Indian Statistical Institute
Transcription
Indian Statistical Institute
Indian Statistical Institute Asymptotic Theory for Some Families of Two-Sample Nonparametric Statistics Author(s): Lars Holst and J. S. Rao Source: Sankhyā: The Indian Journal of Statistics, Series A, Vol. 42, No. 1/2 (Apr., 1980), pp. 19-52 Published by: Indian Statistical Institute Stable URL: http://www.jstor.org/stable/25050211 Accessed: 18/06/2010 09:10 Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=indstatinst. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. 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 [email protected]. Indian Statistical Institute is collaborating with JSTOR to digitize, preserve and extend access to Sankhy: The Indian Journal of Statistics, Series A. http://www.jstor.org : The Indian Journal Sankhy? of Statistics 1 & 2, pp. 1980, Volume 42, Series A, Pts. 19-52. ASYMPTOTIC THEORY FOR SOME FAMILIES OF TWO-SAMPLE NONPARAMETRIC STATISTICS By LARS HOLST* and Uppsala Madison of Wisconsin, University Sweden University, and S. RAO J. University SUMMARY. two the null the ordered Let that hypothesis is studied natives, trie type of test 1,. test are not for the maximum has metric type can Wilcoxon-Mann-Whitney asymptotic test against "optimal" Connections with rank as well as under m 2 h(Sk), fc-1 form based is an example specific are 1. briefly On are methods from alternative, explored Introduction the here the United States interval and limiting and Army under only illustrative and of the those symme alter distinguish of symmetric tests, which the Dixon is shown that tests of the rate nonsym of n'112. investigating one to allow select standard After which tests among tests that can of alter sequence on $*'? hand, suggested some and that the two d.f.s. are the same. hypothesis transformation z->F(z) would permit carrying by a suitable other of the examples type the m 2 hj?(Sjc). *-=l provided. notations Xl9 Sponsored be in the theory at the more converging to this class. which belongs and Yl9 ..., Yn be independent ...,!,?_! continuous distribution two populations with functions to test if two these wish We populations respectively. Let test ... ^ X'~ falling symmetrically It is shown in {Sjc}. efficiency. alternatives any to on these numbers based hypothesis, {Sjt}. Let some functions condi satisfying simple regularity symmetric alternatives, wish < distribution from samples We LetX' of Y-observations asymptotic hypothesis such tests identical. line. the null real-valued relative asymptotic under for the random independent on the real in the that they performance from the hypothesis. this class Among run test and the Dixon it known test, distinguish test theory be , Yn USA sense the well instance includes ... respectively are populations of the statistics Barbara, O studies the null under Santa by Sjt the number statistics be Ylt and parent paper .. , m} have poor asymptotic at a "distance" of n"1^ test an This m 2 hjc(Sjc) which fc-1 of the form natives ,m. for F Denote theory Asymptotic two these and ,Xm-i functions X-observations. k = 1,... X'.), [X'v of families efficiencies . )k= . ) and h( {h*( tions. ... Xl9 distribution continuous of California, random samples from (d.f.s.) F(x) and G(y) are identical, i.e., the A simple probability us to assume that the Contract No, DAAG29-75-C-0024. integral support 20 LARS HOLST J. AND S. RAO of both the probability is the unit interval distributions [0, 1] and that the first of them is the uniform d.f. on [0, 1], For the purposes of this discussion, as can this probability be done of loss transformation without any generality soon. be apparent Thus from now on, we will assume that this reduction and that the first sample has been effected is from the uniform distribution = G o p-1 after the Let G* the the second denote of d.f. ?7(0, 1). sample The null hypothesis to be tested, transformation. probability specifies will = HQ:G*(y) 0 < Xi Let The sample ... < < Dk = = Dkm statistics , ? = 0 and 1. Tests based put X0 X'm in the literature been for the considered have See for instance Define k = for sample. by ... on these sample (1.2) spacings goodnesss-of-fit problems. and Rao and Sethuraman (1975). ..., m 1, Sk (1965) (1953), Pyke Darling from the first ...,m k=\, we where (1.1) are defined for the X-values X??X?_! ... 1. y< 1 be the order X'm_1 < (Dv ..., Dm) spacings 0< y, ? number of in the %'s interval ... [X^_l5 X'k). (1.3) .... is to study various test statistics based on these numbers {Sly Sm} be for testing H0. These called may quantities (since "spacing-frequencies" the frequencies of y's in the sample of the x's) or the spacings they denote Our aim to the gaps in the ranks of the #'s in they correspond as well as the statistics Since the numbers based sample). {Sjc} remain invariant under probability no there is loss of transformations, (since "rank-spaeings" the combined on them in making generality It may be remarked such a transformation here that we in the first vations We and aim will do mv Note more is to study this through assume {nx] and ?? oo, instead as was done of the usual earlier. m obser (m?1) this yields m numbers ..., instead of {Sv Sm} on notation. Tests based simpler {St} have been since sample (m+1), leading to slightly for the two-sample considered Rao (1976). Our on the data, take the asymptotic a nondecreasing throughout, nv ?> in Dixon problem oo and that sequence as v ?> oo, rv ?> Godambe (1961) as m and n tend theory = mv/nv (1940), p, of positive 0 < p < oo. and to infinity. integers ... {mv} (1.4) in (1.2) depend on mv the number of X-values and it is {Djc} defined to label them as {Dkv}. the numbers appropriate Similarly {$#} defined that TWO-SAMPLE NONPARAMETRIC 21 STATISTICS on both mv and nv and should therefore in (1.3) depend be denoted by {Skv}. = Thus we are dealing with variables random of 1, arrays triangular {Dky, k v the to v-th 1. ..., mv} and {Skv, k ? 1, ..., mv} for > 1) (v > Corresponding = let functions and real-valued k be ..., 1, array, satisfying Av( ) {A?v( ), rav} certain regularity conditions 2). Define (see Condition (A) of Section mv Z A*V(S*V) ?