A fixed-source gap above the Holevo rate for visible quantum compression
A finite four-qubit ensemble has Holevo information one bit but requires at least 1.12098 qubits per signal, even with free shared classical randomness.
Yuxuan Zhang · 16 November 2026 · Version 1.0
If Alice knows which mixed quantum state Bob should receive, can she compress it to its Holevo information when they share as much classical randomness as they like? This manuscript gives an explicit counterexample: a fixed ensemble of four-qubit states has Holevo information one bit, while every asymptotically accurate protocol needs at least 1.12098 qubits per signal.
The result covers arbitrary mixed encodings, collective decoding channels, and any amount of independent shared classical randomness. It is a complete proof candidate that has passed sequential internal analytical review and exact diagnostic checks. Independent expert review and publication priority remain unconfirmed.
Full paper (PDF) · Source, checks and review (ZIP) · Version and SHA-256 metadata
The source and the bound
Take the 255 nonidentity Hermitian four-qubit Pauli operators, each with both signs. The resulting 510 labels are equally likely, and the signal is
\[W_{P,b}=\frac{I+bP}{16},\qquad b\in\{-1,1\}.\]Each state is maximally mixed on an eight-dimensional subspace, with entropy three bits. Their average is the maximally mixed four-qubit state, with entropy four. The Holevo information is therefore exactly one bit.
For every blocklength \(n\) and message dimension \(m\), the average block trace-distance error obeys
\[1-\varepsilon_n\le m^{3/4}\left(\frac{53}{170}\right)^{n/2}.\]It follows that the communication rate satisfies
\[R_{\mathrm{vis},q,\mathrm{cr}} \ge \frac23\log_2\frac{170}{53} =1.1209803210\ldots>1.\]Below this lower bound by \(\delta\) qubits per signal, the error is at least \(1-2^{-3\delta n/4}\). The exact optimal rate of this ensemble is still unknown; directly sending the state gives an upper bound of four.
The shared seed is averaged before taking the trace distance:
\[\widehat W_{x^n}=\sum_s q_n(s)\mathcal D_{n,s}(\tau_{x^n,s}), \qquad \varepsilon_n=\frac12\sum_{x^n}p^n(x^n) \|W_{x^n}-\widehat W_{x^n}\|_1.\]This is the error criterion in the QIQCOP problem. No shared entanglement or additional communication is allowed.
Why the bound survives collective decoding
Each target state lies inside the support projector \(Q_{P,b}=(I+bP)/2\). Accurate reconstruction must pass the corresponding support test with high probability. Averaging two copies of these projectors gives a particularly simple operator:
\[\mathbb E_x Q_x\otimes Q_x =\frac{127}{510}I+\frac{4}{255}F,\]where \(F\) exchanges the two copies. For an independent block of labels, the moment expands into operators that exchange any subset of the sites.
For an arbitrary collective decoder, the contribution of each subset is the purity of the Choi matrix of that decoder after tracing out the other output sites. A channel from dimension \(m\) to dimension \(d\) has unnormalized Choi purity at most \(\min(m^2,md)\). Applying both bounds across all subsets gives the tensor-power estimate above.
The estimate holds for every decoder separately. Mixing decoders with a shared random seed preserves it because the support-test acceptance probability is affine in the reconstructed state. No restriction to product decoders or pure messages enters the argument.
Relation to earlier work
Winter’s 2002 Conjecture 23 proposes Holevo-rate visible compression with common randomness at a prescribed rate. The present counterexample allows an unrestricted amount of that resource, so it also contradicts the proposed communication rate.
One-shot separations between information and compression cost were already known. Jain, Radhakrishnan and Sen used random-subspace ensembles and support tests; Bab Hadiashar and Nayak established one-shot incompressibility with general decoding channels. The contribution here is a fixed finite Pauli ensemble with a converse for every tensor power, including arbitrary shared classical randomness.
Hayashi’s formula concerns the unassisted visible rate. Nator and Pereg study classical communication with shared randomness, and Baghali Khanian gives strong-converse bounds in an unassisted compression model. These results address related resource settings. The source package records the theorem-level comparison; the targeted search did not identify an equivalent fixed-source result.
Checks and limits
The review traced the Pauli moment, the partial-Choi identity, the dimension bound and the seed-averaging step for arbitrary blocklength. Finite checks include Pauli basis identities through four qubits, arbitrary-channel moment comparisons, extremal channels and deliberately incorrect versions that the checks reject.
A separately implemented exact calculation uses a complex rational collective decoder with a three-dimensional message and two output qubits. Direct enumeration of all 36 labels and the partial-Choi calculation both give the moment \(10442900/25431849\). This is an internal diagnostic, not an independent proof of the arbitrary-blocklength theorem.
The manuscript was developed with AI assistance for derivation, literature comparison, computations and drafting under the author’s research direction. Its mathematical claim is open to external scrutiny. The website release does not establish first priority or acceptance by the problem catalog.