Дисертації з теми "Quadratic estimates"

Ce mémoire est consacré à des résultats d'analyse harmonique réelle dans des cadres géométriques discrets (graphes) ou continus (groupes de Lie).Soit $\Gamma$ un graphe (ensemble de sommets et d'arêtes) muni d'un laplacien discret $\Delta=I-P$, où $P$ est un opérateur de Markov.Sous des hypothèses géométriques convenables sur $\Gamma$, nous montrons la continuité $L^p$ de fonctionnelles de Littlewood-Paley fractionnaires. Nous introduisons des espaces de Hardy $H^1$ de fonctions et de $1$-formes différentielles sur $\Gamma$, dont nous donnons plusieurs caractérisations, en supposant seulement la propriété de doublement pour le volume des boules de $\Gamma$. Nous en déduisons la continuité de la transformée de Riesz sur $H^1$. En supposant de plus des estimations supérieures ponctuelles (gaussiennes ou sous-gaussiennes) sur les itérées du noyau de l'opérateur $P$, nous obtenons aussi la continuité de la transformée de Riesz sur $L^p$ pour $10$, $1\leq p\leq+\infty$ et $1\leq q\leq +\infty$. Les résultats sont valables en croissance polynomiale ou exponentielle du volume des boules
This thesis is devoted to results in real harmonic analysis in discrete (graphs) or continuous (Lie groups) geometric contexts.Let $\Gamma$ be a graph (a set of vertices and edges) equipped with a discrete laplacian $\Delta=I-P$, where $P$ is a Markov operator.Under suitable geometric assumptions on $\Gamma$, we show the $L^p$ boundedness of fractional Littlewood-Paley functionals. We introduce $H^1$ Hardy spaces of functions and of $1$-differential forms on $\Gamma$, giving several characterizations of these spaces, only assuming the doubling property for the volumes of balls in $\Gamma$. As a consequence, we derive the $H^1$ boundedness of the Riesz transform. Assuming furthermore pointwise upper bounds for the kernel (Gaussian of subgaussian upper bounds) on the iterates of the kernel of $P$, we also establish the $L^p$ boundedness of the Riesz transform for $10$, $1\leq p\leq+\infty$ and $1\leq q\leq +\infty$.These results hold for polynomial as well as for exponential volume growth of balls

Ein nichtparametrisches Schätzverfahren für die quadratische Kovariation von hochfrequent nicht-synchron beobachteter Itô-Prozessen mit einem additiven Rauschen wird entwickelt. Für eine artverwandte Folge von statistischen Experimenten wird die lokal asymptotische Normalität (LAN) im Sinne von Le Cam bewiesen. Mit dieser lassen sich optimale Konvergenzraten und Effizienzschranken für asymptotische Varianzen ableiten. Der vorgestellte Schätzer wird auf Grundlage von zwei modernen Verfahren, für die Anwendung bei nicht-synchronen Beobachtungen zum einen, und einem additiven Rauschen zum anderen, entwickelt. Der Hayashi-Yoshida Schätzer wird in einer neuen Darstellung eingeführt, welche einen Synchronisierungsalgorithmus mit einschließt, der für die kombinierte Methode ausgelegt werden kann. Es wird eine stabiles zentrales Grenzwerttheorem bewiesen, wobei spezieller Wert auf die Analyse des Einflusses der Nicht-Synchronität auf die asymptotische Varianz gelegt wird. Nach diesen Vorbereitungen wird das kombinierte Schätzverfahren für den allgemeinsten Fall nicht-synchroner verrauschter Beobachtungen vorgestellt. Dieses beruht auf Subsampling- und Multiskalenmethoden, die auf Mykland, Zhang und Aït-Sahalia zurück gehen. Es vereint positive Eigenschaften der beiden Ursprünge. Das zentrale Resultat dieser Arbeit ist der Beweis, dass der Schätzfehler stabil in Verteilung gegen eine gemischte Normalverteilung konvergiert. Für die asymptotische Varianz wird ein konsistenter Schätzer angegeben. In einer Anwendungsstudie wird eine praktische Implementierung des Schätzverfahrens, die die Wahl von abhängigen Parametern beinhaltet, getestet und auf ihre Eigenschaften im Falle endlicher Stichprobenumfänge untersucht. Neuen fortgeschrittenen Entwicklungen auf dem Forschungsfeld von Seite anderer Autoren wird Rechnung getragen durch Vergleiche und diesbezügliche Kommentare.
A nonparametric estimation approach for the quadratic covariation of Itô processes from high-frequency observations with an additive noise is developed. It is proved that a closely related sequence of statistical experiments is locally asymptotically normal (LAN) in the Le Cam sense. By virtue of this property optimal convergence rates and efficiency bounds for asymptotic variances of estimators can be concluded. The proposed nonparametric estimator is founded on a combination of two modern estimation methods devoted to an additive observation noise on the one hand and asynchronous observation schemes on the other hand. We reinvent this Hayashi-Yoshida estimator in a new illustration that can serve as a synchronization method which is possible to adapt for the combined approach. A stable central limit theorem is proved focusing especially on the impact of non-synchronicity on the asymptotic variance. With this preparations on hand, the generalized multiscale estimator for the noisy and asynchronous setting arises. This convenient method for the general model is based on subsampling and multiscale estimation techniques that have been established by Mykland, Zhang and Aït-Sahalia. It preserves valuable features of the synchronization methodology and the estimators to cope with noise perturbation. The central result of the thesis is that the estimation error of the generalized multiscale estimator converges with optimal rate stably in law to a centred mixed normal limiting distribution on fairly general regularity assumptions. For the asymptotic variance a consistent estimator based on time transformed histograms is given making the central limit theorem feasible. In an application study a practicable estimation algorithm including a choice of tuning parameters is tested for its features and finite sample size behaviour. We take account of recent advances on the research field by other authors in comparisons and notes.

