---
title: Umegaki Relative Entropy in Quantum Information
url: https://www.emergentmind.com/topics/umegaki-relative-entropy
type: topic
---

# Umegaki Relative Entropy in Quantum Information

Searching arXiv for recent and foundational papers on Umegaki relative entropy to ground the article in cited literature.
Umegaki relative entropy is the standard quantum analogue of Kullback–Leibler divergence. For density operators \(\rho\) and positive semidefinite \(\sigma\), it is defined by
\[
D(\rho\|\sigma):=\operatorname{Tr}\!\big[\rho(\log\rho-\log\sigma)\big]
\]
when \(\operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma)\), and as \(+\infty\) otherwise. In finite dimensions it governs channel coding converses, asymptotic hypothesis testing, thermodynamic irreversibility, resource conversion, and free-energy differences; in operator-algebraic language it is the type I instance of Araki relative entropy [2607.01712; 1712.07283].

## 1. Definition, finiteness, and basic operator-algebraic form

A standard formulation writes
\[
D(\rho\|\sigma):= \operatorname{Tr}\!\big[\rho(\log \rho-\log \sigma)\big]
\quad\text{if}\quad \sigma\gg \rho,
\]
with \(D(\rho\|\sigma)=+\infty\) if \(\sigma\not\gg\rho\), where \(\sigma\gg\rho\) means
\[
\operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma).
\]
Equivalent formulations in the cited literature use \(\ker \sigma \subseteq \ker \rho\), or \(\rho\ll \sigma\), to express the same support condition [1512.02615; 2305.10140].

In finite-dimensional uniqueness results, the same quantity is written as
\[
D(\rho\|\sigma)=\operatorname{Tr}[\rho\log\rho]-\operatorname{Tr}[\rho\log\sigma],
\]
with \(\log\) taken in base \(2\), so the divergence is measured in bits [2607.01712]. In type I operator-algebraic settings, Araki relative entropy reduces to the same Umegaki formula,
\[
S_{\mathfrak M}(\omega\Vert\phi)=\operatorname{Tr}\rho(\log\rho-\log\sigma),
\]
which is the finite-dimensional prototype for the general von Neumann algebraic notion [2605.19106; 1712.07283].

The finiteness condition is structural rather than cosmetic. It is what allows the logarithm of \(\sigma\) to be evaluated on the support of \(\rho\), and it is also the point at which many continuity, variational, and extension theorems bifurcate into faithful and non-faithful regimes. A recurring theme in recent work is that weakening support assumptions often remains technically delicate or explicitly open [1512.02615].

## 2. Variational formulas and the gap to measured relative entropy

A defining feature of Umegaki relative entropy is that, unlike trace distance or fidelity, it is not generally recovered by optimizing a classical divergence over measurements. For a POVM \(M\) with outcome set \(X\), the induced distributions are
\[
P_{\rho,M}(x)=\operatorname{tr}[M(x)\rho],\qquad
P_{\sigma,M}(x)=\operatorname{tr}[M(x)\sigma],
\]
and the measured relative entropy is
\[
D^M(\rho\|\sigma):=\sup_{(X,M)} D\big(P_{\rho,M}\|P_{\sigma,M}\big).
\]
The projectively measured version is
\[
D^P(\rho\|\sigma):=\sup_{\{P_i\}}
\sum_i \operatorname{tr}[P_i\rho]\log\frac{\operatorname{tr}[P_i\rho]}{\operatorname{tr}[P_i\sigma]}.
\]
A key theorem shows
\[
D^P(\rho\|\sigma)=D^M(\rho\|\sigma),
\]
so optimizing over all finite-outcome POVMs gives no advantage over projective measurements. More importantly,
\[
D^M(\rho\|\sigma)\le D(\rho\|\sigma),
\]
and under \(\rho>0,\sigma>0\) equality holds if and only if \([\rho,\sigma]=0\). Thus noncommutativity produces a strict gap:
\[
[\rho,\sigma]\neq 0 \;\Longrightarrow\; D^M(\rho\|\sigma)<D(\rho\|\sigma).
\]
This sharpens Petz’s original conclusion by showing that even arbitrary POVMs do not recover the full Umegaki entropy [1512.02615].

