Controllable Quantum Memory Capacity in Quantum Reservoir Networks with Tunable partial-SWAPs

Abstract: In the field of quantum reservoir computing (QRC), many different computational models and architectures have been proposed. From these models, we identify feedback based models -- which use a feedback mechanism to re-embed classical measurements from the QRC -- and recurrent models -- which use a multi-register approach with memory and readout qubits -- as the two major competing architectures that have been discussed and validated on hardware. In this paper, we advance upon the recurrent architectures, which employ a two register approach to endow the QRC with a fading memory. While these approaches have been validated on hardware and have demonstrated great real-world performance on noisy-intermediate-scale-quantum (NISQ) quantum processing units (QPUs), the exact mechanism through which the memory capacity arises is not completely understood or fully controllable. With this, we augment the recurrent approaches and present a hardware-realizable mechanism, which we call a tunable partial-SWAP, that allows for the direct control of the rate of memory dissipation from a QRN implemented on a gate-based QPU. The theory behind this mechanism is discussed in terms of a controlled amplitude-damping channel and validation experiments using a randomized short-term memory capacity (STMC) recall benchmark and the NARMA-5 dataset are conducted using simulation and IBM QPUs, respectively.

• The paper introduces a tunable partial-SWAP operation to precisely control fading memory in quantum reservoir networks, analogous to the leak rate in classical echo state networks.
• It employs a recurrent architecture with customizable data reuploading and amplitude damping to optimize performance on recall benchmarks and the NARMA-5 task.
• Experimental results on IBM NISQ hardware confirm that optimized partial-SWAP parameters mitigate noise, ensuring robust memory purification and sustained inference accuracy.

Controllable Memory Capacity in Quantum Reservoir Networks via Tunable Partial-SWAPs

Introduction and Motivation

This work introduces a principled approach for endowing quantum reservoir networks (QRNs) with controllable memory capacity through the construction of a tunable partial-SWAP operation. The primary focus is on augmenting the recurrent architecture of QRNs—a leading framework within quantum reservoir computing (QRC)—where memory dynamics have been historically effective but lacked a direct parameter for explicit capacity modulation. This gap in controllability is significant, especially for scaling QML approaches on NISQ hardware.

The authors position their contribution amid two dominant QRN paradigms: feedback-based models, which cyclically re-embed classical measurement outcomes into the quantum state at the cost of higher circuit complexity and potential Hilbert space collapse, and recurrent models, which maintain separate registers for memory and readout, but without a direct, interpretable handle on memory fading rate. Their partial-SWAP-based mechanism clarifies and extends the theoretical basis for controlled memory capacity in QRNs.

Architecture and Methodology

Quantum Reservoir Network Structure

The proposed QRN consists of three key layers:

1. Embedding Layer: Classical time-series input data, with an adjustable context window, is mapped to quantum rotation gates using data reuploading and parametrized entangling layers. This embedding is customizable and supports nonlinear transformations.
2. Memory Exchange Layer: The crux of the contribution is here—a parameterized partial-SWAP operation, denoted as $U_{\text{SWAP}}(\gamma)$, is performed between each pair of memory and readout qubits. $\gamma \in (0, 1]$ serves as a direct analogue to the leak rate in classical echo-state networks (ESNs), controlling the rate of amplitude damping (and hence, information transfer and memory decay) in the quantum memory register.
3. Measure and Reset Layer: Standard projective measurement is performed on readout qubits; the register is then reset, and computation proceeds.

The partial-SWAP operation implements a controlled amplitude damping channel on the memory, achieving a tunable fading memory effect with a single hyperparameter. This mechanism purifies the memory register over time and, by adjusting $\gamma$, enables fine-tuned trade-offs between memory retention and new information integration.

Theoretical Analysis

The amplitude damping effect induced by the partial-SWAP, combined with reset operations, is derived explicitly. The repeated application ensures that, asymptotically, the memory register is driven towards a pure ground state, providing robustness and effective error mitigation for long-running inference tasks.

Experimental Validation

Simulation Results

Short-Term Memory Capacity (STMC)

The memory capacity is quantified using a standard recall benchmark where the QRN must predict delayed versions of a randomly-driven input sequence. The RMSE and $R^2$ scores (delay dependent) are measured and analyzed under various circuit depths (qubit counts), data reuploading repetitions, and partial-SWAP strengths.

Key findings:

• Memory capacity is sharply controlled by the partial-SWAP strength $\gamma$. Optimal recall is consistently observed for intermediate $\gamma$ values ($0 < \gamma < 1$), with performance degrading towards both extremes (full SWAP or no swap).
• Increasing number of qubits enhances memory capacity, supporting the scaling argument for larger QRNs.

NARMA-5 Benchmark

On the nonlinear temporal NARMA-5 task, the partial-SWAP QRN demonstrates strong performance, with the optimal $\gamma$ value for STMC and NARMA-5 tasks converging. Increased embedding nonlinearity (via more data reuploading blocks) is shown to be beneficial here, consistent with the high nonlinearity demands of this task.

Notably, a 12-qubit model achieved a simulation RMSE of 0.0484 on NARMA-5.

Comparison to Classical ESN

Partial-SWAP QRNs were directly compared with classical ESNs with matched hidden state sizes. In nearly every tested configuration, the quantum model exhibited an advantage, particularly at low qubit/node counts. This hints at quantum architectures’ potential for enhanced expressivity in temporal processing, given properly chosen hyperparameters and initialization.

Hardware Experiments

A 12-qubit partial-SWAP QRN was deployed on the IBM ibm_boston backend for the NARMA-5 task. Despite substantial circuit depth (203,133 gates over 1000 steps), the empirical RMSE was 0.0646, just 33% higher than noiseless simulation and well within expectations for NISQ-era hardware. The observed purification effect of the partial-SWAP mechanism appears to mitigate noise accumulation, allowing the quantum network to operate effectively through long sequences.

Implications and Future Trajectories

Practical and Theoretical Significance

The introduction of a tunable partial-SWAP as a single-parameter memory capacity controller for QRNs addresses a substantial limitation in quantum recurrent architectures. This advance:

• Brings quantum recurrent models closer to the convenient and interpretable controllability found in classical ESN frameworks.
• Provides a hardware-realistic, low-overhead means for temporal memory regulation, critical for tasks such as time-series prediction, filtering, and system identification on NISQ devices.
• Demonstrates, through both simulation and experiment, that real-world deployment is feasible and robust to quantum hardware noise due to the built-in purification effect.

Future Directions

As quantum hardware continues to scale and error rates decrease, this work suggests several research avenues:

• Investigation of even deeper quantum reservoirs and alternative embedding schemes to push nonlinear expressivity and memory/accuracy limits further.
• Development of theoretical analysis for the observed memory/expressivity bottleneck around 12 qubits, and extension to larger Hilbert spaces.
• Applications to broader classes of tasks (e.g., quantum control, filtering, or dynamical systems inference) requiring both high nonlinearity and long-term memory.
• Exploration of hybrid quantum-classical reservoir and echo-state architectures leveraging tunable quantum memory for enhanced classical neural computation.

The direct analogy between the partial-SWAP hyperparameter and the leak rate in classical reservoir computing, together with demonstration of the underlying quantum amplitude damping channel, provides a solid foundation for more interpretable and controllable quantum neural architectures.

Conclusion

This work establishes tunable partial-SWAPs as an essential mechanism for achieving controllable fading memory in QRNs. The parameter $\gamma$ provides direct, interpretable access to memory capacity adjustment, enabling robust performance on both linear and nonlinear temporal benchmarks in both simulated and real hardware environments. These results facilitate the deployment of interpretable, scalable, and robust quantum recurrent neural networks, setting the stage for their broader application in quantum machine learning and temporal modeling tasks as quantum hardware matures.

Reference: "Controllable Quantum Memory Capacity in Quantum Reservoir Networks with Tunable partial-SWAPs" (2605.12713)