Cette thèse porte sur l’estimation d'une fonction de régression fournissant la meilleure relation entredes variables pour lesquelles on possède un certain nombre d’observations. Une première partie portesur une étude par simulation de deux méthodes automatiques de sélection du paramètre de laprocédure d'estimation ridge. D'un point de vue plus théorique, on présente et compare ensuite deuxméthodes de sélection d'un multiparamètre intervenant dans une procédure d'estimation d'unefonction de régression sur l'intervalle [0,1]. Dans une deuxième partie, on étudie la qualité del'estimateur PLS1, d'un point de vue théorique, à travers son risque quadratique et, plus précisément,le terme de variance dans la décomposition biais/variance de ce risque. Enfin, dans une troisièmepartie, une étude statistique sur données réelles est menée afin de mieux comprendre la signaturegénique de cellules cancéreuses à partir de la signature génique des sous-types cellulaires constituantle stroma tumoral associé
This thesis deals with the estimation of a regression function providing the best relationship betweenvariables for which we have some observations. In a first part, we complete a simulation study fortwo automatic selection methods of the ridge parameter. From a more theoretical point of view, wethen present and compare two selection methods of a multiparameter, that is used in an estimationprocedure of a regression function on [0,1]. In a second part, we study the quality of the PLS1estimator through its quadratic risk and, more precisely, the variance term in its bias/variancedecomposition. In a third part, a statistical study is carried out in order to explain the geneticsignature of cancer cells thanks to the genetic signatures of cellular subtypes which compose theassociated tumor stroma

