Total Least Squares Fitting the Three-Parameter Inverse Weibull Density

Similar documents
A Note on Estimability in Linear Models

Grand Canonical Ensemble

Review - Probabilistic Classification

Soft k-means Clustering. Comp 135 Machine Learning Computer Science Tufts University. Mixture Models. Mixture of Normals in 1D

10/7/14. Mixture Models. Comp 135 Introduction to Machine Learning and Data Mining. Maximum likelihood estimation. Mixture of Normals in 1D

COMPLEX NUMBER PAIRWISE COMPARISON AND COMPLEX NUMBER AHP

ON THE COMPLEXITY OF K-STEP AND K-HOP DOMINATING SETS IN GRAPHS

Outlier-tolerant parameter estimation

A NEW GENERALISATION OF SAM-SOLAI S MULTIVARIATE ADDITIVE GAMMA DISTRIBUTION*

8-node quadrilateral element. Numerical integration

Lucas Test is based on Euler s theorem which states that if n is any integer and a is coprime to n, then a φ(n) 1modn.

Group Codes Define Over Dihedral Groups of Small Order

Analyzing Frequencies

ST 524 NCSU - Fall 2008 One way Analysis of variance Variances not homogeneous

Econ107 Applied Econometrics Topic 10: Dummy Dependent Variable (Studenmund, Chapter 13)

SCITECH Volume 5, Issue 1 RESEARCH ORGANISATION November 17, 2015

CHAPTER 7d. DIFFERENTIATION AND INTEGRATION

External Equivalent. EE 521 Analysis of Power Systems. Chen-Ching Liu, Boeing Distinguished Professor Washington State University

From Structural Analysis to FEM. Dhiman Basu

Lecture 23 APPLICATIONS OF FINITE ELEMENT METHOD TO SCALAR TRANSPORT PROBLEMS

A Probabilistic Characterization of Simulation Model Uncertainties

The Hyperelastic material is examined in this section.

September 27, Introduction to Ordinary Differential Equations. ME 501A Seminar in Engineering Analysis Page 1. Outline

Folding of Regular CW-Complexes

Optimal Ordering Policy in a Two-Level Supply Chain with Budget Constraint

ON EISENSTEIN-DUMAS AND GENERALIZED SCHÖNEMANN POLYNOMIALS

arxiv: v1 [math.pr] 28 Jan 2019

The Fourier Transform

Decision-making with Distance-based Operators in Fuzzy Logic Control

Physics 256: Lecture 2. Physics

VISUALIZATION OF DIFFERENTIAL GEOMETRY UDC 514.7(045) : : Eberhard Malkowsky 1, Vesna Veličković 2

The Matrix Exponential

te Finance (4th Edition), July 2017.

Discrete Shells Simulation

Consider a system of 2 simultaneous first order linear equations

CHAPTER 33: PARTICLE PHYSICS

The Matrix Exponential

ON RIGHT(LEFT) DUO PO-SEMIGROUPS. S. K. Lee and K. Y. Park

Section 6.1. Question: 2. Let H be a subgroup of a group G. Then H operates on G by left multiplication. Describe the orbits for this operation.

Basic Polyhedral theory

Economics 600: August, 2007 Dynamic Part: Problem Set 5. Problems on Differential Equations and Continuous Time Optimization

2. Grundlegende Verfahren zur Übertragung digitaler Signale (Zusammenfassung) Informationstechnik Universität Ulm

You already learned about dummies as independent variables. But. what do you do if the dependent variable is a dummy?

An Overview of Markov Random Field and Application to Texture Segmentation

Lecture 14. Relic neutrinos Temperature at neutrino decoupling and today Effective degeneracy factor Neutrino mass limits Saha equation

Shortest Paths in Graphs. Paths in graphs. Shortest paths CS 445. Alon Efrat Slides courtesy of Erik Demaine and Carola Wenk

Chapter 6 Student Lecture Notes 6-1

Derangements and Applications

JEE-2017 : Advanced Paper 2 Answers and Explanations

Decentralized Adaptive Control and the Possibility of Utilization of Networked Control System

Fakultät III Univ.-Prof. Dr. Jan Franke-Viebach

Journal of Theoretical and Applied Information Technology 10 th January Vol. 47 No JATIT & LLS. All rights reserved.

EXST Regression Techniques Page 1

cycle that does not cross any edges (including its own), then it has at least

SOME PARAMETERS ON EQUITABLE COLORING OF PRISM AND CIRCULANT GRAPH.

