Let $(X, Y)$ be a pair of random variables such that $X$ is $\mathbb{R}^d$-valued and $Y$ is $\mathbb{R}^{d'}$-valued. Given a random sample $(X_1, Y_1), \cdots, (X_n, Y_n)$ from the distribution of $(X, Y)$, the conditional distribution $P^Y(\bullet \mid X)$ of $Y$ given $X$ can be estimated nonparametrically by $\hat{P}_n^Y(A \mid X) = \sum^n_1 W_{ni}(X)I_A(Y_i)$, where the weight function $W_n$ is of the form $W_{ni}(X) = W_{ni}(X, X_1, \cdots, X_n), 1 \leqq i \leqq n$. The weight function $W_n$ is called a probability weight function if it is nonnegative and $\sum^n_1 W_{ni}(X) = 1$. Associated with $\hat{P}_n^Y(\bullet \mid X)$ in a natural way are nonparametric estimators of conditional expectations, variances, covariances, standard deviations, correlations and quantiles and nonparametric approximate Bayes rules in prediction and multiple classification problems. Consistency of a sequence $\{W_n\}$ of weight functions is defined and sufficient conditions for consistency are obtained. When applied to sequences of probability weight functions, these conditions are both necessary and sufficient. Consistent sequences of probability weight functions defined in terms of nearest neighbors are constructed. The results are applied to verify the consistency of the estimators of the various quantities discussed above and the consistency in Bayes risk of the approximate Bayes rules.

If $X_1, \cdots, X_n$ is a random sample from $F(x - \theta)$, where $F$ is an unknown member of a specified class $\mathscr{F}$ of approximately normal symmetric distributions, then an $M$-estimator of the unknown location parameter $\theta$ is obtained by solving the equation $\sum^n_{i=1} \psi(X_i - \hat{\theta}_n) = 0$ for $\hat{\theta}_n$. A suitable measure of the robustness of the $M$-estimator is $\sup \{V(\psi, F): F \in \mathscr{F}\}$, where $V(\psi, F) = \int \psi^2 dF/(\int \psi' dF)^2$ is (under regularity conditions) the asymptotic variance of $n^{\frac{1}{2}}(\hat{\theta}_n - \theta)$. A necessary and sufficient condition for $F_0$ in $\mathscr{F}$ to maximize $V(\psi, F)$ is obtained, and the result is specialized to evaluate $\sup \{V(\psi, F):F \in \mathscr{F}\}$ when the model for $\mathscr{F}$ is the gross errors model or the Kolmogorov model.

Convergence rates for distributions of simple linear rank statistics are investigated. Both the null hypothesis and near alternatives are considered. The method of proof consists in approximating the characteristic function of the statistic by that of a sum of independent random variables and then applying standard tools.

A new approach to the asymptotic normality of simple linear rank statistics for the regression case studied earlier by Hajek (1968) is provided along with the estimation of the remainder term in the approximation to normality.

This paper proves the asymptotic linearity in the regression parameter of a class of linear rank statistics when errors in the regression model are strictly stationary and strongly mixing. Besides this, several other weak convergence results are proved which yield the asymptotic normality of $L$ and $M$ estimators of the regression parameter under the above dependent structure. All these results are useful in studying the effect of the above dependence on the asymptotic behavior of $R, M$ and $L$ estimators vis-a-vis the least squares estimator. An example of linear model with Gaussian errors is given where it is shown that the asymptotic efficiency of certain classes of $R, M$ and $L$ estimators relative to the least squared estimator is greater than or equal to its value under the usual independent errors model.

For the autoregressive-moving average time series model, the normal theory procedure for setting confidence intervals for the error variance is not robust against nonnormality. This paper proposes three asymptotically robust techniques: they are a "standard-error" procedure, an analog of Box's simple data splitting technique, and the jackknife procedure. The large sample distribution of each of these techniques is derived.

Let $T \subseteqq I$ be sets of real numbers. Let $\{Y(t): t \in I\}$ be a real time series whose mean is an unknown element of a known class of functions on $I$ and whose covariance kernel $k$ is assumed known. For each fixed $s \in I, Y(s)$ is predicted by a minimum mean square error unbiased linear predictor $\hat{Y}(s)$ based on $\{Y(t): t \in T\}$. If $\hat{y}(s)$ is the evaluation of $\hat{Y}(s)$ given a set of observations $\{Y(t) = g(t): t \in T\}$, then the function $\hat{y}$ is called a prediction function. Mean-estimation functions are defined similarly. For certain prediction and estimation problems, characterizations are derived for these functions in terms of the covariance structure of the process. Also, relationships between prediction functions and spline functions are obtained that extend earlier results of Kimeldorf and Wahba (Sankhya Ser. A 32 173-180).

Let $\mathbf{\lambda} = (\lambda_1, \cdots, \lambda_n), \lambda_1 \leqq \cdots \leqq \lambda_n$, and $\mathbf{x} = (x_1, \cdots, x_n)$. A function $g(\mathbf{\lambda, x})$ is said to be decreasing in transposition (DT) if (i) $g$ is unchanged when the same permutation is applied to $\mathbf{\lambda}$ and to $\mathbf{x}$, and (ii) $g(\mathbf{\lambda, x}) \geqq g(\mathbf{\lambda, x}')$ whenever $\mathbf{x}'$ and $\mathbf{x}$ differ in two coordinates only, say $i$ and $j, (x_i - x_j) \cdot (i - j) \geqq 0$, and $x_i' = x_j, x_j' = x_i$. The DT class of functions includes as special cases other well-known classes of functions such as Schur functions, totally positive functions of order two, and positive set functions, all of which are useful in many areas including stochastic comparisons. Many well-known multivariate densities have the DT property. This paper develops many of the basic properties of DT functions, derives their preservation properties under mixtures, compositions, integral transformations, etc. A number of applications are then made to problems involving rank statistics.

