On Stochastic Limit and Order Relationships

Author(s): H. B. Mann and A. Wald

Source: The Annals of Mathematical Statistics , Sep., 1943, Vol. 14, No. 3 (Sep., 1943), pp.
217-226
Published by: Institute of Mathematical Statistics

Stable URL: https://www.jstor.org/stable/2235800

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/2235800?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and

extend access to The Annals of Mathematical Statistics

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 ON STOCHASTIC LIMIT AND ORDER RELATIONSHIPS

BY H. B. MANN AND A. WALD'

Columbia University

1. Introduction. The concept of a stochastic limit is frequently used in

statistical literature. Writers of papers on problems in statistics and probability
usually prove only those special cases of more general theorems which are neces-
sary for the solution of their particular problems. Thus readers of statistical
papers are confronted with the necessity of laboriously ploughing through de-
tails, a task which is made more difficult by the fact that no uniform notation
has as yet been introduced. It is therefore the purpose of the present paper to
outline a systematic theory of stochastic limit and order relationships and at the
same time to propose a convenient notation analogous to the notation of ordinary
limit and order relationships. The theorems derived in this paper are of a more
general nature and seem to contain to the authors' knowledge all previous results
in the literature. For instance the so-called B-method for the derivation of
asymptotic standard deviations and limit distributions, also two lemmas by
J. L. Doob [1] on products, sums and quotients of random variables and a
theorem derived by W. G. Madow [2] are special cases of our results. It is hoped
that such a general theory together with a convenient notation will considerably
facilitate the derivation of theorems concerning stochastic limits and limit dis-
tributions. In section 2 we define the notion of convergence in probability and
that of stochastic order and derive 5 theorems of a very general nature. Sec-
tion 2 contains 2 corollaries of these general theorems which have so far been
most important in applications.
We shall frequently need the concept of a vector. A vector a = (al, * * *, at)
is an ordered set of r numbers a', a . The numbers a', aa are called the
components of a. If the components are random variables then the vector is
called a random vector.
We shall generally denote by a, b constant vectors by x, y random vectors and
by a', , at, x, *. *x their components. Differing from the usual practice we
shallputIaI = (Ja'1, .--,Ia'I)andweshallwritea < bora < bifa' < bi
or a' < bt for every i. This notation saves a great amount of writing, since all
our theorems except theorem 4 are valid for sequences of any number of jointly
distributed variates.
We shall review here the ordinary order notation. In all that follows let f(N)
be a positive function defined for all positive integers N.

l Research under a grant-in-aid from the Carnegie Corporation of New York.

217

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 218 H. B. MANN AND A. WALD

We write
aN = o[f(N)] if lim aN/f(N) = 0.
N-oo

aN = O[f(N)] if I aN ?< Mf(N) for all N and a fixed M > 0.

aN = 0f(N)] if 0 < M'f(N) < I aN <Mf(N) for almost all N
two fixed numbers M > M' > O.
aN = W[f(N)I if 0 < Mf(N) ? | aN I for almost all N and a fixed M > 0.
For instance, log N = o(N) for every e > 0, or sin N/N = 0(1/N), 3 + 4
N/(4 + 8 VN) = 2(V/N) 5/sin N = w(l).
For any statement V we shall denote by P(V) the probability that V holds.

