Polytopic (Nagumo face-by-face LP check) and SOS polynomial
(Prajna-Jadbabaie w/ CSDP) barrier attempts on operation mode.
**Polytopic (barrier_polytopic.jl):** the naive check on
inv2_holds ∩ precursor_tube_bounds fails — 16 of 18 faces can be
crossed under A_cl. This is EXPECTED: safety halfspaces alone form
a set too big for LQR to contract from everywhere. The correct
approach is Blanchini's pre-image iteration (max robustly controllable
invariant set). Sketched in the script; 2-3 days to implement properly.
**SOS (barrier_sos_2d.jl):** a working proof of concept.
CSDP returns OPTIMAL on a 2-state projection of the operation mode
(dn, dT_c) with:
X_entry = |dn| ≤ 0.01, |dT_c| ≤ 0.1
X_unsafe = dn ≥ 0.15 (high-flux-trip direction)
Dynamics = reduced 2×2 A_cl after LQR.
No disturbance (B_w projects to 0 in this subset).
Global decrease condition (-(∇B·f) SOS) instead of Putinar ∂{B=0}.
Result: a degree-4 polynomial B(x) satisfying all three barrier
conditions. Coefficients printed. First non-quadratic barrier
artifact for this plant.
Caveats:
- 2D projection loses precursor coupling.
- Disturbance ignored in this projection.
- Global-decrease is stronger than the Putinar ∂{B=0} condition;
the latter requires bilinear σ_b·B formulation (BMI) and
iterative solvers. Deferred.
- Scaling to 10-state degree-4 gives SDP ~ 1000×1000; CSDP may
choke. Mosek or MOSEK-free SDP (SCS) might handle.
JuMP, HiGHS, SumOfSquares, DynamicPolynomials, CSDP all added to
Project.toml.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
