---
title: Regularized Quantum Hoeffding Divergence
url: https://www.emergentmind.com/topics/regularized-quantum-hoeffding-divergence
type: topic
---

# Regularized Quantum Hoeffding Divergence

Searching arXiv for the cited papers to ground the article in current arXiv records.
Regularized quantum Hoeffding divergence is the asymptotic rate function governing asymmetric binary quantum hypothesis testing under an exponential constraint on the type-II error. In the composite, possibly correlated setting, each blocklength $n$ is associated with convex, compact sets of states $A_n,B_n\subset D(H^{\otimes n})$, and the central result is that, for stable sequences of such sets, the optimal worst-case type-I error exponent is exactly the regularized quantum Hoeffding divergence $H_r^{\infty}(A\|B)$ [2508.12901]. This extends the Hoeffding theory of i.i.d. singleton hypotheses to composite correlated hypotheses, introduces parallel set-based Hoeffding and anti-divergence formalisms, and refines the generalized quantum Stein’s lemma by identifying the full trade-off between type-I and type-II errors [2508.12901].

## 1. Asymmetric testing for composite correlated quantum hypotheses

Let $H$ be a finite-dimensional Hilbert space. For each $n$, the null hypothesis is a convex, compact set $A_n\subset D(H^{\otimes n})$, and the alternative is a convex, compact set $B_n\subset D(H^{\otimes n})$. A two-outcome POVM $\{M_n,I-M_n\}$ with $0\le M_n\le I$ is used in a worst-case manner. The error probabilities are
\[
\alpha(A_n,M_n):=\sup_{\rho_n\in A_n}\operatorname{Tr}[\rho_n(I-M_n)],\qquad
\beta(B_n,M_n):=\sup_{\sigma_n\in B_n}\operatorname{Tr}[\sigma_n M_n].
\]

Given a rate $r>0$, the optimal type-I error under the exponential constraint $\beta(B_n,M_n)\le 2^{-nr}$ is
\[
\alpha_{n,r}(A_n\|B_n):=\min_{0\le M_n\le I}\big\{\alpha(A_n,M_n):~\beta(B_n,M_n)\le 2^{-nr}\big\}.
\]
The associated exponent is
\[
E_n(r):=-\frac{1}{n}\log \alpha_{n,r}(A_n\|B_n),\qquad
E(r):=\liminf_{n\to\infty}E_n(r).
\]

This formulation departs from the standard singleton i.i.d. setting in two ways. First, the hypotheses are composite: one tests between sets rather than fixed states. Second, the sequences may be correlated and need not be i.i.d. The operational question is therefore a worst-case large-deviation problem over state sets whose geometry changes with blocklength [2508.12901].

A key reduction used in the analysis is that composite correlated hypotheses can be transformed into a worst-case pair:
\[
\alpha_{n,r}(A_n\|B_n)=\sup_{\rho_n\in A_n,\sigma_n\in B_n}\alpha_{n,r}(\rho_n\|\sigma_n).
\]
This brings the problem into contact with single-pair asymmetric hypothesis testing while preserving the minimax character of the composite setting [2508.12901].

## 2. Set-based Hoeffding divergences and their regularization

For single states $\rho_n,\sigma_n\in D(H^{\otimes n})$, the finite-$n$ Hoeffding divergence is
\[
H_{n,r}(\rho_n\|\sigma_n):=\sup_{\alpha\in(0,1)}\frac{\alpha-1}{\alpha}\Big(nr-D_{\mathrm{Petz},\alpha}(\rho_n\|\sigma_n)\Big),
\]
where the Petz Rényi divergence is
\[
D_{\mathrm{Petz},\alpha}(\rho\|\sigma):=\frac{1}{\alpha-1}\log Q_\alpha(\rho\|\sigma),\qquad
Q_\alpha(\rho\|\sigma):=\operatorname{Tr}\!\left[\rho^\alpha\sigma^{1-\alpha}\right].
\]

For sets $A_n,B_n\subset D(H^{\otimes n})$, two natural extensions are introduced. The first is the min-divergence extension
\[
H_{n,r}(A_n\|B_n):=\inf_{\rho_n\in A_n,~\sigma_n\in B_n}H_{n,r}(\rho_n\|\sigma_n).
\]
The second is the Petz-based extension
\[
\bar H_{n,r}(A_n\|B_n):=\sup_{\alpha\in(0,1)}\frac{\alpha-1}{\alpha}\Big(nr-D_{\mathrm{Petz},\alpha}(A_n\|B_n)\Big),
\]
with
\[
D(A_n\|B_n):=\inf_{\rho_n\in A_n,~\sigma_n\in B_n}D(\rho_n\|\sigma_n).
\]

