On Interactive Proofs with a Laconic Prover
Total Page:16
File Type:pdf, Size:1020Kb
ON INTERACTIVE PROOFS WITH A LACONIC PROVER Oded Goldreich Salil Vadhan and Avi Wigderson Abstract We continue the investigation of interactive pro ofs with b ounded communication as initiated by Goldreich and Hastad IPL Let L be a language that has an interactive pro of in which the prover sends few say b bits to the verier We prove that the com plement L has a constantround interactive pro of of complexity that dep ends only exp onentially on b This provides the rst evidence that for NPcomplete languages we cannot exp ect interactive provers to b e much more laconic than the standard NP pro of When the pro of system is further restricted eg when b or when we have p erfect completeness we get signicantly b etter upp er b ounds on the complex ityofL Keywords Interactive Pro of systems ArthurMerlin games NP sam pling proto cols statistical zeroknowledge game theory Sub ject classication Q Q Q A Intro duction Interactive pro of systems were intro duce by Goldwasser Micali and Rack o GMR in order to capture the most general way in which one party can eciently verify claims made by another more powerful party That is interactive pro of systems are twoparty randomized proto cols through which a computationally unbounded prover can convince a probabilistic p olynomial time verier of the memb ership of a common input in a predetermined lan guage Thus interactive pro of systems generalize and contain as a sp ecial case the traditional NPpro of systems in which verication is deterministic and noninteractive It is wellknown that this generalization buys us a lot The IP Characteriza tion Theorem of Lund Fortnow Karlo Nisan and Shamir LFKN Sha ArthurMerlin games intro duced by Babai Bab are a sp ecial typ e of interactive pro ofs in which the verier is restricted to send the outcome of eachcoinittosses Suchproof systems are also called public coin and are known to b e as expressive as general interactive pro ofs GS Wewarn that the latter assertion refers to the entire class but not to rened complexity measures such as the total number of bits sentby the prover considered b elow Goldreich Vadhan Wigderson states that every language in PSPACE has an interactive pro of system and it is easy to see that only languages in PSPACE haveinteractive pro of systems It is wellknown that the strong expressive power of interactive pro ofs is largely due to the presence of interaction In particular interactive pro ofs in which a single message is sent like in NPpro ofs yield a complexity class known as MA that seems very close to NP It is interesting to explore what happ ens between these extremes of unb ounded interaction and no interaction That is what is the expressive power of interactive proofs that utilize a bounded but nonzero amount of interaction Prior work regarding interactive pro ofs with b ounded interac tion Interactive Pro ofs with Few Messages The earliest investigations of the ab ove question examined the message complexity of interactive pro ofs ie the number of messages exchanged Sometimes we refer to rounds which are a pair of verierprover messages The Sp eedup Theorem of Babai and Moran BM together with GS shows that the number of messages in an interactive pro of can be always reduced by a constant factor provided the numb er of messages remains at least On the other hand there is a large gap between constantround interactive pro ofs and unrestricted interactive pro ofs As mentioned ab ove all of PSPACE has a general interactive pro of LFKN Sha In contrast the class AM of problems with constantround interactive pro ofs is b elieved to be relatively close to NP Sp ecically AM lies in the second level of the p olynomialtime hierarchy BM cannot contain coNP unless the p olynomialtime hierarchy collapses BHZ and actually equals NP under plausible circuit complexity assumptions AK KvM MV Laconic Provers A more rened investigation of the ab ove question was initiated by Goldreich and Hastad GH who gave b ounds on the complex ity of languages p ossessing interactive pro ofs with various restrictions on the number of bits of communication andor randomness used One of the re strictions they considered and the main fo cus of our investigation limits the number of bits sent from the prover to the verier by some bound b That is what languages can b e proven by laconic provers Since the prover is trying to convey something to the verier this seems to b e the most interesting direction of communication Moreover for applications of interactive pro ofs eg in cryptographic proto cols it mo dels the common On Interactive Pro ofs with a Laconic Prover situation in which communication is more exp ensive in one direction eg if the prover is a handheld wireless device On one