Hierarchies. 1. Lovasz-Schrijver (LS), LS+ 2. Sherali Adams 3. Lasserre 4. Mixed Hierarchy (recently used) Idea: P = conv(subset S of 0,1 n )

Transcription

1 Hierarchies

2 Today 1. Some more familiarity with Hierarchies 2. Examples of some basic upper and lower bounds 3. Survey of recent results (possible material for future talks)

3 Hierarchies 1. Lovasz-Schrijver (LS), LS+ 2. Sherali Adams 3. Lasserre 4. Mixed Hierarchy (recently used) Idea: P = conv(subset S of 0,1 n ) x = x 1,, x n P = prob. distribution over S Want to approximate P by putting linear constraints on x i. Hierarchies try to capture some kind of local correlations

4 Sherali Adams Variable x S intended to model π i S x i x i 2 = x i, so x S = x S {i} if i S Take LP, multiply each constraint by Π i I x i Π j J 1 x j for all I J = S S t x S should satisfy additional constraints Applying this to constraint 1 >= 0 x 12 0 x 1 x 12 0 x 2 x x 1 x 2 + x 12 0 Same as: M S (x) PSD and M S (g l * x) PSD

5 Alternate Description x S,α α 0,1 S Non-negativity x S,α 0 Consistency: x 12, 00 + x 12, 01 = x 1, 0 = x 13, 00 + x 13, 01 Sum up to 1. Pf: x S,α = T S α 1 (0) 1 T x S α 1 1 T (inclusion-exclusion) Alternate view: Define a probability distribution D(S) on each subset S. D S and D T consistent on S T Sherali Adams solution local distributions D(S) that satisfy LP constraints. (will make more sense later)

6 Knapsack Single constraint: Max i p i x i s. t. i a i x i B Sherali Adams: Multiply by Π i I x i Π j J 1 x j for all I J = S S t a i = 1 B = 1.99 x x n 1.99 Can determine x i,j = 0. How? Yet LP has integrality gap about 2, even for linear rounds.

7 Knapsack Gap x x n 2 ε Consider solution x i = p = x S = 0 for 2 S αn 2 ε n 1+α 1 ε Claim: Solution valid for αn rounds of SA. Proof: 1) Valid distribution on S. 2) Constraints: Suffices to check multiplication with π i S 1 x i only i S x i 2 ε (1 x i ) i S

8 Max independent set x i + x j 1 for i, j E Recovers x ij = 0 for edges. (Does not seem to capture clique constraints like Lasserre will do?)

9 Max-cut Basic LP. x ij variables i, j V Max (i,j) E x ij x ij + x jk x ik x ij + x jk + x ki 2 Perhaps could add more valid inequalities E.g. For every set of 5 vertices, at most 6 edges in cut. [Kenyon-Mathieu, de la Vega 07] Gap ½ + ε after log Ω 1 n rounds. Idea: Construct consistent distribution D S on up to 2t +3 vertices (i.e. valid combination of cuts). But objective far from any cut. [Charikar Makarychev Makarychev 09] Even after n γ rounds

10 Lasserre Hierarchy x S variables as in Sherali Adams Multiply by Π i I x i Π j S\I (1 x j ) for all I S, S t Put the constraint that M t x is PSD (previously M S x PSD) Has the following nice implication

11 Recall: Y PSD iff y ij = < v i, v j > for some vectors v 1,, v n (Y = V V T, Cholesky decomposition) M t x matrix with entries are indexed by S,T: x S T So x S T = < v S, v T > Stated as follows: Linear constraints on x_s obtained from g l and from constraint 1>= 0 < v S, v S > = < v T, v T > if S S = T T V φ 2 = 1

12 Interesting consequence x i = < v i, v φ > = < v i, v i > In other words, v i v φ 2 2 = v φ 2 4 Could have also expressed things using x S,α x S,α = < v φ, v S,α > V S,α = ( 1) T T α 1 0 v α 1 1 T

13 Knapsack Let us look at the previous example Max x x n x x n 1.9 Lasserre objective: v i 2 i Claim: Objective 1 As previously x ij = 0. Now implies < v i, v j > = 0 Suppose objective = t >1. 2 Consider i v i tv 0

14 Max independent set One round of Lasserre implies clique inequalities i C x i 1 C: clique As previously x ij = 0. Now implies < v i, v j > = 0 So, i v φ, v i v i 2 1 (as v_i/ v_i orthonormal) But x i = < v i, v φ > = v i 2

15 Max-cut Goemans Williamson SDP get (first level of Lasserre)

16 Mixed Hierarchy Lasserre powerful, but Sherali adams also very useful. Dense max-cut Problems on bounded treewidth graphs Mixed k-hierarchy: Sherali Adams k levels But SDP only on level 1, i.e. X ij = < v i, v j >, x i = < v i, v φ >, 1 =< v φ, v φ > Raghavendra 08: Assuming UGC, best algorithm for any Max-k-CSP is one k-level mixed hierarchy. Idea: Any integrality gap instance can be converted to NP-Hardness proof.

17 Partial Survey How to use the power of hierarchies. How to show that they do not help. Progress on long-standing questions. 1. Sparsest Cut Find S, s.t. min E[S,V\S]/( S V\S ) LP does not give better than log n min x e e E s.t. x ij = 1 and x ij + x jk x ik ARV: Using SDPs (2 nd level hierarchy) get log n Natural formulation + triangle inequalities 2 x ij = v i v j (embedding l_2 square metrics into l_1) (known integrality gaps still far from this)

18 Coloring 3-colorable graphs Using n 0.5 colors (Wigderson) n 0.4 [Blum 91] (blum coloring tools) n 0.25 [Karger Motwani Sudan] n 3/14 [Blum Karger] Sequence of improvements Use ARV structure on space of solutions Recent improvement on Blum coloring tools (Thorup Kawarabayashi 12) Current best around n Belief: n ε approximation using O 1 ε rounds of Lasserre Recently Arora-Ge show... Ratios for special classes Use Local-Global Correlation ideas recently developed.

19 Dense k-subgraph improved best known bound n 1/4 Use Sherali Adams Cannot beat n 1/4 using SA (SODA 12) Strong Lasserre lower bounds: n ε rounds but n γ hardness Unique Games: Local-global correlations Performance of Lasserre rounding related to spectrum of graph. General Techniques: Decomposition Lemma in context of Knapsack Directed Steiner Tree [Rothvoss]

20 How to show lower bounds. Nice techniques developed SA: Max-cut: any local constraint involves at most 2t+3 vertices. If x_s obtained from actual distribution on cuts on V[S] no constraint can be violated. Goal: Design distributions s.t. locally good, but globally bad. LS: (1,x) N(P) if protection matrix Y s.t. (Y 0 Y i )/(1 x i ) P and Y i x i Y 0 = (1, x), Y ii = x i, P x N t (P) if terms above in t-1 th level. Design a lower bound can be viewed as a game against adversary.