Bridging constrained random-phase approximation and linear response theory for computing Hubbard parameters

TL;DR

This paper compares two methods for calculating the Hubbard U parameter in DFT-based methods, revealing their differences, limitations, and how to achieve consistent results by understanding their underlying assumptions.

Contribution

It provides a formal connection between LRT and cRPA, clarifies their discrepancies, and offers guidelines for consistent U calculations using Wannier functions.

Findings

LRT and cRPA U values can differ by up to 30%.

Neglecting exchange-correlation response causes discrepancies.

LRT remains reliable in strongly hybridized systems.

Abstract

The predictive accuracy of popular extensions to density-functional theory (DFT) such as DFT+U and DFT plus dynamical mean-field theory (DFT+DMFT) hinges on using realistic values for the screened Coulomb interaction U. Here, we present a systematic comparison of the two most widely used approaches to compute this parameter, i.e. linear response theory (LRT) and the constrained random-phase approximation (cRPA), using a unified framework based on the use of maximally localized Wannier functions. We show that the U in LRT and cRPA can differ as much as 30%. We demonstrate that this discrepancy arises from two main differences: neglecting the response of the exchange-correlation potential in cRPA and additional excitation channels in LRT. By taking these differences into account, we can achieve near perfect agreement between the two techniques. Moreover, we show that in cases with strong hybridization between interacting and screening subspaces, the application of cRPA becomes ambiguous and can lead to unrealistically small U values, while LRT remains well-behaved. Our work formally connects both methods, sheds light on their strengths and limitations, and emphasizes the importance of using a consistent set of Wannier orbitals to ensure transferability of U values between different implementations.