Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
SOl'fJE PROPERTIES OF A BAYES TWO-STAGE TEST FO,R THE MEAN 1 by Morris Skibinsky Institute of Statistics tJniversity of North Carolina Institute of Statistics Mimeograph Series No. 107 June, 1954 1. This research was supported by the United States Air Force, through the Office of Scientific Research of the Air Research and Development. 11 "Beware the Ja.bberwock, my son 1 The jaws that bite, the claws that catchl Beware the JubJub bird, a.nd shun The frum10us Bandersnatchl" He took his vorpal sword in hand; Long time the manxome foe he sought-• • • • • • • • • • • • • • • • • • • • One, two lOne, two 1 And through and' through The vorpal blade went snicker-snackl He lett 1t dead, and with its head He went galumphing back. .'Twas . . brillig, . . . . .and. .the. .slithy . . . toves ..... Did gyre and gimble in the wabe; All mimsy were the borogoves, And the mome raths outgrabe. Jabberwocky (Lewis Carroll) 111 ACKNOWLEDGMENT This research was suggested by Professor Wassily Hoeffding, to whom I am indebted for his careful and penetrating criticism, and for his kind encouragement. The financial assistance of the United States Air Force is acknowledged, and sincerely appreciated. Last, but certainly not least, I thank my wife, Phyllis, without whose patience and understanding this work could not have been completed. Morris Skibinsky iv TABLE OF CONTENTS Page ACKNOHUDa1ENT INTRODUCTION iii v Chapter 1. GENERAL PROPERTIES OF THE S3COND SAMPLE SIZE FUNCTION IN THE NORMAL CASZ Nature of the Equation Defining the Second Sample Size Function 1. 2. II. An Important Identity ASYr~PTOTIC 1 23 PROPERTIES OF THE BAYES '!WO-STAGE mST 3. III. Asymptotic Dxpression for the Second Sample Size Function 33 4. Expansion of Second Sample Size Function 50 5. Expected Value of Second Sample Size 56 6. Error Probabilities 6, 7. Comparison with One-8taee Test 81 8. A Trivial Asymptotic Solution 67 NON-ASlMPTOTIC CONSIDERATION OF BAYES TtJO-STAGE TEST 9. Further Properties of Seoond Sample Size Function 90 10. Some Exploratory Computations in the Symmetric Case 100 FIBLIOGRAPHY 123 v A fairly general statement Qt Wald's decision problem L-5J 1, for two stages may be exp-essed as follows. Let X be a random variable with frequenc," function, fg(x), where g may be a real or vector valued parameter in some subset,.I'l , of the real line or of a finite dimensional euc1id1an space. Suppose we are given a sequence of independent observations, x l ,x2 ' ••• , on X. On the basis of these observations we are to select one of a finite number of possible alternative courses of action, AO,A l , ••• ,Ak , which comprise a set A. of possible alternatives, by using the following decision rule. We are given numbers, a-m , m=l,2, ••• j functions, pv-m (x ), defined for all points !m in m-space, and for m=l,2, ••• , v=0,1,2, ••• ; and functions,8m+v(Ai'~V)' defined for all Ai € A , all points !m+v in (m+v)-space, and for m=l,2, ••• ,V=O,l,2, ••• ; such that always (0.1) , 1. Numbers in square brackets refer to bibliography. vi and -Rule 1. Take m observations, with probability 2. If the observed sample is x , take a second sample of v ~. -m observations, with probability pv(.!m). 3. If the total observed sample is !m+v' accept alternative The problem is to find sequences q, p, and 8, subject to the indicated restrictions, Which are optimum in some sense. In particular, suppose we are given a loss function, W(O,A i ), defined for all Oe Jl.. , and all Ai EA , non-negative and bounded, representing the 10S8 incurred by accepting alternative Ai when Q is the true parameter; a c.d.f., ~(Q), defined over.fl-; and suppose the cost per observation to be a constant, c. We may then seek those sequences, q, p, 8, for which the average expected loss is minimum, i.e. a Bayes solution. Using the above rule, the expected number of observations required, given'O as the true parameter, 1s vii where EQ denotes the expectation ot a function of the observations on X, when Q is the true pa.r.~ter value. The probability, given 0, that the rule will accept A1 1S Hence, the risk or expected loss incurred by use of the rule, given 0, is 00 (0.5) L Clm m=l and the average risk overJt is, after making the permissible indicated interchange of integration and summation operations, (0.6) 00 r. Clm m=l ( 00 cm+ 1:: v=O viii where Xm+-v stands for the (m+-v)-dimens1onal observation space, and where by fg(~v)' we denote the joint frequency function of the first m+v observation on X. The existance of a sequence, 8, which for any fixed sequences, q, p, and fixed point, !m+v' in (m+-v)space, will minimize (0.6), is immediately apparent. aj(!m+v) Let. f = W(g,Aj)tQ(~+v)dX, ..n... and suppose min(a ,a , ••. ) is unique, then such a sequence is l 2 1 (0.8) 8 m+v (A x j'-m+v ) = , f j=1,2, ••• o , Modifications of (0.8) for which the restriction of a unique minimum may be removed are easily made. To complete our Bayes solution, we need sequences, q and p which will minimize (0.6) when the sequence 8 is as defined by (0.8) or a suitable modification. • Let then the average risk may be written where, for m=1,2, ••• , f r:v+ .X( f [A:.:(Q,A1l&m(A1'~)] tQ(~)dl- Q,Ai lPQ(Ry(Ai '!m) )FQ(!m'dA, ""1,2, • • • . At .n... (0.11) ---fJ. v (x ) .. -m 1 ,yeO .J1. and PQ(Rv(Ai'!m» represents the conditional probability of accepting A., given that the first sample is ~ x and that v l observations are taken in the second. < _ f .1l- Now max W(O,A.)fO(x )0)", eft l -m A i so that~v(!m) is non-negative and has at least one absolute minimum with respect to v, for , x jr;:;,AJE A [W(O,Ai)-W(Q,A ( 0.13) J) If'g(!m)d>.. v < -------.---------c f ..n. f'o(:sm)d.A It cannot have an absolute minimum w.r.t. v, for v greater than this number. For each!m, let V(!m) be a value of' v for which ~ v(!m) is absolutely minimum. One minimizing sequence, p, is then seen to be v (0.14) = vex-m) v=O,l,2, ••• m=1,2, •• If we use this sequence, the average risk may, by (0.10), be written (0.15) where ( 0.16) , )-.1-- m =cm+ m=1,2, ••. xi Suppose there exists a positive integer m = m*, say, such that (0.17) ,( = rt'm* min) m f /"\m ' then a sequence, q, for which the average risk is minimum is , (0.18) ~= c 0 m = m* , m=1,2, ••• , , and this in a formal sense would complete the Bayes solution. Using this solution to our decision rule, gives us, by ( 0.19) By (0.4), the probability, given Q, that the rule will accept Ai 1s (0.20) The minimum average risk over..JL among all such rules is \ xii (0.21) Consider now that the size of the first sample, m, is given and suppose the set A contains only the two alternatives, AO' Al • The above rule may then be simplified as follows. Given functions, pv-m (x ), ~:+ defined for v=O,l, ••• mv (x:+). -mY' and, respectively for all points ~m in m-space and all points !m+v in (m+v)-space, such that always ( 0.22) 1. Take m observations 2. If the observed sample is x , take a second sample of v -m observations, With probability pv(~m)' 3. If the total observed sample is !m+v' accept Al with probability 'm+v(!m+v)' Accept AO With one minus this probability. The sequence, ., is obviously related to the sequence, 8, of the more general rule. Again, suppose we are given the non-negative loss functions, Wi(O) = W(O,A i ), defined for i=O,l, and for all o€J2, representing the loss incurred by accepting Ai when 0 is xiii the true parameter; a c.d.t., ~(Q), detined over Jl. , and suppose the cost per observation to be a constant, c. The average risk, over.A , involved in the use ot this rule is 00 (0.23) cm + Z "=0 f p.(!g,) [ 1 Ccv+Wo(g) 1t g (!",...) d), *m+v A sequence, ., which minimizes this, tor any fixed sequence, p, is clearly seen to be . (1' 11 W (g)tg(!",...)d), o, < 1Wo(g)tg(~.)d), ~ Let • where. is as defined by (0.24), and the relation ot this set to (0.9) is obvious, then the average risk may be written xiv (0.26) cm + r v~oPv (!m) .),z v (!m) dx -m *m where [(cv+Wo(Q) I!-PQ(R,,(.!m) )1+ w1 (Q)PO(R,,(!m» JtO(!m)d>-, .A [ (\oJo(Q) "=1,2, ••. IT-t (x )1+Wl(Q) t(x - m-m- ) )fl'\(x )d>- , m.-m~-m v=o • JL Clearly this is a special case of (0.11), so that by (0.12), it is non-negative and has at least one absolute minimum with respect to " in the interval defined by (0.13). It cannot have an absolute minimum outside of this interval. The sequence, p, which minimizes (0.26) is Just (0.14) for our given value of m, and this again in a formal sense, completes our Bayes solution. Using this solution, the expected size of the second sample becomes ( 0.28) The probability, given Q, that the rule will accept ~ is aJJ.4 tile a1l'1 iIwm ave:raa' :rUk 18 Juat (0-21 our liven val~e t # W1th Dl~ "flaeec1 by ot m. In tai8 paper, we e.re concerned wU_ a particularization of tbe aener..l kyes prol)leo1,ltllned above. In this , ..e, A cOtl$lat. ot two point•.()Jl t.be real lU., say go and Ql' wtth '0< Ql- We pretar e.lterD4ttve We e.re ,lven tbe to11O"UI "0' w.en ap:-lo~l g go' A1 # wb~ 41atr1bution over (0.;0) Our Cll loss functions are , , !, .. h'atl"" ,Xm+\,) j').. Q :II -1- xvi and note that (O.}4) Let W g ),,=- , 1 then the sequence of decision functions for which the average risk is a mintmum 1s, by (0.24) , ( 0.}6) By (0.25) I r )" v >-} rm V=O,l"" xvii so that by (0.27) , v=o As justification for pursuing a Bayes approach. to this problem, it may be noted that Wald and Wolfowitz in their paper on the "Optimum Character of the Sequential Probability Ratio Test ll 1:4 J, proved, for the problem of deciding between two simple alternatives, that for arbitrary apriori probabilities, go' gl' and cost c, every sequential probability ratio test can be regarded as a Bayes solution w.r.t. some values W o' W l' say, of WO' Wl , and hence that , xviii where 8 0 is any sequential ~obab1l1ty rat10 test for deciding between two simple alternatives, 8 11 any other test for the same purpose; Q1(SJ)' 1,3-0,1, 1& the probability, under Sj' of rejecting Hi when it i8 true; Efn 1s the expected number of observations under Sjl when Hi is true (existence assumed). From this it follows, a~o8t ~ed1ately, that It will be shown that in certain special cases, similar properties hold for the Bayes two-stage test. CHAPTER I GENERAL PROPERTIES OF 'IHE SECOND S.Al1PIE SIZE FUNCTION IN mE l!Om1AL CASE 1. Nature of the D~!!