Fix mult_f boundary conventions, window grid, and mg time-domain path - #5
Open
mcdigman wants to merge 1 commit into
Open
Fix mult_f boundary conventions, window grid, and mg time-domain path#5mcdigman wants to merge 1 commit into
mcdigman wants to merge 1 commit into
Conversation
Frequency-domain transforms (both directions) now handle windows that cross the frequency extremes exactly: - forward: conjugate fold-back (X[-jj] = conj X[jj], X[ND-jj] = conj X[jj]) for interior bands instead of zero-truncation - inverse: mirror-lobe fold-back with 2*Re at the self-conjugate bins 0 and ND/2; removed the stray m=Nf tap at delta=-K/2 that the forward never produces (invisible for Meyer, wrong for Gaussian windows) - resolved the center-tap conjugation TODO: -conj(F[0]) was an obfuscated but exactly correct stand-in for F[Nt/2] at even mult_f / odd m; replaced by the uniform index i_midpoint % Nt modified_gaussian window construction: - phitilde_vec_norm sampled on the exact bin grid arange(0,K/2+1)/Nt (the linspace endpoint=False grid was misscaled by ~2/(mult_f*Nt) and dominated round-trip error); prefactor corrected to sqrt(Nf/2) so nrm=1 - phi_vec signature fixed (mult_t kwarg crash broke every time-domain mg call); nu parameter no longer shadowed by nu_init - expansion coefficients now computed lazily and cached instead of at import Public API: wavelet family is selectable via family='meyer' (default, backward compatible, legacy stored-data tests unchanged) or family='modified_gaussian'; mult_f must satisfy 1 <= mult_f <= Nf. Performance: both directions now use length-Nt ffts (alias-folded taps / direct packing) instead of length-K, ~mult_f x less fft work; jitted is 2.9-4.6x faster than the new numpy reference at Nf=2048, Nt=1024, mult_f=8. New: WDMWaveletTransforms/wavelet_reference.py (plain-numpy executable specification of the conventions), tests/wavelet_convention_test.py (56 tests: jit==reference, closed-form atom cross-checks, round-trip convergence to 6e-15 at mult_f=24, boundary/edge cases, packing and normalization guards), benchmark_transforms.py. Consolidation: removed DX_assign_loop_old, commented-out inverse variants, phi_vec_transform, zak_g/zak_g_2filter/frame_bounds diagnostics, _omega_array_loop, _om_dot_helper_old/_convolve and dead mode branches. Modified-Gaussian round trip (Nf=32, Nt=64), previously stuck at ~1e-1 for all mult_f: now 1.7e-3 (mult_f=3), 2.6e-6 (8), 5.5e-12 (16), 6.1e-15 (24). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes frequency-domain transform boundary conventions, the modified-Gaussian window grid, and the mg time-domain path. Adds a plain-numpy reference specification and a comprehensive convention test suite.
2*Reat self-conjugate bins0andND/2, and drops the straym=Nftap the forward never produces. Resolved the center-tap conjugation TODO via the uniformi_midpoint % Ntindex.phitilde_vec_normnow sampled on the exact bin gridarange(0, K/2+1)/Nt(the oldendpoint=Falsegrid was misscaled and dominated round-trip error); prefactor corrected tosqrt(Nf/2). Fixedphi_vecsignature crash (mult_tkwarg /nushadowing) that broke every time-domain mg call. Expansion coefficients computed lazily and cached.family='meyer'(default, backward compatible) orfamily='modified_gaussian';mult_fmust satisfy1 <= mult_f <= Nf.Ntffts instead of length-K, ~mult_f× less fft work; jitted is 2.9–4.6× faster than the new numpy reference at Nf=2048, Nt=1024, mult_f=8.DX_assign_loop_old,phi_vec_transform, zak/frame-bound diagnostics,_omega_array_loop,_om_dot_helper_old/_convolve, commented-out inverse variants).New files
WDMWaveletTransforms/wavelet_reference.py— plain-numpy executable spec of the conventionstests/wavelet_convention_test.py— 56 tests (jit==reference, closed-form atom cross-checks, round-trip convergence to 6e-15 at mult_f=24, boundary/edge cases, packing/normalization guards)benchmark_transforms.pyTest plan
Modified-Gaussian round trip (Nf=32, Nt=64), previously stuck at ~1e-1 for all mult_f: now 1.7e-3 (mult_f=3), 2.6e-6 (8), 5.5e-12 (16), 6.1e-15 (24).
🤖 Generated with Claude Code