---
title: Kubo–Ando Connections in Operator Theory
url: https://www.emergentmind.com/topics/kubo-ando-connections
type: topic
---

# Kubo–Ando Connections in Operator Theory

Searching arXiv for recent and foundational papers on Kubo–Ando connections to ground the article in cited sources.
Kubo–Ando connections are binary operations on the positive cone of a Hilbert-space operator algebra that axiomatize operator means through monotonicity, transformer inequality, and continuity from above. In the normalized case they are Kubo–Ando means, and their decisive structural feature is a one-to-one correspondence with operator monotone functions on \((0,\infty)\). This framework unifies arithmetic, harmonic, geometric, logarithmic, and power-type means, and it has become a standard language for operator inequalities, scalar-to-operator transfer principles, divergence geometry, and several recent developments in quantum information and noncommutative analysis.

## 1. Axiomatic framework and representation theorem

Let \(H\) be a Hilbert space and \(B(H)^+\) the cone of positive operators. A Kubo–Ando connection is a binary operation
\[
\sigma:B(H)^+\times B(H)^+\to B(H)^+
\]
satisfying monotonicity, the transformer inequality, and joint continuity from above. A mean is a connection normalized by \(I\sigma I=I\). In the finite- and infinite-dimensional formulations used in the literature, the central representation theorem states that such means are in bijection with operator monotone functions \(f:(0,\infty)\to(0,\infty)\) satisfying \(f(1)=1\), through
\[
A\sigma B
=
A^{1/2}f\!\bigl(A^{-1/2}BA^{-1/2}\bigr)A^{1/2},
\qquad A>0.
\]
Conversely, every such \(f\) defines a Kubo–Ando mean, and the representing function is unique [2605.11701].

This correspondence immediately identifies the classical examples and places them in a common operator-calculus format.

| Mean | Representing function \(f(t)\) | Operator formula |
|---|---|---|
| Arithmetic | \(\frac{1+t}{2}\) | \(\frac{A+B}{2}\) |
| Geometric | \(t^{1/2}\) | \(A^{1/2}(A^{-1/2}BA^{-1/2})^{1/2}A^{1/2}\) |
| Harmonic | \(\frac{2t}{1+t}\) | \(2(A^{-1}+B^{-1})^{-1}\) |

The same formalism includes weighted arithmetic, weighted geometric, weighted harmonic, and logarithmic means. For a Kubo–Ando mean with representing function \(f\), if \(\alpha=f'(1)\), then the standard operator bounds
\[
A!_\alpha B \le A\sigma B \le (1-\alpha)A+\alpha B
\]
hold. This places every normalized connection between weighted harmonic and weighted arithmetic means and encodes its order-theoretic position through the scalar derivative at the identity [2605.11701].

A second standard representation expresses a connection as an integral of weighted harmonic means against a finite Borel measure. In one formulation,
\[
A\sigma B=\alpha A+\beta B+\int_{(0,\infty)}(A!_\lambda B)\,d\mu(\lambda),
\]
where \(\alpha=\mu(\{0\})\), \(\beta=\mu(\{\infty\})\), and \(A!_\lambda B\) is a weighted harmonic mean. This measure-theoretic description is the bridge between Kubo–Ando theory and Löwner-type integral representations of operator monotone functions.

## 2. Equivalent axiomatizations, scalar connections, and the cone structure

The original continuity axiom can be weakened substantially without changing the class of connections. In particular, joint continuity from above may be replaced by one-sided conditions such as
\[
A_n\downarrow A \implies A_n\sigma X \downarrow A\sigma X
\quad\text{and}\quad
I\sigma A_n \downarrow I\sigma A,
\]
or the corresponding right-sided variant. Under monotonicity and the transformer inequality, these weaker conditions are equivalent to the original Kubo–Ando continuity axiom [1208.4912].

A further refinement is that monotonicity itself can be replaced by joint concavity, or even midpoint concavity, provided the transformer inequality and one of the continuity conditions are retained. Thus, within the axiomatic architecture, concavity is not merely a consequence of the theory; it can serve as an alternative defining principle [1208.4912].

