Four Questions to Ask Before Accepting an Extension Argument
Measure extension arguments can fail even when the final claim looks plausible. A cover may overlap; a proposed sum may use values of a premeasure outside its domain; a uniqueness claim may omit sigma-finiteness; or an extension may be mistaken for a completion. These are different errors, and each calls for a different repair.
Recall the setting. A ring \(\mathcal{R}\) of subsets of \(X\) is closed under finite unions and differences. A premeasure \(\rho:\mathcal{R}\to[0,\infty]\) satisfies \(\rho(\varnothing)=0\) and countable additivity whenever a disjoint sequence \((A_n)\) in \(\mathcal{R}\) has union in \(\mathcal{R}\). Its domain is \(\mathcal{R}\): writing \(\rho(A)\) is meaningful only when \(A\in\mathcal{R}\). An extension is a measure on a larger sigma-algebra that agrees with \(\rho\) on \(\mathcal{R}\). Sigma-finiteness means that \(X\) is covered by countably many members of \(\mathcal{R}\) having finite \(\rho\)-measure.
The Caratheodory-Hahn extension theorem gives an extension to \(\sigma(\mathcal{R})\). Its uniqueness conclusion requires sigma-finiteness of the premeasure. The extension can be constructed by first taking an infimum of costs of countable covers from \(\mathcal{R}\). In that construction, cover costs are sums of premeasure values of the covering sets; they are not automatically the measure of their union.
Failure 1: Treating an Overlapping Cover as Disjoint
Countable additivity applies to disjoint sets. If covering sets overlap, one cannot claim that the sum of their premeasure values equals the premeasure of their union. The sum is still a valid cover cost when every covering set belongs to \(\mathcal{R}\), but equality with the measure of the union needs disjointness.
Worked Example: An Overlapping Cover Counts a Point Twice
Let \(X=\{1,2\}\), let \(\mathcal{R}=\mathcal{P}(X)\), and let \(\rho(A)\) be the number of elements of \(A\). Cover \(E=\{1,2\}\) with \(A_1=\{1,2\}\) and \(A_2=\{2\}\). Both sets belong to \(\mathcal{R}\), and \(A_1\cup A_2=E\), but
The excess is the repeated count of point \(2\). In a covering-infimum construction, this cover has cost \(3\); it does not establish that \(\rho(E)=3\). Indeed, the single-set cover \(E=A_1\) has cost \(2\), so the infimum of these cover costs is at most \(2\). Check: the single-set cover costs \(\rho(E)=2\), while the overlapping cover costs \(2+1=3\). The numerical discrepancy comes exactly from the overlap.
There is a safe way to recover an inequality when the union itself belongs to the ring. Disjointify the cover in its given order. The resulting pieces are contained in the original covering sets, so their measures cannot be larger. This gives subadditivity, not an equality for the original overlapping sets.
Proof. Define \(D_1=A_1\) and \(D_n=A_n\setminus\bigcup_{k=1}^{n-1}A_k\) for \(n\ge2\). Since a ring is closed under finite unions and differences, \(D_n\in\mathcal{R}\). The sets \(D_n\) are pairwise disjoint and have union \(E\). Countable additivity therefore gives \(\rho(E)=\sum_n\rho(D_n)\). For each \(n\), \(D_n\subseteq A_n\), and monotonicity of a premeasure on a ring gives \(\rho(D_n)\leq\rho(A_n)\): write \(A_n=D_n\cup(A_n\setminus D_n)\) as a disjoint union in the ring and use nonnegativity. Summing these inequalities proves the claim, including when the right-hand side is infinite. \(\square\)
Notice the hypothesis \(E\in\mathcal{R}\). If the union is outside the premeasure’s domain, \(\rho(E)\) is not defined, so the proposition cannot be applied with that \(E\).
Failure 2: Using a Premeasure Outside Its Domain
A premeasure is not automatically defined on every subset of \(X\). The covering construction avoids this problem by evaluating \(\rho\) only on covering sets chosen from \(\mathcal{R}\). A proposed expression involving \(\rho(A)\) for \(A\notin\mathcal{R}\) is not an estimate that needs a proof; it is undefined.
Worked Example: Covering the Naturals Without Measuring Them in the Ring
Let \(X=\mathbb{N}=\{1,2,\ldots\}\), let \(\mathcal{R}\) be the ring of finite subsets of \(\mathbb{N}\), and set \(\rho(A)=\#A\) for \(A\in\mathcal{R}\). This is a premeasure: if disjoint finite sets have a finite union, the number of elements in the union is the sum of their numbers of elements. It is sigma-finite because \(X=\bigcup_{n\ge1}\{n\}\) and \(\rho(\{n\})=1<\infty\).
Each singleton is an allowed covering set, so the cover \(\mathbb{N}=\bigcup_{n\ge1}\{n\}\) has cost \(\sum_{n\ge1}\rho(\{n\})=\sum_{n\ge1}1=\infty\). But \(\mathbb{N}\notin\mathcal{R}\), so \(\rho(\mathbb{N})\) is undefined. The valid statement is about the cost of this cover, not about evaluating the premeasure on its union. Check: each of the first \(N\) singletons contributes \(1\), so the partial cost is \(N\); as \(N\to\infty\), the cover’s total cost is \(+\infty\).
The same domain check applies to every covering family: the sets used in its cost must belong to \(\mathcal{R}\). If a subset cannot be covered by members of \(\mathcal{R}\), the infimum over its covers is conventionally \(+\infty\); this still does not define a value of \(\rho\) on that subset.
