---
title: Araki–Yamagami Inequality Overview
url: https://www.emergentmind.com/topics/araki-yamagami-inequality
type: topic
---

# Araki–Yamagami Inequality Overview

Searching arXiv for relevant papers on the Araki–Yamagami inequality and closely related Araki-type inequalities.
The Araki–Yamagami inequality is a sharp perturbation estimate for the positive factor in the generalized polar decomposition of a matrix. In its classical finite-dimensional form, if \(A=QH\) and \(\widetilde A=\widetilde Q\,\widetilde H\) are generalized polar decompositions of complex matrices, then
\[
\|H-\widetilde H\|_F \le \sqrt{2}\,\|A-\widetilde A\|_F,
\]
where \(H=|A|=(A^*A)^{1/2}\), \(\widetilde H=|\widetilde A|\), and \(\sqrt{2}\) is optimal. In current literature, the name also appears in a broader sense, as part of “Araki–Yamagami-type” trace and log-majorization inequalities comparing noncommutative products such as \(A^sB^s\) and \((A^{1/2}BA^{1/2})^s\); however, that broader usage is distinct from the classical Frobenius-norm perturbation theorem [2507.14940][2507.05242].

## 1. Classical statement and polar-decomposition setting

For \(A\in\mathbb C^{m\times n}\) of rank \(r\), the generalized polar decomposition is
\[
A=QH,
\]
where \(Q\in\mathbb C^{m\times n}\) is subunitary, equivalently a partial isometry whose nonzero singular values are all \(1\), and \(H\in\mathbb C^{n\times n}\) is positive semidefinite with \(H=|A|=(A^*A)^{1/2}\). Uniqueness holds when \(\mathcal R(Q^*)=\mathcal R(H)\). For a perturbation \(\widetilde A=A+E\) with decomposition \(\widetilde A=\widetilde Q\,\widetilde H\), the classical Araki–Yamagami inequality controls the positive polar factors by the Frobenius norm of the perturbation:
\[
\|H-\widetilde H\|_F \le \sqrt2\,\|E\|_F.
\]
The result is global, requires no rank assumption beyond finite dimensionality, and the constant \(\sqrt2\) is optimal. In this formulation, the inequality is a Lipschitz bound for the map \(A\mapsto |A|\) in the Frobenius, or Hilbert–Schmidt, norm [2507.14940].

The Frobenius norm is structurally natural in this setting. The identity \(\|X\|_F^2=\operatorname{Tr}(X^*X)\) enables trace and angle arguments, while unitary invariance makes singular-value reductions and SVD-based normal forms canonical. This is also the norm in which the original theorem is stated [2507.14940].

## 2. Geometric interpretation and proof mechanisms

The inequality compares the positive parts of two matrices rather than the matrices themselves. In perturbation-theoretic terms, it isolates the stability of the positive polar factor \(H=|A|\) under additive perturbation \(E\), independently of the behavior of the subunitary factor \(Q\). The estimate is therefore an operator-theoretic analogue of stability for a nonlinear matrix functional, with the notable feature that the optimal global constant is universal.

A recent reproving of the classical result proceeds by combining an angle inequality of Lin–Zhang type,
\[
\cos^2\alpha \le \cos\beta,
\]
where \(\alpha\) is the angle between \(A\) and \(B\) and \(\beta\) is the angle between \(|A|\) and \(|B|\), with an elementary scalar optimization in the parameter \(r=\|A\|_F/\|B\|_F\). This yields
\[
\frac{\||A|-|B|\|_F^2}{\|A-B\|_F^2}\le 2,
\]
and hence the classical inequality. The same analysis confirms that \(\sqrt2\) is not merely admissible but sharp [2507.14940].

