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

# Regularized Quantum Hoeffding Anti-divergence

Regularized quantum Hoeffding anti-divergence is the asymptotic quantity that governs the strong converse side of asymmetric quantum hypothesis testing: when the type-II error is constrained to decay at a rate beyond the relevant relative-entropy threshold, the type-I error necessarily converges to \(1\), and the anti-divergence gives the exponential rate of that convergence. In the correlated-state setting, it is defined from regularized Rényi divergences, i.e. block-normalized limits over many-body states rather than single-copy quantities, and for broad classes of quantum and classical correlated models it coincides with the strong converse exponent [1407.3567]. In later work on composite and correlated hypotheses specified by sets of states, a set-valued regularized quantum Hoeffding anti-divergence was defined and shown to provide a lower bound on the strong converse exponent, extending the framework beyond singleton i.i.d. hypotheses [2508.12901].

## 1. Operational meaning in asymmetric quantum hypothesis testing

The underlying task is binary quantum hypothesis testing between two sequences of quantum states, written in the source material as \(\{p_n\}\) and \(\{o_n\}\), where the states may be correlated rather than i.i.d. The asymptotic question is the trade-off between type-I and type-II error probabilities as the blocklength \(n\) grows. The direct region concerns rates \(r < D(p\|o)\), while the strong converse region concerns rates \(r > D(p\|o)\), where attempts to force the type-II error to decay too fast compel the type-I error to approach \(1\) [1407.3567].

For a prescribed rate \(r > D(p\|o)\), the strong converse exponent is given in the source material by
\[
sc(r \mid p \| o) := \inf \left\{ -\liminf_{n \to \infty} \frac{1}{n} \log (1 - \alpha_n(T_n)) : \limsup_{n \to \infty} -\frac{1}{n}\log \beta_n(T_n) < -r \right\},
\]
with \(\alpha_n(T_n)\) and \(\beta_n(T_n)\) denoting the type-I and type-II errors of a test \(T_n\). Its operational interpretation is explicit: it quantifies how fast the probability of correct acceptance of the null decays once the type-II exponent is pushed beyond the Stein threshold [1407.3567].

The same operational picture persists for set-valued hypotheses. For stable sequences of convex compact sets \(\mathscr A=\{\mathscr A_n\}\) and \(\mathscr B=\{\mathscr B_n\}\), the regularized relative entropy \(D^\infty(\mathscr A\|\mathscr B)\) acts as the boundary between the Hoeffding and strong-converse regimes. For \(r < D^\infty(\mathscr A\|\mathscr B)\), a nontrivial type-I exponent is achievable; for \(r > D^\infty(\mathscr A\|\mathscr B)\), type-I error must converge to \(1\), with rate lower bounded by the regularized quantum Hoeffding anti-divergence [2508.12901].

## 2. Definition from regularized Rényi divergences

For correlated sequences, regularization is essential. The regularized Rényi divergence is defined by the block limit
\[
D_a(p\|o) := \lim_{n \to \infty} \frac{1}{n} D_a(p_n \| o_n),
\]
and the source material stresses that this blockwise limit is critical in the presence of correlations such as those arising on spin chains [1407.3567].

On the strong-converse side, the relevant noncommutative Rényi quantity is the sandwiched, or minimal, Rényi divergence for \(a>1\),
\[
D_a^{*}(\varrho\|\sigma) := \frac{1}{a-1} \log\left( \operatorname{Tr}\left[ \left(\sigma^{\frac{1-a}{2a}} \varrho \sigma^{\frac{1-a}{2a}} \right)^a\right] \right), \quad a>1.
\]
Using its regularization over blocks, the regularized quantum Hoeffding anti-divergence for sequences is written in the source material as
\[
H^{*}(p\|o) := \sup_{a>1} \frac{1}{a} [r - D_a(p\|o)],
\]
and the main result identifies this quantity with the strong converse exponent under the paper’s assumptions [1407.3567].

