Source-linked AI summary
A Quantum Variational Approach to Prototypical Recurrent Unit
Mahyar Sadeghi Garjan, Tommaso Cesari, Michel Barbeau
TL;DR
Quantum recurrent architectures can be resource-intensive or difficult to train, motivating a more compact approach to sequential forecasting. The paper introduces QPRU, a lightweight hybrid quantum–classical recurrent unit, and evaluates it against classical and quantum baselines. QPRU achieves strong forecasting performance while using fewer quantum parameters and shallow circuits.
Problem
Quantum recurrent architectures can remain complex, resource-intensive, or difficult to train because of circuit depth and measurement overhead.
Method
The paper introduces QPRU, a lightweight parameter-efficient quantum recurrent unit inspired by the classical PRU.
Results
QPRU achieves the best reported sinusoidal-benchmark losses and competitive stock-forecasting accuracy against classical and quantum models.
Takeaways & Limitations
QPRU combines strong predictive performance with fewer parameters, reduced quantum resource demands, and shallow circuits.
Abstract
from arXiv · showhide
We introduce a lightweight Quantum Prototypical Recurrent Unit (QPRU) that requires significantly fewer parameters than both classical recurrent architectures, such as Long Short- Term Memory (LSTM) and Gated Recurrent Unit (GRU), and quantum variants, including Quantum LSTM (QLSTM) and Quantum GRU (QGRU). Despite its compact design, the QPRU achieves competitive forecasting performance, matching state-of-the-art baselines while offering important structural and practical advantages, including enhanced scalability and a reduced number of trainable parameters.
1 Introduction
Sequential learning is challenging across multiple application areas, and quantum recurrent models seek efficient temporal modeling with fewer parameters. The paper proposes QPRU as a lightweight alternative and evaluates it against classical and quantum baselines.
- Sequential data learning is a fundamental machine-learning challenge in finance, language, healthcare, and signal analysis.
- Quantum recurrent models use quantum representational capacity and reduced parameterization to model temporal dependencies efficiently.
- QPRU reduces parameter count relative to QLSTM, QGRU, LSTM, QRU, and PRU while maintaining forecasting accuracy on time-series benchmarks.
- The paper reviews related work, introduces variational quantum circuits and QPRU, then evaluates performance against classical and quantum baselines.The paper concludes by summarizing findings and outlining future research directions.
2 Related Work
Prior work spans classical and quantum recurrent models, with quantum approaches showing promise but often imposing substantial resource, hardware, or parameter costs. These limitations motivate more practical quantum sequential architectures and broader empirical comparisons.
- Quantum sequential modeling remains relatively underexplored despite extensive classical research on time-series forecasting.
- Existing quantum approaches include QSegRNNs, QRNNs, continuous-variable quantum RNNs, and quantum reservoir computers for sequential data.
- Quantum-weighted GRU and LSTM variants can improve prediction or forecasting but incur substantial quantum-resource or computational overhead.
- Other methods rely on specialized photonic hardware, scale poorly on large datasets, or remain parameter-heavy and insufficiently compared with classical baselines.
3 Background
Variational quantum circuits combine quantum parameterized circuits with classical optimization for quantum machine learning on NISQ devices. The background describes input encoding, trainable circuit transformations, measurement, and iterative parameter updates.
- Variational quantum circuits are parameterized quantum circuits used for quantum machine learning and optimization.
- A quantum feature map embeds a classical input x_t into a 2^q-dimensional Hilbert space using q qubits.
- Angle encoding represents continuous classical features as qubit rotation angles using gates such as R_x or R_y.
- A parameterized ansatz applies rotations and entangling gates, after which observable measurements produce model outputs.
- Classical optimization updates circuit parameters by minimizing a cost function, alternating quantum evaluations with classical updates.
4 Methodology
QPRU uses variational quantum circuits as nonlinear transformations within recurrent updates, with a compact architecture that shares classical layers and uses fewer circuits. Its two-circuit design reduces quantum-parameter complexity while retaining flexible input and hidden-state dimensionality.
- QPRU uses quantum circuits as nonlinear transformations within recurrent updates inspired by QGRU and QLSTM.
- QPRU concatenates the current input with the previous hidden state, then maps the result through classical layers and VQCs to compute recurrent representations.
