Where Convergence Claims Go Wrong
Many false convergence arguments do not misuse a difficult theorem; they change the order of its quantifiers. In almost-everywhere convergence, the index after which an error is small may depend on the point. In convergence in measure, the error threshold is fixed before the sequence index tends to infinity. Neither dependence can be removed or rearranged without justification. The distinction also explains why almost-everywhere convergence implies convergence in measure on a finite measure space, but not in general on an infinite one.
Throughout, \((X,\mathcal{M},\mu)\) is a measure space, and \(f_n,f:X\to\mathbb{R}\) are measurable real-valued functions. For \(\varepsilon>0\), the error set is \(E_n(\varepsilon)=\{x\in X:|f_n(x)-f(x)|>\varepsilon\}\), which is measurable. The strict inequality matters: points where the error equals \(\varepsilon\) are not in this set.
In the first definition, the exceptional set \(N\) is chosen once, but the index may depend on both \(x\) and \(\varepsilon\). In the second, the threshold is fixed first; the measure of its error set is then sent to zero as \(n\) increases. These are different quantifier patterns. In particular, neither definition asserts that one index works for every point and every threshold.
One Null Set for All Thresholds
A frequent error is to reason separately for each \(\varepsilon\), obtain an exceptional null set \(N_\varepsilon\), and then assume without argument that the exceptional set can be chosen independently of \(\varepsilon\). The countability of the rational numbers supplies the missing step. The following criterion makes that step explicit.
Proof. Suppose first that \(f_n\to f\) \(\mu\)-almost everywhere, and let \(N\) be a null set outside which the pointwise convergence holds. Fix a positive rational \(q\). For every \(x\notin N\), convergence gives an index after which \(|f_n(x)-f(x)|<q\). Thus \(x\) belongs to only finitely many \(E_n(q)\), so \(\limsup_n E_n(q)\subseteq N\). Its measure is zero.
Conversely, suppose \(\mu(\limsup_n E_n(q))=0\) for every positive rational \(q\). The set \(N=\bigcup_{q\in\mathbb{Q},\,q>0}\limsup_n E_n(q)\) is measurable and null, because it is a countable union of measurable null sets. If \(x\notin N\), then for every positive rational \(q\), \(x\) belongs to only finitely many \(E_n(q)\). Given any \(\varepsilon>0\), choose a rational \(q\) with \(0<q<\varepsilon\). For all sufficiently large \(n\), \(x\notin E_n(q)\), so \(|f_n(x)-f(x)|\le q<\varepsilon\). This is pointwise convergence at every \(x\notin N\), with one null set valid for all thresholds. \(\square\)
The proof uses rational thresholds to obtain one exceptional null set, not because the original definition only tests rational errors. Once the result holds for every positive rational \(q\), choosing \(q<\varepsilon\) gives the conclusion for any real \(\varepsilon>0\).
Finite Measure and the Order of Limits
Proof. Fix \(\varepsilon>0\), and write \(E_n=E_n(\varepsilon)\). The set \(A_N=\bigcup_{n\ge N}E_n\) decreases as \(N\) increases, and \(\bigcap_{N=1}^\infty A_N=\limsup_n E_n\). Almost-everywhere convergence implies that this intersection is null: outside the exceptional null set, the error is eventually less than \(\varepsilon\). Since \(\mu(A_1)\le\mu(X)<\infty\), continuity from above for measures gives \[ \lim_{N\to\infty}\mu(A_N)=\mu\left(\bigcap_{N=1}^\infty A_N\right)=0. \] For every \(n\ge N\), \(E_n\subseteq A_N\), so \(0\le\mu(E_n)\le\mu(A_N)\). Given any \(\eta>0\), choose \(N\) with \(\mu(A_N)<\eta\); then \(\mu(E_n)<\eta\) for every \(n\ge N\). This proves \(\mu(E_n)\to0\). The finiteness of \(\mu(X)\) is what permits continuity from above here. \(\square\)
Worked Example: A Fixed Threshold on a Finite Interval
Take \(X=[0,1]\), \(\mathcal{M}\) to be the Lebesgue sigma-algebra, and \(\mu=m\), Lebesgue measure. Let \(f_n=\chi_{(0,1/n)}\) and \(f=0\). For each \(x\in[0,1]\), \(f_n(x)\to0\): at \(x=0\) all terms are zero, and at every \(x>0\), \(1/n\le x\) eventually, so \(x\notin(0,1/n)\).
Fix \(0<\varepsilon<1\). Since the error is either zero or one, the strict-threshold error set is exactly \(E_n(\varepsilon)=(0,1/n)\), and \(m(E_n(\varepsilon))=1/n\to0\). For \(\varepsilon\ge1\), the strict inequality \(|f_n-f|>\varepsilon\) never holds, so \(E_n(\varepsilon)=\varnothing\) and \(m(E_n(\varepsilon))=0\). Thus the sequence converges in measure, as the finite-measure theorem predicts. Check: at \(\varepsilon=1/2\), \(m(E_n(1/2))=1/n\); at \(\varepsilon=1\), the strict inequality excludes points with error exactly one, giving measure zero.
The same proof cannot simply be applied when \(\mu(X)=\infty\). A decreasing sequence of sets can have measures that stay infinite even when its intersection is empty. The failure is not just a gap in that proof: the implication itself can fail.
Worked Example: Almost-Everywhere Convergence Without Convergence in Measure
Let \(X=\mathbb{R}\), \(\mathcal{M}=\mathcal{L}\), and \(\mu=m\). Set \(f_n=\chi_{[n,n+1)}\) and \(f=0\). For each fixed \(x\in\mathbb{R}\), every integer \(n>x\) has \(x\notin[n,n+1)\), so \(f_n(x)=0\) eventually. Hence \(f_n\to0\) everywhere.