Search sequence databases 3 10/25/2016

Reliability of time dependent stress-strength system for various distributions

APPROXIMATE PRICES OF BASKET AND ASIAN OPTIONS DUPONT OLIVIER. Premia 14

Introduction to Arithmetic Geometry Fall 2013 Lecture #20 11/14/2013

On Selection of Best Sensitive Logistic Estimator in the Presence of Collinearity

Lecture 3: Phasor notation, Transfer Functions. Context

Heisenberg Model. Sayed Mohammad Mahdi Sadrnezhaad. Supervisor: Prof. Abdollah Langari

Stress-Based Finite Element Methods for Dynamics Analysis of Euler-Bernoulli Beams with Various Boundary Conditions

HORIZONTAL IMPEDANCE FUNCTION OF SINGLE PILE IN SOIL LAYER WITH VARIABLE PROPERTIES

Naresuan University Journal: Science and Technology 2018; (26)1

NEW APPLICATIONS OF THE ABEL-LIOUVILLE FORMULA

Recall that by Theorems 10.3 and 10.4 together provide us the estimate o(n2 ), S(q) q 9, q=1

Construction of asymmetric orthogonal arrays of strength three via a replacement method

Jones vector & matrices

Supplementary Materials

Epistemic Foundations of Game Theory. Lecture 1

Another converse of Jensen s inequality

UNTYPED LAMBDA CALCULUS (II)

Fakultät III Wirtschaftswissenschaften Univ.-Prof. Dr. Jan Franke-Viebach

Introduction to logistic regression

u x v x dx u x v x v x u x dx d u x v x u x v x dx u x v x dx Integration by Parts Formula

Combinatorial Networks Week 1, March 11-12

Function Spaces. a x 3. (Letting x = 1 =)) a(0) + b + c (1) = 0. Row reducing the matrix. b 1. e 4 3. e 9. >: (x = 1 =)) a(0) + b + c (1) = 0

NON-SYMMETRY POWER IN THREE-PHASE SYSTEMS

LINEAR DELAY DIFFERENTIAL EQUATION WITH A POSITIVE AND A NEGATIVE TERM

3.4 Properties of the Stress Tensor

Α complete processing methodology for 3D monitoring using GNSS receivers

SCRIBE: JAKE LEVINSON

Independent Domination in Line Graphs

Hardy-Littlewood Conjecture and Exceptional real Zero. JinHua Fei. ChangLing Company of Electronic Technology Baoji Shannxi P.R.

Thus, because if either [G : H] or [H : K] is infinite, then [G : K] is infinite, then [G : K] = [G : H][H : K] for all infinite cases.

Cramér-Rao Inequality: Let f(x; θ) be a probability density function with continuous parameter

Abstract Interpretation: concrete and abstract semantics

Ερωτήσεις και ασκησεις Κεφ. 10 (για μόρια) ΠΑΡΑΔΟΣΗ 29/11/2016. (d)

BINOMIAL COEFFICIENTS INVOLVING INFINITE POWERS OF PRIMES. 1. Statement of results

The Equitable Dominating Graph

On the irreducibility of some polynomials in two variables

Relate p and T at equilibrium between two phases. An open system where a new phase may form or a new component can be added

Square of Hamilton cycle in a random graph

Application of Local Influence Diagnostics to the Linear Logistic Regression Models

10. The Discrete-Time Fourier Transform (DTFT)

FEFF and Related Codes

GPC From PeakSimple Data Acquisition

Einstein Equations for Tetrad Fields

Preview. Graph. Graph. Graph. Graph Representation. Graph Representation 12/3/2018. Graph Graph Representation Graph Search Algorithms

Transcription:

EUROPEAN JOURNAL OF PURE AND APPLIED MATHEMATICS Vol. 7, No. 3, 2014, 230-245 ISSN 1307-5543 www.jpam.com Total Last Squars Fttng th Thr-Paramtr Invrs Wbull Dnsty Dragan Juć, Darja Marovć Dpartmnt of Mathmatcs, J.J. Strossmayr Unvrsty of Osj, Trg Ljudvta Gaja 6, HR-31 000 Osj, Croata Abstract. Th focus of ths papr s on a nonlnar wghtd total last squars fttng problm for th thr-paramtr nvrs Wbull dnsty whch s frquntly mployd as a modl n rlablty and lftm studs. As a man rsult, a thorm on th xstnc of th total last squars stmator s obtand, as wll as ts gnralzaton n th l q norm (1 q ). 2010 Mathmatcs Subjct Classfcatons: 65D10, 62J02, 62G07, 62N05 Ky Words and Phrass: nvrs Wbull dnsty, total last squars, total last squars stmat, xstnc problm, data fttng 1. Introducton Th probablty dnsty functon of th random varabl T havng a thr-paramtr nvrs Wbull dstrbuton (IWD) wth locaton paramtrα 0, scal paramtrη>0and shap paramtrβ> 0 s gvn by f(t;α,β,η)= β η β+1 η t α ( η t α )β t>α 0 t α. (1) If α = 0, th rsultng dstrbuton s calld th two-paramtr nvrs Wbull dstrbuton. Ths modl was dvlopd by Erto[6]. Th IWD s vry flxbl and by an approprat choc of th shap paramtrβ th dnsty curv can assum a wd varty of shaps (s Fg. 1). Th dnsty functon s strctly ncrasng on(α, t m ] and strctly dcrasng on[t m, ), whr t m =α+η(1+1/β) 1/β. Ths mpls that th dnsty functon s unmodal wth th maxmum valu at t m. Ths s n contrast to th standard Wbull modl whr th shap s thr dcrasng (forβ 1) or unmodal (for β > 1). Whn β = 1, th IWD bcoms an nvrs xponntal dstrbuton; Corrspondng author. Emal addrsss: jucd@mathos.hr (D. Juć),darja@mathos.hr (D. Marovć) http://www.jpam.com 230 c 2014 EJPAM All rghts rsrvd.

