Professional Documents
Culture Documents
Tasche JournalOfAppliedProbability
Tasche JournalOfAppliedProbability
net/publication/237799633
CITATIONS READS
12 351
1 author:
Dirk Tasche
Swiss Financial Market Supervisory Authority FINMA
66 PUBLICATIONS 3,480 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Dirk Tasche on 04 November 2017.
Printed in Israel
Abstract
rate. We show that the conclusion of the second Borel-Cantelli lemma holds if the series
of the probabilities of the events diverges at a certain rate depending on the mixing rate
of the events. An application to necessary moment conditions for the strong law of large
numbers is given.
SECONDARY 60F15
1. Introduction
holds. This is the assertion of the second Borel-Cantelli lemma. If the assumption of
(Durrett (1991), Theorem (6.6) ch. 1) or by uniform mixing (Iosifescu and Theodorescu
(1969), Lemma 1.1.2'). Rieders (1993) showed by an example that strong mixing without
any further assumption does not guarantee the validity of (1). Hence, the fact that
Yoshihara's (1979, Theorem 1) Borel-Cantelli lemma for strongly mixing events requires
Intuitively, one might expect that the less dependent the events are the better relation
(1) will fit. Our results confirm this intuition in the following manner.
For a fixed rate of convergence of the strong mixing coefficients of the events A,, we
determine how fast the rate of divergence of the series C P(A,) has to be in order to
make (1) valid. In the case of decreasing probabilities (Theorem 2.2) this rate is smaller
than in the general case (Theorem 4.1 and Proposition 4.3). By Theorem 2.2 we deduce
381
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
382 DIRK TASCHE
necessary moment conditions (Corollary 2.3) for the strong law of large numbers in the
case of a strongly mixing stationary sequence. These conditions are useful in Section 3
to show by an example (Proposition 3.1) that the conclusion of Theorem 2.2 is essentially
Throughout this paper we will denote by [x] the integer part sup{z z E : z ? x} of the
real number x, by I(A) the indicator function of the event A and by log+ the function
2. Decreasing probabilities
First let us specify the mixing coefficients we are dealing with in this and the subsequent
sections.
Definition 2.1. Let (Zk),~N be any sequence of random variables on a common prob-
ability space.
are called strong mixing coefficients of (Zk). The sequence (Zk) is called strongly mixing
if lim,n + z (n) = 0.
are called a-mixing coefficients of (Zk). The sequence (Zk) is called !-mixing if
limn - az (n)= 0.
(az(n)) and (Iz(n)) are obviously non-increasing sequences with az(n) ? az(n) for all
n. Thus a-mixing is weaker than strong mixing. We will omit the index Z if there is no
danger of confusion.
For a sequence (Ak)k ~N of events the corresponding mixing coefficients and mixing
properties are defined by considering the random variables Zk := I(Ak). With these
Theorem 2.2. Let (Ak)k E N be a sequence of events such that k 4 P(Ak) is non-increas-
ing. Then P(n,"= _ LU"M, Am,)= 1 is implied by either of the following two conditions.
n=l n=l
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 383
SP(A,)
n= log1(n)
We will prove Theorem 2.2 using some lemmas. But first let us state the announced
corollary whose proof will be deferred to the end of the section. Compare the corollary
to the strong law of large numbers for i.i.d. real random variables (X,), , I which can be
formulated as
in
n-x? n j=1
variables such that limn.n- I.n'=, X exists and is finite a.s. Then we get the following
implications.
(ii) IX, n"5x(n)< o for some r E ]- 1,cj[ = E(XI i(r+)/(r +2)< cc.
essentially.
The proof of Theorem 2.2 is based on the following simple inequality, which is closely
coefficients (5(m)),, , and let k ,---, k, be any positive integers. Define so:= 1 and
s, := 1 +i=, k fobr j 1 ,- ,.-, n. Then for all events B1, B2, B,B with Bj E a(Aj) for
Proof. (By induction on n.) For n= 0 there is nothing to show. Suppose that (4)
holds for some n > 0. By the definition of a(k, +,) and the induction hypothesis we obtain
then
n+l n n+ l
The next lemma supplies us with the probabilistic part of the proof of Theorem 2.2.
Lemma 2.5. Let (k,), , be a sequence of positive integers and let (A,) and the numbers
s be as in Lemma 2.4. Suppose that X&1~ a(k,)< c and that Xf=0 P(A, ,)= ct for every
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
384 DIRK TASCHE
Proof. Denote by A the complement of the event A. Thus we have to show that for
fixed n eN
(5) lim P ( Am = 0.