*v= ... (1.5) *=i and s r;= ... av(^v) fc-i (1.6) on the (mv~l) X-values and the nv 7-values. Though T* is a special case of Tv when on k, we will distinguish these two cases {hkv( )} do not depend since their asymptotic is quite different in the non-null situation. behaviour based It may be noted that here the Wald-Wolfowitz (1940) iun test and the Dixon are of the form test is of the Wilcoxon-Mann-Whitney (1940) T*v while in the combined the form Tv. In fact, any linear function based on the X-ranks can a as case be of also Section 5.) sample, (cf. expressed special Tv. test A few words the m Tv = law A ), {hk( A( v is suppressed suffix : Though the notations about as the functions as well except the quantities m, n, r, Dk, Sk on v, for notational convenience )} depend it is essential. where 2 hk(Sk), T*? fr=i a of random normal = S *=?i and r will h(Sk) stand for (m?n) etc. for instance, The probability (or random vector) X will be denoted by <?(X). mean covariance matrix S will be with ?i and variable distribution while N(0, 0) stands by N(/i, S) throughout zero. at the point For 0 < x < oo, p(x) with mean x and distribution represented distribution Poisson Thus m 7T^) = e-*.^/j!, j = 0,1,2,... for will ... the degenerate the represent (1.7) For p = (n, p) will denote (pv ...,pm), mult n distribution with trials and cell probabilities variable random (r.v.) with density exponential of j. probability multinomial the m-dimensional the Poisson (Pv >Pm)' A negative e~w for w ^ o and zero elsewhere (1.8) 22 LARS will be denoted HOLST AND J. RAO stand for an by W while {Wl7 W2, ...} will of such r.v.'s. The distributed (i.i.d.) sequence throughout and identically independent random variable tj will have a geometric distribution i Pfo=j)=p/(1+P)'+1, for 0 < S. = with p.d.f. ... 0, 1,2, ... (1.9) oo. p< is useful conditional between these distributions following relationship a as r.v. a Let Let r? denote above. later on. IF be exponential negative r.v. which Then the (unconditional) for given W = w, is 7?(w\p). distribution of 7j is The ? = P(v j) = = Err}(Wlp) f-W/P J (w?p)S -^r- c~" dw = p/(l+p)>+1, j = 0, 1, 2, ... ... same as (1.9) above. Thus t? has a geometric on W = w, it has a ^(w/p) distribution. distribution the for any random variable and we write Xn = in probability Also such that P{ XJg(n) | the denote We largest shall > \ Ke} < = Xn if for each Op(g(n)) a sequence consider we write = if conditional ?> 0 op(g(n)) if XJg(n) e > 0, there is a Ks < oo s for all n sufficiently large. Finally [x] will in x. contained integer GUV) Xn, (1.10) of alternatives specified 0< y < y+(Lm(y))lm*, by the d.f.'s ... 1 (1.11) = 0 and S > = . In terms of the J original d.f.'s F and Lm(l) Lm(0) = there is P, while under the alternatives C?, the null hypothesis specifies G as to F the size a sequence of d.f.s. increases. Guthat converge sample is of Indeed Lm( ) (1.11) given by where Lm(y) = m*{Gm{F-\y))-y). is a function ... (1.12) on (0, 1) to which Lm(y) converges. on Lm( For further conditions (B) and (B*). ) and L( ) refer to Assumptions a sense in certain and has been is smooth of alternatives This sequence (1.11) We assume considered that before. there See for instance L(y) Rao and Sethuraman (1975) or Hoist (1972). : In Section of this paper is as follows organization are established. 2.1 gives asymptotic Theorem results The nary 2, some prelimi distribution of TWO-SAMPLE ttONfARAMEtRlC of multinomial functions on limit the while frequencies of non-symmetric distributions n STATISTICS 2.2 Theorem is of which statistics, spacings in Theorem combined 3.1 to obtain independent It is S = |. the limit distribution of Ty under the alternatives (1.11) with distribution clear that putting Lm(y) == 0 in this theorem, gives the asymptotic an asymptotically test for of jTv under HQ. The problem of finding optimal These interest. are a result establishes results a given 3.2. in Theorem is considered Some specific of alternatives sequence are discussed 4 deals with the at the end of this section. Section examples of 4.1 gives statistics distribution the asymptotic T*v. Theorem symmetric 4.2 while 5= Theorem of with the under alternatives T\ \ sequence (1.1) to note tests. It is interesting can alternatives only distinguish symmetric 4 unlike the non-symmetric to the hypothesis at the slow rate of n converging at the more usual can discriminate alternatives which statistics converging _i 2 . on sample spacings depend rate of n Similar results hold for tests based See for instance or not one considers statistics. ing on whether symmetric finds the optimal that Rao and some test among the symmetric classes of test statistics T\ further remarks 2. The following as well functions of this results and Some preliminary 5 contains Section the limit which growth be needed will properties, functions real-valued (1980). results conditions regularity as supply smoothness the next section. (A) : The Condition and Hoist (1975) and Rao and discussion. Sethuraman )} defined {hk( on of the for the {0, 1, 2, ...} satisfy Condition (A) if they are of the form *=l,...,w hk(j)^h(kl(m+l),j), for some function = 0,1,2,. = 0< u < 1, j in u except if any, does not for finitely for defined h(u, j) ? (2.1) 0, 1, 2, ... with the many u the properties (i) h(u,j) is continuous set discontinuity depend and onj. (ii) h(u, j) is not of the form c >j+h(u) for some function A on [0, 1] and a (iii) For real c. number some d> 0, there exist IMn,j)| < cx [n(l-ii)]-i+S. for all 0< u < constants c1 and c2 such that (/2+l) 1 and j = 0, 1, 2, ... ... (2.2) 24 AND HOLST LARS S. J. RA? real-valued functions (A') ; The Condition (A') if they are of the form satisfy gic{x) for some ? function g(u, x) k= x), g(kl(m+l), 1, ..., m u < 0< for defined on {gjc(*)} defined Condition x< 0< and oo 0 <; x < 1 and [0, oo) oo with the u the properties, g(u, (i) in u is continuous x) set if any, discontinuity for finitely except on x, not depend does g(u, x) is not of the form c and a real number c, and (ii) for some (iii) ?> 0, there \g(u9x)\ < c^l-^f We 2.1 Lemma g on function [0, 1] that c2 such cx and and ?