Fast-update in self-learning algorithm for continuous-time quantum Monte Carlo

Abstract

We propose a novel technique for speeding up the self-learning Monte Carlo method applied to the single-site impurity model. For the case where the effective Hamiltonian is expressed by polynomial functions of differences of imaginary-time coordinate between vertices, we can remove the dependence of CPU time on the number of vertices, $n$, by saving and updating some coefficients for each insertion and deletion process. As a result, the total cost for a single-step update is drastically reduced from $O(nm)$ to $O(m^2)$ with $m$ being the order of polynomials in the effective Hamiltonian. Even for the existing algorithms, in which the absolute value is used instead of the difference as the variable of polynomial functions, we can limit the CPU time for a single step of Monte Carlo update to $O(m^2+m\log n)$ with the help of balanced binary search trees. We demonstrate that our proposed algorithm with only logarithmic $n$-dependence achieves an exponential speedup from the existing methods, which suffer from severe performance issues at low temperatures.

arXiv comment6 pages, 1 figure

References in corpus (10)

• Continuous-time Monte Carlo methods for quantum impurity models
• Self-Learning Monte Carlo Method
• Chebyshev-BdG: an efficient numerical approach to inhomogeneous superconductivity
• Chebyshev Matrix Product State Impurity Solver for the Dynamical Mean-Field Theory
• Sub-matrix updates for the Continuous-Time Auxiliary Field algorithm
• Efficient Numerical Self-consistent Mean-field Approach for Fermionic Many-body Systems by Polynomial Expansion on Spectral Density
• Numerical evaluation of Greens functions based on the Chebyshev expansion
• Self-learning Monte Carlo method with Behler-Parrinello neural networks
• Heat transport via a local two-state system near thermal equilibrium
• Low-temperature density matrix renormalization group using regulated polynomial expansion