NOTE-021: Distinguishability and Accessible Information in Quantum Theory

Contribution

The thesis defines each “quantum distinguishability” as a classical distinguishability measure extremized over all POVMs (Eqs. 3.3–3.7). It then does four things with that definition:

1. Fidelity. It solves the fidelity case exactly, identifies the optimal measurement, and derives the fidelity’s alternative representations.
2. Holevo bound. It re-derives the Holevo bound for binary ensembles so that the proof itself specifies a measurement. That yields a new lower bound M(t) and an explicit form of Holevo’s implicit intermediate bound L(t), plus further bounds N(t), R(t) and P(t).
3. Quantum KL. It gives lower bounds on the POVM-optimized (“quantum Kullback”) KL information of the form tr ρ0 ln Λ, generated by operator equations. It also shows that the Umegaki relative entropy upper-bounds it.
4. No-broadcasting. It proves the mixed-state no-broadcasting theorem. The proof rests on a small set of fidelity properties: monotonicity under partial trace, multiplicativity on products, and unitary invariance.

Ch. 2 is expository, as the author says (p. 2: “little of the work represents original research”). It gives operational derivations of the classical measures.

Key insight

Optimizing a classical distinguishability measure over all measurements becomes tractable when the per-outcome term can be bounded, by a Schwarz inequality, below or above by something linear in E_b. POVM completeness Σ_b E_b = 1 then sums the measurement away. The work is in inserting a free unitary (Eq. 3.54), so that the Schwarz equality condition can actually be met by some POVM. Fixing that unitary by polar decomposition (Eqs. 3.60–3.64) produces both the bound’s value and the measurement that attains it.

The same move applied to J”(t) is what the thesis uses to simplify Holevo’s bound. The mutual information’s second derivative contains no logarithm, so it can be minimized over POVMs in the same way. The minimizing measurement is the eigenbasis of the symmetric logarithmic derivative L_ρ(Δ) (Eq. 3.181), and it gives both L(t) and M(t). The operators that achieve these bounds (M for fidelity, L_ρ(ρ0) for the mutual information) behave as operator likelihood ratios: their eigenvalues are (square roots of) ratios of outcome probabilities (Eqs. 3.92, 3.201).

Assumptions

• Finite dimension. Hilbert space of finite dimension D throughout (§1.3.2, §3.5.1: “we retain the finiteness assumptions made by Holevo”). The number of POVM outcomes is unrestricted.
• Binary ensembles only. Every accessible-information result is for two signal states {ρ0, ρ1} with priors π0 = 1−t, π1 = t (Eq. 3.139). Nothing is stated for three or more states.
• Invertibility. Many derivations assume invertibility “for simplicity”:
  • fidelity: ρ1 invertible, with a separate support/null-space construction for the non-invertible case (Eqs. 3.69–3.74);
  • M(t): ρ0 and ρ1 invertible (§3.5.2). §1.2 says instead “when ρ0 and ρ1 have a common null subspace” (Eq. 1.10);
  • the Kronecker-sum formula for L(t): ρ0 and Δ = ρ1 − ρ0 invertible (Eq. 3.243);
  • R(t) and the Umegaki bound;
  • no-broadcasting: ρ0 and ρ1 invertible, “to simplify the exposition” (p. 118). The non-invertible case is not proved here.
• Measurement model. Measurements are POVMs, justified by Neumark’s theorem (§1.3.3). Operations are Kraus maps (Thm 4.1). The physical processes allowed for broadcasting are the most general ancilla-unitary-partial-trace maps (Eq. 4.51).
• Probability interpretation. Probabilities and quantum states are read as Bayesian states of knowledge (§1.3.1–1.3.2). This is interpretive framing. No theorem depends on it, although the argument at p. 120 that the entangled states of Eq. 4.105 are unrealizable is phrased partly in those terms before it is proved formally.
• Classical side. Finite outcome alphabets. Aczél’s uniqueness theorem for the log score requires n ≥ 3 outcomes (Thm 2.3). For n = 2, Muszély’s family (Eq. 2.75) shows the honest-expert payoff is not unique.

Key results

Equation numbers are the thesis’s own. Logs are natural (nats) in Chs. 2–3, except where noted.

