Asymptotic Error Rates in Quantum Hypothesis Testing

Quantum hypothesis testing

• Quantum state discrimination uses optimal measurements on tensor-product states ρ⊗n and σ⊗n, extending likelihood-ratio testing from classical distributions.The equal-prior error probability is connected to the trace norm distance between the states.Source: §1 · pp. 2–3

Quantum Chernoff distance

• The quantum Chernoff distance has operational meaning as a distinguishability measure and avoids the tensor-power inconsistency of the trace norm distance.It characterizes the asymptotic behavior of trace-norm discrimination for many identical copies.Source: §1 · pp. 3–5

Symmetric hypothesis testing

• The quantum Chernoff distance gives the asymptotic error exponent for optimal discrimination of two quantum states prepared in many identical copies.The exponent is achieved by the Holevo-Helstrom strategy and generalizes the classical Chernoff distance.Source: §1 · pp. 3–4

Scope and contributions

• The paper provides a complete, unified treatment of the symmetric and asymmetric quantum hypothesis-testing results, including states with different supports.Its main techniques establish both optimality and achievability of the symmetric bound.Source: §1 · pp. 2, 4

Relation to relative entropy

• The quantum Chernoff distance is presented as more natural than quantum relative entropy for many distance-measure applications because relative entropy is asymmetric and discontinuous.Relative entropy nevertheless remains operationally central for asymmetric testing, where it determines the optimal type-II error rate under a type-I constraint.Source: §1 · p. 5

Asymmetric hypothesis testing

• The paper treats asymmetric testing through the quantum Hoeffding bound, deriving quantum Stein’s Lemma and quantum Sanov’s theorem from it.The Hoeffding construction replaces the classical term ∑i p(i)^(1-s)q(i)^s with Tr[ρ^(1-s)σ^s].Source: §1 · pp. 5–6

Problem formulation

• A two-outcome POVM {1 −Π, Π} defines type-I and type-II errors α and β, with Π representing acceptance of the alternative.The average Bayesian error is Pe = π0α + π1β.Source: §2 · p. 6
• The Bayesian distinguishability problem minimizes Pe = π0α + π1β, with equal priors giving the symmetric special case.The priors π0 and π1 are positive and sum to one.Source: §2 · p. 6

Quantum Neyman–Pearson testing

• The quantum Neyman–Pearson test projects onto the positive part of Tσ − ρ and minimizes α + Tβ among all two-outcome POVMs.If another test has α ≤ α*, its type-II error satisfies β ≥ β*.Source: §2 · p. 7
• The optimal projector lies in the algebra generated by ρ and σ, and commuting hypotheses reduce to the classical maximum-likelihood test.Thus restricting the physical observable algebra does not change the optimal discrimination test.Source: §2 · pp. 7–8

Problem formulation

• Two hypotheses correspond to states ρ and σ, while n-copy discrimination uses POVMs on ρ⊗n and σ⊗n.The states are density operators on a finite-dimensional Hilbert space.Source: §2 · pp. 6, 8

Asymptotic formulation

• The paper defines lower and upper rate limits for error sequences to handle cases where asymptotic error exponents do not converge.These rate limits are applied to type-I and type-II error sequences for measurements Π = (Πn).Source: §2 · p. 8

3. Bayesian Quantum Hypothesis Testing: Quantum Chernoff Bound

• The optimal Bayesian error probability for n copies decays exponentially, with rate equal to the quantum Chernoff distance ξQCB.For finite-dimensional states and nondegenerate priors, the rate limit of P*e,n is ξQCB.Source: §3 · p. 9
• The quantum Chernoff distance is ξQCB = −log inf0≤s≤1 Tr[ρ1−sσs], and it reduces to the classical Chernoff distance when ρ and σ commute.The trace expression is real and non-negative for every s in [0, 1].Source: §3 · p. 9

3.1. Proof of Theorem 1: Optimality Part.

• Optimality follows by mapping density operators to probability distributions while preserving tensor powers and the relevant Chernoff expression.The resulting classical product distributions allow the classical Chernoff bound to lower-bound every quantum test’s error exponent.Source: §3.1 · pp. 9, 10, 12
• The mapping uses eigenvalues and eigenbases of ρ and σ to construct d^2-dimensional probability distributions whose classical discrimination problem bounds the quantum one.The proof applies the classical Chernoff bound and then identifies the resulting expression with ξQCB.Source: §3.1 · pp. 9, 10, 12

3.2. Proof of Theorem 1: Achievability Part.

• Achievability uses a matrix inequality relating trace-norm discrimination to Tr[ρ1−sσs], yielding the matching exponent ξQCB.The proof reduces the inequality to positive operators and establishes it through integral representations and derivative arguments.Source: §3.2 · pp. 12, 13, 14, 16
• The Chernoff quantity Qs = Tr[ρ1−sσs] and trace distance T = ∥ρ − σ∥1/2 obey Qs + T ≥ 1, with equality attainable for suitable commuting states.For ρ = |0⟩⟨0| and σ = (1 − T)|0⟩⟨0| + T|1⟩⟨1|, Qs = 1 − T.Source: §3.2 · p. 13
• The asymptotically optimal projector can be suboptimal at finite n: it is Π(s*) for a minimizing s*, or the Holevo–Helstrom projector when no minimizer exists.The Holevo–Helstrom projector remains optimal for every finite n.Source: §3.2 · p. 14
• The achievability proof handles non-invertible operators by a limiting regularization, covering states that are not faithful.This supports the paper’s full-generality treatment beyond full-support hypotheses.Source: §3.2 · p. 15

