Module: tide.resampling

Signal resampling utilities used by CFL-driven internal sub-stepping.

Functions

• cosine_taper_end
• zero_last_element_of_final_dimension
• upsample
• downsample
• downsample_and_movedim

upsample

Low-pass upsampling on the last dimension.

Typical use:

• Called internally when CFL requires internal dt < user dt.
• Can be used manually for source preprocessing.

downsample

Frequency-limited downsampling on the last dimension.

Typical use:

• Called internally to bring receiver traces back to user sampling interval.

downsample_and_movedim

Convenience wrapper:

• expects receiver_amplitudes shaped [nt, n_shots, n_receivers]
• processes time on the last axis internally
• returns [n_shots, n_receivers, nt_downsampled]

Signatures

upsample(signal, step_ratio, freq_taper_frac=0.0,
         time_pad_frac=0.0, time_taper=False)

downsample(signal, step_ratio, freq_taper_frac=0.0,
           time_pad_frac=0.0, time_taper=False, shift=0.0)

Both functions operate on the final tensor dimension and preserve all leading dimensions. step_ratio=1 is the identity case. A larger ratio changes only the time-axis length.

Round-trip check

fine = tide.upsample(source, step_ratio=3)
recovered = tide.downsample(fine, step_ratio=3)
torch.testing.assert_close(recovered, source, rtol=1e-4, atol=1e-6)

Use a tolerance appropriate for signal bandwidth and taper settings. A signal with energy near the new Nyquist limit cannot be downsampled without loss.

freq_taper_frac softens the spectral cutoff. time_pad_frac reduces circular FFT interaction between trace ends. time_taper=True applies an end taper before padding. These choices affect endpoint samples and should stay consistent between observed and predicted-data preprocessing.