No information transmission through quantum channels above capacity

Scientific Impact Assessment

Core Contribution. This paper resolves a long-standing open problem in quantum Shannon theory: it proves *exponential* strong converse theorems for both unassisted quantum (entanglement-generation) and classical communication over arbitrary finite-dimensional memoryless quantum channels. The central claim is that the channel capacity is a genuine phase-transition point — below capacity, reliable transmission is possible; above it, the entanglement-generation fidelity and classical success probability decay exponentially in the number of channel uses. Prior strong-converse results were confined to restricted channel classes (entanglement-breaking, Hadamard, degradable/antidegradable, generalized dephasing, covariant channels, or stabilizer codes over Pauli channels). The obstruction for the general case was regularization: capacities are non-single-letter, and correlations across channel uses defeated the standard quantum Arimoto approach. The key new ingredient is a set of one-shot integral representations (Theorem 3) that express sandwiched Rényi coherent and Holevo information as integrals of ordinary von Neumann entropy differences on "tilted" states, combined with local support estimates and a Leung–Smith telescoping restoration argument that yields block-length-independent, dimension-local continuity bounds for Rényi capacities (Theorem 4). This delivers the missing asymptotic continuity limα↘1 Qα = Q (and the classical analogue), which plugs directly into the Arimoto bounds.

Methodological Rigor. The paper is a theoretical work with complete, carefully structured proofs. It builds on well-established facts (data processing for sandwiched Rényi divergence, variational/minimax representations, sharp conditional-entropy continuity bounds, Nussbaum–Szkoła weights, Fekete's lemma). Crucially, it does not merely prove the converse: it establishes matching achievability (entanglement-generation exponent in Appendix C via change-of-measure/pinching, and classical via Mosonyi–Ogawa), yielding *exact* strong-converse exponent formulas (Theorem 5). The uniformity of the support estimates in Lemma 12 — where only local dimensions dim B_i, dim E_i enter and the remaining n−1 channel uses do not — is the technical linchpin and appears sound. The extension to nonstationary product channels further demonstrates the robustness of the method. This is an exemplary, self-consistent design.

Potential Impact. Within quantum information theory this is a landmark result that completes the "coding transition" picture (Fig. 1) for the most general unassisted setting. It will be cited as the definitive statement that no information passes above capacity, and the integral-representation/tilted-state technique is a reusable tool that is likely to be adopted for other Rényi-quantity continuity problems (e.g., private capacity, other regularized quantities, resource theories). The impact is concentrated in a specialized subfield rather than broadly across ML/physics, but within that community it is high.

Timeliness & Relevance. The problem has been explicitly "elusive so far," with a steady stream of partial results (2014–2026) chipping away at special cases; several 2026 preprints are cited as immediate precursors. Solving the general case is a natural capstone and is highly timely. The disclosed use of OpenAI Codex/ChatGPT to propose the crucial asymptotic-continuity-via-derivative-bounds idea is itself a noteworthy data point about AI-assisted mathematical discovery, though the authors reworked the argument into the integral representation.

Strengths.

Limitations & Gaps.

Overall. This is a strong, high-quality theoretical contribution that closes a well-known open problem with a novel and reusable method. Its scientific impact should be substantial within quantum information theory, tempered by its specialized audience and the fact that it establishes an expected (if previously unproven) result rather than overturning conventional understanding.