Beaucoup de résultats sur l’estimation d’une matrice d’échelle en analyse multidimensionnelle sont obtenus sous l’hypothèse de normalité (condition sous laquelle il s’agit de la matrice de covariance). Or il s’avère que, dans des domaines tels que la gestion de portefeuille en finance, cette hypothèse n’est pas très appropriée. Dans ce cas, la famille des distributions à symétrie elliptique, qui contient la distribution gaussienne, est une alternative intéressante. Nous considérons dans cette thèse le problème d’estimation de la matrice d’échelle Σ du modèle additif Yp_m = M + E, d’un point de vue de la théorie de la décision. Ici, p représente le nombre de variables, m le nombre d’observations, M une matrice de paramètres inconnus de rang q < p et E un bruit aléatoire de distribution à symétrie elliptique, avec une matrice de covariance proportionnelle à Im x Σ. Ce problème d’estimation est abordé sous la représentation canonique de ce modèle où la matrice d’observation Y est décomposée en deux matrices, à savoir, Zq x p qui résume l’information contenue dans M et une matrice Un x p, où n = m - q, qui résume l’information suffisante pour l’estimation de Σ. Comme les estimateurs naturels de la forme Σa = a S (où S = UT U et a est une constante positive) ne sont pas de bons estimateurs lorsque le nombre de variables p et le rapport p=n sont grands, nous proposons des estimateurs alternatifs de la forme ^Σa;G = a(S + S S+G(Z; S)) où S+ est l’inverse de Moore-Penrose de S (qui coïncide avec l’inverse S-1 lorsque S est inversible). Nous fournissons des conditions sur la matrice de correction SS+G(Z; S) telles que ^Σa;G améliore^Σa sous le coût quadratique L(Σ; ^Σ) = tr(^ΣΣ‾1 - Ip)² et sous une modification de ce dernier, à savoir le coût basé sur les données LS (Σ; ^Σ) = tr(S+Σ(^ΣΣ‾1 - Ip)²). Nous adoptons une approche unifiée des deux cas où S est inversible et S est non inversible. À cette fin, une nouvelle identité de type Stein-Haff et un nouveau calcul sur la décomposition en valeurs propres de S sont développés. Notre théorie est illustrée par une grande classe d’estimateurs orthogonalement invariants et par un ensemble de simulations
Numerous results on the estimation of a scale matrix in multivariate analysis are obtained under Gaussian assumption (condition under which it is the covariance matrix). However in such areas as Portfolio management in finance, this assumption is not well adapted. Thus, the family of elliptical symmetric distribution, which contains the Gaussian distribution, is an interesting alternative. In this thesis, we consider the problem of estimating the scale matrix _ of the additif model Yp_m = M + E, under theoretical decision point of view. Here, p is the number of variables, m is the number of observations, M is a matrix of unknown parameters with rank q < p and E is a random noise, whose distribution is elliptically symmetric with covariance matrix proportional to Im x Σ. It is more convenient to deal with the canonical forme of this model where Y is decomposed in two matrices, namely, Zq_p which summarizes the information contained in M, and Un_p, where n = m - q which summarizes the information sufficient to estimate Σ. As the natural estimators of the form ^Σ a = a S (where S = UT U and a is a positive constant) perform poorly when the dimension of variables p and the ratio p=n are large, we propose estimators of the form ^Σa;G = a(S + S S+G(Z; S)) where S+ is the Moore-Penrose inverse of S (which coincides with S-1 when S is invertible). We provide conditions on the correction matrix SS+G(Z; S) such that ^Σa;G improves over ^Σa under the quadratic loss L(Σ; ^Σ) = tr(^ΣΣ‾1 - Ip)² and under the data based loss LS (Σ; ^Σ) = tr(S+Σ(^ΣΣ‾1 - Ip)²).. We adopt a unified approach of the two cases where S is invertible and S is non-invertible. To this end, a new Stein-Haff type identity and calculus on eigenstructure for S are developed. Our theory is illustrated with the large class of orthogonally invariant estimators and with simulations

Cette thèse a pour objet d'étude les méthodes numériques d'approximation de la densité de probabilité associée à des variables aléatoires admettant des distributions composées. Ces variables aléatoires sont couramment utilisées en actuariat pour modéliser le risque supporté par un portefeuille de contrats. En théorie de la ruine, la probabilité de ruine ultime dans le modèle de Poisson composé est égale à la fonction de survie d'une distribution géométrique composée. La méthode numérique proposée consiste en une projection orthogonale de la densité sur une base de polynômes orthogonaux. Ces polynômes sont orthogonaux par rapport à une mesure de probabilité de référence appartenant aux Familles Exponentielles Naturelles Quadratiques. La méthode d'approximation polynomiale est comparée à d'autres méthodes d'approximation de la densité basées sur les moments et la transformée de Laplace de la distribution. L'extension de la méthode en dimension supérieure à $1$ est présentée, ainsi que l'obtention d'un estimateur de la densité à partir de la formule d'approximation. Cette thèse comprend aussi la description d'une méthode d'agrégation adaptée aux portefeuilles de contrats d'assurance vie de type épargne individuelle. La procédure d'agrégation conduit à la construction de model points pour permettre l'évaluation des provisions best estimate dans des temps raisonnables et conformément à la directive européenne Solvabilité II
This PhD thesis studies numerical methods to approximate the probability density function of random variables governed by compound distributions. These random variables are useful in actuarial science to model the risk of a portfolio of contracts. In ruin theory, the probability of ultimate ruin within the compound Poisson ruin model is the survival function of a geometric compound distribution. The proposed method consists in a projection of the probability density function onto an orthogonal polynomial system. These polynomials are orthogonal with respect to a probability measure that belongs to Natural Exponential Families with Quadratic Variance Function. The polynomiam approximation is compared to other numerical methods that recover the probability density function from the knowledge of the moments or the Laplace transform of the distribution. The polynomial method is then extended in a multidimensional setting, along with the probability density estimator derived from the approximation formula. An aggregation procedure adapted to life insurance portfolios is also described. The method aims at building a portfolio of model points in order to compute the best estimate liabilities in a timely manner and in a way that is compliant with the European directive Solvency II

