Satyen Kale (Yahoo! Research) Joint work with Sanjeev Arora (Princeton)

Slides:



Advertisements
Similar presentations
Geometry and Expansion: A survey of some results Sanjeev Arora Princeton ( touches upon: S. A., Satish Rao, Umesh Vazirani, STOC04; S. A., Elad Hazan,
Advertisements

Arora: SDP + Approx Survey Semidefinite Programming and Approximation Algorithms for NP-hard Problems: A Survey Sanjeev Arora Princeton University
Cognitive Radio Communications and Networks: Principles and Practice By A. M. Wyglinski, M. Nekovee, Y. T. Hou (Elsevier, December 2009) 1 Chapter 12 Cross-Layer.
Iterative Rounding and Iterative Relaxation
1 LP, extended maxflow, TRW OR: How to understand Vladimirs most recent work Ramin Zabih Cornell University.
Routing in Undirected Graphs with Constant Congestion Julia Chuzhoy Toyota Technological Institute at Chicago.
Lasserre Hierarchy, Higher Eigenvalues and Approximation Schemes for Graph Partitioning and PSD QIP Ali Kemal Sinop (joint work with Venkatesan Guruswami)
Near-Optimal Algorithms for Online Matrix Prediction Elad Hazan (Technion) Satyen Kale (Yahoo! Labs) Shai Shalev-Shwartz (Hebrew University)
Cuts, Trees, and Electrical Flows Aleksander Mądry.
Primal Dual Combinatorial Algorithms Qihui Zhu May 11, 2009.
How to Round Any CSP Prasad Raghavendra University of Washington, Seattle David Steurer, Princeton University (In Principle)
Submodular Set Function Maximization via the Multilinear Relaxation & Dependent Rounding Chandra Chekuri Univ. of Illinois, Urbana-Champaign.
Linear Round Integrality Gaps for the Lasserre Hierarchy Grant Schoenebeck.
Fast Regression Algorithms Using Spectral Graph Theory Richard Peng.
IMIM v v v v v v v v v DEFINITION L v 11 v 2 1 v 31 v 12 v 2 2 v 32.
C&O 355 Lecture 23 N. Harvey TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A A A A A A A A A A.
1 LP Duality Lecture 13: Feb Min-Max Theorems In bipartite graph, Maximum matching = Minimum Vertex Cover In every graph, Maximum Flow = Minimum.
C&O 355 Mathematical Programming Fall 2010 Lecture 22 N. Harvey TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A A A.
1 of 56 Linear Programming-Based Approximation Algorithms Shoshana Neuburger Graduate Center, CUNY May 13, 2009.
Geometric embeddings and graph expansion James R. Lee Institute for Advanced Study (Princeton) University of Washington (Seattle)
Graph Laplacian Regularization for Large-Scale Semidefinite Programming Kilian Weinberger et al. NIPS 2006 presented by Aggeliki Tsoli.
Semi-Definite Algorithm for Max-CUT Ran Berenfeld May 10,2005.
Analysis of Network Diffusion and Distributed Network Algorithms Rajmohan Rajaraman Northeastern University, Boston May 2012 Chennai Network Optimization.
Approximation Algoirthms: Semidefinite Programming Lecture 19: Mar 22.
Graph Sparsifiers: A Survey Nick Harvey Based on work by: Batson, Benczur, de Carli Silva, Fung, Hariharan, Harvey, Karger, Panigrahi, Sato, Spielman,
Graph Sparsifiers: A Survey Nick Harvey UBC Based on work by: Batson, Benczur, de Carli Silva, Fung, Hariharan, Harvey, Karger, Panigrahi, Sato, Spielman,
Sparsest Cut S S  G) = min |E(S, S)| |S| S µ V G = (V, E) c- balanced separator  G) = min |E(S, S)| |S| S µ V c |S| ¸ c ¢ |V| Both NP-hard.
Semidefinite Programming
1 Introduction to Linear and Integer Programming Lecture 9: Feb 14.
1 Optimization problems such as MAXSAT, MIN NODE COVER, MAX INDEPENDENT SET, MAX CLIQUE, MIN SET COVER, TSP, KNAPSACK, BINPACKING do not have a polynomial.
Expander flows, geometric embeddings, and graph partitioning Sanjeev Arora Princeton Satish Rao UC Berkeley Umesh Vazirani UC Berkeley ( + survey of other.
Expander flows, geometric embeddings, and graph partitioning Sanjeev Arora Princeton Satish Rao UC Berkeley Umesh Vazirani UC Berkeley.
A general approximation technique for constrained forest problems Michael X. Goemans & David P. Williamson Presented by: Yonatan Elhanani & Yuval Cohen.
1 Traveling Salesman Problem (TSP) Given n £ n positive distance matrix (d ij ) find permutation  on {0,1,2,..,n-1} minimizing  i=0 n-1 d  (i),  (i+1.
Semidefinite Programming Based Approximation Algorithms Uri Zwick Uri Zwick Tel Aviv University UKCRC’02, Warwick University, May 3, 2002.
Multiplicative weights method: A meta algorithm with applications to linear and semi-definite programming Sanjeev Arora Princeton University Based upon:
Distributed Combinatorial Optimization
SDP Based Approach for Graph Partitioning and Embedding Negative Type Metrics into L 1 Subhash Khot (Georgia Tech) Nisheeth K. Vishnoi (IBM Research and.
Approximation Algorithms: Bristol Summer School 2008 Seffi Naor Computer Science Dept. Technion Haifa, Israel TexPoint fonts used in EMF. Read the TexPoint.
Pablo A. Parrilo ETH Zürich Semialgebraic Relaxations and Semidefinite Programs Pablo A. Parrilo ETH Zürich control.ee.ethz.ch/~parrilo.
Dana Moshkovitz, MIT Joint work with Subhash Khot, NYU.
Integrality Gaps for Sparsest Cut and Minimum Linear Arrangement Problems Nikhil R. Devanur Subhash A. Khot Rishi Saket Nisheeth K. Vishnoi.
Arora: SDP + Approx Survey Semidefinite Programming and Approximation Algorithms for NP-hard Problems: A Survey Sanjeev Arora Princeton University.
CS774. Markov Random Field : Theory and Application Lecture 08 Kyomin Jung KAIST Sep
The Multiplicative Weights Update Method Based on Arora, Hazan & Kale (2005) Mashor Housh Oded Cats Advanced simulation methods Prof. Rubinstein.
Expander Flows, Graph Spectra and Graph Separators Umesh Vazirani U.C. Berkeley Based on joint work with Khandekar and Rao and with Orrechia, Schulman.
Semidefinite Programming
C&O 355 Mathematical Programming Fall 2010 Lecture 16 N. Harvey TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A A A.
Robust Optimization and Applications Laurent El Ghaoui IMA Tutorial, March 11, 2003.
Implicit Hitting Set Problems Richard M. Karp Erick Moreno Centeno DIMACS 20 th Anniversary.
Chapter 8 Maximum Flows: Additional Topics All-Pairs Minimum Value Cut Problem  Given an undirected network G, find minimum value cut for all.
C&O 355 Lecture 24 N. Harvey TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AA A A A A A A A A.
Graph Partitioning using Single Commodity Flows
MAP Estimation in Binary MRFs using Bipartite Multi-Cuts Sashank J. Reddi Sunita Sarawagi Sundar Vishwanathan Indian Institute of Technology, Bombay TexPoint.
TU/e Algorithms (2IL15) – Lecture 12 1 Linear Programming.
Approximation Algorithms Duality My T. UF.
Linear Programming Piyush Kumar Welcome to CIS5930.
Approximation Algorithms based on linear programming.
TU/e Algorithms (2IL15) – Lecture 12 1 Linear Programming.
Fernando G.S.L. Brandão MSR -> Caltech Faculty Summit 2016
Fernando G.S.L. Brandão Caltech QIP 2017
Lap Chi Lau we will only use slides 4 to 19
Topics in Algorithms Lap Chi Lau.
Bo-Young Kim Applied Algorithm Lab, KAIST 月
Amir Ali Ahmadi (Princeton University)
A Combinatorial, Primal-Dual Approach to Semidefinite Programs
Structural Properties of Low Threshold Rank Graphs
Graph Partitioning Problems
Embedding Metrics into Geometric Spaces
Topics in Algorithms 2005 Max Cuts
Presentation transcript:

Satyen Kale (Yahoo! Research) Joint work with Sanjeev Arora (Princeton)

Semidefinite Program (SDP): find X s.t. A 1 ² X · 0 A 2 ² X · 0  A m ² X · 0 X º 0 Tr(X) = 1 All eigenvalues are non-negative ´ 9 v 1, v 2, …, v n s.t. X ij = v i ¢ v j A ² B = Tr(AB) =  ij A ij B ij Density matrix

Max Cut [GW’95] Graph Coloring [KMS’98, ACC’06] (a Ç : b) Æ ( : a Ç c) Æ (a Ç b) Æ ( : c Ç b) Constraint Satisfaction [ACMM’05, R ’08, RS ‘09,..] SDP Balanced Partitioning [ARV’04] Control Theory

Algorithms for SDP Ellipsoid method [GLS’81]: O(n 8 ) iterations O(n 8 ) iterations Interior point methods [NN’90, A’95]: O(√n ¢ m 3 ) time O(√n ¢ m 3 ) time Lagrangian Relaxation [KL’95], [AHK’05]: Reduction to eigenvectors Reduction to eigenvectors poly(1/  ) dependence on , limits applicability poly(1/  ) dependence on , limits applicability A 1 ² X · 0  A n ² X · 0  i w i A i ² X · 0 Combinatorial, Primal-Dual algorithms (analogous to flow algorithms via LP) Combinatorial, Primal-Dual algorithms (analogous to flow algorithms via LP)

 Positive semidefiniteness hard to maintain  Rounding algorithms exploit geometric structure (e.g. negative-type metrics)  Matrix operations inefficient to implement

A general scheme that yields fast, combinatorial, primal-dual approximation algorithms for various combinatorial optimization problems using SDP relaxations

Our Results: Primal-Dual algorithmsProblem Previous best: O(√log n) apx Our algorithm: O(√log n) apx Our algorithm: O(log n) apx Undirected Sparsest Cut Õ(n 2 ) [AHK’04] Õ(n 2 ) Õ(m + n 1.33 ) Undirected Balanced Sep. Õ(n 2 ) [AHK’04] Õ(n 2 ) Õ(m + n 1.33 ) Directed Sparsest Cut Õ(n 9.5 ) Õ(m n 2 ) Õ(m 1.33 ) Directed Balanced Sep. Õ(n 9.5 ) Õ(m n 2 ) Õ(m 1.33 ) Min UnCut Õ(n 9.5 ) Õ(n 3 ) --- Min 2CNF Deletion Õ(n 9.5 ) Õ(nm n 3 ) --- Õ(m + n 1.33 ) Õ(m + n 1.33 ) [S’09] Unknown how to achieve using LP!

 Also yields Õ(m) time algorithm for approximating MAXCUT SDP  Gives first near-linear implementation of Goemans-Williamson approximation algorithm  Previous best: Õ(nm) [KL’95]

a 1 ¢ x · 0 a 2 ¢ x · 0  a m ¢ x · 0 x ¸ 0  i x i = 1 Multiplicative Weights Initialize: x 1 = (1/n, …, 1/n) Update: x t+1 = x t £ exp(-  a i t )/  t  t = normalization factor Violation Checker Find i t s.t. a i t ¢ x t >  xtxt aitait Thm: Finds  -feasible x in O(  2 log(n)/  2 ) iterations.  = max ij |a ij | (“width”) Convex combination of constraints allowed:  i y i a i ¢ x t > , where y i ¸ 0,  i y i = 1. Hence Primal-Dual. Analysis uses  t as potential function Coordinate-wise a 1 ¢ x ·  a 2 ¢ x ·   a m ¢ x ·  x ¸ 0  i x i = 1

A 1 ² X · 0 A 2 ² X · 0  A m ² X · 0 X º 0 Tr(X) = 1 Matrix MW Initialize: X 1 = (1/n) I Update: X t+1 = exp(-   t s=1 A i s )/  t  t = normalization factor Violation Checker Find i t s.t. A i t ² X t >  XtXt AitAit Thm: Finds  -feasible X in O(  2 log(n)/  2 ) iterations.  = max i k A i k (“width”) Convex combination of constraints allowed:  i y i A i ² X t > , where y i ¸ 0,  i y i = 1. Hence Primal-Dual. Analysis uses  t as potential function

