Depth-Optimal Native Decomposition Preregistration¶
Date: 2026-05-06
This preregistration defines an offline-first benchmark for reducing compiled Kuramoto-XY circuit depth by targeting backend-native Heron gate structure. It does not submit IBM jobs, reserve backend time, or authorise QPU spend.
Scientific Question¶
Can an XY-aware native decomposition reduce depth, two-qubit gate count, or routing overhead relative to generic Pauli-evolution and current XY compiler baselines without changing the promoted physical observable?
Claim Boundary¶
Supported after successful execution and analysis:
- resource comparison between generic Trotter, current XY compiler, and native-targeted decomposition;
- circuit-equivalence evidence from exact small-system unitary or observable checks;
- backend/layout-specific resource reductions for the sampled transpiler and calibration window.
Blocked even after a positive result:
- hardware coherence improvement without a separately approved QPU run;
- backend-general optimality;
- quantum advantage;
- correctness for all Hamiltonians outside the preregistered Kuramoto-XY/FIM families;
- substitution of the promoted physics circuit if unitary or observable equivalence fails.
Comparator Circuits¶
Every readiness run must compare:
- generic Qiskit Pauli-evolution/Trotter circuit;
- current
scpn_quantum_control.phase.xy_compileroutput; - native-targeted candidate decomposition;
- optionally, provider-transpiler optimized versions of each comparator with identical optimization level and seed list.
No comparator may be removed after seeing resource results.
Native-Target Candidate Rules¶
The candidate decomposition must:
- preserve the intended
XX + YYinteraction angle convention; - record any conversion to
rxx,ryy,rzx,ecr,cx,sx,rz, or backend-supported native operations; - avoid hidden changes to Trotter order, evolution time, or coupling threshold;
- expose a deterministic construction path from committed code;
- emit metadata identifying all approximations and basis-gate assumptions.
If a backend-native operation is unavailable in the selected backend target, the candidate must fall back to a documented basis-gate decomposition and record that fallback.
Offline Readiness Matrix¶
Default no-QPU readiness scope:
| Field | Value |
|---|---|
n |
4, 6, 8 |
| Families | DLA parity, popcount controls, FIM lambda=0 and lambda=4 |
| Depths/reps | representative shallow, promoted, and stress depths |
| Layout methods | default plus layer-selective where available |
| Transpiler seeds | 0, 1, 2, 3, 4 |
| Optimization level | fixed and recorded |
Readiness outputs:
- circuit depth;
- two-qubit gate count;
- basis-gate histogram;
- SWAP or routing-overhead count where available;
- transpilation wall time;
- unitary distance for
n=4where tractable; - observable-level agreement for
n=6,8where full unitary comparison is not practical.
Equivalence Gates¶
Before promoting a resource reduction:
- for
n=4, compare candidate and baseline unitary or statevector evolution up to global phase; - for
n=6,8, compare preregistered observables from exact or high-confidence classical simulation where tractable; - reject if parity survival, magnetisation-sector survival, or selected correlators differ beyond numerical tolerance before noise is applied;
- record the tolerance and simulator used.
Default tolerance:
- unitary/process comparison for
n=4:1e-8absolute Frobenius-normalized distance where feasible; - observable comparison for larger
n:1e-6absolute difference for promoted scalar observables.
Optional Hardware Scope¶
If the offline gates pass and QPU execution is later approved, use a minimal resource-to-noise validation block:
| Field | Value |
|---|---|
n |
4 |
| Families | one DLA parity pair and one FIM pair if still needed |
| Depths | one promoted signal depth and one stress depth |
| Decompositions | current XY compiler, native-targeted candidate |
| Repetitions | 6 per circuit |
| Shots | 4096 |
| Readout states | prepared states plus 0000 and 1111 |
| Readout shots | 8192 |
Circuit ceiling: <= 160 circuits.
IBM-reported QPU-time ceiling: 12 minutes.
Live Readiness Gates¶
Before any hardware submission:
- regenerate all comparator circuits from committed code only;
- live-transpile all comparator circuits on the selected backend/layout;
- reject if native-targeted decomposition fails equivalence gates;
- reject if the candidate increases depth or two-qubit count versus the current XY compiler for the promoted circuit family;
- reject if the candidate only improves one seed and worsens the median across fixed seeds;
- record backend, calibration timestamp, target basis, optimization level, transpiler seeds, circuit count, shot count, and estimated QPU minutes;
- get explicit approval immediately before submission.
Analysis Plan¶
Primary offline endpoints:
- median and worst-case depth delta versus current XY compiler;
- median and worst-case two-qubit gate-count delta;
- routing overhead delta;
- equivalence-check pass/fail;
- seed sensitivity.
Secondary endpoints:
- estimated error-load proxy from gate counts and calibration data where available;
- decomposition-specific basis-gate histogram;
- interaction-layer parallelism achieved.
Hardware endpoints after separate approval:
- parity leakage and exact-state retention by decomposition;
- depth-normalized leakage comparison;
- readout-corrected comparison only where exact-state or full-basis calibration exists.
Falsification Rules¶
The native-decomposition claim is rejected or downgraded if:
- equivalence checks fail;
- resource improvements are not robust across seeds;
- gains disappear after live backend transpilation;
- the candidate reduces depth but increases two-qubit error load enough to worsen the error proxy;
- optional hardware counts show no retention/leakage benefit despite resource reduction.
Negative results are publishable as compiler-boundary evidence for this circuit family.
Output Artefacts¶
Expected paths after offline readiness:
data/phase3_native_decomposition/native_decomposition_readiness_<date>.json;data/phase3_native_decomposition/native_decomposition_transpile_rows_<date>.csv;data/phase3_native_decomposition/native_decomposition_equivalence_rows_<date>.csv;docs/campaigns/phase3_native_decomposition_readiness_<date>.md.
Expected paths after approved hardware execution:
data/phase3_native_decomposition/native_decomposition_counts_<backend>_<timestamp>.json;data/phase3_native_decomposition/native_decomposition_summary_<date>.json;docs/campaigns/phase3_native_decomposition_manifest_<date>.md.
Each artefact must include comparator identity, backend target, basis gates, transpiler seeds, depth and gate summaries, equivalence diagnostics, raw counts where applicable, SHA256 hashes, and reproduction commands.
Submission Boundary¶
This preregistration is complete. Hardware execution remains blocked until offline readiness artefacts, backend selection, budget confirmation, and explicit approval are completed in a separate task.