~l:!Y~ Eq11.aM.on. In the following sections, we consider the particular case outlined in the latter part of our introduction, when 1 N Z (x.-Q) 2 i=l 1 - - 2 I We m = m I Z Xi ' m - 1,2, ••• , i=l that s- m+v Let • let (1.2) s 50 N = 1,2, ••• I cS m +5 v s v m.l.v =Z x , v • 1,2, ••• , i=m+1 i 2 (1.5) i log ~, t N .. sN - QN - N· 1,2, ... , then by (0.33), dt m r m .. 'I.e A. (1.6) .L, ,... , m .. .., 2 , and by (0.34), r Thus,by , v rm+v =- rm , d(s' - = e v -) Q" , " .. 1,2, ... • (0.36) 1 , tm+v > 0 , (1.8) t (x m+" -m+v ).. f " .. 0,1, •.. o, ~ Using (0.38), we have that the size of the second sample for which the average risk is minimum, is given by the integral value of " which minimizes (absolutely) the following function of " • .--J:1 (x ) ( 1 • 9) Gv (tm).. w g" f -m(x) a1 Q 1 -m 00 e -dt m + ..-;;,,-- f 3 where (1.10) A(tm) .. c( ~ + ~ -dt e m) 1 h't( v , t ) .. m (1.11) ~ dv 2 't 1 t "m 2 Now (1.9) can have an absolute minimum with respect to v only for a value of v which satisfies the inequality t 1 e , -dt m t < 0 m- , -> • Hence, an upper bound to any value of v for which (1.9) is absolutely minimum is immediately seen to be , t , t .. < 0 m- > m- 0 4 We shall, for convenience, in the following, drop the m subscript from t m, and write some of the above functions without their arguments, when this practice will cause no confusion. If we regard (1.9) as a continuous function of \/, \/ any non-negative real number, we have d (1.12) - - \/ 2 p;i 1 - ~ e 1 +2 - ~h , We have, first of all, that \I ~ 0\1 G V '<2 >~ (t) =0 [ 22 (d t +1) ~ -l1.. m(t), say • J , if and only if, (1.16) where log v = 2 log TJ 1 2 2 1 (t) - 'lid \I - t \1- .. ft(\I), say , (t) 1j dZ =-- [ ~dt + W a _ ~dt] \-1 e - 2 /2n.;.;.o -1 -1 , We nota here that in all- the work which follows, we assume that d and Z are both positive numhers. (1 19) lim • ~ [ddV G (0 ~ v ~ .. - 00, lim [. v->oo ~ (0)] OV-V ... A(O)>O, 7/1(0)=0, so that vnlen t=O, (1.15) has exactly one root in v. This root is positive ar..d is obviously the value of . v for which Gv (0) is an absolute minimum. On the other hand, l-7han t .; 0, r ~~O L ~ 011 (1.20) v----- \I (t)lJ = ~~o· o· t.C,)v r~-G v (t») . . v-- A(t) > 0 • ThUS, by (1.14), disregarding the case t .. 0, (1.15) has 2 (1. 21) 1 rts ~ in v, i'Jhen I ..( ~ (t) . V=m ClV v o < < > > t) Now 1 (1.22) log In. (t) 2 2 2 ... log 2' + log L(d t +1) d 2' - 17 - 6 is an increasing function ollt J, ~1ith unique minimum t = O. It ->00, as ~ ~ 00 = -00 at and has negatbre second derivative for all t. has a unique maximum at a value of t uhich is ; 0, according as Wo < > WI. It tends to -00, second derivative for all t. as t + -> - 00 and has negative It folIous that there exist two + numbers, call them t- , both of which depend upon the parameters d, Z, W, such that t - < 0 < t+ (1.24) and such that (1.15) has < t+, t ~ 0 t = t or t + or 0 t<t - or>t+ t-< t 2 1 rts. in v, when o In the first case above, the two roots of (1.15) lie one above, one below the inflection point v ... m(t ), so that by (1.14), Gv (t) is relatively maximur~ at the first, relatively minimum at the second. We have discussed the unique root of (1.15) when t • O. +- + 1rlhen t ... t -, the unique root, v + point of zero slope of Gv (t-). .:: 0, v > O. => 7n (t -) , is an inflection + The slope of G (t-) is thus v 7 When t < t- or > t+, the slope of G (t) is > 0, v ~ O. v It follows that when t < t - or > t + , the absolute minimum of G (t) is at v v - - = O. We define the f'lmction larger root of (l.lS), (1. 26) v*(t) = o'lmique root of (1.15), , ( t-< t < t+, t , 0 t · 0 or t - or t + t < t - or > t + It' now for every t, we take vet) to be the value of v for which G (t) is absolutely minimum (choosing the smallest number when v this value is not 1mique), we have clearly that o vet) , GO(t) ~ G *(t) = v t v*(t) , Go(t) > G *(t) v • From the above discussion, it is apparent that Ct': Go(t) > G *(t)} C (t : t- < t < t+) V Recall from (1.15), (1.16) the significance of the equation let us consider, in the following that w W D . vto ,11, 8 then .2-. f (v) .. d l ..We dt t dt I+Wedt _ ~ = 0 v if and only if 2t l+We (1•.31) v .. -d 1..\-1e dt dt -lJ:W(t)" say. We consider" of course, only non-negative values of v" so that when 11<1, the curve (1 . .31) is defined only on the interval 1 o ~ t < d1 log u. In this interval it is convex and has positive slope everywhere. 1rJhen W > 1" the curve (1 •.31) is defined only on the ~terval ~ log ~ < t ~ O. In this interval, it is also convex, but now has negative slope everywhere. In both cases, we have ~ lim'r- ~r(O) = 0, . 1 t-;. 1 vW(t) • 00 d log W ~I 1 I l II 1/: 1..--/ I -or-----~A 0, ..!../!tt.J"!" tA. II \1\.' (?f (;t) Figure 1.1 "', hi >I 9 The role of the function"rW(t) is indicated more fully, by the following analysis W<l , > 0, All " 0 f (,,) (1.33) dt t < -> 0, ~trw(t), < 0, All " , W>l t<O t ~ 1 O~t<a: 1 log W a:1 1 log 1 d log 'V1 < t 1 1 t -d > - logW 1 w ~ t>O • Consider now the equation (1.34) vIhen W < o~ 1, both sides of (1.34) are defined only in the interval 1 t < CT of t which 1 log~. ~ -00 The L. H. S. is a continuous increasing function 1 1 as t -> 0, and ~ +00 as t ~ d log The w. R. H. S. is a continuous, concave, decreasing function of t, equal to a constant, when t ~ 0, and ~ -00 w. 1 1 as t ->d log When vI > 1, both sides of (1.34) are defined only in the interval ~ log ~ < t :: O. The L. H. S. is a continuous decreasing function of t which t 1 1 -.> (i 10g~. ~ -00 as t 0, and -. +00 as The R. H. S. is a continuous concave increasing function of t, equal to a constant when t t ~ ~ 10g~. ~ = 0, and -. -00 as Thus in both cases, the solution to (1.34), call it TW' is unique. 0 10 trJhen W :: 1, (1.30) holds true i f and only i f t = O. In this case (1.33) may be written ~ dt (1.35) f t (v ) < '> 0, all v, t > < 0 , and we may appropriately define (1.36) Lemma 1. (1.37) is an increasing function of t , t - ~ t < Tw has a unique maximum at t=TW,which is equal to lfW(TW) v*(t) [ is a decreasing function of t , TW< t ~ t + Proof. It is obvious from the definition of TlrJ that (1.38) and that this relationship must also hold in the limit as w..;.. 1. ThUS, to prove the lemma, we need only show that (1) v*(t f ~l-t ) < v*(t), all t, tt such that t- * (2) v (t ) < v (t), all t, t I. la) f ~ tt < t ~ TW such that TW ~ t < t t ~ t + \ve first suppose that W < 1 Let t be any number such that t- ~ t < TW' then by (1.33) • 11 where V o is a number such that 1J~(TW) (1.40) ( v,ow<t>, > YO > C , I J} t- < t < 0 - -~)) ........ I \ I I / f I f I I -_._, - - T 1: T- rr,r) w< I .... ---:P: "? ;t I ~"l''dw Figure 1.2 It fol101'1S, using (1.38), that (1.41) v*(T ) > v*(t) > W Ib) V o Now let t be any number such that 0 < t < TW' t such that t- < t' < t, then by (1.33) (1.42) where vI is a number such that , , any number o< t , t < t , _ (1.43) 12 < t < 0 - By the second inequality in (1.40) and the second inequality in (1.41), we have further, that (1.44) -~v ---f(J) r V, I/'d Figure 1.3 Thus, it follows that * * \I (t) > \I (t I ) I lc) Finally, let t, t _ be any two numbers such that t , <; t <; :; 0, then by (1.33) (1.46) £ ,(\I) t <; ft(v) , all \I > 0 , from which it immediately follows that Q.E.D. (1), W<; 1. t 13 2a) Let t be any number such that TW < t ~ t + , then by (1.33) (1.48) where v 2 is a number such that 1 [,riv(t) , TW < t <"d log (1.49 ) [ 00 1 w 11+ , d log W~ t ~ t Figure 1.4 It follows, using (1.38), that (1.,0) 2b) v*(T ) > v*(t) W Now let t be any number such that TW < t < number such that t < t (1.,1) I + :: t , then by (1.33) 1 d1 log W' t 1 any where v is a number such that 3 ~,., V (1.52) U wet) < v:3 < (t ) , t < t [ 00 I '1 1 < d log W 11' d log W~ t := t + By (1.50), (1.49), • _.- , Figure f ()i) Xl 1.5 Thus *' v*(t) > v (t ) (1.54) 2c) FinallyI let t, t < t' ~ t+ , then by I be any numbers such that ~ log ~ ~ t (1.33) from which it follows immediately that * v (t) > v*' (t ) Q.E.D. (2), W< 1. • II. Suppose now that W =- 1. We have by (1.35) that if t, t are any two numbers such that t .. .:: t I I < t :; 0, then f lev) < ft(v), all v > 0 , t so that v*(t) > v*(t t ) If t, t I are any two numbers such that 0 have the same result. III. • ~ t < t I ~ + t , we Q.E.D. (1), (2), W=l. The proof for W > 1 proceeds in a strictly analogous manner to that given for W < 1. Q.E.D. Lemma 2. ,,*(t) J.·s a con t inuous f unc tion 0 f t , t- < t < t+ • .. Proof. The lemma will be proved if for any gi-:-en e > 0 and any arbitrary but fixed t in the indicated intervals, we can prove the following four statements to be true 16 2 2 I ~ o<"'*(t r) - '" * (t)<e 1) t~<rW 3 8>0, • 3 .,. o<t -t<8 Z) t-<t~l j 3) T~t<t+ j 8>0, • 3 • O<t'-t<8 =} O<"'*(t') - \I*(t)<e 4) TW<t~+ 3 l 8>0, • J • 0<t_t <5 ~ 0<\1 *(t) - \I * (t r )<e I 8>0, " '3 • O<t-t <8 ::;> 0<", * (t) - \I * (t r )<e We shall prove, below, only statements 1 and 2. 3 and 4 may be Statements proved in strictly analogous fashion. Proof of Statement 1. We are given that t is an arbitrary but fixed number in the interval, t-~ t < T • W ~l Let = log (1 + Me '" ,(" (T~J) ) Now d ft(\I ) < ~ () log _.d\l 0'" \I > ",*(t) , \I , t -< t < t + , from which it follows that (1.56) * ft(\I (t) ·e ~l ) * < 1 og v (t) + t' "'1. Since ft(v) is a continuous function of t, all t, all v > 0, • 3 ., 2. We use the notation, to mean "such that ". 