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 231 whnβ= 2, t s dntcal to th nvrs Raylgh dstrbuton; whnβ= 0.5, t approxmats th nvrs Gamma dstrbuton. That s th rason why th IWD s a frquntly usd modl n rlablty and lftm studs (s.g. Cohn and Whttn[5], Lawls[18], Murthy t al.[21], Nlson[22]). f(t) β=0.5 t Fgur 1: Plots of th nvrs Wbull dnsty for som valus ofβ and by assumngα=0and η=1.2 β=3 β=2 β=1 In practc, th unnown paramtrs α, β and η of th thr-paramtr nvrs Wbull dnsty (1) ar not nown n advanc and must b stmatd from a random sampl t 1,..., t n consstng of n obsrvatons of th thr-paramtr nvrs Wbull random varabl T. Thr s no unqu way to stmat th unnown paramtrs and many dffrnt mthods hav bn proposd n th ltratur (s.g. Abbas t al.[1], Lawlss[18], Marušć t al.[20], Murthy t al.[21], Nlson[22], Slvrman[26], Smth and Naylor[27, 28], Tapa and Thompson[29]). A vry popular mthod for paramtr stmaton s th last squars mthod. Th nonlnar wghtd ordnary last squars (OLS) fttng problm for th thr-paramtr nvrs Wbull dnsty s consdrd by Marušć t al. [20]. In ths papr w consdr th nonlnar wghtd total last squars (TLS) fttng problm for th thr-paramtr nvrs Wbull dnsty functon. Th structur of th papr s as follows. In Scton 2 w brfly dscrb th TLS mthod and prsnt our man rsult (Thorm 1) whch guarants th xstnc of th TLS stmator for th thr-paramtrc nvrs Wbull dnsty. Its gnralzaton n th l q norm (1 q ) s gvn n Thorm 2. All proofs ar gvn n Scton 3. 2. Th TLS Fttng Problm for th Thr-paramtr Invrs Wbull Dnsty Both th OLS and th TLS mthod rqur th ntal nonparamtrc dnsty stmats ˆf whch nd to b as good as possbl (s.g. Slvrman[26], Marušć t al.[20]). Suppos w ar gvn th ponts(t, y ), = 1,..., n, n>3, whr 0 t 1 t 2... t n ar obsrvatons of th nonngatv thr-paramtr nvrs Wbull random varabl T and y := ˆf(t ) ar th rspctv dnsty stmats.

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 232 Th goal of th OLS mthod (s.g.[2, 3, 8, 11, 13, 14, 19, 25]) s to choos th unnown paramtrs of dnsty functon (1) such that th wghtd sum of squard dstancs btwn th modl and th data s as small as possbl. To b mor prcs, lt w > 0, = 1,..., n, b th data wghts whch dscrb th assumd rlatv accuracy of th data. Th unnown paramtrsα,β andηhav to b stmatd by mnmzng th functonal on th st S(α,β,η)= n w [f(t ;α,β,η) y ] 2 =1 := (α,β,η) 3 :α 0;β,η>0. A pont(α,β,η ) such that S(α,β,η )=nf (α,β,η) S(α,β,η) s calld th OLS stmator, f t xsts. As w hav alrady mntond, ths problm has bn solvd by Marušć t al.[20]. In th OLS approach th obsrvatons t of th ndpndnt varabl ar assumd to b xact and only th stmats y of th dnsty (dpndnt varabl) ar subjct to random rrors. Unfortunatly, ths assumpton dos not sm to b vry ralstc n practc, and many rrors (samplng rrors, human rrors, modlng rrors and nstrumnt rrors) prvnt us from nowng t xactly. In such stuaton, whn also th obsrvatons of th ndpndnt varabl contans rrors, t sms rasonabl to stmat th unnown paramtrs so that th wghtd sum of squars of all rrors s mnmzd. Ths approach, nown as th total last squars (TLS) mthod, s a natural gnralzaton of th OLS mthod (s.g. [7]). In th statstcs ltratur, th TLS approach s nown as rrors-n-varabls rgrsson or orthogonal dstanc rgrsson, and n numrcal analyss t was frst consdrd by Golub and Van Loan[9]. Th TLS mthod can b dscrbd as follows. Lt w, p > 0, = 1,..., n, b som wghts. If w assum that y contans unnown addtv rrorǫ and that t has unnown addtv rrorδ, thn th mathmatcal modl bcoms y = f(t +δ ;α,β,η)+ǫ, = 1,..., n. Th unnown paramtrsα,β andηof dnsty functon (1) hav to b stmatd by mnmzng th wghtd sum of squars of all rrors,.. by mnmzng th functonal (s.g. [4, 7, 10, 17, 24]) T(α,β,η,δ)= n w [f(t +δ ;α,β,η) y ] 2 + =1 n p δ 2 (2) on th st n. A pont(α,β,η ) n s calld th total last squars stmator (TLS stmator) of th unnown paramtrs(α, β, η) for th thr-paramtr nvrs Wbull dnsty, f thr xstsδ n such that =1 T(α,β,η,δ )= nf (α,β,η,δ) n T(α,β,η,δ). Numrcal mthods for solvng th nonlnar TLS problm ar dscrbd n Boggs t al. [4] and Schwtlc and Tllr[24]. As n th cas of th OLS approach, bfor th tratv

