numbers
Gunnar Englund & Timo Koski
Matematisk statistik
KTH
2008
1 Introduction
BorelCantelli lemmas are interesting and useful results especially for proving
the law of large numbers in the strong form.
We consider a sequence events A
1
, A
2
, A
3
, . . . and are intrested in the question
of whether inﬁnitely many events occur or if possibly only a ﬁnite number of
them occur.
We set
F
n
=
∞
¸
k=n
A
k
and G
n
=
∞
¸
k=n
A
k
. (1)
If G
n
in (1) occurs, this means that all A
k
for k ≥ n occur. If there is some
such n, this means in other words that from this n on all A
k
occur for k ≥ n.
With
H =
∞
¸
n=1
G
n
=
∞
¸
n=1
∞
¸
k=n
A
k
we have that if H occurs, then there is an n such that all A
k
with k ≥ n occur.
Sometimes we denote H with liminf A
k
.
The fact that F
n
occurs implies that there is some A
k
for k ≥ n which occurs.
If F
n
in (1) occurs for all n this implies that inﬁnitely many of the A
k
:s occur.
We form therefore
E =
∞
¸
n=1
F
n
=
∞
¸
n=1
∞
¸
k=n
A
k
.
If E occurs, then inﬁnitely many of A
k
:s occur. Sometimes we write this as
E = {A
n
i.o.} where i.o. is to be read as ”inﬁnitely often”, i.e., inﬁnitely many
times. E is sometimes denoted with limsup A
k
.
We need a couple of auxiliary results (the lemmas below) of probability calculus
(found, e.g., on page 3 of [1]) that are basic in the sense that they are derived
directly from the Kolmogorov axioms.
1
First, we consider a sequence of events B
k
∈ F that is increasing to B ∈ F.
This means that
B
1
⊂ B
2
⊂ B
3
⊂ . . . ⊂ B
n
⊂ B
n+1
⊂ . . . ⊂ B
and thus
∪
∞
k=1
B
k
= B.
We can write this as
B
n
↑ B
Lemma 1 If B
n
↑ B, then P(B) = lim
n→∞
P (B
n
).
Proof: We can write ∪
∞
k=1
B
k
= ∪
∞
k=2
(B
k
\ B
k−1
) ∪ B
1
, since B
k
are increasing.
P(B) = P (∪
∞
k=1
B
k
) = P (∪
∞
k=2
(B
k
\ B
k−1
) ∪ B
1
)
But the sets in the decomposition are seen to be disjoint, and hence the axiom
of countable additivity yields
P (∪
∞
k=2
(B
k
\ B
k−1
) ∪ B
1
) =
∞
¸
k=2
P (B
k
\ B
k−1
) +P (B
1
)
= lim
n→∞
n
¸
k=2
P (B
k
\ B
k−1
) +P (B
1
)
Now we observe that since B
k−1
⊂ B
k
, we have
P (B
k
\ B
k−1
) = P (B
k
) −P (B
k−1
) .
Therefore we get a telescoping series
n
¸
k=2
P (B
k
\ B
k−1
) +P (B
1
) = P (B
n
) −P (B
n−1
) +P (B
n−1
) −P (B
n−2
) +. . . +
+P (B
2
) −P (B
1
) +P (B
1
) = P (B
n
) .
In other words we have shown that
P(B) = lim
n→∞
P (B
n
) .
We take next a sequence of events B
k
∈ F that is decreasing. This means that
B
1
⊃ B
2
⊃ B
3
⊃ . . . ⊃ B
k
⊃ B
k+1
⊃ . . . ⊃ ∩
∞
k=1
B
k
and thus we write
B
n
↓ ∩
∞
k=1
B
k
.
Lemma 2 If B
n
↓ ∩
∞
k=1
B
k
, then P (∩
∞
k=1
B
k
) = lim
n→∞
P (B
n
).
