PaperlayerView exact revision on arXiv

Source-linked AI summary

Quality-of-Transmission Estimation in Quantum Networks: Compute, Simulate, or Learn?

Anis Bekri, Amar Abane, Van Sy Mai, Abdella Battou, Saddek Bensalem

arXiv:2610.23827v1quant-phcs.LGcs.NI

TL;DR

Quantum controllers need estimates of candidate-path fidelity, but quantum QoT is stochastic and difficult to measure or represent with a simple optical-style scalar. This article presents a compute/simulate/learn framework and simulation-based study, finding analytical composition sufficient in simple regimes and learned estimation useful when simulation is accurate but too slow for online decisions.

  • Problem

    Quantum control-plane decisions need timely candidate-path fidelity estimates, but probabilistic generation and memory decoherence make fidelity a stochastic target unlike a usually deterministic optical QoT margin.

    Source: Introduction · p. 1

  • Method

    The article frames estimator choice as a compute/simulate/learn decision and illustrates it with a simulator-based study of analytical composition, simulation, and learned surrogates.

    Source: Introduction · p. 2, Putting It Together · p. 4

  • Results

    In the study, analytical composition is accurate enough in a well-provisioned regime, while a learned estimator achieves simulator-grade accuracy at formula-grade speed on familiar and unseen path regimes.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    Learning is most useful when a deployable formula is biased or unfeedable and repeated simulation cannot deliver needed accuracy fast enough or at scale.

    Source: Conclusion · p. 7

    The demonstrated generalization is limited to path and link regimes, not unseen meshed topologies where path fidelity depends on the rest of the network.

    Source: Generalizing to Unseen Path Regimes · p. 6

Abstract

from arXiv · show

Future quantum networks will rely on control-plane decisions such as routing, scheduling, entanglement provisioning, purification, and memory allocation. Each decision depends on a timely and reliable estimate of the quality of the entanglement that a candidate path can deliver, with end-to-end fidelity as a central metric. Inspired by the optical-networking experience with quality-of-transmission (QoT) estimation, this article asks when a quantum controller should compute fidelity analytically, simulate it, or learn a surrogate model. The optical analogy is useful but incomplete: optical QoT estimation often predicts a scalar transmission margin, whereas quantum QoT estimation must account for probabilistic entanglement generation, random memory waiting times, decoherence, state consumption, and measurement constraints that can make fidelity a stochastic target. We propose a compute/simulate/learn framework for quantum QoT estimation and illustrate it with a simulation-based proof of concept. Our study shows that analytical composition can be sufficient in simple regimes, that simulation becomes necessary when stochastic memory effects shape the delivered fidelity distribution, and that lightweight learned surrogates can approximate simulator labels at much lower runtime cost for online control. We close with open challenges for deployment, including stochastic targets, sim-to-real transfer, data-generation cost, physics-informed learning, and generalization beyond individual paths to full network topologies.

Introduction

Quantum control-plane choices depend on timely estimates of the end-to-end fidelity a candidate path can deliver. The article adapts optical QoT estimation to quantum networks by framing when to compute, simulate, or learn, where probabilistic generation and memory decoherence make fidelity stochastic.

Introduction

  • The controller needs a prior estimate of candidate-path quality to choose paths, schedule link generation and swapping, decide on purification, and allocate scarce memories.The central metric is end-to-end fidelity, considered alongside rate and waiting time as quantum QoS quantities.Source: Introduction · p. 1
  • Unlike optical QoT’s usually deterministic scalar margin, a quantum path can deliver a distribution of fidelities because generation and swapping are probabilistic and memories decohere while waiting.The proof of concept targets the distribution’s mean; lower quantiles and threshold-violation probabilities are outside this paper’s scope.Source: Introduction · p. 1
  • Analytical formulas, simulation, and learned predictors offer different estimation options, with reliability depending on regime, runtime, and whether parameters are observable and fresh.Simulation captures waiting and decoherence but is costly for online decisions; a learned model can provide fast predictions after training.Source: Introduction · p. 1
  • The proposed framework adapts the estimator to operating conditions: compose analytically for simple paths, simulate when stochastic memory effects matter, and use a learned model for many online decisions.The study illustrates this compute/simulate/learn framework with a quantum-network simulator-based proof of concept.Source: Introduction · pp. 1–2
  • Entanglement distribution generates elementary links and fuses them through intermediate Bell-state measurements, while swap imperfections and memory decoherence during waiting affect final fidelity.The figure depicts how link quality, swapping, and waiting contribute to end-to-end fidelity.Source: Figure 1 · p. 2

Quantum QoT and the Optical Precedent

The section frames quantum-path QoT around end-to-end fidelity and contrasts it with optical QoT estimation. Under ideal swapping, fidelity composes multiplicatively along a path.