Every connection induces a scalar connection on \(\mathbb{R}^+\) by
\[
(x\tilde{\sigma} y)I=(xI)\sigma(yI),
\]
and if \(f\) is the representing function, then
\[
x\tilde{\sigma} y = x f(y/x), \qquad x,y>0.
\]
The induced scalar connection and the operator connection have the same representing function and the same representing measure. The assignment
\[
\sigma \mapsto \tilde{\sigma}
\]
is an affine order isomorphism between the cone of operator connections and the cone of scalar connections [1208.4912].

The cone of connections itself admits a natural norm. For a connection \(\sigma\),
\[
\|\sigma\|
=
\sup\{\|A\sigma B\|:\|A\|=\|B\|=1,\ A,B\ge0\}
=
\|I\sigma I\|.
\]
With pointwise addition and nonnegative scalar multiplication, the set of connections becomes a normed ordered cone. This cone is isometrically order-isomorphic to the cone of operator monotone functions on \(\mathbb{R}^+\), with norm \(f\mapsto f(1)\), and isometrically isomorphic to the cone of finite Borel measures on \([0,\infty]\), with norm given by total mass. A connection is a mean if and only if its norm is \(1\) [1304.2452].

This three-way equivalence—connections, operator monotone functions, and finite Borel measures—is one of the most structural aspects of the theory. It identifies normalization with unit mass or unit value at \(1\), and it makes convergence of connections equivalent to convergence of their representing functions at \(1\) and of their representing measures in total mass [1304.2452].

## 3. Canonical examples, power means, and operator-monotonicity tests

A central recent development concerns power-type families inside and outside the Kubo–Ando class. For positive definite matrices \(A,B\) and \(p\ge0\), the Kubo–Ando matrix power mean is
\[
P_\mu(p,A,B)
=
A^{1/2}
\left(\frac{1+(A^{-1/2}BA^{-1/2})^p}{2}\right)^{1/p}
A^{1/2}.
\]
Its representing function is
\[
f_p(t)=\left(\frac{1+t^p}{2}\right)^{1/p}.
\]
For \(0<p\le1\), \(f_p\) is operator monotone on \((0,\infty)\), so \(P_\mu(p,\cdot,\cdot)\) is a genuine Kubo–Ando mean. For \(p>1\), the same scalar formula fails to be operator monotone in general, so the corresponding matrix expression no longer need satisfy the Kubo–Ando axioms [2106.05914].

This family yields new characterizations of operator monotone functions. For \(0<p\le1\le q\), if a continuous function \(f\) satisfies one of the inequalities
\[
f(A\# B)\le f(P_\mu(p,A,B)),
\]
\[
f(P_\mu(p,A,B))\le f\!\left(\frac{A+B}{2}\right),
\]
or
\[
f\!\left(\frac{A+B}{2}\right)\le f(P_\mu(q,A,B))
\]
for all positive definite \(A,B\), then \(f\) is operator monotone on \((0,\infty)\). In this sense, monotonicity along chains of geometric, power, and arithmetic means becomes a test for operator monotonicity itself [2106.05914].

The same work studies “naive” matrix power means
\[
M_p(A,B)=\left(\frac{A^p+B^p}{2}\right)^{1/p},
\]
which are generally not of Kubo–Ando form because the defining functional calculus is applied to \(A\) and \(B\) separately rather than through \(A^{-1/2}BA^{-1/2}\). Even so, certain inequalities involving these non-Kubo–Ando means still characterize operator monotone functions. A key example is that if a continuous \(f\) satisfies
\[
f\!\left(\frac{A^{1/2}+B^{1/2}}{2}\right)
\le
f\!\left(\frac{A+B}{2}\right)
\]
for all positive semidefinite \(A,B\), then \(f\) is operator monotone. By contrast, analogous statements for exponent \(2\) do not always characterize operator monotonicity; explicit non-monotone power functions can satisfy such inequalities in that regime [2106.05914].

These results invert the usual Kubo–Ando logic. Classical theory starts from operator monotonicity of \(f\) and builds a mean. The power-mean characterizations start from inequalities between means and deduce operator monotonicity of \(f\). This suggests a broader principle: operator means can serve not only as objects classified by operator monotone functions, but also as probes that detect operator monotonicity.

