---
title: Uhlmann Theorem for Divergences
url: https://www.emergentmind.com/topics/uhlmann-theorem-for-divergences
type: topic
---

# Uhlmann Theorem for Divergences

The Uhlmann theorem for divergences generalizes Uhlmann's classical result on quantum fidelity to a broad class of quantum divergences, revealing deep connections between extension properties of quantum states, data-processing inequalities, and optimization characterizations for quantum divergences—including the Umegaki relative entropy, sandwiched Rényi relative entropies, and measured $f$-divergences. These generalizations yield unified frameworks for quantum information-theoretic analyses, enabling strong converse results, semidefinite programming (SDP) relaxations, and sharp characterizations in quantum hypothesis testing, channel discrimination, and resource theory.

## 1. Classical Uhlmann Theorem and Relative Entropy

A. Uhlmann's original theorem concerns the monotonicity of quantum relative entropy under completely positive, trace-preserving (CPTP) maps. For states $\omega, \nu$ on a unital C*-algebra $\mathfrak{A}$ and a CPTP map $\Phi: \mathfrak{A} \to \mathfrak{B}$, the relative entropy $S(\omega\|\nu)$ satisfies
\[
S(\omega \circ \Phi \| \nu \circ \Phi) \leq S(\omega \| \nu)
\]
with equality if $\Phi$ is a $*$-homomorphism. For density matrices $\rho, \sigma > 0$, this reduces to
\[
S(\rho \| \sigma) = \mathrm{Tr}[\rho (\log \rho - \log \sigma)]
\]
Key features of $S(\rho \| \sigma)$ are non-negativity, joint convexity, lower semicontinuity in the weak-$*$ topology, and monotonicity under CPTP maps—the "data-processing inequality" [2210.13093].

Uhlmann's proof leverages the interpolation of positive quadratic forms using the Pusz–Woronowicz functional calculus. The data-processing property emerges via form-monotonicity and operator interpolation, and the result underpins major constraints in quantum information processing, including resource theories, error correction, and entropy inequalities.

## 2. Generalizations to $f$-Divergences and Operator Convexity

Uhlmann's analytic framework extends beyond Umegaki relative entropy to a family of $f$-divergences, constructed from operator-convex functions $f$ on $\mathbb{R}_+$. The Petz quasi-entropy,
\[
S_f(\rho \| \sigma) = \mathrm{Tr}\left[ \rho^{1/2} f(L_\sigma R_\rho^{-1}) (\rho^{1/2}) \right]
\]
and the class of Wigner–Yanase–Dyson-type quantities inherit data-processing inequalities under CPTP maps [2210.13093]. The connection is mediated by operator-monotonicity, variational forms, and interpolation theorems. This generalizes foundational entropy inequalities (including strong subadditivity) and allows proofs of monotonicity for a wide class of quantum divergences.

## 3. Uhlmann-Type Theorems for Rényi and Measured Divergences

The classical Uhlmann theorem for fidelity ($F(\rho, \sigma)$) finds its full quantum generalization in the context of Rényi divergences and measured $f$-divergences. For $\alpha \in [\tfrac{1}{2}, \infty]$, the sandwiched Rényi divergence is given by
\[
D_\alpha(\rho\|\sigma) = \frac{1}{\alpha-1} \log \mathrm{Tr}\left[ \left( \sigma^{\frac{1-\alpha}{2\alpha}} \rho \sigma^{\frac{1-\alpha}{2\alpha}} \right)^\alpha \right]
\]
with special cases:
- $\alpha = \tfrac{1}{2}$: min-relative entropy (fidelity),
- $\alpha \to 1$: Umegaki relative entropy,
- $\alpha \to \infty$: max-relative entropy.

These $D_\alpha$ satisfy the data-processing inequality and are non-decreasing in $\alpha$. The Uhlmann-type theorems for Rényi divergences are as follows [2502.01749]:
- **Regularized Uhlmann theorem**: For every $\rho_{AB}$, $\sigma_A$, and $\alpha \in [\tfrac{1}{2}, \infty]$,
  \[
  D_\alpha(\rho_A \| \sigma_A) = \lim_{n \to \infty} \frac{1}{n} \min_{\tau_{A^n B^n} \in \mathcal{C}_{A^n B^n}^{\sigma_A^{\otimes n}}} D_\alpha(\rho_{AB}^{\otimes n} \| \tau_{A^n B^n})
  \]
  where $\mathcal{C}_{AB}^{\sigma_A} = \{ \sigma_{AB} \geq 0: \mathrm{Tr}_B \sigma_{AB} = \sigma_A \}$.

- **Measured (single-shot) Uhlmann theorem**: For measured Rényi divergence $D_{\alpha, \mathcal{M}}(\cdot \| \cdot)$,
  \[
  D_{\alpha, \mathcal{M}}(\rho_A \| \sigma_A) \leq \min_{\tau_{AB} \in \mathcal{C}_{AB}^{\sigma_A}} D_{\alpha, \mathcal{M}}(\rho_{AB} \| \tau_{AB}) \leq D_\alpha(\rho_A \| \sigma_A)
  \]
  with equality for $\alpha = \tfrac{1}{2}$ (fidelity) and $\alpha = \infty$ (max-relative entropy).