A related representation given in the same source is
\[
H^*(p\|o) = \sup_{a > 1} \frac{1}{a} [ r - \psi(a) ],
\]
where
\[
\psi(a) := \lim_{n\to\infty} \frac{1}{n} \log \operatorname{Tr}\left[ p_n^a o_n^{1-a} \right].
\]
This emphasizes that the anti-divergence is a regularized transform of an asymptotic Rényi generating function [1407.3567].

For sets of states, the later framework introduces two equivalent finite-\(n\) extensions. One is the supremum of state-wise anti-divergences,
\[
H_{n, r}^*(\mathscr{A}_n \| \mathscr{B}_n) := \sup_{\rho_n \in \mathscr{A}_n, \; \sigma_n \in \mathscr{B}_n} H_{n, r}^*(\rho_n \| \sigma_n),
\]
with
\[
H_{n, r}^*(\rho_n \| \sigma_n) = \sup_{\alpha > 1} \frac{\alpha-1}{\alpha} \left( n r - D_{\text{Sand}, \alpha}(\rho_n\|\sigma_n) \right).
\]
The other is defined directly from the set-valued sandwiched Rényi divergence,
\[
\bar{H}_{n, r}^*(\mathscr{A}_n \|\mathscr{B}_n) := \sup_{\alpha > 1} \frac{\alpha-1}{\alpha}\left( n r - D_{\text{Sand},\alpha}(\mathscr{A}_n \| \mathscr{B}_n) \right),
\]
where
\[
D_{\text{Sand},\alpha}(\mathscr{A}_n \| \mathscr{B}_n) := \inf_{\rho_n \in \mathscr{A}_n, \; \sigma_n \in \mathscr{B}_n} D_{\text{Sand},\alpha}(\rho_n\|\sigma_n).
\]
For convex compact sets, Lemma 11 states that these two definitions coincide, and Lemma 12 gives the regularized asymptotic form
\[
H_{r}^{*,\infty}(\mathscr{A}\|\mathscr{B}) := \lim_{n\to\infty} \frac{1}{n} H_{n, r}^*(\mathscr{A}_n\|\mathscr{B}_n)
= \sup_{\alpha>1} \frac{\alpha-1}{\alpha}\left(r - D_{\text{Sand},\alpha}^{\infty}(\mathscr{A}\|\mathscr{B})\right)
\]
with
\[
D_{\text{Sand},\alpha}^{\infty}(\mathscr{A}\|\mathscr{B}) := \lim_{n\to\infty} \frac{1}{n} D_{\text{Sand},\alpha}(\mathscr{A}_n\|\mathscr{B}_n).
\]
These formulas make the set-based regularization explicit [2508.12901].

## 3. Two general routes to equality with the strong converse exponent

The 2014 analysis develops two distinct approaches to proving that the strong converse exponent equals the regularized Hoeffding anti-divergence for correlated states. The first route assumes a factorization property for both state sequences. The second route assumes existence and differentiability of the regularized sandwiched Rényi divergences in the Rényi parameter for \(a>1\) [1407.3567].

Under the differentiability route, Theorem IV.5 states that if for every \(a>1\) the limit defining \(D_a(p\|o)\) exists and \(a \mapsto D_a(p\|o)\) is differentiable, then for all \(r \ge 0\),
\[
sc(r | p \| o) = H^*(p \| o) = \sup_{a > 1} \frac{1}{a} [ r - D_a(p\|o) ].
\]
The source material lists Markov chains, finitely correlated states, and temperature states of non-interacting fermions and bosons as examples to which this approach applies, while the abstract highlights temperature states of non-interacting fermionic lattice systems and classical irreducible Markov chains [1407.3567].

Under the factorization-property route, Theorem IV.13 states that if both \(\{p_n\}\) and \(\{o_n\}\) satisfy the factorization property, then for all \(r\),
\[
sc(r|p\|o) = H^*(p\|o).
\]
The abstract gives temperature states of translation-invariant finite-range interactions on a spin chain as typical examples. This route is tailored to correlated many-body systems in which block tensor products control the full sequence up to \(n\)-independent multiplicative factors [1407.3567].

