Sufficient conditions are given for the uniqueness of intrinsic and extrinsic means as measures of location of probability measures Q on Riemannian manifolds. It is shown that, when uniquely defined, these are estimated consistently by the corresponding indices of the empirical $\hat Q_n$. Asymptotic distributions of extrinsic sample means are derived. Explicit computations of these indices of $\hat Q_n$ and their asymptotic dispersions are carried out for distributions on the sphere $S^d$ (directional spaces), real projective space $\mathbb{R}P^{N-1}$ (axialspaces) and $\mathbb{C} P^{k-2}$ (planar shape spaces).

A new method of pointwise adaptation has been proposed and studied in Spokoiny [(1998) Ann. Statist.26 1356-1378] in the context of estimation of piecewise smooth univariate functions. The present paper extends that method to estimation of bivariate grey-scale images composed of large homogeneous regions with smooth edges and observed with noise on a gridded design. The proposed estimator $\hat{f}(x)$ at a point x is simply the average of observations over a window $\widehat{U}(x)$ selected in a data-driven way. The theoretical properties of the procedure are studied for the case of piecewise constant images. We present a nonasymptotic bound for the accuracy of estimation at a specific grid point x as a function of the number of pixels n, of the distance from the point of estimation to the closest boundary and of smoothness properties and orientation of this boundary. It is also shown that the proposed method provides a near-optimal rate of estimation near edges and inside homogeneous regions. We briefly discuss algorithmic aspects and the complexity of the procedure. The numerical examples demonstrate a reasonable performance of the method and they are in agreement with the theoretical issues. An example from satellite (SAR) imaging illustrates the applicability of the method.

Thresholding algorithms in an orthonormal basis are studied to estimate noisy discrete signals degraded by a linear operator whose inverse is not bounded. For signals in a set $\Theta$, sufficient conditions are established on the basis to obtain a maximum risk with minimax rates of convergence. Deconvolutions with kernelshaving a Fourier transform which vanishes at high frequencies are examples of unstable inverse problems, where a thresholding in a wavelet basis is a suboptimal estimator. A new "mirror wavelet" basis is constructed to obtain a deconvolution risk which is proved to be asymptotically equivalent to the minimax risk over bounded variation signals. This thresholding estimator is used to restore blurred satellite images.

For various types of noise (exponential, normal mixture, compactly supported, ...) wavelet thresholding methods are studied. Problems linked to the existence of optimal thresholds are tackled, and minimaxity properties of the methods also analyzed. A coefficient dependent method for choosing thresholds is also briefly presented.

Additive regression models have turned out to be useful statistical tools in the analysis of high-dimensional data. The attraction of such models is that the additive component can be estimated with the same optimal convergence rate as a one-dimensional nonparametric regression. However, this optimal property holds only when all the additive components have the same degree of "homogeneous" smoothness. In this paper, we propose a two-step wavelet thresholding estimation process in which the estimator is adaptive to different degrees of smoothness in different components and also adaptive to the "inhomogeneous" smoothness described by the Besov space. The estimator of an additive component constructed by the proposed procedure is shown to attain theone-dimensional optimal convergence rate even when the components have different degrees of "inhomogeneous" smoothness.

We present a general approach to estimating probability measures constrained to lie in a convex set. We represent constrained measures as mixtures of simple, known extreme measures, and so the problem of estimating a constrained measure becomes one of estimating an unconstrained mixing measure. Convex constraints arise in many modeling situations, such as estimation of the mean and estimation under stochastic ordering constraints. We describe mixture representation techniques for these and other situations, and discuss applications to maximum likelihood and Bayesian estimation.

Suppose k-variate data are drawn from a mixture of two distributions, each having independent components. It is desired to estimate the univariate marginal distributions in each of the products, as well as the mixing proportion. This is the setting of two-class, fully parametrized latent models that has been proposed for estimating the distributions of medical test results when disease status is unavailable. The problem is one of inference in a mixture of distributions without training data, and until now it has been tackled only in a fully parametric setting. We investigate the possibility of using nonparametric methods. Of course, when k=1 the problem is not identifiable from a nonparametric viewpoint. We show that the problem is "almost" identifiable when k=2; there, the set of all possible representations can be expressed, in terms of any one of those representations, as a two-parameter family. Furthermore, it is proved that when $k\geq3$ the problem is nonparametrically identifiable under particularly mild regularity conditions. In this case we introduce root-n consistent nonparametric estimators of the 2k univariate marginal distributions and the mixing proportion. Finite-sample and asymptotic properties of the estimators are described.

We propose a new test, based on model selection methods, for testing that the expectation of a Gaussian vector with n independent components belongs to a linear subspace of $\R^{n}$ against a nonparametric alternative. The testing procedure is available when the variance of the observations is unknown and does not depend on any prior information on the alternative. The properties of the test are nonasymptotic and we prove that the test is rate optimal [up to a possible log(n factor] over various classes of alternatives simultaneously. We also provide a simulation study in order to evaluate the procedure when the purpose is to test goodness-of-fit in a regression model.

Model selection using a penalized data-splitting device is studied in the context of nonparametric regression. Finite sample bounds under mild conditions are obtained. The resulting estimates are adaptive for large classes of functions.

This paper presents a saddlepoint approximation to the cumulative distribution function of a random vector. The proposed approximation has accuracy comparable to that of existing expansions valid in twodimensions, and may be applied to random vectors of arbitrary length, subject only to the requirement that the distribution approximated either have a density or be confined to a lattice, and have a cumulant generating function. The result is derived by directly inverting the multivariate moment generating function. The result is applied to sufficient statistics from a regression model with exponential errors, and compared to an existing method in two dimensions. The result is also applied to multivariate inference from a data set arising from a case-control study of endometrial cancer.

Let V be the space of (r,r) Hermitian matrices and let $\Omega$ be the cone of the positive definite ones. We say that the random variable S, taking its values in $\overline{\Omega},$ has the complex Wishart distribution $\gamma_{p,\sigma}$ if $\mathbb{E}(\exp \,\tr (\theta S))=(\det (I_r-\sigma\theta))^{-p},$ where $\sigma$ and $\sigma^{-1}-\theta$ are in $\Omega,$ and where p=1,2,...,r-1 or p>r-1. In this paper, we compute all moments of $S$ and $S^{-1}.$ The techniques involve in particular the use of the irreducible characters of the symmetric group.

Generalized estimating equations are used in regression analysis of longitudinal data, where observations on each subject are correlated. Statistical analysis using such methods is based on the asymptotic properties of regression parameter estimators. This paper presents asymptotic results when either the number of independent subjects or the cluster sizes (the number of observations on each subject) or both go to infinity. A set of (information matrix based) general conditions is developed, which leads to the weak and strong consistency as well as the asymptotic normality of the estimators. Most of the results are parallel to the elegant work of Fahrmeir and Kaufmann on maximum likelihood estimators related to the generalized linear models. The conditions for weak consistency and asymptotic normality are verified for several examples of general interest.