🔢 mathematics

Alternative-mean trace divergences: geometry, data processing, and barycenters

This paper introduces a new class of alternative-mean trace functionals defined via operator monotone functions, proving that they constitute quantum divergences inducing the Bures–Wasserstein metric and satisfying sharp quadratic bounds relative to this metric.

Original authors: Trung Dung Vuong, Hiroki Shudo, Hiroyuki Osaka

Published 2026-09-02

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Trung Dung Vuong, Hiroki Shudo, Hiroyuki Osaka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quantum world, where particles exist in states of probability rather than definite positions, scientists rely on a special kind of geometry to measure how different two states are from one another. Imagine trying to find the distance between two points on a curved surface; the path you take matters, and the surface itself might stretch or shrink depending on how you look at it. In quantum physics, this surface is made of positive definite matrices, which are complex grids of numbers that describe the state of a system. To navigate this landscape, researchers use tools called "means" to average these states and "divergences" to measure the distance between them. One particularly important distance measure, known as the Bures–Wasserstein distance, acts like a ruler that respects the unique curvature of this quantum space. It is essential for understanding how information flows through quantum systems and how to find a central point, or "barycenter," among a collection of quantum states. However, not all ways of measuring distance or averaging states behave well when the system is observed or manipulated. A key question has been determining which mathematical tools remain reliable when a quantum state is processed or simplified, a property known as data processing.

A team of mathematicians has now mapped out a vast new territory of these distance measures, revealing exactly which ones are robust and which ones fail under specific conditions. They introduced a family of new distance formulas based on "alternative means," a method of averaging that involves a specific type of smooth, curved function. By analyzing these formulas, the researchers proved that every member of this family behaves like a valid quantum distance measure, meaning it is always non-negative and zero only when the two states are identical. More importantly, they showed that the local shape of the space defined by these new distances is directly linked to the established Bures–Wasserstein geometry. This connection means that the new tools are not just abstract inventions but are deeply rooted in the known structure of quantum space, scaling the familiar distance by a specific factor determined by the function used.

The study then tackled a critical test of reliability: the data processing inequality. In simple terms, this principle states that if you take a quantum state and process it through a physical operation, the distance between two states should never increase; information cannot be created by processing, only lost or preserved. The researchers discovered that for their new family of distances, this rule holds true for all possible physical operations only in a very narrow and specific set of cases. They found that the only functions that guarantee this reliability are those that can be described by a simple quadratic relationship, effectively reducing the new distance to a constant multiple of the standard Bures–Wasserstein distance. To prove this, they demonstrated that if a function falls outside this specific class, the rule breaks down even in the simplest possible scenario: a two-level quantum system, or qubit, when its off-diagonal information is removed. This finding settles a long-standing debate by showing that the promise of data processing is not a general feature of these alternative means but a rare and rigid property.

Beyond measuring distance, the paper also solved the problem of finding the "center" of a group of quantum states, known as the barycenter problem. When given a set of states, one often wants to find a single state that best represents the group, much like finding the average location of a set of cities. The authors proved that for any collection of quantum states, a best representative always exists, even if we allow the answer to include states on the boundary of the possible space. They derived the precise mathematical condition that this center must satisfy and identified exactly when this center is unique and well-behaved. For a popular family of functions based on powers, they determined that a unique, well-behaved center exists if and only if the power is less than or equal to one-half. If the power is higher, the problem becomes unstable, and multiple different centers can exist for the same set of data, even in the simplest two-dimensional cases. This work provides a complete guide for scientists on which mathematical tools to use for averaging and measuring in quantum systems, ensuring that their calculations remain physically meaningful and mathematically sound.

Technical Summary: Alternative-Mean Trace Divergences

1. Problem Statement

The paper investigates a class of functionals defined on the cone of positive definite matrices, $P_m$ , termed alternative-mean trace divergences. These divergences are constructed using a nontrivial normalized operator monotone function $f: (0, \infty) \to (0, \infty)$ and the geometric mean $A \sharp B$ . Specifically, the functional is defined as:
$\tilde{\Phi}_f(A, B) := \text{Tr}(A \nabla_s B) - \text{Tr}\left( f(A^{-1} \sharp B) A f(A^{-1} \sharp B) \right),$
where $s = f'(1)$ and $A \nabla_s B = (1-s)A + sB$ is the weighted arithmetic mean.

The central problems addressed are:

Geometric Characterization: How does $\tilde{\Phi}_f$ relate to the squared Bures–Wasserstein distance ( $d_{BW}^2$ ), and does it induce a valid Riemannian metric structure?
Data Processing Inequality (DPI): Under what conditions on $f$ does $\tilde{\Phi}_f$ satisfy the exact data-processing inequality under all positive trace-preserving (PTP) maps?
Barycenter Existence and Uniqueness: For a given set of data points in $P_m$ , does the minimization problem associated with $\tilde{\Phi}_f$ admit a unique global minimizer (barycenter), and under what conditions is this minimizer strictly positive definite?

