Let $A$ be a square matrix. The resolvent, $(A-z I)^{-1}, z \in \mathbb{C}$, plays an important role in many applications; for example, in studying functions of $A$, one often uses the Cauchy integral formula,$$f(A)=-\frac{1}{2 \pi i} \int_{\Gamma}(A-z I)^{-1} f(z) d z$$where $\Gamma$ is the boundary of a region $\Omega$ that contains the spectrum of $A$ and on which $f$ is analytic. If $z$ is very close to a simple eigenvalue $\lambda$ of $A$ - much closer to $\lambda$ than to any other eigenvalue of $A$ - then $(A-z I)^{-1} \approx \frac{1}{\lambda-z} x y^*$, where $x$ and $y$ are right and left normalized eigenvectors of $A$ corresponding to eigenvalue $\lambda$. It is sometimes observed, however, that $(A-z I)^{-1}$ is close to a rank one matrix even when $z$ is not very close to an eigenvalue of $A$. In this case, one can write $(A-z I)^{-1} \approx \sigma_1(z) u_1(z) v_1(z)^*$, where $\sigma_1(z)$ is the largest singular value of $(A-z I)^{-1}$ and $u_1(z)$ and $v_1(z)$ are the corresponding left and right singular vectors. We use singular value/vector perturbation theory to describe conditions under which $(A-$ $z I)^{-1}$ can be well-approximated by rank one matrices for a wide range of $z$ values. If $\lambda$ is a simple ill-conditioned eigenvalue of $A$, if the smallest nonzero singular value of $A-\lambda I$ is well-separated from 0 , and if a certain other condition involving the singular vectors of $A-\lambda I$ is satisfied, then it is shown that $(A-z I)^{-1}$ is close to a rank one matrix for a wide range of $z$ values. An application of this result in comparing bounds on $\|f(A)\|$ is described [1] for example, in studying functions of $A$, one often uses the Cauchy integral formula,$$f(A)=-\frac{1}{2 \pi i} \int_{\Gamma}(A-z I)^{-1} f(z) d z$$where $\Gamma$ is the boundary of a region $\Omega$ that contains the spectrum of $A$ and on which $f$ is analytic. If $z$ is very close to a simple eigenvalue $\lambda$ of $A$ - much closer to $\lambda$ than to any other eigenvalue of $A$ - then $(A-z I)^{-1} \approx \frac{1}{\lambda-z} x y^*$, where $x$ and $y$ are right and left normalized eigenvectors of $A$ corresponding to eigenvalue $\lambda$. It is sometimes observed, however, that $(A-z I)^{-1}$ is close to a rank one matrix even when $z$ is not very close to an eigenvalue of $A$. In this case, one can write $(A-z I)^{-1} \approx \sigma_1(z) u_1(z) v_1(z)^*$, where $\sigma_1(z)$ is the largest singular value of $(A-z I)^{-1}$ and $u_1(z)$ and $v_1(z)$ are the corresponding left and right singular vectors.We use singular value/vector perturbation theory to describe conditions under which $(A-$ $z I)^{-1}$ can be well-approximated by rank one matrices for a wide range of $z$ values. If $\lambda$ is a simple ill-conditioned eigenvalue of $A$, if the smallest nonzero singular value of $A-\lambda I$ is well-separated from 0 , and if a certain other condition involving the singular vectors of $A-\lambda I$ is satisfied, then it is shown that $(A-z I)^{-1}$ is close to a rank one matrix for a wide range of $z$ values. An application of this result in comparing bounds on $\|f(A)\|$ is described [1].[-]

In an era where Artificial Intelligence (AI) is permeating virtuallly every single field of science and engineering, it is becoming critical to members of the numerical linear algebra community to understand and embrace AI , and to contribute to its advancement, and more broadly to the advancement of machine learning. What is fascinating and rather encouraging is that Numerical Linear Algebra (NLA) is at the core of machine learning and AI. In this talk we will give an overview of Deep Learning with an emphasis on Large Language Models (LLMs) and Transformers [3, 4]. The very first step of LLMs is to convert the problem into one that can he exploited by numerical methods, or to be more accurate, by optimization techniques. All AI methods rely almost entirely on essentially 4 ingredients: data, optimization methods, statistical intuition, and linear algebra. Thus, the first task is to map words or sentences into tokens which are then imbedded into Euclidean spaces. From there on, the models refer to vectors and matrices. We will show a few examples of important developments in ML, that were heavily based on linear algebra ideas. Among these, we will briefly discuss LoRa [1] a technique in which low-rank approximation was used to reduce computational cost in some models, leading to gains of a few orders of magnitude. Another contribution that used purely algebraic arguments and that had a major impact on LLMs is the article [2]. Here the main discovery is that the nonlinear ""self-attention"" in LLMs can be approximated linearly, resulting in huge savings in computations, as the computational complexity was decreased from $O\left(n^2\right)$ to $O(n)$.The talk will be mostly a survey of known recent methods in AI with the primary goal of unraveling the mathematics of Transformers. A secondary goal is to initiate a discussion on the issue of how NLA specialitst can participate in AI research.[-]

