Fixed-sample optimization in quantum Monte Carlo using a probability density function

We consider parameter optimization for trial functions to be used for fixed-node diffusion Monte Carlo. By employing sample points selected from a positive definite distribution, parameters that determine the nodes of the trial function can be varied with no loss of stability and without sample bias. With CH as a test system, our optimized trial function gives a fixed-node energy lying below that of a MCSCF trial function with the same number of determinants and the same basis set. The present approach sheds light on the important question of how to improve the nodal structure and, thereby, the accuracy of diffusion Monte Carlo.