Hierarchy Higher Eigenvalues and Graph Partitioning Venkatesan Guruswami Carnegie Mellon University Mysore Park Workshop August 10 2012 Joint work with Ali Kemal Sinop ID: 564844
Download Presentation The PPT/PDF document "Lasserre" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Lasserre Hierarchy, Higher Eigenvalues, and Graph Partitioning
Venkatesan GuruswamiCarnegie Mellon University
--- Mysore Park Workshop, August 10, 2012 ---
Joint work with
Ali
Kemal
SinopSlide2
Talk OutlineIntroduction to problems we studyLaplacian eigenvalues and our resultsLasserre hierarchyCase study: Minimum bisectionConcluding remarksSlide3
Graph partitioning problems Minimum bisection: Given edge-weighted graph G=(V,E,W), find partition of vertices into two equal parts cutting as few (in weight) edges as possible:More generally: Minimum -section
Find subset S V of size to minimize cut size G(S) = weight of edges leaving S = |E(S,Sc)|Related problems:Small set expansion: weight vertices by degree, find
S V minimizing G
(S)
with
Vol
(S)= sum of degrees in S =
Sparsest cut (find best ratio cut over all sizes)
1
2
3
4
1
2
3
4
Cost=2Slide4
Many Practical ApplicationsBuilding block for divide-and-conquer on graphsVLSI layoutPacket routing in distributed networks Clustering and image segmentationRoboticsScientific computingSlide5
Approximation AlgorithmsUnfortunately such cut problems are NP-hard. Find an α-factor approximation instead.If minimum cost = OPT, algorithm always finds a solution with value
≤ α OPT.Rounding Algorithm: Solve a convex relaxation and round the “fractional” solution to “integral” solution.
OPT
Algorithm
α
OPT
0
Relaxation
Integrality Gap
Slide6
A notorious problem: Unique GamesGraph G=(V,E)Number of labels k For each edge e=(u,v), A permutationGoal: Label vertices with k colors to minimize number of unsatisfied edges Slide7
Unique Games: ExampleSuppose k=3.
1
2
3
1
2
3
1
2
3
1
2
3
Label Extended Graph
Constraint Graph
A cloud of k=3 vertices
per vertex of GSlide8
Example
A labeling:
1
2
3
1
2
3
1
2
3
1
2
3
Label Extended Graph
Constraint GraphSlide9
Unique Games: ExampleUnsatisfied constraints:(in red’s neighborhood)
1
2
3
1
2
3
1
2
3
1
2
3
Label Extended Graph
Constraint GraphSlide10
Unique Games: ExampleUnsatisfied constraints:(in red’s neighborhood)
1
2
3
1
2
3
1
2
3
1
2
3
Label Extended Graph
Constraint GraphSlide11
Unique Games = Special sparse cut
1
2
3
1
2
3
1
2
3
1
2
3
Label Extended Graph
Constraint Graph
Find subset S of fraction 1/k vertices
in label extended graph,
containing
one vertex in each cloud,
minimizing
(S)Slide12
Unique Games conjecture [Khot’02] > 0
k = k() s.t. it is NP-hard to tell if an instance of Unique Games with k labels has OPT
or
OPT
1-
.
Let OPT = fraction of unsatisfied edges in optimal labeling
i.e., even if a “special cut” with expansion ,
it is hard to find a “special cut” with expansion 1- Slide13
Our workApproximation algorithms for these problems using semidefinite programs from the Lasserre hierarchy
Min-BisectionSmall Set ExpansionSparsest CutMin-Uncut / Max-CutIndependent SetAs well as k-partitioning variants:Min-k-SectionUnique GamesSlide14
Motivations: Algorithmic PerspectiveThese are fundamental, well-studied, practically relevant, optimization problemsYet huge gap:Approximability: or Hardness: not even factor 1.1 known to be NP-hard
Natural goal: Close (or reduce) the gapSDPs one of the principal algorithmic toolsExtend algorithmic techniques to more powerful SDPs.Slide15
Motivations: Complexity Perspective[Khot’02] Unique Games ConjectureUGC has many implications: tight results for all constraint satisfaction and ordering problems, vertex cover,… [
Raghavendra, Steurer’10] Small Set Expansion Conj. (SSEC):O(1)-approximation for Small Set Expansion is hard.Implies the UGC [RS’10], plus (1) hardness for bisection, sparsest cut, minimum linear arrangement, etc. [R,S,Tulsiani’12]Despite these bold conjectures, following is not ruled out:Can 5-rounds of Lasserre Hierarchy SDP relaxation
refute the UGC?
Investigate this possibility…
Identify candidate hard instances (if there are any!)