4. Properties of the Quantum Chernoff Distance

• The quantum Chernoff distance is not a metric because it violates the triangle inequality, but it retains many properties expected of a distance measure.The paper derives these properties from the non-logarithmic quantity Q.Source: §4 · p. 16

4.1. Relation to Fidelity and Trace Distance.

• The Chernoff quantity Q is bounded through fidelity and trace distance, producing a chain of relations that connects operational discrimination with familiar state distances.The bounds combine the trace-distance relation to fidelity with inequalities for Q.Source: §4.1 · p. 17
• The lower bound involving fidelity is sharp when one state is pure.For ρ = |ψ⟩⟨ψ|, the minimizing Chernoff expression equals ⟨ψ|σ|ψ⟩.Source: §4.1 · p. 17
• The upper bound on Q extends beyond finite dimensions, since the proof also works in countably infinite dimensions.The cited proofs establish the bound Q ≤ 1 − T^2.Source: §4.1 · p. 17
• Q equals 1 only for identical states and 0 only for orthogonal states, so ξQCB ranges over [0, ∞] and is infinite on orthogonal states.This differs from relative entropy, which can be infinite whenever supports differ.Source: §4.2 · p. 17

4.2. Range of Q.

• For pure states, Q = F^2 and therefore ξQCB = −2 log F.Unlike fidelity itself, the Chernoff distance is expressed through the logarithm of this squared fidelity.Source: §4.3 · p. 17

4.3. Triangle inequality.

• For the normal shift family N(µ, 1), ξCB = (µ1 −µ2)^2 /8, so it defines a squared metric rather than an ordinary metric.Source: §4.3 · p. 17
• The squared-metric behavior explains why ξCB is not expected to satisfy the triangle inequality even in this structured classical family.Source: §4.3 · p. 17
• The classical Chernoff expression violates the triangle inequality for suitable Bernoulli laws as ε → 0.The comparison between Be(1/2), Be(ε), and Be(1−ε) yields the contradiction.Source: §4.3 · p. 17

4.4. Convexity of Qs as a function of s.

• For positive scalars x and y, s 7→x1−sys is convex because its second derivative is non-negative.Source: §4.4 · p. 18
• When one parameter is zero, s 7→x1−sys remains convex despite being discontinuous at s = 1.Source: §4.4 · p. 18
• By Lieb’s theorem [19], Tr[ρ1−sσs] is jointly concave in (ρ, σ) for each fixed s, making Q jointly concave as a point-wise minimum.Source: §4.5 · p. 18

4.5. Joint concavity of Q in (ρ, σ).

• Joint concavity of Q implies monotonicity under any completely positive trace-preserving map Φ.Source: §4.6 · p. 18

4.6. Monotonicity under CPT maps.

• Q is invariant under unitary conjugations and under adding an ancillary density operator τ.Source: §4.6 · p. 18
• Q is monotone under general completely positive trace-preserving maps by reducing the proof to partial trace through a unitary Stinespring representation.Source: §4.6 · p. 18
• The partial-trace case follows because the partial trace is a convex combination of unitary conjugations, allowing concavity and unitary invariance to establish monotonicity.Source: §4.6 · p. 18
• The bounds Q + T ≥1 and ξQCB = −log Q ≤−log(1 −T ) show continuity with respect to trace distance.Source: §4.7 · p. 19

• The classical Chernoff distance is related to the relative entropy between two probability distributions.Source: §4.8 · p. 19
• The relation is stated in its simplest form when the two discrete probability distributions have coinciding supports.Source: §4.8 · p. 19
• The section connects asymptotic discrimination quantities through the classical distributions p and q.Source: §4.8 · p. 19

• For distributions with different supports, the corresponding classical formulas require modification.Source: §4.8 · p. 19
• The minimizer s* of inf_s∈[0,1] log Q_s can be characterized by a generalized version of the classical condition.Source: §4.8 · p. 19
• For full-support states, Q_s = Tr(ρ^(1−s)σ^s) is analytic and convexity makes its local infimum over [0,1] a global minimum.Source: §4.8 · p. 19
• The minimizing parameter for the quantum expression is obtained by differentiating Q_s and imposing the resulting condition.Source: §4.8 · pp. 19–20
• Although ρ_s is generally not a density operator or self-adjoint, it has positive spectrum and supports the entropy expressions used in the analysis.Source: §4.8 · p. 20
• At the minimizing s, the two relevant relative entropies coincide in the variational characterization of Q.Source: §4.8 · p. 20
• The family ρ_s interpolates between ρ and σ as a quantum analogue of the Hellinger arc, but generally lies outside the state space.Source: §4.8 · p. 20
• The classical identity relating Chernoff distance and relative entropy fails for arbitrary noncommuting quantum density operators.Source: §4.8 · pp. 19–20

