Parameter Estimation for Multivariate Generalized Gaussian Distributions Fred´ Eric´ Pascal, Lionel Bombrun, Jean-Yves Tourneret and Yannick Berthoumieu

Total Page:16

File Type:pdf, Size:1020Kb

Parameter Estimation for Multivariate Generalized Gaussian Distributions Fred´ Eric´ Pascal, Lionel Bombrun, Jean-Yves Tourneret and Yannick Berthoumieu SUBMITTED TO IEEE TRANS. ON SIGNAL PROCESSING 1 Parameter Estimation For Multivariate Generalized Gaussian Distributions Fred´ eric´ Pascal, Lionel Bombrun, Jean-Yves Tourneret and Yannick Berthoumieu. Abstract—Due to its heavy-tailed and fully parametric form, include radar [23], video coding and denoising [24]–[26] or the multivariate generalized Gaussian distribution (MGGD) has biomedical signal processing [25], [27], [28]. Finally, it is been receiving much attention for modeling extreme events interesting to note that complex GGDs have been recently in signal and image processing applications. Considering the estimation issue of the MGGD parameters, the main contribution studied in [29], [30] and that multivariate regression models of this paper is to prove that the maximum likelihood estimator with generalized Gaussian errors have been considered in [31]. (MLE) of the scatter matrix exists and is unique up to a scalar Considering the important attention devoted to GGDs, es- factor, for a given shape parameter β 2 (0; 1). Moreover, an timating the parameters of these distributions is clearly an estimation algorithm based on a Newton-Raphson recursion is interesting issue. Classical estimation methods that have been proposed for computing the MLE of MGGD parameters. Various experiments conducted on synthetic and real data are presented investigated for univariate GGDs include the maximum likeli- to illustrate the theoretical derivations in terms of number of hood (ML) method [32] and the method of moments [33]. In iterations and number of samples for different values of the the multivariate context, MGGD parameters can be estimated shape parameter. The main conclusion of this work is that the by a least-squares method as in [17] or by minimizing a parameters of MGGDs can be estimated using the maximum χ2 distance between the histogram of the observed data and likelihood principle with good performance. the theoretical probabilities associated with the MGGD [34]. Index Terms—Multivariate generalized Gaussian distribution, Estimators based on the method of moments and on the ML covariance matrix estimation, fixed point algorithm. method have also been proposed in [35]–[37]. Several works have analyzed covariance matrix estimators I. INTRODUCTION defined under different modeling assumptions. On the one hand, fixed point (FP) algorithms have been derived and NIVARIATE and multivariate generalized Gaussian dis- analyzed in [38], [39] for SIRVs. On the other hand, in the tributions (GGDs) have received much attention in the U context of robust estimation, the properties of M-estimators literature. Historically, this family of distributions has been have been studied by Maronna in [40]. Unfortunately, introduced in [1]. Some properties of these distributions have Maronna’s conditions are not fully satisfied for MGGDs been reported in several papers such as [2]–[4]. These prop- (see remark II.3). This paper shows that despite the non- erties include various stochastic representations, simulation applicability of Maronna’s results, the MLE of MGGD methods and probabilistic characteristics. GGDs belong to parameters exists, is unique and can be computed by an FP the family of elliptical distributions (EDs) [5], [6], originally algorithm. Although the methodology adopted in this paper introduced by Kelker in [7] and studied in [8], [9]. For has some similarities with the one proposed in [38], [39], β 2 (0; 1], Multivariate GGDs (MGGDs) are a subset of the there are also important differences which require a specific spherically invariant random vector (SIRV) distributions. For analysis (see for instance remark III.1). More precisely, the β > 1, MGGDs are no longer SIRV distributions as illustrated FP equation of [38] corresponds to an approximate MLE for in Fig. ?? (for more details, see [10]). SIRVs while in [39] the FP equation results from a different MGGDs have been used intensively in the image processing problem (see Eq. (14) in [39] compared to Eq. (15) of this arXiv:1302.6498v2 [stat.AP] 24 Feb 2017 community. Indeed, including Gaussian and Laplacian distri- paper). The contributions of this paper are to establish some butions as special cases, MGGDs are potentially interesting for properties related to the FP equation of the ML estimator modeling the statistical properties of various images or features for MGGDs. More precisely, we show that for a given shape extracted from these images. In particular, the