Yuan Zhou, Ryan O’Donnell Carnegie Mellon University Constraint Satisfaction Problems • Given: – a set of variables: V – a set of values: Ω – 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 Example 1: Max-Cut • Vertex set: V = {1, 2, 3, ..., n} • Value set: Ω = {0, 1} • Typical local constraint: (i, j) in E wants σ(i) ≠ σ(j) • Alternative description: – Given G = (V, E), divide V into two parts, – to maximize #edges across the cut • Best approx. alg.: 0.878-approx. [GW'95] • Best NP-hardness: 0.941 [Has'01, TSSW'00] Example 2: Balanced Separator • Vertex set: V = {1, 2, 3, ..., n} • Value set: Ω = {0, 1} • Minimize #satisfied local constraints: (i, j) in E : σ(i) ≠ σ(j) • Global constraint: n/3 ≤ |{i : σ(i) = 0}| ≤ 2n/3 • Alternative description: – given G = (V, E) – divide V into two "balanced" parts, – to minimize #edges across the cut Example 2: Balanced Saperator (cont'd) • Vertex set: V = {1, 2, 3, ..., n} • Value set: Ω = {0, 1} • Minimize #satisfied local constraints: (i, j) in E : σ(i) ≠ σ(j) • Global constraint: n/3 ≤ |{i : σ(i) = 0}| ≤ 2n/3 • Best approx. alg.: sqrt{log n}-approx. [ARV'04] • Only (1+ε)-approx. alg. is ruled out assuming 3SAT does not have subexp time alg. [AMS'07] Example 3: Unique Games • Vertex set: V = {1, 2, 3, ..., n} • Value set: Ω = {0, 1, 2, ..., q - 1} • Maximize #satisfied local constraints: {(i, j), c} in E : σ(i) - σ(j) = c (mod q) • Unique Games Conjecture (UGC) [Kho'02, KKMO'07] No poly-time algorithm, given an instance where optimal solution satisfies (1-ε) constraints, finds a solution satisfying ε constraints • Stronger than (implies) "no constant approx. alg." Open questions Is UGC true? Is Max-Cut hard to approximate better than 0.878? Is Balanced Separator hard to approximate with in constant factor? Easier questions Do the known powerful optimization algorithms solve UG/Max-Cut/Balanced Separator? SDP Relaxation hierarchies • A systematic way to write tighter and tighter SDP relaxations BASIC-SDP r-round SDPOrelaxation (r ) in roughly n time ? … UG(ε) ARV SDP for Balanced Separator GW SDP for Maxcut (0.878-approx.) • Examples – Sherali-Adams+SDP [SA'90] – Lasserre hierarchy [Par'00, Las'01] How many rounds of tighening suffice? • Upperbounds (1 ) – n rounds of SA+SDP suffice for UG [ABS'10, BRS'11] • Lowerbounds [KV'05, DKSV'06, RS'09, BGHMRS '12] (also known as constructing integrality gap instances) (1) – exp((log log n) ) rounds of SA+SDP needed for UG (1) exp((log log n ) ) rounds of SA+SDP needed – for better-than-0.878 approx for Max-Cut (1) (log log n ) – rounds for SA+SDP needed for constant approx. for Balanced Separator From SA+SDP to Lasserre SDP • Are the integrality gap instances for SA+SDP also hard for Lasserre SDP? • Previous result [BBHKSZ'12] – No for UG – 8-round Lasserre solves the Unique Games lowerbound instances From SA+SDP to Lasserre SDP (cont’d) • Are the integrality gap instances for SA+SDP also hard for Lasserre SDP? • This paper – No for Max-Cut and Balanced Separator – Constant-round Lasserre gives better-than0.878 approximation for Max-Cut lowerbound instances – 4-round Lasserre gives constant approximation for the the Balanced Separator lowerbound instances