhenit will simulatemachinee2 on inpr,rts (*,r) wiih z <, lyl. Thus ',veneed only to choosethe codesof the machines in a suitable way. n 4.37 coroliary. ,92is finitely axioma,tiza,ble if ancronly il ,Siz:,s'$ for somei. We ale no|; erbleto decide'',vhet,her: S'2is finiiely axiornaLizabie, lVe shali only slr.ow Lhat it is not, ussuminga plausibleconjectttreLhat, the polynomial IJierar:chycLoes not coJlapse on its fi.niLelevel. This itoes not seem bo be a surprising lesult becauseof bhe relation betr,veen the classesof formulae and comp.l.exity classes El. B* the relation is not so straightfbi:.,,var:ci. ,X,f Inducbionand definability are different things. We shall shovrthab ind.ucbion is relatecl io optimi zation and we shall Lrse this fact in or:r oroof.

4.38'tr'heor.em. Let ? > 0 and suppo se Df_r,* nlt-2. Then f| 7 Si'rt


4.39 corollaly. If the Polynomial Hierarchy does not collapsethen

Tj --lr;+t for i ) 0. (ii Si + 5l+1 fot i > I and,


(ii) ,52 and T2 arc not finitely axiomatizable. (iii) //0 * At is not finitely axiomatizable. (iv) //6 is not finitely axiomatizable. Proof. (i) and (ii) follow using the inclr-rsions (iii) /40 1- ,Q t is contained in 52 in such a w a y t h a t b h e f u n c t i o n s o l 9 z are definable in IAo * At. (iv) Just recall that J71is a single axiom. n Before proving Theorem 4.38 formally we shall sketch the rnain steps of this proof. We shall consider oniy the case i : 0. Let C(r,y) b" a polynomial time preclicate. Considel the optirnization problem of fincling maximal g such that y ( r and c(r,y). In order to simplify noLabion we shall assume that Lhe bouncl A I r is implici t in C'(r,,y) ancl tlrat C(r,O) holds for all r. We shall call y a feasible solution to r il C(*,a).Let 9 be ctefined by

rl q s'le ri s sa

(" 1)

n),) s ( x ,y ,z ) : C ( r ,y ) k ( l u l< l r l - - - - - C ' (z

Since C is polynornial time co.mpubable, <p is D!. iVow the mezrning of ( V z ) 9 @ , y , 2 ) i s t h a b y i s t h , ef e a s i , b l e s o l u t i o n t o r of rnatinzaLsize. In ,S') the sclrema.LlNDtrf proirestha,t for every r there exists a maximal feasible soltrtiony.It T3 : 3L, then this is true niro ir.,?r0,i . e .

r) . S r 1v";(1y)(yz)tp(r,Tt,
. e obtain $ome functions N o r v w e c a n a p p l y T h e o r e m 2 1 . 2 9W fo,.ft,...,.fn compulable in polynomial l,irnesuch that

.J DZI

V. Boundecl Arithmetic

(".2)

zt) V''' l V F ( V r , 2 0 t . . . , z n ) ( t p ( r , / o ( " ) ,r o ) V p ( r , f 1 ( n , z o ) , . . . Y t p ( n ,. f . r . ( r z , o , . . . t z , n , - l ) ,r " ) )

tli.abstrch l,o shor,,r bheory. We wz'r,nL The resi; of the pr:oof is only comple,.ribV allvays were furn.ctions cannoL exisL for all polynomizr,lLime predicates C'.II n for 0 (as i.n-Buss's theorern) bhen or-ir goal would be sirnple. In such a czr,se, exarnple, we cotild construct a maximaL clicl'uein a graph. using a polynomial time com.putable function /s. This is impossible r-tnlessP : A/P. In fact we can talre any /VP problern, nob only optimization proltlems. Let C'(r,y) be definecl, say, by lJ : 0 or y is a Hamillonian circtLit in graph r T'hen foki would be 0, if z is not Harriltonian and /s(r) would be a, lfamiltonian circuil in s otherwise. Bub in general rz need not be 0. So we rnoclify otr.r probiem asj fbiloi,vs: rwi!.I be a string of graphs and y will be a string of 0's and Flamiltonian circuits in corresponding graplr.s.i\ow we assume that lbr Lhis particular C we obtain (e.2) with n - I, (this is suffi.ciently instrucbive). Consider an n' a bwo element string * - (G;,GZ), where both graphs are llamiltonian. Then there are two possibilities: s o l u t i o no r . f o ( " ) : ( n , y z ) w h e r e y r : 1 J 2 : 0 ; (1) /o(") is not a feasible (2) /o(") : (uy y2) ancl y1 is a llarniltonian cycle in G1 or y is a Hamiltoir cycle in G2; In case (Z) fo has prodr-rcednontrivial inforrration. Iir case (1) it is not so, but /1 :must prodtice inlbrmation: if we take zg : (y',0), where yt is some since either i6(r) is HamilLonian cycle in G1, Lhenwe have -g(r,/o("),zs), not tr,feasible solution or z0 is a better olle. Flence we get (Y z1)9@, f {*., zo), zt) , (Ar,az) i s a n o p t i m a l s o l u t i o n .T h i s m e a n s t h a t f t @ , z o ) : i.".,fi(r,ril -, G2 where both yl ancl a2 ate Hamilton cycles. Let us draw an arrow G1 in case (1) and an arrow Gz -- G1 in case (2). f'his indicates that by having a Flamiltonian circuit for the tail \ /e can construct a Hamiltonian circuit fbr the head. Consider all Hamiltonian graphs of size m. Tlne above argument shows that there is an arrow at leasb in one direction between any pair ol them. An easy counting argumenb shows that there exisbsa set of poiynornial size of such graphs which covers the rest. For each of bhesegraphs, we choose a Hamill,olian circrrit in it. Let [n be the set consisbing of these pairs. [1l we can decide in polynomial time r,vhebher has also polynomial size. Using lFll a graph Ci is Hzrmiltonian: try il a,nc[-fi ott ali pzrirs (C|,II) and (/'I,G), FILCompul,aLionswhich use additional polynornial size where l'1 r.uns l,hror-rgh

4. WiI;nessing Irunctions

')DJ

irr.formaLionare called coltrpuLaiion.sv''riLh polynom,ial afluice. T'her.eis :r wr:ll lcnow-na'r,rgurnenL (see fl(arp-l,ipi;on 80]) siro,,vingbhat if every /\/p problem celnbe compr-ttecl using polynomizr.ltime cornpr:.tzrtion with polynorniai ,rc[uice,

';T Nor;v e prove 'Iheorem 4.38in clebai.l. lVe sh.all clivicle th.eploof irrl,oseverzr,l l,ebi ) 0 begiven.r\.for.mLrlaC'(n,y) lerrrLrras. will be calleda.lI! op.ii,mizat,iort probLernitc'(r,y ) ar{fbrm'-rlaserch.thaLc(2, is ) lc'(r,"y)-->y1.nis a0 'c pro'zablein-predicate logic. lVe shall say that y is a length opti:rn,aLsoluri;io1 Lo r if the followinq formu.lallolcls:

El - tItr. then

(Vz)(C(r,y e) ( M < l r l - , - - C ' ( n z,) ) )


4.40 Froposition" For every II! oprimizaLion problem C', Sr+r proves ther,t; C has a length optimal solution to every r.

ProoJ. tfi(x,y) be defined L'et, by tlt(x,u, ) ( 1 a< x ) ( C ' ( r , y ) = klyl >_ . ") By LIND E!-r, we have -t!(r,0) v (lu < lrlX/(",tr,)k.-.tf.t(n,S("))) v tlt(r,,lrl) .
since ',h(r,6) is always true, iL just e,rpresses th.e existence of a length optirnal solution. n

It is an easy exercise to plove thab, in fact, the existence of length op{,imal solutions to tI! opLirnizationproblems is eqtti,aalent to PIND i!a-1 oyer a sufficiently strong base theory, say S$. We are going bouseTheorem4.29.It is convenienL to refer to the conclusion of the theorem, see formula (e.2) above, as a kind of interactive way of c o m p t r t i nga w i tn e ssto l y i n ( Vr ) ( fy) ( Vz) tp( r ,y,z) .Let r be given and stlppose we want to construcLy. Assume (e.2) holds true. First consider z), then take y : .fo("), othe:rwise there is some .zs /b("). If (Vz-)p@,,f0(r), suclr that -tp(r,fo(*),26). V/e shall call this z0 a co.LLnterexumple to ft(r). By (e.2),it must be n ) 0 and we have / V F (V z, zrt. . . ,, zn )(tp (n, , V ... f ( *, z0) zt)
' . Y , p ( " ,t " ( r , z 0 t . . . , z n - L ) .z , n)) . flence vve can repeaL bhe reasoning above and we get that either J(r,to) wil,nessesy or the disjunction calr be redr,Lcecl ftrrther by taliing a countere,varnple z1 Io h(*,zg). I{o.,vever this processcan be repeated at mosb rz birnes, since when we obbain (Yr, zrr)tp(xf , n(r, zot. . .t zn-r), rrr)
r,ve catt sbop.

JDU

V. Bouncled Arithmet,ic

problern. Assume tha.E C(r,y) b" a .trf opbimizzr,tion Ti : ,g;.-". T'hcn. .Let Iij proves thab C' has always length optimal solr,rtions.Tbis staternent can be e;rpressecl i:.r. the form (VcXfyXVz)g(r,y,z) with rp in E!-rr, see (e..1). i.Ience, by Theorem 4.29, we get tirat an optimal solution can be cornputecl using n,?-r, firncbions ancl a fixed number of counLerexamples.This is our first
I

J.emma.

2tr.4-{X,emma. If TA: SL-",; ) 0, |;henfor every tI! optirnizaiionproblem iLs


length optimal solutions can be computeclr-Lsing nr?-n,functions and a fi;recl number of counterercamples.

'.['hecompr-riations with countere:rarrrples clonob clefine a function comple,rity class,sincewe allow the askin.g of counterexamples only for the particular optimizatior: problern that we want to solve. If we alloweclaslcingarbitrarS' clueries of the samelogical complexit.f,we woulcl obtair: just the classtrf*r. But it is easyto prove that optimai solutions can be computedby such fu.ncLions.Thr-isour ne:it steDis a leduction bo a different function class.
r1P 4.42 Definition. A function / belongsto Jllpolg if bhereexisbs someg in u i ( p(k) stich that for and a polynomiil p(r) such that for every fothere is o6 every r, lnl : ft,

l@): g(r,an).

Here cr7, is called a Ttolynomial aduice; it is some extra inlormation given for free for each size of input; the dependence on the size of input is cluiLe arbitrary. 4.4i3 Lernrtra. LeL i > 0. Suppose that for every I/f optimizabion problem its length optimal soluLions can be cornputecl using al+, functions ancl a fixed number of counterexamples. Then for every ,h@,y) in .A! there exists an /

in trf-r, lpoly suchthat N F (Vz)((1a S r)r[(r,a) - ,h@., f ("))) . problemp(u,,u)by .Proof . Let { be given.Defi.ne flf optimization " r) r ,, )& (r)r < (")r). ., & ( V t< l h ( u ) ) ( t 1 . , ( ( ( p ( r , u ) : I h ( u )< l h ( t t ) lunctioni sn n f - r , w h i c hi n t e r a c be some l , e L/ s ( u ) , . . . , f n ( n r u 0 , . . ., u n - 1 . )
'f.orp. Let k be given. lVe shall conl,ively compube length optimal solutions {br inputs of size /c. Once we clescribethe aclvice, sbruct a polynomial zr,dvice |;he clefinition of the funcl;ion o will be clear. l,et

V 1: { r ; l * l : I tk ( 1 y < ")r!@,y)}

4. lViLnessing fnunciior-rs

352

Choosesomefunction tu(r) such that, for r V1, r[(x,u(n)). we shall trse the following noiation. If rrz( //z(u) trre' w(r, I m) d e n o t e st h e se q u e n ce (,r.r.'((t,)o ) ,. ( t) ,) ) . To eachn * 1- tup1e ..,lu( tz o[ eler,ve ments o,f.V1, assignapair (l,u).,0 < / ( n, r-ising bhefollor,vingpi:ocec1t,.re: S t e p 0 : co mp u te/o (" ); i f p (u ,,/o( ") ) atd th( Js( ") ) > 0, then p' r / : Q go to step 1; , : (/o("))e and stop, otherwise ::O S t e pm ( m ,< n ) : c o m p u t .f * ( u , , z u ( u l 0 ) , . . . , u 1 ( u l m - 1 ) ) ;i f p ( n ,f * ( u , , u ( u | 0 ) , . .. , t t ( u | - - t ) ) ) ancl I h ( f r " ( u , u ( u l 0 ) , ., .z . u(ulm- i)))> -, tlrenput l-mancl u0 ) , .. . , u ( u I m - 1 ) ) ) , , , a : (J*@,r.u(| otherwise go to step nz f 1; Step n: If we have reachedthis step, then it necessarily holclsthat p ( u ,f n ( u , - ( " l 0 ) , . . . , u ( u | - t ) ) ) " and Ih(f"(u,,-("|10),...,,u(uln 1 -) ) ): n I . I , thuswepr-rt/:nand U: fn(u,ro(" l 0 ) , . . . , w ( t t|," - 1 ) ) , and stop. Let trs caII y a uitness lbr e if ,1,,@ ,,A). The rneaningof bheabove procectr,rre is the tollowing. Let ,u be an n * 1-tuple and let l,y be assigned to it. T'heir Ir a v i n g ' u v itn e sse fo s r (u )g ,...,(n )t- lr wo can com ptr te y tor ( u) 1. the wil,ness .Vt\Q, For an n-elemenbsnbset Q *" shail say thaL paii: V1 and r e "f (Q,") ts good if for some arrarlgemen { rtg , . . . t r l - 1 x : l + 1 . . . , r n } o l Q , I i s assignedbo r.rby the aboveproceclure, lvhere u is the secluence
( " 0 , . . . t i c l - l r x , r l + ' 1 , .. . , 0 r ) .