2
Proof: We use the axiom of probability of complementary events
P (∩
∞
k=1
B
k
) = 1 −P ((∩
∞
k=1
B
k
)
c
) . (2)
When we apply one of De Morgan
,
s rules we get
(∩
∞
k=1
B
k
)
c
= ∪
∞
k=1
B
c
k
.
Now we observe that if B
k
⊃ B
k+1
, then B
c
k
⊂ B
c
k+1
, i.e., the complement
events of a decreasing sequence of events are an increasing sequence of events.
Thus the ﬁrst lemma implies
P (∪
∞
k=1
B
c
k
) = lim
n→∞
P (B
c
n
) .
Hence we have that
P ((∩
∞
k=1
B
k
)
c
) = P (∪
∞
k=1
B
c
k
)
= lim
n→∞
P (B
c
n
) .
This we shall insert in (2) and get
P (∩
∞
k=1
B
k
) = 1 − lim
n→∞
P (B
c
n
)
= 1 − lim
n→∞
(1 −P (B
n
))
= 1 −1 + lim
n→∞
P (B
n
) = lim
n→∞
P (B
n
) .
This completes the proof.
2 The BorelCantelli results
One notes that F
n
is a decreasing set of events. This is simply so because
F
n
=
∞
¸
k=n
A
k
= A
n
¸
∞
¸
k=n+1
A
k
= A
n
¸
F
n+1
and thus
F
n
⊃ F
n+1
.
Thus the second lemma above gives
P(E) = P(
∞
¸
n=1
F
n
) = lim
n→∞
P(F
n
) = lim
n→∞
P(
∞
¸
k=n
A
k
).
We have, however, by Boole
,
s inequality that
P(
∞
¸
k=n
A
k
) ≤
∞
¸
k=n
P(A
k
)
and this sum → 0 as n → ∞, if the sum
¸
∞
1
P(A
k
) converges. This implies
that we have shown following proposition, also known as the BorelCantelli
lemma.
3
Proposition 1 BorelCantelli lemma
If
¸
∞
n=1
P(A
n
) < ∞ then it holds that P(E) = P(A
n
i.o) = 0, i.e., that with
probability 1 only ﬁnitely many A
n
occur.
One can observe that no form of independence is required, but the proposition
holds in general, i.e., for any sequence of events.
There is a converse to the BorelCantelli lemma obtained if we assume that
the events A
1
, A
2
, . . . are independent .
Proposition 2 Converse BorelCantelli lemma
If A
1
, A
2
, . . . are independent and
∞
¸
n=1
P(A
n
) = ∞,
then it holds that P(E) = P(A
n
i.o) = 1, i.e., it holds with probability 1 that
inﬁnitely many A
n
occur.
Proof: We have by independence
P(
∞
¸
k=n
A
∗
k
) =
∞
¸
k=n
P(A
∗
k
) =
∞
¸
k=n
(1 −P(A
k
)).
Since 1 −x ≤ e
−x
we get 1 −P(A
k
) ≤ e
−P(A
k
)
and
P(
∞
¸
k=n
A
∗
k
) ≤ exp(−
∞
¸
k=n
P(A
k
)).
If now
¸
∞
n=1
P(A
n
) = ∞, then the sum in the exponent diverges and we obtain
P(
∞
¸
k=n
A
∗
k
) = 0.
Thus it holds also that
P(
∞
¸
n=1
∞
¸
k=n
A
∗
k
) = 0,
which implies by De Morgan
,
s rules that
P(
∞
¸
n=1
∞
¸
k=n
A
k
) = 1 −P(
∞
¸
n=1
∞
¸
k=n
A
∗
k
) = 1 −0 = 1
i.e., that inﬁnitely many A
k
:n occur with probability 1.
4
3 Some examples of applications
Exempel 1 Let X
1
, X
2
, X
3
. . . be independent equidistributed with continu
ous distribution. We let
U
n
=
1 if X
n
> X
j
for j = 1, 2, . . . n −1
0 annars.
