Variational Monte Carlo

Content sourced from Wikipedia, licensed under CC BY-SA 3.0.

Variational Monte Carlo (VMC) is a way to find the ground state of a quantum system using the variational principle and Monte Carlo sampling. You start with a trial wavefunction that depends on some parameters, and you adjust those parameters to make the energy as small as possible.

How it works
- The energy you estimate is the average energy of the system when it is described by the trial wavefunction. To compute it, you sample many configurations of the particles according to the probability given by the square of the wavefunction, and you average the local energy for each configuration. The local energy is how much energy the Hamiltonian assigns to that configuration divided by the wavefunction value there.
- By repeating this sampling and averaging, you get an estimate of the energy for a given set of parameters. You then change the parameters to minimize this energy, aiming for the best possible approximation to the true ground state.

Why Monte Carlo?
- The number of possible configurations grows very quickly with the number of particles, making exact, high-dimensional integrals impractical. Monte Carlo sampling turns these huge integrals into averages over a manageable number of samples, giving a scalable way to compute energies in many-body systems.

Choosing a trial wavefunction
- A simple mean-field (product) form is easy but misses important particle correlations.
- A powerful improvement is the Jastrow factor, which explicitly includes pair correlations. The wavefunction becomes a base form multiplied by an exponential of a sum over pair distances, capturing how particles influence each other.
- With Jastrow-type forms, you can achieve a large fraction of the true correlation energy using far fewer parameters than some other methods. For example, in chemistry this approach can reach about 80–90% of the correlation energy with a few dozen parameters, while traditional methods might need tens of thousands.

How the method scales
- The accuracy mainly depends on how good your trial wavefunction is. The computational cost for evaluating energies grows with the system size, but the scaling is typically a modest power of the number of particles, depending on how the wavefunction is written.

Optimizing the wavefunction
- The optimization step is critical and can be affected by statistical noise in the energy and its derivatives. People optimize different objective functions, commonly the energy, the variance of the local energy, or a combination of both.
- Variance minimization has a nice feature: if you could reach the exact ground state, the variance would be zero, so the minimum is well-defined. However, many researchers find energy minimization to be more effective in practice because it often gives better results for other observables and avoids getting stuck in local minima.

Optimization strategies
- Correlated sampling with deterministic optimization: accurate for small systems, but can struggle if changing parameters changes the structure of the wavefunction (such as its nodes) or when the needed density ratios grow with system size.
- Deterministic methods with large data bins: evaluate the cost function and its derivatives using large samples to reduce noise, then optimize.
- Noisy, iterative methods: use stochastic gradient approaches or stochastic reconfiguration to handle fluctuations directly while updating parameters.

Neural networks as wavefunctions
- In 2017, researchers showed that a neural network can be trained, using the VMC objective, to represent the ground state. This idea, called neural network quantum states, has broadened VMC to include very flexible wavefunctions, including those for fermions, improving accuracy for electronic structure problems compared with traditional VMC alone.

In short, Variational Monte Carlo uses a parameterized trial wavefunction to minimize the system’s energy via Monte Carlo sampling. Its success hinges on a good choice of wavefunction (with tools like the Jastrow factor to capture correlations), careful optimization in the presence of statistical noise, and, increasingly, the use of neural networks to represent complex quantum states.