Bayesian Inference in Large-scale Problems

Many modern applications fall into the category of "large-scale" statistical problems, in which both the number of observations n and the number of features or parameters p may be large. Many existing methods focus on point estimation, despite the continued relevance of uncertainty quantification in the sciences, where the number of parameters to estimate often exceeds the sample size, despite huge increases in the value of n typically seen in many fields. Thus, the tendency in some areas of industry to dispense with traditional statistical analysis on the basis that "n=all" is of little relevance outside of certain narrow applications. The main result of the Big Data revolution in most fields has instead been to make computation much harder without reducing the importance of uncertainty quantification. Bayesian methods excel at uncertainty quantification, but often scale poorly relative to alternatives. This conflict between the statistical advantages of Bayesian procedures and their substantial computational disadvantages is perhaps the greatest challenge facing modern Bayesian statistics, and is the primary motivation for the work presented here.

Two general strategies for scaling Bayesian inference are considered. The first is the development of methods that lend themselves to faster computation, and the second is design and characterization of computational algorithms that scale better in n or p. In the first instance, the focus is on joint inference outside of the standard problem of multivariate continuous data that has been a major focus of previous theoretical work in this area. In the second area, we pursue strategies for improving the speed of Markov chain Monte Carlo algorithms, and characterizing their performance in large-scale settings. Throughout, the focus is on rigorous theoretical evaluation combined with empirical demonstrations of performance and concordance with the theory.

One topic we consider is modeling the joint distribution of multivariate categorical data, often summarized in a contingency table. Contingency table analysis routinely relies on log-linear models, with latent structure analysis providing a common alternative. Latent structure models lead to a reduced rank tensor factorization of the probability mass function for multivariate categorical data, while log-linear models achieve dimensionality reduction through sparsity. Little is known about the relationship between these notions of dimensionality reduction in the two paradigms. In Chapter 2, we derive several results relating the support of a log-linear model to nonnegative ranks of the associated probability tensor. Motivated by these findings, we propose a new collapsed Tucker class of tensor decompositions, which bridge existing PARAFAC and Tucker decompositions, providing a more flexible framework for parsimoniously characterizing multivariate categorical data. Taking a Bayesian approach to inference, we illustrate empirical advantages of the new decompositions.

Latent class models for the joint distribution of multivariate categorical, such as the PARAFAC decomposition, data play an important role in the analysis of population structure. In this context, the number of latent classes is interpreted as the number of genetically distinct subpopulations of an organism, an important factor in the analysis of evolutionary processes and conservation status. Existing methods focus on point estimates of the number of subpopulations, and lack robust uncertainty quantification. Moreover, whether the number of latent classes in these models is even an identified parameter is an open question. In Chapter 3, we show that when the model is properly specified, the correct number of subpopulations can be recovered almost surely. We then propose an alternative method for estimating the number of latent subpopulations that provides good quantification of uncertainty, and provide a simple procedure for verifying that the proposed method is consistent for the number of subpopulations. The performance of the model in estimating the number of subpopulations and other common population structure inference problems is assessed in simulations and a real data application.

In contingency table analysis, sparse data is frequently encountered for even modest numbers of variables, resulting in non-existence of maximum likelihood estimates. A common solution is to obtain regularized estimates of the parameters of a log-linear model. Bayesian methods provide a coherent approach to regularization, but are often computationally intensive. Conjugate priors ease computational demands, but the conjugate Diaconis--Ylvisaker priors for the parameters of log-linear models do not give rise to closed form credible regions, complicating posterior inference. In Chapter 4 we derive the optimal Gaussian approximation to the posterior for log-linear models with Diaconis--Ylvisaker priors, and provide convergence rate and finite-sample bounds for the Kullback-Leibler divergence between the exact posterior and the optimal Gaussian approximation. We demonstrate empirically in simulations and a real data application that the approximation is highly accurate, even in relatively small samples. The proposed approximation provides a computationally scalable and principled approach to regularized estimation and approximate Bayesian inference for log-linear models.

Another challenging and somewhat non-standard joint modeling problem is inference on tail dependence in stochastic processes. In applications where extreme dependence is of interest, data are almost always time-indexed. Existing methods for inference and modeling in this setting often cluster extreme events or choose window sizes with the goal of preserving temporal information. In Chapter 5, we propose an alternative paradigm for inference on tail dependence in stochastic processes with arbitrary temporal dependence structure in the extremes, based on the idea that the information on strength of tail dependence and the temporal structure in this dependence are both encoded in waiting times between exceedances of high thresholds. We construct a class of time-indexed stochastic processes with tail dependence obtained by endowing the support points in de Haan's spectral representation of max-stable processes with velocities and lifetimes. We extend Smith's model to these max-stable velocity processes and obtain the distribution of waiting times between extreme events at multiple locations. Motivated by this result, a new definition of tail dependence is proposed that is a function of the distribution of waiting times between threshold exceedances, and an inferential framework is constructed for estimating the strength of extremal dependence and quantifying uncertainty in this paradigm. The method is applied to climatological, financial, and electrophysiology data.

The remainder of this thesis focuses on posterior computation by Markov chain Monte Carlo. The Markov Chain Monte Carlo method is the dominant paradigm for posterior computation in Bayesian analysis. It has long been common to control computation time by making approximations to the Markov transition kernel. Comparatively little attention has been paid to convergence and estimation error in these approximating Markov Chains. In Chapter 6, we propose a framework for assessing when to use approximations in MCMC algorithms, and how much error in the transition kernel should be tolerated to obtain optimal estimation performance with respect to a specified loss function and computational budget. The results require only ergodicity of the exact kernel and control of the kernel approximation accuracy. The theoretical framework is applied to approximations based on random subsets of data, low-rank approximations of Gaussian processes, and a novel approximating Markov chain for discrete mixture models.

Data augmentation Gibbs samplers are arguably the most popular class of algorithm for approximately sampling from the posterior distribution for the parameters of generalized linear models. The truncated Normal and Polya-Gamma data augmentation samplers are standard examples for probit and logit links, respectively. Motivated by an important problem in quantitative advertising, in Chapter 7 we consider the application of these algorithms to modeling rare events. We show that when the sample size is large but the observed number of successes is small, these data augmentation samplers mix very slowly, with a spectral gap that converges to zero at a rate at least proportional to the reciprocal of the square root of the sample size up to a log factor. In simulation studies, moderate sample sizes result in high autocorrelations and small effective sample sizes. Similar empirical results are observed for related data augmentation samplers for multinomial logit and probit models. When applied to a real quantitative advertising dataset, the data augmentation samplers mix very poorly. Conversely, Hamiltonian Monte Carlo and a type of independence chain Metropolis algorithm show good mixing on the same dataset.

A comparison of low-storage strategies for spectral analysis in dissipative systems: filter diagonalisation in the Lanczos representation and harmonic inversion of the Chebychev-order-domain autocorrelation function

We compare the performance of two different low-storage filter diagonalisation (LSFD) strategies in the calculation of complex resonance energies of the HO2, radical. The first is carried out within a complex-symmetric Lanczos subspace representation [H. Zhang, S.C. Smith, Phys. Chem. Chem. Phys. 3 (2001) 2281]. The second involves harmonic inversion of a real autocorrelation function obtained via a damped Chebychev recursion [V.A. Mandelshtam, H.S. Taylor, J. Chem. Phys. 107 (1997) 6756]. We find that while the Chebychev approach has the advantage of utilizing real algebra in the time-consuming process of generating the vector recursion, the Lanczos, method (using complex vectors) requires fewer iterations, especially for low-energy part of the spectrum. The overall efficiency in calculating resonances for these two methods is comparable for this challenging system. (C) 2001 Elsevier Science B.V. All rights reserved.

