Challenges with Differentiable Quantum Dynamics

Differentiable quantum dynamics require automatic differentiation of a complex-valued initial value problem, which numerically integrates a system of ordinary differential equations from a specified initial condition, as well as the eigendecomposition of a matrix. We explored several automatic differentiation frameworks for these tasks, finding that no framework natively supports our application requirements. We therefore demonstrate a need for broader support of complex-valued, differentiable numerical integration in scientific computing libraries.

💡 Research Summary

The paper addresses the practical challenge of making quantum dynamics simulations differentiable, a prerequisite for gradient‑based optimization of quantum‑enhanced metrology protocols. In open quantum systems the state evolution is governed by the Lindblad master equation, a complex‑valued ordinary differential equation (ODE) for the density matrix ρ(t). The performance metric, quantum Fisher information (QFI), is expressed in terms of the eigenvalues and eigenvectors of the final density matrix ρ(x), where x denotes the parameters of a variational quantum circuit (VQC). Consequently, the computational pipeline consists of two implicit operations that must be differentiated: (1) numerical integration of a complex‑valued ODE, and (2) eigendecomposition of a complex, generally non‑Hermitian matrix.

The authors systematically evaluate a range of popular automatic‑differentiation (AD) frameworks—SciPy, TensorFlow Probability, PyTorch (torchdiffeq), JAX, TensorFlow‑diff, NeuroDiffEq, DeepXDE, PyDEns, Diffrax, and DifferentialEquations.jl—against three criteria: support for complex‑valued ODE integration, provision of adjoint or reverse‑mode gradients for the ODE solution, and availability of differentiable eigendecomposition (including gradients w.r.t. both eigenvalues and eigenvectors).

Key findings:

• SciPy offers robust complex ODE solvers (solve_ivp) but lacks any AD interface. Using autograd forces a fallback to odeint, which does not accept complex inputs.
• TensorFlow Probability provides a BDF solver that can integrate complex ODEs, yet its Jacobian routine does not support complex dtypes, and the adjoint sensitivity method is unavailable for complex numbers.
• torchdiffeq (PyTorch) can be patched to accept complex initial conditions, reproducing SciPy’s primal solution, but it does not return gradients with respect to the integration time, limiting full‑parameter backpropagation.
• JAX’s experimental.ode.odeint implements the adjoint method and can differentiate w.r.t. initial state, time, and auxiliary arguments, but the complex‑valued gradient implementation is buggy, leading to incorrect results.
• tfdiff, NeuroDiffEq, DeepXDE, PyDEns all reject complex initial values outright, offering no path to differentiable quantum dynamics.
• Diffrax (a Python front‑end to Julia’s DifferentialEquations ecosystem) successfully integrates complex ODEs, supports DirectAdjoint for full reverse‑mode differentiation, and works seamlessly with JAX for differentiable eigendecomposition.
• DifferentialEquations.jl (Julia) natively handles complex ODEs and eigenproblems with rich AD support, but accessing these capabilities from Python requires additional wrapper code.

The overarching insight is that current mainstream AD libraries are designed primarily for real‑valued scientific computing and do not provide a unified, out‑of‑the‑box solution for complex‑valued ODE integration combined with differentiable eigendecomposition. This gap hampers research in quantum control, quantum sensing, and any domain where complex linear algebra and differential equations intersect.

The authors conclude by calling for the development of a dedicated, open‑source library that (i) offers native complex‑valued ODE solvers with adjoint sensitivity, (ii) implements differentiable eigendecomposition for non‑Hermitian matrices, and (iii) integrates cleanly with existing machine‑learning ecosystems (PyTorch, JAX, TensorFlow). Such a library would enable efficient gradient‑based optimization of VQC‑generated states, accelerate the design of quantum metrology protocols, and broadly benefit scientific computing tasks that rely on complex differential equations.