GPU-Powered Shotgun Stochastic Search for Dirichlet Process Mixtures of Gaussian Graphical Models

Gaussian graphical models (GGMs) are popular for modeling high-dimensional multivariate data with sparse conditional dependencies. A mixture of GGMs extends this model to the more realistic scenario where observations come from a heterogenous population composed of a small number of homogeneous subgroups. In this article, we present a novel stochastic search algorithm for finding the posterior mode of high-dimensional Dirichlet process mixtures of decomposable GGMs. Further, we investigate how to harness the massive thread-parallelization capabilities of graphical processing units to accelerate computation. The computational advantages of our algorithms are demonstrated with various simulated data examples in which we compare our stochastic search with a Markov chain Monte Carlo (MCMC) algorithm in moderate dimensional data examples. These experiments show that our stochastic search largely outperforms the MCMC algorithm in terms of computing-times and in terms of the quality of the posterior mode discovered. Finally, we analyze a gene expression dataset in which MCMC algorithms are too slow to be practically useful.     

Estimates of Orders and Parameters in General ARMA Models

Let X(n)be a time series satisfying the following general ARMA(p,d,r,q)model: E(B)U(B)A(B)X(n)=C(B)W(n),whereC(z)is relatively prime with the polynomial E(z)U(z)A(z),B is the backshiftoperator such that BX(n)=X(n-1),and(W(n),F(n),n≥1)is a sequence ofmartingale differences. For simplicity,we shall assume throughout that the initial values(X(-p-d     

Convergence Rates of the Distributions of Error Variance Estimates in Linear Models

The best possible rate of convergence of the distributions of error variance estimates in linear models, based on the residual sum of squares, is obtained under weakest possible conditions.     

Restricted Maximum Likelihood Estimates in Finite Mixture Models

ＲｅｓｔｒｉｃｔｅｄＭａｘｉｍｕｍＬｉｋｅｌｉｈｏｏｄＥｓｔｉｍａｔｅｓｉｎＦｉｎｉｔｅＭｉｘｔｕｒｅＭｏｄｅｌｓ￥（陈家骅，成平）ＣｈｅｎＪｉａｈｕａ（ＤｅｐａｒｔｍｅｎｔｏｆＳｔａｔｉｓｔｉｃｓ＆ＡｃｔｕａｒｉａｌＳｃｉｅｎｃｅ，Ｕｎｉｖｅｒｓｉ...     

Efficient Computation of First Passage Times in Kou’s Jump-diffusion Model

S. G. Kou and H. Wang [First passage times of a jump diffusion process, Adv. Appl. Probab. 35 (2003) 504–531] give expressions of both the real Laplace transform of the distribution of first passage time and the real Laplace transform of the joint distribution of the first passage time and the running maxima of a jump-diffusion model called Kou model. These authors invert the former Laplace transform by using Gaver-Stehfest algorithm, and for the latter they need a large computing time with an algebra computer system. In the present paper, we give a much simpler expression of the Laplace transform of the joint distribution, and we also show, using Complex Analysis techniques, that both Laplace transforms can be extended to the complex plane. Hence, we can use inversion methods based on the complex inversion formula or Bromwich integral which are very efficent. The improvement in the computing times and accuracy is remarkable.     

Quantile regression with ℓ 1—regularization and Gaussian kernels

The quantile regression problem is considered by learning schemes based on ? ₁—regularization and Gaussian kernels. The purpose of this paper is to present concentration estimates for the algorithms. Our analysis shows that the convergence behavior of ? ₁—quantile regression with Gaussian kernels is almost the same as that of the RKHS-based learning schemes. Furthermore, the previous analysis for kernel-based quantile regression usually requires that the output sample values are uniformly bounded, which excludes the common case with Gaussian noise. Our error analysis presented in this paper can give satisfactory convergence rates even for unbounded sampling processes. Besides, numerical experiments are given which support the theoretical results.     