358

V. Boundecl Arithmetic

a sequence of sttbsets o,f.V1 .Defi.n.e

V t ) . V z )V s " ' ,
in the secluence The j l- 1-st elemenL el.ements. having /V1,l{2, . . .?respectively, as follows.Find an n-elerrent suLrset is cle{inecl Qi 9I,'i such that

dl > ; + r l { " e V i ; ( Q i , x i)s s o o }


ancl Lake

N; - n

V j + , t : Y , \ { " e V i ; ( Q1 , * ) i s g o o d } . We mtr.stshow thai it is always possible to choosesuch a Q i .BV the procedure above, for each n * l-eiement subset {ro r...rrn} of vi we can construct a good pair (Q, r) such that {,0,"',tn}:Qu{r}, (rs, ...,rn). by taking u to be the secg-Lence FIencethere are at least (/lr)

(f ) n-element subseis Q of V5', so goocl pairs. On the other hancl there "t" at least one such Q must form good pairs with at least

( ui \.1t,)
\"-l-t) \") elements.Hence

n*t

Ni+r S iVi
from which we get

-lf;-n

;* ,trui

-F1),

< (*)'', ffi.,'


\n*1/

* h*

(#)'

* (*)'

-,

Hence we get lft < n * 2 after f steps, for

-Fr)/n))-t < ollog2Q\) - o(k)' t < log2(N1).(los2((n


pairs (r, tu(r))' trVe take bhe polynomial size advice to be the seclueuceof wh.ere tr runs trough all elements of

Q t u Q zu " ' u Q t u V '

4. lVitnessing Ifunctions

35g

Now let r be such that

l ' l : k k ( 1 y< d t @ , y ) ,
i.e. e e Vt. T'hen either n Vt ancl so a witness lbr o is in. the aclvice, or (Q5,t) is a good pair for som.ej St. But then we can consbruct a witness fbr r from the wibnessesfor the eJ.ements of Qi (using nf_r, functions in the proceclure above). n 4"44Lernrna. Let i > 0. Suppose that for every ,b@,y) in III there exists an

suchthat f in lle.-rrlpoly

-, ,!@, /r F (Vr)((:v S x)rh@,y) f (*))). ThenElt-r:nl+2.


Proof. Let .4(6) be in il!*r.Withor-Lt loss of gen.erality we can assumethat it can be represented in the form

( V r ) ( l "-l l b l* ( l y < b ) t f t , r , u ) ) ,
where 1 is II!. 'L,el tft(:c,y) be " n! formula such that

, h ( ( b , * ) ,= , v l)" l : l b l- + @ S U B c I ( b , r , a ) ) .
I'et f be a function in z't-rrlpoLaguaranteedby Lemrna 4.48 f.or {, ancl let function 9 in nr?*, ancl polynomial p be guaranteed by Defini tion 4.42 for /. Then we can write .4.(6)in the following D!*, fbrml l l b l- j ( b , t , g ( o , ( b , , ' ) ) ) ) ) . A(b): (1a,)(< ap(l(6,b)l)&(V"Xlr: The right to left implication is trivial, the left to right implication fbllows by talcing an advice a. n This finishes the proof of Theorem 4.38. There is a similar reduction.of the cluestion5i : T;? to a problem,in complexii,y theory. F'irst we sirall define tha,'r, for an opt\rnizalion problem C(*,y), U is an oqt'ti,mal soluLtion io r if

z)(C (Y ( r , y )k ( a { z - + - C ( r , , ) ) ) .
Secondly, we define interacbive cornputations with an unbounded number of counterexamples as the Lhenal,ural extension of the concept above. Of course,
1 1.lo{r"bhaL g is tIl*., defirrable b;, fh6slsln z}.lJ.

360

V. Bor-rncled Arithmetic

ther:e is an implicit polynomial bound to the number of counterexamples, since we consider only polynomial tirne oracle computations. Note that the nttmber of feasible solrtLionsLo r may be exponential in lzl, thus polynomial numbel of corinberexarnplesapparently cannob help to solve some diffic1lt optimization problems (such as the TRAVELLING sALESPERS9N), r,vhile it gives tL:ivial algorithrns if we ash only f.or length optimal soiuiions (as in CLIQUE for in.stance). We irave Lhe fol.lowing counterpart of llheore.m.4.2g. 21..15 Theorem. Leb > 0, Iet S'!, F (Vr)(ly)(y, < t)p(*,U,2), for g in E!*, anill' a term. Then, for a given tr we can compute y using al+, cornputations wit.ir (an unbounded nr-rmberof) counterexamples. T'his theorem can be proved using the proof theoretical method which liluss used fbr h.iswitnessing theorem. We are not going to prove it here. The cor-rnterpartof Lemrna 4.41 is the following. 4.,16 Corollary. llbr i > 1, if SL : Tj, then the optimai solutions for ev-

.,ri.rgnf-r, computations with (an un?ty nlq problemcan be computed bor-rnded numberof) counterexamples

tharS$# r;.

We conjecture that the conclusion of Corollary 4.46 is false, hence also


t

The research o.n witnessing fr-r.nctionsis still going on and we can look for new nice results. The task of proving that Bouncled Arithmeiic is not finitely axiomaLized with,ou,t assu,mptions on compleai,ty classesseemsstill too h.ard. What might be more accessibJ.e is to prove nonconservation results for .93 and Zj using such assumptions.

5. Interpretabil'ity and Consistencu


(a) Infinocnuction
In the last section we were rnainly interested in the strength of theories obtained by restricting the cluantifier complexity of bounded formulae in Lhe schemeof induction. The main theme of this section is investigation of the strength of theories obtained from I Es by adding functions of a different growih raLe.(Thesefunctions have graphsdefinableby fo formulae, so Lhe strengLh is increasedby assuming that they are toial.) We sirall

5 . I n t e r p r . e t a b i l i L ya n d C o n s i s t e n c y

J61

con.sider typical cluestions of rnetamatherntrtics: interpreLations,provabiiity of consistency, conservativity, etc. tr'irst .',ve shall consiclerthe definition of trubh for X6 forrnulae in Bouncleci Arithmetic. As a consequence \,veobtain Lhat lXs * Exp is finitely axiom._ al,izable.Another consequence (which usesfurther tecLrnical resulis) is the existence of an ir:terpretation of I Eo + Q, in e, n : I,2, . .., whicir in tr_rrn impLies thal Q and these theories are equiconsistenb in /fs * A t For ftrrther resuLlts we need a formalization of a weakerform of the cut-elimination tlreorem in I Eg * Eap. The usual form of this blleorem is .unprovable in this bheory.On the other hand Godel's in.completeness theorem can be sbrength-l Enp does not prorrethe ened further for such fragments.For instance 1'Eg consistencyof Q. This and severalother results of this kind will be shown in subsection (f). Finally we shall consider the quesbionof limitecl use of exponenbiation in /X6.

(b) Truth Definitions for Bounded Forrnulae


In Chap. I, Sect. 1 we have shown thai trubh for X9 formulae can be clefinecl by..41 formtrla in 1I1. Becauseof cliagonalization there is no such tr'6 definition. llowever we can omit the main I quarrtifier in the X1 formula and thus we obtain a Xg formula with an addition.al parameter.Then we estimate how large the parameter must be; put another way, 1veare looking for a suitable bounclto the I quantifier in the X1 definition of Xs truth. Sometim.es we sha1l call the X6 folmula obtained in this way a truth definition fbr )-6 formulae, though it is not cluite precise.The consbru.ction will be essentially the same as in Chap. I, only we have to compute the bounds more precisely. Before constructing a formula wiLh a parameter which definesthe truth for Ig formulae, we shall briefly consicler this problem frorn the poinb of view of complexity theory. Assume that there is a definition.of Xs truth described above; supposethtrt it has the cluantifiercomplexity k and the bouncl to the parameler is some function f. T'hen r,vehave bhe tbllowing relation beLween complexity classes X{ S flll 1."" Sect. 2 for clefiniLions). Thus the correspondingciuesLion is: fbr which function g and constant k do we have Xjv E I)ok? RecaLlthat Xfl are just the sets in the Linear Time Flierarchy(Theorem2.16),hence Lhebest that we can achieve(to our present
lcno.,vledge) is

xdvc l)*r;n'.(rn):E.
T'his resLrltcan be usecl to write a definition of l;rubhlbr tr'o formulae basecl on a X6 fbrmula with an exponenLial bouncl to the lengthof the parameter, 'we but again not only need a lbrmula, br-rtalsoproofs of its proper'liesin -f)-0. Furtherrnore we need an esbimate of the constant c. Therefole we presentan e,rpiicit cons brr-rct ion.

362

V. Boundecl Aribhrnetic

We shall Lisethe notal,ion introducecl in Sect. 3. Thr-is fbr insLance rnax(s) is bhe rnaxirnal element of seclttence s. In this subsection we use only the s,ranitarcl aritlrrnebical language -t9. T'he forrnalizabion of synLax is as clescribecl in. Sect. 3; thi-rsif r is the Godel nttmber of an expression, lrl is its length. Our approach is to consiclera botLndedformula literally as a program. The cornptttation of Lerms needs no e;cplanai;ion;Lhe inierpretzr,tion of boundecl ciual:tiiiers is as search.proceclures,Tlr-e crr-rcialbl:ing is Lo esLiinate the size of the numbers th.at ',,yi11 occur in su.cha computabion. Let r-rsconsicieran example firsi;. Leb cz(z) be a boirncled formula

pG),71, (ly < *2111,< y2l,(1u < 22) z , rt),


where B is also a bor-Lnded forrnula. To compu.te i;he bruth .zalue of a(o) we tr-sel,hree F0R Loops)olre fbr each bor-rndedvariable. hlote that the thircl loop lras the range 0 to r2u .5,1Froprosition(/to). (1) For a term I ancl a string of nurnbers s, f (t) < (max(s) -t-}vt . (2) The maxirnal value needed to comr.p'ute the truth of a bounded formula cv on a string s is bor-rnded from above L.y (max(s) -F 21zt"t. hlote in passing thar,l,, bV (1), the rrzrlueof.a cLoseiL term f is bouncled by 'tn 2ltl <2t,Itence it can be clefineclas a total function IEs, while Ibr general terms in -t6 'weneecl /Io -l-J71.Neverthelessthe estimates above are provable in 1'tr'sprovided thai; the upper botind exists. Proof, (l-) is easily proved by incltLctionon the lengLh of f.

(2) This is also proved by induction on bhelengbhof a. We sha1lcompute only tlre induction step for a cltLantifier. Let a(s) be (1U <t(o))p@,s) (for V 'bouncl it is the same).Assume we have (ma.r(s, l) -l2)''u' , to r B (y,s). B y (1 ) w e know that f (o) < (max(s)-F }lU ,
hcnce for a(s) we gel,

-r z1''o' (marc(s, (ma.r(s) -F2)lil ) -vzlztot ( ( ( m a x ( s-)F2 ) l r l ) -F ((nrar'(s ) 2)l'l+1)'to .t

5. Interpretabiliby ancl Corlsistency

JOJ

We shall esbimate the e;rponent:

(lrl -r D2lpl< 2ltl-rl-Flal < 2lol .,


which proves bl-re i::clr,LcLion sl;ep.

FirsL we define the graph of Lhe aul'f'unction '',shich defines the 'yalue of arithm.ebicalterms. A.lsohere we n.eeilan adclitional pararneter:.The meaning of. Valu'e(a,ttz,r.l) is: y is the valr-re of term r computecl on string z, y is tlre parameter. Leb Var(r,,i) be a formalizabion of the conce'pbof bhe i-th variabie. 5.2 Froposition, T'here e;rists a 16 formtra vulu,e(y,*, z,u) and a polynomial p such that it is provable in /Xs that:

(1)

T ' e r r n ( x8 1z )T e r m ( n 2 )e 8 l * t l , , l * z lS , & l " l > p ( l m a x ( z ) lr,) Value(y,trtztu): (z); -- y) k (Value(U , 6 u, z , u ) - - +y : 0 ) k ( V a I u e ( A t , 5 ' ( x t ) , z , u ) k V a l u e ( y 2 t n 1z t, t t ) - - + n : S'(yz)) 8 t ( V a l u , e ( y , r y l - .n 2 , z , u , )k V a l u , e ( y 1 , , n , t) z1 ,z k Value(y2tn2z t ,u,) -, A : lJr* az) k (Vulue(U z 1u) , r 7 ' t ur ? , z , u ) k V a I u , e ( yj I 1 7 1 k Valu,e(y2tn2 zt, u , ) - - +y : A ' r , r A z ) ; -+ (Var(rr,f).*

(2)

Term(r)& 1) , l " l < uk l u l l , l u z> l p ( lm a x ( z )u


8c Value(y t r t .z t u,1)k Value(y2,t, z, u,2)--+yr : Az .

