Continuity of Regularized Channel Rényi Divergences

AI summary

Overview

Research area: Quantum information theory — specifically the mathematics of quantum channel discrimination, channel Rényi divergences, and quantum Stein's lemma.

Technical level: Advanced. The paper is written for researchers comfortable with Stinespring dilations, Choi matrices, Schatten norms, completely positive order, and sandwiched Rényi divergences.

Scope: The paper proves that the regularized, stabilized sandwiched Rényi divergence of finite-dimensional quantum channels converges to the regularized relative entropy as the Rényi order approaches one, and surveys the operational consequences of that continuity.

What This Paper Is About

When you must decide whether an unknown quantum channel is 𝒩 or ℳ using n uses of it — possibly with entangled inputs across uses and adaptive processing between uses — the best achievable tradeoff between the two types of error is governed by Rényi divergences at order α > 1. To convert those bounds into a statement about the optimal exponential error rate, you must take α down to 1 and recover the regularized relative entropy. That limiting step requires continuity at order one, which had remained open because the channel divergence is a supremum over entangled inputs and over arbitrarily many channel uses, so continuity for each fixed input and block length does not suffice.

Key Contributions

1. Proof of continuity at order one (Theorem 1). For any two finite-dimensional quantum channels 𝒩, ℳ with the same input and output spaces, lim_{α→1} D_α^reg(𝒩‖ℳ) = D^reg(𝒩‖ℳ).

2. An amplification principle for the channel hockey-stick divergence. An asymptotic bound strictly below one at a threshold rate is shown to force exponential decay at every higher rate — the mechanism that drives the main theorem.

3. Two new technical lemmas. A fixed-dilation form of Gour's Stinespring filter (Lemma 2), which allows approximations at different λ to be compared because the isometry and environment are held fixed, and a Schatten norm bound (Lemma 3) that controls arbitrarily many copies of an n-use block even under entanglement across copies.

4. Operational applications plus a machine-checked artifact. Continuity is fed into known connections to obtain exponential strong converses, a sharp zero–one testing law, and the subchannel asymptotic equipartition property; a Lean formalization of Theorem 1 is provided, with comparator-checked statement and kernel-replayed proofs.

Main Findings

• Continuity holds in both directions. The left limit (α ↑ 1) follows from monotonicity in α, state-level continuity at one (including support mismatch, where the trace tends to Tr(Π_{supp σ} ρ) < 1 and the divergence goes to +∞), and interchange of the suprema over α, n, and inputs. The right limit (α ↓ 1) is the hard direction and is proved by contradiction.

• The infinite max-relative entropy case is immediate. If D_max(𝒩‖ℳ) = +∞ then supp J_𝒩 ⊄ supp J_ℳ, and a normalized maximally entangled input makes both the relative entropy and every order-α > 1 divergence infinite.

• Bounding sandwich. When D_max(𝒩‖ℳ) = d_∞ < ∞, the CP order gives 𝒩^{⊗n} ≤CP 2^{n d∞} ℳ^{⊗n}, hence 0 ≤ D^reg ≤ D_α^reg ≤ d_∞ for α > 1.

• Gour's filter, fixed-dilation form (Lemma 2). For any λ ≥ 1 and any fixed Stinespring isometry V of 𝒩, there is an operator W_λ on the same environment with Φ_{W_λ} ≤CP λℳ and ‖V − W_λ‖∞ ≤ √(E_λ(𝒩‖ℳ)). W_λ need not be an isometry or contraction; its norm is at most 1 + √(E_λ(𝒩‖ℳ)).

• Schatten norm bound (Lemma 3). If V = Σ_{i=1}^{q} U_i with ‖U_i‖∞ ≤ b_i and Φ{U_i} ≤_CP λ_i ℳ^{⊗n}, then for every α > 1 with s = (α−1)/α, D_α^reg(𝒩‖ℳ) ≤ (2/(ns)) log(Σ_i b_i^{1−s} λ_i^{s/2}).

• Two testing bounds (Lemma 4). For every n and r > 0: E_{2^{nr}}(𝒩^{⊗n}‖ℳ^{⊗n}) ≤ D^reg(𝒩‖ℳ)/r + 1/(nr). And for every α > 1, every ℓ > D_α^reg(𝒩‖ℳ), and every n: E_{2^{nℓ}}(𝒩^{⊗n}‖ℳ^{⊗n}) ≤ 2^{−n(α−1)[ℓ − D_α^reg(𝒩‖ℳ)]}.

• Core implication (Eq. 2.14). If limsup_n E_{2^{nr}}(𝒩^{⊗n}‖ℳ^{⊗n}) < 1, then d_+ ≤ r, where d_+ = inf_{α>1} D_α^reg = lim_{α↓1} D_α^reg.

• Exponential strong converse for adaptive discrimination (Corollary 5). For every r > D^reg(𝒩‖ℳ) there is κ_r > 0 such that any n-use adaptive protocol with type-II error β_n ≤ 2^{−nr} has success probability p_n ≤ 2^{−nκ_r}. The explicit constant is κ_r = ((α−1)/α)[r − D_α^reg(𝒩‖ℳ)] for an α > 1 supplied by Theorem 1.