Classical measures (Ch. 2)

• Bayes error. Pe = Σ_b min{π0p0(b), π1p1(b)} (Eq. 2.11). Cover’s example (Eqs. 2.12–2.17) shows the ranking by one-sample Pe can reverse with two samples.
• Chernoff (Thm 2.1). Pe(N) ≤ λ^N with λ = min_{0≤α≤1} F_α(p0/p1), where F_α = Σ_b p0(b)^α p1(b)^{1−α} (Eqs. 2.18–2.20). This bound is proved; asymptotic tightness is cited to Cover & Thomas.
• Rényi overlaps. F_α (Eq. 2.38) and the Rényi relative information K_α = (α−1)⁻¹ ln F_α (Eq. 2.39).
• Statistical overlap. F(p0,p1) = Σ_b √p0(b)√p1(b) (Eq. 2.40), the Bhattacharyya coefficient. arccos F is the Fisher-metric geodesic distance on the simplex (Eqs. 2.41–2.42).
• KL from “keeping the expert honest” (Thm 2.3, Aczél). For n ≥ 3, Σ p_k F_k(q_k) ≤ Σ p_k F_k(p_k) for all distributions iff F_k(p) = α ln p + γ_k (Eqs. 2.47–2.48). The expected loss from lying is then K(p0/p1) = Σ p0 ln(p0/p1) (Eq. 2.76).
• Method of types (Thms 2.4–2.6).
  • (n+1)^{−B} e^{nH(p1)} ≤ |T(p1)| ≤ e^{nH(p1)} (Eq. 2.95).
  • (n+1)^{−B} e^{−nK(p1/p0)} ≤ P(T(p1)) ≤ e^{−nK(p1/p0)} (Eq. 2.98).
• Shannon entropy. Derived axiomatically via Daróczy’s type-α functional equation (Thm 2.7, Eqs. 2.111–2.131). Given an operational reading through the Kraft inequality and the noiseless coding bound H ≤ l_min ≤ H + 1 (Thm 2.8, Eq. 2.137).
• Mutual information. J = H(π0p0 + π1p1) − π0H(p0) − π1H(p1) = π0K(p0/p) + π1K(p1/p) (Eqs. 2.101, 2.149).

Quantum measures (Ch. 3)

• Definitions (Eqs. 3.3–3.7). Error probability, fidelity, Rényi overlap, quantum Kullback information and accessible information, each extremized over all POVMs.

• Helstrom (Eq. 3.23). Pe(ρ0|ρ1) = π0 + Σ_{γj≤0} γj, where γj are the eigenvalues of Γ = π1ρ1 − π0ρ0 (Eq. 3.12). The thesis re-derives it without variational methods and shows the optimal E0 is unique up to the zero-eigenvalue block. The optimal single-copy measurement depends on the number of copies M: for photon polarizations 0° and 45°, the optimal polarizer angles are 67.5°, 54.54° and 49.94° for M = 1, 2, 3, found numerically (§3.2.2, Fig. 3.1).

• Quantum Chernoff. Defined (Eq. 3.26) but not solved.

• Fidelity theorem (§3.3.1). min_{E} Σ_b √(tr ρ0E_b)√(tr ρ1E_b) = tr√(ρ1^{1/2}ρ0ρ1^{1/2}) (Eq. 3.28; also Eq. 1.15).

  • It is attained by projectors onto the eigenbasis of M = ρ1^{-1/2}√(ρ1^{1/2}ρ0ρ1^{1/2})ρ1^{-1/2} (Eq. 3.67), with eigenvalues m_b = (tr ρ0E_b / tr ρ1E_b)^{1/2} (Eq. 3.92).
  • M = ρ1⁻¹ # ρ0, the Kubo–Ando geometric mean (Eq. 3.106). M is the inverse of the analogous N built with 0 and 1 swapped (Eqs. 3.83, 3.89).
  • The naive Schwarz bound tr(ρ0^{1/2}ρ1^{1/2}) (Eq. 3.43) is shown to be unattainable when [ρ0,ρ1] ≠ 0 (Eqs. 3.45–3.53).
  • For pure states: F = |⟨ψ0|ψ1⟩| (Eq. 3.36), and any orthonormal basis in the span of the two states that straddles them is optimal (Eqs. 3.78–3.79).

• Fidelity properties and representations.

  • Symmetric (Eqs. 3.80–3.82) and unitarily invariant (Eq. 3.31), with 0 ≤ F ≤ 1 (Eq. 3.32).
  • F² = ⟨ψ1|ρ0|ψ1⟩ when ρ1 is pure (Eq. 3.33).
  • Bures distance: d_B² = 2 − 2F (Eq. 3.29). Uhlmann transition probability: F² (Eq. 3.30). Uhlmann purification theorem (Thm 3.1, Eq. 3.35).
  • F = √(min_{G>0} tr(ρ0G) tr(ρ1G⁻¹)) (Eq. 3.94).
  • d_B² = min ‖W0 − W1‖₂² over W_iW_i† = ρ_i (Eqs. 3.100–3.103).
  • Qubit closed form: F = (1/√2)[1 + a·b + √(1−a²)√(1−b²)]^{1/2} (Eq. 3.115), with an optimal Bloch direction n_B = a/√(1−a²) − b/√(1−b²) (Eq. 3.114).
  • Multiplicative on products: F(ρ0⊗σ0, ρ1⊗σ1) = F(ρ0,ρ1)F(σ0,σ1) (Eq. 4.58).