- FCin and FCout are shared across both gates, while VQCupdate and VQCout remain distinct and are trained separately.
- The update equation combines retained previous state and new information using (1 − u_t) ⊙ s_t−1 + u_t ⊙ o_t.
- The architecture reduces computational cost through fewer VQCs, parameter sharing, and quantum entanglement while accommodating mismatched input and hidden-state dimensions.
- 50% fewer quantum parameters than QLSTM and 33% fewer than QGRU make QPRU the most parameter-efficient of the three architectures.
5 Experiments
The experiments compare QPRU with classical and quantum recurrent baselines using standardized training and validation procedures. Evaluation covers synthetic sinusoidal forecasting and AAPL stock-price forecasting, alongside parameter comparisons under fixed architectural settings.
- QPRU is compared with QLSTM, QGRU, classical PRU, GRU, and LSTM after 100 training epochs with 20% validation data.
- Models are trained with MSE loss and RMSProp, using a window size of 3, learning rate 0.01, and hidden-state size 5.
- The study reports ℓ1, ℓ2, and ℓ∞ validation norms with standard deviations and uses PyTorch, PennyLane, Lightning Qubit, and adjoint differentiation.
- Experiment I: Experiment I forecasts the next value in a synthetic sinusoidal sequence using overlapping input windows of length 3 over [0, 18π].
- Experiment II: Experiment II forecasts normalized daily AAPL closing prices from January 1, 2020, to January 1, 2024, using overlapping windows of length 3.
6 Results
QPRU performs strongly on sinusoidal and AAPL stock forecasting, achieving the best reported norm metrics on both tasks and closely tracking the sinusoidal pattern. Hypothesis tests and confidence intervals indicate no statistically significant pairwise differences from the evaluated benchmarks.
- Experiment I: 0.012949, 0.000263, and 0.0282 were QPRU’s lowest ℓ1, ℓ2, and ℓ∞ norms on the sinusoidal benchmark.Values were averaged over 10 independent validation runs.
- Experiment I: QPRU predictions closely track the sinusoidal function across training and validation sets with low absolute error.
- Experiment II: QPRU achieved the lowest stock-forecasting ℓ1 norm (0.046131) and ℓ2 norm (0.003276) among the evaluated models.These metrics were computed on normalized values and averaged over 10 independent runs.
- Experiment II: Validation AAPL forecasting produced denormalized MSE of 4.4370 ± 7.920 and MAE of 1.6265 ± 1.338.The figure reports predictions for both training and validation sets, with a dashed line marking their split.
- Hypothesis Testing: Confidence intervals for all methods overlap, so none of the pairwise differences is statistically significant at the tested levels.The authors describe QPRU as statistically indistinguishable from the established benchmarks within this experimental resolution.
7 Discussion
QPRU combines strong predictive performance with a lightweight design. Across the sinusoidal and AAPL experiments, it uses fewer quantum parameters while maintaining competitive forecasting accuracy.
- Experimental Findings: QPRU obtained the lowest average ℓ1, ℓ2, and ℓ∞ losses on the sinusoidal dataset among the classical and quantum models.The results also report consistently low standard deviations.
- Model Efficiency: QPRU reduces quantum parameters by approximately 50% versus QLSTM and 33% versus QGRU.The authors identify this reduction as an advantage for NISQ-era devices.
- Experimental Findings: On AAPL stock data, QPRU achieved competitive forecasting accuracy, demonstrating generalization beyond the synthetic setting.
- Implications: The results indicate that a simplified hybrid quantum–classical design can model complex temporal patterns using fewer parameters and shallow circuits.
8 Conclusion and Future Work
QPRU is presented as a lightweight recurrent design with reduced quantum resource demands, fewer trainable parameters, stable training, and scalability across sequence-learning tasks. Future work will test broader data, circuit, and hardware settings.
- Conclusion: QPRU offers reduced quantum resource demands, fewer trainable parameters, stable training behavior, and strong scalability across sequence-learning tasks.
- Future Work: Future work will extend QPRU to multivariate time series, alternative encoding and ansatz designs, larger datasets, and actual quantum hardware.These studies are intended to evaluate expressivity, demanding forecasting performance, and practical feasibility and limitations.