2. General theorems on stochastic limit and order relationships.

DEFINITION 1. We write plim XN = 0. (In words XN converges in probability to
N-=

0 with increasing N) if for every e > 0 lim P( I XN I < ) = 1. Further plim

N-0o

XN = X if plim (XN - X) = 0.
N-00

DEFINITION 2. We write XN = o,[f(N)] (XN is of probability order

plim XN/f(N) = 0.
V-Co

DEFINITION 3. We write XN = Op[f(N)] (XN is of probability order O[f(N)]) if

for each e > 0 there exists an A, > 0 such that P( I XN I < Aff(N)) > 1 - e for
all values of N.
DEFINITION 4. xN = 2,[f(N)] if for each e > 0 there exist two numbers A,
> 0 and B. > 0 and an integer NE such that P[AEf(N) < I XN I < BEf(N)] >
1 - efor all N > NE.
DEFINITION 5. XN = co,[(N)] if for every e > 0 there exists an AE > 0 and
integer NE such that P[A,f(N) < I XN I] > 1 - efor all N > NE.
Let E denote a vector space. For any subset E' of E the symbol a C E' will
mean that a is an element of E'.
Since P(x C E1 & x C E2) ? P(x C E1) - P(x Ct: E2) we evidently have
LEMMA 1. If P(x CE1) > 1-e P(x C E2) >1 - e' then P(x C E1;
x C E2) > 1 - e- -'.
We now put o1 = 0o,2 = 0, 03 = Q, 04 = co.
THEOREM 1. For every 6 > 0 let { RN(6)} be a sequence of subs
mensional Cartesian space such that P (XN C RN(e)) > 1 for, all
a certain integer N. . Let { 9N(X) } be a sequence of functions o
such that gN(aN) = 0t[f(N)] for any 6 > 0 and for any sequence
aN C RN(e). Then we have gN(XN) = OMf(N)].
PROOF: For i = 1, 2, 3, there exists a positive integer NR such that I gN(a) I is
a bounded function of a in RN(6) for N > NE. For otherwise we could construct
a sequence {aN} with aN in RN(6) such thati gN(aN) I > Mf(N) for any M and
for infinitely many values of N which contradicts the hypothesis of our theorem.
Hence there exists an NE such that for N > NI the function I gN(a) I is bounded
in RN(E). Let MN(e) be the l.u.b. of I gN(a) I /f(N) in RN(6). We can construct
a sequence { aN} with aN C RN(e) such that I gN(aN) I /f(N) > MN(e)/2 for all
N > NE. Hence for i = 2, 3 the sequence MN(6) must be bounded and for

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 STOCHASTIC LIMIT 219

i = 1 we must have lim MN(e) = 0. Let M (e) be the l.u.b. of MN(e). For
N-oo

i = 3, 4 one shows in exactly the same manner the existence of a g.l.b

I gN(a) I/f(N) if a C RN(e) and for N > N' . Hence for sufficientl
we have

P[ | gN(XN) | < MN(e)f(N)] > 1 - e with lim MN(e) = 0 for i = 1,

N-'a

P[ J 9N(XN) I < M(e)f(N)] > 1 - e for i = 2,

P[M(e)f (N) < I gN (XN) | < M(e)f(N)] 2 1 - e for i = 3,
P[M(e)f(N) < | 9N(XN) >] 2 1 - e for i = 4.
For i = 2 the existence of an M'(e) such that P[ I 9N(XN) I <
for all N follows easily from this result. Hence our theorem i
COROLLARY 1. If xN = Opi [fj(N)] for j = 1, 2, , r and {RN(e)) is a se-
quence of subsets of the k-dimensional space y1, y, 2.. yk such that P[YN C RN(e)] ?
1 - e for sufficently large N, and if { 9N(X1, X2, ... St y,l y2 ... yh ) } is a sequence
offunctions of x1, x2 ... yl, y2 ...yk such that for any e> 0 we havegN(aN, bN)
= Oi[f(N)] for every sequence I{aN, bN}j with a' = O'i [f(N)] (j = 1, 2, * , r) and
bN C RN(e), then 9N(XN, yN) = OM[f(N)].
PROOF: It follows from Lemma 1, the definition of the relation xi = Oi [fVj(N)]
and the hypothesis of our corollary that for any e > 0 there exists a sequence
of subsets {RN(e) } of the space x1, * * *, xi, y1, * * *, yh which satisfies the condi-
tions of Theorem 1 with respect to the sequence of functions {g9N) Hence
Corollary 1 is an immediate consequence of Theorem 1.
Corollary 1 implies inter alia that all operational rules for the ordinary order
and limit relations are also applicable to stochastic limit and order relations.
For instance o[f(N)]/l2 [g(N)] = o[f(N)/g(N)]. Hence also op[f(N)]/Q2p[g(N)] =
opVf(N)lg(N)].
DEFINITION 6. For any N let RN be a region, fN(a) a function defined on RN.
The sequence {fN(a)I will be said to be uniformly continuous with resrpect to { RNI if
the following condition is fulfilled. For every e > 0 there exists a vector a > 0
such that for almost all N

I fN(a + S)-fN(a) ? <e for any < 6, and for any a C RN

THEOREM 2. Let plim (XN - yN) = 0. For every e > 0 let { RN(e) } be a se-
N-oo

quence of subsets of the r-dimensional vector space such that for almost all N we have
P[YN C RN(e)] > 1 - e. If the sequence of functions {fN(a) } is uniformly con-
tinuous with respect to { RN(e) } for every e > 0, then plim [fN(XN) - fN(yN)I = 0.
N-oo

PROOF: We have fN(XN) - fN(YN) = fN(YN + ZN) - fN(YN) where z

for j = 1, ** *, r. Because of the uniform continuity of fN(a) with
RN(e) we see that for every sequence {aN , bN} with aN C RN(e) and
(j =~~~~~~
(j = 1, 2,*f + f), r).

f N(aN + bN) - f N(aN) =

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 220 H. B. MANN AND A. WALD

Hence Theorem 2 follows from Corollary 1.

