ebook img

Sublogarithmic uniform Boolean proof nets PDF

0.15 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Sublogarithmic uniform Boolean proof nets

Sublogarithmic uniform Boolean proof nets ∗ ClémentAubert LIPN–UMR7030,CNRS–UniversitéParis13, 99av.J.-B.Clément,93430Villetaneuse,France AbstractUsingaproofs-as-programscorrespondence,Teruiwasabletocomparetwomodelsof parallelcomputation:Booleancircuitsandproofnetsformultiplicativelinearlogic.Mogbilet.al. gavealogspacetranslationallowingustocomparetheircomputationalpowerasuniformcom- plexityclasses.ThispaperpresentsanoveltranslationinAC0andfocusesonasimplerrestricted notionofuniformBooleanproofnets.Wecanthenencodeconstant-depthcircuitsandcompare complexityclassesbelowlogspace,whichwereoutofreachwiththeprevioustranslations. Introduction Boolean proof nets were introduced by Terui in [8] to study the implicit complexity of proofs nets for Multiplicative Linear Logic [4] comparatively to Boolean circuits. Those two models of paral- lelcomputationweresuccessfullylinkedusingaproofs-as-programsframework,whichmatchesup cut-eliminationinproofnetswithevaluationincircuits. Surprisingly[8]doesnottakeintoaccount uniformity, which guaranteesthattheresources neededtobuilda Boolean circuit isinferiortothe computationalpoweritwilldeliver. [7]and[6]studiedtheBooleanproofnetsinauniformwayand introducedsomenon-determinisminit.AstheirtranslationfromBooleancircuitfamiliestoBoolean proofnetfamiliesisinlogspace(L),itremainedunknowniftheresultswerestillvalidwhenapplied tosublogarithmicclassesofcomplexity,thatistosay AC0andNC1. ByrestrictingtheBooleanproof netsweuse,thispaperoffersanewproofofthecorrespondencebetweencircuitsandproofnetsand extendsittoconstant-depthcircuits. Boolean circuits ([9], section1) and proof nets ([3], section3) are canonical models of parallel computation,butthelatterwasmostlyseenfromtheviewpointofsequentialimplicitcomplexity.To evaluate a proof net is to eliminate its cuts, but to do so with suitable bounds we need to define a parallelelimination. Tryingtoapplythetwousualrulesofrewriting(→ and→ )inparallelleads m a tocriticalpairs,soweareforcedtodefineanewkindofcut(tightening-cut)andanewrewritingrule (→ ).ThesimulationofthisreductionrulebyBooleancircuitsneedsUstConn gatestobemadewith t 2 efficiency. TheproofnetswestudyareforMultiplicativeLinearLogicwithunboundedarity(MLLu,section2) andbecauseofthelinearityofthislogic,weareforcedtokeeptrackofthepartialresultsgenerated by the evaluation that are unused in the result. Boolean proof nets (section4) – as introduced by Terui–haveanexpensivewayofmanipulatingthisgarbage.Inthispaperweintroduceproofcircuits (section5)asarefinementoftheBooleanproofnetsthataresimplertomanipulate. Theyaremade ofpieceswhichtranslategatesandcomposeeasily,sothatwereducethesizeoftheproofnettranslat- ingtheBooleancircuits. Insection6weconcludeourpaperwithourmainresult(theorem2): there existsaconstant-depthreductionfromBooleancircuitfamiliestoproofcircuitfamilies. Soournew frameworkoffersavariantoftheproofsforcomplexityresultsandextendsthemtosmallclassesof complexity,inauniformway. ∗WorkpartiallysupportedbytheFrenchprojectComplice(ANR-08-BLANC-0211-01). Jean-YvesMarion(Ed.):SecondWorkshopon ThisworkislicensedundertheCreativeCommons