7. Beyond one determinant: MP2, CI, and CC
Position in the course: Lesson 7 of 10. Complete the preceding derivation and use the explained exercises to check understanding.
Prerequisites: All molecular lessons; perturbation expansions
Prerequisite lessons: Hartree–Fock from constrained variation · Roothaan equations and a reproducible SCF cycle
1. Define correlation relative to the same problem
At fixed geometry, Hamiltonian and finite basis, correlation recovery relative to optimized HF is \(E_{FCI}-E_{HF}\leq0\). Only the complete-basis limit approaches exact nonrelativistic correlation for that Hamiltonian. MP2 is perturbative, CI variational in a determinant subspace, and standard CC projected with an exponential ansatz. These are not interchangeable accuracy labels.
2. Derive second-order perturbation
Use electronic \(H\) only here, adding \(V_{NN}\) at the end. Expand the state and energy in \(\lambda\), project the first-order equation onto an excited determinant, then project back onto the reference to obtain \(E^{(2)}\). Canonical stationary HF eliminates singles (Brillouin’s theorem); the two-body Hamiltonian connects directly to at most doubles. Ordered spin-orbital sums count each double four times, explaining the factor \(1/4\).
3. CI: linear variation and size consistency
Vary a linear combination of orthonormal determinants. FCI includes all determinants in the chosen one-particle basis, while CISD retains reference, singles and doubles. For two separated fragments, a double excitation on each gives a global disconnected quadruple. Global CISD omits it, so fragment correlation energies need not add. Truncated CI remains variational but is generally not size-consistent.
4. CC: exponential ansatz and projection
CCSD retains \(T_1+T_2\), yet the exponential generates higher disconnected excitations such as \(T_2^2/2\). For separated fragments with commuting cluster operators and factorizing references/equations, the exponential factorizes and explains size extensivity. Standard CC is not variational; strong multireference character can cause failure. CCSD(T) has perturbative triples, not full CCSDT.
Teaching schematic drawn from the equations or algorithm steps in this lesson; not measured data.
5. Worked small-gap breakdown
Use a dimensionless two-determinant teaching Hamiltonian, not molecular MP2 results. For \(\Delta=1,v=0.1\), exact \(E_-=-0.00990195\) while second order is \(-0.01\). For \(\Delta=0.02,v=0.1\), exact \(E_-=-0.0904988\) but second order gives \(-0.5\), badly wrong. At \(\Delta\to0\) the exact solution stays finite at \(-|v|\) while the perturbation diverges. Lower MP2 energy is not proof of reliability.
6. Check yourself
Is finite-basis FCI exact for nature?
Answer and reasoning
No, basis and Hamiltonian approximations remain.
Does CCSD contain only doubly excited determinants?
Answer and reasoning
No, exponentiation generates higher disconnected excitations.
Why can MP2 lie below exact energy without violating the variational theorem?
Answer and reasoning
Its truncated perturbation energy is not a trial-state Rayleigh quotient.
7. Choose a method by the failure it must repair
Correlation is defined relative to a chosen HF limit for the same Hamiltonian; a finite-basis difference also contains representation effects. Single-reference methods work best when one determinant carries most of the relevant state character. Near-degenerate configurations require attention to the reference before increasing a truncation rank. A very negative correction can be an instability warning rather than a sign of high accuracy.
Method comparison needs a balanced question. For a reaction energy, compare reactants and products with consistent basis quality, frozen-core treatment, reference assumptions and numerical thresholds. For a dissociation curve, examine the separated-fragment limit and intermediate geometry, not only the equilibrium minimum. For an excited state, identify whether the state is dominated by a single or multiple excitation and whether the selected method represents that character.
An error budget should separately list finite-basis error, excitation truncation, reference inadequacy, numerical residuals and Hamiltonian approximations. Basis extrapolation cannot repair a missing active orbital; a larger active space cannot by itself supply all external dynamic correlation; a tighter residual cannot make the ansatz physically adequate. These distinctions make convergence experiments interpretable.
Additional exercise: Two methods agree within a small tolerance in one tiny basis. Does that demonstrate a converged prediction? No. They may share a large basis error or the same reference deficiency. Add an independently informative comparison: enlarge the basis, test a different relevant reference space or inspect a known exact limit. Agreement is useful evidence only when its possible shared errors are understood. The following lessons separate MP2, CI and CC so that their formulas, costs and domain of validity can be examined independently.
8. References and further reading
- MSU: Møller–Plesset perturbation
- Florida State University: CI tutorial
- Psi4 official coupled-cluster theory documentation
Derivations and numerical examples here are original teaching constructions, not copied passages or reported research data.
9. Related theory and practice
Quantum mechanics · Density functional theory · Quantum Monte Carlo · Gaussian