The least squares estimate of the parameter matrix $\mathbf{B}$ in the model $\mathbf{y}_t = \mathbf{B'x}_t + \mathbf{u}_t$, where $\mathbf{u}_t$ is an $m$-component vector of unobservable disturbances and $x_t$ is a $p$-component vector, converges to $\mathbf{B}$ with probability one under certain conditions on the behavior of $x_t$ and $\mathbf{u}_t$. When $\mathbf{x}_t$ is stochastic and the conditional expectation of $\mathbf{u}_t$ given $\mathbf{x}_s$ for $s \leqslant t$ and $\mathbf{u}_t$ for $s < t$ is zero, then the least squares estimates are strongly consistent if the inverse of $\mathbf{A}_T = \sigma^T_{t=1} \mathbf{x}_t\mathbf{x}'_t$, where $T$ is the sample size, converges to the zero matrix and if the ratio of the largest to the smallest characteristic root of $\mathbf{A}_T$ is bounded with probability one.

A general finite parameter model for stationary ergodic nondeterministic vector time series is considered. A central limit theorem for parameter estimates, obtained by maximising frequency domain approximations to the Gaussian likelihood, is established. The treatment given extends the central limit theorem of Dunsmuir and Hannan in that the innovations covariance matrix and the linear transfer function need not be separately parameterised. Models for a stationary vector signal observed with stationary vector noise are discussed in relation to the central limit theorem and the conditions imposed for this result are related to this model. Finally, the special case of a scalar autoregressive signal observed with noise is discussed. It is shown that this model may be reparameterised so that the central limit theorem of Dunsmuir and Hannan may be applied.

An infinite distributed lag system is approximated by a truncated one, which can be consistently estimated. We investigate the rate at which the parameter space in the approximate model can be increased as sample size increases, in such a way that these estimates will provide consistent estimates of the underlying system.

Estimation as a reduction of data is usually accompanied by some loss of information. This paper theoretically compares asymptotically efficient estimation methods for parameters in Gaussian linear processes. By means of the concept of "asymptotic information loss" suitably defined, estimates equivalent to the order of $N^{-\frac{1}{2}}$ are differentiated. This problem was studied by C. R. Rao for multinomial distributions and by K. Takeuchi for the exponential family of distributions. They showed that for the i.i.d. case the maximum likelihood estimate is superior to other efficient estimates. This paper extends their results to the Whittle-Walker model of Gaussian linear processes, demonstrating the optimality of the maximum likelihood estimate for that model. In addition, the paper contains a lemma of independent interest. The Craig-Aitken theorem is concerned with the independence of two quadratic forms of a finite-dimensional Gaussian random vector; the theorem is extended to infinite-dimensional Gaussian random vectors.

If the finite population is homogeneous and the statistician's resources permit him to take a sample size $m$ at the most, it is shown here that his best strategy for estimating the population mean is to draw a sample of size $m$ by simple random sampling without replacement and to take the sample mean as the estimate. The strategy is the best in the sense that in the entire class of unbiased strategies subject to the restriction that the size of any observed sample does not exceed $m$, this strategy minimizes for any convex loss function both the maximum and the average risks over the set of parameter points arising from permutations of the labels of the population units. Similar decision-theoretic justification is also derived for the customary strategy for a two-stage cluster-sampling design.

R. L. Anderson proposed a $\chi^2$-type rank statistic for the nonparametric analysis of a randomized blocks design. In this paper the asymptotic distribution of the test statistic is derived under a sequence of alternatives contiguous to the null hypothesis. Using Bahadur's concept of local approximate slope, it is shown that the test is optimal within a large class of rank tests including the tests proposed by Friedman and by Page. The results are extended to BIB designs. Ties are considered.

Let $(X_1, Y_1),\cdots, (X_N, Y_N)$ be i.i.d. rv's where the $X$'s are nonnegative integer-valued. Conditional on $\sigma X_k$ the asymptotic distribution of $\sigma Y_k$ and $\sigma a_k X_k$ are derived by general methods. Some applications are briefly discussed: sampling without replacement, the classical occupancy problem, the Wilcoxon statistic, the Poisson index of dispersion, testing geometric versus Poisson distribution.

