36
Shorter Long Codes and Applications to Unique Games 1 Boaz Barak (MSR, New England) Parikshit Gopalan (MSR, SVC) Johan Håstad (KTH) Prasad Raghavendra (GA Tech) David Steurer (MSR, New England) Raghu Meka (IAS, Princeton)

Shorter Long Codes and Applications to Unique Games

  • Upload
    jenaya

  • View
    33

  • Download
    0

Embed Size (px)

DESCRIPTION

Shorter Long Codes and Applications to Unique Games. Raghu Meka (IAS, Princeton). Boaz Barak (MSR, New England) Parikshit Gopalan (MSR, SVC) Johan Håstad (KTH) Prasad Raghavendra (GA Tech) David Steurer (MSR, New England). Is Unique Games Conjecture true?. - PowerPoint PPT Presentation

Citation preview

Page 1: Shorter Long Codes and Applications to Unique Games

Shorter Long Codes and Applications to Unique Games

1

Boaz Barak (MSR, New England)Parikshit Gopalan (MSR, SVC)

Johan Håstad (KTH)Prasad Raghavendra (GA Tech)

David Steurer (MSR, New England)

Raghu Meka (IAS, Princeton)

Page 2: Shorter Long Codes and Applications to Unique Games

Is Unique Games Conjecture true?

2

Settles longstanding open problems in approximation algorithmsE.g., Max-Cut, vertex cover

Interesting even if notIntegrality gaps: Khot-Vishnoi’04.

UGC ~ Hardness of a certain CSP

Page 3: Shorter Long Codes and Applications to Unique Games

Is Unique Games Conjecture true?

3

Fastest algorithm [ABS10]: .

Best evidence: lowerbound in certain models. [Khot-Vishnoi’04, Khot-Saket’09, Raghavendra-Steurer’09]Captures ABS algorithm – BRS11,

GS11.Best algorithms for most problems!

E.g., Max-Cut, Sparsest-Cut.

Huge Gap!Source of gap: Long code is

actually quite long!

Page 4: Shorter Long Codes and Applications to Unique Games

Our Result

4

Main: Exponentially more efficient “replacement” for long

code.

Not necessarily a blackbox replacment.

Preserves main properties: Fourier analysis, dictatorship testing etc.

Page 5: Shorter Long Codes and Applications to Unique Games

Is Unique Games Conjecture true?

5

Fastest algorithm [ABS10]: .

This Work: Near quasi-polynomial lowerbounds in certain models.

Smaller gap …

Page 6: Shorter Long Codes and Applications to Unique Games

Outline of Talk

6

1. Applications of short code

2. Small set expanders with many large eigenvaluesConstruction and analysis

Page 7: Shorter Long Codes and Applications to Unique Games

Application I: Expansion vs Eigenvalue Profiles

7

S1

Expansion: Spectral:Cheeger Inequalities

Page 8: Shorter Long Codes and Applications to Unique Games

Small Set Expansion

8

Complete graph

Complete graph

Dumbell graph: not expanding … Is it really?Small sets expand!

Page 9: Shorter Long Codes and Applications to Unique Games

When is a graph SSE?Interesting by itselfClosely tied to Unique

Games – RS10

Small Set Expansion (SSE)

9

S1

Spectral:

???

Page 10: Shorter Long Codes and Applications to Unique Games

Core of ABS algorithm for Unique Games

Small Set Expansion (SSE)

10

S

Arora-Barak-Steurer’10Spectral:Atmost

eigenvalues larger than .

1

Page 11: Shorter Long Codes and Applications to Unique Games

Small Set Expansion

11

Question: How many large eigenvalues can a SSE have?

Small set

Small sets expand “Many” bad

balanced cutsBAD CUT

BAD CUT

BAD CUT

Page 12: Shorter Long Codes and Applications to Unique Games

Previous best: Noisy cube – .

Small Set Expansion

12

Question: How many large eigenvalues can a SSE have?

Our Result: A SSE with large eigenvalues.

Corollary: Rules out quasi-polynomial run time for ABS algorithm.

Page 13: Shorter Long Codes and Applications to Unique Games

Application II: Efficient Alphabet Reduction

13

Goemans-Williamson: 0.878 approximation

MAX-CUTGiven G find S maximizing E(S,Sc)

KKMO’04 + MOO’05: UGC true -> 0.878 tight!

Page 14: Shorter Long Codes and Applications to Unique Games

Are we done? (Short of proving UGC …)

Application II: UGC hardness for Max-CUT

14

UGC with n varsalphabet size k

MAX-CUT of sizeKKMO+MOO

KKMO’04 + MOO’05: UGC true -> 0.878 tight!

Page 15: Shorter Long Codes and Applications to Unique Games

Application II: Efficient Alphabet Reduction

15

MAX-CUT is a UG instance with k = 2

Linear UG with n varsalphabet size k

MAX-CUT of size

Page 16: Shorter Long Codes and Applications to Unique Games

Application III: Integrality Gaps

16

SDP Hierarchies: Powerful paradigm for optimization problems.

Which level suffices?

Basic SDP

Optimal Solution

No. Variable Levels

Eg: SDP+SA, LS, LS+, Lasserre, …SDP + SA

KV04: UG, Max-Cut, Sparsest Cut not in O(1) levels.

KS-RS09: UG, Max-Cut, Sparsest Cut not in levels.

This work: UG, Max-Cut, not in levels.

