A slope is cheap
Resting-state scalp EEG shows power spectral density falling as 1/f^α with α typically between *1.0 and 3.0* (measured; interpretation contested), depending on electrode site, arousal state and age. Heart-rate variability, analysed by detrended fluctuation analysis, gives a short-term scaling exponent (DFA α1) that sits near *1.0* — canonical 1/f — in healthy rest and drifts toward *0.5*, the signature of uncorrelated white noise, during heavy exertion or in congestive heart failure *(measured)*. Both numbers are routinely read the same way in the complexity literature: a system living near a critical point, its 1/f spectrum the fingerprint of long-range temporal correlations that emerge from short-range interactions tuned to the edge of a phase transition. That reading survives on an unexamined assumption — that 1/f spectra are hard to produce. They are not.
The thesis here is specific: a 1/f spectrum is what you get, generically and without any tuning, whenever a system contains a sufficiently broad distribution of independent relaxation rates whose characteristic frequencies overlap across the band you happen to be measuring. No criticality, no coupling between elements, no emergent collective mode is required *(established — derivation below)*. That demotes "we found a 1/f slope" from a strong claim about self-organisation to a weak claim about heterogeneity in time constants, and it raises the evidentiary bar for anyone wanting to claim genuine criticality: not a slope on a log-log plot, but a finite-size scaling collapse, with exponents that hold as you vary system size or distance from the putative control parameter.
The arithmetic nobody checks
The superposition-of-Lorentzians argument is old and purely mathematical: sum enough simple relaxation processes, each with its own single exponential decay, and the aggregate spectrum becomes a power law over whatever frequency range the individual relaxation rates span *(established — standard derivation; the fact sheet's framing of this is explicit)*. The qualitative point is important for interpreting both EEG and cardiac aperiodic exponents, because neural tissue and the cardiovascular system are both, trivially, collections of elements with a spread of membrane time constants, ion channel kinetics, or baroreflex loop delays — exactly the raw material the derivation requires.
The integral: from a bank of Lorentzians to exact 1/f
A single relaxation process with time constant τ (an Ornstein-Uhlenbeck-type decay, or equivalently an RC low-pass) has power spectral density S(f;τ) = 4τ / (1 + (2πfτ)²). This is a Lorentzian: flat below 1/τ, falling as f⁻² above it.
Now suppose the system is a superposition of many such independent processes, with relaxation times distributed according to a density p(τ) ∝ 1/τ over a range [τ_min, τ_max] — that is, uniform in log τ, the natural "no preferred scale" distribution. The aggregate spectrum is the integral
S(f) = ∫_{τ_min}^{τ_max} p(τ) S(f;τ) dτ = ∫_{τ_min}^{τ_max} (1/τ) · 4τ/(1+(2πfτ)²) dτ = 4 ∫_{τ_min}^{τ_max} dτ / (1+(2πfτ)²).
Substitute x = 2πfτ, dτ = dx/(2πf):
S(f) = (4/2πf) ∫_{2πfτ_min}^{2πfτ_max} dx/(1+x²) = (2/πf) [ arctan(2πfτ_max) − arctan(2πfτ_min) ].
For any f in the window where τ_min ≪ 1/(2πf) ≪ τ_max — i.e., well inside the overlap band of the relaxation rates — the first arctan saturates to π/2 and the second goes to 0. So S(f) → (2/πf)·(π/2) = 1/f, exactly, with slope 1, independent of the width of [τ_min, τ_max] beyond the requirement that it be wide enough to contain the band of interest.
Nothing in this derivation invokes coupling between the relaxation processes, feedback, or a tuned control parameter. It only requires heterogeneity in rates, spanning at least the measured frequency band. Different distributions of τ (not log-uniform) give other exponents; the same machinery can produce 1/f^α for α away from 1 depending on p(τ)'s shape near the band edges.