3, to mean "there exists"; • 17 we oan find a positive nU1llber, 6, which is , (1.57) o<t -t<6~ * f t fev (t)e ~1 ~ Tw - t and suoh that .>'- ) < log v'(t) + l; 1 But by lemma 1, this impl1e-. that (1.58) * * . ( tt) <v (t)e 1;1 \/ *(t) -<y , which in turn implies that Q.E.D. Proof of statement 2. We are given that t is an arbitrary fixed number in the interval t - < t ~ T ' W Let (1.60) • (1.61) then (1.62) Now ~ \/*(t)e 2 < 7n (t) 18 (1.63) Hence, by the continuity of ft(v)" we can find a positive number, 6" such that o < t - t I < 6 =9 f (1.64) I (m (t» > log t m(t) • But by Lemma 1, this implies that (1.65) Ir~(t) < /)'i, v*(t I ) < v*(t) Thus, by (1.62) (1.66) 0 < v*(t) - v*(t I ) < v*(t)·(l-e -~2 ) ~ v*(Tw).(l-e -t 2 ) De. If, on the other hand, * ft(v (t)e ~2 ) . {~ log v (t) - ~2 > ' we can, by the continuity of ft(v), find a positive number, 6I, such that (1.68) I I O<t - t<6 =* *(t).e~2 ) f I(V t >1og v*(t) -~2 • But by Lemma 1" this implies that v*(t) e ~2 < v*(t I ) < v*(t) , and this leads to the conclusion (1.66). Q.E.D. 19 Lenuna 3. * , t <t < < G (1. 70) v~l-( t) (t) = Go(t) > t t = t··* or t i , , t , t < t·* or t i where t;', tt are two numbers which are dependent upon the parameters d, Z, W, and such that (1.71) t _<t·<O<t·<t * i + Proof. By reference to (1.9), we have that I (1.72) , t<o , t > 0 GO(t) = Ie -dt , and that -dt +~ /2i By Lemma 2, this is a continuous function of t, t-St$t+. is easy to verify that It 20 (1.14) Now previous discussion, see (1.19) - (1.25), has shown that Hence, by (1.12), (1.14), there exists in the open interval, t-<t<O" a unique value of t, call it t." which is dependent upon the parameters d, Z, W of (1.12) and (1.13) and such that for all t in the interval t-~tSO, , (1.16) t < , ); To complete the proof, we have by (1.12) that t >0 • (1.11) t <0 It is easy to verify that (1.'78) d [ dt e dt 1 Gv*(t){t~ > 0 By the discussion referred to above, (1.19) Hence, by (1.11)" (1.18), there exists in the open interval" o<t<t+, a unique value of t, call it t*, which is dependent 21 upon the parameters, d, Z, Wof (1.72) and (1. 73), and such that for all t in the interval (1.80) + , <ct~t .> t Gv*(t)(t) t* <' • This completes the proof. Note that the proof of the above lemma demonstra.tes the + existence of and uniquely defines t* , respectively, as the positive and negative roots of , (1.81) - in the interval t - < t < t + • - Lemma 3 and (1.27) now give us Theorem 1. ( v*(t) v(t) =( 0 , t* < t < t* , t < t* or > t* , where t*, t* are, respectively, the unique negative and positive roots 1n t of (1.81) which are dependent upon the parameters d, z, W, and such that t- < t* < 0 < ti < t + From Lemmas 1 and 2 and Theorem 1, we get Theorem 2. v(t) is a continuous function of t, t* < t < ti . 22 If t* -< T.W -< t* , is an increasing function of t, t* < t < vet) [ has a unique maximum at t = Tw Tw' is a decreasing function of t, Tw < t < t* • If TW < t* vet) is a decreasing function of t, t* < t If Tw> t i < t* , vet) is an increasing function of t, t* < t < ti • The following interesting peculiarity of the second sample size function, vet), is easily deducible from the above results. Theorem 3. The second sample size function, vet), has discontinuities at the points t t , t (1.84) * yeti - 0) - yeti + 0) v{t* + 0) - v(t* - 0) > v (t+), T. > Y*(t-), T.W > t* > Y*(t-), > v*(t+), Tw Recall that (1.86) then Theorem 3 implies , w-< t'* ~ t* T.W < t* , 23 By Theorem 1, we may now modify the statement of our + decision rule as follows. First, compute the numbers" t'i (see (1.81». 1. Take m observations. 3. If the observed sample is x , compute t (1.5) • -m m a) If t < t* , accept A ' mO b) If t m> t i , accept Al • c) If t* < t < ti- , take v(t ) additional observations. m m If (2c) occurs and the observed total sample is !m+v ' compute tm+ v a) b) If t m+ v ~ 0, accept AO' If tm+v > 0, accept Al • Note that in general v(t } will not be integral, in which case m we shall approximate the test by taking the nearest integral value. 2. An Important Identity. In this section, two further lemmas are proved and an important identity established. These results then lead, in the following two sections, to the derivation of an asymptotic expansion for the second sample size function. 24 We define the function o~ (2.1) > 0, Note that for every fixed v ~, which intersects the lines respectively. ~ ~ ~ <~ > o. (2.1) is a linear function of 21 = 0 and ~ = 1 at - 1 ~ d and v ' It for any arbitrary but tixed value of halt open interval, 0 v 1, ~ in the 1, we set ~ and solve for v, we get Consider the non-negative v axis in the t, v plane, (2.4) t = 0, v ~ to correspond to the case, (2.3), tor 0 < ~ ~ 0 , ~ = O. It then follows that 1, plus (2.4), tor family with parameter, ~, 0 ~ ~ ~ ~ = 0, represent a 1, the individual curves of which (ignoring points at infinity) are loci of points in the upper half t, v plane which satisfy (2.2). In particular, note that 7~(t), defined by (1.14), is the particular curve of this family for which ~ = 1. For reasons which will presently become apparent, we now consider the solutions in t ot the equation 25 First, suppose that ~ half open interval 0 is an arbitrary but fixed number in the <~$ 1. (2.6) is an increasing function of Itl, with unique minimum at t t = O. It -> 00 as t -> = -00 + 00 and is concave for all I: o. is increasing to a unique maximum between t and then decreasing. It tends to -00 as t =0 ---> and t + 00 =~ log ~, and is concave for all t. It follows that there exist two and only two values of t, call them t- , t+ , which satisfy (2.5), which must in general ~ ~ depend upon the parameters (2.8) ~, d, Z, W, and such that 26 ~ When = 0, we consider equation {1.16} over (2.4). case, obviouSly, t = 0, + t~ and we define = Note that the numbers, t In this O. -+ ,defined in the discussion above + (1.24), (1.25) are the particular cases t~ of the above solutions to (2.5). Lemma 4. Let ~, d, W be arbitrary but fixed numbers, 0 < ~ S 1, d, W > 0, then t- (Z) is a continuous decreasing function of Z, t+ ~ ~ (Z), a continuous increasing function of Z. Furthermore, + lim t~ (Z) Z-> 0 =0 + , ~ (Z) = + lim Z -> 00. 00 Proof. If we denote the function (2.7) by FZ(t) to indicate its dependence upon Z, we have for any positive 6, Z, (2.11) which quantity is independent of t. (2.12) lim Z->o Fz(t) = -00, Further lim Fz(t) Z->oo = 00, Renee, by the above description of the function, log 71Z ~ (t) , all t. 27 which itself is independent of'Z, we have that t~ (Z) is a decreasing, t + (Z), an increasing function of Z, and that JI. the limiting relations (2.10) hold. I -,// _I '" _ - - - FO:.) - ------ z ........ .ee.~~(;t) / ~_L-l-._"-:' . .--_+-_+-_-----:..I---+-:---\;-~\"!;t £rZtA) X-(%.) f. / t'- Figure 2.1, W< 1 + We shall here only prove continuity tor tJl. (Z). Continuity for t- (Z) may be shown in strictly analogous fashion. any E,"" Z > 0, For let Then, since Fz{t) is a continuous function of Z, all Z all t, we can find a 8 = 8E,Z But this implies that >0 , such that > 0, 28 Lemma 5. Let d, Z, Wbe fixed positive numbers, then t- is a 1.1 + continuous decreasing function of 1.1, t is a continuous 1.1 increasing function of 1.1, 0 ~ 1.1 ~ 1. Proof. + By (2.5), the definition of t~ (see the discussion above (2.8», and the definition of Vt(t) (see (1.26), (1.16)), we have the following identities in 1.1, (2.16) Thus, besides being respectively the unique positive and negative • + ~olutions in t to (2.5), t- may equally as well be considered, 1.1 respectively, as the unique positive and negative solutions in t to ~l (t) 1.1 Now, = Y*(t}, 0 < 1.1 ~ 1. ~ (t) is, for every fixed 1.1, 0 1.1 < 1.1 -< 1, a con- tinuous, increasing, convex function of Itl, with (2.18) /121.1(0)=0 For every fixed t " 0, of 1.1. ?Jl 1.1 (t) ,O<I.1~l. is a continuous decreasing function v*(t), on the other hand, as described by Lemmas 1 and 2, is positive for all t in the interval t l- ~ t ~ t +l ' continuous, increasing to a unique max~um and then decreasing in this interval. Also it is obviously independent of 1.1. v ", .~ --- " --, --~- -,,-'~'-' 29 - l l l (X.) / / , .... " ;: Rence t - _ . )l'(~) -- _. ~~/ Ftpre 1:+ f4 2.a .. is a decrea8in8 function ot ~ O<t"'<t«-~t + 1s an increasing 1.&, t '1.& ~ function ot j.1. By (2.8), (2.9) this is tN.e tor all "'.. includifti the lett endpoint, in the closed interval 0 We sbal1 FOve continuity only similar argument holds Let '" ". -- --?J;f4/ 0> ~ interval.. 0 tor S j.1~ 1. tw.- tj.1+ •. An ess_tully t; . be any arbitrary but fixed number in the balt open - < I,L < 1. Let E be any given positive number. two all inclusive but mutually exclusive possibilities may oc¢v. First, suppose that 0< t + < ~- £ then since t + is an inereasing function of "" we have that I.l Now, suppose that then clearly.. we have that • Hence, since 7Jl j.1 (t) i. a continuous decreas1na function ot ... 30 for all t f. 0, we can find a 6 .. BE > 0, such that ---?!let:) _._- I" / ---- - 7Jl!''' (;t) But this implies that (2.24) Hence, o < t +~ - t + ~ t <£ • This proves the continuity of t + for 0 < Il + t~ ~ to be continuous on the right at (2.26) lim ~-> 0 Hence, we can certainly find a o < ~I ?lz (t) = O. We now prove We have for any t = 00. e = 8e > 0 < 8£ => - 0 ~ ~ -< 1. /')"1. ~I(e) such that > V*{e) But by (2.9) this implies that + o < till - to+ = t +lll - 0 < e. Q.E.D. f. 0 31 By Lemma 5 and (2.16) it now tollows that lim IJ. -> + 1Jl 0 (t-) = "*(0) IJ. , IJ. + and this in turn implies that t~ • 0 are the limiting solutions to (2.17) as IJ. ---> o. By (2.16), (2.29), tor arbitrary but fixed IJ., 0~ IJ. ~ 1, the point· + + (t, v) = (~ , v*(t~) ) lies on a curve of the family (2.3), (2.4) and hence satisfies "the relationship (2.2). This gives us the following identities in IJ.. which may be written , Thus t+=t l >IJ.