[Paper] Kraus Constrained Sequence Learning For Quantum Trajectories from Continuous Measurement

Source: arXiv - 2603.05468v1

Overview

Real‑time reconstruction of quantum states from continuous measurement streams is a cornerstone for quantum feedback control, but traditional stochastic master‑equation (SME) solvers are brittle: they need an exact model, precise system parameters, and they quickly break when those parameters drift. The new paper introduces a Kraus‑constrained sequence learning framework that forces any neural sequence model to output only physically admissible quantum operations, eliminating unphysical predictions while still learning from raw measurement data.

Key Contributions

• Kraus‑structured output layer that maps any hidden state of a sequence model to a completely‑positive trace‑preserving (CPTP) map, guaranteeing valid quantum state updates by construction.
• Model‑agnostic design: the layer is plugged into a variety of backbones (RNN, GRU, LSTM, TCN, ESN, Mamba, Neural ODE) to demonstrate broad applicability.
• Robustness to parameter drift: experiments on stochastic trajectories with time‑varying Hamiltonian and measurement parameters show the constrained models remain stable where unconstrained ones diverge.
• Empirical trade‑off analysis of gating (LSTM/GRU), linear recurrence (ESN), and global attention (Mamba) in the context of quantum state estimation.
• Quantitative gain: the Kraus‑LSTM improves state‑estimation fidelity by ~7 % over its unconstrained counterpart while always respecting positivity and trace constraints.

Methodology

1. Data Generation – Simulated continuous‑measurement records are produced from a stochastic master equation that includes a slowly drifting system parameter (e.g., a fluctuating magnetic field). Each record is paired with the exact conditional quantum state (the ground truth).
2. Sequence Backbone – Any standard sequence model processes the time‑ordered measurement outcomes, producing a hidden representation at each step.
3. Kraus Output Layer
   • The hidden vector is first projected onto a set of Kraus operators ({K_i}).
   • These operators are normalized to satisfy (\sum_i K_i^\dagger K_i = I), ensuring the resulting map is CPTP.
   • The updated density matrix (\rho_{t+1} = \sum_i K_i \rho_t K_i^\dagger) is computed analytically inside the network, so the loss can be back‑propagated through the quantum operation.
4. Training Objective – Mean‑square error (or fidelity‑based loss) between the predicted (\rho_{t+1}) and the ground‑truth state, plus a small regularizer to keep the Kraus set well‑conditioned.
5. Baseline Comparison – Identical backbones without the Kraus layer (i.e., unconstrained dense output) are trained on the same data to isolate the effect of the physical constraint.

Results & Findings

Key observations

• Physical validity: All Kraus‑constrained models never produced a non‑positive density matrix, even after long roll‑outs, whereas unconstrained models occasionally generated negative eigenvalues, causing simulation crashes.
• Stability under drift: When the underlying Hamiltonian parameter was ramped during inference, the constrained models tracked the true state with modest degradation, while unconstrained models diverged dramatically.
• Gating advantage: The LSTM’s internal gates combined with the Kraus layer yielded the best trade‑off between expressive power and regularization, explaining its top performance.

Practical Implications

• Quantum control loops – Developers building real‑time feedback controllers for superconducting qubits, trapped ions, or NV centers can replace fragile SME solvers with a trained Kraus‑LSTM that runs on modest hardware (CPU/GPU) and guarantees physically admissible updates.
• Parameter‑agnostic calibration – Since the model learns directly from measurement streams, it tolerates unknown or slowly varying system parameters, reducing the need for frequent recalibration.
• Edge deployment – The layer’s operations are simple matrix multiplications and normalizations, making it suitable for embedded quantum‑hardware processors where latency is critical.
• Hybrid quantum‑classical pipelines – The approach can be integrated into variational quantum algorithms that require intermediate state estimates, providing a trustworthy classical surrogate for the quantum dynamics.

Limitations & Future Work

• Scalability – The current experiments focus on single‑qubit (2‑dimensional) systems; extending Kraus layers to multi‑qubit Hilbert spaces will increase the number of Kraus operators and computational cost.
• Training data dependence – The model still needs high‑quality simulated (or experimentally calibrated) trajectories for training; generating such data for large systems remains a bottleneck.
• Interpretability – While the Kraus operators are mathematically valid, interpreting what the network has learned about the underlying physics is non‑trivial.
• Future directions suggested by the authors include: hierarchical Kraus layers for larger registers, adaptive selection of the number of Kraus operators during inference, and coupling the framework with reinforcement‑learning agents for autonomous quantum‑control policy discovery.

Authors

• Priyanshi Singh
• Krishna Bhatia

Paper Information

• arXiv ID: 2603.05468v1
• Categories: cs.LG
• Published: March 5, 2026
• PDF: Download PDF