Strong Data-Processing Inequalities (SDPI)

• SDPI are quantitative refinements of the classical data-processing theorem that establish contraction coefficients for divergences under channels.
• They provide robust methods for deriving mixing bounds, privacy amplification, and coding limits across classical, continuous, and quantum regimes.
• SDPI exhibit tensorization and variational characterizations, linking them to fundamental tools like Poincaré and log-Sobolev inequalities in information theory.

A strong data-processing inequality (SDPI) is a quantitative refinement of the classical data-processing theorem for divergences or information measures under a channel, quantifying the contraction or decay of distinguishability measures beyond mere monotonicity. For a Markov kernel $K$ (channel) and a divergence %%%%1%%%%, an SDPI asserts the existence of a contraction coefficient $\eta < 1$ such that $D(PK\|QK) \leq \eta\, D(P\|Q)$ for all relevant $P, Q$, providing robust tools for impossibility results, mixing bounds, privacy amplification, and coding limits across discrete, continuous, classical, and quantum regimes.

1. Formal Definition and Variants of SDPI

A Markov kernel $K:\mathcal{X}\to\mathcal{Y}$ acting on a reference input law $\mu$ contracts divergences and information functionals. The contraction coefficient (SDPI constant) for an $f$-divergence $D_f$ is

$$\eta_f(\mu,K) = \sup_{\nu \neq \mu} \frac{D_f(\nu K\,\|\,\mu K)}{D_f(\nu\,\|\,\mu)} \in [0,1].$$

For KL-divergence and mutual information, the classical Ahlswede-Gács bound is

$I(U;Y) \leq \eta_{KL}(\mu,K) I(U;X)$

for any Markov chain $U \to X \to Y$ with $P_{Y|X} = K$ and $P_X = \mu$ (Raginsky, 2014).

For Rényi divergences of order $\alpha >1$, SDPI constants are formulated as

$$\eta_\alpha(P_X, K) = \sup_{Q_X \ll P_X} \frac{D_\alpha(Q_XK\,\|\,P_XK)}{D_\alpha(Q_X\,\|\,P_X)}.$$

Crucially, for $\alpha=2$, the supremum is always achieved at a boundary (vertex) distribution, allowing efficient computation (Jin et al., 2024).

Quantum analogs utilize divergences such as quantum $\chi^2_\kappa$ or hockey-stick divergences, with the contraction coefficient defined for quantum channels $\mathcal{E}$ acting on density operators $\rho, \sigma$ (Cao et al., 2019, Nuradha et al., 18 Dec 2025).

2. Variational Characterizations and Tensorization

SDPI constants admit variational/differential characterizations. For $\Phi$-divergences,

$$\eta_\Phi(\mu, K) = 1 - \inf_{f\in\mathcal{F}(\mathcal{X}),\, E[f]=1} \frac{E_\mu[\operatorname{Ent}_\Phi[f(X)\mid Y]]}{\operatorname{Ent}_\Phi[f(X)]},$$

where $\operatorname{Ent}_\Phi[f(X)\mid Y]$ is conditional $\Phi$-entropy (Raginsky, 2014). For channels on product spaces, SDPI constants tensorize: $$\eta_\Phi\left(\bigotimes_i \mu_i,\, \bigotimes_i K_i\right) = \max_i\, \eta_\Phi(\mu_i, K_i)$$ and similarly for classical and quantum $\chi^2$-divergences (Cao et al., 2019).

For quantum channels, tensorization holds in full generality for $\kappa = \kappa_{1/2}$, and for $\kappa \geq \kappa_{1/2}$ on quantum-classical channels (Cao et al., 2019).

In the conditional setting, C-SDPI coefficients quantify the average contraction for state-dependent channels and likewise tensorize: $$s(P_X, T_{Y|X,V}\mid P_V) = \max_i s(P_{X_i}, T_{Y_i|X_i,V}\mid P_V)$$ for independent $(X_1,X_2)$ and parallel state-dependent channels (Rahmani et al., 22 Jul 2025).

3. Key Bounds, Examples, and Structural Results

Universal bounds:

• Upper: For any convex $f$ and channel $K$, $\eta_f(K) \leq \theta(K)$ where $\theta(K)$ is the Dobrushin coefficient—the worst-case total variation contraction (Raginsky, 2014).
• Lower: For $\chi^2$-divergence, $\eta_{f}(K) \geq S^2(\mu,K)$, where $S(\mu,K)$ is the Hirschfeld–Gebelein–Rényi maximal correlation (Raginsky, 2014). For Rényi divergence, $\eta_{\alpha}(P_X,K) \geq \eta_{\chi^2}(P_X,K)$ universally for $\alpha >1$ (Jin et al., 2024).
• For $(\varepsilon,\delta)$-local differential privacy mechanisms, the exact hockey-stick contraction is

$$\eta_\gamma(A) \leq \max\left\{ \frac{e^\varepsilon-\gamma + \delta(\gamma+1)}{e^\varepsilon+1},\, \delta \right\}$$

with parallel generalizations for $f$-divergences (Nuradha et al., 23 Jan 2026).

Explicit computation:

• For BSC$(\epsilon)$, $\eta_{KL} = (1-2\epsilon)^2$ and $$\eta_\alpha(\text{BSC}(\epsilon)) = 1 + \frac{1}{(\alpha-1)\ln 2}\ln\big((1-\epsilon)^\alpha + \epsilon^\alpha\big)$$ (Jin et al., 2024, Xu et al., 2015).

