Strange behavior of dwt_coeff_len and dwt_max_level. #877

Activity

1. drkotireddysaripalli commented on Oct 8, 2026

   

   drkotireddysaripalli

   on Oct 8, 2026

   More actions

   With the current source's documented meaning, the original result 2 is consistent. dwt_max_level returns the maximum useful depth for the original signal, using the boundary-effect criterion in its docstring. It is not the number of times an output array can be shortened while a fresh call still returns a positive level.

   For db2, the decomposition filter length is 4, so the documented bound is

   floor(log2(23 / (4 - 1))) = 2
   

   The coefficient-length rule for symmetric is floor((n + 3) / 2). These two calculations give exactly your loop:

   Current array length	Fresh dwt_max_level(n, 4)	Next coefficient length
   23	2	13
   13	2	8
   8	1	5

   The later calls treat 13 and 8 as lengths of new input signals. They do not track the boundary effects already carried through previous decomposition stages, so their values cannot be accumulated into a depth budget for the original 23 samples.

   To predict lengths at the default useful depth, compute that depth once:

   w = pywt.Wavelet("db2")
   n = 23
   depth = pywt.dwt_max_level(n, w)
   lengths = [n]
   for _ in range(depth):
       n = pywt.dwt_coeff_len(n, w.dec_len, "symmetric")
       lengths.append(n)
   # Source formulas give: [23, 13, 8]

   wavedec does this same one-time level selection from the original axis length. For that case its coefficient-list geometry is [cA2, cD2, cD1], with lengths [8, 8, 13].

   An explicit level=3 is a separate choice: the current Python guard returns it while warning that all coefficients experience boundary effects. This warning behavior was deliberately introduced in #403; dwt_max_level is a useful-depth criterion rather than a hard prohibition on a deeper transform.

   I checked the pinned current source, integer length calculations and isolated Python level guard; I did not execute the PyWavelets C/Cython transform or your build, so this does not diagnose the separate __version__ discrepancy. Affiliation: Bharath Healthcare / Pinnacle Blooms Network. Codex assisted with this source inspection, bounded local check and reproducibility-focused reply.