Skip to content

Promote integer samples before finite-difference arithmetic - #239

Merged
pavelkomarov merged 2 commits into
florisvb:masterfrom
antonsoo:fix/integer-finite-differences
Nov 3, 2026
Merged

pavelkomarov merged 2 commits into
florisvb:masterfrom
antonsoo:fix/integer-finite-differences

Conversation

@antonsoo

Copy link
Copy Markdown
Contributor

finitediff allocates a floating derivative buffer but evaluates its subtraction and weighted sums in the input integer dtype first. Those intermediate values can wrap, or raise on negative coefficients with unsigned inputs. For example:

import numpy as np
from pynumdiff import finitediff

x = np.array([100, 90, 80, 70, 60], dtype=np.uint8)
print(finitediff(x, 1.0, order=1)[1])

Before: [246. 246. 246. 246. 246.]. After: [-10. -10. -10. -10. -10.].

Convert integer samples to a floating working buffer before stencil arithmetic. Existing floating inputs retain their current dtype and path; the caller's array is preserved. Add analytic quadratic and signed/unsigned regressions across stencil orders, repeated smoothing and a non-leading time axis.

Validation: all 258 tests pass. Pylint was run against both sources; existing diagnostics remain. Across 360 recorded-PCM configurations, 48 exceptions and 96 incorrect derivative runs become zero; 76 incorrect strongest-derivative selections are corrected. Direct Python-integer stencil sums independently check one-pass derivatives. All 36 float32/float64 control hashes remain exact. The existing convex-solver accuracy warning remains unchanged.

@pavelkomarov

Copy link
Copy Markdown
Collaborator

This gets at a wider issue: I was never testing that integer input and float input would return the same thing. Claude found a few other methods where truncation or overflow could occur, so I've taken a pass over everything now and replaced your test with something simpler that generalizes more.

@pavelkomarov
pavelkomarov merged commit fc6b96e into florisvb:master Nov 3, 2026
1 check passed
@pavelkomarov pavelkomarov self-assigned this Nov 3, 2026
@antonsoo

antonsoo commented Nov 3, 2026

Copy link
Copy Markdown
Contributor Author

Thanks for extending this across the other methods. Replayed the exact merge fc6b96e on three recorded PCM speech clips. The original stencil workload's 48 exceptions and 96 incorrect runs are gone. A separate 504-case replay covering finite differences, Butterworth, kernel smoothing, spectral and wavelet methods removes all 308 integer-versus-float mismatches; all 84 float64 controls are byte-identical to the merge parent. This checks numerical input consistency, not derivative estimation quality.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants