Quantum information · Theory
Quantum communication: a discrimination approach
arXiv:2610.40358v1 ↗PDF ↗Submitted 29 October 2026
- Question
- The two familiar routes to quantum capacity are packing, where Bob distinguishes outputs, and decoupling, where the reference becomes independent of the environment. Decoupling often proves only that a decoder exists. Can entanglement generation be analysed directly as a state-discrimination problem, with an explicit decoder and bounds valid at every block length?
- Result
- For any input and \(\alpha\in[1/2,1)\), random codes built from the source satisfy \(\overline\varepsilon_{\rm EG}(d)\le c_\alpha2^{-\frac{1-\alpha}\alpha(I_\alpha(R\rangle B)_\rho-\log_2d)}\), with decoders derived from Petz recovery (or averaged rotated Petz recovery), where \(I_\alpha\) is the Petz–Rényi coherent information. The coefficient \(c_\alpha\) is sharp. For product sources this gives exponential error decay at every rate below the coherent information, with an explicit exponent.
The paper in detail
Alice retains \(M\), sends \(A\), and Bob must recover a maximally entangled state of Schmidt rank \(d\), with error measured by \(\varepsilon_{\rm EG}=P^2=1-F^2\) with respect to \(\Phi\). Writing \(\{W_z\}_{z=1}^{d^2}\) for the Weyl unitaries on \(M\), Theorem 2.1 states that for pure \(\theta_{MBE}\) three quantities coincide: the best recovery error on \(B\), the best error for discriminating the \(d^2\) equiprobable states \((W_z\otimes I)\theta_{MB}(W_z^\dagger\otimes I)\), and \(\min_{\sigma_E}P^2(\theta_{ME},I_M/d\otimes\sigma_E)\).
That equality is the conceptual centre. The first is entanglement recovery, the second is packing on \(MB\), and the third is decoupling on \(ME\). The joint measurement on \(MB\) is only an analytic tool; the decoder that results acts on \(B\) alone. The paper’s flagged-erasure example makes this concrete: all \(d^2\) hypotheses coincide on the erased branch, so the optimal error is the erasure probability times \(1-1/d^2\) in both pictures.
The code picks a random \(d\)-dimensional subspace of the reference through \(UV_j\), with probability weighted by source mass. The POVM is built from \(\rho_{RB}^\alpha\) rather than \(\rho_{RB}\), which ties the measurement to the exponent, and the Weyl sum normalises it by construction. The denominators are a positive quotient, either symmetric or logarithmic-derivative, corresponding to ordinary and averaged rotated Petz recovery. A tilting inequality bounds the error by a power of the normalisation gap, and operator Jensen moves the average over isometries and over \(U\) inside it.
A Haar second-moment formula then gives \(\frac{d|R|-1}{|R|^2-1}(|R|I_R\otimes\mathrm{Tr}_R\rho^\alpha_{RB}-\rho^\alpha_{RB})\), and the scalar bound follows from superadditivity. The noiseless channel provides a sharp test of the coefficient. What remains open includes optimal exponents, whether correlated inputs improve them, and how to implement the decoders.
Transferable methods
Weyl covariance turns a decoder into a measurement and back
A recovery task and a discrimination task do not look alike, since one outputs a state and the other a label. The Weyl group gives an exact dictionary between them, so a measurement bound becomes a coding bound without losing anything.
A decoder \(\mathscr D\) gives \(T_{MB}=(\mathrm{id}\otimes\mathscr D^*)(\Phi)\ge0\) with \(\mathrm{Tr}_MT=I_B/d\). The Weyl identity \[\sum_{z=1}^{d^2}(W_z\otimes I)X_{MB}(W_z^\dagger\otimes I)=d\,I_M\otimes\mathrm{Tr}_MX_{MB}\] shows that the conjugates of \(T\) form a POVM. Each hypothesis and its test receive the same conjugation, so every correct-outcome probability equals \(\mathrm{Tr}[\theta_{MB}T_{MB}]\), the squared fidelity with \(\Phi\). Conversely, any POVM gives \(T=\frac1{d^2}\sum_z(W_z^\dagger\otimes I)\Pi^z(W_z\otimes I)\), a decoder with the same score.