What “QoT” Means in a Quantum Network

  • Quantum-path QoT is carried chiefly by end-to-end fidelity, which is difficult to predict because it is stochastic, decoheres in memory, is consumed on use, and cannot be measured without destroying the state.Entanglement is distributed by generating elementary links and fusing them through intermediate swaps, which compound imperfections and waiting-time decoherence.Source: What “QoT” Means in a Quantum Network · p. 2
  • A Werner link models an elementary entangled channel as a mixture of a maximally entangled Bell state and uniform noise.Source: What “QoT” Means in a Quantum Network · p. 2
  • With ideal swapping, end-to-end fidelity composes multiplicatively across the path’s link fidelities.The section presents this product as the quantum counterpart of optical GSNR composition and a formula the control plane can use.Source: What “QoT” Means in a Quantum Network · p. 2

Three Ways to Get the Metric

Quantum QoT estimates can come from analytical formulas, simulators, or learned surrogates, each with different costs, accuracy conditions, and parameter needs. The optical precedent contributes a decision discipline, but quantum estimation faces dynamical and unobservable-state challenges.

Three Ways to Get the Metric

  • Analytical formulas are fast but regime-dependent, simulators reproduce stochastic dynamics at runtime cost, and learned surrogates trade offline data and training for rapid feature-based predictions.Table I contrasts the three sources by cost, accuracy, and required parameter knowledge.Source: Three Ways to Get the Metric · p. 2, Table I · p. 3

The Optical QoT Estimation Story

  • Optical QoT estimation layered learned predictors over physics engines such as GNPy to estimate margins for unprovisioned lightpaths and recover capacity.The article imports the decision discipline of computing when physics is trustworthy and learning for online estimates of unseen configurations.Source: The Optical QoT Estimation Story · pp. 2–3
  • Unlike the optical case, the quantum formula is undermined by decoherence over random waiting times and by inputs that no-cloning makes unobservable at runtime.In optics, learning often addresses parameters impractical to probe exhaustively; here the formula’s functional form also cannot represent the dynamical effect.Source: The Optical QoT Estimation Story · p. 3

Learn, Compute, or Simulate? A Decision Grid

The decision grid separates whether learning should replace a cheap analytical formula from whether it should replace an accurate but slow simulator. Formula failure arises when its fixed composition cannot represent dynamics or its inputs are unobservable or stale.

Learn, Compute, or Simulate? A Decision Grid

  • The workflow first asks whether the multiplicative formula is accurate for the path and feedable with cheap, fresh parameters; otherwise it asks whether simulation can run at decision time, using a trained surrogate when online simulation is too costly.The learning workflow collects simulator outputs offline, forms a QoT dataset, and trains a surrogate for online control-plane queries.Source: Learn, Compute, or Simulate? A… · p. 3

Axis A: Learn vs. Compute: When the Formula Fails

  • The formula composes fixed link fidelities as if delivered instantly, so probabilistic generation creates random waiting and memory decoherence that it cannot represent.It is accurate when little decoherence occurs and becomes biased as waiting times grow; the resulting residual is described as learnable.Source: Axis A: Learn vs. Compute: When the… · p. 3
  • Formula idealizations also fail when BSM quality varies across nodes or when interleaved purification makes composition nonlinear and order-dependent.The article does not measure these effects in its study.Source: Axis A: Learn vs. Compute: When the… · p. 3
  • Even a perfect formula may be impractical when active link fidelities or residual memory coherence are unobservable, stale, or costly to characterize.Measuring a live entangled state consumes it, and decoherence on control-round-trip timescales can make measurements out of date; learned predictors instead use stable, cheaply observable features.Source: Axis A: Learn vs. Compute: When the… · p. 3

Axis B: Learn vs. Simulate: The Deployment Case

A learned surrogate’s deployment case is replacing costly simulator runs, not the analytical formula. This distinction makes simulator re-execution the relevant baseline for generalization claims.

