Searcharxiv⌕ Search

arXiv subjects

Pierre Del Moral

Publications and source records attributed to Pierre Del Moral.

At least 19 recordsLinked to original sources

Multilevel Filtering on Manifolds with Light Schrödinger Bridges

In this article we consider the filtering of partially observed diffusions when the observations, made at discrete times, are noiseless. Each observation then confines the hidden state to a manifold defined by the data, whose type, for instance a sphere or an affine subspace, may change from one observation time to the next. In practice, the filter is approximated by particle filters, which require the diffusion to be discretized in time. Multilevel Monte Carlo reduces the cost of this discretization only if the particle approximations at consecutive discretization levels stay close, which standard particle filters do not ensure. We develop a new particle filter based on light Schrödinger bridges, which works between changing manifolds and can be coupled across discretization levels. A key insight is that the previous state enters the light Schrödinger bridge only through a few features, so that the bridge only has to be fitted on the new manifold, regardless of the type of the previous one. We bound the bias of this particle filter and the gap between its approximations at consecutive levels, and show numerically, on spheres and on manifolds that change type, that this gap is much smaller than with standard multilevel particle filters, and that, with the cost of the fits included, a hybrid estimator that uses it at the fine levels reduces the cost at small tolerances.

stat.CO↗

A perturbation theory for Feynman-Kac models and their particle interpretations

We develop a perturbation analysis of Feynman--Kac models, describing how local errors introduced at each step of the Feynman--Kac recursion propagate over time. Under general conditions on the potential functions and the Markov transitions, which are shown to hold for a large class of state-space models, we prove that the global error for any time horizon $T\geq 1$ is bounded by $C\,Δ_T\log(1+Δ_T^{-1})$, where $Δ_T$ is the largest local error; this bound holds both for the prediction and the updated models. These conditions require no Lyapunov function and no stability of the Markov transitions, and allow for non-compact state spaces: confinement is provided by the potential functions. Our perturbation bounds hold in total variation and in the Kolmogorov norm; for the latter, we extend the Hardy--Krause variation and the Koksma--Hlawka inequality to unbounded domains. Our perturbation theory is applied to derive time-uniform bounds for linearized Kalman filters, Euler schemes for partially observed diffusions and nudged state-space models, as well as high-probability error bounds, uniform over whole trajectories, for particle filters and, more generally, for mean field particle methods such as sequential Monte Carlo and diffusion Monte Carlo algorithms.

stat.CO↗

Sequential Entropic Optimal Transport for Bayesian Inverse Problems via Multilevel Monte Carlo

We consider the problem of Bayesian estimation associated to inverse problems governed by partial differential equations (PDEs). In most cases of practical interest, one must use numerical solvers to implement computational methods based on Monte Carlo methods. This inevitably leads to an adaptation of multilevel Monte Carlo (MLMC) methods to reduce computational efforts to achieve a pre-specified mean square error (MSE). MLMC is based on sampling coupled versions of Bayesian posteriors associated to increasingly precise numerical solves of the PDE. In this article we present a novel method to implement MLMC based upon sequential entropic optimal transport couplings learned with light Schrödinger bridges, and correct the transport with Metropolis--Hastings steps that preserve the pairing. We bound the mean square error associated to a version of our new estimation method and show, for an elliptic inverse problem, that to achieve a MSE of $\mathcal{O}(δ^2)$ ($δ>0$) one can have an optimal cost of $\mathcal{O}(δ^{-2})$. The results feature explicit bounds on the effect of the fitted transport and of the Metropolis--Hastings correction. From a numerical perspective, on elliptic, Darcy-flow and Navier-Stokes problems, the method reduces the variance of the most precise solver by a factor of five to ten and is faster than multilevel sequential Monte Carlo [arXiv:1503.07259] at tight accuracies.

