On Helping and Interactive Pro of Systems V Arvind J Kobler R Schuler Department of Computer Science Theoretische Informatik Institute of Mathematical Sciences Universitat Ulm Madras India D Ulm Germany Abstract Weinvestigate the complexity of honest provers in interactive pro of systems This corresp onds precisely to the complexity of oracles helping the computation of robust probabilistic oracle machines We obtain upp er b ounds for languages in FewEXP and for sparse sets in NPFurther interactive proto cols with provers that are reducible to sets of low information content are considered Sp ecicallyiftheverier communicates p only with provers in Pp oly then the accepted language is lowfor In the case that the provers are p olynomialtime reducible to log sparse sets or to sets in strongPlog then the proto col can b e simulated bytheverier even without the help of provers As a consequence we obtain new collapse results under the assumption that intractable sets reduce to sets with low information content Intro duction and overview of results Two extensions of the concept of NP as the class of languages with ecient pro ofs of memb ership Co o namely the class IP of languages that have single prover interactive pro of proto cols and the class MIP of languages with multiprover interactiv e pro of proto cols formal denitions are given in the next section havebeenshown to characterize the complexity classes PSPACE and NEXP resp ectively Sh BFL In the interactive pro of system mo del of computation GMR one or more provers try to convince a probabilistic p olynomial time verier that a given input is a memb er of the considered language The proto col has the following b ehavior if an input is in the language then there are provers that convince the verier to accept with high probability On the other hand if the input is not in the language then no set of provers can convince the verier with more than negligible probability Arthur Merlin games intro duced in Bab are sp ecial cases of singleprover interactive pro of systems in which the random choices of the verier called Arthur are public and therefore visible to Merlin the prover In Bab it is shown that for any constant k ak round interaction AMk can b e substituted bya p two round proto col AM AM and that AM is contained in the second level of the p olynomial time hierarchy Here AMk denotes a k round interaction of messages b etween Arthur and Merlin where the rst message is from Arthur The class MA of languages Part of the work was done while visiting Universitat Ulm Supp orted in part by an Alexander von Humb oldt research fellowship accepted byatwo round proto col starting with a message from Merlin is an apparently smaller class than AM It turned out that the apparently restricted Arthur Merlin games mo del is as p owerful as the general interactive pro of systems In fact anyinteractive pro of system can b e simulated byanArthur Merlin game GS BM with only a constant increase in the numb er of rounds In general the notation AM is used to denote singleprover interactive pro of systems with a constantnumber of interaction whereas IP is used if there is no restriction Following the advances made in LFKN it was nally shown in Sh that PSPACE IP Subsequentlyitwas shown that multiprover interactive pro of systems can accept all of NEXP BFL Intuitively the multiprover mo del is more p owerful since the verier can use one prover to check that the answers of another prover do not dep end on previous queries made to him This is made precise in the pro of of Babai et al of the ab ove mentioned result where it is shown that a p olynomial number of provers can b e substituted bytwoprovers One prover answers all the queries of the original proto col and the second is only used to verify that the answers do not dep end on the history of interaction teractive pro of systems which is also the main An interesting issue that arises in in concern of this pap er is b ounding the complexity of the honest provers that make the verier accept a given language For example it is known Fe Sh that each language in PSPACE has an IP proto col with prover of complexity PSPACE The result of Lund PP et al LFKN implies that P has IP proto cols with prover complexity PP Similarly EXP has MIP proto cols of prover complexity EXP BFL Let IPC and MIPC denote resp ectively the class of languages with prover complexity b ounded by FPC ie the class of functions computable by p olynomialtime b ounded transducers with access to an oracle from the class C In this notation we summarize some of the known results on upp er b ounds for prover complexity Theorem Fe Sh PSPACE IPPSPACE PP LFKNP IPPP BF P IP P NP BFL NEXP MIPEXP BFL EXP MIPEXP A surprising application