Why Localize a Measure-Theoretic Argument?
Some arguments are simplest on a set of finite measure: estimates involving the measure of an exceptional set, or theorems whose hypotheses include finite total measure, can be applied there directly. On a sigma-finite space, one can often prove the desired statement on an increasing sequence of finite-measure pieces and then assemble the conclusions. The assembly step is not automatic: it relies on countability, and the local estimates must fit together in a way that controls the whole space.
Throughout, \((X,\mathcal{M},\mu)\) is a measure space unless a particular space is specified. No completeness is assumed. Functions are real-valued or nonnegative extended real-valued as stated. An assertion holding \(\mu\)-almost everywhere means that its exceptional set is contained in a measurable set of \(\mu\)-measure zero.
The finite-measure condition follows from finite subadditivity: \(\mu(E_n)\le\sum_{j=1}^{n}\mu(C_j)<\infty\). If \(X\) itself has finite measure, the constant sequence \(E_n=X\) is an exhaustion. If \(X\) is empty, the same sequence works and all statements below are vacuous. In general, an exhaustion lets us take a global statement and first inspect it on each \(E_n\).
Two Local-to-Global Principles
Proof. Since \(E_n\) increases to \(X\), the sets \(A\cap E_n\) increase to \(A\). Continuity from below for measures gives \(\mu(A)=\lim_n\mu(A\cap E_n)=0\). For the more general claim, let \(N_n\subseteq E_n\) be a measurable null set outside which the assertion holds on \(E_n\). The union \(N=\bigcup_nN_n\) is measurable and null by countable subadditivity. If \(x\in X\setminus N\), then \(x\in E_k\) for some \(k\), since the exhaustion covers \(X\). Also \(x\notin N_k\), so the assertion holds at \(x\). Thus it fails only on \(N\), a null set. \(\square\)
For statements about functions, the same reasoning applies to a measurable set of failures. For example, if real-valued measurable functions \(f\) and \(g\) satisfy \(f\le g\) almost everywhere on each \(E_n\), then \(f\le g\) almost everywhere on \(X\). The inequalities need not hold at every point; the countable union of the local exceptional sets accounts for all failures.
Worked Example: Combining Local Null-Set Information
Take \(X=\mathbb{R}\), \(\mathcal{M}=\mathcal{L}\), and \(\mu=m\), Lebesgue measure. Let \(E_n=[-n,n]\), so \(m(E_n)=2n<\infty\) and \(\bigcup_nE_n=\mathbb{R}\). Let \(A=\mathbb{Q}\). For each \(n\), the set \(A\cap E_n\) is countable. Every singleton has Lebesgue measure zero, so countable subadditivity gives \(m(A\cap E_n)=0\). The proposition therefore gives \(m(\mathbb{Q})=0\).
Check: Directly, \(\mathbb{Q}=\bigcup_{q\in\mathbb{Q}}\{q\}\) is a countable union of singleton null sets, so it has measure zero as well. The local argument uses the finite-measure exhaustion; the direct argument confirms the same result.
Integral bounds can also pass from the pieces to the whole space, provided the bounds are uniform in \(n\). The required limit step is the Monotone Convergence Theorem: for an increasing sequence of nonnegative measurable functions, its integral is the limit of their integrals. This theorem applies on any measure space; finiteness of the whole space is not required.
Proof. Set \(h_n=h\chi_{E_n}\). Each \(h_n\) is measurable, \(0\le h_n\le h_{n+1}\), and \(h_n(x)\) increases to \(h(x)\) for every \(x\), because every \(x\) belongs to some \(E_n\). By the Monotone Convergence Theorem, \(\int_X h\,d\mu=\lim_n\int_X h_n\,d\mu\). Since \(\int_Xh_n\,d\mu=\int_{E_n}h\,d\mu\le C\) for every \(n\), the limit is at most \(C\). \(\square\)
Worked Applications of Localization
Worked Example: Integrating an Exponential on the Real Line
On \((\mathbb{R},\mathcal{L},m)\), let \(h(x)=e^{-|x|}\), a nonnegative measurable function. Use \(E_n=[-n,n]\). Each piece has finite measure \(2n\), and the sets increase to \(\mathbb{R}\). Symmetry and the antiderivative \(-e^{-x}\) on \([0,n]\) give
The functions \(h\chi_{E_n}\) are nonnegative, measurable, and increase pointwise to \(h\). Monotone Convergence therefore gives
Check: The integral on each half-line is \(\int_0^\infty e^{-x}\,dx=[-e^{-x}]_0^\infty=1\), so adding the two half-lines gives \(1+1=2\), in agreement with the exhaustion calculation.
