Indian Statistical Institute
Transcription
Indian Statistical Institute
Indian Statistical Institute Completeness, Similar Regions, and Unbiased Estimation: Part I Author(s): E. L. Lehmann and Henry Scheffé Source: Sankhyā: The Indian Journal of Statistics (1933-1960), Vol. 10, No. 4 (Nov., 1950), pp. 305-340 Published by: Springer on behalf of the Indian Statistical Institute Stable URL: http://www.jstor.org/stable/25048038 . Accessed: 20/10/2014 12:14 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org. . Springer and Indian Statistical Institute are collaborating with JSTOR to digitize, preserve and extend access to Sankhy: The Indian Journal of Statistics (1933-1960). http://www.jstor.org This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions SANKHY? THE INDIAN JOURNAL OF STATISTICS Edited By : P. C. MAHALANOBIS 10 Vol. 4 1950 Part COMPLETENESS, SIMILAR REGIONS, AND UNBIASED I ESTIMATION-PART By E. L. LEHMANN of California University AND HENRY SCHEFF? Columbia 1. University Introduction is the study of two classical problems aim of this paper of mathematical of of similar and estimation. The reason unbiased the statistics, regions problems is that both are concerned with a family of these two problems for studying together on this family same the measures condition insures a very simple and that essentially The solution of both. of similar region and unbiased concepts an early stage in the development of statistical theory, seems rather difficult On the other hand it ly fruitful. of statistics. systematic development The defined regions were of testing with the problem composite H that a random variable hypothesis to is distributed x, according points = is If the hypothesis ?jpx {P*\?eo>}. the condition cal region A must satisfy Similar Pe Neyman and Pearson In a number replaced of important by Neyman estimate and were both have to justify and Pearson proved extreme either in a completely (1933) in connection one wishes to test the hypotheses. Suppose on values in a "sample X, space" of taking a family some probability of distribution Pex to be tested at level of significance a, the criti for all 6 in o>. (A) < a ... (1.1) of similarity this by the much condition stronger = oc ... for all 6 in <o. (A) Pf cases at introduced the problem reduces by this device 305 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions to that (1.2) of testing Vol. SANKHY?: THE INDIAN JOURNAL OF STATISTICS 10] a simple and furthermore hypothesis, is either uniformly most powerful which 1942: Lehmann, all (among alternative situations cases the most on those test powerful the specific very satisfying (1.1)) frequently strongly depends at which In these the power is maximized and Stein, 1948). (Lehmann therefore the simplifying of value the restriction (1.2) is considerable. Somewhat a again valued the similar regions one among most powerful unbiased (Seheff?, or uniformly in the same hand, the other On 1947). exists there 4 [Part family function remarks ? 3?x analogous of measures A reasonable g(6). the risk for some given weight hold {PBX estimate Given for the problem of point estimation. some real to is it estimate desired \Ge?}, say, which function; seem T would to be one which ... E9[T-g(6)f the expected E9 denotes will depend on the value one clearly chooses T = g(60). where mate value (1.3) this esti Unfortunately, P9*. is d0 If value is this minimized. (1.3) this difficulty would be to replace avoiding calculated of 6 for which One way minimizes minimizes of with (1.3) by ... sup ?9[T-g(e)?. e in restricting consists approach condition of unbiasedness appealing the class of estimates Another EQ(T) = (1.4) by the ... for all 0in?. g(6) rather intuitively (1.5) one minimizing estimates there frequently exists among the unbiased now cannot becomes the variance in 6. of T?uniformly (This clearly (1.3)?which of unbiasedness if we omit the condition unless g(6) is constant). happen It appears that In order completeness be complete to obtain of a family the results of measures of to above, we the notion introduce alluded =--is eaid to fflx <*>}.The family 0X {Mex\6 if = (x) 0 for all 6 in ? d f f(x) Mex = implies f(x) be boundedly 0 except on a set N with complete be a random variable if this Mex(N) implication distribution = holds ... (1.6) 0 for all 6 in <*>.The for all bounded family functions is said to /. Now 3PX, let T be a sufficient in the family P9X for of T. and the distribution denote by (not necessarily 3?x, real-valued) P? function then for any real valued estimable shall show that if 3j?l is complete, We g(6) an that and smallest this estimate is with uniformly unbiased there exists variance, let X with statistic of g(6) which estimate the only unbiased a theorem of immediate consequence We shall show also estimation. unbiased is a function of T of Rao (1945) that whenever ffi similar regions A have a very simple structure (roughly of X falling in A is independent of.), which probability This only. and Blackwell is boundedly that the speaking, was first described ?06 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions ?cnilt is an (1947) on all complete, conditional by Neyman I COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART cases is in these the original hypothesis (1937), and that as a result composite a for the to this method reduced The first to employ finding simple one. essentially was similar P. of L. Hsu (1941). regions totality of a sufficient The applicability of the above remarks hinges on the existence in general there Now is complete (or boundedly complete). are many has the and the question different sufficient arises, as to which statistics, as to the condition. best chance of satisfying the completeness question Actually, statistic which is the appropriate arises also in other statistical sufficient problems. statistic T Speaking one makes intuitively, of a statistical problem plexity condition such that ffi use of sufficient in order statistics information losing one is led to seek without the com to reduce of value. The latter statistic that sufficient being guaranteed by sufficiency, as as to definition the far and hence reduces the statistical possible, problem A sufficient null set qualifications): later with appropriate (to be stated more precisely statistics. sufficient statistic T is said to be minimal of all other if T is a function which a a minimal statistic sufficient this definition we prove that whenever exists, Using is if it can sufficient of completeness statistic equivalent only satisfy the condition the existence the minimal statistic sufficient in a certain sense. We also establish the to of distribu and the sample space is Euclidean a (defined in section 6), density possesses probablity sample generalized and wTe give a method of constructing it, which we show to be valid in this case. We statistic sufficient is of the minimal remark that the result of our construction the minimal tion sufficient statistic when of the statistic of sufficient to the definition by Koopman adopted essentially equivalent (1936) in a more setting. special are also found in the case estimation of unbiased Some results for the problem the minimal where which those characterizes minimum formly end of section variance same -function is not statistic sufficient functions estimable and these A complete: estimates. formal unbiased possessing Finally statistics is obtained theory of uni estimates a justification when testing is given (at the to sufficient (A hypotheses. 4) for the restriction and Rao was case in of estimation the per Blackwell). By given by justification point test and any sufficient we that show decision randomised any functions, given mitting and having statistic statistic there exists a test based only on the sufficient identically the po<wer In this paper ter generality than selves to Euclidean already pected would in sequential to arise in connection test. and for statistical analysis with stochastic processes. arise even came Many sets. with exceptional in the nature of the problems were limited to Euclidean if considerations in the paper are inherent difficulties to on minimum our attention variance while estimates this were that problems null are associated It results given in grea is developed of section 6) the theory (with the exception The reason for not limiting our in statistical is customary papers. that have both for problems is that these are insufficient spaces arisen countered of these as the paper was obtained may of the difficulties We treated, believe and be ex en that most that they spaces. in proof that some of our in papers eaiLer by Rao (1947, 1949) and in an abstract by Seth (1949). 307 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Vol. 10 ] SANKHYA: THE INDIAN JOURNAL OF STATISTICS [Paet 4 The material outlined above forms part I of the present In part II the gene study. ral theory is applied to a number the of more main As special problems. application some theorems on similar regions tests 1942), and 1941), type Bx (Scheff?, (Neyman, most are tests and one-sided extended. powerful (Lehmann, uniformly simplified 1947) These results are obtained and Scheff?. The are solutions which introduced equations by solving the differential by Neyman densities theory of part I is then applied to the families of probability of these equations, and also to some more general families of probabi in a previous results were summarized publication were some of and these obtained and have 1947), (Lehmann independently are also made to some non since then by Ghosh been published Applications (1948). a very simple proof is and For of estimation example, testing. parametric problems similar regions in the non-parametric theorem case, (1943) concerning given of Scheff?'s lity densities. of these Some and Scheff?, this result Halmos is generalized, (1946) concerning and point 2. the is given solution of a problem formulated by estimation. Terminology and Notation a considerable of terminology seems and notation Unfortunately complexity to we minimize this the conventions: Several unavoidable; adopt spaces, following to be denoted will to have be here is the Wx whole Wx, Wt, etc., considered; by space of points x, W* is the space of t, etc. In each space there will be a fixed countably additive etc. Here a family Jfx is said to be count family of sets, Jfx in Wx, Jf* in W\ if additive it contains and with ably Wx, and any set A in jfx its complement Wx?A, with or finite number of sets in union. also their any countable (i.e., denumerable) Jfx The sets in Jfx will be called measurable (jfx). We shall need to define measurable func tions only for the case of real-valued A function/(#) functions: defined on Wx is said to be measurable is in Jfx. A non-negative set (Jfx) if for every real c the set [x\f(x)<c] function Mx defined for all A in Jfx is said to be a measure on Jfx ii it is addi countably = sets in Mx tive, that is, if for any disjoint A Jfx, AX,A2,... (U^. ) ^M^A^). proba = 1. on Jfx is a measure Mx on Jfx for which measures bility measure MX(WX) Probability will usually be denoted etc. by