Five Checks Before Trusting an Approximation
Simple-function approximations are useful only when their basic properties have actually been verified. A sequence can consist of measurable functions but fail to have the claimed range; a formula can take countably many values without being a simple function; and a sequence can converge almost everywhere without converging at every point. These are not technical distinctions: they determine whether a definition applies and what conclusion an approximation justifies.
Throughout, \((X,\mathcal{M},\mu)\) is an arbitrary measure space; no finiteness, completeness, or \(\sigma\)-finiteness is assumed unless stated. A real-valued function \(s\) is a simple function here if it is measurable and takes only finitely many values. Its level sets are the inverse images \(s^{-1}(\{a\})=\{x:s(x)=a\}\). A function is countably valued if its range is finite or countably infinite. The terms “pointwise” and “almost everywhere” will be kept distinct: pointwise convergence requires convergence at every \(x\in X\), whereas convergence \(\mu\)-a.e. permits an exceptional measurable set of \(\mu\)-measure zero.
Countably Many Values Do Not Mean Simple
A common error is to call any function whose range is countable a simple function. Under the finite-valued definition, that is false. Countable range is not itself an obstacle to measurability, however: the decisive information is whether the level sets are measurable.
Proof. Suppose first that \(f\) is measurable. For each \(a\in R\), the singleton \(\{a\}\) is a Borel subset of \(\mathbb{R}\), so \(\{x:f(x)=a\}=f^{-1}(\{a\})\in\mathcal{M}\). Conversely, suppose every such level set is in \(\mathcal{M}\). For any open set \(U\subseteq\mathbb{R}\),
The union is countable because \(R\) is countable, and each set in the union belongs to \(\mathcal{M}\). Thus \(f^{-1}(U)\in\mathcal{M}\) for every open \(U\), which is the definition of measurability for a real-valued function. The argument includes the empty range when \(X=\varnothing\): then every inverse image is empty and measurable. \(\square\)
Worked Example: A Measurable Function That Is Not Simple
On \((\mathbb{R},\mathcal{B}_{\mathbb{R}})\), define \(f(x)=1/n\) when \(x\in[n,n+1)\) for an integer \(n\ge1\), and \(f(x)=0\) otherwise. Check its range, level sets, and measurability.
The range is \(\{0\}\cup\{1/n:n\ge1\}\), which is countable and infinite: the values \(1/n\) are distinct for distinct positive integers \(n\). The level set for \(1/n\) is exactly \([n,n+1)\), a Borel set, and the level set for zero is \((-\infty,1)\). Indeed, the intervals \([n,n+1)\), \(n\ge1\), have union \([1,\infty)\), so their complement is \((-\infty,1)\). Every level set is Borel; the proposition therefore gives measurability. Because there are infinitely many distinct values, \(f\) is not simple under the finite-valued definition. Check: \(f(1)=1\), \(f(2)=1/2\), and \(f(0)=0\), consistent with the specified level sets.
The reverse mistake is to assume that countable range guarantees measurability. On a measure space \((X,\mathcal{M},\mu)\), if \(A\subseteq X\) is not in \(\mathcal{M}\), then \(\chi_A\) takes only the two values \(0\) and \(1\), but is not measurable: its level set \(\{x:\chi_A(x)=1\}\) is \(A\). Thus finite range alone does not suffice either.
Worked Example: A Two-Valued Formula with a Nonmeasurable Level Set
Let \(V\subseteq[0,1]\) be a non-Lebesgue-measurable set, and take \((X,\mathcal{M},\mu)=(\mathbb{R},\mathcal{L},m)\). Define \(f=\chi_V\). Determine whether \(f\) is measurable.
The range is contained in \(\{0,1\}\), and \(\{x:f(x)=1\}=V\). If \(f\) were \(\mathcal{L}\)-measurable, the inverse image of the Borel set \(\{1\}\) would be Lebesgue measurable, forcing \(V\in\mathcal{L}\), contrary to the choice of \(V\). Hence \(f\) is not measurable. This argument uses only the assumed nonmeasurability of \(V\); it does not depend on whether the underlying measure space is complete. Check: the other level set is \(V^c\), and \(V^c\notin\mathcal{L}\) as well, since measurability of \(V^c\) would imply measurability of its complement \(V\).
Overlapping Indicators Must Be Added
A representation such as \(s=\sum_{j=1}^k a_j\chi_{E_j}\) is a pointwise sum. If the sets \(E_j\) overlap, all terms active at a point contribute to the value there. Treating the sets as if they formed a partition can give a false description of the function’s level sets. Disjoint representations avoid that ambiguity, but a sum of indicators is still perfectly well-defined when the sets overlap.
Worked Example: Reading an Overlapping Representation
Let \(X=\{a,b,c\}\), let \(\mathcal{M}\) be all subsets of \(X\), and let \(\mu\) be counting measure. Set \(A=\{a,b\}\), \(B=\{b,c\}\), and \(s=2\chi_A+3\chi_B\). Find the values, level sets, and integral.
At \(a\), \(\chi_A(a)=1\) and \(\chi_B(a)=0\), so \(s(a)=2(1)+3(0)=2\). At \(b\), both indicators equal \(1\), so \(s(b)=2(1)+3(1)=5\). At \(c\), the indicators are \(0\) and \(1\), so \(s(c)=2(0)+3(1)=3\). Consequently, the nonempty level sets are \(\{s=2\}=\{a\}\), \(\{s=5\}=\{b\}\), and \(\{s=3\}=\{c\}\). The integral with respect to counting measure is
As an independent check, add the integrals of the two indicator terms: \(2\mu(A)+3\mu(B)=2(2)+3(2)=10\). The pointwise value at \(b\) is \(5\), not either \(2\) or \(3\); that is exactly the overlap contribution. The calculation also confirms the integral because each of the three points contributes \(2\), \(5\), and \(3\), respectively.
Check the Direction of Monotonicity
Refining a grid does not make every rounded approximation increase. The direction depends on whether values are rounded down or up. For example, for \(x\in[0,1]\), define the upper dyadic rounding \(u_n(x)=2^{-n}\lceil 2^n x\rceil\), where \(\lceil y\rceil\) is the least integer greater than or equal to \(y\). This is an approximation from above, and its monotonic direction is nonincreasing, not nondecreasing.
Proof. Put \(k=\lceil 2^n x\rceil\). Since \(k\ge2^n x\), we have \(2k\ge2^{n+1}x\); because \(2k\) is an integer, the least integer at least \(2^{n+1}x\) is no greater than \(2k\). Therefore \(\lceil2^{n+1}x\rceil\le2\lceil2^nx\rceil\), and division by \(2^{n+1}\) gives \(u_{n+1}(x)\le u_n(x)\). Also \(x\le u_n(x)\), and the ceiling property gives \(u_n(x)<x+2^{-n}\). Since \(2^{-n}\to0\), these inequalities imply \(u_n(x)\to x\). For fixed \(n\), the integer \(\lceil2^nx\rceil\) takes only the finitely many values \(0,1,\ldots,2^n\) on \([0,1]\). Its level sets are intervals, with endpoints included or excluded according to the ceiling rule, and hence are Lebesgue measurable. Thus \(u_n\) is measurable and finite-valued. \(\square\)
Worked Example: Upper Rounding Is Not Increasing
Take \(x=1/3\). Compute the first three upper dyadic roundings and use them to test the claim that the sequence is increasing.
At \(n=1\), \(2^n x=2/3\), so \(\lceil2/3\rceil=1\) and \(u_1(1/3)=1/2\). At \(n=2\), \(2^n x=4/3\), so \(\lceil4/3\rceil=2\) and \(u_2(1/3)=2/4=1/2\). At \(n=3\), \(2^n x=8/3\), so \(\lceil8/3\rceil=3\) and \(u_3(1/3)=3/8\). Thus \(1/2,1/2,3/8\) is not an increasing sequence, because \(3/8<1/2\). It is nonincreasing, as the proposition predicts. Check: the three values are all at least \(1/3\), and their differences from \(1/3\) are \(1/6\), \(1/6\), and \(1/24\); the computed sequence approaches \(1/3\) from above.
Pointwise Means Every Point
Convergence away from a null set is not pointwise convergence on the whole space. This distinction matters even for elementary indicator functions. Conversely, convergence at every point is stronger than convergence almost everywhere, so a statement that establishes only the latter must not be upgraded without checking the exceptional points.
Worked Example: Almost-Everywhere Convergence with One Exception
On \(([0,1],\mathcal{L},m)\), define \(f_n=\chi_{[0,\,1-1/n]}\) for integers \(n\ge2\). Determine the pointwise limit and decide whether \(f_n\to1\) pointwise or almost everywhere.
Fix \(x\in[0,1)\). Since \(1-x>0\), there is an integer \(n\ge2\) with \(1/n\le1-x\). For every such \(n\), \(x\le1-1/n\), so \(f_n(x)=1\); the intervals are nested, so this value remains \(1\) thereafter. At \(x=1\), the condition \(1\le1-1/n\) is false for every \(n\ge2\), and \(f_n(1)=0\) for every \(n\). Hence the pointwise limit is \(\chi_{[0,1)}\), not the constant function \(1\). Each \(f_n\) is a measurable simple function. The exceptional set for convergence to \(1\) is exactly \(\{1\}\), which has Lebesgue measure zero, so \(f_n\to1\) \(m\)-a.e. but not pointwise on all of \([0,1]\). Check: \(m([0,1])=1\), \(m(\{1\})=0\), and \(\int f_n\,dm=1-1/n\), which tends to \(1=\int_0^1 1\,dm\); these integral values do not remove the failure at \(x=1\).
A Practical Diagnostic
When checking a proposed simple-function approximation on a measure space \((X,\mathcal{M},\mu)\), test the claims separately. A finite-valued formula is not measurable until its level sets are known to be in \(\mathcal{M}\). A countably valued measurable function need not be simple. A sum of indicators counts every overlapping term. A monotonicity claim requires comparing consecutive functions at the same point, and a convergence claim must specify whether it is pointwise or only \(\mu\)-a.e.
One particularly misleading shortcut is to infer monotonicity just from the phrase “finer partition.” The dyadic upper-rounding example shows why that fails: the grid becomes finer, but the rounded values move downward. Another is to infer pointwise convergence from an almost-everywhere calculation while ignoring boundary points. Writing out the level sets, the values at intersections, and the exceptional points is usually enough to expose these errors before they enter a proof.
Check Your Understanding
For each question, identify the exact property being tested and distinguish what is known from what still needs verification.
- Can a measurable function on a measurable space have countably infinitely many values without being simple? Give a concrete example and identify its level sets.
- If \(f\) takes values in \(\{0,1\}\), what condition on its level sets is necessary and sufficient for \(f\) to be measurable?
- For \(s=4\chi_A+\chi_B\), where \(A\) and \(B\) overlap, what is the value of \(s\) at a point in \(A\cap B\)? Why is it incorrect to assign that point only one of the two coefficients?
- For the upper dyadic rounding \(u_n(x)=2^{-n}\lceil2^nx\rceil\), is the sequence nondecreasing, nonincreasing, or neither? State the inequalities that establish the answer.
- For \(f_n=\chi_{[0,\,1-1/n]}\) on \([0,1]\), \(n\ge2\), at which point does convergence to the constant function \(1\) fail, and why can the sequence still converge to \(1\) almost everywhere?