L'objectif de ce travail est d'étudier le comportement en temps long d'un modèle de particules avec une interaction de type branchement. Plus précisément, les particules se déplacent indépendamment suivant une dynamique markovienne jusqu'au temps de branchement, où elles donnent naissance à de nouvelles particules dont la position dépend de celle de leur mère et de son nombre d'enfants. Dans la première partie de ce mémoire nous omettons le branchement et nous étudions le comportement d'une seule lignée. Celle-ci est modélisée via un processus de Markov qui peut admettre des sauts, des parties diffusives ou déterministes par morceaux. Nous quantifions la convergence de ce processus hybride à l'aide de la courbure de Wasserstein, aussi nommée courbure grossière de Ricci. Cette notion de courbure, introduite récemment par Joulin, Ollivier, et Sammer correspond mieux à l'étude des processus avec sauts. Nous établissons une expression du gradient du semigroupe des processus de Markov stochastiquement monotone, qui nous permet d'expliciter facilement leur courbure. D'autres bornes fines de convergence en distance de Wasserstein et en variation totale sont aussi établies. Dans le même contexte, nous démontrons qu'un processus de Markov, qui change de dynamique suivant un processus discret, converge rapidement vers un équilibre, lorsque la moyenne des courbures des dynamiques sous-jacentes est strictement positive. Dans la deuxième partie de ce mémoire, nous étudions le comportement de toute la population de particules. Celui-ci se déduit du comportement d'une seule lignée grâce à une formule many-to-one, c'est-à-dire un changement de mesure de type Girsanov. Via cette transformation, nous démontrons une loi des grands nombres et établissons une limite macroscopique, pour comparer nos résultats aux résultats déjà connus en théorie des équations aux dérivées partielles. Nos résultats sont appliqués sur divers modèles ayant des applications en biologie et en informatique. Parmi ces modèles, nous étudierons le comportement en temps long de la plus grande particule dans un modèle simple de population structurée en taille / The aim of this work is to study the long time behavior of a branching particle model. More precisely, the particles move independently from each other following a Markov dynamics until the branching event. When one of these events occurs, the particle produces some random number of individuals whose position depends on the position of its mother and her number of offspring. In the first part of this thesis, we only study one particle line and we ignore the branching mechanism. So we are interested by the study of a Markov process which can jump, diffuse or be piecewise deterministic. The long time behavior of these hybrid processes is described with the notion of Wasserstein or coarse Ricci curvature. This notion of curvature, introduced by Joulin, Ollivier and Sammer, is more appropriate for the study of processes with jumps. We establish an expression of the gradient of the Markov semigroup of stochastically monotone processes which gives the curvature of these processes. Others sharp bounds of convergence, in Wasserstein distance and total variation distance, are also established. In the same way, we prove that if a Markov process evolves according to one of finitely many underlying Markovian dynamics, with a choice of dynamics that changes at the jump times of a second Markov process, then it is exponentially ergodic, under the assumption that the mean of the curvature of the underlying dynamics is positive. In the second part of the work, we study all the population. Its behaviour can be deduced to the study of the first part using a Girsavov-type transform which is called a many-to-one formula. Using this relation, we establish a law of large numbers and a macroscopic limit, in order to compare our results to the well know results on deterministic setting. Several examples, based on biology and computer science problems, illustrate our results, including the study of the largest individual in a size-structured population model

In the modern era of high and infinite dimensional data, classical statistical methodology is often rendered inefficient and ineffective when confronted with such big data problems as arise in genomics, medical imaging, speech analysis, and many other areas of research. Many problems manifest when the practitioner is required to take into account the covariance structure of the data during his or her analysis, which takes on the form of either a high dimensional low rank matrix or a finite dimensional representation of an infinite dimensional operator acting on some underlying function space. Thus, novel methodology is required to estimate, analyze, and make inferences concerning such covariances. In this manuscript, we propose using tools from the concentration of measure literature–a theory that arose in the latter half of the 20th century from connections between geometry, probability, and functional analysis–to construct rigorous descriptive and inferential statistical methodology for covariance matrices and operators. A variety of concentration inequalities are considered, which allow for the construction of nonasymptotic dimension-free confidence sets for the unknown matrices and operators. Given such confidence sets a wide range of estimation and inferential procedures can be and are subsequently developed. For high dimensional data, we propose a method to search a concentration