This methodology uses the Variational Method, parameterised by , in situations where the Energy , calculated as , has integrals that cannot be solved analytically, and as such requires Monte Carlo to be used to calculate the integral. By the variational principle , the ground state energy, for every , with equality if and only if is a ground state. Minimising over therefore gives an upper bound on and an approximate ground state.
Solution
We work in one dimension with units ; the same steps apply verbatim to for particles, which is where Monte Carlo is needed. The relevant equations are
Step 1: the energy as an expectation
Inserting into the energy,
so with samples we estimate , with statistical error , where is the number of effectively independent samples.
Zero-variance property: if is an exact eigenstate, is constant and the estimator has zero variance. The variance of thus measures the quality of the trial function and is itself a possible minimisation target.
Step 2: sampling with Metropolis-Hastings
The normalisation is unknown, but Markov Chain Monte Carlo only needs ratios of . We use the Metropolis–Hastings algorithm:
- From the current , propose with , a symmetric proposal.
- Accept with probability , otherwise keep .
- Discard an initial burn-in segment, and tune so that the acceptance rate is moderate (roughly 30% to 60%).
This chain satisfies detailed balance with respect to , so its samples are distributed as after equilibration. Successive samples are correlated, which is why the error uses , estimated in practice by blocking.
Step 3: the local energy of the trial function
Write with . Then , with and , so
which needs no numerical differentiation. Normalisability requires . Note that this trial function is a single Gaussian in disguise: completing the square, with , so only the width and the centre matter, and the four parameters are redundant. For and the potential term is minimised at for any width, so the optimum is centred. A genuinely richer ansatz is a sum, , which can describe the double-well case , .
Step 4: optimising the parameters
Let . For real and real symmetric ,
so the gradient is estimated from the same samples as the energy. For complex the result is . For the trial function above, , , and .
The algorithm is then: start from ; run the Markov chain of step 2; estimate and from steps 1 and 4; update (or with stochastic reconfiguration, which preconditions by the covariance matrix of the ); repeat until stops decreasing within its error bars. The final estimate should come from a long run at fixed .
Beyond VMC: projector methods
The result of VMC is limited by the trial function. Projector Monte Carlo methods remove this bias using imaginary-time evolution (exactly for bosonic ground states such as this one; for fermions up to the fixed-node approximation). Expanding in eigenstates of ,
so , once normalised, converges to the ground state as whenever . In position space, inserting resolutions of the identity,
where is the imaginary-time propagator (Green’s function). Splitting into many short steps, for which has a simple approximate form (a Gaussian diffusion kernel times ), turns into a high-dimensional integral sampled by random walks. This is diffusion Monte Carlo, usually run with the optimised VMC wave function as for importance sampling.