This matters directly for the two headline numbers above. EEG aperiodic exponents of *1.0–3.0* and the HRV DFA α1 range of *0.5–1.0* are both perfectly consistent with heterogeneous, uncoupled relaxation processes — distributed ion-channel time constants in cortex, distributed autonomic reflex delays in the heart — with no appeal to self-organised criticality needed *(speculative — flagged, as applied specifically to these two systems, since the fact sheet does not establish that this mechanism, rather than coupling-driven criticality, is in fact what generates either signal)*. The derivation shows the mechanism is sufficient to produce the observed slopes; it does not show it is the operating mechanism in cortex or myocardium. That distinction is the whole article.
Kuramoto, done properly
The model that critical-dynamics advocates usually invoke as the contrasting "real" mechanism is the Kuramoto model: N phase oscillators with natural frequencies ω_i drawn from a distribution g(ω), coupled all-to-all,
dθ_i/dt = ω_i + (K/N) Σ_j sin(θ_j − θ_i),
with global coherence measured by the order parameter r e^{iψ} = (1/N) Σ_j e^{iθ_j}, so r runs from 0 (incoherent, phases uniformly scattered) to 1 (fully locked) *(established)*. For a unimodal, symmetric g(ω), the model has an exact mean-field result: synchronisation onsets at a critical coupling Kc = 2/(π g(0)), where g(0) is the density of natural frequencies at the population mean *(established — standard analytic result, Acebrón et al. 2005)*. Below Kc, r = 0 in the N→∞ limit; above it, r rises continuously from zero, following r ∝ √(K−Kc) near onset — a genuine bifurcation with a well-defined critical exponent, derivable from the self-consistency equation rather than fitted to data.
This is what a real phase transition looks like inside a model: a non-analyticity in a thermodynamic-limit observable as a function of a control parameter, with an exponent that is a property of the equations, not of the particular realisation. It is categorically different from "the oscillators synchronised," which can happen smoothly, at finite N, with no singular point at all — the crossover sharpens into a true transition only as N → ∞ *(model-dependent)*. Finite populations show a smeared version of the transition whose sharpness grows with N; whether a finite biological population is "close enough" to that limit for the distinction to be empirically moot is an open question, not a settled one.
Physical synchronisation transitions — the Millennium Bridge's footfall lock-in, arrays of coupled metronomes, electrochemical oscillator arrays — demonstrate that the Kuramoto-type order-parameter growth is not a mathematical curiosity; sharp synchrony onset happens in real coupled-oscillator hardware *(measured)*. But demonstrating synchronisation is not the same as demonstrating that the synchronisation is scale-free or that it generates 1/f statistics; the Kuramoto transition near Kc produces its own characteristic critical slowing and fluctuation signatures, distinct from, and not automatically equivalent to, the superposition-of-Lorentzians 1/f described above.
"Near critical" is not the same claim as "scale-free"
A system poised near a genuine critical point — near Kuramoto's Kc, near a percolation threshold, near a sandpile's marginal stability — exhibits diverging correlation length or time as the control parameter approaches the critical value, and the resulting fluctuations are scale-free with exponents fixed by the transition's universality class. A system merely containing a broad spread of independent time constants also produces scale-free-looking spectra, across whatever band the rates span, with no approach to any special point at all. Both mechanisms can hand you the same log-log slope over the same decade of frequencies. The slope alone cannot distinguish them.
What distinguishes them is behaviour as you change something: system size, coupling strength, or distance from the hypothesised critical point. Genuine criticality predicts that rescaling data by the correct power-law functions of system size collapses curves from different sizes onto one universal function — finite-size scaling collapse. A superposition mechanism predicts nothing of the sort; there is no size-dependent exponent to collapse, because there was never a diverging correlation length driving the spectrum in the first place. The fact sheet's own framing is precise on this point: a power-law distribution is necessary but insufficient for a criticality claim, and genuine criticality additionally requires finite-size scaling collapse and universal exponents *(established, as a methodological standard)*. Almost none of the EEG or HRV literature reporting 1/f slopes performs this test, because doing so requires systematically varying something — electrode array extent, epoch length used as an analogue of system size, or a genuine control parameter like arousal or workload — in a way that most single-session recordings do not support.