Px, P\ A of measures Mx on Jfx will be denoted by fflx. It is convenient to the members of the family by a subscript 6 that takes on values in an abstract = = if the measures are space o),0* Similarly, we may write $x {Mex\6eo>}. {Pgx\fleu} = measures. A set A in for which Mex(A) 0 will be called a null set Jfx probability family index A set will be called a null set for the family jf?lx if it is a null for the measure MQX. measure a set for every in the family. If statement about the points of Wx is true for we shall say it is valid almost everywhere all x in Wx?N, if N is a null set for (MQX) we shall say it is valid almost everywhere (fflx ) ifN is a null set for i?lx, and we shall Mgx, or this by writing abbreviate An arbitrary (a.e. Mf?x) (a.e.0Lx) after the statement. function sets t(x) from Wx in Wt and a family in Wt whose pre-inu^as to a space W* jf?l' of measures ("complete a countably additive Jp of generates family on Wl : the family Jf* consists of all those sets counter are while the measure in Jf*, images") -308 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions I COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART = to Mfx on Jfx is defined for B in Jf* by M9\B) -flf/ on jf* corresponding Mx(t'\B)), = seen where is with It that this defi of JE?. is tie t-l(B) pre-image easily {a; |. (#)e J3} nition o? Jf \ a real-valued if and function f(t ) defined on Wt is measurable (Jffc only if the function is measurable (Jfx). ) f(t(x are disjoint and cover Wx will be called a deocm of Tfx; if all the elements D in 30 are measurable say that 30 is (jfx) we will position a,measurable on a function is defined and a;0 is If Wx, .(#) (jfx) decomposition. = a;0. any point of Wx, we shall say the set {x\t(x) .(a;0)} is a contour of .(#) through a decomposition on W* determines of Wx which may be de t(x ) defined Any function A 30 of sets D which family of t(x). of 201 being the contours any Conversely, given but can a the is not be found function of 30 ranges Wx, (it unique, t(x) decomposition = are in 1:1 correspondence of an}^ two such functions ) such that 30 301 :For example, value at x is the element of function whose set-valued t(x )may be taken to be the we to the above x. form according If for this function 30 containing description t(x) noted 301, the by elements of those sets precisely family JP, it turns out that it consists X n I30 of Jf l additive of elements D of 30 whose union is in Jfx. This countably subfamily xi30 function 3B or an associated t(x ) will be denoted byJf generated by a decomposition if 30 = 301. or jfx,t. 30 are associated We say that a function t(x) and decomposition the countably additive associated and statistics in probability only the decomposition it being immateral which of the different is of importance, For many a function purposes ted with same the will We :Two for decompositions of equivalence sense if there equivalent (J?x) in the strict two kinds ?B and 30' of Wx are positions .ZVfor |lx such that on Wx?N associa functions is chosen. decomposition encounter with the decompositions 3B and W x exists coincide; decom a null set 30 and 30' 130 ' there exists an A' in set A in Jf (S?x) in the weak sense if for every x 130' xl30' XI30 ' ' such that A and A' there exists an A in Jf and for every A' in jf ' Jf differ by a null set for $x, that is, (A? A1) U (A'?A ) is a null set for $x. are equivalent is the sample space of a statistical being considered on Jfx will then be it by Wx. The family of measures problem we shall always denote X variable that the random a family distributions. 3?x of probability By saying set as for we that to Px is distributed mean, usual, any (the "sample") according be used in will teim statistic The of X falling in A is Px A in Jfx the probability (A). of x de function this paper to mean of X, that is, if t(x) is an arbitrary any function If Jfl and $t are defined as above then fined on WX,T = t(X )will be called a statisticl. of If 30t is the decomposition of T falling in any set B in Jf* is P^B). the probability with the to say also that it is associated with Wx associated t(x), it will be convenient sense are equivalent shall say that two statistics statistic T =t{X). We (3?x ) in the strict one When 1While to functions in this whose paper. ?does give for a We statistic of the spaces some it may purposes associated remark be are measurable decompositions however satisfying this that to restrict convenient our construction the definition nothing ( jfx), of the minimal of is gained sufficient restriction. 399 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions statistic, by such statistic for example a restriction in Section 6 Vol. or in the weak consider sense if their the statistical associated decompositions of these two kinds implications for this Fundamental on [Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS 10] is the notion study are. of sufficient the definition to reader may want of equivalence of statistics. The of the We statistic. conditional (1933, p.41) given by Kolmogoroff set of A in T has the value the statistic that Jfx any given P_(_4J. ) a 6 and A, Pe(A\t) is function real-valued of t, measurable point . : For (JP). base it probability each fixed and defined implicitly by the equation = Pf(At\t-\B)) P.(A\t)dP?, \ B ... (2.1) is any set in Jp, and (2.1 ) is regarded as an identity in B. For fixed 6 and is are not if but defined and two determinations A, P6(A\t) uniquely, ge,A (t) f9,A (t) a are of Pe(A/t), is null set for the set Ne,A where P0K A statistic T is unequal they where B said to be a sufficient statistic for |?x if there exists a determination of Pe(A\t) inde on 6, measurable dent of 6, that is, if there exists a function P (A\t) not depending (Jp) for each fixed A, and such that for every 6 in <?>, every A in jp, and any determination = . in a null set for PeK We note that two P0(A\t), P(A\t) B6,A except for P,(A\t) of determinations must be equal except on a null set NA for ?Pfc. By putting P(A\t) B = Wl in (2.1), we see that if T is a sufficient then for all A in Jp. for $x, statistic = Pg{A) We shall have by Kolmogoroff of x such that of the also need (1933, p. 46). Suppose \ w P(A\t)dPe\ ... (2.2) as defined concept of conditional expectation a measurable real-valued (Jp) function <f>{x)is _?,(*) =f * W X ... (2.3) <f>(x)dPQ, <?= the expected value tion Px, is finite. of the statistic under the probability distribu <?>(X), calculated = a not real-valued), then the If T statistic (in general, t(X ) is ? of T to conditional value calculated under be denoted and <_>, t, Px, expected given defined of t, integrable (jft, P/), by l?e(<_?l.)> is a point function implicitly by J <}>(x)dPf= J EQ(s>\t)dP?, ... (2.4) is in Jf*, and (2.4 ) is regarded as an identity in JE?. For fixed 6, two determina are equal except on a null set N0 of P\ the definitions tions of By comparing EQ($\t ) = seen of a set A in is function and is that if <?> <j>A(x) the characteristic (2.1 ) (2.4) it Jfx, then where B P0(A\t) Returning to the general case of = (a.e.P/). Ee(^A\t) <_? with finite E$ ( _>),we ... remark (2.5) that E9(&\t) can also be calculated from the conditional probability PXA1 )if tnis is known for a11 sets ^ in )Jfx, as was proved by Kolmogoroff (1933, p. 48). It then follows from 310 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions the above defi COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART is a sufficient I is finite for |5X, and if EQ($ ) a measureable of 6, of t, independent for all 6 in w, there exists (jP) point function in ?, 6 for that we has shall denote by E(?\t) the property and which which every = on a null set E(<b\l) NQ forPj. EQ(<b\l) except of sufficient nition that if T for the following from the calculus of condi three formulae if : Suppose T is a sufficient is finite statistic for 3?x. (i) Then EQ(Q ) probabilities from follows ... [ E(*\l)dPf. E,(*)= This statistic use shall have We tional statistic = Wfc. (2.4) by taking B (2.6) is a real-valued (?) If f(t) measurable and JE^) are finite for all 0 in ?>,then ... (a.e.*t). E{f(Tm)=f(t)E($\t) (JP) function of ?,and ifE3(f(T)$) p. is proved by Kolmogoroff easily from (Hi) It follows formula This 105). (2.7) (1933, p. 50), differently by Blackwell (2.4) that if cx <^(#) ^ c2 then cx< E(*\t) < c2 (a.e. $<). ... (1947, (2.8) and of conditional that the values remark finally expecta probabilities on statistic. with the associated the a statistic, decomposition depend only tions, given = = a function T! and T Of two statistics t(X) t'(X) we shall say T is of T = a function t(x) \?r(t'(x))(&.e. S?x). t-=\?r(t') on Wtf to Wl suchthat (a.e. i?*) if there exists with the functions associated and ?t' In terms of the decompositions t(x) and 3Bt on the that a Wx?N Af such set for null means exists there S?x decomposition t'(x) this in an element of is is contained that 30t' the subdivides 3Bt, every decomposition JBt/ statis statistic T for $ix will be called a minimal of 3Bt. A sufficient element sufficient We tic for 3?x if, for any conditions Sufficient finding it, are given other for the existence in section 3. Given in Wx, consider for 3PX,T is a function sufficient of minimal statistic, statistic sufficient a Family of = jf?lx integrals {M*\de<?} of the form f(x in general of 6, which (a.e. |?x). a method of of measures of Measures on the additive family ... f J(x)dM.*, x J where of T' and 6. Completeness a family T' jf* of sets (3.1) W (Jfx ). The value (if any ) of this integral will is used; of the family $[x (3.1) is a function on Wx to a func from a function as a transform defined f(x ) the function this transformation that is every of w. Under and measurable ) is real-valued on which measure Mx depend we may regard tion of 6 defined on a part zero on Wx goes into the function where zero on a>. is everywhere Completeness on Wx is the only function zero into the zero means function the that going roughly of the transform. The exact on w ; it is a unicity definition is function property the following : The family ?lx f that of measures f(x)dMx is complete = 0 if forall?inco ... 811 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions (3,2) Vol. 10 ] implies f(x of unbiased SANKHY?: THE INDIAN JOURNAL OF STATISTICS = 0 (a.e. $tx). This definition ) estimation ; for the problem of bounded perty condition (3.2) and is appropriate of completeness of similar regions we require family ffix is called boundedly The completeness: the condition [Part 4 that f(x) is bounded jointly for the problem the weaker pro complete2 if the that f(x) = 0 (a.e. imply 4*tx). note We that of a family example is a slight modification if jfttx is complete of measures which of an example and Savage Mosteller, it is boundedly is boundedly constructed The complete. complete for a different simple following without being complete by Girshick, purpose (1946): *> is the interval 0<#<1, open Suppose Wx is the real line, = x 6 to the to the points the measure and 0, 1, 2..., measure (l?6)26x PQx assigns = ? zero measure to the of set and this of The 1, complement pointa? points. condition (3.2) then becomes Example 3.1: f(-l)d+lf(j)(l-d)2V=0. We is the coefficient of 6l in the Taylor for j = 0, 1, 2..., f(j) Since for for the function ?