Fixed-ray escort representations of sandwiched and $\alpha$--$z$ Rényi divergences on von Neumann algebras

Technical Summary: Fixed-ray escort representations of sandwiched and $\alpha$–$z$ Rényi divergences on von Neumann algebras

Problem Statement
The paper addresses the structural representation of $\alpha$–$z$ Rényi divergences (including the sandwiched Rényi divergence as a special case) for normal states on arbitrary von Neumann algebras. While these divergences are fundamental in quantum information theory and quantum field theory (particularly in Type III algebras where density matrices do not exist), their analytical properties are often difficult to handle directly. Previous integral representations, such as those based on "hockey-stick" or "layer-cake" decompositions, reconstruct divergences from binary-testing data or likelihood thresholds. The authors seek a representation that expresses these divergences as averages of the ordinary Araki relative entropy along a specific trajectory of states, thereby linking the order-dependent Rényi quantity to a probability average of standard relative entropies.

Methodology
The authors employ a combination of finite-dimensional calculus and infinite-dimensional operator algebra techniques, specifically utilizing Haagerup non-commutative $L^p$ spaces.

1. Finite-Dimensional Derivation: In Section 2, the authors derive the integral formula for density matrices. They define a "fixed-ray" in the $(\alpha, z)$ parameter space where $z = c\alpha$. Along this ray, they construct a family of "fixed-ray escort states" $\rho^{(c)}_\beta$. By differentiating the moment function $Q_{\beta, c\beta}$ with respect to the order $\beta$, they establish a differential identity relating the derivative of the normalized moment to the relative entropy of the escort state. Integration of this identity yields the main formula.
2. Operator-Algebraic Extension: To extend this to arbitrary von Neumann algebras (Section 6), the authors cannot rely on standard differentiation of unbounded operators. Instead, they:
   • Construct a compatible weighted Haagerup scale (Lemma 6.1) anchored at a specific element (either $h^{1/c}_\psi$ for lower orders or the sandwiched factor for upper orders).
   • Utilize quasi-Banach interpolation (specifically the all-$p$ interpolation theorem of Gu, Yin, and Zhang [16]) to handle exponents $p < 1$, which is necessary for the full range of parameters.
   • Establish an anchored entropy identity (Proposition 6.4), proving that the logarithmic $L^p$-norm of a path in the Haagerup scale is a convex potential whose derivative is proportional to the Araki relative entropy of the normalized power.
   • Apply the Fundamental Theorem of Calculus for convex functions to integrate this derivative, recovering the integral representation.

Key Contributions and Results

1. Fixed-Ray Integral Representation (Theorem 4.1):
   The primary result is that for normal states $\psi, \omega$ on an arbitrary von Neumann algebra, the $\alpha$–$z$ divergence can be written as:
   $$D_{\alpha,z}(\psi\|\omega) = \int_{I_\alpha} D(\psi^{(c)}_\beta \|\omega) \, d\mu_\alpha(\beta)$$
   where $c = z/\alpha$, $I_\alpha$ is the interval between $1$ and $\alpha$, $\psi^{(c)}_\beta$ is the fixed-ray escort state, and $d\mu_\alpha$ is a specific probability measure.

   • Conditions: For $0 < \alpha < 1$, the formula requires the support inclusion $s(\psi) \leq s(\omega)$. For $\alpha > 1$, finiteness of the divergence implies this condition.
   • Generality: This holds for every $z > 0$ and $\alpha \neq 1$, including parameters outside the standard data-processing inequality (DPI) region.

2. Support-Boundary Correction (Proposition 4.2 & Corollary 4.4):
   When the support condition $s(\psi) \leq s(\omega)$ fails for $0 < \alpha < 1$, the integral representation is modified by an exact boundary term:
   $$D_{\alpha,c\alpha}(\psi\|\omega) = \int_{I_\alpha} D(\tilde{\psi}^{(c)}_\beta \|\omega_e) \, d\mu_\alpha(\beta) - \frac{\alpha}{1-\alpha} \log q_c$$
   Here, $q_c$ is a scalar representing the "mass" of the state compressed onto the support of the reference state. If $q_c = 0$, the divergence is infinite. This term accounts for the "missing support" visible as an endpoint jump.

3. Monotonicity and Convex Geometry (Corollary 4.5):
   The paper proves that the "refined divergence" $R_c(\beta) = D(\psi^{(c)}_\beta \|\omega)$ is non-decreasing along the fixed-ray trajectory. Consequently, the $\alpha$–$z$ divergence is a weighted average of a non-decreasing function, establishing bounds between the Rényi divergence and the ordinary relative entropy. The authors define a convex potential $\Phi_c(\kappa)$ whose derivative is the refined divergence, linking the geometry of the divergence to convex analysis.

4. Stability and Rigidity (Section 4.3):
   Using Pinsker's inequality, the authors derive a stability bound showing that the divergence controls the mean-square trace-norm dispersion of the escort states around their barycenter. They establish that the "Holevo information" of the escort ensemble vanishes if and only if the escort trajectory is constant (almost everywhere).

Operational Implications
The paper reformulates several operational quantities using this integral representation, interpreting them as signed areas or level-crossing statements:

• Strong-Converse Exponents: For the sandwiched line ($z=\alpha$), the exact strong-converse exponent for asymmetric binary hypothesis testing is expressed as the supremum of a signed area under the escort profile (Eq. 4.55).
• Work Extraction: The reliability exponent for work extraction in finite-dimensional thermal operations is reformulated as an optimization over the escort profile (Eq. 4.63).
• Pair Conversion: The conversion rate for restricted classes of state pairs is organized by a family of escort profiles and their associated support-overlap costs (Eq. 4.67).

Significance and Claims
The authors claim that this representation offers a distinct geometric perspective from existing integral formulas (like hockey-stick decompositions). By integrating over the Rényi order itself rather than a likelihood threshold, the representation:

• Unifies the treatment of sandwiched and general $\alpha$–$z$ divergences.
• Extends valid analytical identities to arbitrary von Neumann algebras without requiring global $\sigma$-finiteness or faithfulness on the original algebra (relying instead on support corners).
• Provides a "convex order potential" that simplifies the analysis of optimization problems in quantum information, turning complex order-dependent optimizations into level-crossing problems against a monotone profile.
• Identifies the precise nature of the divergence when support conditions fail, isolating the boundary term exactly.

The paper does not propose new experimental protocols or claim to solve open problems in quantum field theory directly, but rather provides the analytical machinery (the integral representation) that can be applied to such problems, as noted in their companion work regarding quantum null energy conditions.