in- equality based confidence set using a binary search algorithm for the estimation of large sparse covariance matrices. Both sub-Gaussian and sub-exponential concentration inequalities are considered and applied to both simulated data and to a set of gene expression data from a study of small round blue-cell tumours. For infinite dimensional data, which is also referred to as functional data, we use a celebrated result, Talagrand’s concentration inequality, in the Banach space setting to construct confidence sets for covariance operators. From these confidence sets, three different inferential techniques emerge: the first is a k-sample test for equality of covariance operator; the second is a functional data classifier, which makes its decisions based on the covariance structure of the data; the third is a functional data clustering algorithm, which incorporates the concentration inequality based confidence sets into the framework of an expectation-maximization algorithm. These techniques are applied to simulated data and to speech samples from a set of spoken phoneme data. Lastly, we take a closer look at a key tool used in the construction of concentration based confidence sets: Rademacher symmetrization. The symmetrization inequality, which arises in the probability in Banach spaces literature, is shown to be connected with optimal transport theory and specifically the Wasserstein distance. This insight is used to improve the symmetrization inequality resulting in tighter concentration bounds to be used in the construction of nonasymptotic confidence sets. A variety of other applications are considered including tests for data symmetry and tightening inequalities in Banach spaces. An R package for inference on covariance operators is briefly discussed in an appendix chapter.

Pour prédire l'évolution d'un système physique, nous avons besoin d'initialiser le modèle mathématique le représentant, donc d'estimer la valeur de l'état du système au temps initial. Cet état n'est généralement pas directement mesurable car souvent trop complexe. En revanche, nous disposons d'informations du système, prises à des temps différents, incomplètes, mais aussi entachées d'erreurs, telles des observations, de précédentes estimations, etc. Combiner ces différentes informations partielles et imparfaites pour estimer la valeur de l'état fait appel à des méthodes d'assimilation de données dont l'idée est de trouver un état initial proche de toutes les informations. Ces méthodes sont très utilisées en météorologie. Nous nous intéressons dans cette thèse à l'assimilation d'images, images qui sont de plus en plus utilisées en tant qu'observations. La spécificité de ces images est leur cohérence spatiale, l'oeil humain peut en effet percevoir des structures dans les images que les méthodes classiques d'assimilation ne considèrent généralement pas. Elles ne tiennent compte que des valeurs de chaque pixel, ce qui résulte dans certains cas à des problèmes d'amplitude dans l'état initial estimé. Pour résoudre ce problème, nous proposons de changer d'espace de représentation des données : nous plaçons les données dans un espace de Wasserstein où la position des différentes structures compte. Cet espace, équipé d'une distance de Wasserstein, est issue de la théorie du transport optimal et trouve beaucoup d'applications en imagerie notamment.Dans ce travail nous proposons une méthode d'assimilation variationnelle de données basée sur cette distance de Wasserstein. Nous la présentons ici, ainsi que les algorithmes numériques liés et des expériences montrant ses spécificités. Nous verrons dans les résultats comment elle permet de corriger ce qu'on appelle erreurs de position. / Forecasting of a physical system is computed by the help of a mathematical model. This model needs to be initialized by the state of the system at initial time. But this state is not directly measurable and data assimilation techniques are generally used to estimate it. They combine all sources of information such as observations (that may be sparse in time and space and potentially include errors), previous forecasts, the model equations and error statistics. The main idea of data assimilation techniques is to find an initial state accounting for the different sources of informations. Such techniques are widely used in meteorology, where data and particularly images are more and more numerous due to the increasing number of satellites and other sources of measurements. This, coupled with developments of meteorological models, have led to an ever-increasing quality of the forecast.Spatial consistency is one specificity of images. For example, human eyes are able to notice structures in an image. However, classical methods of data assimilation do not handle such structures because they take only into account the values of each pixel separately. In some cases it leads to a bad initial condition. To tackle this problem, we proposed to change the representation of an image: images are considered here as elements of the Wasserstein space endowed with the Wasserstein distance coming from the optimal transport theory. In this space, what matters is the positions of the different structures.This thesis presents a data assimilation technique based on this Wasserstein distance. This technique and its numerical procedure are first described, then experiments are carried out and results shown. In particularly, it appears that this technique was able to give an analysis of corrected position.

