A sequence of functions can be close at each fixed point eventually, or close on most of the space at each late stage, without these being the same statement. Even either kind of closeness may coexist with a persistent integral carried by a narrow spike. Convergence in measure quantifies the size of the exceptional set; uniform integrability prevents a small set from carrying uncontrolled integral mass. Together they give a useful substitute for a single dominating function.
Throughout, is a measure space. Functions are real-valued and measurable, finite almost everywhere; functions in are identified up to null sets. We use countable subadditivity, continuity of measure from above for sets of finite measure, Fatou's lemma, and dominated convergence.
Convergence in measure and subsequences
Definition 1.5.1
(Convergence in measure)
A sequence converges in measure to if, for every ,
The exceptional set can change with . Unlike uniform convergence, the definition permits large errors on a set whose measure tends to zero. Unlike almost-everywhere convergence, it does not demand that a given point eventually avoid all exceptional sets.
Proposition 1.5.2
(Uniqueness of the limit in measure)
If converges in measure to both and , then almost everywhere.
Proof
(Splitting the threshold in half)
For each , lies in the union of and for every . Both measures tend to zero, so . A countable union over gives almost everywhere.
Theorem 1.5.3
(Subsequence theorem)
On any measure space, if in measure, then some subsequence converges to almost everywhere.
Proof
(A summable choice of exceptional sets)
Choose strictly increasing so that
Let be these exceptional sets. For every , the set of points belonging to infinitely many is contained in , whose measure is at most . It is therefore null. Outside it, eventually , so the subsequence converges. No independence of the sets is needed.
Convergence in measure does not imply almost-everywhere convergence of the whole sequence, as the following sequence shows.
Worked example 1.5.4
(Dyadic indicators that converge nowhere)
On with Lebesgue measure, enumerate the half-open dyadic intervals by stages: for each , use all , , before proceeding to the next stage. Let be the indicator of the th interval. Its support measure tends to zero, so in measure. Each point belongs to exactly one interval at every stage and to no other interval in that stage. It therefore has one value at each stage and, for every , at least one value ; both occur infinitely often. The full sequence converges nowhere on .
Egorov: uniform convergence away from a small set
Theorem 1.5.5
(Egorov's theorem)
If and almost everywhere, then for every there is a measurable set with such that uniformly on .
Proof
(One threshold index for each error level)
Include the null set where convergence fails in . For positive integers put
For fixed , these sets decrease as grows, and their intersection is null: at each point of convergence, eventually the error is at most . Since , continuity from above gives . Choose with and add all these sets to . Then . Outside , every satisfies . Given an arbitrary error tolerance, choose with below it; the same works for all points outside .
In particular, on a finite-measure space, almost-everywhere convergence implies convergence in measure. One can also prove this directly: for fixed , the decreasing unions shrink to a null set, and continuity from above bounds each later exceptional set.
Counterexample mode
Finite total measure is not cosmetic
On let . At each fixed point the sequence is eventually zero, but for every : convergence in measure fails. Moreover, no set of finite measure can be removed to make convergence uniform. Uniform convergence outside would force every sufficiently late interval into , whose measure would then be infinite. The finite-measure hypothesis in Egorov's theorem is essential.
Uniform integrability
Definition 1.5.6
(Uniformly integrable family)
A family is uniformly integrable here if
The second condition gives one small-set threshold for the entire family. The first makes the global bound explicit. It does not follow from the small-set condition alone on spaces with atoms: on a one-point space of mass one, the family satisfies that condition whenever , but its norms are unbounded. On any measure space, a family dominated by one satisfies both conditions. Indeed, choose so is small, and use .
Common mistake
(Bounded integrals are not uniform integrability)
Uniform integrability is not just boundedness of the integrals. On , let . Each and in measure, but the sets have measure tending to zero while . The family is not uniformly integrable, and does not tend to zero.
A finite-measure Vitali theorem
Theorem 1.5.7
(Vitali convergence theorem on a finite measure space)
Suppose , , the family is uniformly integrable, and in measure for a measurable finite almost everywhere. Then and
Proof
(Splitting the error at a small threshold)
The subsequence theorem supplies almost everywhere. Fatou's lemma gives
Thus is integrable. Its integral is absolutely continuous with respect to measure: given , choose by dominated convergence with . Then gives .
If , the conclusion is immediate. Otherwise fix . Uniform integrability and the corresponding small-set property for give a such that for every whenever . Choose so that . By convergence in measure, for all sufficiently large , the set has measure less than . Splitting the integral gives
This proves convergence in .
The proof has three hypotheses and uses each on a different piece of the integral.
Proof X-ray
Two controls for two parts of the integral
Convergence in measure makes the set where the error exceeds small. Uniform integrability makes the integral over that set small, uniformly in ; this is why the threshold must not depend on , since changes with . Fatou and the bound in the definition supply integrability of the limit, and with it the same small-set control for . On the complement the pointwise error is at most , and finite total measure turns this into the integral bound . Neither ingredient can replace the other in this split.
Finite total measure is a hypothesis used in the proof: the error bound on the complement is multiplied by . It cannot simply be dropped. On , let . Then uniformly and hence in measure, , and for every measurable . Thus the family satisfies both stated uniform-integrability conditions, yet does not converge to zero in . An infinite-measure extension needs additional control of mass escaping to large regions; one such condition requires for every a finite-measure set with .
Worked example 1.5.8
(No single dominating function is needed)
Choose pairwise disjoint intervals of lengths for . They fit because , for example by comparison with . Let . Then in measure and . The family is uniformly integrable: . Given , choose with . For each , ; for , , uniformly over those finitely many indices. Taking controls both groups.
Yet no integrable dominates every almost everywhere. After combining the countably many exceptional null sets, almost everywhere on every . Disjointness gives
Thus uniform integrability genuinely allows this -convergent family even though the standard dominated convergence hypothesis is unavailable.
Algebra of Convergence in Measure
Throughout, are measurable and finite almost everywhere. Their nonfinite-value sets form a countable union of measurable null sets. Replace all functions by zero on that common null set; the resulting finite measurable representatives have the same convergence in measure and make every sum and product below well defined. Put and .
Sums: no restriction on
Proposition 1.5.9
(Sums preserve convergence in measure)
If and in measure, then in measure.
Proof
(Halving the threshold for each summand)
Fix . If and , then . Hence
By subadditivity, the measure of the left side is at most , which tends to .
Products on a finite measure space
Proposition 1.5.10
(Products preserve convergence in measure when )
If, in addition, , then in measure.
Proof
(Cutting off the large values of the limits)
Expanding gives
Fix and .
Control the tails of and . The sets decrease, as , to , which is null. Because , continuity from above applies. So , and likewise for . Choose with both tails of measure less than .
Bound the three error terms. Off , each term is at most provided:
- and , for the cross terms and ;
- , for the product of errors .
The square-root threshold is the natural choice: if each factor is at most , their product is at most .
Combine. The set is contained in the two tails together with the four sets
Each of these four sets has measure tending to . By subadditivity,
Since is arbitrary, the limit is .
Intuition. A small error multiplied by a huge function can be large. Finite total measure guarantees that and are huge only on a set of small measure. On an infinite space, may have infinite measure for every . The argument still works whenever the tails do become small, for instance when and are bounded. So the product can converge on some infinite spaces; it is simply not guaranteed.
Counterexample on
The finite-measure hypothesis in the product proposition is used only to make the tails of and small, and an unbounded factor on the line defeats it.
Counterexample mode
A uniformly small factor times an unbounded one
Claim: on any measure space, and in measure imply in measure. On with Lebesgue measure let , , and . Since , uniformly, hence in measure, and exactly. On we have . So for every , and in measure. Here has infinite measure for every , which is exactly the step that fails; the repaired statement assumes , or more generally that the tails of and have measure tending to zero.
Solved practice
A finite-measure space can still carry an unbounded factor; the tails then become small and the product converges.
Worked example 1.5.11
(An unbounded factor on a finite interval)
On with Lebesgue measure, let and . The function is finite at every .
The product is , and the limit is . For :
So the exceptional set has measure exactly
Thus the product converges in measure. It does not converge uniformly, since for every .
Practice with solutions
Quick check
(On an arbitrary measure space, show that convergence implies convergence in measure.)
Bound the measure of the exceptional set by an integral.
Solution · Answer
Markov's inequality gives for every .
Quick check
(On , compute the integral of and decide whether in measure.)
For a fixed threshold, find the exceptional set once exceeds the threshold.
Solution · Answer
The integral is for every . For a fixed and , the exceptional set is , of measure , so in measure. Thus convergence in measure does not alone imply convergence of integrals.
Quick check
(Why does the dyadic indicator sequence in the worked example above have an almost-everywhere convergent subsequence although the whole sequence converges nowhere?)
Use the subsequence theorem, then name an explicit subsequence.
Solution · Answer
It converges in measure, so the subsequence theorem applies. Explicitly, with one-based indexing, stages contain intervals, so is the first interval of stage . These indicators tend to zero at every but equal one at for all ; the exceptional point is null.
Exercise 1.5.12
(Let , , and in measure. Prove that in .)
CoreProof
Solution(3)
- 1
Domination gives uniform integrability
, and , which is small uniformly in when is small.
Because: A family dominated by one integrable function satisfies both conditions of the definition.
- 2
Identify the limit
A subsequence almost everywhere, so almost everywhere and is finite almost everywhere and integrable.
Because: The subsequence theorem.
- 3
Apply the finite-measure Vitali theorem
∎All hypotheses hold, so .