QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem

Technical Summary: QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem

Problem Statement
Reliable quantum control in realistic settings is hindered by decoherence—the irreversible coupling of a quantum system to its environment. While existing quantum optimal control methods (e.g., GRAPE, Krotov's method, Reinforcement Learning) effectively optimize control pulses for fidelity, they often treat decoherence as a static cost or apply generic penalties on control fluence (energy) or smoothness. These approaches do not explicitly account for how different control policies route the system through Hilbert space, potentially exposing it to varying degrees of environmental noise. The central challenge is to develop control policies that actively steer the system toward states or subspaces where the observable consequences of decoherence are minimized, rather than merely minimizing the control amplitude.

Methodology
The authors propose QMaxCal, a framework for path-space regularization that operates directly on the distribution of quantum trajectories. The method leverages the Stochastic Schrödinger Equation (SSE) to model open quantum systems under continuous monitoring. In this framework, the system's interaction with the environment generates classical measurement records. Crucially, for two evolutions sharing the same decoherence channels (Lindblad operators), these records differ only in their drift terms, while sharing the same underlying noise structure (Wiener increments).

The core theoretical innovation is the application of Girsanov's Theorem to derive a closed-form, differentiable estimator of the Kullback-Leibler (KL) divergence between the trajectory distributions of a controlled evolution and a reference measure. This allows the optimization objective to penalize the "observable consequences" of control on the decoherence channels.

The framework instantiates this general estimator with two physically motivated reference measures, yielding two distinct regularizers:

1. Wiener KL (KLW): Uses the Wiener measure (pure Brownian motion with zero drift) as the reference. This regularizer drives the system toward the joint kernel of the Lindblad operators (where the drift $\alpha_k = \langle L_k + L_k^\dagger \rangle = 0$). It is most effective when the system can be routed through a decoherence-free subspace (DFS) where the noise vanishes entirely.
2. Drift-Variance (RDV): Uses the closest constant-drift process as the reference. This regularizer minimizes the variance of the drift over time and across noise realizations. It drives the system toward any decoherence-free subspace, even those with non-zero constant drift (where the state is an eigenstate of the Lindblad operators but not necessarily in their kernel).

The total loss function combines the standard final-state fidelity objective with these regularizers and a standard fluence penalty:
$$\mathcal{L}(\theta) = 1 - \mathbb{E}_{P_\theta}[F(\psi(T))] + \lambda_{KLW} KLW + \lambda_{RDV} RDV + \lambda_{flu} Flu$$

Key Contributions

1. Closed-Form KL Estimator: The authors derive a closed-form estimator for the KL divergence over quantum trajectory space (Eq. 8) valid for any pair of evolutions sharing the same Lindblad operators, utilizing Girsanov's theorem.
2. Novel Regularizers: The introduction of KLW and RDV, which penalize the coupling to the environment (via the measurement record drift) rather than the control amplitude itself. These are qualitatively distinct from existing penalties on fluence or smoothness.
3. Empirical Validation: The framework is tested on five open-system benchmarks, including single-qubit amplitude damping, STIRAP (Stimulated Raman Adiabatic Passage), a four-level "diamond" system, a synthetic qubit chain with site-dependent noise, and a six-qubit chain calibrated to the IBM Kingston processor.

Results
The regularizers demonstrate significant improvements over unregularized gradient-based baselines and constrained Reinforcement Learning (PPO) baselines:

• Fidelity Gains: Across benchmarks, the regularizers reduce infidelity by up to 50%. In the IBM Kingston calibrated chain, the drift-variance regularizer achieved a $\sim 16\%$ reduction in infidelity compared to the baseline.
• Robustness to Noise Mismatch: The methods show superior robustness when the training noise model differs from the test environment. For the diamond system, a policy trained at a specific noise level maintained high fidelity under a $2.5\times$ stronger noise mismatch, outperforming the baseline by +27 percentage points (pp).
• Trajectory Routing: In the amplitude damping benchmark, the Wiener KL policy successfully routed the trajectory through the ground state (the kernel of the Lindblad operator), reducing the time-integrated population variance by a factor of 15 compared to the baseline. In the diamond system, the regularizer forced the optimizer to discover a safe indirect route through an auxiliary state, avoiding a lossy direct coupling that trapped the baseline in a local minimum.
• Computational Efficiency: By operating on state vectors (trajectories) rather than density matrices, the SSE-based approach inherits a quadratic memory advantage, allowing the simulation of systems up to 12 qubits on a single GPU, whereas density-matrix solvers run out of memory at this scale.

Significance and Claims
The paper claims that QMaxCal offers a principled way to incorporate decoherence awareness into quantum control without requiring system-specific reward engineering. By penalizing the observable consequences of control on the decoherence channels, the method naturally discovers policies that exploit decoherence-free subspaces or minimize exposure to lossy states.

The authors note that the effectiveness of the regularizers is contingent on the noise structure: they provide the most significant gains when a reachable decoherence-free subspace exists (e.g., amplitude damping, diamond system) or when there is pronounced asymmetry in noise rates (e.g., IBM Kingston chain). In scenarios with small noise or no exploitable DFS, the gains are minimal or absent, consistent with the theoretical derivation that the regularizers vanish exactly on decoherence-free evolutions.

The work is presented as a contribution to the field of Machine Learning applied to quantum optimal control, extending the Maximum Caliber principle to open quantum systems via Girsanov's theorem. The authors explicitly limit their scope to diffusive unravellings (homodyne/heterodyne detection) and Markovian environments, noting that extensions to jump unravellings and non-Markovian settings remain future work.