Axis B: Learn vs. Simulate: The Deployment Case

  • A surrogate’s relevant baseline is re-running the simulator, because analytical formulas accept new path lengths and hop counts directly.The deployment question is whether simulation-based accuracy can be delivered without paying the simulator’s cost at decision time.Source: Axis B: Learn vs. Simulate: The… · pp. 3–4
  • Detailed simulation is too slow for routing loops that must answer in milliseconds and may not be configured for the topology or scale at hand.With appropriate inductive bias, such as graph structure, a surrogate can also predict for networks unseen in training.Source: Axis B: Learn vs. Simulate: The… · p. 4
  • The argument is not that the formula is wrong, but that simulator-level accuracy is needed without paying the simulator’s cost at decision time.The study therefore compares the surrogate against re-simulation on this deployment axis.Source: Axis B: Learn vs. Simulate: The… · p. 4

Putting It Together

The controller should use an analytical formula when it is accurate and supplied with fresh, cheap inputs; otherwise it should simulate when runtime permits or use a trained surrogate for costly online repetition. The proof of concept uses simulator-generated labels and observable path features to predict mean delivered fidelity.

Putting It Together

  • Use the multiplicative formula when it is accurate for the path and its inputs are cheap and fresh; otherwise simulate if runtime permits, or use a trained surrogate for costly online requests.Learning is reserved for cases where a higher-fidelity source is needed but repeated simulation is too costly at decision time.Source: Putting It Together · p. 4
  • The surrogate is trained offline on simulator-evaluated configurations and fidelity labels, then deployed as a fast online estimator.Simulation supplies the training data; learning amortizes that simulation over future control-plane decisions.Source: Putting It Together · p. 4

An Illustrative Simulation-Based Study

  • The study uses SeQUeNCe’s Bell-diagonal pipeline, with probabilistic heralded link generation, random memory waits, depolarizing decoherence, and entanglement swapping.The simulator tracks fidelity analytically for each delivered pair.Source: An Illustrative Simulation-Based Study · p. 4
  • Configurations span 2–40 km links, nominal fidelities of 0.80–0.99, and coherence times of 0.2–5 s; pairs waiting beyond half the coherence time are discarded.The cutoff reflects a controller action and keeps delivered-fidelity distributions unimodal, making their mean a meaningful target.Source: An Illustrative Simulation-Based Study · p. 4
  • The model predicts mean simulator-delivered fidelity from hop count, nominal link fidelities, distances, and coherence-derived decoherence severity.Features are cheap for a control plane to read and keep fresh; labels average three stochastic runs, and the target is simulator fidelity, not real-hardware fidelity.Source: An Illustrative Simulation-Based Study · p. 4
  • The prediction comparison uses only observable physical features, not the analytical estimate, so outperforming the formula would be a genuine comparison.The labels are simulator outputs rather than measurements from real hardware.Source: An Illustrative Simulation-Based Study · p. 4

The Simple Regime: the Formula Suffices

Analytical composition closely matches simulation when memory waiting is negligible, but becomes inaccurate when waiting-driven decoherence shapes a broad fidelity distribution. The proposed workflow uses analytical computation, detailed simulation, or a learned surrogate according to formula validity, input freshness, and runtime constraints.

The Simple Regime: the Formula Suffices

  • MSE is 2 × 10^-7 in the low-decoherence regime, where the uncalibrated formula overestimates simulated fidelity by about 5 × 10^-4.For two- and three-hop paths with 4–6 km links, 0.92–0.96 nominal fidelities, and 50–80 s coherence times, repeated runs agree within 10^-4.Source: The Simple Regime: the Formula Suffices · p. 4
  • The formula’s roughly 5 × 10^-4 overestimate is three orders of magnitude smaller than any fidelity threshold a controller would act on.The residual reflects hardware imperfections and slight decoherence omitted from the ideal formula.Source: The Simple Regime: the Formula Suffices · p. 4

The Complex Regime: Fidelity Becomes a Distribution

  • Per-pair fidelity has a standard deviation above 0.1 on a representative three-hop path when longer links and shorter coherence times make waiting-driven degradation consequential.Pairs assembled quickly arrive fresh, while pairs that wait arrive degraded.Source: The Complex Regime: Fidelity Becomes a… · p. 4
  • The static formula predicts roughly 0.86 where simulator expected fidelity is about 0.64, because it cannot represent waiting-time decoherence.The discrepancy is attributed to decoherence, not measurement noise.Source: The Complex Regime: Fidelity Becomes a… · p. 4
  • Halving the error of an n-run simulator estimate requires four times as many runs, making accuracy costly to obtain by simulation.The learned surrogate amortizes that accuracy cost.Source: The Complex Regime: Fidelity Becomes a… · p. 4
  • The workflow trains a surrogate offline from simulator-generated fidelity labels and uses it online when repeated simulation is too costly.The online flow instead chooses analytical computation when its formula is valid and inputs are fresh, or detailed simulation when runtime allows.Source: Figure 2 · p. 5

Learning the Expected Fidelity

Learned estimators can outperform a calibrated formula for expected fidelity, while lightweight models suffice for the studied path-local task. They also transfer across unseen path and link regimes, though not yet to unseen meshed topologies.

