At low temperatures, however, the behaviour is different. The maximum density of water occurs at 4°C, so water at colder temperatures tends to rise to the surface, which is why lakes and rivers freeze from the top, rather than from the bottom. A layer of ice across the top of the container acts as an insulator and slows further heat loss.
Convection accelerates the early part of the cooling of warm or hot samples. The formation of a cold layer at the surface could slow the later part. These effects therefore may contribute to the Mpemba effect.
More possible artifacts
We should remember the potential importance of artifacts, of which we’ve mentioned one above. There are others: in a typical freezer, the air temperature and the radiation temperature both vary from one place to another, so similar samples at different points may have different rates of heat loss, particularly as they approach their final temperature. Different containers may them selves have different nucleation temperatures. Different preparations of samples may produce different nucleation properties. Nucleation is also probabilistic and history dependent, so repeated freezing of the same sample may produce different freezing times.
Now look at the graphs above, and see how close the two curves are after several times thalf. This proximity means that even small artifacts may also play a role in determining which sample freezes first.
Freezing point depression