+^(a2+l) u < simple lemma, : Let h(u) defined for 0 < in absolute be bounded x < 1 and 0 < for all 0 < following u and many finitely Then the require for some x-\-g(u) constants exist many is stated which ... oo. without (2.3) proof. u < 1, be continuous except for value by an integrable function. i m 2 A(Jfc/(m+l))-> J h(u) du as ra-> oo. D (1/ra)*=i o (2-4) of Tv defined problem, we will obtain the distribution we the statistic consider in two steps. First in (1.5), essentially Tv for given Since the numbers D = {Dv ..., Dm}. values of the X-spacings {Sv ..., Sm} on a we multinomial a the need result multinomial distribution, given D have to the main Turning sums. for We formulate the given on the The expressions 2.1. this part of the result in Theorem of Tv distribution mean of this conditional and variance asymptotic D, are functions of D. limit distributions of functions for in particular, these expressions 3.1 of the next section Theorem lemmas It there, given is clear frequencies S = thus giving that (Sv the of spacings, the combines the required conditional ..., Sm) 2.2, we In Theorem given a general result us to handle allows formulate which mean and variance. asymptotic other these results along with distribution of the vector spacings of Ty. distribution asymptotic the vector D = of (Dv spacing is ..., Dm) TWO-?AMPLE N?NPARAMETRI? 25 STATISTICS on D, Therefore the test statistic ..., Dm). (n, Dv Ty9 conditional as same variable null the random the the distribution hypothesis, mult under Zv where (9^ variance variables, variables = ?iv taking r.v.'s Poisson Theorem be as defined 2.1 m S hk(?kv), (2.5) ... cr2= E(Av), ... var(Av). (2.7) the asymptotic of the multinomial distribution 2 of Hoist case of Theorem as a special (1979) by theorem. that in of of (?kv, hk(?kv)) place (Xky, Ykv) : Let in (2.5), (2.6) k=l be mult(n,pv (<pl5 ...,9^) (2.6) and (2.7). For 0 < q < Av,= thai there exists Assume ... ?T=1 on theorem following Zv can be derived sum mv 2 hk(9k) mean and the asymptotic is mult Since (n, Dv ...,Dm). ...,9^) random terms Poisson in of stated of Zv can be more simply a triangular we random Poisson introduce array of independent v 1 is and set where %kv P(nvpkv) {?lv, ..., gm v}, > Av= The = has a q0 < M 2 ...,pm) and Zv, /iv, andav = 1, set M [mq] and ... ?*(&). (2.8) fc=i 1 such that for M 2 pk->Pq9 q> qQ ... 0<PQ<1, k=l (2.9) and where AQ, BQ and Pq are such Aa -> Ax, TAew a?s j^?> BQ -> ^ 1-0, and Pa -> ... 1. (2.11) 00, ^-^/^-?^OMx-u?. A12-4 that as q-+ D ... (2.12) &6 From ?LARSHOLST AND j. (2.7) an explicit (2.6) and fi(nD) p?nD) ?7(0, 1). m ? S S ?*(j)*r,(nD*). ?=l Thus have we = pk Dk, consider ... (2.14) i=0 oo Wi k=i 2 A?(j) 7r?(a:/r). i=o on spacings have (1965) and Rao based (2.13) we hypothesis, = This is of the form 2 gjc{mDk)where (/?(a;) Statistics by i=0 the null (1.7). Under D are the spacings from = is given ... 2 hk{j) ttj (npjc) fc=l = for the mean expression /?v= 2 the notation using == k 1, ..., m where S. tlA? been considered earlier by Darling LeCam and Sethuraman (1958), Pyke (1975). Most = the discuss when case, i.e., papers, however, gk(x) g(x) symmetric As Pyke out (cf. Section could (1965) pointed 6.2), LeCam's method = to study case. the more Let general non-symmetric {gk{ ), k = 0 < q< be real-valued measurable functions. For 1, let Mv (1953), of these for all k. be used 1, ..., m} [mv q]. Define GQv= M 2 gk(Wk) ... (2.15) fc=i a sequence of i.i.d. Then the r.v.'s. exponential {Wv W2, ...} is states the of theorem distribution statistics asymptotic following explicitly of the type (2.14) and is easily established by checking Assumption (6.6) of where Pyke (1965). Theorem 2.2 : Assume 0< that var (GQv) == o-2(Gqv) < oo for all q and v, ... (2.16) and thatfor each q e (0, 1] / / 0 \ / At B9\\ {Gqv-EG9v)l<r{Gu) \ ... (2.17) Mm ]^N[V \( 0 /), \( Bg q I V S (F*-l)/m* J j / with Aq and Bq such that Aq-* Ax Ba -> Bx = 1 as tf->l-0 as q - 1-0. ... ... (2.18) (2.19) TWO-SAMPLE NONPARAMETRIC as Then, v-^ ? gk(mDk)-EGlv'jl<r(Glv)) ..., Z>) are spacings (Dv corollary gives a simple sufficient following holds. ) in order that the above Theorem : The 2.1 Corollary asymptotic ... 1?J3J). (2.20) 1). Q U(0, from -> N(0, The gk( 27 oo, ??(( wAere STATISTICS normality on the functions condition asserted 2.2 in Theorem holds for any set of functions {gk( )}which satisfy condition (A'). : To prove this Proof corollary, we need to check that the assumptions It can be easily checked condition (2.16) to (2.19) hold when (A') is satisfied. that if g(u, x) satisfies condition then (A'), 00 00 00 x)e~x dx as well S 9?U> x)e~x dx, J g\u, oo satisfy Lemma of Lemma conditions 2.1, as m E(Gqv)lm ?? = var (Gqx)jm= 2.1 in u. Thus as J g(u, x)(x? l)e~x dx o from the definition of Gqv and oo, [mq] q 2 Eg(kl(m+1), Wk) -> JEg(u, W) du (Urn) k=i o ... (2.21) [mq] q (1/m) 2 var (g(kl(m+l),Wk)) -> J (var g(u, W)) du o k=i ... (2.22) and cov (Gqv,2 Wk)\m = M [mq] 2 (1/m) l k=l ~> J cov o from (2.3) of condition Again in q so that (2.18) continuous Finally to check the cov (g(kl(m+l), Wk), Wk) (g(u, W), W) fc=i ... (2.23) (A'), all these limits are finite. and (2.19) are satisfied. asymptotic in (2.17) normality These {a{gk{Wk)-Egk{Wic))+{Wle-\)}= S Jc=i are or equivalently tm0l [ma] S du. g*k(Wk), say ... also of (2.24) LARS HOLST AND J. S. RAO 28 for all real a, we identical case. have only is easily It for the non the Lindeberg condition condition if {gk( )} satisfy (A'), so do to verify seen that {gl( )} defined in (2.24). Let = (Wk) and sfm] * 2 erf. k=l = (r? Eg\ ... (2.25) Since {gl( )} satisfy condition (A'), we have as in (2.22) that = afmgI/? to a finite converges non-zero (l/m)2 o [mq] oo k=i jg\(w)tr"?w o from Lemma constant ... 2.1. (2.26) Now consider [mQ] 2 j (i/^)s < 0;(*)6-*<fe . (mlsfmq])dim) 2 c1[(fc/(m+ i))(i_ t/(m+l)]-i+2' J (a;C2+l)2e-^da; S < {^/^?{(l/m) J { #=i [(?/(m+l))(l-i/(m+l))]-1+25} (xe*+l)*e-*dx). remain bounded because in the first two parentheses As m ?> oo, the quantities in third the 2.1 while the integral of (2.26) and Lemma goes to parenthesis Thus the Linde berg zero for any e> 0 since s[mq^ is of order (\/m) from (2.26). condition is satisfied 3. for (2.24) which distribution Asymptotic define for later use, for theory statistics non-symmetric We the assertion. proves the following additional functions 00 gx(u, x) = 2 = 2 h(u, j) tt^x), ... (3.1) 00 g2(u, x) h2(u, j) tt} {x) ... (3.2) and g3(u, x) = 2 J-O h(u,j)(j-x) n} (x). ... (3.3) STATISTICS TWO-SAMPLE NONPARAMETRIC 29 are well defined (A), these functions h(u, j) satisfies condition Further condition For instance condition gx and gr3 satisfy (A'). since condition of for gx (A') (iii) implies When \gx(u,x)\ for x > 0. (iii) of (A) I 00 = i2 h(u,j)nj(x) < SCiMl-ttjf^a+^^W ;=o _ <c;Ml_w))-*+s(1+/2) (34) To see the role of these the moments of the Poisson distribution. using the representation functions Let of r? given in (1.10). gv g2 and g3, recall over the expectation and variance W while En?w Vn\W denote Em Vw denote over the conditional variance and Then from the 7} given W. expectation definitions of gl9 g2, gz ri) Enh*(u, 7i) some after And Enh(u, = = EwEn, = h(u, r/) h\u, 7i) EwEnlw w = ... Ewgi(u, W\p) Ewg2(u, W/p). (3.5) ... (3.6) calculations, elementary COV(h(u, 7?),7?)