N-+ oo m=n
Since P(Qm=, Am) is non-increasing in N, we only need to prove (5) for an appropriate
Now fix an e > 0. By assumption there is an integer M such that I_ =M+ I 1(km)< e.
M 00_
For the analytical part of Theorem 2.2 we need the following change-of-variable
formula for series, which can be proven by means of the integral test. Its proof will be
omitted.
Lemma 2.6. Let ? be a positive number and let 4): [, oo[ -+ ]0, co[ be a function with
(iii) There exist constants m, M > 0 such that m ? 4)'(x + 1)/1'(x) ? Mfor allx E [~-, o[.
Denote by y the inverse function of 4 and set : = 4(?). Then for every positive and non-
increasing sequence (a(n)),1N the series E,+i ,1 a([y(n)]) converges if and only if the series
Proof of Theorem 2.2. A straightforward application of Lemma 2.6 yields the follow-
ing implications.
n=l n=l
For r= -1:
Sd(n)
n=l n2 n=l
and
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 385
(7) a(n)= O(b") with bE ]O, 1[ =~ ( ) < 00for all a ]0, -log b[.
These observations suggest how to choose kn in Lemma 2.5. First, suppose r > - 1 and
r+l
Z P(An,+[lm)])= 00 z m-1(r+2)P(An+m)=co
S= 1 m = 1
(9)
9 mZ-1/(r+2)p(Am)= co.
P(A,,+,,) holds. Hence, under condition (i), Lemma 2.5 together with (6) and (9) implies
Now, let us consider the situation under condition (ii). Fix an arbitrary number
a E ]0, log b[ and let k := [log+(i)/a]. Then the same reasoning as in (8) leads to
[e-1 1 n+l
(10) sn + - (n + 1)log .
aae
e-1 1 x+l
(x) := - + - (x + 1)log
aae
and again b := -'. Observe that limy,. /'(y+ 1)1'(y)= 1. By Lemma 2.6 we get
m= 1 mf=1I
" P(A,) o
By the definition of y the inequality y(x) ? x/a holds for x ? 3. Hence, 4(y) ? ay for
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
386 DIRK TASCHE
aa
log(4(y)+ 1) = log(ay + 1)
Proof of Corollary 2.3. Because the arguments in the other cases are very similar,
we confine ourselves to considering the case of condition (ii). Since the random variables
n=l
Fix any r> -I and assume E(IX,I(r+'1)/(r+2))=) . Then (12) implies that the
series 1~ n- 1(r+2)P(IX,I >cn) diverges for all c> 0. By Theorem 2.2 it follows that
P(sup, IX,/n = oo)= 1. But sup, IX,/nl I 2 sup, In-' ` =1, XI, and thus n-'C= X
3. An example
We will show that Rieders' (1993) example in a slightly modified form can also serve
Proposition 3.1. For every r > - 1 there exists a sequence (X,),,n of identically dis-
Remark. Unfortunately, the example does not seem to work in the case of an exponen-
tial mixing rate. So the question is still open whether exponential mixing implies the
Proof of Proposition 3.1. Fix r > - 1 and define a probability distribution (Pk)k EN
Pk := ck-(3 +r)(log(k))-2,
p , if i=1,
0, otherwise.
Then (Y,) has a unique invariant distribution u = (Uk)k EN given by j, =( ]jpj)-' and
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 387
By Corollary 1 of Kesten and O'Brien (1976), we know that the strong mixing co-
n=l
Let g: N -+ ]0, oo[ be a non-decreasing function (which will be specified later) and
let X,:=(-j1)'g(Y,,) for all nE N. Then (13) implies (i). Define To:
inf{n > 1: Y,= 1}, Tk:= inf{n > Tk- : Y,= 1}, Rk:= Tk- Tk-l and N(k):=
I{1 ?j < k: Y= 1 }-1 for k ? 1. Then on the event {To<n} the following identity
holds a.s.
1 1To 1 N(n) 1
S- X= X+ E x+x
(14) 1 I N) R n
1 N~n)1
Since g is positive and non-decreasing we have I|=I (- 1)jg(j)l ?g(k) for all k E N.
1 1 "(")+'1 TN("n)
1 R,(n)+l 1 TN(n)+L-n
- (-01)+g(j) - Z (-1)ig(j)
n j=1 n j=1
< - g(RN(,n)+1).