• By-product lower bound on quantum Kullback information. K_B(ρ0/ρ1) = 2 tr[ρ0 ln M] (Eq. 3.93).

• Rényi overlaps (§3.4). Unsolved. Three lower bounds are given:

  • an operator-Hölder/Araki chain that is not linear in E_b (Eq. 3.125);
  • an eigenvalue bound Σ_i λ_{D−i+1}(ρ0)^{1/p} λ_i(ρ1)^{1/q} (Eq. 3.133);
  • an Ando operator-mean bound, Σ_b … ≥ tr[ρ1^{1/2}(ρ1^{-1/2}ρ0ρ1^{-1/2})^α ρ1^{1/2}] (Eq. 3.137).

  The author notes that none of them reproduces the fidelity at α = ½.

Accessible information for a binary channel (§3.5)

J(t) = H(p) − (1−t)H(p0) − tH(p1), and I(t) = max over POVMs (Eqs. 3.144–3.146).

• Holevo upper bound. I(t) ≤ S(t) ≡ S(ρ) − (1−t)S(ρ0) − tS(ρ1), with equality iff [ρ0,ρ1] = 0 (Eq. 3.149; proof pp. 67–72).
  • The proof uses J(0) = J(1) = S(0) = S(1) = 0 (Eq. 3.155) and S” ≤ L” ≤ J” ≤ 0 (Eqs. 3.164–3.165), which give I ≤ L ≤ S (Eq. 3.167).
  • The J” side: J”(t) = −Σ_b (tr ΔE_b)²/tr(ρE_b) (Eq. 3.156) ≥ −tr[Δ L_ρ(Δ)] ≡ L”(t) (Eqs. 3.183, 3.187). Here L_ρ is the symmetric-logarithmic-derivative map ½(ρX + Xρ) = A (Eqs. 3.177, 3.181). Equality holds for projectors onto the eigenbasis of L_ρ(Δ) (Eq. 3.185).
  • The S” side: S”(t) = −Σ Φ(λj,λk)|Δjk|² (Eq. 3.160), with the logarithmic mean Φ(x,y) = (ln x − ln y)/(x − y) ≥ 2/(x+y) (Eq. 3.189).
• Jozsa–Robb–Wootters lower bound. I(t) ≥ Q(t) ≡ Q(ρ) − (1−t)Q(ρ0) − tQ(ρ1), with subentropy Q(ρ) = −Σ_j (Π_{k≠j} λj/(λj−λk)) λj ln λj (Eqs. 3.151–3.152). It is restated, not proved.
• New lower bound M(t). M(t) = tr[(1−t)ρ0 ln L_ρ(ρ0) + tρ1 ln L_ρ(ρ1)] (Eq. 3.203; Eq. 1.10 in bits).
  • It uses the fact that L_ρ(ρ0) = 1 − tL_ρ(Δ) and L_ρ(ρ1) = 1 + (1−t)L_ρ(Δ) commute even when ρ0 and ρ1 do not (Eqs. 3.199–3.200).
  • Qubit form (Eq. 3.267): measure along n0 = (1 − a·c)b − (1 − b·c)a (Eq. 3.266), where c = (1−t)a + tb.
  • For two pure states, n0 coincides with the optimal projective measurement (Eqs. 3.258–3.260). Levitin [110] showed that measurement is optimal among all POVMs, so M(t) = I(t) for pure states. The closed form is Eq. 3.263, with n = √(1 − 4t(1−t)q) and q = |⟨ψ0|ψ1⟩|².
• The Holevo intermediate bound L(t), made explicit. L(t) = −tr[ρ ∫^t L_{ρ(t’)}(Δ)dt’] + c1t + c2 (Eq. 3.230).
  • The integral is computed through the Kronecker sum: vec(H) = 2ρ̃0^{-1/2}(ρ̃0^{-1/2}Δ̃ρ̃0^{-1/2})⁻¹ ln[Ĩ + tρ̃0^{-1/2}Δ̃ρ̃0^{-1/2}]ρ̃0^{-1/2} vec(Δ), with ρ̃0 = ρ0 ⊕ ρ0* and Δ̃ = Δ ⊕ Δ* (Eqs. 3.239–3.244).
  • Qubit closed form (Eq. 3.269): L(t) = (δ/2d²)[−(δ − c·d)ln(δ − c·d) − (δ + c·d)ln(δ + c·d) + β1t + β2], with δ = √((1 − a·b)² − (1−a²)(1−b²)) (Eqs. 3.270–3.272).
  • A doubly infinite Fourier sine series (Eqs. 3.213–3.224) is also worked out and set aside.
• Other upper bounds.
  • N(t). From N”(t) = −tr[L_{ρ^{1/2}}(Δ)²] = −¼ tr[(d√ρ/dt)²] (Eqs. 3.275, 3.279), with I ≤ L ≤ N ≤ S by Φ(x,y) ≥ (2/(√x+√y))² (Eq. 3.281). No closed form: the qubit N” is Eq. 3.290.
  • R(t) (Jensen). R(t) = (1−t) ln tr[ρ0L_ρ(ρ0)] + t ln tr[ρ1L_ρ(ρ1)] (Eq. 3.296); qubit form Eqs. 3.300–3.301. It rests on the per-term bracket −2 ln tr√(ρ^{1/2}ρ_iρ^{1/2}) ≤ K(ρ_i/ρ) ≤ ln tr[ρ_iL_ρ(ρ_i)] (Eq. 3.295).
  • P(t) (purifications). I(ρ0|ρ1) ≤ I(|ψ̃0⟩|ψ̃1⟩) (Eq. 3.306). The best such bound takes Eq. 3.263 with q = F(ρ0,ρ1)² (Eq. 3.308).
• Holevo as an average of Umegaki divergences. I(t) ≤ (1−t)K_U(ρ0/ρ) + tK_U(ρ1/ρ) (Eq. 3.210).

