mathematical · 2019. 2. 12. · [ 1), and kolmorogov and fomin [ 1). the martingale and...

19

Upload: others

Post on 23-Feb-2021

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken
Page 2: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

T R A N S L A T I O N S O F

MATHEMATICAL M O N O G R A P H S V O L U M E 4 7

M. B. Nevelson R. Z. Has'minskff

Stochastic Approximation and Recursive Estimation

fl| American Mathematical Society Providence, Rhode Island

10.1090/mmono/047

Page 3: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

CTOXACTMMECKAH AnnPOKCMMAUMf l

M PEKYPPEKTHOE OUEHMBAHM E

M. B. HEBEJlbCOH , P . 3 . XACbMMHCKM W

M3AaTejibCTBo "MayKa " MocKB a 197 2

Trans lated fro m th e R u s s i a n b y

I srae l Progra m fo r S c i e n t i f i c T r a n s l a t i o n s

Trans la t ion e d i t e d b y B . Si lve r

2000 Mathematics Subject Classification. Primar y 62L20 ; Secondary 60J05 , 93E10 .

Abstract. Thi s boo k i s devote d t o sequentia l method s o f solvin g a

c lass o f problem s t o whic h belongs , fo r example , th e proble m o f findin g a

maximum poin t o f a functio n i f eac h measure d valu e o f thi s functio n con -

tains a rando m error . Som e basi c procedure s o f stochasti c approximatio n

are investigate d fro m a singl e poin t o f view , namel y th e theor y o f Marko v

processes an d martingales . Example s ar e considere d o f application s o f

the theorem s t o som e problem s o f estimatio n theory , educationa l theor y an d

control theory , an d als o t o som e problem s o f informatio n transmissio n i n

the presenc e o f invers e feedback .

Library of Congress Cataloging in Publication Data w ^m ay

Nevel'son, Mikhai l Borisovich . Stochast ic approximatio n an d recursiv e estimation .

(Translations o f mathematica l monograph s ; v . ^7 ) Translat ion o f Stokhasticheskai a approksimatfe£ 3

i rekurrentnoe oftenivanie. Bibliography: p. Includes index . 1. Stochast i c approximation . 2 . Estimatio n

theory. I . Kha s 'minskii, Rafai l Zalmanovich , jo in t author. H . T i t l e . I I I . Ser ies . QA27lf.NWl5 519.2 76-U8298 ISBN 0-8218-1597-0

© 197 6 b y th e America n Mathematica l Society . Al l right s reserved . Reprinted b y th e America n Mathematica l Society , 2000 .

The America n Mathematica l Societ y retain s al l right s except thos e grante d t o th e Unite d State s Government .

Printed i n th e Unite d State s o f America .

@ Th e pape r use d i n thi s boo k i s acid-fre e an d fall s withi n th e guideline s established t o ensur e permanenc e an d durability .

Visit th e AM S hom e pag e a t ht tp: / /www.ams.org /

10 9 8 7 6 5 4 3 0 4 0 3 0 2 0 1 0 0

Page 4: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

TABLE OF CONTENTS Preface 1 Introduction 4 Chapter 1 . Element s of probability and martingales 1 1

1. Probabilit y 1 1 2. Rando m variables. 1 3 3. Conditiona l probabilities and conditional expectations 1 7 4. Independence . Produc t measures 2 0 5. Martingale s and supermartingales 2 2

Chapter 2. Discret e time Markov processes 2 9 1. Marko v processes 2 9 2. Marko v processes and supermartingales 3 2 3. Recursivel y define d processe s 3 4 4. Discret e model of diffusion . 3 7 5. Exi t of sample functions fro m a domain 3 8 6. Serie s of independent rando m variables 4 1 7. Convergenc e of sample functions 4 3 8. Teachin g pattern recognition 4 9

Chapter 3. Marko v processes and stochastic equations 5 3 1. Continuou s time Markov processes 5 3 2. Stochasti c differential equations . 1 5 7 3. Stochasti c integrals 5 9 4. Stochasti c differential equations . II 6 2 5. Ito' s formula 6 4 6. Supermartingale s 6 9 7. Existenc e of solutions in the large 7 0 8. Exi t from a domain. Convergenc e of sample functions 7 3

Chapter 4. Convergenc e of stochastic approximation procedures . 1 7 9 1. Th e Robbins-Monro procedure 7 9 2. Th e Kiefer-Wolfowitz procedur e 8 3 3. Continuou s procedures 8 6 4. Convergenc e of the Robbins-Monro procedure 8 8 5. Convergenc e of the Kiefer-Wolfowitz procedur e 9 4

ui

Page 5: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

CONTENTS

Chapter 5. Convergenc e o f stochasti c approximatio n procedures . I I 10 1 1. Introductor y remark s 10 1 2. Genera l theorems 10 2 3. Auxiliar y result s (continuous time ) 10 6 4. Auxiliar y result s (discrete time ) 11 2 5. One-dimensiona l procedure s 11 8

Chapter 6. Asymptoti c normalit y o f the Robbins-Monr o procedure 12 3 1. Preliminar y remark s 12 3 2. Asymptoti c behavior of solution s 12 9 3. Investigatio n o f the process ^(r) . 13 2 4. Investigatio n o f th e process I2{f) 13 6 5. Asymptoti c normalit y (continuou s time) 14 0 6. Asymptoti c normalit y (discrete time ) 14 7 7. Convergenc e o f moments 15 4

Chapter 7 . Som e modifications o f stochasti c approximatio n procedure s 16 1 1. Statemen t o f th e proble m 16 1 2. Genera l theore m 16 2 3. Auxiliar y result s 16 5 4. Theorem s on convergence and asymptotic normality 16 7 5. Adaptiv e Robbins-Monr o procedure s 16 9 6. Asymptoti c optimalit y 17 5

