Strong Laws of Large Numbers for $r$-Dimensional Arrays of Random Variables

Author(s): R. T. Smythe
Source: The Annals of Probability , Feb., 1973, Vol. 1, No. 1 (Feb., 1973), pp. 164-170
Published by: Institute of Mathematical Statistics

Stable URL: https://www.jstor.org/stable/2959352

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

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 Probability

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
The Annals of Probability
1973, Vol. 1, No. 1, 164-170

STRONG LAWS OF LARGE NUMBERS FOR r-DIMENSIONAL

ARRAYS OF RANDOM VARIABLES

BY R. T. SMYTHE

University of Washington

Let Kr be the set of r-tuples k (ki, k2, * *, k,) with positive integers
for coordinates (r > 1). Let {Xk: k E K,} be a set of i.i.d. random variables
with mean zero, and let < denote the coordinate-wise partial ordering on
Kr. Set Iki = kkk2..- *kr and define, for ke Kr Sk = E jfk Xi. If{Ek: ke K,}
is a set of events indexed by Kr, we say (given a) "Ek f.o." if 3 1(wo) E Kr
such that k $ I implies WGEkc. We say "Ek a.l." if given any IeKr,
3 k > I such that GE Ek. We prove:
(i) If E{IXkI (log+ IXkI)r-l} = oo, then given any A > 0, P{ISkj/ kI >
A a.l.} = 1. Using martingale techniques, we also give a new proof of the
converse result due to Zygmund:
(ii) If E{IXkl (log+ IXkj)r-l} < oo, then given any 6 > 0, P{ISkI/Ikl <
6f.o.}= 1.
For non-identically distributed independent random variables with
mean zero, the usual conditions sufficient for convergence of Sn/n to zero
in the linearly ordered case are also sufficient for matrix arrays.

1. Introduction and notation. We wish to consider the following problem:

Given a probability space (Q, F, P) and an r-dimensional array of independent
random variables with zero mean defined on (Q, F, P), under what conditions
will the sample averages converge to zero?
To make this problem more precise it will be convenient to introduce some
notation. Let Kr be the set of r-tuples k = (k1, k2, ** , kr) with positive
for coordinates (r > 1). Let {Xk: keKr} be a set of independent random
variables with mean zero and let < denote the coordinate-wise partial ordering

on Kr. Set IkI = k.. ... kr and define, for k e Kr:

X~, z - Sk
Sk =E ijs9k Xi, Zk=|k
kIkI
If {Ek: k e Kr} is a set of events indexed by Kr, we say (given w e Q) "Ek occurs
finitely often" (abbreviated "Ek f.o.") if there exists 1(w) e Kr such that k $ I
implies w e Ekc. We say "Ek occurs for arbitrarily large indces" (abbreviated
"Ek a.l.") if, given any 1 e Kr, there exists k > I such that w e Ek. (Note that
when r > 1 this last condition is not equivalent to the condition "Ek occurs
infinitely often", which is weaker.) Our problem can now be stated as follows:
Given s > 0, what conditions on the Xk will guarantee that

P{lZkl >? f.o.}-1 1?

Received February 23, 1972; revised July 20, 1972.
AMS 1970 subject classifications. Primary 60G50; Secondary 60G45, 60F15.
Key words and phrases. Independent, identically distributed random variables; strong law of
large numbers; (reversed) martingale; Borel-Cantelli lemma.

164

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 MULTIDIMENSIONAL SLLN 165

For r = 1, the result reduces in the identically distributed case to Kolmogorov's

strong law of large numbers. For r > 1, the result is of interest in the study of
properties of "Brownian sheets" Xtlt2 ... t with an r-dimensional time paramet
(see [5], [7]); for example, if P{IZkI > e f.o.} 1, one could deduce that
(writing t for (t1, t2, , t.) with probability one,

_ -- 0 as ItI t oo.
ItI

The identically distributed case has been treated in the context of ergodic
theory by Dunford ([4]) and by Zygmund ([9]). Dunford proved that the
integrability of IXkI log+ IXkI was sufficient, when r 2, to insure that
P{IZkj > s a.l.} = 0; implicit in Zygmund's treatment is the result that for
general r, the integrability of Xki (log+ XkI)r1 is a sufficient condition for
convergence.
We show in Section 2 that the sufficient condition given above is also neces-
sary when the random variables are identically distributed. We give also a new
proof of the sufficiency via martingale theory, specifically the inequalities
established by Cairoli ([1]). In Section 3 we discuss the non-identically distri-
buted case.

