A sequence of exceptional sets can be small at every stage and still visit the same point infinitely often. To prove pointwise convergence, it is not enough merely to estimate one exceptional set: we must control the set of points that keep returning to exceptions. The Borel–Cantelli lemmas turn sums of measures into such a statement. Their two directions have different hypotheses, and only the second requires independence.
Let be a measure space.
Definition 3.5.1
(Limit superior and limit inferior of sets)
For measurable sets , define
A point belongs to the limsup exactly when it is in infinitely many ; it belongs to the liminf exactly when it is in every from some index onward. These are measurable by countable unions and intersections. The sets need not themselves have a set-theoretic limit.
Summable exceptions
Lemma 3.5.2
(First Borel–Cantelli lemma)
On any measure space, if , then .
Proof
(Bound the limsup by every tail union)
For each , the limsup is contained in , so monotonicity and countable subadditivity give
The last expression is the tail of a convergent numerical series and tends to zero. No independence or finite total measure of is needed.
Thus almost every point belongs to only finitely many exceptional sets. For example, if for all , this conclusion follows whether the sets are nested, disjoint or dependent. The theorem does not say that the union of all is null: an early exceptional set can have positive measure.
Worked example 3.5.3
(Long runs of ones in binary digits)
On with Lebesgue measure and binary digits (the expansion that is eventually zero at dyadic rationals), let be the set where . Prescribing these digits selects of the dyadic cells of length , so and
By the first lemma, almost every lies in only finitely many : a run of ones occupying exactly positions to occurs for only finitely many . The events are dependent, since and share the digits , yet the first lemma needs no independence. The edge case shows the union is not null: .
Divergent sums and independence
Now assume ; write for .
Definition 3.5.4
(Independent events)
Events are independent if for every finite set of distinct indices ,
This is joint independence of every finite subfamily, stronger than merely requiring pairwise independence. It implies the corresponding product formula for finite intersections of complements: expand and integrate, using the independence of every subcollection.
Lemma 3.5.5
(Second Borel–Cantelli lemma)
If the events are independent and , then .
Proof
(The probability of no occurrence in a tail is zero)
Fix . For , finite independence of the complements and give
as , because the tail of a divergent series of nonnegative terms diverges. The sets on the left decrease in , and their first set has finite measure, so continuity from above gives . The complement of the limsup is the union over of these null events: it consists of points with no occurrences after some index. Hence the limsup has probability one.
Proof X-ray
Where independence enters
The proof attacks the complement: a point outside the limsup has no occurrence in some tail, so it suffices to show that each tail event is null. Independence is used exactly once, to turn the probability of a finite intersection of complements into the product ; this needs independence of the complements, which follows from joint independence of the . The key analytic move is , which converts the product into the exponential of a partial sum, so divergence of the series is what drives the bound to zero. Divergence of the whole series is needed for every , which holds because removing finitely many terms keeps a series divergent. The argument closes with continuity from above, legitimate because is finite, and a countable union of null tail events.
The product identity is the only place where the events interact, and without it the conclusion can fail.
Counterexample mode
A divergent sum alone is not enough
Claim: if , then almost every point lies in infinitely many . On with Lebesgue probability measure, let . Then , but every fixed is outside once , so the limsup is empty. These nested events are not independent: , whereas . A divergent sum does not prevent the same shrinking region from accounting for all the mass; the repair is the independence hypothesis.
From exceptional events to pointwise convergence
Proposition 3.5.6
(Summable exceptions give almost-everywhere convergence)
Let be measurable real-valued functions finite almost everywhere. If for every positive integer ,
then almost everywhere.
Proof
(One null set for each threshold)
The first lemma makes the infinitely-often set null for each . Outside their countable union, for each the error is eventually at most ; given , choose with . The final index may differ for each point and each ; uniform convergence is not asserted.
Common mistake
(One threshold is not enough)
Summability for a single threshold only bounds the eventual error by that threshold. On a probability space let for every and . Then for all , a summable sequence, yet does not converge to anywhere. The hypothesis must hold for every , and the proof needs the countable union over .
A useful sufficient estimate follows from Markov's inequality.
Corollary 3.5.7
(Summable power errors)
If and , then almost everywhere.
Proof
(Markov's inequality at each threshold)
For each , , so the threshold sums of the proposition are at most .
For this is an integral statement; it does not claim that the usual expression is a norm.
Worked example 3.5.8
(Errors of size in mean square)
Suppose for every . Markov's inequality with gives, for each ,
and . The proposition gives almost everywhere. The constant grows with , but each threshold is handled separately, so no uniformity in is needed.
Common mistake
(Pointwise convergence does not force summable exceptions)
The converse of the proposition is false. On with Lebesgue measure, converges pointwise to zero, yet for the exceptional measures sum to . Summability is a strong sufficient condition, not a characterization of pointwise convergence.
A binary-digit example
Worked example 3.5.9
(Every digit value recurs almost surely)
On with Lebesgue probability measure, use the binary expansion that is eventually zero at dyadic rationals. Let be the set of points whose th binary digit is 1. Equivalently, is the union of the half-open dyadic cells of length with odd cell index. Exactly half the cells occur, so .
Any prescribed values of the first digits specify one half-open cell of length , with probability . Summing over unspecified positions proves that any finite set of digit events is independent. Thus , and the second lemma says almost every point has digit 1 infinitely often. The complementary digit-zero events are independent as well and have the same probabilities, so almost every point also has digit 0 infinitely often. Intersecting these two probability-one sets proves both statements simultaneously.
The digit-1 conclusion concerns almost every point: dyadic rationals have eventually zero expansions under the chosen convention, so they are exceptions. In fact the digit-0 conclusion is stronger here: every has infinitely many zeros in its chosen expansion. An expansion eventually consisting only of ones represents a dyadic rational, and the chosen expansion of that rational is eventually zero.
Practice with solutions
Quick check
(If , must ? Give an example or proof.)
Solution · Answer
No. Take and for under Lebesgue measure. The sum is finite but the union has measure ; the limsup is empty.
Quick check
(On let . Calculate the limsup and the sum of measures. Why does the second lemma not apply?)
Solution · Answer
The limsup is empty, while the sum is the divergent harmonic series. The events are nested and not jointly independent, so the second lemma's hypothesis fails.
Quick check
(On with binary digits as above, what does the second lemma say about the events that the th digit is 0?)
Solution · Answer
The events are independent with probability each, so the lemma gives digit 0 infinitely often almost everywhere. Under the chosen expansion convention the stronger statement holds for every point: an expansion eventually all ones is dyadic, for which the chosen expansion is eventually zero.