the irreversibility of diffusion

Stir a drop of cream into coffee and it blends into an even tan. You will never see, however long you wait, the tan coffee spontaneously un-mix back into a sharp drop of cream. Diffusion has a built-in arrow of time: it runs one way only. The irreversibility of diffusion is the mathematical expression of that everyday certainty.

Mathematically, the heat equation u_t = k u_xx is not symmetric under reversing time. If you replace t by -t you get u_t = -k u_xx, the backward heat equation, which is a completely different and ill-posed problem. Forward in time, every Fourier mode decays like e^(-k n^2 t), so detail is steadily destroyed and many different rough pasts all flow toward similar smooth futures — the map from past to future loses information. Trying to invert it means amplifying the lost high-frequency detail by e^(+k n^2 t), which blows up uncontrollably: a microscopic change in the present would correspond to a wild change in the inferred past. This is why you cannot, in general, recover the initial temperature from a later measurement, even though the forward problem is perfectly deterministic.

This irreversibility is the same mathematical fact as the smoothing effect, viewed from the other end, and it has a thermodynamic cousin: diffusion increases entropy and runs toward equilibrium, consistent with the second law. It matters in practice for inverse problems — deblurring, thermal imaging, recovering a source from later data are all ill-posed because they are backward diffusion, and they can only be tamed by regularization (adding extra assumptions). Caveat: irreversibility here is a property of the macroscopic equation, not of the underlying microscopic dynamics, which are time-reversible; the arrow emerges statistically from coarse-graining many particles.