Chapter 8 . Recursiv e estimatio n (discrete time ) 18 1 1. Th e Crame'r-Rao inequality. Efficienc y o f estimate s 18 1 2. Th e Cramer-Rao inequality i n the multidimensiona l cas e 18 6 3. Estimatio n of a one-dimensional paramete r 18 9 4. Asymptoticall y efficien t recursiv e procedure 19 4 5. Estimatio n of a multidimensional parameter 19 7 6. Estimatio n with dependent observation s 20 2

Chapter 9. Recursiv e estimation (continuous time ) 20 7 1. Th e Cramer-Ra o inequalit y 20 7 2. Applicatio n o f th e Robbins-Monro procedur e 21 0 3. Time-dependen t observation s 21 2 4. Som e applications 21 3 5. A modification 22 0

Chapter 10 . Recursiv e estimation with a control parameter 22 1 1. Statemen t o f th e problem 22 1 2. Asymptoticall y optima l recursiv e plan 22 3 3. Tw o examples... . 22 6 4. Continuou s case 22 9

Notes on the literatur e 23 3 Bibliography 23 7 Subject Index 24 2 Main Notation 24 4

IV

Page 6: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

This page intentionally left blank

Page 7: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

NOTES ON THE LITERATURE

Chapter 1

The proo f o f al l theorem s cite d i n § § 1 - 4 ma y b e foun d i n Kolmogoro v [ 1 ] , Halmos

[ 1), an d Kolmorogo v an d Fomi n [ 1). Th e martingal e an d supermartingale concept s ar e du e

to Doob . Theorem s S. l an d 5.l ' an d thei r proof s ar e take n fro m Doo b [ 1 ] .

Chapter 2

§ 1. Fo r mor e detail s abou t propertie s o f Marko v processes an d thei r transitio n func -

tions, see Felle r [ 1 ], Dynkin [ 1 ], Loev e [ 1J, an d others .

§2. Theore m 2. 1 i s essentiall y a special cas e o f Doob' s theore m (Doo b [1 ) ) o n th e

stopping o f a supermartingale. Kolomogorov' s inequalit y fo r supermartingale s ma y als o b e

found i n Doo b [ 1 ].

§3. Theorem s 3. 1 an d 3. 2 ar e wel l known . See , fo r example , Gihman an d Skoroho d

HI. §5. Fo r related results , see, fo r example , AB R [ 1 ] , Chapter IV .

§6. Theorem s 6. 1 an d 6. 2 wer e prove d b y Kolomogoro v [1 ] i n a more genera l setting .

§7. Theore m 7. 1 i s conceptuall y simila r t o Theore m 8 i n Chapte r 4 o f AB R [ 1 ] .

§ 8. Fo r a detailed discussio n o f th e question s relate d t o thi s formulatio n o f th e prob -

lem, see AB R [ 1 ], Chapter 5 .

Chapter 3

§ 1. Fo r mor e detail s abou t th e definitio n an d propertie s o f a continuous Marko v

process, see Dynki n [ 1 ].

§2. Th e limi t passag e fro m equatio n (2.1 ) t o a continuou s Marko v process was con -

sidered b y Bernstei n [ 1 ], who, among othe r things , prove d a theorem o n th e limi t behavio r

of th e transitio n probabilit y a s h -* 0 .

§§3 an d 4 . Thi s definitio n o f th e stochasti c integra l i s du e t o ItS . Fo r a proo f o f

Theorem 4.1 , see, fo r example , Gihma n an d Skorohod [ 2 ] .

§7. A proo f o f Theore m 7. 1 ma y b e foun d i n Has'minski T [ 1 ] .

§8. Simila r results ar e presente d in th e authors ' paper [ 1 ] .

Chapter 4

§ 1. A s mentione d i n th e text , s.a . procedure s fo r locatio n o f th e root s o f regressio n

equations wer e first suggeste d i n th e 19S 1 pape r o f Robbin s an d Monr o [ 1 ] , who prove d

convergence i n mea n squar e o f th e procedur e (S.l ) unde r certai n assumptions . Thi s resul t

233

Page 8: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

234 NOTES O N TH E LITERATUR E

was generalize d b y Wolfowit z [ 1 ], Blu m [ 1J, Chun g [ 1 ], an d others . Blu m [ 1 ), i n particular ,

presented th e firs t proo f o f convergenc e wit h probabilit y 1 . Theore m 1. 1 an d th e metho d o f

proof use d i n th e tex t ar e du e t o Gladyse v [ 1 ]. Multidimensiona l R M procedure s wer e intro -

duced b y Blu m [ 2 ] . Blu m was th e firs t autho r t o appl y th e theor y o f supermartingale s t o

investigate th e convergenc e o f s.a . procedures .

§2. Th e procedure (2.5 ) wa s propose d b y Kiefe r an d Wolfowit z [1 ] i n 1952 . Blu m

[ 2 ] , Derma n [ 1 ] , Burkholde r [1 ] an d other s studie d th e procedur e i n detai l and , i n particular ,

proved tha t i t i s convergen t wit h probabilit y 1 . N o accoun t i s given her e o f mor e genera l s.a .

procedures suc h a s those propose d i n Burkholde r [ 1 ] , Dvoretzk y [ 1 ] , etc . Mor e detaile d

bibliographies ma y b e foun d i n th e revie w article s o f Dvoretzk y [ 1 ] , Fabia n [ 3 ] , Sakriso n

[ 3 ], Schmettere r [ 1J, Logino v [ 1 ], an d others .

