Some hybrid matrix triangle inequalities
Jean-Christophe Bourin, Eun-Young Lee · 2026
Read Some hybrid matrix triangle inequalities by Jean-Christophe Bourin, Eun-Young Lee (2026) 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-2606-29188.txt.
Introduction
Bourin and Lee sharpen a recent matrix inequality of Teng Zhang comparing a sum of matrices with a sum of their symmetrised sizes. For matrices, size is measured by unitarily invariant norms, which depend only on singular values. Zhang's inequality bounds the norm of a sum by a constant times the norm of summed quadratic symmetric moduli, an average of a matrix and its adjoint.
The pair derive operator and eigenvalue refinements, stronger versions that compare the matrices themselves, not just their norms. Their engine is a new polar decomposition of the quadratic symmetric modulus, a factorisation that splits a matrix into a rotation-like part and a size part. One refinement bounds the modulus of the sum by an average of the summed moduli and a unitarily rotated copy. They also treat the maximal modulus tied to Olson's order, with a parallel series of estimates.
Catalog record
| CREATOR | Jean-Christophe Bourin, Eun-Young Lee |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.29188
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Some hybrid matrix triangle inequalities Jean-Christophe Bourin, Eun-Young Lee Abstract Abstract. A recent result due to Teng Zhang compares the sum of m matrices and the sum of their quadratic symmetric moduli: \left\|\sum_{k=1}^{m}A_{k}\right\|\leq\sqrt{2}\left\|\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\| for every unitarily invariant norm. Here |A|_{\mathrm{qsym}} is the quadratic mean of |A| and |A^{*}| . We derive operator and eigenvalue refinements of Zhang’s inequality from a new polar decomposition for the quadratic symmetric modulus. For instance, \left|\sum_{k=1}^{m}A_{k}\right|\leq\frac{\sqrt{2}}{2}\left\{\sum_{k=1}^{m}\left(|A_{k}|_{\mathrm{qsym}}+V|A_{k}|_{\mathrm{qsym}}V^{*}\right)\right\} for some unitary matrix V . We also establish the polar decomposition for the maximal modulus associated with Olson’s order, and derive, as in the quadratic case, a series of estimates. Keywords. Matrix inequalities, log-majorization, unitary orbits, symmetric modulus, Olson’s spectral order. 2020 Mathematics Subject Classification. 15A42, 15A60, 47A30, 47A60. 1 Introduction In [ 10 ] , Zhang initiated a systematic study of matrix triangle inequalities involving different operator moduli. We refer to such estimates as hybrid triangle inequalities. This paper aims to refine one of Zhang’s theorems, stated in the abstract, and thereby contribute to the emerging theory of symmetrized moduli. While Zhang’s theorem is formulated in terms of weak majorization, our main results provide operator inequalities involving unitary orbits. Consequently, they yield refined eigenvalue and norm estimates. We denote by \mathbb{M}_{d} the space of complex d\times d matrices and \mathbb{M}_{d}^{+} its cone of positive semidefinite matrices. For A\in\mathbb{M}_{d} , we write |A|=(A^{*}A)^{1/2} for its absolute value (or right modulus). If A\in\mathbb{M}_{d}^{+} , we denote by \lambda_{1}^{\downarrow}(A)\geq\cdots\geq\lambda_{d}^{\downarrow}(A) its eigenvalues arranged in nonincreasing order. For A,B\in\mathbb{M}_{d}^{+} , the