math.NA↗

On the Kantorovich contraction of Markov semigroups

We develop an operator-theoretic framework for the contraction properties of Markov semigroups with respect to a general class of Kantorovich semi-distances, which includes Wasserstein distances and weighted total variation norms. The framework is based on Kantorovich contraction coefficients, which extend Dobrushin's ergodic coefficient from the total variation distance to general transport costs. Our main result shows that a geometric drift condition, combined with a local contraction condition on the sub-level sets of the Lyapunov function, implies that the Markov operator is a strict contraction with respect to an explicit modification of the underlying Kantorovich cost. This yields, in a unified way, the existence and uniqueness of invariant measures and exponential decay estimates in Wasserstein distances, for both discrete time and continuous time Markov semigroups, as well as for time-varying semigroups. We illustrate the scope of the framework with (i) Markov transitions on models with boundary states, including bounded domains with entrance boundaries, (ii) operator products of a Markov kernel and its adjoint, including two-block Gibbs samplers, (iii) iterated random functions, including Euler-type schemes, (iv) diffusion models, including overdamped Langevin diffusions with potentials that are convex at infinity, jump-diffusions on compact manifolds, and Jacobi and Riccati matrix diffusions, and (v) time-varying and nonlinear semigroups, including Feynman-Kac distribution flows and Sinkhorn bridges. In several of these settings we obtain new contraction estimates, while in others we recover known results through short and elementary arguments.

math.PR↗

On Bridging Mixture Distributions

The construction and the stability of bridges between mixture probability measures are investigated. Given a collection of Markov transitions bridging the components of two mixtures, together with a coupling of the mixture labels, we construct a bridge between the mixtures on an extended state space, and we show that this bridge inherits the entropic optimality of the component bridges. We then turn to Gaussian mixtures bridged by Gaussian Schrödinger bridges, whose parameters are generally unknown and must be estimated. We prove a $2$-Wasserstein continuity theorem between the exact bridge and its plug-in approximation. The analysis rests on a representation of the Riccati fixed point map in inverse coordinates; this map is globally $1$-Lipschitz and requires no matrix inversion. These results are applied to three estimation schemes, with $d$ the dimension and $N$ the number of samples. For Gaussian mixtures estimated by the Expectation-Maximization algorithm, the squared $2$-Wasserstein error is of order $(d\log N/N)^{1/2}+d^2\log N/N$ with probability at least $1-10N^{-1}$. For single Gaussian marginals estimated by their sample moments, it is of order $d^2/N$ in expectation. For the regularized empirical Monge map based on an $ε$-inflation of the sample covariance, the mean squared error is of order $d^2\{(1+ε^{-1})/N+ε^2\}$; as soon as $N$ is a sufficiently large multiple of $d$, it reduces to $d^2/N+d\,ε^2$, up to an exponentially small term. These estimates are illustrated numerically.

math.ST↗

Fluctuations and Long-Time Stability of Multivariate Ensemble Kalman Filters

We develop a self contained stochastic perturbation theory for discrete generation and multivariate Ensemble Kalman filters. Unlike their continuous-time counterparts, discrete EnKF algorithms are defined through a two steps prediction update mechanism and exhibit non Gaussian fluctuations, even in linear settings. In the multivariate case, these fluctuations take the form of non central Wishart type perturbations, which significantly complicate the mathematical analysis. We establish non asymptotic, time-uniform stability and error estimates for the ensemble covariance matrix processes under minimal structural assumptions on the signal observation model, allowing for possibly unstable dynamics. Our results quantify the impact of ensemble size, dimension, and observation noise, and provide explicit bounds on the propagation of stochastic errors over long time horizons. The analysis relies on a detailed study of stochastic Riccati difference equations driven by matrix-valued noncentral Wishart fluctuations. Beyond their relevance to data assimilation, these results contribute to the probabilistic understanding of ensemble-based filtering methods in high dimension and offer new tools for the analysis of interacting particle systems with matrix-valued dynamics.

math.PR↗

New Trends in the Stability of Sinkhorn Semigroups

Entropic optimal transport problems play an increasingly important role in machine learning and generative modelling. In contrast with optimal transport maps which often have limited applicability in high dimensions, Schrodinger bridges can be solved using the celebrated Sinkhorn's algorithm, a.k.a. the iterative proportional fitting procedure. The stability properties of Sinkhorn bridges when the number of iterations tends to infinity is a very active research area in applied probability and machine learning. Traditional proofs of convergence are mainly based on nonlinear versions of Perron-Frobenius theory and related Hilbert projective metric techniques, gradient descent, Bregman divergence techniques and Hamilton-Jacobi-Bellman equations, including propagation of convexity profiles based on coupling diffusions by reflection methods. The objective of this review article is to present, in a self-contained manner, recently developed Sinkhorn/Gibbs-type semigroup analysis based upon contraction coefficients and Lyapunov-type operator-theoretic techniques. These powerful, off-the-shelf semigroup methods are based upon transportation cost inequalities (e.g. log-Sobolev, Talagrand quadratic inequality, curvature estimates), $ϕ$-divergences, Kantorovich-type criteria and Dobrushin contraction-type coefficients on weighted Banach spaces as well as Wasserstein distances. This novel semigroup analysis allows one to unify and simplify many arguments in the stability of Sinkhorn algorithm. It also yields new contraction estimates w.r.t. generalized $ϕ$-entropies, as well as weighted total variation norms, Kantorovich criteria and Wasserstein distances.