of Theorem to classical structural complexity is the follo wing corollary Corollary LFKN BFL For KfPP PSPACE EXPgifKPp oly then K MA p The ab ove result improves previously known collapses to under the same assumption obtained by Meyer Karp and Lipton KL This is particularly interesting since it is p under the assumption NP Pp oly KL cannot b e known that the collapse of PH to p improved to with relativizable pro of metho ds Hel and the LFKN proto col is known to b e not relativizable FS FRS The crux of the pro of of Corollary is as follows if K is contained in Pp oly then by Theorem every language in K has an MIP proto col in which the provers are p olynomial size circuits This proto col can b e simulated by an MA proto col where Merlin simply sends the circuits for the provers to Arthur in the rst round LFKN BFL We observe that the MA proto col in the ab ove pro ofsketch can b e seen as a single prover proto col with prover complexityinPpoly Thus we can identify the classes MIPPp oly and IP Ppoly of interactive pro of systems in which the provers are p olynomial size circuits MIPPp oly IPPp oly MA NP On the other hand in the case of NEXP the NEXP MIPEXP characterization do es not app ear to directly imply any signicant collapse of NEXP under the assumption NEXP Pp oly It is due to the very high upp er b ound on the prover complexityA related result has b een proved by Buhrman and Homer BH who showed that under a p NP NP stronger assumption EXP Pp oly it holds that EXP An imp ortant motivation of the present pap er is to explore further applications of characterizations of complexity classes using interactive pro of systems in order to derive new collapse consequences assuming that the considered class has p olynomial size circuits We note that in Babai et al BFL it is an op en question as to whether the upp er NP b ound of EXP for the prover complexity for NEXP can b e improved The reason for the apparently weak upp er b ound is that all provers must have access to the same tableau of an accepting NEXP computation To wit the high prover complexity is the cost of disambiguating down to one consistent NEXP computation As a rst step weshow that FewEXP languages accepted by NEXP machines that FewEXP have at most exp onentially many accepting paths are contained in MIPNP FewEXP Theorem FewEXP MIPNP FewEXP NP It is to b e noted that NP is a substantially smaller class of provers than EXP FewEXP NEXP Indeed since NP is known to b e contained in P Hem it follows that FewEXP MIPNEXP As a consequence of Theorem we can extend Corollary to the class K FewEXP Corollary If FewEXP Pp oly then FewEXP MA MIPPp oly As observ ed bySchoning Sch there is a close relationship b etween the complexityof provers in interactive pro of systems and the complexity of oracles helping the computation of robust oracle machines A machine M is called robust if LM LM Bforall oracles B and an oracle A onesided helps the computation of M if M with oracle A runs in p olynomial time on all inputs x resp x LM The notion of helping by robust deterministic oracle machines was intro duced bySchoning Sch and was extended to onesided helping byKoKo In a sense the notion of helping by robust machines was a forerunner to the concept of interactive pro of systems Indeed it follows from the probabilistic oracle machine characterization of MIP in FRS that the class MIPC coincides with the class BPP C of languages accepted by probabilistic robust help machines with the onesided help of an oracle from C With this interpretation Theorem FewEXP shows that NP oracles are able to give onesided help to the computation of languages in FewEXP On the other hand the inclusion MIPPp oly MA can b e interpreted as stating that oracles in Pp oly cannot help the computation of languages outside MA An interesting op en question is whether Pp oly provers are useful at all ie whether MIPPp oly contains sets that are not in BPP As follows from the next theorem settling this question is likely to b e very hard since a negative answer would imply that all sparse NP sets are already in BPP which in turn implies that NE is contained in BPE known to b e false in relativized worlds Theorem All sparse NP sets are contained in P Ppoly help In Koitisshown that every set A accepted by a robust deterministic oracle machine p with the onesided help of an oracle from Pp oly ie A P Pp oly is lowfor help Since P Ppoly MIPPp oly the following theorem extends b oth the ab ove help p mentioned result of Ko and the lowness of BPP for Sip ZHSch p Theorem MIPPp oly is low for Considering MIP proto cols with provers reducible to sets of very low information content eg to sets in strongPlog or to log sparse sets we show that such provers cannot help the