2. The identically distributed case.

THEOREM. Let {Xk: k C Kr} be i.i.d. random variables with zero mean.

(1) If E{IXkI (log+ Xkl)r-I} < 00, then for any e > 0, P1[ZkI > e f.o.} = 1.
(2) If E{IXkI (log+ jXkI)r-l} = oo then for any A > 0, P{IZkI > A a.l.}
PROOF. We prove first the necessity of the condition, i.e., (2). Assume that
E IXkI (log+ IXkI)r-1 =oo. Then

*- S- PflX1l > Iky} dkj . dkr

- S1T5 (So- 1 [Iki ?]F(d2)) dkAdk2 .dkr

where F(.) is the distribution function of IXI. By simple integration using
Fubini's theorem the above multiple integral is found to be

50 2(log+ 2)r-lF(d2) + (terms of lower order) = oo

It therefore follows that Lk, *k P{jX1j > Ajkj} = oo for any A > 0; he
that 2Zke .. *Lkr P{IXkI > Ajkl} 0o .
By the (converse) Borel-Cantelli lemma one concludes that P{IXkl > A IkI i.
1; by taking r-dimensional differences of the partial sums this implies that
P{IZkI > A/2r i.o.} = 1 for any A > 0.
To strengthen this to the result that P{IZkI > A a.l.} = 1, it suffices to show
that for arbitrarily large k, T12k P{jXII > A 1lj} = 0o. But if ki > 2, 1 < i < r,
the series EI2k P{IXII > A 11} majorizes, term by term, the series El P{IX/1IlkI >
A 1ll} and by the argument above, this latter series diverges. L

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 166 R. T. SMYTHE

We turn now to (1). F

LEMMA 1. (Zk, Vk) is

PROOF. If j < k it is c
Ejsk E(Xj I Oik) = E(S
E(Xj I <k), proving the

LEMMA 2. Fix kr and

represent the r-tuple
on Kr-I
PROOF. Let SAIm = AiIZ XAm
We have

A, V-=1 g(SAjm) V VA?k g(X Am)

Now

E(Zi,kr I ?k) = ZMr=I E A ik

Using the independence of the Xk, when I ? k the

Er F ( 1 r Si A)

which by Lemma 1 is equal to Ikr 1 SAim/Ikkr = 4k

REMARK. The proof of Lemma 2 can easily be extended to show that if any
two coordinates ki, k, of (k1, k2, * * , kr) are fixed, Zk is a reversed mart
indexed by the r-2-tuples (k1, ..ki, ki+1, * * *, k1 kj+1 ... *, kr), with
pect to the fields 6'kl ki-1, 1,ki~ ? kj-l','kj+l'---kr-
Following Cairoli we will now establish the basic inequalities.

LEMMA 3.

(3) 2P{supk iZki > 21 < A7 SUPk E{IZkI (log+ IZkI)r-1} + Br

(4) E{supk AZkI<} < Apr supk E{IZk },
where Ar, Br and Ap rare universal constants and p > 1.
PROOF. We consider first the case r = 2 and proceed by induction.
From Lemma 2 it follows that (SUPk2 IZk1 k2), Sk1,1) is a positive reverse
submartingale, indexed by k1 (assuming the expectations are finite). By Doob's
inequality, we have, for every i > 0:

(5) iP{supks1k2 IZ(kk2)1 > 21 < SUPkj E{supk2 IZ(kl,k2)1}

But it now follows by a second inequality of Doob ([3] page 317), since
(Z(k1 k2), 91, k2) is a reversed martingale indexed by k2 for fixed k1, that

(6) E{supk2 IZ(ki,k2)I} < A SUPk2 E{IZ(k1,k2)1 log IZ(k1,k2)I} + A

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 MULTIDIMENSIONAL SLLN 167

In a similar fashion it is shown (see again [3] page 317) that

(7) E{SUPkj'k2 IZ(klk2)l } < BP SUPkl,k2 E{IZ(kl,k2) l} for

all p > 1, where BP is a universal constant. This establishes the lemma for
r = 2.
For general r let

her = all positive submartingales (Yk, V'k) such that for k, fixed
(Yk, Vkj .' ki-1,,I, kic, . kr) is a positive submartingale.
By (6) and (7) (and their proofs), the inequalities (3) and (4) are valid for
r = 2 for all submartingales in H2.2
Assume now the validity of the inequalities for all submartingale
If k = (kj, ** , k,-1), then

iP{5UPf(k 4 k_ 4} = 1P{SUPIk (SU1 P kl4 >_ i

By the remark following Lemma 2, supkr |Zk I e Hr1.

