Audenaert's Subadditivity Inequality

• Audenaert's subadditivity inequality is a collection of matrix and quantum information inequalities—ranging from partial-trace norm to concave-function and strong subadditivity formulations—that relate global and marginal quantities.
• It leverages properties of concave and convex functions, eigenvalue majorization, and t-geometric means to provide fidelity-based and entropy refinement bounds in operator norm contexts.
• The framework is extended to multipartite and non-positive matrices, enabling novel applications in entanglement detection, Schmidt number witnesses, and operator-order inequalities.

Searching arXiv for recent and foundational papers on Audenaert's subadditivity inequality and related operator/entropy refinements. Audenaert’s subadditivity inequality denotes a family of closely related matrix and quantum-information inequalities rather than a single universally standardized statement. In its original partial-trace form, for a positive semidefinite bipartite matrix $M$ and Schatten $p$-norms, it asserts

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$

and this inequality is used to prove subadditivity of $q$-entropies for $q>1$. Closely related literature uses the same label for norm subadditivity of concave matrix functions, for commuting-product inequalities involving geometric means, and for strong-subadditivity-type trace functionals. This suggests a unifying theme: subsystem, compression, or marginal quantities are controlled by a global quantity, often with a correction term that is sharp in specific regimes (Rico et al., 24 Jul 2025, Audenaert et al., 2010, Audenaer et al., 2010).

1. Terminological scope and canonical formulations

A recurring source of ambiguity is that “Audenaert’s subadditivity inequality” is used in several nearby but non-identical senses. The literature represented here contains three especially important forms.

Usage	Representative statement	Source
Partial-trace norm inequality	$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$ for $M\ge 0$	(Rico et al., 24 Jul 2025)
Concave-function norm subadditivity	$\|f(A+B)\|\le \|f(A)+f(B)\|$ for $A,B\ge 0$, $f\ge 0$ concave, and unitarily invariant norms	(Audenaert et al., 2010)
Strong-subadditivity trace functional	$p$0	(Audenaer et al., 2010)

In the partial-trace formulation, $p$1 is positive semidefinite, $p$2, $p$3, and $p$4 denotes the Schatten $p$5-norm. No trace-one normalization is required, and the inequality is homogeneous in $p$6. In the context described in (Rico et al., 24 Jul 2025), it is used to prove subadditivity of $p$7-entropies.

In the concave-function formulation, the inequality is part of the Bourin–Uchiyama / Audenaert–Aujla theory of unitarily invariant norm inequalities. In the strong-subadditivity formulation, the object is not a bipartite state but a positive block operator or a compressed operator, and the function $p$8 is required to belong to a class with strong monotonicity or concavity properties.

The terminology therefore should be read contextually. In matrix analysis, it often means norm subadditivity for concave functions or commuting products; in quantum information, it may refer to entropy-type refinements or to strong-subadditivity patterns for trace functionals.

2. Norm subadditivity for concave and convex functions

A central Audenaert-type inequality is the norm comparison

$p$9

valid for positive semidefinite $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$0, non-negative concave $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$1, and every unitarily invariant norm. In eigenvalue form this is

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$2

so the inequality is controlled by Ky Fan majorization rather than by a specific norm alone (Audenaert et al., 2010).

The dual statement is the convex superadditivity inequality. If $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$3 is non-negative, convex, increasing, and $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$4, then

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$5

for all positive semidefinite $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$6 and all unitarily invariant norms. A standard special case is the power function: for $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$7,

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$8

for all Schatten $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1,\qquad p\in[1,\infty],$$9-norms (Audenaert et al., 2010).

The 2010 analysis in (Audenaert et al., 2010) gives a new proof of the concave subadditivity inequality through dominated majorization. The key construction is the vector $q$0, defined by directional derivatives of sums of the $q$1 largest eigenvalues: $q$2 A central formula states that there exists a proper eigenbasis $q$3 of $q$4 such that

$q$5

This reduces a nonlinear eigenvalue comparison to a linear comparison of diagonal entries in the eigenbasis of the dominating matrix (Audenaert et al., 2010).

