Make the linear-algebra epsilons relative, so estimates stop depending on units

[microprediction]

Two absolute epsilons in _linalg.py decided the answer whenever the data was small. Both have units of variance, so both stop being negligible at exactly the scale a user in basis points or log-returns works at. Found by audit; reproduced independently.

cov_to_corrcoef

Added + EPS (1e-12) to sqrt(v_i · v_j). At variances of 1e-12 that halved every entry, so the correlation diagonal — one by definition — came back as 0.5:

variance	diagonal	error
1e+00	1.000000	1e-12
1e-11	0.909091	9.1e-02
1e-12	0.500000	5.0e-01
1e-13	0.090909	9.1e-01

Off-diagonals scale by the same factor, so a true correlation of 0.5 read as 0.045. Each variance is now floored at a fraction of the largest, which keeps the unit diagonal at every scale and still avoids dividing by a genuinely zero variance.

make_pos_def

Clipped eigenvalues up to an absolute 1e-8. Below a data scale of ~1e-4 every eigenvalue is under that, so BlockCovariance, SchurCovariance and SchurLedoitWolfCovariance returned 1e-8 · I — up to nine thousand times the true variance, carrying no correlation information at all. At a data scale of 1e-6 the diagonal now returns 0.81 of truth rather than 4,400× it.

One correction to the audit

It attributed GeodesicEwaCovariance's collapse to this same floor. Making the floor relative fixes that estimator's scale equivariance but not the collapse — the trace still falls to 7.5e-06 against a true 5.99 within ~40 observations. The cause is the geodesic step interpolating toward a rank-one target whose remaining p−1 eigenvalues carry no information, with that direction rotating every step. Filed separately rather than claimed fixed here.

327 tests pass, plus new scale-equivariance regressions across five estimators and four scales; ruff and mypy clean.

🤖 Generated with Claude Code