Electron Correlation and the Correlation Energy

Two electrons of opposite spin never actually crash into each other. The Hartree–Fock wavefunction does not know this. In The Hartree-Fock Energy and Fock Operator, every electron moves in the smeared-out average field of all the others. This mean-field fiction treats the instantaneous dance of electrons dodging one another as if it were a slow, averaged waltz. The energy lost by making this approximation may look small on a scale of hundreds of hartrees. Yet it is exactly the energy that decides whether a bond breaks homolytically. It decides whether a transition state is 2 kcal/mol too high. It decides whether a reaction barrier comes out on the wrong side of an Arrhenius plot. This gap between the true, non-relativistic ground-state energy and the best possible single-determinant (Hartree–Fock) energy is called the correlation energy. It is a deceptively simple definition — just a subtraction. But it hides a rich physical structure. There is a short-range "Coulomb hole" carved out of the electron density around every electron. There is a longer-range "left-right" correlation that becomes essential exactly when bonds stretch. This is precisely when close-shell pictures fail. There is also a whole hierarchy of post-Hartree–Fock methods built to recover it piece by piece. Understanding correlation energy is the doorway into everything that follows in electronic structure theory. Configuration Interaction, Moller-Plesset Perturbation Theory (MP2), and Coupled Cluster Theory are best understood as different strategies for buying back this missing energy. The size of the correlation energy is startlingly regular. For most closed-shell molecules near equilibrium geometry, it comes out to roughly 0.5–1.0 eV per electron pair. This is a few tenths of a percent of the total electronic energy. Yet it is routinely larger than the energy differences chemists actually care about: activation barriers, conformational preferences, binding energies, spin-state splittings. Recovering it systematically, rather than by fortuitous cancellation, is the central technical problem of modern quantum chemistry. This is why "correlated methods" occupy their own hierarchy of cost and accuracy above the mean-field starting point. This article builds the concept from the ground up. It covers how correlation energy is defined operationally, and why Hartree–Fock structurally cannot capture it. It explains the physical distinction between dynamic and static (non-dynamic) correlation. It describes the shape of the Coulomb hole in the pair density. Finally, it shows how these ideas anchor the size and character of corrections needed in real molecules — from stretched H₂ to biradicals to the electron gas.

Key takeaways

Correlation energy is defined as $E_{corr} = E_{exact} - E_{HF}$ , always negative, since HF is variational and exact energy is its lower bound.
Hartree–Fock replaces electron–electron repulsion with an averaged field; each electron feels only the mean positions of the others, missing instantaneous avoidance.
The Coulomb hole is the depletion in the electron pair density around $r_{12}=0$ caused by correlation, on top of the Fermi hole HF already gets right for same-spin electrons.
Dynamic correlation is the local, short-range avoidance dance recovered by adding many small-coefficient excited determinants (MP2, CI, CC).
Static (non-dynamic) correlation appears when one determinant simply cannot describe the state — stretched bonds, diradicals, near-degeneracies — requiring a handful of large-coefficient determinants even qualitatively.
Key idea: the correlation energy is the price of pretending each electron sees only an averaged cloud of the others instead of dodging them instant by instant, and its size and character (dynamic vs static) dictate which post-Hartree–Fock method is required.

