Geodesic Quantum $f$-Divergences

AI summary

Overview

Research area: Quantum information theory — specifically quantum divergence measures, data-processing inequalities, and quantum Markov chains.

Technical level: Advanced. The paper is written for researchers already familiar with quantum relative entropy, Petz recovery maps, operator convexity, and modular operator theory.

Scope (one sentence): The paper constructs a one-parameter family of quantum f-divergences by following the affine-invariant geodesic between the standard (Petz) and maximal relative modular operators, and proves data processing, characterizes its equality cases, and derives the corresponding collapse of quantum Markov chain classes.

What This Paper Is About

Extending a classical divergence such as the Kullback–Leibler divergence to noncommuting quantum states is ambiguous, because operator ordering is not fixed by the requirement of matching the classical formula on commuting states. Two distinguished but generally different endpoints are known: the Umegaki/standard (Petz) construction and the Belavkin–Staszewski (BS)/maximal construction. The paper builds a continuous path between these two endpoints using the affine-invariant geodesic on relative modular operators, then asks whether the intermediate points are still genuine quantum divergences — that is, whether they satisfy data processing under quantum channels.

Key Contributions

1. A new interpolating family. The authors define the geodesic quantum f-divergences, also called (t,f)-divergences, D_f^t for 0 ≤ t ≤ 1, built from the geodesic Γ_t = Δ_{ρ|σ} #t \hat{Δ}{ρ|σ} = L_{ρ^{1−t}} R_{σ^{−1/2} ρ^t σ^{−1/2}} between the standard and maximal relative modular operators. Choosing the logarithmic generator f(x) = x log x gives t-relative entropies D^t joining Umegaki (t = 0) to BS (t = 1); choosing power generators f(x) = x^α gives (t,α)-Rényi divergences D_α^t joining the Petz (t = 0) and geometric (t = 1) Rényi families.

2. Data processing across the whole path. The first main theorem establishes data processing for D_f^t for every finite operator-convex generator f and every t ∈ [0,1], which in particular covers the Rényi orders 0 < α < 1 and 1 < α ≤ 2.

3. Exact equality characterization with an endpoint dichotomy. For invertible states and equality-determining operator-convex generators, equality in the data-processing inequality at any nonmaximal 0 ≤ t < 1 holds exactly when ρ is recovered by the Petz map; at t = 1 the equality class jumps to the generally larger BS class. The quadratic generator (α = 2) is an exception, carrying the BS equality class for every t.

4. Collapse of quantum Markov chains and a monotonicity theory. The three ordered t-conditional mutual informations (one-sided, two-sided, reverse) vanish exactly on invertible quantum Markov chains for 0 ≤ t < 1, and on invertible BS quantum Markov chains at t = 1. The paper further determines monotonicity regimes in t, characterizes a universal monotonicity cone for C¹ generators, and proves strengthened data-processing and reconstruction estimates together with continuity and discontinuity results.

Main Findings

• Classical reduction and endpoints: D_f^0 equals the standard (Petz) f-divergence, D_f^1 equals the maximal f-divergence, and every t gives the classical f-divergence when [ρ, σ] = 0. The Rényi family satisfies D_α^t → D^t as α → 1.

• Usable closed forms: The t-relative entropy is D^t(ρ‖σ) = (1−t) Tr(ρ log ρ) + Tr[ρ log(ρ^{t/2} σ^{−1} ρ^{t/2})], and the Rényi version is D_α^t(ρ‖σ) = (α−1)^{−1} log Q_{α,t} with Q_{α,t} = Tr[σ^{1/2} ρ^{α(1−t)} σ^{1/2} (σ^{−1/2} ρ^t σ^{−1/2})^α]. Both admit a scalar-moment representation D_f^t = ∫ f(λ) dν_t(λ) over a probability measure ν_t from the spectral resolution of Γ_t.

• Monotonicity along the geodesic: D(ρ‖σ) = D^0 ≤ D^t ≤ D^1 = \hat{D}(ρ‖σ), with every inequality between distinct interpolation times strict exactly when ρ and σ do not commute. For Rényi divergences, t ↦ D_α^t is non-decreasing for 0 < α < 2, constant at α = 2, and non-increasing for 2 < α ≤ 3; for α > 3 there is no universal direction in dimension at least three, but an exact pair-of-times criterion is given.