= EWEn]W [h(u, 7?)(7}-W/p)] P^+P)-1 = cov(gi(u, W/p), = W) Ew gs(u, W/p). ... (3.7) Define 1, a-2= From j var ?^ 0 ?u__ ^ the Cauchy-Schwartz 1 ,2 cov \ jo y) du (h(Uj^ ... (t?). / /var (3.8) inequality / 1 , 1 \2 ( J cov (A(w,7/), ??)du < J i2 f J (var h(u, 7/))*(var?/)*dw 1 < var (t?) ( J*var A(w, 9/)dw ] with equality if and only if h(u, j) = number o and some condition satisfying /i(x) and observe = fiv(x) that for 0 < u < c j+h(u) a*2> Thus h(u). = For x (xv ..., xm), we function (A). = 2 g1(kl(m+l), /?(wD) corresponds xk) = *=i ?>vm 2 *=i to the statistic 0 for any 1 for some real function h(u,j) define co 2 hk(j)nj(xk) ;=o in (2.14). ... (3.9) LARS HOLST AND J. S. RAO 30 we Before distribution proceed of Tv under alternatives. Consider in (1.11), given state to theorem the a few words the alternatives, the from F-observations S= i i.e., (1.12), with :GUV) = Gm{F-\y)) = y+Lm(y)lmK 4?> which gives about the 0< y< the of sequence function distribution ... 1. in (3.10) with S = |, (B) ; For the alternatives Assumption there exists a continuous function L(y) such that for 0 ^ y ^ 1, Lm{y) = Also that the suppose some continuous outside limits right derivatives fixed = DI and L\y) = %) exist in [0, 1] and that finite on the open interval and are left and (0, 1). Gm{F-\U'k))-Gm{F-\UU)) = = U'k, k that of a F-observation inside probability falling is given by the uniform spacings hypothesis {Dk}. is given by the alternatives (3.10), this probability under hand, assume the the X-sample, under the null [Xk_v X?), On the other where finite exist of the derivatives Given L'm(y) subset (3.10) as m -> oo. -> ?(y) m*[Gm(-F-%))?y] asymptotic the ... Z>*(l+A*/m*) 1, ..., m are order statistics from (3.11) ?7(0, 1) with U'0 = 0, U'm = 1 and Ak random variable. We proof theorem be completed may be slightly In any the present will case, Theorem 3.1 cr is defined that for some small (3.12) one so that A*? is a well-defined probability now state the main theorem whose of this section, in Lemmas to 3.7. The of this 3.1 conditions weakened but conditions at the expense of added complexity. cases of statistical interest. cover most : Let Vv where ... [Lm(Uk)-Lm(U'k_1)]IDk. 0 with that Dk > Note = = m 2 (hk(Sk)-Ehk(V))lm* k=l in (3.8). ?> 0, In addition - cr to Assumptions \Lm(t)-Lm(s) | < c3(ia-5a) for 0 < s < t< and for (1?e) < s < t <; 1 ... (A) and (3.13) (B) assume 6 ... (3.14) TWO-SAMPLE NONPARAMETRIC where 7/8 < a < 1. Then the alternatives under 31 STATISTICS (3.10), ... ?(Vy)->N(b,l), (3.15) where i b= Proof rewritten, : Observe using c. Jo cov (h(u>7?),7j) l (u) du pl(l+p) first relation the that in constant centering (3.13) may be (1.10) ? m m 2 Ehk(7?)= 2 &=1 fc=l = 2 hk(3)E7?i(W?p) j=0 ... Efiv(W?p) (3.16) and W = (Wl9 ..., Wm) are i.i.d. exponential vector the 2, (Sl9 ..., Sm) given D* is mult in (3.11). D* has the components (n, D*) where the m-vector Using D\ given we write conditional may expectations, in (3.9) ?iv(x) is defined in Section As explained where r.v.'s. E(e^vv) = E E(e^Vv\D*) = ... E(Jv(iD*)Ky(D*)) (3.17) where JV(D*) = exp (itm-i \ji {nD*)-E/i (W?p)]) ... (3.18) and KV(D*)= Now e( from Lemma exp (?im-i[ S hk(8k)-/i(n A*)] )|l>'). ... (3.19) that 3.4, it follows E{Jy(D*)) -> exp {ibt-ct2?2) with 6 and c defined in (3.38) and (3.39) respectively. ?(m-i[pv(nD*)-Eii(Wlp)])-> so that in distribution. one, random JV(D*) converges i.e., for almost every vector KV(D*) By Hence N(b,c) Lemma 3.5, with ... (3.20) probability D*, -> e-***'* (3.21) LARS HOLST AND J. S. RAO 32 with d as defined in (3.43). Combining (3.20) and (3.21), with probability the one, in distribution. since But converges JV(D*)KX(D*) so moments that this also the of the 1, convergence implies product |< |JV(D*)KV(D*) E(JV{D*)KV(D*)) -> exp (ibt-(c+d)t*?2). theorem the continuity Using of the the assertion Lemma : If 3.1 m-* 2 Ar=l theorem for ... and Lemma functions characteristic (3.22) the conditions of Theorem 3.1 hold, then 2 h(kl(m+l), j)Midfy-n?nDk)\ ;=o oo m = m-1 2 A* 2 h(kl(m+l),JXJ-nD^inD^+o^l) where Ak is as defined *=i 2 \hfc(j)\ ? 1 -(j-nDk) n~*2 [ the difference Ak/m* \7Tj{nDk) 2 h%j)7Tj 2 {nDktf [ |exp {jlog(l+Aklmi)-nDkAklm*} ... bracket some elementary of \exp{jlog(l+Aklmi)-nDkAklm* ^l^j^^^A^/m^l2^^)]*. After (3.23) A*/m*}] j=0 ifc=i j=o on inequality 2 hU^inDD-n^nD^l+U-nDk) < m-* 2 ... in (3.12). the Cauchy-Schwartz Applying Proof: the two sides in (3.23), we have m~* 2 3.7, follows. calculations, we see that the term in the second (3.24) square is exp (DkA?lr)-1 -D*A j/r. ... (3.25) TWO-SAMPLE NONPARAMETRIC Since that 2.2 and condition h(u, x) satisfies (A), using Theorem the right hand side of (3.24) can be estimated m-* 2 ^.^/(m = we show Observe that by max(Af/m*) k this max that 33 it is clear (3.25), by + l^^Kl+im^)02^^^ ?=i Now STATISTICS ... Op(l). (Af/m*) goes to zero (3.26) in probability when a> 7/8. (3.14) * | Lm(Ui)-Lm(UU) I/ Dk | A*/m* |= <(U?-U?L1)lm*-Dk ... <D*klm*-Dk since (1953) s< for 0 < (t?s)a (t*?sa) < e we have for any > 0, 1/ min Therefore from oc> probability. Dk) (m2+e = a < 1. Also from Darling 0P(1). : If 3.2 ... lA^/m^l <O??(m(2+8>(1"a)-i). choosing 7/8, by the This proves Lemma 1 and (3.27) max Since t< (3.27) 0< e< (4(1?a))-1?2, max (3.28) |&k?m*\ -? 0 in lemma. the conditions of Theorem 3.1 are satisfied then, for any ?>0, [me] oo ... Urn sup?m-1 2 A* 2Hkl(m+l)9j)(j-nDk)n,(nI>k)\<Kta+? m k=l ?>oo with (3.29) y=0 probability one. : In terms of gs(u, x) defined Proof in (3.29) is (A'), the expression in (3.4) and which satisfies condition [mc] m-i 2 k=>i = A12-5 &kgz(kl(m+l),nDk) [me] 2 [Lm(Uk)~Lm(U'k^)]gz(kl(m+l), nDk)(rnDk)~\ ... (3.30) ?4 LARS ?OLST condition Using (3.14), ANt> J. ?. M and writing = JEtA? [mg], this is M 2 0< it/)*) I(mD*)-1 g8 (*/(m+l), i [LM)-Lm{U^x)] < clCs 2 (J/?-?/?LiXfc/m^l+?AD*) i now make We ... 2]. use of the representation of the (3.31) spacings r.v.'s exponential Wv ... with W2, mean 1. Writing Wk of i.i.d. in terms ? k = 2 Wjjk, the in (3.31) is RHS ?7.A M _ cOH0 __ S [WtHk/M)?(fl$+ W*) W%^{{lc-l)IMYlklM)<>. 1 Wa+C2 rr m = G -6?+'. f^"*'2' if-1! if? ^5) ?Ff-i(*/Jf)?+'-i(?P?+1+ ... '[{?-ii-wtik.wirymkwt)]. as k ?> 00, ?Fj;?> 1 a.s. and Wk? Now {1-(1- ?> 0 a.s. a WklkWtf}HWtlkWt)-> (3.32) so that ... a.s. (3.33) Using the Holder inequality, ~ | Wr< (W+<->W?