Where this lives today

Correlation energy is not an academic bookkeeping exercise — it is the quantity that determines whether a computed number is chemically useful. Drug-discovery pipelines that rely on computed binding affinities must recover enough correlation energy (typically via DFT with dispersion corrections, or MP2/CCSD(T) benchmarks) to get relative binding energies right to within 1 kcal/mol, since HF alone routinely mis-ranks ligand poses by tens of kcal/mol due to missing dispersion — itself almost entirely a correlation effect (the van der Waals attraction between instantaneously correlated fluctuating dipoles is invisible to any mean-field method). Materials scientists modeling transition-metal catalysts or battery cathode materials must worry about static correlation from near-degenerate d-orbitals, routinely requiring multireference or DFT+U treatments rather than plain HF or single-reference correlated methods. In biochemistry, the weak, correlation-driven London dispersion forces that stabilize protein folds, DNA base stacking, and host–guest inclusion complexes are entirely a correlation-energy phenomenon, invisible to Hartree–Fock and only partially captured by standard density functionals without explicit dispersion corrections (Grimme's D3/D4 schemes, exchange-hole dipole moment methods). Spectroscopists computing accurate vibrational frequencies and dissociation energies for benchmarking (e.g., the W1/W2/HEAT thermochemistry protocols used to define reference enthalpies of formation) build up from HF baselines through CCSD(T) precisely to extrapolate the correlation energy to the complete-basis-set limit, since sub-kJ/mol accuracy requires recovering correlation almost exhaustively. And the entire modern hierarchy of post-Hartree–Fock quantum chemistry — Configuration Interaction, Moller-Plesset Perturbation Theory (MP2), and Coupled Cluster Theory — exists for the sole purpose of recovering this one quantity as efficiently, systematically, and accurately as computational resources allow.

Ask the tutor: How does Møller–Plesset perturbation theory formally partition the Hamiltonian to recover a first estimate of this correlation energy?

The physical picture — thorough prose

Picture two electrons in a helium atom. In the exact wavefunction, whenever electron 1 swings close to the nucleus on one side, electron 2 is statistically pushed to the other side — their motions are correlated, in the everyday sense that knowing where one is tells you something about where the other is not. The Hartree–Fock wavefunction, a single Slater determinant built from spin-orbitals each occupied by only one electron at a time, cannot encode this: electron 2's orbital is fixed once and for all as the eigenfunction of an averaged Fock operator. It never "notices" electron 1's instantaneous position — only the time-averaged charge cloud. Formally, this shows up in the electron pair density, $\rho_2(\mathbf{r}_1,\mathbf{r}_2)$ , the probability of finding one electron at $\mathbf{r}_1$ and another at $\mathbf{r}_2$ simultaneously. For any antisymmetric wavefunction — including the HF determinant — same-spin electrons already avoid each other because of the Pauli exclusion principle: exchange antisymmetry forces $\rho_2(\mathbf{r}_1,\mathbf{r}_2)\to 0$ as $\mathbf{r}_1\to\mathbf{r}_2$ for same-spin pairs. This built-in depletion is the Fermi hole, and Hartree–Fock gets it exactly right. What Hartree–Fock gets wrong is the analogous avoidance for opposite-spin electrons, which Coulomb repulsion alone should also enforce. In the true wavefunction there is an additional dip in $\rho_2(\mathbf{r}_1,\mathbf{r}_2)$ as $r_{12}\to 0$ for opposite-spin pairs too — the Coulomb hole — and it is precisely this missing depletion, and the associated lowering of the average $1/r_{12}$ repulsion, that constitutes correlation energy. There are two qualitatively different reasons a single determinant might fail, and distinguishing them is essential for choosing a method. Dynamic correlation is the everyday, local "instantaneous dodging": at every geometry, in every molecule, electrons swerve around each other over length scales of a fraction of an ångström, and this is recovered by mixing in a very large number of excited determinants each with a small coefficient. It is smooth, universal, and roughly transferable — it scales gently with the number of electron pairs. Static correlation (also called non-dynamic or strong correlation) arises instead when the true wavefunction is fundamentally multi-configurational — when two or more determinants are needed with comparable, large coefficients even in a qualitatively correct zero-order description. The textbook example is H₂ stretched to dissociation: at large internuclear distance the restricted HF determinant $|\sigma_g^2\rangle$ forces both electrons into the same bonding-orbital combination, artificially mixing in 50% ionic character (H⁺H⁻) even though the true dissociation limit is two neutral atoms. No amount of small perturbative corrections fixes this — you need the antibonding configuration $|\sigma_u^2\rangle$ mixed in with a coefficient that grows from 0 near equilibrium to comparable magnitude at dissociation. ```

Figure

Fermi hole (built into HF) vs. Coulomb hole (missing from HF) in the electron pair density along r₁₂

``` ```

Figure

Dynamic vs static correlation along the H₂ dissociation coordinate

```

The core ideas — the governing equations and relationships

The correlation energy is defined operationally, not as an a priori physical quantity but as the difference between two well-defined numbers:

E_{corr} \equiv E_{exact} - E_{HF}

where:

$E_{exact}$ is the exact non-relativistic ground-state electronic energy of the molecule in the given basis-set limit (formally, the full-CI energy in a complete basis),
$E_{HF}$ is the Hartree–Fock energy at the same nuclear geometry, obtained self-consistently as in The Self-Consistent-Field Procedure,
both are evaluated within the Born–Oppenheimer, non-relativistic, fixed-nuclei electronic Hamiltonian.

Reading: because Hartree–Fock is a variational method restricted to the space of single Slater determinants, and the exact wavefunction is not restricted this way, the variational principle guarantees

E_{HF} \ge E_{exact} \quad\Longrightarrow\quad E_{corr} \le 0

Correlation energy is always negative by this sign convention: Hartree–Fock always overestimates the true energy (makes it less negative / more positive), because the mean-field approximation cannot lower the energy as much as the true correlated motion can. Chemists sometimes quote $|E_{corr}|$ as a positive "correlation energy recovered," but the algebraic sign in the definition above is unambiguous. The physical origin of $E_{corr}$ can be seen by writing the exact electron-electron repulsion energy in terms of the pair density:

V_{ee} = \frac{1}{2}\int\int \frac{\rho_2(\mathbf{r}_1,\mathbf{r}_2)}{r_{12}}\, d\mathbf{r}_1\, d\mathbf{r}_2

where:

$\rho_2(\mathbf{r}_1,\mathbf{r}_2)$ is the two-particle density (probability density of simultaneously finding electrons at $\mathbf{r}_1$ and $\mathbf{r}_2}$ ),
$r_{12} = |\mathbf{r}_1 - \mathbf{r}_2|$ is the interelectronic distance,
the factor $\tfrac12$ avoids double-counting each pair.

Reading: Hartree–Fock computes this same integral using an uncorrelated (Hartree-product-like, but antisymmetrized) pair density, $\rho_2^{HF}$ , that factors approximately into single-particle densities modified only by exchange:

\rho_2^{HF}(\mathbf{r}_1,\mathbf{r}_2) = \rho(\mathbf{r}_1)\rho(\mathbf{r}_2) - |\rho_1(\mathbf{r}_1,\mathbf{r}_2)|^2

where the second (exchange) term is the Fermi-hole correction, non-zero only between same-spin electrons. Because this factorized form has no opposite-spin depletion at short $r_{12}$ , the HF repulsion energy $V_{ee}^{HF}$ is systematically too large — HF overestimates how close opposite-spin electrons get, hence overestimates their repulsion, hence overestimates the total energy. The correlation energy is essentially the correction to $V_{ee}$ from the missing Coulomb hole:

E_{corr} \approx \frac{1}{2}\int\int \frac{\rho_2^{exact}(\mathbf{r}_1,\mathbf{r}_2) - \rho_2^{HF}(\mathbf{r}_1,\mathbf{r}_2)}{r_{12}}\, d\mathbf{r}_1 d\mathbf{r}_2

(This is not the complete story — orbital relaxation and kinetic-energy contributions also shift slightly between HF and the correlated wavefunction — but it captures the dominant physical mechanism.)

Derivation AWhy a single Slater determinant cannot open a Coulomb hole

Goal. Show explicitly, for the simplest possible case (two electrons, two spatial orbitals), that the Hartree–Fock/single-determinant ansatz gives a factorized pair density with no opposite-spin correlation hole, while a two-configuration wavefunction does.

Step 1 — set up the minimal model. Consider a two-electron singlet system (e.g., a minimal-basis H₂ model) with two orthonormal spatial orbitals $\phi_g$ (bonding) and $\phi_u$ (antibonding). The restricted Hartree–Fock ground state places both electrons, spin-paired, in $\phi_g$ :

\Psi_{HF} = \phi_g(1)\phi_g(2)\,\frac{1}{\sqrt2}\big[\alpha(1)\beta(2)-\beta(1)\alpha(2)\big]

This is a single spatial configuration: both electrons occupy the same orbital, differing only in spin.

Step 2 — compute the HF pair density. Integrating out spin, the spatial pair density for this state is

\rho_2^{HF}(\mathbf{r}_1,\mathbf{r}_2) = 2\,\phi_g(\mathbf{r}_1)^2\,\phi_g(\mathbf{r}_2)^2

This factorizes exactly as $\rho(\mathbf{r}_1)\rho(\mathbf{r}_2)$ up to the normalization constant — the probability of electron 2 being at $\mathbf{r}_2$ does not depend at all on where electron 1 is. There is no mechanism whatsoever, at the level of this determinant, for the presence of electron 1 at some point to reduce the probability of electron 2 being nearby: $\rho_2^{HF}$ shows no dip as $\mathbf{r}_1\to\mathbf{r}_2$ . (Opposite-spin electrons here show zero Fermi hole too, since they carry different spin and the Pauli principle imposes antisymmetry only on identical-spin pairs — consistent with the earlier statement that HF gets the same-spin Fermi hole right but has no opposite-spin hole at all in this simplest example.)

Step 3 — introduce the minimal correlated wavefunction. Now allow the two-configuration ("two-determinant") mixture that constitutes the exact wavefunction in this two-orbital model space (this is the smallest possible configuration-interaction expansion, foreshadowing Configuration Interaction):

\Psi_{corr} = c_g\,\phi_g(1)\phi_g(2) + c_u\,\phi_u(1)\phi_u(2)

(spin singlet factor suppressed), with $c_g^2+c_u^2=1$ and $c_u$ small but non-zero near equilibrium, growing toward $|c_u|\to|c_g|$ at dissociation.

Step 4 — show the pair density now correlates the electrons. Expanding,

\rho_2^{corr}(\mathbf{r}_1,\mathbf{r}_2) \propto c_g^2\phi_g(\mathbf{r}_1)^2\phi_g(\mathbf{r}_2)^2 + c_u^2\phi_u(\mathbf{r}_1)^2\phi_u(\mathbf{r}_2)^2 + 2c_gc_u\,\phi_g(\mathbf{r}_1)\phi_u(\mathbf{r}_1)\phi_g(\mathbf{r}_2)\phi_u(\mathbf{r}_2)

This expression no longer factors into (function of $\mathbf{r}_1$ ) $\times$ (function of $\mathbf{r}_2$ ) for generic $c_u\ne0$ : the cross term couples the two coordinates. Because $\phi_u$ typically has a node where $\phi_g$ peaks (bonding vs. antibonding combinations of atomic orbitals), the cross term is negative when $\mathbf{r}_1\approx\mathbf{r}_2$ near the bond midpoint, actively suppressing the joint probability of finding both electrons at the same place — precisely the Coulomb hole. As $c_u$ grows (bond stretching), this suppression becomes large: physically, one electron localizes preferentially near one nucleus, the other near the other nucleus, exactly as required for correct dissociation into neutral atoms.

Step 5 — connect back to energy. Since $V_{ee} = \tfrac12\int\int \rho_2/r_{12}$ , and the correlated $\rho_2$ is smaller than the HF $\rho_2$ precisely where $1/r_{12}$ is largest (small $r_{12}$ ), $V_{ee}^{corr} < V_{ee}^{HF}$ . The total energy drops correspondingly; this drop, in the variational sense, is (a piece of) the correlation energy. Note also that at $c_u=0$ this reduces exactly to Step 2 — the correct limiting check that the correlated wavefunction collapses back to HF when no admixture is present.

Go deeper · Derivation B · 2 minSize-consistency and the additivity of dynamical correlation energy

Derivation BSize-consistency and the additivity of dynamical correlation energy

Goal. Show why dynamic correlation energy scales approximately with the number of electron pairs — a fact used constantly to estimate correlation energies of large molecules and to judge whether a method is "size-consistent."