§ 3. Th e continuou s analo g o f th e R M procedur e wa s introduced b y Drim l an d Nedom a

[ 1 ] . The y prove d th e convergenc e o f th e procedur e unde r fairl y genera l assumption s o n th e

stochastic processe s appearin g on th e righ t o f th e equation . Unde r certain assumptions ,

Sakrison [ 1 ] prove d th e convergenc e o f th e continuou s analog of th e Kiefer-Wolfowit z pro -

cedure. Th e continuou s procedure s considere d her e wer e define d i n Has'rninsk u [1 ] an d

Nevel'son an d Has'minsk u [ 1 ].

§4. Th e firs t convergenc e theorem s fo r multidimensiona l R M procedures wer e prove d

by Blu m [ 2 ] ; see als o ABR [ 1 ] , Schmetterer [1 ] an d furthe r bibliograph y cite d there . Theo -

rem 4. 4 i s similar t o th e above-mentione d resul t o f Gladyse v [ 1 ]; Theore m 4. 5 i s du e t o

Braverman an d Rozonoe r (se e ABR [ 1 ]). Th e formulatio n o f th e las t theore m raise s th e

question a s to whethe r on e ca n als o weaken th e conditio n Lfl 2(f) < « > in th e proo f tha t

X(t) - + XQ &J . I t turn s ou t tha t thi s i s indee d possible . Fo r example , i n som e case s i t i s

sufficient t o deman d tha t £o"(f ) < °° for som e n > 0 (althoug h on e mus t the n impos e mor e

stringent condition s o n th e "noise") . Th e theorem s prove d her e fo r continuou s R M proced -

ures ar e du e t o th e author s [ 1 ] .

§ 5. Convergenc e theorem s fo r discret e multidimensiona l K W procedures wer e prove d

by Blu m [ 2 ] , Dupa£ [1 ] an d others. Fo r th e continuou s case , see Nevel'so n an d Has'minsk u

[ 1 ] .

Chapter 5

The conjectur e tha t a K W procedure canno t converg e wit h positiv e probabilit y t o

minimum point s o f th e regressio n functio n wa s advance d b y Fabia n [ 1 J, [ 2 ] . Th e result s

presented her e ar e based o n Has'minsk u (1 ] an d Nevel'son [ 1 ] , [ 2 ] . Th e tw o last-mentione d

papers als o prov e mor e genera l theorems. Se e als o Krasulin a [ 1 ] .

Chapter 6

§1 . Theore m 1. 1 fo r th e one-dimensiona l cas e i s proved , e.g. , i n Loev e [ 1 ]; for th e

multidimensional case , see Sack s [ 1 ].

§2. Condition s fo r th e validit y o f th e estimat e EX 2{t) = 0 ( l / f ) a s t - * » , wher e X(f)

is a discret e one-dimensiona l R M process , were firs t obtaine d b y Chun g [ 1 ]. A n analo g o f

Lemma 2. 1 fo r discret e multidimensiona l R M processes was prove d b y Sack s [ 1 ] unde r th e

additional assumptio n E\G(t, x, t j ) | 2 < c < « . Th e fac t tha t (2.9 ) i s a consequence o f (2.8 )

follows fro m a well-known lemm a of Chun g [ 1 ] .

§§3 an d 4 . Lemm a 3. 1 wa s prove d b y Has'minski f [ 2 ] . Lemm a 4. 3 i s du e t o Sack s

Page 9: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

NOTES O N TH E LITERATUR E 235

§5. Theore m 5. 3 fo r M > 0 i s prove d i n Has'minski T [ 2 ] . Th e asymptoti c normalit y

of a R M procedure wa s established , unde r certai n assumptions , b y Holev o [ 1 ].

§6. Theore m 6. 1 wa s prove d b y Sack s under th e followin g additiona l assumptions :

a) th e matri x B ma y b e reduced t o diagona l for m b y a similarity transformatio n involvin g

an orthogona l matrix ; b ) th e nor m o f th e matri x A(t, x) i s bounde d fo r t > 1 an d x € Ej.

Theorem 6. 3 i s new. Asymptoti c normalit y theorem s fo r othe r s.a . procedure s wer e prove d

by Fabia n [ 5 ] , Derman [ 1 ] , Burkhoide r [ 1 ] , DupaS [1 ] an d others . Th e "truncation "

method i n asymptoti c normalit y proof s fo r s.a . procedure s was firs t applie d b y Hodge s an d

Lehman [ 1 ] .

§ 7. Th e firs t convergenc e theorem s fo r moment s o f s.a . processe s wer e prove d b y

Chung [ 1 ] i n th e one-dimensiona l case . Fo r the multidimensiona l case , the resul t o f Theore m

7.2 i s wel l know n (see , fo r example , Schmettere r [ 1J) fo r X > *ta . Bu t i f Conditio n (a ) o f

§ 7 hold s with A < Via, the onl y resul t know n t o u s o n convergenc e o f moment s i s tha t o f

Sakrison [ 2 ] . Constructin g a certai n modificatio n o f th e R M procedure , Sakrison prove d

that th e matri x o f secon d moment s o f th e proces s *Ji(X{i) — XQ) converges t o th e correspon -

ding covarianc e matri x o f a normal law , provided al l moment s E|G(f , x, (J))P t p = 1 , 2 , . . . ,

are bounded .

Chapter 7

§ § 1 - 4 . Alber t an d Gardne r [1 ] establishe d convergenc e an d asymptoti c normalit y

for th e procedur e (1.1 ) an d a more genera l on e i n whic h th e facto r a{t) i s allowe d t o depen d

on pas t observations . Thes e author s als o considere d a multidimensional modificatio n o f (1.1) .

Recently, Cslb i [ 1 ] , [2 ] ha s prove d convergenc e o f truncate d procedure s fo r a broad rang e

of truncations .

§5. Ther e i s a considerable literatur e o n th e optima l choic e o f parameter s fo r R M

procedures an d thei r adaptiv e modifications . Amon g others , we mentio n Chun g [ 1 ],

Dvoretzky [ 1 ] , Keste n [ 1 ] , Cypkin [ 1 ] , [2],an d Stratonovic ' [ 2 ] . Ou r approac h i s base d

on th e wor k o f Vente r [ 1 ] , who prove d Theore m 5. 1 unde r slightly mor e restrictiv e assump -

tions.

§6. Theore m 6. 1 slightl y generalize s a result o f Vente r [ 1 ] ; Theorem 6. 2 i s new .

Chapter 8

§1. Theore m 1. 1 follow s fro m result s o f Chapma n an d Robbin s [1 ] an d Kaga n [ 1 ] .

Theorem 1. 2 wa s prove d b y Ibragimo v and Has'minski i [ 2 ] . Thes e paper s als o deriv e inequal -

ities fo r biase d estimates .

§2. Theore m 2. 1 an d it s analo g fo r biase d estimate s wer e prove d b y Ibragimo v an d

Has'minskii [ 2 ] .

§ 3. Th e us e o f R M procedure s fo r parametri c estimatio n i s discusse d i n Alber t an d

Gardner [ 1 ], wher e a larger clas s o f estimates , no t necessaril y formin g Marko v processes , i s

considered.

§§4 an d 5 . Theorem s 4. 1 an d 4. 5 wer e prove d b y Sakriso n [ 2 ] , [3 ] unde r stronge r

conditions (an d without th e convergenc e o f distributions) . Othe r optimalit y criteri a wer e

considered b y Alber t an d Gardne r [ 1 ] , Cypkin [ 1 ] , and Stratonovic ' [ 2 ] , but thei r procedure s

generally depen d o n pas t observation s (see , fo r example , th e procedur e (0.9 ) i n th e Intro -

duction).

Page 10: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

236 NOTES ON THE LITERATURE

Chapter 9

§1. Fo r inequality (1.7) , see, for example, Van Trees [1]. §2. A procedure equivalent to (2.4) for m linear in the parameter (m(*0) = mxg ,

where m is a nonsingular matrix) was considered by Holevo [1] . H e also considered the case that m is not a square matrix but mm* is nonsingular. Fo r the nonlinear case, Holevo proposed a procedure more complicated than (2.4). Unde r certain assumptions, he established asymptotic normality o f the estimation procedure.

§3. Recursiv e procedures similar to (3.1) ma y be derived from the Bayesian approach, using linearization. Fo r linear systems, similar estimation procedures may be found in Kal-man and Bucy [1] , Lipcer and Sirjaev [1] , and elsewhere.

§4. Th e problem of estimating x0 base d on observations of type (4.7) (estimation o f linear regression coefficients) ha s been considered from other points of view in many papers.

§5. Wit h suitable initial conditions, the procedure (5.3) , (5.4) for a linear function m{t, x) = m{t)x coincide s with the Kalman-Buc y optimal linear filtering equations . I n other words, this system is satisfied, in particular, by the conditional expectation of XQ given Y(s), s < f , provide d the a priori distribution of XQ is Gaussian.

Page 11: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

BIBLIOGRAPHY

M. A . AIZERMAN , fe. M. BRAVERMA N AN D L L ROZONOE R (ABR ) 1. The method of potential functions in machine learning theory, "Nauka" , Moscow ,

1970. (Russian ) ARTHUR E . ALBER T AN D LELAN D A . GARNDER , JR .

1. Stochastic approximation and nonlinear regression, MI T Press , Cambridge, Mass., 1967. M R 3 6 #7261 .

SERGE BERNSTEI N 1. Principes de la theorie des equations differentiates stochastiques, Trud y Fiz.-Mat .

Inst. Steklov . Otd . Mat . S (1934) , 95-124 . DAVID BLACKWEL L AN D M . A . GIRSHIC K

1. Theory of games and statistical decisions, Wiley , Ne w York ; Chapman an d Hall , London, 1954 . M R 16 , 1135 .

JULIUS R . BLU M 1. Approximation methods which converge with probability one, Ann . Math. Statist .

25 (1954) , 382-386 . M R 15 , 973 . 2. Multidimensional stochastic approximation methods, Ann . Math . Statist . 25 (1954) ,

737-744 . M R 16 , 382 . D. L. BURKHOLDER

1. On a class of stochastic approximation processes, Ann . Math . Statist . 27 (1956) , 1044-1059. M R 1 9 , 7 1 .

DOUGLAS G . CHAPMA N AN D HERBERT ROBBIN S 1. Minimum variance estimation without regularity assumptions, Ann . Math. Statist .

22 (1951) , 581-586 . M R 13 , 367 . K. L CHUN G

1. On a stochastic approximation method, Ann . Math. Statist . 2 5 (1954) , 463—483 . MR 16 , 272 .

HARALD CRAME R 1. Mathematical methods of statistics, Princeto n Univ . Press , Princeton , N . J. , 1946 .

MR 8 , 39 . S. CSIBI

1. Simple and compound processes in iterative machine learning, Lecture s delivere d a t the Internationa l Summe r School , Trieste , 1971 .

2. Learning optimal decision functions recursively, Secon d Internat . Sympos . Infor-mation Theor y (Tsahkadsor , 1971) , Abstract s o f papers , Izdat . Akad . Nauk SSSR , Moscow, 1971 , pp . 2 9 2 - 2 9 4. CC A 7 #12990 .

JA. Z. CYPKI N 1. Adaptation and learning in automatic systems, "Nauka" , Moscow , 1968 ; English

transl., Academic Press , New York , 1971 . 2. Foundations of the theory of learning systems, "Nauka" , Moscow , 1970 ; English

transl., Academi c Press , New York , 1973 . CYRUS DERMA N

1. An application of Chung's lemma to the Kiefer-Wolfowitz stochastic approximation procedure, Ann . Math. Statist . 2 7 (1956) , 532-536 . M R 17 , 1218 .

A. G. D'JACKOV AND M. S. PlNSKER 1. 77ie optimal linear method of transmission over a memoryless Gaussian stationary

channel with complete feedback, Problem y Peredac i Informaci i 7 (1971) , vyp. 2 , 3 8 - 4 6 = Problem s of Information Transmissio n 7 (1971), 123-129 . M R 46 #8741 .

237

Page 12: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

238 BIBLIOGRAPHY

J. L. DOOB 1. Stochastic processes, Wiley , New York ; Chapma n an d Hall , London, 1953 . M R

15 ,445 . MlLOSLAV DRIM L AN D JlR I NEDOM A

1. Stochastic approximations for continuous random functions, Trans . Second Pragu e Conf. Information Theory , Publ . House Czechoslova k Acad . Sci. , Prague , 1960 ; Academic Press , New York , 1961 , pp . 145-158 . M R 2 3 #A4171 .

VACLAV DUPA C 1. On the Kiefer-Wolfowitz approximation method, Casopi s Pest . Mat. 82 (1957) , 4 7 -

75. (Czech ; Russian an d Englis h summaries ) M R 19 , 693 . 2. A dynamic stochastic approximation method, Ann . Math. Statist . 3 6 (1965) , 1 6 9 5 -

1702. M R 3 3 #1939 . ARYEH DVORETZK Y

1. On stochastic approximation, Proc . Thir d Berkele y Sympos . Math . Statist.and Prob -ability (1954-55) , vol . I , Univ . o f Californi a Press , Berkeley , Cal. , 1956 , pp . 3 9 -55. M R 18 , 946 .

E B . DYNKIN 1. Markov processes, Fizmatgiz , Moscow , 1963 ; English transl. , Springer-Verlag, Berlin ,

Academic Press , New York , 1965 . M R 3 3 #1886 , 1887 . VACLAV FABIA N

1. Stochastic approximation methods, Czechoslova k Math . J . 1 0 (85 ) (1960) , 123 — 159. M R 22 # 4 0 9 1 .

2. A new one-dimensional stochastic approximation method for finding a local mini-mum of a function, Trans . Third Pragu e Conf . Informatio n Theory , Statistica l Decision Functions , Rando m Processe s (Liblice , 1962) , Publ . House Czech . Acad . Sci., Prague , 1964 , pp . 85-105 . M R 2 9 #6569 .

3. A survey of deterministic and stochastic approximation methods for the minimi-zation of functions, Kybernetik a (Prague ) 1 (1965) , 499-523 . (Czech ; English summary) M R 3 2 #5378 .

4. On the choice of design in stochastic approximation methods, Ann . Math . Statist . 39 (1968) , 457-465 . M R 3 7 # 1 0 3 8 .

5. On asymptotic normality in stochastic approximation, Ann . Math. Statist 3 9 (1968) , 1327-1332. M R 3 7 #6984 .

A. A . FEL'DBAU M 1. Optimal control systems, 2n d rev . ed. , "Nauka" , Moscow , 1966 ; English transl . o f

1st ed. , Academi c Press , New York , 1965 . M R 3 5 #2656 ; 3 2 #9075 . WILLIAM FELLE R

1. An introduction to probability theory and its applications. Vol . II , Wiley, Ne w York, 1966 . M R 3 5 #1048 .

G. M . FIHTENGOL' C 1. A course of differential and integral calculus. Vol . II , 4th ed. , Fizmatgiz , Moscow ,

1959; Germa n transl. , VE B Deutsche r Verla g de r Wissenschaften , Berlin , 1966 . MR 3 9 # l b .

1.1. GlHMAN AND A. V. SKOROHOD 1. Introduction to the theory of random processes, "Nauka" , Moscow, 1965 ; English

transl., Saunders , Philadelphia , Pa. , 1969 . M R 3 3 #6689 ; 4 0 # 9 2 3 . 2. Stochastic differential equations, "Naukov a Dumka" , Kiev , 1968 ; English transl. ,

Springer-Verlag, Ne w York , 1972 . M R 4 1 #7777 ; 4 9 #11625 . E G . GLADYSE V

1. On stochastic approximation, Teor . Verojatnost . i Primenen . 1 0 (1965) , 297 -30 0 = Theor . Probabilit y Appl . 1 0 (1965) , 275-278 . M R 3 2 #3184 .

PAUL R . HALMO S 1. Measure theory, Va n Nostrand , Princeton , N . J. , 1950 . M R 11 , 504 .

Page 13: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

BIBLIOGRAPHY 239

R.Z.HAS'MINSIUI 1. Stability of systems of differential equations under random perturbations of their

parameters, "Nauka" , Moscow , 1969 . (Russian ) M R 41 #3925 . 2. The behavior of the processes of stochastic approximation for large time-variables,

Problemy Peredac i Informaci i 8 (1972) , vyp. 1 , 81-91 = Problem s o f Informatio n Transmission 8 (1972) , 6 2 - 7 0 . M R 4 8 # 1 3 2 1 .

J. L . HODGES , JR., AND E . L . LEHMAN N 1. Two approximations to the Robbins-Monro process, Proc . Third Berkele y Sympos .

Math. Statist , an d Probabilit y (1954-1955) , vol . I , Univ . of Californi a Press , Berke -ley, Ca!. , 1956 , pp . 95 -104 . M R 18 , 947 .

A. S. HOLEVO 1. Estimates for the drift parameters for a diffusion process by the method of sto-

chastic approximation, Studie s i n th e Theor y o f Self-adjustin g System s (V. G . Sragovic, editor) , Vycisl . Centr Akad . Nau k SSSR , Moscow , 1967 , pp . 179-200 . (Russian) M R 3 8 #1800 .

I. A . IBRAGIMO V AN D R . Z . HAS'MINSK H 1. The asymptotic behavior of certain statistical estimates in the smooth case. I : In-

vestigation of the likelihood ratio, Teor . Verojatnost . i Primenen . 1 7 (1972) , 4 6 9 -486 = Theor . Probabilit y Appl . 1 7 (1972) , 445-462 . M R 4 6 #10106 .

2. Information inequalities and superefficient methods, Dokl . Akad . Nau k SSSR 20 4 (1972), 1300-130 2 = Sovie t Math . Dokl . 1 3 (1972) , 821-824 . M R 4 7 # 2 7 3 3 .

3. On the approximation of statistical estimators by sums of independent variables, Dokl. Akad . Nau k SSS R 21 0 (1973) , 1273-127 6 = Sovie t Math . Dokl . 1 4 (1973) , 883-887 . M R 4 9 #10020 .

KlYOSI IT O 1. On a formula concerning stochastic differentials, Nagoy a Math . J . 3 (1951) , 55 -65 .

MR 13 , 363 . KlYOSI IT O AN D MAKIK O NlSI O

1. On stationary solutions of a stochastic differential equation, J . Math . Kyot o Univ . 4 (1964) , 1-75 . M R 3 1 # 1719 .

A.M.KAGAN 1. On the theory of Fisher's information quantity, Dokl . Akad. Nau k SSS R 15 1

(1963), 277-27 8 = Sovie t Math . Dokl . 4 (1963) , 991-993 . M R 2 7 #2032 . R. E . KALMA N AN D R . S . BUC Y

1. New results in linear filtering and prediction theory, Trans . ASME Ser . D . J . Basi c Engrg. 83 (1961) , 9 5 - 1 0 8 . M R 3 8 #3076 .

HARRY KESTE N 1. Accelerated stochastic approximation, Ann . Math . Statist . 29 (1958) , 4 1 - 5 9 . M R

20 # 3 7 1 . J. KlEFE R AN D J . WOLFOWIT Z

1. Stochastic estimation of the maximum of a regression function, Ann . Math . Statist . 23 (1952) , 462-466 . M R 14 , 299 ; erratum, M R 14 , 1278 .

A. N . KOLMOGORO V 1. Grundbegriffe der Wahrscheinlichkeitsrechnung, Springer-Verlag , Berlin, 1933 ; Eng-

lish transl. , Chelsea, Ne w York , 1950 ; 2nd ed. , 1956 . M R 11 , 374 ; 18 , 155 . A. N . KOLMOGOROV AN D S . V. FOMI N

1. Elements of the theory of functions and of functional analysis, 2n d rev . ed. , "Nauka", Moscow , 1968 ; English transl . o f 1s t ed. , vols . 1 , 2 , Grayloc k Press , Al -bany, N . Y. , 1957 , 1961 . M R 3 8 #2559 ; 19 , 44; 22 #9565a .

N. N . KRASOVSK U 1. Certain problems in the theory of stability of motion, Fizznatgiz , Moscow , 1959 ;

English transl. , Stability of motion. Application of Lyapunov fs second method to differential systems and equations with delay, Stanfor d Univ . Press , Stanford, Cal. , 1963. M R 2 1 #5047 ; 2 6 # 5 2 5 8 .

Page 14: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

240 BIBLIOGRAPHY

T. P . KRASULIN A 1. The Robbins-Monro process in the case of several roots, Teor . Verojatnost . i

Primenen. 1 2 (1967) , 386-39 0 = Theor . Probabilit y Appl . 1 2 (1967) , 333-337 . MR 3 5 # 5 0 8 5 .

L M . LiBKIND 1. On the probability of large deviations for a quadratic functional of a stationary

Gaussian sequence, Problem y Peredac i Informaci i 5 (1969) , vyp . 2 , 7 2 - 7 8 = Prob -lems o f Informatio n Transmissio n 5 (1969) , no. 2 , 56—60 .

R. S. LlPCER AND A. N. SlRJAEV 1. Nonlinear filtering of Markov diffusion processes, Trud y Mat . Inst. Steklov . 10 4

(1968), 135-18 0 = Proc . Steklov Inst . Math . 10 4 (1968) , 163-21 8 (1971) . M R 41 #6342 .

2. The absolute continuity with respect to Wiener measure of measures that corre-spond to processes of diffusion type, Izv . Akad . Nau k SSS R 3 6 (1972) , 847-88 9 = Math . USS R Izv . 6 (1972) , 839-882 . M R 4 7 # 1 1 1 9 .

MICHEL LOEV E 1. Probability theory, 2n d ed. , Va n Nostrand , Princeton , N . J. , 1960 . M R 2 3 #A670 .

N . V . LOGINO V 1. Methods of stochastic approximation, Avtomat . i Telemeh . 2 7 (1966) , no . 4 , 1 8 5 -

204 = Automat . Remot e Contro l 2 7 (1966) , 706 -728 . I. G. MALKIN

1. Theory of stability of motion, 2n d rev . ed., "Nauka" , Moscow , 1966 ; Englis h transl. o f 1s t ed. , AE C Transl . #3352 , U . S . Atomic Energ y Commission , Oa k Ridge, Tenn. , 1958 . M R 3 4 #6228 ; erratum, 43 , p . 1695 .

M. B . NEVEL'SO N 1. Certain properties of continuous procedures of stochastic approximation, Teor .

Verojatnost. i Primenen . 1 7 (1972) , 310-31 9 = Theor . Probabilit y Appl . 1 7 (1972) , 298-308 . M R 4 5 #6124 .

2. The convergence of continuous and discrete Robbins-Monro procedures in the case of several roots of the regression equation, Problem y Peredac i Informaci i 8 (1972) , vyp. 3 , 48—57 = Problem s o f Informatio n Transmissio n 8 (1972) , 215—223. M R 48 #1320 .