## 4. Symmetric means, divergence centers, and order determination

For symmetric Kubo–Ando means, there is a variational interpretation in terms of divergence centers. Let \(\sigma\) be a symmetric Kubo–Ando mean with representing function \(f_\sigma\). Define
\[
g_\sigma(x)
=
\int_1^x \left(1-\frac{1}{f_\sigma^{-1}(t)}\right)\,dt,
\]
and then
\[
D_\sigma(A,B)
=
\operatorname{Tr}\,g_\sigma\!\bigl(A^{-1/2}BA^{-1/2}\bigr)
\]
on positive definite operators. The resulting \(D_\sigma\) is an Amari-type divergence, and the binary mean \(A\sigma B\) is the unique minimizer of
\[
X\mapsto \frac12\bigl(D_\sigma(A,X)+D_\sigma(B,X)\bigr).
\]
Thus every symmetric Kubo–Ando mean is a divergence center of its two arguments [2002.11678].

This variational viewpoint extends naturally to weighted multivariate means. For positive definite \(A_1,\dots,A_m\) and a probability vector \(w\), one defines
\[
Q_{\sigma,\mathcal A,w}(X)
=
\sum_{j=1}^m w_j D_\sigma(A_j,X),
\]
and the unique minimizer
\[
\operatorname{bc}(\sigma,\mathcal A,w)
=
\arg\min_{X>0} Q_{\sigma,\mathcal A,w}(X)
\]
is the weighted barycenter associated with \(\sigma\). In the geometric case, this construction recovers the weighted \(\mathcal A\#\mathcal H\)-mean of Kim–Lawson–Lim; more generally, the barycenter lies above the weighted multivariate harmonic mean in Löwner order [2002.11678].

The geometric mean occupies a distinguished place in this framework. The divergence \(D_\sigma\) is symmetric if and only if \(\sigma\) is the geometric mean, and the associated weighted multivariate barycenter is then
\[
\left(\sum_{j=1}^m w_j A_j^{-1}\right)^{-1}
\#
\left(\sum_{j=1}^m w_j A_j\right).
\]
This isolates the geometric mean as the unique symmetric Kubo–Ando mean whose canonical divergence is itself symmetric [2002.11678].

Symmetric means also determine order through their norms. If \(\sigma\) is any symmetric Kubo–Ando mean on \(\mathcal B(H)\), then
\[
\|A\sigma X\|\le \|B\sigma X\|
\quad\forall X\in \mathcal B(H)^{++}
\]
implies \(A\le B\). Equivalently, the norm of every symmetric Kubo–Ando mean is order-determining on \(\mathcal B(H)\) [2301.06355]. This result is conceptually close to the divergence-center interpretation: in both cases, the mean is not merely an interpolant between \(A\) and \(B\), but a structure rich enough to encode order-theoretic information about the positive cone.

## 5. Extensions to unbounded operators, completely positive maps, and channels

Kubo–Ando theory was originally formulated for bounded positive operators, but it has been extended to several unbounded settings natural in von Neumann algebra theory. One line of work generalizes connections to positive \(\tau\)-measurable operators affiliated with a semifinite von Neumann algebra, positive elements in Haagerup \(L^p\)-spaces, and semifinite normal weights. The construction is built around suitable extensions of parallel sum and around analysis of decreasing sequences, and it also yields a version of Ando’s Lebesgue decomposition in noncommutative \(L^p\)-spaces [2101.01176].

A distinct extension lifts Kubo–Ando means from operators to completely positive maps. Fix a \(C^*\)-algebra \(\mathscr A\), a Hilbert space \(\mathcal H\), a CP map \(\Omega:\mathscr A\to B(\mathcal H)\), and \(\Phi,\Psi\le \Omega\). Arveson’s Radon–Nikodym theorem identifies \([0,\Omega]\) with an operator interval in the commutant of a Stinespring representation of \(\Omega\). Transporting a Kubo–Ando mean \(\sigma\) through this order isomorphism yields the relative mean
\[
(\Phi\sigma_\Omega\Psi)(a)
=
V^*\pi(a)\bigl(D_\Phi^\Omega \sigma D_\Psi^\Omega\bigr)V.
\]
Taking \(\Omega=\Phi+\Psi\) gives an intrinsic CP-mean. These means satisfy closure, monotonicity, positive homogeneity, idempotence, CP-transformer inequalities, a \(C^*\)-Jensen inequality, data processing under CP post-processing, and monotonicity with respect to the ambient map [2605.11701].

