Repository navigation
diagnose() at d >= 8: the exponential action now propagates the trajectory — D18 and the resolvent/non-normality layers are the remaining bottlenecks #178
Description
Context
PR #176 (issue #162) moves relaxation-trajectory propagation to the exponential action (expm_multiply) for n = d^2 >= 64. That speeds up the propagation itself 30–46× at n = 256–1024. It does not make diagnose() fast at those sizes, and the PR says so.
Measured (single-threaded BLAS, damped transverse-field Ising chain, bootstrap_B=20)
| d | diagnose() total |
relaxation layer | of which propagation | non-normality | resolvent | transient |
|---|---|---|---|---|---|---|
| 8 | 5.9 s | 2.11 s (fits/bootstrap) | 0.05 s | 2.00 s | 1.17 s | 0.60 s |
| 16 | > 25 min (killed) | — | — | — | — | — |
The profile at d = 8 was also noisy with ResolventConvergenceWarnings: the power iteration in resolvent_norm did not converge in 200 steps at many frequencies.
Candidates, ordered by value/effort
- D18,
lep.initial_state_sensitivity. It materialisessla.expm(L t_eval)once and then applies it only ton_samplesvectors. This is the same pattern Trajectory backend: use matrix-exponential action instead of materialising expm #162 removed. A block actionexpm_multiply(L*t_eval, B)withB = [vec(rho_k)]fits directly.- One caveat: for
n0columns, the deterministic exact-norm branch of SciPy requires||A||_1 <= 63.36 / n0(condition 3.13). The sub-stepping fromnumerics.propagationhas to scale its threshold byn0, or the randomisedonenormestpath, which touches globalnp.randomstate, comes back.
- One caveat: for
- Resolvent / non-normality layers. They dominate at d = 8 and are the likely reason d = 16 does not finish. This needs its own profile at d = 16 (which frequencies, how many
resolvent_normcalls, whether the power iteration hits its cap) before choosing a remedy. transient.py. It computessvdvals(expm(L t))per grid point. The propagator norm genuinely needs more than one action, so the action does not apply directly here. A Krylov or randomised norm estimator would be a methodology change and belongs in its own issue.
Acceptance for item 1
- D18 is unchanged bit-for-bit for
n < 64. - For
n >= 64, dense and action agree within the declared tolerance. - Global RNG state stays untouched: D18 draws its samples from its own
default_rng, andonenormestmust not touch the global state either. - The backend is recorded, as in
RelaxationResult.trajectory_backend.
Activity
perf(#178): add deterministic block exponential action for D18
perf(#178): route D18 Haar ensemble through block action
feat(#178): record D18 propagation backend
fix(#178): keep LepResult dataclass field ordering valid
fix(#178): satisfy sparse union typing in block path
test(#178): pin D18 block determinism and dense parity
docs(#178): document D18 block action semantics
test(#178): persist D18 backend in report JSON
style(#178): normalize first-party import order
fix(#178): attach D18 backend provenance to LepResult
style(#178): separate third-party and local imports
style(#178): satisfy ruff collection literal rule
fix(#178): keep block cost estimate statically typed
perf(#178): use conservative D18-specific auto crossover
docs(#178): describe D18-specific auto policy
test(#178): pin D18-specific auto crossover
docs(#178): record conservative block crossover rationale
marcohost33-maker commentedon Sep 29, 2026
2026-09-30 D18 implementation / crossover falsification
Draft PR #181 implements item 1 on top of merged #176: D18 now has a single-time block exponential primitive and can evolve the Haar RHS ensemble with
expm_multiply(L*t, B)without materialising the propagator.Two details from the implementation changed the naive plan:
n0 = B.shape[1]; using feat(#162): propagate relaxation trajectories by the exponential action #176's one-vector substep threshold for a 10-column Haar block would re-enter randomisedonenormestand consume globalnp.random. perf(#178): propagate D18 Haar ensemble by deterministic block action #181 therefore uses the same safety margin divided byn_rhs, with a regression probe on every block call plus an explicit global-RNG invariance test.expmand applies it to a small block; the trajectory benchmark formed many exponentials. An exploratory check on the same damped-Ising family showed that block action can win at n=64, lose at n=256, and win clearly again by n=1024 on the test environment. That falsifies a simple monotone transfer of the trajectory threshold. perf(#178): propagate D18 Haar ensemble by deterministic block action #181 therefore keeps dense auto below a conservativeBLOCK_ACTION_MIN_DIM=512; forced action remains available for differential evidence.The PR also records
LepResult.initial_state_backend, pins small-system legacy arithmetic, dense/action numerical agreement, report-JSON persistence, and keeps D18 advisory (no classifier change).Current exact-head gates on #181: Quality Contract, QuTiP and CodeQL are green; full Python matrix and reusable pytest are still executing. Do not treat this comment as merge evidence until those exact-head runs finish.
After item 1 lands, item 2 should start from a fresh d=16 profile rather than assuming the d=8 proportions remain representative: record frequency count, non-converged
resolvent_normcalls, iteration counts/cap hits and time by non-normality/resolvent subroutine before choosing an algorithmic replacement.perf(#178): add deterministic block exponential action
perf(#178): use block action for D18 at n>=64
feat(#178): record D18 propagation backend
test(#178): cover D18 block action and RNG isolation
test(#178): compare D18 action within float64 tolerance
docs(#178): record D18 block-action change
style(#178): satisfy import ordering
typing(#178): narrow sparse block dense fallback
marcohost33-maker commentedon Sep 30, 2026
Item 1 (D18) ist als Draft-PR #182 umgesetzt. Kleine Systeme n=d^2<64 behalten den historischen Dense-Pfad exakt. Ab n>=64 werden die Haar-Zustände in derselben RNG-Reihenfolge als Block mit
expm_multiplypropagiert; die Condition-(3.13)-Schwelle wird durch n_rhs geteilt, damit SciPy auf dem deterministischen Exact-Norm-Pfad bleibt und globalesnp.randomunberührt bleibt.LepResult.initial_state_backenddokumentiert den Backend-Wechsel additiv. Neue Tests vergleichen gegen die Dense-Referenz und pinnen RNG-Isolation. Ruff- und mypy-Befunde der ersten CI-Runden wurden bereits behoben; Resolvent/Non-Normality bleiben bewusst separater nächster Profiling-Slice.