Pushkar Tripathi Georgia Institute of Technology Approximability of Combinatorial Optimization Problems with Submodular Cost Functions Based on joint work.

Pushkar Tripathi Georgia Institute of Technology Approximability of Combinatorial Optimization Problems with Submodular Cost Functions Based on joint work with Gagan Goel, Chinmay Karande, and Wang Lei

Motivation Network Design Problem Objective: Find minimum spanning tree that can be built collaboratively by these agents f h g

Additive Cost Function Functions which capture economies of scale cost(a) = 1 cost(b) = 1 cost(a,b) = 2 cost(a) = 1 cost(b) = 1 cost(a,b) = 1.5 How to mathematically model these functions? - We use Submodular Functions as a starting point. Can one design efficient approximation algorithms under Submodular Cost Functions? How to mathematically model these functions? - We use Submodular Functions as a starting point. Can one design efficient approximation algorithms under Submodular Cost Functions?

Assumptions over cost functions Normalized: Monotone: Decreasing Marginal: Submodularity + + Submodular Functions

General Framework Ground set X and collection C 2 X C: set of all tours, set of all spanning trees k agents, each specifies f i : 2 X R + f i is submodular and monotone Find S 1, , S k such that: [ S i 2 C i f i ( S i ) is minimized ORACLE S f(S)

Our Results Lower Bounds : Information theoretic Upper Bounds : Rounding of configurational LPs, Approximating sumdodular functions and Greedy Problem Upper BoundLower BoundUpper BoundLower Bound Vertex Cover 2 2 - 2. log n (log n) Shortest Path O (n 2/3 )(n 2/3 ) O (n 2/3 )(n 2/3 ) Spanning Treen (n) n Perfect Matchingn (n) n Multiple AgentsSingle Agent

Selected Related Work [Grtschel, Lovsz, Schrijver 81] Minimizing non-monotone submodular function is poly-time [Feige, Mirrokni, Vondrak 07] Maximizing non-monotone function is hard. 2/5-Approximation Algorithm. [Calinescu, Chekuri, Pal, Vondrak 08] Maximizing monotone function subject to Matroid constraint: 1-1/e Approximation. [Svitkina, Fleischer 09] Upper and lower bounds for Submodular load balancing, Sparsest Cut, Balanced Cut [Iwata, Nagano 09] Bounds for Submodular Vertex Cover, Set Cover [Chekuri, Ene 10] Bounds for Submodular Multiway Partition

In this talk Submodular Shortest Path with single agent O(n 2/3 ) approximation algorithm Matching hardness of approximation

Submodular Shortest Path s t Given: Graph G, Two nodes s and t f : 2 E R + Submodular, Monotone Goal: Find path P s.t. f(P) is minimized G=(V,E) |V| =n, |E| =m

t Attempt 1: Approximate by Additive function Let w e = f({e}) Idea : w e OPT w e s 2. Pruning: Remove edges costlier than e* 1. Guess e* = argmax{ w e | e 2 OPT } 3. Search: Find the shortest length s-t path in the residual graph ALG diameter(G). w e* diameter(G).OPT e 2 OPT

Attempt 2: Ellipsoid Approximation Johns theorem : For every polytope P, there exists an ellipsoid contained in it that can be scaled by a factor of O( n) to contain P [GHIM 09]: If the convex body is a polymatroid, then there is a poly-time algorithm to compute the ellipse. P

Attempt 2: Ellipsoid Approximation P [GHIM 09]: If the convex body is a polymatroid, then there is a poly-time algorithm to compute the ellipse. S: e 2 S x(e) f(S) e: x(e) 0 f: Submodular, monotone Polymatroid

Approximating Submodular Functions X f : Monotone submodular function g(S) = d e e 2 S g(S) f(S) n g(S) d1d1 d2d2 d6d6 d4d4 d5d5 d3d3 Polynomial time |X| = n

Attempt 2: Ellipsoid Approximation f: 2 E R + Submodular, Monotone STEP 1: STEP 2: Min g(S) s.t. S 2 PATH(s,t) * Minimizing over g(S) is equivalent to minimizing just the additive part [GHIM 09] Analysis : f(P) g(P) g(O) f(O) g(S): = d e P: Optimum path under g O: Optimum path under f {de}{de} EE EE EE

