Publications

2025

• Accelerating Nash Learning from Human Feedback via Mirror Prox
  
  • Tiapkin Daniil
  • Calandriello Daniele
  • Belomestny Denis
  • Moulines Eric
  • Naumov Alexey
  • Rasul Kashif
  • Valko Michal
  • Menard Pierre
  , 2025. Traditional Reinforcement Learning from Human Feedback (RLHF) often relies on reward models, frequently assuming preference structures like the Bradley-Terry model, which may not accurately capture the complexities of real human preferences (e.g., intransitivity). Nash Learning from Human Feedback (NLHF) offers a more direct alternative by framing the problem as finding a Nash equilibrium of a game defined by these preferences. In this work, we introduce Nash Mirror Prox ($\mathtt{Nash-MP}$), an online NLHF algorithm that leverages the Mirror Prox optimization scheme to achieve fast and stable convergence to the Nash equilibrium. Our theoretical analysis establishes that Nash-MP exhibits last-iterate linear convergence towards the $β$-regularized Nash equilibrium. Specifically, we prove that the KL-divergence to the optimal policy decreases at a rate of order $(1+2β)^{-N/2}$, where $N$ is a number of preference queries. We further demonstrate last-iterate linear convergence for the exploitability gap and uniformly for the span semi-norm of log-probabilities, with all these rates being independent of the size of the action space. Furthermore, we propose and analyze an approximate version of Nash-MP where proximal steps are estimated using stochastic policy gradients, making the algorithm closer to applications. Finally, we detail a practical implementation strategy for fine-tuning large language models and present experiments that demonstrate its competitive performance and compatibility with existing methods. (10.48550/arXiv.2505.19731)
  DOI : 10.48550/arXiv.2505.19731
• Bridging multifluid and drift-diffusion models for bounded plasmas
  
  • Gangemi G M
  • Alvarez Laguna Alejandro
  • Massot M.
  • Hillewaert K.
  • Magin T.
  Physics of Plasmas, American Institute of Physics, 2025, 32 (2), pp.023502. Fluid models represent a valid alternative to kinetic approaches in simulating low-temperature discharges: a well-designed strategy must be able to combine the ability to predict a smooth transition from the quasineutral bulk to the sheath, where a space charge is built at a reasonable computational cost. These approaches belong to two families: multifluid models, where momenta of each species are modeled separately, and drift-diffusion models, where the dynamics of particles is dependent only on the gradient of particle concentration and on the electric force. It is shown that an equivalence between the two models exists and that it corresponds to a threshold Knudsen number, in the order of the square root of the electron-to-ion mass ratio; for an argon isothermal discharge, this value is given by a neutral background pressure Pn≳1000 Pa. This equivalence allows us to derive two analytical formulas for a priori estimation of the sheath width: the first one does not need any additional hypothesis but relies only on the natural transition from the quasineutral bulk to the sheath; the second approach improves the prediction by imposing a threshold value for the charge separation. The new analytical expressions provide better estimations of the floating sheath dimension in collisions-dominated regimes when tested against two models from the literature. (10.1063/5.0240640)
  DOI : 10.1063/5.0240640
• Ergodic control of a heterogeneous population and application to electricity pricing
  
  • Jacquet Quentin
  • van Ackooij Wim
  • Alasseur Clémence
  • Gaubert Stéphane
  IEEE Transactions on Automatic Control, Institute of Electrical and Electronics Engineers, 2025. We consider a control problem for a heterogeneous population composed of agents able to switch at any time between different options. The controller aims to maximize an average gain per time unit, supposing that the population is of infinite size. This leads to an ergodic control problem for a “mean-field” Markov Decision Process in which the state space is a product of simplices, and the population evolves according to controlled linear dynamics. By exploiting contraction properties of the dynamics in Hilbert’s projective metric, we prove that the infinite-dimensional ergodic eigenproblem admits a solution and show that the latter is in general non unique. This allows us to obtain optimal strategies, and to quantify the gap between steady-state strategies and optimal ones. In particular, we prove in the one-dimensional case that there exist cyclic policies – alternating between discount and profit taking stages – which secure a greater gain than constant-price policies. On numerical aspects, we develop a policy iteration algorithm with “on-the-fly” generated transitions, specifically adapted to decomposable models, leading to substantial memory savings. We finally apply our results on realistic instances coming from an electricity pricing problem encountered in the retail markets, and numerically observe the emergence of cyclic promotions for sufficient inertia in the customer behavior.
• Maxwell's equations with hypersingularities at a negative index material conical tip
  
