Møller-Plesset Perturbation Theory: Correcting for Electron Correlation

The problem MP theory was built to fix

The Hartree-Fock (HF) method solves the electronic Schrödinger equation by forcing each electron to move in the average field of all the others. That mean-field approximation is remarkably good — it typically recovers roughly 99% of the total electronic energy for many systems — but the missing ~1% is chemistry. Because HF lets an up-spin and a down-spin electron occupy the same spatial orbital while feeling only their smeared-out mutual repulsion, it systematically underestimates how well electrons avoid each other instant-to-instant. The energy difference between the exact non-relativistic solution (in a given basis) and the HF energy is defined as the electron correlation energy, E_corr = E_exact − E_HF, and it is always negative.

Correlation splits conceptually into two flavors. Dynamic correlation is the short-range, instantaneous Coulomb-hole avoidance — the cusp two electrons carve into the wavefunction as they approach — and it is what MP theory is designed to recover. Static (or nondynamic) correlation arises when a single Slater determinant is a qualitatively poor starting point, as in a stretched H₂ bond or a diradical, where two or more configurations contribute comparably. MP theory, being built on one reference determinant, handles dynamic correlation well and static correlation badly — a distinction that governs every success and failure of the method.

The numbers make the stakes concrete: for a molecule like water in a large basis, HF sits roughly 0.15 hartree (about 400 kJ·mol⁻¹) above the true energy. No thermochemistry, no barrier height, no dispersion-bound complex survives an error of that size, so the mean field must be corrected. Møller and Plesset's move was to treat that correction not variationally but perturbatively.

The Rayleigh-Schrödinger derivation, one term at a time

The engine is standard Rayleigh-Schrödinger perturbation theory: partition the full electronic Hamiltonian as Ĥ = Ĥ₀ + λV̂, expand the energy and wavefunction as power series in λ, and collect terms order by order. The genius — and the specific content of the 1934 paper — is the choice of Ĥ₀. Møller and Plesset took the zeroth-order Hamiltonian to be the sum of the one-electron Fock operators, Ĥ₀ = Σᵢ f̂(i). Its eigenfunctions are exactly the Slater determinants built from HF spin-orbitals, and its eigenvalues are sums of orbital energies εᵢ. The reference determinant Φ₀ is the zeroth-order state; its Ĥ₀-eigenvalue is E⁽⁰⁾ = Σ(occupied) εᵢ.

The perturbation V̂ = Ĥ − Ĥ₀ is the fluctuation potential: the difference between the true instantaneous electron-electron repulsion Σ 1/rᵢⱼ and the averaged HF potential already folded into the Fock operators. The first-order energy correction E⁽¹⁾ = ⟨Φ₀|V̂|Φ₀⟩ exactly cancels the double-counting of electron repulsion in E⁽⁰⁾, so that E⁽⁰⁾ + E⁽¹⁾ = E_HF. This is why 'MP1' is just Hartree-Fock and carries no new physics — the first genuine correction appears at second order.

The MP2 energy follows from the second-order sum-over-states formula. By Brillouin's theorem the singly excited determinants have zero matrix element with the HF reference, and triples and higher cannot couple to Φ₀ directly through the two-electron operator V̂ (by the Slater-Condon rules), so at first order only double excitations survive:

E⁽²⁾ = Σ(i<j, a<b) |⟨ij‖ab⟩|² / (εᵢ + εⱼ − εₐ − ε_b)

Here i, j run over occupied and a, b over virtual spin-orbitals, ⟨ij‖ab⟩ is an antisymmetrized two-electron integral, and the denominator is the sum of orbital-energy differences. Every term is negative (occupied energies lie below virtual ones), so MP2 always lowers the energy — it recovers correlation, never adds spurious repulsion. The physical picture is transparent: electrons are excited in pairs out of occupied orbitals into virtuals, letting them dodge one another.