DevelopmentsinImplicitComputationalComplexity(DICE2011) Attribution-Noncommercial-ShareAlikeLicense. EPTCS75,2012,pp.15–27,doi:10.4204/EPTCS.75.2 16 SublogarithmicuniformBooleanproofnets 1 Booleancircuits Boolean circuits (definition2) are of great interest in the study of complexity, for instance because oftheefficiencyoftheirparallelevaluation. Oneoftheirfeaturesisthattheyworkonlyoninputsof fixedlength,andthatforcesustodealwithfamiliesofBooleancircuits–andtherearisesthequestion ofuniformity(definition4). Definition1(Booleanfunction). An-aryBooleanfunctionfnisamapfrom{0,1}n to{0,1}.ABoolean functionfamilyisasequence f =(fn)n∈NandabasisisasetofBooleanfunctionsandBooleanfunc- tionfamilies.Weset: B0={¬,∨2,∧2}andB1={¬,(∨n)nÊ2,(∧n)nÊ2} TheBooleanfunctionUstConn ,givenininputthecodingofanundirectedgraphGofdegreeatmost 2 2andtwonamesofgatessandt,outputs1iffthereisapathbetweensandt inG. B B Definition2(Booleancircuits). Givenabasis ,aBooleancircuitover withninputsC isadirected acyclic finite and labeled graph. The nodes of fan-in 0 are called input nodes andare labeled with x ,...,x ,0,1.Non-inputnodesarecalledgatesandeachoneofthemislabeledwithaBooleanfunc- 1 n B tion from whose aritycoincides with the fan-in of the gate. There is a unique node of fan-out0 whichistheoutputgate.Weindicatewithasubscriptthenumberofinputs:aBooleancircuitC with ninputswillbenamedC . n ThedepthofaBooleancircuitC d(C )isthelengthofthelongestpathbetweenaninputnode n n andtheoutputgate.Itssize|C |isitsnumberofnodes.WewillonlyconsiderBooleancircuitsofsize n nO(1),thatistosaypolynomialinthesizeoftheirinput. C accepts a word w ≡ w ...w ∈ {0,1}n if C evaluates to 1 when w ,...,w are respectively n 1 n n 1 n assigned to x1,...,xn. A family of Boolean circuits is an infinite sequence C = (Cn)n∈N of Boolean circuits,C acceptsalanguageX ⊆{0,1}∗iffforallw∈X,C|w|acceptsw. WenowrecallthedefinitionoftheDirectConnectionLanguageofafamilyofBooleancircuits,an infinitesequenceoftuplesthatdescribesittotally. Definition3(DirectConnectionLanguage[9]). Given(.)asuitablecodingofintegersandC =(Cn)n∈N B afamilyofBooleancircuitsoverabasis ,itsDirectConnectionLanguage–writtenL (C)–isthe DC setoftuples<y,g,p,b>,suchthatfor|y|=n,wehave: g isagateinC ,labeledwithb∈Bifp=ǫ, n elsebisitspth predecessor. Definition4(Uniformity[1]). AfamilyC issaidtobeDLOGTIME-uniformifthereexistsadetermin- isticTuringMachinewithrandomaccesstotheinputtapethatgivenL (C),nandg outputsintime DC O(log(|C |))anyinformation(position,labelorpredecessors)aboutthegateg inC . n n Despite thefact thata DLOGTIME TuringMachinehasmore computationalpower thanacon- stant-depthcircuit, “a consensushasdevelopedamongresearchersincircuitcomplexitythatthisD- LOGTIMEuniformityisthe‘right’uniformitycondition”forsmallcomplexityclasses[5]. Anyfurther referencetouniformityistobereadasDLOGTIMEuniformity. Definition 5 (ACi, NCi). For all i ∈N, given B a basis, a language X ⊆{0,1}∗ belongs to the class ACi(B)(resp. NCi(B))if X isacceptedbyauniformfamilyofpolynomial-size, logi-depthBoolean circuitsoverB ∪B(resp.B ∪B).Weset ACi(;)=ACi andNCi(;)=NCi. 1 0 ClémentAubert 17 2 MLLu RatherthanusingMultiplicativeLinearLogic(MLL)–whichwouldforceustocomposebinarycon- nectives toobtainsn-aryconnectives – we work with MLLuwhich differsonly on thearitiesof the →− ←− connectivesbutbetterrelatestocircuits. Wewrite A (resp. A)foranorderedsequenceofformulae A ,...,A ,(resp. A ,...,A ). 