+' Rl-(1?WtlkWt)-)HWtllcW,)] 1/JP, <[i?wr"p[?fwJo'**'] J U, Uf! Using the fact that in (3.34) converges if ak ? \ H WkiikWje) n 1 as k -? oo, w12 a.s. to the finite limit i a? -? 1 as n -? oo, the RHS TWO-SAMPLE NONPARAMETRIC STATISTICS term the other Similarly we get the desired Lemma can be handled so that -Q : Under 3.3 Wk* W*? in (3.32) involving result. 35 the conditions of Theorem 3.1, m-1 2 &kgz(kl(m+l),nDk) k~i l ?> J l(u) cov (h(u: 7j), 7j) du o : For Proof any fixed ?e] ing of 3 parts , 2 fc-l and ?-[me] 2 can be used In view nDk) A*gf3(?/(m+l), t k-[tnt] The in probability. . Lemma 2 1-* [m(i-a)] will proof the sum in (3.35) as Consist consider shows 3.2 that the that the third first jfe-[m(l-e)j sum is negligible. A similar analysis a.s. by Ke*+?. term is also bounded m-1 (3.35) m [m(l-e)] 2 viz., 0, we may e> ... in probability. (pjl+p) then - be to demonstrate of (3.7), it is enough to show that ? l(u) E(gs(u, W/p) du since complete ... (3.36) e is arbitrary. = except possibly lm(y) exists and is continuous By our assumption L'm(y) If lm(y) is continuous, then bounded for a finite number of points on (?, 1?e). condition ness of lm(y) along with the fact g3(u, x) satisfies (A') allows us to : from the Glivenko-Cantelli 2.2 as follows theorem, apply Theorem max k Theorem Also from A*;?lm(k?m+l) | \?> 0 with probability 2.2, [m(i-?)l 2 m-1 gs(k!(m+l), nDk) k-[me] Hence the sum 1. in (3.36) has [ma-?)] m-1 2 the - same probability 0^(1). limit as lm(kl(m+1)) g3(kl(m+1), nDk) k=[me] which from Theoren Now will not 2.2 is the required if lm( ) has a finite create any problem limit given set of discontinuity since the function in (3.36). inside (6, 1? e), this points is bounded in this intervajt 36 LARS on m. S> Take J. S. RAO in (0, 1) except at y = yQ. By our assump right limits at this point and the point does our 0 so that 0 < y0?8 < y0-}-S < 1. From that lm(y) is continuous Suppose tions lm(y) has finite left and not depend AND HOLST it follows that with probability theorem and the Glivenko-Cantelli assumptions and m is sufficiently A <8 one, \ large. From |k\m?y? | a: | is bounded whenever seen to the sum that the contribution this it is easily arguments by analogous (3.36) of y0 can be made arbitrarily in the neighborhood small a It is obvious of finite that the situation 8 sufficiently small. from such terms by choosing set of discontinuities 3.4 kind can be on m. depend handled This same way, the proof the completes if of : Let JV(D*) be as defined first not set does the discontinuity Lemma 3.3. Lemma the of in (3.18). = exp (Um-* [?y(nD*)-Efi{W?p)]) Then under -> E(JV(D*)) the conditions of Theorem 3.1, ... exp (ibt~ct2/2) (3.37) where i b= ... cov {h{u, r?)9r?)l(u)du p/(l+p) oI (3.38) and c == ?varg^u, 0 JV(D*) In Lemmas ?fiv(nD)] = exp established already in probability converges to b. Thus (?m-*[p,(nD)?E/iv(Wlp)])) g-^u, x) satisfies asymptotic gx{u, W?p))du) . ... / ... (itm-^v(nD*)-/iv(nD)]+[j^v(nD)~E^v(Wlp)]). 3.1 to 3.3, we ?/(exp Since cov(W, \ J0 (3.39) can write We Proof: W?p)du?( normality condition (A'), Corollary that we the first part m~*[/iv(nD*) need only show that -> exp 2V g^k/im+l), k=i nDk) ... (-ct2/2). 2.1 of Section of fiv(nD)= (3.40) 2 holds (3.41) and the TWO-SAMPLE NONPARAMETRIC STATISTICS 37 var (gx(u ,W/p)) and cov (W, gx(u, Wjp)) 2.2. Further by Theorem as functions of Lemma in u, satisfy the conditions 2.1, so that as v ?> oo is assured . m -War A;=l 1 / -> I v^r(g1(u, Wlp))du~( 0 \ , mm v 2 gi(kl(m+l), Wk?p) 7-m~2 \ Nil / 1 cov2 ( 2 Wk, 2 g(k?(m+l), Wk\p) v2 J0 cov (If, gx(u9W?p))du) / ... -c.D Lemma i.e., for : Under 3.5 almost the assumptions (3.42) 3.1, with probability of Theorem one, every D* KV(D*)= El exp [itm-*\ 2 hk(Sk)-/i(nD*))\ D*J j ->exp(-dt2l2) where d = i E[g2(u, WIP)-9l(u, Wlp)fdu-p 0 : The Proof Theorem 2.1 hold theorem Cantelli lemma will be proved by and showing d = Ax?B\. one that with probability M 2 Dl = fc=i U'M+m-* ... ( ) Eg3(u, ' v 0 W?p)du)". that verifying we First have Lm(U'M) ^q^Pq the (3.43) of conditions by the Glivenko ... (3.44) = from [mq] and Uk is the ?-th order statistic ?7(0, 1). Clearly conditions and since PQ = q?? 1 as g ?> 1?, of (2.9) (2.11) of Theorem part a For real numbers and b, consider 2.1 hold. where M h(u>3) It is easy to verify that if h(u,j) = ah(u,j)+bj. satisfies condition ... (A), then (3.45) so does hx(u,j). Consider M = m-*2 W%+l),6)-^iW(w+l),6)) ? k=i ... (3.46) HOLST LARS 38 fl9 ..., ?m are independent it follows that for some positive AND J. S. RAO and ?k is p(nD*k). From ... we constants have cl9 c2, where the assumptions, M V(Q = m-i 2 var (^(?/(m+1), &)) M 2 <m?V1 [(kl(m+l))(l-kl(m+l))Y(nDl)e*+ct k-i Im 2 <c; by the Holder inequality nDl = 6 2 P/im+lJKl-^m+l))]^!):)2^!) *=*i <mr^c1 e c \ ... (nDl)*?m)5 +c8 and Lemma From 2.1. (3.47) the assumption (3.14), -* nDic+niLmiUi??LmiU't^))! < nDk+KxDk m*+K2 D%rr* ... < K3(mDk)+K2(mDk)*m*-*. = the mDk representation c numbers that for > 0, large Using lim is finite with probability one. m^1 2 ?o m ?> it follows (WkIWm), 1 As a > (3.48) by the i with \, we have, using {mDlp one. < m^X 2 i Thus with (X3mZ>*+X2(mZ>*)a7n*-"-fl)C6+1 we will verify (Q < oo. ... (3.49) > 0. ... (3.50) that liminf assumption -> X4 one probability lim sup var By theorem the binomial i probability Now law of [mDjcY mm m-1 2 strong (A), it follows that var(Q there exists an interval [a, b]C (0, 1) and a u < 6, Again from the integers jx ^ j2 such that hx(u,jx) ^ hx(u,j2) for < TWO-SAMPLE NONPARAMETRIC STATISTICS and our assumptions, strong law of large numbers 0 < C < D < oo, with probability one {k: a < k?(m+l) <b, =H= for n Therefore sufficiently var >(Q with similar one. probability fashion for any that 0. large, < *< ?>(w+D Hence it follows seen C < nD\ < D}/m -> Kx > 2 a(m+l) it is easily $0 (3.50) var (A1(A/(m+1),&))/m > K%> 0 is satisfied with probability one. In a that m ?k)\%+'?m< oo. lim sup 2 Elh^kKm+1), *?i the Liapunov Therefore condition M * !a+e/( var (Q) 2 E |h1(k?(m+1), h) i M = m-'2 2 ^ |A1(Jfc/(w+1),(t) !a+8/(var (Q)1+4/2 m i -> 0 as m?> is satisfied with one. probability one. probability var By (Q (?))*) -> #(0, the next -> (3.61) Thus ??Ovar with ... oo, Lemma 1) 3.6, we have that in probability (azAq+2abBqp^+b2qp-1) ?. This verifies that the assumptions of Aq ->AV Bq -J>B1 as q ?