## 3. Sharp spectral refinement

The principal modern refinement replaces the universal constant by a spectral quantity determined by the singular values of \(A\) and \(\widetilde A\). If
\[
A=QH,\qquad \widetilde A=\widetilde Q\,\widetilde H,
\]
with ranks \(r=\operatorname{rank}(A)\), \(s=\operatorname{rank}(\widetilde A)\), and singular values
\[
\sigma_1\ge \cdots \ge \sigma_r>0,\qquad \widetilde\sigma_1\ge \cdots \ge \widetilde\sigma_s>0,
\]
then one defines
\[
F_{r,s}:=\sum_{j=1}^r \sigma_j^2+\sum_{j=1}^s \widetilde\sigma_j^2,\qquad
G_{r,r}:=\sum_{j=1}^r \sigma_j\widetilde\sigma_j.
\]
Writing
\[
t:=\frac{F_{r,s}}{G_{r,r}},
\]
the sharp upper-bound coefficient satisfies
\[
c^2=g_1(t):=t-\sqrt{t^2-2t},
\]
so that
\[
\|H-\widetilde H\|_F \le c\,\|A-\widetilde A\|_F.
\]
By AM–GM, \(t\ge 2\), and \(g_1'(t)<0\), hence \(g_1(t)\le g_1(2)=2\). The classical Araki–Yamagami bound is therefore recovered as the uniform envelope of the sharper spectral estimate. More precisely, the universal constant \(\sqrt2\) is attained only when \(r=s\) and the singular values coincide termwise, \(\sigma_j=\widetilde\sigma_j\) for all \(j\). If \(r\ne s\), or if \(r=s\) but the singular values differ somewhere, then the actual sharp constant is strictly smaller than \(\sqrt2\) [2507.14940].

This refinement changes the interpretation of the classical inequality. The latter is no longer best understood as the intrinsic perturbation constant of the positive polar factor, but rather as the worst-case value obtained after discarding all spectral information. The spectrum-dependent bound is also two-sided: the same work establishes a matching sharp lower bound for \(\|H-\widetilde H\|_F\), showing that the upper estimate is part of an exact optimal trade-off rather than an isolated one-sided inequality [2507.14940].

## 4. Relation to neighboring perturbation inequalities

The same perturbation framework treats the subunitary polar factors \(Q\) and \(\widetilde Q\). For \(r\le s\), sharp upper and lower bounds are given for \(\|Q-\widetilde Q\|_F\) in terms of \(\|E\|_F\) and the singular values of \(A\) and \(\widetilde A\). In the equal-rank case, this refines the Li–Sun estimate
\[
\|Q-\widetilde Q\|_F \le \frac{2}{\sigma_r+\widetilde\sigma_r}\,\|E\|_F
\]
by replacing the endpoint singular-value dependence with an optimized expression involving the full singular-value lists. The refinement is strict except in worst spectral configurations, exactly paralleling the way the spectral upper bound for \(\|H-\widetilde H\|_F\) improves the classical Araki–Yamagami constant [2507.14940].

The same paper places the Araki–Yamagami refinement inside a larger network of Frobenius-norm inequalities. It strengthens Lee’s conjecture by introducing a spectrum-dependent coefficient whose square is
\[
g_2(t):=\frac{1}{\sqrt{t^2+2t}-t},
\]
again with \(t=F_{r,s}/G_{r,r}\), and with the original constant appearing as the worst-case envelope. It also sharpens Kittaneh’s theorem
\[
\||A|-|B|\|_F^2+\||A^*|-|B^*|\|_F^2\le 2\|A-B\|_F^2,
\]
whose normal-matrix corollary is
\[
\||A|-|B|\|_F \le \|A-B\|_F.
\]
In addition, the same convex-analytic machinery yields stronger matrix AM–GM and Cauchy–Schwarz inequalities in Frobenius norm, with sharp coefficients expressed through singular values. A consistent theme is that universal Hilbert–Schmidt inequalities admit nontrivial sharpening once the singular or eigenvalue data are retained rather than suppressed [2507.14940].

## 5. “Araki–Yamagami-type” inequalities in trace and log-majorization form

A separate but related usage of the name occurs in trace and log-majorization inequalities for positive semidefinite matrices. In that literature, “Araki–Yamagami-type” often refers to comparisons among expressions such as
\[
A^{1-s}B^s,\qquad A(A^{-1/2}BA^{-1/2})^s,\qquad (A^{1/2}BA^{1/2})^s,
\]
typically derived from Araki’s log-majorization, BLP, GBLP, or related inequalities. A recent result proves that for \(A,B\ge0\), \(s\in[0,1]\), and any scalar function \(f\) that is nonnegative and nondecreasing on an interval containing \(\operatorname{spec}(A)\),
\[
\operatorname{Tr}\!\big[f(A)A^sB^s\big]\le \operatorname{Tr}\!\big[f(A)(A^{1/2}BA^{1/2})^s\big].
\]
It also establishes the reverse inequality
\[
\operatorname{Tr}\!\big[g(A)(A^{1/2}BA^{1/2})^s\big]\le \operatorname{Tr}\!\big[g(A)A^sB^s\big]
\]
when \(x\mapsto x^s g(x)\) is nonnegative and nonincreasing. These are weighted trace refinements of the Araki-type comparison between \(A^sB^s\) and \((A^{1/2}BA^{1/2})^s\), but they are not the same statement as the classical perturbation inequality for polar factors [2507.05242].

The broader functional framework is visible in generalized Araki–Lieb–Thirring theory. For positive semidefinite matrices \(A\) and \(B\), a continuous nonnegative geometrically concave function \(f\) satisfies
\[
\lambda(f(A)f(B)) \prec_{w(\log)} f^2\!\bigl(\lambda^{1/2}(AB)\bigr)
\]
for all such matrices if and only if \(f\) lies in class \(A\), meaning that \(f\) is geometrically concave and obeys
\[
0\le x f'(x)\le f(x)
\]
wherever \(f'\) exists. The reversed inequality holds for geometrically convex functions in class \(B\), characterized by
\[
x f'(x)\ge f(x).
\]
Although this framework does not state the classical Araki–Yamagami perturbation theorem, it supplies a functional log-majorization template in which many Araki-type and Araki–Yamagami-type comparisons can be organized [1204.3418].

## 6. Reverse inequalities, phase transitions, and current understanding

Recent work on reverse Araki-type inequalities clarifies the limits of possible reversals of the weighted trace inequalities. For arbitrary nonnegative nondecreasing weights \(f\), the reverse inequality
\[
\operatorname{Tr} f(A)A^sB^s \ge \operatorname{Tr} f(A)(A^{1/2}BA^{1/2})^s
\]
holds dimension-freely exactly for
\[
1\le s\le 2.
\]
For every \(s>2\), explicit \(3\times3\) positive definite counterexamples show that the reverse inequality fails, whereas in dimension \(2\) it remains valid for every \(s\ge1\). Thus \(s=2\) is the sharp dimension-free boundary, and dimension \(3\) is the minimal dimension in which failure occurs [2607.01263].

For power weights, the reverse picture is more intricate. In normalized variables, the reverse log-majorization
\[
A^{r+s}B^s \succ_{\log} A^r(A^{1/2}BA^{1/2})^s
\]
holds for all finite dimensions if and only if either \(1\le s\le2\), or \(s>2\) together with
\[
0\le r\le \frac{s}{s-2}.
\]
Equivalently, in the original variables \(0<p\le q\), the sharp condition is \(0\le r\le pq/(q-2p)\) when \(q>2p\), while the whole region \(p\le q\le 2p\) survives for all \(r\ge0\). Beyond the hyperbolic boundary, even the trace inequality fails for \(3\times3\) positive definite matrices. This does not modify the content of the classical Araki–Yamagami perturbation bound, but it materially sharpens the surrounding theory by delineating where reverse Araki-type inequalities can and cannot hold [2607.01263].

Taken together, these developments place the classical Araki–Yamagami inequality in a broader modern landscape. In matrix perturbation theory, it is now understood as the optimal universal member of a sharper spectrum-dependent family for positive polar factors. In operator-inequality and quantum-information-adjacent settings, its name also designates a wider class of trace and log-majorization comparisons whose valid parameter regimes can be subtle and, in reverse directions, sharply dimension-dependent. The resulting picture is technically unified by singular-value reduction, convex-analytic extremality, and log-majorization methods, but the classical theorem remains the benchmark statement: a sharp Frobenius-norm bound for the positive polar factor under perturbation [2507.14940].

Source: https://www.emergentmind.com/topics/araki-yamagami-inequality