In the geometric case, the CP-map construction has a block-positivity characterization: \(\Phi\#_\Omega\Psi\) is the largest CP map \(\Gamma\) such that
\[
\begin{pmatrix}
\Phi & \Gamma\\
\Gamma & \Psi
\end{pmatrix}
\]
is completely positive. In this case the relative and intrinsic geometric means coincide, and
\[
\Phi\#\Psi=0
\iff
[0,\Phi]\cap[0,\Psi]=\{0\},
\]
so the intrinsic geometric mean detects the absence of nonzero common CP submaps. In finite dimensions, these means agree with Choi-matrix means, and in the von Neumann algebraic setting the intrinsic geometric mean agrees with Okayasu’s Pusz–Woronowicz geometric mean [2605.11701].

The two-variable weighted Kubo–Ando geometric means also admit a channel-level formulation through a superoperator perspective. In this setting they are characterized as the only \(2\)-variable operator geometric means that are block additive, tensor multiplicative, and satisfy the arithmetic–geometric mean inequality. The same paper extends this optimality property to quantum channels and introduces the notion of superoperator perspective function, together with basic monotonicity properties under CP supermaps [2503.13379].

Taken together, these extensions show that the Kubo–Ando mechanism is not limited to bounded operator pairs. It survives transport to unbounded positive objects, to CP maps through Radon–Nikodym derivatives, and to channels through Choi and superoperator perspectives. This suggests that the theory is best viewed as an order-theoretic calculus, rather than as a bounded-operator formalism narrowly construed.

## 6. Weak associativity, rigidity, and preserver phenomena

A recent characterization problem concerns symmetric Kubo–Ando means satisfying Molnár’s weak associativity. The corresponding class is defined through representing functions \(f\) satisfying
\[
f(c^2x)=c\,f(x)
\]
for some \(c>0\), \(c\ne1\), together with operator monotonicity, normalization, and symmetry. This class contains the geometric mean, but it is strictly larger: there is an order-preserving bijection between the class and a family of real measurable odd periodic functions bounded in absolute value by \(1/2\), and this yields infinitely many non-geometric Molnár means [2405.20108].

The same analysis shows that a single scaling relation does not isolate the geometric mean, but two incommensurate scaling relations do. More precisely, if a symmetric operator monotone function satisfies the usual symmetry and normalization conditions together with
\[
f(c_1^2x)=c_1 f(x),\qquad f(c_2^2x)=c_2 f(x),
\]
where \(\log c_1/\log c_2\notin\mathbb Q\), then \(f(x)=\sqrt{x}\), so the corresponding mean is the geometric mean [2405.20108].

Rigidity also appears in quantum positivity problems. If \(\sigma\) is a Kubo–Ando mean with representing function \(f\), define \(\kappa_\sigma=-f''(1)\). For non-arithmetic means one has \(\kappa_\sigma>0\), and this curvature forces violations of entanglement-related cone stability. In particular, weighted arithmetic means are the only Kubo–Ando means that preserve the separable cone in all bipartite dimensions; any non-arithmetic mean can violate the PPT condition already in the two-qubit setting and can strictly increase Schmidt number. Through the Choi–Jamiołkowski correspondence, this implies that convex mixing is the uniquely permissible Kubo–Ando operation for preserving entanglement-breaking channels [2605.26272].

