SEQUENTIAL ESTIMATION IN A SUBCLASS OF EXPONENTIAL FAMILY UNDER WEIGHTED SQUARED ERROR LOSS *

Similar documents
ENGI 4421 Central Limit Theorem Page Central Limit Theorem [Navidi, section 4.11; Devore sections ]

ENGI 4421 Central Limit Theorem Page Central Limit Theorem [Navidi, section 4.11; Devore sections ]

Ch. 1 Introduction to Estimation 1/15

Multi-objective Programming Approach for. Fuzzy Linear Programming Problems

5.1 Two-Step Conditional Density Estimator

D.S.G. POLLOCK: TOPICS IN TIME-SERIES ANALYSIS STATISTICAL FOURIER ANALYSIS

ELEG 635 Digital Communication Theory. Lecture 10

Fourier Method for Solving Transportation. Problems with Mixed Constraints

Control Systems. Controllability and Observability (Chapter 6)

MATH Midterm Examination Victor Matveev October 26, 2016

Quantum Mechanics for Scientists and Engineers. David Miller

, the random variable. and a sample size over the y-values 0:1:10.

The Excel FFT Function v1.1 P. T. Debevec February 12, The discrete Fourier transform may be used to identify periodic structures in time ht.

A Study on Estimation of Lifetime Distribution with Covariates Under Misspecification

IJISET - International Journal of Innovative Science, Engineering & Technology, Vol. 2 Issue 12, December

Chapter 3.1: Polynomial Functions

Mean residual life of coherent systems consisting of multiple types of dependent components

[1 & α(t & T 1. ' ρ 1

Solutions: Homework 3

An Investigation of Stratified Jackknife Estimators Using Simulated Establishment Data Under an Unequal Probability Sample Design

Lecture 18. MSMPR Crystallization Model

Certain inclusion properties of subclass of starlike and convex functions of positive order involving Hohlov operator

SOLUTION. The reactor thermal output is related to the maximum heat flux in the hot channel by. Z( z ). The position of maximum heat flux ( z max

x 2 x 3 x b 0, then a, b, c log x 1 log z log x log y 1 logb log a dy 4. dx As tangent is perpendicular to the x axis, slope

Regression Quantiles for Time Series Data ZONGWU CAI. Department of Mathematics. Abstract

Research & Reviews: Journal of Statistics and Mathematical Sciences

Solutions. Definitions pertaining to solutions

A NOTE ON THE EQUIVAImCE OF SOME TEST CRITERIA. v. P. Bhapkar. University of Horth Carolina. and

Lecture 21: Signal Subspaces and Sparsity

Intermediate Division Solutions

Super-efficiency Models, Part II

UNIVERSITY OF TECHNOLOGY. Department of Mathematics PROBABILITY THEORY, STATISTICS AND OPERATIONS RESEARCH GROUP. Memorandum COSOR 76-10

Internal vs. external validity. External validity. Internal validity

Unifying the Derivations for. the Akaike and Corrected Akaike. Information Criteria. from Statistics & Probability Letters,

K [f(t)] 2 [ (st) /2 K A GENERALIZED MEIJER TRANSFORMATION. Ku(z) ()x) t -)-I e. K(z) r( + ) () (t 2 I) -1/2 e -zt dt, G. L. N. RAO L.

are specified , are linearly independent Otherwise, they are linearly dependent, and one is expressed by a linear combination of the others

MATHEMATICS 9740/01 Paper 1 14 Sep hours

Matching a Distribution by Matching Quantiles Estimation

Review of Important Concepts

Examination No. 3 - Tuesday, Nov. 15

~ *The paper was written while second author was on leave from Leningrad

Hypothesis Testing and Confidence Intervals (Part 1): Using the Standard Normal

On the Origin of the Special Relativity Anomalies

Lecture 8. Dirac and Weierstrass

Preliminary Test Single Stage Shrinkage Estimator for the Scale Parameter of Gamma Distribution

Identical Particles. We would like to move from the quantum theory of hydrogen to that for the rest of the periodic table

A Hartree-Fock Calculation of the Water Molecule

A New Method for Finding an Optimal Solution. of Fully Interval Integer Transportation Problems

MATH 320: Probability and Statistics 9. Estimation and Testing of Parameters. Readings: Pruim, Chapter 4

