Adaptive Tensor Network Simulation via Entropy-Feedback PID Control and GPU-Accelerated SVD

Abstract: Tensor network methods, particularly those based on Matrix Product States (MPS), provide a powerful framework for simulating quantum many-body systems. A persistent computational challenge in these methods is the selection of the bond dimension chi, which controls the trade-off between accuracy and computational cost. Fixed bond dimension strategies either waste resources in low-entanglement regions or lose fidelity in high-entanglement regions. This work introduces an adaptive bond dimension management framework that uses von Neumann entropy feedback coupled with a Proportional-Integral-Derivative (PID) controller to dynamically adjust chi at each bond during simulation. An Exponential Moving Average (EMA) filter stabilizes entropy measurements against transient fluctuations, and a predictive scheduling module anticipates future bond dimension requirements from entropy trends. The per-bond granularity of the allocation ensures that computational resources concentrate where entanglement is largest. The framework integrates GPU-accelerated Singular Value Decomposition (SVD) via CuPy and the cuSOLVER backend, achieving individual SVD speedups of 4.1x at chi=256 and 7.1x at chi=2048 relative to CPU-based NumPy for isolated matrix factorisations (measured on an NVIDIA A100-SXM4-40GB GPU with CuPy 13.4.1 and CUDA 12.8). At the system level, benchmarks on the spin-1/2 antiferromagnetic Heisenberg chain demonstrate a 2.7x reduction in total DMRG wall time compared to fixed-chi simulations, with energy accuracy within 0.1% of the Bethe ansatz solution. Integration with the Density Matrix Renormalization Group (DMRG) algorithm yields ground-state energies per site converging to E/N = -0.4432 for the isotropic Heisenberg model at chi = 128. Validation against Amazon Web Services (AWS) Braket SV1 statevector simulator confirms agreement within 2-5% for small systems.

• The paper introduces a dynamic bond dimension allocation framework using entropy-feedback and PID control to improve simulation accuracy and efficiency.
• It employs per-bond granularity and predictive scheduling, reducing memory usage and computational costs while maintaining system fidelity.
• GPU-accelerated SVD integration provides significant wall-time speedups, validated through benchmarks against state-of-the-art simulators and quantum hardware.

Adaptive Entropy-Driven Bond Dimension in Tensor Network Simulation: PID Control and GPU-Accelerated SVD

Overview

The paper "Adaptive Tensor Network Simulation via Entropy-Feedback PID Control and GPU-Accelerated SVD" (2604.03960) introduces a framework for the dynamic allocation of bond dimensions in Matrix Product State (MPS) based tensor network simulations. This framework combines closed-loop control using feedback from the local von Neumann entanglement entropy, a Proportional-Integral-Derivative (PID) controller, and GPU-accelerated Singular Value Decomposition (SVD) to enhance performance both in terms of accuracy and computational efficiency. Integration with the Density Matrix Renormalization Group (DMRG) algorithm demonstrates empirical speedups and error bounds, validated against statevector simulators and quantum hardware.

Technical Contributions

Dynamic Bond Dimension Allocation via Entropy Feedback and PID Control

A persistent bottleneck in tensor network simulation is the choice of the bond dimension $\chi$, which directly controls the representational power and resource requirements of MPS representations. Conventional approaches use a fixed, user-prescribed $\chi$, resulting in inefficiencies: underutilization in low-entanglement regions and loss of fidelity in highly entangled regions. The proposed framework adaptively adjusts $\chi$ individually at each bond, guided by an EMA-smoothed entropy estimate and regulated by a spatially distributed PID controller.

The entanglement entropy at bond $i$, $S_i$, is estimated from the singular value spectrum. The system applies an Exponential Moving Average (EMA) to $S_i$ across DMRG sweeps, reducing susceptibility to noise. The target bond dimension is set as $\chi_i = \lceil \gamma \exp(\bar{S}_i) \rceil$ (with safety margin $\gamma \geq 1$). The PID controller operates with independently tuned parameters per bond, acting on the error between measured $\bar{S}_i$ and a user-specified target entropy, producing a correction $\Delta \chi_i$. Integral anti-windup and predictor-based anticipation of future entropy surges are integrated, ensuring stable, low-latency adaptation.

