sage.dsp.inverse_spectrum_truncation

Filename : inverse_spectrum_truncation.py Description : Short description of the file

Created on 2026-02-09 12:18:56

__author__ = Narenraju Nagarajan __copyright__ = Copyright 2026, ProjectName __license__ = MIT Licence __version__ = 0.0.1 __maintainer__ = Narenraju Nagarajan __affiliation__ = N/A __email__ = N/A __status__ = [‘inProgress’, ‘Archived’, ‘inUsage’, ‘Debugging’]

GitHub Repository: NULL

Documentation:

Functions

inverse_spectrum_truncation_batch(psd, max_filter_len)

GPU-native batched inverse spectrum truncation (IST) for PSD smoothing.

inverse_spectrum_truncation_single(psd, max_filter_len)

CPU version of inverse spectrum truncation for a single PSD (1D tensor).

Module Contents

inverse_spectrum_truncation_batch(psd, max_filter_len, low_frequency_cutoff=None, delta_f=1.0, trunc_method='hann')[source]

GPU-native batched inverse spectrum truncation (IST) for PSD smoothing.

Reproduces the PyCBC inverse_spectrum_truncation routine in PyTorch for batched, device-resident PSDs. IST time-limits the inverse-ASD filter so that it cannot introduce correlations beyond max_filter_len samples, preventing the whitening filter from growing unboundedly long.

Algorithm:

  1. Compute 1 / sqrt(PSD) in the frequency domain (inverse ASD).

  2. IFFT to the time domain and zero out samples beyond max_filter_len/2 from each edge (optionally with a Hann window at the truncation point).

  3. FFT back and square to recover the truncated PSD.

Parameters:

  • psd (torch.Tensor, shape (B, D, F)) – Real-valued one-sided PSDs.

  • max_filter_len (int) – Maximum whitening filter length in samples. Bins beyond this radius are zeroed in the time-domain inverse ASD.

  • low_frequency_cutoff (float or None) – Frequency (Hz) below which the inverse ASD is set to zero.

  • delta_f (float) – Frequency bin spacing in Hz (default 1.0).

  • trunc_method (str or None) – Window applied at the truncation boundary: "hann" applies a half-Hann taper; None uses a hard rectangular truncation.

Returns:

Truncated PSDs ready for use in whitening.

Return type:

torch.Tensor, shape (B, D, F)

inverse_spectrum_truncation_single(psd, max_filter_len, low_frequency_cutoff=None, delta_f=1.0, trunc_method='hann')[source]

CPU version of inverse spectrum truncation for a single PSD (1D tensor).

Parameters:

  • psd (torch.Tensor) – torch.Tensor, shape (F,), real-valued PSD

  • max_filter_len (int) – int, maximum length of the time-domain filter

  • low_frequency_cutoff (float) – float or None, zero out PSD below this frequency

  • delta_f (float) – float, frequency bin spacing of PSD

  • trunc_method (str) – str or None, truncation method (“hann” or None)

Returns:

torch.Tensor, shape (F,), truncated PSD

Return type:

torch.Tensor