Source-linked AI summary
Correlation Detection and an Operational Interpretation of the Renyi Mutual Information
Masahito Hayashi, Marco Tomamichel
TL;DR
The paper asks which Rényi mutual-information and conditional-entropy definitions correspond to operational quantities in composite quantum hypothesis testing. It analyzes correlation-detection and decoupling tests across Hoeffding and strong-converse regimes, showing that the resulting exponents identify different Rényi quantities and connect naturally to channel coding.
Problem
The paper addresses which of the many proposed Rényi mutual-information and conditional-entropy definitions have operational meanings in relevant hypothesis-testing tasks.
Method
It analyzes composite tests distinguishing a fixed bipartite state from product alternatives sharing one marginal, and decoupling tests with a maximally mixed system, using asymptotic error exponents.
Results
The Hoeffding and strong-converse exponents operationally determine Rényi mutual information and conditional entropy in their respective parameter regimes, including sandwiched variants for α > 1.
Takeaways & Limitations
The results give Rényi mutual information and conditional entropy operational interpretations in quantum hypothesis testing, channel coding, and classical probability specializations.
Takeaways & Limitations
The channel-coding formulation assumes an alternative generated by a useless channel whose output is decoupled from the environment.
Abstract
from arXiv · showhide
A variety of new measures of quantum Renyi mutual information and quantum Renyi conditional entropy have recently been proposed, and some of their mathematical properties explored. Here, we show that the Renyi mutual information attains operational meaning in the context of composite hypothesis testing, when the null hypothesis is a fixed bipartite state and the alternate hypothesis consists of all product states that share one marginal with the null hypothesis. This hypothesis testing problem occurs naturally in channel coding, where it corresponds to testing whether a state is the output of a given quantum channel or of a 'useless' channel whose output is decoupled from the environment. Similarly, we establish an operational interpretation of Renyi conditional entropy by choosing an alternative hypothesis that consists of product states that are maximally mixed on one system. Specialized to classical probability distributions, our results also establish an operational interpretation of Renyi mutual information and Renyi conditional entropy.
1. INTRODUCTION
The paper gives composite hypothesis testing an operational role in identifying correlations and interpreting Rényi mutual information and conditional entropy. It characterizes error trade-offs across rate regimes, connects the test to channel coding, and extends the analysis to second-order behavior and classical distributions.
- The test models channel-coding problems by distinguishing a channel output from a useless channel whose output is decoupled from the environment.
- For rates R below I(A : B)ρ, the minimum first-kind error vanishes exponentially, yielding a quantum Stein’s lemma for the composite test.
- The quantum Hoeffding and strong-converse exponents operationally identify different Rényi mutual-information definitions in the regimes α < 1 and α > 1, respectively.
- For rates R above I(A : B)ρ, the first-kind error approaches one exponentially, establishing a strong converse whose exponent is determined by sandwiched Rényi mutual information.
- A related decoupling test operationally characterizes Rényi conditional entropies through Hoeffding and strong-converse exponents, while classical specialization supplies corresponding interpretations for probability distributions.
- The paper combines pinching and irreducible decomposition to derive strong-converse exponents for the composite alternative and also analyzes second-order asymptotics.
2. NOTATION AND PRELIMINARIES
The paper establishes finite-dimensional quantum-system notation and introduces a universal state with permutation and product-unitary invariance. This state supports the later analysis of composite hypothesis testing and Rényi quantities.
- Notation: Quantum systems are modeled by finite-dimensional Hilbert spaces, with tensor-power systems representing n copies and standard operator notation for states, traces, and marginals.The maximally mixed state is denoted πA = 1A/|A|.
- Preliminaries: Pinching maps are defined from spectral decompositions, while permutation invariance and product-unitary invariance organize n-copy operators.These structures are used in the construction of the universal state.
- Universal state: For every n, a universal state exists that is permutation invariant, invariant under n-fold product unitaries, and commutes with all permutation-invariant states.Its construction uses Schur–Weyl duality and block decompositions indexed by Young diagrams.
- Universal state: The universal-state construction is related to prior representation-theoretic constructions, including one whose constant is not optimal.The paper identifies related work while using its own construction for the subsequent arguments.
- Universal state: The universal state's Young-diagram index set has size at most (n + 1)^(d−1), providing a polynomial dimension bound for the construction.The bound follows by comparing Young diagrams with types of length-n strings over d symbols.
