Robust Estimation of a Random Parameter in a Gaussian Linear Model With Joint Eigenvalue and Elementwise Covariance Uncertainties

Many applications require precise estimates of high-dimensional covariance matrices. The standard estimator is the sample covariance matrix, which is conceptually simple, fast to compute and has favorable properties in the limit of infinitely many observations. The picture changes when the dimensionality is of the same order as the number of observations. In such cases, the eigenvalues of the sample covariance matrix are highly biased, the condition number becomes large and the inversion of the matrix gets numerically unstable.
\nA number of alternative estimators are superior in the high-dimensional setting, which include as subcategories structured estimators, regularized estimators and spectrum correction methods. In this thesis I contribute to all three areas. 
\nIn the area of structured estimation, I focus on models with low intrinsic dimensionality. I analyze the bias in Factor Analysis, the state-of-the-art factor model and propose Directional Variance Adjustment (DVA) Factor Analysis, which reduces bias and yields improved estimates of the covariance matrix.
\nAnalytical shrinkage of Ledoit and Wolf (LW-Shrinkage) is the most popular regularized estimator. I contribute in three aspects: first, I provide a theoretical analysis of the behavior of LW-Shrinkage in the presence of pronounced eigendirections, a case of great practical relevance. I show that LW-Shrinkage does not perform well in this setting and propose aoc-Shrinkage which yields significant improvements. Second, I discuss the effect of autocorrelation on LW-Shrinkage and review the Sancetta-Estimator, an extension of LW-Shrinkage to autocorrelated data. I show that the Sancetta-Estimator is biased and propose a theoretically and empirically superior estimator with reduced bias. Third, I propose an extension of shrinkage to multiple shrinkage targets. Multi-Target Shrinkage is not restricted to covariance estimation and allows for many interesting applications which go beyond regularization, including transfer learning. I provide a detailed theoretical and empirical analysis.
\nSpectrum correction approaches the problem of covariance estimation by improving the estimates of the eigenvalues of the sample covariance matrix. I discuss the state-of-the-art approach, Nonlinear Shrinkage, and propose a cross-validation based covariance (CVC) estimator which yields competitive performance at increased numerical stability and greatly reduced complexity and computational cost. On all data sets considered, CVC is on par or superior in comparison to the regularized and structured estimators.
\nIn the last chapter, I conclude with a discussion of the advantages and disadvantages of all covariance estimators presented in this thesis and give situation-specific recommendations. In addition, the appendix contains a systematic analysis of Linear Discriminant Analysis as a model application, which sheds light on the interdependency between the generative model of the data and various covariance estimators.

Summary The primary purpose of this paper is to propose empirical measures of the structural differences between two communities of plants or animals composed of the same species. Structure is defined to consist of; 1) the species in the community, 2) the pattern of interactions as represented by the covariance or correlation matrix of successive observations on each species, and 3) the mean abundances of each species in each of the two communities. Statistical tests are proposed for testing whether the covariance matrices and the vectors of mean densities for each community are equal and empirical measures of the differences between the covariance matrices and mean vectors are proposed. Given unequal covariance or correlation matrices the factor analysis model is used to derive empirical measures of the degree to which each variable of the ecosystem is responsible for the observed defferences in the pattern of interactions in each community. These tests and measures were applied to data gathered by Hunter (1966) on the abundances of six species of Drosophila censused monthly over a period of approximately two and a half years in two adjacent, but different habitats near Bogota, colombia. The two covariance matrices were significantly different indicating different patterns of interactions in the two Drosophila communities.

Using conditional bias in principal component analysis for the evaluation of joint influence on the eigenvalues of the covariance matrix

A diagonally weighted matrix norm between two covariance matrices

Estimating large covariance and precision (inverse covariance) matrices has become increasingly important in high dimensional statistics because of its wide applications. The estimation problem is challenging not only theoretically due to the constraint of its positive definiteness, but also computationally because of the curse of dimensionality. Many types of estimators have been proposed such as thresholding under the sparsity assumption of the target matrix, banding and tapering the sample covariance matrix. However, these estimators are not always guaranteed to be positive-definite, especially, for finite samples, and the sparsity assumption is rather restrictive. We propose a novel two-stage adaptive method based on the Cholesky decomposition of a general covariance matrix. By banding the precision matrix in the first stage and adapting the estimates to the second stage estimation, we develop a computationally efficient and statistically accurate method for estimating high dimensional precision matrices. We demonstrate the finite-sample performance of the proposed method by simulations from autoregressive, moving average, and long-range dependent processes. We illustrate its wide applicability by analyzing financial data such S&P 500 index and IBM stock returns, and electric power consumption of individual households. The theoretical properties of the proposed method are also investigated within a large class of covariance matrices.

