Log-Euclidean Rényi Conditional Mutual Information

Technical Summary: Log-Euclidean Rényi Conditional Mutual Information

Problem Statement
Quantum Conditional Mutual Information (QCMI) is a fundamental measure of conditional independence in quantum information theory and many-body physics, characterizing correlations between systems $A$ and $B$ conditioned on $C$. It vanishes if and only if the tripartite state $\rho_{ABC}$ forms a quantum Markov chain, implying the state can be recovered from its marginal $\rho_{AC}$ via a channel acting only on $C$. While the standard QCMI (corresponding to the limit $\alpha \to 1$ of Rényi generalizations) is well-understood, finding a fully quantum Rényi generalization that satisfies all desirable structural properties has remained an open question. Specifically, a satisfactory Rényi QCMI should:

1. Reduce to the classical Rényi conditional mutual information in the commuting case.
2. Obey the data-processing inequality (DPI) under local quantum channels on $A$ and $B$.
3. Be additive under tensor products.
4. Be monotone in the Rényi parameter $\alpha$.
5. Converge to the standard QCMI as $\alpha \to 1$.

Prior to this work, no known Rényi-type QCMI satisfied all these properties simultaneously due to the obstruction of non-commutativity.

Methodology
The authors introduce two variants of the Log-Euclidean Rényi Conditional Mutual Information (LE-CMI), denoted as $I^\flat_\alpha$ and $I^{\flat,*}_\alpha$. These definitions rely on the log-Euclidean structure of the space of positive definite operators, utilizing the trace of the exponential of a linear combination of logarithms of density matrices.

• Definitions: For a tripartite state $\rho_{ABC}$ and $\alpha \in (0, 1) \cup (1, \infty)$, the non-optimized quantity is defined as:
  $$I^\flat_\alpha(A : B|C)_\rho := \frac{1}{\alpha - 1} \log \text{Tr}\left[ \exp\left( \alpha \log \rho_{ABC} + (1-\alpha)(\log \rho_{AC} + \log \rho_{BC} - \log \rho_C) \right) \right]$$
  The optimized quantity, $I^{\flat,*}_\alpha$, involves an infimum over auxiliary states $\sigma_{BC}, \eta_{AC}$ and a supremum over $\omega_C$ within the same log-Euclidean framework.

• Variational Characterizations: A central methodological tool is the derivation of variational expressions for these quantities. The authors express $I^\flat_\alpha$ and $I^{\flat,*}_\alpha$ as optimizations over auxiliary states $\tau$ involving standard Umegaki relative entropies and conditional mutual informations. For instance, for $\alpha \in (0, 1)$:
  $$I^{\flat,*}_\alpha(A : B|C)_\rho = \inf_{\tau_{ABC}} \left\{ \frac{\alpha}{1-\alpha} D(\tau_{ABC} \| \rho_{ABC}) + I(A:B|C)_\tau \right\}$$
  These variational forms allow the authors to leverage properties of convexity, joint convexity, and the data-processing inequality of standard relative entropies.

• Analytical Tools: The proofs utilize:

  • Sion's minimax theorem to exchange infima and suprema.
  • Joint convexity/concavity of trace exponential functions (specifically $\log \text{Tr} \exp(H + \beta \log A + \beta \log B)$).
  • Fixed-point equations characterizing optimizers (derived via directional derivatives).
  • Multivariate trace inequalities (generalizations of the Golden-Thompson inequality) to establish recovery bounds.

Key Contributions and Results

1. Structural Properties:

   • Data Processing: Both $I^\flat_\alpha$ and $I^{\flat,*}_\alpha$ satisfy the DPI for local quantum channels on $A$ and $B$ for $\alpha \in (0, 1)$.
   • Additivity: The non-optimized quantity $I^\flat_\alpha$ is additive for all $\alpha > 0$. The optimized quantity $I^{\flat,*}_\alpha$ is additive for $\alpha \in [1/2, 2]$. The proof for additivity in the optimized case relies on showing that the tensor product of marginal optimizers is a global optimizer, utilizing fixed-point characterizations.
   • Monotonicity: Both quantities are monotone increasing in the parameter $\alpha$.
   • Convergence: Both quantities converge to the standard QCMI as $\alpha \to 1$.

