approximations for isoperimetric and spectral profile and related parameters
DESCRIPTION
Approximations for Isoperimetric and Spectral Profile and Related Parameters. S. Prasad Raghavendra MSR New England. joint work with. David Steurer Princeton University. Prasad Tetali Georgia Tech. Graph Expansion. d -regular graph G with n vertices. d. - PowerPoint PPT PresentationTRANSCRIPT
Approximations for Isoperimetric and
Spectral Profile and Related Parameters
Prasad RaghavendraMSR New England
S
David SteurerPrinceton University
Prasad TetaliGeorgia Tech
joint work with
Graph Expansion
d-regular graph Gwith n vertices
d
expansion(S) = # edges leaving S
d |S|
vertex set S
A random neighbor of a random vertex in S is outside of S with probability expansion(S)
ФG = expansion(S)minimum|S| ≤ n/2
Conductance of Graph G
Uniform Sparsest Cut Problem Given a graph G compute ФG and find the set S achieving it.
Extremely well-studied, many different contexts pseudo-randomness, group theory, online routing, Markov chains, metric
embeddings, …
S
Measuring Graph Expansion
d-regular graph Gwith n vertices
d
expansion(S) = # edges leaving S
d |S|
vertex set S ФG = expansion(S)minimum|S| ≤ n/2
Conductance of Graph G
CompleteGraph
CompleteGraph
Path Complete graphs with a perfect matching
Typically, small sets expand to a greater extent.
Isoperimetric Profile
Ф(δ) = expansion(S)minimum|S| ≤ δn
Isoperimetric Profile of Graph G
expansion(S) = # edges leaving S
d |S|
• Conductance function – defined by [Lovasz-Kannan] used to obtain better mixing time bounds for Markov chains.
•Decreasing function of δSet Size δ 0.5
Ф(δ)
1
S
Approximating Isoperimetric Profile
Ф(δ) = expansion(S)minimum|S| ≤ δn
Isoperimetric Profile of Graph G
expansion(S) = # edges leaving S
d |S|
S
Uniform Sparsest Cut: Determine the lowest point on the curve.
Set Size δ
0.5
Ф(δ)
1
What is the value of Ф(δ) for a given graph G and a constant δ > 0? Gap Small Set Expansion Problem (GapSSE)(η, δ)Given a graph G and constants δ, η > 0,
Is Ф(δ) < η OR Ф(δ) > 1- η?----Closely tied to Unique Games Conjecture
(Last talk of the day “Graph Expansion and the Unique Games Conjecture” [R-Steurer 10])
AlgorithmTheorem
For every constant δ > 0, there exists a poly-time algorithm that finds a set S of size O(δ) such that
expansion (S) ≤
))/1log()(( O
-- A (Ф(δ) vs ) approximation
)/1log()(
• For small enough δ, the algorithm cannot distinguish betweenФ(δ) < η OR Ф(δ) > 1- η
Theorem [R-Steurer-Tulsiani 10]An improvement over above algorithm by better than constant
factor, would help distinguish Ф(δ) < η OR Ф(δ) > 1- η
A Spectral RelaxationLet x = (x1 , x2 , .. xn) be the indicator function of the unknown least expanding set S
S
SC
1
0
Eji
ji xx),(
2)(Number of edges leaving S =
i
ix2|S| =
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
}1,0{2),(
2
}1,0{min
)(min
|Support(x)| < δn
)(nx nx
Relaxing 0,1 to real numbers
Why is it spectral?Spectral Profile: [Goel-Montenegro-Tetali]
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
min)(
min)( 2),(
2
xxdLxx
xd
xx
T
T
xx
ii
Ejiji
xx nn
1
2),(
2
1
min)(
min
Smallest Eigen Value of Laplacian:
|Support(x)| < δn |Support(x)| < δn
Observation: Λ(δ) is the smallest possible eigenvalue of a submatrix L(S,S) of size at most < δn of the Laplacian L .
