Sum of Squares Lower
Bounds for the Planted
Clique Problem
Aaron Potechin MIT
In collaboration with Boaz Barak, Siu On Chan,
Jonathan Kelner, Raghu Meka, David Steurer, and Avi Wigderson
Talk Outline
• Part I: Planted Clique
• Part II: The Sum of Squares/Lasserre Hierarchy
• Part III: Analyzing the Meka-Wigderson Moment
Part I:
Max-Clique
Largest subset of vertices with all edges present
Classical optimization problem
Worst Case Complexity of
Max-Clique
• NP-hard (made Karp’s list) • NP-hard to approximate by
-Hastad 99, Zuckerman 06
• Unconditional hardness in various models
-Razborov 85 -…
-Tulsiani 09
Cliques in G(n,1/2)
• Largest clique has size roughly • Easy to find cliques of size .
The Planted Clique
Problem
• G(n,1/2) + clique(k)
• Jerrum 92, Kucera 95:
For which k can we find the planted clique?
• Best- Alon et al. 98:
The Planted Clique
Problem
• G(n,1/2) + clique(k)
• Jerrum 92, Kucera 95:
For which k can we find the planted clique?
• Best- Alon et al. 98:
This 5-clique was planted by adding the red edge.
Square-root Barrier
• Jerrum 92: Can’t do using MCMC
(Monte-Carlo Markov Chains)
• Feige-Krauthgamer 00,03: Using
(the rth level of the Lovasz-Schrijver hierarchy) -Can if k is
-Can’t if k is
• Feldman et al. 12: Can’t do using statistical
Why Planted Clique?
• Natural algorithmic problem and test bed
-Spectral algorithms, clustering
• Natural average case hardness candidate
Cryptosystems - [Juels, Peinado 00]
Nash equilibria - [Hazan, Krauthgamer 11]
• Variable for each vertex i in G. • Want if i is not in the clique
• Want if i is in the clique. • Equations:
= for all i. = 0 if
= k
• These equations are feasible precisely when
G contains a k-clique.
Part II:
The Sum of
Squares/Lasserre Hierarchy
• Developed independently by Shor, Nesterov,
Parrillo, and Lasserre.
• Generalization of linear and semidefinite
programming
• Each level gives a more powerful feasibility
test than the last, rth level takes time.
• These tests can often be translated into
approximation algorithms.
• Performance of these feasibility
The Power of the Sum of
Squares Hierarchy
• Strictly stronger than the Lovasz-Schrijver
hierarchy and the Sherali-Adams hierarchy
• Leading candidate for refuting Khot’s
Unique Games Conjecture.
• Captures the known subexponential time
algorithm for Unique Games and can solve many proposed gap instances for other
A Game for the Sum of
Squares Hierarchy
• Setup: Base problem is to determine the
feasibility of a system of polynomial equations over the reals, e.g.
• = for all v. • = k
A Game for the Sum of
Squares Hierarchy
• Two players, Optimist and Pessimist.
• Optimist must claim that the answer is yes
and give some evidence
• Pessimist must try to disprove Optimist’s
evidence.
• Pessimist wins if he/she is able to refute
• What evidence should be required of Optimist? • Choice 1: Optimist must give the value of all
variables.
-To win, Optimist must fully solve the problem.
• Choice 2: No evidence.
-To win, Pessimist must prove infeasibility.
• We want something in the middle.
• For the rth level of the SOS hierarchy,
Optimist must give the expectation values of all monomials up to degree 2r for some
distribution of solutions.
A Game for the Sum of
Squares Hierarchy
Equations for whether G has a triangle:
= for all vertices i of G. = k = 3
= 0 whenever i and j are not adjacent in G.
1
4 3
2
A Game for the Sum of
Squares Hierarchy
Optimist can give the following expectation values (when r = 1):
E[] = E[] = E[] = E[] =
E[] = E[] = E[] = E[] = 3/4 E[] = E[] = E[] =
E[] = E[] = E[] = 1/2.
This corresponds to taking each of the 4 triangles in G with probability 1/4.
1
4 3
2
A Game for the Sum of
Squares Hierarchy
Of course, Optimist could try to lie…
For example, Optimist could give the following pseudo-expectation values:
Ẽ[] = Ẽ[] = Ẽ[] = Ẽ[] =
Ẽ[] = Ẽ[] = Ẽ[] = Ẽ[] = 3/4 Ẽ[] = Ẽ[] = Ẽ[] = Ẽ[] = 3/4
Ẽ[] = Ẽ[] = 0. 1
4 3
2
Detecting Lies
1
4 3
2
G
How can Pessimist detect lies systematically?
1. If the pseudo-expectation values don’t obey the expected equations, it’s bogus!
Let’s check some: (all vertices and edges have pseudo-expectation value 3/4)
3/4 3/4 0 3/4 9/4 3]
Detecting Lies
1
4 3
2
G
How else can Pessimist detect lies?
2. If some square has negative pseudo-expectation value, it’s bogus!
Ẽ[]
Ẽ[] Ẽ[] Ẽ[] Ẽ[] 2Ẽ[] 2Ẽ[] 2Ẽ[] 2Ẽ[] 2Ẽ[] 2Ẽ[]
3/4 3/4 3/4 3/4 0
• We restrict Pessimist to these two methods.
• Optimist wins if he can come up with a pseudo
expectation Ẽ which obeys all of the required equations and has nonnegative expectation on all squares
• All constraints on Ẽ are convex, so we can find
Ẽ (if it exists) with semidefinite programming.
The Moment Matrix
• Each f of degree corresponds to a vector • Ẽ[f2] = fTMf
• Constraint that Ẽ is nonnegative on squares is
satisfied if M is PSD (positive semi-definite)
𝑞
𝑝 Ẽ[p
𝑀
• Does Pessimist have a general strategy too?
• Yes, a Positivstellensatz (sum of squares) proof of
infeasibility.
• Pessimist must find polynomials f and g of degree
at most 2r such that:
1. f = 0 by the problem equations 2. g is a sum of squares
3. -1 = f + g
• This proves the equations are infeasible
• All constraints on f,g are convex, so we can find f,g
(if they exist) with semidefinite programming.
Duality Continued
• Elementary fact: Optimist and Pessimist cannot
both have a winning strategy.
• Reason: apply Ẽ to the equation -1 = f + g
• By convex duality, in virtually all cases, either
Using the Sum of Squares
Hierarchy
Pessimist Wins
Optimist Wins Problem is Feasible
• The SOS hierarchy tells us approximately when
our equations are feasible.
• If we increase r, it becomes harder for Optimist to
Obtaining Approximation
Algorithms
• Let’s say we are trying to optimize some
parameter k. If we have a rounding algorithm which turns a pseudo-expectation Ẽ into an
actual solution (at some cost to k), this gives an approximation algorithm.
Pessimist Wins
Optimist Wins Problem is Feasible
Ẽ A Solution
Two views of the Sum of
Squares Hierarchy
• Relaxation view: Solving our equations exactly
is hard, so we relax this to finding a suitable Ẽ.
• Key questions: When will this relaxation be
feasible? Can we round from Ẽ to a nearly optimal solution?
• Proof complexity view: The sum of squares
hierarchy gives a proof system
(Postivistellensatz proofs) to show infeasibility.
• Key questions: How high does r have to be to
Sum of Squares and
Planted Clique
• Essential question: On a random graph, for
which k does Optimist win the SOS(r) game?
• If Optimist wins, SOS(r) cannot determine
whether or not a clique of size k was planted.
• If Pessimist wins, SOS(r) can determine
whether or not a clique of size k was planted.
• Lower bound strategy: Find a
Previous Work
• Previously, no non-trivial lower bounds were
known for levels.
• In 2013, Raghu Meka and Avi Wigderson
Our Results
• Using Meka and Wigderson’s moment matrix
M, we recover a weaker version of the claimed result.
• In particular, the rth level of the Lasserre/SOS
hierarchy cannot solve the planted clique problem unless
• Note: Meka and Wigderson’s moment matrix
Part III:
The MW Moments
• Idea: Let d = 2r and give each d-clique the
same weight.
Definition: Define to be the number of d-cliques containing V.
Definition: Define and set whenever .
Our goal
• We must show that that the following
moment matrix M is PSD:
𝑥𝑉
𝑥𝑊
[] |𝑉|=|𝑊|=𝑟
Estimating the entries of M
• Think of d as a constant and n >> k >> 1. • If V is a clique,
Picture of M for r = 1 (d = 2)
1 2 3 4 5 6 1
2 3 4 5 6
0 or
Row i, Column i
Picture of M for r = 2 (d = 4)
0 or
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 0 or 0 or
Row ij, Column ij Nonzero if
Row ij, Column ik Nonzero if
i j
i j k
Row ij, Column kl Nonzero if i
j
k
Difficulties in Analyzing M
• Difficulty #1: Only know nonzero entries
approximately.
• Fix: Separate out the discrepancy into a
Picture of M - Δ
0 or
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 0 or 0 or
Row ij, Column ij Nonzero if
Row ij, Column ik Nonzero if
i j
i j k
Row ij, Column kl Nonzero if i
j
k
Difficulties in Analyzing M - Δ
• Difficulty #2: M - Δ has many zero rows and
columns
• Fix: Fill in the zero rows/columns of M – Δ • Calling the resulting matrix M’, the smallest
Filling in the zero rows/columns
• How should we fill in the zero rows/columns
of M – Δ?
• Idea: Only look at edges between row index
Picture of M’
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 0 or 0 or
Row ij, Column ik Nonzero if j k
Row ij, Column kl Nonzero if i
j
k
Decomposition of M’
• We decompose M’ as M’ = E + R • E is the expected value of M’
• R is the random part of M’
• We will show that E is strongly PSD i.e. the
smallest eigenvalue of E
Picture of E
Picture and Analysis of E for r = 1
1 2 3 4 5 6 1
2 3 4 5 6
Row i, Column i
Row i, Column j
Analysis of E for r=2
• We can decompose E using the Johnson
scheme
• Idea: E is sum of PSD matrices, including , so
Picture of R
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 or or 15
Row ij, Column ik Positive if j k
Row ij, Column kl Positive if i
j
k
Decomposition of R
• We write +
Picture of
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56
±
1
Decomposition of
• Difficulty: Structure of is complicated.
• Fix: Partition the entries of based on which
Picture of a piece of
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56
±
1
Row ij, Column ik i < j and i < k
Analysis of
• Idea: can be decomposed as a sum of a
constant number of pieces like the one shown.
Picture of
12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 12 13 14 15 16 23 24 25 26 34 35 36 45 46 56 or
Row ij, Column kl Positive if i
j
k
Analysis of
Putting everything together
• w.h.p and
• Recall that +
• M’ is strongly PSD so long as • This happens as long as
• The nonzero part of M – Δ is a submatrix of
M’, so it is strongly PSD.
• is small, so the nonzero part of M is strongly
Takeaways
• The sum of squares hierarchy gives a series of feasibility tests, each more powerful than the last, whose performance is not well understood. • To show that SOS(r) does not certify infeasibility
of a problem, we must:
• 1. Give a candidate pseudo-expectation which obeys all of the equations for the problem.
• 2. Show that the resulting moment matrix is PSD. • For planted clique, analyzing the MW moment
Further Research and Future Work
• Can we improve this analysis to prove better
lower bounds using the MW moments?
(Spoiler: We think the correct bound is around )
• Are there pseudo-expectations that give
better lower bounds than the MW moment? (Spoiler: We think the MW moments can be improved, at least for r = 2 rounds)
• Can we get a lower bound of for r rounds
Acknowledgements
• Thanks to Raghu for letting me use some of his
Appendix:
Correctness of the MW moments
• Recall the equations for the k-clique problem:
1. = for all vertices i of G. 2. = 0 if
3. = k
• We need to check that Ẽ obeys these
Correctness of the MW moments
• Check #1: .
• Check #2: If g and g has degree at most d
then .
• Recall: for all whenever f has degree at most
Correctness of the MW moments
• Check #3: If g and g has degree at most d
then .
• Note: If V is not a clique then so .
• Check #4: If g and g has degree at most d