Cognitive radio is a technology that aims to maximize the current usage of the licensed frequency spectrum. Cognitive radio aims to provide services for license-exempt users by making use of dynamic spectrum access (DSA) and opportunistic spectrum sharing strategies (OSS). Cognitive radios are defined as intelligent wireless devices capable of adapting their communication parameters in order to operate within underutilized bands while avoiding causing interference to licensed users. An underused band of frequencies in a specific location or time is known as a spectrum hole. Therefore, in order to locate spectrum holes, reliable spectrum sensing algorithms are crucial to facilitate the evolution of cognitive radio networks. Since a large and growing body of literature has mainly focused into the conventional time domain (TD) energy detector, throughout this thesis the problem of spectrum sensing is investigated within the context of a frequency domain (FD) approach. The purpose of this study is to investigate detection based on methods of nonparametric power spectrum estimation. The considered methods are the periodogram, Bartlett's method, Welch overlapped segments averaging (WOSA) and the Multitaper estimator (MTE). Another major motivation is that the MTE is strongly recommended for the application of cognitive radios. This study aims to derive the detector performance measures for each case. Another aim is to investigate and highlight the main differences between the TD and the FD approaches. The performance is addressed for independent and identically distributed (i.i.d.) Rayleigh channels and the general Rician and Nakagami fading channels. For each of the investigated detectors, the analytical models are obtained by studying the characteristics of the Hermitian quadratic form representation of the decision statistic and the matrix of the Hermitian form is identified. The results of the study have revealed the high accuracy of the derived mathematical models. Moreover, it is found that the TD detector differs from the FD detector in a number of aspects. One principal and generalized conclusion is that all the investigated FD methods provide a reduced probability of false alarm when compared with the TD detector. Also, for the case of periodogram, the probability of sensing errors is independent of the length of observations, whereas in time domain the probability of false alarm is increased when the sample size increases. The probability of false alarm is further reduced when diversity reception is employed. Furthermore, compared to the periodogram, both Bartlett method and Welch method provide better performance in terms of lower probability of false alarm but an increased probability of detection for a given probability of false alarm. Also, the performance of both Bartlett's method and WOSA is sensitive to the number of segments, whereas WOSA is also sensitive to the overlapping factor. Finally, the performance of the MTE is dependent on the number of employed discrete prolate spheroidal (Slepian) sequences, and the MTE outperforms the periodogram, Bartlett's method and WOSA, as it provides the minimal probability of false alarm.

This thesis is concerned with the Cauchy problem for the quadratic derivative nonlinear wave equation in two spatial dimensions. Using standard techniques, we reduce local well-posedness in Fourier Lebesgue spaces to bilinear estimates in associated wave Fourier Lebesgue spaces, for which we prove new product estimates. These estimates then allow us to establish local well-posedness in a parameter range that gives improvement over previously known results on the Sobolev scale.

The connection between quadratic estimates and the existence of a bounded holomorphic functional calculus of an operator provides a framework for applying harmonic analysis to the theory of differential operators. This is a generalization of the connection between Littlewood--Paley--Stein estimates and the functional calculus provided by the Fourier transform. We use the former approach in this thesis to study first-order differential operators on Riemannian manifolds. The theory developed is local in the sense that it does not depend on the spectrum of the operator in a neighbourhood of the origin. When we apply harmonic analysis to obtain estimates, the local theory only requires that we do so up to a finite scale. This allows us to consider manifolds with exponential volume growth in situations where the global theory requires polynomial volume growth. A holomorphic functional calculus is constructed for operators on a reflexive Banach space that are bisectorial except possibly in a neighbourhood of the origin. We prove that this functional calculus is bounded if and only if certain local quadratic estimates hold. For operators with spectrum in a neighbourhood of the origin, the results are weaker than those for bisectorial operators. For operators with a spectral gap in a neighbourhood of the origin, the results are stronger. In each case, however, local quadratic estimates are a more appropriate tool than standard quadratic estimates for establishing that the functional calculus is bounded. This theory allows us to define local Hardy spaces of differential forms that are adapted to a class of first-order differential operators on a complete Riemannian manifold with at most exponential volume growth. The local geometric Riesz transform associated with the Hodge--Dirac operator is bounded on these spaces provided that a certain condition on the exponential growth of the manifold is satisfied. A characterisation of these spaces in terms of local molecules is also obtained. These results can be viewed as the localisation of those for the Hardy spaces of differential forms introduced by Auscher, McIntosh and Russ. Finally, we introduce a class of first-order differential operators that act on the trivial bundle over a complete Riemannian manifold with at most exponential volume growth and on which a local Poincar\'{e} inequality holds. A local quadratic estimate is established for certain perturbations of these operators. As an application, we solve the Kato square root problem for divergence form operators on complete Riemannian manifolds with Ricci curvature bounded below that are embedded in Euclidean space with a uniformly bounded second fundamental form. This is based on the framework for Dirac type operators that was introduced by Axelsson, Keith and McIntosh.

