Almost-Everywhere Convergence and Exceptional Sets

Convergence Outside a Null Set

Pointwise convergence requires a sequence to converge at every point of its domain. Almost-everywhere convergence relaxes this requirement: convergence may fail on a measurable set of measure zero. The exceptional set can be nonempty, and its points must still be examined; having measure zero means they can be disregarded for this mode of convergence, not that convergence holds there.

Let \((X,\mathcal{M},\mu)\) be a measure space, not necessarily finite or complete. Suppose \(f_n:X\to\mathbb{R}\) and \(f:X\to\mathbb{R}\) are measurable. All limits in this tutorial are ordinary finite real limits at individual points. Define the failure set by

D=\{x\in X:\ f_n(x)\text{ does not converge to }f(x)\}.

Definition: The sequence \((f_n)\) converges to \(f\) \(\mu\)-almost everywhere, written \(f_n\to f\) \(\mu\)-a.e., if there is a set \(N\in\mathcal{M}\) with \(\mu(N)=0\) such that \(f_n(x)\to f(x)\) for every \(x\in X\setminus N\). Such a set \(N\) is an exceptional set. The functions need not converge at points of \(N\).

The exceptional set need not be specified in advance. In fact, for measurable real-valued functions, the full failure set \(D\) is measurable. This gives a direct test for almost-everywhere convergence.

A Measurable Failure Set

Proposition: Let \((X,\mathcal{M},\mu)\) be any measure space, and let \(f_n,f:X\to\mathbb{R}\) be measurable. Then the set \(D\) where \(f_n\) fails to converge to \(f\) is measurable, and \(f_n\to f\) \(\mu\)-a.e. if and only if \(\mu(D)=0\).

Proof. For positive rational \(q\) and integer \(k\ge1\), define \(A_{q,k}=\bigcup_{n\ge k}\{x:|f_n(x)-f(x)|\ge q\}\). Each set inside this union is measurable because \(f_n-f\) is measurable and the absolute value of a measurable real-valued function is measurable. Thus \(A_{q,k}\) is measurable. We claim that

D=\bigcup_{q\in\mathbb{Q},\,q>0}\ \bigcap_{k=1}^{\infty} A_{q,k}.

If \(x\) belongs to the right-hand side, some \(q>0\) satisfies \(|f_n(x)-f(x)|\ge q\) for arbitrarily large \(n\): membership in every \(A_{q,k}\) means that for every \(k\), at least one \(n\ge k\) has this property. Hence \(f_n(x)\) cannot converge to \(f(x)\). Conversely, if \(f_n(x)\) does not converge to \(f(x)\), the definition of convergence gives an \(\varepsilon>0\) such that for every \(k\), there is an \(n\ge k\) with \(|f_n(x)-f(x)|\ge\varepsilon\). Choose a positive rational \(q\le\varepsilon\). Then \(x\in A_{q,k}\) for every \(k\), so \(x\) belongs to the right-hand side. This proves the set identity. Its right-hand side is a countable union of countable intersections of measurable sets, so \(D\in\mathcal{M}\). Finally, if \(\mu(D)=0\), take \(N=D\) in the definition. If convergence holds outside a measurable null set \(N\), then \(D\subseteq N\); monotonicity of the measure gives \(\mu(D)=0\). \(\square\)

This criterion uses only measurability and the measure axioms; the measure space need not be finite, and completeness is not required. In particular, the exceptional set can be taken to be the measurable failure set itself.

Worked Examples

Worked Example: A Nonempty Exceptional Set on the Real Line

On \((\mathbb{R},\mathcal{L},m)\), where \(m\) is Lebesgue measure, define \(f_n(x)=n\chi_{\{0\}}(x)\) and \(f(x)=0\). Determine where \(f_n\to f\) and whether the convergence is almost everywhere.

For \(x\ne0\), \(\chi_{\{0\}}(x)=0\), so \(f_n(x)=n\cdot0=0=f(x)\) for every \(n\). At \(x=0\), \(\chi_{\{0\}}(0)=1\), so \(f_n(0)=n\), which does not converge to the finite value \(f(0)=0\). Therefore the failure set is exactly \(D=\{0\}\). Each \(f_n\) is measurable because \(\{0\}\in\mathcal{L}\), and \(m(D)=m(\{0\})=0\). The sequence converges to \(f\) \(m\)-a.e., but it does not converge to \(f\) at every point.

Check: Substitution into the definition gives \(|f_n(0)-f(0)|=|n-0|=n\), which does not tend to zero. At every \(x\ne0\), \(|f_n(x)-f(x)|=|0-0|=0\) for all \(n\). These computations give precisely the stated exceptional set.

Worked Example: Pointwise Convergence on a Finite Measure Space

On \(([0,1],\mathcal{L}_{[0,1]},m)\), with Lebesgue measure restricted to \([0,1]\), set \(f_n(x)=x^n\). Find the pointwise limit and verify almost-everywhere convergence.