Dans cette thèse, on étudie des problèmes de propagation du chaos et de limite de champ moyen pour des modèles relatant le comportement collectif d'individus ou de particules. Particulièrement, on se place dans des cas où l'interaction entre ces individus/particules est discontinue. Le premier travail établit la propagation du chaos pour l'équation de Vlasov-Poisson-Fokker-Planck 1d. Plus précisément, on montre que la distribution des particules évoluant sur la droite des réels interagissant via la fonction signe, converge vers la solution de l'équation de VPFP 1d, en probabilité par des techniques de type grandes déviations, et en espérance par des techniques de loi des grands nombre. Dans le second travail, on étudie une variante du modèle de Cucker-Smale, où le noyau de communication est l'indicatrice d'un cône dont l'orientation dépend de la vitesse de l'individu. Une estimation de stabilité fort-faible en distance de M.K.W. est obtenue, qui implique la limite de champ moyen. Le troisième travail a consisté à introduire de la diffusion en vitesse dans le modèle précédemment cité. Cependant, il faut ajouter une diffusion tronquée afin de préserver un système dans lequel les vitesses restent uniformément bornées. Finalement, on étudie une variante de l'équation d'agrégation où l'interaction entre individus est donnée par un cône dont l'orientation dépend de la position de l'individu. Dans ce cas on peut seulement donner une estimation de stabilité fort-faible en distance $W_\infty$, et le modèle doit être posé dans un domaine borné dans le cas avec diffusion. / In this thesis, we study some propagation of chaos and mean field limit problems arising in modelisation of collective behavior of individuals or particles. Particularly, we set ourselves in the case where the interaction between the individuals/particles is discontinuous. The first work establihes the propagation of chaos for the 1d Vlasov-Poisson-Fokker-Planck equation. More precisely, we show that the distribution of particles evolving on the real line and interacting through the sign function converges to the solution of the 1d VPFP equation, in probability by large deviations-like techniques, and in expectation by law of large numbers-like techniques. In the second work, we study a variant of the Cucker-Smale, where the communication weight is the indicatrix function of a cone which orientation depends on the velocity of the individual. Some weak-strong stability estimate in M.K.W. distance is obtained for the limit equation, which implies the mean field limit. The third work consists in adding some diffusion in velocity to the model previously quoted. However one must add some truncated diffusion in order to preserve a system in which velocities remain unifomrly bounded. Finally we study a variant of the aggregation equation where the interaction between individuals is also given by a cone which orientation depends on the position of the individual. In this case we are only able to provide some weak-strong stability estimate in $W_\infty$ distance, and the problem must be set in a bounded domain for the case with diffusion.