Proofs exploit spectral pinching, operator anti-monotonicity, de Finetti/post-selection bounds, and minimax optimization via Sion's theorem. These theorems unify block-regularization and single-shot settings, providing sharp equality conditions in key quantum regimes, and forming the crux of resource-theoretic and hypothesis-testing applications [2502.01749].

## 4. Measured $f$-Divergences and Variational Formulations

Measured $f$-divergences extend these results further. For convex, lower-semicontinuous $f$ with $(0, \infty) \subseteq \mathrm{dom}(f)$, and for quantum states $\rho, \sigma$,
\[
D_f^{\rm meas}(\rho\|\sigma) = \sup_{\mathrm{POVM}\, M} \sum_x P_{\sigma, M}(x) f\left( \frac{P_{\rho, M}(x)}{P_{\sigma, M}(x)} \right)
\]
Adopting the Fenchel conjugate $f^*$, variational expressions take the form [2502.07745]:
\[
D_f^{\rm meas,proj}(\rho\|\sigma) = \sup_{\omega = \omega^\dagger,\, \mathrm{spec}(\omega) \subseteq \mathrm{dom}(f^*)} \left\{ \mathrm{Tr}[\rho \omega] - \mathrm{Tr}[\sigma f^*(\omega)] \right\}
\]
Convex reformulations are available when $f^* \circ \psi$ is operator convex for suitable $\psi$, enabling efficient convex optimization (including SDP relaxation for measured Rényi divergences with $\alpha \in (0, 1/2]$). The class of measured $f$-divergences thereby admits both analytic and computational tractability, and the measured and projective variants coincide when the above operator convexity holds [2502.07745].

## 5. Purification, Extension, and Minimax Structure

The generalized Uhlmann theorem for measured $f$-divergences formalizes the extension property: for suitable $f$ (e.g., with $f^*$ operator convex and operator monotone),
\[
\inf_{\rho_{AR}: \mathrm{Tr}_R \rho_{AR} = \rho_A} D_f^{\rm meas}(\rho_{AR} \| \sigma_{AR}) = D_f^{\rm meas}(\rho_A \| \sigma_A)
\]
and, dually, for $(f^*)^{-1}$ operator concave,
\[
\inf_{\sigma_{AR}: \mathrm{Tr}_R \sigma_{AR} = \sigma_A} D_f^{\rm meas}(\rho_{AR} \| \sigma_{AR}) = D_f^{\rm meas}(\rho_A \| \sigma_A)
\]
For $\alpha$-Rényi divergences, this structure covers all $\alpha \geq 0$, with the classical Uhlmann theorem for fidelity recovered at $\alpha = 1/2$. These identities are realized as min-max saddle-point problems, with Sion's theorem guaranteeing order exchange and compactness of the state set ensuring infimum attainment [2502.07745].

A geometric interpretation follows: for suitable $f$, measured $f$-divergences reduce (on one side) to a maximization over purifications, generalizing the overlap structure to a broader class of information measures.

## 6. Applications and Impact in Quantum Information Theory

These Uhlmann-type theorems for divergences, and the supporting variational/extension frameworks, underpin several central results and operational applications:
- **Hypothesis testing**: The regularized Rényi divergences set strong converse rates for error exponents.
- **Resource theories**: Minimal divergence from free sets, as characterized by support functionals and measured divergences, determines dilution and distillation rates, with SDP-based evaluation directly leveraging measured Uhlmann theorems.
- **Channel capacities**: SDP upper and lower bounds via measured Rényi divergences enable efficient certification of channel properties.
- **Continuity**: Support-function approximation (e.g., via post-selection) provides Fannes-type continuity bounds essential in asymptotic analyses.
- **Strong subadditivity and recovery**: Uhlmann's analytic constructs form the backbone of results such as strong subadditivity and Petz recovery, fundamentally constraining information dynamics under CPTP maps [2210.13093, 2502.07745].

A selection of central results is summarized in the following table:

| Theorem/Result                                      | Setting                                 | Divergence Type/Facts                                      |
|-----------------------------------------------------|-----------------------------------------|------------------------------------------------------------|
| Uhlmann monotonicity/data processing                | CPTP maps, relative entropy             | $S(\Phi(\rho)\|\Phi(\sigma)) \leq S(\rho\|\sigma)$         |
| Regularized Uhlmann theorem for Rényi divergences   | Block-regularized over extensions       | $D_\alpha(\rho_A\|\sigma_A) = \lim_{n\to\infty} \min ...$  |
| Measured Uhlmann theorem for $f$-divergences        | Measured divergence (POVM, PVM)         | Extension property: equality after minimization             |
| Variational formulation for measured divergences    | Convex/SDP program                      | $D_f^{\rm meas}(\rho\|\sigma) = \sup_{\gamma} \ldots$      |

In summary, Uhlmann-type theorems for divergences unify the structure of quantum distinguishability measures, their operational interpretations, and their optimization-theoretic characterizations. They play a foundational role in quantum information science, from mathematical formalism to application in resource conversion, hypothesis testing, and channel analysis [2210.13093, 2502.01749, 2502.07745].

Source: https://www.emergentmind.com/topics/uhlmann-theorem-for-divergences