Appendix. Title. Petr Lachout MFF UK, ÚTIA AV ČR

Similar documents
ROBUST - September 10-14, 2012

Universal Confidence Sets for Solutions of Optimization Problems

Empirical Processes: General Weak Convergence Theory

Kybernetika. Jan Havrda; František Charvát Quantification method of classification processes. Concept of structural a-entropy

Statistics 581 Revision of Section 4.4: Consistency of Maximum Likelihood Estimates Wellner; 11/30/2001

AW -Convergence and Well-Posedness of Non Convex Functions

Econometrica Supplementary Material

Closest Moment Estimation under General Conditions

Epiconvergence and ε-subgradients of Convex Functions

Introduction to Empirical Processes and Semiparametric Inference Lecture 08: Stochastic Convergence

RANDOM APPROXIMATIONS IN STOCHASTIC PROGRAMMING - A SURVEY

DELTA METHOD, INFINITE DIMENSIONAL

Introduction to Empirical Processes and Semiparametric Inference Lecture 09: Stochastic Convergence, Continued

The Golden Ratio and Signal Quantization

Goodness-of-Fit Tests for Time Series Models: A Score-Marked Empirical Process Approach

THEOREMS, ETC., FOR MATH 515

Closest Moment Estimation under General Conditions

Measurable Choice Functions

Stochastic Convergence, Delta Method & Moment Estimators

A convergence result for an Outer Approximation Scheme

Berge s Maximum Theorem

Weak convergence. Amsterdam, 13 November Leiden University. Limit theorems. Shota Gugushvili. Generalities. Criteria

Extremal I-Limit Points Of Double Sequences

Section 45. Compactness in Metric Spaces

Takens embedding theorem for infinite-dimensional dynamical systems

ON THE UNIQUENESS PROPERTY FOR PRODUCTS OF SYMMETRIC INVARIANT PROBABILITY MEASURES

Keywords. 1. Introduction.

SUMMARY OF RESULTS ON PATH SPACES AND CONVERGENCE IN DISTRIBUTION FOR STOCHASTIC PROCESSES

Efficiency of Profile/Partial Likelihood in the Cox Model

SOME CONVERSE LIMIT THEOREMS FOR EXCHANGEABLE BOOTSTRAPS

Estimation of Dynamic Regression Models

Estimation of the Bivariate and Marginal Distributions with Censored Data

University of California San Diego and Stanford University and

SDS : Theoretical Statistics

Asymptotic inference for a nonstationary double ar(1) model

Applied Analysis (APPM 5440): Final exam 1:30pm 4:00pm, Dec. 14, Closed books.

ON THE CONTINUITY OF GLOBAL ATTRACTORS

A Concise Course on Stochastic Partial Differential Equations

OPTIMAL TRANSPORTATION PLANS AND CONVERGENCE IN DISTRIBUTION

An introduction to Mathematical Theory of Control

A SKOROHOD REPRESENTATION THEOREM WITHOUT SEPARABILITY PATRIZIA BERTI, LUCA PRATELLI, AND PIETRO RIGO

Appendix B Convex analysis

Strong Consistency of Set-Valued Frechet Sample Mean in Metric Spaces

Three hours THE UNIVERSITY OF MANCHESTER. 24th January

Generalized Neyman Pearson optimality of empirical likelihood for testing parameter hypotheses

WEAK CONVERGENCES OF PROBABILITY MEASURES: A UNIFORM PRINCIPLE

An elementary proof of the weak convergence of empirical processes

Fundamental Inequalities, Convergence and the Optional Stopping Theorem for Continuous-Time Martingales

On a Class of Multidimensional Optimal Transportation Problems

Stability and Sensitivity of the Capacity in Continuous Channels. Malcolm Egan

The Codimension of the Zeros of a Stable Process in Random Scenery

Weak convergence and Compactness.

