Publications

oordinated truck-drone delivery offers a promising means of reducing carbon emissions in last mile logistics, yet the joint optimization of truck routing and drone operations yields a strongly coupled, NP-hard combinatorial problem. We formulate this problem as a Quadratic Unconstrained Binary Optimization (QUBO) model under a single carbon emission minimization objective and benchmark it against Gurobi, simulated annealing, simulated quantum annealing, and a GPU-based digital annealer. Direct embedding of the full QUBO on D-Wave Advantage quickly becomes infeasible: the logical variable count scales as O(n3), rapidly exceeding hardware capacity and causing actual embedding failure for n ≥ 15 in our experiments. To overcome this scalability barrier, we propose a Hybrid Reverse Annealing (Hybrid RA) framework that decomposes each iteration into fixed-size sub-QUBOs and combines D-Wave quantum local search with classical improvement, keeping physical qubit usage at approximately 450 qubits independent of problem size. Exper iments on instances up to 35 customers show that Hybrid RA achieves 100% feasibility across all tested scales, reduces carbon emissions by 2–8% over a strong greedy baseline, outperforms the within-decomposition Hybrid SA baseline at all 7 tested sizes, and beats the Hybrid 2-opt baseline at 6 of 7 tested sizes; pooled paired Wilcoxon tests confirm both advantages are statistically significant (p < 0.001), supporting quantum-classical hybrid optimization as a practical near-term approach where the primary gain comes from decomposition-enabled scalability and the QPU contributes a statistically significant local-search improvement.

This paper proposes a Quadratic Unconstrained Bi nary Optimization (QUBO) framework for large-scale Automated External Defibrillator (AED) deployment, integrating distance decay probabilistic coverage with demographic weighting in a compact binary formulation. Taipei City, with up to 12,000 demand points and 1,458 candidate sites, serves as the real world testbed. A central structural observation is that a slack penalized budget/cardinality constraint expands into nc 2 all to-all ZZ couplings in the corresponding Ising form, creating an O(n2 c) dense interaction block. We show that this density is consistently associated with (i) practical minor-embedding failure on D-Wave Pegasus P16 beyond a moderate candidate-set size and (ii) degraded feasibility in shallow, penalty-based QAOA circuits, suggesting a shared graph-structure explanation for both observed behaviors.

Abstract—Satellite telemetry anomaly detection benefits from recurrent sequence models, but fully adapting those models is costly when trainable memory is tightly constrained. Quantum Parameter Adaptation (QPA) addresses this problem with a small quantum–classical generator that produces low-rank up- dates for a frozen GRU. We evaluate QPA-LoRA on 54 NASA SMAP telemetry channels against trainable-parameter-matched Quantum-Train (QT) and classical generators. We first correct a global split that silently became channel holdout and pooled resid- ual rankings distorted by channel-scale heterogeneity. With per- channel chronological splits, macro metrics, and command/slope features, Full-GRU macro AUROC rises from 0.53 to 0.60 and residual-inverted events fall from 31% to 13%. Under the most constrained ≈250-parameter budget, QPA-LoRA achieves the best event F1 among matched parameter-generation approaches while using 96.2% fewer trainable parameters than Full GRU and retaining 92.3% of its event-F1 performance. Its median event F1 is 0.403, compared with 0.393 for QT-LoRA and 0.339 for classical generation. This advantage is budget-dependent: classical generation leads near 550 parameters, while all families approach Full GRU near 1100. Motivated by static QPA’s input- invariant adapter, we further introduce State-Adaptive Quantum Parameter Generation (SAGE-QPA), which conditions generated factors and effective rank on telemetry state. Context, gates, and directly encoded quantum probabilities vary measurably, yet five- seed paired evaluation shows no stable F1 gain over QPA. Thus, QPA is most useful in the extreme-compression regime, while state-dependent quantum adaptation can be mechanistically ac- tive without improving prediction

Quantum-Train (QT) compresses a neural model by learning a small hybrid quantum– classical generator that emits the weights of a larger classical network. Standard QT treats the target as a flattened parameter vector and ignores recurrent matrices, gates, and coordi- nates. We introduce Structure-Aware Hierarchical Residual Quantum-Train (SHR-QT), which conditions weight generation on LSTM structure, uses gate-specialized heads, and reserves 50 direct corrections for training-sensitive coordinates. The benchmark uses ten Federal Reserve H.10 currency series from 2005–2025; 30-day, 54-feature sequences predict maximum global FX stress over the next five trading days. Models train on 2005–2017, stop on 2018–2019, develop on 2020–2023, and use a locked 20-seed comparison on the previously characterized 2024–2025 slice. SHR-QT has 480 trainable parameters and generates an 11,297-weight LSTM (95.75% compression); matched Vanilla QT has 478. In the locked comparison, SHR-QT reduces mean absolute error from 0.33557 to 0.32819, an improvement of 2.20% (paired difference −0.00738, 95% CI [−0.01456, −0.00020]; pt= 0.0445; 14/20 wins), although the Wilcoxon test is not sig- nificant (p = 0.1231). A decision-triggered but fixed 50-pair expansion yields a pooled 70-pair improvement of 1.75% (pt= 0.00152; pWilcoxon = 0.0176) and a 7.59-fold reduction in optimizer- seed variance. A training-sensitivity mask outperforms five random masks on development data. However, a matched Classical-SHR generator has lower numerical MAE than SHR-QT (0.32601 versus 0.32819), while both their difference (p = 0.230) and the quantum-by-SHR interaction (p = 0.435) are null. Same-target HAR-Stress and GARCH-Stress baselines also outperform SHR-QT. Thus SHR-QT improves allocation and stability within extreme-compression QT, but the evidence supports neither a quantum-specific advantage nor forecasting superiority

Learning systems may need to operate before labeled data from a newly encountered or under-observed environment have accumulated at scale. Reliability then depends not only on average accuracy, but also on whether the learned decision function changes substantially with the particular small training subset observed. We compare a four-qubit variational classifier with a 33-parameter linear classifier across 30 balanced training subsets and a fixed 200-point unseen evaluation set. At ten training examples, mean accuracy is 0.9988 versus 0.9172, while normalized decision-function drift is 0.0505 versus 0.5007. A seed- block bootstrap gives a 95% confidence interval of [−0.5550,−0.3111] for the quantum- minus-classical drift contrast, and a within-seed block-permutation test that recomputes all pairwise comparisons gives p = 2.0 × 10−5. The stability difference persists against a best-classical envelope that includes a parameter- near MLP, RBF SVM, RBF kernel ridge classifier, and regularized linear classifier, including optimizer-matched Adam configurations. It is nevertheless conditional: predictive differ- ences narrow as labels accumulate, and stability collapses gradually near a critical decision boundary. Representation choice also materially changes transfer to real tabular data. Angle-encoded accuracy exceeds the best-classical envelope on four of five datasets, but a joint accuracy-and-stability advantage appears on only Breast Cancer and Spambase (2/5). We therefore report training-sample-induced decision-function stability as a useful empirical evaluation axis for quantum machine learning, while making no claim of universal quantum advantage, a causal mechanism, complete domain adaptation, or hardware performance.