Why can a probability density function take values greater than 1 when a probability cannot?

## A density is a rate For a continuous random variable X, the probability density function f is defined so that for any interval, ``` P(a <= X <= b) = area under f between a and b ``` Notice what the definition does *not* say: it never says `f(x)` is the probability of anything. `f(x)` is probability **per unit of x**. Its units are the reciprocal of x's units — if X is a latency in milliseconds, f is measured in probability per millisecond. A per-unit quantity has no reason to be capped at 1, in the same way that a speed of 30 metres per second is not "more than all the distance there is". Height becomes probability only after you multiply by a width. ## The worked example Let X be uniform on `[0, 0.5]`. All of the probability, which must total 1, is spread evenly across a width of 0.5. For a rectangle, height times width equals area, so ``` height = 1 / 0.5 = 2 ``` The density is `f(x) = 2` on `[0, 0.5]` and 0 elsewhere. Its value is double the largest probability that exists, and the distribution is entirely legitimate: f is non-negative everywhere, and the total area is `2 * 0.5 = 1`. Keep shrinking the support and the height keeps climbing. Uniform on `[0, 0.01]` has `f(x) = 100`. Uniform on `[0, 0.0001]` has `f(x) = 10000`. There is no ceiling at all: a density may even be unbounded near a point and remain valid, provided the area under it stays finite and equal to 1. ## The unit check that settles it Suppose latency X is measured in milliseconds and has some density `f_ms`. Re-express the same latency in seconds, `Y = X / 1000`. The probability of any physical event has not changed — it is the same latency — but the density has: ``` f_s(y) = 1000 * f_ms(1000 * y) ``` The per-second density is a thousand times larger than the per-millisecond one, because the same probability now sits in intervals that are numerically a thousand times narrower. If density values were probabilities, a change of units would have changed the probabilities. They are not, so it did not: the *areas* are identical. This is the argument to reach for when an interviewer pushes. ## The contrast with a PMF A discrete random variable is described by a probability mass function: ``` p(x) = P(X = x), 0 <= p(x) <= 1, sum over all x of p(x) = 1 ``` Here the values genuinely are probabilities. `p(x) > 1` is impossible, and so is a sum above 1. Almost every candidate who insists that `f(x) <= 1` is silently importing the PMF rules into the continuous case. Saying "PMF values are probabilities, PDF values are rates" out loud is usually enough to show you have the distinction. ## What the actual constraints are A function f is a valid probability density exactly when: 1. `f(x) >= 0` for every x — negative probability is meaningless, and a negative region would let some interval get negative probability. 2. The total area under f over the whole line equals 1 — all the probability is accounted for. That is the complete list. No boundedness, no continuity, no requirement to peak below 1. ## Where the confusion causes real damage - **Misreading a plotted density.** A peak at height 3.8 does not mean "38% of the mass is here". Mass is area; to get a probability you must integrate over a range, or read two CDF values and subtract. - **Comparing peak heights across variables with different units.** One curve looks taller purely because its x-axis is finer. Compare probabilities of comparable intervals instead. - **Panicking at a plot.** A density that spikes above 1 on a narrow, concentrated variable is normal, not evidence that something is broken. The general repair in all three cases is the same: stop reading heights, and read areas — or work with the CDF, whose values *are* probabilities and are always in `[0, 1]`. ## How to say it in an interview "A density is probability per unit of x, so it is only constrained to be non-negative and to integrate to 1 — not to stay below 1. Uniform on `[0, 0.5]` has density 2 everywhere and total area `2 * 0.5 = 1`. A PMF is the one whose values are actual probabilities and therefore capped at 1."