Download presentation

Presentation is loading. Please wait.

Published bySofia Pereira Modified over 4 years ago

1
Tight integrality gaps for vertex-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-Safra02] Unique Games Conjecture [Khot02] No poly-time 2 approximation [Khot-Regev03] Alternative (concrete) approach [ABL02, ABLT06]: Rule out approximations by large subfamilies of algorithms

3
Linear Programming approach min i V v i v i + v j 1, ij E v i {0,1} 0 v i 1 True Optimum Optimal Fractional Solution Integrality Gap:max Easy to see IG 2 for K n : IG = 2 1/n

4
SDP: the ultimate remedy? Vertex Cover on G = (V,E) Tighter relaxation? Smaller integrality gap? min i V (1 + v 0 · v i )/2 (v 0 v i ) · (v 0 v j ) = 0, ij E || v i || 2 = 1, v i R n+1 min i V (1 + x 0 x i )/2 (x 0 x i )(x 0 x j ) = 0, ij E |x i | = 1 Hatami-M-Markakis06: Integrality gap still 2 o(1), even with pentagonal inequalities Semidefinite Programming Relaxations Kleinberg-Goemans98: Integrality gap 2 o(1) Clearly holds in integral case v i { 1,1} (v 0 v i ) · (v 0 v j ) 0, i,j (v i v j ) · (v i v k ) 0, i,j Charikar02: Gap still 2 o(1)

5
Systematic Approach: Lovász-Schrijver Liftings [LS91] Procedures LS 0, LS, LS + for tightening linear relaxations Integral hull in n rounds Optimize over rth round relaxation in n O(r) time Very powerful algorithms obtained through small number of rounds: GW94, KZ97, ARV04 algorithms poly-time in LS + All NP in exponential time May view super-constant rounds lower bounds in LS + models as evidence about inapproximability Initial Linear Relaxation Integral Hull Has PSD constraint Sequence of tighter and tighter SDPs Lift to obtain SDP Relaxation n variables n 2 variables Project back to obtain tighter LP

6
Previous Lower Bounds for Vertex Cover – without SDP constraints (LS) [ABLT06]: Int. gap 2 o(1) after (log n) LS rounds [Tourlakis06]: Int. gap 1.5 o(1) after (log 2 n) LS rounds [STT06b]: Int. gap 2 o(1) after (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 SDP version (LS+)? Int. gap 2-o(1) ? # rounds superconsant? 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 [C02] SDP SDPs ruled out incomparable to SDPs with (generalized) triangle and pentagonal inequalities (e.g., [HMM06]) Theorem: Int. gap 2 O(1/log n/log log n) after O(1) LS + rounds Karakostas [K05] SDP gives 2 (1/log n) approximation Use same graph families as [KG98], [C02], [HMM06] SDP solutions rely on sequence of polynomials applying tensor operations on vectors

11
x k (x i + x j x 0 ) 0 ij E (x 0 x i )(x j – x 0 ) 0 ij E (x 0 x i )(x 0 x j ) 0 v k · (v i + v j v 0 ) 0 ij E (v 0 v i ) · (v 0 v j ) = 0 ij E (v 0 v i ) · (v 0 v j ) 0 x k (x i + x j x 0 ) 0 ij E (x 0 x i )(x 0 – x j ) = 0 ij E (x 0 x i )(x 0 x j ) 0 Y ik + Y jk Y 0k 0 ij E Y 00 Y 0i Y 0j + Y ij = 0 ij E Y 00 Y 0i Y 0j + Y ij 0 Convert vertex cover LP into an SDP? Multiply linear inequalities to get valid quadratic constraints. Crucially, add integrality conditions: (x 0 x i )x i = 0 E.g., Linearize: replace products x i x j with linear variables Y ij Lifted SDP in (n + 1) 2 variables Project resulting convex body back onto n + 1 variables Y 0i x k (x i + x j x 0 ) 0 ij E (x 0 x i )(x i + x j x 0 ) 0 ij E (x 0 x i )(x 0 x j ) 0 LS + lift-and-project: the quick guide min i V x i x i + x j 1 (i,j) E 0 x i 1 i V (x 0 = 1) x i + x j x 0 0 (i,j) E x i 0 i V x 0 x i 0 i V Ye i,Y(e 0 e i ) K Y 0i = Y ii (x 0 x i )x i = 0 (x 0 = 1) Y is PSD Homogenization: cone K = x i