  • Bonnet-Ben Dhia Anne-Sophie
  • Chesnel Lucas
  • Rihani Mahran
  Pure and Applied Analysis, Mathematical Sciences Publishers, 2025, 7 (1), pp.127–169. We study a transmission problem for the time harmonic Maxwell's equations between a classical positive material and a so-called negative index material in which both the permittivity ε and the permeability µ take negative values. Additionally, we assume that the interface between the two domains is smooth everywhere except at a point where it coincides locally with a conical tip. In this context, it is known that for certain critical values of the contrasts in ε and in µ, the corresponding scalar operators are not of Fredholm type in the usual H^1 spaces. In this work, we show that in these situations, the Maxwell's equations are not well-posed in the classical L^2 framework due to existence of hypersingular fields which are of infinite energy at the tip. By combining the T-coercivity approach and the Kondratiev theory, we explain how to construct new functional frameworks to recover well-posedness of the Maxwell's problem. We also explain how to select the setting which is consistent with the limiting absorption principle. From a technical point of view, the fields as well as their curls decompose as the sum of an explicit singular part, related to the black hole singularities of the scalar operators, and a smooth part belonging to some weighted spaces. The analysis we propose rely in particular on the proof of new key results of scalar and vector potential representations of singular fields.
• Understanding the worst-kept secret of high-frequency trading
  
  • Pulido Sergio
  • Rosenbaum Mathieu
  • Sfendourakis Emmanouil
  Finance and Stochastics, Springer Verlag (Germany), 2025. Volume imbalance in a limit order book is often considered as a reliable indicator for predicting future price moves. In this work, we seek to analyse the nuances of the relationship between prices and volume imbalance. To this end, we study a market-making problem which allows us to view the imbalance as an optimal response to price moves. In our model, there is an underlying efficient price driving the mid-price, which follows the model with uncertainty zones. A single market maker knows the underlying efficient price and consequently the probability of a mid-price jump in the future. She controls the volumes she quotes at the best bid and ask prices. Solving her optimization problem allows us to understand endogenously the price-imbalance connection and to confirm in particular that it is optimal to quote a predictive imbalance. Our model can also be used by a platform to select a suitable tick size, which is known to be a crucial topic in financial regulation. The value function of the market maker's control problem can be viewed as a family of functions, indexed by the level of the market maker's inventory, solving a coupled system of PDEs. We show existence and uniqueness of classical solutions to this coupled system of equations. In the case of a continuous inventory, we also prove uniqueness of the market maker's optimal control policy.
• A stochastic use of the Kurdyka-Lojasiewicz property: Investigation of optimization algorithms behaviours in a non-convex differentiable framework
  
  • Fest Jean-Baptiste
  • Repetti Audrey
  • Chouzenoux Emilie
  Foundations of Data Science, American Institute of Mathematical Sciences, 2025. Asymptotic analysis of generic stochastic algorithms often relies on descent conditions. In a convex setting, some technical shortcuts can be considered to establish asymptotic convergence guarantees of the associated scheme. However, in a non-convex setting, obtaining similar guarantees is usually more complicated, and relies on the use of the Kurdyka-Łojasiewicz (KŁ) property. While this tool has become popular in the field of deterministic optimization, it is much less widespread in the stochastic context and the few works making use of it are essentially based on trajectory-by-trajectory approaches. In this paper, we propose a new framework for using the KŁ property in a non-convex stochastic setting based on conditioning theory. We show that this framework allows for deeper asymptotic investigations on stochastic schemes verifying some generic descent conditions. We further show that our methodology can be used to prove convergence of generic stochastic gradient descent (SGD) schemes, and unifies conditions investigated in multiple articles of the literature.
• Adaptive Destruction Processes for Diffusion Samplers
  