Why size consistency makes MP the practical choice

The property that made MP theory a fixture of computational chemistry — and that distinguishes it sharply from the once-popular truncated configuration-interaction methods — is size consistency (and its close cousin, size extensivity). A method is size-consistent if the energy of two infinitely separated fragments A and B equals the sum of their independently computed energies: E(A···B) = E(A) + E(B). Truncated CI, such as CISD, fails this test badly; its correlation energy per particle shrinks as the system grows, so it cannot describe, say, a dissociating dimer or a crystal on the same footing as a monomer.

MP theory is size-consistent and size-extensive at every order, a consequence of the linked-cluster theorem of many-body perturbation theory. In diagrammatic language, only connected diagrams contribute to the energy at each order; the unlinked terms that spoil truncated CI cancel exactly. This is not an accident of MP2 — it holds for MP3, MP4, and beyond, which is why perturbation theory was the natural framework once chemists demanded methods that scale sensibly to real molecules and to the thermodynamic limit.

The price is that MP theory is not variational. Because the energy is a truncated power series rather than an expectation value of a normalized trial function, MP2 (and any finite order) can and does dip below the true energy. For well-behaved closed-shell molecules near equilibrium this is a feature, not a bug — MP2 overshoots correlation only slightly and delivers excellent geometries and vibrational frequencies. But the loss of the variational lower-bound guarantee is a real trade-off, and it becomes dangerous precisely where the perturbation series itself misbehaves.

A worked example: the correlation energy of neon and the noble-gas dimer

Consider the neon atom in a correlation-consistent aug-cc-pVQZ basis. Hartree-Fock gives an energy near −128.547 hartree; the estimated exact non-relativistic energy is about −128.94 hartree, so the correlation energy to be recovered is roughly −0.39 hartree (about −1020 kJ·mol⁻¹). An MP2 calculation captures on the order of −0.32 hartree of that — close to 80% — with the remaining fifth recovered slowly by higher orders and by coupled cluster. This 80% figure is a good rule of thumb across many closed-shell first-row systems: MP2 gets the bulk cheaply, and the last increments are expensive.

The example that made MP2 famous, though, is the noble-gas and van der Waals dimer. Hartree-Fock predicts that two neon atoms, or two methane molecules, feel essentially no attraction — the London dispersion force is a pure correlation effect, invisible to any mean-field theory. MP2 recovers dispersion for the first time at a modest cost: the leading −C₆/R⁶ attraction emerges directly from the double excitations, since instantaneous dipole-induced-dipole coupling is a correlated pair excitation. This is why MP2 became the standard first stab at stacked nucleobases, host-guest complexes, and physisorption energetics.

The cautionary flip side belongs in the same example: MP2 systematically overestimates dispersion, overbinding π-stacked aromatics such as the benzene dimer by several kJ·mol⁻¹ and grossly so for larger conjugated systems. The failure traces to the small HOMO-LUMO gaps of extended π systems, which shrink the MP2 energy denominators and inflate E⁽²⁾. The community response — spin-component-scaled MP2 (SCS-MP2), introduced by Stefan Grimme in 2003, which scales the same-spin and opposite-spin pair contributions by different empirical factors (roughly 1/3 and 6/5) — is now more accurate than plain MP2 at identical cost.

When the series diverges: the limits of MP theory

The comforting picture of a rapidly converging series MP2 → MP3 → MP4 → … → E_exact is often false. Higher orders are not reliably better. MP3 typically recovers only a modest further slice and frequently overcorrects; MP4(SDTQ) is a genuine improvement but jumps to O(N⁷) cost, erasing MP2's economy. Worse, careful studies — notably by Olsen, Christiansen, Koch and Jørgensen in the 1990s — showed that for some ordinary molecules (Ne, HF, the fluorine molecule) the Møller-Plesset series is asymptotically divergent: the terms eventually grow without bound. The divergence is traced to intruder states and to the diffuse-basis behavior of the perturbation, and it means one cannot simply push to higher order to reach the truth.