Rounding EigenvectorsCheeger’s inequality There is a sparse cut of value at mostxxd
LxxT
T
xx n
1
min
Smallest Eigen Value of Laplacian
2
Rounding x
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
min)(
min)( 2),(
2
|Support(x)| < δn |Support(x)| < δn
Lemma: There exists a set S of volume at most δ whose expansion is at most )(2
0
Spectral Profile [Goel-Montenegro-Tetali]
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
min)(
min)( 2),(
2
|Support(x)| < δn |Support(x)| < δn
Unlike eigen values, Λ(δ) is not the optimum of a convex program. We show an efficiently computable SDP that gives good guarantee.
Theorem: There exists an efficiently computable SDP relaxation Λ* (δ) of Λ(δ) such that for every graph G
Λ* (δ) ≤ Λ(δ) ≤ Λ* (δ/2) O(log 1/∙ δ)
RecapTheorem: (Approximating Spectral Profile)There exists an efficiently computable SDP relaxation Λ* (δ) of Λ(δ) such that for every graph G
Λ* (δ) ≤ Λ(δ) ≤ Λ* (δ/2) O(log 1/∙ δ)
Lemma: (Cheeger Style Rounding)There exists a set S of volume at most δ whose expansion is at most )(2
Theorem (Approximating Isoperimetric Profile)For every constant δ > 0, there exists a poly-time algorithm that finds a set S of size O(δ) such that
expansion (S) ≤
))/1log()(( O
Restricted Eigenvalue Problem
Given a matrix A, find a submatrix A[S,S] of size at most δn X δn matrix with the least eigenvalue.
-- Our algorithm is applicable to diagonally dominant matrices
(yields a log(1/δ) approximation).
Approximating Spectral Profile
ii
Ejiji
x xd
xx
n 2),(
2)(min)(
|Support(x)| < δn
SDP Relaxation for
ii
ii xxSupportx 22
|)(|
222
|| nvnvi
ii
i
ii
Ejiji
x xd
xx
n 2),(
2)(min)(
|Support(x)| < δn
Replace each xi by a vector vi:
Eji
ji vvNumerator),(
2||
Ejiiv
),(
2||Denominator = n
Without loss of generality, xi can be assumed positive.This yields the constraint: vi v∙ j ≥ 0
Enforcing Sparsity:
By Cauchy-Schwartz inequality:
SDP Relaxation for
2
,
nvvji
ji
ii
Ejiji
x xd
xx
n 2),(
2)(min)(
|Support(x)| < δn
Minimize (Sum of Squared Edge Lengths)
Subject to
Eji
ji vv),(
2||
nvEjii
),(
2||
Positive Inner Products: vi v∙ j ≥ 0
Average squared length =1
Average Pairwise Correlation < δ
Rounding
Two Phase Rounding:
• Transform SDP vectors in to a SDP solution with only non-negative coordinates.
• Use thresholding to convert non-negative vectors in to sparse vectors.
Making SDP solution nonnegativeLet v be a n-dimensional real vector.Let v* denote the unit vector along direction v.
Map the vector v to the following function over Rn :fv = ||v|| (Square Root of Probability Density Function∙
of n-dimensional Gaussian centered at v*)
)()( *vxvxfv Formally,
Where: Ф(x) = probability density function of a
mean 0, variance σ spherical Gaussian in n-dimensions.
Properties
24/1
2
),(
)(||2
nefEjii
)()( *vxvxfv Where:
Ф(x) = probability density function of a mean 0, variance σ spherical Gaussian in n-dimensions.
Lemma: (Pairwise correlation remains low if σ is small)
2
,
nvvji
ji Average Pairwise Correlation < δSDP Constraint
Pick σ = 1/sqrt{log(1/δ)}
Properties
Lemma: (Squared Distances get stretched by at most 1/σ2) |f1 – f2 |2 ≤ O(1/σ2) |v1 – v2 |2
)()( *vxvxfv Where:
Ф(x) = probability density function of a mean 0, variance σ spherical Gaussian in n-dimensions.
For our choice of σ, squared distances are stretched by log(1/δ) .
With a log(1/δ) factor loss, we obtaining a non-negative SDP solution.
