Conditional contraction coefficients and their applications to quantum networks

AI summary

Overview

Research area: Quantum information theory — specifically the theory of quantum divergences, data-processing inequalities, and channel contraction coefficients, with applications to quantum networks, Markov chains, and quantum machine learning.

Technical level: Advanced. The paper is written in the language of quantum channels (completely positive trace-preserving maps), generalized divergences, relative entropy, conditional mutual information, and semidefinite programming. The definitions are optimization problems over unbounded reference systems, and the results are operator-level mathematical statements.

Scope in one sentence: The paper introduces "conditional contraction coefficients" — a family of contraction coefficients that incorporate arbitrary quantum reference systems by subtracting the distinguishability already present in the reference — and develops their structural theory, their exact correspondence with conditional-mutual-information contraction, and their applications to tensorization, network composition bounds, generalized mixing times, quantum memories, and quantum machine learning.

What This Paper Is About

When a quantum channel acts on a state, it typically destroys some ability to tell that state apart from another one. A contraction coefficient is a single number that quantifies the worst-case fraction of distinguishability a given channel can preserve, and it therefore gives a strengthened version of the data-processing inequality. The standard definition ignores auxiliary ("reference") quantum systems, but naively adding them makes the quantity trivial, because two states can differ only on the reference system and the channel then destroys nothing. The one existing fix — the complete contraction coefficient — forces the two compared states to have identical reference marginals, which is restrictive and does not match many practical network scenarios. This paper's goal is to define a nontrivial contraction coefficient that allows completely arbitrary reference marginals by measuring only the conditional part of the distinguishability.

Key Contributions

1. A new family of coefficients. The authors define the conditional contraction coefficient as the supremum, over all finite reference systems and all pairs of states, of the ratio of "conditional" divergences: the divergence between the channel outputs minus the divergence between the reference marginals, divided by the same quantity for the inputs. They also define the corresponding conditional SDPI (strong data-processing inequality) constants, coefficients with product reference, complete coefficients, expansion coefficients (the infimum analogue), and relative contraction coefficients comparing two channels.

2. A unified ordering and structural theory. They prove the chain of inequalities relating all these notions, plus submultiplicativity under channel composition, lower bounds under tensor products, and convexity for jointly convex divergences. They also give an upper bound on the conditional contraction coefficient via the quantum Doeblin coefficient.

3. An exact operational meaning for the relative entropy. Their central structural result is that for the quantum relative entropy, the conditional contraction coefficient is exactly equal to the contraction coefficient of the conditional mutual information — that is, the supremum over tripartite states of I(C:B|R)/I(C:A|R) with the channel acting from A to B. Notably, C may be quantum, classical, or even binary classical. They also characterize which structural properties of a divergence are needed for such results and discuss extensions to f-divergences.

4. Applications to networks, mixing, and learning. They extend a Polyanskiy–Wu-type bound on mutual-information contraction to quantum network channels decomposed as a composition of two channels, introduce a "replacing time" as a generalization of mixing time, and derive limitations for quantum memories and quantum machine learning.

Main Findings

• Ordering of coefficients. For every non-negative divergence and every channel, the coefficients satisfy 0 ≤ η_D(N) ≤ η^p_D(N) ≤ η^c_D(N) ≤ η^Δ_D(N) ≤ 1. Restricting the conditional coefficient's optimization to states with equal marginals recovers the complete contraction coefficient, which is why η^c_D ≤ η^Δ_D holds.

• Reference systems matter only quantumly, for well-behaved divergences. If the reference system is restricted to be classical and the divergence obeys the extended direct-sum property, then the conditional (and complete) coefficients coincide with the ordinary one. This is presented as a purely quantum phenomenon for certain divergences, with details deferred to Appendix A.

• Trace distance optimizes on orthogonal states. For the trace distance, the optimization defining the conditional contraction coefficient can be restricted to orthogonal input states, giving an expression in terms of the trace distance E₁ of the channel outputs minus that of the reference marginals, normalized by 1 − E₁ of the reference marginals.

• Conditional vs. complete can genuinely differ. For the more general hockey-stick divergence E_γ, the authors give an example where the conditional contraction coefficient is strictly larger than the complete one — even in the classical setting.

