Commutation principles for optimization problems involving strictly Schur-convex functions in Euclidean Jordan algebras

Pedro G. Massey, Noelia B. Rios *, David Sossa · 2025

Read Commutation principles for optimization problems involving strictly Schur-convex functions in Euclidean Jordan algebras by Pedro G. Massey, Noelia B. Rios *, David Sossa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-19780.txt.

Introduction

Pedro G Massey and co-authors prove commutation principles for optimization problems in Euclidean Jordan algebras, published in 2025. The problems involve spectral functions that are strictly Schur-convex, a strong form of symmetry and convexity. The question is when an optimizer must commute, in the operator sense, with a shift term in the problem.

Under set conditions, the authors show that a local optimizer and the shift do indeed commute. They make no extra smoothness assumption on the function. Instead they use the smooth structure of the algebra's symmetry group and tools from majorization of eigenvalues. The principles let them treat several known problems at once, including one about minimizing the condition number in this setting.

Catalog record

CREATORPedro G. Massey, Noelia B. Rios *, David Sossa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-19780

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Commutation principles for optimization problems involving strictly Schur-convex functions in Euclidean Jordan algebras Pedro G. Massey, Noelia B. Rios *, David Sossa Abstract Abstract . In this work we establish several commutation principles for optimizers of shifts of spectral functions in the context of Euclidean Jordan Algebras (EJAs). For instance, we show that under certain assumptions, if \bar{x} is a (local) optimizer of F(x-a) for x\in\Omega , where \Omega\subset\mathcal{V} is a spectral set of an EJA \mathcal{V} , a\in\mathcal{V} and F:\mathcal{V}\rightarrow\mathbb{R} is a strictly Schur-convex spectral function, then a and \bar{x} operator commute. We make no further assumption on the smoothness of F ; instead, we take advantage of the smoothness (Lie structure) of the Automorphism group of \mathcal{V} and make use of majorization techniques for the eigenvalues of elements in EJAs. Our approach allows us to deal with several problems considered in the literature, related to strictly convex spectral functions and strictly convex spectral norms. In particular, we use our commutation principles to analyze the problem of minimizing the condition number in EJAs. Mathematics Subject Classification : 17C20, 15A27, 15A42, 90C26 Keywords : Euclidean Jordan algebra, commutation principle, Lidskii’s inequality, strictly Schur-convex function, condition number. 1 Introduction Commutation principles in Euclidean Jordan algebras (EJAs) consist of a class of optimization problems involving (weakly) spectral sets and functions where it is ensured that each (local) optimizer operator commutes with some element related to the objective function. For instance, let \mathcal{V} be a EJA, consider the map x\in\Omega\mapsto\Theta(x)+F(x), (1) where \Omega\subseteq\mathcal{V} , and \Theta,\,F:\mathcal{V}\to\mathbb{R} . In [ 17 ] , Ramírez, Seeger and Sossa have proved that when \Omega is a spectral set, \Theta is differentiable, and F is a spectral function (see definitions in the next section), if \bar{x} is a local optimizer (minimizer or maximizer) of ( 1 ), then \nabla\Theta(\bar{x}) and \bar{x} operator commute. Gowda and Jeong [ 5 ] further developed this commutation principle by assuming that \Omega and F are weakly spectral (i.e., they are invariant under automorphisms, which is a weaker condition than spectrality). These commutation principles were successfully applied in some variational problems like obtaining a meta-formula for generalized subdifferential of spectral functions [ 12 ] , and deducing bounds for the determinant of the sum of two elements in EJAs [ 20 ] . It has also been investigated strong commutation principles for this class of problems, which consist of commutation principles that provide strong operator commutation. For instance, Gowda [ 4 ] proved that under the assumption that \Omega and F are spectrally defined, and by setting \Theta(x)=\langle x,a\rangle , with a\in\mathcal{V} , \bar{x} is a global maximizer of ( 1 ) \Rightarrow a and \bar{x} strongly operator commute. (2) This