Step 1 — partition the pair density into local contributions. Dynamic correlation is dominated by short-range, local Coulomb-hole physics: each electron pair correlates mainly with its immediate neighbors, not with electrons on the far side of a large molecule. Write the total correlation energy as an approximate sum over electron pairs $(i,j)$ :

E_{corr} \approx \sum_{i<j} \epsilon_{ij}

where $\epsilon_{ij}$ is the pairwise correlation-energy contribution, largest for electrons in the same bond or lone pair (small average $r_{12}$ ) and rapidly decaying for electrons on distant atoms.

Step 2 — consider two non-interacting fragments. Take a "supersystem" consisting of two molecules A and B held infinitely far apart, so they do not interact. The exact energy of the supersystem must equal the sum of the exact monomer energies:

E_{exact}(A\cdots B) = E_{exact}(A) + E_{exact}(B)

This is a basic physical requirement (no interaction ⇒ no coupling of energies) called size-consistency.

Step 3 — check Hartree–Fock is size-consistent (for the mean-field piece). Because the HF determinant for A···B is simply the antisymmetrized product of the separate HF determinants for A and B (when they don't interact), $E_{HF}(A\cdots B) = E_{HF}(A) + E_{HF}(B)$ exactly. HF itself is size-consistent.

Step 4 — require the same additivity of the correlation piece. Subtracting,

E_{corr}(A\cdots B) = E_{corr}(A) + E_{corr}(B)

must also hold if the whole theory is size-consistent. Since cross terms $\epsilon_{ij}$ between an electron on A and one on B vanish at infinite separation (no wavefunction overlap, no shared Coulomb hole), the pairwise sum from Step 1 automatically splits into a sum over A-pairs plus a sum over B-pairs — consistent with Step 4.

Step 5 — the practical consequence. This is why correlation energy for a molecule of $N$ roughly-similar electron pairs (e.g., a long alkane chain) grows roughly linearly with $N$ , at a fairly transferable rate per bond (historically quoted around $-0.02$ to $-0.03\, E_h$ per valence electron pair for many organic bonds at reasonable basis sets, though the exact per-pair value is basis- and method-dependent). It is also why truncated CI methods (like CISD) that are not size-consistent become increasingly poor as molecules grow, motivating size-consistent alternatives such as Coupled Cluster Theory and Moller-Plesset Perturbation Theory (MP2).

Worked examples

Example 1 — Foundational: extracting the correlation energy of a helium atom. The helium atom is the cleanest possible test case: two electrons, one nucleus, a numerically near-exact solution available from highly accurate variational calculations. Take the accepted essentially-exact non-relativistic ground-state energy of He (from Hylleraas-type or explicitly-correlated calculations):

E_{exact}(\text{He}) = -2.90372\ E_h

and the Hartree–Fock limit (infinite basis set) energy:

E_{HF}(\text{He}) = -2.86168\ E_h

Applying the definition directly:

E_{corr}(\text{He}) = E_{exact} - E_{HF} = (-2.90372) - (-2.86168)\ E_h

E_{corr}(\text{He}) = -0.04204\ E_h

Converting to more familiar units ( $1\,E_h = 27.2114$ eV $= 627.51$ kcal/mol):

E_{corr}(\text{He}) = -0.04204 \times 27.2114\ \text{eV} = -1.1441\ \text{eV} = -26.38\ \text{kcal/mol}

Two electrons, one pair, and yet the missing energy is over 26 kcal/mol — vastly larger than a typical rotational barrier or hydrogen-bond strength. This single number, $-0.042\ E_h$ , is the benchmark every correlated two-electron method (full CI in a large basis, explicitly correlated Hylleraas or R12/F12 methods) is checked against, and it establishes the scale: roughly $1$ – $1.1$ eV, or $0.02$ $E_h$ per electron in this tightly-bound system.

Limiting-case check: if helium had only one electron (He⁺, a hydrogenic ion), there would be no electron–electron repulsion at all, hence no correlation energy: $E_{corr}(\text{He}^+) = 0$ exactly, since a one-electron Hartree–Fock wavefunction is the exact wavefunction. This confirms the qualitative point that correlation energy is intrinsically a two-or-more-electron phenomenon, vanishing smoothly as electron number drops to one.