Quantum Kullback information (§3.6)

K(ρ0/ρ1) = max over POVMs of Σ tr(ρ0E_b) ln[tr(ρ0E_b)/tr(ρ1E_b)] (Eq. 3.309).

• Umegaki upper bound. K ≤ K_U = tr(ρ0 ln ρ0 − ρ0 ln ρ1) (Eqs. 3.313, 3.324), obtained from the Holevo bound in the limit t → 1 (Eqs. 3.314–3.323). The thesis claims equality iff the states commute.
• Two lower bounds of the form tr ρ0 ln Λ.
  • K_F = tr[ρ0 ln L_{ρ1}(ρ0)] (Eqs. 3.204, 3.326), where Xρ1 + ρ1X = 2ρ0.
  • K_B = 2 tr[ρ0 ln M] (Eqs. 3.93, 3.328), where Xρ1X = ρ0 (Eq. 3.334).
• A generating scheme for more lower bounds. It is built from operator equations whose eigenvectors satisfy ⟨b|ρ0|b⟩/⟨b|ρ1|b⟩ = λ_b(Λ) (Eqs. 3.336–3.337):
  • a one-parameter family max_α tr[ρ0 ln((1−α)X_α² + αX_α)], with ½α(ρ1X + Xρ1) + (1−α)Xρ1X = ρ0 (Eqs. 3.342–3.343);
  • a general polynomial family (Eqs. 3.344–3.345).
• Ando-based upper bound. K ≤ tr[ρ1^{1/2}ρ0ρ1^{-1/2} ln(ρ1^{-1/2}ρ0ρ1^{-1/2})] (Eq. 3.348).
• Bounds quoted from Hiai–Petz [155]. K_U ≤ (1/p) tr[ρ0 ln(ρ0^{p/2}ρ1^{-p}ρ0^{p/2})] for every p > 0 (Eq. 3.349), and K ≥ tr[ρ0 ln(ρ1^{-1/2}ρ0ρ1^{-1/2})] (Eq. 3.350).

Ch. 4

• Inference–disturbance formalism. Replace the measurement by its Kraus operators {A_bi} with E_b = Σ_i A_bi†A_bi (Thm 4.1, Eqs. 4.17–4.20). Optimize a figure of merit over {A_bi} (Eq. 4.24).
  • A trial relation: P_s + C ≤ max over {A_bi} (Eq. 4.31), where P_s is the equal-prior success probability (Eq. 4.26) and C = ½F(ρ0,ρ0’)² + ½F(ρ1,ρ1’)² is the “clonability” (Eq. 4.29). A trivial ceiling is given in Eq. 4.33. The general case is left unsolved (p. 111: “Presently we have no general solutions”).
  • Worked example: pure states at angle θ, Eve’s POVM fixed to the Helstrom basis, re-preparation restricted to a plane. The least-disturbing re-prepared states are not the guessed states: they are rotated by φo (Eq. 4.45), which peaks at 6.99° at θ = 27.73°.
• No-broadcasting (§4.3).
  • {ρ0, ρ1} can be broadcast (Eq. 4.49) iff [ρ0,ρ1] = 0 (Eq. 4.101). They can be cloned (Eq. 4.50) iff F ∈ {0, 1}, i.e. the states are identical or orthogonal (Eqs. 4.52, 4.81).
  • Proof: fidelity cannot decrease under partial trace (Eq. 4.74) and cannot decrease under E (Eq. 4.79). So broadcasting forces F_A = F(ρ̃0,ρ̃1) = F_B (Eq. 4.80). The equality conditions then force G = H = M (Eqs. 4.89–4.97), and hence [ρ0, M] = 0 (Eqs. 4.98–4.103).
  • Of the candidate broadcast states, the perfectly entangled |ψ̃_s⟩ = Σ_b √λ_sb |b⟩|b⟩ cannot be produced by any physical E when S(ρ0) ≠ S(ρ1) (Eqs. 4.105–4.109).