Statistical and computational techniques for revealing the internal structure that underlies the set of random correlated data exists in a great variety at present; and target decomposition theorems, either in the coherent or incoherent formulation, are well established. In spite of this fact a rather innovative and new concept is presented in this contribution. In turn the Principal Component Analysis (PCA) is considered to possibly add value to existing approaches, and it allows for an interpretation of polarimetric synthetic aperture radar measurements using variables obtained via linear transformation. Starting with the Sinclair backscatter matrix S which will be further transformed into the so called target feature vector by stacking column elements of S and generating the covariance matrices averaged over a certain pixel array, we show, how the Sinclair backscatter matrix is decomposed into the sum of a maximum of four 2 x 2 elementary point scatter matrices which are weighted by the principal components, whereas the variances of these components agree with the eigenvalues of the covariance matrix. This mathematical development defines a decomposition which expresses scattering mechanism from distributed targets in terms of scattering matrices via an incoherent step.

A sample \( X_{1},\ldots ,X_{n} \) consisting of independent identically distributed random vectors in \( \mathbb {R}^{p} \) with zero mean and covariance matrix \( \mathbf {\Sigma }\) is considered. The recovery of spectral projectors of high-dimensional covariance matrices from a sample of observations is a key problem in statistics arising in numerous applications. This chapter describes a bootstrap procedure for constructing confidence sets for the spectral projector \( \mathbf {P}_{r} \) related to rth eigenvalue of the covariance matrix \(\mathbf {\Sigma }\) from given data on the base of corresponding spectral projector \(\widehat{\mathbf {P}}_{r}\) of the sample covariance matrix \(\widehat{\mathbf {\Sigma }}\). This approach does not use the asymptotical distribution of \( \Vert \mathbf {P}_{r} - \widehat{\mathbf {P}}_{r} \Vert _{2} \) and does not require the computation of its moment characteristics. The performance of the bootstrap approximation procedure is analyzed.

Detecting emergence of a low-rank signal from high-dimensional data is an important problem arising from many applications such as camera surveillance and swarm monitoring using sensors. We consider a procedure based on the largest eigenvalue of the sample covariance matrix over a sliding window to detect the change. To achieve dimensionality reduction, we present a sketching-based approach for rank change detection using the low-dimensional linear sketches of the original high-dimensional observations. The premise is that when the sketching matrix is a random Gaussian matrix, and the dimension of the sketching vector is sufficiently large, the rank of sample covariance matrix for these sketches equals the rank of the original sample covariance matrix with high probability. Hence, we may be able to detect the low-rank change using sample covariance matrices of the sketches without having to recover the original covariance matrix. We character the performance of the largest eigenvalue statistic in terms of the false-alarm-rate and the expected detection delay, and present an efficient online implementation via subspace tracking.

We consider Gaussian mixture models in high dimensions, focusing on the twin tasks of detection and feature selection. Under sparsity assumptions on the difference in means, we derive minimax rates for the problems of testing and of variable selection. We find these rates to depend crucially on the knowledge of the covariance matrices and on whether the mixture is symmetric or not. We establish the performance of various procedures, including the top sparse eigenvalue of the sample covariance matrix (popular in the context of Sparse PCA), as well as new tests inspired by the normality tests of Malkovich and Afifi [J. Amer. Statist. Assoc. 68 (1973) 176–179].

Upcoming and existing large-scale surveys of galaxies require accurate theoretical predictions of the dark matter clustering statistics for thousands of mock galaxy catalogs. We demonstrate that this goal can be achieve with our new Parallel Particle-Mesh (PM) Nbody code (PPM-GLAM) at a very low computational cost. We run about 15,000 simulations with ~2 billion particles that provide ~1% accuracy of the dark matter power spectra P(k) for wave-numbers up to k~ 1h/Mpc. Using this large data-set we study the power spectrum covariance matrix, the stepping stone for producing mock catalogs. In contrast to many previous analytical and numerical results, we find that the covariance matrix normalised to the power spectrum C(k,k')/P(k)P(k') has a complex structure of non-diagonal components. It has an upturn at small k, followed by a minimum at k=0.1-0.2h/Mpc. It also has a maximum at k=0.5-0.6h/Mpc. The normalised covariance matrix strongly evolves with redshift: C(k,k')~delta(t)^alpha P(k)P(k'), where delta is the linear growth factor and alpha ~ 1-1.25, which indicates that the covariance matrix depends on cosmological parameters. We also show that waves longer than 1Gpc have very little impact on the power spectrum and covariance matrix. This significantly reduces the computational costs and complexity of theoretical predictions: relatively small volume ~ (1Gpc)^3 simulations capture the necessary properties of dark matter clustering statistics. All the power spectra obtained from many thousands of our simulations are publicly available.

Estimation of the high-dimensional banded covariance matrix is widely used in multivariate statistical analysis. To ensure the validity of estimation, we aim to test the hypothesis that the covariance matrix is banded with a certain bandwidth under the high-dimensional framework. Though several testing methods have been proposed in the literature, the existing tests are only powerful for some alternatives with certain sparsity levels, whereas they may not be powerful for alternatives with other sparsity structures. The goal of this paper is to propose a new test for the bandedness of high-dimensional covariance matrix, which is powerful for alternatives with various sparsity levels. The proposed new test also be used for testing the banded structure of covariance matrices of error vectors in high-dimensional factor models. Based on these statistics, a consistent bandwidth estimator is also introduced for a banded high dimensional covariance matrix. Extensive simulation studies and an application to a prostate cancer dataset from protein mass spectroscopy are conducted for evaluating the effectiveness of the proposed adaptive tests blue and bandwidth estimator for the banded covariance matrix.

