Lebesgue Integral: How It Really Works

Build a usable mental model of the Lebesgue integral by thinking in level sets, measurability, and limits. Learn how simple-function approximation defines the integral, when you can swap limits and integrals, and where common mistakes break proofs.

Integrating by chopping the $x$-axis into tiny intervals feels natural until you meet functions that spike, oscillate, or behave differently on dense sets. The Lebesgue integral fixes that by changing what gets sliced. Instead of tracking where $x$ goes, it tracks how much of the domain lands at each function value. Once you see integration as measuring level sets, the convergence theorems stop feeling like magic tricks and start reading like bookkeeping rules.

Slice by values, not by $x$

Riemann integration asks you to approximate area with vertical rectangles built on subintervals of the domain. Lebesgue integration asks a different question. For each height $t$, how big is the set of points where $f(x)$ is around that height? That flips the geometry from vertical slices to horizontal slices, and it is exactly what you want when the function is ugly in $x$ but well behaved in terms of how often it takes certain values.

A good way to say it without pictures is this. Riemann sums depend on partitions of the domain, while Lebesgue’s construction depends on partitions of the range and the measures of the corresponding preimages in the domain. The phrase preimage matters because it is where measurability enters. You do not measure the set of function values directly. You measure subsets of the domain like $\{x:f(x)>t\}$.

See how the two slicing philosophies behave on the same function.

Once you accept that the domain gets counted by how it maps into value bands, many pathological examples soften. Discontinuities at individual points become irrelevant because single points have Lebesgue measure zero. Dense oscillations become manageable if the sets where the function is large are small in measure.

Mental shift
Lebesgue integration is not about tracking where the graph goes. It is about counting how much of the domain contributes each height.

Measurability is the price of admission

Sigma-algebras and why they show up

To measure sets reliably, you need a collection of sets closed under the operations you use in limits. A sigma-algebra is built for that. It contains the whole space, is closed under complements, and under countable unions. Countable is not decoration. Limits create countably many set operations even when the original definition looks finite.

With that structure, Lebesgue measure becomes a measure defined on a sigma-algebra of subsets of $\mathbb{R}$, with the key features you expect. It agrees with interval length, it is countably additive, and it treats sets of points as size zero when appropriate.

Measurable functions via preimages

A function $f$ is measurable when preimages of basic sets in the codomain land in the sigma-algebra on the domain. A standard test is that $\{x:f(x)>a\}$ is measurable for every real $a$. This is the bridge between slicing by values and measuring subsets of the domain. If those preimages are not measurable, the Lebesgue integral is not defined, because the horizontal slices you want to count are not legal objects.

Explore how the definitions depend on each other.

The payoff is that measurability is stable. If $f_n$ are measurable, then $\sup_n f_n$, $\inf_n f_n$, $\limsup f_n$, and pointwise limits are measurable. That stability is what makes convergence theorems possible.

Building the integral from simple functions

A simple function is a finite linear combination of indicator functions, typically written as $\varphi=\sum_{k=1}^m a_k 1_{A_k}$ with measurable sets $A_k$ and real coefficients $a_k$. For $\varphi\ge 0$, its integral is defined by

$$\int \varphi\,d\mu=\sum_{k=1}^m a_k\,\mu(A_k).$$

Nothing mysterious is happening. You are adding heights times measured widths, except the widths are measures of possibly complicated measurable sets, not interval lengths.

For a nonnegative measurable $f$, the Lebesgue integral is constructed by approximation from below. You look at all simple $\varphi$ with $0\le\varphi\le f$ and define

$$\int f\,d\mu=\sup_{\varphi\le f}\int \varphi\,d\mu.$$

This is why nonnegativity comes first. A supremum over lower bounds is a clean, one sided construction. It cannot produce undefined expressions like $\infty-\infty$.

Watch how increasing simple approximations push the integral upward.

A common practical pattern is to choose $\varphi_n$ that quantize the range. For example, take steps of size $2^{-n}$ and cap at height $n$ so the approximation is bounded and increases with $n$. The integral becomes a limit of sums of the form step height times measure of a level band.

One-sided safety
Approximating from below makes limits behave. You can lose mass, but you cannot create cancellation errors.

Extending to signed functions without breaking things

For a general real-valued measurable function $f$, define its positive part and negative part by $f^+=\max(f,0)$ and $f^-=\max(-f,0)$. Then $f=f^+-f^-$ and $|f|=f^++f^-$. The Lebesgue integral of $f$ is defined when at least one of $\int f^+\,d\mu$ or $\int f^-\,d\mu$ is finite. In that case,

$$\int f\,d\mu=\int f^+\,d\mu-\int f^-\,d\mu.$$

If both $\int f^+$ and $\int f^-$ are infinite, you have the forbidden form $\infty-\infty$. This is not a technicality. It is exactly where conditional cancellation can fake a finite answer depending on how you take limits. Lebesgue refuses to assign a value there because no rearrangement-invariant notion of area exists.