Failure 3: Claiming Uniqueness Without Sigma-Finiteness
Sigma-finiteness is not a technical decoration in the uniqueness theorem. If it is absent, extensions can disagree, even when they agree on every set in the original ring. Here is a complete example.
Worked Example: Two Extensions of the Zero Premeasure
Let \(I\) be an uncountable set, and let \(\mathcal{R}\) be the ring of finite subsets of \(I\). Define \(\rho(A)=0\) for every \(A\in\mathcal{R}\). This is a premeasure: for any disjoint sequence in \(\mathcal{R}\) whose union belongs to \(\mathcal{R}\), both the measure of the union and the sum of the measures are \(0\).
The sigma-algebra \(\sigma(\mathcal{R})\) consists of the countable and the co-countable subsets of \(I\). Indeed, countable unions of finite sets are countable, complements of countable sets are co-countable, and every countable set is a countable union of singletons.
Define \(\mu_0(A)=0\) for every \(A\in\sigma(\mathcal{R})\). Also define \(\mu_\infty(A)=0\) if \(A\) is countable and \(\mu_\infty(A)=\infty\) if \(A\) is co-countable. Both are measures. For \(\mu_\infty\), consider a disjoint sequence whose union is countable: every member is then countable, so the sum and measure of the union are both \(0\). If its union is co-countable, at least one member must be co-countable, since a countable union of countable sets is countable. At most one disjoint member can be co-countable, and the sum of the measures and the measure of the union are both \(\infty\). Thus countable additivity holds in both cases.
Every finite set has measure \(0\) under both extensions, so both extend \(\rho\). Yet \(\mu_0(I)=0\) and \(\mu_\infty(I)=\infty\). The premeasure is not sigma-finite: every member of \(\mathcal{R}\) is finite and has measure \(0\), and a countable union of finite sets is countable, so such sets cannot cover uncountable \(I\). Check: the two values on \(I\) differ, while their values on every finite set agree with \(\rho\).
Failure 4: Calling Every Extension a Completion
An extension is any measure on a larger sigma-algebra that agrees with the original measure on the original sigma-algebra. A completion is a particular extension: it adds every subset of every measurable null set and assigns those subsets measure zero. Thus completion adds sets that were hidden inside null sets; it does not describe every possible enlargement of a measure space.
Worked Example: Two Extensions, Only One Kind of Enlargement
Let \(X=\{a,b,c,d\}\) and \(\mathcal{M}=\{\varnothing,\{a,b\},\{c,d\},X\}\). Define a measure \(\mu\) on \(\mathcal{M}\) by \(\mu(\{a,b\})=0\), \(\mu(\{c,d\})=1\), \(\mu(X)=1\), and \(\mu(\varnothing)=0\). The completion adds \(\{a\}\) and \(\{b\}\), and their measures are \(0\), because they are subsets of the measurable null set \(\{a,b\}\).
Now enlarge the sigma-algebra instead to \(\mathcal{P}(X)\). Define \(\nu_c(A)=1\) if \(c\in A\), and \(0\) otherwise; define \(\nu_d(A)=1\) if \(d\in A\), and \(0\) otherwise. These are Dirac measures. Both extend \(\mu\): each assigns \(0\) to \(\{a,b\}\), \(1\) to \(\{c,d\}\), \(1\) to \(X\), and \(0\) to \(\varnothing\). However, \(\nu_c(\{c\})=1\) while \(\nu_d(\{c\})=0\), and \(\{c\}\) is not added by completion because it is not a subset of a measurable null set. Check: the original measurable atom \(\{c,d\}\) has measure \(1\) under both extensions, while the new set \(\{c\}\) has different values. This is an extension beyond completion, not a completion itself.
A Practical Diagnostic
Before accepting a measure extension calculation, check these points in order:
- Domain: Is every set evaluated by \(\rho\) in \(\mathcal{R}\)?
- Disjointness: Is countable additivity being used only for a disjoint family? If not, can the sets be disjointified, and does the union belong to the ring?
- Hypotheses for uniqueness: Has sigma-finiteness actually been established by a countable cover with finite premeasure?
- Type of enlargement: Is the claim about any extension, or specifically about adding subsets of measurable null sets in the completion?
These checks keep separate the existence of a cover, the value of its cost, a premeasure’s domain, and the hypotheses needed to compare extensions. In particular, the fact that a construction produces an extension does not show that it is the only extension, and the fact that a larger sigma-algebra is used does not make the enlargement a completion.
Check Your Understanding
For each question, identify the precise hypothesis or distinction that controls the conclusion.
- For a ring premeasure, explain why \(\rho(A\cup B)=\rho(A)+\rho(B)\) cannot be invoked when \(A\cap B\ne\varnothing\). What disjointification gives a valid inequality when \(A\cup B\) belongs to the ring?
- In the finite-subset ring on \(\mathbb{N}\), why is \(\rho(\mathbb{N})\) undefined even though the cover by singletons has a well-defined cost?
- For the zero premeasure on finite subsets of an uncountable set \(I\), verify why no countable family of ring sets covers \(I\).
- In the two extensions of the zero premeasure, which measurable set has different extension values, and why does that rule out uniqueness?
- In the four-point example, why is \(\{c\}\) not supplied by completion, although it belongs to the sigma-algebra of a larger extension?