Consider an estimate $\theta^*$ of a parameter $\theta$ based on repeated observations from a family of densities $f_\theta$ evaluated by the Kullback–Leibler loss function $K(\theta, \theta^*) = \int \log(f_\theta/f_{\theta^*})f_\theta$. The maximum likelihood prior density, if it exists, is the density for which the corresponding Bayes estimate is asymptotically negligibly different from the maximum likelihood estimate. The Bayes estimate corresponding to the maximum likelihood prior is identical to maximum likelihood for exponential families of densities. In predicting the next observation, the maximum likelihood prior produces a predictive distribution that is asymptotically at least as close, in expected truncated Kullback–Leibler distance, to the true density as the density indexed by the maximum likelihood estimate. It frequently happens in more than one dimension that maximum likelihood corresponds to no prior density, and in that case the maximum likelihood estimate is asymptotically inadmissible and may be improved upon by using the estimate corresponding to a least favorable prior. As in Brown, the asymptotic risk for an arbitrary estimate “near” maximum likelihood is given by an expression involving derivatives of the estimator and of the information matrix. Admissibility questions for these “near ML” estimates are determined by the existence of solutions to certain differential equations.

This paper shows that the Bayesian bootstrap (BB) distribution of a multidimensional mean functional based on i.i.d. observations has a strongly unimodal Lebesgue density provided the convex hull of the data has a nonempty interior. This result is then used to construct the finite sample BB credible sets. The influence of an outlier on these credible sets is studied in detail and a comparison is made with the empirical likelihood ratio confidence sets in this context.

If we wish to estimate efficiently the expectation of an arbitrary function on the basis of the output of a Gibbs sampler, which is better: deterministic or random sweep? In each case we calculate the asymptotic variance of the empirical estimator, the average of the function over the output, and determine the minimal asymptotic variance for estimators that use no information about the underlying distribution. The empirical estimator has noticeably smaller variance for deterministic sweep. The variance bound for random sweep is in general smaller than for deterministic sweep, but the two are equal if the target distribution is continuous. If the components of the target distribution are not strongly dependent, the empirical estimator is close to efficient under deterministic sweep, and its asymptotic variance approximately doubles under random sweep.

Advances in Markov chain Monte Carlo MCMC methods now make it computationally feasible and relatively straightforward to apply the Dirichlet process prior in a wide range of Bayesian nonparametric problems. The feasibility of these methods rests heavily on the fact that the MCMC approach avoids direct sampling of the Dirichlet process and is instead based on sampling the finite-dimensional posterior which is obtained from marginalizing out the process.

In application, it is the integrated posterior that is used in the Bayesian nonparametric inference, so one might wonder about its theoretical properties. This paper presents some results in this direction. In particular, we will focus on a study of the posterior’s asymptotic behavior, specifically for the problem when the data is obtained from a finite semiparametric mixture distribution. A complication in the analysis arises because the dimension for the posterior, although finite, increases with the sample size. The analysis will reveal general conditions that ensure exponential posterior consistency for a finite dimensional parameter and which can be slightly generalized to allow the unobserved nonparametric parameters to be sampled from a generalized Pólya urn scheme. Several interesting examples are considered.

The problem of detecting a change in the probability distribution of a random sequence is considered. Stopping times are derived that optimize the tradeoff between detection delay and false alarms within two criteria. In both cases, the detection delay is penalized exponentially rather than linearly, as has been the case in previous formulations of this problem. The first of these two criteria is to minimize a worst-case measure of the exponential detection delay within a lower-bound constraint on the mean time between false alarms. Expressions for the performance of the optimal detection rule are also developed for this case. It is seen, for example, that the classical Page CUSUM test can be arbitrarily unfavorable relative to the optimal test under exponential delay penalty. The second criterion considered is a Bayesian one, in which the unknown change point is assumed to obey a geometric prior distribution. In this case, the optimal stopping time effects an optimal trade-off between the expected exponential detection delay and the probability of false alarm. Finally, generalizations of these results to problems in which the penalties for delay may be path dependent are also considered.

Approximating boundaries using data recorded on a regular grid induces discrete rounding errors in both vertical and horizontal directions. In cases where grid points exhibit at least some degree of randomness, an extensive theory has been developed for local-polynomial boundary estimators. It is inapplicable to regular grids, however. In this paper we impose strict regularity of the grid and describe the performance of local linear estimators in this context. Unlike the case of classical curve estimation problems, pointwise convergence rates vary erratically along the boundary, depending on number-theoretic properties of the boundary’s slope. However, average convergence rates, expressed in the $L_1$ metric, are much less susceptible to fluctuation. We derive theoretical bounds to performance, coming within no more than a logarithmic factor of the optimal convergence rate.

We suggest a three-stage procedure to recover discontinuous regression surfaces when noisy data are present. In the first stage, jump candidate points are detected using a jump detection criterion. A local principal component line is then fitted through these points in a neighborhood of a design point. This line provides a first-order approximation to the true jump location curve in that neighborhood. In the third stage, observations on the same side of the line as the given point are combined using a weighted average procedure to fit the surface at that point. If there are no jump candidate points in the neighborhood, then all observations in that neighborhood are used in the surface fitting. If, however, the center of the neighborhood is on a jump location curve, only those observations on one side of the line are used. Thus blurring is automatically avoided around the jump locations. This methodology requires $O(N(k^*)^2)$ computation, where $N$ is the sample size and $k^*$ is the window width. Its assumptions on the model are flexible. Some numerical results are presented to evaluate the surface fit and to discuss the selection of the window widths.