GLUING LEMMAS AND SKOROHOD REPRESENTATIONS

A BOREL SOLUTION TO THE HORN-TARSKI PROBLEM. MSC 2000: 03E05, 03E20, 06A10 Keywords: Chain Conditions, Boolean Algebras.

Serena Doria. Department of Sciences, University G.d Annunzio, Via dei Vestini, 31, Chieti, Italy. Received 7 July 2008; Revised 25 December 2008

s P = f(ξ n )(x i x i 1 ). i=1

Convergence of Feller Processes

is a Borel subset of S Θ for each c R (Bertsekas and Shreve, 1978, Proposition 7.36) This always holds in practical applications.

Zdzis law Brzeźniak and Tomasz Zastawniak

VARIATIONAL PRINCIPLE FOR THE ENTROPY

ON THE COMPLETE CONVERGENCE FOR WEIGHTED SUMS OF DEPENDENT RANDOM VARIABLES UNDER CONDITION OF WEIGHTED INTEGRABILITY

APPROXIMATING BOCHNER INTEGRALS BY RIEMANN SUMS

n [ F (b j ) F (a j ) ], n j=1(a j, b j ] E (4.1)

arxiv:submit/ [math.st] 6 May 2011

II - REAL ANALYSIS. This property gives us a way to extend the notion of content to finite unions of rectangles: we define

Amenability properties of the non-commutative Schwartz space

University of Pavia. M Estimators. Eduardo Rossi

Theoretical Statistics. Lecture 17.

Přednáška 13. Teorie aritmetiky a Gödelovy výsledky o neúplnosti a nerozhodnutelnosti. 12/6/2006 Kurt Gödel 1

WEAK LOWER SEMI-CONTINUITY OF THE OPTIMAL VALUE FUNCTION AND APPLICATIONS TO WORST-CASE ROBUST OPTIMAL CONTROL PROBLEMS

BERNOULLI ACTIONS AND INFINITE ENTROPY

LECTURE 15: COMPLETENESS AND CONVEXITY

Asymptotics of minimax stochastic programs

1 Weak Convergence in R k

Stat 8112 Lecture Notes Weak Convergence in Metric Spaces Charles J. Geyer January 23, Metric Spaces

ITERATED FUNCTION SYSTEMS WITH CONTINUOUS PLACE DEPENDENT PROBABILITIES

Robustness for a Liouville type theorem in exterior domains

A log-scale limit theorem for one-dimensional random walks in random environments

LIMITS FOR QUEUES AS THE WAITING ROOM GROWS. Bell Communications Research AT&T Bell Laboratories Red Bank, NJ Murray Hill, NJ 07974

Bonus Homework. Math 766 Spring ) For E 1,E 2 R n, define E 1 + E 2 = {x + y : x E 1,y E 2 }.

Lecture 4 Lebesgue spaces and inequalities

A note on the σ-algebra of cylinder sets and all that

SOLUTION TO RUBEL S QUESTION ABOUT DIFFERENTIALLY ALGEBRAIC DEPENDENCE ON INITIAL VALUES GUY KATRIEL PREPRINT NO /4

A GENERAL THEOREM ON APPROXIMATE MAXIMUM LIKELIHOOD ESTIMATION. Miljenko Huzak University of Zagreb,Croatia

A Note on Some Properties of Local Random Attractors

7 Complete metric spaces and function spaces

Consistency of the maximum likelihood estimator for general hidden Markov models

Weak convergence and Brownian Motion. (telegram style notes) P.J.C. Spreij

Duality of multiparameter Hardy spaces H p on spaces of homogeneous type

SOME MEASURABILITY AND CONTINUITY PROPERTIES OF ARBITRARY REAL FUNCTIONS

van Rooij, Schikhof: A Second Course on Real Functions

ON A DISTANCE DEFINED BY THE LENGTH SPECTRUM ON TEICHMÜLLER SPACE

Nonparametric one-sided testing for the mean and related extremum problems

Theoretical Statistics. Lecture 1.