12
How LS and LS+ tighten VC Relaxation One round of LS precisely adds odd-cycle constraints: For all cycles C in G of odd length, i C x i (|C|+1)/2 x 1 + x 2 + x 3 2 One round of LS+ adds more: Clique constraints: For all cliques K in G, i K x i |K| – 1 min i V x i x i + x j 1, ij E 0 x i 1 vs. x 1 + x 2 + x 3 3/2

13
Deriving the clique constraints in LS+ 0 x 0 – x i ) (x i + x j – x 0 ) +( (k – x 0 – x i ) i 2 Edge constraint ij Let K be a clique of size k in G SDP condition x i 2 – (k – 1) x 0 2 x i k – 1 After projecting i

14
x K (r) if matrix Y s.t. diagonal is x Y is PSD columns K (r 1) Proving Lower Bounds in LS + Hierarchies I.H. LP relaxation K for G with min VC ~ n: x i + x j 1 ij E (½, ½,…) K (1) K (3) K (2) Int. gap of K is 2 – o(1) (½+, ½+ …) Use inductive proof: find appropriate Ys Protection matrix for x Lemma (LS91):

15
Frankl-Rödl graphs m-dimensional Hamming cube: n = 2 m points V = { 1,1} m (i, j) E iff (i, j) = (1 )m } parameter Theorem: [Frankl-Rödl87] Max Ind.Set size |B(v,n/2(1- ))| m2 m (1 2 /64) m Cor: If = (log m/m) then max IS is o(2 m ) = o(n) Graphs used for int.gaps in [KK91, AK94, KG95, C02, HMM06] (i, j) = (1 )m

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

17
Proof Outline In induction: need vectors v i to define matrix Y ij = v i v j Show v i exist whenever x {0, 1, ½ + } n and > 6 Ensure S {0, 1, ½ + } n where 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 x K (r) if PSD matrix Y s.t. 1. diagonal is x 2. columns K (r 1) 2. Show some set S K (r 1) where columns conv(S) (i, j) = (1 )m VC 1 o(n) x = (½ + )1

18
Back to Frankl-Rödl graphs Natural set {u i } of unit vectors: { 1,1} m (v 0 v i ) · (v 0 v j ) = 0, (i, j) E m 1 Note: u i · u j = 1 2 (i, j)/m Hence (i, j) E u i and u j nearly antipodal Nearly true for v i = u i 2 1 for (i, j) E linear function F of v i · v j (i, j) = (1 )m VC 1 o(n) u i · u j 1 2 1 1 F 1 0 1 v i · v j 1 1 Kleinberg-Goemans: Affine translation on u i to obtain v i F V = { 1,1} m

19
Use Kleinberg-Goemans v i for LS + ? Fact: One round of LS + also requires following ineq: Idea (Charikar): Map u i to w i s.t. F(w i · w j ) 0 F(w i · w j ) = 0 if ij E I.e, when u i · u j = 2 1 How? Use tensoring (v 0 v i ) · (v 0 v j ) 0 i,j equality whenever ij E (i, j) = (1 )m VC 1 o(n) u i · u j 1 2 1 1 F(v i · v j ) 1 0 1 v i · v j 1 1 [KG] affine map on u i linear map F(v i · v j ) F(w i · w j ) 1 0 1 Desired mapping on dot-products

20
Tensoring u, v R n Tensor product: u v R n 2 Value u i v j at coordinate (i, j) [n] 2 Easy fact: (u v) · (u v) = (u · v) 2 Let P(x) = c 1 x t 1 + … + c q x t q Consider map T P (u) = (c 1 u t 1,…, c q u t q ) Example: P(x) = x 2 + 4x T P (u) = (u u, 2u) R n 2 +2n T P (u) · T P (v) = (u · v) 2 + 4(u · v) 2 = P(u · v) Fact: T P (u) · T P (v) = P(u · v) 2 2 P determines dot-product of resulting vectors Positive coefficients

21
Back to finding solution for stronger SDP: Use T P Charikar exhibits appropriate P (i, j) = (1 )m VC 1 o(n) I.e, when u i · u j = 2 1 (v 0 v i ) · (v 0 v j ) 0 i,j equality whenever (i, j) E F(v i · v j ) u i · u j 1 2 1 1 F 1 0 1 Want w i = T P (u i ) s.t. F(w i · w j ) min at (i, j) E u i · u j 1 1 2 1 0 KG C

