Regularized Quantum Chernoff Divergence

• Regularized Quantum Chernoff Divergence is an asymptotic measure that extends the standard Chernoff criterion to account for composite and correlated quantum hypotheses.
• It employs blocklength regularization to address non-multiplicative scenarios, ensuring precise error exponent rates in advanced quantum state discrimination.
• Its operational significance is evident in applications like QKD advantage distillation and composite hypothesis testing, although a direct one-shot channel divergence remains unresolved.

Searching arXiv for relevant papers on regularized quantum Chernoff divergence and closely related Chernoff asymptotics. Regularized quantum Chernoff divergence is an asymptotic distinguishability quantity that arises when the ordinary single-letter quantum Chernoff divergence is insufficient to characterize symmetric quantum hypothesis testing. In its standard state form, the quantum Chernoff divergence between two states is

$$C(\rho,\sigma)=-\log \inf_{0\le s\le 1}\operatorname{Tr}[\rho^s\sigma^{1-s}],$$

equivalently $$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$ (Li, 2015). In i.i.d. state discrimination this quantity is already single-letter and additive on tensor powers, so any natural regularization collapses to the same value (Stasiuk et al., 2022, Li, 2015). By contrast, in broader settings—most notably composite or correlated hypotheses built from sets of states—the asymptotic symmetric error exponent is governed by a genuinely regularized quantity defined from blocklength-$n$ Chernoff divergences and then normalized by $1/n$ (Fang, 18 Aug 2025). For quantum channels, an operational many-use Chernoff exponent exists, but it is not identified in general with a regularized one-shot channel Chernoff divergence (1705.01642).

1. Standard quantum Chernoff divergence and its asymptotic role

For two density operators $\rho,\sigma$, the basic quantum Chernoff quantity is defined as

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$

with the paper on advantage distillation emphasizing the non-logarithmic form

$$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$

so that $\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$ (Stasiuk et al., 2022). Closely related formulations use

$$\xi_{QCB}(\rho_1,\rho_2):= -\log \inf_{0\le s\le 1}\operatorname{tr}\!\left[\rho_1^{\,1-s}\rho_2^{\,s}\right]$$

or

$$C(\rho_i,\rho_j):=\max_{0\le s\le 1}\{-\log\operatorname{Tr}\rho_i^s\rho_j^{1-s}\},$$

which are equivalent notational variants because $$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$0 is decreasing (Nussbaum, 2013, Li, 2015).

Its fundamental operational meaning is the optimal asymptotic Bayesian error exponent in binary discrimination of tensor-power states. In the binary case, the Holevo–Helstrom test satisfies

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$1

(Nussbaum, 2013). The same asymptotic role is expressed in the advantage-distillation setting through

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$2

which identifies the Chernoff divergence with the exponential rate governing the distinguishability of many i.i.d. copies (Stasiuk et al., 2022). A related exposition in polarization theory recalls the same asymptotic form,

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$3

with

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$4

(Ghiu et al., 2010).

This standard single-letter divergence is therefore already an asymptotic per-copy quantity in the i.i.d. binary-state setting. The need for regularization appears only when this single-letter structure ceases to be sufficient.

2. Why regularization is trivial for i.i.d. tensor-power states

In the i.i.d. setting, regularization is mathematically unnecessary because the Chernoff quantity is multiplicative on tensor powers. The advantage-distillation paper states explicitly that

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$5

and therefore the logarithmic Chernoff divergence is additive,

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$6

(Stasiuk et al., 2022). The same conclusion is implicit in the multiple-state discrimination analysis, where tensor-power functional calculus yields

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$7

hence

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$8

(Li, 2015).

A plausible implication is that any natural IID regularization of the form

$$C(\rho,\sigma)=\max_{0\le s\le 1}\{-\log \operatorname{Tr}\rho^s\sigma^{1-s}\}$$9

must equal the ordinary single-copy Chernoff divergence. This inference is stated directly in the QKD work for the $n$0-representation: because $n$1 is multiplicative on tensor powers, “any regularization would collapse to the single-copy quantity” in the IID collective-attacks model (Stasiuk et al., 2022).

This collapse is also reflected in operational theorems beyond the binary case. For multiple i.i.d. hypotheses $n$2, the exact asymptotic exponent is

$n$3

the multiple quantum Chernoff distance (Li, 2015). Thus the large-$n$4 asymptotics remain single-letter even in the multi-hypothesis i.i.d. regime.