2. Faithfulness and Conditional Independence:

   • The optimized quantity $I^{\flat,*}_\alpha$ vanishes if and only if $\rho_{ABC}$ is a quantum Markov chain (exact conditional independence).
   • The non-optimized quantity $I^\flat_\alpha$ also vanishes precisely for quantum Markov chains.

3. Recovery Bounds:
   The paper establishes tight bounds relating the LE-CMI to the distance between the original state and states recovered via rotated Petz recovery maps ($R^{[t]}_{C \to BC}$).

   • Lower Bounds: For $\alpha \in (0, 1)$, $I^\flat_\alpha$ is bounded below by an integral of minimal Rényi relative entropies ($D_{\alpha, 1-\alpha}$) and the measured Rényi relative entropy ($D^M_\alpha$) between $\rho_{ABC}$ and the recovered state.
   • Upper Bounds: For $\alpha > 1$ (full-rank states), $I^\flat_\alpha$ is bounded above by an integral of maximal Rényi relative entropies ($D_{\alpha, \alpha-1}$).
   • Limit $\alpha \to 1$: These bounds yield new estimates for the standard QCMI involving Keyl's rate function ($D^\star$) and its dual (D^\$). The lower bound involving $D^\star$ is shown to be at least as strong as the fidelity-based bound of [17] and strictly stronger in certain cases, though neither bound universally dominates the other.

4. Bounds to Markov and Separable Sets:

   • $I^{\flat,*}_\alpha$ is upper-bounded by the log-Euclidean Rényi relative entropy distance to the set of quantum Markov states.
   • $I^{\flat,*}_\alpha$ is lower-bounded by the Rényi relative entropy distance (measured via one-way LOCC) to the set of separable states.

5. Chain Rules:
   The authors derive a family of chain rules for the LE-CMI involving conditional entropies, mutual informations, and conditional mutual informations. These rules relate quantities with different Rényi parameters ($\alpha, \beta, \gamma$) satisfying specific algebraic relations (e.g., $\frac{\beta}{\beta-1} = \frac{\alpha}{\alpha-1} + \frac{1}{\gamma-1}$). These inequalities are tight as parameters approach 1, recovering standard chain rules. The authors note these relations appear novel even in the classical setting.

6. Duality:
   A duality relation is established for the non-optimized LE-CMI on pure states: $I^\flat_\alpha(A:B|C)_\rho = J_\alpha(A:B|D)_\rho$, where $J_\alpha$ is a dual quantity. However, the authors demonstrate via a counterexample that this dual quantity $J_\alpha$ fails to satisfy the data-processing inequality under local channels. This suggests that closure under duality, a property often desired for conditional entropies, may not be natural for conditional mutual informations.

Significance and Claims
The paper claims to resolve the open problem of constructing a quantum Rényi conditional mutual information that satisfies the core structural properties (DPI, additivity, monotonicity, and convergence) previously unattainable simultaneously. By introducing the Log-Euclidean framework, the authors provide a quantity that:

• Offers a distinct measure of approximate conditional independence, quantified by recovery bounds involving relative entropies to Petz-recovered states.
• Improves upon existing bounds for the standard QCMI, particularly the lower bound involving Keyl's rate function, which is at least as strong as and can strictly improve upon the fidelity-based bound in certain regimes.
• Provides new chain rules and duality insights that deepen the understanding of Rényi generalizations in quantum information.

The authors remain modest regarding future applications, noting that while the quantities are structurally sound, their utility in strong-converse problems or Rényi analogues of squashed entanglement remains a subject for future investigation. They also highlight the difficulty in constructing alternative "Petz-type" or "sandwiched" generalizations due to ordering ambiguities in multivariate trace inequalities, suggesting the Log-Euclidean approach may be a necessary path to maintain permutation symmetry and data processing.