arXiv:2609.18655
Learning to Program Adaptive Non-Local Observables for Machine Learning
Yu-Ting Lee, Samuel Yen-Chi Chen, Huan-Hsin Tseng
Abstract
Quantum neural networks (QNNs) are typically built from variational quantum circuits (VQCs), which are limited by local measurements. Adaptive non-local observables (ANO) address this by jointly optimizing circuit parameters and multi-qubit measurements. However, existing ANO-based VQCs learn only a single static observable that remains invariant across all inputs. We propose QFWP-ANO, a novel architecture which employs a classical hypernetwork to dynamically program VQC parameters and/or non-local observables conditioned on each input. On multivariate time-series forecasting across four ETT datasets, QFWP-ANO achieves the lowest MSE in 16 of 20 settings and second-lowest in the remaining four, surpassing ANO-based and other strong baselines. On reinforcement learning tasks, QFWP-ANO consistently surpasses ANO-VQCs. Our results establish input-conditioned ANO as an effective approach for enhancing QNNs.
AI Summary
The paper asks whether conditioning trainable non-local observables on each input improves quantum neural networks relative to learning one fixed observable per task. It introduces QFWP-ANO, combining a data-reuploading variational circuit with a classical encoder–decoder hypernetwork that generates circuit rotation angles, observable parameters, or both. Unlike recurrent quantum fast-weight programmers, the model processes inputs independently without accumulated memory. Outputs are expectation values over all $\binom{n}{k}$ subsets of $k$ qubits, using trainable Hermitian observables with $4^k$ real parameters. Evaluation covers forecasting on four Electricity Transformer Temperature datasets and reinforcement learning on Acrobot and MiniGrid-SimpleCrossingS9N1, averaged over ten random seeds. Forecasting uses seven qubits, a lookback of 16, five horizons, and a sweep over $k=1,\ldots,7$ and circuit depths $D\in\{1,3\}$. Taking the best variant, QFWP-ANO achieves lowest MSE, including ties, in 16 of 20 settings and second-lowest in four; it improves on the static-observable ANO baseline in 17 settings. Observable-conditioned variants generally outperform angle-only conditioning. Four-qubit reinforcement-learning experiments report faster Acrobot learning and the highest final navigation reward for joint conditioning. The reported gains depend on the setting: Acrobot models ultimately reach similar returns, and forecasting does not improve over the best baseline everywhere. Ablations favor moderate locality, $k\approx2$–$3$; performance deteriorates at $k=7$, attributed to having only one output expectation value. Deeper circuits generally do not improve forecasting performance.
Error found: For ETTm1 at forecasting horizon 48, the reported MSEs imply an 8.3% improvement over the best baseline, rather than the stated 6.6%. The experiments do not isolate a specifically non-local quantum-measurement contribution from the added classical hypernetwork, since the allowed input-conditioned observable includes H(x)=g(x)I and no matched control excluding this classical prediction route is reported.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.