For convex, compact sets $A_n,B_n$, these two extensions coincide at finite blocklength:
\[
H_{n,r}(A_n\|B_n)=\bar H_{n,r}(A_n\|B_n).
\]
This finite-$n$ equality is structurally important because it identifies a variational Petz-based expression with the worst-case single-state Hoeffding rate [2508.12901].

The asymptotic notion arises for stable sequences $A=\{A_n\}_{n\in\mathbb N}$ and $B=\{B_n\}_{n\in\mathbb N}$ satisfying
\[
A_n\otimes A_m\subseteq A_{n+m},\qquad B_n\otimes B_m\subseteq B_{n+m}.
\]
Under this stability, subadditivity of $H_{n,r}$ yields the regularized quantity
\[
H_r^{\infty}(A\|B):=\lim_{n\to\infty}\frac{1}{n}H_{n,r}(A_n\|B_n)
=\inf_{n\ge 1}\frac{1}{n}\,H_{n,r}(A_n\|B_n).
\]
Likewise,
\[
\bar H_r^{\infty}(A\|B)=\sup_{\alpha\in(0,1)}\frac{\alpha-1}{\alpha}\Big(r-D_{\mathrm{Petz},\alpha}^{\infty}(A\|B)\Big),
\]
where
\[
D_{\mathrm{Petz},\alpha}^{\infty}(A\|B):=\lim_{n\to\infty}\frac{1}{n}D_{\mathrm{Petz},\alpha}(A_n\|B_n).
\]

The proven asymptotic relation is
\[
H_r^{\infty}(A\|B)\ge \bar H_r^{\infty}(A\|B).
\]
By contrast, on the strong-converse side the corresponding regularized quantities do coincide exactly. This asymmetry is one of the conceptual distinctions between the direct and strong-converse regimes [2508.12901].

## 3. Generalized Hoeffding bound for stable sequences of sets

The main theorem states that for stable sequences of convex, compact sets in $D(H^{\otimes n})$,
\[
\boxed{~ \liminf_{n\to\infty}-\frac{1}{n}\log \alpha_{n,r}(A_n\|B_n)=H_r^{\infty}(A\|B). ~}
\]
Thus the regularized quantum Hoeffding divergence is not merely a variational quantity; it is the exact optimal exponent for worst-case type-I decay under a prescribed type-II exponential constraint [2508.12901].

The proof uses two complementary directions. The converse proceeds via Audenaert’s inequality and the Petz quasi-entropy $Q_\alpha$, yielding the lower bound on achievable exponents by constructing feasible tests. The achievability direction fixes a blocklength $m$, chooses worst-case states $(\rho_m,\sigma_m)\in A_m\times B_m$, invokes the i.i.d. singleton Hoeffding bound for $(\rho_m^{\otimes n},\sigma_m^{\otimes n})$, and then lets $m\to\infty$. Convexity and compactness ensure existence of optimizers and justify minimax exchanges, while tensor-product stability supplies the subadditivity needed for regularized limits [2508.12901].

In the singleton i.i.d. case $A_n=\{\rho\}^{\otimes n}$ and $B_n=\{\sigma\}^{\otimes n}$, the theorem reduces to the standard quantum Hoeffding exponent
\[
\lim_{n\to\infty}-\frac{1}{n}\log \alpha_{n,r}(\rho^{\otimes n}\|\sigma^{\otimes n})
=H_r(\rho\|\sigma),
\]
with
\[
H_r(\rho\|\sigma)=\sup_{\alpha\in(0,1)}\frac{\alpha-1}{\alpha}\big(r-D_{\mathrm{Petz},\alpha}(\rho\|\sigma)\big).
\]
The same quantity admits the standard Legendre-transform form after the reparametrization $\alpha=1/(1+s)$, $s>0$ [2508.12901].

