Hockey-Stick Divergence: Theory and Applications

• Hockey-stick divergence is an f-divergence defined via thresholded likelihood ratios, enabling precise measurement of statistical distinguishability in both classical and quantum settings.
• It offers integral representations and variational characterizations that underpin advances in differential privacy, robust estimation, and kernel-based testing.
• Its generalized forms and strong data-processing inequalities facilitate rigorous risk lower bounds and enhanced applications in quantum information theory.

The hockey-stick divergence is a pivotal $f$-divergence, parameterized for both classical and quantum settings, that enables fine-grained analysis of statistical distinguishability, differential privacy guarantees, and risk lower bounds in estimation and learning. Defined via thresholded likelihood ratios, it admits powerful variational representations and integral connections to general $f$-divergences. The framework generalizes to measured and quantum contexts, where key analytic and operational properties drive advances in privacy accounting, contraction analysis, and quantum information theory.

1. Formal Definitions and Variants

Classical Hockey-Stick Divergence

Given probability measures $P, Q$ (densities $p, q$), for $\gamma \geq 1$ the classical hockey-stick divergence is: $$E_\gamma(P \| Q) = \int [p(x) - \gamma q(x)]_{+} dx$$ or, equivalently, for discrete domains: $$E_\gamma(P \| Q) = \sum_{x}[P(x) - \gamma Q(x)]_{+}$$ where $[u]_+ = \max\{0, u\}$. It arises from $f_\gamma(t) = \max\{0, t - \gamma\}$, making $E_\gamma$ an $f$-divergence. For $\gamma = 1$, $E_1(P \| Q)$ recovers half the total variation distance (Vandenbroucque et al., 2022, Nuradha et al., 23 Jan 2026).

The generalized two-parameter form is: $$E_{\beta,\gamma}(P \| Q) = \int [\beta p(x) - \gamma q(x)]_{+} dx$$ with $\beta > 0$ and $\gamma \geq \beta$ (Vandenbroucque et al., 2022).

Quantum and Measured Variants

For density operators $\rho, \sigma$ and $\gamma \geq 0$, the quantum version is: $$E_\gamma(\rho \| \sigma) = \sup_{0 \leq M \leq I} \text{Tr}[M(\rho - \gamma \sigma)] - (1 - \gamma)_+$$ which, for $\gamma = 1$, gives trace distance. The measured hockey-stick divergence, $E_\gamma^{\mathcal{M}}$, restricts the supremum to operators within a measurement class $\mathcal{M}$ (Nuradha et al., 21 Jan 2025, Nuradha et al., 18 Dec 2025).

2. Integral Representations and Analytical Properties

Every convex, twice-differentiable $f$-divergence can be expressed as an integral over hockey-stick divergences: $$D_f(P \| Q) = \int_1^\infty f''(\gamma) E_\gamma(P \| Q) d\gamma + \int_0^1 \gamma^{-3} f''(\gamma^{-1}) E_\gamma(Q \| P) d\gamma$$ This holds in the quantum regime as well, with corresponding operator forms (Hirche et al., 2023, Nuradha et al., 18 Dec 2025).

Special cases include the quantum relative entropy (Umegaki entropy): $$D(\rho \| \sigma) = \int_1^\infty \left( \frac{1}{s} E_s(\rho \| \sigma) + \frac{1}{s^2} E_s(\sigma \| \rho) \right) ds$$ Integral regularization yields Petz and sandwiched Rényi divergences in the asymptotic regime (Hirche et al., 2023).

3. Variational, Regularized, and Kernel-Based Estimation

The variational characterization for classical $E_\gamma$ is: $$E_\gamma(P \| Q) = \sup_{g: \mathcal{Z} \to [0,1]} \mathbb{E}_{P}[g(X)] - \gamma \mathbb{E}_{Q}[g(Y)]$$ with optimizer $g^*(x) = \{ dP/dQ(x) \geq \gamma \}$.

Regularized (kernel) estimation in RKHS employs: $$D_{\mathrm{HS},\gamma}^{\tau}(P \| Q) = \sup_{g \in \mathcal{H},\; 0 \leq g \leq 1} \left\{ \mathbb{E}_P[g(X)] - \gamma \mathbb{E}_Q[g(Y)] - \frac{\tau}{2} \|g\|_{\mathcal{H}}^2 \right\}$$ Enables finely adaptive two-sample testing, maximization over kernel and regularization parameters, and data-driven permutation test statistics. These methods control Type I error and asymptotic/non-asymptotic power guarantees (Ribero et al., 27 Jan 2026).

