Kernel Methods in Particle Filtering David Coufal Doctoral thesis - abstract The thesis deals with the use of kernel density estimates in particle filtering. In particular, it examines the convergence of the kernel density estimates to the filtering densities. The estimates are constructed on the basis of an out- put from particle filtering. It is proved theoretically that using the standard kernel density estimation methodology is effective in the context of particle filtering, although particle filtering does not produce random samples from the filtering densities. The main theoretical results are: 1) specification of the upper bounds on the MISE error of the estimates of the filtering densities and their partial derivatives; 2) specification of the related lower bounds and 3) providing a suitable tool for checking persistence of the Sobolev character of the filtering densities over time. In addition, the thesis also focuses on designing kernels suitable for practical use. 1

A sorted array is a fundamental algorithmic concept. Its on-line variant gives rise to the online labeling problem. In the online labeling problem we are given an array of size m and a stream of n integers from the universe {1, ..., r} coming in an arbitrary order. Our task is to maintain all received items in the array in sorted order. The inserted items do not have to be stored consecutively in the array. Since the final order of the items is not known until we see all the items, moves of already inserted items are allowed but should be minimized. We present two algorithms which together provide an optimal solution for almost all values of m as a function of n. We provide tight lower bounds for almost all ranges of m. We introduce a notion of the limited universe and prove lower bounds also in that setting. Some of our lower bounds also apply to randomized algorithms. Powered by TCPDF (www.tcpdf.org)

Cette thèse porte sur des minorants pour des mesures de complexité liées à des sous-classes de la classe P de langages pouvant être décidés en temps polynomial par des machines de Turing. Nous considérons des modèles de calcul non uniformes tels que les programmes sur monoïdes et les programmes de branchement. Notre première contribution est un traitement abstrait de la méthode de Nečiporuk pour prouver des minorants, indépendamment de toute mesure de complexité spécifique. Cette méthode donne toujours les meilleurs minorants connus pour des mesures telles que la taille des programmes de branchements déterministes et non déterministes ou des formules avec des opérateurs booléens binaires arbitraires ; nous donnons une formulation abstraite de la méthode et utilisons ce cadre pour démontrer des limites au meilleur minorant obtenable en utilisant cette méthode pour plusieurs mesures de complexité. Par là, nous confirmons, dans ce cadre légèrement plus général, des résultats de limitation précédemment connus et exhibons de nouveaux résultats de limitation pour des mesures de complexité auxquelles la méthode de Nečiporuk n'avait jamais été appliquée. Notre seconde contribution est une meilleure compréhension de la puissance calculatoire des programmes sur monoïdes issus de petites variétés de monoïdes finis. Les programmes sur monoïdes furent introduits à la fin des années 1980 par Barrington et Thérien pour généraliser la reconnaissance par morphismes et ainsi obtenir une caractérisation en termes de semi-groupes finis de NC^1 et de ses sous-classes. Étant donné une variété V de monoïdes finis, on considère la classe P(V) de langages reconnus par une suite de programmes de longueur polynomiale sur un monoïde de V : lorsque l'on fait varier V parmi toutes les variétés de monoïdes finis, on obtient différentes sous-classes de NC^1, par exemple AC^0, ACC^0 et NC^1 quand V est respectivement la variété de tous les monoïdes apériodiques finis, résolubles finis et finis. Nous introduisons une nouvelle notion de docilité pour les variétés de monoïdes finis, renforçant une notion de Péladeau. L'intérêt principal de cette notion est que quand une variété V de monoïdes finis est docile, nous avons que P(V) contient seulement des langages réguliers qui sont quasi reconnus par morphisme par des monoïdes de V. De nombreuses questions ouvertes à propos de la structure interne de NC^1 seraient réglées en montrant qu'une variété de monoïdes finis appropriée est docile, et, dans cette thèse, nous débutons modestement une étude exhaustive de quelles variétés de monoïdes finis sont dociles. Plus précisément, nous portons notre attention sur deux petites variétés de monoïdes apériodiques finis bien connues : DA et J. D'une part, nous montrons que DA est docile en utilisant des arguments de théorie des semi-groupes finis. Cela nous permet de dériver une caractérisation algébrique exacte de la classe des langages réguliers dans P(DA). D'autre part, nous montrons que J n'est pas docile. Pour faire cela, nous présentons une astuce par laquelle des programmes sur monoïdes de J peuvent reconnaître beaucoup plus de langages réguliers que seulement ceux qui sont quasi reconnus par morphisme par des monoïdes de J. Cela nous amène à conjecturer une caractérisation algébrique exacte de la classe de langages réguliers dans P(J), et nous exposons quelques résultats partiels appuyant cette conjecture. Pour chacune des variétés DA et J, nous exhibons également une hiérarchie basée sur la longueur des programmes à l'intérieur de la classe des langages reconnus par programmes sur monoïdes de la variété, améliorant par là les résultats de Tesson et Thérien sur la propriété de longueur polynomiale pour les monoïdes de ces variétés. / This thesis deals with lower bounds for complexity measures related to subclasses of the class P of languages that can be decided by Turing machines in polynomial time. We consider non-uniform computational models like programs over monoids and branching programs.Our first contribution is an abstract, measure-independent treatment of Nečiporuk's method for proving lower bounds. This method still gives the best lower bounds known on measures such as the size of deterministic and non-deterministic branching programs or formulae{} with arbitrary binary Boolean operators; we give an abstract formulation of the method and use this framework to prove limits on the best lower bounds obtainable using this method for several complexity measures. We thereby confirm previously known limitation results in this slightly more general framework and showcase new limitation results for complexity measures to which Nečiporuk's method had never been applied.Our second contribution is a better understanding of the computational power of programs over monoids taken from small varieties of finite monoids. Programs over monoids were introduced in the late 1980s by Barrington and Thérien as a way to generalise recognition by morphisms so as to obtain a finite-semigroup-theoretic characterisation of NC^1 and its subclasses. Given a variety V of finite monoids, one considers the class P(V) of languages recognised by a sequence of polynomial-length programs over a monoid from V: as V ranges over all varieties of finite monoids, one obtains different subclasses of NC^1, for instance AC^0, ACC^0 and NC^1 when V respectively is the variety of all finite aperiodic, finite solvable and finite monoids. We introduce a new notion of tameness for varieties of finite monoids, strengthening a notion of Péladeau. The main interest of this notion is that when a variety V of finite monoids is tame, we have that P(V) does only contain regular languages that are quasi morphism-recognised by monoids from V. Many open questions about the internal structure of NC^1 would be settled by showing that some appropriate variety of finite monoids is tame, and, in this thesis, we modestly start an exhaustive study of which varieties of finite monoids are tame. More precisely, we focus on two well-known small varieties of finite aperiodic monoids: DA and J. On the one hand, we show that DA is tame using finite-semigroup-theoretic arguments. This allows us to derive an exact algebraic characterisation of the class of regular languages in P(DA). On the other hand, we show that J is not tame. To do this, we present a trick by which programs over monoids from J can recognise much more regular languages than only those that are quasi morphism-recognised by monoids from J. This brings us to conjecture an exact algebraic characterisation of the class of regular languages in P(J), and we lay out some partial results that support this conjecture. For each of the varieties DA and J, we also exhibit a program-length-based hierarchy within the class of languages recognised by programs over monoids from the variety, refining Tesson and Thérien's results on the polynomial-length property for monoids from those varieties.

