Mpemba Effect Explained: Can Hot Water Freeze Faster Than Cold?

At Magamba Secondary School in Tanzania, in 1963, a thirteen-year-old named Erasto Mpemba was racing the clock, not the thermometer. His class was making ice cream  boil the milk, stir in the sugar, let it cool to room temperature, then claim a slot in the school refrigerator's single ice tray. Slots were scarce. Another boy, worried about being shut out, skipped the boiling step entirely and poured his milk in cold. Mpemba, further behind and unwilling to lose his place too, did the opposite: he shoved his mixture into the freezer while it was still hot from the stove. Ninety minutes later the two boys checked. Mpemba's had set into ice cream. The other boy's was still liquid.

He asked his physics teacher to explain it. The answer, as Mpemba later recounted it himself, was blunt: "You were confused, that cannot happen." He let it go, for a while. Years later, at Mkwawa Secondary School in Iringa, a visiting physicist named Denis Osborne from the University College in Dar es Salaam gave a lecture, and Mpemba raised his hand and asked the same question again, this time in sharper form: two identical containers of water, one at 35°C, one at 100°C, placed in a freezer  why does the hotter one sometimes win? His classmates laughed. Osborne didn't. He went back to his lab, ran the experiment with water instead of milk, and got a result close enough to Mpemba's to take seriously  in one run, water starting near 20°C took roughly 100 minutes to begin freezing, while water starting near 80°C began in about 40. In 1969 the two of them published their findings together in Physics Education, under a title that has aged well: "Cool?"

What they'd stumbled into isn't a case of hot water beating cold water on some universal, reliable basis. It never has been. What the Mpemba effect actually names is narrower and stranger: the observation that, under specific and fairly delicate conditions, an initially hotter sample of water can reach some marker of "frozen" before an initially colder one, even though nothing about the process breaks a single law of thermodynamics. The heat still flows downhill, from warm to cold, exactly as it must. What's flexible is everything else about the journey.

Notice, too, that "reach some marker of frozen" is doing a lot of work in that sentence. A sample can be said to freeze when it first touches 0°C, when the first visible ice crystals appear (a moment called recalescence, since the release of latent heat during nucleation actually kicks the local temperature back up toward 0°C for an instant), or when the entire volume has gone solid. These are three different races. A sample that supercools dramatically before nucleating might reach 0°C first and still lose the race to "first ice." One that freezes into a mushy, unevenly cooled block might beat a rival to "first ice" and still lose the race to "fully solid." Much of the confusion in over a century of arguing about this phenomenon traces back to different experimenters, often unknowingly, timing different races.

The puzzle is worth taking seriously precisely because the naive physics argument against it is so clean. Heat loss from an object roughly follows Newton's law of cooling: the rate at which a sample sheds heat is proportional to the temperature difference between it and its surroundings. A pot of water at 80°C sitting in a −18°C freezer does lose heat faster, moment to moment, than one sitting at 20°C  the gap driving the flow is larger. But the hot water also has more total heat to shed before it gets anywhere near freezing. Run the arithmetic forward with fixed properties and a fixed volume of water, and the colder sample should still cross the 0°C line first, because its head start outweighs the hot sample's temporary advantage in cooling rate. This is the correct baseline expectation. The interesting physics lives entirely in the ways a real, physical sample of water refuses to hold its properties fixed while it cools.

None of this is even new territory. Aristotle noted, in his Meteorology, that people in hot climates cooled water quickly by exposing it to the sun rather than leaving it in shade  an observation he used, oddly, to explain seasonal rainfall patterns rather than kitchen physics, but it's the same underlying claim. Francis Bacon wrote in the Novum Organum that "water a little warmed is more easily frozen than that which is quite cold." Descartes touched on something similar in Les Météores. In 1775 the Scottish chemist Joseph Black went further and actually tested it, comparing previously boiled water against unboiled water and finding the boiled sample froze faster even once he'd controlled for evaporation  though he also found that stirring the unboiled sample erased the difference entirely, an early hint that the effect is fragile and easily disrupted by how the experiment is handled.

Evaporation is usually the first explanation people reach for, and it isn't wrong so much as incomplete. Hotter water evaporates faster, which does two useful things: it removes mass, so there's simply less water left to freeze, and evaporation itself is an endothermic process, carrying latent heat away from the sample as molecules escape the surface. In an open, uncovered container, over the span of an hour or two, this can meaningfully shave time off a hot sample's cooling curve. But it cannot be the whole explanation. Cover the container, and evaporation is largely shut off, yet Mpemba-like observations have still been reported in sealed setups. And the mass lost to evaporation over typical freezer timescales is usually just a few percent of the total  real, but rarely large enough on its own to flip a race that, on paper, cold water should win by a wide margin.

Convection tells a subtler story. Water in a container isn't a single uniform temperature; warmer water near the surface is less dense and rises, cooler water sinks, and the resulting circulation constantly cycles warm interior water up to where it can lose heat to the air or the freezer wall. This works cleanly above about 4°C, which is where liquid water is at its densest. Below that threshold, water's behavior inverts  as it cools further toward 0°C, it becomes slightly less dense again, not more. That means a surface layer that has cooled close to freezing can become buoyant, sit on top, and effectively insulate the warmer bulk beneath it from further heat loss. A sample that starts cold spends more of its total cooling time in this awkward, self-insulating regime near 4°C. A sample that starts hot spends more of its time in the vigorously convecting range above it, which can, under the right container geometry, let it shed heat unusually efficiently before it ever nears the freezing point. Henry Burridge and Paul Linden, in a widely cited 2016 study, found that when they tracked this convective behavior carefully using the Rayleigh number (a dimensionless measure of how vigorously a fluid convects), the time samples took to reach 0°C scaled predictably with it  and within that scaling, colder water reached the freezing point first every time, no exceptions.

