3. Algorithm representation
We consider first-order algorithms that solve (2.1) that can be represented as a discrete-time linear time‐varying system in state-space form in feedback interconnection with the potentially nonlinear and set-valued operators \((\partial f_{i})_{i\in\IndexFunc}\) and \((G_{i})_{i\in\IndexOp}\) that define the problem.
Before presenting the algorithm representation, we introduce some notation:
\(\NumFunc = \abs{\IndexFunc}\) denotes the number of functional components in (2.1);
\(\NumOp = \abs{\IndexOp}\) denotes the number of operator components in (2.1);
\(\NumEval_{i}\in\mathbb{N}\) denotes the number of evaluations of \(\partial f_{i}\) and \(G_{i}\) per iteration, for \(i\in\IndexFunc\) and \(i\in\IndexOp\), respectively;
\(\NumEvalFunc = \sum_{i\in\IndexFunc} \NumEval_{i}\) denotes the total number of subdifferential evaluations per iteration;
\(\NumEvalOp = \sum_{i\in\IndexOp} \NumEval_{i}\) denotes the total number of operator evaluations per iteration; and
\(\NumEval = \NumEvalFunc + \NumEvalOp\) denotes the combined total number of evaluations per iteration.
Since we consider algorithms that allow for multiple evaluation of \((\partial f_{i})_{i\in\IndexFunc}\) and \((G_{i})_{i\in\IndexOp}\) per iteration, we define \(\bfcn_{i}:\calH^{\NumEval_{i}}\to\p{\reals\cup\{\pm\infty\}}^{\NumEval_{i}}\) such that
\(\bm{\partial}\bfcn_{i}:\calH^{\NumEval_{i}}\rightrightarrows\calH^{\NumEval_{i}}\) such that
and \(\bm{G}_{i}: \calH^{\NumEval_{i}} \rightrightarrows \calH^{\NumEval_{i}}\) such that
We are now ready to give the algorithm representation: Pick an initial \(\bx_{0}\in\calH^{n}\), an iteration horizon \(K\in\naturals\cup\set{\infty}\), and let
where
are the algorithm variables, and
are matrices containing the parameters of the algorithm at hand.
We close this page with Assumption 3.1 (Well-posedness). It is satisfied by practical algorithms and ensures that the subsequent theoretical results are well-posed.
Assumption 3.1 (Well-posedness).
We assume that the parameter sequence
is chosen such that, for each
there exists a sequence
satisfying (3.1).
In this context, the symbol \(\Pi\) denotes Cartesian products. For interface documentation in AutoLyap, see Algorithms. In particular:
Base algorithms: abstract interfaces for users to implement their own algorithms; see Base algorithms. For an example, see The proximal gradient method.
Concrete algorithms: built-in algorithm implementations; see Concrete algorithms for the full list.