Hansjörg Albrecher (Université de Lausanne)
Randomization and Dimensionality in Risk Modeling abstract
Abstract:
Over the years, randomization has proven to be a powerful tool in the modeling of risks on several levels: for computational purposes, in uncovering connections between different models, but also in the consideration and generation of physical and/or synthetic scenarios in risk management. A second, and in part connected, theme is the parsimonious and structure-preserving refinement of stochastic models via matrix-valued parameters, and related questions concerning the appropriate and effective dimension of models for a given purpose. This lecture will deal with various recent advances in these fields, and also illustrate concrete applications in insurance and finance, including the optimal design of reinsurance treaties and the probabilistic analysis of the profitability of blockchain mining.
10:00 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 43
Prof. Dr. Jaouad Mourtada (ENSAE/CREST)
Distribution-free robust linear regression abstract
Abstract:
We consider the problem of random-design linear regression, in a distribution-free setting where no assumption is made on the distribution of the predictive/input variables. After surveying existing approaches and indicating some improvements, we explain why they fall short in our setting. We then identify the minimal assumption on the target/output under which guarantees are possible, and describe a nonlinear prediction procedure achieving the optimal error bound with high probability.
10:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 19.1
Bogdan Stankov (Université de Lyon)
Exact values of exponential Følner functions and the Coulhon Saloff-Coste inequality abstract
Abstract:
The Følner function of a group is defined on positive integers as the smallest size of a Følner set, the boundary of which is at most 1/n of its size. Its values are finite if and only if the group is amenable. It can be thought of as encoding "how amenable a group is". The functions obtained on the same group by different generating sets are distinct, but asymptotically equivalent. Numerous results on Følner functions are known, however only up to asymptotic equivalence. In this talk, we will consider fixed generating sets and obtain (to our knowledge) the first results on the exact values of exponential Følner functions. We will present possible applications, in particular to the Coulhon Saloff-Coste inequality - in terms of whether its constants can be improved. The inequality in question provides, as a corollary, a lower bound for the Følner function. In a joint work with Christophe Pittet we obtain, for amenable groups of exponential growth, an explicit expression of the optimal multiplicative constant in the Coulhon and Saloff-Coste inequality. We prove that the optimal value over all groups for that constant is between 1 and 2.
10:30 • Université de Genève, Conseil Général 7-9, Room 1-05
Elisa Gorla (Université de Neuchâtel)
Ideals with a radical generic initial ideal abstract
Abstract:
The choice of a term order allows us to associate to any ideal I a monomial ideal, called the initial ideal of I. The initial ideal of I depends not only on the choice of a term order, but also on the system of coordinates. Nevertheless, many properties of I can be inferred from those of its initial ideal(s). For a given term order and in a generic coordinate system, however, the initial ideal of I is always the same and it is then called the generic initial ideal of I. In this talk, I will introduce a family of ideals whose generic initial ideal is independent of the choice of both the term order and of the system of coordinates. These are the multigraded homogeneous ideals which have a radical generic initial ideal. Geometrically, they correspond to varieties in a product of projective spaces and the process of passing to the initial ideals is a flat deformation - often called a Groebner deformation - to a union of linear varieties. Multigraded ideals which have a radical generic initial ideal show interesting rigidity properties, which e.g. allow us to deduce information on their universal Groebner bases. In the talk, I will present examples of ideals which belong to this class and of what we can deduce about them using this machinery. The original work in the talk is joint with Aldo Conca and Emanuela De Negri.
10:30 • Universität Basel, Rheinsprung 21
Adrien Laurent (University of Bergen)
A uniformly accurate integrator for the approximation of penalized Langevin dynamics abstract
Abstract:
n molecular dynamics, Langevin dynamics are used to model the motion of a set of particles. If some constraints are added (that appear, for instance, in the case of covalent bonds between atoms), the dynamics naturally lie on a manifold. In practice, constraints are satisfied up to a parameter ε and the solutions evolves in the vicinity of a manifold.
After a brief overview of the existing methods in R^d and on manifolds, we will introduce a new integrator whose cost and accuracy do not depend on ε. We present numerical experiments to confirm the theoretical findings, in the context of weak convergence and for the invariant measure, on a torus and on the orthogonal group in high dimension and codimension.
14:00 • Université de Genève, Section Mathématiques, Bâtiment: Conseil Général 7-9, Room 1-05
Alisa Govzmann (Universitéit Lëtzebuerg)
14:15 • EPF Lausanne, room MA A1 12 and online
Prof. Dr. Lénaïc Chizat (EPF Lausanne)
Analysis of gradient descent on wide two-layer ReLU neural networks abstract
Abstract:
In this talk, we propose an analysis of gradient descent on wide two-layer ReLU neural networks that leads to sharp characterizations of the learned predictor. The main idea is to study the training dynamics when the width of the hidden layer goes to infinity, which is a Wasserstein gradient flow. While this dynamics evolves on a non-convex landscape, we show that for appropriate initializations, its limit, when it exists, is a global minimizer. We also study the implicit regularization of this algorithm when the objective is the unregularized logistic loss, which leads to a max-margin classifier in a certain functional space. We finally discuss what these results tell us about the generalization performance, and in particular how these models compare to kernel methods.
14:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 19.1
Dr. Mario Schulz (Universität Münster)
Noncompact self-shrinkers for mean curvature flow abstract
Abstract:
In his lecture notes on mean curvature flow, Tom Ilmanen conjectured the existence of noncompact self-shrinkers with arbitrary genus. We employ min-max techniques to give a rigorous existence proof for these surfaces. Conjecturally, the self-shrinkers that we obtain have precisely one (asymptotically conical) end. We confirm this for large genus via a precise analysis of the limiting object of sequences of such self-shrinkers for which the genus tends to infinity.(joint work with Reto Buzano and Huy The Nguyen)
15:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 43 + Streaming
Prof. Dr. Jack Thorne (University of Cambridge, UK)
Equidistribution and reciprocity in number theory
16:30 • Zoom
Prof. Dr. Anders Karlsson (Université de Genève)
From linear to metric abstract
Abstract:
In recent decades metric geometry has become increasingly important and widespread. I will describe certain metric space tools and concepts that are analogous to theorems and notions in functional analysis. Applications to topics in ergodic theory, complex analysis, topology, game theory and deep learning may be mentioned. In particular, even for purely linear problems, metric methods can be useful. The oldest instance might be studies of groups of 2x2 matrices acting instead on the hyperbolic plane. More than that, it turns out that metric notions are in several situations necessary even to formulate a statement that is valid in full generality, almost like a completion of linear spaces with metric notions (on a different level than the usual completion of the rationals by the reals).
17:15 • Université de Fribourg, room Phys 2.52