We give an arithmetic version of Tao's algebraic regularity lemma (which was itself an improved Szemerédi regularity lemma for graphs uniformly definable in finite fields). In the arithmetic regime the objects of study are pairs $(G, D)$ where $G$ is a group and $D$ an arbitrary subset, all uniformly definable in finite fields. We obtain optimal results, namely that the algebraic regularity lemma holds for the associated bipartite graph $(G, G, E)$ where $E(x, y)$ is $x y^{-1} \in D$, witnessed by a the decomposition of $G$ into cosets of a uniformly definable small index normal subgroup $H$ of $G$.[-]

We consider clustering problems that are fundamental when dealing with trajectory and time series data. The Fréchet distance provides a natural way to measure similarity of curves under continuous reparametrizations. Applied to trajectories and time series, it has proven to be very versatile as it allows local non-linear deformations in time and space. Subtrajectory clustering is a variant of the trajectory clustering problem, where the start and endpoints of trajectory patterns within the collected trajectory data are not known in advance. We study this problem in the form of a set cover problem for a given polygonal curve: find the smallest number k of representative curves such that any point on the input curve is contained in a subcurve that has Fréchet distance at most a given r to a representative curve.[-]

The aim of this talk is to present a new variation formulation of the time-dependent many-body electronic Schrödinger equation with Coulombic singularities. More precisely, its solution can actually be expressed as the solution of a global space-time quadratic minimization problem that proves to be useful for several tasks:
1) first, it is amenable to Galerkin time-space discretization schemes, using an appropriate least-square formulation
2) it enables to yield a new variational principle for the construction dynamical low-rank approximations, that is different from the classical Dirac-Frenkel variational principle
3) it enables to obtain fully certified a posteriori error estimators between the exact solution and approximate solutions.
The present analysis can be applied to the electronic many-body time-dependent Schrödinger equation with an arbitrary number of electrons and interaction potentials with Coulomb singularities.[-]

The development of a suitable, efficient and accurate numerical method to solve wave problems is encountered in many academic and industrial applications. The Boundary Integral Equation (BIE) technique, whose discretization is known as the Boundary Element Method (BEM), is an appealing alternative to classical domain method because it allows to handle problems defined on the exterior of bounded domains as easily as those defined in the interior, without the introduction of an artificial boundary to truncate the computational domain. Very recently, an Isogeometric Analysis based Boundary Element Method (IgA-BEM) has been proposed in literature for the numerical solution of frequency-domain (Helmholtz) wave problems on 3D domains admitting a multi-patch representation of the boundary surface. While being powerful and applicable to many situations, this approach shares with standard BEMs a disadvantage which can easily become significant in the 3D setting. Indeed, when the required accuracy is increased, it can soon lead to large dense linear systems, whose numerical solution requires huge memory, resulting also in important computational cost. Recently the development of fast H-matrix based direct and iterative solvers for oscillatory kernels, as the Helmholtz one, has been studied. Here, we investigate the effectiveness of the H-matrix technique, along with a suitable GMRES iterative solver, when used in the context of multi-patch IgA-BEM.[-]

As a toy model for the viscous interaction of planar vortices, we consider the solution of the two-dimensional Navier-Stokes equation with singular initial data corresponding to a pair of point vortices with opposite circulations. In the large Reynolds number regime, we construct an approximate solution which takes into account the deformation of the stream lines due to vortex interactions, as well as the corrections to the translation speed of the dipole due to finite size effects. Using energy estimates based on Arnold's variational characterization of equilibria for the Euler equation, we then show that our approximation remains valid over a very long time interval, if the viscosity is sufficiently small. This is a joint work with Michele Dolce (Lausanne), which relies on previous studies in collaboration with Vladimir Sverak (Minneapolis).[-]