meegkit.tspca

Time-shift PCA.

Functions

tspca(X[, shifts, keep, threshold, weights, ...])	Time-shift PCA.
tsr(X, R[, shifts, wX, wR, keep, thresh])	Time-shift regression.

meegkit.tspca.tspca(X, shifts=None, keep=None, threshold=None, weights=None, demean=False)

Time-shift PCA.

Parameters:

• X (array, shape=(n_times, n_chans[, n_trials])) – Data array.

• shifts (array, shape=(n_shifts,)) – Array of shifts to apply.

• keep (int) – Number of components (shifted regressor PCs) to keep (default=all).

• threshold – Discard PCs with eigenvalues below this (default=1e-6).

• weights (array) – Sample weights.

• demean (bool) – If True, Epochs are centered before computing PCA (default=0).

Returns:

• comps (array) – Principal components array.

• V (array, shape=(n_chans, n_comps)) – PCA weights.

• idx (array) – X[idx] maps to principal comps.

meegkit.tspca.tsr(X, R, shifts=None, wX=None, wR=None, keep=None, thresh=1e-12)

Time-shift regression.

The basic idea is to project the signal X on a basis formed by the orthogonalized time-shifted R, and remove the projection. Supposing R gives a good observation of the noise that contaminates X, the noise is removed. By allowing time shifts, the algorithm finds the optimal FIR filter to apply to R so as to compensate for any convolutional mismatch between X and R.

Parameters:

• X (array, shape=(n_samples, n_chans[, n_trials])) – Data to denoise.

• R (array, shape=(n_samples, n_comps[, n_trials])) – Reference data.

• shifts (array, shape=(n_shifts,)) – Array of shifts to apply to R (default=[0]).

• wX (array, shape=(n_samples, 1, n_trials)) – Weights to apply to X.

• wR (array, shape=(n_samples, 1, n_trials)) – Weights to apply to R.

• keep (int | None) – Number of shifted-R PCs to retain (default=all).

• thresh (float) – Ignore shifted-R PCs smaller than thresh (default=1e-12).

Returns:

• y (array) – Denoised data.

• idx (array) – X[idx] is aligned with y.

• mean (array) – Channel means (removed by TSR).

• weights (array) – Weights applied by TSR.