Deep Learning (DL) as a core subfield of Machine Learning and Artificial Intelligence is at the center of extraordinary technological progress. However, despite of remarkable advances in applications and theoretical analysis, the conceptual and rigorous reasons for the functioning of DL networks are at present not clearly understood (the problem of “interpretability"). In this talk, we present some recent results aimed at the rigorous mathematical understanding of how and why supervised learning works. For underparametrized DL networks, we explicitly construct global, zero loss cost minimizers for sufficiently clustered data. In addition, we derive effective equations governing the cumulative biases and weights, and show that gradient descent corresponds to a dynamical process in the input layer, whereby clusters of data are progressively reduced in complexity ("truncated") at an exponential rate that increases with the number of data points that have already been truncated. For overparametrized DL networks, we prove that the gradient descent flow is homotopy equivalent to a geometrically adapted flow that induces a (constrained) Euclidean gradient flow in output space. If a certain rank condition holds, the latter is, upon reparametrization of the time variable, equivalent to simple linear interpolation. This in turn implies zero loss minimization and the phenomenon known as "neural collapse”. A majority of this work is joint with Patricia Munoz Ewald (UT Austin).

Several recent studies considered the high-friction limit for systems arising in fluid mechanics. Following this approach, we rigorously derive the nonlocal Cahn-Hilliard equation as a limit of the nonlocal Euler-Korteweg equation using the relative entropy method. Applying the recent result about relations between non-local and local Cahn-Hilliard, we also derive rigorously the large-friction nonlocal- to-local limit. The result is formulated for dissipative measure-valued solutions of the nonlocal Euler-Korteweg equation which are known to exist on arbitrary intervals of time. This approach provides a new method to derive equations not enjoying classical solutions via relative entropy method by introducing the nonlocal effect in the fluid equation. During the talk I will also discuss the high-friction limit of the Euler-Poisson system.

The stationary problem for the free boundary incompressible Navier-Stokes equations lies at the confluence of two distinct lines of inquiry in fluid mechanics. The first views the dynamic problem as an initial value problem. In this context, the stationary problem arises naturally as a special type of global-in-time solution with stationary sources of force and stress. One then expects solutions to the stationary problem to play an essential role in the study of long-time asymptotics or attractors for the dynamic problem. The second line of inquiry, which dates back essentially to the beginning of mathematical fluid mechanics, concerns the search for traveling wave solutions. In this context, a huge literature exists for the corresponding inviscid problem, but progress on the viscous problem was initiated much more recently in the work of the speaker and co-authors. For technical reasons, these results were only able to produce traveling solutions with nontrivial wave speed. In this talk we will discuss the well-posedness theory for the stationary problem and show that the solutions thus obtained lie along a one-parameter family of slowly traveling wave solutions. This is joint work with Noah Stevenson.

We generalize the semi-analytic standing-wave framework of Schwartz and Whitney (1981) and Amick and Toland (1987) to finite-depth standing gravity waves. We propose an appropriate Stokes-expansion ansatz and iterative algorithm to solve the system of differential equations governing the expansion coefficients. We then present a more efficient algorithm that allows us to compute the asymptotic solution to higher orders. Finally, we conclude with numerical simulations of the algorithms implemented in multiple-precision arithmetic on a supercomputer to study the effects of small divisors and the analytic properties of rational approximations of the computed solutions. *Joint work with Prof. Jon Wilkening, University of California, Berkeley. **This is joint seminar with the Numerical Analysis Seminar.

We develop a unified framework for the analysis of several well-known and empirically efficient data assimilation techniques derived from various Gaussian approximations of the Bayesian filtering schemes for geophysical-type dissipative dynamics with quadratic nonlinearities. We establish rigorous results on (time-asymptotic) accuracy and stability of these algorithms with general covariance and observation operators. The accuracy and stability results for EnKF and EnSRKF for dissipative PDEs are, to the best of our knowledge, completely new in this general setting. It turns out that a hitherto unexploited cancellation property involving the ensemble covariance and observation operators and the concept of covariance localization in conjunction with covariance inflation play a pivotal role in the accuracy and stability for EnKF and EnSRKF. Our approach also elucidates the links, via determining functionals, between the approximate-Bayesian and control-theoretic approaches to data assimilation. We consider the ‘‘model’’ dynamics governed by the two-dimensional incompressible Navier-Stokes equations and observations given by noisy measurements of averaged volume elements or spectral/modal observations of the velocity field. In this setup, several continuous-time data assimilation techniques, namely the so-called Nudging (AOT)-algorithm, 3DVar, EnKF and EnSRKF reduce to a stochastically forced Navier-Stokes equations. For the first time, we derive conditions for accuracy and stability of EnKF and EnSRKF. Our analysis reveals an interplay between the resolution of the observations (roughly, the ‘richness’ of the observation space) associated with the observation operator underlying the data assimilation algorithms and covariance inflation and localization which are employed in practice for improved filter performance. This is joint work with Dr. Michal Branicki (University of Edinburgh).

We investigate a model for dynamic fracture in viscoelastic materials. The sharp crack interface is regularized with a phase-field approximation, and for the phase-field variable a viscous evolution with a quadratic dissipation potential is employed. A non-smooth penalization prevents material healing.The viscoelastic momentum balance is formulated as a first order system and coupled in a nonlinear way to the non-smooth evolution equation of the phase field. We give a full discretization in time and space, using a discontinuous Galerkin method for the first order system. Based on this, we show the existence of discrete solutions and, as the step size in space and time tends to zero, we prove their convergence to a suitable notion of weak solution of the system. We discuss our modeling approach both at small and at finite strains and point out mathematical challenges. **This is joint work with Sven Tornquist (FUB), Christian Wieners (Karlsruhe), and Kerstin Weinberg (Siegen).

We consider the Kuramoto-Sakaguchi-Fokker-Planck equation (namely, parabolic Kuramoto-Sakaguchi, or Kuramoto-Sakaguchi equation, which is a nonlinear parabolic integro-differential equation) with inertia and white noise effects. We study the hydrodynamic limit of this Kuramoto-Sakaguchi equation. During showing this main result, as a support, we also prove a Hardy-type inequality over the whole real line.