1 n n 1 Definition6(FormulaeofMLLu). GivenαaliteralandnÊ2,formulaeofMLLuare: →− ←− A::=α|α⊥| ⊗n(A)| Mn(A) DualityisdefinedwithrespecttoDeMorgan’slaw: (A⊥)⊥ ≡ A →− ←− (⊗n(A))⊥ ≡ Mn(A⊥) ←− −→ (Mn(A))⊥ ≡ ⊗n(A⊥) Asfortherestofthisarticle,considerthatA,BandDwillrefertoMLLuformulae. A[B/D]denotes AwhereeveryoccurrenceofB isreplacedbyanoccurrenceofD.WewriteA[D]ifB=α. Definition7(SequentcalculusforMLLu). AsequentofMLLuisoftheform⊢Γ,whereΓisamultiset offormulae.TheinferencerulesofMLLuareasfollow: ⊢A,A⊥ ax. ⊢Γ1,⊢A1Γ ,...,.Γ.. ,⊗n(⊢→−AΓ)n,An ⊗n 1 n ←− ⊢Γ,A ⊢∆,A⊥ ⊢Γ,A ⊢Γ,∆ cut ⊢Γ,Mn(←A−) Mn DerivationsofMLLuarebuiltwithrespecttothoserules. MLLuhasneitherweakeningnorcontrac- tion, butadmitsimplicit exchangeandeliminatescuts. Theformulae A and A⊥ in therulecut are calledthecutformulae. 3 Proofnets ProofnetsareaparallelsyntaxforMLLuthatabstractawayeverythingirrelevantandonlykeepthe structureoftheproofs. Weintroducemeasures(definition10)ontheminordertostudytheirstruc- tureandcomplexity,andaparalleleliminationoftheircuts(definition11). Definition8(Links). Weintroduceinfigure1threesortsoflinks–•,⊗n andMn –whichcorrespond toMLLurules. Everylinkmayhavetwokindsofports:principalones,indexedby0andwrittenbelow,andaux- iliaryones, indexedby1,...,n andwrittenabove. Theauxiliaryportsareordered,butaswealways representthelinksasinfigure1,wemaysafelyomitthenumbering. Axiomlinkshavetwoprincipal ports,bothindexedwith0,butwecanalwaysdifferentiatethem(forinstancebynamingthem0 and r 0 )ifweneedto. l Remark1. Thereisnosortcut:acutisrepresentedwithanedgebetweentwoprincipalports. 18 SublogarithmicuniformBooleanproofnets ...... ...... • 1 n n 1 M 0 0 N 00 Figure1: ax-link,⊗n-linkandMn-link Definition9(Decoratedderivationsandproofnet). GivenaderivationofMLLu,wedecorateitinthe followingway: – anindexisassociatedtoeveryformula. Theformulaeintroducedbyanaxiomhavethesame properindex,andtheformulaeintroducedbyalogicalrules(⊗n andMn)haveafreshindex, – adescriptionisassociatedtoeverysequent. Therulesgiveninfigure2indicatehowaproofisdecoratedandhowtobuildaproofnet froma description. ThetypeofaproofnetPisΓifthereexistsadecoratedderivationof⊢Γ⊲D(P):aproofnetalways hasseveraltypes,butuptoα-equivalence (renamingoftheliterals)wemayalwaysassumeithasa uniqueprincipaltype.IfaproofnetmaybetypedwithΓ,thenforeveryAitmaybetypedwithΓ[A]. Byextensionwewillusethenotionoftypeofanedge. Thestructuresobtainedbyfollowingthoserulesrespectcriterionofcorrectness.Forinstancetwo portsofasamelinkmaynotbeconnected,aportmaybeconnectedonlyonceandeveryauxiliary portisconnected.Wedonothavetotakeintoaccountpseudonets. Remark2. Thesameproofnet–asitabstractsderivations–maybeinducedbyseveraldescriptions. Conversely,severalgraphs–asrepresentationsofproofnets–maycorrespondtothesameproofnet: wegetroundofthisdifficultybyassociatingtoeveryproofnetasingledrawingamongthedrawings withtheminimalnumberofcrossings betweenedges, forthesakeofsimplicity. Twographsrepre- sentingproofnetsthatcanbeobtainedfromthesamedescriptionaretakentobeequal. Definition10(Sizeanddepthofaproofnet). Thesize|P|ofaproofnetP isthenumberofitslinks. Thedepthofaproofnetisdefinedwithrespecttoitstype: – Thedepthofaformulaisdefinedbyrecurrence: d(α)=d(α⊥)=1 →− ←− d(⊗n(A))=d(Mn(A))=1+max(d(A ),...,d(A )) 1 n – Thedepthd(π)ofaderivationπisthemaximumdepthofcutformulaeinit. – Thedepthd(P)ofaproofnetP is min{d(π)|πcanbedecoratedas ⊢Γ⊲D(P)forsomeΓ} The depth d(P) of a proof net depends on its type, but it is minimal when we consider the principaltypeoftheproofnet. To make themost of thecomputational power of proof nets, we needtoachieve a speed-up in thenumberofstepsneededtonormalizethem. Ifwetryroughlytoreduceinparallelacutbetween twoax-links,wearefacedwithacriticalpair.[8]avoidsthissituationbyusingatighteningreduction whicheliminatesinonestepallthecutsbetweenaxioms.Wecanthensafelyreducealltheothercuts inparallel. ClémentAubert 19 p • p ax. ⊢p:A,p:A⊥⊲ax p P ...... P 1 n {⊢Γ→−i,pi :Ai⊲D(Pi)}1ÉiÉn ⊗n p1 pn ⊢Γ1,...,Γn,s:⊗n(A)⊲tensorsp1,...,pn(D(P1),...,D(Pn)) ...... Ns ⊢Γ,p:A⊲D(P) ⊢∆,q:A⊥⊲D(Q) P Q cut p q ⊢Γ,∆⊲cutp,q(D(P),D(Q)) P p ...... p n 1 ⊢Γ,p :A ,...,p :A ⊲D(P) n ←−n 1 1 Mn ⊢Γ,s:Mn(A)⊲parpn,...,p1(D(P)) ...... s M s EdgesrepresentingΓor∆arenotdrawn,D(P)isoneofthedescriptionoftheproofnetP. Figure2:Fromdecoratedderivationstoproofnets Definition11(Cutsandparallelcut-elimination). Acutisanedgebetweentheprincipalportsoftwo links.Ifoneoftheselinksisanax-link,twocasesoccurs: iftheotherlink is an ax-link,we takethemaximalchain of ax-linksconnectedbytheirprincipal portsanddefinesthissetofcutsasat-cut, otherwisethecutisana-cut. Otherwiseitisam-cutandweknowthatfornÊ2,onelinkisa⊗n-linkandtheotherisaMn-link. Wedefineonfigure3threerewritingrulesontheproofnets.Forr ∈{t,a,m},ifQmaybeobtained fromP byerasingallther-cutsofP inparallel,wewriteP â Q. IfP â Q,P â Q orP â Q,we r t a m write P âQ. To normalize a proof net P is to apply â until we reach a cut-free proof net. â∗ is definedasthetransitivereflexiveclosureofâ. Theorem1(Parallelcut-elimination[8]). EveryproofnetPnormalizesinatmostO(d(P))applications ofâ. So thetime neededtoevaluate a proof netis relative toitsdepth, andcanbe in the worst case linearinitssize–asfortheBooleancircuits. 20 SublogarithmicuniformBooleanproofnets Forall◦∈{(Mn)nÊ2,(⊗n)nÊ2},◦maybe•in→m. • • • • → ... t • → a ...... ...... ...... ...... M → m N Figure3: t-,a-andm-reductions 4 Booleanproofnets In order to compare the complexities of proof netsandof Boolean circuits, we needto define how proof netsrepresentBoolean values(definition12)andBoolean functions(definition13). Tostudy theminauniformframeworkwedefinetheirDirectConnectionLanguage(definition14). Definition12(Booleantype,0and1[8]). Letb andb bethetwoproofnetsoftype 0 1 B=M3(α⊥,α⊥,α⊗α) respectivelyusedtorepresentfalseandtrue: • • D(b0)=parsq,p,r(tensorrp,q(axp,axq)) b0≡ M N • • D(b1)=parsp,q,r(tensorrp,q(axp,axq)) b1≡ M N →− Wewrite b forb ,...,b fori ∈{0,1}. i1 in As we can see, b and b differ on their planarity: descriptions and proof nets exhibit the ex- 0 1 changesthatwerekeptimplicitinderivations. →− Definition13(Booleanproofnets[8]). ABooleanproofnetwithninputsisaproofnetP(p)oftype ⊢p :B⊥[A ],...,p :B⊥[A ],s:⊗1+m(B[A],D ,...,D ) 1 1 n n 1 m →− →− Given b oflengthn,P(b)isobtainedbyconnectingwithcutsp tob forall1≤j ≤n. j ij ClémentAubert 21 →− P(b)â∗Q whereQ isunique,cut-freeandforsomedescriptionsQ ,...,Q describedby 1 n tensor(D(b ),Q ,...,Q )fori ∈{0,1}. i 1 m →− WewriteP(b)→ b . ev. i →− P(p)representsaBooleanfunction fn ifforallw≡i ...i ∈{0,1}n,P(b ,...,b )→ b . 