Rounding a positive vector solutionLet us pretend the vectors vi are non-negative. i.e., vi(t) ≥ 0 for all t
Rounding Non-Negative Vectors1. Sample t2. Compute threshold
θ = (average of vi(t) over i) * (2/ δ )3. Set xi = max{ vi(t) – θ, 0 } for all i
Observation: |Support(x)| < δ/2 n∙
Observation:
Eji
jiji vvxxE),(
22 ])([
Open Problem
Find integrality gaps for the SDP relaxation:
Current best have δ = 1/poly(logn),
Do there exist integrality gaps with δ = 1/poly(n)?
2
,
nvvji
ji
Minimize (Sum of Squared Edge Lengths)
Subject to
Eji
ji vv),(
2||
nvEjii
),(
2||
Positive Inner Products: vi v∙ j ≥ 0
Average squared length =1
Average Pairwise Correlation < δ
Thank You
Spectral ProfileSecond Eigen Value:
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
min)(
min 2),(
2
|Support(x)| < δn
)(
Spectral Profile
Remarks:• Λ(δ) ≤ minimum of Ф(δ0) over all δ0 ≤ δ• Unlike second eigen value, Λ(δ) is not the optimum of a convex program.• Λ(δ) is the smallest possible eigenvalue of a submatrix of size at most < δn of the matrix L • xi can be assumed to be all positive.
Small Sets via Spectral Profile
Using an analysis along the lines of analysis of Cheeger’s inequality, it yields:
Theorem [Raghavendra-Steurer-Tetali 10]There exists a poly-time algorithm that finds a set S of
size O(δ) such thatexpansion (S) ≤ sqrt (Λ* (δ) O(log 1/∙ δ))
So if there is a set S with expansion(S) = ε, thenthe algorithm finds a cut of size )/1log( O
Similar behaviour as the Gaussian expansion profile
Spectral ProfileSecond Eigen Value:
xxdLxx
xd
xx
T
T
x
ii
Ejiji
x nn
min)(
min 2),(
2
|Support(x)| < δn
)(
Spectral Profile
Replace xi by vectors vi
Subject tovi v∙ j ≥ 0
i
ii
i vnv 22
||
ii
Ejiji
v vd
vv
i2
),(
2
*
||
||min)(
Graph Expansion
d-regular graph G
d
expansion(S) = # edges leaving S
d |S|
vertex set S
A random neighbor of a random vertex in S is outside of S with probability expansion(S)
ФG = expansion(S)minimum|S| ≤ n/2
Conductance of Graph G
Uniform Sparsest Cut Problem Given a graph G compute ФG and find the set S achieving it.
Approximation Algorithms:
•Cheeger’s Inequality [Alon][Alon-Milman] Given a graph G, if the normalized adjacency matrix has second eigen value λ2 then,
•A log n approximation algorithm [Leighton-Rao].•A sqrt(log n) approximation algorithm using semidefinite programming [Arora-Rao-Vazirani].
)1(22)1(
22
G
Extremely well-studied, many different contexts
pseudo-randomness, group theory, online routing,
Markov chains, metric embeddings, …
Limitations of Eigenvalues• The best lower bound that Cheeger’s inequality gives on expansion is
(1-λ2)/2 < ½, while Ф(δ) can be close to 1.
• Consider graph G Connect pairs of points on {0,1}n that are εn Hamming distance away.Then Second eigenvalue ≈ 1- ε and ф(1/2) ≈ εyet Ф(δ) ≈ 1 (small sets have near-perfect expansion)
• A SIMPLE SDP RELAXATION cannot distinguish between- all small sets expand almost completely- exists small set with almost no expansion
A Conjecture
Small-Set Expansion Conjecture:
8η>0, 9 ± >0 such that GapSSE(η, ± ) is NP-hard, i.e.,
Given a graph G, it is NP-hard to distinguish YES: expansion(S) < η for some S of volume ¼ ±NO: expansion(S) > 1- η for all S of volume ¼ ±
Road Map
• Algorithm:– Spectral Profile.