This says simply that U
n
= 1 if X
n
is a ”record”, i.e., the largest value observed
so far. We set A
n
= {U
n
= 1}.
We see that P(U
n
= 1) = 1/n, since the probability that the largest of n values
should occur in the round n is 1/n for reasons of symmetry. Furthermore,
A
1
, A
2
, . . . are independent. We have
P(A
m
∩ A
m+1
∩ . . . A
m+k
) = P(A
m
A
m+1
∩ . . . A
m+k
)P(A
m+1
∩ . . . A
m+k
)
and A
m
and A
m+1
∩ . . . A
m+k
are, of course, independent since A
m
is only
concerned with the relative magnitudes of the m ﬁrst of Xvariables.
We get then, as
∞
¸
n=1
P(A
n
) =
∞
¸
n=1
1
n
= ∞,
that P(A
n
i.o.) = 1 i.e., inﬁnitely many A
n
occur. We get inﬁnitely many
records – a result which perhaps (?) is selfevident. Furthermore, we get
E(
∞
¸
n=1
U
n
U
n+1
) =
∞
¸
n=1
E(U
n
U
n+1
) =
∞
¸
n=1
P(A
n
)P(A
n+1
) =
∞
¸
n=1
1
n(n + 1)
< ∞
and hence
¸
∞
1
U
n
U
n+1
is ﬁnite with probability 1. The conclusion is that there
only occurs a ﬁnitely number of ”double records”, i.e., a record two times in a
row – this result is by no means trivial.
Exempel 2 Let X
1
, X
2
, . . . be independent and equidistributed. Then it holds
that
E(X
1
) =
∞
0
P(X
1
 > x)dx =
∞
¸
n=0
n+1
n
P(X
1
 > x)dx
≤
∞
¸
n=0
P(X
1
 > n) ≤ 1 +
∞
¸
n=1
P(X
n
 > n),
but also
E(X
1
) =
∞
¸
n=0
n+1
n
P(X
1
 > x)dx ≥
∞
¸
n=0
P(X
1
 > n+1) =
∞
¸
n=1
P(X
n
 > n).
If now E(X
1
) < ∞ we see that
¸
∞
n=1
P(X
n
 > n) < ∞ which according
to BorelCantelli lemma implies P(X
n
 > n i.o.) = 0. On the other hand, if
E(X
1
) = ∞and thereby
¸
∞
n=0
P(X
n
 > n) = ∞, the converse BorelCantelli
lemma entails that P(X
n
 > n i.o.) = 1. If E(X
k
) are ﬁnite (respectively
inﬁnite ) X
n
 will be larger than n inﬁnitely many times with probability 0
(respectively 1).
5
4 Proof of the strong form of the law of large
numbers
We let X
1
, X
2
, . . . be independent equidistributed with E(X
i
) = mand Var(X
i
) =
σ
2
< ∞ and deﬁne S
n
= X
1
+ X
2
+ · · · + X
n
. We are interested in showing
the strong form of the law of large numbers (SLLN), i.e., that it holds with
probability one that S
n
/n → m as n → ∞. This means that we want to prove
that
P( lim
n→∞
S
n
n
= m) = 1,
i.e., that there exists a set Ω
0
with P(Ω
0
) = 1 where for every ω ∈ Ω
0
it holds
that
lim
n→∞

S
n
n
−m = 0.
We need in other words to prove that for every ω ∈ Ω
0
and for every ε > 0
there is N(ω, ε) so that if n ≥ N(ω, ε) holds that S
n
/n −m ≤ ε.
It suﬃces to prove that 
S
n
n
−m > ε can occur only a ﬁnite number of times,
i.e.,
lim
N→∞
P(
S
n
n
−m > ε some n ≥ N) = 0.
Note the distinction with regard to the law of large numbers in the weak form,
which says that that for all ε > 0
P(
S
n
n
−m > ε) → 0 as n → ∞.