Per-Bond Granularity and Predictive Scheduling

Contrary to global $\chi$0-adaptation, the framework assigns independent PID controllers to each bond. This granular allocation tailors resources to the local entanglement structure, demonstrated to reduce average $\chi$1 and total tensor storage, particularly in inhomogeneous or open-boundary systems where bulk and boundary entropies diverge. The predictive module uses short-horizon linear extrapolation of the EMA entropies to decrease lag during rapid entropy growth phases, such as in non-equilibrium dynamics.

GPU-Accelerated SVD and Hybrid CPU/GPU Strategy

The computational bottleneck for MPS algorithms is the repeated SVD of large matrices during DMRG sweeps and time evolution. The authors implement GPU offloading using CuPy and cuSOLVER; SVDs for large bond dimensions are batched and parallelized, achieving $\chi$2 (for $\chi$3), $\chi$4 ($\chi$5), and up to $\chi$6 ($\chi$7) wall time speedups relative to CPU-based NumPy/SciPy. For small matrices ($\chi$8), SVD remains on the CPU due to transfer overhead, employing a hybrid dispatch scheme.

Algorithmic Integration with DMRG

The adaptive framework is implemented within the two-site DMRG algorithm. At each bond update, entropy is measured, and the PID-controlled $\chi$9 determines SVD truncation. The variational property of DMRG is preserved, and convergence criteria remain unchanged. Experiments on the $\chi$0 Heisenberg chain validate that the resultant energies are within $\chi$1 of the Bethe ansatz solution, with an average bond dimension reduction of $\chi$2 relative to uniform allocation.

Empirical Results

Benchmarks demonstrate substantive wall-time reductions and memory savings:

• For the 100-site Heisenberg chain, adaptive per-bond PID control with GPU acceleration yields a $\chi$3 speedup over fixed-$\chi$4 CPU DMRG with $\chi$5 energy error.
• The method outperforms ITensor (Julia), TeNPy (Python), and quimb (Python) in wall-clock time at fixed accuracy, attributed to both dynamic allocation and GPU utilization.
• Ablation studies attribute a $\chi$6 speedup to PID-controlled allocation over fixed-$\chi$7, compared with $\chi$8 for threshold-only truncation.
• Cross-validation against the AWS Braket SV1 simulator shows TVD discrepancies $\chi$9, confirming implementation fidelity. Runs on IBM Heron QPUs yield Bell/GHZ fidelities and parameter-shift gradient errors in line with hardware error rates.

The approach is robust to $i$0 PID gain perturbations across both gapped and critical chain Hamiltonians, reflecting well-damped control dynamics.

Implications

Practical Impact

Automating bond dimension provision removes a key empirical tuning step, lowering the technical barrier for large-scale tensor network simulation. The framework's per-bond adaptivity translates directly into fewer SVDs and reduced memory and computational costs. The hybrid CPU/GPU implementation facilitates acceleration for problem sizes both within and beyond single-device RAM, and its Python-centric design allows integration with existing TN libraries.

Theoretical Considerations

The closed-loop PID approach generalizes the relationship between entropy and bond dimension beyond heuristic or open-loop thresholding. Stability analysis using Jury’s criterion confirms controller robustness as the loop gain varies with the entanglement spectrum (saturated, exponential, or power-law decay regimes). The work lays a foundation for feedback-based resource allocation in tensor network algorithms more broadly, including higher-dimensional and non-MPS networks.

Limitations and Future Directions

While the PID controller introduces three gains demanding empirical tuning, the required effort is outweighed by reduced simulation trial runs. The adaptive method’s advantage diminishes in low-entanglement regimes (small $i$1), and gains from GPU offload appear only for sufficiently large matrices. The predictive scheduler is most effective for monotonic entropy growth; further work may employ higher-order forecasting or reinforcement learning controllers. Extension to PEPS, TTN, and other higher-dimensional TN structures is anticipated. Full Lyapunov-based control theoretic analyses and integration into production TN codes (e.g., ITensor, TeNPy) are identified as natural next steps.

Conclusion

This work establishes entropy-feedback PID-controlled, per-bond adaptive bond dimension allocation as a significant improvement over fixed-dimension strategies in MPS simulations, showing both formal control-theoretic grounding and empirical performance gains. GPU-accelerated SVD compounds the benefit for large-scale simulations. The framework’s demonstrated robustness and accuracy, validated against independent classical simulators and quantum hardware, positions it as the state-of-the-art option for automated, scalable simulation of 1D quantum many-body systems. Future extensions to time-dependent and higher-dimensional TNs offer a clear developmental trajectory.