Researchers at Google Quantum AI have developed a new heuristic approach to accelerate Matrix Product State (MPS) simulations, allowing for estimation of expectation values without the traditionally exponential resource demand of storing the full wave function. The work addresses a computational bottleneck at intermediate energy densities, where simulating both wave function and operator evolution proves costly.

This technique rescales MPS results at low bond dimensions using a factor dependent on the fidelity of the MPS wave function, enabling simulations of the two-dimensional Transverse-Field Ising Model on a 7×8 grid with a maximum bond dimension of 4096 on a single A100 GPU, as well as the dynamics of the two-dimensional XY model on grids of size up to 9×9. The team also demonstrated excellent agreement between their MPS simulations and those run on a digital quantum processor, confirming the method’s predictive power.

Matrix Product State Simulation for Out-of-Equilibrium Quantum Systems

Google Quantum AI researchers have developed a method to estimate expectation values of local operators in quantum systems using a bond dimension smaller than previously required for high precision. This advancement addresses a key challenge in simulating the behavior of complex quantum materials, particularly those exhibiting thermalization, where systems converge to a steady state with a fixed energy density. The team, comprised of Salvatore Mandrà, Brayden Ware, Nikita Astrakhantsev, Sergei Isakov, Benjamin Villalonga, Tom Westerhout, and Kostyantyn Kechedzhi, focused on accelerating Matrix Product State (MPS) simulations for two-dimensional quantum spin systems.

The approach allows for efficient computation even when initializing simulations far from the ground state, a scenario that typically generates highly entangled dynamics. Classical simulation of quantum dynamics demands exponential resources, but computations often focus on estimating expectation values of local operators and correlation functions to a defined precision.

The work builds on the understanding that at sufficiently high energy densities, thermalizing systems can be approximated as product operators, simplifying calculations. The new heuristic tackles this bottleneck by enabling MPS simulations to converge more rapidly, reducing the computational demands without sacrificing accuracy. The researchers’ technique uses the observation that error estimates often predict larger errors than those actually observed during early simulation times. This means the simulation reaches a stable and reliable result faster, requiring less computational power and time.

The team demonstrated the method’s effectiveness in simulating the out-of-equilibrium dynamics of the transverse field Ising model, a benchmark for quantum magnetism and critical phenomena, as detailed in a recent experiment. Extrapolating the rescaling to larger system sizes does introduce limitations, as error estimates eventually become significant compared to the physical signal being simulated.

Nevertheless, the approach represents a step toward more efficient quantum simulations. “A Heuristic for Matrix Product State Simulation of Out-of-Equilibrium Dynamics of Two-Dimensional Quantum Spin Systems,” the team states, describes their method for accelerating convergence. The researchers acknowledge contributions from Sergio Boixo, Vadim Smelyanskiy, Michael Foss-Feig, and Andrew Potter for useful discussions and comments during the development of this work. The team’s findings open avenues for exploring more complex quantum systems and phenomena with reduced computational cost, potentially accelerating progress in areas like materials science and fundamental physics.

Heuristic Fidelity Rescaling Accelerates MPS Convergence

Simulations of quantum systems now achieve accelerated convergence through a new heuristic approach, allowing researchers to estimate expectation values using significantly smaller bond dimensions than previously required. This technique, detailed in recent work, enables computations without storing the complete wave function, a process traditionally demanding exponential computational resources. The researchers compare our TFIM results to similar simulations on a digital quantum processor, demonstrating excellent agreement and confirming the predictive power of the method.

While the specific values of fitted parameters depend on the ordering of calculations and how entanglement is distributed and truncated, the validity of the rescaling heuristic remains independent of this choice. The observed power-law scaling of expectation values with fidelity arises from the truncation process and the chaotic nature of the dynamics, not the specific computational ordering. Further analysis indicates the approach generalizes beyond the initial test case.

Researchers observed similar accelerated convergence of MPS in simulations of quench dynamics for the 2D XY model, strengthening the potential for wider application. The team found that even with deviations from a linear fit in some parameters, scaling with the fitted exponent still substantially accelerates convergence.

The work details contributions from several researchers; S.M. designed the rescaling heuristic and developed the TFIM MPS simulator, while B.W. developed the XY model MPS simulator and conducted the 2D XY simulations, and K.K. focused on the theoretical analysis and interpretation of the results. These combined efforts demonstrate a pathway toward more efficient classical algorithms for simulating complex quantum systems, even as challenges remain in optimizing resource usage at intermediate energy densities. The researchers acknowledge that while the absolute computational cost and required bond dimension to reach a given precision depend on the ordering choice, the heuristic’s validity is independent of it.

Two-Dimensional Transverse-Field Ising Model Implementation

This achievement extends to the two-dimensional XY model, with simulations performed on grids up to 9×9, showcasing the technique’s versatility beyond the initial TFIM implementation. The ability to handle these grid sizes without exponential resource scaling represents an advance in simulating quantum dynamics. Researchers compared TFIM simulation results with those obtained from a digital quantum processor, revealing excellent agreement and validating the predictive capability of the new method.

This direct comparison strengthens confidence in the accuracy of the approach and its potential to model quantum systems with greater fidelity. Further analysis demonstrated the method’s adaptability by mentioning the three-dimensional TFIM in the supplemental material. The TFIM Hamiltonian, defined for a general D-dimensional lattice, incorporates Pauli operators and coupling strengths to model interactions between neighboring spins, a fundamental aspect of quantum magnetism.

Simulations focused on calculating the magnetization at fixed parameters, (Δθ, J, h, dt) = (2π/9, -1, 2, 0.25), providing a specific test case for the rescaling heuristic. The choice of the XY model, with its conserved U(1) charge and vortex excitations, further broadened the scope of testing beyond the spin-flip dynamics of the TFIM, revealing the method’s capacity to handle diverse physical phenomena.

The work also highlights the potential for accelerating Matrix Product State (MPS) calculations, a key benefit for simulating out-of-equilibrium dynamics in quantum spin systems. By applying the rescaling heuristic, researchers aim to improve the convergence of MPS simulations, reducing the computational burden associated with accurately representing highly entangled wave functions.

TFIM Dynamics Verified Against Quantum Processor Results

This direct comparison revealed excellent agreement between the MPS predictions and experimental results, bolstering confidence in the method’s ability to accurately model quantum dynamics. Beyond confirming the accuracy of the rescaling heuristic, the simulations offer a glimpse into the limits of current computational resources. Exact tensor network contraction was limited to systems of size up to 6×8, while the MPS method, aided by the rescaling technique, extended simulations to significantly larger grids.

This capability is important for studying systems exhibiting complex behavior, where capturing interactions across larger scales is essential for understanding emergent phenomena. The researchers tracked the total truncated weight in their MPS simulations using the MPS fidelity, F_(MPS), detailed in supplemental material.

Analysis of the order parameter dynamics in the XY model mirrored the accelerated convergence observed in the TFIM simulations, suggesting a general principle applicable to a wider range of quantum spin systems. “We find that this behavior generalizes to other quantum spin systems,” the researchers report, indicating a potentially versatile tool for tackling complex quantum simulations. S.M. focused on the XY model, developing its corresponding MPS simulator, and K.K. The team’s contributions highlight a collaborative effort to address the challenges of simulating quantum many-body systems, pushing the boundaries of what is computationally feasible.

Energy Density Impacts Classical Simulation Efficiency

Simulating quantum systems at high energy densities circumvents a major computational hurdle by allowing researchers to focus on the evolution of local operators rather than the full wave function, a traditionally exponential resource demand. This approach, detailed in recent work, demonstrates that for systems already exhibiting thermalization, tracking how these localized properties change over time requires significantly fewer computational resources than mapping the entire quantum state.

However, achieving this streamlined simulation is not universally straightforward; intermediate energy densities present a persistent challenge. The study highlights that correlation functions, key to understanding system behavior, can be efficiently computed at high energy densities using classical algorithms, but more complex observables still pose difficulties.

The researchers found that the bond dimension required for Matrix Product State simulations scales exponentially with the width of the rectangular grid used in their models, a predictable outcome for states exhibiting short-range entanglement. Importantly, all states examined were within the high-temperature phase of the XY model, a specific quantum system used for testing the new approach. The energy density at the BKT transition is approximately -0.53g, a value not reached by the states simulated in this work.

Application to Two-Dimensional XY Model Dynamics

Google Quantum AI researchers applied this technique, initially developed for the transverse-field Ising model, to a system characterized by a conserved U(1) charge and the presence of vortex excitations, features absent in the earlier TFIM studies. This broadened scope confirms the method’s ability to handle more complex quantum phenomena beyond simple spin flips. The rescaling heuristic’s success with the XY model involved simulating dynamics on grids of size up to 9×9.

This capability allowed for a direct comparison of MPS predictions against exact numerical results obtained through tensor network contraction, up to a system size of 6×8, validating the accuracy of the accelerated simulations. The researchers found that the observed accelerated convergence of MPS calculations mirrored results obtained with the TFIM, suggesting a consistent underlying principle at work across different quantum spin systems.

Further analysis revealed that the technique’s effectiveness isn’t limited to the TFIM; it generalizes to other quantum spin systems as well. The choice of the XY model was deliberate, introducing physical elements not present in the TFIM dynamics. These include the aforementioned conserved U(1) charge and vortex excitations, which represent a different type of order within the system. This deliberate complexity served as a stringent test for the rescaling ansatz, proving its robustness beyond simpler models.

👉 More information
🗞 A Heuristic for Matrix Product State Simulation of Out-of-Equilibrium Dynamics of Two-Dimensional Quantum Spin Systems
✍️ Salvatore Mandrà, Brayden Ware, Nikita Astrakhantsev, Sergei Isakov, Benjamin Villalonga, Tom Westerhout and Kostyantyn Kechedzhi
🧠 DOI: http://link.aps.org/doi/10.1103/qq2m-v44w

More like this