46
Part One · Quantum mechanics

Variational method

Guess any wavefunction at all, compute its average energy, and you have an upper bound on the true ground-state energy. A bad guess gives a high number; a better guess gives a lower one; nothing you guess can ever go below the truth. This one fact powers most of quantum chemistry, and it is the idea behind the algorithm that quantum computers are expected to use first on molecules.

On the plate: the energy of a family of guesses, and the true ground state beneath it, which no guess can cross.

Why every guess is too high

Expand any normalized trial state in the true energy eigenstates, ∣ψ⟩=∑ncn∣n⟩|\psi\rangle = \sum_n c_n|n\rangle (Chapter 9). Its average energy is

⟨H⟩=∑n∣cn∣2En  ≥  ∑n∣cn∣2E0=E0\langle H\rangle = \sum_n |c_n|^2E_n \;\ge\; \sum_n|c_n|^2E_0 = E_0

because every En≥E0E_n \ge E_0 and the weights ∣cn∣2|c_n|^2 add to one. Equality holds only if the guess is the ground state itself. The plate shows a guess as a set of weights on the levels; its energy is their weighted average, which can only sit above the bottom. Established

Hydrogen with the wrong shape

Pretend you don’t know hydrogen’s answer and guess a Gaussian, ψ∝e−αr2\psi \propto e^{-\alpha r^2}, which has the wrong shape both at the nucleus and far away. Its energy, in hartrees, is E(α)=32α−22α/πE(\alpha) = \tfrac32\alpha - 2\sqrt{2\alpha/\pi}. Vary 0.60 and find the bottom.

The best Gaussian, at α=8/9π\alpha = 8/9\pi, gives −4/3π=−0.424-4/3\pi = -0.424 hartree against the exact −0.5-0.5: 85% of the binding energy from a function that is wrong everywhere. Energies are forgiving; an error of size ϵ\epsilon in the wavefunction gives an error of size ϵ2\epsilon^2 in the energy.

Helium

Helium has two electrons that repel each other, and its Schrödinger equation has no exact solution. Try two hydrogen-like 1s orbitals with an adjustable nuclear charge ZeffZ_{\text{eff}}, to let each electron partly hide the nucleus from the other. The energy is E=Zeff2−278ZeffE = Z_{\text{eff}}^2 - \tfrac{27}{8}Z_{\text{eff}} hartree. Set Zeff=Z_{\text{eff}} = 2.00.

At Zeff=2Z_{\text{eff}} = 2, the true charge, you get the first-order perturbation result, −74.8 eV. The minimum, at Zeff=27/16≈1.69Z_{\text{eff}} = 27/16 \approx 1.69, gives −77.5 eV. The measured value is −79.0 eV. Egil Hylleraas, in 1929, added the distance between the electrons to the trial function and got within a fraction of a percent; modern calculations of this kind agree with experiment to more than ten digits.

Let the computer choose

Instead of one adjustable number, take a combination of fixed functions, ψ=∑iciχi\psi = \sum_i c_i\chi_i, and choose all the cic_i at once. Minimizing the energy turns into a matrix eigenvalue problem, Hc=E ScH\mathbf c = E\,S\mathbf c, where SS holds the overlaps of the basis functions. This is the Rayleigh–Ritz method, and it is how almost all quantum chemistry software works.

Use 3 Gaussians of different widths for hydrogen. Each added function can only lower the energy, and by eight the error is under 0.04%. The plate shows the energy falling toward the exact value, the dashed line.

The simplest chemical bond

H₂⁺ is one electron and two protons. Guess a combination of a 1s orbital on each proton, the linear combination of atomic orbitals. The sum piles electron density between the nuclei and binds; the difference has a node there and does not. Move the protons to 2.5 Bohr radii.

The bonding curve has a minimum near 2.5 a0a_0, with a binding energy of 1.76 eV. The exact values are 2.0 a0a_0 and 2.79 eV. The crude guess already explains why a bond exists; better guesses make it quantitative. This is the molecular-orbital picture used for every molecule you will meet in Part Four.

The same idea on a quantum computer

For large molecules, the trial wavefunction itself becomes too big to store on any ordinary computer. The variational quantum eigensolver prepares the trial state on a quantum processor, controlled by adjustable settings, measures its energy, and lets an ordinary computer adjust the settings downhill. Alberto Peruzzo and colleagues ran it first in 2014 on the molecule HeH⁺.

Whether it will beat classical methods for molecules that matter, drug targets and enzymes included, is unsettled: noise in today’s hardware and the difficulty of optimizing many settings are serious obstacles. Part Three examines the evidence. Debated

What you now have

The variational bound and its proof, the Gaussian guess for hydrogen, helium with an effective charge, the Rayleigh–Ritz method, the LCAO picture of H₂⁺, and VQE.

· · ·

Three questions to carry with you.

Why can’t a trial energy fall below E₀?tap
It is a weighted average of eigenvalues, each at least E₀.
Why is Z_eff for helium less than 2?tap
Each electron partly screens the nucleus from the other.
Why does the bonding orbital bind?tap
It concentrates electron density between the nuclei, where it is attracted by both.

And one to work out. With Zeff=27/16Z_{\text{eff}} = 27/16, helium’s variational energy is −(27/16)2-(27/16)^2 hartree. In eV (1 hartree = 27.211 eV)?

E =eV

Sources: W. Ritz, J. reine angew. Math. 135, 1 (1909). E. A. Hylleraas, Z. Phys. 54, 347 (1929). A. Szabo and N. S. Ostlund, Modern Quantum Chemistry. A. Peruzzo et al., Nat. Commun. 5, 4213 (2014). Checked by code: Gaussian hydrogen −0.42441 Ha at α = 0.2829; helium −77.49 eV (variational), −74.83 eV (first order), −79.01 eV (exact); Ritz with 1–8 Gaussians −0.231 … −0.49985 Ha; LCAO H₂⁺ R = 2.49 a₀, D = 1.76 eV.

Next · Chapter 47
WKB approximation

The old Bohr rules weren't wrong. They were the first term of a series in ħ.

Turn the page