PolySpin · Spin glasses and random polynomial systems: structure, algorithms and connections

The project focuses on two areas in the study of random functions in high-dimensions: mathematical Spin Glass theory and random systems of polynomial equations. Research in these fields is currently conducted by two separate mathematical communities. The study of algorithms for solving random systems has so far mostly focused on the well-known 17th problem of Steve Smale posed in 1998, which originally concerns complex polynomials. Mean-field spin glass models, on the other hand, deal with real random polynomial functions.However, Smale also posed a real version of his problem, even more difficult and much less understood. The polynomials in the real version of the problem are exactly the spherical pure p-spin models of spin glass theory. This creates a bridge between the two theories.One part of this project sets out to investigate how this can be exploited, by using the theory of spin glasses to gain insights into real random polynomial systems and the real 17th problem of Smale. We offer a new perspective by viewing the problem of solving a system as a problem of minimizing an appropriate ''energy function'' --- a common, general problem in statistical physics. Most importantly, this approach allows us to build on recent important developments on optimization of spin glasses, and specifically to adapt a Hessian Descent algorithm originally developed for the spherical models to variants of the real 17th problem of Smale.These recent advances on algorithmic optimization were inspired by a new geometric analysis for the celebrated Thouless-Anderson-Palmer (TAP) approach to the mixed p-spin models from 1977. In another part of the project we wish to extend this analysis to various other spin glass models and use it to design new optimization algorithms. Other geometric problems we seek to solve concern the structure and critical points of full-RSB models, relations of the TAP approach to pure states, and properties of the Gibbs measure.