The cleanest sufficient condition is absolute integrability, meaning $\int |f|\,d\mu<\infty$. Then both parts are finite and the integral is stable under the limit operations you usually care about.

Compare how this differs from improper Riemann integration behavior.

The three convergence tools you actually use

The fastest way to remember the convergence theorems is to treat them as rules for when you may swap a limit and an integral. They are not separate tricks. They are three ways to justify the same move.

Monotone Convergence Theorem

If $0\le f_n\uparrow f$ pointwise, then

$$\int f_n\,d\mu\uparrow\int f\,d\mu.$$

No domination, no boundedness, no integrability assumptions beyond nonnegativity. Monotonicity is the entire engine.

Fatou’s lemma

For $f_n\ge 0$,

$$\int \liminf_{n\to\infty} f_n\,d\mu\le \liminf_{n\to\infty}\int f_n\,d\mu.$$

Fatou is what you reach for when you do not have enough structure for equality. It is a lower semicontinuity statement for the integral.

Dominated Convergence Theorem

If $f_n\to f$ pointwise and there exists an integrable $g$ with $|f_n|\le g$ for all $n$, then

$$\int f_n\,d\mu\to\int f\,d\mu.$$

The dominator is the price you pay to allow signed functions and oscillation without monotonicity.

See the hypotheses side by side so you can choose quickly.

A useful habit is to ask two questions before trying to interchange limit and integral. Are the functions nonnegative and monotone? If not, can you produce a single integrable $g$ that bounds them all?

Why Lebesgue beats Riemann on ugly behavior

Many classic counterexamples are not about being exotic. They are about being incompatible with partitions of the domain. Dense discontinuities break Riemann because no interval, however small, avoids the oscillation. Lebesgue ignores pointwise pathology when it lives on sets of measure zero, and it handles dense oscillation when the large values occur on small-measure sets.

Dirichlet-type functions are the canonical demonstration. A function can jump between two values on every interval, making Riemann sums unstable, while Lebesgue can still integrate it because the set where it takes each value has a clean measure.

Open the optional deep dive to see how measure zero changes the outcome.

The deeper lesson is that Lebesgue integration is built to commute with limits. If your problem involves sequences of functions, truncations, approximations, or parameter limits, Riemann integration often makes you fight the definition. Lebesgue makes the definition do the fighting for you.

Where people slip and what to check

Most mistakes are not algebra. They are choosing the wrong theorem, or assuming a hypothesis you do not actually have.

• Nonmeasurable sets: if preimages like $\{x:f(x)>a\}$ are not measurable, you cannot even start.
• Bad dominators: in DCT, $g$ must be integrable and must bound all $|f_n|$, not just eventually or pointwise with exceptions that are not controlled.
• Pointwise vs uniform: pointwise convergence alone does not justify moving limits through integrals. Uniform convergence can help, but it is not a substitute for measurability and integrability checks.
• Fubini and Tonelli: swapping integrals in two variables needs either nonnegativity for Tonelli’s theorem or absolute integrability for Fubini’s theorem. Conditional integrability can fail spectacularly.

Try your own example sequence or function and check which theorem applies.

A good diagnostic when something feels fishy is to look for hidden $\infty-\infty$, or a limit that creates negative mass by cancellation. If either is present, reach for $f^+,f^-$ or for absolute values and see whether integrability survives.

A working toolkit for real problems

When you meet an integral involving limits, parameters, or nasty discontinuities, force it into one of a few patterns.

Pick a theorem by the shape of the sequence

• Monotone up: if you can rewrite $f_n\uparrow f$ with $f_n\ge 0$, use MCT.
• Dominated: if the sequence oscillates or changes sign, hunt for a single integrable $g$ and use DCT.
• Only nonnegative bounds: if you are stuck, apply Fatou to get an inequality, then see what additional estimate would upgrade it to equality.

Rewrite the integrand into measurable pieces

Measurability is usually easy if you use closures under operations. Continuous functions are measurable. Limits of measurable functions are measurable. Indicators $1_A$ are measurable exactly when $A$ is measurable. In computations, rewriting into combinations of these building blocks is often faster than invoking a theorem from scratch.

Two moves show up constantly. Truncate and then pass to a limit, like $f_n=\min(f,n)$ for $f\ge 0$. Or localize to finite measure sets, like $f_n=f\cdot 1_{[-n,n]}$, then let $n\to\infty$. Both are designed to create monotonicity or domination you can prove.

Default move
If you cannot see a theorem, truncate first. Truncation often creates monotone or dominated sequences you can actually verify.