1 / 37

results from

How to get. Hardness of approximation. results from. Integrality gaps. Guy Kindler Weizmann Institute. About this talk. We’ll try to understand some notions, and their relations: Combinatorial optimization problems Approximation: relaxation by semi-definite programs. Integrality gaps

acton-buck
Download Presentation

results from

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. How to get Hardness of approximation results from Integrality gaps Guy KindlerWeizmann Institute

  2. About this talk We’ll try to understand some notions, and their relations: • Combinatorial optimization problems • Approximation: relaxation by semi-definite programs. • Integrality gaps • Hardness of approximation Main example: the Max-Cut problem

  3. Combinatorial optimization problems Input, search space, objective function Example: MAX-CUT

  4. Example: MAX-CUT • input: G = (V,E) • Search space: Partition V=(C, Cc) • Objective function: w(C) = fraction of cut edges • The MAX-CUT Problem:Find mc(G)=maxC{w(C)} • [Karp ’72]: MAX-CUT is NP-complete

  5. Example: MAX-CUT • input: G = (V,E) • Search space: Partition V=(C, Cc) • Objective function: w(C) = fraction of cut edges • The MAX-CUT Problem:Find mc(G)=maxC{w(C)} • -approximation: Output S, s.t. mc(G) ¸ S ¸ ¢mc(G). • History: ½-approximation easy, was best record for long time.

  6. Semi-definite Relaxation Introducing geometry into combinatorial optimization [GW ’95]

  7. Arithmetization v G=(V,E): u xv xu Problem: We can’t maximize quadratic functions,even over convex domains.

  8. Relaxation by geometric embedding v G=(V,E): u xv xu Problem: We can’t maximize quadratic functions,even over convex domains.

  9. Relaxation by geometric embedding v G=(V,E): u xv xu Problem: We can’t maximize quadratic functions,even over convex domains.

  10. Relaxation by geometric embedding v G=(V,E): u xv xu Problem: We can’t maximize quadratic functions,even over convex domains.

  11. xv xu Relaxation by geometric embedding Semi-definiterelaxation v G=(V,E): u Now we’re maximizing a linear function over a convex domain! (unit sphere in Rn)

  12. v G=(V,E): u Relaxation by geometric embedding Is this really a relaxation? (unit sphere in Rn)

  13. v G=(V,E): u xv xu Relaxation by geometric embedding Is this really a relaxation? (unit sphere in Rn)

  14. v G=(V,E): u Analysis by randomized rounding xv xu xu (unit sphere in Rn)

  15. xv xv xu <xu,xv> xu donation to (*) Analysis by randomized rounding arccos(<xu,xv>) • So (unit sphere in Rn)

  16. xv <xu,xv> arccos(<xu,xv>) xu donation to (*) Analysis by randomized rounding • So

  17. Analysis by randomized rounding • So mc(G)¸ GW¢rmc(G) ¸ GW¢mc(G) L.h.s. is tight, iff all inner products are ρ*

  18. An 0.879 approximation for Max-Cut Is GW the best constant here? • The [GW ’95] algorithm: • Given G, compute rmc(G) • Let S=GW¢rmc(G) • Output S. mc(G)¸ S¸ GW¢mc(G) mc(G)¸ GW¢rmc(G) ¸ GW¢mc(G) Is there a graph where this occurs? L.h.s. is tight, iff all inner products are ρ*

  19. Integrality gap Measuring the quality of the geometric approximation [FS ’96]

  20. The integrality gap of Max-Cut On instance G: w rmc(G) mc(G) r-S(G) S(G)

  21. The integrality gap of Max-Cut ¸ GW =GW for some G* On instance G: w rmc(G) The [GW ’95] algorithm: Given G, output IG¢rmc(G) mc(G) r-S(G) S(G)

