Renormalization arguments are used to derive optimal rates of convergence, under integrated squared error loss, for parameter spaces having a certain rectangular structure.

Consider an unknown regression function $f$ of the response $Y$ on a $d$-dimensional measurement variable $X$. It is assumed that $f$ belongs to a class of functions having a smoothness measure $p$. Let $T$ denote a known linear operator of order $q$ which maps $f$ to another function $T(f)$ in a space $G$. Let $\hat{T}_n$ denote an estimator of $T(f)$ based on a random sample of size $n$ from the distribution of $(X, Y)$, and let $\|\hat{T}_n - T(f)\|_G$ be a norm of $\hat{T}_n - T(f)$. Under appropriate regularity conditions, it is shown that the optimal rate of convergence for $\|\hat{T}_n - T(f)\|_G$ is $n^{-(p - q)/(2p + d)}$. The result is applied to differentiation, fractional differentiation and deconvolution.

In this paper, a method for finding global minimax lower bounds is introduced. The idea is to adjust automatically the direction of a local one-dimensional subproblem at each location to the nearly hardest one, and to use locally the difficulty of the one-dimensional subproblem. This method has the advantages of being easily implemented and understood. The lower bound is then applied to nonparametric deconvolution to obtain the optimal rates of convergence for estimating a whole function. Other applications are also addressed.

This paper concerns self-consistent estimators for survival functions based on doubly censored data. We establish strong uniform consistency, asymptotic normality and asymptotic efficiency of the estimators under mild conditions on the distributions of the censoring variables.

Consider the problem of estimating $\mu$, based on the observation of $Y_0, Y_1, \ldots, Y_n$, where it is assumed only that $Y_0, Y_1, \ldots, Y_\kappa \operatorname{iid} N(\mu, \sigma^2)$ for some unknown $\kappa$. Unlike the traditional change-point problem, the focus here is not on estimating $\kappa$, which is now a nuisance parameter. When it is known that $\kappa = k$, the sample mean $\bar{Y}_k = \sum^k_0Y_i/(k + 1)$, provides, in addition to wonderful efficiency properties, safety in the sense that it is minimax under squared error loss. Unfortunately, this safety breaks down when $\kappa$ is unknown; indeed if $k > \kappa$, the risk of $\bar{Y}_k$ is unbounded. To address this problem, a generalized minimax criterion is considered whereby each estimator is evaluated by its maximum risk under $Y_0, Y_1, \ldots, Y_\kappa \operatorname{iid} N(\mu, \sigma^2)$ for each possible value of $\kappa$. An essentially complete class under this criterion is obtained. Generalizations to other situations such as variance estimation are illustrated.

Pollak and Siegmund compared the Shiryayev-Roberts procedure with the CUSUM procedure for detecting a change in the drift of a Brownian motion based on the conditional average delay time. In this paper, the exponentially weighted moving average (EWMA) procedure proposed by Roberts is compared with the Shiryayev-Roberts and CUSUM procedures. The comparison is based on the stationary average delay time as advocated by Shiryayev. The optimal design for the EWMA procedure and its asymptotic properties are studied when the average in-control run length is large. The results show that the EWMA procedure is less efficient than the other two procedures.

We compute the asymptotic distribution of the maximum likelihood ratio test when we want to check whether the parameters of normal observations have changed at an unknown point. The proof is based on the limit distribution of the largest deviation between a $d$-dimensional Ornstein-Uhlenbeck process and the origin.

There has been an increasing interest in modelling regression with heavy-tailed conditional error distributions, mostly in the parametric setting. Nonparametric regression procedures have been studied almost exclusively for the cases where the conditional variance of the regressed variable is finite in the region of interest. We initiate a study of the infinite variance case. Some results in strong uniform consistency of the nearest neighbor estimator with rates are proven. The technique used provides new results and insights when higher conditional moments exist. Some asymptotic distribution theory has also been obtained when the conditional errors are in the domain of attraction of a stable law.

Optimal local smoothing of a curve estimator requires knowledge of various derivatives of the curve in the neighbourhood of the point at which estimation is being conducted. One empirical approach to selecting the amount of smoothing is to employ pilot estimators to approximate those derivatives, and substitute the approximate values into an analytical formula for the desired local bandwidth. In the present paper we study how bandwidth choice for the pilot estimators affects the performance of the final estimator. Our conclusions are rather curious. Depending on circumstance, the pilot estimators should be substantially oversmoothed or undersmoothed, relative to the amount of smoothing that would be optimal if they were to be employed themselves for point estimation. Occasionally, the optimal amount of undersmoothing is so extreme as to render the pilot estimators inconsistent. Here, the resulting local bandwidth is asymptotically random; it is not asymptotic to a sequence of constants.