Consider a space–time stochastic process $Z_ t(x)=S(x) + \xi_t(x)$ where $S(x)$ is a signal process defined on $\mathbb{R}^d$ and $\xi_t(x)$ represents measurement errors at time $t$. For a known measurable function $v(x)$ on $\mathbb{R}^d$ and a fixed cube $ D\subset \mathbb{R}^d$, this paper proposes a linear estimator for the stochastic integral $\int_D v(x)S(x)dx$ based on space–time observations $\{Z_t(x_i); i = 1,\ldots,n; t=1,\ldots,T\}$. Under mild conditions, the asymptotic properties of the mean squared error of the estimator are derived as the spatial distance between spatial sampling locations tends to zero and as time $T$ increases to infinity. Central limit theorems for the estimation error are also studied.

We study integration and reconstruction of Gaussian random functions with inhomogeneous local smoothness. A single realization may only be observed at a finite sampling design and the correct local smoothness is unknown. We construct adaptive two-stage designs that lead to asymptotically optimal methods. We show that every nonadaptive design is less efficient.

Using the approach of finite projective geometry, we make a systematic study of estimation capacity, a criterion of model robustness, under the absence of interactions involving three or more factors. Some general results, providing designs with maximum estimation capacity, are obtained. In particular, for two-level factorials, it is seen that constructing a design with maximum estimation capacity calls for choosing points from a finite projective geometry such that the number of lines is maximized and the distribution of these lines among the chosen points is as uniform as possible. We also explore the connection with minimum aberration designs under which the sizes of the alias sets of two-factor interactions which are not aliased with main effects are the most uniform possible.

Maronna defines affine equivariant $M$-estimators for multivariate location and scatter. They are particularly suited for estimating the pseudo-covariance or scatter matrix of an elliptical population. By defining the bias of a dispersion matrix properly, we consider the maximum bias of an $M$-estimator over an $\varepsilon$ -neighborhood of the underlying elliptical distribution (location known). We find that Tyler’s estimator minimizes the maximum bias.

Consider one-sided testing problems for a multivariate exponential family model. Through conditioning or other considerations, the problem oftentimes reduces to testing a null hypothesis that the natural parameter is a zero vector against the alternative that the natural parameter lies in a closed convex cone $\mathscr{C}$. The problems include testing homogeneity of parameters, testing independence in contingency tables, testing stochastic ordering of distributions and many others. A test methodology is developed that directionalizes the usual test procedures such as likelihood ratio, chi square, Fisher, and so on. The methodology can be applied to families of tests where the family is indexed by a size parameter so as to enable nonrandomized testing by $p$-values. For discrete models, a refined family of tests provides a refined grid for better testing by $p$-values. The tests have essential monotonicity properties that are required for admissibility and for desirable power properties. Two examples are given.

A new test is proposed for the comparison of two regression curves $f$ and $g$. We prove an asymptotic normal law under fixed alternatives which can be applied for power calculations, for constructing confidence regions and for testing precise hypotheses of a weighted $L_2$ distance between $f$ and $g$ . In particular, the problem of nonequal sample sizes is treated, which is related to a peculiar formula of the area between two step functions. These results are extended in various directions, such as the comparison of $k$ regression functions or the optimal allocation of the sample sizes when the total sample size is fixed. The proposed pivot statistic is not based on a nonparametric estimator of the regression curves and therefore does not require the specification of any smoothing parameter.

To obtain test probabilities based on empirical approximations to the distribution of a Studentized function of a mean, we need the approximations to be accurate with sufficiently high probability. In particular, when these test probabilities are small it is best to consider relative errors. Here we show that in the case of univariate standardized means and in the general case of tests based on smooth functions of means, the empirical approximations have asymptotically small relative errors on sets with probability differing from 1 by an exponentially small quantity and that these error rates hold for moderately large deviations. In particular, for standardized deviations of order $n^{1/6}$, the probabilities approximated are exponentially small with exponents of order $n^{1/3}$ and the corresponding relative errors tend to zero on sets whose complements have probabilities of the order of the probabilities being approximated.

A new nonparametric estimation procedure is introduced for the distribution function in a class of deconvolution problems, where the convolution density has one discontinuity. The estimator is shown to be consistent and its cube root asymptotic distribution theory is established. Known results on the minimax risk for the estimation problem indicate the estimator to be efficient.

We consider the classical Wicksell problem of estimating an unknown distribution function $G$of the radii of balls, based on their observed crosssections. It is assumed that the underlying distribution function$G$ belongs to a Hölder class of smoothness $\gamma >1/2$. We prove that, for a suitable choice of the smoothing parameters, kernel-type estimators are asymptotically efficient for a large class of symmetric bowl-shaped loss functions.

The problem of nonparametric function estimation in the Gaussian white noise model is considered. It is assumed that the unknown function belongs to one of the Sobolev classes, with an unknown regularity parameter. Asymptotically exact adaptive estimators of functions are proposed on the scale of Sobolev classes, with respect to pointwise and sup-norm risks. It is shown that, unlike the case of $L_2$-risk, a loss of efficiency under adaptation is inevitable here. Bounds on the value of the loss of efficiency are obtained.