1 / 31

Avner Magen Joint work with Costis Georgiou, Toni Pitassi and Iannis Tourlakis

Tight integrality gaps for vertex-cover semidefinite relaxations in the Lov á sz-Schrijver Hierarchy. Avner Magen Joint work with Costis Georgiou, Toni Pitassi and Iannis Tourlakis University of Toronto. Minimum Vertex Cover. Finding minimum size VC is NP -hard

shlomo
Download Presentation

Avner Magen Joint work with Costis Georgiou, Toni Pitassi and Iannis Tourlakis

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. Tight integrality gaps forvertex-cover semidefinite relaxations in the Lovász-Schrijver Hierarchy Avner Magen Joint work with Costis Georgiou, Toni Pitassi and Iannis Tourlakis University of Toronto

  2. Minimum Vertex Cover Finding minimum size VC is NP-hard Exist simple 2-approximations All known algs are 2  o(1) approximations! Probabilistically checkable proofs (PCPs)  No poly-time 1.36 approximation [Dinur-Safra’02] Unique Games Conjecture [Khot’02]  No poly-time 2   approximation [Khot-Regev’03] Alternative (concrete) approach [ABL’02, ABLT’06]: Rule out approximations by large subfamilies of algorithms

  3. Linear Programming approach Easy to see IG ≤ 2 min SiV vi vi + vj ≥ 1, ij E vi  {0,1} for Kn :IG =2 1/n 0 ≤ vi ≤ 1 True Optimum Integrality Gap: max Optimal Fractional Solution

  4. Semidefinite Programming Relaxations SDP: the ultimate remedy? Vertex Cover on G = (V,E) Tighter relaxation?Smaller integrality gap? Kleinberg-Goemans’98: Integrality gap 2  o(1) min SiV (1 + v0 ·vi)/2 (v0  vi) · (v0  vj) = 0, ij  E || vi ||2 = 1, vi  Rn+1 min SiV (1 + x0xi)/2 (x0  xi)(x0  xj) = 0, ij  E |xi| = 1 Charikar’02: Gap still 2  o(1) Hatami-M-Markakis’06: Integrality gap still 2  o(1), even with “pentagonal” inequalities Clearly holds in integral case vi  {1,1} (v0  vi) · (v0  vj)  0, i,j (vi  vj) · (vi  vk)  0, i,j

  5. Systematic Approach: Lovász-Schrijver Liftings [LS’91] • Procedures LS0, LS, LS+ for tightening linear relaxations • Integral hull in≤ n rounds • Optimize over rth round relaxation in nO(r) time Very powerful algorithms obtained through smallnumber of rounds: • GW’94, KZ’97, ARV’04 algorithms “poly-time” in LS+ • All NP in “exponential time” May view super-constant rounds lower bounds in LS+ models as evidence about inapproximability • Has PSD constraint • Sequence of tighter and tighter SDPs “Lift” to obtain SDP Relaxation Initial Linear Relaxation “Project” back to obtain tighter LP Integral Hull n2 variables n variables

  6. Previous Lower Bounds for Vertex Cover – without SDP constraints (LS) • [ABLT’06]: Int. gap 2  o(1) after W(log n)LS rounds • [Tourlakis’06]: Int. gap 1.5  o(1) after W(log2 n)LS rounds • [STT’06b]: Int. gap 2  o(1) after W(n)LS rounds

  7. Status of SDP variant LS+ Stronger: one round already • Implies clique constraint • More generally, gives n-θ(G) lower bound on VC (so sparse graph are generally not good) • Gives rise to SDPs in the “lift” phases.

