arXiv:2609.28635
Continuity of Regularized Channel Rényi Divergences
Jinzhao Wang, Yuxiang Yang
Abstract
We prove that the regularized, stabilized sandwiched Rényi divergence of finite-dimensional quantum channels converges to their regularized relative entropy as the Rényi order tends to one. The key tool is the channel hockey-stick divergence: Gour's Stinespring approximation bound and a Schatten norm estimate amplify an asymptotic bound below one into exponential decay at higher threshold rates. For channel pairs with finite max-relative entropy, known operational connections then give exponential strong converses for parallel and adaptive discrimination, a sharp zero--one testing law, and the subchannel asymptotic equipartition property.
AI Summary
The paper addresses whether taking the Rényi order to one commutes with optimization over entangled channel inputs and regularization over arbitrarily many uses. For any finite-dimensional channel pair, it proves $\lim_{\alpha\to1}\widetilde D_\alpha^{\mathrm{reg}}(\mathcal N\|\mathcal M)=D^{\mathrm{reg}}(\mathcal N\|\mathcal M)$, where both divergences are stabilized by a reference system. The nontrivial step is right continuity; infinite max-relative entropy is handled separately through Choi-support mismatch. The proof uses the channel hockey-stick divergence $E_\lambda$. A fixed-dilation version of a Stinespring approximation bound produces an operator $W$ satisfying $\Phi_W\le_{\mathrm{CP}}\lambda\mathcal M$ and $\|V-W\|_\infty\le\sqrt{E_\lambda(\mathcal N\|\mathcal M)}$. A Schatten-norm estimate then controls regularized Rényi divergence through decompositions of $V$. Combining approximations at two rates shows that an asymptotic hockey-stick bound strictly below one forces exponential decay at every higher rate, establishing continuity. For pairs with finite $D_{\max}$, known operational connections yield exponential strong converses for parallel and adaptive discrimination: if $\beta_n\le2^{-nr}$ with $r>D^{\mathrm{reg}}$, then acceptance under $\mathcal N$ satisfies $p_n\le2^{-n\kappa_r}$ for some $\kappa_r>0$. Both protocol classes therefore have fixed-error Stein exponent $D^{\mathrm{reg}}$. Further consequences include a sharp off-threshold zero–one testing law and a subchannel AEP, with CP-dominated subchannel approximations converging exponentially in diamond norm. No testing-limit claim is made exactly at the threshold. Smoothing uses trace-nonincreasing maps; trace-preserving smoothing can invalidate the AEP. Three concurrent works also establish continuity, including one treating general process testers.
The proof presented in this paper was proposed by GPT-6 Astra. AI tools also assisted with drafting and revising the manuscript.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 12 pages; comments are welcome; v2: updated references for concurrent works, add a Lean certificate for our proof: https://github.com/JWang226/continuity-of-regularized-channel-renyi-divergence