This paper investigates the density estimation problem in the $L^\infty$ norm for dependent data. It is shown that the iid optimal minimax rates are also optimal for smooth classes of stationary sequences satisfying certain $\beta$-mixing (or absolutely regular) conditions. Moreover, for given $\beta$-mixing coefficients, bounds on uniform convergence rates of kernel estimators are computed in terms of the mixing coefficients. The rates and the bounds obtained are not only for estimating the density but also for its derivatives. The results are then applied to give uniform convergence rates in problems associated with the Gibbs sampler.

This paper obtains some extensions of Gilliland and Hannan's results on equivariance and the compound decision problem. Consider a compound decision problem with restricted component risk and component distributions in a norm compact set of mutually absolutely continuous probability measures. Then the method of proof of a theorem of Gilliland and Hannan translates the results of Mashayekhi on symmetrization of product measures into uniform convergence to zero of the excess of the simple envelop over the equivariant envelope. Our envelope results strengthen, among other things, the results of Datta who obtained admissible asymptotically optimal solutions to the compound estimation problem for a large subclass of the real one parameter exponential family under squared error loss. Sufficient conditions for asymptotic optimality of "delete bootstrap" rules are given and, for squared error loss estimation of continuous functions and for finite action space problems with continuous loss functions, the problem of treating the asymptotic excess compound risk of Bayes compound rules is reduced to the question of $L_1$-consistency of certain mixtures. Examples of estimates satisfying the above consistency condition are provided.

Consider the model where $X_{ij}, i = 1,\ldots, k; j = 1,2,\ldots, n_i; n_i \geq 2$, are observed. Here $X_{ij}$ are independent $N(\theta_i,\sigma^2), \theta_i, \sigma^2$ unknown. Let $X_i = \sum^n_{j = 1}X_{ij}/n_i, \mathbf{X}' = (X_1,\ldots, X_k), \mathbf{\theta}' = (\theta_1,\ldots,\theta_k), V = \sum^k_{i = 1} \sum^{n_i}_{j = 1}X^2_{ij} - n \sum^k_{i = 1}X^2_i$. Let $\mathbf{A}_1$ be a $(k - m) \times k$ matrix of rank $(k - m) \geq 2$ and test $H: \mathbf{A}_1\mathbf{\theta} = \mathbf{0}$ versus $K - H$ where $K: \mathbf{A}_1\mathbf{\theta} \geq \mathbf{0}$. Suppose we assume $\sigma^2$ known and consider a constant size $\alpha$ test $(\alpha < 1/2)$ which is admissible for $H$ versus $K - H$ based on $\mathbf{X}$. Next assume $\sigma^2$ is unknown. Consider the same test but now as a function of $\mathbf{X}/V^{1/2}$ (i.e., Studentize the test). The resulting test is inadmissible. Examples are noted.

We consider a model robust version of the $c$-optimality criterion minimizing a weighted product with factors corresponding to the variances of the least squares estimates for linear combinations of the parameters in different models. A generalization of Elfving's theorem is proved for the optimal designs with respect to this criterion by an application of an equivalence theorem for mixtures of optimality criteria. As a special case an Elfving theorem for the $D$-optimal design problem is obtained. In the case of identical models the connection between the $A$-optimality criterion and the model robust criterion is investigated. The geometric characterizations of the optimal designs are illustrated by a couple of examples.

In a recent paper Pukelsheim and Studden determined the $E$-optimal design for the polynomial regression model on the interval $\lbrack - 1, 1\rbrack$ where the variances of different observations are assumed to be constant. In this note we show that these results can be generalized for polynomial regression models with non constant variances proportional to specific functions.