1 n i1 in ev. f(w) WemayeasilydefinefamiliesofBooleanproofnetsandlanguageacceptedbyafamilyofBoolean proofnets. Thetensorindexedwithsinthetypeistheresulttensor:itcollectstheresultofthecomputation onitsfirstauxiliaryportandthegarbage–herenamedD ,...,D –onitsotherauxiliaryports. 1 m Definition14(DirectConnectionLanguageforproofnets[7]). GivenP=(Pn)n∈NafamilyofBoolean proofnets,itsDirectConnectionLanguage–writtenL (P)–isthesetoftuples<y,g,p,b>where DC for|y|=n:g isalinkinP ,ofsortbifp=ǫelsethepth premiseofg isthelinkb. n If<y,p,0,b>or<y,b,0,p>belongtoLDC(P),thereisacutbetweenbandpinC|y|. 5 Proofcircuits A proof circuit is a Boolean proof net (fact1) made out of pieces (definition15) which represents Boolean functions, constants or duplicates values. Garbage is manipulated in an innovative way, examples1shouldhelptounderstandthemechanismofcomputation. Fromnowoneveryedgerepresentedby isconnectedonitsrighttoanauxiliaryportnum- beredwithanintegerotherthan1oftheresulttensor:itcarriesapieceofgarbage. Definition15(Pieces). Wepresentinthetable1thesetofpiecesatourdisposal. Entriesarelabeled withe,exitswith s andgarbagewith g. Edgeslabeledb –fork ∈{0,1}–areconnectedtotheedge k labeledsofthepieceb . k Table1:Pieces Weset2Éj Éi. • • • • b ≡ N b ≡ N 0 M 1 M s s • b b g 0 1 • • • • • s DUPL1≡ NEG≡ M M e N e N s 22 SublogarithmicuniformBooleanproofnets Table1:Pieces–continuedfrompreviouspage i t im e s i t im e s b b b b g z 0 }| {0 z1 }| 1{ ...... ...... • s • 1 • s . 2 DUPLi ≡ .. N N • si−1 • s • • i M M e N Ifi =2,theedgeaistheedges. g e 1 1 • b 1 • • • M a gj−1 e2 N b •  1 DISJi ≡ i−3times • M • gi−1 b • 1 e j N  • M • s  ei N g e 1 1 b • 0 • • • gj−1 M a e N 2 • b  0 i−3times • M • gi−1 CONJi ≡ N • ei b0 • • s M N e j ClémentAubert 23 ApiecePwithi Ê0entries,j Ê1exitsandkÊ0garbageisoneofthepieceinthetable1,wherei edgesarelabeledwithe ,...,e , j edgesarelabeledwiths ,...,s andkedgesgototheresulttensor. 1 i 1 j WehaveP∈{b0,b1,NEG,{DUPLi}iÊ1,{DISJi}iÊ2,{CONJi}iÊ2}. TocomposetwopiecesP andP weconnectanexitofP toanentryofP .Itisnotallowedto 1 2 1 2 loop:wecannotconnectanentryandanexitbelongingtothesamepiece. Anentry(resp. anexit)thatisnotconnectedtoanexit(resp. anentry)ofanotherpieceissaidto beunconnected. →− Definition 16 (Proof circuits). A proof circuit C (p) with n inputs and one output is obtained by n composingpiecessuchthatnentriesandoneexitareunconnected. Ifnogarbageiscreatedweadd aDUPL1-piececonnectedtotheunconnectedexittoproducesomeartificially.Thenweaddaresult tensorwhosefirstedgeisconnectedtotheexit–whichisalsotheoutputoftheproofcircuit–andthe otherstothegarbage. Wethenlabeleveryunconnectedentrieswithp ,...,p : thosearetheinputs 1 n oftheproofcircuit. →− →− Given b oflengthn,C (b)isobtainedbyconnectingwithcutsp tob forall1≤j ≤n. n j ij Fact1. EveryproofcircuitisaBooleanproofnet. Proof. Weprovethisfactwithacontractibilitycriterion[2],byinductionontheheightofthepieces oftheproofcircuit(countedasthenumberofpiecesfromtheconsideredpiecetotheresulttensor). Asaproofcircuitcanalwaysbetypedwith ⊢B⊥[A ],...,B⊥[A ],⊗1+m(B[A],D ,...,D ) 1 n 1 m itisaBooleanproofnet. This fact establishes that proof circuits normalize and output a value, and that it is possible to