4. Data-Processing Inequalities and Contraction Coefficients

Hockey-stick divergences enjoy strong data-processing inequalities (SDPIs):

• Classical DPI: For any Markov kernel $A$, $E_\gamma(A(P) \| A(Q)) \leq E_\gamma(P \| Q)$.
• Linear SDPI: For $(\varepsilon, \delta)$-DP mechanisms $A$,

$$E_\gamma(A(P) \| A(Q)) \leq \frac{e^\varepsilon - \gamma + \delta(\gamma + 1)}{e^\varepsilon + 1} E_\gamma(P \| Q)$$

• Non-linear SDPI:

$$E_\gamma(A(P) \| A(Q)) \leq \max\left\{\frac{(e^\varepsilon + 2\delta - 1)E_\gamma(P \| Q) - (\gamma - 1)(1 - \delta)}{e^\varepsilon + 1}, \; \delta E_\gamma(P \| Q)\right\}$$

Analogous quantum SDPI bounds hold for quantum channels (CPTP maps), with optimized contraction curves and improved mixing time bounds (Nuradha et al., 23 Jan 2026, Nuradha et al., 18 Dec 2025).

For operator-convex $f$, contraction coefficients $\eta_f$ collapse to the classical value, resolving conjectures by Lesniewski and Ruskai (Hirche et al., 2023, Nuradha et al., 18 Dec 2025).

5. Connections to Differential Privacy and Privacy Loss Accounting

The operational form of $(\epsilon, \delta)$-DP is: $$P(M(x) \in B) \leq e^\epsilon P(M(x') \in B) + \delta \iff E_{e^\epsilon}(P_{M(x)} \| P_{M(x')}) \leq \delta$$ The hockey-stick divergence directly quantifies indistinguishability and is central to privacy loss distribution (PLD)-based accounting (Nuradha et al., 23 Jan 2026, Doroshenko et al., 2022).

PLD-based methods use $E_\gamma$ to express $\delta(\varepsilon)$ via the privacy loss random variable: $$D_{e^\varepsilon}(P \| Q) = \mathbb{E}_{L \sim PLD_{(P, Q)}} [1 - e^{-(L - \varepsilon)}]_+$$ Advanced accounting uses discrete convex–concave envelope methods ("Connect the Dots") to optimally estimate upper/lower bounds for $\delta(\varepsilon)$ in composed privacy mechanisms, outperforming previous bucket rounding and moment-based methods (Doroshenko et al., 2022).

6. Applications to Quantum Information Theory and Quantum Privacy

Quantum hockey-stick divergence underlies critical results in quantum differential privacy, audit frameworks ("quantum pufferfish privacy"), and channel discrimination (Nuradha et al., 21 Jan 2025, Hirche et al., 2023). Measured variants $E_\gamma^{\mathcal{M}}$—where $\mathcal{M}$ is LOCC, PPT, etc.—allow efficient SDP computation and tight privacy parameters for practical adversary models.

In quantum Markov processes, strong SDPI for $E_\gamma$ yields tight mixing time and fixed point convergence guarantees. In channel theory, integral hockey-stick representations lead to new bounds in statistical comparison, amortized divergence, and less noisy orders. Regularized quantum $f$-divergence families unify Petz and sandwiched Rényi divergences (Hirche et al., 2023).

7. Bayesian Risk Lower Bounds in Estimation

Generalized hockey-stick divergences induce sharp finite-sample lower bounds for estimator risk: $$\mathbb{E}[\ell(W, \hat{W})] \geq \rho \left( 1 - \frac{1}{\beta} [E_{\beta, \gamma}(W, \hat{W}) + \gamma L_W(\hat{W}, \rho)] \right)$$ Parameter tuning of $(\beta, \gamma)$ focuses bounds on likelihood ratio tails, outperforming KL-, Hellinger-, and Sibson-type information bounds, especially in non-asymptotic regimes (Vandenbroucque et al., 2022).

8. Computational Approaches and Explicit Examples

Measured and unmeasured hockey-stick divergences for highly symmetric states (Werner, isotropic) and channels (depolarizers) admit closed-form evaluation. SDP formulations enable efficient computation under restricted measurement sets, central for privacy auditing in both classical and quantum infrastructures. Covariant channels allow further reduction (Nuradha et al., 21 Jan 2025).

Table: Comparative Summary of Hockey-Stick Divergence Contexts

Variant	Defining Formula/Domain	Key Applications
Classical $E_\gamma(P \| Q)$	$[p(x) - \gamma q(x)]_+$	DP auditing, risk bounds
Quantum $E_\gamma(\rho \| \sigma)$	$$\sup_{0 \leq M \leq I} \text{Tr}[M(\rho-\gamma\sigma)]$$	Quantum DP, channel discrim.
Generalized $E_{\beta, \gamma}$	$[\beta p(x) - \gamma q(x)]_+$	Fine-grained estimation bounds
Measured $E_\gamma^{\mathcal{M}}$	supremum over $M \in \mathcal{M}_2$	Quantum pufferfish privacy

The hockey-stick divergence, through its various generalizations, measured forms, and analytic properties, enables tight theoretical guarantees and practical tools for privacy, estimation, and statistical testing across classical, quantum, and kernel frameworks.