← Back to blog

Ensemble Equivalence Breaks at the Phase Transition

Your metrics agree until the system is about to fail. The average is the last thing to move.

Published June 2026 · 12 min read

In July 2023, two Danish climate scientists, Peter and Susanne Ditlevsen, published a paper in Nature Communications with a title that made headlines around the world: a warning that the Atlantic Meridional Overturning Circulation, the vast ocean conveyor that drags warm water north and keeps Britain and Scandinavia far milder than their latitude has any right to be, the current whose shutdown is the premise of The Day After Tomorrow, might be approaching collapse, possibly around the middle of this century. What made the claim remarkable was not that the average had moved. The mean strength of the circulation had not dramatically dropped. What had changed was something subtler and, it turns out, far more diagnostic: the system was fluctuating more, and it was recovering from those fluctuations more slowly. The variance was rising and the autocorrelation was climbing while the average sat there looking almost normal. The fluctuations were screaming; the mean was still quiet.

Now, the AMOC timing is genuinely contested, and we'll come back to exactly why, because the contest is part of the lesson, but the method the Ditlevsens used is not fringe. It is the most reliable early-warning signal that physics and dynamical-systems theory know for a system approaching a tipping point, and it carries a message that should reorganize how anyone who runs a production system thinks about their dashboards. The message is this: the average is the last thing to move. If you are waiting for the mean to cross a threshold, you are watching the slowest-moving hand on the clock.

The signature that shows up everywhere

The thing the Ditlevsens were tracking has a name, critical slowing down, and a canonical reference: a 2009 paper in Nature by Marten Scheffer and a long list of co-authors, titled “Early-warning signals for critical transitions.” Scheffer's group showed that as systems of wildly different kinds approach a bifurcation (a tipping point where a small further push flips them into a qualitatively different state) they exhibit the same generic warning signs. The variance of their fluctuations rises. The autocorrelation rises (each moment looks more like the one before it). And the system recovers from disturbances more and more slowly.

The reason to take this seriously is the breadth of the validation. The identical signature has been found ahead of ecosystem collapses, ahead of climate tipping points, ahead of financial market crashes, and ahead of epileptic seizures and asthma attacks. It is not a quirk of oceans. It is what critical systems do on the approach to the edge, and there's a clean mechanism for why it's so universal: near a fold-type tipping point, the mathematical “stiffness” that pulls a system back toward its stable state weakens toward zero, so the system wallows. It takes longer to return after each shove, its wanderings grow larger, and consecutive moments become more correlated. The mean can hold steady through all of this, right up until it can't. The fluctuations are the leading indicator; the average is the lagging one.

Why your monitoring is already inside this theorem

Here is why this is not a charming fact about ice ages but a fact about your pager, and the connection is unusually tight, tighter than an analogy, in the part that matters.

Statistical mechanics proves a deeply reassuring thing about large systems. There are several different ways to compute a system's average quantities, and for a big enough system they all give exactly the same answer. Physicists call it ensemble equivalence, and the engine behind it is simple: fluctuations shrink in relative size as one over the square root of the number of components. With 10²³ molecules in a gas (or a thousand shards, or a million requests an hour) the noise averages away, and the averages converge. This is why your dashboards feel trustworthy. Your two replicas track each other; your sampling methods agree; your correlated metrics stay correlated; your p50 latency is stable and reproducible. The law of large numbers is quietly doing its job underneath every chart.

And I want to be precise about what kind of claim this is, because it changes how much you should believe it. The statement “an outage is a phase transition” is a strong analogy, useful, but the kind of thing you'd check part by part. The statement “your metric is an ensemble average whose fluctuations scale as one over root-N” is not an analogy. It is the same statistics, full stop. Your p50 latency is a sample mean over N requests; the 1/√N law governs it literally, the very equation that governs the gas. So when your metrics agree at scale, that agreement isn't a comforting coincidence. It's a theorem.

