Measure Theory: What Goes Wrong Without It

The Question That Breaks Everything

Measure theory is not an abstraction for its own sake. It exists because of a concrete failure: you cannot assign a consistent “size” to every subset of $\mathbb{R}$.

We want a function $m: 2^{\mathbb{R}} \to [0, \infty]$ satisfying:

1. $m([a,b]) = b - a$ (intervals have their natural length)
2. $m(A + x) = m(A)$ (translation invariance)
3. $m\left(\bigsqcup A_n\right) = \sum m(A_n)$ (countable additivity)

Vitali’s Theorem (1905): No such function exists.

The proof is elementary. Partition $[0,1)$ by the equivalence relation $x \sim y \iff x - y \in \mathbb{Q}$. By the Axiom of Choice, select one representative from each class to form a set $V$. The rational translates $V + q$ (mod 1) are disjoint and partition $[0,1)$. If $m(V) = 0$, the countable union has measure $0 \neq 1$. If $m(V) > 0$, the countable union has measure $\infty \neq 1$. Contradiction.

1. The Resolution: Restrict the Domain

The fix is not to abandon measurement, but to give up on measuring everything. We restrict attention to a $\sigma$-algebra $\mathcal{F} \subsetneq 2^{\Omega}$—a carefully chosen collection of “measurable” sets that is closed under complements and countable unions.

The Borel $\sigma$-algebra $\mathcal{B}(\mathbb{R})$—generated by open sets—is rich enough to contain every set you will ever encounter in practice, yet sparse enough to avoid Vitali pathologies.

The tradeoff: we sacrifice universality (some sets are non-measurable) to gain consistency (the three axioms hold on $\mathcal{F}$).

This is not a technicality. It is a design principle: restrict the questions you are allowed to ask, and the answers become coherent.

2. Why Probability Needs This

In discrete probability, every event is measurable and $P$ is just a sum. The machinery is invisible. The moment we move to continuous random variables, we need it.

What is $P(X \in A)$? If $X$ is a continuous random variable with density $f$, then $P(X \in A) = \int_A f \, dx$. But the Lebesgue integral $\int_A f \, dx$ is only defined when $A$ is Lebesgue measurable.

What is $\mathbb{E}[Y | X]$? The conditional expectation $\mathbb{E}[Y | X]$ is defined as the projection of $Y$ onto the subspace of $\sigma(X)$-measurable functions in $L^2$. Without $\sigma$-algebras, we cannot even state what “information generated by $X$” means—let alone project onto it.

What is a filtration? The tower $\mathcal{F}_0 \subseteq \mathcal{F}_1 \subseteq \cdots$ that drives martingale theory is a sequence of $\sigma$-algebras. Information accumulation is the growth of a $\sigma$-algebra.

3. The Lebesgue Integral: Why It Supersedes Riemann

The Riemann integral partitions the domain into subintervals. The Lebesgue integral partitions the range and measures the preimage of each slice.

Concretely, to compute $\int f \, d\mu$:

• Approximate $f$ by simple functions $\sum a_i \mathbf{1}_{A_i}$
• Define $\int \sum a_i \mathbf{1}_{A_i} \, d\mu = \sum a_i \, \mu(A_i)$
• Take limits

Why is this better? Because it decouples the function from the geometry of its domain. The Riemann integral of $\mathbf{1}_{\mathbb{Q} \cap [0,1]}$ does not exist (the rationals are too scattered). The Lebesgue integral is simply $\mu(\mathbb{Q} \cap [0,1]) = 0$.

More importantly, the Lebesgue integral has superior limit theorems:

• Monotone Convergence: $f_n \uparrow f \implies \int f_n \to \int f$
• Dominated Convergence: $|f_n| \leq g$, $f_n \to f$ a.e. $\implies \int f_n \to \int f$

These are the workhorses of probability theory. Every time you interchange a limit and an expectation, you are invoking one of them.

4. Radon-Nikodym: Densities as Derivatives

Given two measures $\mu$ and $\nu$ on $(\Omega, \mathcal{F})$ with $\nu \ll \mu$ (absolute continuity: $\mu(A) = 0 \implies \nu(A) = 0$), there exists a measurable function $f$ such that:

$$\nu(A) = \int_A f \, d\mu \quad \text{for all } A \in \mathcal{F}$$

The function $f = \frac{d\nu}{d\mu}$ is the Radon-Nikodym derivative—a generalized density.

This single theorem unifies:

• Probability densities: $f(x) = \frac{dP}{d\lambda}(x)$ where $\lambda$ is Lebesgue measure
• Likelihood ratios: $\frac{dQ}{dP}$ is the Radon-Nikodym derivative between two probability measures
• Conditional expectation: $\mathbb{E}[Y | \mathcal{G}]$ is characterized as the Radon-Nikodym derivative of the signed measure $A \mapsto \int_A Y \, dP$ with respect to $P|_\mathcal{G}$

5. The Hierarchy of “Almost”

Measure theory introduces a vocabulary of “almost” that is indispensable:

Statement	Meaning	Strength
a.e. (almost everywhere)	Fails on a set of measure zero	Weakest
$L^p$ convergence	$\int	f_n - f
a.s. (almost surely)	$P(\lim f_n = f) = 1$	Strongest pointwise

The distinctions matter. $L^2$ convergence does not imply pointwise convergence. Almost sure convergence does not imply $L^1$ convergence (without domination). Each notion captures a different kind of “closeness,” and confusing them is a reliable source of errors.

The Takeaway

Measure theory is the type system of analysis. Just as a programming language’s type system prevents you from adding a string to an integer, $\sigma$-algebras prevent you from measuring non-measurable sets or conditioning on events of probability zero without proper care.

You don’t use measure theory because you enjoy abstraction. You use it because without it, the following all break simultaneously: continuous probability, conditional expectation, limit theorems, and the Lebesgue integral. The axioms are the minimum price for a consistent theory of “size” in uncountable spaces.