Jádrové odhady gradientu regresní funkce

LYAPUNOV STABILITY OF CLOSED SETS IN IMPULSIVE SEMIDYNAMICAL SYSTEMS

Probability metrics and the stability of stochastic programs with recourse

Stability of optimization problems with stochastic dominance constraints

A Law of the Iterated Logarithm. for Grenander s Estimator

Jádrové odhady regresní funkce pro korelovaná data

Transcription:

Title ROBUST - Kráĺıky - únor, 2010

Definice Budeme se zabývat optimalizačními úlohami. Uvažujme metrický prostor X a funkci f : X R = [, + ]. Zajímá nás minimální hodnota funkce f na X ϕ (f ) = inf {f (x) x X }. (1)

Definice Hodnota + vznikne přepisem podmínek A X, za nichž minimalizujeme funkci g : A R. Tedy úlohu přepisujeme na tvar kde inf {g(x) x A} inf {f (x) x X }, f (x) = g(x) pro x A, = + pro x A.

Definice Funkce, které nabývají pouze hodnoty + nebo jejichž minimální hodnota je, nejsou z hlediska optimalizace příliš zajímavé. V dalším se proto omezíme pouze na funkce, které jsou vlastní, tj. pro něž ϕ (f ) R.

Definice Definujeme množinu všech minimálních řešení množiny všech ε-minimálních řešení a úrovňové množiny Φ (f ) = {x X f (x) = ϕ (f )}, (2) Ψ (f ; ε) = {x X f (x) ϕ (f ) + ε} ε R, (3) lev (f ) = {x X f (x) } R. (4) Potřebujeme také vhodnou míru vzdálenosti mezi dvěmi množinami. Takovou vhodnou mírou je excess z množiny A X do množiny B X kde d je metrika na X. excess (A, B) = sup inf d (a, b), (5) a A b B

Definice Aproximativním řešením minimalizace funkce f na množině X budeme rozumět θ X, které splňuje bud nebo pro předem zadané ε či. θ Ψ (f ; ε) θ lev (f ),

Definice Zajímáme se o posloupnost minimalizačních úloh minimalizovat funkci f t na množině X pro t N. Zajímá nás asymptotické chování příslušných aproximativních řešení. Jejich konvergence a řád konvergence.

Definice-posloupnost Posloupností aproximativních řešení chápeme posloupnost θ t X, t N, která splňuje bud, pro předem zadaná ε t, t N splňují θ t Ψ (f t ; ε t ), nebo pro předem zadaná t, t N splňují θ t lev t (f t ).

Náhodnost Nyní přidejme závislost sledovaných úloh na náhodě. Budeme uvažovat pravděpodobnostní prostor (Ω, A, Prob) a pro t N Případně f t : X Ω R, ϕ (f t (, ω)) R ω Ω, ε t : Ω R. t : Ω R.

Náhodnost Nyní musíme zobecnit, vysvětlit pojem posloupnosti aproximativních řešení pro případ, kdy úlohy závisí na náhodě. Budeme uvažovat posloupnost zobrazení ˆθ t : Ω X, t N. Máme několik možností, jaké vlastnosti, kvalitu od ní vyžadovat.

Náhodnost-1 Posloupnost ˆθ t : Ω X, t N nazveme náhodná ε-minimální řešení, pokud ˆθ t (ω) Ψ (f t (, ω); ε t (ω)) ω Ω t N, náhodná úrovňově-minimální řešení, pokud ˆθ t (ω) lev t(ω) (f t (, ω)) ω Ω t N.

Náhodnost-2 Posloupnost ˆθ t : Ω X, t N nazveme skoro jistě ε-minimální řešení, pokud existuje Ω 0 Ω takové, že Prob (Ω 0 ) = 1 a ˆθ t (ω) Ψ (f t (, ω); ε t (ω)) ω Ω 0 t N, skoro jistě úrovňově-minimální řešení, pokud existuje Ω 0 Ω takové, že Prob (Ω 0 ) = 1 a ˆθ t (ω) lev t(ω) (f t (, ω)) ω Ω 0 t N.

