2018 IEEE 59th Annual Symposium on Foundations of Computer Science
PPP-Completeness with Connections to Cryptography
Katerina Sotiraki Manolis Zampetakis Giorgos Zirdelis EECS and CSAIL EECS and CSAIL CCIS MIT MIT Northeastern University Cambridge, USA Cambridge, USA Boston, USA [email protected] [email protected] [email protected]
Abstract—Polynomial Pigeonhole Principle (PPP) is an im- the definition of FNP inadequate to capture the intrinsic portant subclass of TFNP with profound connections to complexity of total problems in the appropriate way as it the complexity of the fundamental cryptographic primitives: was first shown in [1]. Moreover, there were evidences for collision-resistant hash functions and one-way permutations. In contrast to most of the other subclasses of TFNP, no complete the hardness of total search problems e.g. in [3]. Megiddo problem is known for PPP. Our work identifies the first PPP- and Papadimitriou [4] defined the class TFNP that contains complete problem without any circuit or Turing Machine given the total search problems of FNP, and Papadimitriou [2] explicitly in the input, and thus we answer a longstanding open proposed the following classification rule of problems in question from [Papadimitriou1994]. Specifically, we show that TFNP: constrained-SIS, a generalized version of the well-known Short Integer Solution problem (SIS) from lattice-based cryptogra- Total search problems should be classified in terms phy, is PPP-complete. of the profound mathematical principles that are In order to give intuition behind our reduction for invoked to establish their totality. constrained-SIS, we identify another PPP-complete problem TFNP with a circuit in the input but closely related to lattice problems. Along these lines, many subclasses for have been We call this problem BLICHFELDT and it is the computational defined. Johnson, Papadimitriou and Yannakakis [1] defined problem associated with Blichfeldt’s fundamental theorem in the class PLS. A few years later, Papadimitriou [2] defined the theory of lattices. the complexity classes PPA, PPAD, PPADS and PPP, Building on the inherent connection of PPP with collision- resistant hash functions, we use our completeness result to each one associated with a profound mathematical princi- construct the first natural hash function family that captures ple in accordance with the above classification rule. More the hardness of all collision-resistant hash functions in a worst- recently, the classes CLS and PWPP were defined in [5] case sense, i.e. it is natural and universal in the worst-case. The and [6], respectively. In Section I-A we give a high-level close resemblance of our hash function family with SIS, leads description of all these classes. us to the first candidate collision-resistant hash function that is both natural and universal in an average-case sense. Finding complete problems for the above classes is impor- Finally, our results enrich our understanding of the con- tant as it enhances our understanding of the underlying math- nections between PPP, lattice problems and other concrete ematical principles. In turn, such completeness results reveal cryptographic assumptions, such as the discrete logarithm equivalences between total search problems, that seemed problem over general groups. impossible to discover without invoking the definition of Keywords-complexity of total search problems; pigeonhole these classes. Since the definition of these classes in [1], [2] principle; collision resistant hash functions; short integer it was clear that the completeness results about problems solutions; that do not have explicitly a Turing machine or a circuit as I. INTRODUCTION a part of their input are of particular importance. For this reason it has been established to call such problems natural The fundamental task of Computational Complexity the- in the context of the complexity of total search problems ory is to classify computational problems according to their (see [7]). inherent computational difficulty. This led to the defini- Many natural complete problems are known for PLS tion of complexity classes such as NP which contains the and PPAD, and recently natural complete problems for decision problems with efficiently verifiable proofs in the PPA were identified too (see Section I-A). However, no “yes” instances. The search analog of the class NP, called natural complete problems are known for the classes PPP, FNP, is defined as the class of search problems whose PWPP that have profound connections with the hardness of decision version is in NP. The same definition extends to important cryptographic primitives, as we explain later in the class FP, as the search analog of P. The seminal works detail. of [1], [2] considered search problems in FNP that are total, i.e. their decision version is always affirmative and thus a Our Contributions. Our main contribution is to provide the solution must always exist. This totality property makes first natural complete problems for PPP and PWPP, and