  8. Integrality gap of 7/6 for LS+ (STT06a) PCP world: Hastad 0.5-hardness for MAX3XOR and the FGLSS reduction imply 7/6-hardness for VC AAT05 proved matching LB (for int. gap) in LS+ world for MAX3XOR STT06b using further ideas from FO06, extend AAT MAX3XOR LB to prove 7/6 int. gap for linear rounds graph family: FGLSS reduction on random MAX3XOR instances Int. gap 7/6 already after one round

  9. Vertex Cover in LS: results so far ABLT ’02,STT ’07 NO YES YES STT ’06 YES NO YES Charikar ’02 YES YES NO New result YES YES YES

  10. Main Result Theorem: Int. gap 2  o(1) for SDPs resulting after (√log n/log log n)LS+ rounds One LS+ round tighter than [C’02] SDP SDPs ruled out incomparable to SDPs with (generalized) triangle and pentagonal inequalities (e.g., [HMM’06]) Theorem: Int. gap 2  O(1/√log n/log log n)after O(1)LS+ rounds Karakostas [K’05] SDP gives 2  (1/√log n) approximation Use same graph families as [KG’98], [C’02], [HMM’06] SDP solutions rely on sequence of polynomials applying tensor operations on vectors

  11. LS+ lift-and-project: the quick guide Homogenization: coneK min SiV xi Convert vertex cover LP into an SDP? Multiply linear inequalities to get valid quadratic constraints. Crucially, add integrality conditions: (x0  xi)xi = 0 E.g., Linearize: replace products xixj with linear variables Yij Lifted SDP in (n + 1)2 variables Project resultingconvex body back onto n + 1 variables Y0i xi + xj  1 (i,j)  E 0  xi  1 i  V xi + xj  x0  0 (i,j)  E xi  0 i  V x0  xi  0 i  V (x0 = 1) (x0 = 1) (x0  xi)xi = 0 Y0i = Yii = xi vk · (vi + vj  v0)  0 ij  E (v0  vi) · (v0  vj) = 0 ij  E (v0  vi) · (v0  vj)  0 Yik + Yjk Y0k  0 ij  E Y00  Y0i  Y0j + Yij = 0 ij  E Y00  Y0i  Y0j + Yij  0 xk(xi + xj  x0)  0 ij  E (x0  xi)(x0 – xj) = 0 ij  E (x0  xi)(x0  xj)  0 xk(xi + xj  x0)  0 ij  E (x0  xi)(xi + xj  x0)  0 ij  E (x0  xi)(x0  xj)  0 xk(xi + xj  x0)  0 ij  E (x0  xi)(xj – x0)  0 ij  E (x0  xi)(x0  xj)  0 Yei ,Y(e0ei)  K Y is PSD

  12. How LS and LS+ tighten VC Relaxation min SiV xi xi + xj ≥ 1, ij  E 0 ≤ xi ≤ 1 One round of LS precisely adds “odd-cycle constraints”: • For all cycles C in G of odd length, SiCxi ≥ (|C|+1)/2 x1 + x2 + x3 ≥ 2 One round of LS+ adds more: • Clique constraints: For all cliques K in G, SiKxi ≥ |K| – 1 vs. x1 + x2 + x3 ≥ 3/2

  13. Deriving the clique constraints in LS+ Let K be a clique of size k in G 0 ≤ S (x0 –xi) (xi + xj –x0) +((k –1)x0 –Sxi) 2 i≠jK iK Edge constraint SDP condition =S xi2 – (k –1) x02 Afterprojecting S xi ≥ k –1 iK

  14. Proving Lower Bounds in LS+ Hierarchies LP relaxation K for Gwith min VC ~n: xi + xj≥1 ij  E (½, ½,…) (½+g, ½+g, …) x K(r)if  matrix Ys.t. • diagonal is x • Y is PSD • “columns”  K(r 1) “Protection” matrix for x Lemma (LS’91): I.H. K(1) K(3) K(2) Int. gap of K is ≥ 2 – o(1)  Use inductive proof: find appropriate Y’s