With diminishing returns in performance with scaling of traditional transistor devices, there is a growing need to understand and improve potential replacements technologies. Sufficient reliability has not been established in these devices and additional redundancy through use of fault tolerance and error correction codes are necessary. There is a price to pay in terms of energy and area, with this additional redundancy. It is of utmost importance to determine this energy cost and relate it to the increased reliability offered by the use of error correction codes. In this thesis, we have determined the lower bound for energy dissipation associated with error correction using a linear (n,k) block code. The bound obtained is implementation independent and is derived from fundamental considerations and it allows for quantum effects in the channel and decoder. We have also developed information theoretic efficacy measures that can quantify the performance of the error correction and their relationship to the corresponding energy cost.

No description available.

No description available.

The task of choosing a preconditioner M to use when solving a linear system Ax=b with iterative methods is often tedious and most methods remain ad-hoc. This thesis presents a randomized algorithm to make this chore less painful through use of randomized algorithms for estimating traces. In particular, we show that the preconditioner stability || I - M-1A ||F, known to forecast preconditioner quality, can be computed in the time it takes to run a constant number of iterations of conjugate gradients through use of sketching methods. This is in spite of folklore which suggests the quantity is impractical to compute, and a proof we give that ensures the quantity could not possibly be approximated in a useful amount of time by a deterministic algorithm. Using our estimator, we provide a method which can provably select a quality preconditioner among n candidates using floating operations commensurate with running about n log(n) steps of the conjugate gradients algorithm. In the absence of such a preconditioner among the candidates, our method can advise the practitioner to use no preconditioner at all. The algorithm is extremely easy to implement and trivially parallelizable, and along the way we provide theoretical improvements to the literature on trace estimation. In empirical experiments, we show the selection method can be quite helpful. For example, it allows us to create to the best of our knowledge the first preconditioning method for kernel regression which never uses more iterations over the non-preconditioned analog in standard settings.

