Quantum-Classical Coexistence Network Tomography

Xuchuang Wang, Joseph C. Chapman, Aneesh Ramaswamy, Matheus Guedes de Andrade, Yu-Zhen Janice Chen, Joseph M. Lukens, Gayane Vardoyan, Don TowsleyarXiv 2026 · arXiv

TL;DR

This paper introduces a quantum network tomography framework designed specifically for quantum-classical coexistence networks (QCNs), where quantum and classical signals share the same fiber infrastructure. By combining a novel direction-dependent noise model with an active traffic-switching protocol, the framework accurately estimates internal link parameters—such as photon loss and Raman-induced depolarization—using only end-to-end measurements.

What Problem Does This Paper Address?

Building dedicated fiber networks exclusively for quantum communication is prohibitively expensive. A practical alternative is the Quantum-Classical Coexistence Network (QCN), which uses wavelength-division multiplexing (WDM) to transmit quantum and classical signals simultaneously down the same optical fiber.

However, high-power classical traffic leaks noise into the quantum channel, primarily through spontaneous Raman scattering (SpRS). Crucially, this noise is direction-dependent: counter-propagating classical traffic (moving against the quantum signal) generates significantly more noise at the receiver than co-propagating traffic. As networks scale, directly measuring the noise of every single internal fiber link becomes operationally impossible. While classical network tomography infers internal link states from edge-to-edge measurements, existing quantum network tomography methods assume static, direction-independent noise, making them inadequate for the unique physics of QCNs.

Methods

The authors address this gap by developing a physical channel model for a coexisting fiber, decomposing the quantum signal evolution into photon loss ($q$), successful transmission ($s$), and direction-dependent depolarization ($d^{(x)}$) across three scenarios: no classical signal, co-propagation, and counter-propagation.

Fig. 1: Model of quantum-classical coexistence fiber. The upper block represents the quantum channel with three possible outcomes: photon loss (ratio q), successful transmission (ratio s), and depolarization (ratio d(x)). The three rows (separated with dashed lines) in the upper block correspond to no classical signal (×, depolarization d(0)), co-propagation (⇒, depolarization d(1)), and counter-propagation (⇐, depolarization d(2)), respectively. The lower block represents the classical channel.

Key methodological innovations include:

  1. Closed-Form Link Tomography (QLT): Instead of computationally heavy iterative optimization or sampling, the authors derive exact, closed-form algebraic estimators to calculate link parameters directly from aggregate photon counts ($T, R, M$).
  2. CPTP Quantum Channel Formulation: They rigorously justify their model by recasting the coexisting fiber as a completely-positive, trace-preserving (CPTP) tensor product of a depolarizing-with-loss signal channel and a Raman-noise-injection channel acting on separate optical modes.
  3. Active Gauge Fixing via Direction Switching: In a star network, static end-to-end probing leaves a mathematical "gauge degeneracy"—an inability to uniquely separate the co- vs. counter-propagating noise ratios. The authors introduce a clever protocol: flipping the classical signal direction on just a single link breaks this symmetry and unlocks full parameter identifiability.
Fig. 3: Link and network tomography illustrations. Blue thick arrows represent classical signals, and other thinner arrows represent quantum signals. In Figure 3a, the quantum and clas- sical signals share a single fiber link via WDM. In Figure 3b, a star-topology network carries two or more classical signals on different wavelengths (from the bottom left node) into the center node, where wavelength switching routes each classical signal to a separate output fiber.
  1. Topology Generalization: They extend the framework to tree topologies using a recursive "peeling" algorithm and to general mesh networks using a log-linear least-squares estimator.

Datasets

The framework was validated using a mix of experimental, emulated, and simulated data:

  • Experimental Testbed Data: Single-link validation utilized a real-world dataset from Chapman et al. [32], featuring polarization-encoded single photons sent over 0.5 km, 5.0 km, and 15 km fibers across 91 different C-band wavelengths.
  • Emulated Star Networks: Because multi-link QCN testbeds do not yet exist physically, the authors evaluated the star-network estimators on emulated multi-link paths constructed by mathematically recombining the single-link experimental data.
  • Synthetic Simulations: Fully synthetic Monte-Carlo simulations were used to validate the peeling and least-squares algorithms on complex tree and cyclic-mesh topologies.