Fractional Dynamics Representation in the Pseudo Phase Plane

Proceedings of the 10th Conference on Dynamical Systems Theory and Applications

Soil spectral library and its use in soil classification

Soil science has sought to develop better techniques for the classification of soils, one of which is the use of remote sensing applications. The use of ground sensors to obtain soil spectral data has enabled the characterization of these data and the advancement of techniques for the quantification of soil attributes. In order to do this, the creation of a soil spectral library is necessary. A spectral library should be representative of the variability of the soils in a region. The objective of this study was to create a spectral library of distinct soils from several agricultural regions of Brazil. Spectral data were collected (using a Fieldspec sensor, 350-2,500 nm) for the horizons of 223 soil profiles from the regions of Matão, Paraguaçu Paulista, Andradina, Ipaussu, Mirandópolis, Piracicaba, São Carlos, Araraquara, Guararapes, Valparaíso (SP); Naviraí, Maracajú, Rio Brilhante, Três Lagoas (MS); Goianésia (GO); and Uberaba and Lagoa da Prata (MG). A Principal Component Analysis (PCA) of the data was then performed and a graphic representation of the spectral curve was created for each profile. The reflectance intensity of the curves was principally influenced by the levels of Fe2O3, clay, organic matter and the presence of opaque minerals. There was no change in the spectral curves in the horizons of the Latossolos, Nitossolos, and Neossolos Quartzarênicos. Argissolos had superficial horizon curves with the greatest intensity of reflection above 2,200 nm. Cambissolos and Neossolos Litólicos had curves with greater reflectance intensity in poorly developed horizons. Gleisols showed a convex curve in the region of 350-400 nm. The PCA was able to separate different data collection areas according to the region of source material. Principal component one (PC1) was correlated with the intensity of reflectance samples and PC2 with the slope between the visible and infrared samples. The use of the Spectral Library as an indicator of possible soil classes proved to be an important tool in profile classification.

Energy-weighted sum rules for mesons in hot and dense matter

We study energy-weighted sum rules of the pion and kaon propagator in nuclear matter at finite temperature. The sum rules are obtained from matching the Dyson form of the meson propagator with its spectral Lehmann representation at low and high energies. We calculate the sum rules for specific models of the kaon and pion self-energy. The in-medium spectral densities of the K and (K) over bar mesons are obtained from a chiral unitary approach in coupled channels that incorporates the S and P waves of the kaon-nucleon interaction. The pion self-energy is determined from the P-wave coupling to particle-hole and Delta-hole excitations, modified by short-range correlations. The sum rules for the lower-energy weights are fulfilled satisfactorily and reflect the contributions from the different quasiparticle and collective modes of the meson spectral function. We discuss the sensitivity of the sum rules to the distribution of spectral strength and their usefulness as quality tests of model calculations.

Spectral Analysis of Buffers in Communication Systems

Muokatun matriisi-geometrian tekniikan kehitys yleimmäksi jonoksi on esitelty tässä työssä. Jonotus systeemi koostuu useista jonoista joilla on rajatut kapasiteetit. Tässä työssä on myös tutkittu PH-tyypin jakautumista kun ne jaetaan. Rakenne joka vastaa lopullista Markovin ketjua jossa on itsenäisiä matriiseja joilla on QBD rakenne. Myös eräitä rajallisia olotiloja on käsitelty tässä työssä. Sen esitteleminen matriisi-geometrisessä muodossa, muokkaamalla matriisi-geometristä ratkaisua on tämän opinnäytetyön tulos.

