DSI epoch export · 11,722,320 epochs · 23 animals
Delta, theta, alpha, beta and gamma are contiguous slices of one spectrum, so asking whether they are related is really asking how many independent measurements you actually have. The answer is about two — and the two that matter are the spectrum’s overall level and its slope.
Mean pairwise r
0.65
Within animal, log10 power. Pooling across animals inflates it to 0.81.
First component
71.4%
One broadband factor takes most of the variance in every animal.
Effective dimensions
1.83
Participation ratio, median of 23 animals. Range 1.26–2.62 within a recording day.
Made by the recipe
0.13%
The pipeline’s own contribution to adjacent-band correlation. It is not an artefact.
Every number here comes from the lab’s epoch export, filtered to channel_index = 1 and features_valid = 1, with RoomD/RN245 dropped as instructed. Correlations are computed inside each animal and then summarised across the 23, because between-animal amplitude spans roughly 4× with no calibration signal — pooling first would manufacture agreement that is not there.
01 — How much information
Within an animal, every band pair is positively correlated, from 0.89 for alpha–beta down to 0.26 for delta–gamma. The gradient is orderly: the closer two bands sit in frequency, the harder they move together.
A single principal component absorbs a median 71.4% of the variance, and the participation ratio — the honest count of how many dimensions are really in play — is 1.83. Five columns, two numbers’ worth of information.
Band-to-band correlation, and what survives conditioning
Left: correlation of log10 band power, median across 23 animals. Right: partial correlation given the other three bands — the direct relationship once shared variance is removed.
Right-hand cells marked • are consistent in at least 22 of 23 animals. Only delta–alpha (+0.03, p = 0.18) is genuinely null.
Conditioning changes the picture sharply. The four neighbouring pairs stay strongly positive (delta–theta +0.52, theta–alpha +0.52, alpha–beta +0.60, beta–gamma +0.68, all 23/23 animals), while the distant pairs fall to zero or go negative: alpha–gamma −0.22, delta–beta −0.14, theta–gamma −0.09.
Two exceptions matter, because they are the only places the data departs from what geometry alone would predict. Theta–beta stays positive at +0.13 despite being two bands apart, in 22 of 23 animals and 23 of 23 recordings. And delta–beta goes negative where a smooth spectrum predicts positive. Those two sign reversals are the real conditional-structure result.
02 — Is it real
Contiguous bands integrated with the trapezoid rule share their edge bins, and a 4-second Hann window smears power across neighbours. Either could manufacture the adjacency gradient, so both were measured rather than assumed.
Neither does. Pushing signals with known independent band content through the exact export recipe induces a correlation of 0.05 at separation 1 and nothing beyond it. Under a realistic generator the recipe’s marginal contribution is +0.0010 ± 0.0004 — 0.13% of the observed 0.758. The shared edge bin adds +0.006. These mechanisms are dismissible.
Correlation falls with distance in frequency — and a two-parameter 1/f model reproduces it
Separation 1 = neighbouring bands (delta–theta … beta–gamma); separation 4 = delta–gamma. The model has a random level and a random slope, and no band coupling anywhere.
Error bars are ±1 SE across the 23 animals. Observed here is the mean across animals from the model-comparison run, so it sits slightly below the median-based matrix in the previous figure. A level-only model with no slope term gives a flat profile (0.65 / 0.63 / 0.63 / 0.62) — the tilt is what creates adjacency.
So the relationship is real signal — but it is not evidence of five distinct oscillations. A generator with two free parameters per epoch, a level and a slope, and no coupling between bands at all, reproduces the entire gradient. One broadband factor accounts for roughly 98% of the off-diagonal correlation.
Where the data does beat the model. Residuals against the 1/f fit are not uniform. Theta–alpha (+0.14), alpha–beta (+0.15) and theta–beta (+0.12) are more coupled than a smooth spectrum predicts, while every gamma pairing is less: beta–gamma −0.23, alpha–gamma −0.25, theta–gamma −0.18.
Gamma is the outlier. It is the band least tied to the others, it barely moves across the day, and it tracks line length — an effective-frequency measure — far more closely than it tracks EEG amplitude. Treat it as carrying something the rest of the spectrum does not, and be cautious about reading it as cortical activity.
The likely reason is movement. A Racine-scale rodent seizure is a motor event, and this is a single subcutaneous biopotential channel with no EMG reference to subtract. Line length is the best ictal detector on the whole page — 0.87, ahead of gamma’s 0.83 — which is what you would expect if part of the high-frequency ictal rise is muscle rather than cortex. It does not weaken detection; it does mean the gamma result cannot be read as a cortical-oscillation finding without a separate EMG channel.