weak majorization, or submajorization, relation A\prec_{w}B means that \sum_{j=1}^{k}\lambda_{j}^{\downarrow}(A)\leq\sum_{j=1}^{k}\lambda_{j}^{\downarrow}(B),\qquad k=1,\ldots,d. (1.1) We adhere to the convention that, for j>d and A\in\mathbb{M}_{d}^{+} , \lambda_{j}^{\downarrow}(A):=0 . A basic property of symmetric norms (i.e., unitarily invariant ones) on \mathbb{M}_{d} is Fan’s dominance principle: if A,B\in\mathbb{M}^{+}_{d} , then \|A\|\leq\|B\|\quad{\mathrm{for\ all\ symmetric\ norms}}\iff|A|\prec_{w}|B|. We refer to [ 4 ] for a background and details on majorization, symmetric norms, and the matrix geometric mean. For a matrix Z\in\mathbb{M}_{d} , its left and right moduli |Z^{*}| and |Z| are the positive parts in the polar decompositions Z=|Z^{*}|U=|Z^{*}|^{1/2}U|Z|^{1/2}=U|Z|. These moduli occur in a number of important operator inequalities. Since neither modulus enjoys an intrinsic preference, it is natural to consider symmetrized versions obtained by averaging them. The symmetric modulus and the quadratic symmetric modulus of Z\in\mathbb{M}_{d} are respectively defined as |Z|_{\mathrm{sym}}:=\frac{|Z|+|Z^{*}|}{2},\qquad|Z|_{\mathrm{qsym}}:=\sqrt{\frac{|Z|^{2}+|Z^{*}|^{2}}{2}}. Note that |Z|_{\mathrm{qsym}}=\sqrt{|{\mathrm{Re\,}}Z|^{2}+|{\mathrm{Im\,}}Z|^{2}} and |Z|_{\mathrm{sym}}\leq|Z|_{\mathrm{qsym}} (1.2) by operator concavity of \sqrt{t} . The quadratic symmetric modulus enjoys remarkably strong subadditivity properties. A striking result of Zhang [ 9 , Theorem 1.9] states that: Theorem 1.1 . Let A,B\in\mathbb{M}_{d} . Then, for some unitaries U,V\in\mathbb{M}_{d} , |A+B|_{\mathrm{qsym}}\leq U|A|_{\mathrm{qsym}}U^{*}+V|B|_{\mathrm{qsym}}V^{*}. This is the symmetrized companion of the celebrated Thompson’s inequality [ 8 ] . Theorem 1.1 was motivated by the following special case [ 1 ] : Corollary 1.2 . Let Z\in\mathbb{M}_{d} . Then, for some unitaries U,V\in\mathbb{M}_{d} , |Z|_{\mathrm{qsym}}\leq U|{\mathrm{Re\,}}Z|U^{*}+V|{\mathrm{Im\,}}Z|V^{*}. Here ( 1.2 ) shows that the left-hand side may be replaced by |Z|_{\mathrm{sym}} . Despite this observation Theorem 1.1 does not hold for the symmetric modulus |\cdot|_{\mathrm{sym}} . Two matrices A,B\in\mathbb{M}_{3} are given in [ 10 ] with \lambda_{1}^{\downarrow}(|A+B|_{\mathrm{sym}})=\sqrt{2}\left\{\lambda_{1}^{\downarrow}(|A|_{\mathrm{sym}})+\lambda_{1}^{\downarrow}(|B|_{\mathrm{sym}})\right\}. Here, in dimension d\geq 3 , the constant \sqrt{2} is optimal. Indeed, a result of [ 2 ] states a submajorization form of the triangle inequality as follows. Theorem 1.3 . Let \{X_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}X_{k}\right|_{\mathrm{sym}}\prec_{w}\sqrt{2}\sum_{k=1}^{m}\left|X_{k}\right|_{\mathrm{sym}}. In his impressive work [ 10 ] , Zhang got the idea to study triangle inequalities between different types of moduli. In particular, he observed the following hybrid version of Theorem 1.3 : Theorem 1.4 . Let \{X_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}X_{k}\right|\prec_{w}\sqrt{2}\sum_{k=1}^{m}\left|X_{k}\right|_{\mathrm{qsym}}. Here both the usual modulus and the quadratic symmetric modulus appear simultaneously. The constant \sqrt{2} is optimal in any dimension d\geq 2 . Theorem 1.4 is remarkable if one keeps in mind the well known submajorization [ 6 , Proposition 2.1] : Proposition 1.5 . Let \{X_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}X_{k}\right|\prec_{w}\sqrt{m}\sum_{k=1}^{m}\left|X_{k}\right| Hence replacing the usual modulus in the right-hand side by the quadratic symmetric modulus yields the pleasant constant \sqrt{2} , regardless of the dimension. The constant \sqrt{m} is optimal whenever m\leq d . If m>d , then the multiplicative constant \sqrt{m} can be replaced by the optimal one \sqrt{d} , see [ 10 , Theorem 1.7] . The purpose of Section 2 is to refine Theorem 1.4 at the operator level. Our main tool is a polar decomposition associated with the quadratic symmetric modulus. It yields unitary-orbit inequalities strengthening Zhang’s submajorization estimate and leads to several eigenvalue consequences. We also establish, in Section 3, a polar decomposition for the maximal modulus associated with Olson’s spectral order. This allows us to obtain estimates for the maximal modulus similar to those of Section 2 for the quadratic modulus. 2 Unitary orbits and log-majorization The following lemma may be regarded as a polar decomposition for the quadratic symmetric modulus. Lemma 2.1 . Let A\in\mathbb{M}_{d} . Then there exists a contraction C\in\mathbb{M}_{d} such that A=\sqrt{2}|A|_{\mathrm{qsym}}^{1/2}C|A|_{\mathrm{qsym}}^{1/2}. Proof. By a limit argument we may suppose invertibility of A . It then suffices to show that C=(|A|^{2}+|A^{*}|^{2})^{-1/4}A(|A|^{2}+|A^{*}|^{2})^{-1/4} is a contraction. Note that the polar decomposition A=|A^{*}|^{1/2}U|A|^{1/2} gives C=(|A|^{2}+|A^{*}|^{2})^{-1/4}\left(|A^{*}|^{1/2}U|A|^{1/2}\right)(|A|^{2}+|A^{*}|^{2})^{-1/4}. Thus C=XY where X=(|A|^{2}+|A^{*}|^{2})^{-1/4}|A^{*}|^{1/2},\qquad Y=U|A|^{1/2}(|A|^{2}+|A^{*}|^{2})^{-1/4}. Hence \|C\|_{\infty}^{2}\leq\|X\|_{\infty}^{2}\|Y\|_{\infty}^{2}=\|X^{*}X\|_{\infty}\|YY^{*}\|_{\infty} , that is \|C\|_{\infty}^{2}\leq\||A^{*}|^{1/2}(|A|^{2}+|A^{*}|^{2})^{-1/2}|A^{*}|^{1/2}\|_{\infty}\||A|^{1/2}(|A|^{2}+|A^{*}|^{2})^{-1/2}|A|^{1/2}\|_{\infty}. Now, since t\mapsto t^{-1/2} is operator decreasing, we have |A^{*}|^{1/2}(|A|^{2}+|A^{*}|^{2})^{-1/2}|A^{*}|^{1/2}\leq|A^{*}|^{1/2}(|A^{*}|^{2})^{-1/2}|A^{*}|^{1/2}=I, and |A|^{1/2}(|A|^{2}+|A^{*}|^{2})^{-1/2}|A|^{1/2}\leq|A|^{1/2}(|A|^{2})^{-1/2}|A|^{1/2}=I. The last three inequalities together give \|C\|_{\infty}^{2}\leq 1. ∎ Now we give a hybrid comparison between the usual modulus and the quadratic symmetric modulus with the help of the geometric mean \# on \mathbb{M}_{d}^{+} . The result is strongly influenced by Zhang’s Theorem 1.4 . Theorem 2.2 . Let \{A_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}A_{k}\right|\leq\sqrt{2}\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}\#V\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}V^{*} for some unitary matrix V\in\mathbb{M}_{d} . Proof. By Lemma 2.1 , for every A\in\mathbb{M}_{d} , the block matrix \begin{bmatrix}\sqrt{|A|^{2}+|A^{*}|^{2}}&A\\ A^{*}&\sqrt{|A|^{2}+|A^{*}|^{2}}\end{bmatrix} is positive semidefinite. Hence so is the sum of m such matrices, \begin{bmatrix}\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}&\sum_{k}A_{k}\\ \sum_{k}A_{k}^{*}&\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\end{bmatrix} where each sum runs over k\in\{1,\ldots,m\} . Now consider the polar decomposition \sum_{k}A_{k}=U|\sum_{k}A_{k}| and perform the unitary congruence \begin{bmatrix}U^{*}&0\\ 