Analysis of a General Family of Regularized Navier–Stokes and MHD Models

We consider a general family of regularized Navier–Stokes and Magnetohydrodynamics (MHD) models on n-dimensional smooth compact Riemannian manifolds with or without boundary, with n≥2. This family captures most of the specific regularized models that have been proposed and analyzed in the literature, including the Navier–Stokes equations, the Navier–Stokes-α model, the Leray-α model, the modified Leray-α model, the simplified Bardina model, the Navier–Stokes–Voight model, the Navier–Stokes-α-like models, and certain MHD models, in addition to representing a larger 3-parameter family of models not previously analyzed. This family of models has become particularly important in the development of mathematical and computational models of turbulence. We give a unified analysis of the entire three-parameter family of models using only abstract mapping properties of the principal dissipation and smoothing operators, and then use assumptions about the specific form of the parameterizations, leading to specific models, only when necessary to obtain the sharpest results. We first establish existence and regularity results, and under appropriate assumptions show uniqueness and stability. We then establish some results for singular perturbations, which as special cases include the inviscid limit of viscous models and the α→0 limit in α models. Next, we show existence of a global attractor for the general model, and then give estimates for the dimension of the global attractor and the number of degrees of freedom in terms of a generalized Grashof number. We then establish some results on determining operators for the two distinct subfamilies of dissipative and non-dissipative models. We finish by deriving some new length-scale estimates in terms of the Reynolds number, which allows for recasting the Grashof number-based results into analogous statements involving the Reynolds number. In addition to recovering most of the existing results on existence, regularity, uniqueness, stability, attractor existence, and dimension, and determining operators for the well-known specific members of this family of regularized Navier–Stokes and MHD models, the framework we develop also makes possible a number of new results for all models in the general family, including some new results for several of the well-studied models. Analyzing the more abstract generalized model allows for a simpler analysis that helps bring out the core common structure of the various regularized Navier–Stokes and magnetohydrodynamics models, and also helps clarify the common features of many of the existing and new results. To make the paper reasonably self-contained, we include supporting material on spaces involving time, Sobolev spaces, and Grönwall-type inequalities.     

Convergence of the reweighted ℓ 1 minimization algorithm for ℓ 2–ℓ p minimization

The iteratively reweighted ? ₁ minimization algorithm (IRL1) has been widely used for variable selection, signal reconstruction and image processing. In this paper, we show that any sequence generated by the IRL1 is bounded and any accumulation point is a stationary point of the ? ₂–? _p minimization problem with 0<p<1. Moreover, the stationary point is a global minimizer and the convergence rate is approximately linear under certain conditions. We derive posteriori error bounds which can be used to construct practical stopping rules for the algorithm.     

Shift invariant preduals of ℓ 1(ℤ)

The Banach space ? ₁(?) admits many non-isomorphic preduals, for example, C(K) for any compact countable space K, along with many more exotic Banach spaces. In this paper, we impose an extra condition: the predual must make the bilateral shift on ? ₁(?) weak*-continuous. This is equivalent to making the natural convolution multiplication on ? ₁(?) separately weak*-continuous and so turning ? ₁(?) into a dual Banach algebra. We call such preduals shift-invariant. It is known that the only shift-invariant predual arising from the standard duality between C ₀(K) (for countable locally compact K) and ? ₁(?) is c ₀(?). We provide an explicit construction of an uncountable family of distinct preduals which do make the bilateral shift weak*-continuous. Using Szlenk index arguments, we show that merely as Banach spaces, these are all isomorphic to c ₀. We then build some theory to study such preduals, showing that they arise from certain semigroup compactifications of ?. This allows us to produce a large number of other examples, including non-isometric preduals, and preduals which are not Banach space isomorphic to c ₀.     

ℓ1-summability of higher-dimensional Fourier series

It is proved that the maximal operator of the ${?}_1$-Fejér means of a $d$-dimensional Fourier series is bounded from the periodic Hardy space $H_p\left({\mathbb{T}}^d\right)$ to $L_p\left({\mathbb{T}}^d\right)$ for all $d/(d+1)<p\le \infty$ and, consequently, is of weak type (1, 1). As a consequence we obtain that the ${?}_1$-Fejér means of a function $f\in L_1\left({\mathbb{T}}^d\right)$ converge a.e. to $f$. Moreover, we prove that the ${?}_1$-Fejér means are uniformly bounded on the spaces $H_p\left({\mathbb{T}}^d\right)$ and so they converge in norm $(d/(d+1)<p<\infty )$. Similar results are shown for conjugate functions and for a general summability method, called $\theta$-summability. Some special cases of the ${?}_1–\theta$-summation are considered, such as the Weierstrass, Picard, Bessel, Fejér, de la Vallée Poussin, Rogosinski and Riesz summations.     