No description available.

This manuscript concerns the performance analysis in signal processing and consists into two parts : First, we study the lower bounds in characterizing and predicting the estimation performance in terms of mean square error (MSE). The lower bounds on the MSE give the minimum variance that an estimator can expect to achieve and it can be divided into two categories depending on the parameter assumption: the so-called deterministic bounds dealing with the deterministic unknown parameters, and the so-called Bayesian bounds dealing with the random unknown parameter. Particularly, we derive the closed-form expressions of the lower bounds for two applications in two different fields: (i) The first one is the target localization using the multiple-input multiple-output (MIMO) radar in which we derive the lower bounds in the contexts with and without modeling errors, respectively. (ii) The other one is the pulse phase estimation of X-ray pulsars which is a potential solution for autonomous deep space navigation. In this application, we show the potential universality of lower bounds to tackle problems with parameterized probability density function (pdf) different from classical Gaussian pdf since in X-ray pulse phase estimation, observations are modeled with a Poisson distribution. Second, we study the statistical resolution limit (SRL) which is the minimal distance in terms of the parameter of interest between two signals allowing to correctly separate/estimate the parameters of interest. More precisely, we derive the SRL in two contexts: array processing and MIMO radar by using two approaches based on the estimation theory and information theory. We also present in this thesis the usefulness of SRL in optimizing the array system.

Le but principal de ce manuscrit est celui de présenter une nouvelle méthode d'interpolation entre des probabilités inspirée du problème de Schrödinger, problème de minimisation entropique ayant des liens très forts avec le transport optimal. À l'aide de solutions au problème de Schrödinger, nous obtenons un schéma d'approximation robuste jusqu'au deuxième ordre et différent de Brenier-McCann qui permet d'établir la formule de dérivation du deuxième ordre le long des géodésiques Wasserstein dans le cadre de espaces RCD* de dimension finie. Cette formule était inconnue même dans le cadre des espaces d'Alexandrov et nous en donnerons quelques applications. La démonstration utilise un ensemble remarquable de nouvelles propriétés pour les solutions au problème de Schrödinger dynamique :- une borne uniforme des densités le long des interpolations entropiques ;- la lipschitzianité uniforme des potentiels de Schrödinger ;- un contrôle L2 uniforme des accélérations. Ces outils sont indispensables pour explorer les informations géométriques encodées par les interpolations entropiques. Les techniques utilisées peuvent aussi être employées pour montrer que la solution visqueuse de l'équation d'Hamilton-Jacobi peut être récupérée à travers une méthode de « vanishing viscosity », comme dans le cas lisse.Dans tout le manuscrit, plusieurs remarques sur l'interprétation physique du problème de Schrödinger seront mises en lumière. Cela pourra aider le lecteur à mieux comprendre les motivations probabilistes et physiques du problème, ainsi qu'à les connecter avec la nature analytique et géométrique de la dissertation. / Main aim of this manuscript is to present a new interpolation technique for probability measures, which is strongly inspired by the Schrödinger problem, an entropy minimization problem deeply related to optimal transport. By means of the solutions to the Schrödinger problem, we build an efficient approximation scheme, robust up to the second order and different from Brenier-McCann's classical one. Such scheme allows us to prove the second order differentiation formula along geodesics in finite-dimensional RCD* spaces. This formula is new even in the context of Alexandrov spaces and we provide some applications.The proof relies on new, even in the smooth setting, estimates concerning entropic interpolations which we believe are interesting on their own. In particular we obtain:- equiboundedness of the densities along the entropic interpolations,- equi-Lipschitz continuity of the Schrödinger potentials,- a uniform weighted L2 control of the Hessian of such potentials. These tools are very useful in the investigation of the geometric information encoded in entropic interpolations. The techniques used in this work can be also used to show that the viscous solution of the Hamilton-Jacobi equation can be obtained via a vanishing viscosity method, in accordance with the smooth case. Throughout the whole manuscript, several remarks on the physical interpretation of the Schrödinger problem are pointed out. Hopefully, this will allow the reader to better understand the physical and probabilistic motivations of the problem as well as to connect them with the analytical and geometric nature of the dissertation.