Learning the Expected Fidelity

  • Tree ensembles achieved roughly 5–6× lower mean-squared error and about 2.5× lower mean absolute error than the per-fold-calibrated formula.A plain linear regressor performed almost as well as the tree ensembles; the neural network was the only learned model that failed to beat the formula.Source: Learning the Expected Fidelity · p. 5
  • The surrogate overtook the formula after roughly 80 labeled configurations and kept improving; once trained, it predicts in well under a millisecond versus seconds for simulation.The study reports that the neural network performed worst, while a plain linear model nearly matched tree ensembles.Source: Learning the Expected Fidelity · p. 5

Generalizing to Unseen Path Regimes

  • Across held-out ring, grid, and backbone families, every model beat the calibrated formula by roughly a factor of two in mean absolute error, with extrapolation nearly as accurate as interpolation.The evaluation trained on two path families and tested on the third.Source: Generalizing to Unseen Path Regimes · p. 5, Figure 4 · p. 6
  • The demonstrated transfer is across path and link regimes, not unseen meshed topologies, because this regime’s fidelity is path-local and graph structure adds no signal.Learning over network structure is left for settings where a path’s fidelity depends on the rest of the network.Source: Generalizing to Unseen Path Regimes · p. 6
  • The study concludes that formulas suffice in simple regimes, while learned estimators bridge formula bias and slow simulation in complex regimes at simulator-grade accuracy and formula-grade speed.The demonstrated capability covers trained configurations and unseen path regimes.Source: Generalizing to Unseen Path Regimes · p. 6

From Estimation to Action

Fidelity estimates support path selection, purification, and memory-allocation decisions. Learned estimators can rank alternatives online, with potential provisioning benefits.

From Estimation to Action

  • Path and swap-order selection need candidate rankings before resources are committed, while purification choices trade delivered fidelity against rate.Memory allocation also affects the fidelity a path delivers through the attempts and waiting time budgeted for a request.Source: From Estimation to Action · p. 6
  • A learned estimator can rank many candidate paths or swap orders online, unlike a simulator, narrowing uncertainty margins and potentially reducing wasted purification and over-allocated memory.The stated provisioning benefit follows from applying the optical-network fidelity-margin concept.Source: From Estimation to Action · p. 6

Open Challenges

Deployment raises challenges in validating learned estimates against hardware, limiting simulation data costs, and extending targets and model structure.

Open Challenges

  • There is no corpus of real quantum-network telemetry, leaving validation and transfer to hardware as an open sim-to-real challenge.Source: Open Challenges · p. 6
  • Active learning and design of experiments must address the surrogate paradox of simulating to learn how not to simulate.Source: Open Challenges · p. 7
  • Because fidelity is a random variable, predicting distributions or quantiles is identified as a route to risk-aware provisioning.Source: Open Challenges · p. 7
  • Physics-informed models could encode Werner and decay structure and learn only the residual for sample efficiency and interpretability.Source: Open Challenges · p. 7
  • Graph neural networks are proposed for structural generalization and joint fidelity, rate, and latency targets.Source: Open Challenges · p. 7

Conclusion

The choice among computing, simulating, and learning depends on the operating regime, not a default preference for machine learning. In the study, learning addressed cases where formulas were biased and simulation was too slow, while optical QoT offered a roadmap whose quantum-specific constraints change which regime dominates.

Conclusion

  • Learning is most useful when a deployable formula is biased or unfeedable and accuracy is needed faster or at a scale repeated simulation cannot provide.The paper frames the choice as a structured decision rather than a default to machine learning.Source: Conclusion · p. 7
  • The learned estimator achieved simulator-grade accuracy at formula-grade speed, including on unseen path regimes, where random waiting-time decoherence biased the formula and slowed simulation.The formula could not represent this dynamical effect, while repeated simulation was too slow for real-time decisions.Source: Conclusion · p. 7
  • Optical QoT provides a roadmap, but decoherence, no-cloning, and state consumption change which estimation regime dominates in quantum networks.Source: Conclusion · p. 7
Loading 2610.23827v1…