We study the asymptotic behavior of the integrated periodogram for $\alpha$-stable linear processes. For $\alpha \epsilon (1, 2)$ we prove a functional limit theorem for the integrated periodogram. The limit is an $\alpha$-stable analogue to the Brownian bridge. We apply our results to investigate some specific goodness-of-fit tests for heavy-tailed linear processes.

Consider the fractional ARIMA time series with innovations that have infinite variance. This is a finite parameter model which exhibits both long-range dependence (long memory) and high variability. We prove the consistency of an estimator of the unknown parameters which is based on the periodogram and derive its asymptotic distribution. This shows that the results of Mikosch, Gadrich, Klüppelberg and Adler for ARMA time series remain valid for fractional ARIMA with long-range dependence. We also extend the limit theorem for sample autocovariances of infinite variance moving averages developed in Davis and Resnick to moving averages whose coefficients are not absolutely summable.

We show that the blockwise bootstrap approximation for the distribution of a studentized statistic computed from dependent data is second-order correct provided we choose an appropriate variance estimator. We also show how to adapt the $BC_a$ confidence interval of Efron to the a dependent case. For the proofs we extend the results of Götze and Hipp on the validity of the formal Edgeworth expansion for a sum to the studentized mean.

The asymptotic properties of the bootstrap in the frequency domain based on Studentized periodogram ordinates are studied. It is proved that this bootstrap approximation is valid for ratio statistics such as autocorrelations. By using Edgeworth expansions it is shown that the bootstrap approximation even outperforms the normal approximation. The results carry over to Whittle estimates. In a simulation study the behavior of the bootstrap is studied for empirical correlations and Whittle estimates.

We study the blind deconvolution problem in the case where the input noise has a finite discrete support and the transfer linear system is not necessarily minimum phase. We propose a new family of estimators built using algebraic considerations. The estimates are consistent under very wide assumptions: The input signal need not be independently distributed; the cardinality of the finite support may be estimated simultaneously. We consider in particular AR systems: In this case, we prove that the estimator of the parameters is perfect a.s. with a finite number of observations.

Data-driven Neyman's tests resulting from a combination of eyman' smooth tests for uniformity and Schwarz's selection procedure are nvestigated. Asymptotic intermediate efficiency of those tests with respect to the Neyman-Pearson test is shown to be 1 for a large set of converging alternatives. The result shows that data-driven Neyman's tests, contrary to classical goodness-of-it tests, are indeed omnibus tests adapting well to the data at hand.

The power of Kolmogorov-Smirnov tests can be increased by transforming the empirical process into a new process that converges to a Wiener process under the null hypothesis and by choosing the transformation in such a way that some families of local alternatives become as noticeable as possible.

Cohen, Sackrowitz and Samuel-Cahn introduced the notion of cone order association and established a necessary and sufficient condition for a normal random vector to be cone order associated (COA). In this paper we provide the following: (1) a necessary and sufficient condition for a multinomial distribution to be COA when the cone is a pairwise contrast cone; (2) a relationship between COA and regular association; (3) a notion of stochastic cone ordering (SCO) of random vectors along with two preservation theorems indicating monotonicity properties of expectations as functions of parameters; and (4) applications to unbiasedness of tests and monotonicity of power functions of tests in cone order-restricted hypothesis-testing problems. In particular, the matrix order alternative hypothesis-testing problem is treated when the underlying distributions are independent Poisson or the joint distribution is multinomial.

It is shown that the maximum likelihood estimator in a model used in the statistical analysis of computer experiments is asymptotically efficient.

This paper contains a collection of results on Latin hypercube sampling. The first result is a Berry-Esseen-type bound for the multivariate central limit theorem of the sample mean $\hat{\mu}_n$ based on a Latin hypercube sample. The second establishes sufficient conditions on the convergence rate in the strong law for $\hat{\mu}_n$. Finally motivated by the concept of empirical likelihood, a way of constructing nonparametric confidence regions based on Latin hypercube samples is proposed for vector means.

We study linear estimators for the weighted integral of a stochastic process. The process may only be observed on a finite sampling design. The error is defined in a mean square sense, and the process is assumed to satisfy Sacks-Ylvisaker regularity conditions of order $r \epsilon \mathbb{N}_0$. We show that sampling at the quantiles of a particular density already yields asymptotically optimal estimators. Hereby we extend the results of Sacks and Ylvisaker for regularity $r = 0$ or 1, and we confirm a conjecture by Eubank, Smith and Smith.

The relationship between admissible incomplete block designs for confidence intervals with maximal coverage probability for treatment-control contrasts and optimal designs for estimation is investigated. For certain types of designs, admissible designs are shown to be precisely those with the number of replications of the control less than or equal to that of an optimal design. Moreover, admissible designs are the Bayes optimal designs for a class of priors.

We consider the problem of finding a nonsequential optimal design for estimating parameters in a generalized exponential growth model. This problem is solved by first considering polynomial regression models with error variances that depend on the covariate value and unknown parameters. A Bayesian approach is adopted, and optimal Bayesian designs supported on a minimal number of support points for estimating the coefficients in the polynomial model are found analytically. For some criteria, the optimal Bayesian designs depend only on the expectation of the prior, but generally their dependence includes the derivative of the logarithm of the Laplace transform of a measure induced by the prior. The optimal design for the generalized exponential growth model is then determined from these optimal Bayesian designs.