C. R´enyi Divergence
This section defines two quantum Rényi divergence families and generalized Rényi mutual-information quantities. It records their data-processing behavior, limiting relation to relative entropy, and polynomial spectral corrections used later.
- Rényi divergence: The paper defines the Rényi relative entropy and sandwiched Rényi divergence for quantum states and positive semidefinite operators, subject to support conditions.Both definitions extend continuously to limiting orders, with infinity assigned when the relevant support condition fails.
- Properties: Data processing holds for the ordinary divergence on α ∈ [0,2] and for the sandwiched divergence on α ∈ [1/2,∞).The paper also records an isometric invariance consequence used in later reductions.
- Rényi divergence: The two divergence families coincide for commuting operators, and both converge to relative entropy as α approaches 1.Several special orders connect these quantities with previously studied min- and max-entropies.
- Properties: Spectral pinching changes Rényi divergences by at most logarithmic terms involving the number of distinct eigenvalues, with a doubled bound for α > 2.For tensor powers, the number of distinct eigenvalues grows polynomially in n, yielding corresponding logarithmic corrections.
- Generalized quantities: The generalized Rényi mutual informations minimize the ordinary or sandwiched divergence between ρAB and τA ⊗ σB over σB.Choosing τA = ρA recovers Rényi mutual information, while choosing τA = 1A recovers Rényi conditional entropy.
B. Characterization of the Minimizers
The section characterizes minimizers and duality relations for generalized Rényi mutual informations, then derives additivity and regularity properties needed for the asymptotic analysis.
- Characterization of minimizers: At α = 1, minimizing D(ρAB∥τA ⊗ σB) over σB yields σB = ρB and I(ρAB∥τA) = D(ρAB∥τA ⊗ ρB).This follows from the chain rule and positivity of relative entropy.
- Characterization of minimizers: For general α, the ordinary Rényi mutual-information minimizer is uniquely characterized by a quantum Sibson identity, whereas the sandwiched minimizer is specified by a nonlinear fixed-point condition.The fixed-point characterization is proved separately in an appendix.
- Duality: A duality relation connects generalized mutual informations for complementary purifications and paired Rényi orders satisfying α + β = 2.The relation generalizes established duality formulas for Rényi conditional entropy and mutual information.
- Additivity: The generalized and sandwiched mutual informations are additive on product states over the ranges established by the paper's lemmas.The proof uses duality relations and product purifications.
- Regularity: The sandwiched generalized mutual information is continuous and monotonically increasing in α, and t ↦ t eI_(1+t) is continuous and convex.These properties follow from its representation as a uniformly convergent limit of classical Rényi divergences.
F. Differentiability in α
The paper formulates a composite hypothesis-testing problem with a fixed bipartite null state and product alternatives sharing a prescribed marginal. It studies type-I error minimization under type-II constraints, while noting a limitation in differentiability analysis.
- Differentiability in α: The continuity argument for the sandwiched Rényi mutual information does not establish differentiability in α.A separate proposition proves continuous differentiability for α ≥ 1, with supporting arguments supplied in appendices.
- Composite hypothesis test: The formulation is deliberately more general so that the two hypothesis-testing problems introduced earlier can be treated as special cases.This unified setup is used to derive operational meanings for both mutual-information and conditional-entropy quantities.
- Composite hypothesis test: The composite alternative consists of states τA ⊗ σB for arbitrary σB, while the null hypothesis is the fixed bipartite state ρAB with ρA ≪ τA.The n-copy problem permits an arbitrary alternative state on B^n, not only product or permutation-invariant states.
- Composite hypothesis test: A binary test operator defines type-I and type-II errors, and the central quantity minimizes type-I error subject to a prescribed upper bound on type-II error.The paper extends the one-copy optimization to n copies and studies its asymptotic behavior.
5. HOEFFDING BOUND
The Hoeffding analysis characterizes the trade-off when the second-kind error decays below the mutual-information rate, with exponents determined by Rényi mutual information.
- The first-kind error decays exponentially when the second-kind error rate R is below I(ρAB∥τA), with exponent given by generalized Rényi mutual information for α < 1.