'Valu,e('y,n,z,u) Proof. i s d e f i n e da s ( l s 1 u ) t p ( y , x , z , s ) , r , v h e r e g(y,r,z,s) is a Xg formula expressing the following conclitions: (1) s is a sequence o f p a i r s ( o o , b o ) ,. . . , ( n r r , b r n ) ; (2) for every i 1m, either a; is a variable, or: ai is a;-F"cr,1r, where 0 < j, h { i, etc. lbr S,0 ancl ,1.; ( 3 ) i l a ; i s t h e T - t h - z a r i a b l et,h e n b ; : ( z ) j ; i f o ; i s a i - l - ' a 6 , l , h e nb ; : b j - F b r ; ebc. for S,0 and 'r.; (4) A:brnandr:cLrn. conditions e,rists,ti:.en it determines y unicluely Clearl.y,if s saLisfying Lh.ese and the conclitions of the proposition a e pr:ovable.(irlamely, using induction, one carnprove that every secluence of pails for a su.bterm of tl is a subsequence of some secttleirce of pairs for /.) Th.r-rs we only have to fi.nd an upper bound to the size of s. Assume s is given. By Proposil,ion 1 (1),

b; ( (rna.r (') -t-Z1@;l

364

V. BoundedArithmebic

Sincethe pa,iringfunction is a polynomial, (o,i,b;)is bounciecl by a poiynomitr.l ( r ) i n u a n d (ma .x(z) < ( q,b;) ( max( z)+ Z;e for sor ne *2 ), hence pol y nom i al q .B y P ro p o si ti o n 3 .30,lul S r ( l( r r ' ax( z)+2) c@l,u) for somepoiy nom i alr , wlrich givesthe req'irecl bou'd lrl S p(lrnax(z)l,u), p a polynomial, The cletailscan be workeclout in a similar way as we clid in Sect. 3. n
5.3 Corollary" The value of an -Ls berm is clefinable by a E["p formula (in faci by a boundecl for:mula in .L2). Proof. Defi.ne ua l ( r , z ) : A : d , yV a l n e ( y ,* , z , 2 P ( lr n a x ( z )l|t, l ); .

tl

Tlre meaningof .T(r,z,u) in the following theorem is: the xs formula r i s satisfiedby a string of numb e'rs z and u is somebound. 5.4 Theorem. Th.ereexistsa Xg formula J-(c, z,,u) and a constant c such that it is provablein /Xs that: =' s -+ (I(z,z,u): ) 2 ) " "& l " l S u . - r ( r : t l " l ) ( m a x ( z+ ( : y X V a l u e ( y , tz ,,u,) k V a l u e ( ys, , 2 , " ) ) ) ) & k(x: (rr -" *z) n (.1(r,z,u): Q@t,z,u) -- f (x2,z,u))))k k ( r : - - o n I- , ( l ( * , 2 , u , ) : - l - ( r t , 2 , u , ) )8 )z & ( t : ( l n t u ( ' f ) r t - + ( . 1 ( r ,z , , u ): ( ) y , a ' ) ( V a l u e ( y,t, t ,z , t L ) ky S a' k f ( r 1, z' ,,r t) ) ) ) ,, where we assumethat r.uis the 7-th variable for sorre j and secluence z' is obtained by replacingits j-th entry by y; the condibions (Tarslci'sconclitions) for (o ancl bounded V' are similar and thereforeomitted. Proof. The proof is sirrrilarto the one above.lVe define 1-(r,z,u) as (lu < zt)t!Q), where r/(s) saysthat s is a secluence consistingof pairs of somesubstitution instancesof subformulaeand their truth values satisfying Tarski's co n d i L i o n s. L e t s b e the sequence of pair s ( oo,bo) ,..., ( ctr r ,bn.). W e s hal l rec1uirethefo11owingcondition.Leta;beoft1reform(l,r.u< (Vtu e 0 ,.,.t& m. H e reD denotes the tlyadicnumer alintloducecl in Sect.:( g) . ( tt tlre valr-te of (lutu 1" t)p(ru) turns out to be TRUE, then ii suffi.ces to injust one ,p(n), but ihis is not a real aclvantage.) clttcle BV Proposition 1(2), the valuesneededbocompute the t:r'utth value of a formula r are bounded bv

(rnax(z) + 2)''''

5. Interplebability ancl Consisiency

365

Thus the a;'s are Godel numbers of subfornrulzre of r r,vith some variabies replacecl by nr-LmeralsD for

n ( ( m a x ( r )- r z ) r ' " t. The length of suchformulaeis estima.ted by c 1, r l.( m a x ( r )- r 2 ) r ' ' 'l r . l " l , wherec1 is a consbanb, sinceth.elength of a numeralis linear in ibs value. The lengthof (a,;, b;) is thus c 2+ l ( m a x ( "+ ) 2 ) r ' " 'l , *l " l ,
for some constant c2. Now we estimate the number rn of.such pairs. We have at most lrl variables in r, and each variable can be replaced by a numeral D with the bound above for n. This gives ((max(z) * z\zt't * r)l'l possibilities and there are ab most lrl subformulae, hence

m 1 lrl 'r. ((max( z) + 21zt"t * t)l'l .


Thus we can estimate the length of s by

(r) + 2)''"'l * lrl2,r ((rnax(z ca,r l(max + r;l*l ( (max(z)+ 2)"''' , ) -r z1zt'r
where ca and c are suitable constants. Hence if u is larger than or eclual to this bound, then there exists some sequence s that witnesses the truth of e. The provability of Tarski's conditions is shown as a'bove. n

F o r a s t ri n g o f va ri a b l e s 2 r,.. . ,,znwe clenote by ( q,.. . , z n \) tr ln e c o oI e o tr : the seqtrerrce zrt. . . t zn. 5.5 Corollary. (1) For every'boundedformula d(2t,. , . , z.n)in trg Lhereexists a constant k such that -o n ( * ( r t " , , . . , 2 n ) = . 1 - ( a(,r t , . . . , r n ) , u ) ) . I Do l-., > 2(*t*(2t,"',zn)*2) (2) Let IUI be a model of IEs,,Iel, a,b,d e IvI,Iet a(z) be a Xo formr-rla. Sr-rppose d is nonstanclarcl ancl [r s 2kt-12)'t. Then I U It sa ( a ) : T ( a , ( " ) , , b ) . Proof. (1) FirsL apply Proposition5.2 to prove the statementfor a atomic using inducbion on the cornplexity of Lerrnsirr it. .fhen tbr generzr,l d use in.iluctionon the logical complexiby of a anclTheorem 5.4. n (2) This is an immeiliate consecluence of (1).

366

V. Bounded ArithmeLic

5.6 Theorern. .IXs * Erp is finitely axiornatizable. Proof . l,eL T' be a finite fragment 1X6 such that ? proves Tarsiri's concliLio.ns of Theorem 5.4 for -f. Then for every a bouncled, 7' proves the formula of Corollary 5.5 (1). LeL tp be clefined by p, u) : | (r, (y, p), LL), V(n.,ll,, (Here we are misusing notation: bo be c1-riLe precise, we should distinguish between zt" pai,r and a i'wo element sequence,bat we do not r,vantbo inbrodrice nevr nobal,ion.)Let Tt heT plus Lhe leasi nr-rmberprinciple for V(rc,y,,p,r:), wlrere y is the induction variable and z, p and u are parameters, plus Erp. Let 'bounclecl a(y,p) be an arbitrary for:mula for which we want bo prove the least n.umber principle (clearly it suffices to have just one parameter). We shall algue inTt. Let p be given and sr-rppose that for some y0 we have a(ys,p). Since we have Erp. we can take u t 2!nax(vo,R))o ,

where fu corresponds to cv. Flence,f.or y 1yo,

a ( a ,p ) : f ( o , ( y ,p ) , r ) : V ( * , a ,p , ? r ) . Thus we can tahe the least y for g(cv, U,p,Lt).

(c) An Interpnetation of J'56 i* Q


Q is a very q/eak theory, e.g. the associabivity of -F is not provable in ib. Eloweverfi:om the point of view of interpretabiliby it is cluite strong. We sha1l show that J'Xg * 9n is interpretable in Q for every n. Moreover th.e form of inberpretation is very naturzr,l:we take the same operations and only restrict the dornain. 'fhe domain of interpretation is, roughly speaking, an initial segment of the original numbers. This suggests an attractive finitist's program, which was pursuecl by Edward Nelson fNelson 86]. Namely, the consistency of Q seems qtr.iteevident; further it is also evident that interpretation preserves consistency, bhus if we develop mathernatics only in theories inLerpretable in against inconsistencies. Later on we sha1l see thab Q, we are saf'e-gtLarcled rve can even interpret in Q some consequences of I Eo -l Erp which are not available in -IXs -F fln (Theorem 5.27 (i)).1
1 L"t us rernark, for those .,vhoare interested in minimal founda.l,ions of mathemalics, thal one can use an extremely weak systern fol set theory inslead of Q, since Q is intelpletable in ii. The sysLem has only two simple axioms:

(3c)(Vy)(-y6 e) ; (Vc)(Vs)(:z)(V")(u e z =(u o v r' : u)) .

5. InLerpretabiliLy anclConsistency

J6,r

The main reason why we consider interprebation of Boun{ecl AriLhrnetic in Q is, however, different. The interprelLability of systems of Bounclecl A.rithmetic gives tts er1'Lticonsistency of them 'with Q, ancl this ecluiconsistelcy is provable in Bor-rndecl A.::iLJ:metic too. FIence Goclel theorem implies bhz-r.L sttch a sysLern cloes not ar,lready prove consisterrcy o,t q. T'he rest of Lhe sttbsectiorr is alinosb enbirely devotecl.Lo the pr:oof of th.e follor'ving theorem. In the .rhoJ.esubsection we shall rrse only inberpreba[ions given by restricting the clomain. Such an inl,erpretation is cleterminecl by a formula r,vith one free variab]e. 5.7 Theorem. For: every n, bhere exists a global inLerpretation of ILg | ,en in 8.

By o global inierpretation we mean the usual concepLof interpretabion. The word globalis useclhere to stressthe fact that we have one translation of the language (but in fact only the domain will be different) sr-rchthat the translations of each one of the infinitely many axioms are provable. On the othe:: hand, local interpretation means that we can choose a different translabionfor every finite subsetof axioms.hlelsonuseslocal interprebations in his book. lVe shall use a local interpretation as an inbermediatestep. lVe shall say that a formr-rla /(r) is inducti,ue in a theory T' if

7 F /(0)&(vr)(/(,) * /(s(r))) .
Recail that / is a cut in 7 if it satisfies moreover I(rc) ka I x -, t(a) , prouably in 7. We shall say thab /(z) is a su,bctt't of /(r) in 7 if moreover

7: ts J(x) -+ .t(r) .
Let Q+ b" Q atrgmented with the following axioms:

(associat'iai,ty of a) (Ieft d'istribu tiuity) (associutiuiiy of .r)

( r - l a )* z : t * ( a* r ) ; * Q + z ) : r U* n z ; @ u ) ": r ( a z ) .

5.8 Lernrna" There exists an inductive formula /(a) which determines an inLerpreLation of Q-F ;" Q. We shall not prove this lemrna.It can be proved using similar triclcsas'we shall irsebelow, but iL is technicallymore complicatecl, sl,art wibh a sincer,ve very weak l,heoryQ. A conrpleteproof can be fbuncl in lNelson86l.

368

V. Bounded Arithrnebic

5.9 Lemryra. Let .I(c) be an incluctiveformulain Q. Then bhere exisbsa sr-rbcut J(x) o f I(r) h Q. Proo!. We sha[]. ttsethe following three easy theoremsof Q: (i) r(0, f i i ) n 1 u : 5 ' ( e )< ^ 9 ( u ) .
\ v / ,

(iii) 0 < r.
LeL I be inductive in Q. Let, I( be clefined by K(t) :/(t) k(Ya,r)(t<aky (r-> z.-r).

We shall show that r'( is inch-rcbive. /f(0=) follo,,vseasily from (i) andfrom 1(0). Suppose Ii(r). Then, clearly, /(S'(z)) and it remains to prove the seconclpalL o f t l r e c o n c l i t i o n f o rK ( . 9 ( r ) ) . L e t z I y < S ( " ) . I t z : 0 , t h e n , b y ( i i i ) , z 1 x . Otlrerwise, by (Q3), , : S("'), for some z/. Then, by (i), (Q5) and (Q1), y cannot be 0. Thr-rs A : ,9(yt), for sorneg'. Thus we get z' 1y' ( ir, using (ii). By /((c), we have z' I r. Using (ii) once again, we get , : S(r') < S("). Hence /( is inductive. Now define J(r) : (Vy < r)Ii(y) . First we shall show that .I is inductive. /(0) follows easily from (i). Assume ./(r) anct let y ( S(r). We neecl to show that /f(y). If y : 0, then I{(y), since y'( is inductive. Suppose that g : S(y'), for some y/. Then, by (ii), we gel At ( r, hence I((yt). Since K is inductive, K(S(A'D, which is /f (y). ThLrs K is indr-rctive.Now we shall orove that

J(r)8tuSn-J(y). Strppose J(r)ky l x,Let z ( 9 be arbitrary. We needLoproveI{(z).By


J(r), we have in particr-rlarI{(n), hence z 1r. Thus we have K("), since

J(").