M B . NEVEL'SO N AN D R . Z . HAS'MINSK H 1. Continuous procedures of stochastic approximation, Problem y Peredac i Informaci i

7 (1971) , vyp. 2 , 58-6 9 = Problem s o f Informatio n Transmissio n 7 (1971) , 1 3 9 -148. M R 4 8 #1319 .

2. On the convergence of the moments of the Robbins-Monro procedure, Avtomat . i Telemeh. 1973 , no. 1 , 96 -10 0 = Automat . Remot e Contro l 3 4 (1973) , 8 5 - 9 0 .

M. S. PlNSKER 1. The probability of error in block transmission in a memoryless Gaussian channel

with feedback, Problem y Peredac i Informaci i 4 (1968) , vyp. 4 , 3—1 9 = Problem s of Informatio n Transmissio n 4 (1968) , no. 4 , 1—14 .

C. RADHAKRISHN A RA O 1. Linear statistical inference and its applications, Wiley , New York , 1965 . M R

36 #4668 . HERBERT ROBBIN S AN D SUTTO N MONR O

1. A stochastic approximation method, Ann . Math . Statist . 2 2 (1951) , 400—407 . MR 13 , 144 .

JEROME SACK S 1. Asymptotic distribution of stochastic approximation procedures, Ann . Math. Statist .

29 (1958) , 373-405 . M R 2 0 #4886 . DAVID J . SAKRISO N

1. A continuous Kiefer-Wolfowitz procedure for random processes, Ann . Math. Statist . 35 (1964) , 590-599 . M R 2 8 #4650 .

Page 15: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

BIBLIOGRAPHY 241

2. Efficient recursive estimation; application to estimating the parameters of a co-variance function, Internat . J . Engrg . Sci . 3 (1965) , 461-483. M R 3 1 #6306 .

3. Stochastic approximation: A recursive method for solving regression problems, Advances i n Communication Systems : Theor y an d Application s (A . V . Balakrishnan , editor), Vol . 2 , Academic Press , New York , 1966 , pp . 51-106 .

J. PlETE R M SCHALKWIJ K 1. Recent developments in feedback communication, Proc . IEE E 5 7 (1969) , 1242 —

1249. J. PlETE R M . SCHALKWIJ K AN D T . KAILAT H

1. A coding scheme for additive noise channels with feedback. I : No bandwidth con-straint, IEE E Trans . Informatio n Theor y IT-1 2 (1966) , 172-182 . M R 4 0 # 3 9 9 1 .

L SCHMETTERE R 1. Multidimensional stochastic approximation, Multivariat e Analysis , I I (Proc . Secon d

Internat. Sympos. , Dayton , Ohio , 1968 ) (P . R . Krishnaian , editor) , Academi c Press , New York , 1969 , pp . 443-460 . M R 4 1 #2844 .

C. E . SHANNO N 1. A mathematical theory of communication, Bel l Syste m Tech . J . 27 , (1948) , 3 9 7 -

423, 623-656 . M R 10 , 133 . A. V . SKOROHO D

1. Studies in the theory of random processes, Izdat . Kiev . Univ . Kiev , 1961 ; English transl., Addison-Wesley , Reading , Mass. , 1965 . M R 3 2 #3082a,b .

R. L STRATONOVI C 1. Conditional Markov processes and their application to the theory of optimal control,

Izdat. Moskov . Univ. , Moscow, 1966 ; English transl. , Amer . Elsevier , Ne w York , 1968. M R 3 3 # 5 3 9 1 ; 3 6 #4912 .

2. Does there exist a theory of synthesis of optimal adaptive, self-learning, and self-adaptive systems'*. Automat , i Telemeh . 1968 , no . 1 , 96-107 = Automat . Remot e Control 2 9 (1968) , 83 -92 . M R 3 9 #5210 .

HARRY L VA N TREE S 1. Detection, estimation and modulation theory. Par t I , Wiley, New York , 1968 .

J. H . VENTE R 1. An extension of the Robbins-Monro procedure , Ann . Math . Statist . 3 8 (1967) ,

181-190 . M R 3 4 #5225 . M. T . WASA N

1. Stochastic approximation, Cambridg e Univ . Press , New York , 1969 . M R 4 0 # 9 7 5 . J. WOLFOWIT Z

1. On the stochastic approximation method of Robbins and Monro, Ann . Math . Statist . 23 (1952) , 4 5 7 - 4 6 1 . M R 14 , 299 .

JOHN M. WOZENCRAF T AN D IRWI N MARK JACOB S 1. Principles of communication engineering, Wiley , Ne w York , 1965 .

K. S . ZlGANGIRO V 1. Transmission of messages in a binary Gaussian channel with feedback, Problem y

Peredaci Informaci i 3 (1967) , vyp. 2 , 98—10 1 = Problem s o f Informatio n Trans -mission 3 (1967) , no . 2 , 7 4 - 7 8 .