math.PR↗

A Contraction Theory for Sinkhorn and Schrodinger Bridges via Log-Sobolev Inequalities

We develop a quantitative contraction framework for Schrodinger and Sinkhorn bridges based on transportation-cost inequalities and Riccati matrix difference equations. Our approach combines logarithmic Sobolev and Talagrand-type inequalities to obtain explicit entropy and Wasserstein contraction bounds for Sinkhorn bridge measures, entropic optimal transport plans, and the associated Markov transport maps. A key feature of the analysis is the interplay between transport-cost inequalities and matrix Riccati difference equations arising in filtering and stochastic control. The results are established under local regularity assumptions on the reference transition, formulated in terms of curvature, Lipschitz continuity, and Fisher-information bounds. Within this general setting, we derive quantitative stability and convergence estimates for Schrodinger bridges and Sinkhorn iterates that are robust with respect to the choice of reference measure. As a main application, we specialize the theory to linear-Gaussian reference transitions, where the Gaussian structure permits sharp constants, refined exponential decay rates, and continuity estimates for Schrodinger bridges, Sinkhorn iterates, barycentric projections, conditional covariances, and proximal sampler semigroups. In this setting, we recover and extend several known contraction results for entropic and Wasserstein distances, and obtain new quantitative bounds that improve previously available rates. Our results provide a unified probabilistic framework for stability, regularity, and convergence of Sinkhorn algorithms.

math.OC↗

Gaussian entropic optimal transport: Schrödinger bridges and the Sinkhorn algorithm

Entropic optimal transport problems are regularized versions of optimal transport problems. These models play an increasingly important role in machine learning and generative modelling. For finite spaces, these problems are commonly solved using Sinkhorn algorithm (a.k.a. iterative proportional fitting procedure). However, in more general settings the Sinkhorn iterations are based on nonlinear conditional/conjugate transformations and exact finite-dimensional solutions cannot be computed. This article presents a finite-dimensional recursive formulation of the iterative proportional fitting procedure for general Gaussian multivariate models. As expected, this recursive formulation is closely related to the celebrated Kalman filter and related Riccati matrix difference equations, and it yields algorithms that can be implemented in practical settings without further approximations. We extend this filtering methodology to develop a refined and self-contained convergence analysis of Gaussian Sinkhorn algorithms, including closed form expressions of entropic transport maps and Schrödinger bridges.

stat.ML↗

Entropic continuity bounds for conditional covariances with applications to Schr\" odinger and Sinkhorn bridges

The article presents new entropic continuity bounds for conditional expectations and conditional covariance matrices. These bounds are expressed in terms of the relative entropy between different coupling distributions. Our approach combines Wasserstein coupling with quadratic transportation cost inequalities. We illustrate the impact of these results in the context of entropic optimal transport problems. The entropic continuity theorem presented in the article allows to estimate the conditional expectations and the conditional covariances of Schr\" odinger and Sinkhorn transitions in terms of the relative entropy between the corresponding bridges. These entropic continuity bounds turns out to be a very useful tool for obtaining remarkably simple proofs of the exponential decays of the gradient and the Hessian of Schrödinger and Sinkhorn bridge potentials.