5".10X,ernrna.Let /(r) be an inducti'rre formula in qr-. Then there exists a cut "/(r) such that J is a subcut of -I closed under 5, -l- anit +, i.. Qr proves

t(s(.)) kr(,.,y) kr(.y) il;.1,;:,';:'Proof. By Lemma 5.9 we may assume that 1 is a cut. We shall use the same construction as in the proof of Theorem 3.5, Chap. III. Take

-, J'(y /o(") = (Vy)(r('u) + ")) ;


J o ( y )+ / s ( y c ) ) . J(x) = (VsX

5. Interpretability ancl Consisbetrcy

369

Then -Is conbainsT, since / is inducLive; ,16is closecluncler by associativiby * of -l-; .I is conta,ined in ,,/6 (since ,/s contains T) ancl hence also in 1; J is closed trncler"J- by left distributivity ancl since./6 is closecluncler -l; J is closeclunder 5', since it contains T ancl it is closeclun.cler -l; J is c.loseduLnc].er 'r by associabirrity of 'r.. To sho,,v that J is closecl clownr,varcls, we use t,he er.ssrurtpl,ion Lher,l, / is tr cuL, le.fl,disLribr-rtivibyzr,nil,.I9being closeclrrncler -F.

5.11 tremrna./f6 is locaily inLerpretable in e. . P r o o fB . y l , e m m a5 . 8r v ec a nt a h e8 - F i n s t e a d o f e . L e t , g 7( r , p ) , . . . , , p n ( r , p ) be given bounded formul.ae.We want to interpreb Q plus inclr-rction for p t ( n , p ) , . .. , p n ( x , p )i n Q + . L e t

I j ( r , p ) = p t ( 0 )k ( V y < 4 ( v i ( a , p ) * , p i T f u ) , p ) ) - , ej @ , p ),
fbrT:1,...,n.Let

I(") : (VpX/r(r,p) & ... k l"(r,p)) .


Since each /i(r, tr) is inductive with respect to z in O*, so is /(r). By Lemrna 5.10 we can find a stibcut J(z) closeclunder * and *. Since J is ciosecluncler S, -F and 'r and it is closed downwar:ds, we have, for: any borrnclecl formuia *(zt',. . . , zk), Q - Ft - J ( " i A . . . Since J(r)--+ I(r) -, Ii@) provably in Q+, for e'ueryj, J determines an interpletation of incluction for every t1:i. AII buL two axioms of Q (even Q+) are open, hence they hold in the inLerpretation. The remaining two axiorns are (Q3), #6 -+ (3y)(r : 5'(y)) " and (QB): the definition of the relation (. The first one is provable using incluction for bouncleclformulae from the others (hint: first prove r 15(r), then r *0 -, (!y S r)(z - S(y))). Hence we can simply suppose that we lrave enough formulae amonEstet,...,en to prove ib. The last one causes no problems, since we can trivially interprel, Q in the theory obLained fi'om n Q bV eliminating ( from the language and deleting (QB) Proof of Theorem 5.'7.By'I'heorern 3.5, Chap. III, we know that each cut in a sufficiently strong bheory can be shortened to a cut closed under un,for a given n. One czr.n easily checlcthat IXg is sLrong enough for that. So, by bhe lasL lemma, lve only have to interpreb 1tr'6 in a finite fragment ? of J'Xs. lVe tzrlte ? so stronp. that kJ(zp) --+ (a(21,...,2n): (*(rt,...,rk)).r).

370

V. Bounded Arithrnetic

(1) ihe properties of the exponenLiation relation are provable in it, (2) Tarslci's conclitions for the formula f' of T'heorem 5.4-are provable in i{,, ancl (3) the leasL ntrmber principle for l(c, (y.,p),r.r),with y as bhe incluction parameber-, is provable in iL. By the sarrle theorem of Chap. III on shortening ctLts we get a cub /(z) sLrch. that

rrI@)*Gil@-22').
By Lernrna 5.10, we can assumethat / is closedunder * ancl *.'V/e claim thaL / determines an interpretation of I Dg. It suffices to prove the ieasL numbel principle for every bounclecl a(y,p) ''rzith just one pararreber. Let p be an element in / and supposethat for some aoin.I we have a(yo,p). Take
o, _ 9(ma;<(yo,p))u

since where k corresponds to a. This is possible,


/ / \\[ 2\ma:<\U0,2))I

2"

,max(/o,P)+'t

and the lasbnumber e;ristsby a property of -[. Now we conclude as in Theorern 5.6: for y ( A 0 t

"t), and we can take the leasl y for J-(a, (,A,p),,u).

a ( a ,p ) : T ( * , ( a ,p ) ,

ll

Tlre concept of consistency Con used in the following theorem is the usual one as used in Chap. III, except that we must use bhe effi.cient coding of sequencesas described in Sect. 3 of this section. 5.12 Theorem. For every n)

. IEo*At F Con(IEo+nX): Con(Q")


-/

o,f I Eg -F O, irnplies bhe consistencyof Q trivially. . The consistency P,roof of I Eo * An can be reducedto the consistencyof Q using T'he consistency an interpretation. We shall analyze this proof, in order to see that it can in IEs * A, be given' be done in IEs * At Let a proof of a conbraclicbion abovewe can transform it inLo a proof Using the interpretation constructect i" Q.Again, it is clear that the interpretation as a relation of a contrardiction by a fg forrnula. Hence we only have to check th.at ihe can be clescribed proof is polynomial in the original one. This can be size of the i;ranslbrmed proveclby showingthat eachapplication of a logical rule in the original proof in the new proof. There is replaced by only a polynomially longer secjLrence

5. Interplebability and Consistency

3'7 |

is only one place w.here \,ve have to modify Llr.einterprel,atioir given. above. When we prove the leasbnr-rmberprinciple tbr: a boundecl formula a, we define parameter u by
u:2(^o*(yo,p))" .

on cl, since a is nol; [-irer]. Now we have to take into accounL ho.v k clepencls Thus insteacl of the double exponential frinction in the definition of f(r.),.,ve take triple e,'(ponential. Flence

Q r I ( r ) - , ( l y ) (\ r A l k a : 2 " ' \ / ,
where cr-Lt.I(y) determines an in.terpretation of a sufficiently strong finibe fragment T o,f.IL'g in Q. Then we replace number k in the defi.nition of u by a more explicit bouncl given by Theorem 5.4: u : 2 ( ^ u * ( y o , p ) ) " 1 ' l. where c is the constant indepen.dent of a from Theorem 5.4. No'w, using the estimate clal o'n)* "' 2(max(a0,il)"' S rzzmax(v one can show thab u exists and "f(u) holds true, and the proof's of bhesetacbs in Q arc of polynomial length in the length of a. llence the induction axioms n are replaced also by polynomially long secluences.

(cl) Cut-Elirninatiom.ancl Herl:rand's'Xheorern in Bounded Arithrnetic


There are two basic measures of the complexity of proofs: the length of proofs and the cluantifi.er cornplexity of proofs. Cui-eliminabion and Herbrand's theorem implies that each first order taubology a has a proof whose cluantifier complexiLy is not larger than the cluanLifiercornplexity of a. The cost for lhe reciuction of the cluantifier complexity is an increase in the size of Lhe proof. T'his is the reason why this theorem is not provable even in /Xg { Eap in ful.l generati|;y. Never:thelessth.ere is a restricted version of this result which is pro.zable in lXo + Erp and which has several applicaiions. We shall prove it in this subsection. The proof will be just an analysis of a classical proof of cub-elimination. T'here is an alternative proof based on Flilberi, style caictLli, bub we shall use a variant of Gen.tzen'sapproach since it is simpler and more transparent. The transformations used in the pr:oof will again be some simple hence clefinable by Io formttlae. Thus we only manipulations with secluences, have to estimate the size of constructed proof's. lVe shali 1rse a slighLly simplifiecl version of Schr,vichtenberg'ssysl,em The proofs will [Schurichtenberg 77] and follow his proof of cut-elimination.

372

V. Bounclecl Aritb'mebic

of pr:oof lcnowleclge we do nol, presuppose ho.arever briefly slcebchecl; be on1.y consulLiTakeuti B0]' It shodd be For more clelaiisthe reaclershotLlcl l,5eory. fi:om tJlis systernLothe systernintroclucecl here that the branslations stressecl a].ecFriteelementary. In parbicular (assuming efin Chap. 0 ancl baclclvard"s in Sect.3), it is possillle bo prove iir ficient cocling,which was introclr-rcecl is proval:lein one systemiff its translation is prov/Xo -l-CI1thaL a sen'cence one. ahle in the oLhe.r
)l<

lVe consicler a system for first orcLerlogic. Ibs langr-iageconsists of logical &, V, -, cluantifiers V, l, va::iables,constants, function symbols conn-ecbives ancl relation symbols. \,v" clo noL distinguish equality as a special (or logical) rela,lion symbol. In orcler to recluce the number of ru1es, and consequently th.e number of cases to be considerecl in the proofs, we allolv negation on15' rg is at atomic fbrmulae. Thus, for a complex formttla g, bhe expression an abbreviation for the equivalent folmtlla obbaineclby applying De ivlorgan rules. For a similar reason we shall treat disjunction V clifferently too' We shall think of the formulae in a disjunction as forming a set rathel bhan In the aribhrnetization, all syntactical objecbs are represented as a secluence. as binary expansion's)'Thus what we ,".1t"rr"", (which in turn are represen.ted *"on is that: in proofs we allow |h"ereplacementof a di,sjunction by anoth'er one which, has the lan-Leelements, For reasons of symrnetry we shall consicler conjunction & to be also a set operation. when we write, for instance, that a formr-rla g has the forrn 'f Y ,h, we mean that g is the disjunction of a set . . .',n/mr'b t r " ' r ' l ' n a r e a t o r r r i cf o r m u l a e ' { l t , , . . . r Y m ,r b t r . . . r r ! r } , w h e r e 7 7 r conjunctions or formulae stai:ting with a cluantifi-er,and we do not assume t l r a i t l r e s e t s { 7 1, . . . , . f m } , , l r h t , , . . . , r h n } a r e d i s j o i n t a n c l7 i s t h e d i s j u n c t i o n o f ? r , . , . , , ^ l r n , o r Y - 1 1 i f . I r - t : I , a n d l : i s t h e d i s j u n c t i o n o l t f t 1 , " ' r t f ; n ,o r

,h:,h

if'n - t'

(usually called proof lr,nes,or steps) of formr-rlae A. proof is a secluence formula is either a logical axiom or follows from preceding ones eaclr. r,vhere by a rttle. There is one a,riom schemain oul system: (A) 'yv VV -(Pt .yYtp,lV rh

ancl fi.ve lules:

(w) (v)
rr\

&^',h) I v (,p 1 Ya ''l v tp (r) r is not fu'eein 7, where the'va.riz-r.ble (Y y)tp (y ) 1v 1v p(r)
7 v (lr)p(r)
where r- is a telm free for x in 9,,

(&)

'

5. Int,erpretability and Consist,ency

3'(3

g, 6Y -g (Cut) l V 7v6
The elemenLs of bhe disjunc{,ions .y ancl 6 are callecl side lormttlae;'y ancl 6 -.rp are ctt't may consist of one formula or can even be voicl. In Ctrt, rp arncl an important rtLle, th,e con'tract'ionrule., ts formu,lae. Let tts point ouL thzr,b only impliciL in our system. A typical. application of it, which. is also needed -.Flerbrand's theorem, is the fbllo.,ving:we clerive by (:) in the proof of