• Reductions within Expansion
• Relationship with Unique Games Conjecture
Gaussian Curve
Gaussian Graph
Vertices: all points in Rd (d dimensional real space)
Edges: Weights according to the following sampling procedure:
• Sample a random Gaussian variable x in Rd
• Perturb x as follows to get y in Rd
Add an edge between x an y
zxy 22)1(
Γε (δ) = Gaussian noise sensitivity of a set of measure δ
= least expanding sets are caps/thresholds of measure δ
= )/1log(
Set Size δ 0.5
Ф(δ)
1
Approximating Expansion Profile
Reductions within Expansion
Reductions within Expansion
Theorem [Raghavendra-Steurer-Tulsiani 10]For every positive integer q, and constants ε,δ,γ, given
a graph it is SSE-hard to distinguish between:
• There exists q disjoint small sets S1 , S2 , .. Sq of size close to 1/q, such that expansion(Si) ≤ ε
• No set of size μ> δ has expansion less than size Γε/2 (μ) -- expansion of set of size μ in Gaussian graph with parameter ε/2
expansion (S) < sqrt (Λ* (δ) O(log 1/∙ μ))
Informal Statement
Set Size δ
1
Set Size δ 0.5
Ф(δ)
1
QualitativeAssumption
GapSSE is NPhard
Quantitative Statement
Given a graph G , it is SSE-hard to distinguish whether, •There is a small set of size whose expansion is ε.
•Every small set of size μ (in a certain range) expands at least as much as the corresponding Gaussian graph with noise ε )/1log( O
Corollaries
Corollary: The algorithm in [Raghavendra-Steurer Tetali 10] has near-optimal guarantee assuming SSE conjecture.
Corollary: Assuming GapSSE, there is no constant factor approximation for Balanced Separator or Uniform Sparsest Cut.
Relation with Unique Games Conjecture
Unique GamesUnique game ¡ :
Each edge (A,B) has an associated map ¼ ¼ is bijection from L(A) to L(B)
graph of size n
AB
label set L(A) ¼
Goal: Find an assignment of labels to vertices such that maximum number of edges are satisfied.
A labelling satisfies (A,B) if ¼ (label of A) = label of B
Referee
sample (A,B,¼)
BA
Player 1 Player 2pick b in L(B)pick a in L(A)
ba
Refereeplayers win if ¼(a) = b
no communication between players
value( ¡ ):maximum success probabilityover all strategies of the players
Unique Games Conjecture [Khot02]
Unique Games Conjecture: [Khot ‘02]8²>0, 9 q >0:
NP-hard to distinguish for ¡ with label set size qYES: value( ¡ ) > 1-² NO: value( ¡ ) < ²
Implications of UGC
UGC
BASIC SDP is optimal for …
Constraint Satisfaction Problems [Raghavendra`08]MAX CUT, MAX 2SAT
Ordering CSPs [GMR`08]MAX ACYCLIC SUBGRAPH, BETWEENESS
Grothendieck Problems [KNS`08, RS`09]
Metric Labeling Problems [MNRS`08]MULTIWAY CUT, 0-EXTENSION
Kernel Clustering Problems [KN`08,10]
Strict Monotone CSPs [KMTV`10]VERTEX COVER, HYPERGRAPH VERTEX COVER
…
If we could show , then a refutation of UGC would imply an improved algorithm for PROBLEM X
“Reverse Reductions”
UGCBASIC SDP is optimal forlots of optimization problems,e.g.: MAX CUT and VERTEX COVER
BASIC SDP optimal
for PROBLEM X
?*
*
Parallel Repetition is natural candidate reduction for [FeigeKinderO’Donnell’07]*
Win-Win Situation
Bad news: this reduction cannot work [Raz’08, BHHRRS’08]
PROBLEM X = MAX CUT
Small Set Expansion and Unique Games
• Solving Unique Games Finding a small non-expanding set in the “label extended graph”
Theorem [Raghavendra-Steurer 10]
Small Set Expansion Conjecture Unique Games Conjecture
Establishes a reverse connection from a natural problem.
Implications of UGC
UGC
BASIC SDP is optimal for …
Constraint Satisfaction Problems [Raghavendra`08]MAX CUT, MAX 2SAT
Ordering CSPs [GMR`08]MAX ACYCLIC SUBGRAPH, BETWEENESS
Grothendieck Problems [KNS`08, RS`09]
Metric Labeling Problems [MNRS`08]MULTIWAY CUT, 0-EXTENSION
Kernel Clustering Problems [KN`08,10]
Strict Monotone CSPs [KMTV`10]VERTEX COVER, HYPERGRAPH VERTEX COVER
…
Uniform Sparsest Cut [KNS`08, RS`09]
Minimum Linear Arrangement[KNS`08, RS`09]
Gap SSE
UGCWith expansion
Most known SDP integrality gap instances for problems like MaxCut, Vertex Cover, Unique games have graphs that are “small set expanders”
Theorem [Raghavendra-Steurer-Tulsiani 10]
Small Set Expansion Conjecture MaxCut or Unique Games on Small Set Expanders is hard.
