On a Theorem of Hsu and Robbins

Author(s): P. Erdos
Source: The Annals of Mathematical Statistics, Vol. 20, No. 2 (Jun., 1949), pp. 286-291
Published by: Institute of Mathematical Statistics
Stable URL: http://www.jstor.org/stable/2236860
Accessed: 29-05-2018 13:28 UTC

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
http://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 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 ON A THEOREM OF HSU AND ROBBINS

BY P. ERDOS

Syracuse University

Let fi(x), f2(x), be an infinite sequence of measurable functions defined

on a measure space X with measure m, m(X) = 1, all having the same distribu-
tion function G(t) = m(x; fk(x) < t). In a recent paper Hsu and Robbins'
prove the following theorem: Assume that

(1) L:tdG(t) = 0,
oo

(2) Let dG(t) < oo.

00

n Ico
Denote by Sn the se
k-I ~~~~~~~~~~~n-il
converges.
n n

It is
k=i k-1
(replace fk(x) by
It was conjectured that the conditions (1) and (2) are necessary for the
00

convergence of E Mn. Dr. Chung pointed it out to me that in this form the
n-1

conjecture is inaccurate; to see this it suffices to put fk(x) = T(1 + rk(x)) where
rk(x) is the kth Rademacher function. Clearly I fk(x) I < 1; thus Mn = 0,
oo0r0??
thus E Mn converges, but t dG(t) 0 0. On the other hand we shall show
nul o
in the present not
In fact we prove
THEOREM I. The necessary and sufficient condition for the convergence of
00

E Mn is that
nl1

(1)~~~~~~ L dGW )< l,

and (2) should hold.
In proving the sufficiency of Theorem I, we can assume without loss of gener-

ality that (1) holds. It suffices to replacefk(x) by (fk(x) - C) where C = L:tdG(t).

co

The following proof of the sufficiency of Theorem I (in other words essentially for
the theorem of Hsu and Robbins) is simpler and quite different from theirs.
Put

(3) ai = m(x; Ifk(x) I > 22),

1 Proc. Nat. Acad. Sciences, 1947, pp. 25-31.

286

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 ON A THEOREM OF HSU AND ROBBINS 287

since the fJ's all have the same distribution, ai c

We evidently have
40 00~~C 00 co co

- -2a
i-o
< , +i(ai-a+)
i-o
< ? dG(t) <i-O
Z 2t+2(ai - ai+)
io
< E i+2 ai.
Thus (2) is equivalent to
00

(4) 22tai i-o < o

Let 2' < n < 2'1. Put
n = (x; Ifk(x) I > 2"2, for at least one k < n),
S(2) = (x; I ,41(x) I > n fkjx) I > n415, for at least two ki < n, k2 < n
= (x; |f (x) > 2'-)
k-1

where the dash indicates that the k with If*(X) I > n4/6 are omitted. We
evidently have

Sn C s($) U S2 U s$3
For if x is not in Sn' U S(2) U S(3), then clearly

| fk(x) < 2'2 + 2i2 < n.

kP-i

00

Thus to prove the convergence of E M. it will suffice to show that

n-1

(5) 2E (m(S(")
u-i
+ m(S(2)) + m(S(8))) < o.
From (3) we obtain that m(S <) <n a* a2 < 2+i- a2. Thus from (4)
00 ~~~00 00

(6) X m(S(") = E )2 m(S(P)) < 22 ' ai < oo.

n-1 i-O 2i?n<2i+i iO
From (4) we evidently have that for large u

m(x; lfk(x) I > u) < 1/42.

Thus since the f's are independent and have the same distribution function it
follows that for sufficiently large n,

m(S(2)) ?<~ii m(X; Ifkl(X) I > /61 Ijfk2(X) I > ?i/5)

I Skl <k2s< n

< (-2) m(X; flf(X)I > n4/1), m(X; I f2(x) I > fl15) <f2 -l616 -

Hence

M00 2)< 0
(7) Z m(Sn2)
n-i
<00.

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 288 P. ERDOS

Put
ffk(x) for I fk(x) I

O otherwise.

Clearly the ft(x) are independent and have the same distribution function
G+(t). Put
00

(8) J t dG+(t) = e, k f (X) - E.