In words: for the law of large numbers in the strong form S
n
/n −m must be
small for all suﬃciently large n for all ω ∈ Ω
0
where P(Ω
0
) = 1.
In tossing a coin we can code heads and tails with 0 and 1, respectively, and
we can identify an ω with a number in the intervall [0, 1] drawn at random,
where binary expansion gives the sequence of zeros and ones. The law of large
numbers says in this case that we will obtain with probability 1 a number
such that the proportion of 1:s in sequence converges towards 1/2. There can
be ”exceptional” ω  for example the sequence 000 . . . is possible, but such
exceptional sequences have the probability 0.
After these deliberations of pedagogic nature let us get on with the proof.
Proof of SLLN: Without restriction of generality we can assume that E(X
i
) =
m = 0, since we in any case can consider X
i
− m. We have V (S
n
) = nσ
2
. By
Chebyshev
,
s inequality it holds that
P(S
n
 > nε) ≤
V (S
n
)
(nε)
2
=
nσ
2
(nε)
2
=
σ
2
nε
2
.
Unfortunately the harmonic series
¸
∞
1
1/n is divergent so we cannot use Borel
Cantelli lemma directly. But it holds that
¸
∞
1
1/n
2
< ∞ and this means that
6
we can use the lemma for n
2
, n = 1, 2, . . . . We have
P(S
n
2 > n
2
ε) ≤
σ
2
n
2
ε
2
.
In other words it holds by BorelCantelli lemma that P(
S
n
2
n
2
 > ε i.o.) = 0
which proves that (with probability 1) S
2
n
/n
2
→ 0. We have in other words ma
naged to prove that for the subsequence n
2
, n = 1, 2, . . . there is convergence
with probability 1. It remains to ﬁnd out what will happen between these n
2
.
We deﬁne therefore
D
n
= max
n
2
≤k<(n+1)
2
S
k
−S
n
2,
i.e., the largest of the deviation fromS
n
2 that can occur between n
2
and (n+1)
2
.
We get
D
2
n
= max
n
2
≤k<(n+1)
2
(S
k
−S
n
2)
2
≤
(n+1)
2
−1
¸
k=n
2
(S
k
−S
n
2)
2
,
where we used the rather crude inequality max(x, y) ≤ (x + y). This
entails
E(D
2
n
) ≤
(n+1)
2
−1
¸
k=n
2
E((S
k
−S
n
2)
2
).
But E((S
k
− S
n
2)
2
) = (k − n
2
)σ
2
≤ 2nσ
2
as n
2
≤ k < (n + 1)
2
and there are
2n terms in the sum and this entails
E(D
2
n
) ≤ (2n)(2n)σ
2
= 4n
2
σ
2
.
With Chebyshev
,
s inequality this gives
P(D
n
> n
2
ε) ≤
4n
2
σ
2
(n
2
ε)
2
=
4σ
2
n
2
ε
2
.
In other words, D
n
/n
2
→ 0 holds with probability 1. Finally this yields for k
between n
2
and (n + 1)
2
that

S
k
k
 ≤
S
n
2 +D
n
k
≤
S
n
2 +D
n
n
2
→ 0.
This means that we have succeeded in proving that S
n
/n → 0 with probability
1. We have done this under additional condition that Var(X
i
) = σ
2
, but with
a painstaking eﬀort we can in fact prove that this additional condition is not
necessary.
Referenser
[1] Z. Brezniak and T. Zastawniak (2007): Basic Stochastic Processes. Spring
er Verlag, London.