Observe that (Rk)kE, is i.i.d. with P(R, =j)=pj for all jE N. Hence, if we assume
1~, g(j)pj to be finite, by the strong law there exists a number hE CR such that
1 m Ri
m--c0 m i= 1 j=l
By the same argument limm,, Rm/m = 0 holds. By the ergodic theorem limn,, N(n)/n=
y, > 0 a.s., and by (14), this implies the almost sure convergence of n-' C_, X,. So all
we have to do in order to obtain (ii) and (iii) is to choose g(k) such that
k=l
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
388 DIRK TASCHE
4. General probabilities
Among the assumptions of Theorem 2.2 the monotonicity of the probabilities appears
unnatural when compared to the assumptions of the second Borel-Cantelli lemma under
independence. Durrett's (1991, Theorem (6.6) ch. 1) version of the lemma suggests a
way to avoid this taint and even to get a stronger conclusion. The price to be paid for
conditions.
Notation. For any sequence (A,),, Nof events in a common probability space and all
nE- let
and
and uniform mixing coefficients and are obviously dominated by them. The idea of the
proof of parts (b) and (c) of the following theorem has been suggested by Rio's (1993)
(a) If liM , c, = oo and E~kl r(k)< oo, then lim ,o SI/c, = 1 almost surely.
(b) If lim,, n-1+(t+2)cn= OO and EX=I k'y(k)< oo for some t O, then limn , SI/c = 1
in probability.
(c) If limn-a c,(log+(n))-'='0 and y(n)=O(b") for some b ]O, 1[, then
Remarks. (1) Part (b) of the theorem can be regarded as a revision of Yoshihara's
(1979) Theorem 1. He stated his result under the assumptions kl=1 (a(k))" '+') < co for
some t > 0 and lim inf,,, n"P(A,) > 0 for some a E ]0, (t + 1)/(t + 2)[ on the mixing
coefficients and the probabilities resp. of the events. These conditions are slightly
stronger than our conditions. Concerning the probabilities this can be shown by a short
calculation. For the comparison of the conditions on the mixing coefficients, see Rio
(2) If limn, S,/c, = 1 in probability, then limk,, Sn/lCnk = 1 a.s. for some subsequence
and therefore P(on, Un,_ Am)= 1. Nevertheless, the question is interesting whether
the conclusion in (b) and (c) can be sharpened to almost sure convergence. Also, we
have to ask whether the rate condition on the mixing coefficients in (a) is really necessary.
In Proposition 4.2 and Proposition 4.3 we provide partial answers to these questions.
(3) Concerning the rate of divergence of I P(A,), the conditions in (b) and (c) are
obviously stronger than the corresponding conditions in Theorem 2.2. To get an intuition
for the gap, observe that in the case of non-increasing probabilities the condition
C",l n-li(t'+)P(A,)= c0 for some e > 0 implies lim sup._, n-1/' 1, P(Ak)= c00.
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 389
(4) It would be nice to find a sequence of strongly mixing events Am such that
P(O,I lU,, Am) equals 1 but S,/c, does not converge in probability.
Proof of Theorem 4.1. We follow as far as possible Durrett's (1991, p. 44) proof. For
var(S,,)
(6c)2
Cn k=1 j=k+I
1 n n l+2M
n - k=1 j=k+1
n-In
k= j=kIP(AkA)-P(A)P(A)I
n-1
k=l
Concerning the second term on the right-hand side of (17), note that
k ? n and kl/"~' +2 j
kk ? min(n,j+2).
This implies
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
390 DIRK TASCHE
n-1 n-kn
f E y(j) ? Ey(j)min(n,jt+2)
(19) "m.
j- '+(j)+n 2 (j).
j=1 j=mn+1
k=n k=n
n-1 n-k
k=1 j=mk+1
j=1 j=mn+l
j=1 j= mn +
(c) Again we have to prove (16). Choose M > 0 such that y(n) Mb" for all n E N.
Z Z IP(Ak nA)-P(Ak)P(A)I
k=1 j=k+l
n-1 k+mk n
n-1 n
( mkP(Ak)+M Z bj-)
k=1 j=k+mk+l
n-1 n - n-1
(21) 1 n 1 1+log n
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 391
k=1 j=k+l
EC P(A)+
which tends to 0 as n -* cG, since limn,- (log n)c, 0=O by assumption. This completes
the proof.
an impression of what can be reached by invoking general limit theorems, we finish our
considerations with two further results on the a.s. convergence of S,/c,. The first proposi-
tion is based on ideas from Blum et al. (1963) and their concept of *-mixing.
