稳健矩阵回归模型和方法研究

随着大数据时代的到来,我们面临的数据越来越复杂,其中待估系数为矩阵的模型亟待构造和求解.无论在统计还是优化领域,许多专家学者都致力于矩阵模型的统计性质分析及寻找其最优解的算法设计.当随机误差期望为0且同方差时,采用基于最小二乘的模型可以很好地解决问题.当随机误差异方差,分布为重尾分布(如双指数分布,t-分布等)或数据含有异常值时,需要考虑稳健的方法来求解问题.常用的稳健方法有最小一乘,分位数,Huber等.目前稳健方法的研究大多集中于线性回归问题,对于矩阵回归问题的研究比较缺乏.本文从最小二乘模型讲起,对矩阵回归问题进行了总结和评述,同时列出了一些文献和简要介绍了我们的近期的部分工作.最后对于稳健矩阵回归,我们提出了一些展望和设想.     

Testing Lack-of-fit for a Polynomial Errors-in-variables Model

When a regression model is applied as an approximation of underlying model of data, the model checking is important and relevant. In this paper, we investigate the lack-of-fit test for a polynomial error-in-variables model. As the ordinary residuals are biased when there exist measurement errors in covariables,we correct them and then construct a residual-based test of score type. The constructed test is asymptotically chi-squared under null hypotheses. Simulation study shows that the test can maintain the significance level well.The choice of weight functions involved in the test statistic and the related power study are also investigated.The application to two examples is illustrated. The approach can be readily extended to handle more general models.     

Upper and lower approximation models in interval regression using regression quantile techniques

In this paper, we propose new interval regression analysis based on the regression quantile techniques. To analyze a phenomenon in a fuzzy environment, we propose two interval approximation models. Without using all data, we first identify the main trend from the designated proportion of the given data. To select the main part of data to be analyzed, we introduce the regression quantile techniques. The obtained model is not influenced by extreme points since it is formulated from the center-located main proportion of the given data. After that, the interval regression model including all data can be identified based on the acquired main trend. The obtained interval regression model by the main proportion of the given data is called the lower approximation model, while interval regression model by all data is called the upper approximation model for the given phenomenon. Also it is shown that, from the lower approximation model (main trend) and the upper approximation model, we can construct a trapezoidal fuzzy model. The membership function of this fuzzy model is useful to obtain the locational information for each observation. The characteristic of our approach can be described as obtaining the upper and lower approximation models and combining them to be a fuzzy model for representing the given phenomenon in a fuzzy environment.     

Two optimization problems in linear regression with interval data

We consider a linear regression model where neither regressors nor the dependent variable is observable; only intervals are available which are assumed to cover the unobservable data points. Our task is to compute tight bounds for the residual errors of minimum-norm estimators of regression parameters with various norms (corresponding to least absolute deviations (LAD), ordinary least squares (OLS), generalized least squares (GLS) and Chebyshev approximation). The computation of the error bounds can be formulated as a pair of max–min and min–min box-constrained optimization problems. We give a detailed complexity-theoretic analysis of them. First, we prove that they are NP-hard in general. Then, further analysis explains the sources of NP-hardness. We investigate three restrictions when the problem is solvable in polynomial time: the case when the parameter space is known apriori to be restricted into a particular orthant, the case when the regression model has a fixed number of regression parameters, and the case when only the dependent variable is observed with errors. We propose a method, called orthant decomposition of the parameter space, which is the main tool for obtaining polynomial-time computability results.     

Partially finite convex programming,Part II: Explicit lattice models

In Part I of this work we derived a duality theorem for partially finite convex programs, problems for which the standard Slater condition fails almost invariably. Our result depended on a constraint qualification involving the notion ofquasi relative interior. The derivation of the primal solution from a dual solution depended on the differentiability of the dual objective function: the differentiability of various convex functions in lattices was considered at the end of Part I. In Part II we shall apply our results to a number of more concrete problems, including variants of semi-infinite linear programming,L ¹ approximation, constrained approximation and interpolation, spectral estimation, semi-infinite transportation problems and the generalized market area problem of Lowe and Hurter (1976). As in Part I, we shall use lattice notation extensively, but, as we illustrated there, in concrete examples lattice-theoretic ideas can be avoided, if preferred, by direct calculation.     

Efcient Estimation of Varying Coefcient Seemly Unrelated Regression Model

In this paper,we propose a class of varying coefcient seemingly unrelated regression models,in which the errors are correlated across the equations.By applying the series approximation and taking the contemporaneous correlations into account,we propose an efcient generalized least squares series estimation for the unknown coefcient functions.The consistency and asymptotic normality of the resulting estimators are established.In comparison with the ordinary least squares ones,the proposed estimators are more efcient with smaller asymptotical variances.Some simulation studies and a real application are presented to demonstrate the finite sample performance of the proposed methods.In addition,based on a B-spline approximation,we deduce the asymptotic bias and variance of the proposed estimators.     