The estimator minefield
Even granting that a real power-law-like spectrum is present, the exponent reported is only as trustworthy as the estimator that produced it, and several common failure modes routinely manufacture or distort slopes *(established, as methodological risks; their prevalence in any specific published estimate is not something this fact sheet can quantify)*:
- DFA results are sensitive to the polynomial order used for detrending; the wrong order can inject an artificial slope that has nothing to do with the underlying correlation structure.
- 1/f and DFA analyses assume quasi-stationarity; running them on genuinely non-stationary segments without proper windowing produces spurious scaling exponents that look like real long-range correlation.
- Fitting a straight line to a log-log PSD plot without reporting confidence intervals or goodness-of-fit can mask data that are, on close inspection, not power-law at all over the claimed range.
- Band-pass or notch filtering, applied upstream for entirely unrelated denoising reasons, can itself manufacture 1/f-like structure in the filtered output, independent of the source signal.
None of these failure modes requires bad faith; they are exactly the kind of thing that slips past a reviewer who is looking at an exponent and a p-value rather than at the raw spectrum, the detrending order, and the fit diagnostics. Given that a correct 1/f slope is already cheap to produce by heterogeneity alone, an incorrectly estimated one adds a second layer of fragility on top of a weak signal.
Scorecard: what has actually been shown
| System / claim | Exponent reported | Estimator | Collapse demonstrated? | Verdict |
|---|---|---|---|---|
| Resting EEG aperiodic activity | α ≈ 1.0–3.0 | PSD log-log fit | No | Consistent with heterogeneous relaxation rates *(speculative — flagged)*; not diagnostic of criticality on its own |
| HRV, healthy rest vs. heart failure / exercise | DFA α1 ≈ 1.0 → ≈ 0.5 | Detrended fluctuation analysis | No | Robust, repeatedly measured shift *(measured)*; mechanism (autonomic balance vs. generic loss of correlated timescales) unresolved |
| Kuramoto model, mean-field limit | r ∝ (K−Kc)^{1/2} near onset | Analytic self-consistency solution | Yes, in the N→∞ limit by construction | Genuine critical phenomenon *(established)*, but it is a model result, not a direct biological measurement |
| Millennium Bridge / coupled metronomes | Sharp order-parameter rise with coupling | Direct measurement of r vs. pedestrian load / coupling | No formal finite-size/universality test reported here | Synchronisation confirmed *(measured)*; scale-free or critical-exponent status untested |
What would change our mind
The thesis is falsifiable in specific, feasible ways, and the burden runs against the criticality reading, not toward it:
- A finite-size (or epoch-length) scaling collapse on cortical or cardiac data: systematically vary a size-like parameter (array coverage, recording duration, or an explicit control parameter such as graded cognitive load or workload) and test whether the spectra collapse onto one universal curve under the scaling predicted by a specific critical exponent, rather than merely each individually fitting a similar slope.
- A direct estimate of the underlying relaxation-time distribution (from single-unit or ion-channel kinetics, or from baroreflex/respiratory loop delays) wide and shaped as predicted by the superposition model, compared against the observed EEG or HRV exponent to see whether the two are quantitatively consistent rather than just qualitatively compatible.
- Manipulation of coupling strength in a biological analogue of Kuramoto — pharmacologically or via closed-loop stimulation that tunes effective coupling between neural populations — with measurement of whether the order parameter shows the predicted non-analytic onset at a specific Kc, versus a smooth, size-independent crossover.
- Pre-registered re-analysis of existing EEG/HRV datasets with explicit detrending-order sensitivity analysis and goodness-of-fit reporting, to establish what fraction of published 1/f and DFA α1 exponents survive the estimator checks listed above.
Until results like these exist for a given system, the correct posture toward a reported 1/f slope — in EEG, in HRV, or elsewhere — is that it is evidence of a broad distribution of timescales, which is interesting and worth characterising, but not evidence of criticality, coupling, or emergence, which remain live, separately-testable hypotheses *(speculative — flagged, where asserted without a collapse)*.