Claims

id	claim	strength	support
C1	The POVM-minimized Bhattacharyya overlap of ρ0, ρ1 equals tr√(ρ1^{1/2}ρ0ρ1^{1/2}), attained by the eigenbasis of M (Eq. 3.67)	strong	proof, §3.3.1, Eqs. 3.54–3.74, invertible and support cases
C2	The optimal fidelity measurement’s eigenvalues are √(likelihood ratios) (Eq. 3.92); M = ρ1⁻¹ # ρ0 (Eq. 3.106)	strong	proof, §3.3.2
C3	Qubit fidelity is (1/√2)[1 + a·b + √(1−a²)√(1−b²)]^{1/2}	moderate	variational calculation “after quite some algebra” (Eq. 3.115), steps omitted; the formula agrees with the standard result
C4	Helstrom’s optimal error probability is π0 + Σ_{γ≤0} γ, and the optimal E0 is unique up to the kernel of Γ	strong	proof, §3.2.1
C5	Holevo bound I(t) ≤ S(t) for binary ensembles, with equality iff [ρ0,ρ1] = 0	strong	proof, §3.5.1 (finite dimension, two states)
C6	Minimizing J”(t) over POVMs yields exactly Holevo’s (implicit) L”(t), attained by the eigenbasis of L_ρ(Δ)	strong for “attained”; assertion for “exactly Holevo’s”	proof, Eqs. 3.169–3.187; the identity with Holevo’s function is stated (p. 71), not shown
C7	M(t) (Eq. 3.203) is a lower bound on I(t), equal to I(t) for two pure states	strong	M(t) is J(t) at a specific measurement (proof); pure-state equality via Eqs. 3.258–3.266 plus Levitin [110] (cited)
C8	M(t) is “often very close to being optimal” for qubits (Eq. 1.10 remark; §3.5.2)	weak	numerical plots only (Figs. 3.2–3.8); Fig. 3.2’s “I(t)” is projective-only, and no exact mixed-state I(t) is shown
C9	I ≤ L ≤ N ≤ S; R(t) and P(t) are upper bounds	strong	proofs, Eqs. 3.167, 3.281–3.283, 3.294–3.296, 3.306–3.308
C10	Holevo and Jozsa–Robb–Wootters are the best bounds expressible in ρ alone for pure-state ensembles (§1.2, p. 4)	assertion	attributed to [6]; not argued in the thesis
C11	Umegaki relative entropy upper-bounds the quantum Kullback information	strong	proof from the Holevo bound, Eqs. 3.314–3.324 (invertible states)
C12	…with equality iff the states commute	weak	one sentence (p. 97) reasoning from Holevo’s equality condition. A strict inequality at every t < 1 need not stay strict in the limit t → 1, so this does not follow as written. The “if” direction is trivial
C13	The operator-equation scheme (Eqs. 3.338–3.345) generates “successively tighter” lower bounds on the quantum Kullback information (abstract)	weak	Eq. 3.343 is ≥ K_F and K_B by interpolation (sound, if Hermitian solutions X_α exist). No higher-order bound is computed or shown to be strictly tighter, and existence of solutions is delegated to the Riccati literature
C14	Two states can be broadcast iff they commute; cloned iff identical or orthogonal	strong for invertible ρ0, ρ1; unproved beyond	proof, §4.3 (invertibility assumed, p. 118); “if” directions by construction (Eqs. 4.53–4.55)
C15	No-broadcasting holds against eavesdroppers acting jointly on whole strings of systems (p. 121)	moderate	the theorem is dimension-free, so it applies to block systems; this is an informal extension
C16	Ch. 2: KL is the unique expected-loss measure from an honest-expert (proper, local) payoff when n ≥ 3	strong	proof of Aczél’s theorem, Thm 2.3 (monotonicity plus Lebesgue differentiability)
C17	Optimal Rényi-overlap POVMs “may well” be projective, with min_A (tr ρ0A^p)^{1/p}(tr ρ1A^{−q})^{1/q} equal to the Rényi overlap	conjecture	stated as a possibility (Eq. 3.131); only a loose lower bound (Eq. 3.133) is proved

