Student's t distribution

The formula

The statistic the distribution describes. Take n independent readings from a normal process with true mean μ; compute the sample mean x̄ and sample standard deviation s. Then ^[1]^[2]:

t = (x̄ − μ) / (s/√n)

Reading: this is the familiar "how many standard errors is the mean off by" ratio — except the denominator uses the estimated s, not the true σ, and that substitution changes the distribution of the ratio. It follows a t distribution with ν = n − 1 degrees of freedom.

The density itself ^[1]^[2]:

f(t) = Γ((ν+1)/2) / ( √(ν·π) · Γ(ν/2) ) · (1 + t²/ν)^(−(ν+1)/2)

Reading: a symmetric bell centered on zero, but with power-law tails — (1 + t²/ν) to a negative power — instead of the normal's e^(−t²/2). Heavier tails means extreme values are far more likely than the Gaussian admits. The shape is controlled entirely by ν: at ν = 1 it is the Cauchy distribution (no mean, no variance), the variance is ν/(ν−2) once ν > 2, and as ν → ∞ it converges to the standard normal ^[1]^[2].

What you actually use it for — the confidence interval on a mean ^[1]^[3]:

x̄ ± t(0.975, n−1) · s/√n

Reading: same recipe as the normal interval, with the Student multiplier replacing 1.96. The multiplier at 95% two-sided: 12.71 for n = 2, 4.30 for n = 3, 2.78 for n = 5, 2.26 for n = 10, 2.05 for n = 30, sliding toward 1.96 as data accumulates.

The GUM version, for uncertainty budgets ^[3]^[4]:

U = k·u_c(y) with k = t_p(ν_eff) and ν_eff = u_c⁴(y) / Σᵢ ( uᵢ⁴(y)/νᵢ )

Reading: expanded uncertainty is coverage factor times combined standard uncertainty, the coverage factor comes off the t distribution, and the Welch–Satterthwaite formula converts a budget of components with different degrees of freedom into one effective ν_eff. A Type A component from n readings carries ν = n − 1; a Type B component from a handbook or a rectangular assumption is usually treated as ν = ∞ ^[3]^[4].

Where you meet it

The cal bench. A load-cell check at a 10.000 lbf deadweight point, five readings: 9.994, 10.006, 9.998, 10.011, 9.991. Mean 10.000, s = 0.0083, SE = 0.0037. The honest 95% interval is ±2.776 · 0.0037 = ±0.010 lbf. Use 1.96 instead and you'd claim ±0.007 — an interval 42% too tight, from a multiplier chosen for data you don't have.
The test stand. Five qual motors, one thrust number each, and the review board wants "mean thrust with 95% confidence." With four degrees of freedom the multiplier is 2.78, not 2. Every mean-of-few-articles claim in a qual package runs through this table, and the difference between 2 and 2.78 has flipped more than one margin from green to yellow.
The uncertainty budget. ISO 17025 cal labs quote U = k·u_c with k = 2 by habit. GUM Annex G says that habit is only licensed when ν_eff is comfortably large. A budget dominated by a Type A term from five readings can land at ν_eff ≈ 8, where the proper 95.45% factor is 2.37 — and a blanket k = 2 delivers about 92% coverage while the certificate says 95 ^[3]^[4].
The screening t-test. Two small batches of parts, did the anneal change the yield strength or not — the t-test that answers it is this distribution applied to a difference of means, and it exists precisely because production experiments come in batches of five, not five hundred.

How it works

The mechanism is a ratio of two random things. If σ were known, (x̄ − μ)/(σ/√n) would be exactly standard normal. But s is computed from the same few readings, so it wanders — and with n = 4 it wanders a lot. When s happens to come in low, the ratio blows up; a small denominator manufactures large t values that a fixed σ never would. Formally, t is a standard normal divided by √(χ²_ν / ν) — an independent chi-squared variable doing the wandering — and that division is what fattens the tails ^[1]^[2]. As n grows, s pins down, the denominator stops moving, and the t distribution collapses onto the normal. NIST's rule of thumb: the approximation is decent past ν = 30 ^[2].

How heavy the tails really are: at ν = 4, P(|t| > 3) ≈ 4.0%. For the normal it's 0.27%. A "three-sigma" excursion is fifteen times more probable than Gaussian intuition says — which is exactly the rate at which small-sample means land outside naive error bars.