class of results allows us, for instance, to simplify the problem of minimizing ( 1 ) to a minimization problem formulated in \mathbb{R}^{n} , with n being the rank of \mathcal{V} , via the eigenvalue map. In this paper, we study commutation principles for the problem of optimizing the map x\in\Omega\mapsto F(x-a), (3) where a\in\mathcal{V} , \Omega\subseteq\mathcal{V} and F:\mathcal{V}\to\mathbb{R} is a spectral function. Observe that, in general, the maps ( 3 ) and ( 1 ) are not related; so, we can not use directly the commutation principle discussed above to characterize the optimizers of ( 3 ). Indeed, despite the fact that F is spectral, the shifted spectral function F(\cdot-a) is no longer spectral. Moreover, we are not imposing a differentiability assumption over F . Some particular cases of ( 3 ) have been studied in the literature. For instance, Jeong and Sossa [ 9 ] have recently proved that when \Omega is weakly spectral and F is a strictly convex spectral function or a strictly convex spectral norm, if \bar{x} is a local optimizer of ( 3 ) then a and \bar{x} operator commute. Their proof relies on a commutation principle developed for ( 1 ) by using nonsmooth analysis. Massey, Rios and Stojanoff [ 14 ] have also obtained a commutation principle for ( 3 ) in the context of Hermitian matrices, which are examples of (simple) EJAs. Their result states that in case \Omega is an orbit of some element of \mathcal{V} and F is a strictly convex spectral norm, if \bar{x} is a local minimizer of ( 3 ) then a and \bar{x} strongly operator commute. Their approach is based on majorization techniques and on the smooth manifold structure of unitary orbits of Hermitian matrices. Similar results have also been applied in some other optimization problems [ 1 ] . Inspired by the works of Massey, Rios and Stojanoff [ 14 ] , and Jeong and Sossa [ 9 ] , our aim is to develop commutation principles and strong commutation principles for the problem of optimizing ( 3 ) with \Omega and F as general as possible. Our proof techniques are mainly inspired by those techniques developed in [ 14 ] ; that is, we will use majorization techniques in EJAs, see [ 8 ] . The spectral function F will be assumed to be strictly Schur-convex. Observe that this class of functions is quite broad and covers classes like spectral functions that are strictly convex, strictly quasi-convex, and strictly convex spectral norms. We also introduce a new strictly Schur-convex function which is used to deal with the problem of minimizing the condition number (see Section 7 ). Our main results are the following commutation principles: Let a,\,b\in\mathcal{V} , \Omega\subseteq\mathcal{V} and F:\mathcal{V}\to\mathbb{R} be a strictly Schur-convex spectral function; (i) Suppose that \Omega is a spectral set. If \bar{x} is a global minimizer (resp. global maximizer) of ( 3 ), then a and \bar{x} (resp. -a and \bar{x} ) strongly operator commute (see Theorem 4.1 ). (ii) Suppose that \Omega is a weakly spectral set. If \bar{x} is a local optimizer (minimizer or maximizer) of ( 3 ), then a and \bar{x} operator commute (see Theorem 5.1 ). (iii) Suppose that \mathcal{V} is simple, and set \Omega as the eigenvalue orbit of b . If \bar{x} is a local minimizer (resp. local maximizer) of ( 3 ), then a and \bar{x} (resp. -a and \bar{x} ) strongly operator commute. Moreover, \bar{x} is a global optimizer of ( 3 ) (see Theorem 6.1 ). Regarding item (i) above, we remark that in this case, if we replace the hypothesis that \bar{x} is a global minimizer (resp. global maximizer) by the weaker hypothesis that \bar{x} is a local minimizer (resp. local maximizer) of ( 3 ) then a and \bar{x} (resp. -a and \bar{x} ) do not necessarily strongly operator commute (see Example 6.4 ). We also point out that if we replace the assumption that \Omega is a spectral set for the (weaker) assumption that \Omega is a weakly spectral set and assume that \bar{x} is a global minimizer (resp. global maximizer) of ( 3 ) then, a and \bar{x} (resp. -a and \bar{x} ) do not necessarily strongly operator commute (see Example 4.2 below). Regarding item (ii), notice that weakly spectral sets are rather small domains (as opposed to spectral sets, that are larger domains) for an spectral function in an arbitrary EJA. Hence, the hypothesis that \bar{x} is a local optimizer within weakly spectral sets

Read the complete text