The theory of Dirichlet processes is applied to the empirical Bayes estimation problem in the binomial case. The approach is Bayesian rather than being empirical Bayesian. When the prior is a Dirichlet process the posterior is a mixture of Dirichlet processes. Explicit estimators are given for the case of 2 and 3 parameters and compared with other empirical Bayes estimators by way of examples. Since the number of calculations become enormous when the number of parameters gets larger than 2 or 3 we propose two approximations for estimators of a particular parameter and compare their performance using examples.

Questions pertaining to the admissibility of fixed sample size tests of hypotheses, when sequential tests are available, are considered. For the normal case with unknown mean, suppose the risk function is a linear combination of probability of error and expected sample size. Then any fixed sample size test, with sample size $n \geqslant 2$, is inadmissible. On the other hand, suppose the risk function consists of the pair of components, probability of error and expected sample size. Then any optimal fixed sample size test for the one sided hypothesis is admissible. When the variance of the normal distribution is unknown, $t$-tests are studied. For one-sided hypotheses and componentwise risk functions the fixed sample size $t$-test is inadmissible if and only if the absolute value of the critical value of the test is greater than or equal to one. This implies that for the most commonly used sizes, the fixed sample size $t$-test is inadmissible. Other loss functions are discussed. Also an example for a normal mean problem is given where a nonmonotone test cannot be improved on by a monotone test when the risk is componentwise.

This paper deals with approximations to Bayesian sequential tests of composite hypotheses. If the distributions of the data form an exponential or truncation family, then such tests may be described by a continuation region in the space of $n$, the sample size, and $M_n$, the sufficient statistics, which are of fixed dimension. In this case Schwarz has been able to describe the asymptotic shape of the continuation region as the sampling cost $c$ approaches zero. We have generalized Schwarz's work by considering more general families of distributions. In this paper the role of $M_n$ is played by the log likelihood function, and we show that the optimal Bayesian stopping rule may be approximated by a stopping rule which depends only on $n, c,$ and two likelihood ratio test statistics.

In multivariate analysis under normality assumptions, many likelihood ratio criteria $(\lambda^{(n)})$ are distributed as $k\prod^m_{i=1} Z^a_{li}(1 - Z_{li})^{b_i}\prod^{m'}_{j=1} Z^{c_j}_{2j}$ for some constants, $k, m, m', a_i, b_i,$ and $c_j$ when their associated null hypotheses are true, where $Z_{ij}$ are independently distributed beta variates. Let $T^{(n)} = -n^{-1} \ln \lambda^{(n)}$. This paper shows that a sequence $\{T^{(n)}\}$ of this kind is asymptotically optimal in the sense of exact slopes. Explicit forms of the exact slopes are obtained.

It has been conjectured by Khatri and Mardia that with probability one MLEs for the parameters of the von Mises-Fisher matrix distribution exist and are unique. We prove that, except for small sample sizes, this conjecture is true, both in the case where the parameter matrix has known rank and in the unrestricted case. The corresponding result for the matrix Bingham distribution is proven also.

An approach to simultaneous treatment of dependence and screening problems is presented. New characterizations of dependence of a random variable $X$ on a random vector $Y$ are obtained by functions $\nu_{X, Y}: (0, 1)\rightarrow \lbrack 0, 1\rbrack$ and $\mu_{X, Y} : (0, 1) \rightarrow \lbrack -1, 1\rbrack$ called respectively screening and monotonic dependence functions. These functions are shown to be appropriate measures of the intensity of connection and concordance of $X$ on $Y$, respectively. The interrelations of $\nu$ and $\mu$ and their relations to the multiple correlation ratio and the multiple correlation coefficient are demonstrated and illustrated by several examples.