distribution parameter β belonging to (0; 1), the MLE of the scatter of wavelet or curvelet coefficients has been shown to be matrix M exists and is unique up to a scalar factor1. An modeled accurately by GGDs [11]–[14]. This property has iterative algorithm based on a Newton-Raphson procedure is been exploited for many image processing applications includ- then proposed to compute the MLE of M. ing image denoising [15]–[18], context-based image retrieval [19], [20], image thresholding [21] or texture classification in The paper is organized as follows. SectionII defines the industrial problems [22]. Other applications involving GGDs MGGDs considered in this study and derives the MLEs of their F. Pascal is with Supelec/SONDRA,´ 91192 Gif-sur-Yvette Cedex, France parameters. Section III presents the main theoretical results (e-mail: [email protected]) of this paper while a proof outline is given in SectionIV. L. Bombrun and Y. Berthoumieu are with Universite´ de Bordeaux, IPB, For presentation clarity, full demonstrations are provided in ENSEIRB-Matmeca, Laboratoire IMS, France (e-mail: lionel.bombrun@ims- bordeaux.fr; [email protected]) J.-Y. Tourneret is with Universite´ de Toulouse, IRIT/INP-ENSEEIHT, (e- 1From the submission of this paper, another approach based on geodesic mail:[email protected]) convexity was proposed in (include reference paper Wiesel). 2 SUBMITTED TO IEEE TRANS. ON SIGNAL PROCESSING the appendices. SectionV is devoted to simulation results elliptical distribution that has received much attention in the conducted on synthetic and real data. The convergence speed literature. Following the results of [44] for real elliptical of the proposed estimation algorithm as well as the bias distributions, by differentiating the log-likelihood of vectors and consistency of the scatter matrix MLE are first inves- (x1;:::; xN ) with respect to M, the MLE of the matrix M tigated using synthetic data. Experimentations performed on satisfies the following FP equation real images extracted from the VisTex database are then N T −1 2 X −gm,β(x M xi) presented. Conclusions and future works are finally reported M = i x xT (5) N h (xT M−1x ) i i in SectionVI. i=1 m,β i i II. PROBLEM FORMULATION where gm,β(y) = @hm,β(y)=@y. In the particular case of an MGGD with known parameters m and β, straightforward A. Definitions computations lead to The probability density function of an MGGD in Rp is N T defined by [41] β X xix M = i : (6) 1 Nmβ T −1 1−β T −1 i=1 xi M xi p(xjM; m; β) = 1 hm,β x M x (1) jMj 2 When the parameter m is unknown, the MLEs of M and m are p for any x 2 R , where M is a p × p symmetric real scatter obtained by differentiating the log-likelihood of (x1;:::; xN ) T matrix, x is the transpose of the vector x, and hm,β (·) is a with respect to M and m yielding so-called density generator defined by β N x xT p β X i i βΓ 2 1 y M = β 1−β ; (7) h (y) = p exp − Nm T −1 m,β p p (2) x M x p 2 β i=1 i i 2 2β m 2m π Γ 2β 2 1 " N # β β X β for any y 2 +, where m and β are the MGGD scale and m = xT M−1x : (8) R pN i i shape parameters. The matrix M will be normalized in this i=1 paper according to Tr (M) = p, where Tr(M) is the trace After replacing m in (6) by its expression (8), the following of the matrix M. It is interesting to note that letting β = 1 result can be obtained corresponds to the multivariate Gaussian distribution. More- N over, when β tends toward infinity, the MGGD is known to 1 X Np T M = xix : (9) converge in distribution to a multivariate uniform distribution N 1−β X β i i=1 yi + yi yj (see (3)). j6=i As mentioned before and confirmed by (9), M can be esti- B. Stochastic representation mated independently from the scale parameter m. Moreover, p Let x be a random vector of R distributed according to an the following remarks can be made about (9). MGGD with scatter matrix Σ = mM and shape parameter β. Gomez´ et al. have shown that x admits the following stochastic Remark II.1 representation [2] When β = 1, Eq. (9) is close to the sample covariance d 1 matrix (SCM) estimator (the only difference between the SCM x = τ Σ 2 u (3) estimator and (9) is due to the estimation of the scale parameter where =d means equality in distribution, u is a random vector that equals 1 for the multivariate Gaussian distribution). For β = 0, (9) reduces to the FP covariance matrix estimator that uniformly distributed on the unit sphere of Rp, and τ is a scalar positive random variable such that has received much attention in [44]–[46]. p Remark II.2 τ 2β ∼ Γ ; 2 (4) 2β Equation (9) remains unchanged if M is replaced by α M where α is any non-zero real factor. Thus, the solutions of (9) where Γ(a; b) is the univariate gamma distribution with pa- (when there exist) can be determined up to a scale factor α. The rameters a and b (see [42] for definition).
