Weighted Relative Entropy Contraction

• Weighted relative entropy contraction is a principle that quantifies the decay of relative entropy under system-specific dynamics using operator convexity and spatial weights.
• It unifies classical and quantum frameworks by linking contraction rates to spectral gaps, curvature, and entropy-dissipation inequalities.
• The approach enables explicit computation of decay rates in models such as unital qubit channels, block dynamics, and conservation laws.

A weighted relative entropy contraction is a general principle asserting that the (possibly weighted or generalized) relative entropy between two evolving states contracts under certain dynamics—stochastic, deterministic, classical, or quantum—often at a rate governed by underlying mixing, spectral gap, or curvature properties. This principle serves as a unifying theme across numerous domains: quantum channel theory, nonequilibrium statistical mechanics, interacting particle systems, conservation laws, and metrics on Markov chains. The weighted nature enters either through operator-convex functions (quantum setting), geometric block decompositions (spin systems), or through assignment of locality or shift weights (hyperbolic conservation laws).

1. Formalism of Weighted Relative Entropy

Weighted relative entropy generalizes the standard Kullback–Leibler divergence by incorporating weights either into the state space decomposition or through embedding via a weight function or operator.

• Classical setting: Given measures $\mu$, $\nu$ on a configuration space $X$, the KL divergence is

$$D(\mu \Vert \nu) = \int_X \log \bigl( d\mu/d\nu(x) \bigr)\; d\mu(x)$$

Weighting arises naturally when entropy is decomposed over regions or blocks, as in the weighted sum $\sum_A \alpha_A D(\mu_A \Vert \nu_A)$.

• Quantum setting: When $g: (0, \infty) \to \mathbb{R}$ is operator-convex, the $g$-divergence is

$$D_g(\rho\Vert\sigma) = \mathrm{Tr}\left[ \sigma^{1/2}\; g\left(L_\rho R_{\sigma^{-1}}\right)\sigma^{1/2} \right]$$

where $L_\rho(X) = \rho X$ and $R_{\sigma^{-1}}(X) = X \sigma^{-1}$. Upon symmetrization, $g(x) + xg(x^{-1}) = (x-1)^2 K(x)$ for some normalized, operator-convex $K$ (Hiai et al., 2015).

• Block-spin or spatial decompositions: In finite-lattice Gibbs setups, the entropy on a volume $V$ is bounded by a weighted block sum:

$$\gamma(\alpha)\;\mathrm{Ent}_V^\tau(f) \leq C \sum_{A\subset V} \alpha_A~\mathbb{E}\left[\mathrm{Ent}_A^\tau(f)\right]$$

with $\gamma(\alpha) = \min_x\, \sum_{A \ni x} \alpha_A$ (Caputo et al., 2020).

• Nonlocal weights (conservation laws): In $\mathbb{R}^m$, the weighted pseudo-distance between a solution $U$ and a fixed shock $S$,

$$D(U(t), S) = \int_{x < \alpha(t)} \eta(U(x)\Vert U_-) dx + a \int_{x > \alpha(t)} \eta(U(x)\Vert U_+) dx$$

depends on the weighting $a$ and shift $\alpha(t)$ (Vasseur, 2013).

2. Contraction Principle and Differential Inequalities

Weighted relative entropy contracts under appropriate dynamics characterized by Markov generators, Lindblad semigroups, or block-update schemes. The contraction behavior is governed by entropy-dissipation inequalities of the form:

$$\frac{d}{dt} D(\mu_t\Vert\nu_t) = -\mathcal{I}(\mu_t, \nu_t)$$

where $\mathcal{I}$ is a (weighted) Fisher information or Dirichlet form.

In the quantum unraveling of Lindblad dynamics, the entropy-dissipation law reads (Ortigueira et al., 28 Nov 2025):

$$\frac{d}{dt} S_{BS}(\rho_t\Vert \sigma_t) \leq -2\, I_{UNR}(\rho_t \Vert \sigma_t) \leq 0$$

Analogous results hold for classical block dynamics and Markov chains with nonnegative curvature (Caputo et al., 2020, Caputo et al., 2024):

$$\frac{d}{dt} D(\mu_t\Vert \nu) \leq -\lambda D(\mu_t\Vert \nu) \implies D(\mu_t\Vert \nu) \leq e^{-\lambda t} D(\mu_0\Vert \nu)$$

3. Spectral Gap, Curvature, and Explicit Contraction Rates

Quantitative control over the contraction rate is typically supplied by a spectral gap, Poincaré inequality, or curvature estimate.

• Quantum semigroups: When the Fokker–Planck generator $L$ has spectral gap $\lambda > 0$, the Belavkin–Staszewski relative entropy decays exponentially with coefficient $2\lambda$ (Ortigueira et al., 28 Nov 2025):

$$S_{BS}(\rho_t\Vert\sigma_t)\;\leq\;e^{-2\lambda t}\,S_{BS}(\rho_0\Vert\sigma_0)$$

• Quantum channels: The $g$-divergence contraction coefficient for CPT maps $\Phi$ is

$$\eta_g(\Phi) = \sup_{\rho\neq\sigma} \frac{D_g(\Phi(\rho)\Vert\Phi(\sigma))}{D_g(\rho\Vert\sigma)}$$

with $0 \leq \eta_g(\Phi)\leq 1$ and explicit evaluation possible for unital qubit channels, amplitude-damping, etc (Hiai et al., 2015).