Proposition 4.2. Let (A,)n,, be a sequence of events such that 1 I1 P(A,) = co. For
every n E N let
P(A)P(B)
(n):= sup -(A)P(B) - 1 : AE a(A,,..., Ak),
BE a(An+k), P(A)P(B)> 0, k E
. I= I I(Ai)
lim =1 a.s.
Proof. Let X,:= I(A,)-P(A,), ~ "= a(A1,..., A,) and b,:= Y= P(A,). Without
loss of generality assume bn > 0 for all n E N. Our mixing assumption implies that
equation (2.22) from Hall and Heyde (1980, p. 40) is satisfied (with f(n) := f(n)). Hence,
by Theorem 2.20 of Hall and Heyde the assertion will follow if we show that
supn b1' = E(IXX I)< oo and 1"~ b-2E(X,2)< Oc. The first relation is obvious. For the
second we compute
N N P(A,)
Z b <2E(X2) b2
,n=1 =,n b 2
1 u bn-b_ 2 1 2
Our last proposition is based on a recent result of Rio (1995) on a.s. convergence of
Proposition 4.3. Let (A,),~ be a sequence of events such that C., P(A,) = c~ and
( 2P (A )t/(t+1)
(22) n < co
for some t ? 0. Suppose that the c-mixing coefficients of (A,) (cf. Definition 2.1) satisfy
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
392 DIRK TASCHE
Z.X= I(A,)
lim =1 a.s.
n -o X1=1 P(Ai)
0 P(A,) <
n=1 (i P(A,))2
For large t, condition (22) is very similar to this inequality, but it does not follow from
(2) There is no clear hierarchy between Theorem 4.1(b) and Proposition 4.3. For
t=O the condition lim,, n-1/(t+2)c, = oo is implied by (22). However, for t E N choose
P(A,) =
0, otherwise
shows that (22) does not follow from lim,_ n-U1(t+2)cn= 00.
=l_ 1P(A,)t/(t+l)
lim = 0.
n -+ (n=1 P (A,))2
Since
for every n E ~, it follows that ',I P(A,)(t+2)/"(t+1)=o, which in turn implies that
Proof of Proposition 4.3. Define the random variables X, and the numbers b, for
lemma, if we prove a.s. convergence of the series X",1 X,/b,. To get this convergence we
will examine the assumptions of Rio's (1995) Corollary 1. In his paper, Rio only uses
the notion of strong mixing. However, a careful inspection of his Lemma 1 reveals that
in Corollary 1 the assumption on the strong mixing coefficients can be replaced by the
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
On the second Borel-Cantelli lemma for strongly mixing sequences of events 393
i=1
First, consider the case t > 0. Then, by Hilder's inequality, for every i e N with the
a(x/2)(Qi (x))2dx
fg 1 )1/(t+ 1) 01 t/(t+1)
1=0
00/(t + 1)
n=0
0i-1 P(Ai)"(t +l)'b~< 00 it remains to show Y"0 (n + 1)'+'(a(n) - (n+ 1))< c. This
follows from
n=0
_ (0) + (t + 1) (n + 1)'i(n), N
n=l1
for all N E N. Concerning the case t= 0, observe that Q, (x) ? 1 and therefore
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
All use subject to JSTOR Terms and Conditions
394 DIRK TASCHE
References
BLUM, J. R., HANSON, D. L. AND KOOPMANS, L. H. (1963) On the strong law of large numbers for a class
HALL, P. AND HEYDE, C. C. (1980) Martingale Limit Theory and its Application. Academic Press, Boston.
IOSIFESCU, M. AND THEODORESCU, R. (1969) Random Processes and Learning Springer, Berlin.
KESTEN, H. AND O'BRIEN, G. L. (1976) Examples of mixing sequences. Duke Math. J1 43, 405-415.
RIEDERS, E. (1993) The size of the averages of strongly mixing random variables. Statist. Prob. Lett. 18,
57-64.
RIo, E. (1993) Covariance inequalities for strongly mixing processes. Ann. Inst. H. Poincard 29, 587-597.
RIo, E. (1995) A maximal inequality and dependent Marcinkiewicz-Zygmund strong laws. Ann. Prob.
23, 918-937.
TASCHE, D. (1995) First-passage time moments of strongly mixing stationary sequences. Preprint 452.
YOSHIHARA, K. (1979) The Borel-Cantelli lemma for strong mixing sequences of events and their applica-
This content downloaded from 155.198.30.43 on Thu, 17 Mar 2016 22:46:52 UTC
View publication stats
All use subject to JSTOR Terms and Conditions