In the following we shall abbreviate "cumulative distribution function" by d.f.
DEFINITION 7. Let fXN} be a sequence of random variables. Let FN be the d.f.
Of XN. Let x have the distribution F. We shall write d oo (XN) = d(x) if lim FN
N-oo

= F in every continuity point of F.

THEOREM 3. Let plim (XN - YN) = 0 and d oo (YN) = d(y); then d oc (XN) =
N-oo

d(y).
PROOF: Let GN, FN be the d.f.'s of XN, YN resp. For any 5 > 0 we have

P(yN < a + 5) 2 P(XN < a; YN < a + 5) 2 P(XN < a; I YN - XN I < 5)

> P(XN < a) - NI - XNI > 8),

P(XN < a) 2 P(XN < a; yN < a - 5 P(YN < a -)

P( XN - NI> 5).
Hence since P(YN < a) = FN(a), P(XN < a) = GN(a), lim P (XN -YN I > 5) =
N-Xo
O we have lim. sup. FN(a + () > lim. sup. GN(a) > lim. inf. GN(a) t lim. inf.
FN(a - 5).
If a + 5 and a - 5 are continuity points of F we have

F(a + () > lim. sup. GN(a) > lim. inf. GN(a) 2 F(a -) .

For any So > 0 there exists a positive a < So such that a - a and a + a are
continuity points of F. Hence we can choose a arbitrarily small and if a is a
continuity point of F we must have

lim. GN(a) = F(a).

THEOREM 4. Let XN, YN be two sequences of one-dimensional vectors and let

plim (XN - YN) = 0. Let FN, GN be the cumulative distribution functions of
N-oo

XN and YN respectively. Let RN(e) be the set of points a for which [ FN(a) -
> e. Let MN(e) be the Lebesgue measure of this set. Then lim MN(e) = 0
N-oo

for every e > 0.

We first prove the following lemma.
LEMMA 2. Let (, e be any arbitrary positive numbers and let f be a distribution
function. The set of points a for which f(a + () - f(a) > e has at most the Lebesgue
measure 5/e.
PROOF: The points a for which f(a + 5) - f(a) > e must have a lower bound a.
Otherwise we could find infinitely many such points whose distance from each
other is more than S. But this contradicts the requirement that f(oo) = 1.
Let a, be the g.l.b. of the a's. Then for any v7 > 0 in the interval (a, < x <
a, + a + q) the value of F increases at least by the amount e. Let now a2 be
the g.l.b. of the a's outside of this interval. We continue our construction by
constructing the interval (a2 < x < a2 + a + n) and so forth. But after at most

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 STOCHASTIC LIMIT 221

1/e such steps the construction mu

- f(a) > e are contained in at most 1/e intervals of length 8 + . Hence since
v was arbitrary the Lebesgue measure of this set is at most 8/e.
VVe come now to the proof of our theorem. We have

P(XN < a) > P(XN < a; YN < a + () > P(XN < a) - P XN YN I > 8),

P(YN ? a + 8) 2 P(XN < a: YN < a + () > P(YN < a + 8)

-P( - YNJ > 6) -P(a ? XN < a +26).
Therefore

P(XN < a; YN < a + 8) = P(XN < a) ONP( XN - YI > I )

= P(YN < a + 6) - NP( XN - YNI > 6) - &NP(a < XN< a + 26),
where O < AN < 1, O < Of < 1. Hence
P(YN ? a + 8) = P(XN < a) + ONP( I XN- YN I > 6)