Which is exactly what makes the exception so dangerous.

The agreement was never the information

Ensemble equivalence has a documented breakdown condition, and it is the worst one imaginable for anybody trying to catch trouble early. The equivalence fails (the different ways of measuring stop agreeing) precisely near phase transitions and in systems with long-range coupling, the regimes where fluctuations stop shrinking and instead become anomalously large. In the same physics where variance rises on the approach to a tipping point, the susceptibility of the system formally diverges, and the variance rises right along with it; the reassuring 1/√N suppression of noise switches off exactly when you need it most.

Translate that to your system and it inverts an intuition you've been running on for years. Your metrics agree right up until the system starts approaching a critical transition (a cascade, a saturation, a runaway) at which point the fluctuations explode, the averages stop agreeing, and the divergence between your metrics becomes the signal. The agreement was never the information. Agreement is the null state; it tells you only that nothing has started to fail yet, which is a far weaker and less comforting statement than “everything is healthy.” The information is in the onset of disagreement, and the variance, which the theory says moves first, is structurally the one statistic your mean-threshold alert is not watching.

So your default alerting strategy (page me when the average crosses a line) is, by construction, designed to fire late. It watches the slowest-moving statistic in the system, the one that goes last. By the time the mean crosses, the transition is already underway.

The cliff you hit every week

You don't have to look to the North Atlantic for this; you live inside a phase transition every time a service saturates. A queue at 80% utilization is fine. At 90%, fine. And then as utilization creeps toward 100%, the mean latency does not rise gently, it goes vertical, because the time a queue takes to drain blows up like 1/(1−ρ) as the load ρ approaches one. That near-vertical wall is a transition-like saturation (a real, finite system shows a sharp crossover rather than a mathematically perfect singularity, but the cliff is real enough to take down your service). And the tell is exactly the one the theory predicts: the variance of the queue length explodes before the mean latency goes vertical. The p99 starts thrashing while the p50 still looks serene, the tail moves first. Meanwhile two metrics that normally rise and fall together (request rate and CPU, or the latencies of two sibling shards) begin to decouple. That decoupling is ensemble equivalence breaking in real time, on your dashboard, minutes before the mean admits anything is wrong.

So the statistics worth watching are not the averages. They are: rising variance, the first mover; slowing recovery, how long the system takes to return to baseline after a small blip; and metric divergence, things that normally agree starting to disagree.

You can poke it on purpose

Critical slowing down hands you something better than patient waiting, and it's the most actionable idea in the whole theory. Because a system near its tipping point recovers from disturbances more slowly, you don't have to wait for a natural fluctuation to reveal your remaining margin, you can create one. Inject a small, controlled perturbation and time how long the system takes to spring back. A recovery time that is creeping upward, run over run, is resilience draining away beneath an average that still looks fine. This is chaos engineering pointed at a sharper question than usual: not “does the failover work?” but “how fast does the system recover, and is that number getting worse?” The relaxation time is a live resilience gauge, and you can read it whenever you like by giving the system a gentle, measured shove, which is a far better use of a controlled disruption than waiting for an uncontrolled one.

Know exactly when the variance lies

Now the part that separates a usable practice from a dangerous superstition, and it's the part most “watch the variance” advice quietly skips. These signals are leading indicators, not crystal balls, and the rigorous version of the idea is the one that says so plainly.

There are three hard limits. The first is false positives: variance and autocorrelation can rise without any transition following. This is not a footnote, it is precisely why the 2023 AMOC warning is contested rather than settled. Later work, including a 2024 study by van Westen and colleagues, showed that even the flagship ocean indicators can throw false alarms under some conditions; the flagship example of the method is also the flagship example of the method's fallibility. The second limit is worse, because it is silent: false negatives. Critical slowing down only precedes one family of transitions, the slow, fold-type kind, where a system gradually loses its resilience. A whole other family (the ones driven by a sudden large shock, or by being pushed too fast to keep up) comes with no warning at all. The variance stays calm, and then the system jumps. “Watch the variance” is blind to an entire class of catastrophe by construction. And the third limit is practical: these signals are data-hungry and noisy. Reliable detection needs a long history (the ocean work leaned on centuries of proxy data), and the empirical track record is genuinely mixed, the slowing-down signature was visible before the 1987 Black Monday crash, but it has been mixed-to-absent before more recent financial crises.