1v (lr)tp(r) v p(r)
7 v (1ll)tP(t) we do not repeat,(1x)tp(r) once more, sinceour clisjuncIn the lower secluent Rule (W) is clearly redundant (and in contrast with (Cui) also tions are se'|,s. harmless),but ib is convenientto have ib in the system. no[ trees(th.elatter being proofs to be sequences, The fact that we clefine .',vill be important later, when we shail consider common in proof theory), weakerthan /Xs * Exp. sysbems We shall follow the notation of Chap. IV, Sect.4 by writing Proof (c\,7) for tlre relation d is a proof of l.By writing 7(c) and then 7(r), we denote the of r in 7. We shall use two measu-res substitubionof r for all free occurrences of the complexity of proofs. The length, denotecllh,(d), is the length of the which codesthe proof d. Of course,lh is clefinedalsofor formttlae and sequence terrns. We assumethat the number of symbolsused in the coding sequences by slh(d),for sorne is finite, thus the Goclelnumber of a proof d is bounclecl for a formula tp, rank(g) as the number of constanbc. Further we c1efine, connectivesancl cluantifiersin g plus 1. The cu,t-rankof.a proof d, denoted by c-rank(d), is 0 if there is no cut in it; otherwiseit is the maximum of the ranks of the cut .formulae. Proof(d,lV(Va)rb(y)); bhenthere exists 5.13Lemma (/Xs +Exp). (i) Suppose ,r/(")) (d',"t v and c'rank(rl') 1c-rank(d),, lh,(dt)< lh(d). r// such that Proof ( i i ) I f P r o o f(d (n ),r(r)), th e n for any ter m r , Pr oof( d( r ) ,r ( ") ) and cr a n l ; ( d ( r ) ) : c-ra n k(d (" )),th (d (r)) < Ih( d( x) ) ' r th( r ) . (iii) Suppose Proof(d,,p0&rp.t), then there exist do,dt such bhab < lh( d) , for i : 0,1. a n d c-ra n k(d ;)S c-rank( d) ,lh( d1) P r o o f( d 6 , t p ;) of variables.) (We iake r and r in (i) and (ii) so that there are no clashes Intuibively, the statements are clearl formally they can be provecl by induction on the length of the proof. The proof'sare easy, bhereforewe omib th.em. Reductlon Lernrna(/to + E"p).Suppose d is a proof of @ such Lhat 5..1-4 thab the part o[ the cp.Sr-rppose a cut fbrmultr, the last interenceis a cut .',viLh has cut-rank less lhan rart'lc(g).Then there proof before the last inf'erence a,nclllt'(dt) < llt'(d)ro,f. L,ristsr//, a proo,f. O, stich that c-ranl;(clt)< c-ra'nlc(d)

Al

tl

V. Bouncled AliLhmebic

.l,et, the last inference be .Proof ' Y Yt P , 6 Y - ' t P

7v6
there is no F,irst slipliose bhat g: is atomic. Then ranlc(tp): .1,hence .1.. 'Y V I of ranlt 0. other cut in the proof. The proof r/ contains a proof dg of side formr-tla in Consicler how cp enters the proof c/6. If ib is by (1V) or as a of bhe forrn (A) by (A), then we can omit it and cleriveTv 6 by (vv). If it is *,? from 6V -tP by weakening (W) ancl ii Vy -V, then we first clerive o V 6V -p. We remove g also trom then use tl:is proof instead of the proof line a Y VY -,tp and extencl them by forming disjunctions the proof lines below o Y p Y of.7 Y 6 ' wiih 6. Tlrus we transform the pr:oof of 7 V tp inbo a proof cll Our sbrategy is the same as above' bub now 2. Suppose tp is (1t){(t). rule (l). (Note thTt' I can be introduced in V c^n bl-in.trodr."d. u.iro'by the conbraction') cl several Limes bv (:); see the example above describing Consider such a derivation in d:

o v $(r) i rr (3r)t1.,@)'
just o.In d we have a proof where it is possible that the lower clisjunciion is gv Lemma 5 (i) and (ii) we get a proof of 6V -p, which is 6V (Vr)(-'l(r)). 13 get oV6' This o f 6 V - , , $ ( r ) . N o w w e u p p l y c u t t o o V $ ( r ) a n c l5 V - , r / ( " ) t o 'yY 6 these changes, we get a proof dt of ctrt has smaller ranl(, L"tr.", afber a,11 of a smaller rank. tpg8t91, then we do the same procedure as in 2 3. If g is a conjr-rncbion e:<ceptthat we now use Lemma 5.13 (iii)' +. ny cluality there are no oLher cases.If, for insl,ance, tp is a clisjr-rnction' then -? is a conjuncbion. d/' This rvill 5. It remains to compr.rte the size of the transformecl proof in I Eg ! fixp ' also be a proof that the above consiclerationscan be carriecl out lines of d lve ifr" proof cil is constructecl in stLcha way that for certain proof -"h?), -trp o-.t or a ploof of 6 V 9; or jurst a proof of 5 V acld a proof of 6 V in the extend th.e disjunction by 6 and remove 9oftom it. The largest i.ncrease --tl.,Q); l,here we have the lollot'ving size occurs when we neecl the proof of 5 V upper bound from Lemma 5.13:

< th(cD2 . lh(d)'t'tt'r'(r)


bhe boundis As there are at most th(ct) piaces where r,yeclo such e:ttensions,

ilralT.

) 0' then,ll,:t" e"risi;s and c-rank(d) (ct,'y) (/tr'6l- Erp).II Proof 5.1-5.l,ernrna ' <222'Ih('t) ancl lh(clt) <c-ranlt'(r/) r//suclr thal Proolkl',l), c-ranlc(cLt')

-l I

I
5. Interpretability and Consisbency

ranlc is of proof cl down to ihe first cttt r,vhose Proof. Talcebheinibial segmelr.b by a proof whoseranlt is lessthan Replace bhissr-Lbproof eclrlalf,o c-runk,(c/). using liem.ma 5.i4. Repeat this procedureuntil .no cut with rank c-ra'nk(cl) with a proof of length lh(d); then in ecpralto c-rank(r/)remains.We sLarL that ther:eare at mosLlh(d) cLr.Ls sbepLhelengLhis raised to the 3, ernc[ eaclr. have to eliminaLe.This eives r,vc a r22*t1'(d) th@t) < th(rt)T'^u) L

Now we shall defi.ne seye::al concepts of provability and consistency. Th.e basic concepts have al.ready been defi.nedin Chap. III; we repeat their de'finition for the reader's convenience. 5..[6 Definition. (0) d is a proof of r i,n T

Proof 7(d,r) =af (1, 5 dX(Vt< lh(s\:r((u);)& *" t)) . (d,(s)s8z'. . . 8z*(o)rn(")-r k Proof
in the th,eory T' (1) Prouabili,ty P ry@) =aS ( 1d) Pr oof y( d,r ) . T ( 2 ) R e s t r i cte d p ro u a b i l i ty i n 'th e tl teor y L?.PryQc, r) = af (1d)(c-ranh(d) ! k k Proofr@, *)) . (3) Cui-freeprouabilityin tlte-theory T CFPrT@) : ay RPrT (0, t) .
(4) .t[erbrane],prouubili;ty in 'th,e'th.eoryT

IIPrT(c)

:ay

"there exists a propositjonal proof of a disjunction of instances of the open part of

finite S'C T" , ffe(fi,9 * r), for some


(see 0.18 for the defi.ni{,ion of the {unction }Je).

3'76

V. Boundecl AriihrneLic

of the theorglt (5) Consistency Cony =df - Pr ' i' 1d:T; . of tlt'etheory'f consis'tency (6) Restric'ted - - 1) . RConT( k) =at - RPr ' r ( |,:,- 0 of th'etlteory T ('7) Cu,t-free consist'ency CFConT ?,tf -CFPr7@ -- t7 . of the (B) Ilerbrund consist,ency 'tlr'eory T .

HConT = aI - HPr t1$ J;

it means prouabilitlJor consirtencyin pure (If we omit the sr-rbscript, logic.) The theoriesthat we shall usewill be either finite or have naturally defined polynomial time computablesetsof axioms.For such theoriesT t'hepredicate a,rpi" an axiom of T" can be expressecl by a E.l formula. Then Proof yQl,r) - A n is also a )Y tormula. an axiom sayingthat the Xs definablefunction2fi is Denote iy Superenp total (seeDefinition III.3.3). 5.17 Theorem F (Vr)(Pr'(t) * CFPr'(x))' (i) IEo -l-Supererp (i i ) F o r e ve ry k, IEs * Exp F ( Vz) ( nPr ' ( E,t) - CFPr ' ( t) ) . Proof. Both statementsfollow immediately from Lernma 5.15. fl

Let us note that it is provablein IEo + Erp that eachpropositionaltautology has a cut-freeproof. This can be shown by formalizing a completeness proof for the propositionalfragment of our system consistingof the scherna (A) and rule (&). Let a propositionaltautology V be given. ConstrucLa proof into parts uni.e. we split conjuncbions o,fp by applying rule (&) back.rarcls, iil only clisjunclions of atomic ancl negated atomic fbrmttlae are leli (recall
bhat negations can occur only at atomic formulae). We rnust shorv that these means that they are tatrtologies" disjulctions are of the fbrm (A) which jr-r.st This is true because rule (&) is sound also in the opposite clirection. In each step of this construction we adcl two formulae of length at most lh(g) and tlre construcbion has at most th(Q steps. Thtts the total size of the proof is estimatecl lrom above by th($ a,2lh('p). Therefore, the construction can be carried out in IL-s I Erp.

5. Int,erprebability and Consistency

3'/7

by shor,virrg thab in "ltrs 1- .Flrpthe ctLt-fi.'ee We conclude this subsecbion llhen we clerive provability ancl the lferbrand provability are eqtLivaler:b. of Q. corollalies abotrt the provability of the consistency
5.18 Middle Sequent Theorem. l,eb p be a first oriler: tautology in prene:c ancl such tJ:.at normctl fornz. Th.en birereexists a proof cl of.tp which is cr-rt-free it can be clivided into two parts clt and r/// snch Lhat d/ is cluantifi.er1ree, while cltt contains only applications of the cluantifier ruies. idea of the proof is sim.ple. First one shows that a,-ciomschema (A) and rule (W) can be restricted so that the newly introduced formttlae mtLsLbe This increasesthe length of proofs only polynornially and no cluanbifi.er-free. Now talie such a cut-free proof of p and switch cluantifier ner,vcuts are need.ed. erndnon-ciuantifier rules to obtain the requirecl form of a proof. The only thing that we have to take care of are possible collisions of variables in rttle (V). Therefore 'we must fi.rst transform the proof into a regu,lar prool , which is a proof in r,vhich each application of rule (V) has its own valiable that is generalized (eigenaari,al:le).To do such a renaming we have to transform the proof into tree form, which means that each prernise is used only once. This may cause at most an exponential increase. Hence it is prorrable in I Eg 1- Exp that a lbrmula cp has stich a proof whenever it has a cut-free proof. Suppose such a proof d of.tp is given. I-,et dt be ihe quantifier-free initial segment of the proof, Iet gl be its last forrnttla. We may also suppose thal, each.formula, e,rcept the last one, is a premise of some application of a rule in c/. Hence tp is obtained from pt by successiveapplications of quantifier rtrles. In this situation we call cpt the middle formu,la of d. (Usually in proof theory sequenbsare used instead of ou.r use of disjuncLion of formulael there{bre the theorem is referred to as the Middle Sequent Theorem.) W" shall show that gt can be easily translbrmed into a tautological Herbrand disjuncbion l{e(tp). Let g be of the forrn ( : r 1 ) ( V y 1 ) . . . ( : r " ) ( Y y " ) r h @ t , ! t , . . . . ,: t n r l J n ). We can Lransfbrm any tbrmula in a prenex normal forrrr inLo such a,forr' adcling durmm-y qr-ranbifiers.Then pt has the fbllo.aring ibrm l-*r 'Ihe

, , ,tL,rL) , V ,t,tti,z\,.
i

wlrereti, are some terms anclz', are variables.Let h, . . . ,,fn h. new fuLuci;ion of p. lilow symboli that wilt correspond to the univelsally cl,Lanbifi.ed varia.bles ,,ve fbllow the derivation.of p lrom gt ,,bttt insteaclof applying t'ne rules, we substitute terms in ,p/ asfoliows.If generalizabion (V) is applieclto the variable ; .., z j . , t l r e n w e su b sb i tu te ( l) is used,then ' ,ve /i (1 i ,...,ti ) Lorz' i. if the r r - r le

ar/a

V. Bounded Arithmetic

c|o 1.oth-ing.In order to pro''re bhzr,tbhe resutlLingclisjr-inctionis a Herllrancl v/e sh.allsho,,vLwo things: clisjLtncbion, (i) ti,...,txi in- them) only afber.' are changecl(i.e. sornething is sLrbsbituLed t|,...',ti - /;, o{ z'i; gen.eralJzation. t.tr.e ( i i ) i f z ' r : z ! , , b h e nj : i ' a n c l l -

is generalized.betbre z?i musL (i) is faLse.Then some zfi ,,vh.icl't SuLppose ; r n n o l , a n r o n E2 1 , . . ' , 2 ] . F l e n c eb e f o r e ' f r o c c u r i n t i t , . . . , t ; x j .C l e z - r , r,lly ""t existenLially by killecl be tt; musL 1.\, . . . in occurrrences its be gene.raiized., , conLradiction, since suc]r .loitifyir.g bheterms inlvhich it occtl::s.BrL[ this is zt. of cluLanLifi.cation universal cluantifi.cationsare done only after: the e;risbential.
7

"j'

l,eb us prove (ii). I:f z'1 occrns in differenL disjuncLs, bhey must be first coni:i:acteclin{,o one beforswe can. generali.ze(see the restriction on (V))' Tv.ro clisjuncts rn.ay become identical in the course of the deriva.tion r// cliffer:e.nb after lve replace some terms ancl free variables in them by boundecl variables. However, clifferent t\,. ... ,txi and. tt,. . , ,tf can be identifiecl in sttch a way be only at'Lervre cluanti:fyz'i $v the same argument as in (i)). Thus (ii) mr-Lst true. Now it is clear that the procedure changes g/ gradualiy into a Herbrand disjuncbion. Condition (i) ensures that the Lerms have the required forrn, copclition (ii) ensures that we replace the same variables by the same berrns, hence 'we preserve the property of being a tautology. In each step the size of the clisjunction increases quadratically (since we substiLute some substrings the size of the resulting lle(tp) can be estimated of a string inbo itself), lr.ence by th(a) hlal ll-t(c112' < 222"

