A Tight Max-Flow Min-Cut Duality Theorem for Non-Linear Multicommodity Flows

TL;DR

This paper extends the classical Max-Flow Min-Cut duality to non-linear multicommodity flows, providing a tight duality theorem, computational methods, and algorithms for approximate solutions in complex network flow problems.

Contribution

It introduces a novel duality theorem for non-linear multicommodity flows, along with methods to compute mutual capacities and efficient algorithms for approximate maximum flows.

Findings

Mutual capacity equals the set of feasible flows, establishing a tight duality.

A method to compute mutual capacity for pairs and sets of cuts.

Efficient algorithms for eps-approximations of maximum flow problems.

Abstract

The Max-Flow Min-Cut theorem is the classical duality result for the Max-Flow problem, which considers flow of a single commodity. We study a multiple commodity generalization of Max-Flow in which flows are composed of real-valued k-vectors through networks with arc capacities formed by regions in \R^k. Given the absence of a clear notion of ordering in the multicommodity case, we define the generalized max flow as the feasible region of all flow values. We define a collection of concepts and operations on flows and cuts in the multicommodity setting. We study the mutual capacity of a set of cuts, defined as the set of flows that can pass through all cuts in the set. We present a method to calculate the mutual capacity of pairs of cuts, and then generalize the same to a method of calculation for arbitrary sets of cuts. We show that the mutual capacity is exactly the set of feasible flows in the network, and hence is equal to the max flow. Furthermore, we present a simple class of the multicommodity max flow problem where computations using this tight duality result could run significantly faster than default brute force computations. We also study more tractable special cases of the multicommodity max flow problem where the objective is to transport a maximum real or integer multiple of a given vector through the network. We devise an augmenting cycle search algorithm that reduces the optimization problem to one with m constraints in at most \R^{(m-n+1)k} space from one that requires mn constraints in \R^{mk} space for a network with n nodes and m edges. We present efficient algorithms that compute eps-approximations to both the ratio and the integer ratio maximum flow problems.