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 233 mnmzaton of th sum of squars t s stll ncssary to as whthr th TLS stmator xsts. In th cas of nonlnar TLS problms t s stll xtrmly dffcult to answr ths quston (s.g.[3, 7, 12, 15 17]). Th dffrnc btwn th OLS and th TLS approach s llustratd n Fg. 2. Gomtrcally, f w = p for all = 1,..., n, mnmzaton of functonal T corrsponds to mnmzaton of th wghtd sum of squars of dstancs from data ponts to th modl curv. y (t,y ) y (t,y ) (t +δ,f(t +δ ;α,β,η)) t (a) OLS (b) TLS Fgur 2: Th dffrnc btwn th OLS and TLS approachs t Our man xstnc rsult for th TLS problm for th thr-paramtr nvrs Wbull dnsty s gvn n th nxt thorm. Thorm 1. Lt th ponts(t, y ), = 1,..., n, n>3, b gvn, such that 0 t 1 t 2... t n and y > 0, = 1,..., n. Furthrmor, lt w, p > 0, = 1,..., n, b som wghts. Thn thr xsts a pont(α,β,η,δ ) n such that.. th TLS stmator xsts. T(α,β,η,δ )= nf (α,β,η,δ) n T(α,β,η,δ), Th proof s gvn n Scton 3. Th followng total l q norm (q 1) gnralzaton of Thorm 1 holds tru. Thorm 2. Suppos 1 q. Lt th ponts and wghts b th sam as n Thorm 1. Dfn n n T q (α,β,η,δ) := w f(t +δ ;α,β,η) y q + p δ q. (3) Thn thr xsts a pont(α q,β q,η q,δ q ) n such that =1 =1 T q (α q,β q,η q,δ q )= nf T q(α,β,η,δ). (α,β,η,δ) n Th proof of ths thorm s omttd as t s smlar to that of Thorm 1. It suffcs to rplac th l 2 norm wth th l q norm. Thrby all parts of th proof rman th sam.

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 234 3. Proof of Thorm 1 Bfor startng th proof of Thorm 1, w nd som prlmnary rsults. Lmma 1. Suppos w ar gvn data(w, t, y ), I :={1,..., n}, n>3, such that 0 t 1 t 2... t n and y > 0, I. Lt w, p > 0, I, b som wghts. Gvn any ral numbr q, 1 q, and any nonmpty subst I 0 of I, lt Σ I0 := w y q + p t τ 0 q, I\I 0 I 0 whr τ 0 argmn x n p t x q. =1 Thn thr xsts a pont n n at whch functonal T q dfnd by (3) attans a valu lss than Σ I0. Summaton I 0 s to b undrstood as follows: Th sum ovr thos ndcs nfor whch I 0. If thr ar no such ndcs, th sum s mpty; followng th usual convnton, w dfn t to b zro. Summaton I\I 0 has smlar manngs. It s asy to vrfy that t 1 mn I0 t τ 0 max I0 t t n. Not that for th cas whn q=2,τ 0 s a wll nown wghtd arthmtc man, and for th cas whn q=1,τ 0 s a wghtd mdan of th data (s.g. Sabo and Sctovs[23]). Proof. Sncτ 0 s an lmnt of th closd ntrval[t 1, t n ], thr xsts r {1,..., n} such that τ 0 (t r 1, t r ], whr t 0 = 0 by dfnton. Lt us frst choos ral y 0 such that 0 y 0 mn I and thn dfn functonsα,β,η :(0,1) by: b β(b) :=τ 0 y 0 b η(b) :=τ 0 b 1/β(b), y (4) α(b) :=τ 0 η(b)b 1/2β(b) =η(b) b 1/β(b) b 1/2β(b). Clarly, functonsβ andηar postv. Furthrmor, by usng th nqualty b 1/β(b) b 1/2β(b) > 0, whch holds for vry b (0,1), t s asy to show that functonαs also postv. Thus, w hav showd that(α(b),β(b),η(b)) for all b (0,1). Lt us now assocat wth ach ral b (0,1) a thr-paramtrc nvrs Wbull dnsty functon β(b) η(b) β(b) β(b) f(t;α(β),β(b),η(b))= t α(b) t α(b) η(b) t α(b) t>α(b) (5) 0 t α(b).