=II t~ = t' > IJ. = IJ.t , IJ. t'" On the other hand, O<t'<t+ - 1 t"'t" I t-1 -< t' -< 0 say. 32 (2.34) Ilt = III ==> Mv*(t)(Il I ) I ~ ft(V) :I -> v=v*(t) v*(t) ·?ltlll(t). But, by the discussion above (2.17), this implies that Ilt = III ===> + t = t~t' 0 ~ 1. < III By (2.29), this holds, as well, in the ltmit as III ---> O. By Lemma. 2, is a continuous function of t, t l(2.35), and by Lemma 5 S t S t +l • By (2.32), (2.33), t~ iS a decreasing function of t J Ilt t has a unique minimum =0 St <0 at Il = 0 is an increasing function of t, 0 < t ~ t~ • f' -=+-_._------~/<. Figure 2.4 We thus arrive, finally, by (2.16), (2.29) at the important identity v*(t) == 7J? v*(o) = Ilt (t) lim t-> 0 J ?Jz t~ ~ t ~ t~, (t) Ilt t I: 0 , CHAPTER II ABlMProTIC PROPERTIES OF BAYES TWO-STAGE TEST 3. ASY!2totic Exiression for the Second Sample Size Function. Equation (2.5) may be written in the form where ~ is any fixed number in the halt open interval 0 < If we now substitute into (3.1) the solutions t + = t-(Z), ~ ~ 5 1, ve set the tollow1n& result, where (,.6) I + + Bote that E:~JJ.(Z) and hence E:~(Z) are positive decreas1ns .... It"... (20) j tor f'unct1oa. respectivel)" at sufficiently larse values + at It~ (Z) I. Bence I by Lemma 4, they are posit1ve decreas 1n8 t\mctiODS ot Z tor sufficiently larae Z. Also 35 Now we define + (3.9) ei.O(z) = ~ .::: + 0 atll} e~(t~(Z» = (3.10) Clearly, then, the relationships (3.5) will hold for Jl=O. By (1.26), (1.16), v*(O) is the unique solution in v to the equation log v = 2 log / dZ (~+~1) --2./2i -11 - 1 d 2v • - '4 36 This may be written Substituting into this equation its unique solution in (3.12) 2 log Z = t \I, we have '2$ O~o+!i1 ) 2 d v*(0) + log v*(0) + 2 log d Now, as Z increases, the R.H.S. of (3.12) must, to maintain the equality, also increase. function of v*(0). But the R.H.S. is an increasing Hence \1*(0) is an increasing function of Z. Further, we have that lim Z~ + 00 v*(0) = 00 + We can now see that €io(Z) and hence €~o(Z) are positive decreasing functions of Z for sufficiently large Z and that lim \1*(0) ~ 00 o. 37 Hence, finally, for all o< - ~ -< 1, ~ in the closed interval + E- 2~ (Z) are positive decreasing functions of Z, for sufficiently large Z, and Z lim ~ 00 + E- (Z) 2~ =0 , 0 _< ~ _< 1 By (2.32) - (2.35), we have the identities (3.16) Thus, by (3.5), we have where o~ t ~ t~ (3.18) E- 2 ~t (Z), t~ ~ t ~ 0 • 38 By some simple manipulation of (3.17), we get,for t~ ~ t ~ t~, t ~ 0, The case, t=O, is excluded to avoid dividing by zero. roots of this quadratic in (3.20) 2(1-Pt(Z)10g Z dlt I ~t are + - 1 - To choose between these roots, we observe that that 0 ~ ~t ~ 1. The 1. ~~ is real and Hence the correct root must be so restricted. First, if either root 1s to be real, we must have 2(1-Pt (Z) log Z d It I - 1 > 1 - or <-1 If the second of these two possibilities were true, both roots would be negative. It follows that the first inequality of (3.21) must hold. However, in this case, the root (3.20) With the positive second term would always be satisfy our requirements. ~ 1 and hence could not There remains only the root with negative second term, which by the first inequality (3.21) is readily seen to satisfy them. Thus, By (2.36), • I!t is continuous and equal to zero at t=O. limit of the R.H.S. above as t -> relationship holds in the limit as t 0 is O. ---> The Hence the above O. Now let Ht) = d\t I log Z ~--:i!'- Using (3.5), we have o = ~(O) ~ t{t) ~ t{t~(Z» Both upper bounds are by (3.14), = l-€- (Z), t~5 21 <1 t ~ 0 for sufficiently large Z. Substituting t(t) into (3.22), we have 40 and this relationship holds in the l1m.1t as t rearrange the R.H.S. ot ~ O. We now 0.25) tor greater convenience. where, again this holds tor t~ in the closed interval t 1- ~ t 5 t +1 and in the limit as t ---> hgain, rearranging the R.H.S., we have where 0, and a.nd we define Finally, we ca.n write 1 ~2 _ t- where l.(l.~(t»~ + 2Pxt(Z) 1+(l-Ht»~ .I • , < t l _t:5 t +l ' 42 and by (2.39), this relationship holds in the limit as t By (3.17), 4(1-Pt(Z) ) log Z t d(j.Lt + 1)2 By ---> o. (3.31) , Thus we get # o. 43 where (3.37 ) p 4 (z) P (z) + P (z) (2. + p (Z» = ---.;t_ _-.::3;..;.t ,3t _ 2 t . (1+ P3t(Z» Substituting (3.36) into 0.33)" we have (3.38) v*(t) ... 1J2. (1- P (Z»log Z, ti ~ t ~ t~ , ~d Il+(l-t;;(t»:2 6t where (3.40) Using (3.14) together with the conclusion above it" we have by (3.18) that (3.41 ) lim P () Pt(Z) > 0, suff. large Z, Z~O t Z = 0, t -l ~ t ~ t + l ' 44 and hence it can be shown in each case for t 1- ~ t ~ t 1+ ' Let (3.43) u(t) =1 + (1 - ~(t»i , then by (3.38) 0.44) 2 2( t ) ( 1v*(t) = ~u d P 6t ( Z» log Z, t 1- ~ t ~ t + • l Substituting this expression into (1.11), we get + h-(v*(t),t) u2(t)(1_ P6t (Z» ~ -1 dt~ >. Q~ 1 ... = --u(-t-)-(l--~P-(-Z-»)""'~-';;';--·(-~ log Z)2, 6t Rearrangem.ent of the R. H. S. gives us 't'lhere 1 ( (1~(t»2 , J(t) =( 1 , - t<O - t > 0 45 vie shall use the following v1e11 known result. e.g., see [2] • 00 (" -ix 2dx (J.48) j e III e _ 1 2 -1 N "fJ y z c .y" jeO J ,I y 2j + (-1) N+ L ~N' N III 0,1,2, ••• , where y is a positive number, , j . 0,1, •.. , (J.49) 00 R = (2N+2)J N 2N+1(N+1)! ( X..2(N+1) e" ix / 2 dX, N = 0,1, ... y and uhere Lemma 6. Let d, Wbe arbitrary but fixed positive numbers. :fe(Z) ~! (1 - 8)10g Z , Let , 46 then for any fixed positive 8 < 1, no matter how small, we can by taking Z sufficiently large, obtain for all t in the closed interval the inequality Proof. By (1.72), the lemma will be proved, if we can show that ~ lim Gv*(t)(t) Z->oo dt 11m Z -> e 0 , - ~8 ~ 0 ~ 0 , . ,-~ 0 ~ t ~ ,) 8 = 0, Gv* ( t ) ( t ) t By (1.74), (1.78), respectively, (3.54) and (3.55) will be true, i f lim Z -> G" e lim Z -> 00 (- v'(:1 6 ) 00 dJ a G v. ,," ( J 8) J 5) ( = 0 '1 s:.) = 0 v We first note that for any fixed positive 8 small, we have by (3.5), (3.8), for ~ = 1, < 1, no matter how that 47 for suff. large Z. Hence, by (3.42) lim Z -> p 00 6,!. J e (Z) mO. We shall prove, here, only (3.56). (3.57) can be proved in strictly analogous fashion. By (1.73) 00 +...!.... /2i ) x2 .-i ax + e $ + h (v*(-J5)'-'J 8 ) By 00 d 18 / e ¥2 dx h- (\1*( - r 5)'· r5) (3.46) = 12(1 - where ~ (Z) log 5 zY , 48 P 6,-1 8 1 - P 6,-1 8 and by (:3.59) p (Z) " Z -> 00 8 lim = o. By (3.48) - (3.50), taking N = 0, Y = (3.61), we have that J 00 .-.;x2dx h-(l'*(-J < ~ [n (l-~8 (Z)log z]-l -l+e Z (Z) 8 e),- '1 e) On the other hand, Hence " (3.66) < P (Z)-8 1 '" 1 "2 [n(l-p (Z)log z]·~. Z 8 • e But, by (3.63), the R.H.S. and hence the L.H.S. Z -> 00. By (3.46) -> 0 as 49 1 (1 - p 6,. (2 log Zr~ r 1 i 8 • ---> 00, 3S Z ---> 00 and (}.60) -> 0.. as Z -> 00. Clearly, this R.B.S. of Finally, by hence the second term (3.44), (3.65), the f1rst term R.B.S. of (3.60) may be written 2 2 i (1 + 8) 2 (1. P 6,. J 8 d and this also ---> 0, as Z ---> 00 ). ~~o leg Z 2. + li1 log 6 zJ Z • Q.E.D. As a direct consequenee of Lemmas 3 and 6, we now have the fol10\Oling. Theorem. 4. Let d, W be arbitrary but fixed positive numbers. ii, t*' be the two numbers defined by Lemma 3. fixed positive 8 Let Then for any < 1, no ma.tter how small, we have, by taking Z sufficiently large that :r8(Z) < ti(Z) tt(Z) < t~(Z) < t*'(Z) < - , J e(Z) ' 50 4. Expansion of the Second Sample Size Function. In this section, l'le expand the function, v*(t) in terms of the parameter Z, to terms of order (3.38), 0(10; Z); our result holding for all t in the closed interval, t ~ J 6(Z). In the following pages, we shall for convenience, and where no confusion is likely, make use of the abbreviated notation indicated below. (4.1) p = p (z) , p t ~ .. ~ ( t) .. (4.2) (4.3) vIe i ~ .. (1 ... ~;)i , u. d = (Z) p , i • it 1,2, .•• • It.I log Z 1+t , w· 1 ... t • shall, :for the same reason, 't-lhere no confusion is likely, drop the arguments from functions else't-mere introduced. For reference to the original definitions of the infinitesimals (4.1), i = 1, •.• ,6, see (3.18), (3.21) ... (3.40), in the previous section. By (3.18), (3.6), 51 where (4.5) By (3.7), (3.15), (3.16), (3.34), we have, for ti ~ t ~ t~ , (4.6) so that by (4.4) JI where (4.8) 52 (4.9 ) = p p 7 + p 71 J 72 , (4.10) f( 2 b log2z + (l-p) 1 2)2 _log Z] J (4.12 ) We note, first of all, that for all Z > 0, t- ~ 1 t ~ t +l ' and that (4.14) lim (Z) p Z~>oo 7t ~ 0 , • Hence k p k = -p _ t (Z)] = o(~), k ~ j+l, logJZ j = 1,2, ...• and this holds for all t in the closed interval ti ~ t ~ t;: . The results which follow depend upon (4.15) and are obtained by straightforward algebra1Q methods. presented without detailed derivation. They are (4.10) and (4.11) can be put in the following form (4.16) Adding these together and using (3.31), we get (4.8) can be put in similar fom. (4.19) 13'" log v - ~ - ~+t P ~u U +0(101 z) , n .!. g It I ~ l'mere (4.20) v ... v (Z) ... t ~ d .u.(W(t»2 • J6 J $4 Hote that here t'le find the remainder to be of the indicated order only for t in ~he closed interval, J 5 is defined by (3.$1) and again 5 ~ t J 5' where ~y be any positive fixed number < 1, no matter how sHall. Substituting (4.18), (4.19) into (4.7), we find after one iteration that for It' (4.21) P = lb~ Z (lOg V - ~) _~l_ 10g'2 Z (22 log V + u w2_~t~ 4u ! where (4.22) • After, progressively expressing P to P in terms of P, we l 5 arrive at the follo''1ing result for P6' valid for It I $ (4.23) P 6 = 1 2 u 10g Z + ",2 2 r1 P + ~ P 4t 1 4 2 2u log Z + J e. o( log12~. i.) Substituting into this our derived expression (4.21) for P, we get, again for It I ~ Je ' 55 (~ V (4 .24) P6 -_ t log log Z log V- ~ + 0 ( lO~2z) Finally, we have from (3.44) that (4.25) v* (t) = 2 '2 d U 2Lrlog Z - P log Z 6 - 7, t- < t < t +1 ' l -- so that the desired expression for this function in the interval I tl (4.26) V*(t) S 15 is = 22u2 ~Og d ~ Z - f- log V _ 2 10g V 2t log Z (~ log V - ~] , S6 5. Expected Value of the Second Sample Size Function. By our original assumption (1.1) and by (1.5) the frequency function of t = t m is , where ~d q We - ~ log A , -imd - ~ log A , = Q { will need the following Lemma 7. 