Suppose that $X_1,\ldots,X_n$ are i.i.d. $\sim F$, and we wish to test the null hypothesis that $F$ is a member of the parametric family $\mathscr{F} = \{F_\theta(x); \theta \in \Theta\}$ where $\Theta \subset \mathbb{R}^q$. The classical Pearson-Fisher chi-square test involves partitioning the real axis into $k$ cells $I_1,\ldots, I_k$ and forming the chi-square statistic $X^2 = \sum^k_{i = 1}(O_i - nF_{\hat{\theta}}(I_i))^2/nF_{\hat{\theta}}(I_i)$, where $O_i$ is the number of observations falling into cell $i$ and $\hat{\theta}$ is the value of $\theta$ minimizing $\sum^k_{i = 1}(O_i - nF_\theta(I_i))^2/nF_\theta(I_i)$. We obtain a generalization of this test to any situation for which there is available a nonparametric estimator $\hat{F}$ of $F$ for which $n^{1/2}(\hat{F} - F) \rightarrow_d W$, where $W$ is a continuous zero mean Gaussian process satisfying a mild regularity condition. We allow the cells to be data dependent. Essentially, we estimate $\theta$ by the value $\hat{\theta}$ that minimizes a "distance" between the vectors $(\hat{F}(I_1),\ldots,\hat{F}(I_k))$ and $(F_\theta(I_1),\ldots, F_\theta(I_k))$, where distance is measured through an arbitrary positive definite quadratic form, and then form a chi-square type test statistic based on the difference between $(\hat{F}(I_1),\ldots,\hat{F}(I_k))$ and $(F_{\hat{\theta}}(I_1),\ldots, F_{\hat{\theta}}(I_k))$. We prove that this test statistic has asymptotically a chi-square distribution with $k - q - 1$ degrees of freedom, and point out some errors in the literature on chi-square tests in survival analysis. Our procedure is very general and applies to a number of well-known models in survival analysis, such as right censoring and left truncation. We apply our method to deal with questions of model selection in the problem of estimating the distribution of the length of the incubation period of the AIDS virus using the CDC's data on blood-transfusion related AIDS. Our analysis suggests some models that seem to fit better than those used in the literature.

This paper is mainly devoted to the following statistical problem: in the case of random variables of any finite dimension and both simple or parametric hypotheses, how to construct convenient "empirical" processes which could provide the basis for goodness of fit tests-more or less in the same way as the uniform empirical process does in the case of simple hypothesis and scalar random variables. The solution of this problem is connected here with the theory of multiparameter martingales and the theory of function-parametric processes. Namely, for the limiting Gaussian processes some kind of filtration is introduced and so-called scanning innovation processes are constructed-the adapted standard Wiener processes in one-to-one correspondence with initial Gaussian processes. This is done for the function-parametric versions of the processes.

The spectral distribution function of a stationary stochastic process standardized by dividing by the variance of the process is a linear function of the autocorrelations. The integral of the sample standardized spectral density (periodogram) is a similar linear function of the autocorrelations. As the sample size increases, the difference of these two functions multiplied by the square root of the sample size converges weakly to a Gaussian stochastic process with a continuous time parameter. A monotonic transformation of this parameter yields a Brownian bridge plus an independent random term. The distributions of functionals of this process are the limiting distributions of goodness of fit criteria that are used for testing hypotheses about the process autocorrelations. An application is to tests of independence (flat spectrum). The characteristic function of the Cramer-von Mises statistic is obtained; inequalities for the Kolmogorov-Smirnov criterion are given. Confidence regions for unspecified process distributions are found.

We place shape theory in the setting of noncentral multivariate analysis, and thus provide a comprehensive view of shape distributions when landmark coordinates are Gaussian distributed. This work allows the statistical analysis of shape to be carried out using standard techniques of multivariate analysis. The paper includes some new results in all dimensions and a general Gaussian approximation to the size-and-shape distribution. We also discuss some inference problems and give a numerical example.

This paper studies the shapes of low dimensional projections from high dimensional data. After standardization, let $\mathbf{x}$ be a $p$-dimensional random variable with mean zero and identity covariance. For a projection $\beta'\mathbf{x}, \|\beta\| = 1$, find another direction $b$ so that the regression curve of $b'\mathbf{x}$ against $\beta'\mathbf{x}$ is as nonlinear as possible. We show that when the dimension of $\mathbf{x}$ is large, for most directions $\beta$ even the most nonlinear regression is still nearly linear. Our method depends on the construction of a pair of $p$-dimensional random variables, $\mathbf{w}_1, \mathbf{w}_2$, called the rotational twin, and its density function with respect to the standard normal density. With this, we are able to obtain closed form expressions for measuring deviation from normality and deviation from linearity in a suitable sense of average. As an interesting by-product, from a given set of data we can find simple unbiased estimates of $E(f_{\beta'\mathbf{x}}(t)/\phi_1(t) - 1)^2$ and $E\lbrack (\|E(\mathbf{x} \mid \beta, \beta'\mathbf{x} = t)\|^2 - t^2)f^2_{\beta'\mathbf{x}}(t)/\phi^2_1(t)\rbrack$, where $\phi_1$ is the standard normal density, $f_{\beta'\mathbf{x}}$ is the density for $\beta'\mathbf{x}$ and the $"E"$ is taken with respect to the uniformly distributed $\beta$. This is achieved without any smoothing and without resorting to any laborious projection procedures such as grand tours. Our result is related to the work of Diaconis and Freedman. The impact of our result on several fronts of data analysis is discussed. For example, it helps establish the validity of regression analysis when the link function of the regression model may be grossly wrong. A further generalization, which replaces $\beta'\mathbf{x}$ by $B'\mathbf{x}$ with $B = (\beta_1,\ldots, \beta_k)$ for $k$ randomly selected orthonormal vectors $(\beta_i, i = 1,\ldots, k)$, helps broaden the scope of application of sliced inverse regression (SIR).