La première partie de cette thèse concerne l'inférence de modèles statistiques non normalisés. Nous étudions deux méthodes d'inférence basées sur de l'échantillonnage aléatoire : Monte-Carlo MLE (Geyer, 1994), et Noise Contrastive Estimation (Gutmann et Hyvarinen, 2010). Cette dernière méthode fut soutenue par une justification numérique d'une meilleure stabilité, mais aucun résultat théorique n'avait encore été prouvé. Nous prouvons que Noise Contrastive Estimation est plus robuste au choix de la distribution d'échantillonnage. Nous évaluons le gain de précision en fonction du budget computationnel. La deuxième partie de cette thèse concerne l'échantillonnage aléatoire approché pour les distributions de grande dimension. La performance de la plupart des méthodes d’échantillonnage se détériore rapidement lorsque la dimension augmente, mais plusieurs méthodes ont prouvé leur efficacité (e.g. Hamiltonian Monte Carlo, Langevin Monte Carlo). Dans la continuité de certains travaux récents (Eberle et al., 2017 ; Cheng et al., 2018), nous étudions certaines discrétisations d’un processus connu sous le nom de kinetic Langevin diffusion. Nous établissons des vitesses de convergence explicites vers la distribution d'échantillonnage, qui ont une dépendance polynomiale en la dimension. Notre travail améliore et étend les résultats de Cheng et al. pour les densités log-concaves. / The first part of this thesis concerns the inference of un-normalized statistical models. We study two methods of inference based on sampling, known as Monte-Carlo MLE (Geyer, 1994), and Noise Contrastive Estimation (Gutmann and Hyvarinen, 2010). The latter method was supported by numerical evidence of improved stability, but no theoretical results had yet been proven. We prove that Noise Contrastive Estimation is more robust to the choice of the sampling distribution. We assess the gain of accuracy depending on the computational budget. The second part of this thesis concerns approximate sampling for high dimensional distributions. The performance of most samplers deteriorates fast when the dimension increases, but several methods have proven their effectiveness (e.g. Hamiltonian Monte Carlo, Langevin Monte Carlo). In the continuity of some recent works (Eberle et al., 2017; Cheng et al., 2018), we study some discretizations of the kinetic Langevin diffusion process and establish explicit rates of convergence towards the sampling distribution, that scales polynomially fast when the dimension increases. Our work improves and extends the results established by Cheng et al. for log-concave densities.

The goal of optimal transport is to map a source probability measure to a destination one with the minimum possible cost. However, the optimal mapping might not be feasible under some practical constraints. One such example is to realize a transport mapping through an information bottleneck. As the optimal mapping may induce infinite mutual information between the source and the destination, the existence of an information bottleneck forces one to resort to some suboptimal mappings. Investigating this type of constrained optimal transport problems is clearly of both theoretical significance and practical interest. In this work, we substantiate a particular form of constrained optimal transport in the context of quantum-to-classical systems by establishing an Output-Constrained Rate-Distortion Theorem similar to the classical case introduced by Yuksel et al. This theorem develops a noiseless communication channel and finds the least required transmission rate R and common randomness Rc to transport a sufficiently large block of n i.i.d. source quantum states, to samples forming a perfectly i.i.d. classical destination distribution, while maintaining the distortion between them. The coding theorem provides operational meanings to the problem of Rate-Limited Optimal Transport, which finds the optimal transportation from source to destination subject to the rate constraints on transmission and common randomness. We further provide an analytical evaluation of the quantum-to-classical rate-limited optimal transportation cost for the case of qubit source state and Bernoulli output distributions with unlimited common randomness. The evaluation results in a transcendental system of equations whose solution provides the rate-distortion curve of the transportation protocol. We further extend this theorem to continuous-variable quantum systems by employing a clipping and quantization argument and using our discrete coding theorem. Moreover, we derive an analytical solution for rate-limited Wasserstein distance of 2nd order for Gaussian quantum systems with Gaussian output distribution. We also provide a Gaussian optimality theorem for the case of unlimited common randomness, showing that Gaussian measurement optimizes the rate in a system with Gaussian source and destination. / Thesis / Doctor of Philosophy (PhD) / We establish a coding theorem for rate-limited quantum-classical optimal transport systems with limited classical common randomness. The coding theorem, referred to as the output-constrained rate-distortion theorem, characterizes the rate region of measurement protocols on a product quantum source state for faithful construction of a given classical destination distribution while maintaining the source-destination distortion below a prescribed threshold with respect to a general distortion observable. This theorem provides a solution to the problem of rate-limited optimal transport, which aims to find the optimal cost of transforming a source quantum state to a destination distribution via a measurement channel with a limited classical communication rate. The coding theorem is further extended to cover Bosonic continuous-variable quantum systems. The analytical evaluation is provided for the case of a qubit measurement system with unlimited common randomness, as well as the case of Gaussian quantum systems.