There is also a base-rate trap waiting for anyone who turns this into a hair-trigger. The overwhelming majority of variance spikes are not the onset of a transition; they're just noise, deploys, traffic, a backup job. An ungated “page me when variance rises” alert will bury you in false pages faster than any mean-threshold alert ever did, you'll have traded a detector that's reliably late for one that's reliably wrong. The escape is corroboration, not a single trigger: treat rising variance, rising autocorrelation, slowing recovery, and metric divergence as a panel that has to largely agree before anyone gets woken up, and keep your ordinary domain-specific signals (saturation percentage, error-budget burn rate, connection-pool headroom) right alongside them, because the boring engineered indicators sometimes beat the elegant generic one.

What to actually change

The shift this asks for is not “delete your threshold alerts.” Those still catch a great deal, and some failures (the shock-driven, no-warning kind) only ever announce themselves there. The shift is to stop drawing comfort from the wrong thing. Your metrics agreeing is not evidence your system is healthy; it is evidence your system has not started to fail yet, and the theorem underneath your monitoring guarantees that the agreement will persist right up to the edge and then break. So add the indicators the physics says move first:

• Plot the variance next to the mean and learn its normal range, so a rise is legible.
• Probe your recovery time deliberately and track whether it's drifting upward.
• Watch the metrics that are supposed to agree, and treat the moment they begin to disagree as the most important thing on the screen, because by the deepest result in this story, that divergence is what moves first when a stable system is quietly becoming an unstable one.

The averages will tell you the building is on fire. If you want the warning while there's still time to do something with it, you have to learn to read the variance, and to be honest about the days it's only the wind.

---

Sources: Peter Ditlevsen & Susanne Ditlevsen, “Warning of a forthcoming collapse of the Atlantic meridional overturning circulation,” Nature Communications (July 2023), the use of rising variance and autocorrelation (critical-slowing-down early-warning signals) to infer loss of resilience, with a contested mid-century central estimate. Marten Scheffer, Jordi Bascompte, William Brock, Victor Brovkin, Stephen Carpenter, Vasilis Dakos, Hermann Held, Egbert van Nes, Max Rietkerk & George Sugihara, “Early-warning signals for critical transitions,” Nature 461:53–59 (2009), rising variance, rising lag-1 autocorrelation, and slowing recovery as generic signatures, validated across ecosystems, climate, finance, and physiology. Ensemble equivalence and its breakdown near phase transitions / under long-range coupling (Campa, Dauxois & Ruffo, Physics Reports, 2009); the divergence of susceptibility and the fluctuation-dissipation relation linking variance to susceptibility; the queueing relaxation time ~ 1/(1−ρ) as a near-critical saturation. Honest limits: false positives in AMOC indicators (van Westen et al., 2024); the absence of critical slowing down before noise-induced and rate-induced transitions; the long-data requirement for reliable detection (≈centuries of proxy data, per Boulton et al., 2014); the mixed empirical record in finance (detectable before the 1987 crash, “limited reliability” for several recent crises). The identity-grade claim is “a metric is an ensemble average obeying 1/√N fluctuation scaling”; “an outage is a phase transition” is a strong analogy, and finite systems show sharp crossovers rather than strictly non-analytic singularities. Variance-based alerting carries its own base-rate problem (most spikes are not transitions), so the recommendation is a corroborating panel of leading indicators plus domain-specific signals, not a single hair-trigger alarm.