For experiments where the strength of association between a response variable $Y$ and a covariate $X$ is different over different regions of values for the covariate $X$, we propose local nonparametric dependence functions which measure the strength of association between $Y$ and $X$ as a function of $X = x$. Our dependence functions are extensions of Galton's idea of strength of co-relation from the bivariate normal case to the nonparametric case. In particular, a dependence function is obtained by expressing the usual Galton-Pearson correlation coefficient in terms of the regression line slope $\beta$ and the residual variance $\sigma^2$ and then replacing $\beta$ and $\sigma^2$ by a nonparametric regression slope $\beta(x)$ and a nonparametric residual variance $\sigma^2(x) = \operatorname{var}(Y \mid x)$, respectively. Our local dependence functions are standardized nonparametric regression curves which provide universal scale-free measures of the strength of the relationship between variables in nonlinear models. They share most of the properties of the correlation coefficient and they reduce to the usual correlation coefficient in the bivariate normal case. For this reason we call them correlation curves. We show that, in a certain sense, they quantify Lehmann's notion of regression dependence. Finally, the correlation curve concept is illustrated using data from a study of the relationship between cholesterol levels $x$ and triglyceride concentrations $y$ of heart patients.

The observation model $y_i = \beta(i/n) + \varepsilon_i, 1 \leq i \leq n$, is considered, where the $\varepsilon$'s are i.i.d. with mean zero and variance $\sigma^2$ and $\beta$ is an unknown smooth function. A Gaussian prior distribution is specified by assuming $\beta$ is the solution of a high order stochastic differential equation. The estimation error $\delta = \beta - \hat{\beta}$ is analyzed, where $\hat{\beta}$ is the posterior expectation of $\beta$. Asymptotic posterior and sampling distributional approximations are given for $\|\delta\|^2$ when $\|\cdot\|$ is one of a family of norms natural to the problem. It is shown that the frequentist coverage probability of a variety of $(1 - \alpha)$ posterior probability regions tends to be larger than $1 - \alpha$, but will be infinitely often less than any $\varepsilon > 0$ as $n \rightarrow \infty$ with prior probability 1. A related continuous time signal estimation problem is also studied.

Suppose that $(X_1, Y_1),\cdots, (X_n, Y_n)$ are i.i.d. bivariate random vectors and that $\xi_p(x)$ is the $p$-quantile of $Y_1$ given $X_1 = x$ for $0 < p < 1$. Estimation of $\xi_p(x)$, when it is monotone in $x$, has been studied in the literature. In the nonparametric conditional quantile estimation one uses only some smoothness assumptions. The asymptotic properties are superior in the latter case; however, monotonicity is not guaranteed. We introduce a new estimator that enjoys both of the above properties.

The background of the paper is the empirical observation from a variety of subject areas that long-range correlations appear to be much more frequent than has been previously assumed. This includes high-quality measurement series which are commonly treated as prototypes of "i.i.d." observations. Evidence is briefly cited in the paper. It has already been shown elsewhere that long-range dependence leads to results that can be qualitatively different from those obtained under short-range dependence, and in particular, that long-range dependence has drastic effects on the naive statistical treatment of absolute constants. The natural question arising from this, also of relevance for statistical practice, is how the long-range dependence affects the statistics for contrasts. The main answer given in this paper is twofold. (i) If the experimental conditions are well mixed as provided by randomization, the levels of tests and confidence intervals derived under the independence assumption are still correct, asymptotically and usually in good approximation for finite samples. (ii) Even under randomly mixed designs there are typically large unnoticed power and efficiency losses due to the long-range dependence. They can be greatly reduced without estimating the correlations by a simple blocking device.