hand we know of interactive pro ofs for several hard problems eg Quadratic Nonresiduosity GMR Graph Nonisomorphism GMW and others GKGG SV in which the communication from the prover to the verier is severely b ounded in fact to one bit On the other hand NP laconic provers exist only for problems in BPP resp BPP in case the pro of system is of the publiccoin typ e GH Furthermore it was conjec tured that NPcomplete problems cannot have general interactive pro ofs with laconic provers but the results in GH fall short of supp orting this conjec ture In this work we provide strong supp ort for this conjecture New results regarding interactive pro ofs with b ounded interac tion Our main fo cus is on laconic provers that is on interactive pro ofs in which the total number of bits sent by the prover is b ounded Laconic Provers Consider interactive pro ofs in which the prover sends at most b bn bits to the verier on inputs of length n Goldreich and NP Hastad GH Thm placed such languages in BPTIMET where polyb T p olyn which clearly implies nothing for languages in NP In contrast weshow that the complements of such languages have constantround interactive pro ofs of complexity T ie the veriers computation time and the total communication is bounded by T In particular NPcomplete problems cannot have interactive pro ofs in which the prover sends p olylogarithmically many bits to the verier unless coNP is in the quasip olynomial analogue of AM In fact assuming NP has constantround interactive pro ofs with logarithmic provertoverier communication we conclude coNP AM As mentioned ab ove this is highly unlikely We obtain stronger results in two sp ecial cases We show that if a language has an interactive pro of of p erfect complete ness ie zero error probability on yes instances in which the prover sends at most bn bits then it is in coNTIMET where T n bn p olyn Thus unless NP coNP NPcomplete languages can not have interactive pro of systems of p erfect completeness in which the prover sends logarithmically many bits We show that if a language has an interactive pro of in which the prover sends a single bit with some restrictions on the error probabilities then it has a statistical zeroknowledge interactive pro of that is is in the Goldreich Vadhan Wigderson class SZK This is a stronger conclusion than our main result b ecause SZK AM coAM as shown by Fortnow For and Aiello and Hastad AH Recalling that Sahai and Vadhan SV showed that any language in SZK has an interactive pro of in which the prover sends a single bit we obtain a surprising equivalence b etween these two classes Interactive Pro ofs with Few Messages We obtain one apparently new result regarding message complexity A question that is left op en by the re sults mentioned earlier is what happ ens in between constant rounds and p olynomially many rounds Phrased dierently can the Sp eedup Theorem of Babai and Moran be improved to show that mnmessage interactive pro ofs can be emulated by and hence are no more p owerful than m nmessage in teractive pro ofs for some m om By combining careful parameterizations of LFKN Sha and BM we observethatsuchanimprovementspeedup is unlikely More precisely for every nice function m we show that there is a language which has an mnmessage interactive pro of but not an omn message one provided that SAT is not contained in the subexp onential ana logue of coAM Additional related work We note that GoldreichandHastad GH have presented signicantly stronger results regarding interactive pro ofs with laconic provers when further restrictions are imp osed on the interactive pro of NP In particular they obtain an upp er b ound of BPTIMET rather than BPTIMET polyb with T polyn for languages p ossessing either of the following kinds of interactive pro ofs publiccoin pro ofs in which the prover sends at most b bits pro ofs in which the communication in both directions is b ounded by b Multiprover interactive pro ofs and PCP The expressivepower of multi prover interactive proofs MIPs and probabilistical ly checkable proofs PCPs with low communication has b een the fo cus of extensive research Much of this research is motivated by the imp ortance of the communication parameter in the applications of MIPPCP to inapproximability In particular Bellare Goldreich and Sudan BGSgive negative results ab out the expressivepower of laconic PCPs and MIPs Since onequery PCPs are equivalent to inter active pro ofs in which the prover sends a single message our results provide bounds on the former On Interactive Pro ofs with a Laconic Prover Knowledge complexityofinteractive pro ofs Our work is also related to work on know ledge complexity Knowledge complexity prop osed by GMR aims to measure how much knowledge is leaked from the prover to the ver ier in an interactive pro of Several