A direct reduction of PPAD Lemke-verified linear complementarity problems to bimatrix games

Publication Date: February 1, 2013

Ilan Adler, Sushil Verma (2013), A direct reduction of PPAD Lemke-verified linear complementarity problems to bimatrix games, arXiv:1302.0067.

Abstract: The linear complementarity problem, LCP(q,M), is defined as follows. For given M,q find z such that q+Mz>=0, z>=0, z(q + M z)=0,or certify that there is no such z. It is well known that the problem of finding a Nash equilibrium for a bimatrix game (2-NASH) can be formulated as a linear complementarity problem (LCP). In addition, 2-NASH is known to be complete in the complexity class PPAD (Polynomial-time Parity Argument Directed). However, the ingeniously constructed reduction (which is designed for any PPAD problem) is very complicated, so while of great theoretical significance, it is not practical for actually solving an LCP via 2-NASH, and it may not provide the potential insight that can be gained from studying the game obtained from a problem formulated as an LCP (e.g. market equilibrium). The main goal of this paper is the construction of a simple explicit reduction of any LCP(q,M) that can be verified as belonging to PPAD via the graph induced by the generic Lemke algorithm with some positive covering vector d, to a symmetric 2-NASH. In particular, any endpoint of this graph (with the exception of the initial point of the algorithm) corresponds to either a solution or to a so-called secondary ray. Thus, an LCP problem is verified as belonging to PPAD if any secondary ray can be used to construct, in polynomial time, a certificate that there is no solution to the problem. We achieve our goal by showing that for any M,q and a positive d satisfying a certain nondegeneracy assumption with respect to M, we can simply and directly construct a symmetric 2-NASH whose Nash equilibria correspond one-to-one to the end points of the graph induced by LCP(q,M) and the Lemke algorithm with a covering vector d. We note that for a given M the reduction works for all positive d with the exception of a subset of measure 0.