Method

The recurring technique, used four times (fidelity §3.3.1; J” §3.5.1; fidelity via G, Eqs. 3.94–3.99; no-broadcasting §4.3), has four steps:

1. Write each outcome’s contribution as an operator inner product in which E_b^{1/2} sits on both sides. Insert a free unitary U; the trace is invariant under it.
2. Apply the Schwarz inequality |tr A†B|² ≤ tr A†A · tr B†B, so that the bound is linear in E_b.
3. Sum over b using Σ E_b = 1.
4. Choose U by polar decomposition to maximize |tr(Uρ0^{1/2}ρ1^{1/2})| (Eq. 3.60: max_U |tr UA| = tr√(A†A)). With this U the Schwarz equality condition becomes an eigen-equation for a Hermitian positive operator: M in the fidelity case, L_ρ(Δ) in the J” case. Its eigenbasis is then an attaining projective measurement.

For the accessible information this is combined with Holevo’s convexity strategy: bound the second derivatives in t, then integrate with zero boundary values at t = 0 and t = 1. L(t) is then integrated using an integration-by-parts identity for the Lyapunov solution (Eqs. 3.225–3.230) and vec/Kronecker-sum linearization (Eqs. 3.232–3.244). The qubit cases (§3.3.3, §3.5.4) instead vary over Bloch directions.

Concepts

• Quantum distinguishability — the value of a classical distinguishability measure at the optimal POVM, with the number of outcomes free (Eqs. 3.3–3.7). There is a different quantity, and in general a different optimal measurement, for each measure.
• Statistical overlap / fidelity F — Σ√p0√p1 classically. Quantum fidelity is its POVM minimum, unsquared (tr√(·)). Jozsa’s “fidelity” [9] is its square (p. 49, p. 115).
• Rényi overlap of order α — F_α = Σ p0^α p1^{1−α}, 0 < α < 1 (Eq. 2.38). Its minimum over α gives the Chernoff exponent λ.
• Kullback–Leibler relative information K(p0/p1) — the expected loss of an expert who reports p1 when he believes p0, under the unique (n ≥ 3) honest-making payoff α ln p + γ_k.
• Quantum Kullback information — K maximized over POVMs (Eq. 3.309). It is distinct from the Umegaki relative entropy K_U, which upper-bounds it.
• Accessible information I(t) — the Shannon mutual information maximized over POVMs for the binary ensemble {(1−t, ρ0), (t, ρ1)} (Eq. 3.146).
• Lowering super-operator G_C — any map with |tr AB| ≤ |tr(C B G_C(A))| (Eq. 3.169). L_ρ, the symmetric logarithmic derivative ½{ρ, X} = A, is the one that attains the bound.
• Subentropy Q(ρ) — Eq. 3.152, the quantity in the Jozsa–Robb–Wootters lower bound.
• Broadcasting — a process whose output ρ̃_s has both marginals equal to ρ_s (Eq. 4.49). Cloning is the product special case ρ_s ⊗ ρ_s (Eq. 4.50).
• Clonability C — the average squared fidelity between input and output states (Eq. 4.29), used as the (non)disturbance measure.
• Operation — a set of Kraus operators {A_bi}, in Kraus’s “effects and operations” terminology.

Connections

The thesis builds directly on the following:

• Wootters’ statistical distance [8] (the pure-state optimal overlap);
• Uhlmann’s transition probability [7] and Jozsa’s mixed-state fidelity [9];
• Helstrom [27, 76] for the error probability;
• Holevo 1973 [5] and Jozsa–Robb–Wootters 1994 [6] for the accessible-information bounds;
• Davies [4] for the number of outcomes an optimal POVM needs (D ≤ N ≤ D²);
• Levitin [109, 110] for the exactly solvable qubit cases;
• Braunstein–Caves [122] for the observation that minimizing J” is formally the quantum Fisher information problem, which is why the symmetric logarithmic derivative appears;
• Wootters–Zurek and Dieks [2, 3] for no-cloning.

The thesis itself collects work published or submitted separately: Fuchs–Caves 1994 PRL [147] and proceedings [148] (the ensemble-dependent bounds), Fuchs–Caves OSID [149] (the fidelity techniques), Fuchs–Peres [111] (information gain versus disturbance), and the Barnum–Caves–Fuchs–Jozsa–Schumacher no-broadcasting manuscript.

