Energy vs. density on paths toward more exact density functionals

TL;DR

This paper proposes a systematic path to improve density functional theory accuracy by analyzing errors in densities and energies, using diverse test sets to benchmark functionals and identify when simpler densities suffice.

Contribution

It introduces a formal framework for measuring progress in density functional theory and demonstrates how diverse test sets can evaluate the significance of density errors.

Findings

Errors in densities are often insignificant for atomic cations up to Z=10.

Oscillating density sensitivity is linked to orbital occupation effects.

Simpler trial densities can be used for large systems with minimal accuracy loss.

Abstract

Recently, the progression toward more exact density functional theory has been questioned, implying a need for more formal ways to systematically measure progress, i.e. a path. Here I use the Hohenberg-Kohn theorems and the definition of normality by Burke et al. to define a path toward exactness and straying from the path by separating errors in \r{ho} and E[\r{ho}]. A consistent path toward exactness involves minimizing both errors. Second, a suitably diverse test set of trial densities \r{ho}' can be used to estimate the significance of errors in \r{ho} without knowing the exact densities which are often computationally inaccessible. To illustrate this, the systems previously studied by Medvedev et al., the first ionization energies of atoms with Z = 1 to 10, the ionization energy of water, and the bond dissociation energies of five diatomic molecules were investigated and benchmarked against CCSD(T)/aug-cc-pV5Z. A test set of four functionals of distinct designs was used: B3LYP, PBE, M06, and S-VWN. For atomic cations regardless of charge and compactness up to Z = 10, the energy effects of variations in \r{ho} are < 4 kJ/mol (chemical accuracy) defined here as normal, even though these four functionals ranked very differently in the previous test. Thus, the off-path behavior for such cations is energy-wise insignificant and in fact, indeterminate because of noise from other errors. An interesting oscillating behavior in the density sensitivity is observed vs. Z, explained by orbital occupation effects. Finally, it is shown that even large normal problems such as the Co-C bond energy of cobalamins can use simpler (e.g. PBE) trial densities to drastically speed up computation by loss of a few kJ/mol in accuracy.