Preserver problems provide a complementary rigidity theory. For symmetric Kubo–Ando means on \(\mathcal B(H)\), the order-determining property of the norm implies that norm comparisons
\[
\|A\sigma X\|\le \|B\sigma X\|,\qquad \forall X>0,
\]
recover Löwner order [2301.06355]. This order-theoretic fact feeds into nonlinear preserver results: if \(\mathcal A,\mathcal B\) are AW\(^*\)-algebras and \(\phi:\mathcal A^{++}\to\mathcal B^{++}\) is a surjective map preserving the norm of a symmetric Kubo–Ando mean, then \(\phi\) extends to a Jordan \(*\)-isomorphism between the ambient algebras [2412.03094].

These results show that symmetric Kubo–Ando means occupy a narrow algebraic corridor. Weak associativity admits a large family, but additional arithmetic or scaling constraints collapse that family sharply; positivity-cone preservation singles out arithmetic means; and norm-preserving symmetries are forced to be Jordan \(*\)-isomorphisms. A plausible implication is that many apparently analytic properties of Kubo–Ando means are in fact disguised algebraic rigidity statements.

## 7. Spectral decomposition, division-algebra correspondences, and algorithmic directions

Recent work has made the spectral structure of Kubo–Ando means more explicit. If \(\sigma\) has representing function \(f\) and \(A,B\in\mathscr P_n(\mathbb D)\) for \(\mathbb D\in\{\mathbb R,\mathbb C,\mathbb H\}\), then \(A\sigma B\) can be written as a finite linear combination of powers of \(A^{-1}B\). More precisely, if \(r=|\sigma(A^{-1}B)|\), then
\[
A\sigma B = \sum_{i=0}^{r-1} c_i (A^{-1}B)^i,
\]
where the coefficients \(c_i\) depend only on \(f\) and the eigenvalues of \(A^{-1}B\). In dimension \(3\), these coefficients can be expressed explicitly through spectral invariants, leading to closed-form decompositions and affine characterizations of the linearizable case [2606.13530].

A related line of work establishes correspondences of Kubo–Ando means over the three real division algebras. Canonical embeddings identify means on \(\mathscr P_n(\mathbb H)\), \(\mathscr P_{2n}(\mathbb C)\), and \(\mathscr P_{4n}(\mathbb R)\), preserving functional calculus, congruence invariance, and the Log–Euclidean metric structure. As an application, every Kubo–Ando mean on \(\mathscr P_2(\mathbb D)\) admits an explicit affine expression
\[
A\sigma B = \alpha_f(X)A+\beta_f(X)B,
\qquad
X=A^{-1/2}BA^{-1/2},
\]
with coefficients determined by the eigenvalues, hence by trace–determinant data of \(X\). This yields explicit formulas for the geometric mean in real, complex, and quaternionic \(2\times2\) settings [2605.27707].

The Kubo–Ando geometric mean has also entered a new majorization and matrix-inequality landscape. For \(A,B\in\mathbb P_n\) and \(a,b\ge0\), one has
\[
\lambda\bigl(a^2A+b^2B+2ab(A\# B)\bigr)
\prec_w
\lambda\bigl(W_{a,b}(A,B)\bigr),
\]
where \(W_{a,b}(A,B)\) is the weighted Bures–Wasserstein expression. This yields norm inequalities for all unitarily invariant norms and refines a two-variable Heron inequality of Bhatia–Lim–Yamazaki [2605.26141].

There is now also an algorithmic direction. Quantum algorithms based on block-encodings, QSVT, and Löwner–Stieltjes or harmonic-mixture representations can compute general Kubo–Ando means
\[
A\sigma_f B
=
A^{1/2} f(A^{-1/2}BA^{-1/2})A^{1/2}
\]
for operator monotone \(f\), with complexity bounds expressed in terms of block-encoding costs, condition numbers, and polylogarithmic dependence on precision. The same framework covers maximal quantum \(f\)-divergences, highlighting a common operator-functional backbone for means, divergences, and entropy-like quantities [2511.10607].

These developments indicate that the modern theory of Kubo–Ando connections now has at least four interacting faces: an axiomatic face built on order and congruence; a geometric face built on barycenters, divergences, and positivity cones; a spectral-combinatorial face built on explicit decompositions and majorization; and an algorithmic face built on resolvent representations and quantum linear-algebra primitives.

Source: https://www.emergentmind.com/topics/kubo-ando-connections