Quantum information · Theory (AI-assisted)

Quantum preparation without disturbance: information costs and hypothesis testing

Pablo Costa Rico, Robert Salzmann, Mario Berta, Ángela Capel

arXiv:2611.18135v1 ↗PDF ↗Submitted 13 November 2026

Question
A device prepares \(\rho_0\) or \(\rho_1\) depending on a memory state \(\omega_0\) or \(\omega_1\), with the same channel for both. How distinguishable must the memories be, measured by the relative entropy \(D(\omega_0\Vert\omega_1)\)? If the memory may be used up, \(D(\rho_0\Vert\rho_1)\) suffices. If the memory must be returned undisturbed, so the device can be used again, the answer was unknown. The question also gives an operational meaning to the Belavkin–Staszewski relative entropy, one of several quantum versions of the Kullback–Leibler divergence.
Result
When the memory must be preserved, the minimum cost is the Belavkin–Staszewski relative entropy \(D_{\mathrm{BS}}(\rho_0\Vert\rho_1)=\mathrm{tr}[\rho_0\log(\rho_0^{1/2}\rho_1^{-1}\rho_0^{1/2})]\), which exceeds \(D\) for noncommuting states, and a classical memory of dimension at most \(\dim S\) attains it. More generally, for any divergence obeying data processing, memory preservation selects its maximal quantum extension. Requiring \(n\) outputs while letting the memory evolve gives costs that rise strictly from \(D\) to \(D_{\mathrm{BS}}\), with an \(O(1/n)\) gap for faithful states. In a testing task where an unknown device emits \(n\) systems per run, the optimal error exponent likewise tends to \(D_{\mathrm{BS}}\).

The paper in detail

A preparation device receives a memory \(\omega_i\) and outputs the system \(\rho_i\); the same channel must work for \(i=0,1\). The memory's information content is measured by a divergence between \(\omega_0\) and \(\omega_1\), and data processing immediately gives \(D(\omega_0\Vert\omega_1)\ge D(\rho_0\Vert\rho_1)\), achieved by simply handing over \(\omega_i=\rho_i\). That device consumes its memory. The paper asks what changes when the memory must come back with the same marginal, so the device can run again, while correlations between the returned memory and the output are allowed.

The answer is the Belavkin–Staszewski relative entropy. It is the largest quantum divergence that reduces to Kullback–Leibler on commuting states, and Matsumoto showed it equals the smallest classical divergence \(D_{\mathrm{KL}}(p_0\Vert p_1)\) over all ways of writing \(\rho_i=\sum_zp_i(z)\tau_z\) with common states \(\tau_z\). The new theorem shows that memory preservation forces exactly this structure, so the optimal reusable device is classical: read a label \(z\) from the memory without disturbing it, and emit \(\tau_z\).

The same argument works for any divergence that obeys data processing. The stationary cost is its maximal extension: the geometric Rényi divergence for Petz Rényi inputs, and explicit maximal versions of trace distance and purified distance. Between the two extremes, requiring \(n\) uses with the memory allowed to drift defines costs \(I_n\) that start at \(D\), increase with \(n\), and converge to \(D_{\mathrm{BS}}\). For relative entropy the increase is strict whenever \(\rho_0\) and \(\rho_1\) do not commute, because equality in data processing would imply a recovery map.

There is a testing version. An unknown device is reset \(m\) times and each time emits \(n\) systems, each with marginal \(\rho\) under one hypothesis and \(\sigma\) under the other, with arbitrary correlations within a run. One test must work for every compatible device. For \(n=1\) this is ordinary Stein testing with exponent \(D(\rho\Vert\sigma)\). As \(n\) grows, the worst-case exponent per run tends to \(D_{\mathrm{BS}}(\rho\Vert\sigma)\), at rate \(O(1/n)\), and this holds however \(n\) grows relative to \(m\).

Questions left open include whether the finite-\(n\) testing exponent equals the finite-\(n\) preparation cost, and what happens with more than two hypotheses, infinite dimensions or non-faithful states, where several of the quantitative bounds fail.

Transferable methods

  1. Koashi–Imoto structure turns a preserved memory into a classical label

    The lower bound must hold for every quantum memory and channel. The Koashi–Imoto theorem describes all channels that leave a family of states unchanged: the memory splits into blocks, each block holds a part that distinguishes the states and a part common to them, and the channel can act only on the common part. Measuring the block label is then a classical memory that does at least as well.

    Let \(V\) be a Stinespring isometry of the device, so the retained-memory channel \(\mathcal T=\mathrm{tr}_S\circ\mathcal R\) fixes \(\omega_0,\omega_1\). Koashi–Imoto gives \[M=\bigoplus_zM_z^L\otimes M_z^R,\quad\omega_i=\bigoplus_zp_i(z)\,a_{i,z}\otimes b_z,\quad V=\bigoplus_zI_{M_z^L}\otimes V_z,\] with \(b_z\) and \(V_z\) independent of \(i\). The output is then \(\rho_i=\sum_zp_i(z)\tau_z\) with \(\tau_z=\mathrm{tr}_{M^R_z}\mathrm{tr}_E(V_zb_zV_z^\dagger)\) common to both hypotheses.

    Measuring the block label is a channel taking \(\omega_i\) to \(\mathrm{diag}\,p_i\), so by data processing \[\mathbb D(\omega_0\Vert\omega_1)\ge\mathbb D_{\mathrm{cl}}(p_0\Vert p_1)\ge\inf_{\rho_i=\sum_zp_i(z)\tau_z}\mathbb D_{\mathrm{cl}}(p_0\Vert p_1)=\widehat{\mathbb D}(\rho_0\Vert\rho_1).\] Conversely, the classical device \(\omega_i=\sum_zp_i(z)\lvert z\rangle\langle z\rvert\), \(\mathcal R(X)=\sum_z\langle z\rvert X\lvert z\rangle\,\lvert z\rangle\langle z\rvert\otimes\tau_z\) preserves its memory and attains any decomposition.

    How to reuse it

    Whenever a protocol must leave a family of quantum states undisturbed, the Koashi–Imoto decomposition reduces it to reading a classical label plus processing state-independent parts. Lower bounds for such protocols then follow from data processing on the label distribution.

  2. The optimal reverse test for the Belavkin–Staszewski entropy

    To identify the classical optimum for relative entropy, one needs the decomposition \(\rho_i=\sum_zp_i(z)\tau_z\) that minimizes \(D_{\mathrm{KL}}(p_0\Vert p_1)\). Matsumoto's answer uses the eigenvectors of the operator ratio \(\rho_1^{-1/2}\rho_0\rho_1^{-1/2}\), giving a memory no larger than the system.

    Editorial derivation: write \(\Gamma=\rho_1^{-1/2}\rho_0\rho_1^{-1/2}=\sum_z\lambda_z\lvert e_z\rangle\langle e_z\rvert\) and set \(q_z=\langle e_z\rvert\rho_1\lvert e_z\rangle\), \(\tau_z=\rho_1^{1/2}\lvert e_z\rangle\langle e_z\rvert\rho_1^{1/2}/q_z\), \(p_z=\lambda_zq_z\). Then \(\sum_zq_z\tau_z=\rho_1\) and \(\sum_zp_z\tau_z=\rho_1^{1/2}\Gamma\rho_1^{1/2}=\rho_0\), and \[D_{\mathrm{KL}}(p\Vert q)=\sum_z\lambda_zq_z\log\lambda_z=\mathrm{tr}\bigl[\rho_1\,\Gamma\log\Gamma\bigr]=\mathrm{tr}\bigl[\rho_0\log(\rho_0^{1/2}\rho_1^{-1}\rho_0^{1/2})\bigr].\] For the last step put \(Y=\rho_0^{1/2}\rho_1^{-1/2}\), so \(\Gamma=Y^\dagger Y\) and \(\Gamma\log\Gamma=Y^\dagger\log(YY^\dagger)\,Y\); then \(\mathrm{tr}[\rho_1Y^\dagger\log(YY^\dagger)Y]=\mathrm{tr}[Y\rho_1Y^\dagger\log(YY^\dagger)]\) with \(Y\rho_1Y^\dagger=\rho_0\) and \(YY^\dagger=\rho_0^{1/2}\rho_1^{-1}\rho_0^{1/2}\).

    Editorial check with random states: in dimension 2, \(D=0.395\), \(D_{\mathrm{BS}}=0.597\) and \(D_{\mathrm{KL}}(p\Vert q)=0.597\); in dimension 4, \(3.033\), \(3.590\) and \(3.590\). Both targets were reconstructed to \(10^{-13}\).

    How to reuse it

    The construction gives an explicit minimal-dimension classical memory for any pair of states and a computable upper bound for any task whose cost is a divergence of a reusable classical record. The gap \(D_{\mathrm{BS}}-D\) measures how much noncommutativity costs when the record must survive use.

Source: Theorem 1.1 and its proof (Theorem A.1, Eqs. (53)–(59)); Eq. (11); Theorems 3.7 and 4.2–4.3. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It gives the Belavkin–Staszewski relative entropy, usually seen only as an upper bound, two operational meanings: the information cost of a reusable preparation and the Stein exponent for testing devices from their outputs. It also explains why quantum memories offer no saving here: preserving two states forces the memory into a classical-label structure.

Limitations. The setting is two hypotheses and finite dimensions. The robustness bound and the convergence rates require faithful (full-rank) targets, and their constants blow up as the smallest eigenvalue \(\lambda\) shrinks, like \(\lambda^{-5/2}\). At finite \(n\), only the inequality between the testing exponent and the preparation cost is proved, not equality. The authors state that ChatGPT 6 Astra assisted with brainstorming, drafting and organizing the literature review, and that they verified the result.

Description based on the full arXiv HTML of v1, read 2026-11-16: Section 1, Section 2.1–2.2, the statements of Proposition 2.1, Theorem 2.2, Definition 3.2, Theorems 3.3 and 3.7, Propositions 3.8–3.9 and Theorems 4.2–4.3, the proof of Theorem A.1 and the AI statement; the remaining appendix proofs (de Finetti extraction, robustness, convergence rates) were not checked. Editorial check: for random states in dimensions 2–4, the classical memory built from the eigenvectors of \(\rho_1^{-1/2}\rho_0\rho_1^{-1/2}\) reproduces both targets to \(10^{-13}\), and its Kullback–Leibler divergence equals \(D_{\mathrm{BS}}\), which exceeds \(D\) in each case.