arXiv:2610.22691
Quantum-Classical Neural Networks for Solving Lindblad Master Equations
Rafael Florencio Díaz, Julio Guerrero García, Elvira Romera Gutiérrez
Abstract
Variational quantum algorithms have emerged as a promising approach for simulating open quantum systems, yet their expressive power and trainability remain poorly understood. We propose a variational framework for solving Lindblad master equations by minimizing a residual cost function using a compositional ansatz obtained by combining a classical neural network with a parameterized quantum circuit. The density matrix is represented through a purification on an enlarged Hilbert space, ensuring positivity, Hermiticity, and unit trace by construction, while all gradients are evaluated exactly via the parameter-shift rule. Beyond the algorithmic formulation, we derive an upper bound for the covering number of the proposed compositional model and compare it with that of purely quantum neural networks. Our analysis shows that the classical parametrization enlarges the effective hypothesis class while providing finer control over its complexity, thereby establishing a theoretical connection between expressivity, covering numbers, effective dimension, and the barren plateau phenomenon. Numerical experiments validate both the accuracy of the proposed method for representative Lindblad master equations and the theoretical predictions regarding gradient concentration. These results provide a unified variational and theoretical framework for the simulation of dissipative quantum dynamics using compositional quantum machine learning models.
AI Summary
The paper studies a residual-based variational approach to Lindblad dynamics, combining a classical neural network with a parameterized quantum circuit. Circuit angles follow $\boldsymbol\theta(t;\mathbf w)=\boldsymbol\theta_0+t\mathbf N_c(t;\mathbf w)$, fixing the initial condition. Tracing out ancillary qubits from the circuit output produces a density matrix that is positive, Hermitian, and normalized. Training minimizes squared discrepancies between observable time derivatives and their Lindblad predictions. First- and second-order parameter-shift formulas are derived for the required gradients. The theoretical analysis gives a covering-number upper bound of the form $\mathcal N(\varepsilon)\leq(A/\varepsilon)^{n_wN_{\rm rg}2^{2k}}$, where $k$ is the total qubit count, $N_{\rm rg}$ the number of rotation gates, and $n_w$ the classical parameter count; $A$ depends on parameter bounds, observable norms, and Lipschitz constants. The paper interprets this exponent as an effective dimension. Assuming approximately uniform sampling on an effective sphere, it argues for gradient second-moment scaling $O((n_wN_{\rm rg}2^{2k})^{-1})$. Dynamics are demonstrated for one dephasing qubit with a $z$-directed Hamiltonian, using one ancilla and four rotation angles. Runs with 34, 64, and 124 classical parameters over 800 gradient-descent iterations show decreasing observable errors, with smaller gains beyond 64 parameters. Separate random-initialization experiments use 50 samples per setting and find stronger gradient suppression for a global projector than for a local Lindblad residual. The paper acknowledges that exponential gradient suppression remains for global observables. Residual witnesses must detect unwanted state components; the concentration argument also relies on its uniform-sampling assumption.
Central claim may not hold: The comparison of covering-number upper bounds does not establish an enlarged hypothesis class, and at fixed time the hybrid circuit's reachable states are contained in those of the freely parametrized circuit. The identification of n_w N_rg 2^(2k) as reachable-state manifold dimension is invalid because the fixed-time parametrization has Jacobian rank at most min(n_w,N_rg). Proposition 1 cannot justify Eqs. 35–36 because the substituted state proportional to O|Psi> is dependent on |Psi>, and the covering-bound exponent is not the Hilbert-space dimension required by the proposition. Eq. 36 and the numerical observations do not establish that classical-parameter tuning prevents vanishing gradients: the equation retains exponential qubit suppression, while the reported mitigation is attributed to local observables. Eq. 27 incorrectly bounds a squared norm of a telescopic sum by the sum of squared norms, although its intended Lipschitz bound is recoverable by a corrected argument.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.