Windowsill Lab · Run M11 · Statistical Physics

A glass that only freezes
at absolute zero

If you cool a magnet whose neighbours can't agree on what they want, does it ever settle — or does it only stop rearranging at absolute zero?

Painterly illustration: a shallow glass dish of metal filings settled into irregular frozen patches under warm lamplight, brass wires between them drawn taut and slack.
ILLUSTRATION (AI-PAINTED) - FROZEN PATCHES, EVERYWHERE DIFFERENT

An ordinary magnet is a crowd that agrees. Cool it and every piece lines up with its neighbours, because every piece wants the same thing. A spin glass is a crowd that can't agree. Before you start, you flip a coin for every pair of neighbours: this pair wants to match, that pair wants to differ. Then you freeze those demands in place forever. Now no arrangement can satisfy everyone — go around a square and you may find three "agree" bonds and one "disagree", and something has to give. Physicists call this frustration, and it is the whole subject.

The natural question is when such a system freezes. In three dimensions it genuinely does, at a real temperature — that experiment is M12, and it is still open. In two dimensions the answer is stranger and completely settled in theory: it never freezes at any temperature above zero. Two dimensions sits exactly at the lower critical dimension, the knife-edge below which no glass phase survives. So M11's correct answer is not a transition temperature. It is a slow, unbounded approach to one that lives at T = 0.

The signature is something called the overlap. Run two independent copies of the same frozen bond pattern — same demands, different starting scramble — and ask how much of the final picture the two copies share. That number is q. Do it hundreds of times, over hundreds of different bond patterns, and you get a distribution of q values. At high temperature the two copies share nothing and q piles up at zero. As you cool, the distribution widens — the copies start finding the same large frozen patches — and it keeps widening all the way down, with no peak ever splitting off, no sudden change, no transition. Just broadening.

The panel below runs two copies of one frustrated lattice side by side, shows you where they agree, and builds the overlap histogram live as you lower the temperature. The histogram spreading — not any number in it — is the result.

Live · two replicas · one frozen ±J bond pattern · checkerboard Metropolis L = 16 AS PUBLISHED — BUT ~1k SWEEPS PER REALIZATION, NOT 60k
T = 2.00

left · the agreement map (ember = the two copies chose the same thing here) with both replicas inset · right · top: P(q), the overlap histogram at this temperature; bottom: its width ⟨q²⟩ against T, filling in as you cool. Cool from 2.00 and watch the histogram spread without ever splitting.

sweeps 0 · bond realizations averaged 1 · samples in this histogram 0 · ⟨q²⟩ · |⟨q⟩| · frustrated plaquettes

The bond pattern rolls to a fresh one every few seconds and the histogram keeps accumulating — P(q) is a property of the disorder average, not of one frozen puzzle. Inside a single realization the sign of q is fixed by the quench, so one puzzle alone builds a lopsided lump. The published run averaged 64 of them, at 60,000 sweeps each; this panel rolls a new one about every second and gets roughly a thousand sweeps from each, so it reaches a narrower ⟨q²⟩ at the cold end than the receipt does. The broadening is the claim, not the endpoint.

0.011 → 0.306⟨q²⟩ as T falls 2.00 → 0.60, promoted run 2026-06-25 MEASURED
0.045max |⟨q⟩| — P(q) stays symmetric MEASURED
64 × 2disorder realizations × replicas
T = 0.6equilibration floor

The published numbers come from the promoted 2026-06-25 run: the width of the overlap distribution grows from 0.011 to 0.306 as the temperature falls from 2.00 to 0.60 — a factor of 27, monotone at all 15 steps, with no crossing and nothing splitting. An independent rerun on 2026-08-13 repeated the sweep and reached 0.335 instead. The two runs agree on the shape and differ on the endpoint, which is worth saying out loud rather than picking the number you prefer: what M11 measures is a trend, and a trend is exactly the kind of claim that survives a 10% disagreement about where it stops.

The Binder cumulant — the standard test for a transition — rises smoothly from 0.01 to 0.71 across the same sweep and never crosses, which is what a system whose critical point sits at absolute zero is supposed to do. The same measurement in three dimensions, where a real transition exists to be found, is M12, and it has not found it yet. The same ±J bond machinery, walked along a line where theory hands you the answer in closed form, is M14.

In two dimensions this system has no glass transition to find — its critical point sits at absolute zero, so what is measured is the approach, not an arrival. The three-dimensional case, which does have a real transition, is M12 and remains open. Single-spin updates cannot equilibrate this lattice below T ≈ 0.5–0.6, so the sweep stops at 0.6 and the claim is the trend toward zero, not the colder tail; one lattice size, so no finite-size collapse is claimed.