approximate identity

Convolution has no true identity element among ordinary functions — there is no function that, convolved with f, returns f unchanged (that role belongs to the Dirac delta, which is not a function). An approximate identity is the next best thing: a family of genuine functions that, as a parameter sharpens them, increasingly behaves like that ideal point-mass, so convolving with them recovers f in the limit.

A family (K_t) is an approximate identity if three conditions hold: each K_t has total integral 1; their L1 norms stay bounded; and as t shrinks, their mass concentrates near 0 (the integral of |K_t| away from any neighborhood of 0 tends to 0). The standard recipe is dilation: take one fixed bump phi with integral 1 and set K_t(x) = (1/t) phi(x/t), which squeezes the bump narrower and taller while keeping its area at 1.

The fundamental theorem: if (K_t) is an approximate identity, then f * K_t -> f as t -> 0, with the mode of convergence depending on f. For f continuous (with compact support) the convergence is uniform; for f in L^p the convergence is in the L^p norm; and at every point of continuity of an integrable f, the convergence is pointwise. The Fejér kernel, the Poisson kernel, and Gaussian mollifiers are the standard examples — this single principle underlies smoothing, density of smooth functions, and the very meaning of recovering a function from its transform.