Spectral retinal image processing and analysis for ophthalmology

Diabetic retinopathy, age-related macular degeneration and glaucoma are the leading causes of blindness worldwide. Automatic methods for diagnosis exist, but their performance is limited by the quality of the data. Spectral retinal images provide a significantly better representation of the colour information than common grayscale or red-green-blue retinal imaging, having the potential to improve the performance of automatic diagnosis methods. This work studies the image processing techniques required for composing spectral retinal images with accurate reflection spectra, including wavelength channel image registration, spectral and spatial calibration, illumination correction, and the estimation of depth information from image disparities. The composition of a spectral retinal image database of patients with diabetic retinopathy is described. The database includes gold standards for a number of pathologies and retinal structures, marked by two expert ophthalmologists. The diagnostic applications of the reflectance spectra are studied using supervised classifiers for lesion detection. In addition, inversion of a model of light transport is used to estimate histological parameters from the reflectance spectra. Experimental results suggest that the methods for composing, calibrating and postprocessing spectral images presented in this work can be used to improve the quality of the spectral data. The experiments on the direct and indirect use of the data show the diagnostic potential of spectral retinal data over standard retinal images. The use of spectral data could improve automatic and semi-automated diagnostics for the screening of retinal diseases, for the quantitative detection of retinal changes for follow-up, clinically relevant end-points for clinical studies and development of new therapeutic modalities.

Spectral analysis of the elliptic sine-Gordon equation in the quarter plane

We study the elliptic sine-Gordon equation in the quarter plane using a spectral transform approach. We determine the Riemann-Hilbert problem associated with well-posed boundary value problems in this domain and use it to derive a formal representation of the solution. Our analysis is based on a generalization of the usual inverse scattering transform recently introduced by Fokas for studying linear elliptic problems.

On spectral invariance of single scattering albedo for water droplets and ice crystals at weakly absorbing wavelengths

The single scattering albedo w_0l in atmospheric radiative transfer is the ratio of the scattering coefficient to the extinction coefficient. For cloud water droplets both the scattering and absorption coefficients, thus the single scattering albedo, are functions of wavelength l and droplet size r. This note shows that for water droplets at weakly absorbing wavelengths, the ratio w_0l(r)/w_0l(r0) of two single scattering albedo spectra is a linear function of w_0l(r). The slope and intercept of the linear function are wavelength independent and sum to unity. This relationship allows for a representation of any single scattering albedo spectrum w_0l(r) via one known spectrum w_0l(r0). We provide a simple physical explanation of the discovered relationship. Similar linear relationships were found for the single scattering albedo spectra of non-spherical ice crystals.

Stratospheric water vapor and climate: sensitivity to the representation in radiation codes

There has been considerable interest in the climate impact of trends in stratospheric water vapor (SWV). However, the representation of the radiative properties of water vapor under stratospheric conditions remains poorly constrained across different radiation codes. This study examines the sensitivity of a detailed line-by-line (LBL) code, a Malkmus narrow-band model and two broadband GCM radiation codes to a uniform perturbation in SWV in the longwave spectral region. The choice of sampling rate in wave number space (Δν) in the LBL code is shown to be important for calculations of the instantaneous change in heating rate (ΔQ) and the instantaneous longwave radiative forcing (ΔFtrop). ΔQ varies by up to 50% for values of Δν spanning 5 orders of magnitude, and ΔFtrop varies by up to 10%. In the three less detailed codes, ΔQ differs by up to 45% at 100 hPa and 50% at 1 hPa compared to a LBL calculation. This causes differences of up to 70% in the equilibrium fixed dynamical heating temperature change due to the SWV perturbation. The stratosphere-adjusted radiative forcing differs by up to 96% across the less detailed codes. The results highlight an important source of uncertainty in quantifying and modeling the links between SWV trends and climate.

Spectral theory of some non-selfadjoint linear differential operators

We give a characterisation of the spectral properties of linear differential operators with constant coefficients, acting on functions defined on a bounded interval, and determined by general linear boundary conditions. The boundary conditions may be such that the resulting operator is not selfadjoint. We associate the spectral properties of such an operator $S$ with the properties of the solution of a corresponding boundary value problem for the partial differential equation $\partial_t q \pm iSq=0$. Namely, we are able to establish an explicit correspondence between the properties of the family of eigenfunctions of the operator, and in particular whether this family is a basis, and the existence and properties of the unique solution of the associated boundary value problem. When such a unique solution exists, we consider its representation as a complex contour integral that is obtained using a transform method recently proposed by Fokas and one of the authors. The analyticity properties of the integrand in this representation are crucial for studying the spectral theory of the associated operator.