0;0 Pg(t)dt r i~ 2 < _/2i t3 _ z .. 1- log Z log Z _ -00 where dqQ ( Proof. ) 1 .. 8! log Z 1 ' 00 fJ8 By (3.48) - (3.50), we have, taking N=O, and recalling the definition 18 , (.~.51) of (5.6) . d (: ~1-8! 16~ z - _.(2i. t3! ~~ Z 1 Ei:. dO) - ~ (1-8+e 2d m ·dqQ 1 og 2 2 Z) log Z :: log Z -1 log Z - ] log Z Q.E.D. As a result of the above lemma, we have, certainly that -18 (5.1) f Pg(t)dt , -00 all finite positive j. 8 we define the notation, EQ , by the identity 58 J (5.8) a J ')( 8 Eg /((t) " - j (t) Pg(t) dt 8 By Theorem 1, , where (5.l0) By Theorem 4, Lemma 1, we have, for sufficiently large Z, 00 0< "'1 < max Lv*(-J a),v*('J a}_7· By (3.38) so that, by (3.59), (5.7), f 1a Pg(t)dt + 59 Renee, (5.14) For the remainder of this section, all relationships in t should be understood to hold for all t in the closed interval It I ~ :re• By (4.26), + 2 l~g Z 5 Eg (;: log v) - n~g z + 5 Eg ( ~~ lOg~ (10~ z) 0 First, we note that (5.16) ~j+l = I d ) j+l \iog Z • ItlJ+1 =o(lcg1..Z) . Itl j +l J j = 1, , 2, •• 0, We shall now consider, one at a time, the terms on the R.R.S. of (5.15). Using (5.16) we have that ~ so = (1 _ ~)~ = 1 _ ~~ _ §£2 _ 0 ( that (5.18) u2 = 2 - ~ + 2t =4 - 1 ~2 2~ - 4 - 0 1 ). 2 log Z t 12 ) • log Z It 13 , ,3 , It 60 Hence, 828 6 U == 4 EO (l) - 2 EO Eg ~ ... ~ 182 '4 Eg + 0 ( . ) 1 2 log Z With the help of Lemma. 7, we have, atter some integration, that E 6 Q _ t 10 ~- logZ + 0 ( 1 ) ZloSZ where 1 de -2m dQ.g + - .& Thus, We may now consider the second term R.K.S. of (5.15). verified that 2 2 ~==4+.J; tu Thus, by (4.22), (4.20), It is easily 61 (5.26) u 2 r log V = 4 log V + .!2L.v + 'log log Z d2 t 2 t = 4 log V + 0 ( u 2 10g2 Z 10~ zJ · t 2 • By (4.22), (4.20), (5.27) log V =i log log Z + log 4 q= + log (W (t»' + log u. d Now, u =1 (5.28) + t u =2 Hence, using may be written 2u = log 2 - t) 1 - ~ ~ + 0 (10~ z). ~2 2~ (4-t)u t 2 (5.26) r 2 (5.,0) P = 2(1 - 41 (5.16), it is easy to show that log u Thus, by 1 1 t2 - 2~ - ~ log V = 2 log log Z + 4 log 8 so that we have q= d + 4 log ('tif (t»' • 62 where Now consider the third term R.H.S. (5.15). We have , so that (5 .~~4) 1 2 log Z = • u 2 ~ log log log Z 1 og ~+ V 2 log V = log Z 2 (lOg log Z \~ §..E d2 + log ('Af(t»i) ) • Hence, (5. 35) 1 2 log Z E 8( U 2 9 \ t2 vr\ log ) = log log Z + log Z 2 (1 8 IiC + \ log Z '\ og d 2 R29) By (5.16), it is clear that the fourth term R.B.S. (5.15) is 0tlo~ z)' 63 Collecting terms (5.24), (5.31), (5.35), we have, consulting (5.14), (5.15), (5.36) EgV(t ) 8 ='2 log d Z /log Z ~ 2 log log Z - "2 '4Q + ~2~";;;;;':1iL.,;;, d d log Z + ( 0 1 where (5.37) 15Q = 414Q - d (5.38) 1 49 J 1Q and 12Q =2 log 2 2 (m + ~) 8€ 2 + 13Q , ' d are defined by (5.22) and (5.32), respectively. With respect to our apriori distribution (0.15), the unconditional expected value of the second sample size function is (5.40) Ev(t) = ~ d where we take log Z Ilog Z _ ~ d ) \log Z £4 + 2 ~Og log Z + £5 2 d log Z 2d l0g Z ' 64 j • 1, 2, ••• Note that n ~4 = 21og 8FC+s ~ ~3 d and, by (5.2), The difference between the expected values of the second sample size function at Q and at QO is l + W1 Wo This difference is, of course, equal to zero, when -- gl =-go = 1. W l When the ratio, W = w- is close to one, it may be con- o venient to use bounds for 1. 2Q • By (5.32), (3.4), we have that 65 6. Error Probabilities. The probability that our Bayes decision rule accepts alternative Al when the true parameter value is Q is given by (6.1) This may be written in the form where, as before, we write t for t • m By Theorem 1, (1.82), the second term in (6.2) may be written t it i 00 J PQ{t) EQ[t..{t) I t } t + -00 Pg{t) EQ[·m{t) t I t Jat. i By (1.8), this gives us (6.4) '* f t -00 00 Pg{t)Pg(t > Olt)dt + [ t t f 00 Pg(t)Pg{t> Olt)dt = t t Pg{t) dt • 66 The first term R.B.S. (6.2) may be written f t tt Pg(t) Egr'm+V(t)(tm+V(t)) if I t ]dt - 1,- t'" ·f t ( 6.5) Pg(t) Pg(tm+v(t) >0 It) dt if t. .t" = f Pg(t)Pg(~v(t) t > Qv(t) • t i t ) dt J '* where -Q is defined by (1.4) and ASv is normally distributed with A mean Qv(t) and variance vet). Note that Sv is a generalization of s~ It follows that (6.6) (1.2), for which v is not restricted to integral values. PQ (~V(t) > ov(t) - tit) 00 1 =- f hQ(t) where , 1 (6.7) h (t) t r.:r;:"\ - ~ d ,v\t, - = (Q-Q)v(t)-t = fV[tJ = -h+ (vet), t), Q=Q1 rvm Q Thus, the first term R.H.S. (6.2) may be put in the form (6.8) ~ r f t 1 -(2i Pg(t) t oo _ ~2 e 2 dx dt i' Hence, + t~l- 00 (6.9) QQ = j t Now, by PQ(t)dt + i -%C 1 r t i' 00 1 Pg(t) J (6.11) 'iX dx dt. hQ(t) (3.46) and Theorem 1, t where e- ~ i' <t <t i- , , 68 By (4.23)" we have (6.12) p 8 = log I log _logt Vl QS2z 2 + 0 ( 1 _ ), 2 log-Z By (6.9), QQ may be written in the following form 1 Q aQj2i (6.13) f 15 -1 5 Pg(t) 00 f e-ix2 dx + IQ(Z) hQ(t) where By Theorem 4, -15 f (6.15) Pg(t) dt -00 7, the R.H.S. of the above inequality is of order less By Lemma than 1 Z log Z • , Hence (6.16) By (6.13), (6.7), (6.10), (3.47) 00 0 r J (6.17) Q" =10 I2i PQO(t) ..18 ._p2 dx dt !2(1-p )l9g Z 8 1 +-!... ~ f 00 f Po (t) 0 0 / 2(t _ix e 2 (l dxc1t + 0 Z log Z 1• 2 -p 8)10gZ For the remainder of this section, all relationships in t should be understood to hold tor all t in the closed interval It I ~ 18 • By (3.48) - (3.50), we have, taking N = 1, 00 (6.18) f 1 2 e-2X dx - (l-j:) )log Z = e 8 1.r-2(1_P )log Z 7- 1 L 8' 1. (2(1-p )log Z)2 8 .J2{ l-p )log Z 8 + Rl ' 10 ~ o < R1 < 3 ~2(1-P ) log 8 Z-l - -(l-p )log Z 8 &. 2 e By (6.12) ( 6 .20 ) - ( 1-p8 ) log Z log V (1 ') log Z + log V - 2t log Z + a log Z 'J =- so that -(looP )log Z (6.21) 8 e c Ye- log V (1) 2t10gZ + 0 log Z Z _ v 110g Z ( - Z log V ~ (1) 1 - 2tlog z) + 0 Z log Z • Thus by (6.19) (6.22) Again, using (6.12), we have that (1 - p8 )-! = 1 + log V + 2 log Z ( 0 1 ) log Z ' (6.23) /2(1- - Hence p 8 -1 1 7 ... 2 1og -. )log Z . 1 ~ + o{~) .Log (., ' 71 I 00 (6.24) 2 .-ix dx =; . Z~ -2~~/i)r /~~gV~(l - 2l~g ~ + 12(l-p )log Z 8 If we multiply out and substitute for v its definition (4.20), the R.R.S. above may be written Using (5.16), we find that (6.26) u = 2 _ ~~ + 0 ( 1 ). t log Z 2 ' $-12LY_ 2 ~ log Z - 0 ( 1 ). log Z Thus we have ! 2 00 .-ix dx d~ (W(t»~ (2 -~t - lO~ z) = 12( I-P )log Z 8 + 0 (z ~g z). (1 + It I + t 2 ) Now consider the inner integral of the second term R.R.S. (6.14). By (3.48) - (3.50), taking N = 1, It I 72 1 - ~2(~2_p )log Z 8 - 7- 1 _---:::--._--:~:__..,.-- (2(t 2_P ) log Z)~ 8 o < El' < 3 ~2(~ 2 _(t 2 _p) 5 -p ) log Z 8 - 7 - - log Z 8 2 e By (6.12) ( 6 .30) - (t 2 - p ) log Z = - log Z 8 + d It I + log V - 2 tOg V log Z +o( log1 Z ) ' so that -(t 2 -p )log (6.31) e 8 Z=Y log edltl e- Z = v V (1)Z 2~10g Z + /lO~ edit I (1 _ log V ) Z 2 log Z ~ .. 01. ~ 1 :\. e d \Z log ZI Thus by (6.29) log 0 It I + RI 1 ' 73 RI _ 1 - ( 0 1 Z log Z ). e d It I By (6.12) (6.33) (t 2- p ) -~ = t (1 + ~Og V).) + 0 8 2t log Z (1 1z) I og Hence + 0 \Z[ 1 log Z ) • e d It I • If we multiply out and substitute for v its definition (4.20), the R.H.S. above may be written + ( 1 ). d It I o Z log Z e 74 Using (5.16), we tind tbat , (6.36) , llis....!..- ~, = 0 I(logZ 1 ). It I Thus we have f 00 (6.38) - 1 ~ + It is easy to verity e-!x 'C''Z 0 2 dx *" = (V(t)t edit! (2 + ~t - lo~ z) 1 ~ ) • (1 + It.I + t2).adltl to€, y. by (5.1), (5.2), that edt Pg o (t). ~1 Pg () t 1 Hence, recalling the definition ot ~(t) (3.4), we have by (6.17), (6.27), (6.}8) 75 PQ (t)dt o + Now by Lemma 1 d Z log Z t PQ 7 J6 f PQ(t)dt = 1 + oil. l~g z) , -16 (6.41) 16 f -16 Hence (t)dt o t PQ(t)dt = qQ + 0 ('~) , -to (6.42) By (0.3 0 ), (0.35), (1.18), (5.2) so that By (6.44), we may write this 11 By (6.13), (6.16) (6.41) Q Q1 = 10 f Pg (t) 1 1 1-- flIji -J O By (6.41), (6.1), (6.10), this may be written o f 00 I Pg1(t) -1e 1 e'"2X 2 dx dt 00 -1 :P (t) Q 1 o f 12 e-'2X dxdt + {2( I-p ) log Z 8 By (6.21), (6.38) and Lemma 1, this becomes Finally, by (6.43), (6.44) 0 ~l~gZ }. 78 By (6.46), (6.51), we have for large Z, taking only the leading terms of the expansions, .... = ~l ~~-~ • 2 gff d Now regard d, g to be arbitrary, fixed, positive, and choose W and Z to be, respectively, the following functions of d, g: 4 , 1 d2gcPO °1 (6.53) WI (d,S) = - °0 g , , Z'(d,g) = 4 2 d glal where °0 , °0 g<-0 (Xl are small given positive numbers. > If °0 , 01 are sufficiently small, and W, Z chosen as above, then we Mve, approximately 79 , (6.54) Let W0' Wi be any values of W0' • Wl' and c I any value of c, such that W'1 -W' = W' o , Since our two-stage test is a Bayes solution, (consider now the test with parameters, d, 8, W', Z·), its average risk is minimum, i.e ., where Ein represents the expected number of observations, at Qi' of any other two-stage test with first sample of size m, and its probability of rejecting Qi' when true, 1 = 0, 1. It follows that if i = 0, 1 then 1 (6.58) m + I. g1 Eg v(t) 1=0 i < 1 L i=O g1 Ei n ai, • 80 Now, by (5.36), for small values of the error probabilities, a., 1. i.e., for large values of Z' , Hence, since go' gl are arbitrary, we have asymptotically for small a i ' (6.60) , 1 = 0, 1 • Thus, asymptotically, for small values of the error probabilities, the Wald property mentioned in our introduction, is seen to hold for Bayes two-stage tests. 81 7. Comparison with ane-Itage Bayes Solution. In this section, we shall contpare one and two-stage Bayes solutions to our problem, for large values of Z, in terms of expected sample size; requiring of Ule one-stage solution, that its error probabilities be the same as those for the two-stage procedure. If we take a single sample of size m + v, the Bayes solution is given by the decision function t l (7.1) 'm + ; (tm + ;) =f m+ ... \I > 0 , o < The probability that we shall accept alternative A , given that l the true parameter value is 0, is 00 (7 • 2) Q ,9 = Pg (tm+v... > 0) = 1 {2n(m+~') J o where <ix, 82 1 - ~ ~ - =E q .Q t - Q m+ v d .. a1 log A , 9 • 91 , • 1 .. ~ n d - 1 d log A , e "" o Q - n =m+v This may be written 00 .. 