In this paper, experimental designs for a rational model, $Y = P(x)/Q(x) + \varepsilon$, are investigated, where $P(x) = \theta_0 + \theta_1 x + \dots + \theta_p x^p$ and $Q(x) = 1 + \theta_{p + 1} x + \dots + \theta_{p+q} x^q$ are polynomials and $\varepsilon$ is a random error. Two approaches, Bayesian D-optimal and Bayesian optimal design for extrapolation, are examined. The first criterion maximizes the expected increase in Shannon information provided by the experiment asymptotically, and the second minimizes the asymptotic variance of the maximum likelihood estimator (MLE) of the mean response at an extrapolation point $x_e$. Corresponding locally optimal designs are also discussed. Conditions are derived under which a $p + q + 1$-point design is a locally D-optimal design. The Bayesian D-optimal design is shown to be independent of the parameters in $P(x)$ and to be equally weighted at each support point if the number of support points is the same as the number of parameters in the model. The existence and uniqueness of the locally optimal design for extrapolation are proven. The number of support points for the locally optimal design for extrapolation is exactly $p + q + 1$. These $p + q + 1$ design points are proved to be independent of the extrapolation point x e and the parameters in $P(x)$. The corresponding weights are also independent of the parameters in $P(x)$, but depend on $x_e$ and are not equal.

This paper aims to solve an often noted incompatibility between graphical chain models which elucidate the conditional independence structure of a set of random variables and simultaneous equations systems which focus on direct linear interactions and correlations between random variables. Various authors have argued that the incompatibility arises mainly from the fact that in a simultaneous equations system (e.g., a LISREL model) reciprocal causality is possible whereas this is not so in the case of graphical chain models. In this article it is shown that this view is not correct. In fact, the definition of the Markov property embodied in a graph can be generalized to a wider class of graphs which includes certain nonrecursive graphs. The resulting class of reciprocal graph probability models strictly includes the class of chain graph probability models. The class of lattice conditional independence probability models is also strictly included. It is shown that the resulting methodology is directly applicable to quite general simultaneous equations systems that are subject to mild restrictions only. Provided some adjustments are made, general simultaneous equations systems can be handled as well. In all cases, consistency with the LISREL methodology is maintained.

We wish to make inferences about the conditional probabilities $p(y|x)$, many of which are zero, when the distribution of X is unknown and one observes only a multinomial sample of the Y variates. To do this, fixed likelihood ratio models and quasi-incremental distributions are defined. It is shown that quasi-incremental distributions are intimately linked to decomposable graphs and that these graphs can guide us to transformations of X and Y which admit a conjugate Bayesian analysis on a reparametrization of the conditional probabilities of interest.

We show that if dynamic sampling is feasible, then there exist surveillance schemes that satisfy a probability constraint on false alarm. Procedures are suggested for detecting a change of a normal mean from 0 to a (unknown) positive value. These procedures are optimal (up to a constant term) when the post-change mean is known, and almost optimal [up to an $o(\log(1/\alpha))$ term when the post-change mean is unknown.

In many areas of application, one searches within finite populations for items of interest, where the probability of sampling an item is proportional to a random size attribute from an i.i.d. superpopulation of attributes which may or may not be observable upon discovery. Here we treat the problem of asymptotically optimal stopping rules for size-dependent searches of this type, as the size of the underlying population grows, where the loss function includes an asymptotically smooth time-dependent cost, a constant cost per item sampled and a cost per undiscovered item which may depend on the size attribute of the undiscovered item. Under some regularity and convexity conditions related to the asymptotic expected loss, we characterize asymptotically optimal rules even when the initial population size and the distribution of size attributes are unknown. We direct especial attention to applications in software reliability, where the items of interest are software faults ("bugs"). In this setting, the size attributes will not be observable when faults are found, and, in addition, our search model allows new bugs to be introduced into the software when faults are detected "imperfect debugging"). Our results extend those of Dalal and Mallows and Kramer and Starr, and are illustrated in the perfect-debugging case on a previously analyzed dataset of Musa.

This paper considers sequential point estimation of the autocorrelations of stationary linear processes within the framework of the sequential procedure initiated by Robbins. The sequential estimator proposed here is based on the usual sample autocorrelations and is shown to be risk efficient in the sense of Starr as the cost per observation approaches zero. To achieve the asymptotic risk efficiency, we are led to study the uniform integrability and random central limit theorem of the sample autocorrelations. Some moment conditions are provided for the errors of the linear processes to establish the uniform integrability and random central limit theorem.

We study approximations to the distribution of the largest Walsh-Fourier coefficient obtained by transforming i.i.d. observations from a non-normal parent distribution. The study is inspired by the use of half-normal plots in screening studies and in the analysis of molecular sequence data. In the latter context, we indicate why continuity corrections are needed.

We show that the maximally selected standardized U-statistic goes in distribution to an infinite sum of weighted chi-square random variables in the degenerate case. The result is applied to the detection of possible changes in the distribution of a sequence observation.

An upper bound for the tail probability $P_{\theta} (\log (L(x_{(n_1, \dots, n_q)}, \Theta)/L(x_{(n_1, \dots, n_q)}, \theta)) \geq t)$ is derived in the case of sampling from q populations. This estimate is used for establishing the Hodges-Lehmann optimality of a test statistic for a hypothesis on exponential distributions.

This paper presents a lower bound, derived from the information inequality for the Bayes risk with respect to truncated priors under quadratic loss. It is discussed in cases where the regularity condition of Brown and Gajek is not always satisfied. A related result for the minimax risk is also given.