Stabilize sliding-window manifold weights

[FlorianPfaff]

Summary

• reject non-finite window_weights during SlidingWindowManifoldMeanSmoother construction
• normalize each active window without overflowing its sum
• avoid JAX/XLA reciprocal underflow for maximum finite weights
• preserve relative weights and add end-to-end regression coverage through smooth()

Root cause

The original implementation normalized weights directly with:

weights / sum(weights)

For individually finite values such as [max_float, max_float / 2], the sum overflows. The first stabilization attempt instead used:

weights / max(weights)

That works with NumPy, PyTorch, and older JAX versions, but the CI-pinned JAX 0.10.2 lowers division by float32.max through reciprocal multiplication. The reciprocal underflows to zero, so both scaled weights become zero and the subsequent normalization produces NaNs. All three JAX matrix jobs failed when the downstream Dirac distribution rejected those non-finite weights; NumPy and PyTorch remained green.

Fix

Validate that every configured weight is finite. For each active window, divide twice by the square root of its largest positive weight before normalizing:

scale_root = sqrt(max(weights))
scaled = (weights / scale_root) / scale_root
normalized = scaled / sum(scaled)

Each divisor remains representable, avoiding JAX's extreme reciprocal underflow while retaining the original weight ratios.

Validation

• reproduced the failure with the CI-pinned JAX 0.10.2: [max_float, max_float / 2] / max_float became [0, 0]
• verified the two-step scaling on JAX 0.10.2, NumPy, PyTorch, and JAX 0.9
• verified maximum finite 2:1 weights normalize to approximately [2/3, 1/3]
• verified trailing-window smoothing of [0, 3] returns approximately [0, 1]
• verified NaN and infinities are rejected

The full GitHub Actions matrix is running on the updated branch.

[github-actions]

✅MegaLinter analysis: Success

Descriptor	Linter	Files	Fixed	Errors	Warnings	Elapsed time
✅ COPYPASTE	jscpd	yes		no	no	13.14s
✅ JSON	prettier	7	0	0	0	0.62s
✅ JSON	v8r	7		0	0	2.52s
✅ MARKDOWN	markdownlint	68	0	0	0	1.07s
✅ MARKDOWN	markdown-table-formatter	68	0	0	0	0.34s
✅ PYTHON	black	1490	182	0	0	56.36s
✅ PYTHON	isort	1490	328	0	0	1.53s
✅ REPOSITORY	betterleaks	yes		no	no	1.61s
✅ REPOSITORY	checkov	yes		no	no	40.73s
✅ REPOSITORY	gitleaks	yes		no	no	8.32s
✅ REPOSITORY	git_diff	yes		no	no	0.26s
✅ REPOSITORY	secretlint	yes		no	no	38.13s
✅ REPOSITORY	syft	yes		no	no	3.43s
✅ REPOSITORY	trivy-sbom	yes		no	no	4.07s
✅ REPOSITORY	trufflehog	yes		no	no	21.27s
✅ YAML	prettier	11	0	0	0	0.38s
✅ YAML	v8r	11		0	0	6.58s
✅ YAML	yamllint	11		0	0	0.32s

Notices

📣 MegaLinter 9.5.0 is out! Discover the new features and security recommendations in the release announcement. (Skip this info by defining SECURITY_SUGGESTIONS: false)

See detailed reports in MegaLinter artifacts

Your project could benefit from a custom flavor, which would allow you to run only the linters you need, and thus improve runtime performances. (Skip this info by defining FLAVOR_SUGGESTIONS: false)

• Documentation: Custom Flavors
• Command: npx mega-linter-runner@9.6.0 --custom-flavor-setup --custom-flavor-linters PYTHON_BLACK,PYTHON_ISORT,COPYPASTE_JSCPD,JSON_V8R,JSON_PRETTIER,MARKDOWN_MARKDOWNLINT,MARKDOWN_MARKDOWN_TABLE_FORMATTER,REPOSITORY_CHECKOV,REPOSITORY_GIT_DIFF,REPOSITORY_GITLEAKS,REPOSITORY_BETTERLEAKS,REPOSITORY_SECRETLINT,REPOSITORY_SYFT,REPOSITORY_TRIVY_SBOM,REPOSITORY_TRUFFLEHOG,YAML_PRETTIER,YAML_YAMLLINT,YAML_V8R

Show us your support by starring ⭐ the repository