Tlre co:nversetransformation, from fle@) into tr,cut-free proof of P, is also form as above. First eliminate tire ner,v easy. Consider a formula p of the san-r.e by replacing maximal terms starting function symbols ft,. . ., fn in H"(d with such a function symbol by new variables. Let us denol,e the resulting formula by V'.Since He(p) is a propositi,onaltaubology, so is g'. Take some proof of ,p'; we have shown above that there is always one thai is cut-free ancl at most exponentially long. Now we only l:ave Lo see that we can apply the clualtifier rrLlesin a suitable order to sLritabieterrns and variables to gei; 5a.Th.e following rules lead to a proot o't tp: (1) Whenever we h.avea disjunct of the form

(Vy ) . . . (:"') (Ya") XVyi+r i)Qn ir-r


, h ( . 4z ,r,...,ti,Lti,x j) : - r , I lj t - r , .. . , r , . , U n ) , apply (:) i;o the Lermti;

5. Inbelprebability and Consislency

379

(2) if l,he abo''reis no{, possible, Lhen consictertbrmr-rlaeof the for:rn

( : r i . p 1 )(V y i -rt). . . (l r" )(V y,) ,lr(tt, t.t,


j ' . . - t r.. , x r r , l l n ) ; ,7i, zi rrS+'rrU

tarlie one of Lher-r.r such l,hn,LLlie Lerm in I:Ie(p) to which :.j corresponds is of ma,xi:mal length, anct gen"eralize25. convelge to V.ThtLs we only clezrr'1.y If we repeat i;his rr-ile i;he clisjurncbions have bo sho,,vth.at th.e rule can be appliecl repeabedLy,until the clisjuncbion becomes p. For (3) there are no restrictions, for (V) we have to shour that zi in only one disjun.cta:nclit cloesnoL occur i.n f 1, ilrr. . . ,t j-'r, z j-r,t,i ol occtr.rs Lhis disjtrncb. The last thing is clear, since zlt . . . ) zj cotrespond to the telms h ( t t ) , . . . , f i ( t t , . . . , t i ) , h e n c eJ i ( t t , . . . , f j ) c a n n o tb e a s u b t e r mo f a n y o f the terms

t t ,f t ( t r ) , . . , t j - r , l i - t ( t r , . . . , ' t i - t ) , t i.
in another clisjttnct. Now suppose (1) is nob possible and suppose zj occl-Lrs Since (1) is nob possible, the clisjunct musb have the followin-g forrn.

. . . (lr")(Vy") (l"r-rr )(Vyr,-rr)


,h('tlL z t l r t r k + r t A k 1 l 1 ,. .. . , i l n , 1 , n.) r ',r r . . . r t t p , \,z\,...,t't , T h e t e r m c o r r e s p o l r d i n g t o z ' * i s f p ( t | , . . . , t ' t ) . I f z i o c c r - r c r e d it n one correthe than be longer to z'1, would then the term corresponcling BuL then /r'(h,. . . ,t j) : sponiting to zi. Hence v/e can have'only j : "'k " ' t ' r : t ' r , . . . , t j : l / * ' F l e n c ei L i s t h e s a m e fn!t,,...,.t'p), thus 7 : k, and formula. We have proveil the following theorem. 5.19 'Iheorem. Ib is provable in 1X6 -l Exp that: a fbrmula has a cut-free proof if and only if it has a Flerbrancl proof. In symbols

ID o -l E xp | (V " )( CFPr ' ( ") = - HPr ' ( t:) ) .

of arithmetic we get bhe proof of th.econsistency IJy fbrmalizing a classical following resulL. 5 . 2 0 ' { ' h e o re m. (i) I Do + SttpererpF ConJyo and for every k:,IEs * f'tp ts RCon'q(lc). (ii) Ir0 + Exp F .HCon"O a,riorn(QB) ProoJ. First we L,ransform Q inbo an open theory Qt: we clelete open anclreplaceaxiotn (QS) bV e]11 funcLio.n the preclecessor clefining_(, a.clcl

380

V. Bouncled Arithmetic

',are ha.re some propositional taul,ology axiorrr. Assr-rme-,HCon(Q/), h.ence insbanceso,f -Q'.BLrt bhis .l:Ie(-Qt), which is a clisjtLnctionof sr-Lbstitution is impossible, since in IEgl Eap we can defi.nel,he bruth lbr open fortlulere. haye /Xo -l- .Unp I HConfi,. Thus ,,.re To prove (i), first i:eclucethe consistency of J tr'o Lo the consistency ol Q' trsing the inl,erpretations of.IEs in Q (Theorerrl 5.12) ancl Q in Q'. Then a:pply cut-eliminzrtion in J..!e * Snperet:p(Theorem 5.17 (i)), ancl bhe ecluivalenceof (Th.eolern5.19). Her:brand consisLency the cuL-freean.c[ RCon$(h is reduced to /i Con'O,(rn),for some rn, by the interpretation of Q in Q'. Applying cut elimination (Theorem 5.17 (ii)), this is f urrtherreducecl to CFCon|, ancl by Theorem 5.-[9 Io -HCon'r,. Thus we get

I D o- t E t p t sR C o n " q ( T ) . ' to H Coni in I Es * 'F)xp is just CFConi 'rhich is ecluivzr'lent Now 'BCon"q(6) n ih" rest of (ii). This showo

of .IXs * EnP (e) The IJ1 'Xheorerns


In this subsection we prove a theorem of A. lVilhie ,,vhich characterizes JI1 sentences provable in 1Io -l Enp as the I11 sentences interpretallle in Q ancl consicier some related problems. We sbart r,vith sorne lernmas. Reca.ll ih-at z denotes the formalized numeral function and that we are using dyadic numerals of Sect. 3. b.Z1 tera1na. For every inductive formula /(r) in Q, there exists k such that

t- (vr) RPr"q(E,/(")). s;1


Proof. Recali that a numeral D, for n ) 0,,is constructed as tbliows. We take n,i- I : ;n6f 2- t.T' henDg i s T anc [, rl l t. .. , n 4 su chth a t ng : I,nk : n and" l b r i : 1 , .. . , k , n ; i s 2 + n;- 1, if n,;: ) Y1' - t' (2,t,n6-) +T, if n;:Zn;-t *ri where T is S(0) ancl2 is (S(0) + S(0)). For a given inituctive .I("), take sorrie is a subcut of -I and /(r) -, J(t,r c), provabiy in Q. Then i;o "/(r) r,vhich show /(n) we needat mosL2k*'I proof steps.The formuiaethat we use are of polynornial size in the len.gthof n, hencethe whole proof has polynomial size.'l'heretbreS) cantbrrnalize Lhisproof. Nobe that the rank of {brmutlae n only by /' useclin the proofcloesnot clependon n, ltenceis cleLermined

5. Intelpretability

and Consistency

381

two termst(tt,'..,rrr) for every k suchbirab Thereexisbs 5.22I'emrna. s ( r t ,. . . , r n ) i n . L 6 : 5 r rl - ( V r 1 , . . .r,, , ) ( ( t ( q , . . . , n n-) s ( r r, . . . , r n ) - ,

and

n ): s ( i r ,. . . , i " ) ) ) R P r f i ( E , r ( i r. ,, .i ,
k ( t ( u 1 , . . . , ,n n ) f t ( * t , . , . , r n ) n R P r i y ( T ,t ( i t , . . . , i n ) * t ( i t , . . . , i " ) ) ) ) .

In the following proof ancl the next lemma we shall denote the function assigning the dyadic numeral to a numtrer also by numeral(r); the dob notahion would be rather confusing in some formulae. Proof. Let 1(r) be a cut in Q on which the operations * and 'r have the properties of ring operations. We can use this cut to construct proofs of instances of associativity, cornmutativity, etc. for numerals. Namel5 by the lemma above, we first show that the nlmerals in cluestion belong to / ancl bhen we apply the law in .I. This cut will be the only factor that influences the cub-rank k of the resulting proof. Now the sentence

t, (il,. .., in) : s(ir,...,in)) RPr'q(E


by induction. We only have to transforrn it into can be easily proved in S21 E.br fbrmula, i.e. we have to find a polynomial bound to the lengih of the " p r o o f o f t ( i 1 , . . . , i n ) : s ( d r, . . . , , i r ) . B y L e m m a5 . 2 I t h e l e n g t ho f a p r o o f of /(r) is polynomial in the length of r. Ne,rt we find polynomial proof'sof

("'1) (".2) ("-3)

-- nlLTneral (5(r1 )), S (nu,m,eral(, r )) n u m e r a l ( r t ) * n u " m e r a l ( n 2 ) : n l L r n e r t , l ( * t* r z ) , n u i n e r a l ( r r ) , F m t m e r a l ( r 2 ) : n r L n L e r a l ( q* ' x ) .

Perhaps Lhe besb way to see a polynomial bouncl bo these statements is the t; where t; ate terms of the form .fiollowing.First construct proofs of i;:

i ' - l - . . . + T * , ( r r> i z > " ' > j ^ ) ,


wher-e 2'" clenotes2,t, (2,F ' ' ' :F2)). Think of t; as a modifi-eil ntrmeral. Then follow the stanclrrra ofgo'rinms fbr -l-' and ,r to obbain Lhe proofs of (e.1-3) for the rnodtf,ed n.umerals.Finally bransform the modified numerals back into r2 e) spectively. original nu,meral(S("r)),numeral(nt1-x2) a;nc|'nu,merul(r1'r.1 As is easily shown by induction, the maximal of these three bounds (the
-t,

382

V. Boundecl AlithmeLic

bouncl for the lengbh of the proof of (e.3) is asymptotically maximal) is a bouncl to the pi:oof of the eclualiLyo.[ general terms. To prove Lhe ineclttality, balrea suitable term tr) Prove t ( i t , . . . , i " ) f - 5 ( u ) : s ( z : 1 ,. '. , i n ) ,
o1'

o ( t r , . . . , i n ) * , 5 ' ( T .:r )t ( i t , , . . . . , i n ) , Lhefact that and r-rse

Q t sl ( r ) U t ( a ) - + x - r S ( Y ) / " .
5.23 Lemma. There exists k suchthaL - r 1) = ( r l nur ner cr l '( v ) V 5 ; F (V y)R P r" q( 8,,( vc) ( rl nu,m er al' ( y n:nurneral'(y+1)))).

Proof. We consbruct a cut / in Q such that

(".a)

-* Q | (Yr,z)(I(z) (r < S(r) = (r < zY r:

S('))))'

First take "I defined bY J ( * ) : a f 0 + r : x k ( V z ) ( S ( z )* r : S ( z + t ) ) . One can easily show that ,/ is inductive. Now take I to be a subcutof J ancl verify (".a).trVeshallwork in 5']. Let anumber g be given. By Lemr''aS'2'L we have Q I l(numerat(y)), henceby (e' ) ' Q l * 1 S ( n u m e r a l ( y ) ) = ( *< n u m e r a l ( v ) V r : S ( n u r n e r a l ( v ) ) ) BvLemmaS.22Qlmr'merat(y+1):S(nu'meral(g))'ThusQproves r ):(, (Vr)(clnttmerat(A-,1 lnumerul(y)V r:n\Lrnerul(y+1))) . n

cut-ranlc. All that we ha.zeused were proofs of boLrnclecl

'weshall fi.x some translation of L2 into tro which. For the lbllov,ringlemma, o,fBASIC 01]solnecut in Q. Such a translation an inter]p::etation clel,ermines shall say exists,since IDO-f fit is interpretable .in Q by I'heorerr 5.12. We of E! (resp. n!r!,it it is the bra-r.nslation that a formula 4; o,fLg is essentzalty o! Lz. a Xf (resp fIP) formr-ila

5. Inierpletability and Consistency

383

of 8). 5.24Lemrnzr(formalized I completeness '.Let bhat (i) 1.,(r,g) b" abottnded formul.ain trs. Then there exists k sLrch

-, RPrfi(E,1=v1,1,1,t, ' F (Vr)((rv) I Eo-r Enp e/))) d@,v)


existsk such that Xf . T'hen bher:e (ii) LeI $(x,y) be esseniially

-, RPriT(E,(1a),h@,y))) . s} t- 1vr;((1y),b@,y)
to prorze only Proof . First we observe thab in both casesit suffi.ces

- EPr'q(E,rh@,ilD (Vr)(t!@,,v) ,
since (lilrh@,y) will tollow by an applical,ion of the rule (1). (i) The task of prolring ,b@,il is similar to the taslc of decicling its truth: 'we shall prove ,b@,y) by proving numerical instances of the open patt o;f.tf;. Thus v,/ecan bound the size of the numerals needed in this proof in the same way as in ihe proof of Theorem 5.4. The numerical instances of the open part of ,r/ follow frorn the nurmerical instances of abomic ancl negated atomic sttbformulae. The existence of bounded cut-rank proof's of atomic sentences with - and their negabions follows from Lemma 5.22.'Ihe atomic sentences with ( and thei:r negations are reduced to atomic sentences with : in the same way as the negations of atomic sentenceswith : had been. We must be a little careful when pasting the proofs of the numerical insbances into a proof of a bounded sentence, since we n.eed a proof of bounded cut-rank. A sentence of the fbrm (ly < t)*(A) follows from sorrle to the value of f . A proof of (Vy < t)t(y) *(tt), where n is less than or ecltLal is constructed from the proofs of a(0), t(T), . . . r*(m), where rn is the valtte - 0,..',rn' For as follows. We prove successively(Vy < n)*(A), for n of 11, n : 0 iL is an immecliate consequenceof *(0); to obtain it fbr n + 7 fi'orn a proof lor n, use Lemma 5.23 ancl a(rz * 1). F'inal1y use Lemma 5.22 to show ',;'/here we proof of (ii) is only a moclificalion of the proof of (i) (ii) 'We aiso neeclan extension of Lemrna 5.23 bounded. replace boundedby sh,arytly lo L2, insteaclof J,g, and uASI.C, instead of {J. Again the proof is essenbially the same: Lranslate the natural polynomial time algorithms tor the evaluation of terms in L2 into a proof using ec;-ralitiesr,vhichhold on a suitable cuL in

v < t ) * ( y ) = ( V y< m ) o ( y ) Q t s( v'fhe

BASIC . We omit the details. fo::mulaand suppose 5.25 Lernnta.Let T/(z) be a'boundecl

IDo-.Eapl(Vc)/(r).