半线性椭圆问题Petrov-Galerkin逼近及亏量迭代

本文考虑了二阶半线性椭圆问题的Petrov-Galerkin逼近格式,用双二次多项式空间作为形函数空间,用双线性多项式空间作为试探函数空间,证明了此逼近格式与标准的二次有限元逼近格式有同样的收敛阶.并且根据插值算子的逼近性质,进一步证明了半线性有限元解的亏量迭代序列收敛到Petrov-Galerkin解.     

A POSTERIORI ERROR ESTIMATE FOR BOUNDARY CONTROL PROBLEMS GOVERNED BY THE PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

In this paper, we discuss the a posteriori error estimate of the finite element approximation for the boundary control problems governed by the parabolic partial differential equations. Three different a posteriori error estimators are provided for the parabolic boundary control problems with the observations of the distributed state, the boundary state and the final state. It is proven that these estimators are reliable bounds of the finite element approximation errors, which can be used as the indicators of the mesh refinement in adaptive finite element methods.     

Efficient estimation of varying coefficient seemly unrelated regression model

In this paper, we propose a class of varying coefficient seemingly unrelated regression models, in which the errors are correlated across the equations. By applying the series approximation and taking the contemporaneous correlations into account, we propose an efficient generalized least squares series estimation for the unknown coefficient functions. The consistency and asymptotic normality of the resulting estimators are established. In comparison with the ordinary/east squares ones, the proposed estimators are more efficient with smaller asymptotical variances. Some simulgtlon＇studies and a real application are presented to demonstrate the finite sample performance of the proposed methods. In addition, based on a B-spline approximation, we deduce the asymptotic bias and variance of the proposed estimators.     

Continuous-Time Generalized Fractional Programming Problems. Part I: Basic Theory

This study, that will be presented as two parts, develops a computational approach to a class of continuous-time generalized fractional programming problems. The parametric method for finite-dimensional generalized fractional programming is extended to problems posed in function spaces. The developed method is a hybrid of the parametric method and discretization approach. In this paper (Part I), some properties of continuous-time optimization problems in parametric form pertaining to continuous-time generalized fractional programming problems are derived. These properties make it possible to develop a computational procedure for continuous-time generalized fractional programming problems. However, it is notoriously difficult to find the exact solutions of continuous-time optimization problems. In the accompanying paper (Part II), a further computational procedure with approximation will be proposed. This procedure will yield bounds on errors introduced by the numerical approximation. In addition, both the size of discretization and the precision of an approximation approach depend on predefined parameters.     

Learning errors of linear programming support vector regression

In this paper, we give several results of learning errors for linear programming support vector regression. The corresponding theorems are proved in the reproducing kernel Hilbert space. With the covering number, the approximation property and the capacity of the reproducing kernel Hilbert space are measured. The obtained result (Theorem 2.1) shows that the learning error can be controlled by the sample error and regularization error. The mentioned sample error is summarized by the errors of learning regression function and regularizing function in the reproducing kernel Hilbert space. After estimating the generalization error of learning regression function (Theorem 2.2), the upper bound (Theorem 2.3) of the regularized learning algorithm associated with linear programming support vector regression is estimated.     

THE DEFECT ITERATION OF THE FINITE ELEMENT FOR ELLIPTIC BOUNDARY VALUE PROBLEMS AND PETROV-GALERKIN APPROXIMATION

1.IntroductionFranketc.of.[l]establishedtheiterateddefectcorrectionschemeforfiniteelemelltofellipticboundaryproblems.FOrlinearellipticboundaryvalueproblem[2--5]havediscllssedtheefficiencyoftheschemebyusillgsuperconvergenceandasymptoticexpansion"lidertheco…     

Ridge regression in two‐parameter solution

We consider simultaneous minimization of the model errors, deviations from orthogonality between regressors and errors, and deviations from other desired properties of the solution. This approach corresponds to a regularized objective that produces a consistent solution not prone to multicollinearity. We obtain a generalization of the ridge regression to two‐parameter model that always outperforms a regular one‐parameter ridge by better approximation, and has good properties of orthogonality between residuals and predicted values of the dependent variable. The results are very convenient for the analysis and interpretation of the regression. Numerical runs prove that this technique works very well. The examples are considered for marketing research problems. Copyright © 2005 John Wiley & Sons, Ltd.     