A new class of bias-robust estimates of multiple regression is introduced. If $y$ and $x$ are two real random variables, let $T(y, x)$ be a univariate robust estimate of regression of $y$ on $x$ through the origin. The regression estimate $\mathbf{T}(y, \mathbf{x})$ of a random variable $y$ on a random vector $\mathbf{x} = (x_1,\cdots, x_p)'$ is defined as the vector $\mathbf{t} \in \mathfrak{R}^p$ which minimizes $\sup_{\|\mathbf{\lambda}\| = 1} \mid T(y - \mathbf{t'x, \lambda' x}) \mid s(\mathbf{\lambda'x})$, where $s$ is a robust estimate of scale. These estimates, which are called projection estimates, are regression, affine and scale equivariant. When the univariate regression estimate is $T(y, x) =$ median $(y/x)$, the resulting projection estimate is highly bias-robust. In fact, we find an upper bound for its maximum bias in a contamination neighborhood, which is approximately twice the minimum possible value of this maximum bias for any regression and affine equivariant estimate. The maximum bias of this estimate in a contamination neighborhood compares favorably with those of Rousseeuw's least median squares estimate and of the most bias-robust GM-estimate. A modification of this projection estimate, whose maximum bias for a multivariate normal with mass-point contamination is very close to the minimax bound, is also given. Projection estimates are shown to have a rate of consistency of $n^{1/2}$. A computational version of these estimates, based on subsampling, is given. A simulation study shows that its small sample properties compare very favorably to those of other robust regression estimates.

In this paper we consider the problem of robust estimation of the scale of the location residuals when the underlying distribution of the data belongs to a contamination neighborhood of a parametric location-scale family. We define the class of $M$-estimates of scale with general location, and show that under certain regularity assumptions, these scale estimates converge to their asymptotic functionals uniformly with respect to the underlying distribution, and with respect to the $M$-estimate defining score function $\chi$. We establish expressions for the maximum asymptotic bias of $M$-estimates of scale over the contamination neighborhood as a function of the fraction of contamination. Using these expressions we construct asymptotically min-max bias robust estimates of scale. In particular, we show that a scaled version of the Madm (median of absolute residuals about the median) is approximately min-max bias-robust within the class of Huber's Proposal 2 joint estimates of location and scale. We also consider the larger class of $M$-estimates of scale with general location, and show that a scaled version of the Shorth (the shortest half of the data) is approximately min-max bias robust in this class. Finally, we present the results of a Monte Carlo study showing that the Shorth has attractive finite sample size mean squared error properties for contaminated Gaussian data.

A variety of exponential models with affine dual foliations have been noted to possess certain rather similar statistical properties. To give a precise meaning to what has been conceived as "similar," we here propose a set of five conditions, of a differential geometric/statistical nature, that specify the class of what we term orthogeodesic models. It is discussed how these conditions capture the properties in question, and it is shown that some important nonexponential models turn out to satisfy the conditions, too. The conditions imply, in particular, a higher-order asymptotic independence result. A complete characterization of the structure of exponential orthogeodesic models is derived.

Let $(\mathbf{X}_t, Y_t)$ be a stationary time series with $\mathbf{X}_t$ being $R^d$-valued and $Y_t$ real valued, and where $Y_t$ is not necessarily bounded. Let $E(Y_0 \mid \mathbf{X}_0)$ be the conditional mean function. Under appropriate regularity conditions, local average estimators of this function can be chosen to achieve the optimal rate of convergence $(n^{-1} \log n)^{1/(d + 2)}$ in $L_\infty$ norm restricted to a compact. The result answers a question raised by Truong and Stone.

The limiting distribution of the normalized autocorrelation coefficient in the case of a unit root is given in a closed form. It is found that high order transcendental functions such as the parabolic cylinder functions are indispensable to express this distribution, thus departing from the simple standard normal distribution that arises in the case of a stable root. Using the formulae derived in this paper, some numerical results available from previous studies are then extended and refined. Finally, the formulae are manipulated analytically to explain the unusual shape of the distribution.

In a heteroscedastic linear model, we establish the asymptotic normality of the iterative weighted least squares estimators with weights constructed by using the within-group residuals obtained from the previous model fitting. An adaptive procedure is proposed which ensures that the iterative process stops after a finite number of iterations and produces an estimator asymptotically equivalent to the best estimator that can be obtained by using the iterative procedure. Theoretical and empirical results of the performance of the adaptive estimator are presented.

For small sample random effects models, results are derived which show in certain cases, and indicate in general, that an estimated random effects variance matrix may be used in the weight matrices without causing undue error in the empirically weighted mean. Exact error variances are derived mathematically for the empirically weighted mean for the two sample case in one and two dimensions. Simulation is used to determine errors for a practical example of six five-variate samples. For estimation of their mean, the differences between the samples are ancillary. The biases of the average and weighted mean estimators conditional on these ancillaries is illustrated in a diagram plotting values obtained by simulation. A curious range anomaly is illustrated which arises if random effects are ignored when present.