Reverse Connections?
Approximating Spectral Profile
RoadmapIntroductionGraph Expansion: Cheeger’s Inequality, Leighton Rao, ARVExpansion Profile: Small Sets expand more than large ones.Cheeger’s inequality and SDPs failGapSSE ProblemRelation to Unique Games Conjecture: Unique Games definition, Applications, lack of reverse reductions. Label extended graph Small setsSmall Set Expansion Conjecture UGCUGC with SSE is easyAlgorithm for SSE: Spectral Profile, SDP for Spectral Profile, Rounding algorithmRelations within expansion:GapSSE Balanced Separator Hardness
NP-hard OptimizationExample:
MAX CUT: partition vertices of a graph into two setsso as to maximize number of cut edges
fundamental graph partitioning problem
benchmark for algorithmic techniques
ApproximationMAX CUT
Trivial approximation
First non-trivial approximation
based on a semidefinite relaxation (BASIC SDP)
Random assignment, cut ½ of the edges
Goemans–Williamson algorithm (‘95)
approx-ratio 0.878
Beyond Max Cut:
Analogous BASIC SDP relaxation for many other problems
Almost always, BASIC SDP gives best known approximation(often strictly better than non-SDP methods)
ApproximationMAX CUT
Trivial approximation
based on a semidefinite relaxation (BASIC SDP)
Random assignment, cut ½ of the edges
Goemans–Williamson algorithm (‘95)
approx-ratio 0.878
Can we beat the approximation guarantee of BASIC SDP?
First non-trivial approximation
Unique Games Conjecture
Unique Games Conjecture [Khot’02] (roughly):it is NP-hard to approximate the value of Unique Games
certain optimization problem:given equations of the form
xi – xj = cij mod qsatisfy as many as possible
Can we beat the approximation guarantee of BASIC SDP for Max Cut?
No, assuming Khot’s Unique Games Conjecture! [KKMO`05, MOO`05, OW’08]
UGC: Is it true?
UGC
?
UGCon expanding
instances
UGCon product instances
[AKKSTV’08, AIMS’10]
[BHHRRS’08,Raz’08]
UGCfor certain SDP
hierarchies
[RaghavendraS’09,KhotSaket’09]
Any non-trivial consequences if UGC is false?
What are hard instances?
“Reverse Reductions”
UGCBASIC SDP is optimal forlots of optimization problems,e.g.: MAX CUT and VERTEX COVER
Win-Win SituationA refutation of UGC implies an improved algorithm for SMALL-SET EXPANSION (better than BASIC SDP!)
SMALL-SET EXPANSION
First reduction from natural combinatorial problem to UNIQUE GAMES
BASIC SDP optimal
for PROBLEM X
Approximating Small-Set Expansion
Expansion profile of G at ±:minimum expansion(S) over all S with volume < ±
SHow well can we approximate the
expansion profile of G for small ± ?
BASIC SDP cannot distinguish between- all small sets expand almost completely- exists small set with almost no expansion
Small-Set Expansion Conjecture:8²>0, 9 ± >0:
NP-hard to distinguish YES: expansion(S) < ² for some S of volume ¼ ±NO: expansion(S) > 1-² for all S of volume ¼ ±
Unique 2-Prover Games
Referee
sample (A,B,¼) from DBA
Player 1 Player 2pick b in L(B)pick a in L(A)
ba
Refereeplayers win if ¼(a) = b
Unique game ¡ :
no communication between players
value( ¡ ):maximum success probabilityover all strategies of the players
distribution D over triples (A,B,¼)- A and B are from U- ¼ is bijection from L(A) to L(B)
universe of size n
AB
label set L(A)
label set L(B)
¼
Approximating Unique Games
Unique Games Conjecture: [Khot ‘02]
8²>0, 9 q >0:NP-hard to distinguish for ¡ with label set size q
YES: value( ¡ ) > 1-² NO: value( ¡ ) < ²
How well can we approximate the value of a unique gamefor large label sets?
