Belavkin–Staszewski Channel Divergence

• Belavkin–Staszewski channel divergence is a maximal f‐divergence defined via an operator logarithm formulation that replaces log(ρ) – log(σ) with log(ρ^(1/2)σ^(–1)ρ^(1/2)), ensuring it exceeds Umegaki’s entropy for non-commuting states.
• It underpins rigorous channel chain rules and data-processing inequalities, providing a structural framework for recovery maps and upper bounds in adaptive channel discrimination and many-body applications.
• Applications of BS divergence include energy-constrained bosonic channel discrimination, Gibbs state analysis, and quantifying information loss under local coarse-graining in quantum systems.

Searching arXiv for recent and foundational papers on Belavkin–Staszewski divergence, maximal $f$-divergences, and channel chain rules. Belavkin–Staszewski channel divergence is the channel-level use of the Belavkin–Staszewski relative entropy, a maximal $f$-divergence that refines the Umegaki relative entropy by replacing $\log\rho-\log\sigma$ with $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$. In the finite-dimensional setting, the underlying state divergence is

$$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$

with the usual support condition, and it coincides with the $\alpha\to1$ limit of the geometric Rényi divergence (Bergh et al., 2021). A distinct point in the literature is that some works use the Belavkin–Staszewski divergence only through data-processing inequalities for channels, without defining an independent channel functional, whereas other works explicitly adopt the standard channel-extension paradigm

$$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$

possibly stabilized over ancillas (Bluhm et al., 2021). The resulting notion connects maximal $f$-divergences, chain rules for channels, recovery-theoretic saturation of data processing, approximate Markov structure, and, more recently, energy-constrained bosonic discrimination (Huang et al., 20 Mar 2026).

1. State-level origin and maximal $f$-divergence structure

The state-level Belavkin–Staszewski relative entropy is the maximal $f$-divergence corresponding to $f$0. In finite dimensions, for strictly positive operators $f$1,

$f$2

and the Belavkin–Staszewski case is

$f$3

so the usual BS relative entropy is a special case of a maximal $f$4-divergence (Gaál et al., 2017).

This maximality is structural. Among divergences that reduce to the classical $f$5-divergence on commuting density operators and are monotone under CPTP maps, the maximal $f$6-divergence is pointwise maximal (Gaál et al., 2017). In the Belavkin–Staszewski specialization this implies that BS divergence is larger than, or equal to, the Umegaki relative entropy, with equality only in the commuting case: $f$7 That comparison is repeatedly exploited in many-body and channel applications, because any upper bound on a BS-based quantity immediately bounds the corresponding Umegaki quantity (Bluhm et al., 2021).

A second foundational representation is via the geometric Rényi divergence

$f$8

Its limit at $f$9 is exactly the Belavkin–Staszewski relative entropy (Berta et al., 2022). The same limit also appears from the sharp Rényi divergence: $\log\rho-\log\sigma$0 as $\log\rho-\log\sigma$1 (Bergh et al., 2021). This places BS divergence at the interface of maximal $\log\rho-\log\sigma$2-divergences, geometric Rényi divergences, and sharp Rényi channel methods.

A further characterization uses pure-state ensemble realizations. The quantity

$\log\rho-\log\sigma$3

was shown to coincide with $\log\rho-\log\sigma$4 for faithful states, with the infimum attained by measures supported on a common, possibly non-orthogonal, basis (Ortigueira et al., 28 Nov 2025). This identifies BS divergence as a minimal classical KL divergence over pure-state ensemble realizations, sharpening the reverse-test perspective for maximal $\log\rho-\log\sigma$5-divergences.

2. Definitions of channel divergence and competing conventions

The standard channel-extension convention takes a state divergence $\log\rho-\log\sigma$6 and defines

$\log\rho-\log\sigma$7

with a stabilized variant obtained by optimizing over $\log\rho-\log\sigma$8 and applying $\log\rho-\log\sigma$9 and $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$0 (Berta et al., 2022). Under this convention, the Belavkin–Staszewski channel divergence is the $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$1 specialization of the geometric Rényi channel divergence: $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$2 The same framework naturally yields stabilized and regularized BS channel divergences (Berta et al., 2022).

Not all works introduce a separate channel functional. In the Gibbs-state literature, channels appear primarily through data processing. There, one studies the deficit

$\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$3

for a CPTP map $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$4, interprets equality through BS recovery, and sometimes refers to this state-pair–dependent quantity as the relevant channel notion, rather than defining a supremum over inputs (Bluhm et al., 2021). This distinction is important: the literature contains both a “supremum-over-inputs” channel divergence and a “DPI-deficit along a fixed channel” usage.

