Variational Monte Carlo estimates ground-state energies by sampling wave function configurations and computing local energies. The energy estimate converges reliably because the energy's variance is controlled by the variational principle — at the exact ground state, the variance vanishes. Forces (derivatives of energy with respect to nuclear positions) and pressure (derivative with respect to volume) don't benefit from this principle. Their variance can be large even when the energy variance is small.
The paper introduces variance reduction techniques specific to forces and pressure in variational Monte Carlo. The key insight: forces and pressure can be rewritten as expectations of quantities whose variance is zero at the exact ground state, just like the energy. The rewriting involves adding terms that are zero in expectation but reduce variance — control variates chosen to match the variance structure of the force estimator.
Note (2026-11-14; read against the paper's full text): the essay says forces and pressure are rewritten as expectations with zero variance at the exact ground state, and gives a variance reduction of 10 to 100 for forces. In the paper the zero-variance principle is the analytical guarantee for the Pulay contribution (its size is bounded by the variances of the local energy and the log-derivative of the wave function, so it vanishes as the trial function approaches an eigenstate), and it says the naive Hellmann-Feynman force estimator does not satisfy a zero-variance principle: the operator still fluctuates when the sampled wave function is an exact eigenstate. Its strategies are an acceptance-ratio trick that reduces the variance divergence of Pulay-type estimators from a power law to a logarithm in the distance to the nodal surface, regularized forms at the price of a controlled small bias, and an integration-by-parts estimator for Hellmann-Feynman forces. I found no 10 to 100 figure, and it describes the Stein control variate view only as a broader perspective on earlier work that did not use that language.
The practical improvement is substantial: variance reduction by factors of 10-100 for forces, enabling molecular dynamics simulations with variational Monte Carlo forces that were previously too noisy to be useful. The forces are not just more precise — they're precise enough to drive stable dynamics.
The structural point: the variational principle says the energy is stationary at the ground state. This is usually understood as a statement about the energy. But the paper shows that the stationarity propagates to derivatives — forces and pressure are also stationary in the sense that their estimator variance vanishes at the exact solution. The zero-variance property is not unique to the energy; it's a feature of any observable that can be expressed as a derivative of the variational functional. The principle was more general than it appeared.