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 235 Ths functon has maxmum at th pont whr α(b)+η(b)(1+1/β(b)) 1/β(b) =τ 0 ǫ(b), ǫ(b) :=η(b) b 1/2β(b) 1+ 1 1/β(b). β(b) It s strctly ncrasng on(α(b),τ 0 ǫ(b)] and strctly dcrasng on[τ 0 ǫ(b), ). Furthrmor, by a straghtforward calculaton, t can b vrfd that Now w ar gong to show that Frst, n vw of (8) and (9), w obtan f(τ 0 +α(b);α(b),β(b),η(b))= y 0, (6) lmβ(b)=, b 0 (7) lm 0, b 0 (8) lmα(b)=0. b 0 (9) lm f(t;α(b),β(b),η(b))=0, t τ 0. (10) b 0 η(b) lm = τ 0 b 0 t α(b) t. β(b) Ifτ 0 t, thn from (7) and (9) t follows radly that lm b 0 η(b) t α(b) = 1 and lm b 0 β(b) lm b 0 η(b) t α(b) β(b)= 0, and thrfor β(b) η(b) β(b) f(t;α(b),β(b),η(b))= lm η(b) β(b) t α(b) = 0. b 0 t α(b) t α(b) Ifτ 0 > t, thn thr xsts a suffcntly grat 0 such that η(b) 0 t α(b) for vry suffcntly small b>0. Now, by usng th nqualty x x (x 0) w obtan η(b) β(b) β(b) 0 β(b), t α(b) b 0, and thrfor, for any b 0 w hav 0f(t;α(b),β(b),η(b))= β(b) η(b) β(b) η(b) β(b) t α(b) t α(b) t α(b)

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 236 Snc 1 η(b) β(b) (0 +1)β(b) η(b) t α(b). t α(b) t α(b) η(b) β(b) (0 +1)β(b) lm η(b) t α(b) = 0, b 0 t α(b) thn from th abov-mntond nqualty t follows that lm f(t;α(b),β(b),η(b))=0, t>τ 0. b 0 Thus, w provd th dsrd lmts (10). Not that f(τ 0 ;α(b),β(b),η(b))= β(b) η(b) β(b) η(b) β(b) τ 0 α(b) τ 0 α(b) τ 0 α(b) = β(b) b b = τ 0 y 0 b b τ 0 α(b) (τ 0 α(b)) b, from whr tang th lmt as b 0 t follows that lm f(τ 0;α(b),β(b),η(b))=. (11) b 0 Du to (9), (10) and (11), w may suppos that b s suffcntly small, so that 0α(b) t 1 (12) 0 f(t ;α(b),β(b),η(b)) y, f t τ 0 (13) f(τ 0 ;α(b),β(b),η(b))>max I y. (14) Lt us now show that for ach I 0 and for vry b (0,1) thr xsts a unqu numbr τ (b) such that (s Fgur 3) t τ (b)τ 0 ǫ(b)τ 0, f t τ 0 t τ (b) t +ǫ(b), f t =τ 0 (15) τ 0 τ (b) t, f t >τ 0 and f(τ (b);α(b),β(b),η(b))= y. (16) Frst, snc th functon t f(t ;α(b),β(b),η(b)) has maxmum at th pontτ 0 ǫ(b) and t s strctly ncrasng on(α(b),τ 0 ǫ(b)] and strctly dcrasng on[τ 0 ǫ(b), ), by usng (4), (6), (13) and (14) w obtan f(t ;α(b),β(b),η(b)) y f(τ 0 ǫ(b);α(b),β(b),η(b)), f t τ 0 f(t ;α(b),β(b),η(b)) y f(τ 0 ;α(b),β(b),η(b)), f t >τ 0 f(t +ǫ(b);α(b),β(b),η(b)) y f(t ;α(b),β(b),η(b)), f t =τ 0.

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 237 Th xstnc of th dsrd numbrsτ (b), I 0, follows from th wll-nown Intrmdat Valu Thorm whch stats that a contnuous ral functon assums all ntrmdat valus on a closd ntrval, whl unqunss follows from monotoncty. f(t) (t, y ) (τ (b), y ) (τ j (b), y j ) (t j, y j ) τ 0 ε(b) τ 0 τ 0 +ε(b) Fgur 3:, j I 0, t τ 0, t j >τ 0 ; 0δ (b)=τ (b) t τ 0 ǫ(b) t τ 0 t ; τ 0 t j τ j (b) t j =δ j (b)0 t Sttng (16) bcoms δ (b) := τ (b) t, f I 0 0, f I\I 0, (17) f(t +δ (b);α(b),β(b),η(b))= y, I 0. (18) Not that only on of th followng two cass can occur: () I 0 =1, or () I 0 >1. Cas (): I 0 =1. In ths cas w havτ 0 = t r. It follows from (15) that 0δ r (b)ǫ(b). Wthout loss of gnralty, n addton to (12)-(14) w may suppos that b s suffcntly small, so that t r 1 +ǫ(b) t r 1+t r t r ǫ(b) 2 and f((t r 1 + t r )/2;α(b),β(b),η(b))mn y. I Du to ths two addtonal assumptons and th fact that th functon t f(t ;α(b),β(b),η(b)) s strctly ncrasng on(α(b), t r ǫ(b)] and strctly dcrasng on[t r ǫ(b), ), w dduc:

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 238 If t t r, thn whras f t > t r, thn 0f(t ;α(b) δ r (b),β(b),η(b))= f(t +δ r (b);α(b),β(b),η(b)) f(t +ǫ(b);α(b),β(b),η(b)) f(t r 1 +ǫ(b);α(b),β(b),η(b)) f((t r 1 + t r )/2;α(b),β(b),η(b))mn I y y, (19) 0f(t ;α(b) δ r (b),β(b),η(b))= f(t +δ r (b);α(b),β(b),η(b)) f(t ;α(b),β(b),η(b)) y. (20) Thus, t follows from (18), (19) and (20) that, for vry b (0,1), T q (α(b) δ r (b),β(b),η(b),0)= n w f(t ;α(b) δ r (b),β(b),η(b)) y q =1 n w y q =Σ I0 Cas I 0 >1. Not that only on of th followng two subcass can occur: () τ 0 t for all I 0, or () τ 0 = t r for som r I 0. =1 r Subcas (): In ths subcas, t follows from (13), (15), (17) and (18) that, for vry b (0,1), T q (α(b),β(b),η(b),δ(b))= f(t ;α(b),β(b),η(b)) y q + I\I 0 w w y q + p t τ 0 q =Σ I0. I\I 0 I 0 I 0 p δ (b) q Subcas (): Assum thatτ 0 = t r for som r I 0. Lt ndx s I 0 b such that t s τ 0. Thn by (15), for vry b (0,1), and thrfor 0δ r (b)ǫ(b) and 0δ s (b) t r ǫ(b) t s p s δ s (b) q + p r δ r (b) q p s t r ǫ(b) t s q + p r ǫ q (b). It can b asly shown that th abov rght-hand sd s lss than p s t r t s q whnvr b s small nough. Thrfor, for vry small nough b w hav T q (α(b),β(b),η(b),δ(b))= f(t ;α(b),β(b),η(b)) y q I\I 0 w

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 239 + p r δ r (b) q + p s δ s (b) q + p δ (b) q Ths complts th proof of th lmma. I 0 \{r,s} w y q + p t τ 0 q =Σ I0. I\I 0 I 0 Proof of Thorm 1. Proof. Snc functonal T s nonngatv, thr xsts T := nf (α,β,η,δ) n T(α,β,η,δ). To complt th proof t should b shown that thr xsts a pont(α,β,η,δ ) n such that T(α,β,η,δ )= T. Lt(α,β,,δ ) b a squnc n n, such that T = lm T(α,β,,δ )= lm = lm t +δ α w y 2 + t +δ >α w f(t +δ ;α 2+,β, ) y p (δ )2 I β η β+1 w t +δ α I β 2 y + p (δ )2. (21) I whr I={1,..., n}. Th summaton t +δ α (or t +δ >α ) s to b undrstood as follows: Th sum ovr thos ndcs nfor whch t +δ α (or t +δ >α ). If thr ar no such ponts t, th sum s mpty; followng th usual convnton, w dfn t to b zro. Thr s no loss of gnralty n assumng that all squncs(α ),(β ),( ),(δ 1 ),...,(δ n ) ar monoton. Ths s possbl bcaus th squnc(α,β,,δ 1,...,δ n ) has a subsqunc(α l,β l,η l,δ l 1,...,δl n ), such that all ts componnt squncs ar monoton; and snc lm T(α l,β l,η l,δ l )=lm T(α,β,,δ )= T. Snc ach monoton squnc of ral numbrs convrgs n th xtndd ral numbr systm, dfn α := lm α, β := lm β, η := lm, δ := lm δ =(δ 1,...,δ n ). Not that 0 α,β,η, bcaus(α,β, ). Also not thatδ for ach = 1,..., n. Indd, f δ = for som, thn t would follow from (21) that T =, whch s mpossbl. To complt th proof t s nough to show that(α,β,η ),.. that 0 α andβ,η (0, ). Th contnuty of th functonal T wll thn mply that T = lm T(α,β,,δ )= T(α,β,η,δ ).