Relation to LIT-007 (Spekkens, Quasi-quantization). The thesis says broadcasting “makes a communication theoretic cut between commuting and noncommuting density operators, and thus between classical and quantum state descriptions” (p. 114). LIT-007’s Table II lists no-broadcasting among the phenomena that epistricted classical statistical theories reproduce. Both hold, because they answer different questions:

• The thesis proves a fact about the quantum formalism: within quantum theory, broadcastability coincides with commutativity.
• LIT-007 disputes the inference that no-broadcasting marks nonclassicality. In a theory of classical distributions subject to an epistemic restriction on the allowed operations, it holds as well.

The thesis’s own gloss, that a broadcaster would have to “read the mind” of whoever assigned the subjective probabilities (p. 120), is a state-of-knowledge reading consistent with the epistemic view LIT-007 develops. A reader should not take the p. 114 sentence as evidence against LIT-007: the thesis never considers restricted classical theories. LIT-007’s table asserts rather than derives that entry (per its status note), so the two cannot be compared at the level of proof.

Bearing on the record

The record has no THEORY documents yet, so there is nothing for this reading to support or contradict. It could source one theory-level claim, stated with care: “within quantum theory, broadcastability of a pair of states is equivalent to their commutativity (C14, proved for invertible states).” Any such document would need to sit beside LIT-007’s point that no-broadcasting does not by itself certify nonclassicality.

For ML practice (the Anthology of the SOTA) the thesis carries no instruction, and none should be drawn from it. Two pieces of Ch. 2 are reusable background for anyone writing theory in that anthology, but both are classical results that the thesis reproduces rather than originates:

• Aczél’s theorem (Thm 2.3). The log score is the unique proper local payoff for n ≥ 3 outcomes, which makes KL the expected regret of misreporting a distribution. This is the classical justification for log-loss/cross-entropy as a scoring rule. The thesis does not make that connection, and it notes that uniqueness fails for n = 2 (Eq. 2.75).
• Chernoff and method-of-types bounds (Thms 2.1, 2.6). Relevant to the classical Bhattacharyya/Rényi/KL divergences.

The original sources, or Cover & Thomas, are the better citations for either. No anthology code should cite this thesis for them.

Limitations

• Scope is binary. Every accessible-information result is for two states. The thesis gives no multi-state bounds, even though its motivating example of a POVM that beats projective measurement uses three states (§1.3.3).
• Invertibility is assumed in places where the general case matters. The no-broadcasting proof assumes invertible ρ0 and ρ1 (p. 118), so as written it does not cover pure states. Pure states are covered separately by no-cloning, but mixed non-full-rank pairs are covered by neither. M(t) and several upper bounds also assume invertibility, and §1.2’s statement under a “common null subspace” condition is not proved in §3.5.2.
• Tightness of the new bounds is shown only by pictures. M(t) = I(t) is established only for pure states. For mixed states the figures compare M against a numerically optimized projective I(t) (Fig. 3.2) or against no exact value at all (Figs. 3.3–3.8). “Often very close to optimal” is therefore unquantified.
• The Ando upper bound is never tighter than the Umegaki bound. This is my identification, not the thesis’s. The Ando-based upper bound (Eq. 3.348) is the Belavkin–Staszewski form tr[σ^{1/2}ρσ^{-1/2} ln(σ^{-1/2}ρσ^{-1/2})]. With A = ρ0^{1/2}ρ1^{-1/2} one checks that it equals tr[ρ0 ln(ρ0^{1/2}ρ1^{-1}ρ0^{1/2})], which is the p = 1 member of the Hiai–Petz family the thesis itself quotes as an upper bound on K_U (Eq. 3.349). So Eq. 3.348 ≥ K_U ≥ K always. The thesis presents it as “another way to get an upper bound” without noting that it is dominated by Eq. 3.324.
• The “successively tighter lower bounds” of the abstract are a programme, not a result (C13).
• The equality clause of the Umegaki bound is argued too quickly (C12).
• Typos a reader will trip on:
  • Thm 2.2 writes λ = K(p_{α*}/p0). Eq. 2.32 shows the correct statement is ln λ = −K(p_{α*}/p0) = −K(p_{α*}/p1).
  • Eq. 2.28’s first line has p0 in both arguments of the min.
  • Eq. 3.24 flips the sign convention of Γ relative to Eq. 3.12, which is harmless.
  • Eq. 3.276 sets the boundary values of N(t) to 1; they must be 0 for N to bound I.
  • Eq. 2.131 writes α → 0 for α → 1.
  • §1.2 uses log₂ where Ch. 3 uses nats.
• Several items are open or stubs: the quantum Chernoff bound (Eq. 3.26), the Rényi overlaps (§3.4), a closed form for L(t) in general dimension, a geometric meaning for U_c (p. 55), and a general inference–disturbance tradeoff (Eq. 4.31). The promised Postscript is missing.