• Block-spin systems: The contraction for weighted block-heat-bath dynamics is dictated by $\gamma(\alpha)$:

$$D(\mu_t\Vert\nu) \leq e^{-\gamma(\alpha)t/C} D(\mu_0\Vert\nu)$$

(Caputo et al., 2020).

• Markov chains with curvature: If $W_d(P\mu, P\nu) \leq (1-\kappa) W_d(\mu, \nu)$ for Wasserstein distance, then equivalently

$D(P\mu\Vert P\nu) \leq (1-\kappa) D(\mu\Vert\nu)$

with $\kappa$ the curvature modulus (Caputo et al., 2024).

4. Structure of Contraction Coefficients and Relationships

Weighted relative entropy contraction coefficients interact with other quantitative distances and monotone metrics:

Coefficient	Inequality chain	Attainment/Special Case
Geodesic	$\leq$ Riemannian	Equates for all operator-convex $K$
Riemannian	$\leq$ $g$-divergence (weighted entropy)	Saturated for unital qubit channels
$g$-divergence	$\leq 1$	Classical case: Dobrushin coeff.
Trace norm	$\leq$ sqrt(Riemannian)	Bounding for all CPT maps

For generalized quantum divergences (e.g., BKM, WY, maximal metric), explicit closed-form contraction coefficients exist for certain channels, but extremal inequalities may be strictly loose outside special cases such as unital qubit or amplitude-damping channels. The conjecture that $K(x) = x^{-1/2}$ maximizes contraction is false; values must be computed case by case (Hiai et al., 2015).

5. Connections to Classical and Quantum Systems

Weighted entropy contraction unifies approaches across disparate models:

• Quantum Lindblad master equations: The minimal KL divergence between unravellings of density matrices realizes the Belavkin–Staszewski entropy, which contracts under Lindblad flow precisely as dictated by classical functional inequalities transferred to the pure-state measure ensemble (Ortigueira et al., 28 Nov 2025).
• Spin systems and interacting particles: Block decomposition with arbitrary $\alpha_A$ weights generalizes log-Sobolev and Shearer inequalities, enabling explicit entropy/mixing control in high-temperature and spatially mixing spin systems (Caputo et al., 2020).
• Conservation laws and shocks: Relative entropy pseudo-distances weighted by shift parameters yield stability and contraction properties for weak solutions, including in the presence of discontinuities or vacuum (Vasseur, 2013).
• Markov chains in geometry: Sectional curvature governs both Wasserstein and entropy contraction, leading to sharp estimates in birth–death processes, zero-range dynamics, and Gibbs samplers. The time-varying MLSI captures scale-dependent curvature improvement (Caputo et al., 2024).

6. Extensions and Generalizations

The contraction principle extends to maximal $f$-divergences by replacing KL with $D_f$ and the BS entropy with $D_f^{\mathrm{max}}$, with parallel contraction-by-gap reasoning holding throughout. For classical Markov semigroups and weighted block dynamics, the proof techniques enable dimension-free spectral estimates, modified log-Sobolev inequalities, and mixing time bounds. In all cases, the contraction coefficients—in particular, the dependence on spectral gap or curvature—can be leveraged for precise rates in hypothesis testing, quantum thermodynamics, and stability analysis.

7. Illustrative Examples and Computability

Explicit contraction rates are available in paradigmatic settings:

• Unital qubit channels: All contraction coefficients for monotone metrics, divergences, and trace norms collapse to $\|T\|^2$ or $\|T\|$ depending on the metric, so optimal rates are directly computable (Hiai et al., 2015).
• Block heat-bath dynamics: For finite-state spaces, total-variation mixing time and log-Sobolev constant are bounded in terms of $\gamma(\alpha)^{-1}$, which is determined by the local block structure and weights (Caputo et al., 2020).
• Birth–death processes: Choosing a metric $d$, the contraction follows from coupling and curvature computation, with the entropy decay rate linked to the maximal mean difference in expected position under the semigroup (Caputo et al., 2024).
• Zero-range mean-field dynamics: Sharp MLSI constants are computable by examining the minimal rate increment, matching the exponential rate of weighted relative entropy contraction (Caputo et al., 2024).
• Conservation laws with shocks: The $L^2$ stability and explicit bounds on the required spatial shift demonstrate the flexibility and robustness of weighted pseudo-distances for nonlinear hyperbolic problems (Vasseur, 2013).

Weighted relative entropy contraction is thus a central unifying concept relating functional inequalities, spectral theory, and stability under time-evolution, accommodating both classical and quantum dynamics, as well as both local and global weighting schemes across a spectrum of models (Ortigueira et al., 28 Nov 2025, Hiai et al., 2015, Caputo et al., 2020, Vasseur, 2013, Caputo et al., 2024).