Self-Interaction and Delocalization Errors

Introduction

An electron repels other electrons, but it should not repel itself. In Kohn–Sham DFT, the classical Coulomb term treats the total density as a continuously charged cloud and therefore includes a self-repulsion contribution. The exact exchange–correlation functional cancels the unphysical part appropriately. Many approximate functionals cancel it imperfectly, which can encourage excess charge spreading. This matters for electron transfer, radicals, anions, dissociation and reaction barriers where the location of an electron changes. The error is a property of the approximation, not a claim that electrons are classical point particles.

Core explanation

The classical Coulomb energy J[n] is calculated from pairs of density elements. If a system contains only one electron, J[n] is nonzero even though there is no second electron with which to repel. In an exact treatment, an exchange–correlation term cancels that unphysical one-electron contribution. Local and semilocal approximations generally do not cancel it perfectly. Perdew and Zunger's original self-interaction study formulated this problem and proposed corrections. A finite fraction of exact exchange in a hybrid may reduce some cases, but partial mixing is not a universal elimination of all self-interaction or delocalization errors.

An illuminating diagnostic concerns a fractional electron number between two neighboring integers while the external potential is held fixed. This is an ensemble construction, not a claim that an isolated measured molecule literally contains 6.3 electrons. For the exact ground-state energy, the energy versus electron number is piecewise linear between adjacent integers under the relevant formal conditions. Many common approximate functionals produce a bowed, convex curve. That curvature can make a fractional distribution of charge across separated fragments artificially favorable compared with an integer-localized distribution. Primary analysis by Mori-Sánchez, Cohen and Yang links convex fractional-charge behavior to delocalization error.

Consider two far-separated equivalent molecular sites sharing one extra electron. Exact physics at infinite separation supports appropriate degenerate integer-localized descriptions and their ensemble, with no spurious energetic reward for smearing a fractional extra charge over both sites. A convex approximate energy curve can make a half-electron on each site look too stable. This can cause a calculation to predict an electron cloud spread across both fragments when a localized charge description is more appropriate. A finite system's true state can itself be delocalized for physical reasons, so one must not label every distributed charge as an error; the diagnostic is whether the model's energy behavior is physically consistent under the stated separation and state conditions.

The same tendency can alter charge-transfer excitation energies, redox potentials, electron affinities and activation barriers. If a transition state spreads charge more than a reactant, an approximation that excessively stabilizes spread charge may lower the computed barrier. If a donor and acceptor are far apart, erroneous fractional transfer can appear even before a real electron-transfer event is energetically favorable. Spin-state and bond-breaking problems add related but distinct static-correlation issues; one correction does not automatically fix them all.

Range-separated hybrids, tuned functionals, explicit self-interaction corrections and other approaches can improve selected failures. Their performance depends on the system and the tuning or parameter choice. A range-separated functional may improve long-range charge transfer yet still have other errors. Comparison against ionization energies, electron affinities, localized-state benchmarks or higher-level wavefunction results can help diagnose a problem. Inspecting charges alone is not enough, because partial atomic charges depend on how the density is partitioned; energy differences and electron-density behavior provide complementary evidence.

Numerical factors can mimic or obscure this error. A too-small basis may artificially localize density, while a diffuse basis may reveal how readily the functional spreads it. An unconverged SCF solution may be trapped in one charge-localized state. Initial guesses and symmetry constraints can alter which self-consistent solution is found. A robust test compares multiple starting states and basis choices with the same geometry and functional, then evaluates whether the resulting physical conclusions survive.

Step-by-step reasoning

1. Identify whether the chemistry involves weakly bound electrons or charge redistribution across fragments. 2. Check the basis and SCF settings so spatial confinement is not a numerical artifact. 3. Compare localized and delocalized initial states when multiple solutions are plausible. 4. Examine density, charge distribution and the energy difference under consistent conventions. 5. Test a functional with a different self-interaction behavior or a suitable higher-level reference. 6. Report whether the predicted barrier, transfer energy or state ordering changes materially.

Visual explanation

Draw energy on the vertical axis and electron number from N to N+1 on the horizontal axis. The exact ensemble result is a straight segment; a typical delocalization-prone approximation bows below it. Next draw two widely separated boxes representing fragments, with an added electron localized in one box versus falsely smeared as half-charge across both. A caption should say fractional charge here is a mathematical ensemble diagnostic, not an observation of a fractional elementary particle.