7
k=1 2 . since Bk are increasing. This means that B1 ⊂ B2 ⊂ B3 ⊂ .First. . . . ⊃ Bk ⊃ Bk+1 ⊃ . and hence the axiom of countable additivity yields ∞ P (∪∞ (Bk k=2 \ Bk−1 ) ∪ B1 ) = k=2 n P (Bk \ Bk−1 ) + P (B1 ) = lim n→∞ P (Bk \ Bk−1 ) + P (B1 ) k=2 Now we observe that since Bk−1 ⊂ Bk . k=1 k=2 P (B) = P (∪∞ Bk ) = P (∪∞ (Bk \ Bk−1 ) ∪ B1 ) k=1 k=2 But the sets in the decomposition are seen to be disjoint. . then P (B) = limn→∞ P (Bn ). . then P (∩k=1 Bk ) = limn→∞ P (Bn ). Therefore we get a telescoping series n P (Bk \ Bk−1 ) + P (B1 ) = P (Bn ) − P (Bn−1 ) + P (Bn−1 ) − P (Bn−2 ) + . Proof: We can write ∪∞ Bk = ∪∞ (Bk \ Bk−1 ) ∪ B1 . k=1 We can write this as Bn ↑ B Lemma 1 If Bn ↑ B. ⊂ B and thus ∪∞ Bk = B. . we have P (Bk \ Bk−1 ) = P (Bk ) − P (Bk−1 ) . . ⊂ Bn ⊂ Bn+1 ⊂ . . This means that B1 ⊃ B2 ⊃ B3 ⊃ . k=1 ∞ Lemma 2 If Bn ↓ ∩∞ Bk . + k=2 +P (B2 ) − P (B1 ) + P (B1 ) = P (Bn ) . . . n→∞ We take next a sequence of events Bk ∈ F that is decreasing. In other words we have shown that P (B) = lim P (Bn ) . ⊃ ∩∞ Bk k=1 and thus we write Bn ↓ ∩∞ Bk . we consider a sequence of events Bk ∈ F that is increasing to B ∈ F .
. also known as the BorelCantelli lemma.Proof: We use the axiom of probability of complementary events ∞ P (∩∞ Bk ) = 1 − P ((∩k=1 Bk )c ) . i. n→∞ n→∞ This completes the proof. by Boole. 2 The BorelCantelli results ∞ ∞ One notes that Fn is a decreasing set of events. 3 . s inequality that ∞ ∞ P( k=n Ak ) ≤ k=n P (Ak ) ∞ and this sum → 0 as n → ∞. then Bk ⊂ Bk+1 . n→∞ This we shall insert in (2) and get c P (∩∞ Bk ) = 1 − lim P (Bn ) k=1 n→∞ = 1 − lim (1 − P (Bn )) n→∞ = 1 − 1 + lim P (Bn ) = lim P (Bn ) . This implies that we have shown following proposition.e. the complement events of a decreasing sequence of events are an increasing sequence of events. Thus the ﬁrst lemma implies c c P (∪∞ Bk ) = lim P (Bn ) . k=1 n→∞ Hence we have that c P ((∩∞ Bk )c ) = P (∪∞ Bk ) k=1 k=1 c = lim P (Bn ) . s rules we get c (∩∞ Bk )c = ∪∞ Bk . however. k=1 (2) When we apply one of De Morgan. We have. k=1 k=1 c c Now we observe that if Bk ⊃ Bk+1 . Thus the second lemma above gives ∞ ∞ P (E) = P ( n=1 Fn ) = lim P (Fn ) = lim P ( n→∞ n→∞ k=n Ak ). This is simply so because Fn = k=n Ak = An k=n+1 Ak = An Fn+1 and thus Fn ⊃ Fn+1 . if the sum 1 P (Ak ) converges.
