Skip to content
LESSON NOTES · 06

6. Variational principle and finite bases

Position in the course: Lesson 6 of 10. Complete the preceding derivation and use the explained exercises to check understanding.

Prerequisites: Eigenstate expansions; integrals; harmonic oscillator

1. State the theorem’s conditions

Choose an admissible nonzero trial state in the energy domain of a Hamiltonian bounded below. Its Rayleigh quotient is an upper bound to the ground energy for the same Hamiltonian and boundary conditions. Lower energy improves the energy estimate, not necessarily every property. Excited-state bounds require orthogonality or a min–max construction.

2. Prove the bound by spectral expansion

For a complete discrete eigenbasis, each coefficient contributes a nonnegative probability weight. The quotient is a weighted average of eigenvalues, none below \(E_0\). A spectral-integral version extends the argument to continuous spectra.

\[ |\phi\rangle=\sum_n c_n|n\rangle,\quad E[\phi]=\frac{\sum_n|c_n|^2E_n}{\sum_n|c_n|^2}\geq E_0. \]

3. Vary a finite nonorthogonal basis

Expand in independent basis functions. Their overlap matrix is positive definite, and normalization uses it. Stationary variation yields a generalized eigenproblem. Symmetric orthogonalization turns this into a usual Hermitian matrix problem without changing energies.

\[ \begin{aligned}|\phi\rangle&=\sum_\mu c_\mu|\chi_\mu\rangle,\quad H_{\mu\nu}=\langle\chi_\mu|H|\chi_\nu\rangle,\quad S_{\mu\nu}=\langle\chi_\mu|\chi_\nu\rangle,\\\delta[c^\dagger Hc-E(c^\dagger Sc-1)]&=0\Rightarrow Hc=ESc,\\H^{\prime}&=S^{-1/2}HS^{-1/2},\quad H^{\prime}u=Eu,\quad c=S^{-1/2}u.\end{aligned} \]

4. Work the Gaussian example completely

For the oscillator choose normalized \(\psi_\alpha=(\alpha/\pi)^{1/4}e^{-\alpha x^2/2}\) with \(\alpha>0\). Evaluate position and momentum second moments; kinetic and potential terms compete. Differentiate the resulting energy and solve for its minimum. The exact energy is obtained because this trial family contains the exact ground state.

\[ \begin{aligned}\langle x^2\rangle&=\frac1{2\alpha},\quad\langle p^2\rangle=\frac{\hbar^2\alpha}{2},\\E(\alpha)&=\frac{\hbar^2\alpha}{4m}+\frac{m\omega^2}{4\alpha},\\\frac{dE}{d\alpha}&=\frac{\hbar^2}{4m}-\frac{m\omega^2}{4\alpha^2}=0\Rightarrow\alpha_*=m\omega/\hbar,\\E(\alpha_*)&=\hbar\omega/2.\end{aligned} \]

5. A dimensionless check

Set \(q=\alpha/\alpha_*\). The arithmetic–geometric mean inequality gives the same bound. Too narrow a Gaussian increases kinetic energy, while too broad a Gaussian increases potential energy. This analytic plot is a teaching example, not executed numerical optimization data.

\[ \frac{E}{\hbar\omega}=\frac{q+q^{-1}}4\geq\frac12. \]

6. Check yourself

Can enlarging a nested basis raise the lowest Ritz eigenvalue?

Answer and reasoning

No, for the same \(H\) and exact optimization, the old trial space remains available.

Does an arbitrary excited-looking trial state bound \(E_1\)?

Answer and reasoning

No; without orthogonality it can contain the ground state.

Variational principle and finite bases — schematic under the stated model assumptions

Teaching schematic drawn from the equations or algorithm steps in this lesson; not measured data.

7. Energy accuracy is not wavefunction accuracy

Let a normalized trial state be \(|\phi\rangle=\sqrt{1-\eta^2}|0\rangle+\eta|1\rangle\) for real \(\eta\). Its energy error is \(\eta^2(E_1-E_0)\), quadratic in the admixture. For an arbitrary observable \(A\), however, its expectation contains a linear term \(2\eta\sqrt{1-\eta^2}\operatorname{Re}A_{01}\). Thus a small variational energy error does not guarantee equally small property errors. If the spectral gap is tiny, even a substantial state admixture can have a tiny energy penalty.

For a gapped nondegenerate ground state with gap \(\Delta\), the spectral expansion gives \(E[\phi]-E_0\geq\Delta(1-|\langle0|\phi\rangle|^2)\). Rearranging bounds the missing ground-state weight by the energy error divided by the gap. This is useful only when the Hamiltonian and its true ground energy and gap are known or bounded; replacing them with uncontrolled approximate values is not a proof.

The Gaussian minimization also illustrates why admissibility matters. Negative \(\alpha\) makes its normalization integral diverge, so it is not a valid alternative with which to challenge the bound. A trial wavefunction must satisfy the same boundary conditions and necessary energy-domain conditions as the target problem. Optimizing parameters cannot repair a forbidden trial domain.

Additional exercise: For \(\eta=0.1\) and unit gap, the trial energy error is \(0.01\). If \(A_{01}=1\) and diagonal elements vanish, its expectation is approximately \(0.199\), showing a much larger first-order property change. Both values are exact consequences of this two-state construction. In practical electronic-structure work, converge the observable of interest rather than assuming that a smooth total-energy convergence curve establishes every density or response property.

8. References and further reading

Derivations and numerical examples here are original teaching constructions, not copied passages or reported research data.

Molecular methods · Quantum Monte Carlo · Gaussian · VASP


Previous lesson · Course overview · Next lesson