  • Gritsaev Timofei
  • Morozov Nikita
  • Tamogashev Kirill
  • Tiapkin Daniil
  • Samsonov Sergey
  • Naumov Alexey
  • Vetrov Dmitry
  • Malkin Nikolay
  , 2025. This paper explores the challenges and benefits of a trainable destruction process in diffusion samplers -- diffusion-based generative models trained to sample an unnormalised density without access to data samples. Contrary to the majority of work that views diffusion samplers as approximations to an underlying continuous-time model, we view diffusion models as discrete-time policies trained to produce samples in very few generation steps. We propose to trade some of the elegance of the underlying theory for flexibility in the definition of the generative and destruction policies. In particular, we decouple the generation and destruction variances, enabling both transition kernels to be learned as unconstrained Gaussian densities. We show that, when the number of steps is limited, training both generation and destruction processes results in faster convergence and improved sampling quality on various benchmarks. Through a robust ablation study, we investigate the design choices necessary to facilitate stable training. Finally, we show the scalability of our approach through experiments on GAN latent space sampling for conditional image generation. (10.48550/arXiv.2506.01541)
  DOI : 10.48550/arXiv.2506.01541
• Efficient treatment of the model error in the calibration of computer codes: the Complete Maximum a Posteriori method
  
  • Kahol Omar
  • Congedo Pietro Marco
  • Le Maitre Olivier
  • Denimal Goy Enora
  International Journal for Uncertainty Quantification, Begell House Publishers, 2025. <div><p>Computer models are widely used for the prediction of complex physical phenomena. Based on observations of these physical phenomena, it is possible to calibrate the model parameters. In most cases, such computer models are mis-specified, and the calibration process must be improved by including a model error term. The model error hyperparameters are, however, rarely learned jointly with the model parameters to reduce the dimensionality of the problem. Sequential and non-sequential approaches have been introduced to estimate the hyperparameters. The former, such as the Kennedy and O'Hagan (KOH) framework, estimates the model error hyperparameters before calibrating the model parameters. The latter, such as the Full Maximum a Posteriori (FMP), introduces a functional dependence between the model parameters and the model error hyperparameters. Despite being more reliable in some cases (bimodality e.g.), the FMP method still fails to estimate correctly the posterior distribution shape. This work proposes a new methodology for treating the model error term in computer code calibration. It builds upon the KOH and FMP framework. Called the Complete Maximum a Posteriori (CMP) method, it provides a closed-form expression for the marginalization integral over the model error hyperparameters, significantly reducing the dimensionality of the calibration problem. Such expression re- lies on a set of assumptions that are more general and less stringent than the ones usually employed. The CMP method is applied to four examples of increasing complexity, from elementary to real fluid dynamics problems, including or not bimodality. Compared to the true reference solution and unlike the KOH and FMP, the CMP method correctly captures the shape of the posterior distribution, including all modes and their weights. Moreover, it provides an accurate estimate of the distribution tails</p></div> (10.1615/Int.J.UncertaintyQuantification.2025056317)
  DOI : 10.1615/Int.J.UncertaintyQuantification.2025056317
• PSWF-Radon approach to reconstruction from band-limited Hankel transform
  
  • Goncharov Fedor
  • Isaev Mikhail
  • Novikov Roman
  • Zaytsev Rodion
  Applied and Computational Harmonic Analysis, Elsevier, 2025. We give new formulas for reconstructions from band-limited Hankel transform of integer or half-integer order. Our formulas rely on the PSWF-Radon approach to super-resolution in multidimensional Fourier analysis. This approach consists of combining the theory of classical one-dimensional prolate spheroidal wave functions with the Radon transform theory. We also use the relation between Fourier and Hankel transforms and Cormack-type inversion of the Radon transform. Finally, we investigate numerically the capabilities of our approach to super-resolution for band-limited Hankel inversion in relation to varying levels of noise.
• Inverse scattering for the multipoint potentials of Bethe-Peierls-Thomas-Fermi type
  