- For R ≥ I(ρAB∥τA), the first-kind error decays slower than exponentially; for R < I0(ρAB∥τA), it can decay faster than exponentially.
- The achievability proof uses permutation-invariant tests, universal states, additivity, and a minimax argument, while optimality invokes the converse quantum Hoeffding bound.
6. STRONG CONVERSE EXPONENT
The strong-converse analysis treats rates above the mutual information, where the first-kind error approaches one exponentially and its exponent is governed by sandwiched Rényi mutual information.
- When R exceeds I(ρAB∥τA), the first-kind error converges to one exponentially, with exponent determined by sandwiched Rényi mutual information for s > 1.
- The stated strong-converse result is restricted to rates sufficiently close to I(ρAB∥τA), although the valid range can be extended.
- The special case τA = ρA gives an operational interpretation of sandwiched Rényi mutual information eI↓α(A|B)ρ for α > 1.
- Achievability combines pinching, a classical Neyman–Pearson test, and large-deviation analysis based on a Gärtner–Ellis theorem variant.
C. Proof of Achievability
The achievability proofs construct tests for the below-threshold and above-threshold regimes, then derive the corresponding exponent bounds through asymptotic and convex-analytic arguments.
- For the strong-converse regime, the proof constructs tests satisfying the second-kind error constraint and analyzes the first-kind error through random variables and asymptotic cumulant generating functions.
- Convexity and a minimax theorem transform the optimization into the strong-converse exponent expression.
- Optimality fixes an auxiliary state, applies a converse bound, and maximizes over that state to obtain the desired result.
7. STEIN’S LEMMA AND SECOND ORDER
The paper derives Stein-type and strong-converse consequences, then analyzes second-order behavior, finding convergence of the first-kind error to a Gaussian-determined constant.
- The error-exponent and strong-converse results directly yield a variant of Stein’s lemma and its strong converse.
- In the second-order regime, the first-kind error converges to a constant when the second-kind error vanishes at the specified exponential rate.
- The limiting expression involves the cumulative standard normal distribution Φ.
- Second-order achievability uses the Section 6 test with a threshold containing nI(ρAB∥τA), a square-root term, and a logarithmic correction.
- Optimality follows from an existing second-order expansion for binary quantum hypothesis testing.
Appendix A: A Minimax Theorem
Appendix A develops a minimax theorem using weaker concavity and convexity notions, then applies it under compactness, semicontinuity, and mixed concavity assumptions.
- The theorem uses 1/2-concavelike and 1/2-convexlike conditions defined through interpolating points in X and Y.
- König’s minimax theorem requires compact Hausdorff Y, lower semicontinuity in y, and the corresponding generalized concavity and convexity assumptions.
- Proposition 21 applies the minimax theorem to a concave function on X and a 1/2-convexlike, lower-semicontinuous function on compact Hausdorff Y.
- The proof verifies generalized concavity for g(x,y) by selecting an intermediate x between arbitrary x1 and x2.
Appendix C: Characterization of Minimizers
Appendix C characterizes optimizing marginal states through differentiability, strict curvature, and fixed-point conditions for the relevant trace functionals.
- The optimization can be restricted to compact sets of states whose eigenvalues are bounded away from zero, with nonempty optimizer sets.
- Mα(B) equals the fixed-point set Fα(B), identifying the optimizing marginal states with solutions of the nonlinear fixed-point map.
- The derivative and spanning arguments rely on full-support marginals and on spanning the traceless Hermitian operators with admissible perturbations.
- The optimizer condition is equivalent to vanishing directional derivatives, which forces the associated operator to be proportional to the identity.
- Strict concavity for α < 1 and strict convexity for α > 1 provide uniqueness of the optimizer when a fixed point exists.
Appendix D: Differentiability of the R´enyi Mutual Information
Appendix D proves continuous differentiability of the Renyi mutual information by combining smooth optimization, definite Hessians, and a separate analysis at α = 1.
- The auxiliary envelope result gives continuous differentiability of an optimized function when the minimizer is interior and the Hessian is positive definite.
- For α away from 1, the result follows from twice continuously Fréchet differentiable objectives and definite Hessians at interior optima.
- At α = 1, the proof computes the derivative using the limiting optimizer and the identities supplied by earlier lemmas.
- Continuity at α = 1 follows after showing that the optimizing marginal converges to ρB, using the implicit function theorem.