A gambler seeks to maximize the expected utility earned upon reaching a goal in a game where he is allowed at each stage to stake any amount of his current fortune. He wins each bet with probability $w$. In the discounted case the utility at the goal is $\beta^n$ where $\beta$, the discount factor, is in $(0, 1)$ and $n$ is the number of plays used to reach the goal. In the rapid case the utility at the goal is 1 and the gambler seeks to minimize his expected playing time given he reaches the goal. Here all optimal strategies are characterized when $w \leqq \frac{1}{2}$ for the discounted case and when $w < \frac{1}{2}$ for the rapid case. It is shown that when $w < \frac{1}{2}$ the set of rapidly optimal strategies coincides with the set of optimal strategies for the discounted case.

We show that in the mixed model of the analysis of variance, there is a sequence of roots of the likelihood equations which is consistent, asymptotically normal, and efficient in the sense of attaining the Cramer-Rao lower bound for the covariance matrix. These results follow directly by an application of a general result of Weiss (1971, 1973) concerning maximum likelihood estimates. This problem differs from standard problems in that we do not have independent, identically distributed observations and that estimates of different parameters may require normalizing sequences of different orders of magnitude. We give some examples and comment briefly on likelihood ratio tests for these models.

Let $X$ be an observation from a $p$-variate normal distribution $(p \geqq 3)$ with mean vector $\theta$ and unknown positive definite covariance matrix $\not\Sigma$. It is desired to estimate $\theta$ under the quadratic loss $L(\delta, \theta, \not\Sigma) = (\delta - \theta)^tQ(\delta - \theta)/\operatorname{tr} (Q\not\Sigma)$, where $Q$ is a known positive definite matrix. Estimators of the following form are considered: $\delta^c(X, W) = (I - c\alpha Q^{-1}W^{-1}/(X^tW^{-1}X))X,$ where $W$ is a $p \times p$ random matrix with a Wishart $(\not\Sigma, n)$ distribution (independent of $X$), $\alpha$ is the minimum characteristic root of $(QW)/(n - p - 1)$ and $c$ is a positive constant. For appropriate values of $c, \delta^c$ is shown to be minimax and better than the usual estimator $\delta^0(X) = X$.

For estimating the mean in the one parameter exponential family with quadratic loss, Karlin (1958) gave sufficient conditions for admissibility of estimators of the form $aX$. Later, Ping (1964) and Gupta (1966) gave sufficient conditions for admissibility of estimators of the form $aX + b$ for the same problem. Zidek (1970) gave sufficient conditions for the admissibility of $X$ for estimating an arbitrary piecewise continuous function of the parameter, say $\gamma(\theta)$, not necessarily the mean. In this paper it is shown that Karlin's argument yields sufficient conditions for the admissibility of estimators of the form $aX + b$ for estimating $\gamma(\theta)$. The results are then extended to the case when the parameter space is truncated.

Let $X_1, \cdots, X_n$ be i.i.d. $F$, and estimate the median of $F$ by the median $T_\beta$ of $\beta X_i + (1 - \beta)X_j, i \neq j$, where $\beta$ is a fixed positive constant. Then $T_\beta$ is the solution of a $U$-statistic equation from which its asymptotic normality is readily derived. The asymptotic relative efficiency of $T_\beta$ is computed for a few cdfs $F$ and seen to be reasonably high for unintuitive choices such as $\beta = .9, \beta = 2$, and also to be remarkably constant for $\beta > 1$. Moreover, the influence curves and breakdown points of $\{T_\beta: \beta > 0\}$ are derived and indicate that the good robustness properties of the Hodges-Lehmann estimator $(\beta = \frac{1}{2})$ are shared by the entire class. Monte Carlo estimates of the variance of $T_\beta$ for sample sizes $n = 10, 20$, and 40 indicate that some of these estimators perform as well as those discussed in the Princeton Robustness Study when the underlying $F$ is double-exponential or Cauchy.

The collection of variance-covariance matrices for any linear model may be represented, without altering relationships among linear unbiased estimators, as a compact convex subset of nonnegative definite matrices throughout the relative interior of which all matrices are positive definite.

It is shown that the Fedorov procedure for finding $L$-optimal designs always converges to an $L$-optimal design, even when the limiting design is singular. Two lemmas are given first. An example illustrates the case of singular limiting design.

This paper is concerned with the minimum variance estimation of a time-dependent population mean, assuming that one is restricted to the case of linear unbiased estimators. A number of results are given for a new rotation sampling model (RSM), in which unequal sample sizes are used on each occasion. Also results corresponding to the special case of sampling with a fixed sample size on all the occasions are derived. Finally the optimum structure of the suggested model is discussed and a comparison of this sampling scheme with Patterson's and Eckler's schemes is made.

For the one-sample $t$-test a new form of the exact distribution of the test statistic $t^2$ is obtained when sampling from a distribution which is a mixture of two normal distributions. A numerical example is provided to show that the size of the test can differ greatly when sampling from distributions having the same skewness and kurtosis. Contours of equal size are plotted for a particular case in a certain cross section of the parameter space.

The independence of $U = \min (X, Y)$ and $V = X - Y$ or $W = |X - Y|$ is studied where $X$ and $Y$ are not assumed to be independent. The bivariate exponential distribution of Marshall and Olkin is characterized as the distribution with exponential marginals where $U$ is exponential and independent of $V$.

The von Mises, or circular normal density on the unit circle is the hitting density of a two-dimensional Brownian motion starting at the origin, with constant drift velocity and direction. The concentration and location parameters of the density have a natural relation to the drift parameters.