Accelerated optimization for composite regularizers

This optimization method minimizes functions of the form f(x) + h(Bx) where

• f is a strongly smooth function
• h is a nonsmooth function whose proximity operator is easy to compute
• B is a linear map.

An example of such optimization problems is regularization with penalties such as the composition of the L₁ norm or Group Lasso with a linear map.

The algorithm combines a fixed-point method with Nesterov acceleration. See Efficient First Order Methods for Linear Composite Regularizers.