k=n Since 1 − x ≤ e−x we get 1 − P (Ak ) ≤ e−P (Ak ) and ∞ ∞ P( k=n A∗ ) k ≤ exp(− k=n P (Ak )). There is a converse to the BorelCantelli lemma obtained if we assume that the events A1 .e. . for any sequence of events. . are independent . P( n=1 k=n Ak ) = 1 − P ( n=1 k=n A∗ ) = 1 − 0 = 1 k i. n=1 then it holds that P (E) = P (An i. then the sum in the exponent diverges and we obtain ∞ P( k=n A∗ ) = 0. ... . Proof: We have by independence ∞ ∞ ∞ P( k=n A∗ ) = k k=n P (A∗ ) = k (1 − P (Ak )). it holds with probability 1 that inﬁnitely many An occur.. A2 . that with n=1 probability 1 only ﬁnitely many An occur.e. that inﬁnitely many Ak :n occur with probability 1. are independent and ∞ P (An ) = ∞. i.e. i.o) = 1.Proposition 1 BorelCantelli lemma If ∞ P (An ) < ∞ then it holds that P (E) = P (An i. Proposition 2 Converse BorelCantelli lemma If A1 . . k n=1 k=n which implies by De Morgan s rules that ∞ ∞ ∞ ∞ . 4 . If now ∞ n=1 P (An ) = ∞. .e. One can observe that no form of independence is required. k Thus it holds also that P( ∞ ∞ A∗ ) = 0..o) = 0. i. but the proposition holds in general. A2 .
i. . inﬁnitely many An occur. . 5 . . Am+k )P (Am+1 ∩ .e. We set An = {Un = 1}. as ∞ ∞ 1 P (An ) = = ∞. Furthermore. We have P (Am ∩ Am+1 ∩ .. if E(X1 ) = ∞ and thereby ∞ P (Xn  > n) = ∞. . Exempel 2 Let X1 . of course.3 Some examples of applications Exempel 1 Let X1 .o. Am+k ) = P (Am Am+1 ∩ . i. . On the other hand. . The conclusion is that there only occurs a ﬁnitely number of ”double records”. X3 . This says simply that Un = 1 if Xn is a ”record”. . . be independent and equidistributed. we get ∞ ∞ ∞ ∞ E( n=1 Un Un+1 ) = n=1 ∞ 1 E(Un Un+1 ) = n=1 P (An )P (An+1 ) = n=1 1 <∞ n(n + 1) and hence Un Un+1 is ﬁnite with probability 1. We get inﬁnitely many records – a result which perhaps (?) is selfevident. . .. independent since Am is only concerned with the relative magnitudes of the m ﬁrst of Xvariables. be independent equidistributed with continuous distribution. . We see that P (Un = 1) = 1/n. A2 . Am+k are..e. .) = 0. 2. are independent. a record two times in a row – this result is by no means trivial. . . the converse BorelCantelli n=0 lemma entails that P (Xn  > n i. If now E(X1 ) < ∞ we see that P (Xn  > n) < ∞ which according to BorelCantelli lemma implies P (Xn  > n i. Furthermore. X2 . If E(Xk ) are ﬁnite (respectively inﬁnite ) Xn  will be larger than n inﬁnitely many times with probability 0 (respectively 1). We get then. but also ∞ n+1 ∞ ∞ E(X1 ) = n=0 n P (X1 > x)dx ≥ n=0 ∞ n=1 P (X1  > n + 1) = n=1 P (Xn  > n). We let Un = 1 0 if Xn > Xj for j = 1. .) = 1. Am+k ) and Am and Am+1 ∩ . the largest value observed so far. .) = 1 i.e. Then it holds that ∞ ∞ n+1 E(X1 ) = 0 P (X1  > x)dx = n=0 ∞ n P (X1 > x)dx ∞ ≤ n=0 P (X1  > n) ≤ 1 + n=1 P (Xn  > n). since the probability that the largest of n values should occur in the round n is 1/n for reasons of symmetry. n − 1 annars. .o. n n=1 n=1 that P (An i. .o. X2 . . A1 .
