Source-linked AI summary
Quantum hypothesis testing and the operational interpretation of the quantum Renyi relative entropies
Milan Mosonyi, Tomohiro Ogawa
TL;DR
The paper asks which quantum extension of Rényi relative entropy has the appropriate operational meaning in hypothesis testing. It analyzes direct and strong-converse regimes, showing that the new extension governs the α>1 strong-converse exponent while the traditional extension governs α<1 direct testing, with measurement attainability and CPTP monotonicity as side results.
Problem
Quantum Rényi relative entropy has multiple inequivalent extensions because non-commuting quantum states prevent a unique extension of the classical definition.
Method
The paper analyzes binary quantum hypothesis testing using direct and strong-converse error exponents, pinching, and asymptotic measurement attainability.
Results
The new quantum Rényi relative entropies have an operational interpretation as generalized cutoff rates in the α>1 strong-converse domain.
Takeaways & Limitations
The operationally relevant definition depends on α: the traditional divergence is supported for α<1, while the new divergence is supported for α>1.
Abstract
from arXiv · showhide
We show that the new quantum extension of Renyi's α-relative entropies, introduced recently by Muller-Lennert, Dupuis, Szehr, Fehr and Tomamichel, J. Math. Phys. 54, 122203, (2013), and Wilde, Winter, Yang, Commun. Math. Phys. 331, (2014), have an operational interpretation in the strong converse problem of quantum hypothesis testing. Together with related results for the direct part of quantum hypothesis testing, known as the quantum Hoeffding bound, our result suggests that the operationally relevant definition of the quantum Renyi relative entropies depends on the parameter α: for α<1, the right choice seems to be the traditional definition, whereas for α>1 the right choice is the newly introduced version. As a sideresult, we show that the new Renyi α-relative entropies are asymptotically attainable by measurements for α>1, and give a new simple proof for their monotonicity under completely positive trace-preserving maps.
I. INTRODUCTION
The introduction frames quantum Rényi divergences as non-unique extensions whose operational relevance depends on α, and connects the new α>1 version to strong-converse hypothesis testing.
- Motivation: Quantum Rényi divergences admit inequivalent extensions because quantum states are non-commutative operators rather than probability distributions.Positivity and monotonicity under completely positive trace-preserving maps are identified as basic requirements for an extension.
- Motivation: Umegaki’s relative entropy is operationally characterized by quantum Stein’s lemma as the optimal type-II error decay rate when type-I error vanishes.This establishes it as the information-theoretically relevant non-commutative extension of classical relative entropy.
- Prior operational interpretation: For α<1, traditional quantum Rényi relative entropies describe generalized cutoff rates in the direct quantum hypothesis-testing problem.The connection follows from the quantum Hoeffding bound, which characterizes the trade-off between exponential error-decay rates.
- Main contribution: For α>1, the paper identifies the new Rényi divergences with the converse side of binary quantum state discrimination.When the type-II error is forced to decay faster than the relative-entropy rate, the type-I error approaches one exponentially, with exponent given by the converse Hoeffding divergence.
- Main contribution: The new α>1 divergences are asymptotically attainable by measurements, and the paper gives a simple proof of their monotonicity under CPTP maps.The proof uses monotonicity under pinching together with asymptotic attainability.
II. PRELIMINARIES
The preliminaries establish operator, state, spectral, pinching, and asymptotic tensor-power notation used in the hypothesis-testing analysis.
- Operator notation: The paper works with linear operators, positive semidefinite operators, and density operators on a finite-dimensional Hilbert space.Density operators are positive semidefinite operators with trace 1.
- Measurement notation: A finite-valued POVM is a collection of positive operators summing to the identity, representing measurements on the Hilbert space.The measurement elements are indexed by a finite set and satisfy 0≤M_i.
- Map inequalities: Positive trace-preserving maps and their adjoints support the trace variational identities used in the subsequent testing arguments.These identities relate maximization after applying a map to the positive part of the original operator.
- Pinching: Pinching with respect to a Hermitian operator projects another operator into the operator’s commutant and is controlled by the number of distinct eigenvalues.The paper introduces the pinching operation and the associated pinching inequality for positive semidefinite operators.
- Asymptotic notation: For n copies, the paper writes ρ_n=ρ^⊗n and σ_n=σ^⊗n, then pinches ρ_n with respect to σ_n.The number v_n of distinct eigenvalues of σ_n grows at most polynomially as v_n≤(n+1)^dim H.
- Spectral notation: Spectral decompositions and projections define positive spectral parts and the support convention for powers of positive semidefinite operators.The convention sets operator powers to zero outside the support and defines A^0 as the support projection.
III. PROPERTIES OF THE NEW R´ENYI RELATIVE ENTROPIES
The section establishes asymptotic measurement attainability and monotonicity properties for the new quantum Rényi relative entropies, primarily for α > 1, under support assumptions where required.
- Asymptotic attainability: The proof of asymptotic attainability uses monotonicity under pinching by the reference state and then derives monotonicity under measurements.The argument combines the pinching lemma with classical monotonicity and tensor-power measurements.
- Assumptions and limitations: The support condition supp ρ ⊆ supp σ is necessary for the measurement monotonicity lemma and is not generally valid without it.For orthogonal supports, Fα(ρ∥σ) = −∞ while the trivial measurement produces a finite classical quantity; under the support condition, the relevant quantity is finite except for zero operators.
- Asymptotic attainability: The new Rényi relative entropies are asymptotically attainable by measurements for α > 1 in the limit of infinitely many copies.The maximization can be taken over POVMs on tensor powers, and a concrete binary Neyman–Pearson test can replace the general measurement maximum.
- Further properties: For states satisfying supp ρ ⊆ supp σ, the function Fα(ρ∥σ) is convex in α for α ≥ 1.This follows by viewing Fα(ρ∥σ) as a pointwise limit of convex functions associated with tensor powers.
- Monotonicity: The new Rényi relative entropies are monotone under completely positive trace-preserving maps for α > 1.The proof applies the asymptotic measurement characterization to tensor powers of the map and its adjoint.
- Further properties: The monotonicity conclusions also hold for trace-preserving linear maps whose tensor powers are positive, a weaker condition than complete positivity.The section additionally derives joint convexity of the associated Q quantity from monotonicity.
A. Simple Quantum Hypothesis Testing
Simple quantum hypothesis testing studies the trade-off between type I and type II errors for distinguishing tensor-power states. The direct and strong-converse regimes characterize how error probabilities behave under exponential constraints.
- Problem: A two-outcome POVM decides between null state ρ^n and alternative state σ^n from measurement outcomes on n copies.The test accepts H0 or H1, with the complementary POVM element determined by the first outcome.
- Error trade-off: The type I and type II errors cannot generally be made unconditionally small, creating an asymptotic trade-off.The direct analysis fixes a constraint on one error while optimizing the other.
- Direct regime: For any fixed ε in (0,1), quantum Stein’s lemma identifies D(ρ∥σ) as the asymptotic type II error exponent while type I error vanishes.A suitable test sequence satisfies n^-1 log β_n(T_n) = −D(ρ∥σ) and α_n(T_n) → 0.
- Direct regime: When the required type II exponent r is below D(ρ∥σ), the optimal type I error vanishes exponentially at the converse Hoeffding rate.This rate is expressed using the traditional quantum Rényi relative entropies and the Hoeffding divergence.
- Strong-converse regime: If r exceeds D(ρ∥σ), the type I error approaches one exponentially, defining the strong-converse regime.The strong-converse exponent instead tracks the decay of the success probability 1 − α_n(T_n).
- Assumption: The strong-converse property requires supp ρ ⊆ supp σ; without it, tests can have zero type II error while type I error vanishes.The paper therefore assumes this support inclusion for the remainder of the analysis.
B. Exponents for the Neyman-Pearson tests
The paper analyzes Neyman–Pearson tests through threshold exponents and establishes their asymptotic behavior in the non-trivial interval between relative and max-relative entropy.
- Setup: Neyman–Pearson tests S_n(a) are analyzed using a trade-off threshold a to determine type I success and type II error probabilities.The analysis excludes the trivial case ρ = σ and focuses initially on a < D_max(ρ∥σ).
- Threshold range: S_n(a) is zero exactly when a ≥ D_max(ρ∥σ), where D_max(ρ∥σ) is defined by ρ ≤ e^γσ.Thus the max-relative entropy marks the upper boundary of the nontrivial threshold range.
- Asymptotic exponents: The asymptotic inequalities for Neyman–Pearson error exponents become equalities when D(ρ∥σ) < a < D_max(ρ∥σ).Outside this interval, the paper identifies trivial or separately characterized regimes.
- Measurement attainability: For every α > 1, a threshold a_α yields post-measurement binary states whose new Rényi divergence grows asymptotically as exp(nF(α)).The construction uses the asymptotic type I and type II rates of the Neyman–Pearson tests.
C. The strong converse exponent
The paper determines the exact strong-converse exponent for binary quantum hypothesis testing and expresses it through the newly introduced quantum Rényi relative entropies.
- Main theorem: The main goal is to prove that the strong-converse exponent equals the converse Hoeffding bound.One inequality follows from Rényi-divergence monotonicity; equality is established using Neyman–Pearson asymptotics.
- Assumptions: The proof assumes ρ ≠ σ and supp ρ ⊆ supp σ, ensuring that the strong-converse regime is nontrivial.The argument treats separate regions of the imposed type II exponent r.
- Piecewise characterization: The resulting strong-converse exponent is piecewise described through the convex rate function φ and the max-relative entropy boundary.For r below the threshold r_max it uses the solution of r − a_r = φ(a_r), while beyond r_max it becomes r − D_max(ρ∥σ).
- Achievability: The construction uses modified Neyman–Pearson tests whose type II error exponent is r and whose type I success exponent is r − a.The corresponding asymptotic relations are given explicitly for Tr σ^nT_n(r,a) and Tr ρ^nT_n(r,a).
- Operational interpretation: The new Rényi relative entropies are essentially Legendre–Fenchel transforms of the operational strong-converse exponent.This provides a direct operational interpretation for α > 1.
- Consequences: Explicit Neyman–Pearson error-rate calculations provide measurement attainability for the new Rényi divergences, beyond the alternative information-spectrum route.The result also yields an operational proof of the Lieb–Thirring inequality.
D. Representation as cutoff rates
The paper connects generalized cutoff rates with strong-converse exponents and uses this operational representation to derive monotonicity of the new Rényi divergences.
- Cutoff-rate definition: The generalized κ-cutoff rate C_κ(ρ∥σ) is defined as the smallest r_0 satisfying an asymptotic strong-converse condition.The construction assumes supp ρ ⊆ supp σ and ρ ≠ σ.
- Rate structure: The strong-converse exponent is monotone increasing and convex as a function of the imposed type II exponent r.These properties support its representation through convex duality.
- Representation: The operational interpretation identifies the new Rényi relative entropies with a transform of the strong-converse exponent and cutoff-rate quantities.The equality is obtained through the convex function ˜ψ and its derivative-based characterization.
- Monotonicity: Monotonicity of the new Rényi divergences follows as an immediate consequence of their operational interpretation.The resulting monotonicity statement applies for α > 1 under the stated positivity condition on tensor powers of the map.
- Proof mechanism: The proof compares optimization over tests before and after applying the map through its positive unital Hilbert–Schmidt dual.This transfers tests on the output system into admissible tests for the input system.
V. CONCLUSION
The paper determines the exact strong converse exponent for binary quantum hypothesis testing and identifies the operationally relevant quantum Rényi divergence according to α. It also contrasts this with the α<1 direct-testing regime and outlines the proof strategy.
- V. CONCLUSION: The exact strong converse exponent for binary quantum hypothesis testing is expressed using the new quantum Rényi relative entropies D(new)_α for α > 1.The result gives these divergences a direct operational interpretation as generalized cutoff rates.
- V. CONCLUSION: For α < 1, previous quantum hypothesis-testing results identify the traditional Rényi relative entropies D(old)_α as operationally relevant.
- V. CONCLUSION: The converse Hoeffding divergence is shown to be optimal using monotonicity of D(new)_α under measurements for α > 1.
- V. CONCLUSION: Achievability follows from quantum Neyman–Pearson tests, pinching, classical large-deviation theory, and asymptotic attainability results.The tests are chosen using a suitable trade-off parameter, with modifications for sufficiently large rates.
Appendix A: Monotonicity and attainability properties of the R´enyi divergences
Appendix A organizes monotonicity and attainability properties of quantum Rényi divergences across classes of maps and asymptotic procedures. It derives several new properties for D(new)_α with α > 1 and records limitations and open questions.
- Properties and map classes: The appendix studies monotonicity under positive, completely positive, stochastic, measurement, pinching, and tensor-power-positive maps.These properties are denoted MON, SMON, EPPMON, MMON, and PMON, with map classes specified in the appendix.
- Attainability properties: Asymptotic attainability is defined through sequences of maps on tensor-power states, including measurements and pinching with respect to σ⊗n.
- Limitations and prior results: The old divergences fail MON for α > 2, while monotonicity of the new divergences has been established over broader parameter ranges by several methods.For the old divergences, the failure is linked to the lack of operator convexity of x 7→x^α for α > 2.
- New-divergence results: For α > 1, PMON yields MMON and AAP, which in turn imply AAM and the stronger monotonicity property EPPMON.
- Extension properties: The new Rényi relative entropies provide the smallest possible quantum extension of the classical Rényi relative entropies under mild conditions.