+ NO[F(a + 26) - FN(a)1

where I ON 1, I ON I < 1.
By hypothesis we have P( I XN - YN I > 1/rn) < 1/m for almost all N and
every integer m. Hence we can choose a sequence I{N} with SN > 0 in such a
way that lim AN = 0, lim P( I XN - YN I > AN) = 0. We can then choose Ne
N--oo N-o

so that P( I XN - YN I ? 8N)
that except for a set of measure at most 6 AN/( we have FN(a + 26N) - FN(a) <
e/3. Similarly the set of points for which gN(a + AN) - gN(a) > E/3 has at most
the Lebesgue measure 3 SN/E. Hence, except in a set of points whose measure
is at most 9 AN/E, we have

I GN(a) - FN(a) I < E,

and this completes the proof of Theorem 4.
THEOREM 4a. Let plim (XN - YN) = 0. Let FN, GN be the distribution func-
N-Xoo

tions of XN, yNrespectively. Furthermore, let RN(E) be the set of points inside an
r-dimensional cube where I FN- GN I > E and let MN(E) be the Lebesgue measure
of RN(e), then lim MN(f) = 0.
N--oo

We prove first
LEMMA 2a. Let 8 = ( 82 62 , * ) > 0 and max. 8' = d. Let I be the cube
defined by (-A < xi < A, i- 1, 2, * * r). Let furthermore f be a d.f. Then
the Lebesgue measure of the points a in I for which f(a + 8) - f(a) > e is at most
dr2 Ar-/E.
PROOF: Let fi(xl), f2(x2), * fr(x7) be the marginal distributions of x1, x2, ...
xr respectively. It follows from Lemma 2 that the linear Lebesgue measure of
those numbers at for which fi(at + ) - fi(at) > e/r is smaller than rd/e. We
form the set (xi = a' & x C I) for every such at and for i = 1, 2, ... r. The

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 222 H. B. MANN AND A. WALD

Lebesgue measure of the sum R(E) of all these sets is at most r2dA?1/E. We
shall show that R(e) contains all points a inside I for which f(a +'5) - f(a) > e.
We have

f(aw + A1, a2 +2,... a + 82, aa + at) (a",= + a2 + *+ Ar

where Ai f(a 1 a21 . . . ai-I ai + Si, .. *d a + 6") - fWl1 * * * a', ai+1 + a , *
a" + Sr). If f(a + 8) - f(a) > e then we must have for at least one i

Ai > e/r.

But Ai is the probability of a subset of the set T = (a' < x' < a' + 5') and
fi(a' + 8') - fi(a') is the probability of T itself. Hence

e/r < Ai < fi(a' + 5) -fi(a'),

and if (a1, a2, a?) is in I then it is contained in R(e). Hence Lemma 2a is
proved.
The proof of Theorem 4a using Lemma 2a is similar to that of Theorem 4 and
therefore it is omitted.
The Jordan measure of a set R with respect to the distribution function F is
defined as follows. We consider only intervals whose boundary points are
continxuity points of F. We cover R with the sum I of a finite number of inter-
vals. (The intervals themselves may also be infinite. For instance the sets
a < x < oo, a < x < X are also considered intervals.) We consider M(I) =

J dF for every I covering R. The g.l.b. of all such M(I) is called the exterior

Jordan measure M(R) of R. Similarly we consider all sums I of a finite number

of intervals which are contained in R. The l.u.b. of f dF is called the interior

Jordan measure M(R) of R. If M(R) = M(R) then M(R) is called the Jordan
measure of R.
LEMMA 3. Let FN(x) be a sequence of d.f.'s such that lim FN(x) = F(x) in every
N-0o
continuity point of F(x). Let h(x) be a bounded function such that the discontinuity

points of h(x) have the Jordan measure 0 with respect to F and such that f h(x) dFN
,* +00 ,* +Go. +X

(x) and h(x)

0 dF(x) exist.
fN-ao Then lim f h(x)dFN(x)a
= h(x) dF(x).
PROOF: There is only an enumerable set of hyperplanes parallel to the plane
xi = 0 which have positive probability with respect to F. Hence we can find
for every 8 an interval net whose cells have a diameter at most 8 and such that
the boundary points of every cell are continuity points of F.