Das Thema dieser Dissertation ist die Anwendung von Variationsmethoden auf Evolutionsgleichungen parabolischen und hyperbolischen Typs. Im ersten Teil der Arbeit beschäftigen wir uns mit Reaktions-Diffusions-Systemen, die sich als Gradientensysteme schreiben lassen. Hierbei verstehen wir unter einem Gradientensystem ein Tripel bestehend aus einem Zustandsraum, einem Entropiefunktional und einer Dissipationsmetrik. Wir geben Bedingungen an, die die geodätische Konvexität des Entropiefunktionals sichern. Geodätische Konvexität ist eine wertvolle aber auch starke strukturelle Eigenschaft und schwer zu zeigen. Wir zeigen anhand zahlreicher Beispiele, darunter ein Drift-Diffusions-System, dass dennoch interessante Systeme existieren, die diese Eigenschaft besitzen. Einen weiteren Punkt dieser Arbeit stellt die Anwendung von Gamma-Konvergenz auf Gradientensysteme dar. Wir betrachten hierbei zwei Modellsysteme aus dem Bereich der Mehrskalenprobleme: Erstens, die rigorose Herleitung einer Allen-Cahn-Gleichung mit dynamischen Randbedingungen und zweitens, einer Interface-Bedingung für eine eindimensionale Diffusionsgleichung jeweils aus einem reinen Bulk-System. Im zweiten Teil der Arbeit beschäftigen wir uns mit dem sog. Weighted-Inertia-Dissipation-Energy-Prinzip für Evolutionsgleichungen. Hierbei werden Trajektorien eines Systems als (Grenzwerte von) Minimierer(n) einer parametrisierten Familie von Funktionalen charakterisiert. Dies erlaubt es, Werkzeuge aus der Theorie der Variationsrechung auf Evolutionsprobleme anzuwenden. Wir zeigen, dass Minimierer der WIDE-Funktionale gegen Lösungen des Ausgangsproblems konvergieren. Hierbei betrachten wir getrennt voneinander den Fall des beschränkten und des unbeschränkten Zeitintervalls, die jeweils mit verschiedenen Methoden behandelt werden. / This thesis deals with the application of variational methods to evolution problems governed by partial differential equations. The first part of this work is devoted to systems of reaction-diffusion equations that can be formulated as gradient systems with respect to an entropy functional and a dissipation metric. We provide methods for establishing geodesic convexity of the entropy functional by purely differential methods. Geodesic convexity is beneficial, however, it is a strong structural property of a gradient system that is rather difficult to achieve. Several examples, including a drift-diffusion system, provide a survey on the applicability of the theory. Next, we demonstrate the application of Gamma-convergence, to derive effective limit models for multiscale problems. The crucial point in this investigation is that we rely only on the gradient structure of the systems. We consider two model problems: The rigorous derivation of an Allen-Cahn system with bulk/surface coupling and of an interface condition for a one-dimensional diffusion equation. The second part of this thesis is devoted to the so-called Weighted-Inertia-Dissipation-Energy principle. The WIDE principle is a global-in-time variational principle for evolution equations either of conservative or dissipative type. It relies on the minimization of a specific parameter-dependent family of functionals (WIDE functionals) with minimizers characterizing entire trajectories of the system. We prove that minimizers of the WIDE functional converge, up to subsequences, to weak solutions of the limiting PDE when the parameter tends to zero. The interest for this perspective is that of moving the successful machinery of the Calculus of Variations.