The Matrix Exponential Matrix exponential: exp(A) = I + A + A 2 /2! + A 3 /3! + … Always PSD: exp(A) º 0 Always PSD: exp(A) º 0 Tricky beast: Tricky beast: exp(A+B) = exp(A) exp(B) Computation: Computation: O(n 3 ) time O(n 3 ) time Computing exp(A)v: Õ(m) time (arises as soln of ODE) Computing exp(A)v: Õ(m) time (arises as soln of ODE) Golden-Thompson inequality: Tr(exp(A+B)) · Tr(exp(A)exp(B)) Matrix MW Initialize: X 1 = (1/n) I Update: X t+1 = exp(-   t s=1 A i s )/  t  t = normalization factor

Input Quadratic Program SDP: Vector variables Reduction SDP Solver Rounding Algorithm SDP Opt: X X ij = v i ¢ v j v1v1 vnvn v2v2 Relaxation Primal-Dual SDP

Reduction Relaxation Rounding Algorithm v1v1 vnvn v2v2 Matrix MW X y 1,…, y m Primal-Dual SDP SDP: Vector variables Input X ij = v i ¢ v j Quadratic Program

 Cut (S, S’) is c-balanced if |S|, |S’| ¸ cn  Min c-Balanced Separator: c-balanced cut of min capacity  Numerous applications:  Divide-and-conquer algs  Markov chains  Geometric embeddings  Clustering  Layout problems …… G = (V, E) S S’

¼ k v i – v j k 2 = 0 if i, j on same side, = 1 otherwise SDP min  i,j 2 E ¼ k v i – v j k 2 8 i: k v i k 2 = 1  i, j ¼ k v i – v j k 2 ¸ c(1-c)n 2 8 ijk: k v i –v j k 2 + k v j –v k k 2 ¸ k v i –v k k 2 v i = -1v i = 1 G = (V, E) S S’

SDP (C1)  i,j 2 E ¼ k v i – v j k 2 <  (C2) 8 i: k v i k 2 = 1 (C3)  i, j ¼ k v i – v j k 2 ¸ c(1-c)n 2 (C4) 8 ijk: k v i –v j k 2 + k v j –v k k 2 ¸ k v i –v k k 2 To find violated constraints, given X (thus, v i ):  Check (C2) and (C3)  To check (C1) and (C4): find multicommodity flow s.t. 1. Total flow from any node is at most d = Õ(  /n) 2. Total flow on any edge is at most 1 3.  i,j f ij k v i – v j k 2 ¸ 4  f ij = total flow between i, j Bounds width  Binary search parameter Either (C1) or (C4) violated

X ij = v i ¢ v j v1v1 vnvn v2v2 d d d d d d d To find flow s.t.  ij f ij k v i – v j k  ¸ 4  Thm: This yields either: 1.a flow s.t.  ij f ij k v i – v j k 2 ¸ 4  2.a cut of value O(log(n)  ) k v i k 2 = 1 ¼  ij ¼ k v i – v j k 2 ¸ c(1-c)n 2 Max flow

X ij = v i ¢ v j v1v1 vnvn v2v2 Thm: This yields either: 1.a flow s.t.  ij f ij k v i – v j k 2 ¸ 4  2.a cut of value O(√log(n)  ) d d d d d d d d Multicommodity flow To find flow s.t.  ij f ij k v i – v j k  ¸ 4  k v i k 2 = 1 ¼  ij ¼ k v i – v j k 2 ¸ c(1-c)n 2

Derandomization Deterministic Expander Graphs Constructions [K’07, WX ’07] Deterministic Approximation via Covering SDPs [K ’07, WX ’07] Quantum Computing QIP = PSPACE [JJUW ’10] and related results [JUW ’09, JW ’09] Learnability of Quantum States [A ’06, K ’07] Machine Learning Online Variance Minimization [WK ’06] Online PCA [WK ’06] Other SDP based Approximation Algorithmss Near Linear-Time Approximation Algorithms for 2-CSPS [S ’10] Near-Linear Time Balanced Graph Partitioning [OV ’11]

 Make SDP solver efficient in practice  Sampling, rescaling techniques might be useful  ``Combinatorialize’’ interior-point methods?  Near linear-time algs for approx max flow?