A common finite-measure estimate controls the size of a level set by an integral. Its local form is useful even when the full space has infinite measure.
Proof. Put \(B=E\cap\{|f|>t\}\), which is measurable. On \(B\), \(t\chi_B\le |f|\chi_E\), and this inequality also holds outside \(B\) because its left side is zero. Monotonicity of the nonnegative Lebesgue integral yields \(t\mu(B)=\int t\chi_B\,d\mu\le\int_E|f|\,d\mu\). Dividing by \(t>0\) proves the estimate. \(\square\)
Worked Example: Applying a Level-Set Estimate Piece by Piece
Let \(X=\mathbb{N}=\{1,2,\ldots\}\) with its power-set sigma-algebra and counting measure \(\mu\). This space is sigma-finite: \(E_n=\{1,\ldots,n\}\) has measure \(n\), and the sets exhaust \(X\). Define the real-valued function \(f(k)=2^{-k}\), and take the strict level set \(B=\{k:f(k)>1/8\}\). Since \(2^{-1}=1/2>1/8\), \(2^{-2}=1/4>1/8\), and \(2^{-3}=1/8\) is not strictly greater than \(1/8\), we have \(B=\{1,2\}\) and \(\mu(B)=2\).
For \(n\ge1\), the local integral is the finite geometric sum
The level-set estimate with \(t=1/8\) gives \(\mu(B\cap E_n)\le8(1-2^{-n})\). For \(n=1\), the left side is \(1\) and the right side is \(8(1-1/2)=4\). For \(n\ge2\), the left side is \(2\), while \(8(1-2^{-n})\ge8(1-1/4)=6\). Thus the estimate holds on every piece. As \(n\) increases, \(B\cap E_n\) increases to \(B\), and the bounds tend to \(8\), giving \(\mu(B)\le8\). Directly counting the two elements of \(B\) gives \(\mu(B)=2\), which verifies the result and shows that the estimate need not be sharp.
What the Local Bounds Do—and Do Not—Give
The uniformity of a bound matters. In the integral proposition, one fixed finite \(C\) bounds \(\int_{E_n}h\,d\mu\) for every \(n\). Bounds that grow with \(n\) can show that each restriction is integrable without showing that \(h\) is integrable on the whole space. For example, on \((\mathbb{R},\mathcal{L},m)\), take \(h(x)=1\) and \(E_n=[-n,n]\). Then \(\int_{E_n}h\,dm=2n<\infty\) for each \(n\), but these values have no finite uniform bound. Indeed, \(\int_{\mathbb{R}}1\,dm=\infty\).
Common mistake: “Each \(E_n\) has finite measure and \(h\) is integrable on each \(E_n\), so \(h\) is integrable on \(X\).” The example \(h=1\) on \(\mathbb{R}\) disproves this: all the local integrals \(2n\) are finite, but their limit is infinite. The correction is to retain the actual local values and pass to the limit; for a finite global integral bound, the local bounds must remain bounded by one common finite constant.
There is a parallel caution for almost-everywhere arguments. A countable collection of local exceptional null sets still has null union, which is why a sigma-finite exhaustion is effective. An uncountable union of null sets need not be null. Thus the proof must use the countable exhaustion, not merely a claim that every point lies in some finite-measure set.
Check Your Understanding
For each question, identify the exhaustion and the precise step that passes from local information to a statement on the full space.
- Given a sigma-finite measure space and a finite-measure cover \(C_1,C_2,\ldots\), define an increasing finite-measure exhaustion.
- If \(A\) is measurable and \(\mu(A\cap E_n)=0\) for every set in an exhaustion, why does \(\mu(A)=0\)?
- What uniformity is needed in the local bounds \(\int_{E_n}h\,d\mu\le C_n\) to conclude that \(\int_Xh\,d\mu<\infty\) by the integral proposition?
- For counting measure on \(\mathbb{N}\), compute \(\mu(\{k:2^{-k}>1/4\})\), taking care with the strict inequality.
- Why does integrability of \(h=1\) on every interval \([-n,n]\) not imply integrability on \(\mathbb{R}\)?