• Relative entropy equals conditional mutual information contraction. The identity η^Δ_D(N) = sup over ρ_RCA of I(C:B|R)/I(C:A|R) extends the known classical correspondence between relative-entropy contraction and mutual-information contraction to the setting with quantum side information. The paper also recalls the earlier identity η_D(N) = sup over classical-quantum states of I(U:B)/I(U:A), and summarizes all relations in Fig. 1 and Corollary 3.

• Tensorization of SDPI constants. The paper gives an approximate tensorization result for relative-entropy SDPI constants, yielding exact tensorization in several special cases, including the generalized depolarizing channel with respect to its fixed point. It shows that tensorization of the mutual-information SDPI constant is equivalent to a particular relative-entropy contraction tensorization result from prior work, and proves exact tensorization for the conditional mutual information: η_CMI(N₁⊗N₂, σ₁⊗σ₂) = max{η_CMI(N₁, σ₁), η_CMI(N₂, σ₂)}.

• Network composition bound. For a network channel written as N_{A→BE} = R_{D→E} ∘ M_{A→BD}, they prove η_MI(N_{A→BE}) ≤ η_R · η_MI(M_{A→BD}) + (1 − η_R) · η_MI(M_{A→B}), generalizing a bound of Polyanskiy–Wu to the quantum setting, and derive implications including new bounds for tensor-product channels.

• Replacing time generalizes mixing time. They define t^D_rep(N, δ) as the smallest n such that a divergence-based distance between (id_R ⊗ N^n)(ρ_RA) and σ_R ⊗ ω_A, minus the reference divergence, is at most δ for all reference systems and all states. They bound it in terms of the conditional contraction coefficient and show it is asymptotically equivalent to mixing.

• Tighter Doeblin-based upper bound. For any jointly convex divergence, η^Δ_D(N) ≤ 1 − α₊(N), where α₊ is the quantum Doeblin coefficient (computable as a semidefinite program). Since η^c_D ≤ η^Δ_D, this is a tighter lower bound on 1 − α₊(N) than the previous complete-coefficient bound, and it holds generally rather than case by case.

• Composition and convexity lemmas. SDPI constants and contraction coefficients are submultiplicative under composition of channels (η(M∘N) ≤ η(M)·η(N) for each of the three flavours); tensor products give lower bounds of the form η(M⊗N) ≥ max{η(M), η(N)}; and mixtures of channels satisfy η(λN₁ + (1−λ)N₀) ≤ λη(N₁) + (1−λ)η(N₀) for jointly convex divergences.

• Expansion coefficients can be trivial. The authors note that the usefulness of expansion coefficients depends strongly on the divergence, and that even the standard expansion coefficient is often equal to zero for many divergences.

Methodology in Plain English

The authors work at the level of definitions and proofs, building a general framework rather than running experiments. The core move is a change of bookkeeping: instead of asking how much total distinguishability survives a channel, they ask how much excess distinguishability survives once you account for the part that was already sitting in the reference system. Concretely, they subtract the divergence between the reference marginals from both the numerator (output side) and denominator (input side) of the usual ratio. This "conditional divergence" was previously used under the name conditional relative entropy.

They then check which mathematical properties of a divergence are actually needed for the framework to behave — non-negativity, faithfulness, additivity on tensor products, the direct-sum property for classical registers, and joint convexity — and prove the resulting lemmas (ordering, composition, tensor products, convexity) at that level of generality. Specializing to concrete divergences (relative entropy, trace distance, hockey-stick divergences, f-divergences) produces sharper statements. For the relative entropy, the key step is a variational argument identifying the supremum over state pairs with a supremum over tripartite states of a ratio of conditional mutual informations, which the authors stress is insensitive to whether the conditioning system C is quantum, classical, or binary. Upper bounds come from an operator-algebraic tool, the quantum Doeblin coefficient, which can be computed by semidefinite programming and which measures how much of a channel's action can be written as a convex mixture with a fully depolarizing (replacing) map.

No numerical experiments, datasets, or benchmark figures are reported in the paper content; all results are analytic.

Why This Matters

Impact on research. Complete contraction coefficients force equal reference marginals, which excludes most states that arise in distributed or networked settings. The conditional coefficient removes that restriction while remaining nontrivial, and it slots into a clean hierarchy whose top element is exactly the conditional mutual information contraction. That gives researchers a single framework covering no-reference, product-reference, equal-marginal, and arbitrary-reference regimes, together with a route to SDPI tensorization results that were previously out of reach for CMI. The framework also produces a strictly tighter Doeblin-based bound and clarifies when reference systems genuinely change the answer (a quantum-only effect for well-behaved divergences).

Real-world applications (as framed by the paper):

• Quantum networks: bounding how much correlation survives when a channel is split into a broadcast part and a subsequent processing part, as in N_{A→BE} = R_{D→E} ∘ M_{A→BD}.
• Quantum memories: deriving fundamental limits on how long quantum information can be stored.
• Quantum machine learning: establishing limitations on what such models can achieve.
• Noisy intermediate-scale devices and error mitigation: the contraction-coefficient toolkit has previously been applied to quantum computing, error mitigation, and differential privacy, which the paper cites as motivation.
• Discrete-time quantum Markov chains: the "replacing time" gives a mixing-time-like convergence measure that accounts for correlations with an external reference.

Industry relevance. Contraction coefficients are the standard currency for reasoning about noise accumulation, privacy leakage, and convergence of noisy iterated processes, so a version that handles side information is directly relevant to quantum networking hardware, quantum memory design, and privacy analysis of quantum algorithms. The Doeblin bound being an SDP makes it a candidate for practical numerical certification.

Future Directions

• Which divergences admit the relative-entropy-style identity? The authors identify structural properties sufficient for the main results and discuss extensions beyond relative entropy, but the precise boundary — for instance for general f-divergences — remains to be pinned down.

• Closing the gap between conditional and complete coefficients. The hockey-stick example shows strict separation even classically, so characterizing exactly when η^Δ strictly exceeds η^c, and how large the gap can be, is open.

• Full tensorization theory for SDPI constants. The paper obtains approximate tensorization for relative entropy with exactness in special cases (including generalized depolarizing channels at their fixed point) and exact tensorization for mutual information and CMI; broader exactness criteria are left open.

• Open problems in the conclusion. Section 9 concludes the paper and briefly discusses open problems, though the truncated content does not enumerate them individually.

• Tighter computable bounds. Since contraction coefficients are generally hard to compute, improving the Doeblin-type SDP upper bound, and developing matching lower bounds, is a natural next step.

Target Audience

Quantum information theorists and mathematically inclined researchers working on channel capacities, strong data-processing inequalities, divergence contraction, and tensorization; researchers in quantum networking and distributed quantum information who need to track correlations with external reference systems; and mathematically mature graduate students or industry researchers in quantum error mitigation, quantum privacy, or quantum machine learning who want rigorous bounds on information loss under noise. Readers should be comfortable with quantum channels, the quantum relative entropy, and optimization-based definitions.

Authors’ abstract

Contraction coefficients quantify the loss of distinguishability induced by a channel and provide a strong form of the data-processing inequality. While standard contraction coefficients ignore auxiliary quantum systems, existing extensions based on complete contraction coefficients require the compared states to have identical reference marginals. In this work, we introduce conditional contraction coefficients, a novel family that incorporates arbitrary quantum reference systems by subtracting the distinguishability already present in the reference system. We develop a general framework for contraction coefficients with such quantum side information, including the corresponding strong-data-processing-inequality (SDPI) constants, expansion coefficients, and relative contraction coefficients. For the trace distance, we show that the optimization can be restricted to orthogonal input states. For the quantum relative entropy, we prove that its conditional contraction coefficient is exactly equal to the contraction coefficient of the conditional mutual information, extending the classical correspondence between relative-entropy contraction and mutual-information contraction to the setting with quantum side information. More generally, we identify structural properties of divergences required for these results and discuss extensions beyond the relative entropy. These results establish a unified framework for analyzing information contraction in quantum network settings, where quantum side information and distributed correlations are intrinsic features of the information-processing task. Applications include an extension of the Polyanskiy-Wu bounds on mutual information contraction, new perspectives on mixing times, and fundamental limits on quantum memories.