PhronesisThe kernel
Grant the physics first, because it is settled.
A system resting in a valley — a basin of attraction — answers a small push by rolling back to the floor. How fast it rolls back depends on how steep the walls are. Steep walls, quick return. As the system nears a tipping point — a bifurcation, where that valley flattens and gives way to another — the walls go slack. The same push takes longer to recover from. This is critical slowing down, and it is not a metaphor; it is a mathematical property of a system approaching such a transition.
Scheffer and colleagues gathered the signals in one review (Scheffer et al., 2009). Their own summary sentence names three: “the phenomenon of critical slowing down leads to… slower recovery from perturbations, increased autocorrelation and increased variance” as a system approaches a bifurcation. The slowing return is the root of all three. A system that recovers sluggishly carries more of its past state into its present — that is what rising lag-1 autocorrelation measures, and it is why the paper reads that autocorrelation as, in its words, the slowness of recovery made visible. The wandering that a flattening basin permits is what rising variance measures. One fact, three faces.
The range is the point. The same signature has been sought in lakes tipping to turbid, in ice-core and climate records before abrupt shifts, in physiological collapse — a generic early-warning signal, not a quirk of one system. Scheffer frames the three markers as possible consequences of an approaching transition — possible, not guaranteed. Grant it at exactly that strength.
The reading
Now the framework's turn, and it is offered as a proposal — a way of reading the kernel, not a change to it.
The framework already carries a picture of a system resting at some ground state, with displacement ξ measuring the gap from that ground, and D(ξ) the pressure to return (the sheet). Critical slowing down is a statement about the shape of the valley around that ground. A deep basin has steep walls: a knock is answered fast, the return is crisp, resilience is high. A shallow basin has slack walls: the same knock lingers, the return drags, resilience is low. The proposal is a single sentence — recovery time is a readout of basin depth. How fast you come back from a small perturbation reports how deep your ground state currently sits, without your having to measure the depth directly.
That is the appeal of it. Basin depth is hidden; you cannot see the valley walls. But you can knock the system and time the return, and the return time carries the depth out to where it can be read. The signal is cheap where the structure is expensive.
A steep valley answers fast. A flat one lets the knock linger. Recovery time reads the wall you cannot see.
Van de Leemput and colleagues carried the same dynamical-systems picture into mood (van de Leemput et al., 2014). Read depressed and healthy states as alternative stable basins with a tipping point between them, and the resilience markers should rise as a person nears the crossing. In their data, “approaching transitions between depressed and normal states correlates with elevated temporal autocorrelation, variance, and correlation between emotions.” The slowing return, read from repeated self-reports of feeling over time. That is the reading applied to a person — and exactly where it needs its limits said out loud.
The limit
Early-warning signals are real and imperfect, and the note is only honest if it says both.
They do not give timing. A rising signal says the basin is shallowing; it does not say when, or whether, the system will actually tip. The clock is not in the signal.
They demand good data. The markers need long, clean, well-sampled time series. Short or noisy records give unstable estimates — a trend that vanishes when you resample.
They carry error rates that are often left unquantified. Boettiger and Hastings put the caution precisely: proposed detection methods “hardly ever characterize their expected error rates,” and those rates “can be quite severe for common indicators even under favourable assumptions” (Boettiger & Hastings, 2012). False alarms and missed warnings are both live; the null and the tipping distributions overlap more than the tidy examples suggest. Testing a method only on systems already known to have tipped stacks the deck — the same authors call that the prosecutor's fallacy. Reliability across fields stays contested, and this is mainstream caution, not a fringe complaint.
The mood application inherits every one of these and adds its own. Van de Leemput's evidence is observational — experience-sampling and twin-registry data, not a controlled experiment — and it drew a published challenge: Bos and De Jonge argued the idea still needs empirical proof, and the authors' reply conceded the support rests partly on between-person rather than within-person prospective data. Promising and preliminary. Not settled prediction.
So for a person, hold the reading at the right size. Recovery time is a noticing signal, not a prediction and not a diagnosis. That a low mood is taking longer to lift than it used to is worth noticing — it is information about the shape of the ground you are standing on. It is not a verdict about what comes next, and nothing here is therapy or diagnosis. The limits page holds that line.
The claim, at its true size
Near a tipping point, a system returns more slowly from small perturbations — critical slowing down — and the slowing shows up as longer recovery time, higher variance, higher autocorrelation. That is established and cited. Reading the slowing return as a resilience gauge — recovery time as a readout of how deep the ground currently sits — is the proposal, offered to be argued with. And the whole thing, applied to a person, is an analogy made honest: a way to notice that return is dragging, not an instrument that predicts a fall.
A live version sits at /field/recovery: knock it, watch it come back, watch the return slow as the basin flattens.
Kin to Recovery — the return timed live — and Settle, the return itself; alongside The Stress Cloud and Anxiety as the Signature of Displacement for the framework's other near-critical readings, and Blame the Grain.
Rests on: Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, Dakos V, Held H, van Nes EH, Rietkerk M, Sugihara G, “Early-warning signals for critical transitions,” Nature 461(7260), 53–59 (2009), doi:10.1038/nature08227 — critical slowing down and its three markers (slower recovery, rising variance, rising lag-1 autocorrelation), framed there as possible signals of an approaching bifurcation; van de Leemput IA, Wichers M, Cramer AOJ, Borsboom D, Tuerlinckx F, Kuppens P, van Nes EH, Viechtbauer W, Giltay EJ, Aggen SH, Derom C, Jacobs N, Kendler KS, van der Maas HLJ, Neale MC, Peeters F, Thiery E, Zachar P, Scheffer M, “Critical slowing down as early warning for the onset and termination of depression,” PNAS 111(1), 87–92 (2014), doi:10.1073/pnas.1312114110 — the mood application, observational and contested: see Bos & De Jonge, PNAS (2014), doi:10.1073/pnas.1323672111, and the authors' Reply, doi:10.1073/pnas.1323835111; Boettiger C & Hastings A, “Quantifying limits to detection of early warning for critical transitions,” J R Soc Interface 9(75), 2527–2539 (2012), doi:10.1098/rsif.2012.0125, with the companion “Early warning signals and the prosecutor's fallacy,” Proc R Soc B 279(1748), 4734–4739 (2012), and Dakos V et al., “Robustness of variance and autocorrelation as indicators of critical slowing down,” Ecology 93(2), 264–271 (2012), doi:10.1890/11-0889.1 — the limits: no timing, need for long clean series, non-trivial and often unquantified error rates, reliability still contested. These results are established prior art and cited as such. What is proposed is only the reading of recovery time as a resilience readout in displacement terms — offered to be argued with, and offered for a person as a noticing signal, not a diagnosis or a prediction.