Presentation is loading. Please wait.

Presentation is loading. Please wait.

Sum of Squares, Planted Clique, and Pseudo-Calibration

Similar presentations


Presentation on theme: "Sum of Squares, Planted Clique, and Pseudo-Calibration"โ€” Presentation transcript:

1 Sum of Squares, Planted Clique, and Pseudo-Calibration
Joint work with Boaz Barak, Sam Hopkins, Pravesh Kothari, Jonathan Kelner, and Ankur Moitra

2 Planted Clique Lower Bound
Theorem [BHKKMP, FOCS 2016]: If the input graph ๐บ is chosen randomly from the Erdรถs- Rรฉnyi distribution ๐บ(๐‘›, 1 2 ), then with high probability the degree ๐‘‘ sum of squares hierarchy gives a value of at least ๐‘› 1 2 โˆ’๐‘ ๐‘‘ ๐‘™๐‘œ๐‘”๐‘› for the size of the largest clique in ๐บ.

3 Talk Outline Part I: The Sum of Squares Hierarchy
Part II: Planted Clique Part III: Pseudo-calibration

4 Part I: The Sum of Squares Hierarchy

5 The Sum of Squares Hierarchy
Developed independently by Shor, Nesterov, Parrillo, and Lasserre. Generalization of linear and semidefinite programming Strictly stronger than the Lovasz-Schrijver hierarchy and the Sherali-Adams hierarchy Captures the Goemans-Williamson algorithm for MAX-CUT, the Goemens-Linial algorithm for sparsest cut (analyzed by Arora, Rao, Vazirani), and the Arora-Barak-Steurer subexponential time algorithm for Unique Games Leading candidate for refuting Khotโ€™s Unique Games Conjecture.

6 Positivstellensatz Proofs
Setup: Want to know if equations ๐‘  1 ๐‘ฅ 1 ,โ€ฆ, ๐‘ฅ ๐‘› =0, ๐‘  2 ๐‘ฅ 1 ,โ€ฆ, ๐‘ฅ ๐‘› =0, โ€ฆ can be solved simultaneously over โ„. Degree ๐‘‘ Positivstellensatz proof of infeasibility: Equation of the form โˆ’1= ๐‘– ๐‘“ ๐‘– ๐‘  ๐‘– + ๐‘— ๐‘” ๐‘— 2 where for all ๐‘–, deg ๐‘“ ๐‘– + deg ๐‘  ๐‘– โ‰ค๐‘‘ and for all ๐‘—, deg ๐‘” ๐‘— โ‰ค ๐‘‘ 2

7 Degree ๐‘‘ Sum of Squares Stengleโ€™s Positivstellensatz: If equations are infeasible, there exists a Positivstellensatz proof of infeasibility. However, degree could be very high Degree ๐‘‘ Sum of Squares hierarchy: Returns NO if there is a degree ๐‘‘ Positivstellensatz proof of infeasibility, otherwise returns YES. Fundamental questions: At what degree is there a Positivstellesatz proof of infeasibility? If there is no degree ๐‘‘ proof, how do we show it?

8 Pseudo-expectation values
Pseudo-expectation values: linear mapping แบผ from polynomials of degree โ‰ค๐‘‘ to values in โ„ satisfying the following conditions: แบผ 1 =1 แบผ ๐‘“ ๐‘  ๐‘– =0 whenever deg ๐‘“ + deg ๐‘  ๐‘– โ‰ค๐‘‘ แบผ ๐‘” 2 โ‰ฅ0 whenever deg ๐‘” โ‰ค ๐‘‘ 2 Intuition: แบผ โ€œlooks likeโ€ the expected values over a distribution of solutions.

