Files
MeanField/experiments/SHAPE_FIELD_WEAK_RESPONSE_2026-09-18.md

161 lines
9.5 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Shape–field weak response investigation — 2026-09-18
## Prespecified scope
Follow up the large, nearly cancelling shape/field Newton direction for n=1.5,
J=0 without a central-density constraint. This study distinguishes within-element
physical field transport from surface variation and tests the specific hypothesis
that the Q3 displacement/Q4 reference geometry mismatch contributes to weak
response. Cancellation by itself is not evidence of a gauge or bad element orders.
The baseline branch is `research/shape-field-weak-response-20260918`, based on
`db854c0f7cdf5137e377273499f0ac5b6053b752`. The isolated order variant will be
`research/shape-field-displacement-q4-20260918`. Preserve original checkout,
immutable executables, source revisions, mesh/input hashes and serial run records.
## Observation and controls
The diagnostic prepares a frozen state with fixed normalization, independently
verifies its full-J linear defect, saves every value-block Jacobian contribution,
and checks restoration of accepted buffers and operator action. Three small raw
Taylor trials test the same original operator. Physical observations use independent
grid functions and the prepared displacement; no accepted state is modified.
At common quadrature orders 10 and 14, record physical scalar pullback changes,
transport `grad(f).v`, and Eulerian changes `delta f - grad(f).v`. Density and
potential gradients are broken/local: facet jumps and RT Piola variations are not
included, so these observations cannot certify a pure relabelling nullspace.
Surface radial parameters are unweighted nodal samples. An independent affine
identity projection checks ambient displacement representation, separately from
the actual r-squared radial extension. Small-span SVDs describe only the sampled
block directions, never the full Jacobian rank.
Baseline frozen cases: the recorded n=1.5 seed and late directions at eta=0.03,
and a fresh cold solve at the six-update coupled endpoint. Then a fresh six-update
baseline with gravity-Schur inverse, gravity-curvature globalization, restart 160,
maximum 200 Krylov iterations, eta=0.03, and a fresh final direction. Match this
live run with the Q4 variant from the same analytic seed, including a Q4 seed
probe. Cases, builds and expensive audits run serially. A failed linear gate or
preparation remains a failure; do not silently relax controls.
## Order intervention (conditional on baseline observation)
Only change `Displacement::vectorOrder` from `3 + uniform increment` to
`4 + uniform increment` on the variant branch; keep uniform increment zero.
The loaded reference geometry remains Q4, density/potential L2 Q2, gravity RT
index 2, and enthalpy H1 Q3. The shared displacement collection means the scalar
surface trace also becomes Q4. This introduces shape modes and changes derived
quadrature orders and cost; it is not a pure interpolation-error intervention.
Document exact source line, date, rationale, DOFs, and compiled quadrature changes.
Use a separate build directory and new analytic seeds: existing coefficient
snapshots cannot be transferred unchanged between orders.
## Decision criteria
Compare verified linear accuracy, accepted step lengths, within-variant growth
of weak-direction indicators, common physical norms, virials, positivity/EOS
metrics and elapsed cost. Across orders, coefficient norms and discrete residuals
are not physical error measures. Q10/Q14 agreement diagnoses integration
sensitivity, not convergence certification. Virial <=1e-5 remains required but
cannot replace internal pointwise checks. Six updates are a bounded mechanistic
comparison, not evidence of high-n/high-rotation convergence.
## Results
Baseline observations are recorded below. No production default has been changed on this baseline branch.
### Reproducible baseline
Evidence root: `../../diagnostics/shape_field_weak_response_2026-09-18` relative
to this checkout. Frozen probes `baseline_probes_v1` use source `d79db94`; the
six-update control `q3_live_v1` uses `2c4a5e0`. Both include immutable binaries,
mesh/input hashes, commands, clean-source provenance, and independent elapsed
and CPU clocks. All full-J linear gates passed. Independent audits
`baseline_probes_audit_v2.json` and `q3_live_audit_v1.json` pass; accepted buffers,
factors and restored actions agree exactly. Q3 curved/tensor/geometry regression:
5 cases, 21,368 assertions passed.
The first checker version omitted the new cache-preservation invariant from its
schema. Version 2 requires it. A small finite synthetic matrix reproduced
NumPy/Accelerate matmul warnings; direct finite-checked contractions avoid that
backend path, without changing tolerances. Raw FEM artifacts remain unchanged.