Given $G \in \mathscr{S}$ a space of cumulative distribution functions and a weak (reflexive and transitive) order relation $\prec$ on $\mathscr{S}$, the subclass of $\mathscr{S}$ given by $\{F \in \mathscr{S} \mid F < G\}$ is called an ordered family of distributions. Suppose $\pi_1,\cdots, \pi_k$ represent $k$ populations with distributions only known to belong to some specified ordered family. The general problem is to design an experiment to select $\pi_i$'s having large (small) $\alpha$-quantities. A preferred population is defined to be any population with $\alpha$-quantile "near" the $t$th largest $\alpha$-quantile and a correct selection occurs if the subset of populations selected contains at least a prespecified number, $r$, of preferred populations. The design problem is solved for both fixed and random subset size selection procedures under star and tail ordering. Tables of the sample sizes required to guarantee prespecified minimum probabilities of correct selection are given for the case of selecting from continuous IFRA populations. Comparisons are made with the optimal procedure for selecting from exponential populations. Properties of the proposed rules are discussed.

A population consisting of an unknown number of distinct species is searched by selecting one member at a time. No a priori information is available concerning the probability that an object selected from this population will represent a particular species. Based on the information available after an $n$-stage search it is desired to predict the conditional probability that the next selection will represent a species not represented in the $n$-stage sample. Properties of a class of predictors obtained by extending the search an additional $m$ stages beyond the initial search are exhibited. These predictors have expectation equal to the unconditional probability of discovering a new species at stage $n + 1$, but may be strongly negatively correlated with the conditional probability.

We consider the class of regression functions $\mathscr{M}(F, G) = \{m(x) = E\lbrack Y\mid X = x\rbrack, (X, Y) \in \Pi(F, G)\}$ where $\Pi(F, G)$ denotes the set of random vectors with marginal distributions $F$ and $G$. A characterization of $\mathscr{M}(F, G)$ is given together with a representation for the projection operator it induces in an appropriate Hilbert space.

Assuming that $X$ has a two-dimensional normal distribution certain conditional distributions of $X$ given that $X$ lies on a hyperbola or a parabola are found. Two of these distributions, related respectively to the parabola and the hyperbola, resemble the von Mises distribution, which can be obtained as a conditional distribution of $X$ given that $X$ lies on a circle. It is, however, proved that the assumptions leading to the conditional distribution in the hyperbolic case are not analogous to those leading to the von Mises distribution.

Let $\{X_n\}$ be a Markov chain having a stationary transition function and assume that the state set is an arbitrary set in a Euclidean space. The state transition law of the chain is given by a function $F(y|x) = P\lbrack X_{n+1} \leqslant y|X_n = x\rbrack$, which is assumed defined and continuous for all $x$. In this paper we give a statistical procedure for determining a function $F_n(y\mid x)$ on the basis of the sample $\{X_j\}^n_{j=1}, n = 1, 2,\cdots,$ and prove that if the chain is irreducible, aperiodic, and possesses a limiting distribution $\pi$, then with probability 1, $\sup_y|F_n(y|x) - F(y|x)| \rightarrow_n0$ for every $x$ such that any open sphere containing $x$ has positive $\pi$ probability. This result improves upon a study by Roussas which gives only weak convergence. We demonstrate that a certain clustering algorithm is useful for obtaining efficient versions of our estimates. The potential value of our methods is illustrated by computer studies using simulated data.

Estimators are given for the two main parameters of the branching process with random environments, namely, the mean of the environmental mean (the average number of offspring per individual) and the mean of the logarithm of the environmental mean. The latter parameter indicates the certain extinction or possible explosion of the process. The consistency and asymptotic normality of the estimators are shown in the context of Smith and Wilkinson. Consistent estimators are also given for the corresponding variances and thus confidence intervals are obtained for the two main parameters of the system.

The normative procedure for the design of an experiment is to select a utility function, assess the probabilities, and to choose that design of maximum expected utility. One difficulty with this view is that a scientist typically does not have, nor can be normally expected to have, a clear idea of the utility of his results. An alternative is to design an experiment to maximize the expected information to be gained from it. In this paper we show that the latter view is a special case of the former with an appropriate choice of the decision space and a reasonable constraint on the utility function. In particular, the Shannon concept of information is seen to play a more important role in experimental design than was hitherto thought possible.

A technique used in pseudorandom number generation is to combine two or more different generators with the goal of producing a new generator with improved randomness properties. We study such a class of generators and show that in a strong sense the combined generator does offer improvement. Our approach applies results from majorization theory.