9 Duality Degree ๐‘‘ Positivstellensatz proof: โˆ’1= ๐‘– ๐‘“ ๐‘– ๐‘  ๐‘– + ๐‘— ๐‘” ๐‘— 2
โˆ’1= ๐‘– ๐‘“ ๐‘– ๐‘  ๐‘– + ๐‘— ๐‘” ๐‘— 2 Pseudo-expectation values: แบผ 1 =1 แบผ ๐‘“ ๐‘– ๐‘  ๐‘– =0 แบผ ๐‘” ๐‘— 2 โ‰ฅ0 Cannot both exist, otherwise โˆ’1=แบผ โˆ’1 = ๐‘– แบผ[๐‘“ ๐‘– ๐‘  ๐‘– ] + ๐‘— แบผ[๐‘” ๐‘— 2 ] โ‰ฅ0 If strong duality holds (which we will assume), one or the other will exist

10 The Moment Matrix ๐‘ž ๐‘,๐‘ž are monomials of degree at most ๐‘‘ 2 . ๐‘ แบผ[๐‘q] ๐‘€ Each ๐‘” of degree โ‰ค ๐‘‘ 2 can be viewed as a vector in the basis of monomials แบผ[ ๐‘” 2 ] =๐‘”๐‘‡๐‘€๐‘” แบผ ๐‘” 2 โ‰ฅ0 whenever deg ๐‘” โ‰ค ๐‘‘ 2 โ‡”M is PSD (positive semi-definite)

11 Lower Bounds Strategy for SOS
To prove a lower bound on SOS, we must Construct pseudo-expectation values แบผ and the corresponding moment matrix ๐‘€ Show that แบผ obeys the required equalities and that ๐‘€ is PSD (this is the hard part).

12 Semidefinite Programs for SOS
Can search for the moment matrix ๐‘€ (and the corresponding pseudo-expectation values แบผ) with a semidefinite program of size ๐‘› ๐‘‚(๐‘‘) . The conditions that แบผ 1 =1 and แบผ ๐‘“ ๐‘  ๐‘– =0 whenever deg ๐‘“ + deg ๐‘  ๐‘– โ‰ค๐‘‘ give linear constraints on entries of ๐‘€. ๐‘€ must be PSD SOS gives a hierarchy of increasingly powerful (and large) semidefinite programs

13 Optimization with SOS Often have equations with parameter(s)
Want to optimize over green region, SOS optimizes over the blue and green regions. Equations are feasible Infeasible but no proof Positivstellensatz proof of infeasibility As we increase the degree, the blue region shrinks.

14 Approximation Algorithms with SOS
For many problems, there is a method for rounding the pseudo-expectation values แบผ into an actual solution (with worse parameters). This gives an approximation algorithm. Optimal Solution Equations are feasible แบผ A Solution Infeasible but no proof Positivstellensatz proof of infeasibility

15 For more info on SOS Princeton sum of squares seminar website Harvard sum of squares seminar website

16 Part II: Planted Clique

17 The Planted Clique Problem
G(n,1/2) + clique(ฯ‰) Jerrum 92, Kucera 95: For which ฯ‰ can we find the planted clique? Best- Alon et al. 98: ฯ‰= โ„ฆ( ๐‘› ) a j b i c h d g e f Can you find the 5-clique?

18 The Planted Clique Problem
G(n,1/2) + clique(ฯ‰) Jerrum 92, Kucera 95: For which ฯ‰ can we find the planted clique? Best- Alon et al. 98: ฯ‰= โ„ฆ( ๐‘› ) a j b i c h d g e f This 5-clique was planted by adding the red edge.

19 Equations for ๐œ”-Clique
Variable ๐‘ฅ ๐‘– for each vertex i in G. Want ๐‘ฅ ๐‘– =0 if i is not in the clique Want ๐‘ฅ ๐‘– =1 if i is in the clique. Equations: ๐‘ฅ ๐‘– 2 = ๐‘ฅ ๐‘– for all i. ๐‘ฅ ๐‘– ๐‘ฅ ๐‘— = 0 if ๐‘–,๐‘— โˆ‰๐ธ(๐บ) ๐‘– ๐‘ฅ ๐‘– = ๐œ” These equations are feasible precisely when G contains a ๐œ” -clique.

20 Part III: Pseudo-Calibration

21 Choosing แบผ: First attempt
How should we choose แบผ? First attempt: give every clique of size ๐‘‘ the same weight. Definition: Define ๐‘ฅ ๐‘‰ = ๐‘–โˆˆ๐‘‰ ๐‘ฅ ๐‘– ๐ธ ๐‘ฅ ๐‘‰ โ‰ˆ 2 |๐‘‰| ๐œ” |๐‘‰| ๐‘› |๐‘‰| if ๐‘‰ is a clique, ๐ธ ๐‘ฅ ๐‘‰ =0 otherwise. This gives non-trivial lower bounds, but cannot give the full lower bound

22 Choosing แบผ: Second attempt
Second attempt: See what went wrong and fix it This works for degree ๐‘‘=4 but is ad-hoc and would be very complicated for higher degrees.