> 1 one. 2.1 are satisfied with From Theorem the definition (3.43) probability of d as well as the expressions and for and (3.55) (3.54) Aq Bq, it follows where d= which proves Lemma Under the 3.6 ^1-?2 ... (3.52) lemma. : Given D*, let (?l9 ..., i;m) be independent 3.1 Theorem of and %k be ^(nDj). the assumptions m"1 m 2 var ^(?/(m-J-1), i ?*)) -> a^Aq+2abBqp^+b^f-1 ... (3.53) 40 ?. s. rao where in probability = Aq and and holst Lars q = Bq p-i ? E(g3(u, o Wjp)) ... du. to those in Lemma 3.1, that, 2 A2(?/(m+l),j)[7r^Z>*)-^(nD^)]^0 fc=i i=o 2.2, we Theorem Using get M q m-1 2 A2(?/(m+l),j)7T^Z)^-> o j (Eg2(u,Wjp)) du fc=i Therefore in probability. M q ?*) -> oJ (%2(^,Wjp)) du. m-1 2 EWikKm+l), i other The terms can be handled analogously = c+d where (3.5), : From c+d = (3.39) and the definitions (3.7), we (3.6) and ... (3.43) of c and d and from identities get / 1 J var (g^t*,-W?) du? o W v2 1 J cov (If, ^(w, Wjp)) du)/ \ /0 Eg3(u,Wjp)?S / 1 1 \ v))) d^ J Fw(???|w(H% y))) du+ 0J Ew{Vn w(h(u, 0 -(!+/>) == (3.56) (3.8) respectively. + 0} E\g?u,Wlp)-gt(u, Wlp)fdu-p( = the assertion. proves (T* (3.43) and in (3.39), c, d, tr2 are defined Proof which : 3.7 Lemma By calculations for instance ? M m-1 2 in probability. it follows (3.54) (3.55) = Proof : Recall from (3.45) that hx(u,j) ah(u,j)+bj. similar ... Wjp)]* du f? EMu> Wlp)-gi(u, [ ? cov {h(u,y), y) pj(l+p) 1 r J var (A(w,9/))dw? 0 i du^ y? cov (A(w,?/), 9/)du L 0J* J p2j(l+p) = <r2. ... (3.57) STATISTICS TWO-SAM?LE NONPARAMETRIC lemmas These lemma following Lemma gives : A 3.8 the origin to 3.7 : We condition sufficient have 0< = Lm(0) 0 < L'Ju) < for 0 < s< 3.1. of Theorem proof condition for (3.14) to hold. to hold (3.14) for : Under 3.1 c -u*-1 for The in a neighborhood of u < 0< (3.58) c(t?-s?)lot-(Lm(t)-Lm(s)). !> 0, the assertion follows. the null (1.1), of Vy defined in (3.13) is N(0, ... e. ? t< du = s| (cu^-L'Ju)) 0 and L'm(u) Corollary This the complete sufficient a simple is that Proof Since 3.1 4l hypothesis the asymptotic distribution 1). is a direct of Theorem 3.1 and is obtained consequence by 1 u 0 in This corollary the null distri < < l(u) ^ 0, (3.15). regarding of Vv can also be reformulated in the following form using interesting taking bution Lemma result 2.1. Corollary variables with random 7?x,r\2, ... be a sequence of i.i.d. geometric the asymptotic Then null distribution given in (1.9). of : Let 3.1' p d.f. m im S hk(Sk) is N(E[ i var 2 x i hk(7ik) ' , regression See finding (3.10), given coefficient ? also Holst the optimal i.e., a given \ v m 2 2 v i hk(7ik)-~? i 7?k' where ? is the by , m = / m cov | m 2 hk(7jk),2 v i 7/fcj ,m > Ivar ? 2 t/^J. now 2. We of the problem consider (1979), example choice of the function h(u, j) for a given alternative sequence sequence of functions Lm(u) with the property Lm(u) = m*(Gm(F-l(u))?u) as m-> -> L(w) oo. ... (3.59) : If the sequence of alternatives is such that the assumptions of test 3.1 are fulfilled, most powerful the Theorem then an asymptotically (AMP) of the simple alternative hypothesis against (3.10) is to reject H0 when Theorem 3.2 m 2 l(k?m+l)Sk > ?;=i a 12-6 c ... (3.60) LARS HOLST AND J. S. RAO 42 I is the derivative where statistic of this optimal l(kj(m+l))(Sk-ljp)) under I Arnr* \ the alternatives test which I ?l\u) (3.10) . / ->X(/>-1 rejects HQ when 1 . i the asymptotic c is determined /r it follows that (r?) ... . that ... l(u)-j. the AMP by (1972), we have 3.1 of Hoist follow distributions asymptotic 3.1 for the above special case. this result, of the -i* the 3.1 and Corollary power (3.63) 1 var (A(w, ?/), 7/) dw > / = ... J*(var h(u, t?))du 1 ,2 J cov h(u,j) on (3.62) J Z2(^) dw,-cra / , o j cov (?(^, 7/), ?/) l(u) du / as in Lemma argument when is maximized From (3.61) ... ?=l the same results , that 3.1, it follows m 2 h(kj(m-j-l), Sk) > Theorem - Jf The ... (3.59) satisfying 2 l(kj(m+l))(Sk~ljp)) k=i Ph = quantity <r2) du)(l+p)jp2 m : From Proof Using ->X(0, distribution with or*= while asymptotic \m i cQlm-t 2 under H0 The in (3.59). of L, mentioned is given by (3.64) this (3.65) directly from Theorem test of level a is explicitly given by :Reject HQ if r rn- r , i i ni J l*(u)duj(l+p)p^ ]> Aa j V_l{kj{m+\){Sk-ljp) j j \m ( Aa is the 3.2 we Theorem where standard normal of the iV(0, upper a-percentile find that the asymptotic power c.d.f. is given by the expression Also from 1) distribution. of this test in terms of the ... *(-*?+( ... (3.66) Sl2Wduj(l+p)f). (3.67) TWO-SAMPLE Relative seen from Theorem in using h(u, j) = it is easily Furthermore (ARE) Efficiency NONPARAMETRIC [0, 1] instead of the optimal h(u,j) = e= 43 STATISTICS 3.1 that the Pitman >j for d(u) some Asymptotic function d on l(u) j is ... (3.68) / f d(u)l(u)du \2j { J d2(u)du-( } d(u)du\2}j J Z2(^)cfel : translation some appli now consider alternatives. We Example on we cations of the above results tests. shall look at First non-symmetric the translation alternatives. Let Xv be absolutely continuous ..., Xm_1 i.i.d. random Let Yl9 ..., Yn be i.i.d. with variables with distribution F. 3.A. d.f. G. to test wish We :G(x) = F(x) #0 against the of translation sequence :G(x) = A&> Let f(x) = Lm(u) = and iff'(x) exists at the continuity Uu) tically now illustrate optimal test Example normal L(u), -> l(u)= - 6f\F-\u))lf{F-\v)). based 3.2 may on ... to obtain (3.70) #'s then, many finitely be used (3.71) the asympto {Sk}. or normal der Waerden ... say. score type : test) For the d.f. F(x) we = of fr(F~\u)) how Theorem (3.69) as m ?> oo except for at most we have statistic (A van : ... F(x-d\m?). -> -df(F~\u)) is continuous points = Then m\Gm(F-\u))-u\ And We Gm(x) continuous. be Ff(x) alternatives = = <J>(x) (2n)~* ? J aoexp (- \t^)dt, - oo < x< oo find -f'(F-\u))lf(F-\u)) It is easy to check are 3.1 Hence satisfied. that the the AMP T= = 4>-H?). conditions required regularity on test is based the statistic m S (D-^m+l))?* ... of Theorem (3.72) 44 from Theorem 3.2. LARS AND J. under m J x2y(x)dx = 1and 2 O-^Jk/im+l)) = 0, fc=l -ao 0 the null that hypothesis (T/w*) -> N(0, From Theorem 3.2, ? 0( Aa+0(1+/?)-*), To find test based (3.73) test relative of the Wilcoxon efficiency Since (3.68). only need to calculate (3.72), we .J (2u-l)Q)-1(u)du = y(x) dx f 2(?(x)-l)x 0 ? oo? oo 1 and ... (l+p)?p2). test power for a one-sided asymptotic same as that of the Student's ?