00

We have from (8) that f gk(x) dm = 0, and by (1) that e 0 as n -* o. We

evidently have

f (Z)gk(X) dm = dm + 6 lk ' q*(x) g'(x) dm.

k=1 gkk ?kk< 1: in

Now since max I9k(x) I < n4/

fug*(x) dm < (n415 + E) f9g(x) dm < c-n9

and

2 g(X) _g2(X) dnm - 2f(x) dm f a(x) dm < cr .

Thus
n \4
g Ek(z)) dm <C30l1.
;gk=1

Hence

(9) m (x; Y, gk(x) > n/16) < c4 n- (715)

Thus fro-m (8), (9), 1 f+(x) < I gk(x) I + 1/16 (for e < 1/16) and n/8
we have

Xn f (x) > 2) = m (x; fE f(x) > )

< m(x; | Egk(x) j> n/1-6 < C4 n

or

(10) nr(()<c
Thus finally from (6), (
of the sufficiency of The

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 ON A THEOREM OF HSU AND ROBBINS 289

Next we prove the necessity of Theorem I, in other words we shall show that if
00

? Mn converges then (1') and (2) hold.

n-1

First we prove (2). The following proof was suggested by Dr. Chung, who
simplified my original proof. By a simple rearrangement we see that (2) is
equivalent to
co

(11) ~~~E n WUG() < Xo

n-1 Itl >cn

for any c > 0; while

(12) f t I dG (t) < oo oo

is equivalent to

(13) E dG(t) < oo

nl I tl>cn
for any c > 0. Now we have clearly,

(x; Ifn(x) I > 2n) C Sn- U Sn.

Hence

n f tl>2ndG(t)
nto It>2n no < E (m(Sn-1) + m (Sn)) < 0o.
Thus we obtain (12). Since the terms of this series is non-increasing it follows
that

(14) n f dG(t) -+ 0.
I tl>2n

Our assumption being that ^ Mn < oo we have Mn -O 0 as n --+ o. It follows

that there is a constant p > 0 independent of k and n such that

m(x; fz(x) < n) > p.

Now, writing set intersections as products, we have

U (x; Ifk-(x) I > 2n)*( x; | fi(x)] < n) C Sn-

k=k

Writing this for a moment as

U (RkTk) C Sn
k=1

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 290 P. rERDOS

where Rk = (x; I fk(x) I > 2n

we have

Mn- mU(u ( T)U (Rk Tk)

M (Sn) >
kl /

n
=M mU (R1 Tj)' (Rk.1 Tk..)' Rk Tk)
kl-

- 2 m((R1T) T m( (R1U U Tk-1)URk Tk)

k-i

> ,I m(RlT*) Rk(- Rk TO)

n

> , l m(Rk-Tk) - m((RI U ***U Rk l)Rk)}

k-I

> {-m(Tk)
k-I
- (k - l)m(R)))m(Rp) m
n m

> F, {p -nm(]?,) IM(Rk)2 (p- 0(1))M(Rk)-

k-l k-1

> tn
_ p' Y2m(Rk) = np' f dG(t)
k-i I tj >2n

by (14) since m(RD) = f dG(t), nm(R1) -O 0 as n - o.

Itj >2n

Thus

Sn f dG(t) ? , g Mn < 0
n I t>2n Pt n

Hence we have (11), which is equivalent to (2). Th

By virtue of (2) we can put
co

oo
tG(t) = C.

If C > 1, then it follows from (2) and Tschebycheff inequality that M. -n 1 as

n -+ oo, thus C ? 1. But if C = 1, we conclude from (2) and the central limit
theorem that Mn does not tend to 0. Hence C < 1, and (1') is proved.
By similar methods we can prove the following results: Let 2 < c < 4. Put

M(c) = m (x; | fk(x) > n 21)

Then the necessary and sufficient condition for the convergence of alX2
k-I

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms
 ON A THEOREM OF HSU AND ROBBINS 291

is that
co , 00

t tdG(t)- = 01 t lCdG (t) < oo

Z ~~~~~~~~~oo

If c < 2 then the necessary and sufficient condition for the convergence of
00 p*0

n-1
E
coo
is that f

Finally we can prove the following result: Assume that L t dG(t) = 0 and

I t4 dG(t) < oo. Then there exists a constant r so that

(17) m m[;| E f (X)| > njt 2 (log nyt < 0.

The case of the Rademacher functions shows that (17) can not be improved
very much, in fact only the value of r could be improved.

This content downloaded from 140.113.213.2 on Tue, 29 May 2018 13:28:16 UTC
All use subject to http://about.jstor.org/terms