Random variables

The formula

X: Ω → ℝ A random variable is a function from the set of possible outcomes Ω to the real numbers. Roll the die, pull the resistor, run the fatigue coupon — X is the number that comes back ^[4].

E[X] = Σ x_k · p(x_k) (discrete) The expected value of a discrete variable: each possible value weighted by its probability. A fair die gives E[X] = 3.5 — a value the die itself can never show.

E[X] = ∫ x · f(x) dx (continuous) Same idea with a density instead of a mass function: the probability-weighted center of the scatter ^[2].

Var(X) = E[(X − μ)²] = E[X²] − μ², σ = √Var(X) Variance is the mean squared distance from the mean; σ is its square root, back in the original units ^[7]. For that fair die, E[X²] = 91/6 ≈ 15.17, so Var = 15.17 − 3.5² ≈ 2.92.

E[aX + bY] = a·E[X] + b·E[Y] Expectation is linear, always — whether X and Y are independent, correlated, or the same variable twice.

Var(X + Y) = Var(X) + Var(Y) (independent X, Y only) Variances add; standard deviations root-sum-square. Drop the independence requirement and this equation quietly becomes wrong.

E[g(X)] = ∫ g(x) · f(x) dx ≠ g(E[X]) in general Push a random variable through a nonlinear function and the mean of the output is not the function of the mean. Only linear g gets a pass.

Where you meet it

The parts bin. Every 1 kΩ resistor pulled off the reel is one draw from a distribution, not a copy of the number on the label. The datasheet tolerance is a statement about that distribution's spread. Design against the label value alone and the population, not the schematic, decides your yield.
The tolerance stack at a design review. Four mating parts, each held to ±0.1 mm. Worst case says the stack can be off by ±0.4 mm. Treat each dimension as an independent random variable and the statistical stack is √(4 · 0.1²) = ±0.2 mm — half the worst case. Which number you defend at the board depends on whether you can actually defend the independence.
Averaging on the test stand. A load-cell reading is true value + random error, and the error is a random variable with its own σ. Average n independent readings and the mean's scatter is σ/√n: 25 averages buys a factor of 5, not 25. The square root is why averaging stops paying long before your patience runs out.
The reliability slide. Fatigue life, bearing life, MTBF — none of these are numbers. Each is a random variable, and the single figure on the chart is one summary of its distribution (a mean, an L10, a B-basis allowable). A review board that argues about "the" life of a part is arguing about which percentile to quote.

How it works

The random variable is the load path between physics and statistics. The experiment lives in Ω — every micro-detail of a solder joint, a wind gust, a grain boundary. You will never enumerate Ω. The random variable X collapses all of it to one number per outcome, and the distribution of X is the complete engineering description: everything you can compute — means, spreads, exceedance rates, percentiles — comes from the distribution, not from any single draw ^[4].

Two species. A discrete random variable takes countable values — failure counts, bit errors, particles per cubic foot — and carries a probability mass function that assigns probability to each value directly. A continuous one takes values on a continuum — voltage, stress, time to failure — and carries a density instead, with the odd consequence that P(X = c) = 0 for any exact value c ^[4]. Your DMM will never read exactly 4.99700 V twice, and the theory agrees. Probability for continuous variables lives in intervals, never at points.

The measurement model most engineers actually use is reading = true value + bias + ε, where ε is a zero-mean random variable. Bias is systematic — averaging never touches it; only calibration does. The random part ε is what statistics can attack. Conflating the two is the classic uncertainty-budget error: a hundred averages of a miscalibrated instrument give you a very precise wrong answer.

The gotchas worth money:

Variance addition assumes independence. Two channels reading the same test article share temperature drift, supply ripple, and the same technician. Correlated errors do not RSS away; in the worst case they add straight. An RSS'ed uncertainty budget built on secretly correlated terms is optimistic by construction, and it fails in the direction that hurts.
Nonlinearity moves the mean. Take P = V²/R with V a random variable of mean 10 V and σ = 0.5 V across a 50 Ω load. Naive: 10²/50 = 2.000 W. Correct: E[V²] = μ² + σ² = 100.25, so mean power is 2.005 W. Small here — but the shift scales with variance and curvature, and for strongly nonlinear responses (dB conversions, thermal radiation going as T⁴, life models exponential in stress) plugging the mean into the formula is not computing the mean of the answer.
Sample statistics are random variables too. The x̄ and s you compute from 30 units are one draw from the distribution of x̄ and s. Run the lot again and both numbers move. Every confidence interval in the trade exists because someone took this seriously.
The distribution is an assumption, not a fact. Calling something Gaussian because scatter usually looks Gaussian works fine near the middle and can be badly wrong in the tails — which is exactly where spec limits, first-failure times, and 1-in-10,000 requirements live. Weakest-link failures want a Weibull; counts want a Poisson (whose mean and variance are the same number); waiting times with no memory want an exponential, where P(T > t) = e^(−λt) and 36.8% of units outlive the mean life.

The mistake underneath all of these is treating a measured number as the quantity itself. The quantity is the random variable; the number is one sample of it. Every calculation downstream — margins, yields, budgets — is a calculation about the distribution, whether the spreadsheet admits it or not.

History

The object showed up before the name did, at a gambling table. In 1654 Blaise Pascal and Pierre de Fermat, trading letters over how to fairly split the stakes of an interrupted game, worked out that a future gain should be valued in proportion to the chance of getting it — the first real computation with an unknown-but-distributed quantity ^[2]. Christiaan Huygens heard about the correspondence on a visit to Paris and, back in Holland, wrote De ratiociniis in ludo aleae (1657), the first printed book on probability, building the whole treatment around the notion of expectation ^[1]^[2].

For the next two centuries probability stayed a theory of events and their chances. Pafnuty Chebyshev, working in St. Petersburg in the mid-1800s, changed the unit of thought: his 1867 paper on mean values proved a general law of large numbers by manipulating random quantities and their expectations directly, and he is credited as the first to think systematically in terms of random variables ^[3]^[4]. Kolmogorov himself later wrote that Chebyshev was the first to clearly grasp and use the notions of a random quantity and its expectation ^[3].

The rigor arrived in 1933. Andrey Kolmogorov's slim monograph Grundbegriffe der Wahrscheinlichkeitsrechnung rebuilt probability from axioms the way Euclid built geometry ^[5]^[6], and inside that framework the random variable finally got its modern definition: a measurable function from a probability space to the reals ^[4]. The measure-theoretic fine print rarely matters at the bench, but it is what makes every downstream tool — conditional expectation, limit theorems, stochastic processes — provable instead of plausible.

Related tools

/tools/rms-peak — the RMS of a zero-mean signal is the σ of its amplitude as a random variable
/tools/crest-factor — peak-to-RMS ratio is a statement about the amplitude distribution's tails
/tools/thermal-noise-floor — thermal noise voltage as the canonical Gaussian random variable
/tools/snr-enob — quantization error modeled as a uniform random variable with RMS q/√12 ≈ 0.289·q
/tools/bearing-life-l10 — bearing life as a Weibull random variable; L10 is its 10th percentile
/tools/half-life-decay — the exponential survival function in decay form

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →