Here, we describe algorithms for minimizing nonsmooth functions and composite nonsmooth functions, which are the sum of a smooth function and a (usually elementary) nonsmooth function. We start with the subgradient descent method, whose search direction is the minimum-norm element of the subgradient. We then discuss the subgradient method, which steps along an arbitrary direction drawn from the subdifferential. Next, we describe proximal-gradient algorithms for nonsmooth composite optimization, which make use of the gradient of the smooth part of the function and the proximal operator associated with the nonsmooth part. Finally, we describe the proximal point method, a framework optimization that is valuable both as a fundamental method in its own right and as a building block for the augmented Lagrangian approach described in the next chapter.

First derivatives (gradients) are needed for most of the algorithms described in the book. Here, we describe how these gradients can be computed efficiently for functions that have the form of arising in deep learning. The reverse mode of automatic differentiation, often called “back-propagation” in the machine learning community, is described for several problems with nested-composite and progressive structure that arises in neural network training. We provide another perspective on these techniques, based on a constrained optimization formulation and optimality conditions for this formulation.

Here, we define subgradients and subdifferentials of nonsmooth functions. These are a generalization of the concept of gradients for smooth functions, that can be used as the basis of algorithms. We relate subgradients to directional derivatives and to the normal cones associated with convex sets. We introduce composite nonsmooth functions that arise in regularized optimization formulations of data analysis problems and describe optimality conditions for minimizers of these functions. Finally, we describe proximal operators and the Moreau envelope, objects associated with nonsmooth functions that are the basis of algorithms for nonsmooth optimization described in the next chapter.

In this section, we discuss fundamental methods, mostly based on gradient information, that yield descent, that is, the function value decreases at each iteration. We start with the most basic method, the steepest-descent method, analyzing its convergence under different convexity/nonconvexity assumptions on the objective function. We then discuss more general descent methods, based on descent directions other than the negative gradient, showing conditions on the search direction and the steplength that allow convergence results to be proved. We also discuss a method that also makes use of Hessian information, showing that it can find a point satisfying approximate second-order optimality conditions and finding an upper bound on the number of iterations required to do so. We then discuss mirror descent, a class of gradient methods based on more general distance metrics that are particularly useful in optimizing over the unit simplex – a problem that arises often in data science. We conclude by discussing the PL condition, a generalization of the strong convexity condition that allows linear convergence rates to be proved.