tboudreaux/MeanField
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Record matched central-constraint control results

Browse Source
This commit is contained in:
Emily Boudreaux
2026-09-16 15:04:33 -04:00
parent a1d7d4e679
commit 4ec1db230f
1 changed files with 53 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
experiments/CENTRAL_CONSTRAINT_HYPOTHESIS_2026-09-16.md
View File

@@ -20,19 +20,62 @@ residual after that source is removed. This is separate from the late Krylov
failure: the initial no-phase J=.01 control retained that failure while reducing
the unbordered residual, with negligible change in virial balance.
The next matched control is n=1, J=.05 with and without the constraint, using the
same mesh, seed, normalization, preconditioner, and solver settings. Record formal
solve status separately from the retained state's physical quality. Compare
energy and force virial errors against the required 1e-5 milestone, mass and J,
actual angular velocity, the unbordered residual, and location-resolved central
enthalpy/EOS/Bernoulli errors. The phased baseline's sampled maximum EOS mismatch
is about .934% of the density scale despite virial errors near 3e-7; its cause
requires this control. Higher-spin no-phase cases follow if the matched control
is satisfactory. J or angular velocity in reference units is not itself a
measured fraction of critical rotation.
## Completed matched controls
At n=1, J=.05, the phase comparison uses the same mesh, Lane–Emden seed,
normalization, preconditioner, restart length 40, and stopping controls. Physical
measurements below use quadrature order 18 and each run's last accepted state.
| Measurement | Central constraint, R40 | No central constraint, R40 | No central constraint, R160 |
|---|---:|---:|---:|
| Sampled maximum EOS mismatch / density scale | .00933951 | .000430594 | .000430592 |
| Sampled Bernoulli range | .0145751 | .00130398 | .00130398 |
| Unbordered normalized residual | 4.54227e-6 | 6.03609e-9 | 3.38036e-11 |
| P(h) energy virial error | 3.01617e-7 | 3.06913e-7 | 3.07200e-7 |
| P(h) force virial error | 3.18803e-7 | 3.24099e-7 | 3.24382e-7 |
| Absolute mass error | 1.22944e-11 | 1.23133e-11 | 1.74860e-13 |
| Absolute angular-momentum error | 1.71698e-12 | 1.72180e-12 | 1.54182e-14 |
Removing the phase lowers the sampled maximum EOS mismatch by 21.69 times and
the Bernoulli range by 11.18 times, while preserving virial errors well below
the required 1e-5 milestone. Actual angular velocity changes only from
.185891548249 to .185891547385. These matched observations support removing
the n=1 phase to improve interior consistency. The extrema are quadrature
samples, not pointwise error certificates or errors against an independent
rotating reference solution.
Both R40 runs accept five full Newton steps and then fail the same late linear
convergence gate: after 200 linear iterations their true relative residuals are
about .505 against the requested .03. Phase removal does not cure that failure.
The separate no-phase R160 run accepts seven full Newton steps and meets the
requested nonlinear relative tolerance 1e-8 with absolute tolerance zero. Its
final normalized residual is 3.38036e-11. The larger Krylov restart window and
removal of the physical constraint address different observed problems.
Data provenance in the audit workspace: the phased baseline is
`diagnostics/campaign_2026-09-16/n1_J0p05`; the matched and larger-window controls
are `diagnostics/numerical_campaign_2026-09-16/controls/n1_J0p05_no_central_R40`
and `n1_J0p05_no_central_R160`. Measurements come from `final_physical.csv`;
iteration outcomes and settings come from `newton.csv` and `metadata.txt`.
These controls use the existing diagnostic driver's optional model constraint;
they do not establish that this example/regression commit builds or passes tests.
Higher-spin and higher-index controls remain separate work. J or angular
velocity in reference units is not itself a measured fraction of critical
rotation. These results do not establish high-rotation robustness or change the
optional n=3 constraint semantics described above.
## Focused regression status
The added user-API regression covers the previously missing explicit-seed,
no-phase context construction and solver borrowing/release path. Existing tests
already cover the square fixed-J root and its Jacobian column. This lifecycle
test performs no Newton solve and makes no physical-convergence claim. Retain
the existing phased n=3 tests; unphased n=3 convergence is not a requirement.
Build and focused tests are **pending** for this branch. The new test is
`An Explicit Lane Emden Seed Builds A Fixed Angular Momentum Context Without Fixing Central Density`
(`"[user-api][context][seed]"`). Also run the existing
`Fixed Angular Momentum User API Generates Rotation And Its Composable Solver Border`
and `Fixed Angular Momentum Root Uses Its Generated Angular Velocity In Every Physical Row`
tests. No compilation or numerical runs were performed in this worktree.