of Statistics and Probability Asymptotic Distribution of One Order Statistic 21 3 Asymptotic Theory of Likelihood Ratio Test Statistics
Statistical asymptotics draws from a variety of sources including (but not restricted to) probability theory, analysis (e g Taylor's theorem), and of
Asymptotic theory (or large sample theory) aims at answering the question: what happens as we gather more and more data? In particular, given random sample,
The asymptotic theory of statistical inference is the study of how well we may succeed in this pursuit, in quantitative terms Any function of the data,
Review of probability theory, probability inequalities • Modes of convergence, stochastic order, laws of large numbers • Results on asymptotic normality
Asymptotic Theory of Statistics and Probability, Springer Serfling, R (1980) Approximation Theorems of Mathematical Statistics, John Wiley, New
To celebrate the 65th birthday of Professor Zhengyan Lin, an Inter- national Conference on Asymptotic Theory in Probability and Statistics
In Chapter 5, we derive exact distributions of several sample statistics based on a random sample of observations • In many situations an exact statistical
22869_6S0273_0979_1988_15668_6.pdf
254 BOOK REVIEWS
BULLETI
N (Ne w Series ) O F TH
E AMERICAN MATHEMATICAL SOCIETY Volume 18, Number 2, April 1988 ©1988 American Mathematical Society 0273-0979/88 $1.00 + $.25 per page
Asymptotic
theory of statistical inference, b y B . L . S . Prakas a Rao . Joh n Wile y an d Sons , Ne w York , Chichester , Brisbane , Toronto , Singapore , 1987
, xi v + 43
8 pp. , $49.95 . ISB N 0-471-84335- 0
Statistics
, generall y speaking , addresse s th e proble m o f ho w t o deter min e fro m dat a knowledg e o f th e underlyin g mechanism , presume d ran dom , whic h produce s tha t data . Usuall y th e mechanis m i s idealize d a s a probabilit y la w whic h i s assume d t o belon g t o a collectio n o f possibl e laws . I f w e hav e abundan t data , w e expec t tha t w e ca n determin e fairl y accuratel y th e unknow n law , o r som e aspec t o f it , sa y th e mean , // , i n whic h w e ar e interested . Th e asymptoti c theor y o f statistica l inferenc e i s th e stud y o f ho w wel l w e ma y succee d i n thi s pursuit , i n quantitativ e terms . Any functio n o f th e data , whe n th e amoun t o f dat a i s n , i s calle d a "statistic " o r estimato r fi{n) of , e.g. , th e mea n /i . Th e sequenc e {p>{nj} i s sai d t o b e consisten t fo r \i i f p,(n) converge s t o / i a s n goe s t o infinity . Th e sequenc e i s sai d t o b e asymptoticall y norma l (regrettably , languag e i s abuse d thi s way ) i f fi(n) - /i ca n b e normalize d s o tha t th e la w o f th e re sultin g sequenc e converge s t o a norma l distribution . Proof s tha t particu la r estimator s hav e thes e an d othe r nic e propertie s i n variou s version s an d setting s compris e muc h o f th e wor k o f classica l an d moder n asymptoti c statistics . I n purel y mathematica l terms , th e subjec t i s abou t convergenc e o f se quence s o f function s o r measure s i n variou s senses ; i n particula r it s tool s ar e draw n fro m tha t par t o f rea l analysi s an d measur e theor y calle d probabilit y theory . Unti l rathe r recently , som e woul d sa y "classically" , a larg e portio n o f prob abilit y theor y deal t wit h operation s o n sequence s o f independen t rando m variables , an d statistica l model s assume d tha t dat a consiste d o f sequence s o f independen t observations . A s probabilit y theor y bega n t o focu s o n othe r processes - Marko v processe s i n th e 50
s an d 60s
, an d stationar y tim e serie s i n th e 60
s an d 70s
, mathematica l statistic s bega n t o dea l wit h model s wher e observation s wer e assume d t o follo w thes e patterns . Th e las t te n o r fifteen year s hav e produce d a stron g thrus t o f activit y i n sev era l area s associate d wit h stochasti c processes : stochasti c integrals , stochas ti c analysis , stochasti c differentia l equations , wea k an d stron g convergenc e o f stochasti c processes , etc . A clas s o f processe s receivin g a lo t o f attentio n i s th e ver y broa d clas s calle d semimartingales . Durin g th e sam e tim e perio d ther e ha s bee n a burs t o f activity , partl y i n respons e t o computin g powe r an d convenience , i n technique s o f dat a analysis , statistica l softwar e packages , an d adaptiv e statistica l procedures . Th e subjec t o f asymptoti c statistics , buoye d up , perhaps , b y th e prosperit y o f it s neighbors , ha s take n of f energeticall y i n a numbe r o f fres h directions . I n suc h a situatio n i t i s a darin g ste p t o writ e a boo k whos e state d ai m i s t o brin g u p t o dat e th e interfac e betwee n probabilit y theor y an d asymptoti c
BOOK REVIEWS 255
statistics . I t i s indee d t o thro w larg e stone s int o a torren t wit h th e ai m o f providin g som e kin d o f bridge .
Prakas
a Rao' s first an d longes t chapte r collect s a variet y o f probabilit y topics , eac h introduce d wit h severa l line s o f orientation . A numbe r o f point s receiv e specia l emphasis . On e i s recen t wor k o f She u an d Ya o o n a momen t inequalit y fo r embeddin g times . Anothe r i s th e extensiv e developmen t i n th e relate d area s o f absolut e continuit y an d contiguit y o f measures . Th e exposi tio n extend s fro m Kakutani' s well-know n resul t t o recen t wor k o f Liptse r an d
Shiryaye
v an d other s o n contiguit y o f stochasti c proces s measures . Th e topic s i n thi s chapte r ar e indee d primaril y concerne d wit h stochasti c processe s eve n thoug h th e statistica l conten t o f th e boo k is , a s state d i n th e preface , mostl y abou t independen t dat a a s oppose d t o mor e genera l stochasti c proces s data . I t i s implici t her e tha t th e potentia l fo r us e o f thes e topic s i n statistic s i s fa r ahea d o f thei r exploitation . Particularl y o n thi s account , i t woul d hav e bee n convenien t t o hav e som e forwar d indexin g i n th e for m o f additiona l note s i n th e "Remarks " whic h appea r a t th e en d o f eac h section , tellin g u s wher e i n thi s boo k o r elsewher e statistica l applicatio n o r significanc e o f th e probabilit y result s ma y b e found . Th e remainin g chapter s ar e abou t asymptoti c statistics , an d her e Prakas a Ra o ha s don e a larg e jo b i n assemblin g an d selectin g piece s o f wor k fro m a n enormou s literatur e spannin g th e fifteen-year perio d beginnin g abou t 1971
. A s a guidelin e fo r emphasi s h e ha s use d insight s gaine d throug h hi s ow n re searc h o n severa l topics . A s a patter n o f expositio n fo r eac h topic , h e ha s selecte d a particula r autho r o r author s whos e writin g lend s itsel f t o pre sentatio n i n boo k form , whos e method s ar e attractiv e an d innovativ e an d whos e wor k i s importan t fo r tha t topic , an d presente d tha t person' s wor k a s a kin d o f featur e article . I n thi s wa y h e ha s manage d t o cove r a n impres siv e arra y o f topic s withou t gettin g bogge d dow n wit h differen t approache s an d interplays . Th e expens e i s tha t ther e i s littl e integratio n o r amalga matio n o f th e wor k presente d o r o f th e literature . W e have , fo r instance ,
Strasse
r o n globa l an d loca l asymptoti c bound s fo r risk , Sweetin g o n max imu m likelihoo d estimatio n fo r processes , Deshaye s an d Picar d o n a partic ula r change-poin t problem , Khmaladz e o n goodness-of-fi t an d s o on . Eac h o f thes e work s represent s a larg e development . Th e collectio n i s extensiv e an d provide s a usefu l introductio n t o man y topic s an d acces s t o thei r lit erature . Separat e referenc e list s a t th e en d o f eac h chapte r an d larg e au tho r an d subjec t indice s mak e thi s a source-boo k fo r orientatio n an d reference . Ther e ar e occasiona l sign s tha t th e author' s ow n revie w o f th e literatur e ma y no t b e ver y thoroughgoing . Fo r example , followin g Moor e [2 ] h e point s ou t a difficult y abou t chi-squar e test s wher e bin s ar e base d o n a n estimate d parameter . Bu t thi s proble m ha s bee n resolve d b y Dzhaparidz e an d Nikuli n [1] , an d recen t literatur e i n thi s directio n i s no t explored . Eve n a cursor y stud y o f th e boo k wil l yiel d a n impressio n o f th e magnitud e o f th e tas k i t attempt s an d als o o f area s wher e researc h effor t i s especiall y invited . I t i s evident , fo r instance , tha t th e exploitatio n o f result s abou t rat e o f convergenc e an d o f "martingale " method s fo r provin g wea k convergenc e an d fo r constructin g statistic s ha s onl y jus t begun .
256 BOOK REVIEWS
REFERENCE
S 1 . K . O . Dzhaparidz e an d M . S . Nikulin , On a modification of the standard statistics of
Pearson,
Theor y Probab . Appl . 1 9 (1974) , 851-853
. 2 . D . S . Moore , A chi-squared statistic with random cell boundaries, Ann . Math . Statist . 4 2 (1971) , 147-156
.
PRISCILL
A E . GREENWOO D
BULLETI
N (Ne w Series ) O F TH
E AMERICAN MATHEMATICAL SOCIETY Volume 18, Number 2, April 1988 ©1988 American Mathematical Society 0273-0979/88 $1.00 + $.25 per page
Invariant
manifolds, entropy and billiards; smooth maps with singularities, b y Anatol e Kato k an d Jean-Mari e Strelcyn , wit h th e collaboratio n o f F .
Ledrappie
r an d F . Przytycki . Lectur e Note s i n Mathematics , vol . 1222
,
Springer-Verlag
, Berlin , Heidelberg , Ne w York , London , Paris , Tokyo , 1986
, vii i + 28
3 pp. , $23.60 . ISB N 0-387-17190- 8 Man y dynamica l system s arisin g i n physics , meteorology , chemistry , biol ogy , engineerin g an d othe r fields exhibi t chaoti c behavior . Ther e i s n o precis e definitio n o f "chaos " ; however , i n simpl e terms , chaoti c behavio r mean s tha t a typica l orbi t seem s t o wande r aimlessl y i n th e phas e spac e wit h n o iden tifiabl e patter n an d it s futur e i s unpredictabl e althoug h th e syste m itsel f i s deterministi c i n nature . Th e onl y know n caus e o f chao s i s hyperbolicity . Sup pos e w e begi n movin g alon g a hyperboli c orbi t wit h th e spee d prescribe d b y th e syste m an d observin g th e relativ e motio n o f nearb y orbit s tha t star t o n a codimensio n 1 transversa l t o ou r orbit . The n i n a n appropriat e coordinat e syste m th e relativ e motio n u p t o first-order term s wil l b e th e sam e a s i n a neighborhoo d o f a saddl e poin t x = Arc , y = My. Th e eigenvalue s o f A hav e strictl y negativ e rea l parts ; th e eigenvalue s o f M hav e strictl y positiv e rea l parts . Thes e rea l part s ar e calle d Lyapuno v characteristi c exponent s (LCEs ) an d giv e u s th e exponentia l rate s wit h whic h nearb y trajectorie s mov e t o o r awa y fro m ou r orbit . I f al l orbit s ar e hyperbolic , al l LCE s ar e uniforml y sep arate d fro m 0 an d al l estimate s ar e uniform , the n w e hav e a n Anoso v system ; a goo d exampl e i s th e geodesi c flow o n a compac t surfac e o f curvatur e - 1 . D . Anoso v an d Ya . Sina i studie d suc h system s abou t 2 0 year s ago . The y constructe d invarian t familie s (o r foliations ) o f stabl e an d unstabl e manifold s an d use d the m t o prov e ergodicit y (i.e. , chaos ) fo r Anoso v system s preserv in g a n absolutel y continuou s measure . I n th e mid-70 s Ya . Pesi n generalize d th e whol e theor y fo r smoot h nonuniforml y hyperboli c dynamica l systems , i.e. , system s fo r whic h LCE s ar e no t bounde d awa y fro m 0 an d som e ma y actu all y equa l 0 . H e followe d a simila r pat h b y constructin g an d usin g th e stabl e an d unstabl e manifold s an d prove d tha t th e measure-theoreti c entrop y equal s J2j ƒ Vj dm , wher e th e //j;' s ar e positiv e exponent s an d m i s a n absolutel y con tinuou s invarian t measure . Late r D . Ruelle , R . Man e an d other s simplifie d an d generalize d som e o f Pesin' s results . I n thei r boo k A . Kato k an d J . M . Strelcy n generalize th e Pesi n theor y fo r th e cas e o f a dynamica l syste m wit h singularities . A n exampl e o f suc h a sys te m an d on e o f th e mai n motivation s fo r th e boo k i s a billiar d system , i.e. ,