That same paper also marks an important limitation. It presents counterexamples to conjectures that Ando’s inequality for operator convex functions could more generally hold for ordinary convex, non-negative functions. The positive “sum” inequalities therefore extend from operator-monotone or operator-concave settings to scalar concavity and convexity, but analogous “difference” inequalities do not extend so broadly.

3. Commuting products and $q$6-geometric means

Another standard usage of “Audenaert’s inequality” is a commuting-product norm inequality. If $q$7 and $q$8 for each $q$9, then for any unitarily invariant norm,

$q>1$0

In (Hoa, 2015) this is identified as a special case of a more general inequality for $q>1$1-geometric means.

For positive semidefinite matrices $q>1$2, $q>1$3, $q>1$4, and any unitarily invariant norm,

$q>1$5

Here

$q>1$6

is the weighted geometric mean, extended by continuity to singular positive semidefinite matrices (Hoa, 2015).

The Audenaert inequality is recovered by setting $q>1$7, $q>1$8, and using the commuting assumption $q>1$9. In that case

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$0

The 2015 result therefore places Audenaert’s inequality inside a broader operator inequality derived from three ingredients: Bourin–Uchiyama’s matrix subadditivity inequality for convex functions, concavity of the $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$1-geometric mean, and a Hiai–Ando log-majorization inequality (Hoa, 2015).

This generalization is structurally significant because the theorem itself requires no commutativity. Commutativity enters only when one identifies $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$2 with $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$3 and thus recovers Audenaert’s original commuting-product formulation.

4. Entropic refinements and fidelity-based remainder terms

In quantum information, subadditivity is usually the von Neumann entropy inequality

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$4

with equality if and only if $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$5. The gap

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$6

is the mutual information, and

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$7

is a relative entropy (Carlen et al., 2014).

The quantitative version emphasized in (Carlen et al., 2014) is a remainder term for subadditivity. The paper proves the Rényi-type lower bound

$$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$8

and specializing to $$\|\operatorname{tr}_A[M]\|_p+\|\operatorname{tr}_B[M]\|_p\le \|M\|_p+\|M\|_1$$9, $M\ge 0$0 gives a fidelity-based lower bound for the mutual information. The same paper also recalls the Pinsker inequality

$M\ge 0$1

so that

$M\ge 0$2

The authors emphasize that the Pinsker-based bound is not sharp except in the small-distance regime, while the fidelity-based bound is sharp in highly correlated cases (Carlen et al., 2014).

Audenaert is not mentioned by name in (Carlen et al., 2014), but the paper explicitly fits the pattern of an Audenaert-style refinement: it quantifies deviation from equality in subadditivity, uses an overlap or fidelity-like quantity between $M\ge 0$3 and $M\ge 0$4, and gives a non-trivial remainder term that vanishes precisely on product states. The two-particle fermionic reduced-state example shows exactness of the Rényi-based bound, whereas the Pinsker estimate remains much smaller (Carlen et al., 2014).

This part of the literature makes clear that “subadditivity inequality” in the Audenaert orbit is not restricted to norm inequalities. It also includes stability statements: if the entropy gap is small, then the state is close to a product state in a fidelity sense; conversely, if the overlap with the product state is small, the mutual information must be large.

5. Strong subadditivity, trace functionals, and operator extensions

A more structural generalization is the notion of a strongly subadditive function. Let $M\ge 0$5 be written in $M\ge 0$6 block form with respect to a decomposition $M\ge 0$7, and let

$M\ge 0$8

Then $M\ge 0$9 is called strongly subadditive if

$\|f(A+B)\|\le \|f(A)+f(B)\|$0

for all such $\|f(A+B)\|\le \|f(A)+f(B)\|$1. A sufficient condition is the theorem of Hiai–Petz–Audenaert: if $\|f(A+B)\|\le \|f(A)+f(B)\|$2 is differentiable and $\|f(A+B)\|\le \|f(A)+f(B)\|$3 is matrix monotone on $\|f(A+B)\|\le \|f(A)+f(B)\|$4, then $\|f(A+B)\|\le \|f(A)+f(B)\|$5 satisfies this strong subadditivity inequality (Audenaer et al., 2010).