  • Kuo Pei-Cheng
  • Novikov Roman
  Inverse Problems, IOP Publishing, 2025, 41 (6), pp.065021. <div><p>We consider the Schrödinger equation with a multipoint potential of the Bethe-Peierls-Thomas-Fermi type. We show that such a potential in dimension d = 2 or d = 3 is uniquely determined by its scattering amplitude at a fixed positive energy. Moreover, we show that there is no non-zero potential of this type with zero scattering amplitude at a fixed positive energy and a fixed incident direction. Nevertheless, we also show that a multipoint potential of this type is not uniquely determined by its scattering amplitude at a positive energy E and a fixed incident direction. Our proofs also contribute to the theory of inverse source problem for the Helmholtz equation with multipoint source.</p></div> (10.1088/1361-6420/ade282)
  DOI : 10.1088/1361-6420/ade282
• Taking Advantage of Multiple Scattering for Optical Reflection Tomography
  
  • Wasik Thomas
  • Barolle Victor
  • Aubry Alexandre
  • Garnier Josselin
  , 2025. <div><p>Optical Diffraction Tomography (ODT) is a powerful non-invasive imaging technique widely used in biological and medical applications. While significant progress has been made in transmission configuration, reflection ODT remains challenging due to the ill-posed nature of the inverse problem. We present a novel optimization algorithm for 3D refractive index (RI) reconstruction in reflection-mode microscopy. Our method takes advantage of the multiply-scattered waves that are reflected by uncontrolled background structures and that illuminate the foreground RI from behind. It tackles the ill-posed nature of the problem using weighted time loss, positivity constraints and Total Variation regularization. We have validated our method with data generated by detailed 2D and 3D simulations, demonstrating its performance under weak multiple scattering conditions and with simplified forward models used in the optimization routine for computational efficiency.</p><p>In addition, we highlight the need for multi-wavelength analysis and the use of regularization to ensure the reconstruction of the low spatial frequencies of the foreground RI.</p></div>
• Byzantine-Robust Gossip: Insights from a Dual Approach
  
  • Gaucher Renaud
  • Dieuleveut Aymeric
  • Hendrikx Hadrien
  , 2025. Distributed learning has many computational benefits but is vulnerable to attacks from a subset of devices transmitting incorrect information. This paper investigates Byzantine resilient algorithms in a decentralized setting, where devices communicate directly in a peer-to-peer manner within a communication network. We leverage the so-called dual approach for decentralized optimization and propose a Byzantine-robust algorithm. We provide convergence guarantees in the average consensus subcase, discuss the potential of the dual approach beyond this subcase, and re-interpret existing algorithms using the dual framework. Lastly, we experimentally show the soundness of our method.
• Long time behavior of a degenerate stochastic system modeling the response of a population face to environmental impacts
  
  • Collet Pierre
  • Ecotière Claire
  • Méléard Sylvie
  Electronic Communications in Probability, Institute of Mathematical Statistics (IMS), 2025, 30 (none). We study the asymptotics of a two-dimensional stochastic differential system with a degenerate diffusion matrix. This system describes the dynamics of a population where individuals contribute to the degradation of their environment through two different behaviors. We exploit the almost one-dimensional form of the dynamical system to compute explicitly the Freidlin-Wentzell action functional. That allows to give conditions under which the small noise regime of the invariant measure is concentrated around the equilibrium of the dynamical system having the smallest diffusion coefficient. (10.1214/24-ECP650)
  DOI : 10.1214/24-ECP650
• ExceedGAN: Simulation above extreme thresholds using Generative Adversarial Networks
  
  • Allouche Michaël
  • Girard Stéphane
  • Gobet Emmanuel
  Extremes, Springer Verlag (Germany), 2025. This paper devises a novel neural-inspired approach for simulating multivariate extremes. Specifically, we propose a GAN-based generative model for sampling multivariate data exceeding large thresholds, giving rise to what we refer to as the ExceedGAN algorithm. Our approach is based on approximating marginal log-quantile functions using feedforward neural networks with eLU activation functions specifically introduced for bias correction. An error bound is provided {on the margins}, assuming a $J$th order condition from extreme value theory. The numerical experiments illustrate that ExceedGAN outperforms competitors, both on synthetic and real-world data sets.
• Crediting football players for creating dangerous actions in an unbiased way: the generation of threat (GoT) indices
  