Recommended publications
  • 4.2 Variance and Covariance
    4.2 Variance and Covariance The most important measure of variability of a random variable X is obtained by letting g(X) = (X − µ)2 then E[g(X)] gives a measure of the variability of the distribution of X. 1 Definition 4.3 Let X be a random variable with probability distribution f(x) and mean µ. The variance of X is σ 2 = E[(X − µ)2 ] = ∑(x − µ)2 f (x) x if X is discrete, and ∞ σ 2 = E[(X − µ)2 ] = ∫ (x − µ)2 f (x)dx if X is continuous. −∞ ¾ The positive square root of the variance, σ, is called the standard deviation of X. ¾ The quantity x - µ is called the deviation of an observation, x, from its mean. 2 Note σ2≥0 When the standard deviation of a random variable is small, we expect most of the values of X to be grouped around mean. We often use standard deviations to compare two or more distributions that have the same unit measurements. 3 Example 4.8 Page97 Let the random variable X represent the number of automobiles that are used for official business purposes on any given workday. The probability distributions of X for two companies are given below: x 12 3 Company A: f(x) 0.3 0.4 0.3 x 01234 Company B: f(x) 0.2 0.1 0.3 0.3 0.1 Find the variances of X for the two companies. Solution: µ = 2 µ A = (1)(0.3) + (2)(0.4) + (3)(0.3) = 2 B 2 2 2 2 σ A = (1− 2) (0.3) + (2 − 2) (0.4) + (3− 2) (0.3) = 0.6 2 2 4 σ B > σ A 2 2 4 σ B = ∑(x − 2) f (x) =1.6 x=0 Theorem 4.2 The variance of a random variable X is σ 2 = E(X 2 ) − µ 2.
    [Show full text]
  • On the Scale Parameter of Exponential Distribution
    Review of the Air Force Academy No.2 (34)/2017 ON THE SCALE PARAMETER OF EXPONENTIAL DISTRIBUTION Anca Ileana LUPAŞ Military Technical Academy, Bucharest, Romania ([email protected]) DOI: 10.19062/1842-9238.2017.15.2.16 Abstract: Exponential distribution is one of the widely used continuous distributions in various fields for statistical applications. In this paper we study the exact and asymptotical distribution of the scale parameter for this distribution. We will also define the confidence intervals for the studied parameter as well as the fixed length confidence intervals. 1. INTRODUCTION Exponential distribution is used in various statistical applications. Therefore, we often encounter exponential distribution in applications such as: life tables, reliability studies, extreme values analysis and others. In the following paper, we focus our attention on the exact and asymptotical repartition of the exponential distribution scale parameter estimator. 2. SCALE PARAMETER ESTIMATOR OF THE EXPONENTIAL DISTRIBUTION We will consider the random variable X with the following cumulative distribution function: x F(x ; ) 1 e ( x 0 , 0) (1) where is an unknown scale parameter Using the relationships between MXXX( ) ; 22( ) ; ( ) , we obtain ()X a theoretical variation coefficient 1. This is a useful indicator, especially if MX() you have observational data which seems to be exponential and with variation coefficient of the selection closed to 1. If we consider x12, x ,... xn as a part of a population that follows an exponential distribution, then by using the maximum likelihood estimation method we obtain the following estimate n ˆ 1 xi (2) n i1 119 On the Scale Parameter of Exponential Distribution Since M ˆ , it follows that ˆ is an unbiased estimator for .