Náhodnost-3 Posloupnost ˆθ t : Ω X, t N nazveme striktně asymptotická ε-minimální řešení, pokud existuje Ω 0 Ω takové, že Prob (Ω 0 ) = 1 a ω Ω 0 ˆθ t (ω) Ψ (f t (, ω); ε t (ω)) t N dostatečně velká, striktně asymptotická úrovňově-minimální řešení, pokud existuje Ω 0 Ω takové, že Prob (Ω 0 ) = 1 a ω Ω 0 ˆθ t (ω) lev t(ω) (f t (, ω)) t N dostatečně velká.

Náhodnost-4 Posloupnost ˆθ t : Ω X, t N nazveme slabě asymptotická ε-minimální řešení, pokud lim t + Prob ( ω Ω : ˆθt (ω) Ψ (f t (, ω); ε t (ω)) slabě asymptotická úrovňově-minimální řešení, pokud ( ) lim Prob ω Ω : ˆθt (ω) lev t(ω) (f t (, ω)) t + ) = 1, = 1.

Konvergence Cílem je ukázat asymptotické chování náhodných aproximativních řešení. Jelikož však nepředpokládáme měřitelnost, je teorie konvergncí trochu složitější. Budeme uvažovat čtyři typy konvergence: as, as, v Prob. a v distribuci. Definice těchto konvergencí, jejich vlastnosti a vztahy jsou uvedeny v [VaartWellner(1996)]. Definition of convergences

Konvergence Je třeba si uvědomit, že stačí vyšetřovat pouze náhodná ε-minimální řešení. Všechny ostatní typy lze na ně převést.

Převod Nejdříve si povšimněme přímého vztahu mezi ε-optimálním řešením a level-optimálním řešením. Ψ (h; ε) = lev ϕ(h)+ε (h), (6) lev (h) = Ψ (h; ϕ (h)). (7)

Konstrukce Uvažujme posloupnost zobrazení ˆθ t : Ω X, t N a ε t : Ω R +,0, t N. Zvolme libovolně, ale pevně, α t : Ω R + s vlastností lim t + α t (ω) = 0 pro každé ω Ω. Nyní dokážeme vybrat ξ t : Ω X, t N tak, že Položme dále ξ t (ω) Ψ (f t (, ω); ε t (ω) + α t (ω)) ω Ω, t N. η t (ω) = ˆθ t (ω) když ˆθ t (ω) Ψ (f t (, ω); ε t (ω)) = ξ t (ω) jinak. Konstrukce dává posloupnost η t, t N, která splňuje η t (ω) Ψ (f t (, ω); ε t (ω) + α t (ω)) ω Ω, t N. Nyní porovnáme asymptotické vlastnosti ˆθ t, t N a η t, t N.

Konstrukce-vztahy Lemma Necht ˆθt : Ω X jsou skoro jistě ε-minimální řešení nebo striktně asymptotická ε-minimální řešení. Pak, existuje Ω 0 Ω, Prob (Ω 0 ) = 1 tak, že pro každé ω Ω 0 ˆθ t (ω) = η t (ω) t N dostatečně velká. Následně pro každé ω Ω 0, ) Ls (ˆθt (ω), t N = Ls (η t (ω), t N).

Konstrukce-vztahy Označme ε(ω) = lim sup t + ε t (ω). Když vezmeme náhodná zobrazení τ t : Ω R, pak pro každé ω Ω 0 [ ] lim τ t(ω) f t (ˆθ t (ω), ω) f t (η t (ω), ω) = 0, t + [ ( ) lim τ t(ω) excess {ˆθ t (ω)}, Ψ (f t (, ω); ε(ω)) t + excess ({η t (ω)}, Ψ (f t (, ω); ε(ω))) ] = 0.

