# Mathematics

## Optimal Coffee shops, Numerical Integration and Kantorovich-Rubinstein duality

Suppose you want to open up 7 coffee shops so that people in the downtown area have to walk the least amount to get their morning coffee. That’s a classical problem in Optimal Transport, minimizing the Wasserstein distance between the sum of 7 Dirac measures and the (coffee-drinking) population density. But in reality things are trickier. If the 7 coffee shops go well, you want to open an 8th and a 9th and you want to remain optimal in this respect (and the first 7 are already fixed). We find optimal rates for this problem in ($W_2$) in all dimensions. Analytic Number Theory makes an appearance and, in fact, Optimal Transport can tell us something new about $\sqrt{2}$ . All of this is also related to the question of approximating an integral by sampling in a number of points and a conjectured extension of the Kantorovich-Rubinstein duality regarding the $W_1$ distance and testing of two measures against Lipschitz functions.

## Deep kernel-based distances between distributions

Optimal transport, while widespread and effective, is not the only game in town for comparing high-dimensional distributions. This talk will cover a set of related distances based on kernel methods, in particular the maximum mean discrepancy, and especially their use with learned kernels defined by deep networks. This set of distance metrics allows for effective use in a variety of applications; we will cover foundational properties and develop variants useful for distinguishing distributions, training generative models, and other machine learning applications.

## Kac goes to work: Stochastic processes as probes of the architecture of plant root systems

The past decade has seen a rapid development of data-driven plant breeding strategies based on the two significant technological developments. First, the use of high throughput DNA sequencing technology to identify millions of genetic markers on that characterize the available genetic diversity captured by the thousands of available accessions in each major crop species. Second, the development of high throughput imaging platforms for estimating quantitative traits associated with easily accessible above-ground structures such as shoots, leaves and flowers. These data-driven breeding strategies are widely viewed as the basis for rapid development of crops capable of providing stable yields in the face of global climate change. Roots and other below-ground structures are much more difficult to study yet play essential roles in adaptation to climate change including as uptake of water and nutrients. Estimation of quantitative traits from images remains a significant technical and scientific bottleneck for both above and below-ground structures. The focus of this talk, inspired by the analytical results of Kac, van den Berg and many others in the area of spectral geometry, is to describe a computational and statistical methodology that employs stochastic processes as quantitative measurement tools suitable for characterizing images of multi-scale dendritic structures such as plant root systems. The substrate for statistical analyses in Wasserstein space are hitting distributions obtained by simulation. The practical utility of this approach is demonstrated using 2D images of sorghum roots of different genetic backgrounds and grown in different environments.

## Bandit learning of Nash equilibria in monotone games

Game theory is a powerful framework to address optimization and learning of multiple interacting agents referred to as players. In a multi-agent setting, the notion of Nash equilibrium captures a desirable solution as it exhibits stability, that is, no player has incentive to deviate from this solution. From the viewpoint of learning the question is whether players can learn their Nash equilibrium strategies with limited information about the game. In this talk, I address our work on designing distributed algorithms for players so that they can learn the Nash equilibrium based only on information regarding their experienced payoffs. I discuss the convergence of the algorithm and its applicability to a large class of monotone games.

## Scaling Optimal Transport for High dimensional Learning

Optimal transport (OT) has recently gained lot of interest in machine learning. It is a natural tool to compare in a geometrically faithful way probability distributions. It finds applications in both supervised learning (using geometric loss functions) and unsupervised learning (to perform generative model fitting). OT is however plagued by the curse of dimensionality, since it might require a number of samples which grows exponentially with the dimension. In this talk, I will explain how to leverage entropic regularization methods to define computationally efficient loss functions, approximating OT with a better sample complexity. More information and references can be found on the website of our book “Computational Optimal Transport” https://optimaltransport.github.io/

## Initial value problems viewed as generalized optimal transport problems with matrix-valued density fields

The initial value problem for many important PDEs (Burgers, Euler, Hamilton-Jacobi, Navier-Stokes equations, systems of conservation laws with convex entropy, etc…) can be often reduced to a convex minimization problem that can be seen as a generalized optimal transport problem involving matrix-valued density fields. The time boundary conditions enjoy a backward-forward structure of “ballistic” type, just as in mean-field game theory.

## Large orbit closures of translation surfaces are strata or loci of double covers II

Any translation surface can be presented as a collection of polygons in the plane with sides identified. By acting linearly on the polygons, we obtain an action of GL(2,R) on moduli spaces of translation surfaces. Recent work of Eskin, Mirzakhani, and Mohammadi showed that GL(2,R) orbit closures are locally described by linear equations on the edges of the polygons. However, which linear manifolds arise this way is mysterious.

In this lecture series, we will describe new joint work that shows that when an orbit closure is sufficiently large it must be a whole moduli space, called a stratum in this context, or a locus defined by rotation by π symmetry.

We define "sufficiently large" in terms of rank, which is the most important numerical invariant of an orbit closure, and is an integer between 1 and the genus g. Our result applies when the rank is at least 1+g/2, and so handles roughly half of the possible values of rank.

The five lectures will introduce novel and broadly applicable techniques, organized as follows:

An introduction to orbit closures, their rank, their boundary in the WYSIWYG partial compactification, and cylinder deformations.

Reconstructing orbit closures from their boundaries (this talk will explicate a preprint of the same name).

Recognizing loci of covers using cylinders (this talk will follow a preprint titled “Generalizations of the Eierlegende-Wollmilchsau”).

An overview of the proof of the main theorem; marked points (following the preprint “Marked Points on Translation Surfaces”); and a dichotomy for cylinder degenerations.

Completion of the proof of the main theorem.

## Combinatorial structures in perturbative quantum field theory

I will give an overview of a few places where combinatorial structures have an interesting role to play in quantum field theory and which I have been involved in to varying degrees, from the Connes-Kreimer Hopf algebra and other renormalization Hopf algebras, to the combinatorics of Dyson-Schwinger equations and the graph theory of Feynman integrals.

For other events in this series see the quanTA events website.

## UBC/ PIMS Mathematical Sciences Young Faculty Award Colloquium: Liam Watson

What is Khovanov homology, and when is it boring?

Khovanov homology, though relatively young, is difficult to survey in an hour. This talk will nevertheless attempt to do so, by focussing on the problem of characterizing thin links—those links with simplest-possible Khovanov homology. This is a story that is still unfolding; I will describe some progress that is part of a joint project with Artem Kotelskiy and Claudius Zibrowius.

## Large orbit closures of translation surfaces are strata or loci of double covers

Any translation surface can be presented as a collection of polygons in the plane with sides identified. By acting linearly on the polygons, we obtain an action of GL(2,R) on moduli spaces of translation surfaces. Recent work of Eskin, Mirzakhani, and Mohammadi showed that GL(2,R) orbit closures are locally described by linear equations on the edges of the polygons. However, which linear manifolds arise this way is mysterious.

In this lecture series, we will describe new joint work that shows that when an orbit closure is sufficiently large it must be a whole moduli space, called a stratum in this context, or a locus defined by rotation by π symmetry.

We define "sufficiently large" in terms of rank, which is the most important numerical invariant of an orbit closure, and is an integer between 1 and the genus g. Our result applies when the rank is at least 1+g/2, and so handles roughly half of the possible values of rank.

The five lectures will introduce novel and broadly applicable techniques, organized as follows:

- An introduction to orbit closures, their rank, their boundary in the WYSIWYG partial compactification, and cylinder deformations.
- Reconstructing orbit closures from their boundaries (this talk will explicate a preprint of the same name).
- Recognizing loci of covers using cylinders (this talk will follow a preprint titled “Generalizations of the Eierlegende-Wollmilchsau”).
- An overview of the proof of the main theorem; marked points (following the preprint “Marked Points on Translation Surfaces”); and a dichotomy for cylinder degenerations.
- Completion of the proof of the main theorem.