The method of stochastic discrimination (SD) introduced by Kleinberg is a new method in statistical pattern recognition. It works by producing many weak classifiers and then combining them to form a strong classifier. However, the strict mathematical assumptions in Kleinberg [The Annals of Statistics 24 (1996) 2319-2349] are rarely met in practice. This paper provides an applicable way to realize the SD algorithm. We recast SD in a probability-space framework and present a concrete statistical realization of SD for two-class pattern recognition. We weaken Kleinberg's theoretically strict assumptions of uniformity and indiscernibility by introducing near uniformity and weak indiscernibility. Such weaker notions are easily encountered in practical applications. We present a systematic resampling method to produce weak classifiers and then establish corresponding classification rules of SD. We analyze the performance of SD theoretically and explain why SD is overtraining-resistant and why SD has a high convergence rate. Testing results on real and simulated data sets are also given.

The paper shows how to convert statistical prediction sets into worst case hedging strategies for derivative securities. The prediction sets can, in particular, be ones for volatilities and correlations of the underlying securities, and for interest rates. This permits a transfer of statistical conclusions into prices for options and similar financial instruments. A prime feature of our results is that one can construct the trading strategy as if the prediction set had a 100 % probability. If, in fact, the set has probability $1-\alpha$, the hedging strategy will work with at least the same probability. Different types of prediction regions are considered. The starting value $\A$ for the trading strategy corresponding to the $1-\alpha$ prediction region is a form of long term value at risk. At the same time, $\A$ is coherent.

We introduce a new scatter matrix functional which is a multivariate affine equivariant extension of the mean deviation $E(|x-\mbox{Med}(x)|)$. The estimate is constructed using the data vectors (centered with the multivariate Oja median) and their angular distances. The angular distance is based on Randles interdirections. The new estimate is called the zonoid covariance matrix (the ZCM), as it is the regular covariance matrix of the centers of the facets of the zonotope based on the data set. There is a kind of symmetry between the zonoid covariance matrix and the affine equivariant sign covariance matrix; interchanging the roles of data vectors and hyperplanes yields the sign covariance matrix as the zonoid covariance matrix. It turns out that the symmetry relies on the zonoid of the distribution and its projection body which is also a zonoid.) The influence function and limiting distribution of the new scatter estimate, the ZCM, are derived to consider the robustness and efficiency properties of the estimate. Finite-sample efficiencies are studied in a small simulation study. The influence function of the ZCM is unbounded (linear in the radius of the contamination vector) but less influential in the tails than that of the regular covariance matrix (quadratic in the radius). The estimate is highly efficient in the multivariate normal case and performs better than the regular covariance matrix for heavy-tailed distributions.

A class of projection-based depth functions is introduced and studied. These projection-based depth functions possess desirable properties of statistical depth functions and their sample versions possess strong and order $\sqrt{n}$ uniform consistency. Depth regions and contours induced from projection-based depth functions are investigated. Structural properties of depth regions and contours and general continuity and convergence results of sample depth regions are obtained.

Affine equivariant multivariate medians induced from projection-based depth functions are probed. The limiting distributions as well as the strong and order $\sqrt{n}$ consistency of the sample projection medians are established. The finite sample performance of projection medians is compared with that of a leading depth-induced median, the Tukey halfspace median (induced from the Tukey halfspace depth function). It turns out that, with appropriate choices of univariate location and scale estimators, the projection medians have a very high finite sample breakdown point and relative efficiency, much higher than those of the halfspace median.

Based on the results obtained, it is found that projection depth functions and projection medians behave very well overall compared with their competitors and consequently are good alternatives to statistical depth functions and affine equivariant multivariate location estimators, respectively.

A general Wishart family on a symmetric cone is a natural exponential family (NEF) having a homogeneous quadratic variance function. Using results in the abstract theory of Euclidean Jordan algebras, the structure of conditional reducibility is shown to hold for such a family, and we identify the associated parameterization $\phi$ and analyze its properties. The enriched standard conjugate family for $\phi$ and the mean parameter $\mu$ are defined and discussed. This family is considerably more flexible than the standard conjugate one. The reference priors for $\phi$ and $\mu$ are obtained and shown to belong to the enriched standard conjugate family; in particular, this allows us to verify that reference posteriors are always proper. The above results extend those available for NEFs having a simple quadratic variance function. Specifications of the theory to the cone of real symmetric and positive-definite matrices are discussed in detail and allow us to perform Bayesian inference on the covariance matrix $\Sigma$ of a multivariate normal model under the enriched standard conjugate family. In particular, commonly employed Bayes estimates, such as the posterior expectation of $\Sigma$ and $\Sigma^{-1}$, are provided in closed form.

