What does a gamma distribution with shape k and rate lambda represent?

## The family The gamma distribution is a flexible two-parameter law for a positive continuous quantity. Its density is ``` f(x) proportional to x^(k - 1) * exp(-lambda * x) for x > 0 ``` with **shape** `k > 0` and **rate** `lambda > 0`. The proportionality constant is `lambda^k / Gamma(k)`, where `Gamma(k)` is the gamma function — a continuous extension of the factorial satisfying `Gamma(n) = (n - 1)!` for positive integers. That constant is what allows non-integer shapes, and it is where the family gets its name. Moments: ``` mean = k / lambda variance = k / lambda^2 SD = sqrt(k) / lambda ``` ## The waiting-time reading The interpretation that makes the family memorable applies when k is a positive integer. If single events occur independently at a constant rate `lambda`, the wait for the first one is exponential with mean `1 / lambda`. The wait for the k-th one is a gamma with shape k and the same rate: you are stacking k independent exponential waits end to end. That immediately explains the moments — k waits of mean `1/lambda` give a total mean of `k / lambda`, and k independent variances of `1/lambda^2` add to `k / lambda^2`. The integer-shape case is often called the Erlang distribution. For non-integer k there is no "count of events" story, but the density is still perfectly well defined and the same moment formulas hold. ## What each parameter actually does **Shape k changes the form of the curve.** - `k < 1`: the density is unbounded as x approaches 0 and decreases throughout — mass crowds towards zero with a long tail. - `k = 1`: the density is `lambda * exp(-lambda x)`, exactly the exponential, peaking at zero. - `k > 1`: the density starts at 0, rises to a single interior mode at `(k - 1) / lambda`, then decays. As k grows the shape becomes steadily more symmetric and bell-like, because the relative spread `SD/mean = sqrt(k)/k = 1/sqrt(k)` shrinks. **Rate lambda changes only the scale.** If X is gamma with shape k and rate lambda, then `c * X` is gamma with shape k and rate `lambda / c`. Changing the rate stretches or compresses the horizontal axis and rescales the height to keep the area at 1; the silhouette is untouched. That is the precise sense in which shape is shape and rate is scale. ## The hazard connection The hazard rate — the instantaneous chance the event happens now given it has not yet — is constant for `k = 1`, increasing for `k > 1` and decreasing for `k < 1`, approaching lambda in both directions as time grows. This is why the gamma is the natural generalisation when a constant-hazard model is too rigid: it lets you express "the longer this has run, the more likely it is to finish" (`k > 1`) or the opposite (`k < 1`) with one extra parameter. ## The parameterisation trap There are two conventions in wide use: ``` rate form: shape k, rate lambda -> mean = k / lambda, variance = k / lambda^2 scale form: shape k, scale theta -> mean = k * theta, variance = k * theta^2, theta = 1 / lambda ``` Both describe the same distribution, and mixing them up inverts your answer by a factor of `lambda^2` in the variance. Say which convention you are using before you quote a number; interviewers notice when a candidate does this unprompted. ## The beta, for contrast The beta distribution is the other standard two-parameter continuous family, and it is easy to confuse with the gamma because both have two positive shape parameters. The difference is the support. A beta lives on the unit interval `[0, 1]`, which makes it the natural way to describe a proportion, a rate between 0 and 1, or an uncertain probability. Its density is proportional to `x^(alpha - 1) * (1 - x)^(beta - 1)`, its mean is `alpha / (alpha + beta)`, and larger `alpha + beta` concentrates it more tightly around that mean. `Beta(1, 1)` is exactly the uniform distribution on `[0, 1]`. So: gamma for an unbounded positive magnitude such as a duration, beta for a bounded fraction. ## What an interviewer is checking Three things, usually. First, that you can say what the two parameters do without reciting a density. Second, that you know `k = 1` collapses to the exponential — the single most useful special case. Third, that you handle the rate-versus-scale ambiguity carefully rather than confidently quoting the wrong moment. The topic itself is a differentiator rather than a screener, so a clear, correctly qualified answer counts for more than exhaustive detail.