Recap. Approximating by linear functions : Works for graphs with small diameter Approximating by ellipsoid functions : Works for sparse graphs n/2 Dense Graph with large diameter

Algorithm for Shortest Path STEP 1: Pruning - Guess edge e* = argmax {w e | e OPT path} - Remove edges costlier than w e*

Algorithm for Shortest Path STEP 1: Pruning - Guess edge e* = argmax {w e | e OPT path} - Remove edges costlier than w e* STEP 2 : Contraction - if v, s.t. degree(v) > n 1/3, contract neighborhood of v - repeat

s t s t Dense connected component

Algorithm for Shortest Path STEP 1: Pruning - Let w e = f({e}) - Guess edge e* = argmax {w e | e OPT path} - Remove edges costlier than w e* STEP 2 : Contraction - if v, s.t. degree(v) < n 1/3, contract neighborhood of v - repeat STEP 3 : Ellipsoid Approximation - Calculate ellipsoidal approximation (d,g) for the residual graph

Algorithm for Shortest Path STEP 1: Pruning - Let w e = f({e}) - Guess edge e* = argmax {w e | e OPT path} - Remove edges costlier than w e* STEP 2 : Contraction - if v, s.t. degree(v) < n 1/3, contract neighborhood of v - repeat STEP 3 : Ellipsoid Approximation - Calculate ellipsoidal approximation (d,g) for the residual graph STEP 4 : Search - Find shortest s-t path according to g.

Algorithm for Shortest Path STEP 1: Pruning - Let w e = f({e}) - Guess edge e* = argmax {w e | e OPT path} - Remove edges costlier than w e* STEP 2 : Contraction - if v, s.t. degree(v) < n 1/3, contract neighborhood of v - repeat STEP 3 : Ellipsoid Approximation - Calculate ellipsoidal approximation (d,g) for the residual graph STEP 4 : Search - Find shortest s-t path according to g. STEP 5 : Reconstruction - Replace the path through each contracted vertex with one having the fewest edges.

s t Path having fewest edges

Analysis s t P1 P2 R

Bounding the cost of P1 s t P1 P2 Has at most n 4/3 edges R f(P 1 ) E(R).g(P 1 ) E(R).g(OPT) E(R).f(OPT) n 2/3 f(OPT)

Bounding the cost of P2 s t Diam(G i ) | G i |/ n 1/3 f(P 2 ) (dia(G 1 ) +.. +dia(G k ) ) w e* (|G 1 | / n 1/3 + . ) w e* (n / n 1/3 ) w e* n 2/3 f(OPT) G1G1 G2G2 G3G3

In this talk Submodular Shortest Path with single agent O(n 2/3 ) approximation algorithm Matching hardness of approximation

Information Theoretic Lower Bound Polynomial number of queries to the oracle Algorithm is allowed unbounded amount of time to process the results of the queries Not contingent on P vs NP f S1 f(S1) S2 f(S2) S3 f(S2)

General Technique Cost functions f, g satisfying OPT( f ) >> OPT( g ) f (S) = g(S) for most sets S A any randomized algorithm f(Q ) = g( Q ) with high probability for every query Q made by A. Probability over random bits in A.

Yaos Lemma f(Q) = g(Q) with high probability for every query Q made by randomized algorithm A. f and a distribution D from which we choose g, such that for an arbitrary query Q, f(Q) = g(Q) with high probability

Non-combinatorial Setting X : Ground set f(S) = min{ |S|, } D : R X, |R| = g R (S) = min{| S R c | + min( S R, ) } D : R X, |R| = g R (S) = min{| S R c | + min( S R, ) }

Optimal Query Claim : Optimal query has size Case 1 : |Q| < Probability can only increase if we increase |Q|

Case 2 : |Q| > Probability can only increase if we decrease |Q| Optimal query size to distinguish f and g R is

Distinguishing f and g R = (1+ ) E[|Q R|] f and g are hard to distinguish Chernoff Bounds

