Strengthening the quantum data processing inequality for maximal $f$-divergences via unraveling

Technical Summary: Strengthening the Quantum Data Processing Inequality for Maximal $f$-Divergences via Unraveling

Problem Statement
The paper addresses the quantitative refinement of the Quantum Data Processing Inequality (DPI) for maximal $f$-divergences, specifically focusing on the Belavkin–Staszewski (BS) relative entropy and the broader class of maximal $f$-divergences. While the standard DPI asserts that these divergences are non-increasing under completely positive trace-preserving (CPTP) maps, it does not quantify the rate or magnitude of this decrease for specific dynamical evolutions. The authors investigate channels generated by Lindblad equations with Hermitian jump operators, a setting modeling noisy quantum systems affected by classical noise. The goal is to establish explicit lower bounds on the loss of divergence (dissipation) and derive exponential decay rates for the distance between two evolving quantum states.

Methodology: The Unraveling Lift
The core methodological innovation is "lifting" the quantum problem to a classical stochastic setting via stochastic unraveling. Instead of analyzing the quantum semigroup directly, the authors map the pair of quantum states $(\rho_t, \sigma_t)$ to a pair of probability measures $(\mu_t, \nu_t)$ on the space of pure states $\mathcal{P}(\mathfrak{h})$.

1. Stochastic Unraveling: The Lindblad equation is represented by a Stochastic Schrödinger Equation (SSE) driven by Brownian motions. For Hermitian jump operators, this SSE preserves the norm of the state vector, and the evolution of the law of the pure states follows a classical Fokker–Planck equation (FP).
2. The Bridge (Unravel $f$-Divergence): The authors define the unravel $f$-divergence $D^{\text{UNR}}_f(\rho\|\sigma)$ as the infimum of classical $f$-divergences over all ensembles (probability measures on pure states) that realize the quantum states $\rho$ and $\sigma$.
3. Common Basis Construction: A critical technical step involves constructing a "common basis" for any pair of states $(\rho, \sigma)$ (where $\sigma$ is faithful). The authors prove that the infimum defining $D^{\text{UNR}}_f$ is attained on a specific pair of discrete measures supported on this common basis. Crucially, this minimizing pair consists of pure states, making it admissible initial data for the stochastic unraveling.
4. Transfer of Estimates: By proving the DPI and dissipation inequalities for the classical measures $(\mu_t, \nu_t)$ evolving under the FP equation, and utilizing the equality $D^{\text{UNR}}_f = D^{\max}_f$ on the common basis, the results are transferred back to the quantum states.

Key Contributions and Results

1. Strengthened Data Processing Inequality (Theorem 4.1)
The paper establishes a strengthened DPI where the loss of maximal $f$-divergence is bounded below by the time integral of a relative $f$-Fisher information of the unraveling:
$$D^{\max}_f(\rho_t\|\sigma_t) + 2 \int_0^t I_f(\mu_s\|\nu_s) \, ds \leq D^{\max}_f(\rho_0\|\sigma_0)$$
Here, $I_f$ is a classical Fisher information associated with the Fokker–Planck generator. This result quantifies the "dissipation" of information as an additive, time-integrated term rather than a static bound.

2. Exponential Decay under Bracket Generation (Theorem 4.16, Corollary 4.18)
Assuming the jump operators generate a Lie algebra that acts transitively on the pure states (the bracket-generating condition, satisfying Hörmander's condition), the authors derive exponential decay estimates.

• Rate: The decay rate reaches twice the spectral gap ($2\lambda_{\text{sym}}$) of the symmetrized generator of the FP equation.
• Delay: The full rate is achieved only after an explicit time delay $t_f$, which depends on the transition kernel bounds of the unraveling.
• Hamiltonian Role: The spectral gap $\lambda_{\text{sym}}$ is independent of the Hamiltonian $H$, but the delay and mixing constants depend on it.

3. Decay without Functional Inequalities (Theorem 4.22)
For systems where the jump operators alone do not span the tangent space (non-bracket-generating), the authors provide a second decay estimate relying solely on two-sided bounds on the transition kernel of the unraveling.

• Rate: This approach yields exponential decay at any rate $2\lambda$ where $\lambda$ is below the spectral gap of the full generator (including the Hamiltonian drift).
• Significance: This demonstrates that the Hamiltonian can supply a decay rate that the dissipative jump operators alone cannot achieve, a result not accessible via the symmetrized generator analysis.

4. Equivalence of Maximal and Unravel Divergences (Theorem 3.2)
The paper proves that for any operator-convex $f$, the maximal $f$-divergence $D^{\max}_f(\rho\|\sigma)$ is exactly equal to the classical $f$-divergence of the specific pair of measures supported on the common basis. This validates the "lift" strategy and shows that the infimum over the continuum of measures is attained on a discrete set of $n$ atoms.

Significance and Claims
The authors claim that their approach provides a strengthened version of the DPI by converting the abstract monotonicity into an explicit dissipation integral. The significance lies in:

• Classical Reduction: Reducing the analysis of quantum relative entropies to classical probability theory on the manifold of pure states, allowing the use of tools like Poincaré inequalities, logarithmic Sobolev inequalities, and kernel bounds.
• Explicit Rates: Providing concrete, albeit sometimes delayed, exponential decay rates for maximal $f$-divergences, including the BS entropy, which are relevant for quantum channel capacities and hypothesis testing.
• Hamiltonian Enhancement: Showing that in non-bracket-generating systems, the coherent dynamics (Hamiltonian) can restore ergodicity and provide decay rates that the dissipative part alone cannot.

Limitations and Open Questions
The paper explicitly notes the following constraints and open issues:

• Hermitian Jump Operators: The results rely on Hermitian jump operators to ensure the existence of a norm-preserving classical unraveling. General Lindbladians would require different unraveling techniques.
• Kernel Bounds: The decay estimates assume the existence of two-sided bounds on the transition kernel (Property K). While hypoellipticity guarantees smoothness, strict positivity (the lower bound) is assumed or observed numerically in examples but not proven generally.
• Intrinsicness: The dissipation term and the derived delays are functionals of the unraveling measures and transition kernels, not intrinsic functionals of the quantum states $\rho_t$ and $\sigma_t$ alone. The authors state it is unclear if a remainder expressible solely in terms of the states is available.
• Finite Dimensions: The analysis assumes a finite-dimensional Hilbert space, utilizing the compactness of the pure state manifold.