Second-Order Methods Compress Quantum Circuits and Improve Optimisation of Spin Chains

Isabel Nha Minh Le and colleagues have achieved up to a four-order-of-magnitude improvement in fidelity when optimising tensor networks, compared to conventional methods like naive Trotterization. Optimisation of these networks previously faced limitations due to slow convergence and susceptibility to local minima, but the team developed an analytical Hessian-vector product kernel. This new kernel enables scalable second-order optimisation without constructing the full Hessian matrix, using recursive tangent-state propagation.

A new computational method refines complex systems called tensor networks, sharply improving their accuracy over current techniques. This method avoids calculating a large, complex matrix, streamlining the optimisation process for these networks and offering a practical benefit to quantum circuit compression. The technique refines how computers handle tensor networks, a complex web of interconnected variables used to represent high-dimensional data. Current optimisation methods often struggle with slow progress and getting stuck in suboptimal solutions, but this new computational approach bypasses a major bottleneck by avoiding the need to calculate a large matrix. This innovation utilises a mathematical shortcut, a Hessian-vector product kernel, to efficiently calculate how the system changes, and was successfully tested on quantum circuit compression, achieving up to a four-order-of-magnitude improvement in accuracy compared to existing techniques.

Scalable second-order optimisation enhances fidelity of quantum spin chain simulations

A four-order-of-magnitude improvement in fidelity was achieved when optimising time-evolution circuits for spin chains, exceeding the accuracy of naive Trotterization methods. Now, more intricate and powerful circuits become viable, as previous optimisation techniques struggled to reliably compress quantum circuits beyond a certain complexity. The team at [Institution Name] developed an analytical Hessian-vector product kernel, a computational shortcut that avoids the prohibitive cost of calculating a full Hessian matrix, a large representation of a system’s curvature.

The transverse-field Ising model demonstrated the benefits of this new optimisation technique, simulating a chain of 50 sites with specific magnetic field settings. A reference unitary generated via a fourth-order Trotterization with 20 repetitions established a baseline for comparison in this benchmark. Further validation involved a Heisenberg chain of 40 sites, parameterised with specific values for spin interactions and magnetic fields, and evolved for a time of 0.25 using the same high-order Trotter scheme. These simulations highlighted the method’s ability to handle complex systems and provided a quantitative measure of its performance against established techniques, including a comparison with Riemannian ADAM using circuits composed of 11 layers enforcing translational invariance to simplify the analysis.

Analytical Hessian-vector products via recursive tangent-state propagation

This advance in optimising tensor networks, a complex web of interconnected variables used to represent high-dimensional data, was underpinned by recursive tangent-state propagation. An analytical Hessian-vector product kernel was devised as a mathematical shortcut for calculating how the system changes, rather than calculating a full Hessian matrix, a computationally expensive undertaking. This kernel efficiently computes the action of the Hessian on a vector without building the entire matrix, streamlining calculations for large systems and resembling the use of a gradient to find the steepest path downhill. The technique uses a two-pass algorithm, propagating information through the network to estimate curvature efficiently, with a bounded virtual bond dimension keeping computational demands manageable.

Recursive tangent-state propagation balances fidelity and efficiency using bounded virtual bond

Optimising the complex tensor networks underpinning calculations remains a substantial hurdle, despite increasingly precise simulations being delivered across fields like materials science and quantum chemistry. The team acknowledges reliance on a ‘bounded virtual bond dimension’ to manage computational load, although this work offers a strong advance. This constraint introduces a potential trade-off between accuracy and efficiency, and the abstract provides limited detail on how this balance shifts with increasingly complex systems or different network architectures.

Nevertheless, the substantial gains demonstrated, a four-order-of-magnitude improvement in fidelity over standard methods, remain a striking achievement and warrant further investigation into mitigating the limitations of the bounded virtual bond dimension. This advance opens questions regarding the application of this kernel to different tensor network structures and the potential for extending it to two-dimensional simulations. By calculating how a system changes without needing to build a large, computationally expensive matrix, the method circumvents a key limitation of previous approaches, delivering substantially improved optimisation, particularly for quantum circuits, and representing a major step towards more reliable simulations and a pathway to tackling previously intractable problems in quantum physics and related fields.

The researchers developed a new method for optimising tensor networks, achieving up to a four-order-of-magnitude improvement in fidelity compared to naive Trotterization. This advance bypasses the need to construct a full Hessian matrix, a computationally expensive process for large systems, by efficiently calculating its action on a vector. The technique utilises recursive tangent-state propagation with a bounded virtual bond dimension to maintain scalability. The authors suggest further work could explore applying this kernel to different tensor network structures and two-dimensional simulations.