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 240 It rmans to show that(α,β,η ). Th proof wll b don n fv stps. In stp 1 w wll show thatα t n. In stp 2 w wll show thatβ 0. Th proof thatη wll b don n stp 3. In stp 4 w prov thatη 0. Fnally, n stp 5 w show thatβ. Stp 1. Ifα t n, from (21) t follows that T = n =1 w y 2 + I p δ 2. Snc accordng to Lmma 1 (for q=2and I 0 ={1}) thr xsts a pont n n at whch functonal T attans a valu smallr thanσ I0 and sncσ I0 n =1 w y 2 + I p δ 2, ths mans that n ths way (α t n ) functonal T cannot attan ts nfmum. Thus, w hav provd thatα t n. Bfor contnung th proof, lt us ntroduc som notaton and ma on rmar. Frst lt us dfn Iα, f I α I 0 := {1}, othrws whr I α :={ I : t +δ =α }. Lt us not that Lmma 1 wth q=2 mpls that whrτ I0 = I 0 p t I 0 p. T w y 2 + p (t τ I0 ) 2 =:Σ I0, (22) I\I 0 I 0 By tang an approprat subsqunc of(α,β,,δ ), f ncssary, w may assum that f t +δ α, thn t +δ α for vry. Smlarly, f t +δ >α, w may assum that t +δ >α for vry. Du to ths, now t s asy to show that from (21) t follows that T w y 2 + lm t +δ α + I t +δ >α w β η β+1 t +δ α β 2 y p δ 2. (23) Stp 2. Ifβ = 0, thn by usng th nqualty x x (x 0) w obtan 0 β η β+1 t +δ α whrfrom t follows radly that β η β+1 lm t +δ α β β t +δ α, f t +δ >α, β = 0, f t +δ >α. Now, from (23) t follows that T w y 2 + I\I 0 p δ 2 I 0

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 241 = w y 2 + p (t α ) 2 I\I 0 I 0 w y 2 + p (t τ I0 ) 2 =Σ I0, (24) I\I 0 I 0 whch contradcts (22). Thrfor, n ths way (β = 0) functonal T cannot attan ts nfmum. Thus, w hav provd thatβ 0. Th last nqualty n (24) follows drctly from a wll-nown fact that th quadratc functon x I 0 p (t x) 2 attans ts mnmum I 0 p (t τ I0 ) 2 at pontτ I0. Stp 3. Lt us show thatη. W prov ths by contradcton. Suppos on th contrary thatη =. Wthout loss of gnralty, w may thn assum that f t +δ >α, thn η for all. Thn from th nqualty x x (x 0) t follows that f t +δ >α, thn Thus, f t +δ >α, thn 0 β η β+1 t +δ α β β η β t +δ α,. 1 η 2β t +δ α t +δ α β β η β = η β t +δ t +δ α t +δ α α η β t +δ α. (25) η Furthrmor, snc lm t +δ = andβ η 0, w hav lm β α = η and thrfor lm 2β η β t +δ t +δ α α = 0, so that from (25) t follows that β η β+1 lm t +δ α Puttng th abov lmts nto (23), w mmdatly obtan T w y 2 + p δ 2 Σ I0, I\I 0 I β = 0, f t +δ >α. whch contradcts (22). Hnc w provd thatη. So far w hav shown thatα t n,β 0 andη. By usng ths, n th nxt stp w wll show thatη 0. Stp 4. Lt us show thatη 0. To s ths, suppos on th contrary thatη = 0. Thn only on of th followng two cass can occur:

D. Juć, D. Marovć/Eur. J. Pur Appl. Math, 7 (2014), 230-245 242 () η = 0 andβ (0, ), or () η = 0 andβ =. Now, w ar gong to show that functonal T cannot attan ts nfmum n thr of ths two cass, whch wll prov thatη 0. Cas ():η = 0 andβ (0, ). In ths cas w would hav lm β η β+1 t +δ α β β η β = lm t +δ α t +δ α = 0, f t +δ >α and hnc from (23) t would follow that T w y 2 + t +δ α I p δ 2 Σ I0 whch contradcts assumpton (22). Cas ():η = 0 andβ =. Snc 0, thr xsts a ral numbr L> 1 and suffcntly grat 0 such that f t +δ >α and > 0, thn /(t +δ α )1/L. Wthout loss of gnralty, w may assum that 0 = 1. Thus, f t +δ >α, thn 0 β η β+1 t +δ α 1 β t +δ α L β Furthrmor, snc from (26) t follows that β β η β = η β t +δ t +δ α t +δ α α β. (26) β lm L β = 0 and lm β η β+1 lm t +δ α β = 1, β β = 0, f t +δ >α. Fnally, from (23) w obtan T t +δ w y 2 α + I p δ 2 Σ I0, whch contradcts assumpton (22). Ths mans that n ths cas functonal T cannot attan ts nfmum. Thus, w hav provd thatη 0.

REFERENCES 243 Stp 5. It rmans to show thatβ. W prov ths by contradcton. Suppos that β =. Argung as n cas () from stp 4, t can b shown that β η β lm t +δ α t +δ α β = 0, f 0 η t +δ 1. (27) α η If t +δ > 1, thn thr xsts a suffcntly grat α 0 such that η 0. Now, by usng th nqualty x x (x 0) w obtan β β η 0β t +δ α,, and thrfor β η β 0 t +δ α t +δ α 1 η (0+1)β t +δ α t +δ α β β. (28) η Snc lm β =, w hav that and thrfor from (28) t follows that ( lm 0+1)β t +δ α β η β lm t +δ α t +δ α β = 0 β = 0, f From (23), (27) and (29) w would obtan T t +δ w y 2 α + I p δ 2 contradcts (22). Thus, w hav provd thatβ and compltd th proof. η t +δ 1. (29) α > Σ I0, whch Rfrncs [1] B. Abbas, A.H.E. Jahrom, J. Arat, and M. Hossnouchac. Estmatng th paramtrs of Wbull dstrbuton usng smulatd annalng algorthm. Appld Mathmatcs and Computaton, 183:85-93, 2006. [2] D.M. Bats and D.G. Watts. Nonlnar rgrsson analyss and ts applcatons. Wly, Nw Yor, 1988.

