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.