A unified approach for minimizing composite norms (Q2452377)

The problem under consideration consists in minimizing \(\mu _{1}\left\|\sigma (\mathcal{F}(X)-G)\right\| _{\alpha }+\mu _{2}\left\| \mathcal{C}(X)-d\right\| _{\beta }\) subject to \(\mathcal{A}(X)-b\in \mathcal{Q}\); here \(\mathcal{F}:\mathbb{R}^{m\times n}\rightarrow \mathbb{R}^{r_{1}\times r_{2}},\) \(\mathcal{C}:\mathbb{R}^{m\times n}\rightarrow \mathbb{R}^{p}\) and \(\mathcal{A}:\mathbb{R}^{m\times n}\rightarrow \mathbb{R}^{q}\) are linear maps, \(G,\) \(d\) and \(b\) are vectors with the appropriate dimensions, \(\sigma\) is used to denote singular values, \(\left\| \cdot \right\| _{\alpha }\) and \(\left\| \mathcal{\cdot }\right\| _{\beta }\) are suitable norms, \(\mu _{1}\) and \(\mu _{2}\) are nonnegative numbers, and \(\mathcal{Q}\) is a nonempty closed convex subset of \(\mathbb{R}^{q}\). To solve this problem, a first-order augmented Lagrangian algorithm is proposed, for which convergence results are proved and implementation details are discussed. Numerical experiments comparing the proposed algorithm with other methods are reported. The authors also explain how their algorithm can be extended to solve the more general problem in which the objective function contains a third term of the type \(H(X)\), with \(H:\mathbb{R}^{m\times n}\rightarrow \mathbb{R}\) being strongly convex.