-test. the the the Pitman on RAO S. facts ? = J (O-1^))2^ have the Using 1 we HOLST = 2 f 1/3, o test versus Wilcoxon to the optimal dx = (<p(x))2 = J (<S>-\u)Y du 1, using o scores the normal formula (3.68) test statistic has (3.72) : scale alternatives. Example random variables under scale ..., Yn be Yv i.i.d. Next the scale Lm(u) = the consider continuous absolutely Let Xv ..., Xm_x be i.i.d. = 0. We wish to test alternatives. = F(0) 0(0) G(y) with ... F(x), (3.75) alternatives :G(x) = H["> If the density we (3.74) as asymptotic properties der Waerden's rank tests. H0:G(x)= against ... same the of test type and van Fisher-Yates-Terry-Hoeffding positive and F(x) 77-* the ARE e = 3/tt. 3.B is 1 = J (2u-~l)2du The a of level f(x) = F'(x) m^GUF-^uV-u)-* Gm(x) = is continuous, L(u) ... F(x(l+6?m^)). = then - (3.76) as m ?> oo, df(F~\u)) . F~\u). ... (3.77) TWO-SAMPLE And if f'(x) exists to analogous NONPARAMETRTO is continuous and 45 STATISTICS for finitely except many then points, (3.71), F-\u)jf(F-\u))} -> if*)= -d[l +f'(F-\u)) Uu) where statistics /' exists. Optimal case of translation alternatives. : Example distribution F(x) on based score or exponential (Savage = (1?e~x) for x > 0 we l(u) = - {Sk} ... can be derived For test). (3.78) just as in the the exponential find ... #(l+log(l-*?)). of Theorem 3.1 can be verified assumptions is given from Theorem 3.2, by The and hence (3.79) an optimal statistic m T= ... 2 log (l-kj(m+l))(Sk-ljp). (3.80) fc=i i Since oj = (l+log(l?u))2du 1 we that get c?{Tjm*)-> N(0, and the asymptotic that ... (l+p)jp2) is power ... ?(-Aa+d(l+p)-i). The ARE statistic of Wilcoxon statistic is an approximation T (3.81) to T relative to the Savage (3.82) in (3.80) above is 3/4. statistic (see Lehmann, has which 4. This i.e., the test for the above power asymptotic deals with the (3.82) as the statistic theory distribution Asymptotic statistics on 2 Xkj is the test based 2 k=*l same section situation class for of statistics symmetric symmetric T 1975 n m?1 p. 103). The UMP The fc=l Yk in (3.80). statistics in {Sv ..., Sm], of the form T*= ?h(Sk) ... (4.1) 46 LARS AND HOLST J. S. RAO tests are an important subclass Such symmetric for some given function h(j). tests and hence are suited for testing the equality invariant of the rotationally is also statistics of two circular populations. Clearly this class of symmetric if the in the last section. Indeed by the asymptotic theory discussed not the section the function last with of does function hk(j) k, i.e., vary h(u,j) we is of then obtain the symmetry of and is a function u, j independent only i But since J l(u) du = 0, it follows from Theorem in the numbers [Sl9 ..., Sm}. o covered distribution of?7* under the sequence 3.1 that the asymptotic 3.1 and Corollary Thus the null hypothesis. that under of alternatives (3.10) coincides with are that alternatives cannot the of statistics type (4.1) distinguish symmetric at a of n~* and 'distance' Therefore more zero power we :G&y) = y+Lmiy)?mW, 4? close such against have efficiency comparisons, Let with 8 = 1/4 in (1.11). alternatives distant have to make in order to alternatives. consider the 1 0< y < with Lm(u) For this assumptions = we will make situation, symmetric : (B*) Assumption is a function L(u), 0 < : Assume u < ... mV\Gm{F-\u))-u). is twice Lm is twice 1, which (4.2) the following slightly stronger on [0, 1] and differentiable differentiate continuously there and such that ?(0) where Notice V = define D? We observe ?(1) I and to analogous Dk(l+Aljmv*) that under max 1^ k^m 0, sup V denote smooth \Lm(u)-L(u)\ = - L" = for such that sup We = first the sup =o(l), and also hold = o(l). (4.3) of L. derivatives second \L'm(u)-l{u)\ ... o(l) : ... (4.4) (3.12) = A*c with the above |AJ| < and the following alternatives, (3.11) = \L?(u)--L\u)\ conditions, we \lm(u)\ < K < oo. regularity sup 0^ m^1 ... [Lm(Uh)-Lm(Uk_x)]IDk. (4.5) have ... (4.6) NONPARAMETRIC TWO-SAMPLE The STATISTICS : Suppose 4.1 Theorem that there exist constants c? and satisfy that (4.7) (h(Sk)-Eh(7i))?m^cr ... (4.8) and Assumption^*) c2 such ... \Mj)\ <c1(/2+l)/oraZZj. Let Lm(u) of the symmetric distribution theorem gives the asymptotic following T* under the alternatives (4.2). statistics 47 let m P;= 2 fc=i where a2 = t? is and the alternatives var (h(7?))?[cov random geometric ... (h(r?), r?)]2?var (r?) variable defined in (4.9) Then (1.9). under the (4.2) 1) as <?(V*)->N(A, ... v -> oo (4.10) where to : Following Proof show that the method m~* We ... l2(u)du\ cov(h(7?), 7i(7}-l)-47ilp)p2l2(l+p)2(T. A=[\ used [fly(nD*)~ju,v(nD)] in the proof ?>A of Theorem 3.1, (4.11) it suffices in probability. have m-*[ju,v(nD*)-fiv(nD)] oo m = m-* 2 2 hU^n^nDD-rr^nDfc)] oo m 2 h(j)n}(nDk)[( 1+ A*/??1'*)/ exp -( raZ>*A?t/m1'4) = m-* S fc=l ?=o _l_j_WjDA;A*/m1/*-{j(j-l)-2>2)&+(?I>i)2}A|/m* m +w-?>4 2 fc=i ?? +WI-1 2 oo 2 h(j) TrunD?iJ-nDjk) A* y=o 00 Vm)n}{nDk){j(3-l)-23nDk+{nDk)*} ?%\2. ... (4.12) LARS HOLST AND J. S. RA? 48 some After direct it may calculations, two terms on the RHS of be verified tedious of the first limits probability but (4.12) are that the zero while that of the third term is under - J l\u)du) ( the assumptions, (cov (%), v(v-l)-mP) thus = or putting hk(j) h(j) -y- k null result on the asymptotic 1 in Theorem 0, 0 < u < 3.1', we get the following the symmetric statistics. for ) the proof. completing = l(u) Taking in Corollaries 3.1, distribution -p2j2(l+p)2 4.1 4.1 : Let V*v be as defined in (4.8). Then under the null hypothesis Corollary a N(0, 1) distribution (1.1) F* has asymptotically h( ) satisfies if the function condition (4.7). 3.2, we now in Theorem As function ) for the h( the conditions the form of Theorem : Reject H0 when on the optimal choice of the case. symmetric : For 4.2 Theorem a result consider the sequence of alternatives given by (4.2), satisfying test is of most powerful 4.1, the asymptotically (AMP) m 2 Sk(8k-l)>c. ... (4.13) k=l Proof test of the : From Theorem 4.1, it follows (4.1) is a maximum Observe that form maximized. when that the asymptotic A given the quantity cov (r?, r?(7j?l)??7ijp) = 0 power of a in (4.11) is ... (4.14) and var where ? ? (h(r?)) is the usual cov2 (h(y), 7j)j var linear ? Therefore 2p-2(\+p)2Aj we can regression = cov ? (r?) var ... (h(7j)??ri) (4.15) coefficient ... (h(r?), 7/)/var (r?). (4.16) rewrite = Jo l2(u)du cov (h(ri)-?ri, r){<r?-l)-4?///>)/[var (%)-^)]* = cor (%)-/??/, < [var (ti(V-l)-4?/p)]* 7i(t)-1)-4:7iIp) = [var M?/-l)-4^//>)]* 2p~*(l+p) ... (4.17) TWO-SAMPLE NONPARAM?TRIC 49 STATISTICS with equality in (4.17) if and only if for some a and real numbers b. = maxi 0 4.1, we Theorem further have off ( [ 2^ ?*(?