384

V. Bor-rnded Aribhrnetic

Then there exists ,b such that

- 2f,)-, ,h@D , IEo t (Vc)((:y)(y


tl'(")), for anv k' nob prove (Vr)((ly)(y :2i) ---n .Proof. Suppose/fs cloes IEo- F { ( :yXV - 2"h) k- - { ( c) | k e tV} ( t a new T h e n , b y co mp a ctn ess, is consistent.Take amodel IVI o'fthis Lheorya,ndan initial segmen'u constanb) of it

K : {cte tuIl(lk' N)(tvItr a 32")} .


'I'hen 1( is a model o,t I Do Exp, but K tr --nlt(c),because is boundecl. n 1/ Novr we are ready for the theorems. formula. Then the following are ecluivb.ZGTheorem. Let $(r) he a bouLnded alent: (1) There exists a cut /(r) in Q such that

Q F (vr)(/(r) -+ d@)).
(2) There exists an interpretation of Q + $n)$(r) (3 ) IE o + E xp | (V a) /( r ) . io Q.

of cuts closedunder of shortenings ProoJ.(1) =+ (2) followsfrom the existence


* and i''

(Z) + (1) LeL r be an interpretationof I + (V";$(r) in Q. We shall construct a cut /(r) in /Xs suchthat

. IEsr (Vr)(/(r)-, ,b@D


This is suffi.cient, since J'Xg has an interpretation in Q determined by a cut. The cut /(t ) is defined by the formalization of the following formula: "There exists an isomorphism of the initial segment [0, r] onto an initial segmenbin the sense of interpretation 1". Here we take intervals as strucbures with:, (, ancl *, ,Fas ternary relations. An isomorphism is a number which codes a sequence of pairs. lVe leave it to the reacler l,o write clolvn an explicit fbrmula lb; f and chlck that it is a cut in ,IXs. Let us only note that the extension of an isomorphism with clomain 10,r] to an isomorphism with domain [0, S(')] 'is obtained by aclding to it a pair (S(r), z), where z is the , successorof the image of s. Sr-tchan e:ctensionof a sequence is possible in /16 (but not in by the Q). Now the fact that (Vr)r/(r) holds in the sense of I is transferred isomorphisms onto -I. (1) + (3) Suppose

. Q F (Vr)(/(r) -' ,b@))

5. Intelpletability

and Consist,ency

;lBs

By l , e m m a 5.2 1 I Eo j- ErTt| (Vr)fi Prft(F, r(r)) , for some constant /c.Hence r E o + E rp F ( Vz) R/,r [ ( T,{ ( r ) ) . o,f.Q with respecl,bo By T'heorem 5.20, IL'O * Erp provesthe con.sistency proot,sof cut-ranlc /c. Thus we have

I E o- l E x pF ( V , ) - ; P " h ( T , - r h @ D .
By Lemma5.2a(i) Enp| (Yr)(-n[@)-t RPrq(m,'-{(i))) rLlo -tinstead for some m. We can asstlme that k : nr, (otherwise talce max(k,*) of /c and rn). Flence (3) follows from the last two stateme.nts. (3) + (1) Suppose "IX6 * Eap | (Vr)rl',@). BV the lemma above we have some k such Lhat

: 2f) '- ,h@D . IEo F (Vr)((ry)(a


Take ctrts -I, J in Q such that ,.I determines an inberpretabion of I\s and -I is a strbcut o,f J such that in Q

-2Xk t(a)). Q I I ( x )* ( 3 y ) ( y
Then we have Q F (Vr)(I(r) --' ,h@)). tr

in trs. Then the following are formr-tla 5.2? Tlreorem. Let {(n) be a boundecl equivalent: ( 1 ) I Eo - r E xp | (V r)/(r). ( 2 ) F o r s o me/c sr1t_1vr ;Rpr "q( E,r lr @) ) . I:t{,,(r) is the translationof a.trf forrnulatp(r) into -[,s,then (1) a,nd(2) are
equivalent to:

(3) For some k S i +- R Con' q( k)F ( Vr ) e( r ) . ProoJ.Q) =) (2) Suppose ID1 t Ex)pl- (Vr)/(r), rlt botrnded.Their bv T , h e o |e m 5 .2 6 w e h a ve Q r (vr)(.1( x)- r h@D tbr sor necut - Iir . Q.By Xr

completeness

| nprb(L (v"x/(r) -, /("))) , s')

386

V. Bouncled Alithmetic

fo r so mej .B V L e mma5.21

r-(vr;RPr$(m,/(r)) ,s$
rn. Talce& to be the m.aximumof 7 and rn, then we ha,ie (2). for so:me (Z) --- (t) Bv definil;ion

(".5)

-' (vr)-'Ji -r RCon$(hF (vr)liPr'q(T,rb@D . Prfi(T,-,./,(i)) s').

in /Xe "+'Erp,thus (e.5) is provable By Theorem 5.20,RCon'q(E)is provable in..[Eto1- Erp. Sincewe assum.(2), we have

. IEo-l Expts(Vr)-.g.Prb6,-.rh@D
B y L e m m a5 . 2 a ( i )

-, ,h@D . I Eo-l Eapts(Vr)(-ftPrb(T,-r/("))


(2) implieu(1). FIence that ,r/ is the translation of a l7f formula cp(r) into ,ts. Now assttme (Z) =+ (3) BV Lemma 5.24(ii) we have

-, ,!@)), F (Vr)(-RPr$(E,-,rh@)) 5',21


Then using (e.5) we get

F (vz)/(z) . s) -vBCon$(E)
from (2). But 9@) is ecluivalentto {:(a) in S}, thus we have (3)(3) + (1)'Ihis is because/)lo* Erp "contains" S$ and,provesnCon"q(E).

o! IEo* Enp over' say, S'} We wouldlike to axiomatize I/1 consec{ttences (put oLherwise, we want to find some nice basis). Tlie ecluivalenceof (1) with a formula of (3) gives only a partial answer: since RCon$(E) can be written ai,s es RCon"g(T), lc : 0, 1, . . . with { in n!, the s,etof senbenc the fbrm (vr)/(z) of I Es + Etpaxiornabizesthe Vj?f consecltlences

Theorerns (f) Incompleteness


theorems did not hold in if the incompleteness It wouicl be paraclo,rical wealcsysbems.Of course tormalization of syntax in weaker theories is a therefore we cannot apply classicalproofs of the second more clifficuli l,aslc, oul, that Lhe seconcl theorem cluite directly. IL bur-ns Godel incompleteness

5. Inlelpretability

ancl Coirsistency

387

A typical exatnple is the can be even sLrengLhened. bheorem incompleteness in of Q in IEo -F Eap. The results presenbecl unprovabiliLyof the consisbency theorems proved above and of Lhe of bh.e this subsectionwill be corolla:ries following theorem. Iiobe bhabin the following theore* (i) holds for every finiiely a.riomal.ized here. this generalbheorem shall not pr:ove theory containingQ, SNc secltreniial /X1, staternent (i) For finite theoriesin bhelanguageof arithmetic exbencling 'followsfrom Theorem 3.20(1) Chap. III. Statement (ii) gener:alizes TlreoChap. III, to weakertheorieslhoweverthe assumptionabonL tire rem 3.1..1, here. It is possibleto use here the same assumpis sLronger a,riom.abizability III, but we shall not pro're this bion about the ax,iornatizabilityas in Cher.p. Cont' (T) is Recall that -B strengLhening. (V z)(t(z) kc-ra n l ,:(t) ST, - n - ,Pr oof|( ".,0=T) ) .
5.28 Theotr'em. (i) For every k there exisbs a cttt I in Q sttch that

e I RConS fD;
(ii) Let ? be a consistent theory containin"g Q ancl having a Xf axiornatization. Then for every cut, J in 7 there exists rn such that

-(? l- ncon?f @)) .


(i) Bv Theorem5.20(ii), for everyk ProoJ. IEo+Eupl RCon'q(E).
By Tlreorem 5.26,,ihis implies the fi.rsbsbatement. (ii) The proof of this sLaternent will be similar to the proof of Theorem 3.11, C[ap. I11. RoLrghly spealcing,we only have l,o talce rn su'ffrciently large. key role in the rest of this section, we shall pr-ove it Since bhis resurlbplays zr, in detail. Because of the shortening techniclues, we may stlppose thal, ../ has sorTle actditional properties. Thus 'we shall assume that J is a cut in Q and it cleterrninesan interpretation of S$ in Q.In particular, it is provable in Q, lrence also in T r Lhat for any tr,vosequences in -I, th.eir concatenation lies zr,lso in J. 'We'wouicl lilce to use the provability condibions of 2..16, Chap. III, for .RPr.y(m,,r). We shall prove the first tr,vo.Th.e third one, which is folinerls'zecl' Ponens, is clearly false, since Modu,s Ponens increases the cr-rL-ranlc. .tulorl,us proof follow the usuzr,l nz, and Lhen we shzr,ll Thus .,veshall first take a sLriberble

383

V. Boundecl Alithmetic

of the second in.completenesstlr.eorem and check that ihe thircl condition is \vVeshall clivicle the actu.ally appliecl only Lo formutlae of ranlc at mosl, rn. proof inbo several.claims. sentence p, if T F tp, bhen T F CFPr'{ f'a' Clalnr 1" For ever:y' .Proof. Su.pposeT I p. By cut-eliminatj.on vre ha.re a cLtt-free proof of cp

an of Q, Q F CFPr+(A). Since,.I cletermines inT. By the I-'-completeness fl of the claim' get the conclusion of Q in T, r've interpretal,ion
/c si-rchLhat Ciaim 2. There e.:<ists

' (r,y)-, RPrft{ :r I RPr?f 6,ErQ 1*;911


.l,et us first write clownRPrf @, y) more expliciLly: Proof (f.D ( J(d ) k J( x) 8zJ( 0 k c- r ank( d)I t 8t Pr oof7( d,v) ) ' for some /c1 5.2t1, By Lem.rna S $ l ra n k(d )< rkPr oofb@,y) - , ---+ < i k.Pr oof|( d ',i l ) . RPr ' 6( T1,r onk( cl)
By Lemma5.2'1, for some /c2

. k J(i) k J(i/D J(rt) sl r aerq(Tz,


Tb obtain Aer.f (i,i) *e needonly the rulesfor & and I which do not - max(kr,kz) for &. Hence the cuL-ranlt. increase

' ar$ 1" s) r arri@ , v1:. , y)-- RPri.(n,


in. ?. r\s r | 6ilJ, Flerewe can replaceQ by ?, sinceQ is conta,ined we get

Aer! 1* , T | (RPri.@, ,s111r il -, RPr?yft,


which is in fact the statementof the claim. By 2.I, Chap. 1II, we have some6(t ) sucl:thah I

' f 6(r): -RP''t' (r, a(rN) "

5. Inietpletability

anrJ Consistency

339

we can choosern.I.'eL fi") ir the numeralof the fbrmula 6( numeral(n)).No.,v


m be such that (1) rn ) h, where /c is from Claim , . (2) rn ) rank(nPr{ (r,,y)); (3) rn ) rank(A(")). Note that we thus have also nz ) rank(RPrf (^,g/)) anclnt 2 r'unk(6(m)). Let us clefine also A : t r o\r n) j

. d :ctf Afr?r{ 1nr,A1

Claims. T I A - -RPri/ 1wt,A7,


Proof. By definition we have

rts6@)--,RPrtrl1m,ffi7.
Thus we only need

r tsi6 : 5(*),
L-l

of Q, hence of.T. which follows from the X-completeness Clairn 4" -(T I A).

? lProof. SupposeT I A. Then, by Claim I, T I CFPrlr (A), ,,vhence C l a i m 3 th i s means that ? is inconsistent,which is a conn P r ? / @, 2 ).e y
tradiction.

Claim 5. ? F ItPr2l (m,') k RPr{ (m, =r) -+ -Rcont' Qn).