Examples listed in (Audenaer et al., 2010) include $\|f(A+B)\|\le \|f(A)+f(B)\|$6, $\|f(A+B)\|\le \|f(A)+f(B)\|$7, $\|f(A+B)\|\le \|f(A)+f(B)\|$8 for $\|f(A+B)\|\le \|f(A)+f(B)\|$9, $A,B\ge 0$0 for $A,B\ge 0$1, and $A,B\ge 0$2 for $A,B\ge 0$3. Equality is described explicitly in several cases. For $A,B\ge 0$4, equality holds if and only if

$A,B\ge 0$5

for invertible $A,B\ge 0$6. For $A,B\ge 0$7, equality holds if and only if $A,B\ge 0$8 (Audenaer et al., 2010).

The operator extension of strong subadditivity in (Kim, 2012) lifts the scalar entropy inequality to an operator inequality. For a tripartite state $A,B\ge 0$9, with $f\ge 0$0 acting on the full space by tensoring with identities,

$f\ge 0$1

Taking the full trace recovers the usual strong subadditivity of von Neumann entropy (Kim, 2012).

The 2025 reanalysis in (Luijk et al., 30 Jul 2025) identifies the mathematical structure behind this operator inequality as Connes’ theory of spatial derivatives. In the finite-dimensional setting the basic operator statement is

$f\ge 0$2

with $f\ge 0$3 full rank. In the general setting of an inclusion $f\ge 0$4 of von Neumann algebras, the corresponding inequality is

$f\ge 0$5

In the case of standard representations, this reduces to monotonicity of the relative modular operator (Luijk et al., 30 Jul 2025).

These strong-subadditivity results are not identical to the original partial-trace norm inequality, but they belong to the same conceptual family. They replace a scalar norm comparison by block-matrix, operator-valued, or modular-theoretic control, and they clarify the equality structure through Markov-type conditions.

6. Operator-order and multipartite generalizations

Later work pushes Audenaert-type inequalities in two directions: operator order under spectral assumptions, and norm inequalities beyond positivity.

For strictly positive operators $f\ge 0$6 satisfying the spectral disjointness conditions

$f\ge 0$7

the power inequalities

$f\ge 0$8

and

$f\ge 0$9

hold in operator order. The paper presenting these inequalities does not mention Audenaert by name, but it explicitly places them in the same family as subadditivity and superadditivity inequalities for powers of positive operators, and stresses that they are stronger than earlier norm versions because they hold at the operator level (Moradi et al., 2019).

A different extension removes positivity. For an arbitrary matrix $p$00, with partial traces

$p$01

the 2025 multipartite generalization proves

$p$02

For $p$03, this has exactly the form of Audenaert’s original inequality, but now it holds for arbitrary, possibly non-positive, matrices. If $p$04, the rank-refined estimate is

$p$05

The central technical ingredient is a majorization relation for Kronecker sums, and the framework is extended further to arbitrary unitarily invariant norms through a template

$p$06

where $p$07 is defined through the dual symmetric gauge and a vector built from local singular-value data (Rico et al., 24 Jul 2025).

The applications in (Rico et al., 24 Jul 2025) show that these are not merely formal generalizations. The paper proves that Werner states are $p$08-copy undistillable for all $p$09, and it derives new Schmidt-number witnesses and $p$10-positive maps from the rank-refined inequalities. Those applications indicate that the modern scope of Audenaert-type subadditivity now includes partial traces of general complex matrices, multiple tensor factors, rank-sensitive bounds, and entanglement-theoretic consequences (Rico et al., 24 Jul 2025).

Taken together, these developments show that Audenaert’s subadditivity inequality is best understood as a cluster of results centered on a single structural idea: subadditivity is not only a scalar statement about entropy or norms, but also a matrix inequality with commuting-product, concave-functional, fidelity-based, block-compression, operator-order, and modular-theoretic realizations.