Continuous and Gaussian channels:

• For the Gaussian convolution (heat flow), if $\mu$ is strongly log-concave, the KL/chi-square SDPI coefficients satisfy

$$\eta_{\chi^2}(\mu, s) \leq \frac{1}{1 + s/C_P(\mu)};\quad \eta_{KL}(\mu, s) \leq \frac{1}{1 + s/(\mu)},$$

with Poincaré and log-Sobolev constants controlling rates of contraction; convexity of $s\mapsto \eta_{\chi^2}(\mu,s)$ holds under log-concavity (Klartag et al., 2024).

4. Applications: Mixing, Learning, Privacy, Reliable Computation

SDPIs underlie rigorous mixing time bounds for Markov chains, MCMC, and Langevin dynamics. For the Proximal Sampler in strongly log-concave targets,

$$I(\rho_{k+1}^X\|\nu^X) \leq (1+\alpha\eta)^{-2} I(\rho^X_k\|\nu^X)$$

yielding exponential convergence and sharp iteration complexity for sampling in relative Fisher information (Wibisono, 8 Feb 2025).

In privacy, SDPIs quantify privacy amplification: post-processing by a channel with contraction $\eta_{TV}<1$ strictly reduces the Rényi differential privacy parameter (Grosse et al., 20 Jan 2025). In quantum settings, non-linear SDPIs for hockey-stick divergences yield tighter mixing times and privacy parameters under multiple sequential private quantum channels (Nuradha et al., 18 Dec 2025).

For learning and memorization, SDPI-based lower bounds establish that any classifier achieving constant accuracy must memorize at least $\Omega(d/n)$ bits of training information, revealing sharp trade-offs between sample size and memorization for high-dimensional problems (Feldman et al., 2 Jun 2025).

Reliable Boolean computation with noisy circuits (Evans–Schulman and von Neumann frameworks): the SDPI constant for a BSC$\left(\delta\right)$ gate appears in the necessary condition for reliable computation with $k$-fan-in gates,

$\delta < \tfrac{1}{2} - \tfrac{1}{2\sqrt{k}},$

i.e., $k (1-2\delta)^2 \geq 1$ (Yang, 2024, Sun, 20 Jul 2025, Zhou et al., 2021).

5. Advanced Topics: Nonlinear SDPI, Reverse Pinsker, Functional Inequality Connections

Non-linear SDPIs replace the flat contraction bound by input-dependent curves, e.g., for quantum hockey-stick divergences,

$$F_{\gamma'}^{\mathcal{N}}(t) \leq \max\{ \frac{(\gamma + 2\delta -1) t - (\gamma'-1)(1-\delta)}{\gamma+1}, \delta t \}$$

allowing strictly sharper bounds except at the worst-case (Nuradha et al., 18 Dec 2025).

Pinsker-type inequalities offer optimal reverse and improved direct inequalities linking Rényi/f-divergences to total variation, allowing precise contraction bounds in the cross-channel setting under distribution restrictions (Grosse et al., 20 Jan 2025).

SDPIs are deeply connected to Poincaré, log-Sobolev, and $\Phi$-Sobolev inequalities. For reversible Markov chains, SDPI constants upper/lower bound log-Sobolev and hypercontractivity constants, controlling concentration of measure and mixing rates (Raginsky, 2014, Caputo et al., 2024).

6. Counterexamples, Limitations, and Open Directions

While many universal upper bounds hold for $f$-divergences, for Rényi divergences $\eta_\alpha(K) \leq \eta_{\mathrm{TV}}(K)$ fails in general due to rare-event amplification by the channel (Jin et al., 2024). There exist Markov chains where continuous-time contraction rates far exceed discrete-time ones and multi-step contraction is not always comparable to one-step (Caputo et al., 2024).

Quantum tensorization fails for SDPI constants for certain divergences beyond $\chi^2_\kappa$. The maximal family of quantum channels admitting full tensorization is open (Cao et al., 2019).

Efficient computation of SDPI constants for general channels and divergences remains challenging; specialized techniques (doubling trick, operator Jensen, majorization) address special cases (Rahmani et al., 22 Jul 2025, Sason, 2021). For non-asymptotic bounds in coding and compression, majorization-based SDPIs offer refined analyses for list-decoding and source coding (Sason, 2021).

7. Synthesis: Role and Scope in Information Theory

Strong data-processing inequalities provide finer-grained limits on information flow, contraction, and distinguishability in complex systems: Markov chains, learning algorithms, privacy mechanisms, circuits, quantum devices. Their universality, tensorization, and variational structure yield powerful and interpretable tools for lower bounding risk, quantifying mixing, designing privacy-preserving algorithms, and certifying reliability in noisy computation; ongoing research addresses sharp bounds for continuous, nonlinear, quantum, and composite settings (Raginsky, 2014, Wibisono, 8 Feb 2025, Grosse et al., 20 Jan 2025, Jin et al., 2024, Zhou et al., 2021, Yang, 2024, Sun, 20 Jul 2025, Xu et al., 2015, Caputo et al., 2024, Nuradha et al., 18 Dec 2025, Cao et al., 2019, Nuradha et al., 23 Jan 2026, Feldman et al., 2 Jun 2025, Rahmani et al., 22 Jul 2025, Klartag et al., 2024, Sason, 2021).