For finite-dimensional channels, a Choi-operator form is available. The unconstrained BS channel divergence can be written as

$\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$5

where $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$6 is the operator relative entropy and $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$7 are the Choi operators (Huang et al., 20 Mar 2026). This formula is the direct channel analogue of the state-level operator-relative-entropy representation.

The axiomatic perspective on channel divergences supplies a broader frame. A channel divergence is any function on pairs of channels that is monotone under superchannels, and a channel relative entropy additionally satisfies additivity and normalization (Gour, 2020). That work does not treat BS explicitly, but it implies a classification principle: any admissible BS-type channel relative entropy reducing to classical KL on classical channels must lie between minimal and maximal channel extensions and, in the classical case, must collapse to the unique KL channel divergence (Gour, 2020). This suggests that BS channel divergence is non-unique at the quantum channel level unless one fixes a specific extension convention.

3. Chain rules, data processing, and composition inequalities

A central development is the channel chain rule for geometric Rényi divergences. For positive maps $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$8 and states $\log(\rho^{1/2}\sigma^{-1}\rho^{1/2})$9,

$$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$0

Taking $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$1 yields the corresponding Belavkin–Staszewski chain rule

$$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$2

at least under the continuity assumptions standard for the geometric family (Berta et al., 2022). This places BS channel divergence in the same formal class as sandwiched and geometric Rényi channel divergences.

The same framework gives a stabilized chain rule and, by specialization to composition, a subadditivity inequality

$$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$3

This is the natural composition law for BS channel divergence and is directly relevant to adaptive protocols, circuit layers, and sequential channel simulation (Berta et al., 2022).

The amortized variant also fits this pattern. For the geometric Rényi family, the amortized channel divergence equals the stabilized one, and the same structure is expected at $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$4: $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$5 again under the same continuity passage from $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$6 to $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$7 (Berta et al., 2022). A plausible implication is that amortization does not increase BS distinguishability beyond entangled single-use optimization, but that statement is presented in the source as an $$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$8 specialization rather than as an independently re-proved BS theorem.

The many-body literature emphasizes a different but related chain-rule mechanism. For a conditional expectation or a partial trace, the quantity

$$\widehat D(\rho\Vert\sigma) = \operatorname{Tr}\!\left[\rho\,\log\!\left(\rho^{1/2}\sigma^{-1}\rho^{1/2}\right)\right],$$9

is treated as a channel-induced BS loss (Capel et al., 2024). This object governs conditional mutual informations, approximate recovery, and Markovianity, and is therefore a second operationally important sense in which “Belavkin–Staszewski channel divergence” appears.

4. Equality, recovery maps, and saturation of the BS data-processing inequality

The data-processing inequality for BS entropy,

$\alpha\to1$0

admits a precise equality characterization. Equality holds if and only if

$\alpha\to1$1

which defines the BS recovery map

$\alpha\to1$2

This map is trace-preserving but generally not positive (Bluhm et al., 2019).

The strengthened data-processing inequality quantifies deviation from that recovery condition. For arbitrary quantum channels $\alpha\to1$3, one has a lower bound of the form

$\alpha\to1$4

with $\alpha\to1$5 (Bluhm et al., 2019). Thus the BS-DPI deficit controls the Hilbert–Schmidt distance to BS recoverability.

A complementary upper bound was later derived for conditional expectations and extended to arbitrary channels. For states $\alpha\to1$6 and a channel $\alpha\to1$7,

$\alpha\to1$8

is bounded above by an operator-norm expression involving the mismatch between $\alpha\to1$9 and its processed version through $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$0 (Capel et al., 2024). Together with the lower bound, this sandwiches BS-DPI loss between explicit recovery-type quantities.

The tripartite specialization yields a BS analogue of quantum Markov chains. For $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$1, equality of BS-CMI is equivalent to

$$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$2

and this defines a Belavkin–Staszewski quantum Markov chain (Bluhm et al., 16 Jan 2025). The associated symmetric BS recovery map is completely positive, though not trace-preserving, and acts as an exact recovery map on BS-Markov states (Bluhm et al., 16 Jan 2025).

This recovery perspective matters at channel level because any BS channel divergence inherits its structural rigidity from the state-level equality theory. If a channel-level supremum is attained at an input $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$3, then saturation under post-processing reduces to equality of BS-DPI for the corresponding output states, hence to BS recoverability of that state pair. This suggests that BS channel reversibility is mediated by the same recovery map that governs state-level equality.

