You roll a die, toss a handful of coins, wait for a bus, measure the height of the next person through the door. In every case a chance experiment happens somewhere in the background — in a space of raw outcomes — and yet the thing you actually care about is a number: the total showing, the count of heads, the minutes waited, the centimetres. A random variable is the bridge. It is the rule that reads a raw outcome and reports the number you were watching for.

Measure theory lets us say precisely what that "rule" is, and — crucially — which rules are allowed. The one-sentence answer, the whole of this page in a nutshell:

That is the entire definition. No new machinery — a random variable is just a measurable function whose domain happens to carry a probability measure. Everything else on this page (its law, its distribution function, discrete versus continuous, the \sigma-algebra it generates) is a consequence of that single line.

Why measurability is exactly the right condition

Here is the point of the definition — not a technicality, but the very thing that makes probability work. You want to ask questions like

But P only knows how to assign probabilities to events — the members of \mathcal{F}. The expression P(X \le x) is really shorthand for the probability of the set of outcomes on which X comes out at most x:

For that right-hand side to mean anything, the set X^{-1}((-\infty,x]) has to be something P can weigh — it must live in \mathcal{F}. Measurability is precisely the promise that it always does. Measurability is not red tape; it is the exact condition that makes the probabilities you want to compute well-defined. An arbitrary function X:\Omega\to\mathbb{R} might send a perfectly innocent interval back to a non-event, a set with no assigned probability, and then "P(X \le x)" would be gibberish.

This is why the cumulative distribution function below is built from \{X \le x\}: those rays are the seeds from which the entire distribution grows.

Worked example: X = number of heads in three tosses

Nothing makes the abstraction concrete like building the whole object by hand. Toss a fair coin three times. The sample space is the eight equally likely sequences:

Take \mathcal{F} = 2^{\Omega}, the full power set (every subset is an event), and P the uniform measure, P(\{\omega\}) = \tfrac18. Define the random variable

Is it measurable? On a power-set \sigma-algebra every subset is an event, so X^{-1}(B)\in\mathcal{F} automatically — a function on a discrete probability space is always a random variable. Now read off the preimages of the single values:

Weigh each preimage with P and you have pushed the probability from \Omega out onto the number line. That collection of weights is the law of X:

Drag the threshold in the figure below. It shows the raw outcomes up top, the map X carrying each to its head-count on the real line, and the sliding preimage X^{-1}\bigl((-\infty,x]\bigr) = \{X \le x\} lighting up as an event inside \Omega. Its probability is the cumulative distribution function F_X(x), read out live.

Watch the readout jump. Between the atoms the count of caught outcomes is constant, so F_X is flat; each time x crosses an integer k the preimage swallows a fresh clump of outcomes and F_X leaps by exactly P(X=k). That staircase — flat, jump, flat, jump — is the distribution function of a discrete random variable.

The law: a probability measure that lives on \mathbb{R}

The example did something profound in passing. We started with a probability measure P on the abstract space \Omega and ended with a collection of weights on \mathbb{R}. That new object is a bona fide probability measure in its own right — the pushforward of P along X, written \mu_X = P \circ X^{-1} and called the law or distribution of X:

It is a genuine probability measure on (\mathbb{R}, \mathcal{B}): it is non-negative, it gives the whole line total mass \mu_X(\mathbb{R}) = P(X^{-1}(\mathbb{R})) = P(\Omega) = 1, and it is countably additive because X^{-1} turns disjoint Borel sets into disjoint events and P is countably additive. Measurability of X is exactly what makes \mu_X well-defined on all of \mathcal{B} — the preimage never leaves \mathcal{F}.

This is the great simplification of the subject. Once you have \mu_X, you can forget \Omega entirely. Every probabilistic question about X — its mean, its variance, the chance it lands in any set — is answered by \mu_X alone, on the familiar real line, with no memory of the underlying coin tosses or bus stops that produced it.

The cumulative distribution function F_X

The law \mu_X assigns a number to every Borel set — a lot of information. It can be repackaged into a single function of one real variable, the cumulative distribution function (CDF), built from exactly the rays we said were enough:

Remarkably, F_X carries the whole distribution — because the rays generate \mathcal{B}, knowing F_X pins down \mu_X on every Borel set. And F_X is not any old function; the measure-theoretic definition forces three properties, and these three properties in turn characterise CDFs completely.

For the three-coin example the CDF is the staircase you watched being built:

Note the closed left endpoints: at x=1 the value is already \tfrac48, not \tfrac18. That is right-continuity in the flesh — F_X includes the jump at the point where it happens. The height of each jump is the point mass P(X=k) = F_X(k) - F_X(k^-).

Discrete, continuous, and the density in between

The shape of F_X sorts random variables into families, according to how the law \mu_X is spread out over the line.

• Discrete. \mu_X is concentrated on a countable set of atoms \{x_i\} with masses p_i = P(X=x_i) summing to 1. The CDF is a pure staircase (our coin example, dice, counts).
• Continuous. F_X has no jumps; the law spreads smoothly, and (in the absolutely continuous case) there is a density f_X \ge 0 with F_X(x) = \int_{-\infty}^{x} f_X(t)\,dt and P(X\in B)=\int_B f_X\,d\lambda.

Where does a density even come from, in the measure-theoretic picture? It is a Radon–Nikodym derivative. A density exists exactly when \mu_X is absolutely continuous with respect to a reference measure — Lebesgue measure \lambda in the continuous case, counting measure in the discrete case — and then

(The Radon–Nikodym theorem guarantees this derivative exists whenever \mu_X \ll \lambda; that is the whole story of "why densities integrate to give probabilities." Some laws are neither discrete nor continuous — mixtures, or the exotic Cantor distribution — but every law splits into these pieces.)

What X knows: the \sigma-algebra it generates

There is one more object riding along with every random variable, and it becomes indispensable the moment you meet conditional expectation. Among all \sigma-algebras on \Omega that make X measurable, there is a smallest one, gathered from the preimages of the Borel sets:

It is a \sigma-algebra (preimage commutes with the set operations), and by construction it is contained in \mathcal{F} precisely because X is a random variable. Intuitively \sigma(X) is the information carried by X: the events you can decide knowing only the value of X, and nothing more about which \omega occurred.

For the three-coin count, \sigma(X) is generated by the four preimages \{X=0\},\{X=1\},\{X=2\},\{X=3\}. Knowing X tells you how many heads fell — so you can decide the event "at least two heads" — but it can never distinguish TTH from THT from HTT, since X reports the same 1 for all three. That blind spot, made precise, is what \sigma(X) encodes, and it is the launch pad for conditioning.