No consensus opinion
on its validitySlide16
Talk OutlineIntroduction to problems we studyLaplacian eigenvalues and our resultsLasserre hierarchy
Case study: Minimum bisectionConcluding remarksSlide17
Laplacian and Graph Spectrum
12
3
4
rows and cols
indexed by V
λ
2
: Measures expansion of the graph through
Cheeger’s
inequality
.
λ
r
: Related to small set expansion
[
Arora
, Barak, Steurer’10]
,
[Louis,Raghavendra,Tetali,Vempala’11], [
Gharan
, Lee,Trevisan’11]
.
0 = λ1 ≤
λ2 ≤ … ≤ λ
n ≤2 and λ1 + λ
2 + … + λn = n,Slide18
Our Results IBy rounding r/O(1) round Lasserre hierarchy SDPs, in time we obtain approximation
factorNote we get approximation scheme0 = λ1 ≤ λ2 ≤ … ≤ λn ≤2 and λ1 + λ2
+ … + λn = n,
Minimum Bisection*
Small Set Expansion*
Uniform Sparsest Cut
Their k-way generalizations*
* Satisfies constraints within factor of 1
o(1)
For r
=n,
λ
r
>1,
λn-r <1
Minimum Uncut (min. version of Max Cut)
More generally, our methods apply to quadratic integer programming problems with positive
semidefinite
objective functionsSlide19
Our Results II: Unique GamesFor Unique Games, a direct bound will involve spectrum of label extended graph, whereas we want to bound using spectrum of original graph.We give a simple embedding and work directly on the original graph.We obtain factor in time
Concurrent to our work, [Barak-Steurer-Raghavendra’11] obtained factor in time (using weaker Sherali-Adams SDP).
Combining with
[Arora-Barak-Steurer’10]
“higher order
Cheeger
”
thm.
UG with
compl. 1- O(1
) is easy for n
levels of Lasserre HierarchySlide20
InterpretationOur results show that these graph partitioning problems are easy on many graphsAlso hints at why showing even weak hardness results has been elusivePoints to the power of
Lasserre hierarchyCould be a serious threat to small set expansion conjecture, or even UGC.Recent work [Barak,Brandao,Harrow,Kelner,Steurer,Zhou’12] shows O(1) rounds enough to solve known gap instancesSlide21
Our Results IIIIndependent set: O(1)
approximation in nO(r) time when r’th
largest
eigenvalue
n-r
1 + O(1/
d
max)
O(d
max/t) approximation if n-r
1 + 1/t
Normalized
Laplacian
eigenvalues
0 = λ1
≤ λ2 ≤ … ≤ λ
n ≤ 2 and λ1 + λ
2 + … + λn = n,
[Arora-Ge’11] Given 3-colorable graph, find independent set of size
n/12 in nO
(r) time if
n-r < 17/16. Slide22
Talk OutlineIntroduction to problems we studyLaplacian eigenvalues
and our resultsLasserre hierarchyCase study: Minimum bisectionConcluding remarksSlide23
Basic Lasserre Hierarchy RelaxationQuadratic IP formulation for a k-labeling problem: For each S of size ≤ r, and each possible labeling f : S {0,1,…,k-1}
Boolean variable representing: with all implied pairwise consistency constraints.(SDP Relaxation) Replace with
r
= number of
rounds/levels
of
Lasserre
hierarchy;
Resulting SDP can be solved in
n
O(r) timeSlide24
ConsistencyLasserre Relaxation for Minimum Bisection
Relaxation for consistent labeling of all subsets of size r:
Marginalization
Distribution
Partition
SIze
Cut cost
Given d-regular graph G, find subset U of size
minimizing
G
(U)
Intended integral value of
x
u
(1) : 1 if u
U, and 0 otherwise.
MinimizeSlide25
Intuition Behind Lasserre RelaxationFor each S, the vectors xS(f) give a local distribution on labelings f : S
{0,1}Prob. of f = Inner products of vectors (xS(f))S,f represent joint probabilities (give a psd moment matrix)Division corresponds to conditioning:Slide26
Previous Work on Lasserre HierarchyFew algorithmic results known before, including:[Chlamtac’07], [Chlamtac, Singh’08]
nΩ(1) approximation for 3-coloring and independent set on 3-uniform hypergraphs, [Karlin, Mathieu, Nguyen’10] (1+1/r) approximation of knapsack for r-rounds.Some known integrality gaps are:[Schoenebeck’08], [Tulsiani’09] Most NP-hardness results carry over to Ω(n) rounds of Lasserre.[
Bhaskara, Charikar
, G.,
Vijayaraghavan
, Zhou’12]
Densest k-
subgraph (n(1) integrality gap for (n) rounds of Lasserre
hierarchy)Slide27
Rounding Lasserre RelaxationFor regular SDP [Goemans, Williamson’95] showed that with
hyperplane rounding:No analogue known for rounding Lasserre RelaxationHere: an intuitive local propagation based rounding frameworkAnalysis via projection distance, and connections to ``column selection” in low-rank matrix approximationSlide28
General Rounding Framework
(Seed Selection) Choose appropriate seed set S.(Seed Labeling) Choose wp . (Propagation) Perform randomized rounding.so that the output satisfies:(i.e., match the conditional prob. for label for i, given S got labeling f)
Insp
ired
by
[
Arora
,
Kolla
, Khot, Steurer
, Tulsiani, V
ishnoi’08] algorithm for Unique Games on expanders: propagation from a single node chosen uniformly at randomSlide29
Talk OutlineIntroduction to problems we studyLaplacian eigenvalues
and our resultsLasserre hierarchyCase study: Minimum bisectionConcluding remarksSlide30
Case Study: Minimum BisectionWe will present some details of the analysis of rounding for the minimum bisection problem on d-regular unweighted graphs (for simplicity).We will show that it achieves factor .