  15. “Frankl-Rödl” graphs m-dimensional Hamming cube: n = 2m points parameter V = {1,1}m (i, j) E iff(i, j) = (1  )m } Theorem: [Frankl-Rödl’87] Max Ind.Set size  |B(v,n/2(1- ))| m2m(1  2/64)m (i, j) = (1  )m Cor: If  = (√log m/m) then max IS is o(2m) = o(n) Graphs used for int.gaps in [KK91, AK94, KG95, C02, HMM06]

  16. What’s so wonderful about them?... Start with a perfect matching n/2 o(n) • Vertex Cover= +O( ) • ``Geometric’’ vertex cover = n/2 Perturb : edges connect vertices of Ham. Dist. (1-)n

  17. Proof Outline (i, j) = (1  )m VC 1  o(n) x K(r) if  PSD matrix Ys.t. • diagonal is x • “columns”  K(r1) x = (½ + )1 In induction: need vectors vi to define matrix Yij = vi vj • Show vi exist whenever • x {0, 1, ½ + }n and •  > 6 • Ensure S {0, 1, ½ + }nwhere     O()  (/) round lower bound for x = (½ + )1 Constantand = (√log m/m)  Int. gap 2  o(1) after (√log n/log log n) rounds 2’.Show some set SK(r1) where “columns” conv(S)

  18. V = {1,1}m (i, j) = (1  )m Back to Frankl-Rödl graphs VC 1  o(n) 1 √m Natural set {ui} of unit vectors: {1,1}m Note:ui · uj = 1  2(i, j)/m 2 1 for (i, j)  E Hence (i, j)  E  uianduj nearly antipodal Nearly true for vi =ui (v0  vi) · (v0  vj) = 0, (i, j)  E ui · uj vi · vj linear function F of vi · vj Kleinberg-Goemans: Affine translation onuito obtain vi 1 1 1 F 0 21 F 1 1 1

  19. F(vi · vj) (i, j) = (1  )m Use Kleinberg-Goemans vi for LS+? Fact: One round of LS+ also requires following ineq: (v0  vi) · (v0  vj)  0 i,j equality whenever ij  E I.e, when ui · uj = 2 1 VC 1  o(n) ui · uj vi · vj F(wi · wj) F(vi · vj) 1 1 1 1 linear map • Idea (Charikar): Map ui to wi s.t. • F(wi · wj)  0 • F(wi · wj) = 0 if ij  E Desired mapping on dot-products [KG] affine map on ui 0 0 21 How? Use tensoring 1 1 1 1

  20. Tensoring u, v Rn Tensor product: uv Rn2 Value uivj at coordinate (i, j)  [n]2 Easy fact:(uv) · (uv) = (u · v)2 Let P(x) = c1xt1 + … + cqxtq Consider map TP(u) = (c1ut1,…, cqutq) Example:P(x) = x2 + 4x TP(u) = (uu, 2u)  Rn2+2n TP(u) · TP(v) = (u · v)2 + 4(u · v)2 = P(u · v) Fact: TP(u) · TP(v) = P(u · v) Positive coefficients 2 2 P determines dot-product of resulting vectors

  21. F(vi · vj) (i, j) = (1  )m (v0  vi) · (v0  vj)  0 i,j equality whenever (i, j)  E Back to finding solution for stronger SDP: Use TP VC 1  o(n) I.e, when ui · uj = 2 1 ui · uj Want wi = TP(ui)s.t. F(wi · wj) min at (i, j)  E 1 1 KG 21 0 ui · uj 1 0 1 C 21 F 1 1 Charikar exhibits appropriate P

  22. I.H. Charikar sol’n gives one round LS+ lower bound x = (½ + )1 x K(r) if  PSD matrix Ys.t. • diagonal is x • “columns”  K(r1) Charikar vectors define Yij = vi·vj that: • Diagonal is x = (½ + )1 • “Columns” K VC = 1  o(n) Must have seq of polynomials Can Charikar vectors show “columns” K(1)? Values distributed like polynomial of Gaussian • Problems: • (1) “Columns” not of form (½ + )1 • (2) Charikar’s vectors work only for one value

