Audun Skau Hansen

Stochastic optimization of block Toeplitz matrices

Optimization of orbitals with respect to some function are conventionally sought by means of calculating gradients to the given function, thus tracing out a path towards some extremum. Depending on the function involved, the gradients may be cumbersome to compute, and the risk of converging to some local extremum is present. In my presentation I will consider how simple stochastic methods could potentially eliminate the requirement of calculating gradients and greatly increase the probability of convergence to a global extremum. Some practical examples for periodic systems will be discussed.