1 / 20

A Linear Round Lower Bound for Lovasz-Schrijver SDP relaxations of Vertex Cover

A Linear Round Lower Bound for Lovasz-Schrijver SDP relaxations of Vertex Cover. Grant Schoenebeck Luca Trevisan Madhur Tulsiani UC Berkeley. Integer Optimum. Optimum of Program. Minimum Vertex Cover.

lindsey
Download Presentation

A Linear Round Lower Bound for Lovasz-Schrijver SDP relaxations of Vertex Cover

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Linear Round Lower Bound for Lovasz-Schrijver SDP relaxations of Vertex Cover Grant Schoenebeck Luca Trevisan Madhur Tulsiani UC Berkeley

  2. Integer Optimum Optimum of Program Minimum Vertex Cover • For G = (V,E) find the smallest subset of vertices containing at least one endpoint of every edge. • Integrality Gap = MaxG = 2 – o(1) for both LP Minimize u2V xu xu2 [0,1] 8 u 2V xu + xv¸ 1 8 (u,v) 2 E SDP (Lovasz -function) Minimize u2Vz0¢zu k zuk 1 8u 2 V, k z0k= 1 (z0 - zu)¢(z0 - zv) = 0 8 (u,v) 2 E

  3. Integrality Gaps with more constraints • For the complete graph (Kn) on n vertices: LP Optimum = n/2; Integer Optimum = n-1 Integrality Gap = 2 – 2/n • What if we add the constraint xu+xv+xw¸ 2 for every triangle (u,v,w) in G? Performance ratio for Kn is 3/2, but the integrality gap still remains 2-o(1). • What if add constraints analogous to the one above for every odd cycle? What if we add xu + xv + xw + xz 3 for every clique of size 4? Size 5? • One needs to prove integrality gaps from scratch every time new constraints are added.

  4. Automatically generating “natural” constraints • LS/LS+ hierarchies define define “cut operators” applied to (convex) solution space. • Operators can be iteratively applied to generate tighter LP/SDP relaxations. • Relaxation obtained by r cuts (rounds) solvable in nO(r) time. • Constant number of rounds produce most known LP/SDP relaxations.

  5. y y1z(1) y2z(2)  ynz(n) The Lovasz-Schrijver Hierarchy • Goal: Only allow convex combinations of 0/1 solutions. Probability distributions! y = (y1, y2, …, yn)  S Marginal distribution! Ask for conditionals. z(n) = (z(n)1, z(n)2, …, 1) w(n) = (w(2)1, w(n)2, …, 0) ynz(n)+ (1-yn)w(n) = y z(1) = (1, z(1)2, …, z(1)n) w(1) = (0, w(1)2, …, w(1)n) y1z(1)+ (1-y1)w(1) = y z(2) = (z(2)1, 1, …, z(2)n) w(2) = (w(2)1, 0, …, w(2)n) y2z(2)+ (1-y2)w(2) = y … • Y = YT • Diagonal(Y) = y • Yi/yi  S, (y –Yi)/(1-yi)  S • Y is p.s.d. LS Yij = Pr[i=1 ^ j=1] LS+ Y =

  6. Prover-Adversary Game Showing y  LSr(VC) can be viewed as “almost” a 2-player game y  LSr(VC) u (0 < yu < 1) z(u) ,w(u)  LSr-1(VC) z(u) ,v  Prover Adversary

  7. 1 1/2 2/3 1 ? 1 1/2 ? 2/3 1 2/3 1 2/3 1 1/2 ? ? 2/3 1 1 1/2 0 1/2 1 1 2/3 1 1 1 0 2/3 Prover-Adversary Game (contd…) 2/3 x 1/3 x Ha! 1/2 x 1/2 x

  8. Work on LS+ for Vertex Cover GK’98 1 round 2 -  Charikar’02 1 round + triangle inequality 2 -  AAT’05 (n) rounds (VC in k-uniform hypergraphs) k-1- FO’06 1 round (random 3XOR) 7/6 -  This paper (n) rounds 7/6 -  GMPT’06 (√(logn/loglogn)) 2 - 

  9. 001 010 110 000 011 101 111 100 The Graphs • Random 3XOR formula (m = cn clauses) (FO’06) • Show solutions for Independent Set (y1, …, yn)  LSr(VC)  (1-y1, …, 1-yn)  LSr(IS) x1 + x2 + x3 = 1 x3 + x4 + x5 = 0 1/4 1/4 1/4 1/4 1/4 1/4 1/4 1/4

  10. Properties of random 3XOR formulas • Any assignment satisfies at most (1/2 + )m clauses. For VC: Integer optimum  4m - (1/2+)m = (7/2-)m (IS  assignment, vertices included  clauses satisfied) Fractional optimum = 4m – ¼  4m = 3m Integrality gap  7/6 -  • Two clauses share at most one variable (constant probability). • k clauses involve at least 1.9k variables for k  n (AAT’05, other works in proof complexity) (implies no small unsatisfiable subset)

  11. 1/2 1/2 1/2 1 1/2 The conditional distributions 1/4 0 1/4 1 1/4 1/4 A 1/4 0 1/4 1/4 1/4 0 B A B ? Expansion prevents propogation of effects from conditioning.

  12. 0 0 0 0 1/3 0 0 1 1 0 0 1/3 1/3 1 0 0 The conditional distributions (contd…) • Conditional distribution for yi = 0 is a convex combination of distributions for yj = 1 for other j’s. • All values are 0, 1/4, 1/2 or 1 and neighbors of 1 are 0. Solutions are fractional independent sets. • Need to maintain expansion even after variables are assigned. = 1/3 + 1/3 + 1/3

  13. Maintaining the expansion • Problem: Adversary fixes variables – may cause loss of expansion. Solution: We fix more variables - all variables in a maximal non-expanding subset of clauses, say B. • Adversary fixes t (= 1 or 2) clauses, we fix |B| (k, 1.9)  (k – t – |B|,1.9) • If y = iy(i), then y(i) S  i y S. • Express y as uniform distribution over consistent assignments to clauses in B. B

  14. Maintaining the expansion (contd…) • If B is consistent, k clauses & l vars  2l-k solutions. • For C  B (with 3 variables), if every assignment to C is consistent with B, every assignment appears in 2k-l-2 solutions (solutions form affine subspace). • B consistent with any assignment to C as contradictions have very low expansion. 2l-k-2 2l-k-2 1/4 1/4 1/4 1/4 2l-k-2 2l-k-2

  15. B Assigning clauses consistently • Let S  B be minimal unsatisfiable subset after fixing an assignment to C  B. • S involves at most 1.5|S| variables as each must occur twice. Adversary : 1 (or 2) clauses, at most 3 (or 4) variables C : 1 clause, at most 3 variables S : |S| clauses, at most 1.5|S| variables 1.5|S| + 3 + 3  1.9 (|S| + 2)  |S|  5

  16. How many rounds? • Start with (n, 1.95)-expansion. • At ith round, adversary fixes set Si, we fix Ti. 1.95(i |Si| + |Ti|)  #(fixed vars)  3i|Si| + 1.9i|Ti| • Stop after r rounds if n - i (|Si| + |Ti|)  4 n – 44r  4

  17. But the title says LS+! • Still need to show matrix Y at each round is p.s.d. • A matrix Y Rm  m is p.s.d. iff v1, …, vm Rm such that Yij = vivj  i,j • Need to exhibit a vector for every vertex with above property (also shows symmetry).

  18. 001 00 010 111 100 11 The vector solutions (FO’06) • Divide variables into equivalence classes: 3-variable eqn  different classes 2-variable eqn  same class • One coordinate for each class. vi has  yi in the coordinate corresponding to the classes contained and 0 elsewhere. One extra coordinate = yi in vi. x1 + x2 + x3 = 1 x3 + x4 = 0 Classes: (x1), (x2), (x3 & x4) vi vj= yiyj(1 + nagree – ndisagree) 0 if contradict = yi yj no shared vars 2yi yj shared agree (¼, ¼, ¼, ¼) (¼, -¼, ¼, -¼) (½, 0,0, - ½) (¼, ¼, ¼, ¼) (¼, ¼, -¼, -¼) (½, 0,0, ½)

  19. Conclusions/Open Problems • 2- gap for (n) rounds. • Reduction from CSPs  LS+ rounds  partial assignments. 2- gap using other CSPs? • Interpretation as probability distributions still does not give a natural interpretation of the vector solutions. Other natural ways of looking at the hierarchy? • Similar results for Sherali-Adams hierarchy? • LS+ results for other problems? • Sparsest Cut • Khot’s SDP for Unique Games

  20. Thank You Questions?

More Related