Proof.If a and -ra are derivable using proofs of cut-rank of at most trr, t h e n so are a V0 : T and -ra V0 : 1 (acld 0 - T to each proof line as a side -T formtrla). Since m ) rank(RPr?f (r,y)) : rank(a), we get a proof of 0 '1-'he by taking the concatenation of these proofs and applying the cub rule. n closure properties of ..I ensure that such a proof will be also in J. Clairn 6. 7 F RPr{ Proof. By Claim 2 (m,Z) ---" RPr{ (rn, a).

r I nPr\{ (m,A) -, Rtr\{ (m,RPr{ (ffi,Z)) .


of a and l-completeness By the defini{,ion

TI ItPr\l(*,21:a.

rl

390

V. BoundeclAriihmetic

C l a i r nT " T t s E P r { ( r n . , A ) n

RPrf (m,-cv).

Proof.By Clairn 3 and the defir:ition of a, r'' l- A -- -a' BY Claim 1, this impJiesT F .RPrt' (^,-A -, -'cv). If we have proofs o,f A ancl A --+ -a of we ci,n easil,ycombine them inbo a pr:oofof cut-rank rn of -cz, cu-ranlc ?77) that Lhis m ) ru,nl:,(A).Again, bhe closureproperties of J' ensr-rre becatrse n proof is in J Loo. Now we can finish the proof of part (ii) of the theorem. By claims 3, 6, ancl 5. we have in 7 ''(

-,A -> Rrr{ (n,A) -, @Prf @, *) 8, RPr?f (m) , (m,"-*)) -> -,RCon?1
i.e. T ts RCon{ (*) A. }Ience, by Claim 4, RCon};t (nz) is not provable in

5.?9 Corollaly. Cong is not provablein JXs * Eup. noL prorre Conff for any cut -[' Flence . By the theorem above, Q cloes Proof n by Tlreorem5.26 IEo + Erp does not prove Confi. b.30 Corollary. Cut elimination (unrestricted),and Herbrand's Theorem are not provable in I Do * Enlt. Proof. By Theorem 5.20(ii) /t0 + Erp proves cut-free and Herbrand o,fQ.If we had cut-elirninationor Herbrand's Theorem in I f)o -l consisLency of Q to the restrictedone. n Erp,'we could reducethe ordinary consistency Theot*ern.IEs + Exp * Conlro doesnot prove Con"12o4E*'p' 5.31. by Proof. Let J be the cub in I Eo + Enp clefrnecl

J ( * ): a r F i l @ : 2 9 ) '
\ffe shall show that

(f.1)

IDo + E:cpl- Con"{

Suppose 1ot. Let us work in .l-Xo -l Et:p,let d be a proof of a contraclicLion 'I'hen d has cut-rank ab most lh(d). We have in J'. t0.t ,ttr,(d) < j 'd 's'rttt'(a) be nonstandalcl and d is exponentially larger tb'an lh(d); hence since c/ mr-Lst such a number exists. Thus lve can apply cut elirnination to olltain a cub-lree

5. Inter:preta,bility and Consistency

391

proof of a conbracliction itt 8 (upply Lemill.& 5.I5 l,h(d)-times and use tJr.e ineclualiby , Z z + r n ( a )2 r r ' o ' " ' ) , Br-it this is nob possible by 5.20 (ii). of l,he l,heoretnis lalse. Thus No,,v suppose bhe sLa'uelr.e.Lrt IEo -l I)rp I Conl"o -, ConlyorExp. Since Con12o an.d Cong ale ecluivalenbin. /Xg -f Exp, we have .tDo * Erp F Coni2 -> Coniroy-Eup. that pro'vably in -Ixg { Erp,, a u'sing an obvious Lemma 5.25, we can decltLce proof d o'L a contradiction in I EO + Erp -can be transformed into a proof dl of a contracliction in Q such lihat c{ < 2fl f.or a suita'ble standard k. L e t I ' b e of "I such that a sr-rbcub

l rlo-r Erp F /(") -' J(2T).


Hence a proof d of a contracliction in /Xo + Et:p such th"al, r/ is in .f can be transformecl into a proof clt of a contradiction in Q such that d' is in ..I. Thus, by (f.f ) no such proof can exist, i.e. -IXg * Erp proves Conllyo-pocp. B u t t h i s n is in conbradiction to Theorem 5.28 (ii).

5.32 Corollary" IEo* Enp is not iuterpretablein 1trs. Proo!. SupposeI flo -l Erp is inLerpretablein I Es. It, is well-i<nownthat if 7 LhenT is consistenttoo. This simple is interpretablein ,9 and 5 is consistenb, theorem can be easily fbrma[zed in /Xo + Ury, if. T is finitely axiomatizecl. Si:nceIEo+ ErTt is finitely axiornatizable(Theorem 5.6), as one can check provably in .Itr'6* Erp, we get /tr'o t Etp ts Con'7r, -, ConlyotEap, 5.31. r,vhichcontradicts T'heorern

5"33 '{'heorerrr" For every k there e;cistsrn such that RCon'q(m) is noL provable

in J'I'o* 9n * RCon\(h) for any n. . We shalluseTheorem5.28again.Let l' be a cub in Q such ihab Proof Q I RCo'nbt(f)

392

Arithmetic V. Bor-rndecl

'We an interpretation of I Eg. Let J be a sulrcul. that / deberrnines can s'Lrppose that of J. such. I(22*) . Q I J(n) --+ '\rVe lrnovrthat, florsomem, Q*-.R Con'j (n) is consistent.Let fuI be a mocl.eJ by o[ i;histheory. l,eb I'f be its initial segnientclefinecl : {ct e IVI;(lb ltuISlZte lr)(a < cri(b))} . .11 "Let ct K; Firsb ',ve shall show thab /( is closed under all functions arn. 'we can supposethat a is nonstandard. Then, for .withoLrt loss of generality b e JM, a < w;(b).'Let n be given. Then somenonstanclarcl ) r *r *( n,;) ( um' x( ",i)( b) ) o n (a) I wn( co;( b) S ( ur'ru*(",,;)+r(b) e K' prove th.e last inequaliiy, use incluction on j to show that 1_.o ai@5@\ S ,j+r(") , for every sufficiently large r (hence for every nonstanclard r)' No',v we neecl that .F( g IIUI.This follows from the definition of the cuts and from bhe fbllowing inequality, which is also easily provable by induction on j: ,i@) < 22', for r sufficientlYlarge '

IEs, { is a model of 1Xs *un, in .IM, which is a model of. SinceJ'i is containecl luI F --R Con{ (m), we have for every n. sjinceQ I RCo""{(E) and.

U -ACon'q(m) K F RConb@> ,
which finishes the proof. tr

This theorerngivesus another pieceof important information about VII| of IEg:L Er,p: they are not "finitely basecl".Furbher we have consequences this consecluence. over IEo+ {Qni n //}' Coro[lary,.lDo -l Erp is not -Il1 conselvative 5.3,tr In flct this is all ihat we lcnowabout -ll1 conservativitybetweensystemsof plausiblethai /Io * At is not -[t ar-ithmetic.In particular it seems bor,rnclecl over -Ifg, but we cannot even prove that IEO+ {A"; n lr/} is conservati've which might separateI Eo* flt over /Xs. A .I/1 sentence conservative nob ./7'1 Bertrancl's postulate. It is the theorem (due to from ilo ir the so-ca1lecl Sylvester) which says that th.ereis a prime between n and 2n, for every n'

5. [r-rLerpretabilityancl Consisl;ency

393

in. This senten.ceis provable i.n.[tr'g * At br-rt no p::oo[ of ib has been foLr.nd is an impo.rtant open pr:oltlem. /tr0. WheLher such a proof e,'cisl,s

(g) Or the Lirnitecl Use of llxponerrtia.tion


tJren it seems Lhat a1].such, senIf we are in.berestedonly in .[1 senl,ences, te:nces wh.ich are also truLy' mathematical results are pr:o'vablein Itrg 1- Erp for an ordinary'rnaLh(.,veconsider sentenceslilce Con(Q) Lo be u.ninteresLing ernatician). In fact the proofs use exponenbiation at most on.ce.This m.eans the set of all that for instance if lve constluct soure graph, l,vemay conside.': graphs of a given size, bul, usually we do not consicler the seb of subsets, etc. Pr-rt otherwise, we use l,he finiLe potver set operation only once. I'his phenomenon is worthwhile to formalize, l>ut it cannot be done directly in first order logic. In first order theory we have either to accept axiom Erqt, and tlren we have also all finite iterations of 2*,, or we do not have exponentiation at all. A possible approach is to use higher order bounded arithmebics. We shall use a different approach. Our aim is, roughly speaking, to sho.,v bhat there is an infinil,e hierarchy according to the number of appl.ications of bhe exponentiation axiorn. So what does it really mean thai (Vr)p(c) is provable using e,xponenbiation that 2o exists; fbrmal[y only once? This means that rve prove tp(r) assr-nning

I E o - v , 0 1F ( V r ) ( ( : y X v: 2 " ) - - , p ( n ) ) .
IL is convenient to consicler IEg-F At instead.of /X6, since thus we obbain a concept which is more robtLst: in .IX6 -F ,Qt, 20 e,rists iff 2u- exists, etc. There that (Vr)ta(r) is provable using is an ecluivalent and shorter way to e,Kpress only once, which is e;cpon.entiabion

. I Eo+ a1ts(Yx)e(l"l)
Sirrilarly, k appiications correspond to the plovabiliby of (Vr)9(lr,l(tu)), r,vhere lrl(i') denotes the k times iberated function r * lrl. Finatly let us also mention a model theoretical characterization of provby abiliby using expone.nbiationonly once. Let f(z) be clefi.ned

:2*) I(*) : af GY)@


'LeL K be the class of models of the forrn

1i - {cte N|; IuIF IDs* Atk IvIF J(o)i ,


are in clues;l,ion with the operabionsinheritecl from lrzl. T'hen the senl,ences true in all models of /(. the senl,ences iursb

394

r\rithmetic V. Bor-ruded sr-Lchl,J:at for e'rery nz b" a bor.rncled.for:muLla

L,et $(r,y)

F (Vy)d(ifr,y) = -RCon"iy(m) .Itro-l-J21

is not Lhat(Vy)rh(m,lvl(/')) tn such. e,tisl,s Fbr everyk th.ere 5.35Lerrrrna. provable in /trs -FAt.
in Q st-ch that .I deLermii:esan inLerpretation ol. Proof. Let /(r) be a cr-rb IZo 1- t\. Let "/(c) be a su.bcutof -i(r) such that

' Q I J(x)-+ I(zfr)


By Theorem 5.28, there ex.istsrn such that

-(8 | trConf (rn)).


Then is provablein IEo 1- J71. that (Yy)rlt(m,lyl(e)) I\ow snppose

rb(*,lvl(/')))' , I F ((VE)
hence, by the construction of ..I,

, 0 r ((vv)Q(m,y))r
which is a contradiction. b.BG Tlaeoreln.For every k bhereexists a boundeclfblmula qa(r) such ihaL ll

I E o+ a t | ( V " ) , p ( 1 r ; ( r + t ) ; ,
but

- ( / t r s * a r F ( v r ) p ( l r l ( A ).; ;

proof . Let, rn be strch that (Vy)d (*,la1(l') ; is nolr provable in I Do )- J}r. Since in If)o+ Exp, iL must be provable RCon"nQn), i.". (Vy)',h@r,y),is provabJ.e "Jripott"rrtizrtion a finite number of times, see Lemma 5.25' Th'us, fbr usiLg solne 72,

IEo -r th t (Vz)$('n-t',lcltn't;.
H e n c ef b r s o m e i , k < i 1 n ,

r* / , 1 * 1 ( ; - r -,t ) ; I E o1 f r t t s( V " ) ,(

5. InLelprebability and Consistency

395

anc[

(;) -(/X6 -t-,A1 F (Vr)ILOn,, lr 1 1;.

'Ialce p ( c ) t o b e g ( m , l r l ( d - t ' ) ; , Lhen the th.eor:emfollo'ws.

Tire insta,nr:es of the schema.PIlPL'g are./71 and Lhey are provable using one exponenbiation. AcLually, mosl of the pr:oofs that recli.rire exponentiation can be reclucecl to PH.PEo. Sbil] ib seems thaL PHPIg is not su.Ifrcient to prove all senbencesderivable using one e;Kponentiation. We are not abLe to cleri'veP.|I.PEs in. /f6 -l ,Qt, so it is possible that some instances could be ttsed as more nabr-tral. flormulae in ihe theorem. above Ior k: 0. However we clo noL know which. The poinb is thr-rt tlr.e lcnown mathematical staber:o.ents whiclr can be cteriveclr.rsingPHP Dg can be derived also from a weaker version of P|HP which is already provable in -IIs -F At (Bertrand's posbulate is an example). The relatecl problems of whether I Es]_9,;is -Il1 conservabiveover I Eg-lg j, fbr 7 ( i, are stil1 open. They are typical in the sense thaL they ar.'e relaLed to open problems in complexity theory. For instance, the cluestion whether IEo 1-,Q1 is ff1 conservative over /.N6 seelns to be formally related to the open problem whether Pl-l - l-inH (though we are not able lo proae arry relation). New techniques will need to be clevelopeclin order to solve such problems. Solving such problems in Boundecl .Arithmetic might be the first step in solving the correspon.dingpersistent problems in complexiby theory.

You might also like