Band power is the amount of signal energy that sits within a given frequency band. For EEG, it is the single most common summary of a recording — because it collapses a long, high-rate time series into a small number of interpretable values, one per band per channel.
There are two flavors, and they answer different questions.
Absolute band power
Absolute band power is the raw integrated power in a band, in units of µV² (or µV²/Hz if you report power density rather than total band power). It reflects the actual amplitude of the signal in that band.
Because absolute values depend on electrode impedance, skin preparation, skull thickness, reference choice, and hardware gain, they are hard to compare directly across subjects, sessions, or sites. Two recordings from the same brain on different days can differ in absolute power by a factor of two or more for reasons that have nothing to do with the brain.
Relative band power
Relative band power expresses each band as a fraction (or percentage) of the total power across all bands:
rel_alpha = P_alpha / (P_delta + P_theta + P_alpha + P_beta + P_gamma)
Because it is a ratio, most of the multiplicative nuisance factors (impedance, gain, skull) divide out. Relative power is what most cross-subject comparisons should use.
Absolute values depend on amplifier gain and electrode impedance. Relative values divide by the sum of the bands shown — change the denominator and every percentage moves.
How the numbers are computed
The standard pipeline is:
- Preprocess. High-pass filter to remove drift (typically 0.5 Hz or 1 Hz), notch filter line noise (50 or 60 Hz), and reject or interpolate obvious artifact epochs.
- Estimate the power spectral density (PSD). Welch's method is the workhorse: split the signal into overlapping segments (commonly 2–4 s, 50% overlap), window each segment (Hann is standard), take the FFT, square the magnitude, and average across segments. This trades frequency resolution for a lower-variance estimate. The multitaper method is a more sophisticated alternative that controls the bias–variance trade-off explicitly.
- Integrate over each band. Numerically integrate the PSD (usually via the trapezoidal rule) between the low and high edges of the band. That integral is the absolute band power in µV².
- Normalize if you want relative power. Divide by the sum of the band powers you consider "total."
The denominator problem
Relative band power depends entirely on what you sum in the denominator. This is the part that trips people up.
If your total is delta + theta + alpha + beta + gamma, those bands must exactly tile the frequency range with no gaps and no overlaps. If Alpha is defined as 8–12 Hz and Beta as 13–30 Hz, the 12–13 Hz slice belongs to neither band, is silently dropped from the denominator, and every percentage you report is slightly inflated. If Beta is 12–30 Hz, the 12 Hz bin is double-counted.
See the sibling article, EEG Frequency Bands, for the full picture.
Another common choice: use the total power across a broad analysis range (for example 0.5–45 Hz) as the denominator, rather than the sum of named bands. This handles gaps between bands cleanly, but it means "relative alpha" now includes contributions from everything in that range, including sub-band ripples and any artifact that survived preprocessing. Report which convention you used.
What "total" includes and excludes
Regardless of which normalization you pick, the total is:
- Bounded by your PSD's frequency range. If you Welch-estimate up to the Nyquist frequency but only integrate named bands to 30 Hz or 50 Hz, everything above that is excluded — including line noise if the notch was imperfect.
- Sensitive to preprocessing. A high-pass filter at 1 Hz removes real delta power below that. An aggressive line-noise notch can eat neighboring frequencies.
- Not immune to artifact. Muscle activity inflates beta and gamma. Eye movement inflates delta. Sweat drift inflates the lowest frequencies. Artifact rejection choices therefore change every band's relative contribution.
Practical checklist
- State the band edges you used. Every time.
- State whether "total power" is sum of bands or integral over a broad range.
- Report the PSD method (Welch parameters: window length, overlap, window function; or multitaper parameters).
- Give both absolute and relative power where you can — they answer different questions.
BandForm computes absolute and relative band power per channel via Welch, and lets you configure the band edges so the boundaries match your lab's convention.