Suppose that an order restriction is imposed among several p-variate normal mean vectors. We are interested in testing the homogeneity of these mean vectors under this restriction. This problem is a multivariate extension of Bartholomew's [Biometrika} 46 (1959) 36-48]. When the covariance matrices are known, this problem has been studied by Sasabuchi, Inutsuka and Kulatunga [Hiroshima Math. J. 22 (1992) 551-560], Sasabuchi, Kulatunga and Saito [Amer. J. Math. Management Sci. 18 (1998) 131-158] and some others. In the present paper, we consider the case when the covariance matrices are common but unknown. We propose a test statistic, study its upper tail probability under the null hypothesis and estimate its critical points.

In this article, we consider the case when the number of observations n is less than the dimension p of the random vectors which are assumed to be independent and identically distributed as normal with nonsingular covariance matrix. The central and noncentral distributions of the singular Wishart matrix $S=XX'$, where X is the $p \times n$ matrix of observations are derived with respect to Lebesgue measure. Properties of this distribution are given. When the covariance matrix is singular, pseudo singular Wishart distribution is also derived. The result is extended to any distribution of the type $f(XX')$ for the central case. Singular multivariate beta distributions with respect to Lebesgue measure are also given.

Feedforward neural networks, projection pursuit regression, and more generally, estimation via ridge functions have been proposed as an approach to bypass the curse of dimensionality and are now becoming widely applied to approximation or prediction in applied sciences. To address problems inherent to these methods--ranging from the construction of neural networks to their efficiency and capability--Candès [Appl. Comput. armon. Anal. 6 (1999) 197-218] developed a new system that allows the representation of arbitrary functions as superpositions of specific ridge functions, the ridgelets.

In a nonparametric regression setting, this article suggests expanding noisy data into a ridgelet series and applying a scalar nonlinearity to the coefficients (damping); this is unlike existing approaches based on stepwise additions of elements. The procedure is simple, constructive, stable and spatially adaptive--and fast algorithms have been developed to implement it.

The ridgelet estimator is nearly optimal for estimating functions with certain kinds of spatial inhomogeneities. In addition, ridgelets help to identify new classes of estimands--corresponding to a new notion of smoothness--that are well suited for ridge functions estimation. While the results are stated in a decision theoretic framework, numerical experiments are also presented to illustrate the practical performance of the methodology.

In this paper we develop a general theory of local asymptotics for least squares estimates over polynomial spline spaces in a regression problem. The polynomial spline spaces we consider include univariate splines, tensor product splines, and bivariate or multivariate splines on triangulations. We establish asymptotic normality of the estimate and study the magnitude of the bias due to spline approximation. The asymptotic normality holds uniformly over the points where the regression function is to be estimated and uniformly over a broad class of design densities, error distributions and regression functions. The bias is controlled by the minimum $L_\infty$ norm of the error when the target regression function is approximated by a function in the polynomial spline space that is used to define the estimate. The control of bias relies on the stability in $L_\infty$ norm of $L_2$ projections onto polynomial spline spaces. Asymptotic normality of least squares estimates over polynomial or trigonometric polynomial spaces is also treated by the general theory. In addition, a preliminary analysis of additive models is provided.

Consider the regression of a response Y on a vector of quantitative predictors $\X$ and a categorical predictor W. In this article we describe a first method for reducing the dimension of $\X$ without loss of information on the conditional mean $\mathrm{E}(Y|\X,W)$ and without requiring a prespecified parametric model. The method, which allows for, but does not require, parametric versions of the subpopulation mean functions $\mathrm{E}(Y|\X,W=w)$, includes a procedure for inference about the dimension of $\X$ after reduction. This work integrates previous studies on dimension reduction for the conditional mean $\mathrm{E}(Y|\X)$ in the absence of categorical predictors and dimension reduction for the full conditional distribution of $Y|(\X,W)$. The methodology we describe may be particularly useful for constructing low-dimensional summary plots to aid in model-building at the outset of an analysis. Our proposals provide an often parsimonious alternative to the standard technique of modeling with interaction terms to adapt a mean function for different subpopulations determined by the levels of W. Examples illustrating this and other aspects of the development are presented.

In the common trigonometric regression model, we investigate the optimal design problem for the estimation of the individual coefficients, where the explanatory variable varies in the interval $[-a,a]$, $0 <a \le \pi$. It is demonstrated that the structure of the optimal design depends sensitively on the size of the design space. For many important cases, optimal designs can be found explicitly, where the complexity of the solution depends on the value of the parameter a and the order of the term, for which the corresponding coefficient has to be estimated. The main tool of our approach is the reduction of the problem for the trigonometric regression model to a design problem for a polynomial regression. In particular, we determine the optimal designs for estimating the parameters corresponding to the cosine terms explicitly, if the design space is sufficiently small, and prove that under this condition all optimal designs for estimating the parameters corresponding to the sine terms are supported at the same points.