Multivariate statistical analysis. A high-dimensional approach (Q1347365)

The book presents new possibilities to treat some problems in multivariate analysis when both the sample size and the dimension of the observed vectors are tending to infinity and are of the same order. The methods are based on a special treatment of spectral properties of sample variance matrices and of related functionals in high-dimensional situations in combination with the multivariate techniques based on the decision functions approach by A. Wald. The author points out that for high-dimensional problems, under a weak restrictions on variable dependence, the standard quality functions of regularized multivariate procedures prove to be independent of the distibutions. The book is divided into the Introduction and 11 chapters. The Introduction presents historical aspects and the line of developments. Chapter 1 recalls some fundamentals of classical multivariate analysis when observations follow the normal distribution, including spectral properties of large Wishart matrices. Chapter 2 presents a finite-dimensional distribution-free theory of spectral properties of large sample covariance matrices. These results are extended to large pooled sample covariance matrices in Chapter 3. Chapter 4 studies a specific normalization effect made by a large number of restrictively dependent estimators. In Chapter 5, the problem of estimating inverses of covariance matrices of high dimensions is treated. Chapter 6 concerns the problem of estimation of normal mean vectors with minimum quadratic risk, while Chapter 7 treats the problem of estimating mean vectors for arbitrary populations with some moment restrictions. In Chapter 8, the quadratic risk of linear regressions with a large number of random predictors is considered. Chapter 9 deals with discriminant analysis for normal populations with identical covariance matrices using a class of regularized linear discriminant functions. In Chapter 10, the procedures developed in Chapter 9 are extended to arbitrary populations with some moment restrictions. The last Chapter 11 develops a theory of weighting and selection of independent variables in high-dimensional discriminant analysis. Chapter headings: Introduction; 1. Spectral properties of large Wishart matrices; 2. Resolvents and spectral functions of large sample covariance matrices; 3. Resolvents and spectral functions of large pooled sample covariance matrices; 4. Normal evaluation of quality functions; 5. Estimation of high-dimensional inverse covariance matrices; 6.Epsilon-dominating component-wise shrinkage estimators of normal means; 7. Improved estimators of high-dimensional expectation vectors; 8. Quadratic risk of linear regression with a large number of random predictors; 9. Linear discriminant analysis of normal populations with coinciding covariance matrices; 10. Population free quality of discrimination; 11. Theory of dicsriminant analysis of an increasing number of independent variables. The author claims that the proposed methods are useful in some particular problems of multivariate analysis, which is certainly true, but before these ideas are applied to some real data sets still quite a lot has to be done. The starting point of the book are papers by A.N. Kolmogorov published in the period 1970-1974. A substantial part of the book is based on results obtained by the author and by Girko. The book can be of interest for specialists in mathematical statistics, econometrics, reliability and risk analysis with special interest in multivariate analysis. To understand the material, quite advanced knowlwedge of theoretical results on spectra of high-dimensional random matrices is useful.