The qutrit value of the determinant dimension witness with independent devices is 3√3/4
Preprint, version 2 (17 November 2026); not peer reviewed. Version 2 adds a section on finite statistics; the theorems and proofs are unchanged from version 1. Bowles, Quintino and Brunner (Phys. Rev. Lett. 112, 140407 (2014)) introduced a nonlinear dimension witness for prepare-and-measure experiments with independent devices: four preparations, two binary measurements, and W2 = |det D|, where D is the 2×2 matrix of differences of outcome probabilities. Classical bits give W2 = 0, qubits and classical trits give at most 1, and systems of dimension four reach the algebraic maximum 2. For qutrits they reported, from numerical analysis, W2 ≤ 1.299. We prove that the qutrit maximum is exactly 3√3/4 = 1.29903…, over all mixed states and all binary measurements. More precisely, for measurements given by rank-one projectors (or their complements) with overlap c, the maximum is √(1−c)(1+√c) for every c in [0,1], and it is largest at c = 1/4. The proof is elementary: it reduces the problem to the plane of the two measurement vectors and a convex body of probability pairs, and it shows where dimension three is used. We also give a short proof of the classical-trit bound 1. Consequently, if the noise does not depend on the preparation and scales every probability difference by the same visibility η (white noise on the states is an example), the optimal qutrit strategy excludes qubits and all classical trits exactly when η > (4/(3√3))^(1/2) = 0.87738…. This threshold holds only for that noise model; it is not a tolerance to arbitrary experimental imperfections. For finite data we give a rigorous interval test. Under independent, identically distributed trials it certifies W2 > 1 at confidence 1 − 10⁻³, with probability 0.9, using about 650 trials per setting for noiseless data and about 2.4·10⁴ per setting at η = 0.9. A plain z-test on the determinant estimator is shown not to be a valid certificate. The proof was checked symbolically, its extremal cases were certified by a branch-and-bound computation, and it was confirmed by two blind re-derivations by a separate AI instance, with negative controls. These re-derivations are cross-checks by a separate instance of the same language model, not verification by a human referee. Files: the preprint, and a reproducibility archive with: the preregistrations and addenda, the proof notes and the literature check; exact SymPy checks of every formula in the note; the certified branch-and-bound check; both blind re-verifications with their specifications, reports and manifests; the finite-statistics study (new in version 2); all run logs, including crashed, capped, superseded and degenerate-control runs, labelled as such; verify.sh. The theorems rest on the written proofs; the scripts check identities, explicit strategies, finite enumerations, finitely many parameter values and simulated error rates.
Authors
- Vasily Psykovsky (ORCID: https://orcid.org/0009-0008-0416-1377)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24302455
- Primary Topic
- Quantum Mechanics and Applications
- Type
- preprint