1 ~x e 2 dx, y.~ where 1 (7.6) 1 1 i ~ 1 -- ~ y. = -(-1) dn + - log A·n ~ 2 . d . , i=O,l. i=O,l, 83 ... To determine what value of n is required so that the error probabilities (7.5) associated with this one-stage procedure will be equal to the error probabilities Qg , 9 of our two-stage o 91 Bayes solution, we set 00 1 2 -I'm 1 e "''2 x dx :02 Q 9 0 This gives us 1 (7.8) YO 1 =~ ~ dn . + 1 d log 1 ...'" '2' ~'n . = X(QeO ) =X where we define the function X(a), by the identity 00 -I2i 1 X(a) 0 ,say, 84 Solving for (7.8) for ~, we get (7.10) If we expand the radical on the right hand side above, we get - = ~(Xo 4 2 - log A) - O(X-2 ) o d (7.11) n . Some investigation shows that for positive x (7.12) X . 2(-1) x = 2 log x - log log x - log 4n + 0 (log i loS x) og x By (6.46), (7.13) so that setting , 85 , we get 2 2 log x = 2 log Z + 2 log (7.16) log log x = log d go 1 4 E + o(log z) l log Z + O(10~ Z) Substituting into (7.12), we get X2 = 2 log Z 0 jiog Z + 2 2 log go 8 !!l d rn + O(log 10~ ~) log Substituting this, in turn, into (7.11), we get, recalling the definition (0.35) of A, 86 If now, on the other hand, we set 00 -/2n 1 12 e- , X 1 , dx .. Q ~l we get 1 (7.20) -Yl .. 1 ~ ~ I I -dn - a log l·n ~ =0 Q~ X(l - ) I - Solving for nand proceding as above, we arrive at exactly the same result, (7.18). A comparison of the leading terms on the right hand sides of ('1.18) and (5.36) now clearly shows that the average number of observations required by the two-stage procedure is asymptotically equal, for large Z, to the number of observations required by the above one-stage plan, i.e. (7.21) lim Z ->00 m + Eev(t) =0 - n 1 87 Now the one-stage plan is a degenerate two-stage procedure of the type (with first sample of given size m) we have been considering, - which has its second sample of constant size, v, for all possible observations in the first sample. since it minimi~es Our two-stage Bayes solution, the average risk With respect to the class considered, will require not more observations on the average, than any such one-stage plan With the same error probabilities. The result (7.21) indicates, however, that the larger the value of Z, the slighter will be any improveme~t over the one-stage set up that results. A similar comparison with the Sequential Probability ratio test, using Vald's approximation to the expected value shows that the ratio of expected values, two-stage over sequential probability ratio, approaches 8. 4, in the limit as Z ---> 00. A Trivial Asymptotic Solution. We have seen, by the results of section 7, that the two- stage Bayes solution to our problem is, for large values of the parameter, Z, little better, in terms of the average number of observations required, than a one-stage Bayes solution with the same power. We now examine the possibility of an asymptotic solution to our problem when d = Q1 - eO becomes small, other 88 parameters remaining fixed. That this leads to a trivial result, is indicated by the following Lemma 7 Let Z, W be arbitrary, fixed,positive numbers, then lim d ->0 vet) == 0 (8.1) , all t. Proof By (1.22) lim ~ 2 d _> 0 log fll(t) = log t (8.2) which is an increasing function of = -00 at It I (8.3) = O. It I , with unique minimum By (1. 2), (1.17), d I' ~> 0 f t ()7?(t» = -00 , all t . + It now follows from the definition of t- (see discussion above (1. 24», that 89 + t- "" 0 d->O lim (8.4) • Thus, by theorem 1, d J.im.> 0 v(t) ~ 0 , all t. Q.E.D. By some simple additional calculation, we then have f (8.5) li.rn _ lim d ---> 0 Qg - d ---> 0 E Q 'm(t)= ( 0 : , A > = < • 1 CHAPTER III NON-ASYHPTOTIC CONSIDERATION OF BAYES '!WO-STAGE TEST 9. Further Properties of Second Sample Size Function. Throughout this section we shall regard the parameters m and d to be arbitrary, positive, fixed, m, an integer, and consider the second sample size and its related functions in tenls of their dependency upon the loss ratio, W '" Wl!t'10 and the ratio, Z =min(vl ,W )/c. 0 l By theorem 1, we have for arbitrary fixed W, Z > 0, the inequalities , (9.1) + + ,..,,. th e pro of 0 f 1 e~una, 3 t* - t* (W, Z) are the respective where~~ _ unique solutions in t to the equations , +, and by (1.21) - (1.24), t- + = t-(W,Z) are determined uniquely as the positive and negative solutions in t to the equation 91 Let , y ... dt then by (1.23) 1 (9.5) ft(?Jz (t» ... 2 log dZ 2 vTn - I Y' ... (y2+1)~-K(y,W)"'f(y,W,Z),say, where for all y, all W~ 0 (2 10g(W+e-Y), 0:: 2 log r",r +W e- Y] ... -~-l 1 \2 w~ 1 10g(1~-Y),W~ 1 ,Y?fJ (9.6) K(y/oJ) "" =(2 l2 log(l+W"Y) ,0 .:: W ~l ,~ log(}"Y) ,\4 .:: 1 o. 92 Note that for arbitrary fixed y, K(y,W) is a continuous function of~. - For y > 0, it increases from -2y at W• 0, to a unique maximum • 2 log(l+e- Y), at W• 1. zero in the limit as W---> 00. It then decreases, approaching The picture for y ~ ° is immediately seen from the relationship K(-y,W) - 1 K(y,t:1) This clearly holds for all y, all W > 0, and in the limit as itl" 0, (9.8) 00.. Hence, over the same range K( 'y I ,0) ~ K(y,lV) :: K(y,l) It follows, by (9.5), that for arbitrary fLxed Z and over the same range of values of y and W, and (9.10) - - f(y,l,Z) < f(y,1v,Z) < f( Iy I ,O,Z) 93 It may be inferred from this, and from the description of K{y,W) that for fixed y,Z, f(y,W,Z) has a unique minimum at W .. 1. By (1.22), the R.H.S. of (9.3), (9.11) and this function is clearly independent of W. Now (9.12) are the unique positive and negative values of y at which each member of the family of curves f(y,W,Z) intersects the curve g(y). Hence by (9.9) ... (9 .13) - . t-(W,Z) ,. -t"'{~, Z) , w,Z > 0 Unique intersections at finite positive and negative values of y occur also in the limit as W---> 0, 00. Now both upper and lower bounding curves of (9.l0) are symnetric about y intersections with g{y) are thus respectively at = O. Their 94 By the nature of g(y) already described in terms of the variable - t (see (1.22», it then follows that for Z > 0, W > 0, and in the limit as W- > 00 (9.15) t+(l,Z) < ~ t-(W,Z) < t+(O,Z) + - - For arbitrary fixed Z > 0, t+(W,Z) has a unique minimum and t-(W,Z) has a unique maximum, at W• 1. Both functions are monotone for W< 1, and monotone (in the opposite sense) for W> 1, and both can be shown to be continuous By (9.1), in W. + we then see that t*(W,Z) are certainly bounded functions of W. The function v*(t) = v*(t,W,Z) is defined by (1.26) to be the larger root in v of the equation (9.16) , + and the unique root in v of this equation when t = t -. Let , Z ::II dZ 212it J 95 then by (1.16), (1.17), (9.6), we 2 (9 .18) C (.;) + log 'V z t +v m~ write this equation 2 • 2 10 g z - 2C It I z t -K( 2C i'W) Let t Y - -z , x v = '"'"2 z then the equation becomes (9,20) .!(Cx + x Iy I )2 = -log x - K(2Cy,W) If we apply the transformations (9.17), (9.19) to both sides of equation (9.3) and then solve for y, the two solutions in y which are obtained will be (9.21) • 96 + where y-(W,C) are functions of Wand C only. (9.22) v*(t,W,Z) • We then have that 2 z x(y,W,C) , where x(y,W,C) is the larger root in x of (9.20) whenever and the unique root in x for v at the endpoints of this interval. In keeping with the definition (1.26), we define x(y,W,C) to be identically zero for all y outside, and not bordering, the interval (9.23). For arbitrary fixed C > 0, let S(C) be the region in the y,W plane, W ~ 0, defined by (9.23) and including the bounding + curves y = Y-(';J',C). It can be shown by some simple arguments which involve the equation (9.20), that x(y,W,C) forms a continuous bounded surface over S(C). For fixed Y, the surface is monotone in VI, decreasing when 'H < 1, increasing ~lhen W > 1. For fB,ed W, its behavior as Y value may be inferred, by (9.22), from LmJmas 1 and 2. Now for arbitrary Z > 0, let SI(Z) be the region in the . + t,W plane, W~ 0, buunded' by the t axis and the curves, t - t-(W,Z). 97 Reversing the transformations (9.17), (9.19), we have, by the above, that v*(t,W,Z) is continuous and bounded over 8'(Z). For fixed t, it is monotone in W, decreasing when W < 1, increasing when W> 1. As a function of t, for fixed W, it is described by lemmas 1 and 2. In addition, we note that by (9.20), (9.1) x(-y,W,C) - 1 x(y,~, C) so that by (9.22) v*(-t,W,Z) = v~~t, 1 W' Z). The sy1metry of v* in t when W = 1 is just a special case of (9.25). On the other hand, using the results of section 2, in particular, the identity (2.38), (2.39), it is easy to show that for arbitrary fjxed t,W in the region 8 1 (Z), v*(t,W,Z) is a continuous increasing function of Z, all Z > O. (9.26) 1m Z ---> 0,00 v*(t,W,Z) - 0 , In addition by lemraa 4, Further, 00 for arbitrary fixed W, t+(W,Z) increases, t-(W,Z) decreases, continuously with Z. By (2.10), 8'(Z) reduces 98 to the positive W axis in the limit as Z - > 0, while as Z - > it becomes the entire upper half t,ll plane. 