Page 17: Shorter Long Codes and Applications to Unique Games

Outline of Talk

17

1. Applications of short code

2. Small set expanders with many large eigenvaluesConstruction and analysis

Page 18: Shorter Long Codes and Applications to Unique Games

Long Code and Noisy Cube

18

Long code: Longest code imaginableWork with noisy cube – essentially the

same

Eg., is hypercube

Page 19: Shorter Long Codes and Applications to Unique Games

Noisy Cube is an SSE

19

Powerful: implies KKL for instance

Our construction “sparsifies” the noisy cube

Thm: Noisy cube is a SSE.

Page 20: Shorter Long Codes and Applications to Unique Games

Better SSEs from Noisy Cube

20

Idea: Find a subgraph of the noisy cube.

Natural approach: Random subsetComplete failure: No edges!

Our Approach: pick a linear codeNeed: bad rate, not too good distance!But not too bad… want local testablity of dual

Page 21: Shorter Long Codes and Applications to Unique Games

Locally Testable Codes

21

Input:Pick

Accept if

TestingDistance: DQuery Comp.: Good soundness:

Parameters

Page 22: Shorter Long Codes and Applications to Unique Games

SSEs from LTCs

22

Given

Page 23: Shorter Long Codes and Applications to Unique Games

Thm: Given If

SSEs from LTCs

23

Symmetry across coordinates.Fraction of non-

zero coordinates.

Page 24: Shorter Long Codes and Applications to Unique Games

Instantiate with Reed-Muller (RM) CodesC = RM code of degree Dual = RM of degree Testability: Batthacharya-Kopparty-

Schoenbeck-Sudan-Zuckerman’10

SSEs from RM Codes

24

Thm: Graph has vertices and large eigenvalues and is a SSE.

Vertices: degree poly’s over Edges: if where affine functions.

Page 25: Shorter Long Codes and Applications to Unique Games

Analyzing expansion

25

When do small sets expand?

Need: Indicators of small sets are far from span of top eigenvectors

First analyze noisy cube.

Page 26: Shorter Long Codes and Applications to Unique Games

Analyzing expansion for noisy cube

26

Is (essentially) a Cayley graph.

Eigenvectors: Characters of

Hamming weight

Eige

nval

ues

0 1 2

N eigenvalues Exponential decay: Large eig. -> weight small

Need: Indicators of small sets far from span of low-weight characters

Follows from (2,4)-hypercontractivity!

Page 27: Shorter Long Codes and Applications to Unique Games

SSEs from LTCs

27

Eigenvecs -> CharactersLarge eval -> low-weight(2,4)-Hypercontractivity

Cayley GraphLocal TestabilityK-wise independence

SSE for Noisy Cube SSE for

Page 28: Shorter Long Codes and Applications to Unique Games

N eigenvalues Threshold decay: Large eig. -> “weight” small

28

Edges of :

A Cayley graph!

Eige

nval

ues

0 1 2

Proof of Expansion

Smoothness, low query com. of Soundness of

𝜒𝑤 (𝑥 )= (−1 )⟨𝑤 , 𝑥 ⟩

Page 29: Shorter Long Codes and Applications to Unique Games

Proof of Expansion

29

Eigenvecs -> CharactersLarge eval -> low-weight(2,4)-Hypercontractivity

Cayley GraphLocal TestabilityK-wise independence

SSE for Noisy Cube SSE for

Fact: is (D-1)-wise independent. QED.

Page 30: Shorter Long Codes and Applications to Unique Games

Open Problems

30

Prove/refute the UGCProof: Larger alphabets?Refute: Need new algorithmic ideas or maybe

stronger SDP hierarchies

Question: Integrality gaps for rounds of Lasserre hierarchy?

Very recent work - Barak, Harrow, Kelner, Steurer, Zhou : Lasserre(8) breaks current instances!

Page 31: Shorter Long Codes and Applications to Unique Games

Open Problems

31

Is ABS bound for SSE tight?Need better LTCs

Page 32: Shorter Long Codes and Applications to Unique Games

32

Thank you

Page 33: Shorter Long Codes and Applications to Unique Games

Long Code d-Short Code

Dict. testing: Noisy cube Dict. testing: RM testerAnalysis: Maj. is stablest Analysis: SSE, Maj. is

stablest

Take Home …

33

Using Long code? Try the “Short code” …

Page 34: Shorter Long Codes and Applications to Unique Games

Sketch for Other Applicatons

34

Dictatorship testing for long code/noisy cube[Kahn-Kalai-Linial’88, Friedgut’98,

Bourgain’99, Mossel-O’Donnel-Oleszkiewicz’05], ...

Focus on MOO: Majority is StablestInvariance principle for low-degree

polynomials

Page 35: Shorter Long Codes and Applications to Unique Games

35

P multilinear, no variable influential.

MOO’05: Invariance principle for Polynomials

Need . Can’t prove in general … … but true for RM code!RM codes fool polynomial threshold functions

PRG for PTFs [M., Zuckerman 10].

Corollary: Majority is stablest over RM codes.Corollary: Alphabet reduction with quasi-polynomial blowup.

Page 36: Shorter Long Codes and Applications to Unique Games

Integrality Gaps for Unique Games, MAX-CUT

36

Idea: Noisy cube -> RM graph in [Khot-Vishnoi’04], [KKMO’05] etc.,

Analyze via Raghavendra-Steurer’09Thm: vertex Max-Cut instance resisting:

rounds in SDP+SA (compare to ))