Earlier work by Mosonyi and Ogawa had already established, for general sequences of single hypotheses, that the strong converse exponent can be obtained from a regularized Hoeffding anti-divergence under either a factorization property or differentiability of a regularized Rényi divergence; typical examples include temperature states of translation-invariant finite-range interactions on a spin chain, temperature states of non-interacting fermionic lattice systems, and classical irreducible Markov chains [1407.3567]. The 2025 development extends the scope from single correlated sequences to composite correlated sets.

## 4. Strong converse, anti-divergence, and the Stein threshold

For single states, the Hoeffding anti-divergence is defined by
\[
H^*_{n,r}(\rho_n\|\sigma_n):=\sup_{\alpha>1}\frac{\alpha-1}{\alpha}\Big(nr-D_{\mathrm{Sand},\alpha}(\rho_n\|\sigma_n)\Big),
\]
where the sandwiched Rényi divergence is
\[
D_{\mathrm{Sand},\alpha}(\rho\|\sigma):=\frac{1}{\alpha-1}\log\operatorname{Tr}\Big[\big(\sigma^{\frac{1-\alpha}{2\alpha}}\rho\,\sigma^{\frac{1-\alpha}{2\alpha}}\big)^\alpha\Big].
\]

For sets,
\[
H^*_{n,r}(A_n\|B_n):=\sup_{\rho_n\in A_n,~\sigma_n\in B_n}H^*_{n,r}(\rho_n\|\sigma_n),
\]
and
\[
\bar H^*_{n,r}(A_n\|B_n):=\sup_{\alpha>1}\frac{\alpha-1}{\alpha}\Big(nr-D_{\mathrm{Sand},\alpha}(A_n\|B_n)\Big).
\]
For convex, compact $A_n,B_n$ these coincide for each $n$, and for stable sequences the regularizations also coincide:
\[
H_r^{*,\infty}(A\|B)=\bar H_r^{*,\infty}(A\|B)
=\sup_{\alpha>1}\frac{\alpha-1}{\alpha}\Big(r-D_{\mathrm{Sand},\alpha}^{\infty}(A\|B)\Big).
\]

The strong-converse lower bound is
\[
\boxed{~ \liminf_{n\to\infty}-\frac{1}{n}\log\big(1-\alpha_{n,r}(A_n\|B_n)\big)\ge H_r^{*,\infty}(A\|B)=\bar H_r^{*,\infty}(A\|B). ~}
\]
Hence, if the type-II exponent is forced above the relevant threshold, the type-I error approaches one at least exponentially fast [2508.12901].

This threshold is furnished by the generalized quantum Stein’s lemma for sets:
\[
\lim_{n\to\infty}-\frac{1}{n}\log\beta_\varepsilon(A_n\|B_n)=D^{\infty}(A\|B),
\qquad
D^{\infty}(A\|B):=\lim_{n\to\infty}\frac{1}{n}\inf_{\rho_n\in A_n,\sigma_n\in B_n}D(\rho_n\|\sigma_n).
\]
The Hoeffding theory refines this statement into a full trade-off curve. For $0<r<D^{\infty}(A\|B)$,
\[
\liminf_{n\to\infty}-\frac{1}{n}\log\alpha_{n,r}(A_n\|B_n)\ge \bar H_r^{\infty}(A\|B)>0,
\]
with equality
\[
E(r)=H_r^{\infty}(A\|B).
\]
For $r>D^{\infty}(A\|B)$,
\[
\liminf_{n\to\infty}-\frac{1}{n}\log\big(1-\alpha_{n,r}(A_n\|B_n)\big)\ge \bar H_r^{*,\infty}(A\|B)>0.
\]
Under the polar assumption used in the proof of the generalized Stein’s lemma, one further has continuity of the regularized Petz Rényi family at $\alpha\uparrow 1$:
\[
\sup_{\alpha\in(0,1)}D_{\mathrm{Petz},\alpha}^{\infty}(A\|B)=D^{\infty}(A\|B).
\]
This identifies the Stein exponent as the sharp threshold separating the direct and strong-converse regimes [2508.12901].

