arXiv:2610.18785
Relative-Entropy Geometry and Convergence in Nonconvex Wiretap Optimization
Farzin Salek, Masahito Hayashi
Abstract
We develop and analyze an entropic descent method for a nonconvex optimization problem on probability simplices arising in classical and quantum secure communication. The optimization variables describe a randomized encoder, allowing its input distribution and stochastic preprocessing to be optimized jointly. An exact relative-entropy identity yields a channel-independent majorization bound and an explicit mirror-descent iteration with a fixed unit step size. We establish quantitative descent, convergence of objective values, and stationarity of accumulation points on their support faces for arbitrary finite classical channels, without assuming degradedness or less noisiness. For strictly positive transition matrices, a square-root parametrization and a Kurdyka--Łojasiewicz argument establish finite length of the transformed trajectory and convergence of the entire sequence, including limits on the simplex boundary. For less-noisy channels, a restricted convex formulation admits global convergence with an explicit \(O(1/t)\) objective-error bound. We extend the relative-entropy construction to noncommuting quantum outputs and derive monotone alternating updates for covariant encoders, including correlated preprocessing over a fixed block. Numerical experiments compare the method with analytic benchmarks and constrained optimization, reveal multiple attracting values, and demonstrate strict improvements from stochastic preprocessing.
AI Summary
The paper studies numerical optimization of classical wiretap secrecy capacity and classical–quantum single-letter private information, jointly varying the auxiliary distribution and stochastic encoder $p_{X|U}$. This joint optimization is generally nonconvex; restricting to $U=X$ can miss positive achievable rates. An exact relative-entropy identity bounds the objective’s linearization error by $D(p\|q)$, independently of the channels. For $F(p)=I(U;Z)-I(U;Y)$, this yields the explicit unit-step entropic update $p^+\propto p\exp(-\Omega[p])$, where $\Omega$ is the simplex gradient, and the descent guarantee $F(p^+)\le F(p)-D(p\|p^+)$. For arbitrary finite classical channels, objective values converge, successive iterate differences vanish, and accumulation points are stationary on their support faces. With strictly positive transition matrices, square-root coordinates and a Kurdyka–Łojasiewicz argument establish finite trajectory length and convergence of the entire sequence. For less-noisy channels restricted to $U=X$, convergence is global, with objective error at iteration $T$ bounded by $D(p^*\|p^{(0)})/T$. The relative-entropy construction extends to noncommuting quantum outputs, giving descent and support-face stationarity, and supplies monotone alternating updates for covariant encoders, including fixed-block correlated preprocessing. Numerical checks use analytic capacities, binary convex-envelope benchmarks, and constrained optimization; a noncommuting example achieves approximately $0.0004744481$ bits/use while its best no-prefixing rate is zero. The nonconvex guarantees do not certify local or global optimality; full-support fixed points can fail to be local minima. Quantum results concern single-letter or fixed-block achievable rates, not generally the regularized capacity. Benchmarks are finite-resolution, and block costs grow exponentially.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.