You can write the Schrödinger equation for the helium atom in a single line — and then you are stuck. Two electrons repelling each other while both orbit the nucleus is a three-body problem with no closed-form solution, and helium is the easy hard case: real molecules, solids, and everything quantum chemistry cares about are worse. Diagonalising the Hamiltonian exactly is simply off the table.

And yet chemists routinely quote ground-state energies of huge molecules to astonishing accuracy. How? By clever guessing. The variational method turns "guess the answer" into a rigorous, systematic technique: propose a trial wavefunction, compute the average energy it would have, and — this is the magic — know for certain that this number can never be smaller than the true ground-state energy. Every guess gives an honest upper bound, so a better guess is simply one that gives a lower number. It is the single most important approximation tool in quantum mechanics, and the engine underneath density-functional theory, Hartree–Fock, and most of computational physics.

The variational principle

Fix a system with Hamiltonian \hat H whose true (and unknown) energy levels are E_0 \le E_1 \le E_2 \le \dots, with E_0 the ground-state energy we are after. Take any normalisable state |\psi\rangle at all — a wild guess, it does not have to solve anything — and form its average energy, the Rayleigh quotient:

Read it slowly, because it is doing a lot of work. It says the expectation of energy over literally any guess sits at or above E_0 — never below. You cannot cheat your way under the true ground energy no matter how badly you guess. So the whole game becomes: make E[\psi] as small as you can, and you are squeezing down onto E_0 from above.

Why it is true — a two-line proof

The result looks almost too strong to be free, but the proof is short and worth doing once by hand. The energy eigenstates \{|n\rangle\} of \hat H form a complete orthonormal basis (Postulate 2), so expand the trial state in them:

Sandwich the Hamiltonian. Because the |n\rangle are orthonormal, all the cross terms vanish and we get a clean weighted sum of the energy levels:

Now the punchline. Every E_n \ge E_0 by definition of the ground state, and every weight |c_n|^2 \ge 0, so replacing each E_n by the smallest one only shrinks the sum:

Divide through by \langle\psi|\psi\rangle > 0 and out drops E[\psi] \ge E_0. The inequality is saturated only when all the weight sits on the ground level — that is, when the guess is the ground state. That is the entire principle: a weighted average of numbers can never be less than the smallest number in the list.

The method: guess, then minimise

A single guess gives a single bound; the trick is to guess a whole family at once. Choose a trial function that carries one or more adjustable variational parameters \alpha — a "knob" you can turn — write down the trial energy as a function of that knob,

and then minimise over the parameter. Since every value of \alpha gives a legitimate upper bound, the smallest one is the best (tightest, lowest) bound the family can offer. So you solve

for the optimal \alpha^\star, and report E(\alpha^\star) as your estimate of E_0. Enrich the family with more parameters and the bound can only get better — this is exactly how industrial electronic-structure codes work, optimising thousands of parameters to press the estimate down toward the truth. The graph below is this minimisation made visible.

Watching the bound tighten

Here is the variational search for the harmonic oscillator's ground state, using the Gaussian family we work out below. In tidy units the trial energy takes the shape E(\alpha) = A\,\alpha + B/\alpha — a kinetic piece that grows with \alpha and a potential piece that shrinks with it. Their tug-of-war makes a valley. Drag the \alpha slider and watch the trial energy (the coloured dot on the curve) slide down toward the dashed floor at E_0 — it kisses the floor at the bottom of the valley and, no matter how hard you push, never dips below it. That dashed line is the true ground energy the principle guarantees you cannot beat.

The bottom of the valley sits where dE/d\alpha = A - B/\alpha^2 = 0, i.e. \alpha^\star = \sqrt{B/A}, and the minimum value there works out to the neat E(\alpha^\star) = 2\sqrt{AB}. Keep that little formula in your pocket — the quiz uses it, and it is the answer to a surprising number of one-parameter variational problems.

Worked example 1 — a Gaussian nails the oscillator

Take the one-dimensional harmonic oscillator, \hat H = -\dfrac{\hbar^2}{2m}\dfrac{d^2}{dx^2} + \tfrac12 m\omega^2 x^2, and guess a normalised Gaussian trial function with width parameter \alpha > 0:

Computing the two expectation values (standard Gaussian integrals) gives a kinetic average that grows with \alpha and a potential average that falls with it:

This is exactly the A\alpha + B/\alpha shape from the graph, with A = \hbar^2/2m and B = m\omega^2/8. Minimise:

Substitute back and the two terms become equal at

That is not just a good bound — it is the exact ground-state energy, and \alpha^\star reproduces the exact ground-state wavefunction. Why so perfect? Because the true ground state of the oscillator genuinely is a Gaussian, so it lives inside our trial family, and by the "equality iff" clause the method finds it dead on. When your family contains the answer, the variational method returns the answer.

Worked example 2 — a guess that misses (but only just)

The oscillator is a flattering case. To see the method in its honest role — an estimate, not an oracle — try a trial function that is the wrong shape. Put a particle in the infinite square well of width L (walls at x = 0 and x = L) and, instead of the true sinusoidal ground state, guess a simple tent — a triangle rising to a peak at the centre and back down to the walls:

The tent has the right qualitative features (it vanishes at both walls, it is humped in the middle) but the wrong curvature — it has a sharp kink at the peak where the true ground state is smoothly rounded. Evaluating the Rayleigh quotient for this shape gives

The tent overshoots the true ground energy by a mere 1.3% — above it, exactly as the principle promises, but strikingly close for such a crude guess. That is the everyday experience of the method: even a rough trial function, as long as it has the right gross features, lands you within a few percent of the truth, always on the high side. Round off the kink (add a parameter that smooths the peak) and the bound tightens further toward E_0.

What it can and cannot promise

The power of the variational method is its guarantee, but that guarantee is narrow, and knowing its edges keeps you honest:

• Only an upper bound. You learn that E_0 \le E(\alpha^\star), and nothing rigorous about how far below. A loose guess gives a loose (but still valid) bound; you never get a lower bound for free.
• Only as good as the family. The estimate can never be better than the best member of your trial family. If the true ground state is nowhere near anything your parameters can build, even the optimised bound stays well above E_0. Choosing a physically sensible trial shape is the real art.
• Excited states need care. Naively the principle only bounds the ground state. But if you restrict the trial state to be orthogonal to the (known or approximated) ground state, the same argument bounds the first excited energy E_1 from above — and so on up the ladder, orthogonalising as you climb.
• It never lies low. Whatever you do, no guess and no optimisation can push the estimate below E_0. That one-sidedness is the feature: every variational number is a certified ceiling on the true ground energy.