Riemannian optimization is a powerful and active area of research that studies the optimization of functions defined on manifolds with structure. A class of functions of interest is the set of geodesically convex functions, which are functions that are convex when restricted to every geodesic. In this talk, we will present an accelerated first-order method, nearly achieving the same rates as accelerated gradient descent in the Euclidean space, for the optimization of smooth and g-convex or strongly g-convex functions defined on the hyperbolic space or a subset of the sphere. We will talk about accelerated optimization of another non-convex problem, defined in the Euclidean space, that we solve as a proxy. Additionally, for any Riemannian manifold of bounded sectional curvature, we will present reductions from optimization methods for smooth and g-convex functions to methods for smooth and strongly g-convex functions and vice versa.

This talk is based on the paper https://arxiv.org/abs/2012.03618.

-

A link for this talk will be sent to our mailing list a day or two in advance.  If you are not on the list and wish to be sent a link, please contact @email.

Abstract: In this talk we try to establish an analytic framework for studying Set-Valued Backward Stochastic Differential Equations (SVBSDE for short), motivated largely by the current studies of dynamic set-valued risk measures for multi-asset or network-based financial models. Our framework will be based on the notion of Hukuhara difference between sets, in order to compensate the lack of “inverse” operation of the traditional Minkowski addition, whence the vector space structure, in traditional set-valued analysis. We shall examine and establish a useful foundation of set-valued stochastic analysis under this algebraic framework, including some fundamental issues regarding Aumann-Itˆo integrals, especially when it is connected to the martingale representation theorem. We shall identify some fundamental challenges and propose some extensions of the existing theory that are necessary to study the SVBSDEs. This talk is based on the joint works with C¸ a˘gın Ararat and Wenqian Wu.

Following the risk-taking model of Seel and Strack, n players decide when to stop privately observed Brownian motions with drift and absorption at zero. They are then ranked according to their level of stopping and paid a rank-dependent reward. We study the optimal reward design where a principal is interested in the average performance and the performance at a given rank. While the former can be related to reward inequality in the Lorenz sense, the latter can have a surprising shape. Next, I will present the mean-field version of this problem. A particular feature of this game is to be tractable without necessarily being smooth, which turns out to offer a cautionary tale. We show that the mean field equilibrium induces n-player ε-Nash equilibria for any continuous reward function— but not for discontinuous ones. We also analyze the quality of the mean field design (for maximizing the median performance) when used as a proxy for the optimizer in the n-player game. Surprisingly, the quality deteriorates dramatically as n grows. We explain this with an asymptotic singularity in the induced n-player equilibrium distributions. (Joint work with M. Nutz)