representBooleanfunctionswiththem. Familiesofproofcircuitsandacceptationofalanguagebyafamilyofproofcircuitsaredefinedas usual. Examples1. Wepresentbrieflytwoexamplesofcomputation: thenormalizationofapieceb con- 0 nectedtoaNEG-piece(figure4,page24)andhowtheconditionalworks(figure5,page25). Condi- tionalis thecore of thecomputation, we findthispatternin everypiece except NEG andthecon- stants. Remark 3. TocomposetwoproofcircuitsC andC ,we remove theresult tensorofC ,identifythe 1 2 1 unconnected exit of C with the selected input of C , and recollect all the garbage with the result 1 2 tensorofC .Wethenlabeltheunconnectedentriesanewandobtainaproofcircuit. 2 Definition 17 (PCCi (resp. mBNi, [7])). A language X ⊆ {0,1}∗ belongs to the class PCCi (resp. mBNi) if X is accepted by a polynomial-size, logi-depth uniformfamily of proof circuits (resp. of Booleanproofnets). If we add "UstConn -pieces" to the set of pieces, we may easily define PCCi(UstConn ) and 2 2 remarkthatforalli ∈N,PCCi ⊆PCCi(UstConn ). [8]provesthatthereexistsBooleanproofnetsof 2 constant-depthandpolynomialsizethatrepresentUstConn ,sowedonotdefinemBNi(UstConn ) 2 2 becauseitwouldbeeasytoprovethatthisclassisequaltomBNi foralli ∈N. Remark4. Byfact1wehavetriviallythatforalli ∈N,PCCi ⊆mBNi. →− →− →− Lemma1. ForallproofcircuitC (p)andall b, thecutsatmaximumdepthinC (b)arebetween n n theentryofapieceandavalue(aconstantb orb ,oraninputb forsome1≤j ≤n). 0 1 ij 24 SublogarithmicuniformBooleanproofnets • • • • • • • • → • • m N N M M M N (cid:21) a • • • • ≡ N N M M Figure4:b connectedtoNEGnormalizestob 0 1 →− Proof. ForeverypieceP ofC (b)anycutconnectinganentryisalwaysofdepthsuperiororequal n tothemaximaldepthofthecutsconnectingtheexits.ThecutsofP thatdonotconnectanentryor anexitofapiecearealwaysofdepthinferiororequaltocutsconnectingtheentries. Thedepthsofthecutformulaeslowlyincreasefromtheexittotheentry,andastheentriesthat arenotconnectedtootherpiecesareconnectedtovalues,thislemmaisproved. 6 Results ByusingourproofcircuitsweproveanewtheinclusionsbetweenACi andlogicalclassesofcomplex- ityandextendthisinclusiontosublogarithmicclassesofcomplexity. Definition18(Problem:TranslationfromACi toPCCi). Input: L (C)forC afamilyofBooleancircuitsinACi. DC Output: L (C)forC afamilyofproof circuitsin PCCi, such thatfor DC →− →− alln∈N,forall b ≡b ,...,b , C (b)→ b iffC (i ,...,i ) i1 in n ev. j n 1 n evaluatesto j. Theorem2. Foralli ∈N,translationfrom ACi toPCCi belongstoAC0. Proof. ThetranslationfromC toC isobvious,itreliesoncoding: foreveryn,afirstconstant-depth circuit associatetoeverygateofC thecorrespondingpiecesimulatingitsBoolean function. Ifthe n fan-outofthisgateisk >1,aDUPLk-pieceisassociatedtotheexitofthepiece,andthepiecesare connectedasthegates.TheinputnodesareassociatedtotheinputsofC .Asecondconstant-depth n circuit recollects the only free exit and the garbage of the pieces and connects them to the result tensor. The composition of these two Boolean circuits produces a constant-depth Boolean circuit thatbuildsproofcircuits. It is easy to check thatCONJk, DISJk and NEG represent ∧k, ∨k and ¬ respectively. DUPLk duplicates a value k times, b and b represent 0 and 1 by convention. The composition of these 0 1 piecesdoesnotraiseanytrouble:C effectivelysimulatesC oneveryinputofsizen. n n

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.