math.PR↗

On the Particle Approximation of Lagged Feynman-Kac Formulae

In this paper we examine the numerical approximation of the limiting invariant measure associated with Feynman-Kac formulae. These are expressed in a discrete time formulation and are associated with a Markov chain and a potential function. The typical application considered here is the computation of eigenvalues associated with non-negative operators as found, for example, in physics or particle simulation of rare-events. We focus on a novel \emph{lagged} approximation of this invariant measure, based upon the introduction of a ratio of time-averaged Feynman-Kac marginals associated with a positive operator iterated $l \in\mathbb{N}$ times; a lagged Feynman-Kac formula. This estimator and its approximation using Diffusion Monte Carlo (DMC) have been extensively employed in the physics literature. In short, DMC is an iterative algorithm involving $N\in\mathbb{N}$ particles or walkers simulated in parallel, that undergo sampling and resampling operations. In this work, it is shown that for the DMC approximation of the lagged Feynman-Kac formula, one has an almost sure characterization of the $\mathbb{L}_1$-error as the time parameter (iteration) goes to infinity and this is at most of $\mathcal{O}(\exp\{-κl\}/N)$, for $κ>0$. In addition a non-asymptotic in time, and time uniform $\mathbb{L}_1-$bound is proved which is $\mathcal{O}(l/\sqrt{N})$. We also prove a novel central limit theorem to give a characterization of the exact asymptotic in time variance. This analysis demonstrates that the strategy used in physics, namely, to run DMC with $N$ and $l$ small and, for long time enough, is mathematically justified. Our results also suggest how one should choose $N$ and $l$ in practice. We emphasize that these results are not restricted to physical applications; they have broad relevance to the general problem of particle simulation of the Feynman-Kac formula.

math.PR↗

On Time Uniform Wong-Zakai Approximation Theorems

We consider the long time behavior of Wong-Zakai approximations of stochastic differential equations. These piecewise smooth diffusion approximations are of great importance in many areas, such as those with ordinary differential equations associated to random smooth fluctuations; e.g. robust filtering problems. In many examples, the mean error estimate bounds that have been derived in the literature can grow exponentially with respect to the time horizon. We show in a simple example that indeed mean error estimates do explode exponentially in the time parameter, i.e. in that case a Wong-Zakai approximation is only useful for extremely short time intervals. Under spectral conditions, we present some quantitative time-uniform convergence theorems, i.e. time-uniform mean error bounds, yielding what seems to be the first results of this type for Wong-Zakai diffusion approximations.

math.PR↗

Bayesian Parameter Inference for Partially Observed Diffusions using Multilevel Stochastic Runge-Kutta Methods

We consider the problem of Bayesian estimation of static parameters associated to a partially and discretely observed diffusion process. We assume that the exact transition dynamics of the diffusion process are unavailable, even up-to an unbiased estimator and that one must time-discretize the diffusion process. In such scenarios it has been shown how one can introduce the multilevel Monte Carlo method to reduce the cost to compute posterior expected values of the parameters for a pre-specified mean square error (MSE). These afore-mentioned methods rely on upon the Euler-Maruyama discretization scheme which is well-known in numerical analysis to have slow convergence properties. We adapt stochastic Runge-Kutta (SRK) methods for Bayesian parameter estimation of static parameters for diffusions. This can be implemented in high-dimensions of the diffusion and seemingly under-appreciated in the uncertainty quantification and statistics fields. For a class of diffusions and SRK methods, we consider the estimation of the posterior expectation of the parameters. We prove that to achieve a MSE of $\mathcal{O}(ε^2)$, for $ε>0$ given, the associated work is $\mathcal{O}(ε^{-2})$. Whilst the latter is achievable for the Milstein scheme, this method is often not applicable for diffusions in dimension larger than two. We also illustrate our methodology in several numerical examples.

stat.CO↗

On the Mathematical Theory of Ensemble (Linear-Gaussian) Kalman-Bucy Filtering

The purpose of this review is to present a comprehensive overview of the theory of ensemble Kalman-Bucy filtering for continuous-time, linear-Gaussian signal and observation models. We present a system of equations that describe the flow of individual particles and the flow of the sample covariance and the sample mean in continuous-time ensemble filtering. We consider these equations and their characteristics in a number of popular ensemble Kalman filtering variants. Given these equations, we study their asymptotic convergence to the optimal Bayesian filter. We also study in detail some non-asymptotic time-uniform fluctuation, stability, and contraction results on the sample covariance and sample mean (or sample error track). We focus on testable signal/observation model conditions, and we accommodate fully unstable (latent) signal models. We discuss the relevance and importance of these results in characterising the filter's behaviour, e.g. it's signal tracking performance, and we contrast these results with those in classical studies of stability in Kalman-Bucy filtering.We also provide a novel (and negative) result proving that the bootstrap particle filter cannot track even the most basic unstable latent signal, in contrast with the ensemble Kalman filter (and the optimal filter). We provide intuition for how the main results extend to nonlinear signal models and comment on their consequence on some typical filter behaviours seen in practice, e.g. catastrophic divergence.

