How to measure attunement

Almost every signal worth studying is really two signals stacked together: something fast, and a slower shape riding on top of it. A voice has a pitch and a cadence. A flame has a flicker and a guttering. Footsteps have an impact and a gait. Usually the slow shape carries the relationship you are hunting, and the fast one is merely the carrier it is written on.

So the first move is not a comparison at all. It is a reduction: turn each recording into one slow trace of how much is happening. Do that to two unlike signals and they become the same kind of object, sampled at the same rate, and every method further down this page will work on them without knowing or caring where they came from.

There are really two choices here, and running them together is the mistake worth avoiding. The first is what each recording must become before it is a trace at all. A sound and an EEG already are one, but an ECG is not, since its information is in when beats arrive rather than how large they are, so it reduces to instantaneous rate.

The second choice is the one that is easy to miss: having got a trace, you still decide what about it to compare. Three options, and all three are ordinary:

• ·The trace itself, with nothing extracted.
• ·A frequency component over time: how much energy sits in one band, moment to moment. Band-pass, then take the envelope.
• ·Complexity over time: the scaling exponent recomputed in a sliding window, so that “how irregular is this” becomes a series rather than a single number.

Every one of those comes out the same shape: one value per moment. That is precisely why any of them can be handed to any method further down this page, and it is the vertical axis of the grid in step 07.

In one sentence
Choosing the observable is a modelling decision, not a preprocessing step. It encodes what you think matters about the signal, and applying the same transform to everything because it worked once is the fastest way to measure something real about a quantity nobody cares about.

Entrainment is not really about amplitudes agreeing. It is about a constant relationship in phase, one thing keeping time with another whatever the offset between them happens to be. Phase-locking value asks that and nothing else: is the gap the same now as it was a moment ago?

Coherence asks a stricter version, requiring phase and amplitude to agree together. That strictness has a real cost. On signals whose loudness wanders, which is most natural ones, it refuses to see relationships a phase-only measure finds easily.

Blind to · relationships with no rhythm to hold on to. Phase is only meaningful for something that oscillates.

In one sentence
Two signals can be perfectly locked and completely uncorrelated at the same time. Choosing the estimator that matches the kind of regularity a signal actually has is not a technicality. It is most of the analysis.

A correlation of zero means “no linearrelationship”. It does not mean the two are unrelated, and reading it that way discards every relationship that bends. Mutual information asks the general question instead, and does not care what shape the answer takes.

Blind to · direction and sign. It will tell you the two are bound together, never that more of one meant less of the other. Granger causality and transfer entropy belong to this family and add the direction back.

In one sentence
Generality is not free: this measure reports dependence that is not there whenever signals are smooth, which is exactly why the last part of this page exists.

The strangest family, and the one that best fits what attunement between a body and a place might actually be. Two things can be deeply related without ever lining up in time. What they share instead is the statistical structure of their fluctuations: how variability at fine scales relates to variability at coarse ones.

Blind to · timing, entirely. Two perfectly matched signals need never coincide, and this measure would not notice if they did.

In one sentence
Attunement need not mean simultaneity. Two people walking together do not synchronise step for step, yet the scaling of their gait variability converges.

Every estimator above returns a number greater than zero even when there is nothing there. Smooth signals correlate by accident. Autocorrelated signals manufacture mutual information out of nothing. A recording that looks generously long contains far fewer independent observations than it does samples.

The defence is to build a null on purpose. Destroy the relationship while keeping everything else intact: same spectrum, same autocorrelation, same distribution. Measure again, and repeat until you know the shape of nothing. Then ask where your real number falls in it.

In one sentence
A coupling value on its own is not a result. The result is where it sits relative to a null you built deliberately.

It is tempting to look for the most sensitive method and use it everywhere. The four demos above show why that instinct fails. Build a pair coupled purely in phase and the linear measure reports nothing. Build a pair coupled through a fold and the phase measure reports nothing. Match two signals in scaling alone and every timing-based measure returns zero, correctly, because there is no timing relationship to find.

In one sentence
The real work happens before any estimator runs: deciding what kind of relationship you think is there, and therefore what would count as evidence. A method chosen after seeing the data is not a measurement. It is a preference.

That is also the honest reason to show four rather than name a favourite. Attunement between a living thing and its surroundings almost certainly is not one phenomenon. A body may track a rhythm in one respect, ignore it in another, and share a texture with it in a third, at once, and all of it real.