  23. Making non-uniform “columns” uniform values distributed like polynomial of Gaussian “Columns” we want to continue from not of form (½ + )1 Def [STT]:x  K is -saturated if for all ij  E so that xi, xj<1 there is surplus: xi + xj 1 + 2 Lemma [STT]:x is -saturated  there exists set of vectors x(i) {0, 1, ½ + }n in K s.t. x  conv({x(i) }).  Can convert “columns” to (essentially) (½ + )1 IF “columns” are -saturated Will be safe to “ignore” 0/1

  24. Is saturation good enough? Goal: matrix Y for x with “column” saturation  O() Recall P(x) definesTP(u) such that TP(u) · TP(v) = P(u · v) deg(PC) =O(1/) Fact: Yhas “columns” s.t. some edges never have surplus Problem: saturation of “close by” edges? = o(m) ~ P(1)P(1-1/m) ≤ P’(1)/m For all P P Saturation  Bad saturation zone Normal. Ham. Dist. from blue edge Necessary: deg(P) ≥  · m ! The blue edge

  25. Want column saturation   O() Precise technical property needed for P: For all vertices k and all edges ij : | P(ui · uk ) + P(uj · uk) |  O() y 1  [1, 1] 11/m Need | P(x) + P(y) |  O() over R 12 R But ui · uj = 2  1 for all edges ij, so 12 1 x • |ui·uk+uj·uk|  2 1 21 11/m • |ui·ukuj·uk|  2(1-) 21 Red points correspond to 0-1 edges  Ignored in saturation calculation 1 Domain of P(x) + P(y)

  26. Defining the sequence of tensoring polynomials arbitrary  > 6 So far: • There must be a seq of polys dep. on , , m. • Polynomials must have large degree. Let x {0, 1, ½ + }n Take P(x) = (-O()) x(x  1)m/ +  x 1/ + (1- -O()) x Properties: • Minimum at ui · uj, ij  E • P’(1) >m  • Works as long as  > 6 • The “Columns” of Y that is produced by using TP,m(ui) have saturation   O()  KG P ui · uj 1 21 0 1 C

  27. (i, j) = (1  )m Putting everything together VC 1  o(n) r = (/) x = (½ + )1 Induction: Have x {0, 1, ½ + }n where  > 6 • Define Y using TP,m(ui)  “Columns” have saturation     O() [STT]  Exists S K  {0, 1, ½ + }ns.t. “columns” conv(S) Induction Hypothesis  S K(r 1) Takeconstant and = (√log m/m) x K(r) if  PSD matrix Ys.t. • diagonal is x • “columns”  K(r1) 2’.Show some set SK(r1) where “columns” conv(S)  x K(r)

  28. Requiring that ||vi-vj||2 is l1? • As is, no l1 inequalities are not implied. • The results of [HMM] (showing that metric-cut ineqaities and pentagonal inequalities hold) suggest the examples are still good. • Need to • Give Sherali Adams LB • introduce dij = ||vi-vj||2 • Add more reqs the LS+ proof need to satisfy.

  29. Sherali-Adams [SA’90] Lift-and-Project Idea: Keep “lifting” but never project! Simulate third, fourth, etc, degree products with linear vars Only known integrality gap [FK’06]: W(log n)SA rounds  int. gap ≤2  e for MAX-CUT SA+ lower bound would inequalities for lifted variables  Triangle, pentagonal, etc., inequalities derivable E.g., x1x2x3 Y123 LP not SDP version

  30. Relations to Unique Games Conjecture (UGC) • LS+ lower bounds may provide evidence of inapproximability • UGC [Khot’02] implies optimal inapproximability results for Vertex Cover, MAX-CUT, etc  Strong LS+, SA+ lower bounds for VC, MAX-CUT

  31. Thanks

More Related