error function

The bell curve is the most important shape in statistics, but the area under part of it cannot be written with elementary functions — there is no finite combination of polynomials, exponentials and trig functions whose derivative is e^{-x^2}. So mathematicians simply named the needed antiderivative. The error function, erf(x), is the running area under the Gaussian bell, scaled to settle at 1.

Precisely, erf(x) = (2 over square root of pi) times integral from 0 to x of e^{-t^2} dt. The prefactor 2/(square root of pi) is chosen so that erf(infinity) = 1 and erf(-infinity) = -1; the function is odd, with erf(0) = 0, and it rises steeply near the origin then flattens out. Although it has no elementary closed form, it is perfectly well-defined, infinitely differentiable, and easy to compute — its Taylor series is erf(x) = (2 over square root of pi) times (x minus x^3/3 plus x^5/10 minus ...), and tabulated or built-in routines evaluate it to full machine precision.

The error function is the cumulative normal probability in disguise: the probability that a standard normal variable lies within a range is a shifted, scaled erf, which is why it governs measurement error, signal-to-noise calculations, and quality control. It also solves the heat and diffusion equations — the temperature profile spreading from a sudden step is exactly an error function in the similarity variable x over square root of (4 D t). Saying the integral of e^{-x^2} is non-elementary does not mean it is uncomputable; it means we give the answer a name, just as we once named the logarithm.