Example 2 — Applied: correlation energy and the bond dissociation of N₂. Nitrogen's triple bond is a workhorse test of correlation recovery because RHF badly mismodels bond breaking. Suppose, using a reasonably large basis set, a calculation gives (values representative of standard correlation-consistent basis results, chosen here as clean non-round numbers for the exercise): At the equilibrium bond length $R_e = 1.10$ Å:

E_{HF}(R_e) = -108.9927\ E_h, \qquad E_{FCI}(R_e) = -109.2649\ E_h

E_{corr}(R_e) = -109.2649 - (-108.9927) = -0.2722\ E_h

At a stretched geometry, $R = 2.20$ Å (roughly twice $R_e$ , deep in the dissociating region):

E_{HF}(2.20\,\text{Å}) = -108.7451\ E_h, \qquad E_{FCI}(2.20\,\text{Å}) = -109.1467\ E_h

E_{corr}(2.20\,\text{Å}) = -109.1467 - (-108.7451) = -0.4016\ E_h

The correlation energy has grown in magnitude from $-0.2722\ E_h$ to $-0.4016\ E_h$ — an increase of $0.1294\ E_h$ (about $81$ kcal/mol) simply from stretching the bond. This growth is the signature of static correlation switching on: at equilibrium, the $\sigma$ and $\pi$ bonding orbitals are well described by a single determinant, and the correlation is mostly dynamic (short-range Coulomb-hole physics, roughly constant per pair). As the bond stretches toward dissociation, the bonding and antibonding combinations of each broken bond become near-degenerate, and RHF is forced to keep 50% ionic (N⁺≡N⁻-like) character that the true wavefunction discards — hence the ballooning discrepancy. This is exactly why single-reference methods (HF, or even MP2/CCSD built on an RHF reference) systematically fail to describe bond-breaking curves correctly beyond a certain stretch, while methods admitting multiple reference configurations (multireference CI, CASSCF-based approaches) are needed in that region — the topic that Configuration Interaction generalizes properly.

Limiting-case check: at $R\to\infty$ (complete dissociation to two ground-state N atoms, each a well-behaved open-shell doublet-derived quartet system describable by a single UHF-like determinant), the correlation energy per fragment should relax back toward something closer to twice the atomic correlation energy of N — i.e., the "excess" correlation induced by the artificial RHF ionic mixing disappears once the molecule is allowed to separate into genuinely unpaired-spin atoms treated with an unrestricted reference. The huge intermediate bump in $|E_{corr}|$ is a diagnostic of static correlation, not a permanent feature of the dissociated limit.

Example 3 — Exam-level: partitioning correlation energy using Møller–Plesset perturbation theory, and checking additivity. A common exam question asks you to estimate how much of the correlation energy a low-order method like MP2 recovers, and to verify size-consistency numerically. Suppose for a diatomic-like model system with an RHF reference, MP2 recovers the following at second order (foreshadowing Moller-Plesset Perturbation Theory (MP2)):

E_{MP2}^{(2)} = -0.1980\ E_h

while the full correlation energy (from FCI in the same basis) is

E_{corr}^{FCI} = -0.2400\ E_h

The fraction of correlation energy recovered by MP2 is

f_{MP2} = \frac{E_{MP2}^{(2)}}{E_{corr}^{FCI}} = \frac{-0.1980}{-0.2400} = 0.825

so MP2 recovers 82.5% of the correlation energy in this basis — typical of MP2's known behavior of capturing roughly 80–90% of dynamic correlation at modest cost, while missing higher-order and any static-correlation contributions. Now suppose this system is actually two identical, well-separated (non-interacting) fragments X, each independently contributing $E_{corr}^{FCI}(\text{X}) = -0.1200\ E_h$ and $E_{MP2}^{(2)}(\text{X}) = -0.0990\ E_h$ . Because MP2 is size-consistent (unlike truncated CI), the composite system's numbers must be exactly additive:

E_{corr}^{FCI}(\text{X}\cdots\text{X}) = 2\times(-0.1200) = -0.2400\ E_h \quad\checkmark

E_{MP2}^{(2)}(\text{X}\cdots\text{X}) = 2\times(-0.0990) = -0.1980\ E_h \quad\checkmark

matching the composite values given above exactly — confirming both that size-consistency holds and that the fraction recovered, $f_{MP2}=0.825$ , is identical whether computed for one fragment or for the non-interacting pair:

f_{MP2}(\text{X}) = \frac{-0.0990}{-0.1200} = 0.825

This invariance of the recovered fraction under simple doubling is exactly what size-consistency guarantees, and it is a standard sanity check applied when benchmarking a new correlated method: compute it on a monomer and a well-separated dimer of that monomer, and confirm the energies (and derived correlation energies) simply double.

Limiting cases, scope & common traps

Regime	Behavior of $E_{corr}$	Dominant character
One-electron systems (H, He⁺, hydrogenic ions)	$E_{corr}=0$ exactly	none — no electron–electron repulsion to mismodel
Closed-shell molecule near equilibrium geometry	Modest, roughly $-0.02$ to $-0.04\ E_h$ per electron pair	almost entirely dynamic
Stretched/breaking single or multiple bonds	Grows sharply, RHF diverges qualitatively from FCI	static (non-dynamic) correlation dominates
Diradicals, transition metals with near-degenerate d-orbitals	Large, single-reference methods qualitatively wrong	static correlation, often multi-configurational ground state
Homogeneous electron gas / metals	Correlation energy per electron finite, computed via many-body theory (e.g. RPA)	dynamic, long-range screening effects included
Large basis-set limit	$\|E_{corr}\|$ increases with basis size (more virtual orbitals to correlate into)	dynamic correlation converges slowly, as $\sim 1/L^3$ in angular momentum $L$

The chemist behind it

The concept of electron correlation as understood today — the gap between the true many-electron wavefunction and any single-determinant mean-field approximation — crystallized through the combined efforts of quantum chemists working out the mathematical structure of the many-electron problem in the 1950s, most influentially Per-Olov Löwdin, who gave the term "correlation energy" its modern, precise operational definition.

Origins · 1 minPer-Olov Löwdin and how this was discovered

Per-Olov Löwdin

1916–2000 · Uppsala University and University of Florida, mid-20th-century quantum chemistry

Löwdin was a Swedish quantum chemist whose 1955 papers in Physical Review formalized the definition $E_{corr}=E_{exact}-E_{HF}$ that remains standard today, alongside foundational work on natural orbitals, the density matrix formalism, and orthogonalization procedures (Löwdin orthogonalization) used throughout electronic structure theory. He founded the Sanibel Symposia (later Quantum Theory Project meetings), which became — and remain — a central recurring forum where correlation methods from configuration interaction through coupled-cluster theory were developed, debated, and disseminated among theoretical chemists worldwide. His insistence on rigorous, basis-set-independent formal definitions gave the field a shared, precise vocabulary before the era of routine large-scale computation.

History

- 1929 — Dirac notes the many-electron Schrödinger equation is, in principle, exactly soluble but computationally intractable, framing the practical need for approximations.

- 1930 — Hartree and Fock (independently, then jointly refined) establish the self-consistent mean-field method that will later be shown to systematically miss correlation.

- 1955 — Löwdin formally defines "correlation energy" as the difference between the exact and Hartree–Fock energies, giving the field a precise, citable operational quantity.

- 1959 — Boys introduces Gaussian-type orbitals and early configuration-interaction machinery, enabling practical numerical estimates of

E_{corr}

for real molecules.

- 1980s–1990s — Systematic post-Hartree–Fock hierarchies (MPn, CI, coupled cluster) mature into routine tools for recovering dynamic correlation quantitatively, while multireference methods mature to handle static correlation.

Is correlation energy the same thing as "electron correlation"?