We first determine a closed finite interval I such that L dF(x) > 1 - and

such that the boundary points of I are continuity points of F. We further

determine a sum I' of a finite number of open intervals such that I' contains all

discontinuity points of h, f dF(x) < and such that the boundary of I' does

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 STOCHASTIC LIMIT 223

not contain any discontinuity points of F. All this is possible by hypothesis

and because the set of hyperplanes with positive probability is enumerable.
Let R be the subset of I consisting of all points of I which are not contained in I'.
R is a closed set and can be decomposed into a finite number of intervals. The
function h is continuous in R and therefore uniformly continuous. We can
therefore cover R by a finite set of intervals such that the variation of h in every
interval is less than e and such that the boundary points of each interval are
continuity points of F. Let A1 , 2 * * Ik be such a finite set of intervals. Let
xi be any point in Ij. We have

IHN I = L h(x) dFN(x) h(x) dF(x) = Z / [h(x) - h(xi)] dFN(x)

-, f [h(x) h(xi)] dF(x) + h(x Ff dFN(x) dF(x)

O#R ZZR
+ L h(x) dFN(x) h(x) dF(x)

< 6 + 6 + , h(xi) [f dFN(x) dF(x)]

+ max. h(x)f dFN (X) + ]

But lim f dFN(x) > 1-e. Hence

lim. sup. HN < 2e + 26 max. h(x) .

Since e was arbitrary, we must have lim HN = 0.

N-oo

We are now prepared to prove

THEOREM 5. Let doo (XN) = d(x). Let g(x) be a Borel measurable function
such that the set R of discontinuity points of g(x) is closed and P(x CI R) = 0.
Then doo[g(xv)] = d[g(x)].
PROOF: Let FN be the d.f.'of XN, F the d.f. of x, FN,, F, the d.f.'s of g(XN),
g(x) resp. Then lim FN = F in every cont. point of F. Let h(x) be defined
N-oo
as follows:

h(x) = 1 if g(x) < a,

h(x) = 0 if g(x) > a.

The discontinuities of h are contained in the set M of all points where g(x) = a
and is continuous or where g(x) is discontinuous. The set R of discontinuity
points of g(x) is closed and of measure 0 with respect to F. We can therefore
subtract from M a sum R* of a finite number of open intervals of arbitrarily
small measure with respect to F which contains all discontinuity points of g(x).
This difference set M' is closed and contains only points where g(x) = a and

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
224 H. B. MANN AND A. WALD

x 4: R. If a is a continuity point of F, then the B

to F is 0. Since M' is closed, its Jordan measure is also 0. Hence the Jordan
measure of the discontinuity points of h(x) is 0 if a is a continuity point of F,.

Since g(x) is Borel measurable, f h(x) dFN(x) = FN,(a) and f h(x) dF(x)
= F,(a) exist for every a. Hence by Lemma 3 lim FN,(a) = F,(a) in every
N-*OO

continuity point of F, and this proves our theorem.

3. Corollaries and applications. COROLLARY 2. If plim (XN - YN) = 0,

N -oo

d oo (YN) = d(y) and if f is continuous except in a set R for which lim P(YN C R)
N-.
= 0 then plimf(xN) - f(YN) = 0.

PROOF: Let I be a closed interval such that P(YN C I) > 1 - e/2. Let I' be
a sum of open intervals containing all discontinuity points of f(x) in I and such
that P(yAN C I') < e/2 for sufficiently large N. The set J of points of I which
are not points of I' is a closed set. Hence f is uniformly continuous in J and
P(YN C J) > 1 - e for sufficiently large N. In Theorem 2 we put RN(E)-J,
fN = f. Then all conditions of Theorem 2 are satisfied and it follows that plim
N-oo

[ff(N) - f(Yv)] = 0.
If, moreover, the set of discontinuity points of f is closed then by Theorems
3 and 5 doo [f(XN)] = d oo [f(YN)] = d[f(y)].
Special cases of Corollary 2 have been proved by J. L. Doob and WV. G.
Madow (2).
Theorem 5 is very useful in deriving limit distributions.
It follows for instance from Theorem 5 that if d so (XN) = d(x), d oo (YN) -
d(y), where x, y are independently and normally distributed with mean 0 and
equal variances, then d oo (XN/YN) = d(x/y). That is to say the distribution of
XN/yN converges to a Cauchy distribution.
It also follows from Theorem 5 that under very general conditions the limit
distribution of t = VN(V - p)/s is normal. (x = sample mean, u = population
mean, 52 = sample variance.) For we have under very general conditions d so
V/N(x - ) =-d(t), plim s = a, where t is normally distributed with vari-
ance a .
Applying Theorem 5 it can also easily be shown that under very general
conditions the limit distribution of T2 is a chi-square distribution if the means
of all variates are 0. Hotelling's 72 (the generalized Student ratio) for a
p-variate distribution is defined as follows:

T2 = NE EAii t? where IfAif = Ifsii fV'- = X

i-i j=1

where siJ is the sample covariance between x?