• Monotonicity in the Rényi order: For every fixed t, α ↦ D_α^t(ρ‖σ) is non-decreasing, and strictly so unless ρ = σ.

• The geodesic path is constant at order two: At α = 2 the entire interpolating path coincides, with D̄_2(ρ‖σ) = \hat{D}_2(ρ‖σ) = log Tr(ρ² σ^{−1}), even for noncommuting states.

• Basic structural properties: Normalization D_f^t(ρ‖ρ) = f(1); nonnegativity and state separation when f(1) = 0 and f is scalar convex on an interval containing {1} ∪ spec(Γ_t) (strictly convex gives ρ = σ iff the divergence vanishes); unitary invariance; exact flag rules; continuity; tensor additivity of D^t and normalized Rényi divergences; joint convexity of unnormalized (t,f)-divergences for operator-convex generators, with D_α^t jointly convex below order one and jointly quasi-convex above order one.

• Equality hierarchy: The rigidity theorem gives several equivalent formulations of the Petz equality condition — via likelihood ratios, half powers, the full modular functional calculus, Petz recovery, and preservation of Umegaki relative entropy — with the Rényi fixed-point diagram included (Theorems 7.2, 7.7, 7.9).

• Quantum Markov chain collapse: For 0 ≤ t < 1, the one-sided, two-sided and reverse t-QMC classes all equal the set of invertible quantum Markov chains; at t = 1 all three equal the invertible BS quantum Markov chains, a class that can be strictly larger, witnessed by the Werner-state construction. The paper attributes the abrupt collapse to modular rigidity: all non-KMS weights impose the same modular covariance, while the KMS weight may have a larger symmetry class.

• Conditional mutual informations are not ordered: The three t-CMIs are universally incomparable; the paper proves an additive projective comparison for the two forward orientations, a genuine multiplicative comparison at Rényi order two, and shows that the analogous BS multiplicative factor cannot exist in general.

• Strengthened data processing: The DPI deficit has a positive resolvent remainder compressible into a single intrinsic Hilbert–Schmidt recovery defect, yielding a quartic remainder for D^t, order-dependent remainders for D_α^t, and explicit tripartite lower bounds.

• Continuity and discontinuity: Continuity bounds are proved for conditional t-entropies, the three t-CMIs, and general (t,f)-divergences when the smallest eigenvalues are bounded below by fixed positive constants; for the nonoptimized quantities at any fixed t > 0, no dimension-only continuity modulus can hold uniformly as the smallest eigenvalues approach zero.

• Universal monotonicity cone: For C¹ generators with g(x) = f(x)/x, membership in the universal cone 𝔐 (non-decreasing t ↦ D_f^t for every finite-dimensional invertible pair) is characterized by an exact ordinary-matrix condition involving a nonnegativity test over all dimensions, all a > 1, and positive definite P, K with Tr P^a = Tr(P^{a−1} K^{−1}) = 1. On the finite operator-convex slice this is equivalent to a canonical-measure polar condition ∫{[0,1]} κ{ρ,σ,t}(u) dν_f(u) ≥ 0.

• Relation to the (α, z) family: The exact universal intersections of the geodesic Rényi family with the (α, z) family, even allowing real t, are determined, and their value ranges on the common DPI region are compared.

• Capacity-per-unit-cost interpretation: Using the canonical finite experiment (p_t, q_t) from the companion paper, which satisfies D^t(ρ‖σ) = D_KL(p_t‖q_t), the classical zero-cost-symbol theorem yields a capacity-per-unit-cost interpretation for every D^t, including its BS specialization.

Methodology in Plain English

The authors treat the space of operators as a Hilbert space and look at two well-known positive operators that encode a pair of states: the standard relative modular operator and its "maximal" commutant counterpart. Between two positive operators there is a natural geodesic — the affine-invariant one — and the authors take the midpoint-like path along it, parametrized by t from 0 to 1. Feeding that path through a function f via functional calculus, and taking an inner product against σ^{1/2}, produces the new divergence.

Because the construction is a convex interpolation between two operators, much of the analysis reduces to studying a single scalar probability measure ν_t supported on the spectrum of the path operator. Monotonicity in t, nonnegativity, the α-dependence of the Rényi normalization, and the comparison with existing families all become statements about moments of this measure. Data processing is then attacked by combining operator-convexity arguments with the Petz recovery map, and the equality analysis proceeds by tracking when data processing can be tight — which reduces to a modular covariance condition. The quantum Markov chain results follow by specializing data processing to partial traces in three different orderings of the three systems.

