Geodesic Quantum $f$-Divergences

Technical Summary: Geodesic Quantum f-Divergences

Problem Statement
In quantum information theory, extending classical divergences (such as the Kullback–Leibler divergence or $f$-divergences) to non-commuting quantum states introduces a fundamental ambiguity regarding operator ordering. While the Umegaki relative entropy (standard) and the Belavkin–Staszewski (BS) relative entropy (maximal) are well-established endpoints, there exists a lack of a unified geometric framework that interpolates between them while preserving key information-theoretic properties like the Data Processing Inequality (DPI). Existing families, such as the $\alpha$-$z$ Rényi divergences, organize various quantum extensions but do not arise from a direct geometric interpolation of the underlying modular operators. This paper addresses the construction and characterization of a new family of divergences derived from the affine-invariant geodesic connecting the standard and maximal relative modular operators.

Methodology
The authors construct the geodesic quantum $f$-divergences, denoted $D^t_f$, by interpolating between two specific positive operators on the Hilbert–Schmidt space:

1. The standard relative modular operator: $\Delta_{\rho|\sigma} = L_\rho R_{\sigma^{-1}}$.
2. The maximal (commutant) relative modular operator: $\hat{\Delta}_{\rho|\sigma} = R_{\sigma^{-1/2}\rho\sigma^{-1/2}}$.

The interpolation is defined via the weighted geometric mean (Kubo–Ando mean) of these operators:
$$\Gamma_t := \Delta_{\rho|\sigma} \#_t \hat{\Delta}_{\rho|\sigma} = L_{\rho^{1-t}} R_{B_t}, \quad \text{where } B_t = \sigma^{-1/2}\rho^t\sigma^{-1/2}.$$
The divergence is then defined via functional calculus on this geodesic modular operator:
$$D^t_f(\rho\|\sigma) := \langle \sigma^{1/2}, f(\Gamma_t)[\sigma^{1/2}] \rangle_{HS}.$$
The paper analyzes specific generators $f$:

• Logarithmic generator ($f(x) = x \log x$): Yields the $t$-relative entropies $D^t$, interpolating between Umegaki ($t=0$) and BS ($t=1$).
• Power generators ($f(x) = x^\alpha$): Yields the $(t, \alpha)$-Rényi divergences $D^t_\alpha$, interpolating between Petz ($t=0$) and geometric ($t=1$) Rényi divergences.

The analysis relies on finite-dimensional operator theory, spectral calculus, the Daleckii–Krein formula for derivatives, and the theory of operator-convex functions. The authors also utilize a companion paper's construction of a canonical finite classical experiment to provide operational interpretations.

Key Contributions and Results

1. Data Processing Inequality (DPI):
   The primary result is the proof that $D^t_f$ satisfies the DPI for every finite operator-convex generator $f$ and every interpolation parameter $t \in [0, 1]$. Specifically:

   • $D^t$ satisfies DPI for all $t$.
   • $D^t_\alpha$ satisfies DPI for $0 < \alpha < 1$ and $1 < \alpha \leq 2$.
     This justifies the terminology "quantum divergences" for the new family. The proof employs a transformer contraction argument adapted from Petz's proof for the Umegaki entropy, utilizing the contractivity of the map $V_{\sigma, \Phi}$ on the Hilbert–Schmidt space.

2. Equality and Recovery Conditions (The Endpoint Dichotomy):
   The paper establishes a sharp dichotomy in the equality conditions for the DPI:

   • Non-maximal range ($0 \leq t < 1$): For equality-determining operator-convex generators, the equality condition $D^t_f(\rho\|\sigma) = D^t_f(\Phi(\rho)\|\Phi(\sigma))$ is equivalent to the Petz sufficiency condition (recoverability by the Petz map). This holds for all such $t$, meaning the recovery class does not interpolate; it remains the standard Petz class throughout the interior.
   • Maximal endpoint ($t = 1$): The equality condition changes to the generally larger Belavkin–Staszewski (BS) class.
   • Quadratic Exception ($\alpha = 2$): The quadratic generator yields a divergence independent of $t$ that possesses the BS equality class for all $t$.
     This result implies that for invertible states, the set of quantum Markov chains (QMCs) defined by vanishing conditional mutual information collapses to the standard QMC class for all $t < 1$, while at $t=1$ it expands to the BS-QMC class.

3. Monotonicity in the Interpolation Parameter:
   The behavior of the divergences with respect to $t$ is characterized as follows:

   • Relative Entropies ($t$-entropies): Strictly increasing in $t$ for non-commuting states.
   • Rényi Divergences ($D^t_\alpha$):
     • Non-decreasing for $0 < \alpha < 2$.
     • Constant for $\alpha = 2$.
     • Non-increasing for $2 < \alpha \leq 3$.
     • No universal monotonicity direction for $\alpha > 3$ in dimensions $\geq 3$.
       The paper provides a complete characterization of the "universal monotonicity cone" for general $C^1$ generators using an ordinary-matrix test (Theorem 5.10) and a polar condition on the canonical representing measure for operator-convex generators (Theorem 5.11).

4. Quantitative Remainders and Continuity:

   • Strengthened DPI: The authors derive quantitative lower bounds for the DPI deficit. For $D^t$, the deficit is bounded below by a quartic term of an intrinsic Hilbert–Schmidt defect (related to the likelihood ratio). For Rényi divergences, the bounds depend on the order $\alpha$.
   • Continuity: The paper establishes continuity bounds for the divergences and conditional quantities under the assumption of positive lower bounds on the smallest eigenvalues of the states. It demonstrates that without such bounds, the conditional geodesic entropies and forward conditional mutual informations are discontinuous at rank-deficient states for any $t > 0$.

5. Operational Interpretation:
   Using the canonical classical experiment constructed in the companion paper, the authors prove a capacity-per-unit-cost interpretation. The geodesic relative entropy $D^t(\rho\|\sigma)$ is shown to be exactly the capacity per unit cost of a specific binary-input memoryless channel derived from the modular record of the state pair.

Significance and Claims
The paper claims that the geodesic construction provides a unified framework that:

• Bridges Endpoints: It connects the Umegaki and BS relative entropies (and Petz and geometric Rényi divergences) in a geometrically natural way.
• Preserves Structure: It retains the data-processing property and tensor additivity, which are essential for information-theoretic applications.
• Reveals Rigidity: It uncovers a "rigid" structure in the equality conditions, showing that the recovery class does not interpolate smoothly but rather exhibits an abrupt transition at the maximal endpoint. This challenges the intuition that intermediate points might possess intermediate recovery properties.
• Distinguishes from Existing Families: The paper demonstrates that the geodesic family is structurally distinct from the $\alpha$-$z$ family, with different universal intersection properties and monotonicity regimes.

The authors note that the results are stated for invertible states, with extensions to non-invertible cases being local and dependent on support conditions. The paper does not claim to resolve all open problems regarding sharp continuity moduli or the maximal data-processing range for non-convex generators, but it provides a complete characterization for the operator-convex slice and establishes the foundational properties of the new family.