mpemba
hot or cold, who freezes first?
Does hot water freeze before cold water? Sometimes, by chance. And never more than half the time, unless the containers are unfair or the water remembers being hot.
The question
In 1963 a Tanzanian schoolboy, Erasto Mpemba, noticed that his ice cream mix froze faster when he put it in the freezer hot. He published the observation with the physicist Denis Osborne in 1969, and it has carried his name since. Aristotle, Bacon and Descartes had remarked on something similar. Sixty years of experiments have proposed evaporation, dissolved gas, convection and frost as causes, and careful measurements have since cleared most of them away. Burridge and Linden found no evidence that hot water cools to 0 °C any faster than cold water.
What survives in ordinary water is supercooling. Water usually cools well below 0 °C before it freezes, and the moment the first ice crystal forms is random. If that is the whole story, a single race between a hot cup and a cold one says very little. The thing to measure is a probability: how often the hot cup freezes first. This project asks what can be said about that number in general, and what it looks like under the best current model of nucleation.
The lottery
The model is Klimov and Finkelstein's. A cup cools by Newton's law towards the freezer temperature, and ice then nucleates as a random event whose rate climbs extremely steeply with supercooling: at about 7 K below freezing the wait is minutes, at 6 K it is weeks. Run it below. One race is a coin with a bias you cannot see. A thousand races show the bias.
Why one half is a ceiling
On its way down, the hot cup has to pass through the temperature the cold cup started at. Suppose heating leaves nothing behind, so that from that moment the hot cup behaves like a cold cup that started late. Suppose too that the two containers are interchangeable, for instance identical cups handed out at random. Then from the crossing point the race is between two equally matched contestants, one of which has a head start, and the hot cup wins at most half the time.
The argument uses no cooling law and no model of nucleation, and it survives every cup having its own random cooling rate. So a hot-first rate above one half, in a fair design, means heating changed something other than the temperature: water lost to evaporation, gas driven out, or frost melted under a hot tin so that it sits better on the shelf. In a design where the containers differ on purpose, a rate above one half needs no memory at all.
A window half a kelvin wide
Under the model the probability also has a floor, set by one number. Call Da the nucleation rate at the freezer temperature divided by the cooling rate. When Da is small, ice forms long after both cups have settled at the freezer temperature, and the hot cup wins nearly half the time. When Da is large, ice forms while the cups are still cooling, and the cold cup almost always gets there first. The effect lives around Da = 1, which for these parameters is 6.95 K of supercooling. Da changes by a factor of e for every 0.12 K, so the whole visible effect fits inside about 0.56 K of freezer temperature. A bigger cup moves the window only slightly: to 6.60 K for 35 mL and 6.25 K for a litre.
- Da
- 1.52
- P(hot first)
- 0.218
- floor
- 0.101
- mean wait at freezer temp
- 11 min
with differing containers here: 0.3 K → 0.36 · 0.5 K → 0.40 · 1 K → 0.43
Real containers differ by more than that. Different tubes nucleate at different temperatures, typically by half a kelvin to a kelvin, and Brownridge's vials by several. With a spread of 0.5 to 1 K the window stretches across several kelvin, and at −7.5 °C the hot cup wins about 27 to 37 percent of the time, where identical containers would give 0.4 percent. Where the effect shows up depends mostly on how much the containers differ, and only secondarily on water.
What the experiments show
Read against the ceiling, the published record sorts itself. Every hot-first rate above one half comes from a design in which the containers were not interchangeable, and the designs that come close to fair all report rates below one half.
| Study | Containers | Hot first |
|---|---|---|
| Brownridge 2011 | a pair of vials picked because their freezing points differ by about 5.5 °C | 28 / 28 |
| Burridge & Hallstadius 2020 | one beaker sanded inside, on purpose | 7 / 9 in the rough one, 0 / 7 in the smooth |
| Klimov & Finkelstein 2025 | closed vials, roughly fair | about a third near −7 °C |
| Janni et al. 2026 | tins and beakers, random placement | 46 / 176 samples |
| Burridge & Linden 2016 | fair | 0 / 3 |
Burridge and Hallstadius published their raw time series, which allow a sharper count than their paper gives. Over every pairing of a rough and a smooth beaker, whatever the starting temperatures, the rough beaker froze first in 18 of 20. The container decided the race and the starting temperature did not.
Janni, Botero Ampudia and Dahlberg have the largest recent data set, 176 samples in five setups. Simulating their exact protocol with no memory at all reproduces some setups and not others. The one that stands out is the only setup where tins stood on a metal baking sheet: 21 hot-first counts against a no-memory interval of 3 to 12. That fits a candidate for memory first raised in a 1969 letter to New Scientist: a hot tin melts the frost beneath it and improves its own contact with the cold. Their summary statistics cannot settle it, so for now it is a hypothesis that needs their per-sample data.
Where it stands
The paper is in draft, written for the American Journal of Physics. The per-sample data from Janni et al. have been requested, and the setup on the metal sheet will be tested directly if they arrive. The experiment that would settle the memory question is cheap: freeze the same sealed tube many times after hot and cold pre-treatment, in random order, so each container is only ever compared with itself. A home version is designed and has not been run yet.
Both figures run the paper's model. The race is a browser port that reproduces the paper's exact probabilities (0.392 at −6.8 °C, 0.218 at −7.0 °C, 0.067 at −7.2 °C) to within sampling error; the window is computed by the paper's own Python.