00, To relate the second sample size function to the result (9.22), we apply the transformations (9.17), (9.19) to the L.R.S. of both equations (9.2). The unique solutions in y obtained will be respectively + + y*(W,C) (9.27) -!z t*(W,Z) , - + where y~~(W,C) are functions of Wand C only. By theorem 1, the second sample size function may then be written -= (oz2x(y,W,C) * t , y (W,C) < y < Y (\l,C) , otherwise (9.28) v(t,W,Z) Finally, it can be shown that the points of discontinuity + of the second sample size function, t*(W,Z) are continuous in W and Z. First, 00 00 1 2 e-dt e dy+- 2Y I'm 1 2 "2Y dy e 99 - is clearly continuous in v, t, W, and Z; v, Z > 0, W > 0, all t. v*<t, w, Z) is positive and continuous in t, W, Z everywhere in [Ct, w, Z)a t-(W, Z) ~t ~ t+(W, Z), Z > 0, W ~ 01 . / Hence Gv*(t)(t) is continuous in t, 1~, Z over this range. As was indicated in the proof of lemma 3, Gvit(t)(t) is Jl1Onotone decreasing in t for any fixed lv, Z in (9.30). We may regard this function, tor fixGd Z as a continuous surface over the t, W plane. The trace of this surface in the plane of height 1 above the t, U plane will because of the continuity of the surface, its monotonicity in t for all - 1'1 > 0, and because of (9.1), describe a curve which determines a single valued continuous function of ~,. This curve will, 'i by (9.2), of course, be t (VI, Z). Exactly the same argument, with Z and W interchanged, shen'ls the oontinuity and single * valuedness of t (W, Z) in Z. Silllilarly, since e dtGv*(t)(t) is monotone increasing in t for all fixed W, Z in (9.30), ~ we oan infer the continuity of t (U, Z) in W and Z. 100 10. Some E~loratory CODi!tations in the ~ymmetric Case. To obtain a more exact idea of the behavior or our t,.,o-stage Bayes solution for intermecliate valuss ot the parameters, we investigate, by means of computation, the symmetric case ot equal 10s8es and equal apriori probabilities. i.e. in this section, we shall cons1der that W W-J-l, o Since W8 are takina W• 1, we cease, tor eonwnience, to indicate it as an argtRent in tunctions dependent on it. We shall, hOl-leVer, l-rhere necessary tor unclerstanding, indicate functional dependence upon other parameters. When W • 1, the solutions to equations (9.2), (9.3) are Nspeotively symmetrioal about t (10.2) t+ • - t- • ~ , ::I O. say, t t .. - t 1. e. * - t * , say , and both tMB4l ftUIIberc -d8pand ~ upon Z and d. By (9.21), (9.27) .J ,.10.,3. ) 1 + -; ~ • 'Y '" - - ill( • '\"1' say, '* iIt* · Yt. •• Y • Y*, say • 101 i~ V1 and V' depend only on the parameter C defined by (9.17). By (9.1), lore have, indicating the functional dependence on C, (10.4) • From section 9, lore have that by using the t: ansformation 4 (10.5) , = 2CV y • dt equation (9.3) may be written The L. H. S. of (10.6) is a continuous, increasing, concave function of I y I which ... is independent of C. -00, 00, as Y ...... 0, 00, and which Hence the solution of (10.6) in IYI , namely (10.7) approaches zero as C ... O. Substituting this solution into the L. H. S. of (10.6) and solving the resulting identity in C for v1 (C), we get VICO) .i e .t + e 1 (e) + (0) & 2 102 where & 1 (0), & (c) ... 0, as C .... O. Thus 2 Rewriting (10.6) in the torm (lO.9) 1 1 ·l+1)" + log t<4C2y2+1)":Z -lJ + 2101(1+8..2Ctt~ -Jt)g 202 -0 2Civt +(41J 2 we find that the L. H. S. is, tor fixed C, a continuous, :Sncreasing, concave !'unction of t y I which ... 1"( I.... 0, 00, -00, 00, a8 while tor fixedy, it is a continuous increasing function of O. Hence the solution, f y t• Tl(C) is a decreasing function of C. It follows, by (10.6), that 1 e""""}. is a least upper bound for VI(C),and hence, by (10.4), j an upper bound for y*(C). By (9.28), the second sample size function v(t), ~y be written M< "'*(0) (10.10) .. > x.{ Y,e) is the larger root in x of the equation j 103 for all V such that I y I y 1= Yl(O), 1< v1(0), the unique root '£-men and identically zero for I y I> Yl(O). Olearly, this function depends on Y and 0, only and is symmetric about Y = o. ApplYing the transformations (9.17), (9.19) to the L. H. S.Js of equations (9.2), we have, for W· 1, 00 2 (10.12) -2...(l+e Oy) x(O,y) + - .f2n 1 J2n 1 2 20 Y -"'2Y e dy + !...- fOO '-if1 2 ffn / h- e h+ where + (10.13) h- = 0 J'x{y,tn Y - + Setting (10.12) equal to 1, we obtain y*(O) as the unique solution in y. Our decision rule for deciding between QO and 91 may now be written: Take m observations. If the observed sample is Xl"." v=--Zt Xm, form the function a 1 - Z r m~ JC. _ 1a l 1 1 m (QO + 9 ) ] - -2 1 . dy , 104 If y ~ -y*(C), accept QO' ever, I y 1< ylf-(C), If y ~ y*(C), accept 91 , take v :I If, how- z2x (y,C) additional observations. If, now, the total observed sample is Xl"" ,Xm+v , from t he function If this is negative, accept GO' positive, accept 9 , We need not 1 consider the case t + mv = 0, since this has probability zero, Table I C y*(C) .01 .2536 .10 .2535 1.00 .2373 3.00 .1833 5.00 .1.472 10.00 .J007 20.00 .0644 50.00 .0334 100.00 .0196 105 In table I, we give, tor some selected values ot C, the values ot the function "(*(C) rounded at the 4th decimal place. These numbers were obtained by setting (10.12) equal to 1 and using this equation in conjunction with (10.11). Standard tables, which are listed in the bibliography, were used for this and the remaining tables in this section. From these results, it would appear that y*(C) 1s a decreasing function of C with L.U.B. less than .26 which is well below the upper bound given by (10. 8) . Table II: y/C o .01 1 3 x(y,C) , 10 ,0 100 .249994 .203888 .100859 .0,82611 .023601, .00194291 .000602768 .01S .249769.20.3663 .1006.34 .0580360 .0233768 .00173058 .000414810 .030 .249092 .202967 .0999,78.0,7.3602 .02770,4 .00115491 .045 .247961 .201855 .0988263.05623lO .0215930 .060 .246.367 .200262 .0972332.05464.30 .0200387 .075 .244304 .198191 .0951618.0525847 .0180149 .090 .241757 .195649 .0926143.0,00341 .0154112 .105 .238711 .192600 .0895485.0469495 .120 .235l44 .189027 .0659346.0432499 .135 .231031 .184902 .0817166.0387638 .150 .226338 .180186 .0168041 .165 .221021 .174829 .0710398 106 Table II (continued) y/c .01 1 3 10 5 .180 .215025 .168762 .0641129 .195 .208277 .161888 50 100 .210 .200613 .154065 .225 .192071 .145010 .240 .182252 y*(C) .172008 .136558 .0623741.0342506 .0129849 .000950055.000285660 In table II, we have tabulated, for some of the vauee of C in table I, and for values of y, from y = 0, at intervals of length .015 up to y = y*(C), the function, x(y,C). Entries were computed from (10.11) to six significant figures, the last figure being rounded. The expected value of the second s~lp1e size function and the probabilities of wrong decisions may be easily found in terms of integrals which involve x(y,C). In this symmetric case, we have, of course, (10.14) EQ vet) = Eg vet) = Ev(t), say; Q OleO = l-Q 91 = Q, say. 107 Also by (10.1) W (10.15) A--=1 g By (5.1), (5.2), we have, since vet) is symmetric about t :: 0, t* (10.16) Ev(t) :: 1 ~ [ 11 vet) e - 2m( t + ~d) 2 dt • -t Applying the transformations (9.17), (9.19), and the additional transformation (10.17) we get after some simple manipulation, y*(C) 2 (10.18) d Ev(t) ~ 8c 3 ----- e r ;rn lr2J . -i(fY) x(y,C)cosh(cy) e ~ o 2 dy • 108 On the other hand, by (6.9) .(10.19) Q=--L /2ilm f .1..( t e 2m Imd) ~ 2 dt" 1 2n t* where hQ (t) is defined by (6.7). o j rm t 00 -l~ 1 (t+1md) e 2m 2 -t* 2/ 00 .1y2 e 2 dydt, hQ (t) o Applying the same transformations as above, we arrive at the following result dr. Clearly, (10.18) and (10.20) are functions of C and r, only. We now compare the two-stage tsst of 9 v.S. 9 which is 0 1 defined by the second sample size function (10.10) and the decision function (1.8), with the analogous one-stage Bayes solution and 109 • with the sequential probability ratio test, in terms of expected sample size, requiring of these tests that they have error probabilities equal to Q. By (10.15), (7.5), the probability that the one-stage solution accepts Q when Q is true, or vice versa, is given by l O 00 1 2 Y e2 (10.2l) where n is the one-stage sample size. dy , Setting this equal to Q gives us 1 .!2 (10.22) where X (a), 0 ~ .:l~2 W! =X(Q) =X0' say, a ~ 1, ts defined by 00 (10.23 ) 1 -ffn f X(a) 12 e 2 dy= a - - y 110 Thus (10.24) Now consider the sequential probability ratio test of Q O Wa1d ~3_7 v.s. Ql' which has symmetric error probabilities, Q. gives the following approximation (a lower hound) to the expected number of observations required bv this test. (10.25) - 2 E = ~ (1 - 2Q) log d l-Q ~ . Multiplying both sides by d 2 , we have (10.26) Since, as we have shmvn, Q is a function of C and r, that (10.24) and 10.26) are also functions of C and only, it follows r, Now our two-stage procedure is a Bayes solution. average risk (10.27) only. Hence the 111 is a minimum among all two-stage tests with first sample of size m. When , (10.263) i.e. when (10.1) is satisfied, the average risk is (10.29) c t m + Ev ( t) ] + t:1 Q , and this is minimum among all symmetric two-stage tests with first sample of size m. By continuity results of the previous section, Q, (10.19), is seen to be a continuous function of the parameter, Z, other parameters being fixed. By the asymptotic results of section 6, we have that (10.30) Now (10.31) lim Z -> 0 ". 0 00 ~ 112 Hence, Q is a continuous function of C, and lim (10.32) C ---> 00 Q =0 • On the other hand, by (10.20), 00 (10.33) I lim Q =.1...C -> 0 I'8i 1 2 "2 Y dy. e r It follows that for fixed r, given any a in the interval 00 (10.34) o f 1 <a<- -.pn e 1 2 -'2 y dy , r l'11e can always find a value of C such that (10.35) Q(c,r) = a Since r > 0, this implies that, given any a in the interval o ~ a <~, we can always find a pair, (C,r), such that (10.35) 113 holds. Let (Ci~,r*) be a particular pair such that (10.35) holds. Recall the definition of Z (1.18). In the symmetric case, (10.28), this is Z=~c (10.36) Now let W*, 0* be any two positive numbers such that (10.31) w* *c * "'" Z "'" 4C*l'?R d Also, let m~(o, diE- be any two positive numbsrs, the first, integral, such that (10.38) '* "2'1 d* \1m'" r~r Consider the two-stage Bayes procedure with starred parameters and let E'>kv(t) be its expected second sample size. is, of course, a. The error probability Let S be any other two stage test with first sample of size m* for deciding between go and Q + d*, ESn, its O expected total sample size at these two points, and a(S), the 114 probability of wrong decisions in its 11se, and supnose that (10.39) a(S) :: a • Since we must have * c* 1-* '-m + E ,,( t) (10.40) ] + tv*C& :: * c*S E n + W a( S) , it follows that * * (10.41) S m +Ev(t):;En Thus, the Bayes solution in the symmetric case provides a lower bound for the average number of observations required by double samplin~; schemes with fixed first sample size. Table III c r Q .01 .05 .4800 .25 .4013 .4 .3446 .5 .3085 2.5 .0062 2 d (m+Ev) .010337 .250471 2... dn .010055 .250055 2 d i 11. R B R .006399 .99489 1.563 .014 .157995 .99918 1.582 .036 .640442 .640056 .399752 .99991 1.601 .042 6 .618047 .99995 1.618 .041 1.0 32 1.0455 25.0832 25.0375 10.0248 .99997 2.494 .055 115 Table III (continued) r C • 4. 