Hence by the induction assumption:

(8) 1P{5UPI4,k Zj kr I i 4 Ar-1 supk E{supkr IzzkrI (log+ IZkrj) } +

But it is clear that for fixed k, 4kIr I is a positive supermartingale index
since t(log+ t)r-2 is a convex increasing function for r > 2, ZkkrI (log+
is also a positive submartingale.
Employing Doob's inequality noted earlier and the inequality

t(log+ t)-2 log+ ([t log+ tyr-2 < (r - 1 )t[log+ t]r1

which holds for integers r > 2 and all t > 0 we have

(9) E{supk |Zj jE (logl ZIk )r} )

< A SUPkr E{Z4,krl (log+ Z4,krI) log r [jZjkl (log1 0 Zi krl ]} + A
< A(r - 1) SUPkr E{fZAkrl (log+ 14k r} + A
From (8) and (9) we immediately deduce (3) and (4).

LEMMA 4. If the right-hand side of (3) (resp. (4)) is finite, then limk-.. Zk exists
in the L' (resp. LP) sense (with the usual definition of convergence along a directed
set).

PROOF. It is well known that uniform integrability of a (forward or reversed)

nartingale implies L' convergence along a directed set ([4] page 86).
We can now deduce a.s. convergence by a modification of Cairoli's argument.
Let a be a positive constant and set

X(a) = a X < -a

X IXI < a
a X > a.

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 168 R. T. SMYTHE

Then {Xk- XkIcI: k C Kr} is an array of i.i.d. random variables.

Let Sk'a) L js~kXjk7 and define Zk(a) accordingly. Let 1 denote the r-t
(1, 1, . ., 1). By retracing the arguments leading to (3) for the array
{ Xk - Xk(a)} we get

(1 0) P{SUPk Zk - Zk(a) > 41

< (A/i) SUPk E{ Zk - Zk(a) (log+ Zk - Zk (a) ) + A
< (A/i) E{IX1 - Xlcl) (log+ IX1 - X1(a) )r-1j + A
The last inequality is a consequence of Jensen's (conditional) inequality, since

is Sk (a)
E{Xi - Xi (a) |ag(Sk - Sk(c))1 - Sk _ Sk

Now take a sequence {Ckl I 0 and a sequence {ak} T oo such that

( 1 1 ) P{supk Zk Zk(aj) > ej <2-.
This is possible by (10), since

P{supk Zk - Zk(ij)l > eJ P{supk C Zk - Zk(a)i > Ce,}

< (A/csj)E{c IX1 - Xl(ai)l (log+ c IX1 - Xl(atj) lr-1+ (A/ce)
which for fixed es can be made <2-i by choosing c and aj sufficiently
By (11) and the Borel-Cantelli lemma, P{supk AZk - Zk4)l > ej i
Thus if we can show that Zk(a) converges a.s. it will follow that Zk converges
a.s.

By Lemma 4 we know that Zk(,) converges in LP for p > 1. Denote the limit
by XYa); then X(') C LP. Choose a sequence of indices k'i) such that as j] 00,
k* 00, *, ik 00. Fix j; if 6~' (a denotes the field VI2k So', then for
k > k), {Zk(t) - X(a), ?k(7a)} is a martingale. By (4):

E{supk~ki Zk(a) - X(a') } < BP{supk!k(j) E jZk(a) - X(a) p}

< BPEt|Zk (a - X(a) I pi

(since E{Zk, - X | Sk (a) - Zk( ) - X(a)). But E{jZk'j, - X0a|P} j t ? and

supk~k(i) AZk) -X(a)IP is decreasing in j; hence suPkNk Zk(a) - X(a) I-->
as j T o0 so Zk -+ X(a) a.s. Since Zk clearly ->0 in probability, it follow
X(a) = 0 a. s.
We have now shown a.s. convergence in the following sense:

(12) For any s > 0, PF{ZkI > s a.l.} = 0.