Open questions

Each open question below comes with the result that would close it.

• Quantum Chernoff exponent (Eq. 3.26). Closed by an explicit formula for min_α min_E Σ(tr ρ0E)^α(tr ρ1E)^{1−α} for single-copy measurements, or a demonstration that collective measurements change it.
• Quantum Rényi overlap F_α (Eq. 3.116). Closed by an operator-Hölder inequality that is linear in E_b and has an attainable equality condition, together with a proof or refutation that the optimal POVM is projective (Eq. 3.131).
• General L(t). Closed by finding a basis that diagonalizes ρ̃0^{-1/2}Δ̃ρ̃0^{-1/2} (Eq. 3.244), which would make L(t) closed-form in any dimension. N(t) would also follow from integrating −¼ tr[(d√ρ/dt)²].
• Mixed-state tightness of M(t). Closed by comparing M(t) against the true POVM accessible information for mixed qubit pairs, for instance numerically using Davies’ bound of at most D² outcomes.
• Quantum Kullback information. Closed by showing whether the α-family of Eq. 3.343, or its higher-order generalizations, converges to K(ρ0/ρ1), and by determining whether K = K_U holds only for commuting pairs.
• No-broadcasting without invertibility. Closed by a proof covering non-full-rank ρ0 and ρ1 with overlapping supports.
• Inference–disturbance. Closed by an explicit maximum of P_s + C over all operations (Eq. 4.31), even for two pure qubit states, removing the §4.2.3 restrictions on Eve’s POVM and re-preparation plane.

Corrections to the seeded skim

• The thesis has no trace-distance results and no fidelity–trace-distance inequalities. The dossier calls it a “foundational source for the Fuchs–van de Graaf-era toolkit (fidelity, trace-distance relations)”, but the text never mentions the trace distance and states none of its inequalities. The only trace-norm-like quantity is Helstrom’s error probability, Pe = π0 + Σ_{γj≤0} γj for Γ = π1ρ1 − π0ρ0 (Eq. 3.23). It is written as a sum of negative eigenvalues and never as ½(1 − ‖π1ρ1 − π0ρ0‖₁). The inequalities 1 − F ≤ D ≤ √(1 − F²) come later (Fuchs & van de Graaf, 1999). This thesis is the source for the fidelity half only.

• Ch. 5 is not annotated. It is an unannotated list of 528 references under three headings: §5.1 quantum problem, items 1–212; §5.2 classical distinguishability/information theory, 213–363; §5.3 matrix inequalities and operator techniques, 364–528. None of them is cited in the text. The works actually cited form a separate 165-entry bibliography. §1.1 (p. 3) miscounts Ch. 5 as “527 books and articles”; the list runs to 528.

• Much of Ch. 3 had already been published. It is not primary to the thesis:

  • the fidelity-as-optimal-overlap result is cited in §4.2 (p. 107) and §4.3 (p. 115) as Fuchs & Caves [149] (Open Systems & Information Dynamics, “to be published”);
  • the L(t)/M(t) ensemble-dependent bounds are Fuchs & Caves, PRL 73, 3047 (1994) [147], plus the proceedings paper [148];
  • the no-broadcasting section is “based largely on a manuscript submitted to Physical Review Letters”, with Barnum, Caves, Jozsa and Schumacher (p. 113, fn 3);
  • the §4.2 example is a variant of Fuchs & Peres [111].

  The thesis’s own claim to novelty for the fidelity (p. 49) is narrower than the dossier implies. It concedes that the projective-measurement version could already be assembled from Wootters [8], Alberti [82–84] and Uhlmann [7]. What it claims as new is the statistical interpretation, the simpler and more general derivation over all POVMs, and the identification of the optimal measurement.

• Figures and bounds. The comparison figures are Figs. 3.2–3.8, not 3.2–3.7. They also plot the purification bound P(t) and the Jensen bound R(t), which the dossier does not mention.

  • The curve labelled I(t) in Fig. 3.2 is the optimum over orthogonal projective measurements, found numerically (caption, p. 83). It is not the POVM accessible information.
  • Figs. 3.3–3.8 plot no exact I(t) at all, except in the pure-state cases, where M(t) = P(t) = I(t).

• The “Postscript” never appears. §1.3 (p. 8) promises one; the text has none.

• The dossier’s other points check out:

  • the unsquared fidelity convention (explicit at p. 49, where Jozsa’s fidelity is noted to be the square);
  • Eq. 3.36 as the pure-state reduction;
  • the no-broadcasting collaboration;
  • year 1995 / arXiv January 1996.