Proof overview • Integrality gap instance – SDP completeness: good vector solution – Integral soundness: no good integral solution • Show the instance is not integrality gap instance for Lasserre SDP – no good vector solution – we bound the value of the dual of the SDP – interpret the dual as a proof system (”SOSd/sum-of-squares proof system") – lift the soundness proof to the proof system What is the SOSd proof system? Polynomial optimization • Maximize/Minimize p (x ) • Subject to q1 ( x) 0, q2 ( x) 0, qm ( x) 0 r1 ( x) 0, r2 ( x) 0, rm' ( x) 0 all functions are low-degree n-variate polynomials • Max-Cut example: 2 Maximize (i,j)EE ( xi x j ) s.t. xi (1 xi ) 0, i Polynomial optimization (cont'd) • Maximize/Minimize p (x ) • Subject to q1 ( x) 0, q2 ( x) 0, qm ( x) 0 r1 ( x) 0, r2 ( x) 0, rm' ( x) 0 all functions are low-degree n-variate polynomials • Balanced Separator example: Minimize E ( xi x j ) 2 (i,j)E s.t. xi (1 xi ) 0, i E[ xi ] i 1 2 , E [ x ] 3 3 i i Certifying no good solution • Maximize • Subject to p (x ) q1 ( x) 0, q2 ( x) 0, qm ( x) 0 r1 ( x) 0, r2 ( x) 0, rm' ( x) 0 • To certify that there is no solution better than , simply say that the following equalities & inequalities are infeasible p(x) q1 ( x) 0, q2 ( x) 0, qm ( x) 0 r1 ( x) 0, r2 ( x) 0, rm' ( x) 0 The Sum-of-Squares proof system • To show the following equalities & inequalities are infeasible, q1 ( x) 0, q2 ( x) 0, qm ( x) 0 r1 ( x) 0, r2 ( x) 0, rm' ( x) 0 • Show that 1 f ( x ) q ( x ) h( x ) i 1... m i i • where h(x) is a sum of squared polynomials, including ri (x)'s • A degree-d "Sum-of-Squares" refutation, where d max {deg( f i ) deg( qi ), deg( h)} i Positivstellensatz Subject to some mild technical conditions, every infeasible system has such a refutation Caveat: fi’s and h might need to have high degree. Lasserre SDP and SOSd proof system A degree-d SOS refutation O(d)-round Lasserre SDP is infeasible Summary • Defined the degree-d SOS proof system • Remaining task Integral soundness proof low-degree refutation in the SOS proof system Example 1 • To refute x2 x (1 x ) 0 • We simply write 1 x(1 x) ( x 2) ( x 1) 2 • A degree-2 SOS refutation One-slide How-to Thm: Max-Cut of this graph is ≤ blah Proof: … Invariance Principle … … Majority-Is-Stablest… “Check out these polynomials.” Thm: Min-Balanced-Separator in this graph is ≥ blah Proof: … hypercontractivity… “Check out these polynomials.” Example 2: Max-Cut on triangle graph • To refute ( x1 x2 ) 2 ( x2 x3 ) 2 ( x3 x1 ) 2 2 x1 (1 x1 ) 0, x2 (1 x2 ) 0, x3 (1 x3 ) 0 • We "simply" write ... ... Example 2: Max-Cut on triangle graph (cont'd) ( x1 x2 ) 2 ( x2 x3 ) 2 ( x3 x1 ) 2 2 ( x1 x2 x2 x3 x1 x3 x2 ) 2 ( x1 x2 1) 2 ( x2 x3 1) 2 x1 (1 x1 )( x22 x32 2 x2 x3 1) x2 (1 x2 )( x1 x32 2 x1 x3 2 x1 2 x3 3) x3 (1 x3 )( x1 x2 2 x1 x2 1) • A degree-4 SoS refutation Latest results • Our theorem on Max-Cut is improved by [DMN’12] – Constant-round Lasserre SDP almost exactly solves the known instances • Constant-round Lasserre SDP solves the hard instances for Vertex-Cover [KOTZ’12] Open question • Does constant-round Lasserre SDP solve the known instances for all the CSPs? – I.e. SOS-ize Raghavendra’s theorem. Thank you!