    [Show full text]
  • A Study of Non-Central Skew T Distributions and Their Applications in Data Analysis and Change Point Detection
    A STUDY OF NON-CENTRAL SKEW T DISTRIBUTIONS AND THEIR APPLICATIONS IN DATA ANALYSIS AND CHANGE POINT DETECTION Abeer M. Hasan A Dissertation Submitted to the Graduate College of Bowling Green State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY August 2013 Committee: Arjun K. Gupta, Co-advisor Wei Ning, Advisor Mark Earley, Graduate Faculty Representative Junfeng Shang. Copyright c August 2013 Abeer M. Hasan All rights reserved iii ABSTRACT Arjun K. Gupta, Co-advisor Wei Ning, Advisor Over the past three decades there has been a growing interest in searching for distribution families that are suitable to analyze skewed data with excess kurtosis. The search started by numerous papers on the skew normal distribution. Multivariate t distributions started to catch attention shortly after the development of the multivariate skew normal distribution. Many researchers proposed alternative methods to generalize the univariate t distribution to the multivariate case. Recently, skew t distribution started to become popular in research. Skew t distributions provide more flexibility and better ability to accommodate long-tailed data than skew normal distributions. In this dissertation, a new non-central skew t distribution is studied and its theoretical properties are explored. Applications of the proposed non-central skew t distribution in data analysis and model comparisons are studied. An extension of our distribution to the multivariate case is presented and properties of the multivariate non-central skew t distri- bution are discussed. We also discuss the distribution of quadratic forms of the non-central skew t distribution. In the last chapter, the change point problem of the non-central skew t distribution is discussed under different settings.
    [Show full text]
  • On the Meaning and Use of Kurtosis
    Psychological Methods Copyright 1997 by the American Psychological Association, Inc. 1997, Vol. 2, No. 3,292-307 1082-989X/97/$3.00 On the Meaning and Use of Kurtosis Lawrence T. DeCarlo Fordham University For symmetric unimodal distributions, positive kurtosis indicates heavy tails and peakedness relative to the normal distribution, whereas negative kurtosis indicates light tails and flatness. Many textbooks, however, describe or illustrate kurtosis incompletely or incorrectly. In this article, kurtosis is illustrated with well-known distributions, and aspects of its interpretation and misinterpretation are discussed. The role of kurtosis in testing univariate and multivariate normality; as a measure of departures from normality; in issues of robustness, outliers, and bimodality; in generalized tests and estimators, as well as limitations of and alternatives to the kurtosis measure [32, are discussed. It is typically noted in introductory statistics standard deviation. The normal distribution has a kur- courses that distributions can be characterized in tosis of 3, and 132 - 3 is often used so that the refer- terms of central tendency, variability, and shape. With ence normal distribution has a kurtosis of zero (132 - respect to shape, virtually every textbook defines and 3 is sometimes denoted as Y2)- A sample counterpart illustrates skewness. On the other hand, another as- to 132 can be obtained by replacing the population pect of shape, which is kurtosis, is either not discussed moments with the sample moments, which gives or, worse yet, is often described or illustrated incor- rectly. Kurtosis is also frequently not reported in re- ~(X i -- S)4/n search articles, in spite of the fact that virtually every b2 (•(X i - ~')2/n)2' statistical package provides a measure of kurtosis.