To strengthen this to the statement that P{IZkI > s f.o.} = 1 we proceed as
follows. Let s > 0 be given, and take r = 2. On each of the lines x =I and
y = j, the one-dimensional strong law holds a.s.; denoting by Q. the
Q formed by deleting the null sets of non-convergence, we have P(Q0) = 1. For
? C Q0, find k = (k1, k2) satisfying (12); by taking k' > k if necessary we can

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 MULTIDIMENSIONAL SLLN 169

(since there are only a finite number of lines x = l for l < k1 and y =j for
j < k2) conclude that
I ZI(w)l < IS for all I I k'.
Now assume inductively that P{jZk4 > s f.o.} 1 fo
for each of the r - 1 dimensional hyperplanes defi
there is a null set on whose complement jZkI > e f.o.;
section of the sets of convergence over these hyperplanes. For w C Q0, find
k = (k1, k kr) such that (12) holds; by taking k' > k if necessary we again
get

IZ (0)I < S for I ! k',

hence that P{IZkI > s f.o.} = 1, proving (1).
3. The non-identically distributed case. In this section we consider independent
random variables with zero mean, but we no longer assume identical distribu-
tions. We will simply outline some of the results here, without proof.
One of the keys to an analysis of this case is the following analog of
Kolmogorov's inequality for the one-dimensional case ([6]):
THEOREM (Wichura) Let

Mn = maxjsSi, a E(Sn2).
Then

E(Mn2) < 4r<X2

Using this result one can be prove half of the Three Series Theorem for matrix
arrays, i.e. that the convergence of the three series is sufficient to insure a.s.
convergence of the random series Ek Xk. (The other half is valid also, but we
do not have an elementary proof-what is needed is an analog of Kolmogorov's
second inequality.) Proceeding as in the one-dimensional case one can then
prove, for example, the following analogue of a well-known theorem (e.g. [2]
page 17):

THEOREM. Let p be a positive, even continuous function on RI such that

increases, ~p(x)/x increases, and pD(X)/X2 decreases. Then if

Ek E((p(Xk)) < 00
p jkj

it follows that PF{lZk > s f.o.} = 1 for any s > 0.

The usual sufficient conditions for convergence in the one-dimensional case are
thus sufficient for n-dimensional arrays as well, e.g., if supk E IXkIP < oo for
for some p > 1, we have convergence.
Finally we remark that the approach via Wichura's inequality and the Three
Series Theorem can be used to give a different and more classical proof of the
strong law in Section 2. Similar techniques can be employed to yield strong
laws for more general partially ordered arrays.

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms
 170 R. T. SMYTHE

Acknowledgment. The author wishes to thank Professor Ron Pyke for

introducing him to this area of research and for many helpful conversations
concerning this work, and Professor Ted Harris for the references [4] and [9].

REFERENCES

[1] CAIROLI, R. (1970). Une Inegalit6 pour Martingales a Indices Multiples et ses Applications,
Seminaire de Probabilites IV, Springer-Verlag Lecture Notes in Mathematics 124 1-28.
[2] CHUNG, K. L. (1968). A Course in Probability Theory. Harcourt, Brace and World, New
York.
[3] DOOB, J. L. (1953). Stochastic Processes. Wiley, New York.
[4] DUNFORD, N. (1951). An individual ergodic theorem for noncommutative transformations.
Acta Sci. Math. Szeged. 14 1-4.
[5] MEYER, P. A. (1966). Probability and Potentials. Ginn-Blaisdell, New York.
[6] PARK, W. J. (1970). A multi-parameter Gaussian process, Ann. Math. Statist. 41 1582-1595.
[7] WICHURA, M. J. (1969). Inequalities with applications to the weak convergence of random
processes with multidimensional time parameters, Ann. Math. Statist. 40 681-687.
[8] YEH, J. (1960). Wiener measure in a space of functions of two variables. Trans. Amer.
Math. Soc., 95 433-450.
[9] ZYGMUND, A. (1951). An individual ergodic theorem for noncommutative transformations,
Acta Sci. Math. Szeged. 14 103-1 10.
DEPARTMENT OF MATHEMATICS
UNIVERSITY OF WASHINGTON
SEATTLE, WASHINGTON 98105

This content downloaded from

196.200.146.5 on Tue, 25 Apr 2023 11:28:08 UTC
All use subject to https://about.jstor.org/terms