The artifact checker passes 38 tiny algebra/corruption fixtures. The angular
analyzer independently checks its degree-8 recurrence, orthogonality, cubic
symmetries, rank, and file binding on analytic fixtures.
| Observation | Seed | Old late state (12 updates) | Fresh control (6 updates) | Coupled endpoint (6 updates) |
|---|---:|---:|---:|---:|
| Normalized residual | 2.73659e-4 | 1.39670e-5 | 1.49720e-5 | 1.24583e-5 |
| Sampled block-span minimum singular response | 5.80574e-2 | 1.43543e-4 | 8.16558e-4 | 6.94894e-4 |
| Sampled surface radius span / nodal mean | 3.42403e-5 | 1.24122e-2 | 1.09849e-2 | 1.11125e-2 |
The largest sampled singular response remains approximately 22.1–22.6. The SVD
uses the six nonzero disjoint value-block pieces of each p, normalized by their
coefficient norms. It is not a full-J spectrum or rank test. In particular,
comparing the 12-step late state with six-step endpoints does not demonstrate a
causal benefit of the coupled correction: the matched six-step control has a
slightly stronger minimum response than the coupled endpoint.
Late q14 Eulerian/pullback change norm ratios are density 0.05769, enthalpy
0.18870, potential 0.01288, versus seed 0.3243, 0.3080, 0.1993. This is substantial
transport compensation but not physically invisible motion. Broken DG gradients
omit jumps. The maximum q10/q14 relative difference in the three original cases'
Eulerian squared integrals is 1.69e-5, an integration sensitivity observation.
The radius observations expose a validation gap: the old late state has nodal
radii 0.997674–1.010093, while its q18 P(h)-based virial error is 4.06909e-6.
The fresh six-step control has q18 virial error 4.22590e-6 and reproduces the old
six-step residual trajectory. Neither is converged or physically certified;
negative density samples remain. Virial success can coexist with appreciable
nonrotating surface asphericity at these unfinished states.
`q3_surface_modes_v1.json` contains unweighted fits, not surface-integral power.
The 12-step late direction is 99.9995107% cubic-invariant in sampled squared
norm, but degree <=8 spherical harmonics leave 0.893679 of its nodal norm.
Degree <=2 leaves 0.981646. Thus this is predominantly finer cubic-symmetric
structure, not simple expansion, translation, or quadrupolar deformation.
This is consistent with mesh/discretization coupling, without proving its cause.
The affine control on the actual stellar mesh gives Q3 relative value error
9.01277e-5 and gradient error 1.23472e-3. This warrants the isolated Q4 comparison,
but adding surface modes could also worsen weak response.
### Structural interpretation and a rejected shortcut
At J=0, bulk Bernoulli rows are unweighted moments. Their direct shape derivative
at fixed reference-field coefficients is `integral(q * B * div(v))`; the EOS
shape column similarly contains its local closure imbalance. Thus pointwise
near-equilibrium shape information increasingly relies on gravity and mechanical
coupling. Small projected residuals alone do not prove small pointwise imbalance.
Current shape rows are the prescribed lift's transpose applied to volume force.
For regular surface h~s, density~s^n and pressure~s^(n+1), weakening near-surface
coupling as n increases is plausible. DG jumps and the independent RT field
prevent concluding exact discrete redundancy from the continuum identities.
Relevant assembly: `libmeanfield/impl/operators/prepared_hydrostatic_equilibrium.cpp`
(lines 864–911, 1354–1397), `operators/kernels/barotropic_closure_kernels.cpp`
(line 655), and `operators/prepared_stellar_equilibrium.cpp` (lines 806–811).
A naive replacement of mechanical shape rows by surface Bernoulli moments is
rank deficient at the desired pointwise B=0, J=0 root with the current trace
spaces. The actual closed stellar surface has 96 quads, 192 edges, 98 vertices.
Continuous surface Qp has 96*p*p+2 degrees of freedom; broken Q(p-1) potential
traces supply only 96*p*p. Q3 surface therefore has 866 rows versus 864 DG Q2
stellar trace coefficients, after eliminating enforced h=0 trace variations.
C adds no independent trace dimension; the direct shape measure term vanishes
with B. Surface-only Q4 widens this hypothetical deficit to 1538-864=674.
This concerns that proposed replacement, not a proven deficiency of the current
mechanical Jacobian. Any boundary formulation requires a compatible trace/test
space and explicit rank/stability analysis.
An invertible transport coordinate/preconditioner change preserves roots and may
improve linear algebra, but cannot itself change the exact Newton direction or
remove a true weak physical response. Shape representation, step control, and
operator formulation remain distinct hypotheses.
### Q4 intervention
Pending on its isolated branch. Its rationale is to remove measured ambient
affine representation error and test the effect on the observed weak direction;
the additional surface modes and quadrature changes are explicit confounders.