Real-world analogy

Imagine a bookkeeping system that incorrectly charges a household for transferring money between two accounts it already owns, then gives an artificial discount when the money is split evenly. The discounted split could be chosen for accounting reasons rather than economic reality. Approximate DFT can similarly reward an unphysical distribution of charge. The analogy cannot reproduce the quantum-mechanical possibility of genuine delocalization or the formal definition of fractional electron ensembles.

Real-world example

A researcher computes electron transfer between distant donor and acceptor fragments. A semilocal functional yields fractional charge on both fragments even where available spectroscopy indicates integer-like localized states. An alternative range-separated functional and a wavefunction benchmark yield more localized density and a different transfer energy. The researcher checks that the outcome is stable against diffuse-basis changes and does not present the first fractional charge as a directly measured electron count.

Why?

Why does energy curvature favor charge spreading? If an approximate energy for fractional occupation lies below the straight line connecting integer endpoints, dividing one electron into fractional occupations on two equivalent distant sites can lower the approximate total energy. The exact piecewise-linear relation has no such artificial reward. This is an energy argument, not merely a visual preference for localized orbital plots.

Common misconception

“Any fractional atomic charge proves self-interaction error.” Atomic partial charges are model-dependent and real covalent bonds distribute density. The error concerns unphysical energy and density behavior under a defined comparison. Another mistake is saying hybrid DFT always removes it completely. A third is equating an SCF solution's numerical stability with the correct charge state. Finally, fractional electron number in the formal diagnostic is not a claim that an electron is physically divisible.

Worked example

Suppose a hypothetical exact ensemble energy between N and N+1 electrons is linear: at halfway, E exact(N+0.5) is the average of E(N) and E(N+1). An approximate functional instead gives a halfway energy 0.10 eV below that average. For two infinitely separated equivalent sites sharing one extra electron, placing half of the extra electron on each gains about 0.20 eV in the simple additive model compared with the two integer endpoints' average. That artificial stabilization can favor a smeared state. The values are invented; real fragment interactions and symmetry must be considered before diagnosing an actual calculation.

Quick check

1. Why is the classical Coulomb self-energy unphysical for a one-electron system? Answer: There is no second electron for that electron to repel, so an exact exchange–correlation treatment must cancel the self term. 2. What does a convex energy curve below the straight fractional-charge line tend to favor? Answer: It can spuriously stabilize fractional charge distributed across separated fragments.

Exam focus

Define one-electron self-interaction and explain how exact exchange–correlation cancels it in principle. State the piecewise-linear fractional-electron diagnostic with fixed external potential and distinguish it from a physically fractional electron. Relate convex curvature to delocalization and possible changes in barriers or charge transfer. Name basis, SCF solution and functional checks before attributing a result to one error mechanism.

Advanced insight

The energetic derivative can change discontinuously at an integer electron number in the exact theory. Approximate functionals that smooth or curve the energy may therefore produce misleading orbital gaps and charge-transfer behavior even if a total energy at an integer looks reasonable. Static-correlation error has a related formal diagnostic involving fractional spin but is not identical to fractional-charge delocalization. Separating these failures helps avoid using one broad “DFT error” label for chemically different problems.

Summary

Approximate density functionals may incompletely cancel an electron's unphysical self-repulsion. Their fractional-electron energy can bow below the exact piecewise-linear behavior, favoring excessive charge spreading. This can affect electron transfer, anions, barriers and other small energy differences. Diagnosis needs controlled basis and SCF checks plus independent benchmarks; a partial charge value alone is insufficient.

Practice questions

1. Is a formal N+0.5 electron ensemble a claim that an individual electron is physically half a particle? Answer: No. It is a mathematical ensemble used to examine energy behavior between integer electron numbers. 2. Does a larger diffuse basis necessarily remove delocalization error? Answer: No. It improves spatial representation but the functional's energy curvature may remain or become more visible. 3. Why can a charge-spread transition state have a barrier that is too low with a delocalization-prone functional? Answer: The approximation may over-stabilize the spread-charge transition state relative to more localized reactants. 4. What should accompany an atomic partial-charge analysis? Answer: Energy and density comparisons under consistent conditions, plus functional and basis sensitivity checks.