5. 1. 2 d (m+Ev) 2dn .04317 64.01015 64.0483 12 6 .0 267 100.0 14 100.0480 d 2E 20.7166 30.1290 R 1 R s R 6 .999999 3.069 .0 2 6 .999999 3.319 .0 1 .05 .3258 .815 .816 .507 .999 1.609 .002 .25 .3249 .7271 .8251 .5125 .881 1.419 .314 .4 .2961 .9466 1.1412 .7060 .825 1.305 .455 .5 .2709 1. 2389 1.4885 .9071 .832 1.366 .429 .1455 4.0842 4.4613 2.5110 .915 1.621 .193 1.5 .0621 9.0302 9.4561 4.7562 .955 1.699 .091 16.0097 16.4543 1.3322 .973 2.183 .049 2.5 .0056 25.4535 10.1596 .962 2.461 .030 4. 25.0025 .04298 64.0411 .06270 100.0610 64.4529 20.8385 .993 3.071 .010 100.4505 30.247 .996 3.306 .006 1, 2. 5. 3 Q .0213 .25 .1722 3.5420 3.5763 2.0590 .990 1.720 .023 .4 .1742 3.2609 3.5172 2.0280 .921 1.608 .172 .5 .1695 3.1822 3.6563 2.1006 .670 1.515 .305 .1061 4.8506 6.2281 3.3592 .719 1.444 .480 1.5 .0477 9.3113 11.1271 5.4181 .837 1.719 .318 16.0982 18.0900 7.8753 .890 2.044 .195 1. 2. .0161 116 Table III (continued) r C Q d2(m+E\I) 2- d n d2i R 1 R s R ( 2.5 .0046 25.02,6 .04242 64.0312 4. 5. 5 10.6373 .924 2.353 .125 66.0531 21.2609 .969 3.010 .046 102.05 31.06 .980 3.220 .029 .25 .11.41 5.79 5.81 3.16 .997 1.830 .008 .4 .1167 5.4977 5.6807 3.1036 .968 1.771 .071 .5 .1169 5.2362 5.6725 3.0997 .923 1.689 .170 .0819 5.8002 7.7577 4.0430 .148 1.435 .521 1.5 .0385 9.6116 12.5018 5.9399 .113 1.628 .432 16.2134 19.4130 8.3032 .835 1.953 .288 2.5 .0039 25.0558 .04204 64.0327 4. .06187 100.0524 5. 28.3615 11.0144 .883 2.275 .191 67.3174 21.5943 .951 2.964 .013 103.30 30.99 .968 3.227 .046 1. 2. 10 .06220 100.0511 27.0730 .0138 .4 .063 9.28 9.34 4.14 .993 1.959 .01.4 .5 .0641 8.951 9.261 4.616 .961 1.915 .066 .0525 1.892 10.507 5.177 .751 1.525 .491 1.5 .0268 10.510 14.897 6.198 .106 1.546 .542 .0100 16.487 21.653 9.008 .761 1.830 .409 30.500 11.638 .824 2.159 .287 69.402 22.144 .922 2.890 .1l4 31.52 .949 3.113 .073 1. 2. 2.5 .0029 25.128 .04155 64.0362 4. .06143 100.0555 5. 105.31 117 In table III, we have tabulateo the function, (10.20), (10.18) plus d2m = 4r 2, (10.24), and (10.26). Computations were made for five values of C and selected values of t61 Gregory's formula of numerical integration r. was employed to evaluate the integrals in (10.18) and (10.20). All entries are rounded at the last decimal place given. ~lald and ,{rlolfowitz 11J (see introduction), have shown that the sequential probability ratio test minimizes, simultaneously'at go and"gl ' the eh~ected number of observations among all sequential tests v1i.th the same or smaller error probabilities. Both one and two-stage tests may be regarded as degenerate sequential tests. Also, the one-stage test may be regarded as a degenerate two-stafe test. We thus have that E< m + Ev(t) < n (10.42) and this inequality is clearly evidenced by the results in table III. To compare the two-stage procedure l~ith its one-stage and sequential probability ratio test analogues, we have computed the following two expectation ratios (10.43) , R s = m + Ev(t) E 118 If we consider the IIdistance ll (using expected number of observations as criterion) between the one-stage and sequential probability ratio teats to be 1, we can use the measure n- rm + Ev(t) 1- R =.... (10.44) n- E to indicate the fraction of this IIdistance II which lies between the one and tHo-stage tests. These l1lAasures of comparison have been tabulated in table III. In decisions between two simple alternatives QO' ~l' the problem usually specifies the distance, d, between QO and Ql' as trell as the probabilities, a, ~, which indicate the degree of allowable error. Table III, though inadequate for more exhaustive investigation, provides some insight into the behavior of our two-stage test, under these restrictions, in the symmetric case, a =~. Ex~lination of table III indicates that for fixed 0, the expected total sample size has a minimum w.r.t. f, and is monotone on either side of the minimum. This minimizing value of f, call it fCC), appears to increase with increasing C, while the corresponding value of Q decreases. If this can be taken as representing the actual case, then the "bast" parameter point, (C,r), is of the form (C,f(C», where C satisfies the equation Q (c,r(o» = a , 119 and a is the desired error probability. If we make allowance for the inadequacy of the computations, the following table indicates the situation, approximately. TABLE IV Approximate C r{c) Q .01 .05 .48 1. .2, .32 3. .5 .5 .17 5. 10. .11 .05 1. For a fixed value of d, selection of r in this way would insure that 'toTe took the rest value of the first sample size, m. Practical use of these tables of course requires further • extension and greater accuracy in table IV, which in turn requires extension of table III to a ppropriate values of C and r. HO'l1ever, lore may illustrate the above by an example. Suppose we desire to test Q against Q +,1 with a probability of error O O approximately equal to .05. Table IV indicates that C .... 10 and r = 1 (the approximate value of r for 't-Thich m + Ev( t) is minimum at C = 10) have associated with them the desired error probability. 120 Since d = .1, 'tie would take a first sample of size, m .a 4(1)2 • .01 400 Consulting table III, we see that the teat total of approximately 789 observations. ~dll require an average In terms of this expected number of observations, the two-stage test would lie about one-half of the way between the analogous one-stage and sequential probability ratio tests, which would require approximately 1051, and on the average 518 observations, respectively. A comp~isQP of computed and asymptotic formula values of x(y,C) indicates a fairly close correspondence between the two for the larger values of C. of x (table II) at y = 0, e.g., at C = 100" the computed values .015, .0196 are, respectively, .0360, .0341, .0329, while the corresponding values obtained by using (4.26) are .0352, .0333, .0321. Comparison was also made at C = ,0 and C = 10, where the correspondence was found not quite e.g., at C • 10" y • 0, table II gives .02360, (4.26) gives .01693. Computations of d2Ev and Q were made only up to so good. C ... 10. At this value, computed and asymptotic formula values of d2Ev are reasonably close for r ~ .5. e.g., for r = .4, .5, 2 table III gives d Ev = 9.3, 9.0, respectively, while formula (5.36) gives 8.8, 8.4. For larger values of r, larger values of C are needed for reasonable comparison. The leading terms of the 121 asymptotic formula for Q (6.46) are independent of r and give for C =- 10; Q z:; .036. 'Ibis is not too bad an approximation < lS (see table III), but for larger values of r, C is again too small for a reasonable comparison. for r = 10 The results obtained here are, with an exception indicated below, not directly oomparable to other two-stage prooedures discussed in the literature. Owen CIJ has proposed a double sample test, whioh at least for the example cited above, seems almost as good, in terms of expected number of observations, as the Bayes test. For« = .05, a first sample one-half the size of the analogous one-stage sample, is taken. Depending on its outcome, following a certain intuitively proposed rule, one or the other alternatives may be acoepted, or an additional sample, equal in size to the analogous one-stage sample, taken, and then the decision betl'1een alternatives made. Computations in the paper, indicate that in the symmetric case, the ratio of expected total sample size to the analogous one-stage sample size will be .757 for this procedure. This is but slightly higher than the value, .751, 'l'1hich may be obtained for this ratio from the above example. As indicated by (10.41), the Bayes solution in the symmetric case provides a lower bound for the average number 122 of observations required b1 double sml~ling schemes with fixed first sampling size and in this sense it is best for deciding between two simple alternatives. The laborious computations necessary to exhibit these tests, their average sample sizes, and error probabilities, with sufficient accuracy for practical application, is a serious drawback to their use. • 123 BIBLIOBRAPHY 1. 2. 0\-1en, D., "A Double 8am:ple '!!est FTocedure, II Annals of Hathematical Statistics, 24 (1953), 449-457. • Rosser, J.B., Theory and ~lication ?! je 1 2 x -~ o dx, N~w York, New York" t1ap1et.on House, (1948), 41-43. 3. vlald, A., ~equential Analysis, New Yorle, John Hiley and Sons, 1947. 4. Hald, A. and lIolfowitz, J., 1I0ptir,mln Character of the Sequential Probability Ratio TGst, II Annals of Mathematical Statistics, 19 (1948), 326-339. 5. TJa1d, A., Statistical Decision Functions, New York, John and Sons, 1950. ~liley 6. Tlhittaker, E. and Robinson, G., The Calculus of Observations, London, B1ackie and Son-1~~ited, (~4), 143-145. 7. Tab1Gs of Circular and dxr,rbo1ic Sines and CoSines, Federal Horks Agenoy, ~j9. 8. Tables of the 1b:ponential Funotion, eX, National Bureau ot Standards ApplIed ~·;ratfiematics "Series 14, 1951. 9. Tables of 1\1atura1 Lo:t;arithms, 3, 4, Federal viorks Agenoy, 1939 • 10. Tables of i~ormal Probability Function, National Bureau of i;)ta11. . dards Applied ~athematics Series 23, 1953. .