The dictionary tells us how to build the POVM. Taking the quotient of \(W_z\rho^\alpha W_z^\dagger\) by the sum of all conjugates, which is the right-hand side above, gives \(\sum_z\Pi^z=I\) automatically on the support. Using the quotient \(X^{-1/2}YX^{-1/2}\) gives Petz recovery, and the logarithmic derivative gives an averaged rotated Petz recovery.
Editorial check: the Weyl identity held to \(10^{-14}\) for random positive \(X\) on \(d=2,3\) with a qubit \(B\). For a Bell state through qubit dephasing with \(p=0.1\), recovery error equals \(p=0.1\) (the Bell states being orthogonal apart from \(Z\)), and the flagged-erasure values \(p(1-1/d^2)\) are \(0.15\) at \((d,p)=(2,0.2)\) and \(0.0889\) at \((3,0.1)\).
How to reuse it
Whenever a target state is invariant under a group representation, try rewriting fidelity as success probability of discriminating the group orbit. This brings in the pretty-good-measurement toolbox, and it is exact when the group forms a 1-design.
A sharp tilting inequality tested by the identity channel
Coding proofs often waste a constant when converting a measurement bound into an exponent. The noiseless channel shows that the coefficient \(c_\alpha\) here is the smallest possible at each order, and the proof reaches it.
For \(\alpha\in[1/2,1)\), the tilting lemma in Appendix A bounds the error operator by a positive power: \(1-T\) applied to \(\rho\) is controlled by \(c_\alpha\mathrm{Tr}[\rho^\alpha(\text{gap})^{(1-\alpha)/\alpha}]\), with \[c_\alpha=\Bigl(\frac{1-\alpha}\alpha\Bigr)^{\frac{1-\alpha}\alpha}\Bigl(\frac{2\alpha-1}\alpha\Bigr)^{\frac{2\alpha-1}\alpha}\in[\tfrac12,1].\] The exponent \((1-\alpha)/\alpha\in(0,1]\) makes the map operator concave, which permits operator Jensen across an isometry and across the Haar average.
For the identity channel with \(d>|A|\), the discrimination converse is short: the \(d^2\) hypotheses live in dimension \(d|A|\), so a POVM has average success at most \(\frac1{d^2}\sum_z\mathrm{Tr}\,\Pi^z=|A|/d\), and the exact error is \(1-|A|/d\). Eq. (18) gives \(c_\alpha(d/|A|-1)^{(1-\alpha)/\alpha}\). Choosing \(\alpha=d/(d+|A|)\) makes \((1-\alpha)/\alpha=|A|/d\), and the bound then equals \(1-|A|/d\) exactly.
Editorial check: for \((d,|A|)=(3,2),(5,2),(7,3)\) this gives 0.3333, 0.6000, 0.5714, matching \(1-|A|/d\), and it gives 0 for \((4,4)\). \(c_\alpha\) stays in \([1/2,1]\), reaching \(1/2\) near \(\alpha=2/3\). For a Bell state through dephasing with \(p=0.1\), \(I_\alpha\) rises from 0.322 at \(\alpha=1/2\) to 0.531 near \(\alpha=1\), which is the coherent information \(1-h(0.1)\), consistent with the claimed continuity at \(\alpha=1\).
How to reuse it
After proving an achievability bound with a constant, test it on a noiseless channel with non-integer dimension ratios. If the bound can touch the exact converse, the constant cannot be improved, and you can stop looking for a better one.
Context & limitations
Why it matters. It explains why packing and decoupling both work: they have the same optimal error as a duality of three problems. The bound holds for finite \(n\) and for correlated channels, and it mirrors the Arimoto-type strong-converse bound. Having explicit recovery maps gives a concrete object to study for decoder design, even though implementing them is a separate problem.
Limitations. The exponent is achievability only. It is not claimed to be optimal, and optimising across correlated inputs is not done. The code is random, chosen before use, and the existence of a good code does not make the decoder efficient. Regularisation remains in the capacity formula, so the result does not settle capacity for any specific channel. This entry covers Sections 1–3; the binary-testing bound in Section 4 and the appendices were not read closely.