BASIC SDP cannot distinguish between games with value ¼ 1 and ¼ 0for large label sets
Approximating Unique Games
Unique Games Conjecture:8²>0, 9 q >0:
NP-hard to distinguish for ¡ with label set size qYES: value( ¡ ) > 1-² NO: value( ¡ ) < ²
Our main theorem:
Small-Set Expansion Conjecture ) Unique Games Conjecture
Reduction: Small-Set Expansion Unique Games
Task: find non-expanding set of volume ¼ ±
graph G
A
B
Refereesample R = 1/± random edges M
A = one half of each edgeB = other half of each edge
BA
Player 1 Player 2pick b 2 Bpick a 2 A
ba
Refereeplayers win if (a,b) 2 M
P( ) < ²
CompletenessSmall Non-Expanding Set (Partial) Strategy
graph G
AB
S
Suppose expansion(S) < ²
Strategy for Player 1:pick a 2 A if a is unique intersection with Sotherwise, refuse to answer
With constant probability, |A Å S |= {a}Conditioned on this event:
other half of a’s edge outside of S
partial game value > 1 - ²
referee allows players to refuse for few queries
SoundnessStrategy Small Non-Expanding Set
standard trick: can assume both players have same strategy
Suppose:players win with prob > 1-²
Idea: strategy distribution over sets- sample R-1 vertices U- output S = { x | Player 1 picks x if A=U+x }
Easy to show:
E volume(S) = 1/R = ±E # edges leaving S
E d |S|< ²
Problem:volume(S) might not be concentrated around ±
Are we done?No!
SoundnessStrategy Small Non-Expanding Set
Idea: Referee adds ²-noise to A and B+ ²-noise+ ²-noise
New distribution over sets:- sample R-1 vertices U- S = { x | players pick x if A = U+x+noise
with probability > ½ }
Intuition:players cannot distinguish x and noise
Can show: 8 U. volume(S) <
= ± / ²
² R random vertices
# noise vertices1
Suppose:players win with prob > 1-²
Summary
UGCBASIC SDP is optimal forlots of optimization problems,e.g.: MAX CUT and VERTEX COVER
BASIC SDP optimal for SMALL-SET
EXPANSION
more reverse reductions?
Open Questions (+ Subsequent Work)
Hardness results based on Small-Set Expansion Conjecture? Reductions between Expansion Problems [Raghavendra S Tulsiani’10]
Thanks!Questions?
noise here vs. noise in other hardness reductions?
Better algorithms for UNIQUE GAMES via SMALL-SET EXPANSION? 2npoly(²) algorithm for (1-², ½)-UG via SSE [Arora Barak S’10]
APPENDIX
Rounding a UG strategy to a small non-expanding set
Unique Games
decorated bipartite graph
uv
label set L(v)label set L(u)
bijection ¼uv: L(u) L(v)
Referee
random edge (u,v)
vu
Player 1 Player 2pick j in L(v)pick i in L(u)
ji
Refereeplayers win if ¼uv(i) = j
Unique Games instance ¡ :
each player knows only half of referee’s edge
value( ¡ ):maximum success probabilityover all strategies of the players
ApproximationMAX CUT
Trivial approximation
First non-trivial approximation
based on a semidefinite relaxation (BASIC SDP)
Random assignment, cut ½ of the edges
Goemans–Williamson algorithm (‘95)
approx-ratio 0.878
General constraint satisfaction problems (CSPs):
for every CSP: approximation matches integrality gap of certain SDPnear-linear running time [S’10]
simple generic approximation algorithm [RaghavendraS’09]
Can we beat the approximation guarantee of BASIC SDP (for Max Cut)?
Combinatorial OptimizationExample:
MAX CUT: partition vertices of a graph into two setsso as to maximize number of cut edges
CAIDA at UCSD analyzes business relationshipsamong Autonomous Systems in the internet using MAX 2SAT
¼ MAX CUTA concrete practical application: