Resource Theory of Asymmetric Distinguishability

• Resource theory of asymmetric distinguishability is a framework treating differences in quantum states as a usable resource under free operations.
• It employs quantum boxes—ordered pairs of states inspired by the Neyman–Pearson setting—with monotones like min-, max-, and quantum relative entropies to quantify resource value.
• The theory connects one-shot tasks such as distillation and dilution to asymptotic reversibility, showing that the quantum relative entropy governs the interconversion rate.

The resource theory of asymmetric distinguishability takes distinguishability itself as a resource and studies how an ordered pair of quantum states can be manipulated when the same physical operation is applied to both components. In its basic form, the resource object is a quantum box $(\rho,\sigma)$, interpreted as two different experimental scenarios: in the first, $\rho$ is prepared; in the second, $\sigma$ is prepared. The theory was systematically developed in 2019, building on an earlier proposal of Matsumoto, and identifies bits of asymmetric distinguishability as a canonical currency, exact and approximate one-shot operational tasks in terms of min- and max-relative entropies, and an asymptotically reversible regime in which the quantum relative entropy is the universal interconversion rate (Wang et al., 2019).

1. Quantum boxes and free manipulation

A quantum box is an ordered pair of quantum states acting on the same Hilbert space, $(\rho,\sigma)$. The order matters: the framework is asymmetric because only the first state is required to be reproduced exactly or approximately in the operational tasks, while the second state functions as a benchmark or noise state. This mirrors the Neyman–Pearson setting, in which one controls Type I error for one hypothesis and minimizes Type II error for the other (Wang et al., 2019).

The free operations are completely positive trace-preserving maps applied componentwise,

$$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$

This is free because the same channel can be applied without learning which state was prepared; no side information is used. Isometric channels and appending an ancillary state are reversible by free operations, in the sense described in the foundational formulation of the theory (Wang et al., 2019).

Any quantity $M(\rho,\sigma)$ obeying the data-processing inequality under CPTP maps,

$$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$

is a resource monotone. In the basic state theory, the relevant monotones include the quantum relative entropy, the min-relative entropy, the max-relative entropy, their smooth versions, and certain Rényi divergences. This places the theory squarely within the general resource-theoretic paradigm: convertibility is constrained by monotonicity, and operational tasks reveal which monotones are complete for which regimes (Wang et al., 2019).

A closely related semiring perspective replaces normalized state pairs by more general dichotomies, i.e. pairs of positive semidefinite operators with a preorder defined by a variant of relative submajorization. In that formulation, direct sum and tensor product make the set of equivalence classes of dichotomies into a commutative semiring, and asymptotic convertibility is characterized by spectral points that are additive under direct sum and multiplicative under tensor product (Perry et al., 2020).

2. Currency and central divergences

The unit resource is one bit of asymmetric distinguishability, given by the box

$\left(|0\rangle\langle 0|,\pi\right),$

where

$$\pi=\frac{|0\rangle\langle 0|+|1\rangle\langle 1|}{2}.$$

In the Neyman–Pearson asymmetric hypothesis testing interpretation, this currency captures “zero Type I error, controlled Type II error,” since measuring $\sigma_Z$ and declaring $\rho$0 when the outcome is $\rho$1 yields Type I error $\rho$2 and Type II error $\rho$3 (Wang et al., 2019).

More generally, the canonical compressed $\rho$4-bit box is

$\rho$5

For integer $\rho$6, this box is equivalent by free CPTP maps to

$\rho$7

and the equivalence holds for all real $\rho$8 as well. This normalized convention makes $\rho$9 the exact number of bits present (Wang et al., 2019).

The central divergences are the quantum relative entropy,

$\sigma$0

when $\sigma$1 and $\sigma$2 otherwise; the max-relative entropy,

$\sigma$3

and the min-relative entropy,

$\sigma$4

Their smooth one-shot versions are defined using normalized trace-distance smoothing. The smooth min-relative entropy is the hypothesis-testing divergence,

$\sigma$5

while the smooth max-relative entropy is

$\sigma$6

The paper works primarily with the normalized trace distance $\sigma$7 for smoothing, motivated by operational interpretability (Wang et al., 2019).

The Rényi relatives organize these quantities. The Petz Rényi divergence satisfies $\sigma$8 and $\sigma$9, while the sandwiched Rényi divergence satisfies $(\rho,\sigma)$0 and $(\rho,\sigma)$1. Both families are monotone in $(\rho,\sigma)$2 in the ranges stated in the foundational development (Wang et al., 2019).