• Stein's lemma for channels (Corollary 5, Eq. 3.3). For ε ∈ (0,1) and β_{n,ε} the infimum of type-II errors among n-use adaptive protocols with p_n ≥ 1 − ε, lim_{n→∞} −(1/n) log β_{n,ε} = D^reg(𝒩‖ℳ). Achievability uses block repetition, state Stein's lemma, and discarding leftover channel uses.

• Sharp testing threshold. The threshold for the channel hockey-stick divergence to decay exponentially is exactly d_+ = D^reg(𝒩‖ℳ) — a zero–one law at the threshold, plus the subchannel asymptotic equipartition property.

• Three concurrent proofs exist. Gao, Ji, and Liu obtain a uniform Stinespring approximation through iterative rate reduction; Schmitt and Sutter use a variational formula with spectral confinement; Zhu and Wang use a two-rate, three-amplitude argument built from a positive Choi slack and also treat general process testers, including indefinite causal order.

Methodology in Plain English

The proof strategy is a proof by contradiction built on two different ways of bounding the same quantity — the channel hockey-stick divergence E_λ, which measures the best advantage a tester can get when a false negative is penalized by a factor λ.

First, the authors suppose the limit from above (d_+) is strictly larger than the regularized relative entropy (d), and pick a rate r strictly between them. A relative-entropy testing bound then shows that at rate 2^{nr} the hockey-stick divergence stays asymptotically below one — call its square-root limsup b, so b < 1. Meanwhile, a Rényi testing bound shows the same divergence vanishes at every rate above d_+.

Second, they take a Stinespring isometry V_n of 𝒩^{⊗n} and apply their fixed-dilation version of Gour's filter twice: once at rate r to get an operator X_n with Φ_{X_n} ≤CP 2^{nr}ℳ^{⊗n} and ‖V_n − X_n‖∞ ≤ b_n, and once at a higher rate d_+ + t^{−2} to get W_n. Writing V_n = X_n + Y_n + Z_n splits the exact isometry into a cheap approximation, a correction of norm below b + o(1) < 1, and a vanishing leftover.

Third, a Schatten-norm triangle-inequality argument bounds arbitrarily many tensor copies of this block — crucially allowing inputs entangled across all copies — which bounds the regularized divergence. Setting α = n/(n−t), fixing t and letting n → ∞ removes the leftover, then letting t → ∞ drives the correction rate down to d_+. Because d_α ≥ d_+, the resulting inequality collapses to 1 ≤ b, contradicting b < 1. The order of limits is deliberate: taking n → ∞ first avoids needing uniform convergence in t.

Why This Matters

Continuity at order one was the missing technical link between Rényi-indexed bounds and the operational relative-entropy threshold. Fawzi and Fawzi characterized the adaptive strong-converse exponent via a channel Rényi chain rule but left right continuity at one open; Fang, Gour, and Wang connected that continuity to the exponential strong converse. This paper closes the gap and, in doing so, upgrades weak (vanishing-error) strong converses into exponential ones.

Potential real-world applications (not evaluated in the paper):

• Security proofs for quantum key distribution, where channel discrimination exponents translate into key-rate and finite-size bounds.
• Certification and benchmarking of quantum hardware, where distinguishing a device's actual channel from an ideal or adversarial one is the core task.
• Quantum illumination and target-detection protocols, which are channel discrimination problems in disguise.
• Quantum network design, where channel AEP results inform compression and resource-accounting for repeated channel use.

Industry relevance: The results apply to any setting where an adversary's channel must be distinguished from a benign one under many uses — a standard model in quantum cryptography and device verification. The paper's own provenance (proof proposed by GPT-6 Astra, with a Lean formalization whose theorem statement was comparator-checked and whose proofs were replayed in the Lean kernel) is itself notable for AI-assisted mathematical research and for the emerging practice of machine-checked quantum information theory.

Future Directions

• Removing the finite-max-relative-entropy assumption. The operational corollaries are stated under D_max(𝒩‖ℳ) < ∞; extending them to pairs with infinite max-relative entropy is a natural and explicitly acknowledged boundary of the current results.
• Infinite-dimensional channels. The paper works entirely with finite-dimensional spaces, and the approximation and Schatten-norm machinery would need new tools beyond that setting.
• Beyond subchannel smoothing. Gour showed that the channel AEP with trace-preserving smoothing fails in general, which is why subchannel smoothing is used; characterizing exactly which channel pairs admit stronger smoothing notions remains open.
• Process testers and indefinite causal order. Zhu and Wang treat general process testers; extending the continuity and amplification framework to that regime, and formally verifying those extensions in Lean, is a logical next step.

Target Audience

Quantum information theorists working on channel discrimination, strong converses, and Rényi divergence asymptotics; mathematically inclined researchers interested in Stinespring approximation and Schatten-norm techniques; formal-verification researchers following Lean formalizations of quantum information results; and cryptographers who rely on exponential strong converses for channel-based security proofs.

Authors’ 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.