2575-8454/18/$31.00 ©2018 IEEE 148 DOI 10.1109/FOCS.2018.00023 thus solve a longstanding open problem from [2]. Beyond 1) the computational problem BLICHFELDT associated that, our PPP completeness results lead the way towards with Blichfeldt’s theorem, which can be viewed as answering important questions in cryptography and lattice a generalization of Minkowski’s theorem, is PPP- theory as we highlight below. complete, 2) the cSIS problem, a constrained version of the Short UNIVERSAL COLLISION-RESISTANT HASH FUNCTION. SIS PPP PWPP Integer Solution ( ), is -complete, Building on the inherent connection of with 3) we combine known results and techniques from lattice collision-resistant hash functions, we construct a natural H theory to show that most approximation lattice prob- hash function family cSIS with the following properties: lems are reducible to BLICHFELDT and cSIS. - Worst-Case Universality. No efficient algorithm can These results create a new path towards a better under- H find a collision in every function of the family cSIS, standing of lattice problems in terms of complexity classes. unless worst-case collision-resistant hash functions do not exist. COMPLEXITY OF OTHER CRYPTOGRAPHIC ASSUMP- Moreover, if an (average-case hard) collision-resistant TIONS. Besides lattice problems, we discuss the relationship hash function family exists, then there exists an effi- of other well-studied cryptographic assumptions and PPP. ciently samplable distribution D over HcSIS, such that Additionally, we formulate a white-box variation of the (D, HcSIS) is an (average-case hard) collision-resistant generic group model for the discrete logarithm problem [27]; hash function family. we observe that this problem is in PPP and is another natural - Average-Case Hardness. No efficient algorithm can candidate for being PPP-complete. find a collision in a function chosen uniformly at random from HcSIS, unless we can efficiently find A. Related Work short lattice vectors in any (worst-case) lattice. In this section we discuss the previous work on the The first property of HcSIS is reminiscent of the existence complexity of total search problems, that has drawn attention of worst-case one-way functions from the assumption that from the theoretical computer science community over the P = NP [8]. The corresponding assumption for the existence past decades. We start with a high-level description of the of worst-case collision-resistance hash functions is assuming total complexity classes and then discuss the known results FP = PWPP, but our hash function family HcSIS is the first for each one of them. natural definition that does not involve circuits, and admits PLS. The class of problems whose totality is established this strong completeness guarantee in the worst-case. using a potential function argument. H The construction and properties of cSIS lead us to the Every finite directed acyclic graph has a sink. first candidate of a natural and universal collision-resistant PPA. The class of problems whose totality is proved hash function family. The idea of universal constructions through a parity argument. of cryptographic primitives was initiated by Levin in [9], Any finite graph has an even number of odd-degree who constructed the first universal one-way function and nodes. followed up by [10], [11]. Using the same ideas we can PPAD. The class of problems whose totality is proved also construct a universal collision-resistant hash function through a directed parity argument. family as we describe in Appendix A. The constructed hash All directed graphs of degree two or less have an function though invokes in the input an explicit description even number of degree one nodes. of a Turing machine and hence it fails to be natural, with the PPP. The class of problems whose totality is proved definition of naturality that we described before. In contrast, through a pigeonhole principle argument. our candidate construction is natural, simple, and could have Any map from a set S to itself either is onto or practical applications. has a collision. COMPLEXITY OF LATTICE PROBLEMS IN PPP. The hard- Using the same spirit two more classes were defined after ness of lattice problems in NP ∩ coNP [12] has served as [2], in [5] and [6]. the foundation for numerous cryptographic constructions in CLS. The class of problems whose totality is established the past two decades. This line of work was initiated by using both a potential function argument and a the breakthrough work of Ajtai [13], and later developed parity argument. in a long series of works (e.g. [14], [15], [16], [17], [18], PWPP. The class of problems whose totality is proved [19], [20], [21], [22], [23], [24], [25], [26]). This wide use through a weak pigeonhole principle. of search (approximation) lattice problems further motivates Any map from a set S to a strict subset of S has their study. a collision. We make progress in understanding this important re- Recently, a syntactic analog PTFNP of the semantic class search front by showing that: TFNP has been defined in [28], and a complete problem for
149 first natural PPA-complete problem, without a circuit as part of the input, has been identified in [7]. This illustrates an interesting relation between PPA and complexity of social choice theory problems. CLS-completeness. The CLS class was defined in [5] to capture the complexity of problems such as P-matrix LCP, computing KKT-points, and finding Nash equilibria in con- gestion and network coordination games. Recently, it has been proved that the problem of finding a fixed point whose existence invokes Banach’s Fixed Point Theorem, is CLS- complete [50], [51]. TFNP and cryptography. The connection of TFNP and cryptography was illustrated by Papadimitriou in [2], where he proved that if PPP = FP then one-way permutations cannot exist. In [52], a special case of integer factorization was shown to be in PPA ∩ PPP. This was generalized in [6] by proving that the problem of factoring integers is in PPA ∩ PPP under randomized reductions. Recently, Figure 1: The classes PPP and PWPP in the TFNP world. strong cryptographic assumptions were used to prove the average-case hardness of PPAD and CLS [53], [54], [55]. In [56] it was shown that average-case PPAD hardness this class has been identified. It has also been shown that does not imply one-way function, whereas in [57] it was all the classes we described above are subsets of PTFNP. shown that any hard on average problem in NP implies Oracle separations between all these classes are known [29], the average case hardness of TFNP. Finally, in [58] it is with the only exception of whether PLS is contained in proved that the existence of multi-collision resistant hash PPAD. functions is equivalent with a variation of the total search RAMSEY PLS-completeness. The class PLS represents the complex- problem , which is not known to belong to any ity of local optimization problems. Some important problems of the above complexity classes. Interestingly, they prove RAMSEY that have been shown to be PLS-complete are: Local Max- that another variation of called colorful-Ramsey C RAMSEY PWPP Cut [30], Local Travelling Salesman Problem [31], and ( - )is -hard. Although this an important C RAMSEY Finding a Pure Nash Equilibrium [32]. Recently, important result, the problem - still invokes a circuit in PWPP results for the smoothed complexity of the Local Max-Cut the input and in not known to be in , hence does problem were shown in [33], [34]. not resolve the problem of identifying a natural complete PWPP PPAD-completeness. Arguably, the most celebrated appli- problem for . TFNP cation of the complexity of total search problems is the and lattices. In [59] it was shown that the characterization of the computational complexity of finding a computational analog of Minkowski’s theorem (namely MINKOWSKI PPP Nash equilibrium in terms of PPAD-completeness [35], [36]. )isin , was conjectured that it is also PPP This problem lies in the heart of game theory and economics. -complete. The authors justified their conjecture by EQUAL SUMS The proof that Nash Equilibrium is PPAD-complete initiated showing that - , a problem from [2] that is PPP MINKOWSKI a long line of research in the intersection of computer conjectured to be -complete, reduces to . science and game theory and revealed connections between Additionally, they show that a number theoretic prob- DIRICHLET MINKOWSKI the two scientific communities that were unknown before lem called reduces to , and PPP (e.g. [37], [38], [39], [40], [41], [42], [43], [44], [45]). thus is in . In [60] it is proven that the problem PPA PPA NUMBER-BALANCING is equivalent to a polynomial ap- -completeness. -complete problems usually arise as the undirected generalizations of their PPAD-complete proximation of Minkowski’s theorem in the 2 norm (via analogs. For example, Papadimitriou [2] showed that Cook reductions for both directions). PPAD Sperner’s Lemma in a 3-D cube is -complete and B. Roadmap of the paper later Grigni [46] showed that Sperner’s Lemma in a 3- We give here a brief description of the results contained in manifold consisting of the product of a Möbius strip and this paper. First we briefly describe the PPP-completeness a line segment is PPA-complete. Since Möbius strip is of BLICHFELDT that illustrates some of the basic ideas non-orientable, this indeed is a non-directed version of behind our main result that cSIS is PPP-complete. Then, the Sperner’s Lemma. Similarly, other problems have been we present a brief description of our main theorem and its showed to be PPA-complete, all involving some circuit as proof. We also describe the PPP-completeness of a weaker an input in their definition [47], [48], [49]. Recently, the
150 version of cSIS and its relation with the definition of the a solution to BLICHFELDT becomes trivial if the input first natural universal collision resistant hash function family representation of S has length proportional to its size, i.e. in the worst-case sense. This proof also provides the first one can iterate over all element pairs of S. The problem candidate for a collision resistant hash function family that becomes challenging when S is represented succinctly. We is both natural and universal in the average-case sense. introduce a notion for a succinct representation of sets that Finally we mention, for completeness of our exposition, we call value function. Informally, a value function for a set other lattice problems that are already known to belong S is a small circuit that takes as input log(|S|) bits that to PPP and PWPP and more general other cryptographic describe an index i ∈{0,...,|S|−1}, and outputs si ∈ S. assumptions that belong to PPP and PWPP. We give a proof overview of our first main theorem, and For the complete proofs of the statement that we present highlight the obstacles that arise, along with our solutions. in this paper we refer to the full version of our paper. Theorem II.2. The BLICHFELDT problem is PPP- II. OUR RESULTS complete. Before we describe our results in more detail we define the PPP MEMBERSHIP OF BLICHFELDT OVERVIEW. We de- class PPP more formally. The class PPP contains the set of note with [n] the set {0,...,n − 1}. We define the map problems that are reducible to the PIGEONHOLE CIRCUIT σ : Zn →P(L) ∩ Zn that reduces any point in Zn modulo problem. The input to PIGEONHOLE CIRCUIT is a binary n n n the parallelepiped to P(L) ∩ Z , i.e. (mod P(L)). Using circuit C : {0, 1} →{0, 1} and its output is either an n n n σ we can efficiently check the membership of any v ∈ Z x ∈{0, 1} such that C(x)=0, or a pair x, y ∈{0, 1} in L, by checking if σ maps v to the origin. Observe that ¯such that x = y and C(x)=¯ C(¯y). ¯ ¯ ¯ ¯ ¯ ¯ if σ(x)=σ(y) then x − y ∈L. Our first and technically most challenging result is to P L |P L ∩ Zn| identify and prove the PPP-completeness of two problems, It is well known that vol( ( )) = ( ) , hence the input requirement for S is equivalent to |S|≥|P(L) ∩ both of which share similarities with lattice problems. For n L⊆Zn Z |. Notice that the points of S after applying the map σ, our exposition, a lattice can be viewed as a finitely n Zn L L either have a collision in P(L) ∩ Z or a preimage of the generated additive subgroup of . A lattice = (B):= S B · Zn is generated by a full-rank matrix B ∈ Zn×n, called origin exists in . It follows by a pigeonhole argument that a solution to BLICHFELDT always exists. For the rest of basis. In the rest of this section we also use the fundamental |S| |P L ∩ Zn| n parallelepiped of L defined as P(L):=B · [0, 1)n. this part we assume that = ( ) and let = log(|S|). We construct a circuit C : {0, 1}n →{0, 1}n that on A. BLICHFELDT is PPP-complete. input an appropriate index i, evaluates the value function BLICHFELDT We define the problem as the computa- of S to obtain si ∈ S, and computes σ(si). The most tional analog of Blichfeldt’s theorem. challenging part of the proof is to construct an efficient map n×n σ S |P L ∩Zn| Theorem II.1 (Blichfeldt’s Theorem [61]). Let B ∈ Z from ( ) to [ ( ) ] in the following way. We define D L × L ×···× L be a set of n-dimensional linearly independent integer vec- an appropriate parallelepiped =[ 1] [ 2] [ n] n L tors and a measurable set S ⊆ R .Ifvol(S) > det(L(B)), where the i are non-negative integers, and a bijection π P L ∩ Zn → D D then there exist x, y ∈ S with x = y and x − y ∈L(B). : ( ) . Because is a cartesian product, a natural efficient indexing procedure exists as described in the full version of our paper. This allows to map π(σ(si)) BLICHFELDT Problem. n n×n to j ∈ [|P(L) ∩ Z |]. The circuit C outputs the binary INPUT:Ann-dimensional basis B ∈ Z and a set S ⊆ n decomposition of j. It follows that any x such that C(x)=0 Z described by the value function representation (s, VS). corresponds to a vector x ∈ S such that¯ σ(x)=0.Onthe¯ ¯ OUTPUT:Ifs