Christos Papadimitriou gave an overview talk at the NIPS conference last week and mentioned the new result by Chen and Deng showing that, given the payoff matrix, the complexity of computing a 2-Player Nash Equilibrium is PPAD-complete.
NIPS is mainly an AI conference and some of my Machine Learning
friends, who have a good understanding of NP-completeness, went around
asking who knew what PPAD-complete meant. They got many blank stares
and a few confusing answers.
So what is PPAD? Let's start with TFNP (Total Functions in
NP). Consider a polynomial-time computable predicate P(x,y) where for
every x there is at least one y such that P(x,y) is true. A TFNP
function is the problem of given an input x, finding a y such that
The interesting TFNP functions come from combinatorial theorems that
guarantee solutions. One such theorem, called the Polynomial Parity
Argument (PPA), is given a finite graph
consisting of lines and cycles (every node has degree at most 2), there
is an even number of endpoints.
The class PPAD is defined using directed graphs based on PPA.
Formally PPAD is defined by its complete problem (from an earlier post):
Given an exponential-size
directed graph with every node having in-degree and out-degree at most
one described by a polynomial-time computable function f(v) that
outputs the predecessor and successor of v, and a vertex s with a
successor but no predecessors, find a t≠s that either has no
successors or no predecessors.
Besides two player Nash, arbitrary k-player Nash and a discrete
version of finding Brouwer fixed points are also complete for PPAD.
So how do we explain this Nash Equilibrium result say when we try to
explain the result to economists and AI folk. Just say Nash
Equilibrium for 2-player games has the same complexity as k-player
Nash and finding fixed points and we have some evidence that no efficient
solutions exist for such problems.