0&I\end{bmatrix}\begin{bmatrix}\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}&\sum_{k}A_{k}\\ \sum_{k}A_{k}^{*}&\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\end{bmatrix}\begin{bmatrix}U&0\\ 0&I\end{bmatrix} to get the positive semidefinite block-matrix: \begin{bmatrix}U^{*}\left\{\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\right\}U&\left|\sum_{k}A_{k}\right|\\ \left|\sum_{k}A_{k}\right|&\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\end{bmatrix} By the maximal characterization of the geometric mean, \left|\sum_{k}A_{k}\right|\leq\left(U^{*}\left\{\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\right\}U\right)\#\left\{\sum_{k}\sqrt{|A_{k}|^{2}+|A_{k}^{*}|^{2}}\right\} Using X\#Y=Y\#X and setting V=U^{*} completes the proof. ∎ From the matrix AGM inequality we obtain: Corollary 2.3 . Let \{A_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}A_{k}\right|\leq\frac{\sqrt{2}}{2}\sum_{k=1}^{m}\left(|A_{k}|_{\mathrm{qsym}}+V|A_{k}|_{\mathrm{qsym}}V^{*}\right) for some unitary matrix V\in\mathbb{M}_{d} . Corollary 2.3 clearly implies Zhang’s Theorem 1.4 . Besides the series of eigenvalue inequalities of the form ( 1.1 ), Corollary 2.3 adds some new ones, by application of the classical Weyl inequalities: Corollary 2.4 . Let \{A_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, for all i,j=0,1,\ldots , \lambda_{1+i+j}^{\downarrow}\left(\left|\sum_{k=1}^{m}A_{k}\right|\right)\leq\frac{\lambda_{1+i}^{\downarrow}\left(\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right)+\lambda_{1+j}^{\downarrow}\left(\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right)}{\sqrt{2}}. Another consequence of Theorem 2.2 is a weak log-majorization \prec_{\mathrm{wlog}} on \mathbb{M}_{d}^{+} given in the next corollary. Recall that, for A,B\in\mathbb{M}_{d}^{+} , we write A\prec_{\mathrm{wlog}}B to mean that the following d inequalities hold: \prod_{j=1}^{p}\lambda_{j}^{\downarrow}(A)\leq\prod_{j=1}^{p}\lambda_{j}^{\downarrow}(B),\qquad p=1,\ldots,d. It is well known that A\prec_{\mathrm{wlog}}B\Rightarrow A\prec_{w}B . Corollary 2.5 . Let \{A_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then, \left|\sum_{k=1}^{m}A_{k}\right|\prec_{\mathrm{wlog}}\sqrt{2}\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}. Proof. We have \left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}\#V\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}V^{*}=\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}^{1/2}\!\!W\!\cdot\!V\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}^{1/2}\!\!\!V^{*} for some unitary W\in\mathbb{M}_{d} . Using Theorem 2.2 , we infer \left|\sum_{k=1}^{m}A_{k}\right|\leq\sqrt{2}\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}^{1/2}\!\!W\!\cdot\!V\left\{\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right\}^{1/2}\!\!\!V^{*}. (2.1) This combined with Horn’s multiplicative inequalities for singular values yields \prod_{j=1}^{p}\lambda_{j}^{\downarrow}\left(\left|\sum_{k=1}^{m}A_{k}\right|\right)\leq\prod_{j=1}^{p}\lambda_{j}^{\downarrow}\left(\sqrt{2}\sum_{k=1}^{m}|A_{k}|_{\mathrm{qsym}}\right) for all p=1,\ldots,d . This is exactly our weak log-majorization relation \prec_{\mathrm{wlog}} . ∎ From another classical singular value inequality of Weyl, we obtain from ( 2.1 ): Corollary 2.6 . Let \{A_{k}\}_{k=1}^{m} be in \mathbb{M}_{d} . Then,