Estimating a bivariate extreme-value copula is equivalent to estimating A, its associated Pickands function. This function A: [0,1] \( \rightarrow \) [0,1] must satisfy some constraints : $$\max\{1-t, t \} \leq A(t) \leq 1, \hspace{3mm} t\in[0,1]$$ $$\text{A is convex.}$$ Many estimators have been proposed to estimate A, but few satisfy the imposed constraints. The main contribution of this thesis is the introduction of a simple correction technique for Pickands function estimators so that the corrected estimators respect the required constraints. The proposed correction uses a new property of the extreme-value random vector and the convex hull of the obtained estimator to guaranty the respect of the Pickands function constraints. The second contribution of this thesis is to present a nonparametric bayesian estimator of the Pickands function based on the form introduced by Capéraà, Fougères and Genest (1997). The estimator uses Dirichlet processes to estimate the cumulative distribution function of a transformation of the extreme-value bivariate vector. Analysis by simulations and a comparison with popular estimators provide a measure of performance for the proposed correction and bayesian estimator. The analysis is done on 18 bivariate extreme-value distributions. The correction reduces the mean square error on all distributions. The bayesian estimator has the lowest mean square error of all the considered estimators.

The subject of our thesis is intersection between two fields: The Extreme Value Theory (EVT) and copulas. The object of EVT is to find the limit law of the maximum of a sample. Due to the results of EVT, we can model extreme phenomena. In addition, there is a bivariate variant of EVT. The bivariate variant of EVT uses a family of copulas called extreme value copulas to account for the connection between the two extreme events. Any copula with extreme values depends on a Pickands function. The object of our thesis is to estimate the Pickands function from data. We have found a way to build a Pickands function using cubic splines. From this construction, we obtain an extended family of Pickands functions in which we perform our statistical inference. We chose the Bayesian approach to build the estimator and the MCMC methods for the estimates. The method was applied on simulated and real data.

Markov chain Monte Carlo (MCMC) methods commonly use chains that respect the detailed balance condition. These chains are called reversible. Most of the theory developed for MCMC evolves around those particular chains. Peskun (1973) and Tierney (1998) provided useful theorems on the ordering of the asymptotic variances for two estimators produced by two different reversible chains. In this thesis, we are interested in non-reversible chains, which are chains that don’t respect the detailed balance condition. We present algorithms that simulate non-reversible chains, mainly the Guided Random Walk (GRW) by Gustafson (1998) and the Discrete Bouncy Particle Sampler (DBPS) by Sherlock and Thiery (2017). For both algorithms, we compare the asymptotic variance of estimators with the ones produced by the Metropolis- Hastings algorithm. We present a recent theoretical framework introduced by Andrieu and Livingstone (2019) and their analysis of the GRW. We then show that the DBPS is part of this framework and present an analysis on the asymptotic variance of estimators. Their main theorem can provide an ordering of the asymptotic variances of two estimators resulting from nonreversible chains. We show that an estimator could have a lower asymptotic variance by adding propositions to the DBPS. We then present empirical results of a modified DBPS. Through the thesis we will mostly be interested in chains that are produced by deterministic proposals. We show a general construction of the delayed rejection algorithm using deterministic proposals and one possible equivalent for non-reversible chains.

In this thesis, we are mainly interested in the estimation of parameters. It consists of three papers in which we discuss estimation problems in specific models. In the first paper, we consider the Farlie-Gumbel-Morgenstern parametric family of copulas. The observations come from a distribution that involves both the copula and the margins, through the decomposition of the Sklar's theorem. Margins are unknown. In the estimation of a function of the parameter of the copula, pseudo-likelihood is often used instead of the likelihood. It is an approach that will generally produce non-efficient estimators, especially for small sample sizes. In pseudo-likelihood, margins are replaced by estimators that depend on the ranks. We propose to use the rank-likelihood, which is the likelihood based on the distribution of the rank statistics. This approach can be complex in practice because we work on the permutation group and we calculate multiple integrals. However, it is possible to do the calculations for the Farlie-Gumbel-Morgenstern family of copulas. We compare the estimators obtained from these approaches using the Bayesian method. The results of numerical studies are presented. Some estimators found in the literature are ratios of normal random variables. In the second paper, we are interested in the distribution of the ratio of two normal random variables. Several authors have addressed this subject. The new parametrization that we introduced allows us to obtain new results faster and effortlessly. Firstly, we show that the expression of the density of the ratio can be written as a mixture of densities belonging to a new family that we create. We find some properties, and we give the analytic expression of the characteristic function for the new family. This new family is in fact a generalization of the family of densities of Student distributions having odd degrees of freedom. We obtain convergence results generalizing the one for Student distributions when the degree of freedom tends to infinity. Similar results of convergence as well as some properties are developed for the distribution of the ratio. We use a Bayesian approach to estimate the ratio of the means of two normal random variables. Graphical illustrations and simulation results are presented. In the third paper, we extend the family of distributions, resulting from the difference of two independent random variables having a gamma distribution, by considering the case where these random variables are positively correlated. The densities in this family are found. Their complex forms make tedious the estimation methods based on the likelihood. The estimation method based on the characteristic function is used, and the continuous approach is compared to the discrete approach. Moreover, we developed two simple algorithms for generating data from two positively correlated and identically distributed random variables having a gamma distribution. The best equivariant estimator for the scale parameter is obtained in the case of independence.

The extension of subcopulas is an important domain. One of possible applications is the nonparametric estimation of a copula: it consists of the smoothing of a subcopula (the empirical copula) while preserving the copulas properties. In Chapter 2, we present an extension of the empirical copula based on the tensor product of splines functions. Our estimators are bona fide estimators of the copula. Chapter 3 tackles the problem of finding all possible extensions of a given subcopula. This subject has been treated in the literature but these characterizations do not apply on very general spaces. Chapter 4 deals with the following problem: finding the expression of the upper bound of the extensions of a finite subcopula in dimension 3.

This thesis focuses on the presentation of the Bernstein estimators which are recent alternatives to conventional estimators of the distribution function and density. More precisely, we study their various properties and compare them with the empirical distribution function and the kernel method estimators. We determine an asymptotic expression of the first two moments of the Bernstein estimator for the distribution function. As the conventional estimators, we show that this estimator satisfies the Chung-Smirnov property under conditions. We then show that the Bernstein estimator is better than the empirical distribution function in terms of mean squared error. We are interested in the asymptotic behavior of Bernstein estimators, for a suitable choice of the degree of the polynomial, we show that the Bernstein estimators are asymptotically normal. Numerical studies on some classical distributions confirm that the Bernstein estimators may be preferable to conventional estimators.

Thèse numérisée par la Division de la gestion de documents et des archives de l'Université de Montréal.

In this master thesis, we consider simulation methods based on antithetic variates for estimate integrales of f(x) on interval (0,1] where f is monotonic function, not a monotonic function or a function difficult to integrate. The main idea consists in subdividing the (0,1] in m sections of which each one is subdivided in l subintervals. This method is done recursively. At each step the variance decreases, i.e. The variance obtained at the kth step is smaller than that is found at the (k-1)th step. This allows us to reduce the error in the estimation because the estimator of integrales of f(x) on interval [0,1] is unbiased. The objective is to optimize m.

Thèse numérisée par la Direction des bibliothèques de l'Université de Montréal.

Thèse numérisée par la Direction des bibliothèques de l'Université de Montréal.

Thèse numérisée par la Direction des bibliothèques de l'Université de Montréal.

Mémoire numérisé par la Direction des bibliothèques de l'Université de Montréal.

Thèse numérisée par la Direction des bibliothèques de l'Université de Montréal.