STRUCTURES IN MIKE 21. Flow over sluice gates A-1

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 5

The Sampling Distribution of the Maximum. Likelihood Estimators for the Parameters of. Beta-Binomial Distribution

Bayesian and E- Bayesian Method of Estimation of Parameter of Rayleigh Distribution- A Bayesian Approach under Linex Loss Function

Lecture 7: Properties of Random Samples

Review for cumulative test

Extreme Value Theory in Civil Engineering

The beta density, Bayes, Laplace, and Pólya

ELEG 635 Digital Communication Theory. Lecture 11

Unbiased Estimation. February 7-12, 2008

Comparative analysis of bayesian control chart estimation and conventional multivariate control chart

ε > 0 N N n N a n < ε. Now notice that a n = a n.

HE ATOM & APPROXIMATION METHODS MORE GENERAL VARIATIONAL TREATMENT. Examples:

Sx [ ] = x must yield a

Copyright 1978, by the author(s). All rights reserved.

Inference in the Multiple-Regression

After the completion of this section the student. V.4.2. Power Series Solution. V.4.3. The Method of Frobenius. V.4.4. Taylor Series Solution

REVERSIBLE NON-FLOW PROCESS CONSTANT VOLUME PROCESS (ISOCHORIC PROCESS) In a constant volume process, he working substance is contained in a rigid

Some vector-valued statistical convergent sequence spaces

Greedy Algorithms. Kye Halsted. Edited by Chuck Cusack. These notes are based on chapter 17 of [1] and lectures from CSCE423/823, Spring 2001.

RMO Sample Paper 1 Solutions :

Fixed Point Approximation of Weakly Commuting Mappings in Banach Space

A NEW NOTE ON LOCAL PROPERTY OF FACTORED FOURIER SERIES

Section 11 Simultaneous Equations

ECE 8527: Introduction to Machine Learning and Pattern Recognition Midterm # 1. Vaishali Amin Fall, 2015

Pulse Shaping in Non-Coherent DLL Tracking of CDMA Signals

On the convergence rates of Gladyshev s Hurst index estimator

Fourier Series & Fourier Transforms

Asymptotic analysis of the Bell polynomials by the ray method

Statistica Sinica 6(1996), SOME PROBLEMS ON THE ESTIMATION OF UNIMODAL DENSITIES Peter J. Bickel and Jianqing Fan University of California and U

COURSE INTRODUCTION: WHAT HAPPENS TO A QUANTUM PARTICLE ON A PENDULUM π 2 SECONDS AFTER IT IS TOSSED IN?

BIO752: Advanced Methods in Biostatistics, II TERM 2, 2010 T. A. Louis. BIO 752: MIDTERM EXAMINATION: ANSWERS 30 November 2010

Bayesian Estimation for Continuous-Time Sparse Stochastic Processes

Active redundancy allocation in systems. R. Romera; J. Valdés; R. Zequeira*

Study of Energy Eigenvalues of Three Dimensional. Quantum Wires with Variable Cross Section

Topics. Homework Problems. MATH 301 Introduction to Analysis Chapter Four Sequences. 1. Definition of convergence of sequences.

In SMV I. IAML: Support Vector Machines II. This Time. The SVM optimization problem. We saw:

On the Variability Estimation of Lognormal Distribution Based on. Sample Harmonic and Arithmetic Means

DANIELL AND RIEMANN INTEGRABILITY

Weighted Approximation by Videnskii and Lupas Operators

Claude Elysée Lobry Université de Nice, Faculté des Sciences, parc Valrose, NICE, France.

Math 778S Spectral Graph Theory Handout #3: Eigenvalues of Adjacency Matrix

MASSIVELY PARALLEL SEQUENCING OF POOLED DNA SAMPLES-THE NEXT GENERATION OF MOLECULAR MARKERS

A remark on p-summing norms of operators

A General Iterative Scheme for Variational Inequality Problems and Fixed Point Problems

THE ASYMPTOTIC PERFORMANCE OF THE LOG LIKELIHOOD RATIO STATISTIC FOR THE MIXTURE MODEL AND RELATED RESULTS

Markov processes and the Kolmogorov equations

On Equivalence of Martingale Tail Bounds and Deterministic Regret Inequalities

Exact Inference on the Random-Effects Model for. Meta-Analyses with Few Studies

Topologie. Musterlösungen

Hyun-Chull Kim and Tae-Sung Kim

Transcription:

Iraia Jural f iee & Tehlgy Trasati A Vl.. A Prited i The Islami Republi f Ira 007 hiraz Uiersity QUTIAL TIMATIO I A UBCLA OF POTIAL FAMIL UDR WIGHTD QUARD RROR LO *. MATOLLAHI ** M. JAFARI JOZAI AD. MAHLOOJI Departmet f tatistis Faulty f mis Allameh Tabataba'i Uiersity Tehra I. R. f Ira mail: ematllahi@atu.a.ir Abstrat I a sublass f the sale-parameter expetial family e sider the sequetial pit estimati f a futi f the sale parameter uder the lss futi gie as the sum f the eighted squared errr lss ad a liear st. Fr a fully sequetial samplig sheme sed rder expasis are btaied fr the expeted sample size as ell as fr the regret f the predure. The frmer researhes Gamma ad xpetial distributis a be dedued frm ur geeral results. Keyrds equetial estimati stppig rule regret aalysis expetial family trasfrmed hi-square distributi. ITRODUCTIO The prblem f sequetial estimati refers t ay estimati tehique fr hih the ttal umber f bseratis used is t a degeerate radm ariable. I sme prblems sequetial estimati must be used beause predure that uses a preassiged radm sample size a ahiee the desired bjetie fr example the estimati f the parameter /p i a sequee f Berulli trials. I ther prblems a predure hih uses a preassiged radm sample size may exist but a sequetial estimati predure may be better i sme ays. The prblems f sequetial aalysis ere first studied i the 90s by Barard ad Wald h itrdued the equetial Prbability Rati Test PRT idepedetly Wald ad Wlfitz pred its ptimality ad Haldae ad tei 5 shed h sequetial methds a be used t takle sme usled prblems i pit ad iteral estimati. There is a large bdy f literature this subjet ad it is grig rapidly. Fr a summary f results as ell as a list f referees see Lai. equetial estimati f the sale parameter f xpetial ad Gamma distributis hae bee sidered by tarr ad Wdrfe 7 Wdrfe 8 Gsh ad Mukhpadhyay 9 Isgai ad U 0 Isgai et al. ad U et al.. Uder squared errr lss tarr ad Wdrfe 7 sidered the risk effiiet estimati f the sale parameter i expetial distributi ad studied sme f the first rder prperties f the sequetial predure. Als fr the sequetial estimati f a futi f the expetial parameter U et al. gae the stppig rule ad a suffiiet diti t get a sed rder apprximati t the risk f the sequetial predure. Fr the estimati f the sale parameter the sale iariat squared errr lss is mre apprpriate tha the squared errr lss. I additi there are seeral ases fr hih the estimati f a futi f the sale parameter is desired. it is atural t use δ as a apprpriate relatie squared errr lss Reeied by the editr August 005 ad i fial reised frm April 8 007 Crrespdig authr

90. ematllahi / et al. futi hih is a speial frm f eighted squared errr lss futi. Hee i this paper e sider the lss as the sum f the eighted squared errr lss ad a liear st hih iludes sale iariat squared errr lss. We btai a sequetial pit estimati f a futi f the sale parameter i a sublass f the sale parameter expetial families f distributis hih ilude xpetial Weibull Gamma rmal Ierse Guassia ad sme ther distributis ad determie a stppig rule uder the lss futi. Als a sed rder apprximati t the expeted sample size ad the risk f the sequetial predure as the st per bserati teds t zer are gie. We sh that the results btaied by U et al. i xpetial distributi ad Wdrfe 8 i Gamma distributi uder squared errr lss are speial ases f ur results. I eti the sublass f distributis is itrdued ad the sequetial estimati prblem is speified. I seti e derie asymptti expasis f the expeted sample size ad regret assiated ith the prpsed predure. me speial ases f ur results are gie i seti.. QUTIAL TIMATIO I A UBCLA OF POTIAL FAMIL Let... be a sequee f idepedet ad idetially distributed radm ariables frm a x distributi ith desity g here g is k ad τ is a uk sale parameter. I sme ases τ τ the abe desity redues t Tx fx x e > 0 r here x is a futi f x τ is a k alue ad T is a mplete suffiiet statisti fr ith Gamma - distributi. xamples f distributis f the frm are. xpetial β ith β T x a x. Gamma β ith k ad β T x Γ a. Ierse Gaussia λ ith T x π x λ. rmal 0 σ ith σ T x π β β β 5. Weibull β ith k β ad T x β x. Rayleigh β ith β T x x 7. Geeralized Gamma λ p ith k p ad p p T x x λ Γ p / 8. Geeralized Laplae λ k ith k k ad k a k λ T x. k Γ/ k if... be a radm sample f size frm distributi the the jit desity f... is gie by Tx i / i 0 fx x e > Iraia Jural f iee & Tehlgy Tras. A Vlume umber A prig 007