Not quite — "electron correlation" is the physical phenomenon (electrons avoiding each other beyond what a mean-field picture predicts), while "correlation energy" is the specific numerical quantity, $E_{exact}-E_{HF}$ , used to measure how much of that phenomenon is missing from a Hartree–Fock treatment. You can discuss electron correlation qualitatively (Fermi holes, Coulomb holes, static vs. dynamic character) without ever computing a number, but correlation energy is always a defined difference between two specific energies at a specific geometry and basis set.

Why doesn't Hartree–Fock just include electron correlation directly, since it does include electron repulsion?

Hartree–Fock includes the average Coulomb repulsion (via the Coulomb operator $J$ ) and exact exchange between same-spin electrons (via the operator $K$ ), both built from single-electron densities. But it represents the full wavefunction as a single Slater determinant, which mathematically forces the pair density to be (nearly) a product of one-electron densities. There is no term in the single-determinant ansatz that lets electron 2's instantaneous position depend on electron 1's instantaneous position beyond what exchange antisymmetry already provides — that flexibility requires mixing in additional determinants, which by definition takes you beyond Hartree–Fock.

What is the difference between dynamic and static correlation in practical terms?

Dynamic correlation is recovered by including many excited determinants each with a small coefficient — it is smooth, local, and roughly proportional to the number of electron pairs, well-suited to perturbative or coupled-cluster treatments built on a single good reference determinant. Static correlation appears when a handful of determinants have comparably large coefficients even in a qualitatively correct zero-order wavefunction — typically near bond-breaking, in diradicals, or in transition metals with near-degenerate orbitals — and requires multireference methods (CASSCF and its extensions) rather than single-reference perturbation theory or coupled cluster built from one determinant.

Does a bigger basis set always increase the magnitude of the correlation energy?

Yes, up to the complete-basis-set limit. A minimal basis set has too few virtual orbitals available to build the excited determinants needed to open the Coulomb hole, so calculated $|E_{corr}|$ grows steadily as the basis is enlarged (more diffuse and polarization functions, higher angular momentum shells), converging slowly — often described as converging like an inverse power of the highest angular momentum included. This slow convergence is why correlated calculations often extrapolate results from a sequence of basis sets (e.g., cc-pVDZ, cc-pVTZ, cc-pVQZ) toward an estimated complete-basis-set limit rather than trusting one finite-basis number directly.

Why is dispersion (van der Waals attraction) considered a correlation effect?

Dispersion arises from instantaneous, correlated fluctuations of electron density on two separate fragments: a fleeting dipole on one fragment induces a complementary fleeting dipole on the other, and their mutual attraction survives time-averaging even though each fragment's average dipole is zero. This is exactly the kind of instantaneous electron-electron correlation a single, time-averaged Hartree–Fock determinant cannot represent — HF dispersion energy between two neutral, non-overlapping fragments is essentially zero. Recovering dispersion requires a correlated method (MP2, coupled cluster) or an empirically corrected density functional, which is why plain HF and uncorrected local/GGA functionals notoriously fail for stacked aromatic rings, noble-gas dimers, and layered materials.

Can correlation energy ever be exactly zero for a many-electron system?

Only in the strict absence of electron–electron repulsion — practically, only for one-electron systems (H atom, He⁺, hydrogenic ions), or in artificial models where the interaction is switched off entirely. Any system with two or more interacting electrons has a strictly negative, non-zero correlation energy, because the exact ground state always lowers its energy below the best single-determinant description by allowing electrons to avoid each other instantaneously rather than only on time-average.

How does correlation energy relate to the choice between restricted and unrestricted Hartree–Fock?

Restricted Hartree–Fock (RHF) forces paired electrons into the same spatial orbital, which is efficient near equilibrium geometries but becomes increasingly poor as bonds stretch, artificially inflating the apparent correlation energy because RHF is fighting an unphysical ionic-character constraint. Unrestricted Hartree–Fock (UHF) allows different spatial orbitals for different spins and can partially mimic static correlation at dissociation (via spin polarization), often giving a qualitatively better dissociation curve — but at the cost of spin contamination (the UHF determinant is no longer a pure spin eigenstate), which introduces its own error not cleanly separable from "correlation energy" in the Löwdin sense.