Contribution
It gathers, in one place and with simulations whose code is public (github.com/voytekresearch/OscillationMethods), the ways in which standard spectral methods applied to field potentials return numbers that are easy to misread as oscillatory physiology. Its frame: “these are not failures of the algorithms per se, which do, mathematically, exactly what they should; the potential pitfalls lay in how we interpret their outputs” (p. 6).
Key insight
Most methods assume a sinusoidal, stationary, oscillation-present signal. Neural field data are bursty, non-sinusoidal and sit on dynamic aperiodic activity, so a measure can be numerically correct and still not mean what it is read to mean. For coupling: “Numerically, these values are not objectionable … However, there is possible fallacy in the interpretation, if this relationship is taken to reflect significant coupling between independent rhythms, when in fact no such interaction between multiple rhythms need exist” (p. 25).
Assumptions
- Simulated data made to match known properties of field recordings, with NeuroDSP (p. 7); “ground-truth properties of physiological data are not known” (p. 37), so the simulations show possibility, not prevalence.
- Univariate focus (power, phase); multivariate measures such as cross-frequency coupling are discussed as inheriting the problems (pp. 7, 38).
Key results
The seven considerations (Box 1, p. 8):
- Oscillations may be absent (pp. 9–12). Aperiodic 1/f^χ activity has power at all frequencies (χ typically 0–4), so “just because there is power in a frequency band does not imply that there is an oscillation” (p. 9); filtering aperiodic noise yields rhythmic-looking traces (Fig. 1C).
- Peak frequencies vary (pp. 13–16): a canonical 8–12 Hz band misestimates an 8 Hz alpha (Fig. 2).
- Aperiodic activity is dynamic (pp. 17–19): a change in the 1/f exponent produces correlated band-power changes across all bands (Fig. 3); the theta/beta ratio in ADHD “appears to be driven by aperiodic activity” (p. 18).
- Bursts (pp. 21–23): averaging bursty trials creates “an illusion of sustained activity” (p. 21).
- Waveform shape (pp. 24–27): narrowband filtering shifts the peaks and troughs of asymmetric waves (Fig. 5A); non-sinusoidal shape appears as harmonics (Fig. 5B–C), which can be read as separate coupled rhythms; and PAC measures “are greatly influenced by waveform shape”, giving significant coupling from one non-sinusoidal 10 Hz rhythm (Fig. 5D–E). Diagnostic: “narrow ranges at exactly multiples of the base frequencies indicative of possible coupling caused by waveform shape” (p. 26).
- Overlapping sources (pp. 28–31): volume conduction makes split peaks and spurious connectivity.
- SNR (pp. 32–35): low power corrupts phase estimates and so PLVs and cross-frequency coupling (Fig. 7F; Aru et al. 2015).
Reproduction of #5. A 60 s, 10 Hz signal of sharp-peaked shape, exp(3 cos φ), with white noise, gave a Tort modulation index between 8–12 Hz phase and 25–60 Hz amplitude of 0.021, against 0.0001 for a pure cosine with the same noise; with a 15–25 Hz amplitude band (one harmonic) and with an 80–250 Hz band, both stayed at noise level (≤ 0.00006). So shape-induced PAC is real and confined to bands spanning several harmonics. A strictly periodic wave did not reproduce the 15–25 Hz case of Fig. 5D, which presumably relies on the filter or on variability in the simulated rhythm that the caption does not specify.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Band power does not imply an oscillation in that band | strong | the spectrum of aperiodic signals (Fig. 1); elementary |
| C2 | Changes in aperiodic activity can masquerade as changes in several oscillations | strong | simulation (Fig. 3); empirical cases cited (p. 18) |
| C3 | Non-sinusoidal waveforms produce significant PAC without interacting rhythms | strong | simulation (Fig. 5D–E); reproduced here for a multi-harmonic band |
| C4 | Coupling confined to exact harmonic multiples suggests shape; broad coupling suggests genuine nesting | moderate | heuristic, citing Cole et al. 2017, Vaz et al. 2017 (p. 26) |
| C5 | Low SNR biases PLV and coupling measures | strong | simulation (Fig. 7) |
Concepts
- Aperiodic activity — the 1/f-like component of field data, with “different putative generators” from oscillations (p. 17).
- Waveform shape — rise/decay asymmetry, peak sharpness; measured cycle by cycle in the time domain (p. 25).
- Phase-amplitude coupling — treated as a measure of dependence that may or may not reflect “nested oscillations” (p. 25).
Connections
- Kramer, Tort and Kopell 2008 (sharp edges and spurious coupling), Cole et al. 2017, Lozano-Soldevilla et al. 2016, Giehl et al. 2021 and Vaz et al. 2017 are the PAC-and-shape literature it rests on (p. 25).
- Voytek is senior author here and the source of the PAC method Szczepanski et al. used (LIT-1982, NOTE-1528).
- “Park and colleagues (2017)”, which A5 attributed to this line of caution, is not cited in this paper. The only 2017 entries near the name are Cole et al. 2017 on Parkinson’s beta, whose title contains “Parkinson’s”. It may be a garbled memory of that paper.
Bearing on the record
- Jennings 2026 (LIT-1963, NOTE-1508), on PAC. Jennings reads PAC as “evidence that low frequency waves constrain high frequency waves, a form of power or control” (p. 28). This paper shows that a PAC measure can be significant with no second rhythm at all. A measured PAC therefore needs shape controls before it is evidence of one rhythm modulating another, let alone of a macro-level power (THEORY-1314).
- On 1/f scaling. Jennings offers the brain’s self-organization “into scale-free patterns” as what “allows for the emergence of the subject” (pp. 27–29), noting in fn. 20 that such background activity was once treated as noise. Here 1/f activity is the aperiodic background, present in every recording, with putative generators in postsynaptic currents (p. 9; Gao et al. 2017). It is a physiological feature worth studying, but its presence is not by itself evidence of a higher-level constraint.
- On Szczepanski et al. (THEORY-1317). That study’s 80–250 Hz coupling to a 2–5 Hz phase is broad and far from low harmonics, so by C4 it is not the typical shape artefact; but no shape control was run. A5’s caution applies in principle and is not shown to defeat it.
- The registration’s summary held up: “non-sinusoidal waveforms produce apparent coupling” is the paper’s #5.
- No instruction for machine-learning practice.
Limitations
- Simulations show what can go wrong, not how often it does in published findings; the paper does not re-analyse any empirical PAC result.
- Read from the preprint; the published version may differ.
Open questions
- What fraction of reported attention-related PAC survives cycle-by-cycle shape controls and aperiodic correction?