equetial estimati i a sublass f 9 here x Π xi ad i T i Gamma. Csider the estimati f a futi f i the sale parameter say here is a psitie alued ad three times tiuusly differetiable futi f. Gie a sample... f size e at t estimate by ˆ uder the lss futi L ˆ ˆ here > 0 is the k st per uit sample ad is a psitie alued ad t times tiuusly differetiable eight futi. peial ases f the lss are saled iariat squared errr lss ad squared errr lss ˆ L ˆ ˆ L ˆ 5 hsig ad respetiely. te that is the usual estimatr f i.e. ML UMVU hee usig iariae prperty f maximum likelihd estimatrs it is reasable t estimate by ˆ. The risk futi is gie by R R ˆ L ˆ ˆ. We at t fid a apprpriate sample size that ill miimize the risk R. Usig Taylr expasi f abut e btai ˆ Var as ad hee R as. fr suffiietly large R hih leads t the fllig lemma. Lemma -. The risk futi ad fr this alue R i miimized at 0 7 R. 8 ie is uk e a t use the best fixed sample size predure 0. Further there is fixed sample size predure that ill attai the miimum risk R. Thus it is eessary t fid a sequetial 0 prig 007 Iraia Jural f iee & Tehlgy Tras. A Vlume umber A

9. ematllahi / et al. samplig rule. We use the fllig stppig rule if m 9 here m is the pilt sample size. if e estimate by ˆ the the risk is gie by R ˆ. 0 I the ext seti e derie sed rder apprximati t the expeted sample size ad the risk f the abe sequetial predure R as 0.. COD ORDR APPROIMATIO I this seti e shall iestigate sed rder asymptti prperties f the sequetial predure. Let Kt t> 0 t t t t t} t ad Z K / K. The frm 9 the stppig rule bemes m Z > if : }. Usig Taylr expasi f abut ad the relatis K t t t t t t K t t t ad K t K t t t t t t t t t t t t t t t t t t t t t t e btai the fllig lemma. T i Lemma -. Let i fr i... ad i the i Z ψ here t K K ψ K K ad is a radm ariable lyig betee ad. Iraia Jural f iee & Tehlgy Tras. A Vlume umber A prig 007

equetial estimati i a sublass f 9 Let t t t if : > 0} ad ρ 5 t Fllig assumptis f Aras ad Wdrfe amely A Z m is uifrmly itegrable ε fr sme 0 < ε 0 < here x max x0 m A P ψ < ε} < fr sme 0 < ε < e btai the fllig apprximati t fr all 0 but t uifrmly i. Therem -. If A ad A hld true the here ρ l as 0 l K. 7 K Prf: Obiusly P as ad sie is a radm ariable lyig betee ad therefre P K p K ad hee l. Als K K d W 0 as. ψ K d ψ l. W as. K The rest f the prf is the same as the prf f Therem f U et al. ith replaigσ σ ξ ad h by ψ ad K respetiely is mitted. We shall assess the regret R. By Taylr's therem 8 t prig 007 Iraia Jural f iee & Tehlgy Tras. A Vlume umber A

9. ematllahi / et al. here is a radm ariable lyig betee ad. Let ad hse 0 > 0 suh that. We impse the fllig assumpti A : Fr sme a > ad u > sup 0< 0 au u < ad sup <. 0< 0 au Frm A A ad A e hae the fllig therem. Therem -. If A A ad A hld true the as 0 R 7 5 } 9 Prf: Frm 8 e hae R 5 }. 0 Fllig the prf f Therem f U et al. he 0 e btai Iraia Jural f iee & Tehlgy Tras. A Vlume umber A prig 007

equetial estimati i a sublass f prig 007 Iraia Jural f iee & Tehlgy Tras. A Vlume umber A 95 7 } l 5 9 } } } } } W } 5 ad } 5.