5. Continuity, structural bounds, and comparison with other channel divergences

BS entropy is jointly convex and satisfies continuity bounds on full-rank domains, but its stability is more delicate than Umegaki’s. For fixed full-rank $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$4,

$$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$5

where $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$6 and $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$7 is the minimal eigenvalue of $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$8 (Bluhm et al., 2023). The explicit $$D(\mathcal N\Vert\mathcal M):=\sup_\rho D(\mathcal N(\rho)\Vert \mathcal M(\rho)),$$9 dependence reflects the singular sensitivity of BS-based quantities near rank deficiency.

For derived quantities, BS-conditional entropy is discontinuous on rank-deficient states, while BS-mutual information can diverge with $f$0 or $f$1 (Bluhm et al., 2022). Consequently, continuity statements for BS channel functionals usually require full-rank output constraints or work on compact full-rank subsets.

The same ALAFF-based continuity machinery can nevertheless be lifted to optimized channel quantities. A notable example is the BS-Rains information of a channel, which satisfies an explicit continuity bound under perturbations of the channel in $f$2 norm (Bluhm et al., 2022). This shows that BS-based channel upper bounds in capacity theory vary continuously under small channel perturbations, despite the stronger eigenvalue sensitivity at state level.

Comparison inequalities organize BS channel divergence within a hierarchy. Since

$f$3

the same ordering lifts to the corresponding channel divergences: $f$4 Likewise, BS is bounded above by geometric Rényi divergences of order $f$5 (Huang et al., 20 Mar 2026). This makes BS channel divergence an intermediate object: stronger than Umegaki channel relative entropy, but below higher-order geometric Rényi channel divergences.

The preserver theory for maximal $f$6-divergences adds a rigidity statement. Any bijection on positive operators preserving the BS divergence must be a unitary or antiunitary conjugation (Gaál et al., 2017). A plausible implication is that exact output-space symmetries of BS channel divergence are correspondingly rigid, though the paper formulates that result at state level.

6. Applications: Gibbs states, learning, ETH, and bosonic channel discrimination

In one-dimensional Gibbs states, BS entropy has turned out to be especially effective because it admits operator-norm control: $f$7 This bound converts locality estimates on Gibbs marginals directly into bounds on BS mutual information and hence on ordinary mutual information (Bluhm et al., 2021). The resulting exponential decay of BS mutual information and superexponential approximate BS recoverability under partial trace make BS divergence a technical bridge between Araki-type operator estimates and information-theoretic decay statements.

That same viewpoint was sharpened into BS conditional mutual informations for Gibbs states. In translation-invariant 1D systems, the one-sided, two-sided, and reversed BS-CMIs decay superexponentially with the size of the buffer region, and the associated symmetric BS recovery maps can be concatenated to build MPO approximations and efficient learning procedures from local data (Capel et al., 2024). Here the “channel divergence” is the BS loss across partial trace or conditional expectation, rather than a supremum over channel inputs.

The subsystem ETH literature uses a related strategy. It connects quantum variance to BS relative entropy, proves small BS entropy between local eigenstate reductions and thermal reductions, and thereby derives subsystem ETH bounds with algebraic speed of convergence in translation-invariant systems (Huang et al., 2023). This shows that BS divergence can quantify local indistinguishability between dynamical many-body states and Gibbs states.

The most explicit recent channel-discrimination application is energy-constrained bosonic discrimination. For channels $f$8 and an input Hamiltonian $f$9, the energy-constrained BS channel divergence is

$f$0

It admits the Choi-operator form

$f$1

and satisfies an energy-constrained chain rule strong enough to bound fully adaptive discrimination protocols (Huang et al., 20 Mar 2026).

Specifically, for the final states of an $f$2-use adaptive protocol,

$f$3

which yields the asymptotic error-exponent upper bound

$f$4

for energy-constrained asymmetric channel discrimination (Huang et al., 20 Mar 2026). In bosonic dephasing and loss-dephasing models, measured relative entropy, Umegaki relative entropy, and geometric Rényi divergence admit truncated SDP formulations, while the BS divergence is accessed through the $f$5 limit of geometric Rényi quantities (Huang et al., 20 Mar 2026).

These applications illustrate the two dominant operational roles of Belavkin–Staszewski channel divergence. In many-body theory it quantifies information loss under local coarse-graining and recovery. In channel discrimination it serves as a single-letter upper bound on the best adaptive error exponent, particularly under physically motivated constraints such as average energy.