These two methods are structurally different. One exploits a quasi-product structure; the other relies on regularized Rényi smoothness. A plausible implication is that the anti-divergence is robust under different kinds of correlation, provided the asymptotic Rényi theory remains sufficiently regular.

## 4. Set-valued and composite extensions

The set-valued extension treats each hypothesis as a sequence of sets \(\mathscr A_n,\mathscr B_n \subset \mathscr D(\mathcal H^{\otimes n})\), thereby encompassing composite, adversarial, and correlated testing scenarios. The framework assumes stable sequences of convex compact sets, with stability understood as closure under tensor products, and the regularized anti-divergence is formulated by optimizing over the hardest pairs in the sets or, equivalently for convex compact sets, by inserting the set Rényi divergence into the \(\alpha\)-optimization [2508.12901].

The main strong-converse statement in this setting is Theorem 14:
\[
\liminf_{n\to \infty} -\frac{1}{n} \log (1-\alpha_{n, r}(\mathscr{A}_n \| \mathscr{B}_n)) \geq H_{r}^{*,\infty}(\mathscr{A}\|\mathscr{B}),
\]
where \(\alpha_{n,r}\) is the minimal worst-case type-I error under the constraint that the worst-case type-II error is at most \(2^{-nr}\). The paper therefore provides a lower bound on the strong converse exponent in terms of the regularized quantum Hoeffding anti-divergence, rather than the equality statement available for the correlated singleton sequences treated in the earlier work [2508.12901].

The same source emphasizes a sharp transition at the regularized relative entropy. For \(r < D^{\infty}(\mathscr{A}\|\mathscr{B})\), the regularized Hoeffding divergence governs the achievable type-I exponent. For \(r > D^{\infty}(\mathscr{A}\|\mathscr{B})\), the anti-divergence governs the exponential approach of type-I error to \(1\). This refines the generalized quantum Stein’s lemma for composite correlated hypotheses [2508.12901].

| Framework | Structural assumptions | Asymptotic anti-divergence statement |
|---|---|---|
| Correlated state sequences \(\{p_n\},\{o_n\}\) | Factorization property, or existence and differentiability of regularized sandwiched Rényi divergences | \(sc(r|p\|o)=H^*(p\|o)\) |
| Stable sequences of convex compact sets \(\mathscr A,\mathscr B\) | Stability under tensor products; convex compactness for minimax equalities | \(\liminf -\frac1n\log(1-\alpha_{n,r}) \ge H_r^{*,\infty}(\mathscr A\|\mathscr B)\) |

A common misconception is that composite generalization merely replaces single states by worst-case pairs without changing the asymptotic structure. The set-based theory shows that regularization, stability, and minimax equivalences are central, and that the strongest available conclusion in the cited work is a lower bound in the strong converse regime rather than a universal exact formula [2508.12901].

## 5. Explicit models and computable instances

The earlier work gives physically relevant correlated models in which the anti-divergence can be identified explicitly. For translation-invariant finite-range Gibbs states on a spin chain, the finite-volume Gibbs state is written as
\[
\omega_n(A) := \mathrm{Tr}(A e^{-\beta H_n})/Z_n.
\]
These states satisfy the factorization property, so Theorem IV.13 applies. Theorem V.2 states that for any two such Gibbs states at temperatures \(\beta_1,\beta_2\), and for any \(r\),
\[
sc(r|p\|o) = H^*(p\|o).
\]
This provides a direct strong-converse formula for a class of correlated thermal states [1407.3567].