  • Baouan Ali
  • Coustou Sebastien
  • Lacome Mathieu
  • Pulido Sergio
  • Rosenbaum Mathieu
  Journal of Quantitative Analysis in Sports, De Gruyter, 2025. We introduce an innovative methodology to identify football players at the origin of threatening actions in a team. In our framework, a threat is defined as entering the opposing team's danger area. We investigate the timing of threat events and ball touches of players, and capture their correlation using Hawkes processes. Our model-based approach allows us to evaluate a player's ability to create danger both directly and through interactions with teammates. We define a new index, called Generation of Threat (GoT), that measures in an unbiased way the contribution of a player to threat generation. For illustration, we present a detailed analysis of Chelsea's 2016-2017 season, with a standout performance from Eden Hazard. We are able to credit each player for his involvement in danger creation and determine the main circuits leading to threat. In the same spirit, we investigate the danger generation process of Stade Rennais in the 2021-2022 season. Furthermore, we establish a comprehensive ranking of Ligue 1 players based on their generated threat in the 2021-2022 season. Our analysis reveals surprising results, with players such as Jason Berthomier, Moses Simon and Frederic Guilbert among the top performers in the GoT rankings. We also present a ranking of Ligue 1 central defenders in terms of generation of threat and confirm the great performance of some center-back pairs, such as Nayef Aguerd and Warmed Omari.
• Partial regularity for optimal transport with $p$-cost away from fixed points
  
  • Goldman Michael
  • Koch Lukas
  Proceedings of the American Mathematical Society, American Mathematical Society, 2025. We consider maps $T$ solving the optimal transport problem with a cost $c(x-y)$ modeled on the $p$-cost. For Hölder continuous marginals, we prove a $C^{1,\alpha}$-partial regularity result for $T$ in the set $\{\lvert T(x)-x\rvert&gt;0\}$.
• Signature volatility models: pricing and hedging with Fourier
  
  • Abi Jaber Eduardo
  • Gérard Louis-Amand
  SIAM Journal on Financial Mathematics, Society for Industrial and Applied Mathematics, 2025, 16 (2). We consider a stochastic volatility model where the dynamics of the volatility are given by a possibly infinite linear combination of the elements of the time extended signature of a Brownian motion. First, we show that the model is remarkably universal, as it includes, but is not limited to, the celebrated Stein-Stein, Bergomi, and Heston models, together with some path-dependent variants. Second, we derive the joint characteristic functional of the log-price and integrated variance provided that some infinitedimensional extended tensor algebra valued Riccati equation admits a solution. This allows us to price and (quadratically) hedge certain European and path-dependent options using Fourier inversion techniques. We highlight the efficiency and accuracy of these Fourier techniques in a comprehensive numerical study. (10.1137/24M1636952)
  DOI : 10.1137/24M1636952
• Objective assessment of cardiac function using patient-specific biophysical modeling based on cardiovascular MRI combined with catheterization
  
  • Gusseva Maria
  • Castellanos Daniel Alexander
  • Veeram Reddy Surendranath
  • Hussain Tarique
  • Chapelle Dominique
  • Chabiniok Radomír
  AJP - Heart and Circulatory Physiology, American Physiological Society, 2025, 329 (5), pp.H118-H1191. Synthesizing multi-modality data, such as cardiovascular magnetic resonance imaging (MRI) combined with catheterization, into a single framework is challenging. Different acquisition systems are subjected to different measurement errors. Coupling clinical data with biomechanical models can assist in clinical data processing (e.g., model-based filtering of measurement noise) and quantify myocardial mechanics via metrics not readily available in the data, such as myocardial contractility. In this work we use a biomechanical modeling with the aim 1) to quantitatively compare model- and data-derived signals, and 2) to explore the potential of model-derived myocardial contractility and distal resistance of the circulation (Rd) to robustly quantify cardiovascular physiology. We used 51 ventricular catheterization pressure and cine MRI volume datasets from patients with single-ventricle physiology and left and right ventricles of patients with repaired tetralogy of Fallot. Ventricular time-varying elastance (TVE) metrics and linear regression were used to quantify the relationship between the maximum value of TVE (Emax) and maximum time derivative of ventricular pressure (max(dP/dt)) in data- and model-derived pressure and volume signals at p&lt;0.05. Pearson’s correlations were used to compare model-derived contractility and data-derived Emax and max(dP/dt), and model-derived Rd and data-derived vascular resistance. All data and model-derived linear regressions were significant (p&lt;0.05). Model-derived max(dP/dt) vs. data-derived Emax produced higher R2 than data-derived max(dP/dt) vs. data-derived Emax. Correlations demonstrated significant relationships between most data- and model-derived metrics. This work revealed the clinical value of biomechanical modeling to assist in clinical data processing by providing high-quality pressure and volume signals, and to quantify cardiovascular pathophysiology. (10.1152/ajpheart.00232.2025)
  DOI : 10.1152/ajpheart.00232.2025
• Learning homogenized hyperelastic behavior for topology optimization of lattice structures
  