We have dcc (Ai,) = d(cr"), where I fif a II-1
1, 2, * * * p then d co (VN ~i) = d(ni) where the X

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 STOCHASTIC LIMIT 225

with covariance matrix I o-i. Hence

d c (T2) = d [,Z nim] = d (Z ,)X

where the ' are normally and independently distributed with variance 1.
Hence the distribution of T2 converges to a chi-square distribution with p degrees
of freedom.
If the samples are drawn from a sequence of populations 17N) all with the
same covariance matrix and such that lim VIN/iN = i where /LiN iS the mean
N-

value of the ith variate in the Nth populatio

way that the limit distribution of 72 is a non-central square distribution with
p degrees of freedom.
The limit distribution of T72 has been derived by W. G. Madow (2).
COROLLARY 3. Let XN, YN be r-dimensional vectors d to (YN) = d(y) and XN -
YN = Opff(N)] with lim f(N) = 0. Let g(x) be a function admitting continuous
N-*o

jth derivatives except in a set R uwith lim P(YN C R) = 0. Let

N-*o

T1(x, a) = E a (xi - ai) + ... + [7 (x -a ) \ ] 9a

then

g(XN) - g(yN) - Tj(XN , YN) = OPUf(N)]'}.

Since the jth derivatives are continuous except in a set of limit measure 0
we can determine a closed set R(e) on which they are uniformly continuous and
so that P(YN C R(e)) > 1 - e for sufficiently large N. Then for every sequence
with aN- bN = O(f(Nl)), bN C R(e) we have
g(aN) - g(bN) - T,(aN, bN) = o[f(N)l .
Hence Corollary 3 follows from Theorem 1.
Corollary 3 was first proved by W. G. Madow [2] and J. L. Doob [1] for the
important case that YN is a constant.
The following example will illustrate Corollary 3. Let x, y be normally and
independently distributed random variables with mean 0 and variance 1; ZNV},
{ZN sequences of random variables with plim \/N ZN = plim T/N z = 1.
N-*oo N-+,w

Let XN = X + ZN, YN = Y + Z
Y3/3 + 2x - 2y + 5. Applying C
g(x, y) = [41/V\N], ZN = O(1
Corollary 3 for j = 1 we have

9(XN, YN) - g(X, Y) - (X2 + 2)ZN - (Y2 - 2)z/ = op(l1/x/N) .

Multiplying by -\IN we have
[g(XN, YN) - g(X, Y)] \N - [(x + 2)ZN + (Y - 2)zN] N = op(l) .

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms
 226 H. B. MANN AND A. WALD

This is equivalent to

plim[\NV(9(XN, YN) - g(X, Y))] = x2 + y2X

N-oo

Hence the distribution of /VN(9(XN, YN) - 9(x, y)) converges to the chi-square
distribution with 2 degrees of freedom.
If plim XN = a and I ON is a sequence of numbers with limr N = 0 such that
N-oo N-0oo
d o [(XN- aI)/O
if g admits cont

from 0, then putting (X) = 9,, we have

9(XN) - 9(a) = g(x - a') + * + (x4 - ad) + op(g) .

Hence applying, Theorems 3 and 5 we have

(i) d o [V(XN)- g(a)] _ d(gq,l + * + grr).

a'N

That is to say the distribution of [g(XN) - g(a)]/ON converges to the distribution

p

of E gj4j in all continuity points of the latter. A corresponding result can be

i.1

obtained from Corollary 3 if all first derivatives are 0 at x = a and at least

one second derivative is different from 0 and so forth.
A method of deriving limiting distributions and limit standard deviations based
on (i) is known as the 6-method and has been extensively applied in statistical
literature.

REFERENCES

[1] J. L. DOOB, "The limiting distribution of certain statistics." Annals of Math. Stat., Vol.
6 (1935).
[2] W. G. MADOW, "Limiting /distribution of quadratic and bilinear forms." Annals of
Math. Stat., Vol. 11 (1940).

This content downloaded from

140.113.114.44 on Mon, 08 May 2023 13:04:42 +00:00
All use subject to https://about.jstor.org/terms