Work

Pre-registration and result (16 Nov 2026): cycle-count large-deviation rate functions do NOT predict simulated-annealing time-to-solution on 100-node 3-regular QUBO instances

Year: 2026
Type: article
Abstract: Version 2. Version 1 of this record held the protocol and code alone and was published before the confirmatory run returned; this version adds the result, so the deposit order is itself part of the record. CLAUSE: the mean absolute error of a d-regular rate-function surrogate predicting log10 time-to-solution on 100-node 3-regular QUBO instances (d=3) is not at least 15% lower than that of the matched Erdos-Renyi rate-function baseline. VERDICT: FALSIFIED. Over 240 generated instances and 50 paired cross-validation folds, held-out MAE was 0.2475 for the degree-constrained arm and 0.2513 for the Erdos-Renyi arm: a 1.52% relative reduction against the 15% required. The direction is real (one-sided Wilcoxon p = 0.00016; the degree-constrained arm won 35 of 50 folds) and an order of magnitude too small. THE MORE INFORMATIVE RESULT: an intercept-only predictor scores 0.2490, between the two arms. The Erdos-Renyi baseline is worse than predicting the mean, the degree-constrained arm beats the mean by 0.6%, and a spectral-gap-only surrogate scores 0.2500. On this instance family, cycle-count large-deviation rate functions carry essentially no information about annealing time-to-solution under either null, so the margin the clause asks about was never reachable. This was visible only because the intercept-only and spectral-gap baselines were pre-registered as reported-regardless quantities. CONTROLS, all passed: enumerated cycle counts against the trace identities; planted targets in BOTH directions (88.1% and 87.7% MAE reduction), so the harness is not biased toward either arm and the measured 1.5% is a small real effect rather than a resolution floor; split-half reliability of the success probability (Spearman 0.956); shuffled-target leakage check (0.04% deviation from intercept-only). The shuffled-target control failed at a 16-instance pilot and was calibrated against instance count before the confirmatory run, without reference to the verdict axis; no amendment was made and the registered protocol ran as frozen. SCOPE: one instance family (3-regular, n=100, +/-1 couplings), one annealing schedule (T 2.0 to 0.05 over 1,000 sweeps, 200 restarts), cycle lengths 3-6. The reference optimum is the best energy found, not a proven ground state. Nothing here concerns quantum annealing hardware, other degrees or other instance sizes.
Language: English
Open Access status: green
OpenAlex ID: W7234298047
FWCI:
Cites:
Cited by:
Related to: 10