The proof is organized around parallel variational formulas. For measured relative entropy,
\[
D^M(\rho\|\sigma)=1+\max_H \Bigl(\operatorname{tr}[\rho H]-\operatorname{tr}[\sigma e^H]\Bigr),
\]
while for Umegaki relative entropy,
\[
D(\rho\|\sigma)=1+\sup_H \Bigl(\operatorname{tr}[\rho H]-\operatorname{tr}[e^{\log\sigma+H}]\Bigr).
\]
The distinction lies in
\[
\operatorname{tr}[e^{\log\sigma+H}]
\quad\text{versus}\quad
\operatorname{tr}[\sigma e^H].
\]
If \(\sigma\) and \(H\) commute, the terms coincide; in general, Golden–Thompson gives
\[
\operatorname{tr}[e^{\log\sigma+H}] \le \operatorname{tr}[\sigma e^H],
\]
which makes the Umegaki objective at least as large as the measured one. The strict gap is therefore an operator-theoretic manifestation of noncommutativity, and it formalizes the statement that some quantum distinguishability is lost under measurement [1512.02615].

## 3. Operational characterizations and uniqueness

The data-processing inequality does not single out Umegaki relative entropy: many divergences satisfy DPI, including Petz Rényi, sandwiched Rényi, max-relative entropy, and hypothesis-testing divergence. A stronger criterion is provided by optimal binary discrimination. For priors \(t\in[0,1]\), the optimal guessing probability is
\[
P_{\rm guess}^{(t)}(\rho,\sigma)=
\max_{0\le \Lambda\le I}
\Bigl\{
t\,\operatorname{Tr}(\Lambda\rho)+(1-t)\operatorname{Tr}((I-\Lambda)\sigma)
\Bigr\}.
\]
This induces the quantum Lorenz preorder \((\rho,\sigma)\succ_L(\rho',\sigma')\), meaning that \((\rho,\sigma)\) performs at least as well in every such binary game. A quantum Lorenz divergence is any scalar \(D\) monotone under this preorder. The central uniqueness theorem states: if \(D\) is a normalized additive QLD whose classical restriction is Lorenz continuous, then for every finite-dimensional pair with \(\operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma)\),
\[
D(\rho\|\sigma)=\operatorname{Tr}[\rho\log\rho]-\operatorname{Tr}[\rho\log\sigma].
\]
In this sense, Umegaki relative entropy is the unique normalized additive distinguishability measure compatible with the odds revealed by optimal binary measurements. The collapse to \(\alpha=1\) is explicitly described as a purely quantum consequence of noncommutativity: classically, additive monotones remain a continuum of Rényi mixtures [2607.01712].

This operational viewpoint complements more traditional asymptotic statements. Umegaki relative entropy is described as having “operational significance as the threshold rate for asymmetric binary quantum hypothesis testing” [1512.02615]. A recent additivity criterion goes further by studying optimization problems of the form
\[
\mathfrak D(\rho^{\otimes n})=\min_{\sigma^{(n)}\in S_n}D(\rho^{\otimes n}\|\sigma^{(n)}).
\]
Under tensor-stability and multiplicative-polar assumptions on \(\{S_n\}\), regularization is unnecessary if and only if a single-copy optimizer \(\sigma_0\) satisfies
\[
\sup_{t\in\mathbb R,\ \tau\in S_1}
\operatorname{Tr}\!\Big[
\tau\,\sigma_0^{-\frac12(1+it)}\rho\,\sigma_0^{-\frac12(1-it)}
\Big]\le 1.
\]
When this holds,
\[
\min_{\sigma^{(n)}\in S_n}D(\rho^{\otimes n}\|\sigma^{(n)})
=
n\min_{\sigma\in S_1}D(\rho\|\sigma),
\]
so the asymptotic Umegaki quantity is decided at the single-copy level [2507.05696].

## 4. Rényi connections, geometric extensions, and continuity theory

The Umegaki entropy sits at \(\alpha=1\) in several Rényi families. In particular,
\[
D_1(\rho\|\sigma)=D(\rho\|\sigma).
\]
The comparison between measured and fully quantum relative entropies extends to Rényi divergences: for noncommuting full-support states, the sandwiched Rényi relative entropy is strictly larger than the measured Rényi relative entropy for \(\alpha\in(\tfrac12,\infty)\), and strictly smaller for \(\alpha\in[0,\tfrac12)\). Thus the Umegaki case belongs to a larger strict-gap phenomenon driven by noncommutativity [1512.02615].

A different extension starts from Umegaki as a seed. Using Kubo–Ando weighted geometric means, one defines
\[
D^{Um,\#_\gamma}(\rho\|\sigma)
=
\frac{1}{1-\gamma}\,D_U(\rho\|\sigma\#_\gamma \rho),
\qquad \gamma\in(0,1).
\]
For full-rank \(\rho,\sigma\), this family is monotone increasing in \(\gamma\), satisfies
\[
D^{Um,\#_\gamma}(\rho\|\sigma)\searrow D_U(\rho\|\sigma)
\quad (\gamma\searrow 0),
\]
and
\[
D^{Um,\#_\gamma}(\rho\|\sigma)\nearrow D^{\max}(\rho\|\sigma)
\quad (\gamma\nearrow 1).
\]
In the barycentric Rényi construction, choosing Umegaki in all slots reproduces the log-Euclidean Rényi divergence in the two-variable case [2207.14282].

Continuity theory for \(D(\rho\|\sigma)\) is necessarily support-sensitive. For strictly positive finite-dimensional states, one upper bound is
\[
S(\rho\|\sigma)
\le
\|\rho-\sigma\|_1\,\lambda_\rho\,
\frac{\log(\alpha_\rho)-\log(\alpha_\sigma)}{\alpha_\rho-\alpha_\sigma}
\le
\frac{\lambda_\rho}{\alpha}\,\|\rho-\sigma\|_1,
\]
where \(\lambda_\rho\) is the largest eigenvalue of \(\rho\), \(\alpha_\rho,\alpha_\sigma\) are the smallest eigenvalues, and \(\alpha=\min\{\alpha_\rho,\alpha_\sigma\}\) [1904.00795]. A more systematic ALAFF-based theory proves almost concavity for Umegaki relative entropy, recovers almost-tight Alicki–Fannes–Winter type bounds for conditional entropy, mutual information, and conditional mutual information, and yields direct continuity estimates for \(D(\rho\|\sigma)\) in the first input, the second input under lower domination assumptions, and jointly in both inputs on appropriate support-compatible sets [2208.00922; 2305.10140].

## 5. Relative entropy to subalgebras, models, and constrained structures

One important generalization replaces the second state by a structured set. For a finite von Neumann algebra \((M,\operatorname{tr})\) and a subalgebra \(N\subset M\) with trace-preserving conditional expectation \(E:M\to N\), the relative entropy to the subalgebra is
\[
D(\rho\|N):=\inf_{\sigma\in S(N)}D(\rho\|\sigma).
\]
At \(p=1\), the minimizer is exactly the conditional expectation:
\[
D(\rho\|N)=D(\rho\|E(\rho))=H(E(\rho))-H(\rho).
\]
This is special to the Umegaki case; for general sandwiched Rényi parameters \(p\neq1\), the minimizer need not be \(E(\rho)\). In finite-dimensional or II\(_1\)-factor settings,
\[
-\log \lambda(M:N)=\sup_{\rho\in S(M)}D(\rho\|N),
\]
and for II\(_1\) subfactors this becomes
\[
\log[M:N]=\sup_{\rho\in S(M)}D(\rho\|N),
\]
identifying Jones index with maximal Umegaki relative entropy to the subalgebra [1909.01906].

A parallel role appears in many-body correlation theory. For a Gibbs family \(\mathcal E\),
\[
\mathrm d_{\mathcal E}(\rho):=\inf_{\sigma\in\mathcal E}D(\rho\|\sigma).
\]
The projection theorem gives
\[
\mathrm d_{\mathcal E}(\rho)=D(\rho\|\pi_{\mathcal E}(\rho))
=H(\pi_{\mathcal E}(\rho))-H(\rho),
\]
where \(\pi_{\mathcal E}(\rho)\) is the maximum-entropy state compatible with the relevant constraints. For \(k\)-local Gibbs families \(\mathcal E_k\), the many-body correlation quantity \(c_k\) is exactly \(\mathrm d_{\mathcal E_k}\), and irreducible \(k\)-party correlation becomes
\[
C_k(\rho)=D(\pi_{\mathcal E_k}(\rho)\|\pi_{\mathcal E_{k-1}}(\rho)).
\]
This makes Umegaki relative entropy the geometric backbone of hierarchical-model correlation measures [1406.0833].

