Defeating Barren Plateaus with Task-Aligned Symmetry

Quantum Recurrent Neural Networks (QRNNs) are designed to process sequential data by iteratively applying a variational quantum block. A critical concern for these models is the "barren plateau" (BP) phenomenon, where gradients vanish exponentially, making training impossible as sequence length increases. This paper investigates whether the recurrent architecture itself is sufficient to avoid BPs or if specific architectural constraints are required.

To determine the conditions for trainability, the authors analyze the gradient variance of QRNNs using two models: one without parameter sharing (where each time step uses independent parameters) and one with parameter sharing (where the same parameters are reused across all time steps). They employ Haar measure integration to derive the scaling of gradient variance for the non-sharing case and use trigonometric-moment representations to evaluate the sharing case. They further validate these analytical proofs with numerical simulations using multi-layer ansatz circuits.

The study reveals that the recurrent topology alone does not prevent barren plateaus. In the absence of parameter sharing, the gradient variance decays exponentially with sequence length, confirming that QRNNs are inherently untrainable in this configuration. By contrast, imposing time-translation symmetry through parameter sharing collapses the effective degrees of freedom. This constraint changes the scaling of the gradient variance from exponential decay to a polynomial dependence, effectively suppressing the barren plateau and enabling the training of longer sequences.

This work provides a rigorous design principle for quantum machine learning. It demonstrates that the "curse of dimensionality" in Hilbert space can be mitigated by injecting task-aligned inductive biases, similar to how classical deep learning relies on architectural symmetries. By identifying parameter sharing as a dual-purpose mechanism—simultaneously ensuring trainability and improving generalization—the authors offer a constructive path for building scalable quantum models for sequential data.