Key Insights and Ideas

  • Multiplicative, Not Additive Noise: In standard quantum network models, noise probabilities typically add up. However, because Raman noise injection inflates the photon population, the received photon ratios compose multiplicatively along a path. This requires a conceptual shift to log-linear tomographic solvers.
  • The Asymmetry of Traffic Direction: The data confirms that counter-propagating classical traffic inflates the depolarization probability by a factor of 1.8× to 3.2× compared to co-propagating traffic. This insight directly informs routing strategies: co-propagating quantum and classical channels yields measurably higher fidelity.
  • Active vs. Passive Tomography: Passive edge probing is fundamentally insufficient for QCNs due to gauge degeneracy. Actively reconfiguring the classical traffic direction on a single link is the secret key to unlocking full network visibility.

Key Results

  • High Accuracy at Low Computational Cost: On single-link experimental data, the closed-form QLT estimator closely tracks the accuracy of full Bayesian Process Tomography (BPT) but replaces heavy sampling algorithms with instantaneous $O(1)$ arithmetic.
  • Super-Exponential Sample Complexity: The paper demonstrates that the number of probes required scales super-exponentially with network depth because the success signal shrinks geometrically. For example, achieving a 5% relative error at a network depth of 5 requires a massive sample size of $\sim 2.5 \times 10^{11}$ probes. This highlights a critical practical bottleneck for deep network tomography.

Limitations and Future Work

  • Depolarization-Only Approximation: The model assumes isotropic depolarization. While highly accurate, a small residual fidelity gap remains when compared to experimental data because the model ignores non-unital or coherent noise components.
  • The Deep Network Bottleneck: Due to super-exponential sample complexity, purely end-to-end tomography is impractical for very deep networks. The authors suggest operators confine end-to-end tomography to shallow paths (depth $\le$ 2–3) and characterize links individually where direct access is available.
  • Simplified Noise Sources: The current framework focuses primarily on Raman scattering and does not yet account for other non-linear fiber effects like four-wave mixing or direct classical-to-quantum crosstalk.
  • Physical Multi-Link Validation: A physical experiment splicing multiple coexisting links together remains an important next step to fully validate the composition laws outside of emulation.

Summary updated

Abstract

Quantum-classical coexistence networks (QCNs) share optical fiber between quantum and classical signals via wavelength-division multiplexing, offering a practical path to quantum communication over existing telecom infrastructure. However, co- and counter-propagating classical traffic introduce distinct depolarization noise, complicating channel characterization. We develop a tomography framework that infers per-link channel parameters of a QCN from end-to-end measurements alone. We first model each coexisting fiber by decomposing the signal evolution into photon loss, successful transmission, and three direction-dependent depolarization components. We then derive closed-form link-level estimators, and extend the approach to star-topology networks through a system of multiplicative equations across end-node pairs, together with a simple classical-signal-direction-switching protocol that resolves the remaining unknowns. On single-link experimental testbed data, we recover per-link depolarization probabilities accurately, with estimated process fidelities closely tracking the Bayesian-process-tomography baseline across multiple fiber lengths and wavelengths; residual gaps reflect the depolarization-only approximation. Absent a multi-link coexistence testbed, we validate the star-network estimators on emulated paths built from measured single-link channels. We further extend the framework in two directions: (i) a channel model that factorizes the coexisting fiber into a depolarizing-with-loss signal channel and a Raman-noise-injection channel on separate optical modes -- a completely-positive, trace-preserving tensor product -- whose link observables reduce exactly to our basic model; and (ii) a generalization to arbitrary topologies via a peeling algorithm (trees) and a least-squares estimator (meshes), validated by Monte-Carlo simulations on tree and cyclic-mesh networks.
BibTeX
@misc{wang2026quantum,
  author = {Xuchuang Wang and Joseph C. Chapman and Aneesh Ramaswamy and Matheus Guedes de Andrade and Yu-Zhen Janice Chen and Joseph M. Lukens and Gayane Vardoyan and Don Towsley},
  title = {{Quantum-Classical Coexistence Network Tomography}},
  year = {2026},
  eprint = {2608.09364},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.09364},
}
Figure