Page 16: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

SUBJECT INDE X

Admissible control , 22 1 asymptotically optimal , 23 0

Asymptotic efficiency, 18 6 normality, 12 3 variance, 18 5

Borel set , 1 2 Chapman-Kolmogorov relation , 3 0 Condition (A) , 3 6

(B), 7 2 (a), 15 5

Control parameter , 22 1 Convergence

almost sure , 1 5 in distribution , 1 5 in probability , 1 5

Diffusion coefficient, 3 8 matrix, 6 7

Distribution o f a vector, 1 4 Drift

coefficient, 3 8 vector, 6 7

Estimate, 18 1 asymptotically efficient , 186 , 21 0

strongly, 185 , 188 , 20 9 weakly, 186 , 18 8

asymptotically unbiased , 18 2 efficient, 185 , 18 8 maximum likelihood , 5 recursive, 190 , 21 0 strongly consistent , 18 2 unbiased, 18 2 weakly consistent , 18 2

Expectation, 1 5 conditional, 17 , 1 8

relative t o a random vector , 1 9 finite, 1 5

Fisher information , 18 2 matrix, 18 6

Function Borel, 1 3 characteristic (o f a set), 1 7 distribution, 1 4 Ljapunov, 8 9 measurable, 1 3

Gaussian whit e noise , 6 3 Generating operator , 3 1

differential, 6 7 Independent

events, 2 0 random variable an d a-algebra , 2 0 sets o f rando m vectors , 2 0 0-algebras, 20 vectors, 2 0

Inequality Cebysev's, 1 5 CrameV-Rao (information) , 18 2 Kolmogorov's (fo r supermartingales) , 33 -3 4

Information content, 6 , 18 2 Fisher, 18 2 inequality, 18 2 matrix (Fisher) , 18 6

Integral, Lebesgue , 1 4 Itd's formula fo r stochasti c differentials , 66—6 7 Kiefer-Wolfowitz procedure , 83ff . Lebesgue integral , 1 4 Lemma

Fatou's, 1 6 Ljapunov's, 13 3

Martingale, 22 , 23 , 6 9 Measurable space , 1 2 Measure, 1 2

absolutely continuous , 1 6 complete, 1 2 finite, 12 Lebesgue, 1 2 probability, 1 3 a-finlte, 1 2

Observation process , 20 7 Plan o f experiment , 22 1

asymptotically optimal , 22 3 Probability, conditional , 17 , 1 8

relative t o a random vector , 1 9 Probability space , 1 3 Product

of measures , 2 1 of spaces , 2 0

Random variable, 1 3 independent o f th e future , 6 1 vector-valued, 1 3

242

Page 17: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

SUBJECT INDE X 243

Recursive procedure , 19 0 truncated, 16 2

Regularity condition , 7 1 Regularization, 3 1 Robbins-Monro procedure , 4 , 79ff . Sample functio n o f a stochastic process , Separating hype r plane, 5 0 Set o f elementar y events , 1 2 o-algebras, 1 2

Borel, 1 2 generated b y a random vector , 1 9 imbedding o f on e i n another , 1 2 intersection of , 1 2 minimal, 1 2

Signalling rate , 21 6 Space o f elementar y events , 11 , 1 2 Stable

matrix, 13 2 point, 10 6

Standard Wiene r process , 5 5 with value s i n £ j , 5 6

Statistic, 18 1 Stochastic

approximation procedure s 4 , 79ff . Kiefer-Wolfowitz, 83ff . Robbins-Monro, 4 , 79ff .

differential (Ito) , 6 3 equivalence, 1 4

of stochasti c processes , 5 4 strong, 6 2

integral (Ito's) , 6 0 22, 5 3 process , 5 3

discrete, 2 2 Markov, 30 , 5 4

homogeneous, 5 4 measurable, 5 3 regular, 7 1 separable, 5 3 stationary (i n th e narro w sense) , 14 4

Supermartingale, 23 , 6 9 Theorem

Fubini's, 2 1 Ito's, 6 6 Kolmogorov's, 5 6 Lebesgue's, 1 6 Radon-Nikodym, 1 6

Transition densit y o f a Markov process , 5 4 Transition function , 30 , 5 4

homogeneous, 3 1 temporally homogeneous , 5 4

Transition probabilit y o f a Markov process , 2 9

Page 18: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken

MAIN NOTATIO N

11*11 = yfcijbff- nor m of matri x B = (tyy)) , p . 12 5

C2: clas s o f function s V(t, x) continuousl y differentiabl e wit h respec t t o t an d twic e con -

tinuously differentiabl e wit h respec t t o x, p . 5 7

C2' se t o f function s V(x) € C 2 wit h bounde d second-orde r partia l derivatives , p . 9 2

Dr: domai n o f definitio n o f generatin g operator , p . 3 1

D£ ' x ' : domai n o f definitio n o f generatin g operato r L a t poin t (t, x), p . 3 1

Eji euclidea n /-space , p . 1 2

7;: monoton e famil y o f a*lgebras , pp . 36 , 5 9

/(x): Fishe r informatio n matrix , pp . 182 , 18 6

/ : identit y matrix , p . 12 4

L: generatin g operator , pp . 31 , 6 7

1*2fo b]: se t o f measurabl e rando m function s /(f , u>) , t € [a, b], a ; e ft, ^-measurabl e fo r

every f , suc h tha t / £ / (f , w)d f < « • with probabilit y 1 , p . 5 9

)lf: a-algebr a o f event s generate d b y rando m variable s X(u), u < t, pp . 30 , 7 3

U: closur e o f se t U, p . 16 2

U€(B): e-neighborhoo d o f se t B , p . 3 9

U€.RW = K €<*> n <* : W < * * ' P - 4 0

K€(B) = EJ\U €{B), p . 3 9

1: domai n o f value s o f unknow n paramete r x, p . 18 1

X(t) ~9l(m , 5) : asymptoti c normalit y o f X(t) 9 p . 12 3

*W = <y x y w ) ,P-202 2: domai n o f value s o f contro l paramete r z , p . 22 1 Vg/Cx): vector wit h coordinate s \f(x + ce/ ) - / ( x — ce/)]/2c, p . 8 6

dv/bx: matri x wit h element s dty/dxy , p . 10 7

r^: firs t exi t tim e fro m domai n G , pp . 32 , 7 3

4>(B): clas s o f function s define d o n p . 4 0

244

Page 19: MATHEMATICAL · 2019. 2. 12. · [ 1), and Kolmorogov and Fomin [ 1). The martingale and supermartingale concepts are due to Doob. Theorems S.l and 5.l' and their proofs are taken