Obtaining factor requires some additional ideas.Slide31
Lasserre SDP for Min -section
Vector xS(f)
for each S of size ≤ r, and each possible labeling f : S
{0,1}
MinimizeSlide32
Rounding AlgorithmGiven optimal solution to r’=O(r) round Lasserre SDP:Choose suitable seed set
S of size rDetails laterPartition S by choosing f with probability Propagate to other nodes:For each node v independentlyWith probability include v in U.
Return U.Slide33
AnalysisPartition SizeEach node is chosen into U independentlyFixing S,f, expected size of U equalsBy Chernoff
, with high probability 33Slide34
AnalysisBy our rounding, for fixed seed set S,After some calculations, we have the following bound on number of edges cut:
Normalized Vector for xS(f)
≤ OPT
Call this matrix
SSlide35
Matrix ΠSRemember {x
S(f)}f are orthogonal. is a projection matrix onto span{xS(f)}f .For any
Let P
S
be the corresponding projection matrix.Slide36
Picking the seed setThe final bound is:Define X = matrix with columns {Xu = xu(1)}
Want seed set S to minimize Projection distance to span of columns { X
u : u S}Slide37
Column selectionGiven matrix X Rm x n, pick r columns S to minimize
Introduced by [Frieze, Kannan,Vempala’04]; studied in many works since.For any S of size r this is lower bounded by:[G.-Sinop] Can efficiently find set S of columns so thatAnd this bound is tight.
Error of best rank-r approximation of XSlide38
Relating Performance to Graph SpectrumCan show: Worst case is when best rank-r approximation of X is obtained by first r eigenvectors of graph Laplacian.Using Courant-Fischer theorem,ThereforeSlide39
Few words about column selectionGiven matrix X Rm x n, pick t
columns S to minimize[G.-Sinop] Can efficiently find set S of columns so that
Error of best rank-r approximation of XSlide40
Proof Idea Goal: (Min. projection dist.)Observe Choose S with probability Volume Sampling [Deshpande
, Rademacher, Vempala, Wang’06]Converts sum-of-ratios to ratio-of-sums.Slide41
Proof Idea (contd.) where are eigenvalues of XT X, is rth
symmetric form. Expected projection distance achieved by volume sampling equalsSlide42
Schur Concavity is a Schur-Concave function.F() F() if majorizes For a fixed prefix sum
F() is maximized bySubstituting back:
Best rank-r approximation
error.Slide43
Talk OutlineIntroduction to problems we studyLaplacian eigenvalues
and our resultsLasserre hierarchyCase study: Minimum bisectionConcluding remarksSlide44
SummaryRounding for Lasserre hierarchy SDPs for certain QIPs + analysis based on column selectionApproximation scheme-like guarantees for several graph partitioning problemsnO(r) time to solve r-levels of hierarchy. Rounding framework only looks at 2
O(r) nO(1) bits of solution. Can also make runtime 2O(r) nO(1) [G.-Sinop, FOCS’12]Lasserre SDP seems very powerfulOnly very weak integrality gaps known for the studied problems Slide45
Open questionsCan O(log n) rounds of Lasserre hierarchy refute SSE conjecture? Refute UGC? Currently no candidate hard instances for even 5 rounds
0.878 approx. for Max-Bisection? (0.85 [Raghavendra-Tan’12])Integrality gaps for Lasserre hierarchy beating NP-hardness (or matching UG/SSE-hardness) results [Tulsiani’09] For Max k-CSP, clique/coloring[G.-Sinop-Zhou’12] Balanced separator, uniform sparsest cut
[Bhaskara,
Charikar
, G.,
Vijayaraghavan
, Zhou’12]
Densest k-subgraph (n
(1) integrality gap for (n) rounds of
Lasserre hierarchy)