Stabilize tensor-train relative truncation

[FlorianPfaff]

Bug

TensorTrain.from_dense(...) computed the tensor Frobenius norm with an unscaled Euclidean norm, while _choose_rank(...) computed singular-value tail energies by squaring the raw singular values.

For large but finite tensors, those intermediate squares overflow even when the tensor entries, singular values, requested relative tolerance, and correct decomposition are all finite. A diagonal 2 x 2 tensor with entries 1e200 raises under strict NumPy overflow handling; without strict handling, the infinite tolerance can incorrectly collapse the decomposition to rank one.

Fix

• compute the dense tensor norm after scaling by the largest magnitude;
• compute singular-value tail energies in normalized coordinates;
• compare against the correspondingly scaled local tolerance;
• preserve the existing zero-tolerance and max_rank behavior.

Regression coverage

Add a focused tensor-train regression using diag(1e200, 1e200) under np.errstate(over="raise", invalid="raise"). The decomposition must retain TT ranks (1, 2, 1) and reconstruct the finite input after scale normalization.

Validation

• independently reproduced overflow in both the old dense norm and old singular-value square path;
• isolated patched calculation returns the finite Frobenius norm sqrt(2) * 1e200 and selects rank two;
• final diff is limited to the tensor-train helper and its focused regression test;
• the full repository test matrix is authoritative.

[github-actions]

✅MegaLinter analysis: Success

Descriptor	Linter	Files	Fixed	Errors	Warnings	Elapsed time
✅ COPYPASTE	jscpd	yes		no	no	26.5s
✅ JSON	prettier	7	0	0	0	1.15s
✅ JSON	v8r	7		0	0	3.29s
✅ MARKDOWN	markdownlint	68	0	0	0	1.82s
✅ MARKDOWN	markdown-table-formatter	68	0	0	0	0.72s
✅ PYTHON	black	1865	87	0	0	90.11s
✅ PYTHON	isort	1865	139	0	0	2.78s
✅ REPOSITORY	betterleaks	yes		no	no	2.46s
✅ REPOSITORY	checkov	yes		no	no	54.49s
✅ REPOSITORY	gitleaks	yes		no	no	17.32s
✅ REPOSITORY	git_diff	yes		no	no	0.31s
✅ REPOSITORY	secretlint	yes		no	no	79.25s
✅ REPOSITORY	syft	yes		no	no	5.82s
✅ REPOSITORY	trivy-sbom	yes		no	no	6.59s
✅ REPOSITORY	trufflehog	yes		no	no	31.26s
✅ YAML	prettier	11	0	0	0	0.75s
✅ YAML	v8r	11		0	0	11.27s
✅ YAML	yamllint	11		0	0	0.66s

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

[FlorianPfaff]

Superseded by #5251, which reapplies the same focused tensor-train fix directly onto current main so the full workflow matrix runs with the updated APT setup. Closing this stale duplicate to keep one canonical PR.