Approximation Schemes via Sherali-Adams Hierarchy for Dense Constraint Satisfaction Problems and Assignment Problems Yuichi Yoshida (NII & PFI) Yuan Zhou (CMU) Constraint satisfaction problems (CSPs) • In Max-kCSP, given: – a set of variables: V = {v1, v2, v3, …, vn} – the domain of variables: D – a set of arity-k “local” constraints: C • Goal: find an assignment α : V D to maximize #satisfied constraints in C ìï üï max í å pi1,i2 ,… ,ik (a (vi1 ), a (vi2 ),… , a (vik ))ý a :V®D ïî(i1,i2 ,… ,ik )ÎC ïþ Constraint satisfaction problems (CSPs) • In Max-kCSP, given: – a set of variables: V = {v1, v2, v3, …, vn} – the domain of variables: D – a set of arity-k “local” constraints: C • Goal: find an assignment α : V D to maximize #satisfied constraints in C • Example: MaxCut , Max-3SAT, UniqueGames, … – D = {0, 1} – p(i,j) = 1[vi ≠ vj] Assignment problems (APs) • In Max-kAP, given – a set of variables V = {v1, v2, v3, …, vn} – a set of arity-k “local” constraints C • Goal: find a bijection π : V {1, 2, …, n} (i.e. permutaion) to maximize #satisfied constraints in C ìï üï max í å pi1,i2 ,… ,ik ( p (vi1 ), p (vi2 ),… , p (vik ))ý p :V®[n] ïî(i1,i2 ,… ,ik )ÎC ïþ Assignment problems (APs) • Examples – MaxAcyclicSubgraph (MAS) • π(u) < π(v) – Betweenness – – • π(u) < π(v) < π(w) or π(w) < π(v) < π(u) MaxGraphIsomorphism (Max-GI) • (π(u), π(v)) ∈ E(H), where H is a fixed graph DensekSubgraph (DkS) • (π(u), π(v)) ∈ E(Kk), where Kk is a k-clique Approximate schemes • Max-kCSP and Max-kAP are NP-Hard in general • Polynomial-time approximation scheme (PTAS): for any constant ε > 0, the algorithm runs in nO(1) time and gives (1-ε)-approximation • Quasi-PTAS: the algorithm runs in nO(log n) time • Max-kCSP/Max-kAP admits PTAS or quasi-PTAS when the instance is “dense” or “metric” PTAS for dense/metric Max-kCSP • Max-kCSP is dense: has Ω(nk) constraints. – PTAS for dense MaxCut [dlV96] – PTAS for dense Max-kCSP [AKK99, FK96, AdlVKK03] • Max-2CSP is metric: edge weight ω satisfies ω(u, v) ≤ ω(u, w)+ω(w, v) – PTAS for metric MaxCut [dlVK01] – PTAS for metric MaxBisection [FdlVKK04] – PTAS for locally dense Max-kCSP (a generalized definition of “metric”) [dlVKKV05] Quasi-PTAS for dense Max-kAP • Max-kAP is dense: – roughly speaking, the instance has Ω(nk) constraints • In [AFK02] – (1-ε)-approximate dense MAS, Betweenness in nO(1/ε^2) time – (1-ε)-approximate dense DkS, Max-GI, Max-kAP in nO(log n/ε^2) time Previous techniques • • • • • Exhaustive search on a small set of variables [AKK99] Weak Szemerédi’s regularity lemma [FK96] Copying important variables [dlVK01] A variant of SVD [dlVKKV05] Linear programming relaxation for “assignment problems with extra constraints” [AFK02] • In this paper, we show: The standard Sherali-Adams LP relaxation hierarchy is a unified approach to all these results! Sherali-Adams LP relaxation hierarchy • A systematic way to write tighter and tighter LP relaxations: [SA90] • In an r-round SA LP relaxation, – For each set S = {v1, …, vr} of r variables, we have a distribution of assignments μS = μ{v1, …, vr} – For any two sets S and T, marginal distributions are consistent: μS(S∩T) = μT(S∩T) • Solving an r-round LP relaxation takes nO(r) time. Our results • Sherali-Adams LP-based proof for known results – O(1/ε2)-round SA LP relaxation gives (1-ε)-approximation to dense or locally dense Max-kCSP, and Max-kCSP with global cardinality constraints such as MaxBisection – O(log n/ε2)-round SA LP relaxation gives (1-ε)approximation to dense or locally dense Max-kAP • New algorithms – Quasi-PTAS for Maxk-HypergraphIsomorphism when one graph is dense and the other one is locally dense Our techniques • Solve the Sherali-Adams LP relaxation for sufficiently many rounds (Ω(1/ε2) or Ω((log n)/ε2)) • Randomized conditioning operation to bring down the pair-wise correlations • Independent rounding for Max-kCSP • Special rounding for Max-kAP Conditioning operation • Randomly choose v from V, sample a ~ μv • For each local distribution μ{v1, …, vr}, generate the new local distribution μ{v1, …, vr}|v=a • r-round SA solution (r-1)-round SA solution • Essentially from [RT12]: – After t steps of conditioning, 1 – on average, μ{v1, …, vk} is only -far from μ{v1} x … x μ{vk} t Independent rounding for Max-kCSP After Ω(1/ε2) steps of conditioning, on average, μ{v1, …, vk} is only ε-far from μ{v1} x … x μ{vk} Sample each v from μ{v}, and we have éë p(v ,… ,v ) (v1,… , vk )ùû = éë p(v ,… ,v ) (v1,… , vk )ùû ± e E 1 k 1 k (v1,… ,vk )~m{v ,… ,v } (v1,… ,vk )~m{v } ´… ´m{v } E 1 k 1 k Therefore, E [ rounding solution] = [ LP value] ± e × nk This is a (1-O(ε))-(multiplicative) approximation because of the density Rounding for Max-kAP • Independent sampling does not work: – objective value is good, but resulting assignment might not be permutation because of collisions • Our special rounding: – View {μ{v}(w)}v,w as a doubly stochastic matrix, therefore a distribution of permutations – Distribution supported on one permutation ✔ Similar operation – Two permutations? Merge them in [AFK02] – Even more permutations? Pick arbitrary two, merge them, and iterate Merging two permutations 1. View the two permutations as disjoint cycles 2. Break long cycles (length > n1/2) into short ones (length ≤ n1/2) 3. In each cycle, choose Permutation 1/Permutation 2 independently Analysis • Step 2: modified O(n1/2) entries of Permutation 2, affecting O(n-1/2)fraction of the constraints n1/2 Merging two permutations 1. View the two permutations as disjoint cycles 2. Break long cycles (length > n1/2) into short ones (length ≤ n1/2) 3. In each cycle, choose Permutation 1/Permutation 2 independently Analysis • Step 3: value of the constraints where each variable from a distinct cycle is preserved because of independence – all but n-1/2-fraction of them n1/2 Merging two permutations 1. View the two permutations as disjoint cycles 2. Break long cycles (length > n1/2) into short ones (length ≤ n1/2) 3. In each cycle, choose Permutation 1/Permutation 2 independently Analysis • Conclusion: In this way, we get a permutation whose objective value is at least (1 – O(n-1/2)) * [Indep. Sampling] ≥ (1 – O(n-1/2)) (1 – O(ε)) [Val of LP] n1/2 Future directions • Can we solve the Sherali-Adams LP faster (as in [GS12]) to get a PTAS for dense assignment problems? Thanks