PhD students

Wednesday 20/05 – MSA 2.240 :

• • 12:30 – 14:00 : Lunch + Poster Session

• • 14:00 – 14:45 : Javier Fernandez Piriz – University of Luxembourg
    
    Grassmannians and representations of Lie groups

    Grassmannians are objects endowed with rich geometrical structures that have been studied in algebraic geometry since the 19th century. A useful way to understand these spaces is through the seemingly unrelated theory of representations of Lie groups. The goal of this talk is to present a brief overview of the interplay between these fields and to motivate how computers are useful in answering many related questions.

• • 14:45 – 15:30 : Rodolphe Abou Assali – IECL
    
    Steklov problems and spectral inequalities in planar domains

    Classical spectral problems, such as the Dirichlet and Neumann problems, focus on the analysis of eigenvalues and eigenfunctions with applications to heat conduction, sound propagation, and vibrational modes in domains with boundaries. Other well-known problems are the Steklov and biharmonic Steklov problems with various boundary conditions. Kuttler and Sigillito established fundamental inequalities relating the eigenvalues of these problems in planar domains. These results were later extended to the scalar case on Riemannian manifolds by Hassannezhad and Siffert. We recently generalized these inequalities to the setting of differential forms. In this talk, we present these spectral problems and the Kuttler-Sigillito inequalities in planar domains, and briefly discuss their generalization.

• • 15:30 – 16:00 : Break

• • 16:00 – 16:45 : Quirijn Boeren – University of Luxembourg
    
    Cusps in the AdS/CFT correspondence

    The AdS/CFT correspondence is a powerful tool in theoretical physics, relating string theories on hyperbolic (Anti-de Sitter) manifolds to a conformal field theory on a boundary manifold. It provides some of the most promising models of quantum gravity. As often in theoretical physics the theory struggles with divergences. I will walk you through one such divergence, caused by a construction from hyperbolic geometry: a manifold with cusp—a puncture at infinite distance—can generate infinite summands to the relation, producing a divergence.

• • 16:45 – 17:30 : Valentin Clarisse – IECL
    
    General relativity and Gregory-Laflamme instability

    The Einstein equations are central to general relativity. They relate the geometry of spacetime to the distribution of matter within it. As we will see later, they form a particularly challenging system of partial differential equations to study. The first major breakthrough in mathematical relativity was achieved by Y. Choquet-Bruhat, who proved in 1952 the local-in-time existence of solutions to the Einstein equations viewed as an evolution problem. More recently, in 1993 and 1994, R. Gregory and R. Laflamme numerically demonstrated the instability of certain types of black string extensions in dimensions greater than or equal to $5$. In 2012, R.M. Wald and S. Hollands developed a fairly general method and criterion for studying the linear stability of black holes, which can be applied to establish Gregory–Laflamme-type instabilities. The article we will focus on, which is more accessible, comes from the doctoral thesis of Sam C. Collingbourne. It was submitted in 2020 and is entitled The Gregory-Laflamme Instability of the Schwarzschild Black String Exterior. It provides a direct mathematical proof of the Gregory–Laflamme linear instability in dimension $5$.

Thursday 21/05 – MSA 2.240 :

• • 9:00 – 9:45 : Katarzyna Szczerba – University of Luxembourg
    
    AI-informed Non-linear Cox Regression for Time-to-event Analysis

    The Cox proportional hazards model is the most commonly used method for multivariate survival analysis. Despite its many advantages, such as simplicity and interpretability, it has a serious drawback: it fails to capture non-linear relationships. In this study, we propose AI-informed Non-linear Cox Model, a method that uses insights from a highly predictive machine learning model, extracted with an interpretable machine learning tool, to integrate non-linear relationships into the traditional Cox model via means of splines. On simulated data with a deliberately introduced non-monotonic relationship between the predictor and the outcome variable, the AI-informed Cox model outperformed the traditional proportional hazards (PH) Cox model. Its concordance index (C-index) was also comparable to that of the best-performing machine learning model – gradient boosted Cox model. Similar results were observed when the models were applied to a prospective dataset in running.

• • 9:45 – 10:30 : Yingtong Hou – IECL
    
    Butcher series: from ordinary differential equations to Rough Path Theory and Regularity Structures

    In this talk, I will give a gentle introduction to Butcher series (B-series), Rough Path Theory, Regularity Structures, and their underlying Hopf algebras. Rough Path Theory and Regularity Structures provide pathwise frameworks for solving rough differential equations (RDEs) and singular stochastic partial differential equations (SPDEs), respectively. We will see that all these pathwise solution ansatz are obtained from iterating Taylor expansions. Therefore, Rough Path Theory and Regularity Structures can be viewed as generalisations of B-series designed for solving ordinary differential equations (ODEs). I will present the derivation of B-series-type solution ansatz for ODEs, RDEs, and SPDEs. Rooted trees and Hopf algebras appear naturally in encoding the expansions of solution ansatz. No prior background knowledge in rough analysis is required. Familiarity with Taylor expansions will be sufficient.