Spectral gap for Glauber type dynamics for a special class of potentials

We consider an equilibrium birth and death type process for a particle system in infinite volume, the latter is described by the space of all locally finite point configurations on Rd. These Glauber type dynamics are Markov processes constructed for pre-given reversible measures. A representation for the ``carré du champ'' and ``second carré du champ'' for the associate infinitesimal generators L are calculated in infinite volume and for a large class of functions in a generalized sense. The corresponding coercivity identity is derived and explicit sufficient conditions for the appearance and bounds for the size of the spectral gap of L are given. These techniques are applied to Glauber dynamics associated to Gibbs measure and conditions are derived extending all previous known results and, in particular, potentials with negative parts can now be treated. The high temperature regime is extended essentially and potentials with non-trivial negative part can be included. Furthermore, a special class of potentials is defined for which the size of the spectral gap is as least as large as for the free system and, surprisingly, the spectral gap is independent of the activity. This type of potentials should not show any phase transition for a given temperature at any activity.

On the relationship between a 2 x 2 matrix and second-order scalar spectral problems for integrable equations

A simple proof is given that a 2 x 2 matrix scheme for an inverse scattering transform method for integrable equations can be converted into the standard form of the second-order scalar spectral problem associated with the same equations. Simple formulae relating these two kinds of representation of integrable equations are established.

Representation theorems for indefinite quadratic forms and applications

Diese Arbeit widmet sich den Darstellungssätzen für symmetrische indefinite (das heißt nicht-halbbeschränkte) Sesquilinearformen und deren Anwendungen. Insbesondere betrachten wir den Fall, dass der zur Form assoziierte Operator keine Spektrallücke um Null besitzt. Desweiteren untersuchen wir die Beziehung zwischen reduzierenden Graphräumen, Lösungen von Operator-Riccati-Gleichungen und der Block-Diagonalisierung für diagonaldominante Block-Operator-Matrizen. Mit Hilfe der Darstellungssätze wird eine entsprechende Beziehung zwischen Operatoren, die zu indefiniten Formen assoziiert sind, und Form-Riccati-Gleichungen erreicht. In diesem Rahmen wird eine explizite Block-Diagonalisierung und eine Spektralzerlegung für den Stokes Operator sowie eine Darstellung für dessen Kern erreicht. Wir wenden die Darstellungssätze auf durch (grad u, h() grad v) gegebene Formen an, wobei Vorzeichen-indefinite Koeffzienten-Matrizen h() zugelassen sind. Als ein Resultat werden selbstadjungierte indefinite Differentialoperatoren div h() grad mit homogenen Dirichlet oder Neumann Randbedingungen konstruiert. Beispiele solcher Art sind Operatoren die in der Modellierung von optischen Metamaterialien auftauchen und links-indefinite Sturm-Liouville Operatoren.

EEG reactivity and EEG activity in never-treated acute schizophrenics, measured with spectral parameters and dimensional complexity

Our approaches to the use of EEG studies for the understanding of the pathogenesis of schizophrenic symptoms are presented. The basic assumptions of a heuristic and multifactorial model of the psychobiological brain mechanisms underlying the organization of normal behavior is described and used in order to formulate and test hypotheses about the pathogenesis of schizophrenic behavior using EEG measures. Results from our studies on EEG activity and EEG reactivity (= EEG components of a memory-driven, adaptive, non-unitary orienting response) as analyzed with spectral parameters and "chaotic" dimensionality (correlation dimension) are summarized. Both analysis procedures showed a deviant brain functional organization in never-treated first-episode schizophrenia which, within the framework of the model, suggests as common denominator for the pathogenesis of the symptoms a deviation of working memory, the nature of which is functional and not structural.