Why This Matters

Impact on research. The paper shows that a large, natural geometric interpolation between the two most important quantum relative entropies does not produce a genuinely new continuum of recovery classes: for every nonmaximal t the recoverable structure is exactly Petz's, and only at the maximal endpoint does the strictly larger BS class appear. This is a rigidity statement about how quantum conditional independence can be deformed, and it constrains how far any interpolation-based generalization of sufficiency can reach.

Real-world applications (as suggested by the paper's framing):

• Quantum channel estimation and discrimination, where relative entropies set the achievable error exponents.
• Quantum capacity and communication-cost analysis, via the capacity-per-unit-cost interpretation derived from the canonical finite experiment.
• Approximate recoverability and error correction in near-Markovian many-body states, where small CMI quantifies recoverability.
• Thermodynamics of nonequilibrium quantum systems, since the Umegaki divergence relative to a Gibbs state equals a nonequilibrium free-energy difference up to the inverse-temperature factor.

Industry relevance. Tools for bounding how much information a quantum channel destroys, and for certifying when a process is reversible, feed directly into quantum error correction, device benchmarking, and the design of quantum communication protocols. The strengthened DPI remainders give quantitative, computable error terms that are more useful in practice than bare inequality statements.

Future Directions

1. Non-invertible states. The standing convention restricts the main results to invertible states. The authors state that extensions are result-dependent and explicitly do not claim extensions for strict equality characterizations, recovery and fixed-point results, analyticity, real-t extrapolations involving negative powers, or estimates containing inverse spectral bounds — leaving these as open problems.

2. Beyond equality-determining generators. Affine or spectrally sparse, non-equality-determining generators require separate treatment, so the full equality classification for arbitrary operator-convex f is not settled.

3. The companion paper's programme. The Hessian-metric structure, the induced BKM–RLD interpolation, the intrinsic role of t as a boundary index, the comparison with the Hasegawa–Petz interpolation, the restricted hypothesis-testing consequences, and the Busemann and horospherical geometry are developed in a separate companion paper rather than here.

4. Sharpness of continuity. Since no dimension-only continuity modulus can hold as smallest eigenvalues approach zero for nonoptimized quantities at fixed t > 0, identifying which restricted settings restore such bounds remains open.

Target Audience

Advanced graduate students and researchers in quantum information theory, mathematical physics, and operator theory who work on quantum divergences, data-processing inequalities, sufficiency and recovery, or quantum Markov chains. Readers interested in the statistical and geometric (Hessian, horospherical) side will need the companion paper, as those developments are not contained here.

Authors’ abstract

We introduce the geodesic quantum $f$-divergences $D_f^t$, $0\leq t\leq1$, obtained from the affine-invariant geodesic between the standard and maximal relative modular operators. They reduce to the classical $f$-divergence for commuting states. The logarithmic generator yields geodesic relative entropies joining the Umegaki and Belavkin-Staszewski entropies, while the power generators yield $(t,α)$-Rényi divergences joining the Petz and geometric families. Our first main result is data processing of $D_f^t$ for every finite operator-convex generator $f$ and every $t\in[0,1]$. In particular, this gives DPI for the geodesic relative entropies and for the $(t,α)$-Rényi divergences when $0&lt;α&lt;1$ or $1&lt;α\leq2$. Our second main result identifies equality in the DPI: for invertible states, every equality-determining operator-convex generator has, at each nonmaximal parameter $0\leq t&lt;1$, exactly the Petz sufficiency class; at $t=1$, this changes to the generally larger maximal, or BS, class, which also coincides with the equality class of the divergences associated with the quadratic generator for every $t$. Our third main result is the corresponding collapse of invertible geodesic quantum Markov chains: for each of the three ordered conditional-mutual-information constructions, the $t$-quantum Markov chains are precisely the quantum Markov chains for $0\leq t&lt;1$, whereas at $t=1$ they are the invertible BS quantum Markov chains, a class that can be strictly larger. We also determine parameter-monotonicity regimes and the intersections with the $(α,z)$ family. We prove strengthened data-processing and reconstruction estimates; we compare the three conditional orientations; we establish continuity bounds under positive lower-eigenvalue assumptions together with complementary discontinuity results; and we give a capacity-per-unit-cost interpretation.