We study Langevin dynamics with noise projected onto the directions orthogonal to an isometric group action. This mathematical model is introduced to shed new light on the effects of symmetry on stochastic gradient descent for over-parametrized models. Our main result identifies a novel form of implicit regularization: when the initial and target density are both invariant under the group action, Langevin dynamics with projected noise is equivalent in law to Langevin dynamics with isotropic diffusion but with an additional drift term proportional to the negative log volume of the group orbit. We prove this result by constructing a coupling of the two processes via a third process on the group itself, and identify the additional drift as the mean curvature of the orbits.

The theory of force-free electrodynamics (FFE) provides a robust framework for modeling the magnetospheres of compact objects, where the electromagnetic field's energy density dominates the surrounding plasma. Central to this theory is the existence of two-dimensional integral submanifolds, or field sheets, which foliate the spacetime. While it is established that every null force-free field possesses an associated 2-D null geodesic foliation [see Govind Menon, Classical Quantum Gravity 38, 045011 (2021) for details], the converse -identifying which null geodesic congruences can support a force-free solution-remains a nontrivial computational challenge. In this paper, we extend the foliation-based approach to null FFE by addressing two primary obstacles to the existence of a solution: the equipartition of null mean curvature and the involutivity of the field sheet distribution. We prove a general existence theorem demonstrating that for any given null geodesic congruence, there always exists a local rotation of a 2-D basis transverse to the geodesic congruence that satisfies the equipartition condition. Furthermore, we establish that a shear-free null geodesic congruence is sufficient to guarantee the existence of an arbitrary function of three variables such that any choice of such a function will generate a null field sheet foliation. Additionally, each unique foliation will be associated with a null force-free field that further contains an arbitrary function of two variables. These results are formally linked to the vanishing of the shear tensor, providing a coordinate-independent geometric criterion for the existence of null FFE solutions. We illustrate these theorems with explicit examples in Schwarzschild and Kerr geometries and present new, nontrivial exact null solutions in flat spacetime and for the C-metric.

We derive robust long-time a-priori estimates for the Navier-Stokes equation in a two-dimensional infinite strip which are uniform in the Reynolds number. These estimates provide several new scale invariant upper bounds for the size of the mixing layer in the Kelvin-Helmholtz instability.

Diffusion limits provide a framework for the asymptotic analysis of stochastic gradient descent (SGD) schemes used in machine learning. We consider an alternative framework, the Riemannian Langevin equation (RLE), that generalizes the classical paradigm of equilibration in R^n to a Riemannian manifold (M^n, g). The most subtle part of this equation is the description of Brownian motion on (M^n, g). Explicit formulas are presented for some fundamental cones.

This work introduces a construction of conformal processes that combines the theory of branching processes with chordal Loewner evolution. The main novelty lies in the choice of driving measure for the Loewner evolution: given a finite genealogical tree $\mathcal{T}$, we choose a driving measure for the Loewner evolution that is supported on a system of particles that evolves by Dyson Brownian motion at inverse temperature $\beta \in (0,\infty]$ between birth and death events. When $\beta=\infty$, the driving measure degenerates to a system of particles that evolves through Coulombic repulsion between branching events. In this limit, the following graph embedding theorem is established: When $\mathcal{T}$ is equipped with a prescribed set of angles, $\{\theta_v \in (0,\pi/2)\}_{v \in \mathcal{T}}$ the hull of the Loewner evolution is an embedding of $\mathcal{T}$ into the upper half-plane with trivalent edges that meet at angles $(2\theta_v,2\pi-4\theta_v,2\theta_v)$ at the image of each edge $v$. We also study the scaling limit when $\beta\in (0,\infty]$ is fixed and $\mathcal{T}$ is a binary Galton-Watson process that converges to a continuous state branching process. We treat both the unconditioned case (when the Galton-Watson process converges to the Feller diffusion) and the conditioned case (when the Galton-Watson tree converges to the continuum random tree). In each case, we characterize the scaling limit of the driving measure as a superprocess. In the unconditioned case, the scaling limit is the free probability analogue of the Dawson-Watanabe superprocess that we term the Dyson superprocess.

We study the Riemannian geometry of the Deep Linear Network (DLN) as a foundation for a thermodynamic description of the learning process. The main tools are the use of group actions to analyze overparametrization and the use of Riemannian submersion from the space of parameters to the space of observables. The foliation of the balanced manifold in the parameter space by group orbits is used to define and compute a Boltzmann entropy. We also show that the Riemannian geometry on the space of observables defined in [2] is obtained by Riemannian submersion of the balanced manifold. The main technical step is an explicit construction of an orthonormal basis for the tangent space of the balanced manifold using the theory of Jacobi matrices.

We construct an analogue of Dyson Brownian motion in the Siegel half-space H that we term Siegel Brownian motion. Given \beta in (0,\infty], a stochastic flow for Z_t in H is introduced so that the law of the eigenvalues \lambda_t of the cross ratio matrix R(Z_t,iI_n) is determined by the Ito differential equation corresponds to stochastic gradient ascent of a function S. S turns out to be the log volume of isospectral orbit in H and can be understood as a Boltzmann entropy. In the limit \beta=\infty, the group orbits evolve by motion by minus a half times mean curvature.

This article provides an expository account of training dynamics in the Deep Linear Network (DLN) from the perspective of the geometric theory of dynamical systems. Rigorous results by several authors are unified into a thermodynamic framework for deep learning. The analysis begins with a characterization of the invariant manifolds and Riemannian geometry in the DLN. This is followed by exact formulas for a Boltzmann entropy, as well as stochastic gradient descent of free energy using a Riemannian Langevin Equation. Several links between the DLN and other areas of mathematics are discussed, along with some open questions.

