arXiv:2609.28440
Sharp pairwise reduction for quantum hypothesis testing
Kuan-Yi Lee, Ludovico Lami
Abstract
We determine the optimal universal coefficient in a reduction of multiple to binary quantum hypothesis testing. Specifically, for every finite ensemble in any Hilbert space dimension (finite or infinite), we prove that the error probability of the global pretty good measurement (PGM) is at most four times the sum of the optimal binary error probabilities. By constructing a family of regular-simplex ensembles, we further show that the coefficient four is optimal, even for arbitrary global measurements. This result improves on the pairwise bounds established by Cheng and Liu [arXiv:2606.06246 (2026)] and entails an explicit, sharp guarantee for the standard PGM itself. Our proof is also simpler: rather than constructing sequential measurements and applying a union bound, our proof relies on purely matrix-analytic techniques, combining a direct block Gram matrix analysis of the PGM error and a refined inequality between quantum Hellinger distance and Bures $χ^2$-divergence. Our analysis also yields a refined, one-shot pairwise Chernoff bound and explicit sufficient number of copies for a desired discrimination error.
AI Summary
The paper asks how tightly multiple quantum hypothesis testing can be reduced to binary discrimination, and whether an explicit measurement attains the best universal bound. For a finite ensemble $A_i=p_i\rho_i$, it proves $\operatorname{Err}^{\star}_N\leq\operatorname{Err}^{\mathrm{PGM}}_N\leq4\sum_{i<j}\operatorname{Err}^{\star}(A_i,A_j)$, with binary errors retaining the original, unnormalized priors. The coefficient is independent of dimension, priors, and ensemble size. This improves the previously established coefficient eight and applies directly to the standard pretty good measurement (PGM). The upper bound represents the PGM error through off-diagonal blocks of a block Gram matrix’s square root. A coefficient-one inequality between squared quantum Hellinger distance and Bures $\chi^2$-divergence reduces this expression to pairwise harmonic quantities, each bounded by the corresponding Helstrom error. Uniformly weighted regular-simplex ensembles show that no coefficient below four works universally, even for an optimal global measurement; the ratio approaches four as the number of hypotheses grows. The analysis also gives $\operatorname{Err}^{\mathrm{PGM}}_N\leq4\sum_{i<j}s_{ij}^{s_{ij}}(1-s_{ij})^{1-s_{ij}}\operatorname{Tr}A_i^{1-s_{ij}}A_j^{s_{ij}}$. For independent copies, choosing $s_{ij}=1/2$ yields the sufficient condition $n\geq\lceil\log((N-1)/\delta)/\log(1/\epsilon)\rceil$, where $\epsilon=\max_{i<j}\operatorname{Tr}\sqrt{\rho_i}\sqrt{\rho_j}<1$, for PGM error at most $\delta$. The results are stated for finite ensembles, including trace-class states in infinite dimension. The bound does not assert ensemble-specific optimality of the PGM. Optimal coefficients for fixed ensemble size, a proposed hierarchy of subset reductions, and extensions to continuous ensembles remain open.
Error found: The stated trace-class extension proves continuity only of the optimal discrimination error, leaving the infinite-dimensional PGM guarantee without the additional, valid continuity argument it needs.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 11 pages