If \(x=0\), then \(x^n=0\) for every positive integer \(n\). If \(0<x<1\), powers of \(x\) tend to zero. At \(x=1\), \(x^n=1\) for every \(n\). Hence the pointwise limit is \(f=\chi_{\{1\}}\): it is zero on \([0,1)\) and one at \(1\). This sequence converges to \(f\) at every point, so its failure set is empty and the convergence is \(m\)-a.e. The space has finite measure because \(m([0,1])=1\).

Check: At \(x=\tfrac12\), substitution gives \(f_n(\tfrac12)=(\tfrac12)^n\to0=f(\tfrac12)\); at \(x=1\), substitution gives \(f_n(1)=1=f(1)\) for every \(n\). The endpoint values agree with the pointwise limit, including the point where the limiting function is nonzero.

Worked Example: Failure on a Positive-Measure Set

On the same finite measure space \(([0,1],\mathcal{L}_{[0,1]},m)\), define \(g_n(x)=(-1)^n\) for every \(x\in[0,1]\). Can this sequence converge almost everywhere to a measurable real-valued function \(g\)?

Fix any \(x\in[0,1]\). Along even indices \(n=2k\), \(g_{2k}(x)=(-1)^{2k}=1\). Along odd indices \(n=2k+1\), \(g_{2k+1}(x)=(-1)^{2k+1}=-1\). Since these subsequences have distinct limits, the sequence does not converge at \(x\). This reasoning applies to every \(x\), so the failure set is \(D=[0,1]\). Its measure is \(m(D)=m([0,1])=1\), not zero. The sequence therefore does not converge almost everywhere to any real-valued function.

Check: At any proposed pointwise limit value \(g(x)\), convergence would require both constant subsequences, with values \(1\) and \(-1\), to tend to \(g(x)\). A constant sequence has its constant value as its limit, so this would require \(g(x)=1\) and \(g(x)=-1\) simultaneously, which is impossible. Thus no points can be removed as a null exceptional set to obtain convergence.

Uniqueness of an Almost-Everywhere Limit

Theorem: Let \((X,\mathcal{M},\mu)\) be any measure space, and let \(f_n,f,g:X\to\mathbb{R}\) be measurable. If \(f_n\to f\) \(\mu\)-a.e. and \(f_n\to g\) \(\mu\)-a.e., then \(f=g\) \(\mu\)-a.e.

Proof. By the two almost-everywhere convergence assumptions, there are measurable sets \(N_f,N_g\in\mathcal{M}\) with \(\mu(N_f)=\mu(N_g)=0\), such that the respective convergences hold outside those sets. Their union \(N=N_f\cup N_g\) is measurable and has measure zero, since \(\mu(N)\le\mu(N_f)+\mu(N_g)=0\). For every \(x\in X\setminus N\), the real sequence \(f_n(x)\) converges both to \(f(x)\) and to \(g(x)\). Uniqueness of the limit of a real sequence gives \(f(x)=g(x)\). Thus \(\{x:f(x)\ne g(x)\}\subseteq N\), and this set has measure zero. Therefore \(f=g\) \(\mu\)-a.e. \(\square\)

The conclusion is equality almost everywhere, not necessarily equality at every point. For example, on \(([0,1],\mathcal{L}_{[0,1]},m)\), the constant sequence \(f_n=0\) converges everywhere to the zero function, and also converges almost everywhere to \(\chi_{\{1\}}\): the two proposed limits differ only at the null set \(\{1\}\). This is consistent with the theorem.

A Common Pitfall

A null exceptional set does not mean that the sequence converges at every point in that set. In the real-line example, it would be incorrect to write \(n\chi_{\{0\}}(x)\to0\) pointwise “because \(\{0\}\) has measure zero.” At \(x=0\), direct substitution gives \(n\chi_{\{0\}}(0)=n\), which does not converge to zero. The correct conclusion is that convergence holds for every \(x\notin\{0\}\), and the set where it fails has Lebesgue measure zero.

It is also important to distinguish a null set from a set that is merely small or finite. The definition requires an exceptional set of measure exactly zero. In the alternating-sign example, failure occurs on all of \([0,1]\), whose measure is one, so no choice of a null exceptional set can make the sequence converge outside it.

Almost-everywhere convergence records what happens after excluding a negligible set, while retaining the pointwise convergence requirement at every remaining point. The measurable failure-set criterion makes this distinction precise, and uniqueness ensures that any two almost-everywhere limits agree except possibly on a null set.

Check Your Understanding

For each question, identify the pointwise behavior first, then determine the relevant failure set and its measure.

On \((\mathbb{R},\mathcal{L},m)\), let \(h_n(x)=\chi_{\{1/n\}}(x)\). For which \(x\) does \(h_n(x)\) fail to converge to zero?
On \(([0,1],\mathcal{L}_{[0,1]},m)\), define \(u_n(x)=x^{2n}\). Find the pointwise limit, including its value at both endpoints, and state whether convergence is everywhere or only almost everywhere.
Explain why the failure set in the measurable-failure-set proposition is measurable, indicating where countability is used.
If a sequence of measurable real-valued functions converges almost everywhere to both \(f\) and \(g\), what can be concluded about the set where \(f\ne g\)?
On a finite measure space, can a sequence that fails to converge at every point of a set of positive measure converge almost everywhere? Justify your answer from the definition.