REFERENCES 244 [3] Å. Björc. Numrcal Mthods for Last Squars Problms. SIAM, Phladlpha, 1996. [4] P.T. Boggs, R.H. Byrd, and R.B. Schnabl. A stabl and ffcnt algorthm for nonlnar orthogonal dstanc rgrsson. SIAM Journal on Scntfc and Statstcal Computaton, 8:1052-1078, 1987. [5] A.C. Cohn and B.J. Whttn. Paramtr Estmaton n Rlablty and Lf Span Modls. Marcl Dr Inc., Nw Yor and Basl, 1988. [6] P. Erto. Nw Practcal Bays stmators for th 2-Paramtr Wbull dstrbuton. IEEE Transactons on Rlablty, R-31:194-197, 1982. [7] W.A. Fullr. Masurmnt Error Modls. Wly, Nw Yor, 2006. [8] P.E. Gll, W. Murray, and M.H. Wrght. Practcal Optmzaton. Acadmc Prss, London, 1981. [9] G.H Golub and C.F. Van Loan. An analyss of th total last squars problm. SIAM Journal on Numrcal Analyss, 17:883-893, 1980. [10] S. Van Huffl and H. Zha. Th Total Last Squars Problm. Elsvr, North Holland, Amstrdam, 1993. [11] D. Juć. On th l s -norm gnralzaton of th NLS mthod for th Bass modl. Europan Journal of Pur and Appld Mathmatcs, 6:435-450, 2013. [12] D. Juć and D. Marovć. On nonlnar wghtd rrors-n-varabls paramtr stmaton problm n th thr-paramtr Wbull modl. Appld Mathmatcs and Computaton, 215:3599-3609, 2010. [13] D. Juć and D. Marovć. On nonlnar wghtd last squars fttng of th thrparamtr nvrs Wbull dstrbuton. Mathmatcal Communcatons, 15:13-24, 2010. [14] D. Ju, M. Bnšć, and R. Sctovs. On th xstnc of th nonlnar wghtd last squars stmat for a thr-paramtr Wbull dstrbuton. Computatonal Statstcs and Data Analyss, 52:4502-4511, 2008. [15] D. Juć, K. Sabo, and R. Sctovs. Total last squars fttng Mchals-Mntn nzym ntc modl functon. Journal of Computatonal and Appld Mathmatcs, 201:230-246, 2007. [16] D. Juć, R. Sctovs, and H. Späth. Partal lnarzaton of on class of th nonlnar total last squars problm by usng th nvrs modl functon. Computng, 62:163-178, 1999. [17] D. Juć and R. Sctovs. Exstnc rsults for spcal nonlnar total last squars problm. Journal of Mathmatcal Analyss and Applcatons, 226:348-363, 1998.

REFERENCES 245 [18] J.F. Lawlss. Statstcal modls and mthods for lftm data. Wly, Nw Yor, 1982. [19] D. Marovć, D. Juć, and M. Bnšć. Nonlnar wghtd last squars stmaton of a thr-paramtr Wbull dnsty wth a nonparamtrc start. Journal of Computatonal and Appld Mathmatcs, 228:304-312, 2009. [20] M. Marušć, D. Marovć, and D. Juć. Last squars fttng th thr-paramtr nvrs Wbull dnsty. Mathmatcal Communcatons, 15:539-553, 2010. [21] D.N.P. Murthy, M. X, and R. Jang. Wbull Modls. Wly, Nw Yor, 2004. [22] W. Nlson. Appld lf data analyss. Wly, Nw Yor, 1982. [23] K. Sabo and R. Sctovs. Th bst last absolut dvatons ln - proprts and two ffcnt mthods for ts drvaton. ANZIAM Journal., 50:185-198, 2008. [24] H. Schwtlc and V. Tllr. Numrcal mthods for stmatng paramtrs n nonlnar modls wth rrors n th varabls. Tchnomtrcs, 27:17-24, 1985. [25] G.A.F. Sbr and C.J. Wld. Nonlnar Rgrsson. Wly, Nw Yor, 1989. [26] B.W. Slvrman. Dnsty stmaton for Statstcs and Data Analyss. Chapman & Hall/CRC, Boca Raton, 2000. [27] R.L. Smth and J.C. Naylor. A comparson of maxmum llhood and Baysan stmators for th thr-paramtr Wbull dstrbuton. Bomtra, 73:67-90, 1987. [28] R.L. Smth and J.C. Naylor. Statstcs of th thr-paramtr Wbull dstrbuton. Annals of Opratons Rsarch, 9:577-587, 1987. [29] R. A. Tapa and J. R. Thompson. Nonparamtrc probablty dnsty stmaton. Johns Hopns Unvrsty Prss, Baltmor, 1978.