These constrained formulations connect to several broader programs. One preliminary categorical approach shows that, when the relevant relative entropies are finite, finite-dimensional quantum relative entropy behaves as an affine functor on a category of noncommutative statistical hypotheses, satisfying a chain-rule identity under composition [2105.04059]. A distinct inferential program derives minus Umegaki relative entropy as the entropy functional for rationally updating density matrices from a prior \(\hat\varphi\), leading to posteriors of the form
\[
\hat\rho=\frac{1}{Z}\exp\!\Big(\log\hat\varphi+\sum_i\alpha_i\hat A_i\Big)
\]
under expectation-value constraints [1710.09373].

## 6. Operator-algebraic, field-theoretic, and geometric developments

In operator-algebraic and field-theoretic settings, Umegaki relative entropy reappears through Araki’s modular formulation. In type I one has
\[
S(\omega,\omega')=\operatorname{Tr}(p\log p-p\log p'),
\]
while in general von Neumann algebras
\[
S(\omega,\omega')=\langle \Omega,\log \Delta_{\omega,\omega'}\,\Omega\rangle.
\]
For chiral conformal field theory, this framework makes mutual information a relative entropy:
\[
F(A,B)=S(\omega,\omega_A\otimes_2\omega_B).
\]
For free fermions the mutual information of two disjoint intervals has an explicit cross-ratio form, and for finite-index subnets the singular adjacent-interval limit acquires a correction involving \(-\tfrac12\log\mu_{\mathcal B}\), linking entropy duality to global dimension and subfactor index [1712.07283].

A complementary CFT approach computes Rényi relative entropies by replica methods and then takes the \(\alpha\to1\) limit to obtain Umegaki relative entropy. In that setting the relative entropy is ultraviolet finite, is written as
\[
S(\rho\|\sigma)=\Delta\langle H_\sigma\rangle-\Delta S,
\]
and provides an upper bound on trace distance via Pinsker’s inequality. Explicit formulas were obtained for one-dimensional CFTs at zero and finite temperature using correlation functions and twist operators [1404.3216]. In holography, the same identity underlies the first-law relation
\[
\Delta S=\Delta\langle H\rangle
\]
for infinitesimally nearby states, while positivity and monotonicity of relative entropy become nontrivial tests of the RT/HRT prescription [1305.3182].

Recent operator-theoretic work also reexpresses Umegaki relative entropy through the distribution of observables. For spectral decompositions of \(\rho\) and \(\sigma\), the associated Nussbaum–Szkola distributions \(P,Q\) satisfy
\[
D(\rho\|\sigma)=D(P\|Q),
\]
and, under \(\operatorname{supp}\rho\subseteq\operatorname{supp}\sigma\),
\[
D(\rho\|\sigma)
=
\int_0^\infty \lambda\,\mu^{\rho,-\log\sigma}(d\lambda)
-
\int_0^\infty \lambda\,\mu^{\rho,-\log\rho}(d\lambda).
\]
This gives a finite- and infinite-dimensional realization of Umegaki entropy as a difference of first moments of the distributions of \(-\log\sigma\) and \(-\log\rho\) under the state \(\rho\) [2203.01964].

A further geometric development arises from modular self-duality. For an antiunitary involution \(J\) and a one-parameter family \(\rho(g)\), the natural comparison functional at a modularly self-dual point is the symmetrized Umegaki relative entropy
\[
\mathfrak S_J(g)
=
S(\rho(g)\|\rho_J(g))
+
S(\rho_J(g)\|\rho(g)),
\qquad \rho_J(g)=J\rho(g)J.
\]
At the self-dual point \(g_\star\),
\[
\mathfrak S_J(g_\star)=0,\qquad \partial_g \mathfrak S_J(g)|_{g=g_\star}=0,
\]
and the Hessian is governed by Bogoliubov–Kubo–Mori geometry:
\[
I_J(g_\star)=8\,F_{\mathrm{BKM}}(g_\star).
\]
This identifies the symmetrized Umegaki entropy as the finite-dimensional prototype of a broader modular comparison principle extending to type III algebras via symmetrized Araki relative entropy [2605.19106].

Umegaki relative entropy therefore occupies several simultaneously active roles: it is the canonical noncommutative extension of classical relative entropy, the unique additive scalar compatible with full binary-testing geometry, the quantity that remains strictly larger than any measurement-optimized classical KL divergence on noncommuting pairs, and the analytic core of a wide range of constructions involving subalgebras, hierarchical models, additivity problems, conformal field theory, and modular geometry [1512.02615; 2607.01712].

Source: https://www.emergentmind.com/topics/umegaki-relative-entropy