    [Show full text]
  • On the Efficiency and Consistency of Likelihood Estimation in Multivariate
    On the efficiency and consistency of likelihood estimation in multivariate conditionally heteroskedastic dynamic regression models∗ Gabriele Fiorentini Università di Firenze and RCEA, Viale Morgagni 59, I-50134 Firenze, Italy <fi[email protected]fi.it> Enrique Sentana CEMFI, Casado del Alisal 5, E-28014 Madrid, Spain <sentana@cemfi.es> Revised: October 2010 Abstract We rank the efficiency of several likelihood-based parametric and semiparametric estima- tors of conditional mean and variance parameters in multivariate dynamic models with po- tentially asymmetric and leptokurtic strong white noise innovations. We detailedly study the elliptical case, and show that Gaussian pseudo maximum likelihood estimators are inefficient except under normality. We provide consistency conditions for distributionally misspecified maximum likelihood estimators, and show that they coincide with the partial adaptivity con- ditions of semiparametric procedures. We propose Hausman tests that compare Gaussian pseudo maximum likelihood estimators with more efficient but less robust competitors. Fi- nally, we provide finite sample results through Monte Carlo simulations. Keywords: Adaptivity, Elliptical Distributions, Hausman tests, Semiparametric Esti- mators. JEL: C13, C14, C12, C51, C52 ∗We would like to thank Dante Amengual, Manuel Arellano, Nour Meddahi, Javier Mencía, Olivier Scaillet, Paolo Zaffaroni, participants at the European Meeting of the Econometric Society (Stockholm, 2003), the Sympo- sium on Economic Analysis (Seville, 2003), the CAF Conference on Multivariate Modelling in Finance and Risk Management (Sandbjerg, 2006), the Second Italian Congress on Econometrics and Empirical Economics (Rimini, 2007), as well as audiences at AUEB, Bocconi, Cass Business School, CEMFI, CREST, EUI, Florence, NYU, RCEA, Roma La Sapienza and Queen Mary for useful comments and suggestions. Of course, the usual caveat applies.
    [Show full text]
  • A Multivariate Student's T-Distribution
    Open Journal of Statistics, 2016, 6, 443-450 Published Online June 2016 in SciRes. http://www.scirp.org/journal/ojs http://dx.doi.org/10.4236/ojs.2016.63040 A Multivariate Student’s t-Distribution Daniel T. Cassidy Department of Engineering Physics, McMaster University, Hamilton, ON, Canada Received 29 March 2016; accepted 14 June 2016; published 17 June 2016 Copyright © 2016 by author and Scientific Research Publishing Inc. This work is licensed under the Creative Commons Attribution International License (CC BY). http://creativecommons.org/licenses/by/4.0/ Abstract A multivariate Student’s t-distribution is derived by analogy to the derivation of a multivariate normal (Gaussian) probability density function. This multivariate Student’s t-distribution can have different shape parameters νi for the marginal probability density functions of the multi- variate distribution. Expressions for the probability density function, for the variances, and for the covariances of the multivariate t-distribution with arbitrary shape parameters for the marginals are given. Keywords Multivariate Student’s t, Variance, Covariance, Arbitrary Shape Parameters 1. Introduction An expression for a multivariate Student’s t-distribution is presented. This expression, which is different in form than the form that is commonly used, allows the shape parameter ν for each marginal probability density function (pdf) of the multivariate pdf to be different. The form that is typically used is [1] −+ν Γ+((ν n) 2) T ( n) 2 +Σ−1 n 2 (1.[xx] [ ]) (1) ΓΣ(νν2)(π ) This “typical” form attempts to generalize the univariate Student’s t-distribution and is valid when the n marginal distributions have the same shape parameter ν .
    [Show full text]
  • A Family of Skew-Normal Distributions for Modeling Proportions and Rates with Zeros/Ones Excess
    S S symmetry Article A Family of Skew-Normal Distributions for Modeling Proportions and Rates with Zeros/Ones Excess Guillermo Martínez-Flórez 1, Víctor Leiva 2,* , Emilio Gómez-Déniz 3 and Carolina Marchant 4 1 Departamento de Matemáticas y Estadística, Facultad de Ciencias Básicas, Universidad de Córdoba, Montería 14014, Colombia; [email protected] 2 Escuela de Ingeniería Industrial, Pontificia Universidad Católica de Valparaíso, 2362807 Valparaíso, Chile 3 Facultad de Economía, Empresa y Turismo, Universidad de Las Palmas de Gran Canaria and TIDES Institute, 35001 Canarias, Spain; [email protected] 4 Facultad de Ciencias Básicas, Universidad Católica del Maule, 3466706 Talca, Chile; [email protected] * Correspondence: [email protected] or [email protected] Received: 30 June 2020; Accepted: 19 August 2020; Published: 1 September 2020 Abstract: In this paper, we consider skew-normal distributions for constructing new a distribution which allows us to model proportions and rates with zero/one inflation as an alternative to the inflated beta distributions. The new distribution is a mixture between a Bernoulli distribution for explaining the zero/one excess and a censored skew-normal distribution for the continuous variable. The maximum likelihood method is used for parameter estimation. Observed and expected Fisher information matrices are derived to conduct likelihood-based inference in this new type skew-normal distribution. Given the flexibility of the new distributions, we are able to show, in real data scenarios, the good performance of our proposal. Keywords: beta distribution; centered skew-normal distribution; maximum-likelihood methods; Monte Carlo simulations; proportions; R software; rates; zero/one inflated data 1.