03 — Across the day
The bands do not each keep their own schedule. They are driven by a single transition tied to the room lights — a rapid fall around 17:50 and a roughly 90-minute rise beginning near 05:00 — and one component takes 97.8% of each animal’s 24-hour band structure.
But the gain on that switch is steeply graded by frequency. Alpha loads at 0.54 and swings 0.64–1.30 of its own daily median; gamma loads at 0.07 and barely moves. Line length runs the other way entirely, rising at night while total power falls.
24-hour profile of each band, normalised to its own daily median
Animal-equal-weight hourly medians over 11.7 M epochs. The shaded block is the light phase, 06:00–18:00 — the rest phase for a nocturnal rodent, established independently from the camera’s infrared switch.
Line length is drawn for contrast, not as a band. It is an effective-frequency measure and moves opposite to spectral power.
Static correlation cannot see the second, smaller thing happening here. Inside the light block the spectrum slowly tilts: delta loses 0.23 percentage points of spectral share per hour (in 22 of 23 animals) while beta gains 0.10 and alpha 0.05 (both 23 of 23). Across the dark block it is flat. That drift is unanimous across the cohort and independent of the switch itself.
04 — During seizures
Seizures tilt the spectrum hard and consistently, and the effect grows with frequency. Gamma rises 2.90 interictal standard deviations and beta 2.23, while delta (+0.10) and theta (+0.44) have intervals that cross zero — the slow bands do not reliably move at all.
How far each band moves during a seizure
Median shift in log10 power, in units of that animal’s own interictal standard deviation, with animal-clustered 95% intervals (n = 19). Line length and total power are shown for reference, not as bands.
Delta and theta are the two whose intervals cross zero — the slow bands do not reliably move at all. The effect grows monotonically with frequency.
Ictal spectral tilt, and how little it changes with Racine stage
Left: shift in centred-log-ratio share, all ictal epochs against interictal baseline. Right: the same measure by stage, on the 8 animals that contribute every stage — a balanced comparison.
Bars are medians with animal-clustered 95% intervals. The stage panel is restricted to animals present in all four stages; the unbalanced version of this chart shows a dramatic ladder that is entirely a between-animal artefact.
Severity is where it stops. Across Stage 2 to Stage 5 the gamma shift runs 1.06 / 1.04 / 1.27 / 1.25 — a total span of 0.23 z against an ictal-versus-interictal effect of 2.90. The bands detect the event; they do not grade it.
That shows up directly in discrimination. Every band, every log-ratio, and a 13-feature leave-one-animal-out model all land between 0.46 and 0.55 for severe-versus-mild once session is controlled for — while the same features reach 0.55–0.87 for simply detecting that a seizure is happening.
Detecting a seizure versus grading one
Within-session AUROC. Left dot: severe (Stage 4+5) versus mild (Stage 2+3). Right dot: ictal versus interictal. Same features, same epochs, same protocol.
The dashed marker at 0.927 is session severe-rate alone — a number containing no physiology whatsoever. Pooled, it beats every real feature; within session it is exactly 0.5.
The pooled number is the trap. Score severity on session identity alone — how often that session’s seizures happen to be severe, with the EEG never consulted — and pooled AUROC reaches 0.927. Within session it collapses to exactly 0.5, because there is nothing in it. Any severity figure from this export that is not computed within session is measuring which session a clip came from.
Seizures also change how the bands relate. Four of ten pairwise correlations drop significantly after correction — theta–alpha −0.16, theta–beta −0.20, delta–alpha −0.24, alpha–beta −0.07 — and the share explained by one factor falls from 0.77 to 0.67.
That is easy to misread as desynchronisation, and it is not. In absolute terms both components grow: the shared broadband factor’s variance rises 1.42× (in 13 of 17 animals) and the band-specific residual rises 2.54× (in all 17). Seizures add differential energy faster than they add common energy, so the proportion attributable to one factor falls while synchrony itself rises.
05 — What to do with it
Two things this page does not cover. Spikes events are entirely absent — all 94 Spikes clips have every overlapping epoch blanked by the export’s validity gate, so that class contributes nothing here and its behaviour is unknown. And nothing here looks at the pre-ictal window: every ictal number describes what the bands do during a seizure, not whether they forewarn one. The second is answerable from this export; the first is not, without a re-export.
One caveat on the headline. The nearest-neighbour chain is not a discovered property of rodent EEG. Integrating any smooth log-spectrum over five contiguous bands produces it, with no physiology involved — a synthetic smooth-spectrum null reproduces the chain in full. What genuinely departs from that null is narrower: the positive theta–beta edge and the negative delta–beta edge, plus gamma’s systematic under-coupling.