*-l)-2m//| and that the asymptotic from the r?(ri?l) and (4-18) under H0, 1) ... (4.19) of level a is }nu)du)j(i+p)]. proof we see cor2(h(7?)? ?y} t?(t?? 1)??7\\p). above = im^pp-^l+p)])^^, ?[-k+( Further, that for a test power by h(7?) Q J* Z2(w)du?(l+p). ? Using is maximized A Thus that the in using ARE Hh(Sk) instead of 2 Sk(Sk? 1) is e= m ... (4.20) m The statistic 2 SI which is equivalent to 2 Sk(Sk? 1) was proposed by *=1 k**l Dixon (1940). Blumenthal test while and Weiss Blum and Weiss (1957) also show (1976) discuss the ARE of this (1963) and Rao the consistency properties. LMP is asymptotically (1957) consider that the Dixon test "linear" alternatives with density {l+c(y?|)}, shown that Dixon For test a nonnegative is indeed AMP if we r, integer 1 f alternatives I0 against but we have 0<y<l(|c|<2) against Blum of the form (4.2). define for x= r h(x)=< ... (4.21) otherwise, then T*v is the r. statistic This A12-7 = Qw(r), the proporiton statistic has been discussed m 2 fc=i h(Sk) of values in Blum are equal to among {Sk} which and Weiss (1957) from the point 50 LARS AND HOLST J. RAO S. of consistency. the asymptotic Our results establish normality of alternatives (4.2). After H0 as well as under the sequence tions we find from Corollary 4.1 that under the null hypothesis A ~> N(0, nQm(r)-pl(I+p)r+1]) of Qm{r) under some computa ... or2) (4.22) where ^2 = W(l+P)r+1}[l-(W(l+P)r+1){l+(^- -. (4-23) run test Let U be the number (1940) is related to Qm(0). of runs of X's and 7's in the combined The hypothesis H0 is rejected sample. of Qm(r), it follows easily that when Ujm is too small. From the definition The Wald-Wolfowitz < \(Ujm)-2(njm)(l-Qm(0))\ Thus the asymptotic thus have, distribution and we under the ARE as shown in Rao (4.24) of 2p(l?Qm(0)) H0, ... -> N(0, 4p/(l+p)3). t?(mK(Ujm)-2j(l+p)]) Therefore as that same is the of Ujm ... 1/m. of the run-statistic the Dixon's against statistic (4.25) is pj(l+p) (1976). 5. Further remarks and discussion in this paper gives is interesting to note that the theory developed on are asymptotically to the corresponding which equivalent {Sk} tests in all the known in Section 3. For a unified discussed examples It tests rank based to the theory of rank tests see Chernoff and Savage (1958) or Hajek approach and Sidak that in general, (1967). We given any rank test, one conjecture can construct a test of the form (3.60) which the same has asymptotically null distribution here seems and to lead power. to much If this is the test case, statistics then which the are theory linear presented in {Sk} as simpler to the corresponding rank tests based on score functions. compared optimal one can derive are linear in the ranks the fact that tests linear in Using {Rk}, {Sk} the asymptotic of statistics of the form (3.60) from rank theory. distributions 3.1 nor the fact that tests results of Theorem general are seems to follow from rank theory. such (3.60) asymptotically optimal, Further these two between groups of tests is under investigation. relationships It may also be remarked Theorems that the theory presented here, especially 4.1 and 4.2, covers many other tests that are not based on ranks as, for instance, But neither the more as the run test and the median test and seems to be more general to that extent. TWO-SAMPLE The theorems statistics also be to applied the censored samples in the X-sample. in the same way [(m-i)ff] T= 2 ... l(k\(m+l))(Sk-l\p). ?=i Under test similar study are that the For censored. instance, suppose samples order statistic at the right by the [(m?l)q]-th X[{m_1)qj, as in Theorem Under the same assumptions 3.2, we obtain is given by test statistic that optimal when are can here presented 51 STATISTICS NONPARAMETRIC (5.1) H0, ?(T\m?) -? N(0, f/ l2(u)du-(f l(u)du))*}(p+l)/p2) and the on results Sobel in rank censoring or Johnson (1960) ... (5.2) J l2(u)du(/ l(u)du^\ (l+p)}V2)). ( l2(u)duj{[/ 0(-Aa+ For is power asymptotic The Acknowledgement. led to many improvements instance, Rao, and Savage (1972). in the presentation comments the referee whose to thank wish authors for see, theory and Mehrotra of this paper. References J. R. Blum, 28, S. Blumenthal, two-sample and D. A. J. W. P. Math. Ann. Statist., and L. (1972) metrika, : Two R. sample A. with associated the of certain efficiency non related to the random of an division interval. the two-sample : On (1967) two that samples are from the same 199-204. 11, :A problem heuristic method for constructing 1091-1107. 32, : Theory normality of Rank and Tests, efficiency Academic for certain Press, New goodness York. of fit tests. Bio 137-145. 59, -(1979) and normality the hypothesis testing Statist., Statist,, : Asymptotic statistics 972-994. 29, of problems for criterion Z. Statist., test two of 239-253. 24, (1961) Math. v/ / Sidak, a class Math. Ann. Ann. J. : On : A (1940) V. tests. Johnson, tests. two-sample 1513-1523. : Asymptotic (1958) Math. Ann. Statist., population. Godambe, I. R. Savage, (1953) Math. Ann. Holst, of certain normality asymptotic Math. 34, Statist., Ann. statistics. parametric Darling, Hajek, : Consistency (1957) : The (1963) problem. H. Chernoff, Dixon, L. and Weiss, 242-246. and Mehrotra, problem limit conditional with K. censored G. data, theorems (1972) Ann. with : Locally Math. Ann. applications, most Statist., rank powerful 823-831. 43, 7, Statist., tests 551-559. for the two 52 Le LARS L. Cam, Pub. E. Lehmann, San Pyke, L. R. Rao, J. : (1976) J. S. and V. R., statistics A. tion. Paper Revised (Tun RAO S. intervalle par des pris points au hasard. 7-16. Statistical Based Methods on Ranks, Inc., Holden-Dey, Jour. tests Some L. Roy. based (1980) : Statist. on arc Some Ser. Soc, lengths results B, for the on the 27, circle. 395-449. Ser. B, Sankhy?, asymptotic 38, Submitted theory. spacings 329-338. publication. J. Sethtjraman, of random Wald, la division 7, Paris, Nonparametric Spacings. : S. and Holst, for U. Univ. (1975): (1965) Rao, Rao, sur th?orems Statist. J. Francisco. J. S. Rao, : Un (1958) Inst. AND HOLST variables I. R. Savage, : The two and Wolfowitz, Ann. Math. received : August, subject and sample J. Statist., : January, 1979. (1975) : Weak to perturbations Sobel, 11, : On scale of empirical factors. Ann. : Contributions (1960) case. Ann. Math. Statist., a test whether 147-162. 1979. and M. censored (1940) convergence two samples to the 31, functions distribution Statist., 3, 299-313. theory of rank order 415-426. are from the same popula