.for example the sequence 000 . . There can be ”exceptional” ω . since we in any case can consider Xi − m. that there exists a set Ω0 with P (Ω0 ) = 1 where for every ω ∈ Ω0 it holds that Sn lim  − m = 0. n→∞ n We need in other words to prove that for every ω ∈ Ω0 and for every ε > 0 there is N(ω. n In words: for the law of large numbers in the strong form Sn /n − m must be small for all suﬃciently large n for all ω ∈ Ω0 where P (Ω0 ) = 1. . where binary expansion gives the sequence of zeros and ones. ε) holds that Sn /n − m ≤ ε. P ( lim n→∞ n i. Proof of SLLN: Without restriction of generality we can assume that E(Xi ) = m = 0. This means that we want to prove that Sn = m) = 1. is possible. We have V (Sn ) = nσ 2 . . After these deliberations of pedagogic nature let us get on with the proof.e. The law of large numbers says in this case that we will obtain with probability 1 a number such that the proportion of 1:s in sequence converges towards 1/2. We are interested in showing the strong form of the law of large numbers (SLLN). (nε)2 (nε)2 nε Unfortunately the harmonic series ∞ 1/n is divergent so we cannot use Borel1 Cantelli lemma directly. X2 . which says that that for all ε > 0 P ( Sn − m > ε) → 0 as n → ∞. Sn It suﬃces to prove that  − m > ε can occur only a ﬁnite number of times. n i. But it holds that ∞ 1/n2 < ∞ and this means that 1 6 ..4 Proof of the strong form of the law of large numbers We let X1 . be independent equidistributed with E(Xi ) = m and Var(Xi ) = σ 2 < ∞ and deﬁne Sn = X1 + X2 + · · · + Xn . By Chebyshev. In tossing a coin we can code heads and tails with 0 and 1. s inequality it holds that V (Sn ) nσ 2 σ2 P (Sn  > nε) ≤ = = 2. . ε) so that if n ≥ N(ω. that it holds with probability one that Sn /n → m as n → ∞. . i. but such exceptional sequences have the probability 0.e. respectively. N →∞ n Note the distinction with regard to the law of large numbers in the weak form.e. Sn lim P ( − m > ε some n ≥ N) = 0. and we can identify an ω with a number in the intervall [0. 1] drawn at random..
2. This entails (n+1)2 −1 2 E(Dn ) ≤ k=n2 E((Sk − Sn2 )2 ). Dn /n2 → 0 holds with probability 1. n = 1.o. 7 . But E((Sk − Sn2 )2 ) = (k − n2 )σ 2 ≤ 2nσ 2 as n2 ≤ k < (n + 1)2 and there are 2n terms in the sum and this entails 2 E(Dn ) ≤ (2n)(2n)σ 2 = 4n2 σ 2 . n = 1. We have P (Sn2  > n2 ε) ≤ σ2 . n2 ε2 In other words it holds by BorelCantelli lemma that P ( Sn 2  > ε i. Finally this yields for k between n2 and (n + 1)2 that  Sn2  + Dn Sn2  + Dn Sk ≤ ≤ → 0. s inequality this gives P (Dn > n2 ε) ≤ 4n2 σ 2 4σ 2 = 2 2. . but with a painstaking eﬀort we can in fact prove that this additional condition is not necessary. (n2 ε)2 nε In other words. . . k k n2 This means that we have succeeded in proving that Sn /n → 0 with probability 1.. Zastawniak (2007): Basic Stochastic Processes. It remains to ﬁnd out what will happen between these n2 .we can use the lemma for n2 . London. We have done this under additional condition that Var(Xi ) = σ 2 . Springer Verlag. n ≤k<(n+1) i. . there is convergence with probability 1. We deﬁne therefore Dn = 2 max 2 Sk − Sn2 .e. . Brezniak and T. y) ≤ (x + y). the largest of the deviation from Sn2 that can occur between n2 and (n+1)2 .) = 0 n2 2 which proves that (with probability 1) Sn /n2 → 0. We have in other words managed to prove that for the subsequence n2 . where we used the rather crude inequality max(x. . Referenser [1] Z. 2. With Chebyshev. . We get (n+1)2 −1 2 Dn = n2 ≤k<(n+1)2 max (Sk − Sn2 ) ≤ k=n2 2 (Sk − Sn2 )2 .