We study the Rayleigh--Taylor instability for two miscible, incompressible, inviscid fluids. Scale invariant estimates for the size of the mixing zone and coarsening of internal structures in the fully nonlinear regime are established following techniques introduced for the Saffman--Taylor instability in [G. Menon and F. Otto, Comm. Math. Phys., 257 (2005), pp. 303--317]. These bounds provide optimal scaling laws and reveal the strong role of dissipation in slowing down mixing.

Force-Free Electrodynamics (FFE) is known to describe the highly magnetized plasma around pulsars and astrophysical black holes. The equations describing FFE are highly non-linear and the task of finding a well-defined analytical solution has been unyielding. However, FFE can also be understood in terms of 2-dimensional foliations of the ambient 4-dimensional spacetime. The study of the foliations can provide significant insights into the structure of the force-free fields and such foliations can be exploited to generate new null and non-null solutions. In this paper, we present several nonnull solutions to the FFE equations in the Kerr spacetime obtained by applying the aforementioned methods of non-null foliations.

Force-free electrodynamics is the theoretical paradigm used to describe electromagnetic fields in a region where the inertia of plasma is negligible compared to the strength of the electromagnetic field. While these fields are studied extensively around accreting black holes in an attempt to describe energy extraction, force-free fields also routinely appear in the study of cosmological magnetic fields. Despite this, there are no systematic studies of exact force-free fields in an expanding universe. In this paper, we use geometric methods to find a wide variety of force-free solutions in a fixed Friedmann-Lemaitre-Robertson-Walker metric background. The method we use can be directly generalized to any arbitrary electrically neutral space-time, and hence, provides a powerful tool to study force-free fields in general.

We study the Rayleigh-Taylor instability for two miscible, incompressible, inviscid fluids. Scale-invariant estimates for the size of the mixing zone and coarsening of internal structures in the fully nonlinear regime are established following techniques introduced for the Saffman-Taylor instability in [10]. These bounds provide optimal scaling laws and reveal the strong role of dissipation in slowing down mixing.

This paper introduces a probabilistic formulation for the isometric embedding of a Riemannian manifold (M^n,g) into Euclidean space ℝ^q. We show that an embedding u: M →ℝ^q is isometric if and only if the intrinsic and extrinsic constructions of Brownian motion on u(M)⊂ℝ^q yield processes with the same law. The equivalence is first established for smooth embeddings and it is followed by a renormalization procedure for C^1,α embeddings, α >12. This formulation is based on a gedanken experiment that relates the intrinsic and extrinsic constructions of Brownian motion on an embedded manifold to the measurement of geodesic distance by observers in distinct frames of reference. This viewpoint provides a thermodynamic formalism for the isometric embedding problem that is suited to applications in geometric deep learning, stochastic optimization and turbulence.

This paper introduces two explicit schemes to sample matrices from Gibbs distributions on $\mathcal S^{n,p}_+$, the manifold of real positive semi-definite (PSD) matrices of size $n\times n$ and rank $p$. Given an energy function $\mathcal E:\mathcal S^{n,p}_+\to \mathbb{R}$ and certain Riemannian metrics $g$ on $\mathcal S^{n,p}_+$, these schemes rely on an Euler-Maruyama discretization of the Riemannian Langevin equation (RLE) with Brownian motion on the manifold. We present numerical schemes for RLE under two fundamental metrics on $\mathcal S^{n,p}_+$: (a) the metric obtained from the embedding of $\mathcal S^{n,p}_+ \subset \mathbb{R}^{n\times n} $; and (b) the Bures-Wasserstein metric corresponding to quotient geometry. We also provide examples of energy functions with explicit Gibbs distributions that allow numerical validation of these schemes.

We construct Dyson Brownian motion for $\beta \in (0,\infty]$ by adapting the extrinsic construction of Brownian motion on Riemannian manifolds to the geometry of group orbits within the space of Hermitian matrices. When $\beta$ is infinite, the eigenvalues evolve by Coulombic repulsion and the group orbits evolve by motion by (minus one half times) mean curvature.

The deep linear network (DLN) is a model for implicit regularization in gradient based optimization of overparametrized learning architectures. Training the DLN corresponds to a Riemannian gradient flow, where the Riemannian metric is defined by the architecture of the network and the loss function is defined by the learning task. We extend this geometric framework, obtaining explicit expressions for the volume form, including the case when the network has infinite depth. We investigate the link between the Riemannian geometry and the training asymptotics for matrix completion with rigorous analysis and numerics. We propose that implicit regularization is a result of bias towards high state space volume.

An interplay between algorithmic and mechanistic viewpoints is necessary to quantify the rates of transition between different states in developmental pathways. An ab initio method that uses only information theory and geometry to model the conformational entropy of linkages is discussed. This approach also reveals some ties between biological questions and the foundations of mathematics.

In this paper, we present a covariant formalism that connects solutions to force-free electrodynamics in the non-null case and foliations of spacetime. In doing so, we are also able to derive an expression of the general non-null current density vector. Just as in the null case, solutions in the non-null case can give rise to a dual solution, however, as is shown below, this can happen only when the solution describes a vacuum field. All theorems are illustrated with previously known solutions.

This paper provides an introduction to an information theoretic formulation of the embedding problem for Riemannian manifolds developed by the author. The main new construct is a stochastic relaxation scheme for embedding problems and hard constraint systems. This scheme is introduced with examples and context.