The mistake people make, in the order it costs money:

Using 1.96 with small n. The classic. At n = 5, an interval built with 1.96 covers the true mean about 87.8% of the time, not 95%. The certificate is wrong roughly one time in eight instead of one in twenty, and nothing on the page reveals it. This is the whole reason the distribution exists.
Applying the multiplier to the wrong spread. t·s/√n bounds the mean. It says nothing about where the next unit lands — that's a tolerance or prediction interval, which is wider still. Briefing a mean interval as a fleet-dispersion bound compounds two errors in the tight direction.
Forgetting where the degree of freedom went. ν = n − 1 because s is measured around x̄, not around the unknown μ — one piece of information already spent. Pool data across conditions or fit a line first and more degrees of freedom disappear; a regression with p fitted coefficients leaves n − p.
Expecting t to fix everything. The t interval corrects for an estimated σ under roughly normal, independent readings. It does not rescue correlated data (autocorrelated samples inflate the effective s/√n problem the same way they always do), and it does not remove bias — a calibration offset sails through untouched. Exact normality matters less than people fear for the mean itself, but a wild outlier in five readings distorts both x̄ and s at the same time, and no table multiplier repairs that.
k = 2 as a reflex. In GUM budgets, k = 2 is shorthand for "ν_eff is large." Check it. A worked case: Type A u = 0.8 units at ν = 4, Type B u = 0.5 at ν = ∞ gives u_c = 0.94 and ν_eff = 7.7; truncate to 7 and Table G.2 says k = 2.43 for 95.45% ^[3]. Quoting k = 2 there understates U by 18%.

One inversion worth knowing: the same heavy tails that widen your error bars make the t distribution useful as a noise model. Sensor channels with occasional spikes are often modeled as t-distributed with low ν rather than Gaussian, because a fit that expects rare large excursions doesn't get dragged around by them. The distribution that protects small samples also describes glitchy data.

History

The curve was on paper before the man it's named for was out of school. In 1876 the German mathematician Jacob Lüroth, working on the theory of observational errors, derived what is now recognized as the t distribution as a Bayesian posterior — a result that sat unnoticed until historians excavated it in 1996 ^[1]^[9].

The version engineers inherited came out of a brewery. William Sealy Gosset joined Arthur Guinness & Son in Dublin in 1899, fresh from Oxford, as a chemist and brewer ^[5]^[6]. His problem was industrial: barley and hop experiments yielded samples of three, four, five — and the large-sample error theory of the day had nothing rigorous to say at those sizes. Guinness sent him to Karl Pearson's laboratory at University College London for the 1906–07 session to work it out ^[5]^[6]. The result was "The Probable Error of a Mean," published in Biometrika in 1908 ^[7]^[8]. Guinness, burned once by an employee publication that gave away trade information, barred its scientists from publishing under their own names or mentioning the company — so the paper appeared under the pseudonym "Student" ^[1]^[5].

Gosset's original statistic wasn't quite the modern t. He worked with a ratio he called z, built with the standard deviation computed over n rather than n − 1; the modern form is t = z·√(n−1) ^[7]. Ronald Fisher saw what the tables were really for, recast the statistic to fit his framework of degrees of freedom, and put the letter t into print — the symbol itself appears to have originated in Gosset's correspondence with Fisher around 1923 ^[7]. Fisher's 1925 textbook Statistical Methods for Research Workers and his paper "Applications of 'Student's' Distribution" carried the method to working scientists, and the name "Student's distribution" stuck ^[1]^[7]. Gosset once mailed Fisher a copy of his tables with a note that Fisher was "the only man that's ever likely to use them" ^[7]. The tables now sit in every test report, cal certificate, and statistics package on the planet.

Related tools

/tools/strain-gauge-bridge — a handful of repeated bridge readings is the textbook small-n case; the mean's error bar takes a Student multiplier.
/tools/snr-enob — noise figures computed from a finite record are estimates with their own degrees of freedom, not gospel.
/tools/rms-peak — the RMS of a short noise capture is an s, and an s from little data is exactly what this distribution prices.
/tools/adc-resolution — deciding whether averaging bought you a real resolution gain from a short test record is a small-sample mean question.

Sources

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