## 5. Variational structure, special cases, and computation

On the direct side,
\[
\bar H_r^{\infty}(A\|B)=\sup_{\alpha\in(0,1)}\frac{\alpha-1}{\alpha}\Big(r-D_{\mathrm{Petz},\alpha}^{\infty}(A\|B)\Big).
\]
Because this is a supremum of affine functions of $r$ with negative slopes, the trade-off function $E(r)$ is non-increasing and convex in $r$. On the strong-converse side,
\[
\bar H_r^{*,\infty}(A\|B)=\sup_{\alpha>1}\frac{\alpha-1}{\alpha}\Big(r-D_{\mathrm{Sand},\alpha}^{\infty}(A\|B)\Big),
\]
which is again a supremum of affine functions in $r$ and is non-decreasing in $r$. Differentiability is not generally guaranteed for composite sets, but the convexity structure yields continuity except at possible kinks corresponding to changes of the maximizing $\alpha$ [2508.12901].

Several special cases are explicitly identified. For i.i.d. singleton hypotheses, the generalized bounds reduce to the Hayashi–Nagaoka/Audenaert quantum Hoeffding exponent and the Mosonyi–Ogawa strong-converse exponent. In the classical or commuting case, $Q_\alpha$ reduces to classical log-moment generating functions, and the bounds coincide with classical Hoeffding bounds extended to sets. For composite sets such as energy-constrained sets, convex hulls of finite ensembles, and adversarially generated c–q families, semidefinite representations are available; since $Q_\alpha$ is jointly concave for $\alpha\in(0,1)$, one can compute $D_{\mathrm{Petz},\alpha}(A_n\|B_n)$ and hence $\bar H_{n,r}(A_n\|B_n)$ by convex optimization, for example using QICS [2508.12901].

These exponents have direct implications for adversarial quantum channel discrimination, composite testing in many-body systems, and the design of asymptotically optimal tests under constraints. A plausible implication is that the passage from single hypotheses to stable convex-compact sets preserves enough variational structure to keep the asymptotic trade-off analyzable and, in important examples, computationally tractable [2508.12901].

## 6. Related developments, misconceptions, and open questions

A recurrent misconception is that asymptotic binary tests should suffice in the same way on both sides of the error-exponent trade-off. That is correct for the strong-converse regime but not for the Hoeffding regime. For $\alpha>1$, the regularized measured Rényi divergence and the regularized test-measured Rényi divergence coincide with the sandwiched Rényi divergence, so asymptotically two-outcome tests suffice. For $\alpha<1$, however, even for commuting states the regularized quantity attainable using two-outcome measurements is in general strictly smaller than the Rényi divergence, and the operationally correct Hoeffding exponent is the Petz-generated
\[
H_r(\rho\|\sigma)=\sup_{0<\alpha<1}\frac{1-\alpha}{\alpha}\big(D_\alpha(\rho\|\sigma)-r\big),
\]
rather than a test-measured regularization [2201.05477].

This direct/strong-converse separation is mirrored in the broader literature on correlated states. Mosonyi and Ogawa showed that for general sequences of quantum states the strong converse exponent is equal to the Hoeffding anti-divergence under either a factorization-property approach or a differentiability-based approach [1407.3567]. In a different direction, the divergence $D^{\#}_{\alpha}$, defined via a convex optimization program for $\alpha>1$, has the property that its regularization equals the sandwiched Rényi divergence. This yields converging upper bounds on regularized sandwiched Rényi divergences for channels, a chain rule, an amortization-collapse statement, and a characterization of strong converse exponents for channel discrimination in which adaptive strategies offer no advantage [2007.12576].

Within the composite-set framework itself, two limitations are explicit. First, although
\[
H_r^{\infty}(A\|B)\ge \bar H_r^{\infty}(A\|B),
\]
the full equality in the regularized direct regime remains open; establishing it would require a minimax theorem in the regularized setting. Second, on the strong-converse side, the available result is a lower bound on the exponent of $1-\alpha_{n,r}$, and obtaining matching upper bounds remains an important challenge [2508.12901].

The resulting picture is structurally sharp. The regularized relative entropy $D^{\infty}(A\|B)$ is the threshold, the regularized quantum Hoeffding divergence $H_r^{\infty}(A\|B)$ governs the subthreshold direct regime, and the regularized Hoeffding anti-divergence $H_r^{*,\infty}(A\|B)$ controls the superthreshold blow-up of type-I error. This suggests a unified large-deviation description of asymmetric quantum state discrimination that accommodates convex uncertainty, long-range block structure, and non-i.i.d. correlations without abandoning the Rényi-variational framework [2508.12901].

Source: https://www.emergentmind.com/topics/regularized-quantum-hoeffding-divergence