[seonghobae]

Fixes the marginal._gh part of #2110.

Problem

numpy.polynomial.hermite_e.hermegauss overflows its weight formula from q = 371 (numpy 2.5.3). The weights then sum to 0 or NaN, so marginal._gh returned NaN weights. _gh feeds the NumPy marginal estimators, fitstats and linking, so any of them evaluated at a study-setting node count of 371 or more got NaN quadrature.

Change

• New private helper fast_mlsirm/_gauss_hermite.py::probabilists_gauss_hermite(q).
  • Where the normalized hermegauss rule is finite, it returns that rule unchanged, so results for q <= 370 stay bit-identical.
  • Otherwise it falls back to the Golub–Welsch Jacobi eigen-rule. That rule mirrors the Rust core quadrature.rs::gauss_hermite_probabilists (alpha_k = 0, beta_k = k, symmetrized, unit-sum weights).
  • The docstring carries APA 7 citations with the same locators already verified in quadrature.rs.
• marginal._gh calls the helper. The incorrect comment that hermegauss handles any q is removed. There is still no node-count cap (fix(quadrature): remove unsourced numeric defaults for quadrature node counts in monotonicity/bifactor APIs #1929).

Not in this PR

• estimators/mmle.py::gauss_hermite_nodes validates n_nodes <= MAX_GAUSS_HERMITE_NODES = 100, so it cannot reach q >= 371 today. That cap itself conflicts with the fix(quadrature): remove unsourced numeric defaults for quadrature node counts in monotonicity/bifactor APIs #1929 maintainer rule. Removing it changes a public validation contract, so it should be a separate decision.
• polytomous.py stays out of scope; it belongs to fix(polytomous): finite Gauss-Hermite rule for high q in monotonicity checks #2109.

Verification (local, Python 3.12, numpy 2.5.3, macOS arm64)

• RED on the unmodified code: the new tests/test_gauss_hermite_high_q.py gave 6 failed / 9 passed. Every q >= 371 case failed with non-finite weights.
• GREEN: the same file gives 15 passed.
• Checks covered: q ∈ {371, 400, 481, 961, 1000} give finite, ascending, exactly symmetric weights that sum to 1. The N(0, 1) moments E[x²] = 1, E[x⁴] = 3 and E[x⁶] = 15 are reproduced. For q ∈ {1, 2, 7, 21, 121, 241, 370} the output is bit-identical to normalized hermegauss.
• Regression check: 36 existing test files touching marginal, fitstats, linking, paper features, Rust parity and cov_e, plus the new file, gave 697 passed, 1 skipped, exit 0.
• ruff is not installed in the local venv, so it was not run. Hosted lint will cover it.

Summary by CodeRabbit

• Bug Fixes
  • Improved numerical reliability of Gaussian quadrature at high node counts, with finite, symmetric nodes and normalized weights.
  • Preserved validation for invalid node counts and accurate quadrature results across tested counts.
• Tests
  • Added coverage for high-count quadrature behavior, weight stability, and normal-moment accuracy.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: ContextualWisdomLab/fast-mlsirm/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: 32f8b442-5032-463f-b148-a2665fb644aa

📥 Commits

Reviewing files that changed from the base of the PR and between a0d7958 and 3b97e83.

📒 Files selected for processing (3)

• python/fast_mlsirm/_gauss_hermite.py
• python/fast_mlsirm/estimators/marginal.py
• tests/test_gauss_hermite_high_q.py

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The change adds a shared probabilists’ Gauss–Hermite helper with a Golub–Welsch fallback. The marginal estimator delegates node and weight generation to the helper. New tests cover high node counts, numerical properties, lower-count equivalence, and invalid counts.

Changes

Gauss–Hermite quadrature

Layer / File(s)	Summary
Quadrature generation and fallback python/fast_mlsirm/_gauss_hermite.py	Adds probabilists_gauss_hermite, which returns normalized NumPy results when they pass validity checks and otherwise uses a Golub–Welsch eigendecomposition. The fallback symmetrizes nodes and weights and handles q == 1 directly.
Marginal estimator integration and validation python/fast_mlsirm/estimators/marginal.py, tests/test_gauss_hermite_high_q.py	Updates _gh to use the shared helper while retaining its q >= 1 validation. Tests check high-count numerical properties, exact results against normalized hermegauss for selected counts, log weights, and invalid counts.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix

Merge Risk: ⚪ Minimal · up to 3b97e

No actionable issue remains before merge, subject to normal test and lint checks.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 44.44% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 9 functions across 3 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the fix: keeping Gauss–Hermite weights finite for node counts q ≥ 371.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Fix all pre-merge checks with AI

✨ Finishing Touches 💡 2

📝 Generate docstrings 💡

• Commit to this branch
• Create a new PR

🛠️ Fix failing CI checks 💡

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}