31 distributions, live

The Distribution Bench

Every distribution worth knowing, with its shape under your hands.

Each card holds a picture of the shape, sliders that reshape it, and preset buttons for situations worth seeing. Hover anywhere on a chart (or tap it on a phone) to read off the exact chance at that point. Drag a slider and the curve answers immediately; every slider says underneath what it means in real terms. Hover or tap a preset to find out what it demonstrates, and tap any dotted word for a plain definition. Tap a title to open the full write-up: what it models, where it goes wrong, and what it turns into.

Nothing matches that. Try “count”, “waiting”, “heavy tail”, or a name.

Why the tail is the whole argument

Four symmetric curves centred on zero, three of them with the same variance. What separates them is not the middle but the edge. Switch the vertical axis to logarithmic and look at six units out: the normal has all but disappeared while the Cauchy is still going, a gap of six orders of magnitude.

How they turn into each other

Almost none of these were invented separately. Follow an arrow to see the assumption that converts one into another — most of applied statistics lives on these edges.

Scroll sideways to follow the chain.

Picking one

Start from the thing you measured, not from the curve you like.

A single yes or noBernoulli
Successes in a fixed number of triesBinomial
Successes drawn from a finite pool, no replacementHypergeometric
Events per hour, rare and independentPoisson
Events per hour, but more erratic than Poisson allowsNegative binomial
Tries until the first successGeometric
Time until the next event, no memoryExponential
Time until the k-th eventGamma / Erlang
Time until failure, with ageing or burn-inWeibull
A sum of many small independent effectsNormal
A product of many small effects; prices, incomesLog-normal
A proportion or an unknown probabilityBeta
Anything between two bounds, nothing more knownUniform
The largest value in each block or yearGumbel
Wealth, city size, file size — the top dominatesPareto
A mean estimated from a small sampleStudent's t
A sample variance, or a goodness-of-fit totalChi-square
A ratio of two variancesF

Past one dimension

These need more than an x-axis, so they get a sentence instead of a curve.

  • Multinomial — the binomial with more than two outcomes. Counts across k categories from n draws; every marginal is a binomial.
  • Multivariate normal — the normal with a covariance matrix instead of a variance. The backbone of regression, Kalman filters and Gaussian processes.
  • Dirichlet — the beta with more than two parts. A distribution over probability vectors that sum to one; the conjugate prior for the multinomial, and the base of topic models.
  • Wishart — a distribution over covariance matrices; the multivariate chi-square.
  • Von Mises — the normal wrapped onto a circle, for angles, wind directions and times of day.
  • Copulas — a way to keep each variable's own marginal distribution while specifying the dependence between them separately. Their misuse in credit models is a well-documented chapter of 2008.
  • Mixtures — two or more distributions weighted together. Any bump-shaped-but-not-quite dataset is usually a mixture, and fitting a single curve to it produces a mean nobody in the data actually has.
  • The empirical distribution — your data, each point weighted 1/n. Resampling from it is the bootstrap, and it assumes nothing at all.