The primary focus of this thesis is to consider Kato square root problems for various divergence-form operators on manifolds. This is the study of perturbations of second-order differential operators by bounded, complex, measurable coefficients. In general, such operators are not self-adjoint but uniformly elliptic. The Kato square root problem is then to understand when the square root of such an operator, which exists due to uniform ellipticity, is comparable to its unperturbed counterpart. A remarkably adaptable operator-theoretic framework due to Axelsson, Keith and McIntosh sits in the background of this work. This framework allows us to take a powerful first-order perspective of the problems which we consider in a geometric setting. Through a well established procedure, we reduce these problems to the study of quadratic estimates. Under a set of natural conditions, we prove quadratic estimates for a class of operators on vector bundles over complete measure metric spaces. The first kind of estimates we prove are global, and we establish them on trivial vector bundles when the underlying measure grows at most polynomially. The second kind are local, and there, we allow the vector bundle to be non-trivial but bounded in an appropriate sense. Here, the measure is allowed to grow exponentially. An important consequence of obtaining quadratic estimates on measure metric spaces is that it allows us to consider subelliptic operators on Lie groups. The first-order perspective allows us to reduce the subelliptic problem to a fully elliptic one on a sub-bundle. As a consequence, we are able to solve a homogeneous Kato square root problem for perturbations of subelliptic operators on nilpotent Lie groups. For general Lie groups we solve a similar inhomogeneous problem. In the situation of complete Riemannian manifolds, we consider uniformly elliptic divergence-form operators arising from connections on vector bundles. Under a set of assumptions, we show that the Kato square root problem can be solved for such operators. As a consequence, we solve this problem on functions under the condition that the Ricci curvature and injectivity radius are bounded. Assuming an additional lower bound for the curvature endomorphism on forms, we solve a similar problem for perturbations of inhomogeneous Hodge-Dirac operators. A theorem for tensors is obtained by additionally assuming boundedness of a second-order Riesz transform. Motivated by the study of these Kato problems, where for technical reasons it is useful to know the density of compactly supported functions in the domains of operators, we study connections and their divergence on a vector bundle. Through a first-order formulation, we show that this density property holds for the domains of these operators if the metric and connection are compatible and the underlying manifold is complete. We also show that compactly supported functions are dense in the second-order Sobolev space on complete manifolds under the sole assumption that the Ricci curvature is bounded below, improving a result that previously required an additional lower bound on the injectivity radius.

Face recognition is a challenging problem in computer vision. Difficulties such as slight differences between similar faces of different people, changes in facial expressions, light and illumination condition, and pose variations add extra complications to the face recognition research. Many algorithms are devoted to solving the face recognition problem, among which the family of nonnegative matrix factorization (NMF) algorithms has been widely used as a compact data representation method. Different versions of NMF have been proposed. Wang et al. proposed the graph-based semi-supervised nonnegative learning (S2N2L) algorithm that uses labeled data in constructing intrinsic and penalty graph to enforce separability of labeled data, which leads to a greater discriminating power. Moreover the geometrical structure of labeled and unlabeled data is preserved through using the smoothness assumption by creating a similarity graph that conserves the neighboring information for all labeled and unlabeled data. However, S2N2L is sensitive to light changes, illumination, and partial occlusion. In this thesis, we propose a Semi-Supervised Half-Quadratic NMF (SSHQNMF) algorithm that combines the benefits of S2N2L and the robust NMF by the half- quadratic minimization (HQNMF) algorithm.Our algorithm improves upon the S2N2L algorithm by replacing the Frobenius norm with a robust M-Estimator loss function. A multiplicative update solution for our SSHQNMF algorithmis driven using the half- 4 quadratic (HQ) theory. Extensive experiments on ORL, Yale-A and a subset of the PIE data sets for nine M-estimator loss functions for both SSHQNMF and HQNMF algorithms are investigated, and compared with several state-of-the-art supervised and unsupervised algorithms, along with the original S2N2L algorithm in the context of classification, clustering, and robustness against partial occlusion. The proposed algorithm outperformed the other algorithms. Furthermore, SSHQNMF with Maximum Correntropy (MC) loss function obtained the best results for most test cases.

Thesis (Ph. D.)--Florida State University, 2006.
Advisor: Xu-Feng Niu, Florida State University, College of Arts and Sciences, Dept. of Statistics. Title and description from dissertation home page (viewed Sept. 18, 2006). Document formatted into pages; contains ix, 116 pages. Includes bibliographical references.

碩士
國立中央大學
電機工程學系
104
Improved robust observer-based servo designs are proposed in this thesis for the linear systems subject to unknown disturbances. First, the poles of the error dynamic system of the state observer integrated with the unknown input estimator for the continuous-time minimum phase system subject to unknown input disturbance (UID) are optimally assigned to lie to the left of some vertical line in the s-plane with prescribed degree of relative stability. Similar merit has been also applied to the servo design. Consequently, restrictions on the estimation of UID with low frequencies and servo control for slow time-varying command inputs presented in literature have been released to the cases for the UID with high frequencies and drastic time-varying command inputs, so that a more wide range unknown input estimations and servo designs can be achieved. In contrast with the above-mentioned merits, the proposed approach for the continuous-time systems has been also extended to the discrete-time version for the minimum phase and/or non-minimum phase systems. Especially, the new current-output observer/UID esitmator-based servo design for the discrete-time system with an unknown disturbance is proposed. Furthermore, based on the equivalent input disturbance (EID) principle, the proposed approaches are applicable to the class of mismatched input disturbances.

