Random Variables

The measurable-function definition of a random variable is not merely a formality — it is the load-bearing structure underneath essentially all of modern quantitative science. In statistics, every estimator (a sample mean, a maximum-likelihood estimate) is itself a random variable, a measurable function of the data, and its sampling distribution is exactly the induced law described here. In stochastic processes and mathematical finance, an entire process $\{X_t\}$ is a family of random variables indexed by time, and questions about stopping times or filtrations hinge on which $\sigma$-algebra each $X_t$ is measurable with respect to — directly generalizing the single-variable definition above. In information theory and machine learning, random variables model both noisy data and internal randomness of algorithms (e.g. random initializations, stochastic gradient noise), with measurability ensuring quantities like cross-entropy loss are well-defined expectations. In statistical physics, macroscopic observables (energy, magnetization) are random variables on a microstate space, exactly mirroring the abstract-$\Omega$-to-$\mathbb{R}$ structure introduced here. In cryptography, the security of a scheme is often phrased via random variables representing an adversary's view or output, with measurability guaranteeing that "probability of successful attack" is a well-posed number rather than an ill-defined heuristic.

Ask the tutor: How does the choice of $\sigma$-algebra $\mathcal{F}$ change which functions on $\Omega$ qualify as random variables?

Think of $\Omega$ as the true, possibly enormous or unstructured, space of outcomes: a sequence of coin flips, a random path of a particle, a hand of cards. The probability measure $\mathbb{P}$ lives on $\Omega$ and encodes everything about the experiment's uncertainty. A random variable $X$ is a reading device: it looks at an outcome $\omega$ and reports a single real number $X(\omega)$, deterministically and always the same way. Roll the same two dice showing the same faces twice — you get the same sum both times. The variable does not "choose" a value; the outcome does, and $X$ just reads it off. Why require measurability? Because the only thing we know how to compute probabilities for are elements of $\mathcal{F}$. If we want to ask "what is $\mathbb{P}(X \le 8)$?", this question is only meaningful if the underlying set of outcomes $\{\omega : X(\omega) \le 8\}$ is itself an event — a member of $\mathcal{F}$ that $\mathbb{P}$ knows how to measure. Measurability guarantees this for every real threshold $x$, and (as the theorem below shows) that guarantee automatically extends to every Borel set, which is essentially every set of reals one could reasonably want to consider.

(Proof sketch, stated for completeness — not the focus of this article: one shows $\{X + Y < x\} = \bigcup_{q \in \mathbb{Q}} \big(\{X < q\} \cap \{Y < x - q\}\big) \in \mathcal{F}$ using countability of $\mathbb{Q}$ and closure of $\mathcal{F}$ under countable unions and finite intersections; measurability of $g(X)$ follows from $\{g(X) \in B\} = X^{-1}(g^{-1}(B))$ and part (2) of the theorem above applied twice, since $g^{-1}(B)$ is Borel whenever $g$ is Borel measurable and $B$ is Borel.)

Example 1 — A biased, non-symmetric coin-flip random variable. Let $\Omega = \{H, T\}$ with $\mathcal{F} = \{\emptyset, \{H\}, \{T\}, \Omega\}$ and $\mathbb{P}(\{H\}) = 0.37$, $\mathbb{P}(\{T\}) = 0.63$. Define

Check measurability directly from the definition: for any $x \in \mathbb{R}$,

Every one of these four possible sets ($\emptyset, \{T\}, \Omega$) lies in $\mathcal{F}$, so $X$ is a random variable. Its law is the Borel measure

a two-point (Bernoulli-type) distribution on $\mathbb{R}$, concentrated at $-1.9$ and $5.2$ with the given asymmetric weights — illustrating that the numeric values assigned by $X$ are irrelevant to which weights attach to which outcome; only the underlying $\mathbb{P}$ on $\{H,T\}$ determines the probabilities. Now compute a genuine probability statement: $\mathbb{P}(X > 0) = \mathbb{P}(\{X > 0\})$. Since $\{X > 0\} = \{H\}$ (as $5.2>0$ but $-1.9 \not > 0$), and $\{H\} \in \mathcal{F}$, this is a legitimate event with $\mathbb{P}(X>0) = 0.37$. Had $X$ mapped $H$ and $T$ to values with $\{X>0\}$ landing outside $\mathcal{F}$, this computation would be undefined — which cannot happen here precisely because $\mathcal{F}$ is the full power set of $\{H,T\}$, but would be a genuine risk on a coarser $\sigma$-algebra (see the counterexample below).

Example 2 — Sum of two non-identical dice, an explicit induced law. Let $\Omega = \{1,\dots,6\}^2$ (36 outcomes), $\mathcal{F} = 2^\Omega$ the power set, and suppose the two dice are not fair in the same way: die 1 shows each face with probability $1/6$ (fair), but die 2 is loaded so that face $6$ has probability $1/4$ and faces $1$–$5$ each have probability $3/20$ (check: $5 \times 3/20 + 1/4 = 15/20 + 5/20 = 1$). Assume independence of the two dice, as defined in Independence of Events, so

Define $X(i,j) = i + j$. This is a random variable on $(\Omega, 2^\Omega, \mathbb{P})$ trivially, since $\mathcal{F} = 2^\Omega$ makes every subset of $\Omega$ an event, so every preimage automatically lies in $\mathcal{F}$. Compute $\mathbb{P}(X = 9)$ explicitly. The pairs $(i,j)$ with $i + j = 9$ and $i,j \in \{1,\dots,6\}$ are

Using independence and the two different marginal laws:

Summing with a common denominator of $120$: $\tfrac{1}{24} = \tfrac{5}{120}$ and each $\tfrac1{40} = \tfrac{3}{120}$, so

This exhibits the full pipeline: $X$ is measurable (trivially, since $\mathcal{F}$ is the power set), the event $\{X=9\}$ is identified explicitly as a subset of $\Omega$, and $\mathbb{P}_X(\{9\})$ is computed from the underlying (here asymmetric) measure $\mathbb{P}$ on $\Omega$ — precisely the content of the theorem above applied to the single point $B = \{9\}$.