Tree Tensor Network Reservoir Computing: Hierarchical Ensemble with Invariant Phase Boundaries

Traditional reservoir computing (RC) methods, such as Echo State Networks (ESN), often struggle with tasks requiring high-order nonlinear processing or long-range contextual dependencies. This paper investigates whether the hierarchical, multi-scale structure of Tree Tensor Networks (TTN) can serve as a more effective random reservoir for time-series prediction than standard sparse random matrices.

The authors introduce TTN-RC, a quantum-inspired architecture that replaces the standard reservoir with a Tree Tensor Network. To manage the tendency of these networks to exhibit exponential output concentration or divergence, the authors propose a hierarchical ensemble method. This method partitions a fixed-size reservoir into multiple independent sub-reservoirs, allowing for better control over the effective tree depth. The authors also derive a closed-form expected contraction rate based on the reservoir Jacobian and apply mean-field theory to characterize the reservoir's stability.

TTN-RC demonstrates competitive or superior performance compared to conventional ESNs on NARMA (Nonlinear Autoregressive Moving Average) benchmarks, particularly in tasks requiring higher-order nonlinear processing. Theoretically, the authors identify a critical phase-boundary at a tensor-element standard deviation of σ_T = √2. In the limit of large sub-reservoir sizes, this value serves as a convergence point for stability, mean-field statistics, and the transition between concentration and divergence. This provides a clear design principle for tuning hyperparameters in tensor-network-based reservoirs.

This work bridges the gap between tensor network theory and reservoir computing, offering a rigorous statistical-physics interpretation of reservoir topology. By providing a concrete stability criterion (σ_T = √2), the paper offers practitioners a systematic way to initialize and scale hierarchical reservoirs, potentially unlocking better performance for complex temporal modeling tasks that exceed the capabilities of standard ESNs.