Cette thèse est organisée en trois chapitres. Les deux premiers proposent une approche régularisée pour l’estimation du modèle de données de panel dynamique : l’estimateur GMM et l’estimateur LIML. Le dernier chapitre de la thèse est une application de la méthode de régularisation à l’estimation des élasticités de l’offre de travail en utilisant des modèles de pseudo-données de panel. Dans un modèle de panel dynamique, le nombre de conditions de moments augmente rapidement avec la dimension temporelle du panel conduisant à une matrice de covariance des instruments de grande dimension. L’inversion d’une telle matrice pour calculer l’estimateur affecte négativement les propriétés de l’estimateur en échantillon fini. Comme solution à ce problème, nous proposons une approche par la régularisation qui consiste à utiliser une inverse généralisée de la matrice de covariance au lieu de son inverse classique. Trois techniques de régularisation sont utilisées : celle des composantes principales, celle de Tikhonov qui est basée sur le Ridge régression (aussi appelée Bayesian shrinkage) et enfin celle de Landweber Fridman qui est une méthode itérative. Toutes ces techniques introduisent un paramètre de régularisation qui est similaire au paramètre de lissage dans les régressions non paramétriques. Les propriétés en echantillon fini de l’estimateur régularisé dépend de ce paramètre qui doit être sélectionné parmis plusieurs valeurs potentielles. Dans le premier chapitre (co-écrit avec Marine Carrasco), nous proposons l’estimateur GMM régularisé du modèle de panel dynamique. Sous l’hypothèse que le nombre d’individus et de périodes du panel tendent vers l’infini, nous montrons que nos estimateurs sont convergents and assymtotiquement normaux. Nous dérivons une méthode empirique de sélection du paramètrede régularisation basée sur une expansion de second ordre du l’erreur quadratique moyenne et nous démontrons l’optimalité de cette procédure de sélection. Les simulations montrent que la régularisation améliore les propriétés de l ’estimateur GMM classique. Comme application empirique, nous avons analysé l’effet du développement financier sur la croissance économique. Dans le deuxième chapitre (co-écrit avec Marine Carrasco), nous nous intéressons à l’estimateur LIML régularisé du modèle de données de panel dynamique. L’estimateur LIML est connu pour avoir de meilleures propriétés en échantillon fini que l’estimateur GMM mais son utilisation devient problématique lorsque la dimension temporelle du panel devient large. Nous dérivons les propriétes assymtotiques de l’estimateur LIML régularisé sous l’hypothèse que le nombre d’individus et de périodes du panel tendent vers l’infini. Une procédure empirique de sélection du paramètre de régularisation est aussi proposée. Les bonnes performances de l’estimateur régularisé par rapport au LIML classique (non régularisé), au GMM classique ainsi que le GMM régularisé sont confirmées par des simulations. Dans le dernier chapitre, je considère l’estimation des élasticités d’offre de travail des hommes canadiens. L’hétérogéneité inobservée ainsi que les erreurs de mesures sur les salaires et les revenus sont connues pour engendrer de l’endogéneité quand on estime les modèles d’offre de travail. Une solution fréquente à ce problème d’endogéneité consiste à régrouper les données sur la base des carastéristiques observables et d’ éffectuer les moindres carrées pondérées sur les moyennes des goupes. Il a été démontré que cet estimateur est équivalent à l’estimateur des variables instrumentales sur les données individuelles avec les indicatrices de groupe comme instruments. Donc, en présence d’un grand nombre de groupe, cet estimateur souffre de biais en échantillon fini similaire à celui de l’estimateur des variables instrumentales quand le nombre d’instruments est élevé. Profitant de cette correspondance entre l’estimateur sur les données groupées et l’estimateur des variables instrumentales sur les données individuelles, nous proposons une approche régularisée à l’estimation du modèle. Cette approche conduit à des élasticités substantiellement différentes de ceux qu’on obtient en utilisant l’estimateur sur données groupées.
This thesis is organized in three chapters. The first two chapters propose a regularization approach to the estimation of two estimators of the dynamic panel data model : the Generalized Method of Moment (GMM) estimator and the Limited Information Maximum Likelihood (LIML) estimator. The last chapter of the thesis is an application of regularization to the estimation of labor supply elasticities using pseudo panel data models. In a dynamic panel data model, the number of moment conditions increases rapidly with the time dimension, resulting in a large dimensional covariance matrix of the instruments. Inverting this large dimensional matrix to compute the estimator leads to poor finite sample properties. To address this issue, we propose a regularization approach to the estimation of such models where a generalized inverse of the covariance matrix of the intruments is used instead of its usual inverse. Three regularization schemes are used : Principal components, Tikhonov which is based on Ridge regression (also called Bayesian shrinkage) and finally Landweber Fridman which is an iterative method. All these methods involve a regularization parameter which is similar to the smoothing parameter in nonparametric regressions. The finite sample properties of the regularized estimator depends on this parameter which needs to be selected between many potential values. In the first chapter (co-authored with Marine Carrasco), we propose the regularized GMM estimator of the dynamic panel data models. Under double asymptotics, we show that our regularized estimators are consistent and asymptotically normal provided that the regularization parameter goes to zero slower than the sample size goes to infinity. We derive a data driven selection of the regularization parameter based on an approximation of the higher-order Mean Square Error and show its optimality. The simulations confirm that regularization improves the properties of the usual GMM estimator. As empirical application, we investigate the effect of financial development on economic growth. In the second chapter (co-authored with Marine Carrasco), we propose the regularized LIML estimator of the dynamic panel data model. The LIML estimator is known to have better small sample properties than the GMM estimator but its implementation becomes problematic when the time dimension of the panel becomes large. We derive the asymptotic properties of the regularized LIML under double asymptotics. A data-driven procedure to select the parameter of regularization is proposed. The good performances of the regularized LIML estimator over the usual (not regularized) LIML estimator, the usual GMM estimator and the regularized GMM estimator are confirmed by the simulations. In the last chapter, I consider the estimation of the labor supply elasticities of Canadian men through a regularization approach. Unobserved heterogeneity and measurement errors on wage and income variables are known to cause endogeneity issues in the estimation of labor supply models. A popular solution to the endogeneity issue is to group data in categories based on observable characteristics and compute the weighted least squares at the group level. This grouping estimator has been proved to be equivalent to instrumental variables (IV) estimator on the individual level data using group dummies as intruments. Hence, in presence of large number of groups, the grouping estimator exhibites a small bias similar to the one of the IV estimator in presence of many instruments. I take advantage of the correspondance between grouping estimators and the IV estimator to propose a regularization approach to the estimation of the model. Using this approach leads to wage elasticities that are substantially different from those obtained through grouping estimators.