Exact statevector baseline for the kingston DLA-parity data — 2026-07-21¶
Status: completed. Addresses audit findings B-9/B-10 (exact classical reference
+ mitigation-ablation context for the promoted DLA-parity data).
Code: src/scpn_quantum_control/analysis/dla_parity_exact_baseline.py,
runner scripts/run_dla_parity_exact_baseline.py, tests
tests/test_dla_parity_exact_baseline.py (40).
Artifact: data/dla_parity_exact_baseline/dla_parity_exact_baseline.json.
Source data: data/phase2_dla_parity/ (promoted Phase-2 reduced A+G, ibm_kingston, n=4).
What the DLA-parity campaign measures¶
The promoted runs execute the Kuramoto-XY Trotter circuit
H_XY = Σ K_nm (X_n X_m + Y_n Y_m) + Σ ω_n Z_n
(K[i,j] = 0.45·exp(-0.3·|i−j|), ω = linspace(0.8, 1.2, n), t_step = 0.3)
and report parity leakage — the fraction of shots that end in the opposite
excitation-number parity to the prepared state — for an even (0011) and an odd
(0001) initial state across ten depths (2…50).
The exact reference¶
Every gate in the circuit conserves total excitation number: rz trivially, and
the paired rxx(θ)·ryy(θ) on each nearest-neighbour edge because XX and YY
commute on a pair and XX+YY is the number-conserving hopping generator. So the
ideal, noiseless final state never leaves the prepared parity sector and its
parity leakage is exactly zero.
The module computes this directly by statevector (feasible for n ≤ ~14). The
reconstruction is verified bit-for-bit against the campaign builder
(scripts/phase1_mini_bench_ibm_kingston.build_xy_trotter_circuit): statevector
fidelity = 1.0. The exact leakage is < 1e-9 at every depth, for even and odd
sectors, and across widths n = 2…8.
Exact vs hardware¶
| Depth | exact leakage | hardware leakage (even / odd) |
|---|---|---|
| 2 | 0 | 0.084 / 0.082 |
| 6 | 0 | 0.153 / 0.147 |
| 20 | 0 | ~0.30 |
| 50 | 0 | 0.422 / 0.422 |
Hardware leakage grows monotonically with circuit depth to 42 % at depth 50,
while the ideal reference is pinned at 0. All observed leakage is device noise
(it accumulates with gate count), and the even/odd DLA-parity asymmetry
(asymmetry_relative, ~0.01–0.09 and vanishing at large depth) sits inside that
noise floor rather than reflecting coherent parity-selective dynamics.
Mitigation-ablation context (B-10)¶
The promoted packs ran resilience_level = 0 (no ZNE/DD) and the Phase-2 readout
mitigation self-reports full_confusion_matrix_available = false (an
approximation, not a full inversion). This exact baseline supplies the ideal
leakage = 0 reference that a ZNE/DD ablation compares against; running the
ablation on real hardware is a separate owner-gated QPU submission (AUD-5/7),
now unblocked with a rigorous classical reference in place.
Authored by Anulum Fortis & Arcane Sapience (protoscience@anulum.li) Seat: 7f6b