[Paper Review] TensorCircuit: a Quantum Software Framework for the NISQ Era

Shi‐Xin Zhang, Jonathan Allcock+14 more|arXiv (Cornell University)|May 20, 2022

Quantum Computing Algorithms and Architecture4 citations

TL;DR

TensorCircuit is an open-source quantum software framework built on tensor network contraction, enabling high-performance simulation of large-scale parameterized quantum circuits in the NISQ era. It leverages machine learning backends for automatic differentiation, JIT compilation, and hardware acceleration, achieving up to 600-qubit simulations with low-depth, one-dimensional connectivity and orders-of-magnitude speedup over existing simulators.

ABSTRACT

TensorCircuit is an open source quantum circuit simulator based on tensor network contraction, designed for speed, flexibility and code efficiency. Written purely in Python, and built on top of industry-standard machine learning frameworks, TensorCircuit supports automatic differentiation, just-in-time compilation, vectorized parallelism and hardware acceleration. These features allow TensorCircuit to simulate larger and more complex quantum circuits than existing simulators, and are especially suited to variational algorithms based on parameterized quantum circuits. TensorCircuit enables orders of magnitude speedup for various quantum simulation tasks compared to other common quantum software, and can simulate up to 600 qubits with moderate circuit depth and low-dimensional connectivity. With its time and space efficiency, flexible and extensible architecture and compact, user-friendly API, TensorCircuit has been built to facilitate the design, simulation and analysis of quantum algorithms in the Noisy Intermediate-Scale Quantum (NISQ) era.

Motivation & Objective

• Address the growing need for efficient simulation of large, complex parameterized quantum circuits in the Noisy Intermediate-Scale Quantum (NISQ) era.
• Overcome limitations of existing simulators in handling variational quantum algorithms with high parameter counts and deep circuits.
• Integrate modern machine learning engineering practices—automatic differentiation, JIT compilation, and hardware acceleration—into a quantum simulation framework.
• Enable scalable simulation of up to 600 qubits with moderate circuit depth and 1D connectivity, supporting advanced quantum algorithm development and analysis.

Proposed method

• Builds a flexible tensor network engine for efficient quantum circuit simulation using tensor contraction.
• Uses industry-standard machine learning frameworks (e.g., TensorFlow) as backends to enable automatic differentiation and JIT compilation.
• Supports vectorized parallelism and hardware acceleration via integration with ML libraries, improving performance on GPUs.
• Implements advanced features such as conditional measurements, post-selection, and Pauli string expectation evaluation with vmap for batched operations.
• Provides QuOperator and QuVector classes for representing quantum states and operators, enabling modular and extensible circuit construction.
• Offers customizable contraction path finding and subtree reconfiguration to optimize tensor network contraction performance.

Experimental results

Research questions

• RQ1 Can a tensor network-based quantum circuit simulator achieve orders-of-magnitude speedup over existing simulators while supporting variational quantum algorithms?
• RQ2 To what extent can integration with machine learning frameworks enhance performance and usability in quantum circuit simulation?
• RQ3 How efficiently can tensor network contraction simulate large-scale quantum circuits with up to 600 qubits and low-dimensional connectivity?
• RQ4 Can advanced features like batched gradient computation, Monte Carlo noise simulation, and conditional measurements be efficiently supported in a pure Python framework?
• RQ5 What is the performance ceiling of a simulator that combines tensor network methods with modern ML engineering practices in the NISQ context?

Key findings

• TensorCircuit successfully simulates circuits with up to 600 qubits and moderate depth using a 1D ladder layout of two-qubit gates, demonstrating scalability beyond typical simulators.
• For a 600-qubit transverse field Ising model (TFIM) with 7 layers and 4900 gates, the framework achieved a ground energy accuracy of 99.4% with no hyperparameter tuning.
• The time per computational step for the 600-qubit case was 18.2 seconds, including both energy expectation value and gradient computation, using an NVIDIA A100 40GB GPU.
• The framework achieved orders-of-magnitude speedup in simulation tasks compared to common quantum software, particularly for variational quantum eigensolver (VQE) workflows.
• The use of SVD decomposition on two-qubit gates reduced the bond dimension to d=2 for XX-type gates, enabling more efficient tensor network contraction.
• Batched operations via vmap and vectorized_value_and_grad enabled efficient evaluation of cost functions and gradients over multiple input states or circuit parameters, accelerating optimization.