Konstrukce-vztahy Lemma Necht ˆθt : Ω X jsou slabě asymptotická ε-minimální řešení. Pak, ) lim Prob (ˆθt = η t = 1. t +

Konstrukce-vztahy Označme ε(ω) = lim sup t + ε t (ω). Následně, když vezmeme náhodná zobrazení τ t : Ω R, pak ] Prob τ t [f t (ˆθ t ) f t (η t ) 0, t + ( ) τ t [excess {ˆθ t }, Ψ (f t ; ε) excess ({η t }, Ψ (f t ; ε)) ] Prob t + 0.

Appendix-Topology We have to recall a few from topological terminology. Definition For a sequence η n, n N in a metric space W, we denote the set of its cluster points by Ls (η n, n N), i.e. Ls (η n, n N) = { ψ W subsequence s.t. lim } η n k = ψ. k + Back

Appendix-Topology Definition For a sequence A n, n N subsets of a metric space W, we denote the set of its cluster points by Ls (A n, n N), i.e. Ls (A n, n N) = { ψ W } subsequence A nk and a sequence η k A nk s.t. lim η k = ψ. k +

Appendix-Topology Definition We say that a sequence η n, n N in a metric space W is compact if each its subsequence possesses at least one cluster point. Compact sequence in metric space possesses an equivalent description. Lemma Let η n, n N be a sequence in a metric space W. Then, the following statements are equivalent: 1. The sequence is compact. 2. There is a compact L W tak, že η n L for all n N. 3. The set {η n n N} Ls (η n, n N) is compact.

Appendix-Topology Lemma Let η n, n N be a sequence in a metric space W and K W be a compact. If for every open set G K there is an n G N tak, že η n G for all n N, n n G. Then the sequence is compact and Ls (η n, n N) K.

Appendix-Nonmeasurable mappings This auxiliary section contains all necessary theory na nonmeasurable mappings. Definitions and relations are taken from [VaartWellner(1996)], chapter 1.9., p.52-56. All proofs can be found in [VaartWellner(1996)], also. Definition For a set B Ω its outer probability is defined as Prob (B) = inf {Prob (A) B A, A A} and the inner probability is Prob (B) = Prob (Ω \ B).

Appendix-Nonmeasurable mappings Definition For a mapping T : Ω R the outer integral is defined as E [T ] = inf { E [U] U T, U : Ω R measurable and E [U] exists } and the inner integral is E [T ] = E [ T ].

Appendix-Nonmeasurable mappings Lemma For any mapping T : Ω R there exists a measurable funkce T : Ω R with 1. T T ; 2. T U a.s. for every measurable U : Ω R with U T. The funkce T is called a minimal measurable majorant of T and is not uniquely defined. The funkce T = ( T ) is called a maximal measurable minorant of T.

Appendix-Nonmeasurable mappings Lemma For any T : Ω R 1. E [T ] = E [T ] if one of these integral exists; 2. E [T ] = E [T ] if one of these integral exists.

Definition Let X be a metric space with metric d and X n, X : Ω X, n N be arbitrary maps. X n, n N converges almost surely to X if ( ) Prob lim d (X n, X ) = 0 = 1. n + We use notation X n as n + X. X n, n N converges almost uniformly to X if for every ε > 0 there exists a measurable set A ε Ω with Prob (A ε ) 1 ε and d (X n, X ) 0 uniformly na A ε, i.e. sup ω Aε d (X n (ω), X (ω)) 0. We use notation X n au n + X.