Dissolved substances add another layer. Heating water drives off dissolved atmospheric gases, chiefly nitrogen, since gas solubility drops as temperature rises. In water with significant mineral content, the same principle causes dissolved calcium and magnesium bicarbonate  the stuff behind limescale, or "kettle fur"  to precipitate out as the water is heated, converting Ca(HCO3)2 into solid CaCO3 that sticks to the container wall. Physicist Jonathan Katz argued in 2006 that whatever solutes remain in never-heated water get concentrated ahead of the advancing ice front as the sample freezes, depressing the local freezing point there and reducing the temperature gradient that drives heat out of the freezing water, slowing the whole process down. It's a coherent mechanism, and it makes a testable prediction: the effect should be stronger in hard water and vanish in distilled water. But other experimenters have used water pre-boiled to strip out dissolved gases in both the hot and cold samples and still reported a Mpemba-like result, which means dissolved solutes can't be a universal requirement, whatever role they play in specific setups.

Supercooling might be the deepest rabbit hole of all, and it's where the story stops being about heat transfer and starts being about the physics of how a solid actually begins to exist. Water doesn't have to freeze the instant it touches 0°C. In the absence of a suitable nucleation site  a scratch, a dust particle, a dissolved impurity, some irregularity that lets water molecules lock into the beginnings of a crystal lattice  liquid water can persist several degrees below its nominal freezing point, sometimes far below it. David Auerbach, testing this directly in 1995, found some samples supercooling to between −6°C and −18°C before ice finally appeared. He also found that his initially hot samples tended to supercool less than his initially cold ones, meaning they began forming visible ice closer to 0°C. That sounds like an explanation, until you notice the complication: because the hot samples still had farther to travel to reach 0°C in the first place, the reduced supercooling didn't reliably translate into an earlier appearance of ice overall, in his data. Supercooling matters enormously, but which direction it pushes the race is not settled by temperature alone.

There's a further wrinkle here that undercuts the whole idea of "identical containers" doing identical things. James Brownridge, running repeated trials with the same vessels, found that a given container tends to nucleate ice at a strikingly consistent supercooled temperature every time  as if the nucleation point were a fixed property of that specific piece of glass, tied to whatever microscopic defect happens to sit on its surface. Two containers from the same production batch, same shape, same volume, poured from the same tap, can behave completely differently near the freezing point simply because one has a nucleation site the other lacks. Much of what looks like unpredictability in a kitchen freezer experiment may have less to do with temperature history than with which glass you happened to grab.

Put all of this together and the disagreements running through six decades of Mpemba literature start to make more sense. One study measures time to 0°C, sidestepping supercooling; another measures time to visible ice, which supercooling dominates. One uses open beakers where evaporation matters; another uses sealed containers where it can't. One controls for container geometry with obsessive care; another treats "same size cup" as close enough. Burridge and Linden's 2016 paper is the most pointed challenge on record  after controlling convection, geometry, and measurement technique as tightly as they could, they found no reliable case of hotter water reaching 0°C first, and suggested that many earlier "confirmations" may trace back to a stray thermometer placed where a convective plume passed unpredictably, or to samples that weren't as identical as claimed.

Yet the phenomenon hasn't been laid to rest, because two physicists took the question somewhere water can't complicate it. In 2017, Zhiyue Lu and Oren Raz showed mathematically, using the general theory of how systems governed by Markov processes relax toward equilibrium, that an initial state which looks "hotter" by simple temperature can, in a more abstract statistical sense, actually sit closer to the eventual equilibrium state  and can therefore relax toward it faster, sometimes even exponentially faster, a case they termed the strong Mpemba effect. In 2020, Avinash Kumar and John Bechhoefer at Simon Fraser University tested the prediction not with water but with a single glass bead 1.5 micrometers across, suspended in a fluid and held in a laser-sculpted potential energy landscape that let them dial in a precise effective "temperature" and watch it relax. Under carefully chosen conditions, a bead started hotter reached thermal equilibrium with its surroundings exponentially faster than one started at an intermediate temperature  a clean, repeatable, quantitatively predicted version of exactly the crossover Mpemba stumbled into by accident with a bowl of milk. A related demonstration turned up the same year in granular fluids, systems of vibrating particles that "cool" through collisions rather than molecular heat transfer, reinforcing that this crossover behavior isn't a quirk unique to H2O.

So the effect is real, in the sense that matters most to a physicist: it's a legitimate, well-understood feature of how systems relax toward equilibrium, proven cleanly in systems built specifically to remove water's complications. What remains genuinely unresolved is whether that same crossover is what anyone is actually watching happen in an ordinary glass of tap water set beside another in a home freezer. Ice formation is a phase transition, gated by a stochastic nucleation event on a microscopic surface defect nobody can see or characterize in advance  and no theory yet lets you predict, for a specific container pulled from a specific cupboard, exactly when that gate will open.