Venkatesan Guruswami Yuan Zhou (Carnegie Mellon University) (Boolean) Constraint Satisfaction Problems • Given: – a set of variables: V = {1, 2, 3, ..., n} – a set of values: Ω = {0, 1} – a set of "local constraints": E • Goal: find an assignment σ : V -> Ω to maximize #satisfied constraints in E • Examples (prob. name and typical local constraint) – Max-Cut: σ(i) ≠ σ(j) – Max-3LIN: σ(i)+σ(j)+σ(k) = 0/1 (mod 2) – Max-3SAT: σ(i) + σ(j) + σ(k) >= 1 (Boolean) Constraint Satisfaction Problems (cont'd) • Given: – a set of variables: V = {1, 2, 3, ..., n} – a set of values: Ω = {0, 1} – a set of "local constraints": E • Goal: find an assignment σ : V -> Ω to maximize #satisfied constraints in E • α-approximation algorithm: always outputs a solution of value at least α*OPT Approximability of some Boolean CSPs Max-Cut Approximability Random assignment >= 0.878... 1/2 [GW95] Max-3LIN < 1/2 + ε 1/2 approx. resistant 7/8 approx. resistant [Hastad01] Max-3SAT < 7/8 + ε [Hastad01] • Approximation resistant: when random assignment is the best approximation algorithm Bounded occurrence CSPs • B-bounded occurrence: each variable appears in at most B constraints • Theorem. [Hastad00] B-bounded occurrence Boolean CSPs admit (random + Ω(1/B))approximation algorithm ==> Not approximation resistant Ordering CSPs • Given: – a set of variables: V = {1, 2, 3, ..., n} – "local constraints" E, on the order of related variables • Goal: find an ordering σ : V -> [n] to maximize #satisfied constraints in E Ordering CSPs (cont'd) • Example – Maximum Acyclic Subgraph (MAS) • Constraints: for each (i, j) э E, σ(i) < σ(j) 1 2 3 4 5 5 ordering constraints, OPT = 4 Ordering CSPs (cont'd) • More Examples – Maximum Acyclic Subgraph (MAS) • Constraints: for each (i, j) э E, σ(i) < σ(j) – k-ary monotone constraint • (i1, i2, ..., ik) э E, σ(i1) < σ(i2) < ... < σ(ik) – Betweenness • (i, j, k) э E, σ(i) < σ(j) < σ(k) or σ(k) < σ(j) < σ(i) Approximability of ordering CSPs • Theorem. [GMR08, CGM09, CGHMR11] Assuming the Unique Games Conjecture, every ordering CSP is approximation resistant. • Bounded occurrence ordering CSPs? • Theorem. [Berger-Shor97] The B-bounded occurrence maximum acyclic subgraph problem admits a (1/2+Ω(1/√B))-approximation algorithm Our results • Goal. Every bounded occurrence ordering CSP is not approximation resistant (generalization of Hastad's theorem for CSPs) • Theorem. Every B-bounded occurrence monotone ordering CSP can be approximated by (1/(k!) + Ω(1/B)) – A generalization of Berger-Shor • Theorem. Every 3-ary bounded occurrence CSP is not approximation resistant Technical Part : Proof of Theorems Proof sketch • Step 1. Find t-ordering instead of full ordering – t-ordering: a mapping σt : V -> [t] n variables: 1 2 3 4 ... ... n σt: t bins: 1,3,7 random assignment full ordering: 3 < 7 < 1 < 2,5 < 2<5 < 4,6,8 < 4<8<6 • Step 2. Extend t-ordering to full ordering by random (within each bin) Proof sketch (cont'd) • Step 1. Find t-ordering instead of full ordering – t-ordering: a mapping σt : V -> [t] • Step 2. Extend t-ordering to full ordering by random (within each bin) • Problem. What kind of t-ordering do we want? (Take MAS as example,) in Step 2, constraint σ(i) < σ(j) is satisfied w.p. 1 0 1/2 when σt(i) < σt(j) w(σt(i), σt(j)) = when σt(i) > σt(j) when σt(i) = σt(j) regular CSP w( t (i ), t ( j )) with domain • Answer. To maximize ( i , j )E size t ! Proof sketch (cont'd) Ordering CSP I final ordering random t-ordering CSP It (regular CSP) (variant of) Hastad's alg. t-ordering for It • Theorem. [Hastad00] Given an B-bounded occurrence CSP instance It, there is an algorithm finding a solution of value at least rand(It) + Ω(opt(It) - rand(It))/B • Goal. Suffices to show that for some constant t, opt(It) - rand(I) = Ω(opt(I) - rand(I)) Negative news for t = 2 • Take MAS for example 1 2 opt(I) = n-1 2 bins: 3 4 ... ... n B=[n]\A A < |{i: iэA, i+1эB}|+ opt(I2) = max A |{i: i,i+1эA}|/2+|{i: i,i+1эB}|/2 <= n/2 What about t = 3 ? • Take MAS for example 1 2 3 4 ... ... n opt(I) = n-1 3 bins: 1,4,7,... < 2,5,8,... opt(I3) >= (n-1) * 2/3 < 3,6,9,... In general... • Lemma. t = 4 works for – monotone bounded occurrence ordering CSPs – every 3-ary bounded occurrence ordering CSP I.e., for any instance I from the two cases above, opt(I4) - rand(I) = Ω(opt(I) - rand(I)) • Remark. t = 3 might also work -- but we do not have a proof. Proof sketch of the lemma • Write objective value of I4 as the maximum value of a function over Boolean cube f : {-1, 1}2n -> R≥0 (encode each of the n values with 2 Boolean bits) f ( x) fˆ ( S ) S ( x) • Fourier expansion. S [ 2 n ] • Observation. rand ( I 4 ) E[ f ] fˆ ( ) • Definition. adv(f) fˆ ( S ) S • Technical Lemma. [Hastad00] If f has constant degree and is "B-occurrence bounded", there is an algorithm finding x such that f(x) = E[f] + Ω(adv(f))/B Proof sketch of the lemma (cont'd) • Technical Lemma. [Hastad00] If f has constant degree and is "B-occurrence bounded", there is an algorithm finding x such that f(x) = E[f] + Ω(adv(f))/B • Lemma. For monotone/3-ary ordering CSPs, adv(f) = Ω(opt(I) - rand(I)) • Proof. Fourier analysis, and... stare at the Fourier spectrum of the pay-off functions in the 4-ordering instances... Conclusion & open questions • It is easy to beat random assignments for many bounded occurrence ordering CSPs • Hard instances for ordering CSPs cannot be bounded occurrence • Question 1. Algorithm for all bounded occurrence ordering CSPs? • Question 2. Improve the Ω(1/B) bound? -Where Berger-Shor gets Ω(1/√B). – Maybe monotone constraint is the first step? Thanks!