3. Multiple-state Chernoff distances and the single-letter regime

For an ensemble $n$5, the multiple quantum Chernoff bound is defined as

$n$6

(Nussbaum, 2013), equivalently

$n$7

(Li, 2015). This quantity is the worst pairwise Chernoff distance and acts as the benchmark asymptotic exponent in multiple-state Bayesian discrimination.

Earlier results established only upper bounds or partial attainability for mixed-state ensembles. One paper recalled the general upper bound

$n$8

and a known universal lower bound of $n$9 (Nussbaum, 2013). It then proved exact attainability under the separation condition

$1/n$0

for some least favorable pair (Nussbaum, 2013). Subsequently, the general finite-dimensional i.i.d. problem was solved completely: for arbitrary priors independent of $1/n$1,

$1/n$2

(Li, 2015).

These results are directly relevant to regularization because they show where regularization is not needed. In the finite-dimensional i.i.d. state regime, the asymptotic exponent is exactly the single-letter minimum pairwise quantum Chernoff distance. No separate regularized divergence must be introduced to obtain the correct exponent (Li, 2015).

4. Genuine regularization for sets of states and correlated hypotheses

A genuinely regularized quantum Chernoff divergence appears when the hypotheses are not fixed states but sequences of sets of states, possibly with arbitrary correlations. In that setting, one no longer has exact additivity, and the blocklength-$1/n$3 Chernoff divergence must be normalized and regularized across $1/n$4 (Fang, 18 Aug 2025).

For sets $1/n$5, the set Chernoff divergence is defined by

$1/n$6

For sequences $1/n$7 and $1/n$8, the regularized quantities are

$1/n$9

and when the limit exists,

$\rho,\sigma$0

(Fang, 18 Aug 2025).

The associated quasi-divergence between sets is

$\rho,\sigma$1

and its regularized version is

$\rho,\sigma$2

when the limit exists (Fang, 18 Aug 2025).

Under stability under tensor product, regularization is not merely formal but structurally necessary. The paper recalls that the ordinary Chernoff divergence is additive on identical tensor powers but only subadditive in general tensor products: $\rho,\sigma$3 and explicitly notes that nonadditivity occurs for set-based problems arising in resource theory (Fang, 18 Aug 2025). Under the stability assumption, one obtains the Fekete-type identity

$\rho,\sigma$4

(Fang, 18 Aug 2025).

This is the clearest formal notion of a regularized quantum Chernoff divergence in the available literature: an asymptotic blocklength-regularized worst-case Chernoff rate for stable sequences of convex compact sets of states.

5. Minimax structure and operational characterization

The set-based theory has a minimax structure that parallels the ordinary single-state Chernoff formula. For convex sets,

$\rho,\sigma$5

and for stable sequences of convex sets,

$\rho,\sigma$6

(Fang, 18 Aug 2025). The finite-block identity uses Lieb concavity and a minimax interchange over states and $\rho,\sigma$7, while the regularized identity uses a second asymptotic minimax interchange over blocklength and $\rho,\sigma$8 (Fang, 18 Aug 2025).

The operational setting is a worst-case Bayesian discrimination problem over sets. For $\rho,\sigma$9 hypotheses represented by sets $$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$0 and priors $$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$1, the minimum error probability is

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$2

A central minimax lemma states

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$3

for convex sets, with attainment for compact sets (Fang, 18 Aug 2025). This means discriminating sets of states is no harder than discriminating their worst-case elements.

For binary composite hypotheses $$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$4, $$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$5, the exact asymptotic theorem is

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$6

under finite-dimensionality and the assumptions that the sequences are stable, convex, and compact (Fang, 18 Aug 2025). For multiple composite hypotheses, the corresponding exponent is the regularized pairwise minimum

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$7

under an additional permutation-invariance assumption in the achievability proof (Fang, 18 Aug 2025).

This establishes the regularized quantum Chernoff divergence as an exact asymptotic symmetric hypothesis-testing exponent in the composite correlated setting. In this sense it is the proper extension of the ordinary quantum Chernoff divergence beyond i.i.d. simple hypotheses.

6. Channel discrimination, operational exponents, and non-equivalence to a regularized channel divergence

For quantum channels, the asymptotic symmetric/Bayesian exponent has been studied operationally, but the available channel results stop short of defining a general one-shot channel Chernoff divergence whose regularization equals the operational exponent (1705.01642). The channel-discrimination setting allows completely general $$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$8-use strategies: adaptive sequential use, arbitrary ancillas, memory, interleaving quantum operations, and joint final measurements. The paper defines

$$\xi(\rho, \sigma) \coloneqq -\log \left( \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right) \right),$$9

where $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$0 is the infimum average Bayesian error over all achievable output-state pairs under such strategies (1705.01642).

