Variational Quantum Dimension Reduction for Recurrent Quantum Models

TL;DR

This paper introduces a variational quantum dimension reduction method for recurrent quantum models, effectively compressing memory while preserving dynamics, and demonstrating significant fidelity improvements over existing truncation techniques.

Contribution

The authors propose a novel variational framework that reduces memory size in recurrent quantum models using parameterized circuits guided by a new fidelity metric, enhancing scalability.

Findings

Achieves up to three orders of magnitude smaller QFDR compared to matrix product state truncation.

Requires only trajectory samples, avoiding explicit state reconstructions.

Enables scalable, data-driven learning of minimal recurrent quantum architectures.

Abstract

Recurrent quantum models (RQMs) realize sequential quantum processes through repeated application of a unitary operation on a memory system coupled with a series of output registers. However, such models often rely on unnecessarily large memory spaces, introducing redundancy and limiting scalability. Here, we introduce a \textit{variational quantum dimension reduction} framework that identifies and removes irrelevant memory degrees of freedom while preserving the recurrent dynamics of the target model. Our approach employs two parameterized quantum circuits: a decoupling unitary $V(\theta_1)$ that isolates the essential memory subspace; and a compressed recurrent unitary $\tilde{U}(\theta_2)$ that reconstructs the dynamics in the reduced space. The optimization is guided by a unified cost function combining decoupling fidelity and dynamical accuracy, evaluated using the \textit{Quantum Fidelity Divergence Rate} (QFDR), a metric that quantifies long-term fidelity per time step. Applied to a cyclic random walk model, our framework achieves up to three orders of magnitude smaller QFDR compared to variational matrix product state truncation, while requiring only trajectory samples rather than explicit state reconstructions. This establishes a scalable, data-driven paradigm for learning minimal recurrent quantum architectures, enabling variational circuit optimization and quantum process compression for near-term quantum devices.