Abstract
The computational cost of quantum Monte Carlo (QMC) calculations of realistic periodic systems depends strongly on the method of storing and evaluating the many-particle wave function. Previous work has demonstrated the reduction of the $O(N^3)$ cost of evaluating the Slater determinant with planewaves to $O(N^2)$ using localized basis functions. We compare four polynomial approximations as basis functions—interpolating Lagrange polynomials, interpolating piecewise-polynomial-form (pp-) splines, and basis-form (B-) splines (interpolating and smoothing). All these basis functions provide a similar speedup relative to the planewave basis. The pp-splines have eight times the memory requirement of the other methods. To test the accuracy of the basis functions, we apply them to the ground state structures of Si, Al, and MgO. The polynomial approximations differ in accuracy most strongly for MgO and smoothing B-splines most closely reproduce the planewave value for the variational Monte Carlo energy. Using separate approximations for the Laplacian of the orbitals increases the accuracy sufficiently to justify the increased memory requirement, making smoothing B-splines, with separate approximation for the Laplacian, the preferred choice for approximating planewave-represented orbitals in QMC calculations.
Introduction
Quantum Monte Carlo (QMC) methods are essential for high-precision electronic structure calculations in materials science. However, their application to periodic systems is often limited by computational expense. In these systems, electron orbitals are typically represented using a planewave basis. The primary bottleneck is the evaluation of the Slater determinant, which scales as $O(N^3)$ with the number of electrons $N$ when using planewaves. To make QMC practical for large-scale simulations, it is necessary to transition to localized basis functions that can reduce this scaling to $O(N^2)$.
Methodology
This study evaluates four localized polynomial approximation methods designed to replace spatially extended planewaves:
- Interpolating Lagrange Polynomials: A local approximation utilizing the 64 nearest grid points.
- Interpolating Piecewise-Polynomial-Form (pp-) Splines: These ensure continuous derivatives across grid points but require significant memory (eight coefficients per grid point).
- Interpolating B-splines: A basis-form spline mathematically equivalent to pp-splines in value but significantly more memory-efficient, requiring only one coefficient per grid point.
- Smoothing B-splines: These are constructed to exactly reproduce the Fourier (planewave) coefficients of the original orbital rather than simply interpolating grid points.
An additional refinement investigated is the use of a separate polynomial approximation for the Laplacian ($\nabla^2\phi$). While calculating the Laplacian directly from the orbital spline is memory-efficient, it often introduces significant errors in the kinetic energy calculation. Creating a dedicated spline for the Laplacian, though increasing memory usage by 2x to 5x, aims to recover planewave-level precision.
Benchmarking and Results
The methods were tested on three periodic systems: covalent diamond silicon (Si), metallic fcc aluminum (Al), and ionic rock-salt magnesium oxide (MgO). While all methods performed adequately for Si and Al, MgO served as a rigorous "litmus test" due to its sharp orbital variations.
As shown in the figure above, smoothing B-splines consistently outperformed Lagrange and interpolating splines across all error metrics—orbital value, gradient, and Laplacian. The error reduction is particularly pronounced at lower planewave cutoffs, which are common in practical simulations.
All polynomial methods achieved a similar, substantial speedup over the original planewave basis. For a 64-atom system, the evaluation time was reduced by approximately one order of magnitude. However, the accuracy of the smoothing B-splines in reproducing the Variational Monte Carlo (VMC) energy and local energy fluctuations makes them the superior choice.
Conclusion
The transition from $O(N^3)$ to $O(N^2)$ scaling is critical for the future of QMC in material science. This research identifies smoothing B-splines, combined with a separate approximation for the Laplacian, as the optimal strategy. This combination provides the best balance of computational speed, memory efficiency (compared to pp-splines), and the high accuracy required to match original planewave results. While the separate Laplacian approximation increases memory overhead, the gain in precision justifies the cost for complex systems like MgO.
Summary updated