• • 10:30 – 11:00 : Break

• • 11:00 – 11:45 : Luís Maia – University of Luxembourg
    
    Fractional Brownian Fields at H=0: Constructions and Limit Theorems

    Fractional Brownian motion and fractional Brownian fields become singular at the endpoint H=0: the usual covariance degenerates. In this talk, I will explain two normalization that recover a meaningful object when $H=0$. The first, due to Neuman and Rosenbaum, treats one-dimensional fractional Brownian motion by subtracting a local average and rescaling. The second, due to Hager and Neuman, extends this idea to higher-dimensional fractional Brownian fields. In both cases, the normalized fields converge to log-correlated Gaussian distributions. I will then discuss results on Hermite functionals of these fields, both on fixed domains and on growing domains.

• • 11:45 – 12:30 : Juan Mardomingo-Sanz – IECL
    
    Slow-fast limits of stochastic particle systems arising in telomere biology

    The ends of linear chromosomes, called telomeres, shorten at each cell replication, eventually driving the cells to a senescent state when they become too short. The enzyme telomerase, present in cancerous cells and some unicellular organisms, elongates the telomeres and allows cells to continue replicating. Recent experiments show that if this enzyme is inactivated some rare survivors (ALT), which elongate their telomeres without telomerase, will appear and will eventually invade the cultures. I will present a simple stochastic particle system which accounts for the emergence and invasion of these ALT cells under an appropriate scaling with different speeds for each cell type.

• • 12:30 – 14:00 : Lunch

• • 14:00 – 14:45 : Szabolcs Buzogany – University of Luxembourg
    
    Galois and torsion-Kummer representations of elliptic curves

    The absolute Galois group $G_Q$ is the group of all isomorphisms from the field of all algebraic numbers to itself and remains a central object in contemporary number theory.
    A common way of studying $G_Q$ is to study its quotients, by the means of defining a group homomorphism between $G_Q$ and a well-studied group. Examples of these maps are Galois (respectively torsion-Kummer) representations, where the codomain is associated with n-torsion (respectively n-division points) of an elliptic curve. In this talk I will provide a gentle introduction to these representations.

• • 15:00 – 17:30 : Scavenger Hunt in the city (Luxembourg)

• • 19:00 : Social Dinner at Brasserie du Cercle (Luxembourg City)

Friday 22/05 – MSA 3.500 :

• • 9:00 – 9:45 : Gautier Schanzenbacher – IECL
    
    An Introduction to Hyperbolic Geometry: Surfaces, Geodesics, and Entropy

    For centuries, mathematicians tried to prove Euclid’s fifth axiom (the parallel postulate) using only the first four. In the 19th century, Gauss showed that replacing this axiom leads to a new, consistent geometry: non-Euclidean geometry. In particular, if we suppose that there are infinitely many lines parallel to a given line passing through a single point, we obtain Hyperbolic Geometry. In this talk, I will start from these foundations to define hyperbolic surfaces. We will then explore the world of curves, geodesics, and homotopy classes to understand the concept of entropy of the geodesic flow of a hyperbolic surface in the simplest way possible.

• • 9:45 – 10:30 : Francesco Tognetti – University of Luxembourg
    
    Who cares about coinduction?

    Everyone is familiar with the concept of (proof by) induction, though not as many are familiar with its dual. In this talk you will get an overview of what coinduction is, when it arises naturally and how it’s used throughout various areas of mathematics.

• • 10:30 – 11:00 : Break

• • 11:00 – 11:45 : Musbahu Idris – IECL
    
    Algorithmic Aspects of Newman Polynomials and Their Divisors

    A Newman polynomial is a polynomial with coefficients in ${0,1}$ and constant term $1$. We investigate which integer-coefficient polynomials divide a Newman polynomial, focusing on those with small Mahler measure. Using mixed-integer linear programming, we determine the divisibility status of all $8,438$ known polynomials with Mahler measure less than $1.3$. We further exhibit new polynomials that divide no Newman polynomial, improving the best known upper bound on a conjectural universal constant $\sigma$ to approximately $1.419$.

• • 11:45 – 12:30 : Francisco Pina – University of Luxembourg
    
    Statistics of Interacting Particle Systems

    Interacting particle systems can be seen as a system of N SDEs describing the evolution of a collection of agents whose behaviour depends not only on their own dynamics, but also on their interactions with the rest of the system. Such models arise in many different contexts, and a typical example is opinion dynamics, where the evolution of an individual’s opinion is influenced by the opinions of others.
    In this talk, we present the mathematical framework of interacting particle systems and discuss how statistical methods can be used to estimate the interaction law governing the system from observed particle trajectories. In particular, we introduce a nonparametric approach for estimating the underlying interaction function.

