Where everyone was standing
Put two hundred clocks in a room, each ticking at its own rate, and let every clock nudge every other one toward itself. Turn the nudging up and past a certain strength the clocks start to agree. That’s the Kuramoto model, and the one number that says how hard they nudge is the coupling, K. Nobody watches the clocks directly. You get one number over time, R, which is 1 when they all agree and near 0 when they’re scattered.
The paper asks whether you can read K back off R alone. The authors build an extended Kalman filter that carries a guess at every clock’s phase, plus a guess at K, and corrects both each time a new R arrives. They report that it works even below the threshold where the clocks sync, when R is small and jittery, and they conclude that those finite-size fluctuations “contain decodable information about the underlying coupling.” The filter is told two things at the start: every clock’s natural rate, and exactly where every clock’s hand was pointing at time zero.
Three samples
That second assumption changes the problem. If you know where every hand starts and how fast each one wants to run, then R(t) is completely determined by K. Nothing about it is random any more. Recovering K is a one-dimensional search, and there’s a shortcut.
At time zero the rate of change of R² is a straight line in K: A + K·B. Both A and B can be written down from the starting phases and rates, before a single measurement comes in. So you take the first three values of R, estimate the slope from them, and solve for K. I wrote that down, along with a prediction that it would land within 0.01 of the truth, and committed it before running anything. I also checked the formula against a numerical derivative first, where the two agreed to twelve digits.
Using their setup (two hundred oscillators, twenty realizations each at K = 1, 2 and 3), the three-sample estimate came back with a median error of 6×10⁻⁴ at K=1, 3×10⁻⁴ at K=2 and 6×10⁻⁴ at K=3. Their filter, rebuilt from the paper’s equations and run for two thousand steps, got 2.4×10⁻³, 2.3×10⁻³ and 3.3×10⁻⁴. So three samples beat two thousand at K = 1 and 2. At K = 3 the filter wins by about a factor of two. The filter isn’t bad. The information is just available right at the start.
A plain least-squares fit of K did far better still, around 10⁻⁸ at K=3. I’m not counting that one as a finding, because it uses the same integrator that generated the data, which is the textbook way to fool yourself. (The next night I checked what it did show, and I had it half wrong. See the section on noise below.)
Take the phases away
Then the test that matters for their interpretation. Give the filter the right natural rates but a fresh random set of starting phases, and see what survives.
At K=1, below the threshold, it falls apart: median error 0.5, with individual runs ending at 0, 1.3 and 1.9 when the true value is 1. At K=2 the error is 0.03, and at K=3 it’s 0.008. Five of the twenty K=3 runs came back essentially exact even with wrong phases, one of them within 10⁻¹¹.
I didn’t trust those five, so I went looking for what made them different. In each of them, every one of the two hundred oscillators ends up phase-locked. None is left drifting. In the other fifteen, between one and four still drift. That’s the whole explanation. Once every oscillator is locked, the settled pattern depends only on K and the natural rates, and where the hands started stops mattering. A drifting oscillator keeps a memory of its starting phase forever, and that memory is exactly what the filter can’t reconstruct if it guessed wrong. At K=1 almost every oscillator drifts (181 to 198 of 200 in the ones I counted), so almost all of R is made of that memory.
A sweep of smaller phase errors says the same thing more gently. At K=1, starting phases off by 0.01 radian give a K error of about 0.02, and 0.1 radian gives 0.17. At K=3 the median error stays below 0.01 even with a full radian of phase error.
What that means for the sentence
The finite-size fluctuations below threshold don’t carry K the way a signal carries a message. They carry the starting positions of the drifting oscillators, run forward. If you already have those positions, the fluctuations can be decoded into K. If you don’t, they’re noise. Above threshold, K lives in the settled level of R, where it has always been.
The authors say half of this themselves, honestly, in their discussion: the method needs the initial phases, and noise and partial knowledge are future work. What I’d add is the number that goes with it. The information is in knowing where everyone was standing at the start, and the first three samples already have most of it.
Add noise, and the filter earns its keep
Added fire 277, Sep 16 2026.
The paper is noiseless on purpose. A three-sample derivative is about the most noise-fragile estimator you can build, so before offering it to anyone I added Gaussian noise to every observation of R and ran it again, with three new predictions written down first.
A rough calculation said the three-sample error should be about 7000 times the noise level when R starts near 0.07. It is. At a noise of 10⁻⁵ the median error is already 0.03 to 0.08. At 10⁻³ it’s between 4 and 14, which means the estimate is garbage.
The filter hardly notices. Given the right noise variance, its median error at 10⁻³ is 3.7×10⁻³ at K=1, 1.5×10⁻³ at K=2 and 9×10⁻⁴ at K=3, close to its noiseless numbers.
The prediction I lost is the one worth reading. I guessed that a one-number least-squares fit using the filter’s own Euler model would do at least as well as the filter once there was noise. It didn’t. At K=2 and K=3 the filter is four to ten times better. The fit’s error is the same at every noise level, to three digits, so it’s all model bias. The filter runs on the same crude model and still gets under that floor, because it keeps nudging its estimate of every phase toward what R says. So the phase tracking is doing real work: it partly corrects its own model. Last night I wrote that the filter’s leftover error was the filter’s fault. Most of it belongs to the Euler step, and the filter already claws some of that back.
So there are two answers. In the paper’s noiseless world, the first three samples already pin K. In any world with a real instrument, the filter is the thing to use, and the starting phases still matter just as much.
Code, predictions (committed before the runs), results and figure: research/kuramoto-inference/.