For the fixed threshold \(\varepsilon=1/2\), the strict error set is \(E_n(1/2)=[n,n+1)\), because the error equals one on that interval and zero elsewhere. Therefore \(m(E_n(1/2))=1\) for every \(n\), not a quantity tending to zero. The sequence does not converge to zero in measure. Check: each interval has length \((n+1)-n=1\), so its Lebesgue measure is exactly one. This shows why finiteness of the whole space cannot be omitted from the preceding implication.
Keep the Threshold Fixed While Taking the Limit
Convergence in measure says that for each fixed \(\varepsilon>0\), the measures \(\mu(E_n(\varepsilon))\) tend to zero. It does not say the same when the threshold is allowed to shrink with \(n\). The order is: choose \(\varepsilon\), then let \(n\to\infty\). Substituting a sequence \(\varepsilon_n\to0\) changes the claim.
Worked Example: A Moving Threshold Gives a Different Claim
On \(X=[0,1]\) with Lebesgue measure \(m\), let \(f_n(x)=1/n\) for all \(x\), and let \(f=0\). Fix any \(\varepsilon>0\). If \(n\ge 1/\varepsilon\), then \(1/n\le\varepsilon\), so the strict error set \(\{|f_n-f|>\varepsilon\}\) is empty. Thus \(f_n\to0\) in measure (in fact, uniformly).
Now choose the moving threshold \(\varepsilon_n=1/(2n)\). At every \(x\in[0,1]\), the error is \(1/n>1/(2n)\), so \(\{|f_n-f|>\varepsilon_n\}=[0,1]\) and its measure is one for every \(n\). Check: substituting the actual values gives \(1/n-1/(2n)=1/(2n)>0\), and \(m([0,1])=1\). There is no contradiction: convergence in measure never asserted convergence of these moving-threshold measures.
A related quantifier error is to turn the point-dependent index in almost-everywhere convergence into a single index that works for every point. That stronger conclusion is uniform convergence, and almost-everywhere convergence does not imply it. For example, on \([0,1]\) with Lebesgue measure, \(f_n(x)=x^n\) converges to zero at every \(x<1\) and to one at \(x=1\), hence almost everywhere to \(f=\chi_{\{1\}}\). But for every \(n\), the supremum of \(|f_n-f|\) on \([0,1]\) is one: values \(x^n\) approach one as \(x\) approaches one from below. Thus no common index makes the error uniformly small. The exceptional null set is fixed, but the convergence index need not be.
Convergence in Measure Does Not Supply Pointwise Convergence
The converse direction also fails: convergence in measure does not generally imply almost-everywhere convergence of the full sequence. Small error sets at each index may move around so that a point is hit infinitely often. The finite-space example below makes the quantifiers visible.
Worked Example: Small Error Sets That Visit Every Point
Let \(X=[0,1)\) with Lebesgue measure. For each integer \(k\ge1\), use a block of \(2^k\) functions, one for each \(j=0,\ldots,2^k-1\), given by \(f_{k,j}=\chi_{[j/2^k,(j+1)/2^k)}\). List the blocks in increasing order of \(k\) to form one sequence, and let \(f=0\). Every interval has measure \(2^{-k}\).
Fix \(0<\varepsilon<1\). For every function in block \(k\), its strict error set is its defining interval, since the error is one there and zero outside. Its measure is \(2^{-k}\). As the sequence index tends to infinity, its block number \(k\) tends to infinity, so these measures tend to zero. For \(\varepsilon\ge1\), every strict error set is empty. Hence \(f_n\to0\) in measure.
However, each \(x\in[0,1)\) lies in exactly one interval of every block, so the sequence takes the value one at \(x\) infinitely many times. For every \(k\ge2\), the block also has \(2^k-1\ge1\) other intervals, and their indicators take the value zero at \(x\). Thus the sequence takes both zero and one infinitely often at every \(x\); it does not converge anywhere. Check: each block covers \([0,1)\) by disjoint half-open intervals, so there is exactly one value-one term per block at each point, while the block lengths \(2^k\) ensure the individual error-set measures \(2^{-k}\) tend to zero.
A Practical Quantifier Check
Before accepting a convergence argument, write down what is fixed and what may depend on what. For almost-everywhere convergence, the null set is independent of the threshold, while the eventual index can depend on the point and the threshold. For convergence in measure, the threshold is fixed before taking the limit in \(n\). To pass from almost-everywhere convergence to convergence in measure, also check that the space has finite measure. To infer almost-everywhere convergence from convergence in measure, more is needed: the typewriter example shows that the full sequence may fail even pointwise convergence.
Common mistake and correction. The incorrect claim “for every \(\varepsilon>0\), the measures of the error sets tend to zero, so this also holds for \(\varepsilon_n=1/(2n)\)” substitutes a varying threshold for a fixed one. In the constant-function example, the fixed-threshold measures are eventually zero, whereas the moving-threshold measures are all one. The correction is to use the definition with each fixed \(\varepsilon\), or to invoke a separate theorem whose hypotheses justify a stronger conclusion.
Check Your Understanding
For each question, identify the quantifier or hypothesis that controls the conclusion.
- In the rational-threshold criterion, why is taking a countable union of null sets legitimate, and where is countability used?
- Which step in the proof that almost-everywhere convergence implies convergence in measure uses \(\mu(X)<\infty\)?
- For \(f_n=1/n\) on \([0,1]\), compute the strict error set at the fixed threshold \(\varepsilon=1/4\) for \(n=4\), and at the moving threshold \(\varepsilon_4=1/8\).
- Why do the translated indicators on \(\mathbb{R}\) converge pointwise but not in measure?
- In the block construction on \([0,1)\), what happens infinitely often at each point, and why does that not prevent convergence in measure?