Hardness of learning submodular functions Set = n 1/2 log n Optimal query size = = n 1/2 log n |R| = = n 1/2 log n E[ Q R] = log 2 n = (1+ ) E[ Q R] = (1+ ) log 2 n Super logarithmic Corollary : Hard to learn a submodular function to a factor better than n 1/2 /log n in polynomial value queries. f and g are indistinguishable f(R) = min{ |R|, } = |R| = = n 1/2 log n g R (R ) = min{| R R c | + min( R R, ) } = = log 2 n

Randomly chosen set may not be a feasible solution in the combinatorial setting. Eg. Randomly chosen set of edges rarely yield a s-t path. Difficulty in Combinatorial Setting Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution. Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution.

Base Graph G ...st n 2/3 levels n 1/3 vertices

Functions f and g . st Y B f(S) = f( S B ) & g(S) = g( S B )

Functions f and g . st Y B f(S) = min( |S B|, )

Functions f and g Y B .. st .. g R (S) = min{| S R B| + min( S R B, )} Uniform random subset of B of size Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution. Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution.

Functions f and g . st Y B g R (S) = min{| S R B| + min( S R B, ) Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution. Solution : 1.Do not choose R randomly from the entire domain X. 2.Use a subset of R as a proxy for the solution. R = n 2/3 log 2 n

Setting the constants Set = n 2/3 log 2 n Optimal Query size = = n 2/3 log 2 n = log 2 n f and g are indistinguishable f(OPT) = min{ |R|, } = |R| = = O( n 2/3 log 2 n) g R (OPT ) = min{| R R c | + min( R R, ) } = = log 2 n Theorem : Submodular Shortest Path problem is hard to approximate to a factor better than O(n 2/3 )

Problem Upper BoundLower BoundUpper BoundLower Bound Vertex Cover 2 2 - 2. log n (log n) Shortest Path O (n 2/3 )(n 2/3 ) O (n 2/3 )(n 2/3 ) Spanning Treen (n) n Perfect Matchingn (n) n Multiple AgentsSingle Agent n: # of vertices in graph G Whats the right model to study economies of scale? Summary

Newer Models Discount Models f h g E R

Task: Minimize sum of payments Cost Payment f(a) + f(b) + f(c) . Sub modular functions

Approximability under Discounted Costs[GTW 09] ProblemLower BoundUpper bound Edge CoverO(log n) Spanning TreeO(log n) Shortest Pathn Minimum Perfect Matching O(poly log n)n O(log n) O(poly log n) O(log n)

Shortest Path : O(log c n) hardness Set Cover Instance U S s t Agents - Cost of every edge is 1 1 1 Claim : Set cover of size |S| Shortest path of length |S|

Hardness Gap Amplification s t s t Original Instance Harder Instance Replace each edge by a copy of the original graph. Edges of the same color get the same copy. Edges of different colors gets copies with new colors(agents)

Claim : The new instance has a solution of cost 2 iff the original instance has a solution of cost . For any fixed constant c iterate this construction c times to further amplify the lower bound to O(log c n).

Why is it so hard to distinguish f and g ? Observation: f R (S ) is at most g(S ) for any set S. Case 1: Small size queries - |Q | n This probability can only increase if we increase |Q |

Case 2: Large size queries - |Q | n This probability can only increase if we decrease |Q |

Combinatorial Optimization C - Ground set f - Valuation function over subsets of C X - Collection of some subsets C having a special property Task - Find the set in X that has minimum cost under a given valuation function.

General Technique cont. f S2 S3 S1f(S1) = g(S1) f(S2) = g(S2) f(S3) = g(S3) A cannot distinguish between f and g Output is at least OPT( g ) OPT( g ) OPT( f )

Plan Fix a cost function f Fix a distribution D of functions such that for every g in D OPT(f ) >> OPT (g) For an arbitrary query Q, f(Q) = g(Q) with high probability

Optimal size queries Queries of size n- |Q | = n

Pushkar Tripathi Georgia Institute of Technology Approximability of Combinatorial Optimization Problems with Submodular Cost Functions Based on joint work.

Documents

e approximation

submodular cost functions

e r submodular

p slide

submodular shortest

guess e

talk submodular shortest

log n log n shortest