5.1. The Classical Hoeffding Bound.

• The classical HBCL theorem characterizes the optimal tradeoff: for each r > 0, βR ≥ r is achievable with αR = e(r), while βR > r implies αR ≤ e(r).The curve e(r) is non-increasing and convex, with e(0) = H(q∥p) and e(H(p∥q)) = 0.Source: §5.1 · p. 21

5.2. Nonequivalent hypotheses.

• For nonequivalent classical hypotheses, conditioning on the intersecting support reduces testing to equivalent conditional measures and yields the shifted exponent −log ψ0 + ˜e(r + log ψ1).For 0 ≤ r < −log ψ1 when ψ1 < 1, the exponent is ∞ and αn = 0 is achievable for every n.Source: §5.2 · pp. 23–24
• The general classical exponent is infinite on [0, −log ψ1), equals H(˜q∥p) at −log ψ1, decreases convexly to H(p∥q), then is constant at −log ψ0.This extends the HBCL theorem to arbitrary finite-support probability measures and explains the support-induced shift of the rate curve.Source: §5.2 · pp. 24–25

5.3. Quantum Hoeffding Bound.

• For quantum states, eQ(r) has the same support-conditioned form, with ψ0 = Tr[ρ supp σ] and ψ1 = Tr[σ supp ρ].Its critical behavior is inherited from the classical conditional exponent, including convex decrease and a constant tail.Source: §5.3 · p. 25
• When both quantum hypotheses are pure, the error-exponent curve degenerates to a rectangular form, taking eQ(r) = ∞ or eQ(r) = −log ψ1.More generally, Sσ(ρ∥σ) remains finite outside supp σ even when the usual relative entropy S(ρ∥σ) is infinite.Source: §5.3 · pp. 26–27
• The quantum HBCL theorem states that test projections achieve αR = eQ(r) with βR ≥ r, while every sequence with βR > r satisfies αR ≤ eQ(r).Thus every attainable rate pair lies on or below the quantum error-exponent curve, and every point on that curve is achievable.Source: §5.3 · p. 27
• Unlike the classical case, no general quantum variational formula of the form eQ(r) = infτ{S(τ∥ρ) : S(τ∥σ) ≤ r} is known.The passage states that this remains unresolved even when both states are faithful.Source: §5.3 · p. 27

5.4. Proof of Optimality of the Quantum Hoeffding Bound.

• For every measurement sequence, the type-I error exponent is bounded by eQ(r) whenever the type-II exponent is at least r, proving Hoeffding optimality.The proof combines classical Hoeffding bounds with monotonicity and continuity of e(r), including separate cases for the optimizing rate.Source: §5.4 · p. 29

5.5. Proof of Achievability of the Quantum Hoeffding Bound.

• The projector test based on the positive part of a−b achieves the quantum Hoeffding bound, and the proof applies without assuming faithful states, including non-orthogonal pure states.The construction uses Qs = Tr[ρ^(1−s)σ^s] and chooses x to obtain the required error-exponent inequalities; the achievability argument is mainly due to Hayashi.Source: §5.5 · pp. 29–30

5.6. Quantum Stein’s Lemma and quantum version of Sanov’s Theorem.

• Quantum Stein’s lemma gives the optimal type-II error exponent S(ρ∥σ), independently of fixed 0 < ǫ < 1, by combining Hoeffding achievability with Proposition 1 in.The result was first obtained by Hiai and Petz, with its optimality later strengthened by Ogawa and Nagaoka.Source: §5.6 · pp. 30–31
• In the composite-null setting, quantum Sanov’s theorem identifies the type-II exponent with the relative entropy distance S(Γ∥σ), which is also an upper bound under the type-I constraint.The theorem follows from the quantum HBCL theorem and the equivalence established in Theorem 1 of.Source: §5.6 · p. 31

Theorem 6

• Theorem 6 bounds the relevant positive-operator expression for 0 ≤ s ≤ 1, specializing to a fidelity–Q_s relation for quantum states.For A = σ and B = ρ, the left-hand side becomes F(ρ, σ), while the right-hand side becomes Q_s(ρ, σ)^1/2.Source: A. Proofs of Bounds on Q · p. 32

Direct proof

• A direct proof of inequality (30) is given for general positive operators and works in infinite dimensions.This proof bypasses inequality (29), broadening the stated setting beyond the finite-dimensional presentation.Source: A. Proofs of Bounds on Q · p. 32

Theorem 7

• Theorem 7 states a further inequality for positive operators, derived by decomposing two operators into their sum and difference.The proof introduces S = P + Q and D = P − Q, with P = (S + D)/2 and Q = (S − D)/2.Source: A. Proofs of Bounds on Q · p. 32

Proof techniques

• The operator bounds rely on trace-norm estimates, Hölder’s inequality, positivity, and repeated use of a relation for positive operators.The derivation proceeds through square-root substitutions and then squares the resulting inequality.Source: A. Proofs of Bounds on Q · pp. 32–33