Microwave signal processing using an analog quantum reservoir computer

The study addresses whether near-term noisy intermediate-scale quantum (NISQ) devices can deliver practical advantages for machine learning without facing barren plateaus that hinder training in parameterized quantum neural networks. Building on reservoir computing, quantum reservoir computing (QRC) places all learning in a final classical linear layer, bypassing gradient-based training issues. Prior QRC demonstrations largely used gate-based digital circuits that discretize temporal inputs and introduce input bottlenecks. The authors propose and demonstrate an analog, continuous-time QRC using a superconducting microwave circuit (oscillator coupled to a qubit) that can directly ingest weak, continuous microwave signals. The purpose is to show effective feature generation via quantum nonlinear dynamics and measurement back-action for classifying time-independent and time-dependent signals, including RF modulation schemes and filtered noise, and to highlight a potential path toward hybrid quantum sensing-computational advantage in processing ultra-low-power microwave signals.

The paper situates the work within two NISQ research directions: demonstrating supremacy versus achieving advantageous applications. It reviews challenges in quantum machine learning due to barren plateaus in variational circuits and presents QRC as an alternative that performs learning via a classical linear readout. Prior experimental QRC implementations relied on digital, gate-based approaches that discretize inputs and face input bottlenecks and sampling noise. The authors reference reservoir computing in classical ML, digital-analog quantum computation, and prior QRC/proposals using single nonlinear oscillators or photonic platforms. They also connect to quantum sensing advances in superconducting circuits (quantum microwave radiometry, quantum-enhanced detection) as enabling technologies for a hybrid sensing-computation advantage. Finally, they discuss sampling noise challenges and eigentask analyses relevant to physical reservoirs, motivating their moment-based feature encoding approach.

Hardware and model: The reservoir is a continuous-variable quantum system comprising a 3D aluminum cavity oscillator coupled to a transmon qubit with a separate readout resonator. The cavity resonance is ~6 GHz with ~2 kHz linewidth. Dynamics are governed by a qubit–oscillator Hamiltonian in a rotating frame with a cross-Kerr-like interaction, plus time-dependent drives on both the oscillator and qubit. Non-unitary dynamics arise from interleaved projective measurements.

Protocol: The reservoir operation consists of repeated blocks of (i) entangling unitary evolution via conditional displacements and qubit rotations while injecting the analog input into the oscillator, followed by (ii) a projective qubit measurement and (iii) an oscillator parity measurement. Four rounds of unitary-plus-measurements are executed before a fast reset; this is repeated N times to collect measurement trajectories. The parity measurement projects onto even/odd Fock subspaces, introducing non-Gaussian, non-commuting measurement back-action between rounds, which together with the entangling unitaries generates complex correlations across the trajectory.

Input encoding: Analog microwave inputs are sent resonantly to the cavity. For time-independent inputs, the in-phase and quadrature (I,Q) components yield a constant displacement; for time-dependent signals (e.g., digitally modulated RF), the input varies during the unitary block. Control pulses include π and π/2 rotations on the qubit and conditional displacements on the oscillator placed before/during/after input injection to ensure qubit-state-independent sensitivity and to imprint a geometric phase on the qubit proportional to the enclosed phase-space area, enabling phase extraction for static signals.

Measurements and trajectories: Each run produces a measurement trajectory (bitstring) of length M (here M=8 measurements per reset across tasks). After four unitary-measurement rounds the system is reset and the procedure repeats, generating N trajectories. Repeated measurements both read out features and induce non-unitary dynamics enhancing nonlinearity.

Feature construction: Rather than estimating the full trajectory distribution (which is expensive in shots), the authors build feature vectors from central moments of the measurement outcomes up to third order. They compute the mean (first-order), covariance (second-order), and third-order central moments, but truncate spatial-temporal correlations to a neighborhood of at most three measurement steps apart to reflect finite reservoir memory. For M=8 and up to third-order with truncation, the feature vector dimension is 94. This leverages favorable noise scaling of low-order moments while capturing essential correlations. For time-dependent tasks with real-time sampling, shots count the number of reservoir resets during continuous signal reception (inputs differ each shot).

Training and classifier: A classical linear readout is trained on the feature vectors using either pseudo-inverse (to minimize MSE) or backpropagation with a softmax output (choosing the better-performing option per task). The trained linear layer maps features to class scores. No quantum parameters are trained.

Key parameters and task settings: For the static spiral classification, max input amplitude displaces the oscillator to a coherent state with ~0.3 photons per 1 μs input round; four unitary-measurement rounds per reset; 8 measurements total per sample. For RF modulation classification, 10 classes are generated at ~2 symbols/μs (2 MSps); signals are streamed without repeating identical sequences; classification is evaluated as a function of shots acquired in real time. For filtered-noise classification, six classes are created by moving-average filtering white noise with Gaussian, Lorentzian, or inverse-power-law windows at two widths (50 ns and 600 ns), normalized to equalize photon-number distributions across widths; performance is evaluated versus shots (up to ~2000 shots ≈ 10 ms of signal). A qubit dephasing study artificially reduces T2 via readout resonator population to probe the role of entanglement on performance.

