Calculate relative spectral power in fixed frequency bands for an input time-series vector
Source:R/fft_features.R
fft_features.RdSplits the frequency range from 0 to 0.5 cycles per sample (the Nyquist frequency) into 100 equal-width bands and returns the fraction of the series' spectral power in each, as fft_power_1 (0 to 0.005 cycles per sample) to fft_power_100 (0.495 to 0.5). The bands sum to 1.
Arguments
- y
numericvector of values- squared
Deprecated and ignored: the features are relative power, so there is no magnitude/squared-magnitude choice
- phase
Booleanspecifying whether to also return each band's phase as its cosine and sine. Only meaningful when the series share a time origin. Defaults toFALSE
Details
The power spectrum is the periodogram of the mean-removed series, computed with the FFT. A series of length n has periodogram ordinates at frequencies k/n, k = 1, ..., floor(n/2); each ordinate's power is spread evenly over its frequency cell of width 1/n and the result is integrated over each band. Because the bands are defined in normalised frequency rather than by FFT coefficient index, each feature measures the same frequency range for series of any length (n >= 2), and no coefficients are missing for short series. Short series have a coarse frequency resolution (1/n), so neighbouring bands then share power from the same ordinate. As the features are relative power, they don't depend on the series' mean or scale (see moments for those). A constant series has no spectral power, so all its features are NaN.
With phase = TRUE, each band's phase is also returned, as fft_phase_cos_b and fft_phase_sin_b (the cosine and sine of the phase, which represent it without the jump at +/-180 degrees). The band's phase is the argument of the sum of its complex FFT coefficients, each weighted by the fraction of its frequency cell in the band, so strong coefficients dominate. Phase is measured relative to the first sample, so shifting a series in time changes it: these features are only meaningful when the series share a time origin (e.g., event-locked trials or aligned segments). Bands with negligible power have no well-defined phase and give NaN.