Georgia Institute of Technology College of Sciences

In this talk, we study solvers for geometrically embedded graph structured block linear systems. The general form of such systems, PSD-Graph-Structured Block Matrices (PGSBM), arise in scientific computing, linear elasticity, the inner loop of interior point algorithms for linear programming, and can be viewed as extensions of graph Laplacians into multiple labels at each graph vertex. Linear elasticity problems, more commonly referred to as trusses, describe forces on a geometrically embedded object.We present an asymptotically faster algorithm for solving linear systems in well-shaped 3-D trusses. Our algorithm utilizes the geometric structures to combine nested dissection and support theory, which are both well studied techniques for solving linear systems. We decompose a well-shaped 3-D truss into balanced regions with small boundaries, run Gaussian elimination to eliminate the interior vertices, and then solve the remaining linear system by preconditioning with the boundaries.On the other hand, we prove that the geometric structures are ``necessary`` for designing fast solvers. Specifically, solving linear systems in general trusses is as hard as solving general linear systems over the real. Furthermore, we give some other PGSBM linear systems for which fast solvers imply fast solvers for general linear systems.Based on the joint works with Robert Schwieterman and Rasmus Kyng.

Beginning with Szemerédi’s regularity lemma, the theory of graph decomposition and graph limits has greatly increased our understanding of large dense graphs and provided a framework for graph approximation. Unfortunately, much of this work does not meaningfully extend to non-dense graphs. We present preliminary work towards our goal of creating tools for approximating graphs of intermediate degree (average degree o(n) and not bounded). We give a new random graph model that produces a graph of desired size and density that approximates the number of small closed walks of a given sparse graph (i.e., small moments of its eigenspectrum). We show how our model can be applied to approximate the hypercube graph. This is joint work with Santosh Vempala.

We study stable marriage where individuals strategically submit private preference information to a publicly known stable marriage algorithm. We prove that no stable marriage algorithm ensures actual stability at every Nash equilibrium when individuals are strategic. More specifically, we show that any rational marriage, stable or otherwise, can be obtained at a Nash equilibrium. Thus the set of Nash equilibria provides no predictive value nor guidance for mechanism design. We propose the following new minimal dishonesty equilibrium refinement, supported by experimental economics results: an individual will not strategically submit preference list L if there exists a more honest L' that yields as preferred an outcome. Then for all marriage algorithms satisfying monotonicity and IIA, every minimally dishonest equilibrium yields a sincerely stable marriage. This result supports the use of algorithms less biased than the (Gale-Shapley) man-optimal, which we prove yields the woman-optimal marriage in every minimally dishonest equilibrium. However, bias cannot be totally eliminated, in the sense that no monotonic IIA stable marriage algorithm is certain to yield the egalitarian-optimal marriage in a minimally dishonest equilibrium – thus answering a 28-year old open question of Gusfield and Irving's in the negative. Finally, we show that these results extend to student placement problems, where women are polygamous and honest, but not to admissions problems, where women are both polygamous and strategic. Based on joint work with Craig Tovey at Georgia Tech.

Optimization problems arising in decentralized multi-agent systems have gained significant attention in the context of cyber-physical, communication, power, and robotic networks combined with privacy preservation, distributed data mining and processing issues. The distributed nature of the problems is inherent due to partial knowledge of the problem data (i.e., a portion of the cost function or a subset of the constraints is known to different entities in the system), which necessitates costly communications among neighboring agents. In this talk, we present a new class of decentralized first-order methods for nonsmooth and stochastic optimization problems which can significantly reduce the number of inter-node communications. Our major contribution is the development of decentralized communication sliding methods, which can skip inter-node communications while agents solve the primal subproblems iteratively through linearizations of their local objective functions.This is a joint work with Guanghui (George) Lan and Yi Zhou.