Symmetry Breaking with the SCAN Density Functional Describes Strong Correlation in the Singlet Carbon Dimer
Page 2 of 10
(2)
is the positive orbital kinetic energy density. Rung 1 of the ladder (e.g., LSDA1,2.4) uses only the local
spin densities, rung 2 (e.g., PBE5 GGA) adds the gradients of the spin densities, and rung 3 (e.g., SCAN6
or r2SCAN7 meta-GGA) adds the kinetic energy densities.
There is a nearly straight line8 from the exact density functional theory to its first-principles
approximations (e.g., LSDA, PBE, or SCAN), which proceeds from exact (if impractical) expressions for
to the derivation of mathematical properties of the exact functional that can be satisfied by an
approximation on a given rung, and finally to imposing those constraints on the approximate functional.
Appropriate norms, or systems in which a given rung can be accurate for the exchange energy alone and
the correlation energy alone, can be safely fitted. Appropriate norms for all three rungs include electron
gases of uniform spin densities. Appropriate norms for rung 3 include closed-subshell atoms. LSDA
inherits 8 exact constraints from its uniform gas norm, PBE satisfies 11 exact constraints appropriate to
the GGA level, and SCAN satisfies all 17 known exact constraints that a meta-GGA can satisfy. The
energies of molecules and solids, often fitted by empirical functionals, are not appropriate norms, because
of an understood but imperfect and uncontrolled cancellation between errors in the approximate exchange
and correlation energies. While empirical functionals interpolate between bonds, first-principles
approximations predict9 them, typically working better than empirical functionals for artificial molecules
that are unlike those commonly fitted.
SCAN with a long-range van der Waals correction was more accurate than other tested
functionals from the first three rungs, and more accurate than some hybrid functionals employing a
fraction of exact exchange, in a 2017 test10 on the GMTKN55 suite of 55 molecular test sets. SCAN is
also accurate for insulating solids. For some strongly-correlated sd-bound solids, such as the cuprates11
and manganese dioxides12, symmetry-broken SCAN is accurate without a “Hubbard-like:”+U correction
(sometimes interpreted as a self-interaction correction13). In other strongly-correlated transition-metal
oxides, symmetry-broken SCAN requires a +U correction, but one significantly smaller than PBE
requires14. Range-separated GGA hybrid functionals with short-range exact exchange can describe15 many
strongly-correlated solids better than LSDA or GGA and without a material-dependent parameter like U.
Such functionals16 and other nonlocal functionals17 can also describe polaronic symmetry breaking.
For equilibrium bonds, a self-interaction correction seems to be much more needed for sd
than for sp bonds. The Perdew-Zunger self-interaction correction18 is first-principles, but is not reliably
accurate due to the lobedness of its localized one-electron densities19. Perhaps in the future an improved
self-interaction correction to a SCAN-like functional will lead to a reliable and widely-useful description
of strong correlation via symmetry breaking. For now, any strongly-correlated equilibrium sp bond is of
special interest as a test of the ability of symmetry breaking to describe strong correlation. Such bonds are
rare, but the singlet ground state of the molecule C2 is known to be strongly correlated. This molecule has
an avoided crossing of two energy surfaces near its equilibrium bond length, and has been studied carefully
with the full configuration interaction quantum Monte Carlo (FCIQMC) correlated-wavefunction
approach.20
Symmetry breaking (in both wavefunction and density functional theories) and strong
correlation have attracted the interest of many, notably Gustavo E. Scuseria and Richard L. Martin15,22,23.
Strong correlation arises when degenerate or nearly degenerate Slater determinants are strongly mixed by
the electron-electron pair interaction. This leads to correlation or exchange-correlation energies more
negative than those that a standard approximate density functional can produce. A standard approximate