For translation-invariant quasi-free fermionic lattice states, the source material considers symbols \(Q\) and \(R\) describing states \(w_Q\) and \(w_R\). Under the assumptions \(cI \leq Q, R \leq (1-c)I\), the regularized sandwiched Rényi divergences exist and are explicitly computable using generalized Szegő’s theorem for block Toeplitz operators. Theorem V.5 states that for all \(r\),
\[
sc(r | w_Q \| w_R) = H^*(w_Q \| w_R) = \sup_{a > 1} \frac{1}{a} \left[ r - D_a(w_Q \| w_R) \right],
\]
where
\[
D_a(w_Q\|w_R) = \int_{[0,2\pi)^v} d^v x\ \frac{1}{2\pi} \log\left(q(x)^a r(x)^{1-a} + (1 - q(x))^a (1 - r(x))^{1-a} \right).
\]
Here \(q(x)\) and \(r(x)\) are the symbols, i.e. Fourier transforms, of \(Q\) and \(R\) [1407.3567].

The same framework also recovers known results for classical i.i.d. models and irreducible Markov chains, as noted in Appendix D of the paper. This places the regularized quantum Hoeffding anti-divergence on a continuum from classical correlated processes to quantum many-body systems [1407.3567].

## 6. Relation to other divergences and later developments

The anti-divergence sits within a broader split between the direct and strong-converse regimes. One later perspective proves that a “parent” quantum Rényi divergence obtained from an integral representation regularizes to the Petz Rényi divergence for \(0<\alpha<1\) and to the sandwiched Rényi divergence for \(1<\alpha<\infty\). The source explicitly notes that this split matches operational regimes: Petz governs error exponents for \(\alpha<1\), while sandwiched governs strong converse exponents for \(\alpha>1\) [2306.12343].

A closely related operational observation is that, for \(\alpha>1\), the regularized measured Rényi divergence and the regularized test-measured Rényi divergence both coincide with the sandwiched Rényi divergence, and sequences of two-outcome measurements suffice to attain the regularized value. For \(\alpha<1\), by contrast, the regularized test-measured quantity is in general strictly smaller, even for commuting states. This asymmetry clarifies why sandwiched Rényi divergences appear naturally on the anti-divergence side of binary testing [2201.05477].

Computationally, a convex-optimization-based divergence \(D_\alpha^\#\) for \(\alpha>1\) has the property that its regularization equals the sandwiched Rényi divergence. That result yields a converging hierarchy of upper bounds on regularized sandwiched Rényi divergences for channels and is used to characterize the strong converse exponent for channel discrimination in terms of
\[
\sup_{\alpha > 1} \frac{\alpha-1}{\alpha}\big(r - \widetilde{D}^{\mathrm{reg}}_\alpha(\mathcal N\|\mathcal M)\big).
\]
This suggests a computational route to strong-converse quantities structurally analogous to regularized Hoeffding anti-divergence [2007.12576].

The composite setting also shows that the governing divergence need not be invariant under changes in hypothesis structure. In a different composite Hoeffding problem, involving a thermal equilibrium state versus a probe state subject to unknown dephasing in the energy eigenbasis, the optimal Hoeffding exponent is given exactly by the reverse sandwiched Rényi divergence on a single copy. The source emphasizes that passing from simple to composite hypotheses can fundamentally change which quantum divergence determines the operational limit of discrimination [2605.02203].

A further extension allows an inconclusive outcome in hypothesis testing. There, anti-divergence quantities built from sandwiched Rényi divergences quantify the rate at which the probability of conclusive outcomes must decay when one targets error exponents beyond those achievable with vanishing inconclusiveness. This preserves the strong-converse character of anti-divergence, but in a postselected testing geometry [2510.07601].

Taken together, these developments support a precise picture. Regularized quantum Hoeffding anti-divergence is the canonical asymptotic quantity for the strong-converse side of asymmetric discrimination when regularized \(\alpha>1\) Rényi theory is the relevant asymptotic language. What varies across frameworks is not the existence of a strong-converse object, but the structural assumptions under which it is exact, the level at which regularization must be performed, and the particular Rényi divergence that survives as the operationally controlling limit.

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