/(?1)6>(1?O)'2. |0|<1, see that origin 6(l-d)-2 = series about the ?jdi, ro it follows that (3.2) is satisfied if and only if j /?)-?i/(?1). = ... 0,1,2,.... = (3.3) = 0 for j is bounded, satisfies if f(x) Hence ?1, 0, 1, 2,..., that is, f(j) (3.2) and = On thfe other if f(x) is and thus 3?x hand, complete. boundedly 0(a.e. $x), f(x) = 0 satisfies 0, then it satisfies (3.2) but not the condition f(x) (3.3) and /(?1)_t_-. is not complete. (a.e. $x), hence $x in general completeness of jf?lxdoes not imply the same for Jflflx,but of a subset Mi all null sets for jRf are null sets for ??tx. hold provided It is worth We while shall now to note give that some simple is discussed completeness jpxwhose family in statistics; has been used extensively or bounded that this completeness does implication In each case the of completeness. examples measures that will be a family of probability a it is of interest that in number of these examp for a transform to the problem of unicity case the space Wx literature. In every on be taken as Wx and the additive may family rjp reduces of completeness les the question in the mathematical that has been treated will be the real line ?oo <.*:<+oo, set of space w of points 6 may be taken as an appropriate is complete in which the family of measures All the examples real numbers. may be _. the subsets without certain modified proper completeness. destroying by by replacing one con some means of which in Part of theorems that here II, We mention general by a in product cerns completeness may be proved for large number of spaces, completeness the class of Borel 2 The publication property The sets. here called bounded completeness was called completeness (1947). 312 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions by us in an earlier I COMPLETENESS,-S?MILAR REGIONS AND UNBIASED ESTIMATION?PART of measures families and Here line. in n-dimensional Euclidean use we in examples elsewhere the d Px?d[Xx = to indicate the measure that is absolutely Px from these spaces "Nikodym examples on the real notation derivative" g(x ) with continuous respect to the measure that px and P*(A)= \ g(x)dixx A for all A in Jfx. If |?x 3.2: Example 0 and unit variance, mean is a family then = dPx?dx of normal - (27r)-*exp distributions probability t with (x-0)2 [-!<*-*) ] If the condition satisfied (3.2).is by f(x) r\ f(x) J ior ?oo<6<oo. exp exp and (?\x2) (a.e. $x). ? Thus Example 0 we variance find r? i i -^x2-\-0x \dx= ... 0 (3.4) of the function/(#) transform Laplace (3.4) is the bilateral of this transform it follows that f(x) theorem the unicity = 0 on measure and L denotes Lebesgue where Jfx, hencef(x) Now (?\x2), exp we from 0(a.e. L), is complete. $x For 3.3: a family 3?x of normal with distributions zero mean and have d Pex?dx = (2*0)-* exp ( - x2 -j? (-?-) J for 0 < 0< oo. ... (3.5) or even boundedly for every 0 the density complete because family is not complete is an of x, and hence (3.2 )will be satisfied by any f(x), which (3.5) is an even function of x and such that its product with odd function (3.5) is integrable (L) for all 0>O. t = x2we find measures transformation the a new of set we to If transform P9l by ffi This fl (2ntd)-1- dPJIdt^ < 20 10 Condition (3.2) written for $t l exp(- for t< 0. instead of 3PXgives [ /(?)r*exp?-? for 0> 0. Letting r = transform we find/(i) -^t)iort>09 - * )*'=? the unicity theorem (20 )~l,and applying = 0 is complete. that is, ffi (a.e. ffi), for the unilateral 313 ? This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions (3-6) Laplace 3.4: Example we a For have J?x of family = dPx\dx reasons 3?xmay again be symmetry = x2, the completeness before, we let. seen of the Stieltjes with distributions Cauchy zero median for 6>0. \nd\l+x2\d\-x For property [ Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS Vol. 10] not of the to be resulting If, as complete. from the unicity boundedly 3P* follows transform f fm ? t+6 butions a For 3.5: Example if 0 is a half distri (chi-square ifa>0, f2-M~iexp(-te)/r(0) <? I 0 if a;< 0, = follows completeness distributions gamma integer), dP;?dx 0>O, of family the unicity from of the Mellin property transform 00 0 3.6: Example If is the uniform Pex distribution on the interval (0, 6), fll6if0<x<d, dPexldx= i ^0 elsewhere, the 6>0, of ^x completeness follows from a theorem : If of Lebesgue 9 0J for all 6 in an interval, a function which f(x), then f(x) is involved The 3.7: Example family f(x)dx.= 0 = 0 (a.e. _L) on the interval. in this case, is its indefinite of $x uniform The transform (3.1) of integral. on distributions the intervals i?, 0+1), [Ii?d<x<0+19 dPe ?oo<0<oo, is not boundedly x?dx= <{ ?^0elsewhere, : If f(x) complete is any periodic function 1 and i [ f(x)dx o it is easily seen = 0, that 314 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions with period COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART I + TO = 0 f"00/(*)d P.' for Px all 0. so that of Poisson distributions, 3j?x is the family ? x zero to the 0, 1, 2,..., and measure e~e0x?x\ to the points set. The condition be written (3.2) may 3.8: Example the measure assigns of this complement for 0>O. and $x series of zero expansi?n it is complete. distributions of binomial family Px constant 0, 0<#<1, probability If $x trials with independent of 0* in the power coefficient 0 (a.e. $x), 3.9: Example to n is the Since/(j) that f(x) = follows Suppose is the corresponding the pro assigns bability x = to the points 0, 1,..., n and implied by the theorem of the argument values zero probability elsewhere. Completeness of degree n vanishes that if a polynomial forw+1 zero. it is identically of 3?x is distinct distributions of the 3.10: Let 3?x be Example family hypergeometric in the of defectives for fixed lot size N and fixed sample size n where 0 is the number == the in the sample, is the probability of a; defectives 0, 1,..., N; ifPx lot, 0 Pex assigns discrete probabilit?s (!) (?=9 / (?) to the x = points In 0, l,...,n. = M Condition then (3.2) (3.8). ables find Hence successively /(0) |5X is complete. 0ifr<0orr>m. = = ?,... m f<4) (??ZI) 0, /(l) = 0, ..., f(n) = 0 by putting 0 = 0, 1, ..., w in ... be a sequence of vari random Let Xx, X2, independent 1 and 0 with probability p and l?p capable of taking on the values on these variables where the a sequential scheme Consider sampling Example each being respectively. on whether decision already understand becomes I and we (3.7) we been taken 3.11: or not ismade to take according an A^+P1 observation, N to the value of 21X{. when Let N observations have n be the total number i=l 315 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions of n is arandorn in one experiment? taken observations for of bounded completeness problem rule and the values stopping 0<j?<1, and Mosteller S Savage to be (1946), x{ 1=1 / (N such a point scheme being was N \ __. #i? N? n ? and let X ? 2 XL. the solved Wolfowitz I as (1946), coordinates if it has to a fixed |JX corresponding family as follows in a series of and of papers (1947). Savage a point in a plane, by For and the sequen under probability so that, tial considered but is not a stopping for this procedure point is observed, in this point another will be taken. observation As was shown when is and sufficient the above papers, a necessary condition for bounded completeness N to the same value of that, given any pair of accessible points P1,P2 corresponding sum N is the of the of all the coordinates points (for any point P, P), lying on the line define segment accessible. connecting The minimality Px accessible variable i=l The Girshick, [Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS Vol. l? ] and P2 and positive valued integer having of completeness in this section is concept developed of a sufficient at end of introduced the secton statistic, is a sufficient is a minimal sufficient Theorem If T 3.1: complete, and if U sense. (;PX) in the weak The proof statistic for 3?x such statistic for 3?x, then T can be based of this theorem are coordinates also to that related of 2, by is boundedly and U are equivalent that ffi on the following = = are two sufficient statistics t(X) and T' If T tf(X) for as and A x are equal in Jp, and considered ?Px, if for every functions P(A\t'), of P(A\t) in the weak sense. (a.e. ?Px), then T and T' are equivalent the lemma To prove let jfxlt = t~x(ix), and define Jpi^ analogously. We a set A' in Jpi*' which shall prove that to any set A in Jpfc there corresponds differs Lemma 3.1: by a null from A set for 3?x. Let <j>A(x)be the = (?>A(x)depends on x through t(x), say (?>A(x) then P(A\l) and = characteristic g(t(x)). By of the function (2.5)and E(g(T)\t) =g(t) (a.e.$0, Ux) (a.e.W) set _4; (2.7), thus P(A\t(x)) = But by hypothesis P(A\t(x)) P(A\tf(x)) (a.e. $x). ... as the set in Jfxt' where the right member of (3.9) equals 1. Now ol (3.9) equals for $x such that on Wx?N member the right and the lemma is established. of A and A' in Wx?N coincide, Define A' a null set N the parts = 316 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions (3.9) there exists hence (?>A(%>}\ COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART Theorem for which 3.1 will an if we show that if there were thus be proved Ax in Jfx were not equal then could not be ffi (a.e. $x), P(Axl{u(x))y the real-valued Consider measurable (jfx) function v(x) defined by and P(AJt(x)) boundedly I complete. v(x) =P(Ax\t(x)-P(Ax[u(x)). We note that and that (a.e. $*), |v(a:)|<l the set in Jfx where If V = v(X), it follows from (2.2) that set for $\ v(x) ^ 0 E.(V) for all ? in ?>. Since for $x is a minimal U V is also a function and hence so that the result sufficient of T We (a.e. $'). is a measurable of v(x) does not invalidate and (3.10), is not (3.10) of T redefine 0 is not a null set for 3?f. This redefinition boundedly 4. ... thus fwt/<i)dzy for all ? in o), and hence ffi =0 a null (a.e. 3?x) v(x) on a null set of t, say f(t ), with the properties can thus function (jfl) that |/(01< 1, and the set in jp wheref(t) ^ it is a function statistic is not Similar = o complete. Regions in jfx is said to be a similar region of size oc for the family |?x of probabi on Jfx if Pfl%4) = a for all 0 in w. Neyman (1937) noted that if T is a lity measures for J?x and if the set A has the property statistic sufficient A set A = ... P(A\t)*(a.e. $'), then A is a similar region of size a for |5X; this follows in jfx has the Neyman did not (4.1 ). Neyman set A structure with (4.1) from (2.2). sufficient respect to the under what conditions, that a whall'say statistic T if it satis We statis given a sufficient investigate : T to structure this is of all for have this T for similar with tic ^x respect |?x, regions since there one wants in the Neyman-Pearson tests, theory of optimum importance one therefore needs to know the and of all similar to choose the "best" regions, fies totality Theorem of such regions. A partial answer is given by the following corollary to 4.1. is boundedly a sufficient and if ffi statistic for $x, 4.1: Corollary // T is then a set A in Jfx is a similar region for |?x if and only if it has theNeyman complete, structure with respect to T. is that sets A arises of similar regions problem tests of a hypo as possible critical in Jfx are being considered regions of statistical is rejected by : If the sample X falls in the critical region A the hypothesis thesis a similar region If A is is accepted. the hypothesis if it falls in WX~A, the test, while in $x. distribution for all is constant the of rejection for |5X, probability probability x such that of points to employ a third category it is convenient For many purposes or one does not is observed when one of these points accept, always reject always The situation in which the 317 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions [Part 4 SANKHY? :THE INDIAN JOURNAL OF STATISTICS 10 ] Vol. but rejects according to a chance method (say, with the help of a table of random mum the probability is a predetermined of rejection number bers), for which <?>(x),0<$[x) whole < 1. One may then extend to of the function the the definition sample <?>(x ) = 0 on the = 1 on the and set set, space Wx, by setting acceptance <p(x) <?>(x) rejection which and is thus led to the notion of critical function and may Stein, 1948), (Lehmann as a special case of the randomized of Wald be regarded functions decision (1947): x 0 1. which A critical function for of is measurable function < <f>(x)^ (Jp ) <?>(x) any one rejects the hypothesis with proba Its use in testing hypotheses is that when X=x a of the random to random pro process (?>(x) according independent bility statistically cess governing X. calculated under value of rejection is then the expected defined Critical regions by (2.3). PQX, namely, JS?(?_>)as are characteristic to the special case of critical functions which correspond of sets in Wx, The that probability take on only is, which 0 and the values is constant and equal to a for all 6 in <owe shall Eq(<P) critical function of size a for 3?x. Clearly, similar regions are characteristic case of similar critical functions which special is its characteristic If A is a set in Jp and <f>A a if similar critical function if and only 0A is <?>(x)is a to the correspond of sets : functions with that is a similar region for 3?x func shall We say that a critical to a sufficient statistic T for |?x if a if critical function <f>is the charac structure with <j>(x)has the Neyman respect = a (a.e. ?J?*). It is obvious from (2.5) that E(<P\t) of a set A in Jp then $ has the Neyman teristic function structure say then A function, for $x. tion and only if A has the Neyman follows from functions 1. If similar of <_> =<?>{X) seen to are structure to T. respect with to T if respect Therefore Corollary 4.1 4.1: statistic 3?x,a necessary andsufficient If T is a sufficient for all structure similar with respect critical to have the for Neyman functions for $* i? that $fc be boundedly complete. Theorem condition to T shall We of size oc for for all ? prove 3?x. in ?, Then first from all 6 /(.) = E&\b)-*. ... Since O<0(.c)<l, (a.e. $*) by (2.8). 0<l?(*)Ji)^l determination of E ($|.) in (4.3) and then redefining/(.) assume /(.) to be bounded. is boundedly Since ffi (a.e. ^e), that Next complete is, </>(x)has the Neyman we there shall would (4.2) Jv/(t)aP0? in w, where prove exist function = * = 0 ... for critical (2.6), \^E(*\t)dP? so that <j>(x) is a similar Suppose sufficiency. necessity similar structure by critical with showing functions (4.3) a particular by taking a on null set for $)fc, we may = 0 (4.2) implies /(.) complete, Hence respect that to T. if ffi were (of every 318 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions not boundedly not size a, 0<a<l) COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART having there the Neyman structure with exists a bounded measurable to T. respect function (jP) is not If$fc f(t) such I complete boundedly that ?JWu for all 0 in w, but A, ={t|/(i)^0} Suppose |f(t)\ <M, and define = flW cf(t)+z, is not a null set for$'. where 0<c<Jf-1min Then {a, 1?a}. O<0(i)<l, ... {J W tg(<)dPe'=a (4.4) for all 0 in ?, and ... [?(^aonif. as critical take Now ?($!*) = function 0(a) = Then g(t(x)). E(g(T)\i), and hence by (2.7) E($\t) We (4.5) = (a.e. $*). fl<*) ... (4.6) is a similar critical function of size a for $*, (2.6), (4.6) and (4.4) that <j>(x) have the Neyman structure wsth respect to T. (4.6) and (4.5) that it does not see from and from be appropriate there exist for which It may butions with respect to a sufficient a simple example of a family $x do have which not the Neyman similar regions statistic T. to give here of distri structure be a random 4.1: Let X = (XVX2,..., sample of size n>l Xn) Example The on the intervals distribution space Wx (0? J, 0+J). from the uniform sample as Borel sets the additive in Wx, the as a Euclidean family jp w-space, may be taken ? of the sample is the uniform P bo<#< + oo. The distribution and toas the real line Qx ? = i Let tx(x) = min cube on the 7i-dimensionsl l,2,...,n. distribution \x- 0\<%, = max %i>h(x) i dimensional probability &.. The Euclidean t= transformation space W\ and t(x) the = (^(cr), t2(?)) maps statistic T = ?(X) = Tfx into the i two (?\, has the Ta) density fc(?a-?i)n"a if ?-|<^i<<2<^+i, [0 otherwise. = is known 6.4) that T is a sufficient (it also follows from Example dpi' dtxdt2. It is not boundedly 5.3 that W for $x; it will be shown in Example complete. statistic with to T can be based on structure the respect not Similar Neyman having regions a continuous distribution of 0, there independent the range R =t T^T r Since B has where 319 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Vol. 10] exists for any a(0<a<l) The set R<ra. [Part 4 THE INDIAN JOURNAL OF STATISTICS SANKHY?: a constant r? such = Aa all for that 6 the probability is a that {x\t2(x)-~tx(x)<ra)} is thus a similar Since the characteristic of the set Aa can be expres function region. sed as a function takes only the values of., it follows from (2.5) and (2.7) that P(Aa\t) 0 and 1 (a.e. i?1-), that is, Aa does not have the Neyman structure with respect to T. we of similar illustration take n = 1, an even simpler If, in this example a can to not the structure with respect sufficient statistic possessing regions Neyman measure a and be given. On the interval (0,1) in Wx take any Borel set _4 xof Lebesgue 1 so A is and coincides function then define that its characteristic periodic with period is a sufficient statistic for $x and P(A\x) = 0 that of A x in (0,1 ). Then X = X with or 1 (a.e. $x). conclude We statistics: section this ? be a set Let the by noting in the ?-space following containing "justification" of sufficient and <i>,let 3?x={Pex\deQ}, given any critical function sup is sufficient for iPx. Then T =t(X) pose <?>(x) on x there exists a critical function for testing H0:de(?, <?>x(x) depending only through as <?>(x). The proof is similar to that of Rao the same power function t(x), and having Let ijf(t) = __?(<_?].) and let <Px(x)= ijr(t(x));. \?r is and Blackwrell's theorem (Section 5): of 6 in Q since T is sufficient for |?x. The power of rejecting independent (probability the statistic ) is H0) of <?>(x = E9(^) and by (2.6) this equals Eg(&). be remarked It should E9(f(T)) =Ee(E(*\T)), that an even stronger was given of sufficient justification by Halmos statis and Savage of statistical inference for all problems (1949) means a use of the statistic and out sufficient of that random variables by pointed a statistic to construct it is possible the same distri with known distribution having as Their that however there exists the original bution presupposes argument sample. tics who measure is a probability that probability P(A\t) of this supposition has been established (1948, p.399) by Doob is a Euclidean The for statistics sufficient space. justification of the a determination (a.e. $*) ; the validity in the case that Wx given by us above out any restriction conditional for the problem on the sample 5. of testing has the advantage of being valid with space. Ujsbiased Estimation will denote the family of all distributions to ={PQx\deo)} a a of statistical is restricted attention inference. which priori in problem particular as statistics solutions of of real-valued unbiased esti In considering problems possible to those with finite second moments. we shall restrict ourselves Let V be the mation = statistics V for is which class of all real-valued v(X ) v(x ) measurable (Jfx) and E9(V2) ?. in w. is V all For also 6 for in in finite 0 all for We shall say finite %J,EQ( V) is any In this section $x 920 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION-PART a real-valued that function is estimable g(0) = g(0) for all 0 in <o. An unbiased variance will be called a minimum is a minimum V variance = g(0) 'Eg(V) (?) estimate variance a V exists if there in V I that such Eq(V) is of uniformly of g(0) which minimum of g(0), that is, a statistic V in estimate estimate of (7(0 ) if for all 0 in <*,and (n) Var? (F)<Var^( V ) for.all 0 in o>and all F' in V statisfying (i). the notion applying the following easily By obtains to a result of completeness of Rao and Blackwell, one that $* there exists a sufficient statistic T for 3?x such If is complete variance has a minimum then every estimable function estimate, and a statis ic V in V is a minimum variance estimate of its expected value if and only if it is a function of T (a.e. 3?x). 5.1: Theorem The is a con of Rao this theorem (1945) and Blackwell (1947) of which is a sufficient statistic for |lx and V in V an unbiased = <*> if i/r is defined by i/r(f) = E(V\t), \?r(T) is also an unbiased result states sequence of g(0), then that if T estimate estimate of g(0) with Var^ (^)< Var^ (V), equality holding for all 0 if and only if F is a function of ?7 (a.e. $x). To prove then there exists function; g(0) is an estimable Let * be defined as above. To see that * is a mini 5.1 suppose Theorem = a V in V for which g(0). Eg(V) of g(0 ), suppose! Vr in V is any other estimate == and define Then f'(t) E(V"\t). by the Rao-Blackwell unbiased estimate of g(0) and for all 0. Var?(^')<Var?(F') mum variance of g(0), estimate = theorem ?' i}f'(T) is an = Since 0, E0(*'-*) unbiased J w for all 0 in <o, and hence it follows from the Thus (a.e. $fc). Var?(*)= Var?(^')<Var?(Fr), estimate of g(0). last part of the theorem The in the Rao-Blackwell theorem. equality The application of Theorem 5.1 is obtained is illustrated 5.1: Example 3?x is the Sappo3e family of size n>l from a normal population with It is well-known - (27r02)-i* of distributions t2{x) = S xjn, i=l = S condition for of a random 0X and variance 02. With = ... Wx, djax dxn, and dxx dx2 S (*i-0i)a that with tx(x) the mean - ~ exp from by sample x = in the Euclidean (xvx2,..., xn) a point space 0 = (0V 02) a point a>= in the half plane {0|02>O}, dPo'ld/f = of $l that \?r(t) i?rf(t) is a minimum variance completeness and so * ?i-^?x) , 321 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions ... (5.1) Vol. 10 ] [Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS T = = statistic t(X) is a sufficient sufficient and that statistic), (TVT2) as a minimal 6.2 it will (in Example and _T2 are independently for $x be derived Tx distributed with joint probability density ! J--f->r '' ^ if t2 > o, (2?2) \?(nMV*Th^\ 0 if ?2<0. is complete. II of this study it will be proved that $[ From the com of the sample mean at once it follows functions that the following are estimates of the indicated minimum variance func variance T2?n In Part of $' pleteness sample -\ancl tions of 0-and -""-of. 8t: cn ^"l+VnA/^a from mined and x; _Y(n-l) = of the population, (lower) normal respectively, where 2^(i(n-l))r(in); 100 p percent point so tables that probability ?f tne y/0t, .-yT,..02and where ap is deter aP (2?)-? f-00 xp(-?Z2)dZ = ?>. 5.1 that any real are all special cases of the statement from Theorem following a is minimum variance variance of its estimate valued of (TVT2) with finite function value. expected These Theorem 5.1 : can be extended easily to the case of simultaneous estimation Cramer (1946), we express the concen of 6 if, following of several real valued-functions of concentra in terms of its ellipsoid its mean, about estimate tration of an unbiased = of 6, and let g(d) functions tion. Let gL(6),..., (gx(6), ...,gm(6)). gm(6) be m real-valued = an of g(6) if V{ is an unbiased estimate is unbiased shall say that V We ...Vm) (Vv of g(6) V will be said to be an estimate The statistic of g{(6) for i = l,...m. estimate for of g(d), if estimate if it is an unbiased concentration with maximum E9(V2)?x> = estimate concentration the if unbiased other for in and U, and all 6 <o, i any l,...,m of U. T for |?x such that 3Pfe statistic sufficient = ..., then every estimable function gm(0)) for which gx(6), ..., g(6) (gx(6), concentration. has an estimate with maximum linearly independent, The : If there 5.2 Theorem is complete, gm(d), 1 are in that is contained of F elliposid proof of this theorem follows m quadratic forms: * Pointed out Let II a exists easily from the following be two positive result3 concerning definite quadratic m a.-y^y., to us by Professor E. W. ___ bi,yiy-: Barankin. 322 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART m m and forms, 2 let I 2 a'^y^., h ri be b^yy^ i,M m the inverse corresponding forms. Then m S i, j=l S ?ij-Mj < i, j-1 fttf-Mj for all real y's implies m m 2 *, j=i let F = Now let ?7 = and Then for any m of s i, j=i b3 any unbiased (Vv...,Vm) (Uv...,Um). &ij#i2/j for all real y's. 2 "" > ftij/A2/j estimate of g(0), m S ^Fj i=i real yv...ym, = let ^ is an unbiased m 2/i??i(0)3 and by Theorem the variances Comparing estimate m is a minimum 5.1, S yiUi i-l i<L i=l two of these ^F^?7), estimates m we estimate variance of S y^S?). find m are the covariances and Further pf (Ui, U. ) and (Fi5Fj ) respectively. Au.j A.^ so are 1 are linearly and since gi(0),...,grm(0), Fl5...,Fm,l Uv...,Um,l, independent and hence the above quadratic definite. forms are positive It follows that where more m m S i,j-i which the theorem proves since the = A^^y. and m+2 shall now answers which minimum any shall develop about questions which but estimates, statistic T, three the may defined class ra+2, are defined respectively. a single real-valued estimates variance of estimating of minimum the problem a formal theory to the class V of statistics sufficient = Xv'hj^. once more certain variance In addition for consider of 0, and we function S i,j=l *,j=l We of U and F of concentration m 2 equations i,j-i the ellipsoids m by S Av^ Vj2/i?/j > of estimable functions possessing in specific problems. to apply of this section we define, at the beginning be difficult classes ?T, ^Tft, and V?1. Here T will in to be be expected theory may all class of the be for 3?x. Let VT simplest We define $T? as the subclass of sta are funct'ons in V which of T (a.e. $x). statistics = 0 for all 0 in o>. If $fc is complete, VT? consists tistics F in VT for which E$(V) = 0 In statistical F = v(X) for which v(x) (a.e. $x), and conversely. only of statistics are estimates of T which of those functions consists say the VT? language we may subclass be the to defined is of zero (unbiased, with finite variance). VTX Finally, that FF? is in ?T? for every F? in VT?. the condition F satisfy of ?T whose members general not be real-valued, if T is a minimal and the application statistic sufficient of the 323 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Vol. [Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS 10] = = 0 for 0, or Cov9(VV?) every all real constants.4 contains An is that condition equivalent Eg(VVQ) class Vo in The every VT?. #/ always 5.3: Theorem A statistic 0 in *>and 10 is a minimum in variance estimate of its a it is The class and member all estimable Jg-A S expected of func if only if of tions g(6) possessing minimum variance estimates is thus obtained by applying the operator V value E9 to themembers of V^. strict If we identify all V in V which are equivalent ($x) in the then the correspondence sense, between is 1:1. <? and Vt1 = To prove the theorem variance suppose first that V v(X) in V is a minimum = = estimate of its expected value. Define Then i?r(t) E(V\t). v(x) (a.e. $x) f(t(x)) Rao and is Blackwell's and be thus V in Let Vo element of #T? theorem, by any ?T. = and let U V + XVo, where ? is a constant. estimate Clearly U is also an unbiased of 0(0) =E0(V), and = V&Ye(U)-VM9(V) A2V&t9(V?)+2?Cov9(VV?). This function of ? cannot be negative since F is a minimum esti variance quadratic = and it is easily found in either of the cases mate, 0, that consequently )>0 or Var^F0 F? )= 0 for all 6 in <*and Vo in ?_?. is Hence in V VTl. Cov^F Next g(d) = is in V?1 V that suppose = E9([V). Then if w(t) and Also is a minimum F The last Lemma V = V1 with statement 5.1: tion one for ;> 1; hence p(d) 4 it for Pg should be defined of (a function <& is the minimum the efficiency is defined is one seems the members = Var?(F)+Var?(F?) > Var^(F). 5.3 follows are F1 from minimum the following variance estimates of then g(6) <_.. all 6 in - to us of />(#) in the and ?^,1, then variance to the = 1, and of Theorem light that there be defined estimate variance lower bound of g($). 5.3 that lh(d)+mO)+ih(0)P(d). exist the as opposed (absolute, V is unbiased only when e) may = Var^F+JF1) constants Ag,Bg that of the minimum relative 0 for all = = let h(6) = If for some d, 0, then Var^(F) Var^F1). h(6) = V1. one for this 6, V' = If h(d)>0, let p(6) be the correla g(0) of F and F1, form the unbiased U = ^(V + V1), and note that estimate probability coefficient probability = EQ(W?V) this A(0)<Var0(U) Thus is in VT, W-V estimate. in Theorem V and If probability To prove with variance of estimate now have Var^f/) > Var^IF) Var?(F+F0) Hence unbiased is also unbiased, and V*rg(W) < E(V\t), W =w(T) theorem. Var?(C7) by Rao and Blackwell's 6 in o), and so Vo = IF? F is in #T?. We is another U that is a member and g($) as feel # for the variance that may "efficient if be efficiency less than of unbiased estimate" class Vare(F) shculd > not 0, if Var?(7) be defined 1, as is the case estimates estimate <g of the theorem 1 given by ; its efficiency = in such the Cramer-Rao 324 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions reserved V of g($) for example lity. that the be should of an "efficiency" of the Varfl(^r)/Var?(F) We estimate term to relative) such 0, where a way if that efficiency inequa COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART V1 = Ae+BeV. Hence E9{V*) = I ... A0+B?E9(V); = = (5.2) = = so 1. If -1, 2g(0) from (5.2), Var^F1) BQ Ag thatP^2 'B9*Vnr0(V), = = and with Hence and to one, F1 0; 2g(0)- V U=g(6). equal Var^E/) probabilityPQ = we see that but From this contradiction hence from (5.2), 1, Var6>(C/)>A(0)>O. BQ = = F1 with to one. thus and F 0, equal probability Ae PQ also, 5.3 in the case of Theorem shall now give twro examples of the application statistic is and not complete. $5fc (actually minimal) = x in Example 5.2 : We 3.1. take t(x) In defining statistics Example = ? x of definition matters the the at function the points F ?v(X), 1, v(x) only = V is d3fin3? th3 class of fun3fcions for which series the 0, 1, 2,... VF power v(x) by We where is a sufficient T in0 j=0 has a radius ^>1. From of functions of convergence T0t? is defined by the class *(j) for such = v(?l) series functions the radius = of of convergence = (5.3) is equal series above the ... 1, 2,...); 0, l(po, if the above to v(x) satisfying ^Txis defined by the class of functions Finally all vQ for and such that v(x)v?(x ) satisfies condition (5.3), (x) satisfying (5.3 ) all v?(x) such ... ~jv(-l)v?(-l) v?(j) =-jvO(-l). we find that V = v(X) is in ^t1 (5.5), (5.4) and <? of extimable two-parameter (i v(j)=v(-l) Then for F in 3?T\ Eg(V) class = = = 0}+v(-l)P${X ^ v(0)Pe{X =?(o )(i-ey+v{-1 ){2d-e*). minimum g{6 )possessing of quadratic functions = that only In this constants 0} variance estimates is thus cx+c2(i-0)2, of g(0 ) is the statistic estimate variance the minimum = X 0 and the value cx if ^ 0. cx+c2 if X 5.3: if 1,2,...), functions family (5.5) if and only and to prove is, (5.4) ... = g(0) Example that that with v(0) arbitrary. the (j 0). Combining The v(x ) satisfying -j*>(-l) v(j)v0(j) for see that 3.1 we of Example calculations the we example have minimum shall see that variance V(X ) taking on the value it is sometimes estimates, 325 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions without possible completely Vol. [Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS 10] the not be very difficult in the present this would (although of distributions of a random let 3PXbe the family As in Example 4.1, example). on the interval We from a rectangular sample of size n>l 6+\). population (d?\, = 4.1 and consider define T From (4.7) we see the class ^T?. (TVT2) as in Example co that a statistic of the type F is in VT? if and only if f(Tv T2) (a.e. $x) Eg(V2)< and determining for all 6 in o. = class VT? (?oo, may of a non-trivial ct2 J -foo). we be constructed; r0+? J e-\ We B~\ shall indicate in passing remark 0 ... f(t1J2)(t2-tlr*dt1dt2= that how certain there are (5.6) also of solutions periodic (5.6) solutions non-periodic kind. Let on the 45? line, constituting is a right triangle with hypotenuse the upper Ag = to at of a unit and sides with square Sg centered parallel (6,0), (tvt2) = 0 Hold and in the axes. define a Borel-measurable 60 fixed for the moment, Ag = function/ t2) such that the condition (5.6) is satisfied for 6 60, while =f(tv so that left half 0< ? f We the next extend the definition f2(t2?Qn-2dtxdt2<oo. to the of / square by Sg defining / in symmetrically lower right half, /(Mi)=/Ma)> and we then extend the definition of/ to the strip by the periodicity 0<?2?tx<l condition (v,/* Rt1+v,t2+p)=f(tvt2) With the aid of a figure is satisfied showing the triangles estimates A0 0,?l, and ?2,...). A? it is easy to see that (5.6) for all 6, while Ee(V>) We = = Eeo(V*)<ao. can now prove that the only are constants. Suppose f(Tv estimable T2) isin^Tx functions with so that minimum iff(tv variance t2) satisfies (5.6) and E9[f(TvT2)]2<\v(de<*), so does ... the product It will suffice to show f(tvt2)f(tx,t2). is a constant (a.e. L), where L denotes measure; Lebesgue be the same for all Let 0Q be any value of 6, henceforth Ag. (5.7) that in every the constant held fixed, 326 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions \?r(tv t2) Ag, must then let A+ be the COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART part of that it is impossible where Ago and ^>0, that L{A+) define f(tL, t2) in A? let A_ of the part be be and L{AJ) both A0 where i?r<0. If positive. We I shall show we they were, could from = Kkh){h~tiY-2 llL(A+)inA+, ? I ... i -\/L(A_)mA_, I 0 elsewhere ^ in (5.8) A0O. = Since / satisfies (5.6) for 0 above. described the method to an / statisfying 0O, it can be extended or have If A + A_ every points within and the condition is not bounded in this way (5.6) for all 0 by of the e-distance (5.7) may not be are if As and if L(A+) and L(A_) + and Ae__ are the positive, then for sufficiently small e, L(AS+) the strip 0<1?2?tx<e, jjarts of ^4+ and A_ outside now If the above definition modified and L(Ae _ ) will also be positive. of/ is by replac and __ satisfies all 0. But in and A (5.6) (5.7)for (5.8)by Ae+&nd Ae_, then/ ing A + while in the rest of then ^/cannot (5.6) for 0=0O for in A*+ and Ae__, ^/>0, satisfy 45? line, the / defined satisfied. However, Ae , \?rf= 0, and thus mz-hY^dt^?. f ?A We now have where (a.e. $?x) is in V^1, then the sets in T2) f((Tv both where have positive i?r(tv t2) < 0, cannot f2) > 0, and constant for every But if F is in V^1 so is V?c c, and so it ^(?1? measure. Lebesgue follows that for every Aeo if F = that shown c the sets in AQ where i/r>c and where i?r<c cannot both From this it can be shown that for some constant c, i/f=c(eL.e. L) positive measure. = c Thus then the V is in if F and this proves that V?1 in (a.e. $x ). only estimable A0O, are the constants. estimates variance with minimum functions have 6. CONSTRUCTION AND OF EXISTENCE MINIMAL SUFFICIENT to impose any restrictions necessary in this section we shall assume However, So far it has not been p9(x) integrable (jp, STATISTICS on the ?J9X family that the measures measures. of probability is in with respect to some measure continuous in J?x are all absolutely ?ix on Jp which union of sets in jp of that Wx is a countable of 0 and has the property dependent existence for the all 0 in ?>of a to is equivalent finite measure assuming ?ix. This function ?ix) such that for all A in jp Pex(A) =^p0(x)d?ix. We shall refer to this situation by saying with respect to ?ix (it is the "Nikodym pQ(x) includes of course the two This situation a generalized probability density mentioned in section derivative'' dP0x/d/ix 3). cases with fixed sample size important known there exists 327 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Vol. 10 ] [ Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS as the continuous on a Euclidean case, where px is Lebesgue measure space Wx, and the discrete5 of the sample point X are included case, where for all 6 the possible positions in a fixed countable set {x1} (i = 1,2,3,...) In the latter case we may of 6. independent take for /?x (A ) the number of points x'1 in A, and for the probability that X = x{ pg(xl), when the probability The generalized distribution of the sample is density pg(x) Pgx. is not uniquely and two determined determinations for the same by Pgx jux; however, 6 are (a.e. equal ?jlx a.e. and Pgx). For most of the families there exists considered statisticians S?x ordinarily by a generalized or "most some to and the with if respect '"simplest" density pg(x) ?ix, for 5PXcan be natural" is used, a minimal sufficient of statistic determination pg(x) an operation to found by applying to the family p = below. be described # [pg(x ){dea} on which determination of the does depend the result of this operation Unfortunately can which This introduces how certain measure-theoretic difficulties, family p is used. ever be surmounted and by applying as we shall it to some show examples later. We of families the operation # begin by defining of distributions of some statistical interest. these f(x ) on Wx, and suppose functions on values in The result A,i={fg(x)\6eA}. of the operation For of Wx, to be denoted # # on f is a ?ecomposition any point ((). by x? in 1FX the element D of & (t ) containing is defined as the set of all x?, written D(x?), on 6, and such not depending there exists a function k(x, x?) ^0, points x for which = we may all 6 in A. Roughly say that D(x?) k(x, x?)f9(x?)?oT speaking, tha,tf9(x) of 6. We note that if consists of all x for which the ratio (a:0) is independent fg(x)\fg x' is in D(x?), then x? is in D(x')\ also, that Let f denote are indexed a family of real-valued 6 taking by a subscript D? = {x\f9(x) functions = 0 for all 6eA} of # (f). It may be shown (as in the proof of measurable number of a countable (Jfx ) functions is an element consists of Theorem then (6.1)) that if f a measurable is #(*i) (Jp) decomposition. consider Example binomial 6.1: population now five examples = is ..., Xn) Suppose X (XX,X2, on with 6, X{ taking parameter 6 and l?d. Let /ix assign measure a the random values case In each the operation of applying #. as of Borel the taken a and be Wx is Euclidean Jfx may n-space, family on as measure be taken will but Example Jp. 6.1, /.x Lebsgue We sets. sample 1 and In all a from 0 with 1 to each of the 2n points probabilities respective zero or 1, and measure of the points in the set W+x consisting (xx, x2,...,xn) with x{=0 take Jp as the family of all subsets of W+x, In this example we might to Wx? W+x. 6 In ^ 9 _4 rs present for discrete the general case, while can of course be non-denumerable. \J0A0 each The of X values $ the possible case of non-denumerable constitute A treatment. 32& This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions is not a countable included set in the I COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART or the family of all subsets of Wx, instead Interval O<0<1, For ? we take the open j^O For D? this left member ?(x, ? a. = i-l and tinct this conclusion x? not (6.1) x is in D(x?) and hence a:0)^(x0), * =jc(XyXo^ i if and in D?, and only if only ... (e.2) if ... E ^?, (6.3) i-1 if (6.2 )were required 0 and 1. The between even be valid would in Wx. ... if xeW+x, of 0 if is independent of (6.2) L i [0/(1?0)] The p0(x) sets otherwise. For Wx?W+X. = such that p0(x) specification, there exists Jc(x, x?) for ! j-i = family of all Borel the determination the usual fn 77 0^(1-0)1^ = p(x) of and to hold only for two dis of 0 instead of all values resulting decomposition of (6.1) may thus be des the application of the operation & to the family {p0(x)} : The decomposition cribed as follows is equivalent (J?x ) in the strong sense to that of this. For anyx?in the statistic with function associated or 1:1 any 2-Xj W+* the values from element D(x?) consists of the n-\-2 elements D, contains The same decomposition of two {Pex) consisting points (s^0/ namely (6.3). The decomposition n sets D where 2 xi=v(v=0, 1,..., n). i=l in W + satisfying D? and the-%4-1 is obtained if the operation elements. or more # is applied to any subset of 6.2: Let X be a random with from a normal Example population sample mean 5.1. With the determination (5.1 )of the proba 02 as in Example 0xand variance we of the fraction is the find D? the Since denominator bility density p0(x ) empty set. cannot vanish, D(x?) of 0. This is the set where this ratio is independent Pq{x)Ip0(x?) is seer. 13 be the same as the set where is independent of 0, namely, the set where 2^i The tic decomposition (2X., 2 X{2); i = 2^? i and 2^? = 2^?. t $> is thus that associated by the operation is the same as that associated with the statistic induced this i with the (TX,T2) statis defined in Example 5.1. It may be verified the same result if in this case that one obtains the operation to a set of any of # is applied three pf the members {p0(x )}, say for ? = * are not the three collinear. 1,2,3, providing points (0?5 02i)> (0V 02) (0^, 0?i) 329 4 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions 6.3: Suppose X is a random Example with median -?oo<0< 6, and <t> ?{61 +co}. bution [ Part 4 SANKHY?: THE INDIAN JOURNAL OF STATISTICS Vol. 10 ] of n from a Cauchy sample the usual We take for pg(x ) distri deter mination = T.-" n p$(x) is the set where Again D? is empty and D(x?) This be written ratio may n it .?1 the ratio is independent pg(x)lpg(x0) of 6. [(o-xf-iKO-xf+i] n [(g-X}-i)(0-Xi+i)] i= If this quotient is independent ^?1. must have and denominator in 6 in the numerator where -0)2]-i. [l+(a- of 6, the polynomials of degree 2n we conclude equal roots, from which of (x\,x2?,...,xnQ). be a permutation The ..., xn) must decomposition (xvx2, as that associated with the set of ??order statistics" of induced by ? may be regarded a is of where the sample, rearrangement namely Z1^iZ2^i...^tZn (Zv Z2,...,Zn), same be for of found subset The result would any ...,Xn. Xx, X2, {p9(x)} corresponding to 2n+l values of d. distinct that 6.4: As in Examples 4.1 and Example on (d?\, 6+\), distribution the uniform 1 if With for all i, as 0 otherwise: |#i?d\ <\ n from as see that pg(x) = 0 otherwise. pg(x) i a random of sample determine pg(x) = max x{, we x-v t2(x) We i 1 if 0-i<?1(a:)<?2(^)<0+i ? 5.3 let X be ? oo<?<oo. = min tx(x) ... (6.4) Therefore D? = {x\t2(x)-tx(x)^l}. vanish for the same for any x? and x, x is in D(x?) if and only if pg(x) and pg(x? ) < if and only if (6.4) is that for x? not in _D?, a; is in D(x?) set of 6 in >. It follows w \ true for all 6 in which satisfy Now x is in D(x?) if and only if i. (x) = t. (x?), j == 1, 2. ^ is thus equivalent The decomposition imposed by the operation (3PX) in the strong 4.1 same and of 5.3. The sense to that associated with the statistic Examples (_T1?T2) a to if the operation subset of denumerable is obtained ft is applied decomposition It follows {p?x)} that for a;0 not in D?, to a subset corresponding of _. everywhere dense in c.. and independent Zn) (Z1,Z2,...,Xm) (Zm+1,..., means normal and with from and variances random samples populations 6X dx+S, of the Behrens-Fisher The null hypothesis then problem 62 and ?3, respectively^ == ? a so is for known that where the the d oonstant, <PXof S0, S0 family specifies Example 6.5: Denote by 330 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions I COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION-PART distributions probability Euclidean space under satisfying the null hypothesis ~oo<01<+oo, 02>O, of a three-dimensional is the subset If we determine 03>O. the density as ^0*0 1 it is easy to show that of The rrieas?fe-theoretie a minimal statistic sufficient n & to this three-parameter of applying the operation the statistic with associated is the decomposition the result functions density four real (Tx, jP2, T3-,TA) with the two sample variances. family i m components of the consisting two sample ? the operation in applying difficulty sufficient If a minimal is the following: means and to construct statistic for the sense. If family $JX exists then it is unique up to equivalence ($x) in the strong to the with a measures respect 3$x of probability density possesses generalized pe(x) for fixed 0 differ of two determinations measure p0(x) /?x, then p0(x) is not unique; ? = is if 0. Now on a set is in Jp and for which which non-denumerably /ix(A0) A0 wmay not be a null set for ?ax\ it need not even be mea 0 in union of the for infinite, A0 of the operation from the application The decomposition surable (jp). #(p) resulting determina on which densities # to the family of generalized depends p={pg(x)[de<?