• • 12:30 – 14:00 : Lunch

A Newman polynomial is a polynomial with coefficients in {0,1} and
constant term 1. We investigate which integer-coefficient polynomials
divide a Newman polynomial, focusing on those with small Mahler measure.
Using mixed-integer linear programming, we determine the divisibility
status of all 8,438 known polynomials with Mahler measure less than 1.3.
We further exhibit new polynomials that divide no Newman polynomial,
improving the best known upper bound on a conjectural universal constant
σ to approximately 1.419.

We consider branching Brownian motion (BBM), a random process
that describes the evolution of a particle population, reproducing and
moving independently. Beyond obvious biological motivations and its link
with the F-KPP equation, BBM can be seen as a toy model for spin
glasses, such as the Sherrington-Kirkpatrick model. In this perspective,
we will introduce the Gibbs measures of BBM. We will study some of their
properties, including their connection with the so-called additive
martingales. We will also study the maximal particle of BBM (or, from
the perspective of statistical physics, the ground state of the system).
A new martingale then appears, that is, the derivative martingale. If
time allows, we will briefly present an ongoing work with Gabriel Flath and Julien Berestycki,
in which we obtain an almost sure path localization of the derivative
martingale.

We study a model of random permutations, which we call random standardized permutations, based on a sequence of i.i.d. discrete random variables. This model generalizes others, such as the riffle-shuffle and the major-index-biased permutations. We first establish an exact result on the joint distribution of the number of cycles of given lengths, involving the notion of primitive words. Thanks to this result obtained via combinatorial methods, we obtain convergence in distribution as the size of the permutation tends to infinity . This talk will be an opportunity to introduce (or recall) the method of moments, a very useful tool for proving convergence in distribution, particularly for combinatorial objects. We will present a few limit results on the distribution of “small” and “large” cycles of the permutation, as well as on the total number of cycles.

La géométrie birationnelle est la branche qui étudie les espaces “presque partout isomorphes”. Dans ce cadre, j’introduirai une opération fondamentale: l’éclatement, qui permet de modifier localement un espace en remplaçant un point par un ensemble de directions. Je présenterai ensuite une version plus flexible de cette construction, les éclatements à poids, où les différentes directions sont prises en compte de manière non uniforme.

J’introduirai brièvement les variétés horosphériques, qui fournissent un cadre particulièrement bien adapté aux actions de groupes et à une description combinatoire de la géométrie. Je terminerai par un aperçu des contractions divisorielles horosphériques, qui s’avèrent être, dans ce contexte, toutes données par des éclatements à poids.

The Wasserstein space $\mathscr{P}(M)$ associated with a closed Riemannian manifold is defined as the space of probability measures on the manifold, endowed with the so-called Otto metric, which provides it with the structure of a formal infinite-dimensional Riemannian manifold. In this talk I will describe the geometric features of this space, emphasizing its connections with optimal transport theory and some classical PDEs. I will then introduce the group of diffeomorphisms $\mathscr{D}(M)$, viewed as an Inverse Limit Hilbert Lie group, and present the Riemannian submersion structure that relates $\mathscr{D}(M)$ and $\mathscr{P}(M)$. The space $\mathscr{P}_\infty(M) \subset \mathscr{P}(M)$ of smooth positive measures is of particular interest. The geodesic convexity of such a space highly depends on the geometry of the base manifold. I will review some significant developments on this topic, mainly due to Ma, Trudinger, Wang, Loeper and Villani. If time permits, I will try to introduce some of the topics of my PhD thesis which focuses on the study of random paths on $\mathscr{P}(M)$ and its tangent bundle $T\mathscr{P}(M)$.

It is well known about the Riemann zeta function that $\zeta(2i) \in \mathbb{Q}\pi^{2i}, i\geqslant 1$, and thus all these numbers are linearly independent over $\mathbb{Q}$, since $\pi$ is transcendental.
However, very little is known about the numbers $\zeta(2i+1), i\geqslant 1$. It was proved in 1978 that $\zeta(3)$ is irrational, and in 2000 that infinitely many of these numbers are irationnal.

The aim of this talk is to introduce the core ideas used to prove such a result, and to elaborate on the new ideas that allowed many generalizations since the 2000s (better bound on the proportion of irrational numbers, broader class of functions rather than just $\zeta$, …)

We study the residual stream of multi-head Transformers in which the attention weights are i.i.d.\ random matrices across layers and heads. We identify critical scaling laws linking the depth $L$, the residual scale $\eta$, and the number of heads $H$, and show that different joint limits yield distinct homogenized effective models. To formalize these limits, we leverage the theory of stochastic modified equations. We apply this framework to Transformers at initialization and derive effective dynamics that clarify the roles of additional parameters, including the inverse temperature $\beta$, the embedding dimension $d$, and the context length $n$.

In this talk, we’ll consider various situations (favorable, semi-favorable and unfavorable) for the computation of the partition function. For each situation, an overview of the techniques used to simulate and run the inference for the models will be discussed. More specifically, we will briefly discuss : the inverse of the cumulative distribution function, maximum likelihood, Gibbs sampler, maximum pseudo-likelihood, Metropolis-Hastings algorithms and maximum likelihood MCMC.