Measure and integration are where real analysis becomes a coherent system rather than a bag of clever \epsilon tricks. The definitions are chosen so that the theorems you want to be true actually become true, while the counterexamples tell you what cannot be demanded.
This article is a guided tour through measure and integration using worked examples. The theme is not to cover everything, but to show how the central moves repeat: build measurable sets, approximate complicated objects by simple ones, and use convergence theorems that have precise hypotheses.
Worked example: indicator functions and the meaning of an integral
For a measurable set E in [0,1], the indicator function 1_E is defined by 1_E(x)=1 if x∈E and 1_E(x)=0 otherwise.
A core fact is:
∫_0^1 1_E(x) dx = μ(E),
where μ is Lebesgue measure, which agrees with length on intervals.
This single identity is the bridge between sets and functions. Many problems about integrals become problems about the sizes of level sets.
Worked example: building measurable sets from open intervals
A standard measurable collection on ℝ is the Borel sets: start with open intervals and close under countable unions, countable intersections, and complements.
Example pattern: if E_n are open sets, then ⋃_n E_n is open, and its complement is closed, hence Borel. If F_n are closed sets, then ⋂_n F_n is closed, hence Borel. By iterating these closures you build extremely complicated sets while keeping measurability automatic.
This is a key habit: you do not prove measurability by explicit formulas; you prove it by closure properties.
Worked example: simple functions approximate measurable functions
A simple function is a finite linear combination of indicators:
s(x)=∑_{j=1}^m a_j 1_{E_j}(x),
where E_j are measurable and a_j are real numbers.
The Lebesgue integral is built by first defining the integral of nonnegative simple functions as
∫ s dμ = ∑_{j=1}^m a_j μ(E_j),
then defining the integral of a nonnegative measurable f as the supremum of ∫ s over simple s with 0 ≤ s ≤ f, and then extending by linearity to integrable functions.
A concrete approximation you can always keep in mind is range binning. Let f:[0,1]→[0,1] be measurable. For each n, partition [0,1] into 2^n bins of width 2^{-n} and define
s_n(x) = ∑_{k=0}^{2^n−1} (k/2^n) · 1_{ {x : k/2^n ≤ f(x) < (k+1)/2^n} }(x).
Then s_n is simple, 0 ≤ s_n ≤ f, and s_n(x) increases pointwise \to f(x) as n increases. This is not an ad hoc trick. It is one reason the integral behaves well under monotone limits.
Worked example: monotone convergence in action
Let f_n(x)=1_{[0,1−1/n]}(x) on [0,1]. Then f_n(x) increases pointwise \to 1_{[0,1)}(x). By the monotone convergence theorem,
∫_0^1 f_n(x) dx → ∫_0^1 1_{[0,1)}(x) dx.
Compute directly:
∫_0^1 f_n(x) dx = 1 − 1/n → 1,
and
∫_0^1 1_{[0,1)}(x) dx = 1.
The theorem matches calculation because the integral is defined in a way that respects increasing approximation by simple functions.
Worked example: dominated convergence and why it needs domination
Consider f_n(x)=sin(nx)/n on [0,2π]. Pointwise, f_n(x)→0. Also |f_n(x)| ≤ 1/n ≤ 1 for all x and all n, so the constant function g(x)=1 dominates every f_n and is integrable. Dominated convergence gives
∫_0^{2π} f_n(x) dx → 0.
You can compute the integral directly:
∫_0^{2π} sin(nx)/n dx = [−cos(nx)/n^2]_0^{2π} = 0.
So the theorem matches calculation. The pattern is what matters: oscillation is harmless when amplitude is uniformly controlled by an integrable bound.
In contrast, spike sequences show that pointwise convergence without any integrable domination does not allow exchanging limit and integral.
Worked example: L^p norms show different kinds of control
On [0,1], define h_n(x)=√n · 1_{[0,1/n]}(x). Then
||h_n||_1 = ∫_0^1 |h_n| dx = √n · (1/n) = 1/√n → 0,
but
||h_n||_2^2 = ∫_0^1 |h_n|^2 dx = n · (1/n) = 1,
so ||h_n||_2 = 1 for all n, and
||h_n||_∞ = √n → ∞.
This one example teaches three distinct notions of convergence:
- h_n → 0 in L^1
- h_n does not approach 0 in L^2
- h_n does not approach 0 uniformly
Different norms control different operations. If your argument needs worst-case error control, L^1 is not enough. If your argument needs quadratic energy control, L^2 is the right scale. If you need pointwise stability across the domain, the sup norm matters.
Worked example: differentiating under the integral sign as a limit exchange
A standard real-analysis question is when you can pass a derivative inside an integral:
d/dθ ∫_a^b F(x,θ) dx = ∫_a^b ∂_θ F(x,θ) dx.
A practical sufficient condition is:
- F(·,θ) is integrable for each θ
- ∂_θ F(x,θ) exists for almost every x
- and |∂_θ F(x,θ)| ≤ g(x) for an integrable g, uniformly for θ in a neighborhood
Then dominated convergence applied to the difference quotient gives the result.
Concrete example: I(θ)=∫_0^1 x^θ dx for θ > −1.
