P2.1 Spherical potentials: backend-agnostic namespace-swap#893
Merged
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## feat/backends #893 +/- ##
===============================================
Coverage 99.92% 99.92%
===============================================
Files 228 228
Lines 34213 34461 +248
Branches 712 716 +4
===============================================
+ Hits 34186 34434 +248
Misses 27 27 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
jobovy
force-pushed
the
feat/P2.1-spherical
branch
from
7628b5a to
9432cca
Compare
The P2.1 migration of PowerSphericalPotential._evaluate switched to an xp.where-based form, so _evaluatePotentials now returns a 0-d numpy scalar (with .shape == ()) instead of a Python float for scalar input. This made fE/fQ take the array (reshape) branch instead of the scalar `return out[0]` branch, dropping coverage of that line in isotropicPowerLawdf.py (L126) and osipkovmerrittPowerLawdf.py (L152) and failing codecov/project (+2 misses). Add direct scalar-input assertions (matching the existing constantbetaPowerLawdf pattern) to restore 100% coverage. Test-only change; numpy code path is byte-identical. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The backend rewrite added safe_h = where(h==0, 1, h) to avoid 1/0, but at the exact shell edge R==a the projected surface density genuinely DIVERGES (the 1/sqrt(a^2-R^2) limb singularity), so 1/(2 pi a h) -> inf is the physically correct value and what the original numpy code returned. Masking h there silently replaced the divergence with a finite ~1/(2 pi a). Drop safe_h and use h directly: inf at R==a (matches original, byte-identical for all R), the only consequence being a singular gradient exactly at the edge (which is real). The R>a branch remains NaN-safe via safe_arg (sqrt(negative) guard) above. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Migrate the remaining numpy-only / value-branching compute methods of the spherical-potential family to the galpy.backend (xp) namespace layer, so they run and differentiate under numpy / jax / torch: - BurkertPotential._surfdens, ._revaluate - HernquistPotential._surfdens, JaffePotential._surfdens, NFWPotential._surfdens - NFWPotential._evaluate, NFWPotential._mass - SphericalPotential._mass - PowerSphericalPotentialwCutoff._dens, ._ddensdr, ._d2densdr2 The four cuspy _surfdens methods used numpy.sqrt(R^2-a^2+0j) complex arithmetic plus an `if Rma == 0` scalar branch for the R == a removable singularity. These are rewritten with a complex sqrt built from a NaN-safe real argument (xp.astype(safe, complex128)), xp.real for the generic branch, and an xp.where selecting the closed-form edge limit at R == a. Per the safe-denominator rule (jax/torch where evaluates both branches), every 1/Rma, 1/(R^2-a^2), 1/(r^2-a^2) in the dead branch is guarded so it cannot inject NaNs into reverse-mode autodiff. NFW._evaluate and Burkert._revaluate used scipy.special.xlogy purely for its 0*inf -> 0 convention as r -> infty (the prefactor 1/r, 2/x is never zero for finite r); these are reproduced exactly with a NaN-safe xp.where, so no scipy dependency remains. All migrated numpy paths are bit-identical (atol=0) to the originals, including the R == a edge and the r == 0 / r == inf limits. Left numpy-only (genuine scipy.special, no backend-agnostic replacement yet): base TwoPowerSphericalPotential._evaluate/_Rforce/_zforce/_R2deriv/_Rzderiv/ _mass (hyp2f1, gamma), PowerSphericalPotential._surfdens (hyp2f1), PowerSphericalPotentialwCutoff._evaluate/_Rforce/_zforce/_R2deriv/_z2deriv/ _Rzderiv/_mass (gamma/gammainc/hyp1f1). tests: extend test_backend_spherical.py with numpy-vs-jax-vs-torch parity for the migrated _surfdens (grid spanning R<a, R==a, R>a), an edge-finiteness check, grad-vs-finite-difference off the edge plus grad-finiteness at the edge (exercising the safe-where guards), the PowerSphericalPotentialwCutoff closed- form densities, and add _evaluate to the NFW and Burkert AD cases. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 10, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 12, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 13, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 15, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 17, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 19, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 19, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jobovy
added a commit
that referenced
this pull request
Jun 25, 2026
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of the P2.x potential namespace-swap series (→ #892 integration branch).
Migrates the spherical family to backend-agnostic compute (numpy/jax/torch) via the PR0 convention (
xp = get_namespace(...); private compute layer only).Migrated:
SphericalPotential(the 3D→1D reduction layer — covers the whole family), Dehnen(Core)Spherical, Hernquist, Jaffe, NFW, TwoPowerSpherical (base dens/z2deriv), PowerSpherical/Kepler, HomogeneousSphere, SphericalShell, Burkert.Deferred to Pspecial (scipy.special/interpolate): TwoPower
_evaluate(hyp2f1/gamma), NFW_evaluate(xlogy), the_surfdensof Hernquist/Jaffe/NFW/Burkert (data-dependent complex-sqrt branch), Einasto, interpSpherical.Hazards fixed: torch lacks
power→**; in-placeout[r2==0]=-inf→safe-denominatorxp.where; piecewiseif r<a→shape-polymorphicxp.where;numpy.fabs→xp.abs(a real cross-namespace bug caught by the test).Tests:
tests/test_backend_spherical.py— 95 passed, 15 skipped (numpy/jax/torch parity rtol 1e-12 + grad-vs-FD + grad==−Rforce). Numpy path byte-identical (max diff 0.0 on spot checks).🤖 Generated with Claude Code