3. One-shot distillation and dilution

The first pair of operational tasks are distillable distinguishability and distinguishability cost. In exact one-shot distillation, one asks for the maximum number of bits of asymmetric distinguishability that can be extracted from $(\rho,\sigma)$3 while reproducing the first output with no error. The exact answer is

$(\rho,\sigma)$4

This is realized by the measurement channel

$(\rho,\sigma)$5

Conversely, the exact one-shot distinguishability cost is

$(\rho,\sigma)$6

with an explicit realization from the operator inequality $(\rho,\sigma)$7 (Wang et al., 2019).

Allowing error $(\rho,\sigma)$8 in reproducing the first state leads to the approximate one-shot equalities

$(\rho,\sigma)$9

These equalities give the smooth min- and max-relative entropies a direct resource-theoretic meaning. The same work also shows that these quantities, and the approximate box transformation feasibility problem, admit semidefinite-program formulations, enabling efficient computation in finite dimensions (Wang et al., 2019).

Exact one-shot distillation and exact one-shot cost are additive on i.i.d. boxes:

$$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$0

At the same time, the exact one-shot theory is generally irreversible because $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$1 in general. A standard example is $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$2 and $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$3 with $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$4: then $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$5 is finite, while $$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$6 because the relevant support condition fails. This shows that finite exact distillation can coexist with infinite exact formation cost (Wang et al., 2019).

A later note reformulated these one-shot tasks in terms of error exponents and strong converse exponents. For distillation and dilution, the corresponding one-shot exponents can be evaluated by semidefinite programming, satisfy data-processing inequalities under positive trace-preserving maps, and are related to each other by explicit trade-off inequalities. In the i.i.d. regime, the distillation error exponent and strong converse exponent are expressed by variational formulas involving Petz and sandwiched Rényi divergences, respectively (Wilde, 2022).

4. Asymptotic reversibility and interconversion rates

The asymptotic i.i.d. regime is the central structural result of the theory. Defining

$$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$7

and

$$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$8

one obtains

$$\Lambda : (\rho,\sigma) \mapsto (\Lambda(\rho),\Lambda(\sigma)).$$9

Thus, in the asymptotic i.i.d. limit, the resource theory is reversible, and the quantum relative entropy is the universal rate (Wang et al., 2019).

The asymptotic equipartition property links the one-shot quantities to this reversible rate. For every fixed $M(\rho,\sigma)$0,

$M(\rho,\sigma)$1

The same development states second-order refinements

$M(\rho,\sigma)$2

with the $M(\rho,\sigma)$3 corrections identifiable through the relative entropy variance and normal approximation (Wang et al., 2019).

The notable application is the asymptotic conversion between two i.i.d. boxes. For $M(\rho,\sigma)$4 and $M(\rho,\sigma)$5 with the stated support conditions, the optimal asymptotic conversion rate is

$M(\rho,\sigma)$6

Achievability proceeds by distilling approximately $M(\rho,\sigma)$7 bits of asymmetric distinguishability from the source and then diluting those bits to the target. Converse bounds use data processing together with Rényi pseudo-continuity bounds and yield an exponential strong converse (Wang et al., 2019).

The semiring treatment of dichotomies gives a different asymptotic regime. There, the asymptotic conversion rate in the strong-converse sense is

$M(\rho,\sigma)$8

so the spectral points governing the asymptotic preorder are exactly the sandwiched Rényi quasi-entropies of order $M(\rho,\sigma)$9 and $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$0 (Perry et al., 2020). This does not contradict the reversible $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$1 theory; it identifies a distinct strong-converse asymptotic preorder rather than the vanishing-error asymptotic rate of the original reversible formulation.

5. Extensions: channels, strategies, catalysis, and partial information

The state theory was extended to quantum channel boxes, where the basic object is a pair of channels $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$2 and the free operations are arbitrary quantum superchannels. One-shot exact and approximate distillation and dilution again coincide with channel min- and max-relative entropies and their smooth versions. In asymptotic settings, however, the structure is richer than for states: exact sequential cost equals the channel max-relative entropy, sequential distillable distinguishability equals the amortized channel relative entropy, parallel distillation is given by the regularized channel relative entropy, and the theory simplifies for environment-seizable and classical–quantum channel boxes (Wang et al., 2019).

The same pattern extends to quantum strategies or combs. In that setting, the one-shot operational quantities are the smooth strategy min- and max-relative entropies, and semidefinite programs characterize the normalized strategy distance, distillable distinguishability, and distinguishability cost. Numerical work on generalized amplitude damping channels showed significant adaptive advantages over non-adaptive strategies for finite numbers of channel uses (Katariya et al., 2020).

A different single-shot extension introduces catalytic relative majorization with correlations. For commuting pairs of states, catalytic convertibility with a catalyst returned locally unchanged is possible if and only if the Umegaki relative entropy does not increase:

$$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$3

under the quasi-classical assumptions stated in the theorem. This gives relative entropy a single-shot catalytic operational meaning beyond its i.i.d. role in Stein’s lemma (Rethinasamy et al., 2019).

Two later generalizations broaden the objects beyond single pairs. One replaces pairs by multiple-state boxes $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$4, with a CPTNI relative submajorization preorder characterized asymptotically by componentwise sandwiched Rényi divergences. In that setting, the strong converse exponent for a composite null hypothesis against a simple alternative is the maximum of the corresponding pairwise exponents (Bunth et al., 2020). Another replaces a single pair by sets of states $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$5, thereby treating partial information and composite hypotheses. For convex sets, one-shot approximate distillation and dilution are given by worst-case smoothed min- and max-divergences, while asymptotic distillation and dilution are both governed by a regularized divergence $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$6, yielding an asymptotically reversible interconversion theory for state sets (Yao et al., 2 Oct 2025).

For process tensors and more general quantum combs, a technical issue arises: Choi divergences need not satisfy the relevant data-processing inequality under superprocesses, whereas generalized comb divergences

$$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$7

do satisfy it. This makes generalized divergences, rather than Choi divergences, the natural candidates for monotones in resource theories of asymmetric distinguishability for multitime processes (Zambon, 2024).

6. Hypothesis testing, asymmetry, and broader formulations

The original operational meaning of the theory is asymmetric hypothesis testing. Given null hypothesis $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$8 and alternative $$M(\rho,\sigma) \ge M(\Lambda(\rho),\Lambda(\sigma)),$$9, the minimal Type II error $\left(|0\rangle\langle 0|,\pi\right),$0 achievable with Type I error $\left(|0\rangle\langle 0|,\pi\right),$1 satisfies

$\left(|0\rangle\langle 0|,\pi\right),$2

Within the resource theory, this is exactly the distillation rate: the number of currency bits that can be extracted from $\left(|0\rangle\langle 0|,\pi\right),$3 is governed by $\left(|0\rangle\langle 0|,\pi\right),$4, and the cost to simulate the pair from currency bits is governed by the same quantity in the asymptotic limit (Wang et al., 2019).

This relation to asymmetric distinguishability also appears in other resource-theoretic settings. In the resource theory of asymmetry underlying the Wigner–Araki–Yanase theorem, asymmetric distinguishability quantifies how well the orbit of a state under a symmetry group encodes the group element under symmetric processing. Perfect symmetric simulation of an asymmetric projective measurement is possible only if the resource state is perfectly asymmetric, meaning that its orbit states are mutually orthogonal. The information-theoretic account of the WAY theorem therefore treats asymmetry as an encoding resource and connects it to measurement programmability and discrimination bounds [(Marvian et al., 2012); (Ahmadi et al., 2012)].

An abstract categorical formulation casts resource theories as quantale modules and includes a resource theory of distinguishability in which hypothesis-independent post-processings are free. In that setting, cost, yield, robustness, weight, and relative-entropy-based “minimal distinguishability from free resources” all arise from general monotone constructions, and matrix majorization, markotopes, and zonotopes supply convertibility criteria for classical distinguishability resources (Gonda, 2021).

A recurring misconception is that reversibility is universal across all regimes. The literature does not support that statement. Exact one-shot formation can be irreversible even for simple state pairs, channel asymptotics can require regularization or amortization, strong-converse asymptotic preorders are governed by sandwiched Rényi divergences rather than only by $\left(|0\rangle\langle 0|,\pi\right),$5, and Choi-divergence-based process monotones can fail data processing under superprocesses (Wang et al., 2019, Wang et al., 2019, Perry et al., 2020, Zambon, 2024). The unifying principle is narrower and more precise: when the free operations apply identically to both hypotheses, the relevant divergence is whichever monotone exactly captures the operational task and asymptotic regime under consideration.