Direct computation gives I(θ)=1/(θ+1) and I'(θ)=−1/(θ+1)^2.
Inside the integral, ∂_θ x^θ = x^θ ln x, so the candidate is
∫_0^1 x^θ ln x dx.
Justifying the interchange is a domination problem: you need an integrable bound for |x^θ ln x| that holds uniformly for θ in a compact interval. This is a standard estimate, and one clean route uses the substitution x=e^{−t} \to turn the integral into a convergent integral on [0,∞) with an exponential weight. The key lesson is structural: interchanging derivative and integral is not symbolic; it is a limit exchange argument, and limit exchange is justified by uniform domination.
Worked example: almost everywhere is a feature, not a loophole
Many theorems conclude that something holds almost everywhere, meaning outside a set of measure zero. This is not a concession. It is a recognition that measure-zero exceptions do not affect integrals and L^p norms.
Example: define f(x)=1 on rational numbers in [0,1] and f(x)=0 on irrational numbers in [0,1]. The set of rationals has measure zero, so ∫_0^1 f(x) dx = 0. Yet f is discontinuous everywhere. This shows that measurability and integrability do not enforce pointwise regularity. Real analysis separates questions:
- If you care about integrals and averages, measure theory is the right tool.
- If you care about pointwise continuity, you need additional hypotheses.
The thread that ties the examples together
Every example above is a different face of one idea: the right notion of approximation depends on the operation you want to control.
- To control integrals, you need monotonicity, domination, or L^1 convergence.
- To control pointwise structure like continuity, you often need uniform convergence or equicontinuity.
- To control derivatives, you need uniform control on derivatives or domination applied to difference quotients.
- To control worst-case error, you need sup-norm bounds.
Measure and integration give real analysis its backbone because they provide a stable way to pass to limits, but they do so only under hypotheses that prevent mass from hiding in places your chosen notion of convergence cannot see.
If you work these examples until you can reproduce the core estimates without looking, you will have a practical command of the subject: not as a list of results, but as a disciplined method for deciding what is true, why it is true, and which hidden assumption would be required if it is not.
Worked example: Riemann integrable versus Lebesgue integrable
On a bounded interval, every Riemann integrable function is Lebesgue integrable and the integrals agree, but the Lebesgue integral handles limits more robustly. A simple illustration is the sequence of step functions that approximate a measurable function from below as in the range-binning construction. The point is not that step functions are special, but that the approximation can be arranged to be monotone, and monotonicity unlocks monotone convergence.
A practical habit is: if you can build an increasing sequence of simple functions s_n with s_n ↑ f, then you can compute or estimate ∫ f by computing ∫ s_n and taking a limit. This method naturally respects sets of measure zero, so you do not need to track behavior on negligible exceptional sets.
Worked example: convergence in measure is weaker than L^1 but still useful
A sequence f_n converges \to f in measure on [0,1] if for every ε>0,
μ({x : |f_n(x) − f(x)| > ε}) → 0.
This captures the idea that large errors occur on sets whose measure becomes small, but it does not force the average size of the error to vanish. Spikes again clarify the distinction.
Take the spike family f̃_n from earlier with integral 1 and pointwise limit 0. For any fixed ε>0, the set where f̃_n(x) > ε is essentially (0,1/n], whose measure is 1/n. So f̃_n → 0 in measure. Yet ∫ f̃_n = 1 for all n, so there is no convergence of integrals.
This shows why dominated convergence asks for domination and not merely convergence in measure: convergence in measure controls where the spikes are, but not how tall they can be.
Worked example: Fubini on a rectangle with a simple function
Let E be a measurable \subset of [0,1] and consider the function on the unit square [0,1]×[0,1],
F(x,y)=1_E(x).
This function does not depend on y. Its integral over the square is
∫_0^1 ∫_0^1 1_E(x) dy dx.
Compute the inner integral first: ∫_0^1 1_E(x) dy = 1_E(x) because the inner integral is over y and the integrand is constant in y. Therefore,
∫_0^1 ∫_0^1 1_E(x) dy dx = ∫_0^1 1_E(x) dx = μ(E).
If you reverse the order, you get the same result:
∫_0^1 ∫_0^1 1_E(x) dx dy = ∫_0^1 μ(E) dy = μ(E).
This is a toy computation, but it encodes the idea behind Fubini: under appropriate integrability hypotheses, iterated integrals agree with the integral over the product domain. In more advanced problems, the role of the toy calculation is to remind you what is being claimed, and which step uses integrability rather than algebra.
Worked example: absolute continuity as the integration-friendly notion of regularity
In the classical setting, a differentiable function with integrable derivative satisfies
f(b) − f(a) = ∫_a^b f'(x) dx.
In Lebesgue theory, the natural hypothesis for this identity is absolute continuity. An absolutely continuous function can be reconstructed from its derivative almost everywhere, and the derivative is integrable. This is one of the places where almost everywhere language is not a weakness but the correct interface between pointwise change and integral control.
The practical takeaway is: when you want to interchange differentiation and integration or recover a function from its derivative, the right regularity class is not merely continuous and not merely differentiable at many points, but absolutely continuous, because that notion is built to behave well under integration.