} that it can be shown 6.1) tion is chosen for p. As long as w is countable (Theorem a sufficient minimal with is associated the decomposition (Jp), and #(p)is measurable from seen that the two decompositions it is easily #(p) resulting sense. the in and of p are equivalent strong determinations two different 3?x) (?ix of p could <*> determination the of choice a is not if countable, However, pathological of two the is not measurable equivalence ? (p)that lead to a decomposition (Jfx), and This of p could not be proved. determinations two different from ?(p) resulting to the lead ? need not necessarily of the operation that the application shows statistic and for ^x, sufficient minimal This statistic. difficulty is resolved as follows: We restrict to families ourselves $x for sense to be in a certain is separable densities {pg[x)}of- generalized There a is Euclidean space. this separability defined below, always holds when Wx the operation in the whole set, and we apply dense subset of is then a countable {ve{x)} we noted that in every subset. <&to this countable (Recall that in the above examples 0 to a suitable countable the operation case the same result was obtained by applying below it follows that the result On the basis of the theorems subset of the densities). is a measurable subset of the densities of applying the operation # to the countable for the whole statistic sufficient is associated with a minimal which (Jp) decomposition which the family set J?x of distributions. Theoiem jp, possessing the operation ? 6.1: The Suppose result needed 3?x a generalized probability to a particular is applied for countable subsets is on measures of probability to ?ix. If with respect density function p0(x) the determination resulting of the family {p9(x)}9 is a countable set 331 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Vol. SANKHY? :THE INDIAN JOURNAL OF STATISTICS 10] [Part 4 a minimal is measurable is associated with decomposition (Jp) and sufficient the statistic for 3PX. The decompositions operation ft to two differ resulting from applying are equivalent in the strong sense. ent determinations of the family (/ix and ^x) {p9(x)} The (1949) of this theorem proof a theorem generalizing to be a sufficient statistic generalized with density of Neyman = _>,(*) for a suitable determination (Jp) and h(x) is integrable on and Savage this generalized pg(x) of (Jp, p*). and {fi(x)}, ft(l) - (6-5) .,(*(*))*(*) 6.1 let us denote prove Theorem some determination (?, and hyf^x) particular ?= of Halmos (1935): To .... Write result6 condition A necessary and for the statistic T=t(X) sufficient = a a measures set |?x of probability for {Pgx \de<*}possessing such that and h(x) respect to ?ix is that there exist functions gg(t) 6.2: Theorem rests on the following density, where gg(t(x)) of the by 6{ the elements of the generalized density^ for the decomposition generated by is measurable countable i= i (x), the operation set 1,2, ft f. If Df = then ft (?) is measurable (Jp). {x\fi(x)=0}9 = (Jp) since f{(x) is measurable Di? is measurable (Jp); hence D? C\iDi? ^smeasurable now a that x? i for which is not smallest in exists there _D?. Then (Jp). Suppose Define Di (x?) as the set where fi(x)^0 and say I=I(x?). fi(x?)=?0, shall We show first that fi(x)\fi(x)=fi(xO)\f1(x% Since f{(x) and/i Wx where fx(x) ^0, in Wx? Dj?, the part of (x) are measurable (Jp) so is their quotient and hence the part D{(x?) of Wx?D-? where the quotient has a cons = tant value is measurable is measurable If (]i A(x?) (Jp). D(x?) (Jp). Therefore for any point x? in Wx, D(x?) denotes is of ft(t) containing the element x?, it easily verified that for x? not in5?, D(x?) = D(x?), and for x? in J9?, D(x?) =D?. This completes the proof that ft(t) is measurable (Jp). Let be another of the generalized then determination densities; f'=-{/'.(a;)} is the set where f\ is also is in Hence N=ViNi Jfx and /^(.ZVJ^O. Ni (x)^f?x), a null set for so are ft (t) and ft (V). This t and i' are identical hence ?ix. On Wx?N, if N{ proves the strong Next we equivalence (ux and 3PX) of the two decompositions. shall prove contains if D(x) is the element of ft (i) which x, and that, we if T?t(X) is a statistic could take with v ft(t),?in associated particular t(x)=D(x) ?then 6.2. To this end we be factored in the form (6.5) of Theorem f{(x) may choose in each element D of #(f) a single point x?=x(D)If the set of points x? thus a function chosen is denoted x?=i?r(x) from TFX into TF?, by W?, this choice determines 6We paper exactly are containing the form to Professors indebted this stated result here; Halmos and long before the The publication. connection is discussed Savage reader for will giving not find in the Appendix us the a manuscript copy of their 6.2 there in above Theorem to the present 333 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions paper, COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION-PART I since x in Wx determines D=D(x) in # (f), and D determines x?=x(D)=X(I){x))=?f(x)> Take any D in #(f) and hold it fixed. Then for all x inD and x?=x(D)> where i= and k(x,x?)^=0, This 1,2,... may be written fi(x)^k(x,xlr(x))fi(x(D)^ For a in WX~D? define define of Define h(x)=0. now We have /(#). ?(x)=Jfc(a;, f(x))^0, where 9i(t)=fi(x(D)), for x in D? and D the elements of # - fi{*)=m*))M*)> (t) are the contours (?.?) and {a:|A(x)=0}=D0. 6.2 except that we do not form (6.5) of Theorem (6.6) is of the factored so as to satisfy can be determined the measurability the factors know yet whether this let us denote in proving For brevity conditions and integrability of the theorem. is a union of elements D of # (f); the by <?t the family of sets in WX each of which Now of members <gt need not GM be in jp. is a sequence {GJ show that all Oi are shall now K{ = {x\\{x) = = x\fx(x)=f2(x) It is clear that i =1,2,..., For ...=fi_x(x)=0> in<?t. a; in Gk(k = this with (6.6), we (H*-*,) ...n^i-in 1,2,...) A(z) Combining get we = have from (6.6) ... gk(Kx))h(x)^o. for x in 6?k and ?= <?? (i = 1,2,...) - define *(*) = 2-*A(*). - (6-9) and in i((?k) tbfine ft(0 (6.7) 1, 2,..., /i(*) = [U^WS^W^Y In and we U iGi = Wx ?D?, = so Hx is in <St. Let 0}, = 0 = } H{ U D?, so Jft is also in <?t. Finally, 0}={x\gtf(x))h(x) For /i(*)#0}. of disjoint sets in Jfx with in <gt: Let Hi = {x\g^x)) oriin?in is also define = 2fi(*)IW*). 333 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions (6-8) Vol. [Part 4 SANKHY? :THE INDIAN JOURNAL OF STATISTICS 10] and gt(t) noting from (6.7) that gk(t)^0 in t(Ck). This defines h(x) in Wx-D? for t(Wx-D?) ?= in D? 1,2,...; in define ... h(x)=0, (6.10) in t(D?), g^t) = 0. To see that there is no inconsistency in our definition of g^t) it is important to note that the sets t(D?) and t(Gk)(k = are This all 1,2,...) disjoint: is a consequence of D? and Gx,02,... sets in <6t. From being disjoint (6.8) we have and = /_(*) ... g?t(x))h(x). (6.9) and (6.10) we see that h(x) is measurable sets 6?k and D?, and since these form a countable disjoint in Jfx, h(x ) is measurable That it is also integrable (Jp ) in Wx. (6.11) From = |A(*)|?/** J TFX J[ f with We T = To denominator non-vanishing now have it is possible tf(x) is contained If T' the generalized that to find a null in a contour is sufficient g^xj) density, if T' = 6.2 to be satisfied set N" to prove and hence it is minimal. for 3?x, statistic t'(X ) is any other sufficient of contour for $?x such that in Wx?N" every say/.'(a;), then by Theorem such that for i = (Jfx)- The = 1, 2, a determination exists 6.2 there of ..., gi'(t'(*)W(x), V ? determination {/'i(~)} may ? = previous determination? {f^x)} to the extent that if JV?' iV/is W'-N''. that (Jp) (jfx) in TFX. is measurable oft(x). for $x, is measurable h'(x) in Wx?D?, all the hypotheses of Theorem statistic for iPx; it remains only /i'(*) where Gk shown a sufficient .(X)is do this we must show then (Jfx,/?x ) follows by sets from in J_)?while from (6.11) g{ (t(x)) is the quotient of measurable Since g{(t(x))=0 functions of Wx covering the = %2-*< oo . fk(x)dp* S 2-k f > of each h(x)dv?=t kn J\ fk(x)dp JF-Z)0 4 in (Jp) set for ?f. Hence N' = Ui-^Yisanullsetfor/i* Let N0 = {x\h'(x) = 0}. Then for all ?, /,'(?) a null from 3?x, and/?' the then {^{fiix)^^)}, and = differ =/j on 0 on N0, J N i 0 and so _V0 is null WV^/i'i*) =f,(x) We shall now set for $x. LJ^" = N' U N0,N" is a null set for $x, and on and *'(*) ^0. show that in a con every contour of ?'(a;) is contained of and let A(x?) be the part in Fx-r'iV in Wx?N", in Wx?N" tour of t(x). Let x? be any point the contour of t'(x) containing x?, = A(x?) {x\xeWx-Nn, t'(x) = t'(x?)}. 334 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART I For all x in A (x?) = m=fif(x) 9iW))h'(x); in particular, fi(x?) = 9iV(x?W(xQ). Since h'(x?) ^0, f{(x) for all x in A(x?), = from containing x?. which -#0. is contained in D(x?), A(x?) (6.12) that in Wx?N", This the Theorem of 6.1. proof completes In proving an existence may be non-denumerably whose for minimal theorem we infinite, shall sufficient employ the statistics contour of t(x) for families the following two 3?x lemmas is obvious. proof Lemma 6.1: If the statistic T is sufficient for measures (6.12) where k(x,x?) ^h!(x)\h'(x?) It follows ... k(x,xQ)f^) it is sufficient for any 3?x. of subfamily thefamily 3$x of probability 3PX, if it is a minimal for the family If T is a sufficient statistic null set for 3?x is a null set 3$xx of 3Px, and if every sufficient statistic for the subfamily for 3?x, then T is a minimal sufficient statistic for J?\ 6.2: Lemma Our sarily of the existence extension case countable theorem from the countable to the not neces of a function of separability space with = one. and of order <7?g(x) are two f(x) It/ as (jp, /?x), we define their distance integrable the notion involves in the mean pect to convergence on Wx, valued functions *,(/,?) = res real [ *\f{x-g(xWx. Jw each integrable f= a family of real-valued functions, {f0(x)\0eu>}, a of that for every subset such countable f fixed tx (Jp,/?x), is separable (?ix ) if there exists in f there is a sequence {g{ \i= 1,2...} of functions S? (/,?i)->0 as ?