The dominant, everyday failure is static correlation. Whenever the HF reference is a poor single-determinant description — a stretched σ bond en route to homolysis, a transition-metal complex with near-degenerate d orbitals, a singlet diradical, an antiaromatic species — one or more MP2 denominators (εᵢ + εⱼ − εₐ − ε_b) approaches zero. The corresponding term explodes, and the correlation energy diverges toward −∞ instead of converging. Symptoms include:

• Bond dissociation: restricted MP2 gives a qualitatively wrong, non-parabolic potential energy curve as any covalent bond is broken.
• Small-gap systems: extended π systems, biradicals, and many open-shell transition-metal species where the HOMO-LUMO gap collapses.
• Spin contamination: unrestricted UMP2 built on a spin-contaminated UHF reference converges erratically and slowly.

The practical diagnostic is the T₁ diagnostic (from a companion coupled-cluster run) or simply the size of the HF-orbital gap and the largest MP2 amplitudes: when a handful of double-excitation amplitudes dominate, the single-reference assumption has failed and one must switch to a genuinely multireference method — CASSCF followed by CASPT2 or NEVPT2, the multireference generalizations of MP-type perturbation theory.

History, place in the hierarchy, and what MP2 is used for today

Møller and Plesset's 1934 paper was ahead of its instruments — with no computers, second-order corrections for many-electron molecules were unthinkable, and the method lay largely dormant for four decades. Its revival came in the 1970s when John Pople's group implemented analytic MP2 in the Gaussian program and married it to systematic basis sets; the closely related many-body perturbation theory (MBPT) of Kelly, Bartlett, and others supplied the diagrammatic underpinnings. Pople shared the 1998 Nobel Prize in Chemistry (with Walter Kohn) 'for his development of computational methods in quantum chemistry,' the ecosystem in which MP2 became a household name. The modern acceleration — resolution-of-the-identity (RI-MP2) and local and explicitly correlated F12 variants — has pushed routine MP2 to hundreds of atoms.

MP2 occupies a specific rung on the ab initio hierarchy that Pople sketched: HF → MP2 → MP3 → MP4 → CCSD → CCSD(T) → full CI, with cost and accuracy rising together. In practice most chemists skip MP3 and MP4 entirely and treat MP2 as the affordable correlated method for large systems and CCSD(T) as the benchmark for small ones. Against density-functional theory, MP2 is more principled about dispersion than a bare functional but is comparable in cost to hybrid DFT only for modest systems; the two are complementary rather than competing.

Where MP2 earns its keep today: geometry optimizations of medium organic molecules (its equilibrium bond lengths are typically within 1-2 pm of experiment); reaction and conformational energetics where DFT dispersion is suspect; noncovalent interactions in the SCS- and F12-corrected forms; and as the reference correlation source inside double-hybrid density functionals such as B2PLYP (Grimme, 2006), which literally add a scaled MP2 correlation term to a DFT calculation. Ninety years after two physicists fed Hartree-Fock into Rayleigh-Schrödinger theory, MP2 remains the first honest correction most computational chemists reach for.

MP2 versus the coupled-cluster benchmark CCSD(T): the two most-used correlated methods contrasted on the properties that decide which you run.

Property	MP2	CCSD(T)
Underlying theory	Rayleigh-Schrödinger perturbation, truncated at 2nd order	Exponential cluster ansatz, singles+doubles + perturbative triples
Formal cost scaling	O(N⁵)	O(N⁷) (the (T) step)
Excitations included	Doubles only (at 2nd order)	Singles, doubles (iterative) + connected triples (perturbative)
Variational?	No	No
Size-consistent?	Yes	Yes
Typical accuracy vs. exact	Good geometries; overbinds dispersion; ~5-15 kJ·mol⁻¹ errors	'Gold standard': ~4 kJ·mol⁻¹ (chemical accuracy)
Breaks down when	Reference is multireference / near-degenerate; small HOMO-LUMO gap	Strong static correlation (bond breaking, diradicals)