arXiv:2609.40203
Generalized Reimpell-Werner Iteration
Shihao Ru, Bikun Li, Weibo Gao, Liang Jiang
Abstract
Quantum measurements and channels determine how information is extracted, encoded, and transmitted in quantum protocols. Optimizing their performance often requires numerical methods that remain practical as Hilbert space dimensions increase. The Reimpell-Werner iteration offers a practical approach to these tasks through repeated matrix updates that respect the constraints. Here, we generalize this iteration to linear objectives with arbitrary Hermitian cost matrices. We prove that the iterates converge to a global optimum whenever the initialization satisfies suitable support overlap conditions. For each fixed problem, choice of iteration parameters, and admissible initialization, $\mathcal{O}(1/\varepsilon)$ iterations suffice asymptotically to bring the objective value within $\varepsilon$ of the optimum. These results provide a rigorous foundation for the iteration and broaden the class of optimization problems to which its convergence guarantees apply.
AI Summary
The paper addresses global convergence of the Reimpell–Werner (RW) iteration for channel optimization and the Ježek–Řeháček–Fiurášek (JRF) iteration for measurement optimization. It generalizes RW to the SDP $\max_{X\succeq0}\{\operatorname{tr}(CX):\operatorname{tr}_A X\preceq D\}$ for arbitrary Hermitian $C$ and PSD $D$. A slack matrix and a common spectral shift make the augmented cost positive semidefinite without changing the optimization problem. Under explicit initial-support conditions, satisfied by positive-definite seeds, every update is well defined and the matrix iterates converge to a global optimum. The limiting normalizer yields the unique dual optimizer of the support-reduced, normalized problem. Specializations establish convergence of the original RW iteration and of JRF for arbitrary finite mixed-state ensembles under the stated support conditions. The proof combines polar normalization on a Stiefel manifold, a Łojasiewicz gradient inequality, and a spectral argument establishing limiting dual feasibility and closing the duality gap. For each fixed instance, shift, and admissible seed, objective accuracy $\varepsilon$ requires asymptotically $O(1/\varepsilon)$ iterations; a computable dual certificate requires $O(1/\varepsilon^2)$. Half-gradient-exponent cases have logarithmic bounds. Explicit slow families have objective iteration exponents approaching one and unbounded matrix-error exponents as dimension grows. The bounds have instance-dependent constants and onset, so they do not establish complexity as dimension varies. Runtime comparisons exclude reference-bound construction and iterative certificate evaluation. Extension to general linear constraints remains open because the normalization exploits partial-trace structure; optimal shift and initialization choices are also unresolved.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 24 pages, 2 figures, 1 table