<div>The rise of artificial intelligence (AI) hinges on the efficient training of modern deep neural networks (DNNs) for non-convex optimization and uncertainty quantification, which boils down to a non-convex Bayesian learning problem. A standard tool to handle the problem is Langevin Monte Carlo, which proposes to approximate the posterior distribution with theoretical guarantees. However, non-convex Bayesian learning in real big data applications can be arbitrarily slow and often fails to capture the uncertainty or informative modes given a limited time. As a result, advanced techniques are still required.</div><div><br></div><div>In this thesis, we start with the replica exchange Langevin Monte Carlo (also known as parallel tempering), which is a Markov jump process that proposes appropriate swaps between exploration and exploitation to achieve accelerations. However, the na\"ive extension of swaps to big data problems leads to a large bias, and the bias-corrected swaps are required. Such a mechanism leads to few effective swaps and insignificant accelerations. To alleviate this issue, we first propose a control variates method to reduce the variance of noisy energy estimators and show a potential to accelerate the exponential convergence. We also present the population-chain replica exchange and propose a generalized deterministic even-odd scheme to track the non-reversibility and obtain an optimal round trip rate. Further approximations are conducted based on stochastic gradient descents, which yield a user-friendly nature for large-scale uncertainty approximation tasks without much tuning costs. </div><div><br></div><div>In the second part of the thesis, we study scalable dynamic importance sampling algorithms based on stochastic approximation. Traditional dynamic importance sampling algorithms have achieved successes in bioinformatics and statistical physics, however, the lack of scalability has greatly limited their extensions to big data applications. To handle this scalability issue, we resolve the vanishing gradient problem and propose two dynamic importance sampling algorithms based on stochastic gradient Langevin dynamics. Theoretically, we establish the stability condition for the underlying ordinary differential equation (ODE) system and guarantee the asymptotic convergence of the latent variable to the desired fixed point. Interestingly, such a result still holds given non-convex energy landscapes. In addition, we also propose a pleasingly parallel version of such algorithms with interacting latent variables. We show that the interacting algorithm can be theoretically more efficient than the single-chain alternative with an equivalent computational budget.</div>

Depuis l’article fondateur de Jordan, Kinderlehrer et Otto en 1998, il est bien connu qu’une large classe d’équations paraboliques peuvent être vues comme des flots de gradient dans l’espace de Wasserstein. Le but de cette thèse est d’étendre cette théorie à certaines équations et systèmes qui n’ont pas exactement une structure de flot de gradient. Les interactions étudiées sont de différentes natures. Le premier chapitre traite des systèmes avec des interactions non locales dans la dérive. Nous étudions ensuite des systèmes de diffusions croisées s’appliquant aux modèles de congestion pour plusieurs populations. Un autre modèle étudié est celui où le couplage se trouve dans le terme de réaction comme les systèmes proie-prédateur avec diffusion ou encore les modèles de croissance tumorale. Nous étudierons enfin des systèmes de type nouveau où l’interaction est donnée par un problème de transport multi-marges. Une grande partie de ces problèmes est illustrée de simulations numériques. / Since 1998 and the seminal work of Jordan, Kinderlehrer and Otto, it is well known that a large class of parabolic equations can be seen as gradient flows in the Wasserstein space. This thesis is devoted to extensions of this theory to equations and systems which do not have exactly a gradient flow structure. We study different kind of couplings. First, we treat the case of nonlocal interactions in the drift. Then, we study cross diffusion systems which model congestion for several species. We are also interested in reaction-diffusion systems as diffusive prey-predator systems or tumor growth models. Finally, we introduce a new class of systems where the interaction is given by a multi-marginal transport problem. In many cases, we give numerical simulations to illustrate our theorical results.