Venkatesan Guruswami (CMU) Yury Makarychev (TTI-C) Prasad Raghavendra (Georgia Tech) David Steurer (MSR) Yuan Zhou (CMU) MaxCut and Goemans-Williamson alg. G = (V, E) A Objective: edges( A, B) max |E| B • The GW SDP relaxation [GW95] 1 vi , v j 1 max | E | (i , j )E 2 subject to – 0.878-approximation – (1 ) vs (1 ) approximation || vi ||2 1, i V Finding almost-perfect MaxCut • (1 ) vs (1 ) approximation – Bipartite graph recognition algorithm (robust version against noise) – Optimal under Unique Games Conjecture [KKMO07, MOO10] MaxBisection G = (V, E) A Objective: B | V | even edges( A, B) max |E| | A || B | • Approximating MaxBisection? – No easier than MaxCut • Reduction: take two copies of the MaxCut instance MaxBisection (cont'd) G = (V, E) A Objective: B | V | even edges( A, B) max |E| | A || B | • Approximating MaxBisection? – No easier than MaxCut – Strictly harder than MaxCut? – Approximation ratio: 0.7028 [FJ97, Ye01, HZ02, FL06] – Approximating almost perfect solutions? Not known Finding almost-perfect MaxBisection • Question – Is there a (1 ) vs (1 o (1)) approximation algorithm for MaxBisection? • Answer. Yes. • Our result. – Theorem. There is a (1 ) vs (1 1/ 3 log 1 ) approximation algorithm for MaxBisection. – Theorem. Given a (1 ) satisfiable MaxBisection instance, it is easy to find a (.49, .51)-balanced cut of value (1 ) . The rest of this talk... • Theorem. There is a (1 ) vs (1 1/ 20 log n) approximation algorithm for MaxBisection. Approach Approach -- SDP • The standard SDP (used by all the previous algorithms) 1 vi , v j 1 max | E | (i , j )E 2 • Integrality gap OPT < 0.9 , subject to || vi ||2 1, i V v iV SDP = 1 i 0 Our approach A simple fact • Fact.(1 / 2 ,1 / 2 ) -balanced cut of value c of value c 2 . bisection • Proof. Get the bisection by moving fraction of random vertices from the right side to the left side. • Only need to find almost bisections. Almost perfect MaxCuts on expanders • λ-expander: for each S V, such that vol( S ) vol(V ) / 2 , we have edges( S , V S ) , where vol( S ) di iS vol( S ) • Key Observation. The (volume of) difference between two (1 ) cuts on a λ-expander is at most 2 / vol(V ) . • Proof. cut( A, B) 1 cut(C , D) 1 C X edges( X Y ,V X Y ) 2 vol(V ) A B vol( X Y ) 2 / vol(V ) Y D Almost perfect MaxCuts on expanders • λ-expander: for each S V, such that vol( S ) vol(V ) / 2 , we have edges( S , V S ) , where vol( S ) di iS vol( S ) • Key Observation. The (volume of) difference between two (1 ) cuts on a λ-expander is at most 2 / vol(V ) . • Approximating almost perfect MaxBisection on expanders is easy. – Just run the GW alg. to find the MaxCut. The algorithm (sketch) • Decompose the graph into expanders – Discard all the inter-expander edges • Approximate OPT's behavior on each expander by finding MaxCut (GW) – Discard all the uncut edges • Combine the cuts on the expanders – Take one side from each cut to get an almost bisection. (subset sum) Expander decomposition • Cheeger's inequality. Can (efficiently) find a cut of sparsity if the graph is not a -expander. • Corollary. A graph can be (efficiently) decomposed into -expanders by removing log n edges (in fraction). • Proof. – If the graph is not an expander, divide it into two parts by sparsest cut (cheeger's inequality). – Process the two parts recursively. The algorithm • Decompose the graph into 1/ 20 log n edges. – Lose 1/10-expanders. • Apply GW algorithm on each expander to approximate OPT. – OPT(MaxBisection) = (1 ) – GW finds (1 ) cuts on these expanders 1/ 2 1/ 10 1/10 different from behavior of OPT • / – Lose 1 / 2 edges. • Combine the cuts on the expanders (subset sum). • • ( 2 1 1/10 1 , 2 1/ 20 -balanced cut of value ( 1 log n) ) a bisection of value (1 1/ 20 log n) 1/10 Eliminating the log n factor • Another key step. • Idea. Terminate early in the decomposition process. Decompose the graph into 1/10-expanders or subgraphs of n vertices. • Corollary. Only need to discard 1/ 20 log 1 edges. • Lemma. We can find an almost bisection if the MaxCuts for small sets are more biased than those in OPT. Finding a biased MaxCut • Lemma. Given G=(V,E), if there exists a cut (X, Y) of value (1 ) , then one can find a cut (A, B) of value (1 ) , such that | A || X | | V .| • SDP. 1 maximize subject to v ,v |V | iV 0 i 1 vi , v j 1 1 | E | ( i , j )V 2 || vi ||2 1 2 2 -triangle inequality | vi v j , v0 | , i {0} V || vi v j ||2 2 , i, j V • Rounding. A hybrid of hyperplane and threshold rounding. Future directions • (1 ) vs (1 ) approximation? • "Global conditions" for other CSPs. – Balanced Unique Games? The End. Any questions?