The central theorem is structural: $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$1 Equivalently, if finite-use perfect discrimination is possible then $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$2 for all sufficiently large $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$3 and $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$4; otherwise the error decays at most exponentially, ruling out super-exponential decay (1705.01642). The proof gives explicit upper bounds

$$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$5

depending on which obstruction to perfect finite-use discrimination applies (1705.01642).

What is crucial for the notion of regularized quantum Chernoff divergence is the paper’s explicit limitation: it “does not define a channel analogue of the state quantum Chernoff divergence” and does not prove that the operational exponent equals a regularized one-shot channel Chernoff quantity (1705.01642). The many-use optimization is built directly into the definition of the exponent. Thus, for channels, one may speak of an inherently many-use operational Chernoff rate, but not of a formally identified regularized quantum Chernoff divergence in the same sense as for stable sequences of state sets.

A related modern perspective comes from the study of test-measured Rényi divergences. For $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$6, the regularized test-measured Rényi divergences generally fail to recover the standard Rényi divergences, even in the classical case, but at the special point $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$7 one has

$$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$8

so the regularized test-measured quantity at $$Q(\rho, \sigma) \coloneqq \inf_{0 < s < 1} Tr \left( \rho^{s} \sigma^{1-s} \right),$$9 does recover the standard Chernoff divergence (Mosonyi et al., 2022). This sharp distinction suggests that not every regularized measurement-restricted Rényi construction yields the correct Chernoff theory, although the Chernoff point itself is exceptional (Mosonyi et al., 2022).

7. Applications, interpretations, and conceptual boundaries

One application of the state Chernoff divergence appears in advantage distillation for QKD and DIQKD. In the IID collective-attacks model, Eve’s relevant conditional states become tensor powers $\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$0 and $\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$1, and the security threshold is governed by

$\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$2

with matching sufficient and necessary conditions up to an extra structural assumption in the necessary direction (Stasiuk et al., 2022). The important point for regularization is that the protocol reduces directly to tensor-power discrimination, so the ordinary single-copy Chernoff quantity already captures the many-copy asymptotics (Stasiuk et al., 2022).

Another application appears in generalized composite testing and resource theory. For a pure state $\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$3 and a convex set of free states $\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$4, the maximum overlap

$\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$5

satisfies

$\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$6

providing an operational meaning for this overlap in symmetric hypothesis testing (Fang, 18 Aug 2025). The same paper lists regularized identities for specific free sets such as stabilizer, incoherent, separable, and PPT states (Fang, 18 Aug 2025). This shows that regularized Chernoff divergences naturally encode asymptotic resource distinguishability when nonadditivity prevents a purely single-letter description.

A further conceptual boundary is provided by new Rényi-type constructions that could support Chernoff-style optimization but do not yet deliver a Chernoff theory. A recent cumulant-based quantum relative Rényi functional introduces a regularized full-rank extension

$\xi(\rho,\sigma)=-\log Q(\rho,\sigma)$7

but it does not define a Chernoff divergence or prove a binary symmetric-testing exponent theorem (Meunson et al., 30 Jun 2026). This suggests a broader landscape in which regularized Rényi-like functionals may furnish ingredients for Chernoff-type quantities without yet establishing operational Chernoff meanings.

A common misconception is that “regularized quantum Chernoff divergence” is always just a synonymous reformulation of the ordinary state Chernoff divergence. The literature supports a more differentiated view. In i.i.d. simple-state discrimination, regularization collapses trivially to the single-letter divergence (Stasiuk et al., 2022, Li, 2015). In composite or correlated state-set discrimination, regularization is essential and defines the exact asymptotic exponent (Fang, 18 Aug 2025). For channels, the operational many-use exponent is known, but it is not generally identified with a regularized one-shot channel Chernoff divergence (1705.01642).

In summary, the regularized quantum Chernoff divergence is best understood as a family of closely related notions rather than a single universal formula. In the ordinary i.i.d. state setting it coincides with the standard quantum Chernoff divergence because of multiplicativity on tensor powers. In the composite correlated setting it becomes a genuine blocklength-regularized worst-case asymptotic rate. For channels, the closest available object is an operational many-use Chernoff exponent whose formal divergence interpretation remains incomplete.