. ematllahi / et al. Iraia Jural f iee & Tehlgy Tras. A Vlume umber A prig 007 9 ubstitutig it 0 e get 7 R 5 9 7 5 as 0 this mpletes the prf.. PCIAL CA I this seti e sider sme speial ases f the results btaied i seti. These speial ases are. Fr xpetial β -distributi ith β T ad Therem - ad - ith i.e. ith the lss futi 5 bemes the Therem ad f U et al. respetiely. the expeted sample size ad the regret btaied by U et al. are speial ases f ad 9 respetiely. Als he ad the regret bemes R hih is the result btaied by Wdrfe 8.. Fr Gamma - distributi ith k T ad i estimati f uder the lss futi 5 i.e. the regret bemes R hih is the result btaied by Wdrfe 8.. Fr sale iariat squared errr lss e hae ad frm 9 the regret bemes. } } } } 7 R Therefre fr the regret bemes R. te that the lss futi is mre apprpriate fr estimatig the sale parameter tha squared errr lss. The results f seti ad a be exteded t sme ther distributis hih d t eessarily belg t a sale family suh as Paret r Beta distributis. A family f distributis that iludes these distributis as speial ases is the family f trasfrmed hi-square distributis hih is rigially itrdued by Rahma ad Gupta. They sidered the e parameter expetial family x h b x a e x f 7 ad shed that b a has a j Gamma - distributi if ad ly if. j b b 8

equetial estimati i a sublass f 97 I ase j is a iteger a b flls a hi-square distributi ith j degrees f freedm. They alled the e parameter expetial family 7 hih satisfies 8 the family f trasfrmed hisquare distributis. Fr example Beta Paret xpetial Lgrmal ad sme ther distributis belg t this family f distributis see Table f Rahma ad Gupta. it is easy t sh that if diti 8 hlds the the e parameter expetial family 7 is i the frm f the sale parameter expetial family ith j T a ad. Hee b ith these substitutis e a exted the results f seti ad t the family f trasfrmed hisquare distributis. RFRC. Barard G. A. 9. my i samplig ith speial referee t egieerig experimetati. Brith. Mi. upply Teh. Rep. QC/R/7 I.. Wald A. 95. equetial test f statistial hyptheses. A. Math. tat 7-8.. Wald A. & Wlfitz J. 98. Optimum harater f the sequetial prbability rati test. A. Math. tat. 9-9.. Haldae J. B.. 95. O a methd f estimatig frequeies. Bimetrika -5. 5. tei C. 95. A t sample test fr a liear hypthesis hse per is idepedet f the ariae. A. Math. tat. -58.. Lai T. L. 00. equetial Aalysis: me lassial prblems ad e halleges. tatistia iia 0-08. 7. tarr. & Wdrfe M. 97. Farther remarks sequetial estimati: the expetial ase. A. Math. tat. 7-5. 8. Wdrfe M. 977. ed rder apprximati fr sequetial pit ad iteral estimati. A. tatist. 5 98-995. 9. Gsh M. & Mukhpadhyay. 989. equetial estimati f the peretiles f expetial ad rmal distributis. uth Afria tatist. J. 7-. 0. Isgai. & U C. 99. equetial estimati f a parameter f a expetial distributi. A. Ist. tatist. Math. 77-8.. Isgai. Ali M. & U C. 00. equetial estimati f the pers f rmal ad expetial sale parameters. equetial Aalysis 77-8.. U C. Isgai. & Lim D. 00. equetial pit estimati f a futi f the expetial sale parameter. Austria Jural f tatistis 8-9.. Aras G. & Wdrfe M. 99. Asymptti expasis fr the mmets f a radmly stpped aerage. The Aals f tatistis 50-59.. Rahma M.. & Gupta R. P. 99. Family f trasfrmed hi-square distributis. Cmm. tatist. Thery Methds 5-. prig 007 Iraia Jural f iee & Tehlgy Tras. A Vlume umber A