  22. The integrality gap of Max-Cut =GW for some G* On instance G: • Using ¢rmc(G) to approximate mc(G), the factor =IG cannot be improved! • On G* the algorithm computes mc(G*) perfectly! w rmc(G) The [GW ’95] algorithm: Given G, output IG¢rmc(G) mc(G) r-S(G) S(G)

  23. (unit sphere in Rq) The Feige-Schechtman graph, G* arccos(ρ*) Vertices: Sn Edges: u~v iff <u,v>=* • rmc(F)¸(1-*)/2 • [FS]: mc(G*)=arccos(ρ*)/ • so

  24. From IG to hardness A geometric trick may actually be inherent [KKMO ’05]

  25. Thoughts about integrality gaps *under some reasonable comlexity theoretic assumptions • The integrality gap is an artifact of the relaxation. • The relaxation does great on an instance for which the integrality gap is achieved. • And yet, sometimes the integrality gap is provably* the best approximation factor achievable: • [KKMO ’04]: under UGC, GW is optimal for max-cut. • [HK ’03], [HKKSW], [KO ’06], [ABHS ’05], [AN ’02]

  26. Thoughts about integrality gaps And the IG instance G* is used in the hardness proof • The integrality gap is an artifact of the relaxation. • An algorithm does great on an instance for which the integrality gap is achieved. • And yet, sometimes the integrality gap is provably* the best approximation factor achievable: • [KKMO ’04]: Under UGC, GW is optimal for max-cut. • [HK ’03], [HKKSW], [KO ’06], [ABHS ’05], [AN ’02]

  27. A recipe for proving hardness Take the instance G*: w rmc(G*) mc(G*) r-S(G*) S(G*)

  28. A recipe for proving hardness Add ‘teeth’ to S(G*)which achieve rmc(G*). w rmc(G*) mc(G*) r-S(G*) S(G*)

  29. A recipe for proving hardness Now combine several instances, such that finding a point which belongs to all teeth becomes a hard combinatorialoptimization problem. w rmc(G*) mc(G*) r-S(G*) S(G*)

  30. A recipe for proving hardness Now combine several instances, such that finding a point which belongs to all teeth becomes a hard combinatorialoptimization problem. w

  31. A recipe for proving hardness Determining whether mc(G`)=mc(G*) or whethermc(G`)=rmc(G*) is intractable. Factor of hardness: mc(G*)/rmc(G*)=IG !! w

  32. Vertices: Sn Edges: u~v iff <u,v>=* Adding teeth to Feige-Schechtman (unit sphere in Rq)

  33. xi w.p. ½ + ½ρ -xi w.p. ½ - ½ρ y: yi = (unit sphere in Rq) Adding teeth to Feige-Schechtman Vertices: {-1,1}n Edges: u~v iff <u,v>=*: Vertices: Sn Edges: u~v iff <u,v>=* a random edge (x,y): x~{-1,1}n, • E[<x,y>] = ρ

  34. Adding teeth to Feige-Schechtman • mc = arccos(ρ)/ ? • For C(x) = x7, • w(C) = Pedge[x7 ≠ y7] = (1-ρ)/2 !!

  35. Adding teeth to Feige-Schechtman • mc = arccos(ρ)/ ? • For C(x) = x7, • w(C) = Pedge[x7 ≠ y7] = (1-ρ)/2 !! • For C(x) = sign(xi) = Maj(x), w(C) = P[Maj(x)≠Maj(y)]≈ (arccos ρ)/π no influential coordinates

  36. Adding teeth to Feige-Schechtman • mc = arccos(ρ)/ ? • [for axis parallel cut]:w(C)=(1-ρ)/2 • [MOO ‘05]:If i, Ii(C)<, w(C) (arccos ρ)/π +o(1) • Ratio between weight of ‘teeth’ cuts and regular cuts is ≈ GW(for ρ = ρ*)

  37. Conclusion We tried to understand some notions, and their relations: • Combinatorial optimization problems • Approximation: relaxation by semi-definite programs. • Integrality gaps • Hardness of approximation

More Related