    [Show full text]
  • Generalized Inferences for the Common Scale Parameter of Several Pareto Populations∗
    InternationalInternational JournalJournal of of Statistics Statistics and and Management Management System System Vol.Vol. 4 No. 12 (January-June,(July-December, 2019) 2019) International Journal of Statistics and Management System, 2010, Vol. 5, No. 1–2, pp. 118–126. c 2010 Serials Publications Generalized inferences for the common scale parameter of several Pareto populations∗ Sumith Gunasekera†and Malwane M. A. Ananda‡ Received: 13th August 2018 Revised: 24th December 2018 Accepted: 10th March 2019 Abstract A problem of interest in this article is statistical inferences concerning the com- mon scale parameter of several Pareto distributions. Using the generalized p-value approach, exact confidence intervals and the exact tests for testing the common scale parameter are given. Examples are given in order to illustrate our procedures. A limited simulation study is given to demonstrate the performance of the proposed procedures. 1 Introduction In this paper, we consider k (k ≥ 2) independent Pareto distributions with an unknown common scale parameter θ (sometimes referred to as the “location pa- rameter” and also as the “truncation parameter”) and unknown possibly unequal shape parameters αi’s (i = 1, 2, ..., k). Using the generalized variable approach (Tsui and Weer- ahandi [8]), we construct an exact test for testing θ. Furthermore, using the generalized confidence interval (Weerahandi [11]), we construct an exact confidence interval for θ as well. A limited simulation study was carried out to compare the performance of these gen- eralized procedures with the approximate procedures based on the large sample method as well as with the other test procedures based on the combination of p-values.
    [Show full text]
  • A Study of Ocean Wave Statistical Properties Using SNOW
    A study of ocean wave statistical properties using nonlinear, directional, phase- resolved ocean wavefield simulations Legena Henry PhD Candidate Course 2 – OE [email protected] abstract We study the statistics of wavefields obtained from phase-resolved simulations that are non-linear (up to the third order in surface potential). We model wave-wave interactions based on the fully non- linear Zakharov equations. We vary the simulated wavefield's input spectral properties: peak shape parameter, Phillips parameter and Directional Spreading Function. We then investigate the relationships between these input spectral properties and output physical properties via statistical analysis. We investigate: 1. Surface elevation distribution in response to input spectral properties, 2. Wave definition methods in an irregular wavefield with a two- dimensional wave number, 3. Wave height/wavelength distributions in response to input spectral properties, including the occurrence and spacing of large wave events (based on definitions in 2). 1. Surface elevation distribution in response to input spectral properties: – Peak Shape Parameter – Phillips' Parameter – Directional Spreading Function 1. Surface elevation distribution in response to input spectral properties: Peak shape parameter Peak shape parameter produces higher impact on even moments of surface elevation, (i.e. surface elevation kurtosis and surface elevation variance) than on the observed odd moments (skewness) of surface elevation. The highest elevations in a wavefield with mean peak shape parameter, 3.3 are more stable than those in wavefields with mean peak shape parameter, 1.0, and 5.0 We find surface elevation kurtosis in non-linear wavefields is much smaller than surface elevation slope kurtosis. We also see that higher values of peak shape parameter produce higher kurtosis of surface slope in the mean direction of propagation.