Standard errors for the cox proportional hazards cure model

Standard errors for the maximum likelihood estimates of the regression parameters in the logistic-proportional-hazards cure model are proposed using an approximate profile likelihood approach and a nonparametric likelihood. Two methods are given and are compared with the standard errors obtained from the inverse of the joint observed information matrix of the regression parameters and the nuisance hazard parameters. The observed information matrix is derived and is shown to be an approximation of the conditional information matrix of the regression parameters given the hazard parameters. Simulations indicate that the standard errors obtained from the inverse of the observed information matrix based on the profile likelihood and the full likelihood are comparable and appropriate. The coverage rates for the logistic regression parameter are generally good. The proportional hazards regression parameter show reasonable coverage rates under ideal conditions but lower coverage rates when the incidence proportion is low or when censoring is heavy. The three methods are applied to a data set to investigate the effects of radiation therapy on tonsil cancer.     

A Minmax Regret Linear Regression Model Under Uncertainty in the Dependent Variable

This paper analyzes the simple linear regression model corresponding to a sample affected by errors from a non-probabilistic viewpoint. We consider the simplest case where the errors just affect the dependent variable and there only exists one explanatory variable. Moreover, we assume the errors affecting each observation can be bounded. In this context the minmax regret criterion is used in order to obtain a regression line with nearly optimal goodness of fit for any true values of the dependent variable. Theoretical results as well as numerical methods are stated in order to solve the optimization problem under different residual cost functions.     

Testing in linear composite quantile regression models

Composite quantile regression (CQR) can be more efficient and sometimes arbitrarily more efficient than least squares for non-normal random errors, and almost as efficient for normal random errors. Based on CQR, we propose a test method to deal with the testing problem of the parameter in the linear regression models. The critical values of the test statistic can be obtained by the random weighting method without estimating the nuisance parameters. A distinguished feature of the proposed method is that the approximation is valid even the null hypothesis is not true and power evaluation is possible under the local alternatives. Extensive simulations are reported, showing that the proposed method works well in practical settings. The proposed methods are also applied to a data set from a walking behavior survey.     

Some Properties and Improvements of the Saddlepoint Approximation in Nonlinear Regression

We summarize properties of the saddlepoint approximation of the density of the maximum likelihood estimator in nonlinear regression with normal errors: accuracy, range of validity, equivariance. We give a geometric insight into the accuracy of the saddlepoint density for finite samples. The role of the Riemannian curvature tensor in the whole investigation of the properties is demonstrated. By adding terms containing this tensor we improve the saddlepoint approximation. When this tensor is zero, or when the number of observations is large, we have pivotal, independent, and ² distributed variables, like in a linear model. Consequences for experimental design or for constructions of confidence regions are discussed.     

Asymptotic Solutions of Nonrelativistic Equations of Quantum Mechanics in Curved Nanotubes: I. Reduction to Spatially One-Dimensional Equations

We consider equations of nonrelativistic quantum mechanics in thin three-dimensional tubes (nanotubes). We suggest a version of the adiabatic approximation that permits reducing the original three-dimensional equations to one-dimensional equations for a wide range of energies of longitudinal motion. The suggested reduction method (the operator method for separating the variables) is based on the Maslov operator method. We classify the solutions of the reduced one-dimensional equation. In Part I of this paper, we deal with the reduction problem, consider the main ideas of the operator separation of variables (in the adiabatic approximation), and derive the reduced equations. In Part II, we will discuss various asymptotic solutions and several effects described by these solutions.     

Lagrange插值在—重积分Wiener空间下的同时逼近平均误差

在加权L_p范数逼近意义下,确定了基于扩充的第二类Chebyshev结点组的Lagrange插值多项式列,在一重积分Wiener空间下同时逼近平均误差的渐近阶.结果显示,在L_p范数逼近意义下,Lagrange插值多项式列逼近函数及其导数的平均误差都弱等价于相应的最佳逼近多项式列的平均误差.同时,在信息基复杂性的意义下,若可允许信息泛函为标准信息,则上述插值算子列逼近函数及其导数的平均误差均弱等价于相应的最小非自适应信息半径.     

基于最小二乘近似的模型平均方法

本文提出基于最小二乘近似的模型平均方法.该方法可用于线性模型、广义线性模型和分位数回归等各种常用模型.特别地,经典的Mallows模型平均方法是该方法的特例.现存的模型平均文献中,渐近分布的证明一般需要局部误设定假设,所得的极限分布的形式也比较复杂.本文将在不使用局部误设定假设的情形下证明该方法的渐近正态性.另外,本文...