High dimensional covariance matrices have attracted much attention of statisticians and econometricians during the past decades. Vast literature is devoted to the research in high dimensional covariance matrices. However, most of them are for constant covariance matrices. In many applications, constant covariance matrices are not appropriate, e.g. in portfolio allocation, dynamic covariance matrices would make much more sense. Simply assuming each entry of a covariance matrix is a function of time to introduce a dynamic structure would not work. In this paper, we are going to introduce a class of high dimensional dynamic covariance matrices in which a kind of additive structure is embedded. We will show the proposed high dimensional dynamic covariance matrices have many advantages in applications. An estimation procedure is also proposed to estimate the proposed high dimensional dynamic covariance matrices. Asymptotic properties are built to justify the proposed estimation procedure. Intensive simulation studies show the proposed estimation procedure works very well when sample size is finite. Finally, we apply the proposed high dimensional dynamic covariance matrices, together with the proposed estimation procedure, to portfolio allocation. The results look very interesting.

Summary Ensemble optimization (referred to throughout the remainder of the paper as EnOpt) is a rapidly emerging method for reservoir-model-based production optimization. EnOpt uses an ensemble of controls to approximate the gradient of the objective function with respect to the controls. Current implementations of EnOpt use a Gaussian ensemble of control perturbations with a constant covariance matrix, and thus a constant perturbation size, during the entire optimization process. The covariance-matrix-adaptation evolutionary strategy is a gradient-free optimization method developed in the “machine learning” community, which also uses an ensemble of controls, but with a covariance matrix that is continually updated during the optimization process. It was shown to be an efficient method for several difficult but small-dimensional optimization problems and was recently applied in the petroleum industry for well location and production optimization. In this study, we investigate the scope to improve the computational efficiency of EnOpt through the use of covariance-matrix adaptation (referred to throughout the remainder of the paper as CMA-EnOpt). The resulting method is applied to the waterflooding optimization of a small multilayer test model and a modified version of the Brugge benchmark model. The controls used are inflow-control-valve settings at predefined time intervals for injectors and producers with undiscounted net present value as the objective function. We compare EnOpt and CMA-EnOpt starting from identical covariance matrices. For the small model, we achieve only slightly higher (0.7 to 1.8%) objective-function values and modest speedups with CMA-EnOpt compared with EnOpt. Significantly higher objective-function values (10%) are obtained for the modified Brugge model. The possibility to adapt the covariance matrix, and thus the perturbation size, during the optimization allows for the use of relatively large perturbations initially, for fast exploration of the control space, and small perturbations later, for more-precise gradients near the optimum. Moreover, the results demonstrate that a major benefit of CMA-EnOpt is its robustness with respect to the initial choice of the covariance matrix. A poor choice of the initial matrix can be detrimental to EnOpt, whereas the CMA-EnOpt performance is near-independent of the initial choice and produces higher objective-function values at no additional computational cost.

Sensitivity analysis of uncertain nonlinear systems plays an important role in analyzing chemical engineering process models. Even though a variety of different techniques exists for analyzing nonlinear systems, these methods often do not update the parameter distribution using available measurement data. This forms the motivation behind this work as a sensitivity analysis procedure is introduced that deals with nonlinear systems and uncertainty in model parameters, and also incorporates the use of available data. The technique computes two covariance matrices: a covariance matrix that describes the effect that changes of the parameters values, sampled from a prior distribution, have on the outputs, and a covariance matrix that determines the effect of available data on the posterior distribution of the parameters using a Markov chain Monte Carlo method. The information contained in these two covariance matrices is then simultaneously evaluated by balancing the two matrices. The results returned by this technique are the directions in parameter space that have the largest impact on the system behavior according to the given parameter uncertainty as well as the contribution of individual parameters to these important directions in parameter space. The presented technique is applied to three examples.

The rotated delta, which is argued to be a theoretically better-grounded distance measure, has failed to receive any empirical support for its superiority. This study revisits the rotated delta—which is more commonly known as the Mahalanobis distance in other areas—with two different covariance matrices that are estimated from training data. The first covariance matrix represents the between-author variability, and the second the within-author variability. A series of likelihood ratio-based authorship verification experiments was carried out with some different distance measures. The experiments made use of the documents arranged from a large database of text messages that allowed for a total of 2,160 same-author and 4,663,440 different-author comparisons. The Mahalanobis distance with the between-author covariance matrix performed far worse compared to the other distance measures, whereas the Mahalanobis distance with the within-author covariance matrix performed better than the other measures. However, superior performance relative to the cosine distance is subject to word lengths and/or the order of the feature vector. The result of follow-up experiments further illustrated that the covariance matrix representing the within-author variability needs to be trained using a good amount of data to perform better than the cosine distance: the higher the order of the vector, the more data are required for training. The quantitative results also infer that the two sources of variabilities—notably within- and between-author variabilities—are independent of each other to the extent that the latter cannot accurately approximate the former.