fix: natural_logpdf clamps a genuine -inf to -1.8e308

[Jammy2211]

Overview

AbstractMessage.natural_logpdf (autofit/messages/interface.py:98) reduces through

xp.nan_to_num(log_base + eta_t - log_partition, nan=-xp.inf)

The nan=-xp.inf is deliberate — an out-of-support NaN is zero density. But posinf/neginf are left at their defaults, and nan_to_num replaces a genuine -inf with negative float max. The call therefore does the opposite of its intent for exactly the inputs that already had the right answer:

value reaching the reduction	intended	actual
NaN	-inf	-inf ✅
-inf	-inf	-1.7976931348623157e+308 ❌

Measured on main @ 6e2d8c8

LogGaussianPrior(0.4, 1.3):

value	message.logpdf	why
-1.0	-inf	log(-1) is NaN → nan=-inf applies
0.0	-1.7976931348623157e+308	log(0) is -inf → default neginf clamps

That asymmetry is the proof of mechanism, and the clamped value is exactly -sys.float_info.max.

UniformPrior and LogUniformPrior return -inf correctly outside their boxes, so this is not visible for every family — it needs an expression that reaches -inf rather than NaN.

Why it matters

-1.8e308 is finite, and isfinite is the test a lot of this codebase branches on:

• optax.apply_if_finite — the mechanism autofit/non_linear/clipper.py is built around. That module's entire premise is that leaving the prior support makes the objective non-finite.
• Two such terms summed overflow to -inf while one does not, so behaviour depended on how many parameters were out of support.

Plan

• Pass neginf=-xp.inf, posinf=xp.inf so only NaN is replaced.
• Cover both halves of the reduction, pin that in-support values are untouched, and assert the general property that no prior family reports a finite density off its support.

Origin

Loose end from #1532, noted in its Shipped comment and not filed at the time. PyAutoMind prompt: draft/bug/priors/natural_logpdf_clamps_neginf.md.

Deliberately out of scope

TruncatedGaussianPrior's message returns finite logpdf well outside its limits (-8.20 at -1.0 for a (0, 3) support). That is separate looseness in TruncatedNormalMessage, not this clamp — the prior-level log_prior_from_value is correct there, which is why the P6 property tests pass. Not addressed here, and the new general-property test deliberately excludes it rather than papering over it.

[Jammy2211]

Shipped.

• PR fix: natural_logpdf clamps a genuine -inf to -1.8e308 #1534 MERGED as 5c391fd, head 1e3b1a0, +90/-1 over 2 files.
• CI green on every leg: unittest (3.12) 1m47s of test time, unittest (3.13) 2m10s, unittest-nojax 58s, docs / docs-build 1m35s.
• Full suite 2203 passed / 36 skipped; baseline measured on the branch point 6e2d8c8 with changes stashed: 2190 / 36.
• Label pending-release — not released.

What landed

autofit/messages/interface.py — natural_logpdf now passes neginf=-xp.inf, posinf=xp.inf, so only NaN is replaced. A genuine -inf survives the reduction instead of becoming -1.7976931348623157e+308.

test_autofit/messages/test_natural_logpdf_neginf.py — 13 tests covering both halves of the reduction (the -inf half that was broken, the NaN half that was correct and is now pinned so fixing one cannot break the other), in-support values across 8 points, and the general property over Uniform / LogUniform / LogGaussian that no family reports a finite density off its support. Verified by inversion: reverting the fix fails 2 of the 13.

Still open, unfiled

TruncatedGaussianPrior's message returns finite logpdf well outside its limits — -8.20 at -1.0 for a (0, 3) support. That is separate looseness in TruncatedNormalMessage, not this clamp, and the prior-level log_prior_from_value is correct there (which is why the P6 property tests pass). The new general-property test excludes it explicitly rather than asserting something weaker across all families. Worth its own task if the message-level API is ever trusted directly.

---

Generated by Claude Code