2. Methodology

The authors employ a combination of operator theory, matrix analysis, and optimization techniques:

Operator Monotonicity and Functional Calculus: The analysis relies heavily on the properties of normalized operator monotone functions, specifically their representation via the Kubo–Ando integral formula involving a probability measure $\mu_f$ .
Riccati Equations and Geometric Means: The geometric mean $A \sharp B$ is treated as the unique positive definite solution to the Riccati equation $RAR = B$ . This allows the authors to express the divergences in terms of the spectrum of $R = A^{-1} \sharp B$ .
Taylor Expansions and Hessian Analysis: To establish the Riemannian structure, the authors perform second-order expansions of the functional near the diagonal ( $A=B$ ) to compute the diagonal Hessian and compare it with the Bures–Wasserstein metric.
Perturbation and Pinching Arguments: To characterize the data-processing inequality, the authors utilize diagonal pinching maps on qubit states (2x2 matrices). They construct specific counterexamples using real-symmetric faithful qubit states to test the failure of the inequality.
Convexity and Matrix Concavity: The barycenter problem is analyzed by studying the convexity of the objective function $E_f(X) = \sum w_j \tilde{\Phi}_f(A_j, X)$ . The authors introduce "residual certificates" involving the function $G(x) = f(x)^2 - cx^2$ to establish strict convexity and boundary behavior.

3. Key Contributions and Results

A. Geometric Comparison and Quantum Divergence

The paper establishes a sharp global two-sided comparison between the alternative-mean divergence and the squared Bures–Wasserstein distance.

Inequality: For any nontrivial normalized operator monotone function $f$ with $s=f'(1)$ , the following holds for all $A, B \in P_m$ :
$s(1-s) d_{BW}(A, B)^2 \leq \tilde{\Phi}_f(A, B) \leq (1 - s + s^2) d_{BW}(A, B)^2.$
The constants are optimal uniformly over the class of functions with a fixed derivative $s$ .
Quantum Divergence Status: The diagonal Hessian of $\tilde{\Phi}_f$ at $A=B$ is shown to be $2\gamma_f g_{BW}^A$ , where $\gamma_f = s(1-s) - f''(1) > 0$ and $g_{BW}^A$ is the Bures–Wasserstein metric. Consequently, $\tilde{\Phi}_f$ is a quantum divergence in the sense of Bhatia–Gaubert–Jain (smooth, nonnegative, vanishing only on the diagonal, with a positive definite Hessian).

B. Data Processing Rigidity

The paper provides a complete classification of functions $f$ for which $\tilde{\Phi}_f$ satisfies the exact data-processing inequality (DPI) under all PTP maps.

Equivalence of Classes: The condition for exact DPI under all PTP maps is equivalent to the condition under all completely positive trace-preserving (CPTP) maps.
Characterization: Exact DPI holds if and only if $f(x)^2$ is a polynomial of degree at most two. Specifically, $f$ must be of the form:
$f(x) = \sqrt{((1-u) + ux)((1-v) + vx)}, \quad u, v \in [0, 1].$
Consequence: In this class, $\tilde{\Phi}_f$ is a constant multiple of the squared Bures–Wasserstein distance.
Testing: The validity of the DPI is completely determined by testing diagonal pinching on faithful qubit states. If the inequality fails, it fails for a pair of real-symmetric faithful qubit states under diagonal pinching.
Power Family Application: For the power family $f_t(x) = x^t$ , exact DPI holds if and only if $t = 1/2$ . For $t \neq 1/2$ , the inequality fails under diagonal pinching.

C. Barycenter Theory

The paper develops the theory for the barycenter problem $\min_{X \succeq 0} \sum w_j \tilde{\Phi}_f(A_j, X)$ .

Existence: A global minimizer (closed-cone minimizer) always exists for any finite data set.
Scalar Uniqueness: For scalar data ( $m=1$ ), the minimizer is unique and lies strictly within the range of the data values, regardless of $f$ .
Positive Definiteness and Uniqueness:
- The authors introduce a residual certificate: If $G(x) = f(x)^2 - cx^2$ is nonnegative, nonconstant, and matrix-concave of order $m$ , then the minimizer is unique and strictly positive definite.
- If $G(x)$ is operator monotone (dimension-free certificate), the uniqueness and positive definiteness hold for all matrix sizes.
Sharpness for Power Families: For the power family $f_t(x) = x^t$ $f_{t} (x) = x^{t}$ , the property that every data set admits a unique positive definite minimizer holds exactly when $0 < t \leq 1/2$ $0 < t \leq 1/2$ .
- For $1/2 < t < 1$ , the paper constructs a counterexample using two commuting $2 \times 2$ positive definite matrices where the minimizer is not unique (specifically, there are at least two distinct minimizers related by a symmetry transformation).

4. Significance and Claims

The paper claims to extend the understanding of quantum divergences beyond the standard Kubo–Ando and perspective-based constructions.

Geometric Insight: It demonstrates that a broad class of alternative-mean divergences are intrinsically linked to the Bures–Wasserstein geometry, serving as global approximations with controlled error bounds.
Rigidity of Data Processing: The work highlights a "rigidity" phenomenon where the requirement for exact data processing under general PTP maps forces the divergence to collapse to a scalar multiple of the Bures–Wasserstein distance. This contrasts with other divergence families where a wider range of parameters might satisfy monotonicity.
Sharpness of Barycenter Conditions: The paper provides a definitive answer for the power family regarding the existence of unique positive definite barycenters, showing that the range $0 < t \leq 1/2$ is not just sufficient but necessary. It explicitly demonstrates that non-uniqueness can arise even for simple commuting data sets when $t > 1/2$ .

The authors note that while they have characterized the exact-DPI class and the power-family barycenter behavior, the full characterization of all normalized operator monotone functions $f$ that guarantee unique positive definite barycenters for arbitrary data sets in every matrix size remains an open problem.