  • Ribeiro Nogueira Breno
  • Allaire Grégoire
  , 2025. (10.5281/zenodo.14900138)
  DOI : 10.5281/zenodo.14900138
• Finite elements for Wasserstein $W_p$ gradient flows
  
  • Cancès Clément
  • Matthes Daniel
  • Nabet Flore
  • Rott Eva-Maria
  ESAIM: Mathematical Modelling and Numerical Analysis, Société de Mathématiques Appliquées et Industrielles (SMAI) / EDP, 2025, 59 (3), pp.1565-1600. Wasserstein $\bbW_p$ gradient flows for nonlinear integral functionals of the density yield degenerate parabolic equations involving diffusion operators of $q$-Laplacian type, with $q$ being $p$'s conjugate exponent. We propose a finite element scheme building on conformal $\mathbb{P}_1$ Lagrange elements with mass lumping and a backward Euler time discretization strategy. Our scheme preserves mass and positivity while energy decays in time. Building on the theory of gradient flows in metric spaces, we further prove convergence towards a weak solution of the PDE that satisfies the energy dissipation equality. The analytical results are illustrated by numerical simulations. (10.1051/m2an/2025035)
  DOI : 10.1051/m2an/2025035
• Volume growth of Funk geometry and the flags of polytopes
  
  • Faifman Dmitry
  • Vernicos Constantin
  • Walsh Cormac
  Geometry and Topology, Mathematical Sciences Publishers, 2025, 29 (7), pp.3773-3811. We consider the Holmes-Thompson volume of balls in the Funk geometry on the interior of a convex domain. We conjecture that for a fixed radius, this volume is minimized when the domain is a simplex and the ball is centered at the barycenter, or in the centrally-symmetric case, when the domain is a Hanner polytope. This interpolates between Mahler's conjecture and Kalai's flag conjecture. We verify this conjecture for unconditional domains. For polytopal Funk geometries, we study the asymptotics of the volume of balls of large radius, and compute the two highest-order terms. The highest depends only on the combinatorics, namely on the number of flags. The second highest depends also on the geometry, and thus serves as a geometric analogue of the centro-affine area for polytopes. We then show that for any polytope, the second highest coefficient is minimized by a unique choice of center point, extending the notion of Santaló point. Finally, we show that, in dimension two, this coefficient, with respect to the minimal center point, is uniquely maximized by affine images of the regular polygon. (10.2140/gt.2025.29.3773)
  DOI : 10.2140/gt.2025.29.3773
• A holographic uniqueness theorem for the two-dimensional Helmholtz equation
  
  • Nair Arjun
  • Novikov Roman
  The Journal of Geometric Analysis, Springer, 2025, 35 (4), pp.123. We consider a plane wave, a radiation solution, and the sum of these solutions (total solution) for the Helmholtz equation in an exterior region in $\mathbb R^2$. We consider a straight line in this region, such that the direction of propagation of the plane wave is not parallel to this line. We show that the radiation solution in the exterior region is uniquely determined by the intensity of the total solution on an interval of this line. In particular, this result solves one of the old mathematical questions of holography in its two-dimensional setting. Our proofs also contribute to the theory of the Karp expansion of radiation solutions in two dimensions. (10.1007/s12220-025-01949-x)
  DOI : 10.1007/s12220-025-01949-x