• Time-independent spiral classification: With inputs displacing the oscillator to ~0.3 photons per 1 μs round, the QRC achieves >97% accuracy at 10^3 shots using 8 measurements. A purely linear classifier on raw (I,Q) cannot exceed ~67% (near 50% random baseline). A 64-dimensional, two-layer digital reservoir is needed to match QRC performance (simulation).
• Role of coherence: Reducing qubit T2 during execution (via readout resonator photons) degrades performance markedly; effects appear once T2 approaches the reservoir duration. In the T2→0 limit, entanglement with the oscillator is effectively removed and accuracy drops sharply, indicating the importance of coherent entangling dynamics.
• RF modulation classification (10 classes): Using real-time streaming at ~2 symbols/μs, the reservoir surpasses 90% accuracy with fewer than 2000 shots and 8 measurements; a linear classifier on input data achieves only ~20% even with long signals. Confusion matrices at 512 and 10^3 shots are nearly diagonal.
• Filtered-noise classification (6 classes): The QRC attains 93% accuracy with ~2000 shots (~10 ms of signal). Visualization via SVD shows clear separation of noise classes in feature space. The main remaining confusion at 2000 shots is between 50 ns inverse-power-law and 600 ns Gaussian classes.
• Importance of higher-order correlations: For short-coherence signals (50 ns), the mean (first moment) contributes most, followed by second-order correlations; third-order adds little. For long-coherence signals (600 ns), third-order nonlocal correlations dominate and can nearly reach ~90% accuracy alone, demonstrating sensitivity to long-range temporal structure beyond the measurement rate.
• Analog, few-photon regime: The device processes and classifies signals comprising only a few photons within any single run, highlighting potential for hybrid quantum sensing–computation advantages in low-SNR microwave environments.

The results demonstrate that a continuous-time, analog superconducting QRC can directly ingest and process weak microwave signals, generating rich nonlinear features through entangling dynamics and non-commuting measurement back-action. High accuracies across distinct tasks—nonlinear static classification, multi-class RF modulation recognition, and discrimination of stochastic processes—show that the moment-based feature encoding captures essential temporal correlations with modest measurement counts (M=8) and shots. Dephasing tests confirm that coherence and entanglement materially contribute to performance, connecting physical resources to ML outcomes. While the present system is classically simulable and does not claim computational quantum advantage, the ability to classify few-photon signals suggests a path to a different kind of advantage: hybrid sensing–computation, where quantum hardware jointly detects and processes weak analog signals more effectively than classical counterparts. Analog operation avoids discretization bottlenecks and directly interfaces with microwave inputs, potentially outperforming gate-based approaches in scenarios with limited sampling rates or fast signal features.

This work introduces and experimentally validates an analog quantum reservoir computer based on a cavity–transmon system that processes continuous-time microwave signals using interleaved entangling unitaries and projective measurements. By constructing compact feature vectors from low-order central moments, the approach achieves strong performance on diverse tasks: >97% accuracy on a nonlinear static spiral task, >90% on 10-class RF modulation recognition with fewer than 2000 shots, and 93% on 6-class filtered-noise discrimination in ~10 ms of data. The study links performance to quantum coherence and highlights the role of higher-order correlations for long-coherence signals. Looking ahead, improvements in oscillator coherence, larger reservoirs (more modes/qubits), increased measurement counts M, and enriched feature orders could expand capability. Open questions include expressiveness for time-dependent inputs, optimal trade-offs among reservoir size, measurements, feature dimension, and shot budgets, and extensions to spatiotemporal inputs. Integrating such analog QRCs with quantum microwave sensors could yield quantum smart sensors that achieve hybrid sensing–computational advantages on ultra-low-power signals.

• No claim of computational quantum advantage: the reservoir size is small and simulations are feasible.
• Inputs were synthesized at room temperature and attenuated; the setup did not integrate a quantum sensor front-end in this work.
• Performance depends on the number of shots; sampling noise limits accuracy in few-shot regimes.
• Limited measurement depth (M=8) and truncated moment features (up to third order, local correlations within three steps) constrain expressivity and memory.
• Coherence requirements: degradation in qubit T2 comparable to protocol duration harms performance; oscillator coherence limits the number of effective measurements.
• Some tasks required finite input photon numbers (~0.3 photons per 1 μs round for the spiral task) to reach high accuracy within practical shot counts.
• Comparisons to classical baselines are limited to linear classifiers and a simulated digital reservoir for one task; broader benchmarking against state-of-the-art classical ML is not exhaustive.