Filter off: No filter for categories
We consider the adjoint-free random-plane iteration algorithm for finding the top eigenvector of a matrix proposed in (SIAM J. Matrix Anal. Appl. 47 (2026), 1349-1374). Unlike standard methods such as the power method or gradient descent, the above algorithm lacked linear convergence guarantees in the presence of a spectral gap. In this paper, we establish explicit spectral-gap-dependent global linear convergence rates in expectation, providing quantitative guarantees for the empirically observed convergence.
Joint work with Oleh Melynk, https://arxiv.org/abs/2610.42071.
TBA
TBA __________________________ Invited by Prof. Christian Hainzl
We consider a novel robust utility maximisation problem under bounded cumulative transition-rate uncertainty within the class of non-Markovian marked point processes on a finite state space. The admissible biometric scenarios are described by path-dependent cumulative transition rates restricted by path-dependent upper and lower bounds, allowing for model uncertainty beyond the classical Markovian setting. Utility is maximised over the class of admissible controls, while Nature chooses a least favourable biometric scenario. We establish a martingale optimality principle providing a general sufficient condition for optimality. As a first application, we characterise worst-case and best-case prospective reserves for life and health insurance contracts with reserve-dependent payments. These reserves are described by a non-standard backwards stochastic differential equation, for which we prove a novel existence and uniqueness result. The corresponding extremal biometric scenarios admit an intuitive path-wise interpretation in terms of the sums-at-risk: transitions increasing the reserve are assigned their upper transition-rate bounds, while transitions decreasing it are assigned their lower bounds. As a second application, we study a robust consumption-insurance problem under biometric uncertainty. For power utility preferences, we derive explicit path-dependent feedback forms for optimal consumption and transition-contingent insurance coverage. The resulting worst-case transition rates depend endogenously on the utility-valued sums-at-risk and relative risk aversion. Our results extend classical Markovian robust reserving and consumption-insurance models to a genuinely non-Markovian setting. Based on joint work with Philipp Carsten Hornung (University of Copenhagen), cf. [1].
TBA
Lean is a popular proof assistant (also called interactive theorem prover) and programming language introduced by Leonardo de Moura in 2013. I will briefly review the historical development of proof assistants since the 1960s. Then I will present Lean, Lean's Mathlib library of formally verified mathematics, and Mathlib's fast-growing AI-developed extension Tau Ceti. The talk will include a live demonstration of Lean and Mathlib. I will mention some success stories about formalized mathematics using Lean. I will conclude by describing some of my team's research on developing and using Lean. ____________________________ Invited by Prof. Gitta Kutyniok
t.b.a.
We consider the numerical minimization of nonlocal energies with singular power-law kernels. Their discretizations couple every grid point to every other one, so the cost of a single evaluation decides what can be computed.
First, interaction energies with competing attractive and repulsive power-law terms, as in generalized Gamow-type problems and collective behavior models, minimized over densities with a pointwise upper bound. We follow the Wasserstein gradient flow, discretized with stabilized Q1 finite elements and a logarithmic barrier. The velocity is a convolution and can be computed by FFT. The computed minimizers are characteristic functions of sets. As the mass grows, the components lose convexity well before a discontinuous transition to an annulus, with a narrow window in which both configurations are stable.
Second, fractional total variation denoising with L1 fidelity, for which a grid discretization Gamma-converges. Since the sign of u(x) - u(y) enters every term of the double sum, neither FFT nor the classical fast multipole method applies. Sorting the values within each box of the multipole quadtree and forming prefix sums of the moments restores an O(N log N) evaluation for any piecewise polynomial nonlinearity, so that grids with tens of millions of pixels become feasible. Whether a small feature survives denoising then turns out to depend on its neighbours.
Joint work with K. Bessas (Pavia), A. Bonito (College Station), J.-P. Kniesel (Freiburg), L. Striet (Freiburg), S. Wojtowytsch (Pittsburgh)
Balanced least-squares assignments (LSAs) are a common method for partitioning datasets, as the resulting clusters satisfy a valuable separation property: every pair of clusters can be separated by a hyperplane within a cell complex. We are interested in applications where there is a need to gradually transition from an initial LSA to a target LSA while retaining separability throughout. Such applications naturally arise in resource redistribution efforts and can be modeled as a sequence of balanced LSAs. The computation of each step of the transition is possible through classic but expensive parametric optimization techniques where a linear program (LP) is solved to identify the next LSA.
We devise a routine to combine progress in a linear transition of sites with a desired change in cluster sizes. This shortens the sequence of LSAs by making progress toward both objectives simultaneously, thereby eliminating intermediate LSAs that would otherwise be required. Additionally, we reduce the size of LPs being solved in each iteration. Through a collection of upper-envelope computations, we identify a small candidate set of data points that can be moved between clusters. This leads to a coreset algorithm in which one may restrict the setup and solution of LPs in each iteration to that candidate set. Computational experiments confirm that these techniques reduce the number of steps in the transition while improving overall computational efficiency.