When a statistical model has a hierarchical structure such as multilayer perceptrons in neural networks or Gaussian mixture density representation, the model includes distribution with unidentifiable parameters when the structure becomes redundant. Since the exact structure is unknown, we need to carry out statistical estimation or learning of parameters in such a model. From the geometrical point of view, distributions specified by unidentifiable parameters become a singular point in the parameter space. The problem has been remarked in many statistical models, and strange behaviors of the likelihood ratio statistics, when the null hypothesis is at a singular point, have been analyzed so far. The present paper studies asymptotic behaviors of the maximum likelihood estimator and the Bayesian predictive estimator, by using a simple cone model, and show that they are completely different from regular statistical models where the Cramer-Rao paradigm holds. At singularities, the Fisher information metric degenerates, implying that the cramer-Rao paradigm does no more hold, and that he classical model selection theory such as AIC and MDL cannot be applied. This paper is a first step to establish a new theory for analyzing the accuracy of estimation or learning at around singularities.

Despite of the rich literature in discriminant analysis, this complicated subject remains much to be explored. In this article, we study the theoretical foundation that supports Fisher's linear discriminant analysis (LDA) by setting up the classification problem under the dimension reduction framework as in Li(1991) for introducing sliced inverse regression(SIR). Through the connection between SIR and LDA, our theory helps identify sources of strength and weakness in using CRIMCOORDS(Gnanadesikan 1977) as a graphical tool for displaying group separation patterns. This connection also leads to several ways of generalizing LDA for better exploration and exploitation of nonlinear data patterns.

How do you carry out data analysis\ulcorner There are few texts and little theory. One approach could be to use a pattern language, an idea which has been successful in field as diverse as town planning and software engineering. Patterns for data analysis are defined and discussed, illustrated with examples.

Local linear smoothing enjoys several excellent theoretical and numerical properties, an in a range of applications is the method most frequently chosen for fitting curves to noisy data. Nevertheless, it suffers numerical problems in places where the distribution of design points(often called predictors, or explanatory variables) is spares. In the case of univariate design, several remedies have been proposed for overcoming this problem, of which one involves adding additional ″pseudo″ design points in places where the orignal design points were too widely separated. This approach is particularly well suited to treating sparse bivariate design problem, and in fact attractive, elegant geometric analogues of unvariate imputation and interpolation rules are appropriate for that case. In the present paper we introduce and develop pseudo dta rules for bivariate design, and apply them to real data.

This paper concerns a problem of classification based on training dta. A framework of information geometry is given to elucidate the characteristics of discriminant functions including logistic discrimination and AdaBoost. We discuss a class of loss functions from a unified viewpoint.

Estimation of the edge of a distribution has many important applications. It is related to classification, cluster analysis, neural network, and statistical image recovering. The problem also arises in measuring production efficiency in economic systems. Three most promising nonparametric estimators in the existing literature are introduced. Their statistical properties are provided, some of which are new. Themes of future study are also discussed.

Symmetric estimation and recursive mean adjustment are considered to construct tests for the doble unit root hypothesis for both parametric and semiparametric time series models. It is shown that simultaneous application of symmetric estimation and recursive mean adjustment yields the most powerful test. Moreover, size property of the semiparametric test based on the simultaneous application is bet among all semiparametric tests.

The paper is a tutorial introduction to quantum information theory, developing the basic model and emphasizing the role of statistics and probability.

New statistical methods are ended to analyzed data in a multi-scale way. Some multi-scale extensions of stand methods, including novel visualization using dynamic graphics are proposed. These tools are used to explore non-standard structure in internet traffic data.