->oo. g{ in tx for which fg is dense subset tx of this definition to say that the countable It will be convenient We shall that say remark that if Wx is a Euclidean space and Jp is the family of Borel sets (?ix) in t. We to ?ix is separable: densities with respect in Wx then any family {p0(x)} of generalized all of functions of the separability7 is a consequence This integrable (?ix) of the family a a in of metric of the and family space. any subfamily separable separability (jp, ?ax), the operation # in each of Examples that the result of applying We can now conclude 6.1 to 6.5 was just made to give about Euclidean obtained being composition of the generalized i This (1948, Chelsea, may a minimal densities, be proved sufficient statistic: This follows from (i) the remark the same de about in the examples, (ii) the remarks to is certain countable # if the operation applied theorem. and (Hi) the following Wx, from the Approximation Theorem on page 4 of Ergodentheorie N.Y.). 335 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions subsets by Hopf Vol. 10] SANKHY?: THE INDIAN JOURNAL OF STATISTICS 6.3: Theorem the family i/ $x measures of probability p= a generalized density p0(x) with respect to ux, and if thefamily [Part 4 on Wx possesses of densities {p0(x) |0e<*>} is separable (jax), then there exists a minimal sufficient statistic for* 3?x, and it may tructed by applying the operation & to any countable set px dense, (?ix) in p. be cons let tx = {f{(x)}(i = 1,2,...) be a particular determination a of countable dense (?ix) in p, and let 3?xx be the countable px of densities subfamily of to 6.2 we must 3$x corresponding px. In order to apply Lemma subfamily x is a null that if N set for $xx show it is a null set for $x. Let be any P$ a sequence and let member of $x, let f0(x) = in tx for which {grj be dPgox?djix, To prove the theorem 0 as ?-> ^/?(/o^i) -> oo. We have J?N for all i. = 0 9i(x)dux Now II = L I| ?*N h^x- I??N gid/i*!I f?JNMx)dj*= Ij VfN t f^x I .< [J N \U-9i\dj? (f9-gMr? I I*J( N Hence = 0 f /?(*)i/?x J N and N the ing remains is a null set for 3$x. Suppose T=t(X) of dent sufficient # to tx. To prove T operation to prove that T is sufficient only a real-valued exists there that is a minimal and 0, such function statistic for is a minimal $3f, obtained sufficient statistic for T P(A\t), 3?\ That measurable is sufficient (Jp) for fixed by apply for *$x it 'or means 3?x A, indepen that P*{At\t-*{B)) = ... ] t P(A\t)dP? (6.13) to 3?xx. in :Jp, B in Jf*, and 0 in av where &v is the subset of w corresponding to show that if (6.13) is valid for all 0 in u>v for J?x it thus suffices To prov? T sufficient of t~\B). then it is valid for all 0 in c*. Let us write B' as an abbreviation for all A We may transform (6.13) to ( J AAB' Where n(A, x) = P(A\t(x))9 " ? JB' tdP^=?n{A,x)dP^ or ^^ ? Ane' We note t><w(ii,a;)<l (a.e. $x*) by = ?/?X ^ f b' (2.5) and xWx)dPx- (2.8), and hence - (a.e. $*). 336 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions (6-14) Let 0? 0? COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART = be any 6 in -, write as i -> oo. in fx for which (">(?>g^-^O From f ... fJ b' *?,_>* gidf= J aab' we let {gr?}be a sequence g(x), and i>0o(~) I (6.15) shall prove =f gdftx J f J A/IB' i?>oo and by letting are the corresponding that showing the of members limits ... ngd/f B' of the and right (6.16) of left members (6.15) (6.16). \gi-g\dn* ,gidp*-J\ AAB' gdfiA^ J[ AilB' | [ I J AAB' <S_-(0i>0)->O. < This completes [ , \9i-9\fox<*?(gi,9)->0. J B of Theorem the proof 6.3. APPENDIX With torization of which only notational changes, on result the sufficient and statistics the and Savage stated by Halmos densities of generalized probability their a is different is 6.2 above Theorem version, following slightly fac (1949), additive 1 : Let <?l be countably Jp and families of sets in the Corollary = t t(x) be a function from Wx onto Wt such that for spaces Wx and W{, respectively. Let be a family Let (#) every set B in <fe\ tr1 (B) is in Jp, and let <?Y ^t'1 Mx={Mgx\de?>} to with the continuous on Jp with all measure respect finite measure of finite Mgx absolutely to be a sufficient statis condition for T=t(X) and sufficient on Jp. Then a necessary ?ix = be in the form tic8 for ?lx is that for every 6 in <*>, factorable fg(x) fg =gg h, dMgxjdpx are and and is measurable where 0^gg integrable (Jfx, px), h(x)=Q (f?3^), 0-^A 0^ggh on every null set for 0lx. (a.e. /.x) we have obtained arise from the restrictions of the simplifications take for (gt the family JP of all sets in W* whose pre-images always Some first, we 8 If Jf?lx Savage define these definitions the only case is n?t statistic, here considered a family of probability conditional except by probability, to remark us?the that definitions T measures, in the is not statistic sufficient and case agree that 0[x with is a a random variable. for this case. We family of ours. 337 5 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions need probability that, are Halmos not in and consider measures? Vol. 10] $*> and that, secondly, we 6.2 follows from Corollary assume #ftx to be a family of probability six steps: of the following measures. Theorem 1 by means that t(x) satisfies assumption in <6* are in Jfx is automatically for fulfilled = take <gt Jp as just mentioned. always Io. [ Part 4 SANKHYA: THE INDIAN JOURNAL OF STATISTICS the The condition an arbitrary of sets pre-images function t(x) because we that out by Halmos and Savage that in their work the assumption 2?. It is pointed on /?x that that Wx is a countable ?ix(Wx) is finite may be replaced by the assumption union of sets of finite measure (/?x). 3?. The is non-negative and integrable that assumption (Jp, /p) may ggh of our assumption is a be dropped because of probability that ?lx measures, family = so that and be assumed non-negative ggh fg may = 1, \wx9ehdpx on of non-negativeness since we be dropped, gg and h may as h and then and and redefine may \gf?\ gg |h| without fg non-negative or the integrability the product of of h. the measurability affecting gg, ggh, to the assump 5?. The assumption that ge(af) ismeasurable (<?xlt) is equivalent 2 in the same ismeasurable where tion that it is of the form (<?*) by Lemma gg(t) g9(t(x)), our choice of <?*= Jp, is measurable and Savage. But with paper by Halmos gg(t) 4?. The condition assume (45*)if and only if g9(t(x)) is measurable (Jp). the assumption h(x)=0 (a.e. ?ix) on every null set for $lx, proof condi a little more effort. If h satisfies all the remaining be dropped may requires all the also satisfies tions of Corollary 1 we shall show that h can be defined which =' 0 and furthermore conditions (a.e. /?x) on every null set for fflx. remaining h(x) 6?. The that = for which /e>0. a particular of determination g h denote dMgx?dfix Let/e a countable subset 0lxx of 0LX, 7 of Halmos Lemma there exists B and Savage, = = are null sets for ?fft*. sets for all null such that 0XX say 0li dx, #2,...}, {Mgx\6 = 0 for all i. Then Denote _40 is a null set for fflxx by _4? the set in Jp where fg, (x) and hence for $lx. Define f h(x) for x in Wx? ^ 0 for a; in _40. _40, For all A in Jp, = f 9ehdlf= M A A) = Mgx(A-A0)Ja J[ a-ao g9hd,ix = f gfi?p= ( 9$df> J A-Ao J = greA determination and. hence as a bility A = instead thus use g h (a.e. /.x). We may fg ggh^0 the condition and ? obviously satisfies of dMx\dpx, (Jp, ?ix) since h did. 338 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions of of /? = ggh integra I COMPLETENESS, SIMILAR REGIONS AND UNBIASED ESTIMATION?PART It remains to prove only Let N' be the part (a.e. jix) on N. Now if N that of N where 0 = M g X(N')= Let AT. be the of part N' where = 0. X *# 0, and thus If ^. = Ofor all ?. Hence N'?N" ggi 0 on N', and ?y= therefore /?x(Ar/) so W-N" to prove = 0 = 0. h(x) ?ix (N') g0 hd/ix. so geh>0, is contained On 0. = ?ix(N") in A0, set. = /^(JVj) then is the empty for j?lx then set null 0; we need h^ f JV" = U^, a is and 0, ? = = 2V'? Ari, 0, gr^? on JV' and iV", But 0 on N'?N". Since N' contains unbiased sequential = N' N", and #", 0. References David Blackwell, Math. J. Doob, M.N. sampling Paul Halmos, P. P. of B. Soc, of "Application Annals of Math. the the theorem distributions a admitting theory 32, 62-69. der Mathematik Ergebnisse New Chelsea, by to the Biometrika, standpoint", power-function der Wahrscheinlichkeitsrechnung. Reprinted \ 7, 34-43. Stat., Radon-Nikodym binomial certain 225-241. 20, Stat., from Stat., 17, 13-23. of Math. for estimates "Unbiased (1946): estimation. of Math. 329-338. 8, Sankhy?, York (1946). Trans. statistic. sufficient Math. Amer. Soc, and L. H. Scheff?, of tests On optimum 18, 473-494. (1947): On (1947): the of problem one with hypotheses composite similar Annals constraint. Broc. regions. Acad. Nat. of Sei., 382-386. 33, E. Lehmann, tions. J. bility. J. mass and L. (1941): On and Bull. a statistical E. Information Math. (1917): Minimum 280-283. S. Trans. of composite estimation series A, Soc. accuracy 37, no. and the On (1933): and variance of statistical London, Roy. Soc, tests Distribu I. Normal hypotheses, London, problem series in attainable on based the classical of proba theory 333-380. 236, in routine arising problem 12, 46-76. Stat., of Math. Annals Pearson, * Phil. (1945): Calcutta Roy. Soc. powerful 19, 495-516. Stat., of a theory Outline Trans. Most (1948): of Math. (1937): Phil. hypotheses. C. R. C. Stein, Annals production. J. 43, (1949): Grundbegriffe 2, no. 3 On (1936): Stat., E. C. R. of unbiased of variance (1933): O. E. L. Lehmann, JIao, Math. Amer? Trans. probabilities. 399-409. Math. Neyman, J. statistics," Grenzgebiete, L.J. Annals applications," Annals "Analysis ihrer 39, Lehmann, transition regions. Savage, theory L. Savage, A. Koopman, Rao, The (1941): und Neyman, (1946): and Kolmogoroff, Neyman, problems R. of similar and P., with sufficient P. L. the problems On A., Mosteller, Halmos, Hsu, of Princeton. of Statistics, of Markoff properties Asymptotic (1948): M. Girshick, Annals estimation. 393-421. 63, Ghosh, Methods Mathematical (1946): (1948): and expectation 105-110. 18, Harald Cramer, Conditional (1947): Stat., and analyses of A, the the 231, most in sampling tests efficient of inspections of statistical 289-337. estimation of statistical parameters. 3. 81-91. the estimation of several parameters. Proc. 339 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions Camb. Phil. Soc, Vol. 10 ] Rao, SANKHY?: THE INDIAN JOURNAL OF STATISTICS C. R. L. Savage, J. Stat. Henry Scheff?, of Math. of Math (1949) Stat., Abraham 15, Wolfowitz, theory On a measure (1943): Math. the of estimates. variance minimum for unbiased bionomial sequential constraint. Annals problem arising in the theory of tests. non-parametric Annals 14, 227-233. abstract: 20, (1947): (1946): 45, Annals Certain of properties the multiparameter unbiased estimates. Annals 142. Foundations of a general theory of sequential decision functions. Econometrica 279-313. J. Soc, estimation. one with hypotheses composite testing PhU. Proc.Camb. 280-293. 13, Stat., G. R. Wald, On (1942): and theorem uniqueness 295-297. 18, Stat., Henry Scheff?, statistics A (1947): of Math. Seth, Sufficient (1949): 213-218. [Part 4 On sequential binomial estimation," Annals of Math. Stat. 840 This content downloaded from 128.118.88.48 on Mon, 20 Oct 2014 12:14:17 PM All use subject to JSTOR Terms and Conditions 17, 489-493. of