compress

compress(data, epsilon=150.0, alpha=28.0, floor_k=64)

Compress an LFP snippet using adaptive SVD and wavelet-packet thresholding.

Parameters

Name	Type	Description	Default
data	ndarray of shape (nc, ns)	LFP data matrix, float32 or float64. Rows are channels, columns are time.	required
epsilon	float	SVD threshold multiplier. rank = #{k : sv[k] > epsilon × sigma_noise}. Default 150.	150.0
alpha	float	WP threshold multiplier per component: tau_k = alpha × sigma_noise / sv[k]. Set to 0 to skip wavelet-packet stage. Default 28.	28.0
floor_k	int	Survival floor on the dominant mode: the top retained row (k == 0) keeps at least its floor_k largest-magnitude WP coefficients. This guarantees the reconstruction is never identically zero. Without it, low-SNR chunks (sv[0] / sigma_noise below ~alpha / max\|coeff\|, empirically ~140) have every coefficient thresholded and decompress to exact zero — silently destroying real low-amplitude LFP. High-SNR recordings are unaffected: their dominant row keeps far more than floor_k coefficients, so the floor never triggers and the output is bit-for-bit identical to floor_k = 0. Set to 0 to disable. Default 64 (kills the zeroing on affected recordings while leaving the benchmark set unchanged).	64

Returns

Name	Type	Description
	LFPCompressed

Notes

The floor never resurrects saturation-muted spans: an all-zero input gives sv = 0 so U_scaled = 0 and the output is exactly zero whatever floor_k is. Zero output over a muted span is correct; the floor only targets the pathological all-zero from over-thresholding genuine low-amplitude LFP.