Norm-one projections onto subspaces of finite codimension in ℓ1 andc 0

We study 1-complemented subspaces of the sequence spaces ₁ andc ₀. In ₁, 1-complemented subspaces of codimensionn are those which can be obtained as intersection ofn 1-complemented hyperplanes. Inc ₀, we prove a characterization of 1-complemented subspaces of finite codimension in terms of intersection of hyperplanes.Work prepared under the auspices of GNAFA-CNR (National Council of Research) and Minister of Public Instruction of Italy.     

Finite Representability of ℓp in Quotients of Banach Spaces

A typical result of the paper states that if X is a Banach space with a basis and for some 1pq, the spaces _p and _q are finitely block representable in every block subspace of X, then every block subspace of X admits a block quotient Z such that for every r[p,q], the space _r is finitely block representable in Z. Results of a similar nature are also established for ^N _p-block-sequences and asymptotic spaces.     

GLMMLasso: An Algorithm for High-Dimensional Generalized Linear Mixed Models Using ℓ1-Penalization

We propose an ?₁-penalized algorithm for fitting high-dimensional generalized linear mixed models (GLMMs). GLMMs can be viewed as an extension of generalized linear models for clustered observations. Our Lasso-type approach for GLMMs should be mainly used as variable screening method to reduce the number of variables below the sample size. We then suggest a refitting by maximum likelihood based on the selected variables only. This is an effective correction to overcome problems stemming from the variable screening procedure that are more severe with GLMMs than for generalized linear models. We illustrate the performance of our algorithm on simulated as well as on real data examples. Supplementary materials are available online and the algorithm is implemented in the R package glmmixedlasso.     

Distribution of Increasing ℓ-sequences in a Random Permutation

This paper examines the distribution of the number, k, of increasing -sequences in a random permutation of . A new solution is determined based on the compositions of n which requires, at most, summands. This solution easily yields existing results for the special case and provides an alternate form for the case . The expected number of increasing -sequences in a random permutation is determined and it is shown that the limiting distribution is degenerate about 0 for 2$$ " align="middle" border="0"> . An alternate algorithm to determine the exact distribution is presented, based on the partitions of n, which is easy to implement and efficient for small n. Applications in non-parametric statistics and graph theory are discussed.     

Gaussian radial‐basis functions: Cardinal interpolation of ℓp and power‐growth data

Suppose λ is a positive number. Basic theory of cardinal interpolation ensures the existence of the Gaussian cardinal function \(L_\lambda (x) = \sum\nolimits_{k \in \mathbb{Z}} {c_k \exp ( - \lambda (x - k)^2 ),x \in \mathbb{R}} ,\) satisfying the interpolatory conditions \(L_\lambda (j) = \delta _{0j} ,j \in \mathbb{Z}.\) The paper considers the Gaussian cardinal interpolation operator

$(\mathcal{L}_\lambda {\text{y}})(x): = \sum\limits_{k \in \mathbb{Z}} {y_k L_\lambda (x - k),{\text{ y}} = (y_k )_{k \in \mathbb{Z}} ,{\text{ }}x \in \mathbb{R}} ,$

as a linear mapping from ℓ^p(ℤ) into L ^p(ℝ), 1≤ p ∞, and in particular, its behaviour as λ→0⁺. It is shown that \(\left\| {\mathcal{L}_\lambda } \right\|_p \) is uniformly bounded (in λ) for 1 < p < ∞, and that \(\left\| {\mathcal{L}_\lambda } \right\|_1 \asymp \log (1/\lambda )\) as λ→0⁺. The limiting behaviour is seen to be that of the classical Whittaker operator

$\mathcal{W}:{\text{y}} \mapsto \sum\limits_{k \in \mathbb{Z}} {y_k \frac{{\sin \pi (x - k)}}{{\pi (x - k)}}} ,$

in that \(\lim _{\lambda \to 0^ + } \left\| {\mathcal{L}_\lambda {\text{y}} - \mathcal{W}{\text{y}}} \right\|_p = 0,\) for every \({\text{y}} \in \ell ^p (\mathbb{Z}){\text{ and }}1 < p < \infty .\) It is further shown that the Gaussian cardinal interpolants to a function f which is the Fourier transform of a tempered distribution supported in (-π,π) converge locally uniformly to f as λ→0⁺. Multidimensional extensions of these results are also discussed.