22
Charikar soln gives one round LS + lower bound Charikar vectors define Y ij = v i · v j that: 1. Diagonal is x = (½ + )1 2. Columns K x K (r) if PSD matrix Y s.t. 1. diagonal is x 2. columns K (r 1) I.H. x = (½ + )1 Can Charikar vectors show columns K (1) ? VC = 1 o(n) Must have seq of polynomials Problems: (1) Columns not of form (½ + )1 (2) Charikars vectors work only for one value Values distributed like polynomial of Gaussian

23
Making non-uniform columns uniform Columns we want to continue from not of form (½ + )1 Def [STT]: x K is -saturated if for all ij E so that x i, x j < 1 there is surplus: x i + x j 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 values distributed like polynomial of Gaussian

24
Goal: matrix Y for x with column saturation ( ) Recall P(x) defines T P (u) such that T P (u) · T P (v) = P(u · v) deg(P C ) = O(1/ ) Fact: Y has columns s.t. some edges never have surplus Problem: saturation of close by edges? Saturation Normal. Ham. Dist. from blue edge Necessary: deg(P) · m ! For all P P Bad saturation zone The blue edge ~ P(1) P(1-1/m) P(1)/m Is saturation good enough? = o(m)

25
Want column saturation O( ) Precise technical property needed for P: | P(u i · u k ) + P(u j · u k ) | O( ) For all vertices k and all edges ij : [ 1, 1] But u i · u j = 2 1 for all edges ij, so Need | P(x) + P(y) | O( ) over R Red points correspond to 0-1 edges Ignored in saturation calculation 1 1 1 1 1 2 2 1 R 1 1/m x y Domain of P(x) + P(y) |u i ·u k +u j ·u k | 2 |u i ·u k u j ·u k | 2(1- )

26
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) = ( x x 1) m/ + x 1/ + (1- x Properties: Minimum at u i · u j, ij E P(1) > m Works as long as > 6 The Columns of Y that is produced by using T P,m (u i ) have saturation O( ) u i · u j 1 12 1 0 KG C P arbitrary > 6 Defining the sequence of tensoring polynomials

27
Putting everything together Induction: Have x {0, 1, ½ + } n where > 6 Define Y using T P,m (u i ) Columns have saturation O( ) [STT] Exists S K {0, 1, ½ + } n s.t. columns conv(S) Induction Hypothesis S K (r 1) Take constant and = (log m/m) x K (r) if PSD matrix Y s.t. 1. diagonal is x 2. columns K (r 1) 2. Show some set S K (r 1) where columns conv(S) x = (½ + )1 r = ( / ) (i, j) = (1 )m VC 1 o(n) x K (r)

28
Requiring that ||v i -v j || 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 d ij = ||v i -v j || 2 Add more reqs the LS+ proof need to satisfy.

29
Sherali-Adams [SA90] Lift-and-Project Idea: Keep lifting but never project! Simulate third, fourth, etc, degree products with linear vars Only known integrality gap [FK06]: (log n) SA rounds int. gap 2 for MAX-CUT SA + lower bound would inequalities for lifted variables Triangle, pentagonal, etc., inequalities derivable E.g., x 1 x 2 x 3 Y 123 LP not SDP version

30
Relations to Unique Games Conjecture (UGC) LS + lower bounds may provide evidence of inapproximability UGC [Khot02] implies optimal inapproximability results for Vertex Cover, MAX-CUT, etc Strong LS +, SA + lower bounds for VC, MAX-CUT

31
Thanks

Similar presentations

OK

Integrality Gaps for Sparsest Cut and Minimum Linear Arrangement Problems Nikhil R. Devanur Subhash A. Khot Rishi Saket Nisheeth K. Vishnoi.

Integrality Gaps for Sparsest Cut and Minimum Linear Arrangement Problems Nikhil R. Devanur Subhash A. Khot Rishi Saket Nisheeth K. Vishnoi.

© 2019 SlidePlayer.com Inc.

All rights reserved.

To make this website work, we log user data and share it with processors. To use this website, you must agree to our Privacy Policy, including cookie policy.

Ads by Google