Definition X n, n N converges outer almost surely to X if d (X n, X ) We use notation X n as n + 0 for some versions of d (X n, X ). as n + X. X n, n N converges in outer probability to X if for every ε > 0 Prob (d (X n, X ) > ε) 0. We use notation X n Prob n + X. Random approximate solutions

Appendix-Nonmeasurable mappings There are known relations between all of these notions of convergences. Lemma Let X be a metric space, X : Ω X be Borel measurable and X n : Ω X, n N be a sequence of arbitrary maps. Then, X n X n as n + X implies X n as n + X if and only if X n Prob n + X and X n au n + X. as n + X ;

Definition Let X be a metric space. The sequence X n : Ω X, n N of arbitrary maps is called asymptotically measurable if E [f (X n )] E [f (X n )] n + 0 for every bounded continuous funkce f : X R.

Definition Let X be a metric space. The sequence X n : Ω X, n N of arbitrary maps is called strongly asymptotically measurable if f (X n ) f (X n ) as n + 0 for every bounded continuous funkce f : X R.

Appendix-Nonmeasurable mappings Definition Let X be a metric space, X n : Ω X, n N be an asymptotically measurable sequence and X : Ω X be random variable. We say that X n, n N converges in distribution to X if E [f (X n )] E [f (X )] n + for every bounded continuous funkce f : X R. D We use notation X n X. n +

Appendix-Nonmeasurable mappings Lemma Let X be a metric space, X : Ω X be Borel measurable and separable, and X n : Ω X, n N be a sequence of arbitrary maps. Then, as X if and only if X n n + asymptotically measurable. X n as n + X and X n, n N is strongly

Appendix-Nonmeasurable mappings Lemma Let X be a metric space, X : Ω X be Borel measurable and separable, and X n : Ω X, n N be a sequence of arbitrary maps, which is asymptotically measurable. Then, X n as n + X implies X n Prob n + X.

Appendix-Nonmeasurable mappings Lemma Let X be a metric space, X n : Ω X, n N be an asymptotically measurable sequence and X : Ω X. Then If X n X n X n If X n D n + X then X is measurable, i.e. it is a random variable. as n + X implies X n Prob n + X implies X n D n + X. D n + X. D n + X and X c is nonrandom then X n Prob n + X.

Portmanteau lemma Lemma Let X be a metric space, X n : Ω X, n N be an asymptotically measurable sequence and X : Ω X be random variable. Then the following is equivalent: X n D n + X. E [f (X n )] n + E [f (X )] f C b (X ). E [f (X n )] n + E [f (X )] f C b (X ).

Portmanteau lemma lim inf Prob (X n G) Prob (X G) G G (X ). n + lim sup Prob (X n F ) Prob (X F ) F F (X ). n + Random approximate solutions

References Hoffmann-Jørgensen, J.: Probability with a View Towards to Statistics I, II. Chapman and Hall, New York, 1994. Kelley, J.L.: General Topology. D. van Nostrand Comp., New York, 1955. Lachout, P.: Stability of stochastic optimization problem - nonmeasurable case. Kybernetika 44,2(2008), 259-276. Lachout, P.: Stochastic optimization sensitivity without measurability. In: Proceedings of MMEI held in Herĺany (2007), 179-184. Lachout, P.; Liebscher, E.; Vogel, S.: Strong convergence of estimators as ε n -minimizers of optimization problems. Ann. Inst. Statist. Math. 57,2(2005), 291-313.

References Lachout, P.; Vogel, S.: On Continuous Convergence and Epi-convergence of Random Functions. Part I: Theory and Relations. Kybernetika 39,1(2003), 75-98. Robinson, S.M.: Analysis of sample-path optimization. Math. Oper. Res. 21,3(1996), 513-528. Rockafellar, T.; Wets, R.J.-B.: Variational Analysis. Springer-Verlag, Berlin, 1998. Vajda, I.; Janžura, M.: On asymptotically optimal estimates for general observations. Stochastic Processes and their Applications 72,1(1997), 27-45. van der Vaart, A.W.; Wellner, J.A.: Weak Convergence and Empirical Processes Springer, New York, 1996.