math.ST↗

A Lyapunov approach to stability of positive semigroups: An overview with illustrations

The stability analysis of possibly time varying positive semigroups on non necessarily compact state spaces, including Neumann and Dirichlet boundary conditions is a notoriously difficult subject. These crucial questions arise in a variety of areas of applied mathematics, including nonlinear filtering, rare event analysis, branching processes, physics and molecular chemistry. This article presents an overview of some recent Lyapunov-based approaches, focusing principally on practical and powerful tools for designing Lyapunov functions. These techniques include semigroup comparisons as well as conjugacy principles on non necessarily bounded manifolds with locally Lipschitz boundaries. All the Lyapunov methodologies discussed in the article are illustrated in a variety of situations, ranging from conventional Markov semigroups on general state spaces to more sophisticated conditional stochastic processes possibly restricted to some non necessarily bounded domains, including locally Lipschitz and smooth hypersurface boundaries, Langevin diffusions as well as coupled harmonic oscillators.

math.PR↗

On the exit-problem for self-interacting diffusions

We study the exit-time from a domain of a self-interacting diffusion, where the Brownian motion is replaced by $σB_t$ for a constant $σ$. The first part of this work consists in showing that the rate of convergence (of the occupation measure of the self-interacting process toward some explicit Gibbs measure) previously obtained in \cite{kk-ejp} for a convex confinment potential $V$ and a convex interaction potential can be bounded uniformly with respect to $σ$. Then, we prove an Arrhenius-type law for the first exit-time from a domain (satisfying classical hypotheses of Freidlin-Wentzell theory).

math.PR↗

Log-Normalization Constant Estimation using the Ensemble Kalman-Bucy Filter with Application to High-Dimensional Models

In this article we consider the estimation of the log-normalization constant associated to a class of continuous-time filtering models. In particular, we consider ensemble Kalman-Bucy filter based estimates based upon several nonlinear Kalman-Bucy diffusions. Based upon new conditional bias results for the mean of the afore-mentioned methods, we analyze the empirical log-scale normalization constants in terms of their $\mathbb{L}_n-$errors and conditional bias. Depending on the type of nonlinear Kalman-Bucy diffusion, we show that these are of order $(\sqrt{t/N}) + t/N$ or $1/\sqrt{N}$ ($\mathbb{L}_n-$errors) and of order $[t+\sqrt{t}]/N$ or $1/N$ (conditional bias), where $t$ is the time horizon and $N$ is the ensemble size. Finally, we use these results for online static parameter estimation for above filtering models and implement the methodology for both linear and nonlinear models.

stat.CO↗

Stochastic Epidemic Models inference and diagnosis with Poisson Random Measure Data Augmentation

We present a new Bayesian inference method for compartmental models that takes into account the intrinsic stochasticity of the process. We show how to formulate a SIR-type Markov jump process as the solution of a stochastic differential equation with respect to a Poisson Random Measure (PRM), and how to simulate the process trajectory deterministically from a parameter value and a PRM realisation. This forms the basis of our Data Augmented MCMC, which consists in augmenting parameter space with the unobserved PRM value. The resulting simple Metropolis-Hastings sampler acts as an efficient simulation-based inference method, that can easily be transferred from model to model. Compared with a recent Data Augmentation method based on Gibbs sampling of individual infection histories, PRM-augmented MCMC scales much better with epidemic size and is far more flexible. PRM-augmented MCMC also yields a posteriori estimates of the PRM, that represent process stochasticity, and which can be used to validate the model. If the model is good, the posterior distribution should exhibit no pattern and be close to the PRM prior distribution. We illustrate this by fitting a non-seasonal model to some simulated seasonal case count data. Applied to the Zika epidemic of 2013 in French Polynesia, our approach shows that a simple SEIR model cannot correctly reproduce both the initial sharp increase in the number of cases as well as the final proportion of seropositive. PRM-augmentation thus provides a coherent story for Stochastic Epidemic Model inference, where explicitly inferring process stochasticity helps with model validation.

stat.ME↗