Georgia Institute of Technology College of Sciences

Equilibrium computation is among the most significant additions to the theory of algorithms and computational complexity in the last decade - it has its own character, quite distinct from the computability of optimization problems. Our contribution to this evolving theory can be summarized in the following sentence: Natural equilibrium computation problems tend to exhibit striking dichotomies. The dichotomy for Nash equilibrium, showing a qualitative difference between 2-Nash and k- Nash for k > 2, has been known for some time. We establish a dichotomy for market equilibrium. For this purpose. we need to define the notion of Leontief-free functions which help capture the joint utility of a bundle of goods that are substitutes, e.g., bread and bagels. We note that when goods are complements, e.g., bread and butter, the classical Leontief function does a splendid job. Surprisingly enough, for the former case, utility functions had been defined only for special cases in economics, e.g., CES utility function. We were led to our notion from the high vantage point provided by an algorithmic approach to market equilibria. Note: Joint work with Jugal Garg and Ruta Mehta.

The problem of quantifying the amount of information loss due to a random transformation (or a noisy channel) arises in a variety of contexts, such as machine learning, stochastic simulation, error-correcting codes, or computation in circuits with noisy gates, to name just a few. This talk will focus on discrete channels, where both the input and output sets are finite. The noisiness of a discrete channel can be measured by comparing suitable functionals of the input and output distributions. For instance, if we fix a reference input distribution, then the worst-case ratio of output relative entropy (Kullback-Leibler divergence) to input relative entropy for any other input distribution is bounded by one, by the data processing theorem. However, for a fixed reference input distribution, this quantity may be strictly smaller than one, giving so-called strong data processing inequalities (SDPIs). I will show that the problem of determining both the best constant in an SDPI and any input distributions that achieve it can be addressed using logarithmic Sobolev inequalities, which relate input relative entropy to certain measures of input-output correlation. I will also show that SDPIs for Kullback-Leibler divergence arises as a limiting case of a family of SDPIs for Renyi divergence, and discuss the relationship to hypercontraction of Markov operators.

The rank of a bimatrix game (A, B) is defined as the rank of (A+B). For zero-sum games, i.e., rank 0, Nash equilibrium computation reduces to solving a linear program. We solve the open question of Kannan and Theobald (2005) of designing an efficient algorithm for rank-1 games. The main difficulty is that the set of equilibria can be disconnected. We circumvent this by moving to a space of rank-1 games which contains our game (A, B), and defining a quadratic program whose optimal solutions are Nash equilibria of all games in this space. We then isolate the Nash equilibrium of (A, B) as the fixed point of a single variable function which can be found in polynomial time via an easy binary search. Based on a joint work with Bharat Adsul, Jugal Garg and Milind Sohoni.

(algebraic statistics reading seminar)