convolution

Convolution is a way of blending two functions into a third by sliding one across the other and accumulating the overlap. Intuitively, it is smearing: if g is a little blurring kernel, then f convolved with g is a blurred, smoothed version of f, where each output value is a weighted average of nearby input values. This is exactly what happens when a camera lens blurs a sharp image or a moving average smooths noisy data.

Formally, the convolution of f and g is (f * g)(x) = integral of f(y) g(x - y) dy. It is commutative, associative, and distributes over addition — it behaves like a multiplication. The smoothing is real: convolving with a smooth kernel makes the result at least as smooth as the kernel, regardless of how rough f is, because derivatives can be moved onto the smooth factor.

Its starring role in harmonic analysis comes from one identity: the Fourier transform turns convolution into ordinary pointwise multiplication. That is, the transform of f * g equals f-hat times g-hat. A tangled integral on one side becomes a plain product on the other — which is why filtering, differential equations, and probability (where the density of a sum of independent variables is a convolution) all become tractable in the frequency domain.