23 Choosing แบผ: Pseudo-calibration
Planted distribution: Each vertex is in the clique with probability ๐œ” ๐‘› , ๐‘ฅ ๐‘– =1 if ๐‘–โˆˆ๐‘๐‘™๐‘–๐‘ž๐‘ข๐‘’, 0 otherwise Idea: ๐ธ should match the planted distribution on low degree tests

24 Fourier Characters ๐œ’ ๐ธ Definition: Given a set ๐ธ of possible edges of ๐บ, define ๐œ’ ๐ธ ๐บ =โˆ’ 1 |๐ธโˆ–๐ธ(๐บ)| ๐‘ฅ 1 ๐‘ฅ 2 ๐‘ฅ 4 ๐‘ฅ 3 ๐บ Example: If ๐ธ={ ๐‘ฅ 1 , ๐‘ฅ 2 , ๐‘ฅ 1 , ๐‘ฅ 3 ,( ๐‘ฅ 1 , ๐‘ฅ 4 )} then ๐œ’ ๐ธ ๐บ =โˆ’1 as ๐ธโˆ–๐ธ ๐บ =1

25 Pseudo-calibration For all ๐‘‰,๐ธ, ๐ธ ๐บโˆผ๐บ ๐‘›, ๐œ’ ๐ธ ๐ธ [ ๐‘ฅ ๐‘‰ ] = ๐ธ ๐บ,๐‘ฅโˆผ๐บ ๐‘›, ๐พ ๐œ” [ ๐‘ฅ ๐‘‰ ๐œ’ ๐ธ ] Right hand side is 0 (over the random part of G) unless ๐‘‰โˆช๐‘‰ ๐ธ โŠ†๐‘๐‘™๐‘–๐‘ž๐‘ข๐‘’, in which case the right hand side is 1. Each vertex is in the clique with probability ๐œ” ๐‘› , so the right hand side has value ฯ‰ ๐‘› ๐‘‰ ๐ธ โˆช๐‘‰

26 Pseudo-calibrated moments
โˆ€๐‘‰,๐ธ, ๐ธ ๐บโˆผ๐บ ๐‘›, ๐œ’ ๐ธ ๐ธ [ ๐‘ฅ ๐‘‰ ] = ๐œ” ๐‘› |๐‘‰โˆช๐‘‰(๐ธ)| ๐ธ ๐‘ฅ ๐‘‰ = ๐ธ: ๐‘‰ ๐ธ โˆช๐‘‰ <ฯ„ ฯ‰ ๐‘› ๐‘‰ ๐ธ โˆช๐‘‰ ๐œ’ ๐ธ ๐บ Note: We only have ๐‘– ๐‘ฅ ๐‘– โ‰ˆฯ‰ with these moments.

27 Properties of ๐œ’ ๐ธ Recall: ๐œ’ ๐ธ ๐บ =โˆ’ 1 |๐ธโˆ–๐ธ(๐บ)|
If we define the inner product ๐‘“,๐‘” = ฮ• ๐บ [ ๐‘“ ๐บ ๐‘” ๐บ ] where ๐‘“,๐‘” are functions of the input graph ๐บ, then ๐œ’ ๐ธ 1 ๐œ’ ๐ธ 2 = ๐œ’ ๐ธ 1 ฮ” ๐ธ 2 (ฮ” = symmetric difference)

28 Current/Future work Can we apply these techniques to prove sum of squares lower bounds for other planted problems? Can we find general conditions for the pseudo-calibrated ๐ธ which imply a sum of squares lower bound? Can we prove our lower bound with an ๐ธ which fully respects the equation ๐‘ฅ ๐‘ฅ ๐‘– =ฯ‰ ?

29 Acknowledgements Thanks to the National Science Foundation, Microsoft Research, and the Simons Foundation for supporting this research.

30 Thank you!


Download ppt "Sum of Squares, Planted Clique, and Pseudo-Calibration"

Similar presentations


Ads by Google