    [Show full text]
  • Sampling Student's T Distribution – Use of the Inverse Cumulative
    Sampling Student’s T distribution – use of the inverse cumulative distribution function William T. Shaw Department of Mathematics, King’s College, The Strand, London WC2R 2LS, UK With the current interest in copula methods, and fat-tailed or other non-normal distributions, it is appropriate to investigate technologies for managing marginal distributions of interest. We explore “Student’s” T distribution, survey its simulation, and present some new techniques for simulation. In particular, for a given real (not necessarily integer) value n of the number of degrees of freedom, −1 we give a pair of power series approximations for the inverse, Fn ,ofthe cumulative distribution function (CDF), Fn.Wealsogivesomesimpleandvery fast exact and iterative techniques for defining this function when n is an even −1 integer, based on the observation that for such cases the calculation of Fn amounts to the solution of a reduced-form polynomial equation of degree n − 1. We also explain the use of Cornish–Fisher expansions to define the inverse CDF as the composition of the inverse CDF for the normal case with a simple polynomial map. The methods presented are well adapted for use with copula and quasi-Monte-Carlo techniques. 1 Introduction There is much interest in many areas of financial modeling on the use of copulas to glue together marginal univariate distributions where there is no easy canonical multivariate distribution, or one wishes to have flexibility in the mechanism for combination. One of the more interesting marginal distributions is the “Student’s” T distribution. This statistical distribution was published by W. Gosset in 1908.
    [Show full text]
  • The Smoothed Median and the Bootstrap
    Biometrika (2001), 88, 2, pp. 519–534 © 2001 Biometrika Trust Printed in Great Britain The smoothed median and the bootstrap B B. M. BROWN School of Mathematics, University of South Australia, Adelaide, South Australia 5005, Australia [email protected] PETER HALL Centre for Mathematics & its Applications, Australian National University, Canberra, A.C.T . 0200, Australia [email protected] G. A. YOUNG Statistical L aboratory, University of Cambridge, Cambridge CB3 0WB, U.K. [email protected] S Even in one dimension the sample median exhibits very poor performance when used in conjunction with the bootstrap. For example, both the percentile-t bootstrap and the calibrated percentile method fail to give second-order accuracy when applied to the median. The situation is generally similar for other rank-based methods, particularly in more than one dimension. Some of these problems can be overcome by smoothing, but that usually requires explicit choice of the smoothing parameter. In the present paper we suggest a new, implicitly smoothed version of the k-variate sample median, based on a particularly smooth objective function. Our procedure preserves many features of the conventional median, such as robustness and high efficiency, in fact higher than for the conventional median, in the case of normal data. It is however substantially more amenable to application of the bootstrap. Focusing on the univariate case, we demonstrate these properties both theoretically and numerically. Some key words: Bootstrap; Calibrated percentile method; Median; Percentile-t; Rank methods; Smoothed median. 1. I Rank methods are based on simple combinatorial ideas of permutations and sign changes, which are attractive in applications far removed from the assumptions of normal linear model theory.
    [Show full text]
  • A Note on Inference in a Bivariate Normal Distribution Model Jaya
    A Note on Inference in a Bivariate Normal Distribution Model Jaya Bishwal and Edsel A. Peña Technical Report #2009-3 December 22, 2008 This material was based upon work partially supported by the National Science Foundation under Grant DMS-0635449 to the Statistical and Applied Mathematical Sciences Institute. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. Statistical and Applied Mathematical Sciences Institute PO Box 14006 Research Triangle Park, NC 27709-4006 www.samsi.info A Note on Inference in a Bivariate Normal Distribution Model Jaya Bishwal¤ Edsel A. Pena~ y December 22, 2008 Abstract Suppose observations are available on two variables Y and X and interest is on a parameter that is present in the marginal distribution of Y but not in the marginal distribution of X, and with X and Y dependent and possibly in the presence of other parameters which are nuisance. Could one gain more e±ciency in the point estimation (also, in hypothesis testing and interval estimation) about the parameter of interest by using the full data (both Y and X values) instead of just the Y values? Also, how should one measure the information provided by random observables or their distributions about the parameter of interest? We illustrate these issues using a simple bivariate normal distribution model. The ideas could have important implications in the context of multiple hypothesis testing or simultaneous estimation arising in the analysis of microarray data, or in the analysis of event time data especially those dealing with recurrent event data.
    [Show full text]