Efficient implementation of the parquet equations -- role of the reducible vertex function and its kernel approximation

Gang Li, Nils Wentzell, Petra Pudleiner, Patrik Thunstr, Karsten HeldarXiv 2015 · DOI · arXiv

Abstract

We present an efficient implementation of the parquet formalism which respects the asymptotic structure of the vertex functions at both single- and two-particle levels in momentum- and frequency-space. We identify the two-particle reducible vertex as the core function which is essential for the construction of the other vertex functions. This observation stimulates us to consider a two-level parameter-reduction for this function to simplify the solution of the parquet equations. The resulting functions, which depend on fewer arguments, are coined "kernel functions". With the use of the "kernel functions", the open boundary of various vertex functions in the Matsubara-frequency space can be faithfully satisfied. We justify our implementation by accurately reproducing the dynamical mean-field theory results from momentum-independent parquet calculations. The high-frequency asymptotics of the single-particle self-energy and the two-particle vertex are correctly reproduced, which turns out to be essential for the self-consistent determination of the parquet solutions. The current implementation is also feasible for the dynamical vertex approximation.

Introduction

In the study of strongly correlated electron systems, the parquet formalism stands as a comprehensive framework, ensuring self-consistency at both single- and two-particle levels. Despite its theoretical elegance, the computational application of the parquet equations has been severely limited by the high dimensionality and slow decay of the two-particle vertex functions in Matsubara frequency space. This paper introduces a significant algorithmic advancement: a "kernel approximation" that simplifies these complex functions, making the parquet equations computationally tractable without sacrificing physical accuracy.

The Complexity of Vertex Functions

The primary obstacle in the parquet formalism is the two-particle vertex function, which depends on three independent variables. In the imaginary frequency domain, these functions do not decay rapidly, creating an "open boundary" problem. Standard truncation methods often lead to unphysical results because they fail to capture the high-frequency "tails" necessary for self-consistency.

Major structures of the vertex function in frequency space.

As shown in Figure 1, the vertex function $F_d(k, k')$ exhibits a complex landscape consisting of a smooth background and sharp diagonal structures. The authors identify the reducible vertex—the part of the interaction that can be split by cutting two propagator lines—as the "engine" of the parquet equations. If this component is approximated correctly, the rest of the vertex structure follows naturally.

Methodology: The Kernel Approximation

To address the dimensionality and boundary issues, the authors propose a two-level parameter reduction for the reducible vertex function. This reduction results in "kernel functions" that depend on fewer arguments:

  1. Level 1: A single-frequency kernel captures the broad background of the vertex.
  2. Level 2: A two-frequency kernel accounts for the specific "stripes" or diagonal structures in frequency space.

By using these kernels, the implementation can analytically account for the high-frequency asymptotics. This allows the solver to achieve high precision with as few as 60 frequency points, whereas traditional methods might require hundreds or thousands to achieve similar stability.

Validation and Results

The implementation was benchmarked using the 2D Hubbard model, comparing results against numerically exact data from Dynamical Mean-Field Theory (DMFT) obtained via a Continuous-Time Interaction Expansion (CT-INT) solver.

Comparison of self-energy results between DMFT and parquet methods.

As illustrated in Figure 4, the kernel-based parquet implementation accurately reproduces the DMFT self-energies. The method demonstrates remarkable stability, converging for interaction strengths up to $U=8$, a regime where previous implementations often struggled. Furthermore, the implementation respects fundamental physical constraints such as crossing symmetry and correctly reproduces high-frequency asymptotics, which are essential for self-consistent solutions.

Conclusion

The introduction of kernel functions for the reducible vertex represents a major step forward in making the parquet equations a practical tool for many-body physics. By solving the open boundary problem and reducing computational memory requirements, this method opens the door to studying more complex models, including multi-orbital systems and those with significant spatial correlations via the Dynamical Vertex Approximation (D$\Gamma$A).

Summary updated

Abstract

We present an efficient implementation of the parquet formalism which respects the asymptotic structure of the vertex functions at both single- and two-particle levels in momentum- and frequency-space. We identify the two-particle reducible vertex as the core function which is essential for the construction of the other vertex functions. This observation stimulates us to consider a two-level parameter-reduction for this function to simplify the solution of the parquet equations. The resulting functions, which depend on fewer arguments, are coined "kernel functions". With the use of the "kernel functions", the open boundary of various vertex functions in the Matsubara-frequency space can be faithfully satisfied. We justify our implementation by accurately reproducing the dynamical mean-field theory results from momentum-independent parquet calculations. The high-frequency asymptotics of the single-particle self-energy and the two-particle vertex are correctly reproduced, which turns out to be essential for the self-consistent determination of the parquet solutions. The current implementation is also feasible for the dynamical vertex approximation.
BibTeX
@misc{li2015efficient,
  author = {Gang Li and Nils Wentzell and Petra Pudleiner and Patrik Thunstr and Karsten Held},
  title = {{Efficient implementation of the parquet equations -- role of the reducible vertex function and its kernel approximation}},
  year = {2015},
  eprint = {1510.03330},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1510.03330},
}
Figure