How Many Photons Does It Take to Validate a Quantum Switch?

Quantum networks of the future will route entanglement through switches. Entanglement is a shared quantum correlation between distant nodes, and it is created link by link and then joined at a central device. A body of queueing theory now tells us how much traffic such a switch can carry and which scheduling rules reach that limit. Every one of these results has been checked only against the authors' own Monte Carlo simulations, never against hardware. The obstacle is not only that multi-user testbeds are scarce. The signals these theories rest on are counts of rare photon detections: a few hundred to a few thousand per second on real fiber. Particle physicists have long treated scarce counts as the binding constraint when they test a theory against data. This proposal applies the same discipline to quantum networking. Before anyone claims a switch "reaches 99.8% of capacity," we should ask how many detection events it would take to tell that apart from 99%, and whether real links even behave like the theory's assumptions over that many events.

We propose to build a finite-count validation protocol for quantum switch theory. It replays measured single-link detection traces as the client links of a simulated multi-user switch. It then asks three statistically framed questions. First, can a predicted capacity boundary be confirmed or refuted with a realistic event budget? Second, do real link traces violate the "independent, identically distributed arrivals" assumption enough to move that boundary? Third, do physics-aware scheduling gains survive when the scheduler has to estimate link quality from the same scarce counts?

Core Claim

Asymptotic capacity results for quantum switches are only empirically meaningful together with an event budget: the number of heralded link successes needed to resolve a stated capacity difference at a chosen confidence. On real optical links that budget is large compared with the time over which link parameters drift. This creates a sensitivity floor below which theoretical claims cannot be checked, and below which the scheduler's own decisions become noisy. Two things are new. The first is quantifying this floor for the specific published switch results. The second is using recorded hardware traces, rather than synthetic Bernoulli link generation, to test whether the published stability boundaries hold.

The scaling is simple. Suppose a switch serves request traffic at rate $\lambda$ against an effective service rate $\mu$ set by link generation. The queue backlog then drifts at rate $\lambda-\mu$ and fluctuates with a standard deviation that grows like $\sigma\sqrt{t}$, where $\sigma^2$ is the per-unit-time variance of arrivals plus services. Declaring the queue stable or unstable at confidence level $z$ (in standard deviations) therefore takes an observation time of roughly

$$t^\star \approx \frac{z^2 \sigma^2}{(\lambda-\mu)^2},$$

which is the usual first-passage argument for a drifted random walk. Equivalently, a Poisson link producing $N$ successes estimates its own rate to relative precision $1/\sqrt{N}$. Resolving a fractional capacity difference $\delta$ at $z$ standard deviations therefore needs

$$N \gtrsim (z/\delta)^2$$

events. The quantities here are $N$ (counted heralded successes) and $\delta$ (the relative gap between two capacity claims). For $\delta = 0.002$, the gap between the 99.8% stability figure reported by On the Capacity of the Quantum Switch with and without Entanglement Decoherence and full capacity, at $z=2$, this gives about $10^6$ events per load point. At the roughly 400 bits/s secret key rate reported at 3.5 km on IDQuantique Clavis3 hardware by High-Fidelity Coherent-One-Way QKD Simulation Framework for 6G Networks, that is on the order of forty minutes per link per load point. Over that interval drift, recalibration, and detector saturation are plausible, so the i.i.d. assumption is itself under test. The key rate is used here only as a proxy for usable detection rate; the raw heralding rate would be higher, and the protocol will use raw counts where they are available.

Prior Work and Its Roles

The switch theory supplies the predictions we test. On the Capacity Region of a Quantum Switch with Entanglement Purification derives capacity regions for a star switch that either purifies links before joining them (Purify-and-Swap) or joins them first (Swap-and-Purify). It proves Max-Weight scheduling throughput-optimal and reports that a noisy switch must attempt generation about three times as often as a noiseless one. On the Capacity of the Quantum Switch with and without Entanglement Decoherence casts capacity as a convex program. Its dual-gradient scheduler (PDGA) uses queue lengths as feedback and is stable to 99.8% of capacity. Full Exploitation of Limited Memory in Quantum Entanglement Switching handles fewer memories than clients. It reports its MEW scheduler stable at 70% and 99% load and unstable at 120%, with 7 clients and 4 memories. All three assume stationary i.i.d. link generation and one-slot or infinite memory lifetimes, and all three validate only in simulation. We keep their models and policies unchanged and change what drives them.

Entanglement Distribution Delay Optimization in Quantum Networks with Distillation adds nitrogen-vacancy-center physics to switch delay optimization. It reports 30% lower delay than physics-agnostic models and 5–11% fidelity gains. That paper assumes the physical parameters are known exactly. Our third question asks whether such gains persist when the parameters must be estimated online from scarce counts.

Two papers motivate the gap and also limit what we can claim. The COW QKD study is the only work in the file that validates a network simulation against hardware: QBER about 3% and key rates of about 2000 bits/s at 1 km falling to about 400 bits/s at 3.5 km. It explicitly stops at point-to-point links. Accurate and Scalable Simulation of Cavity-Based Networks in Modular Quantum Architectures calibrates the NetSquid simulator against Lindblad master-equation solutions, the standard equations for how a quantum system loses coherence to its environment. It finds that loss along a multi-hop chain is non-monotonic and that fixed-error simulator models are over-optimistic, but it never touches hardware. We borrow its calibration idea and replace the theoretical reference with measured traces.

The methodological lens comes from particle physics. The CLEO-c measurement of the $D^+$ decay constant (Improved Measurement of B(D+→μ+ν)) confronted lattice QCD with about 50 events at 15% statistical error. The later analysis (Precision measurement of B(D+→μ+ν)) reduced this to 4.4% with roughly three times the data. Only then did the comparison with the lattice prediction become sharp. Telescope Array's hotspot analysis shows a further hazard: scanning many configurations inflates apparent significance. That applies directly when one sweeps many load points looking for an instability. Quantum Characterization, Verification, and Validation explains why full tomography cannot verify multi-qubit network states and why single-number summaries such as randomized benchmarking are used instead. It constrains our design: we validate network-level observables (queue stability, delivered rate, delivered fidelity estimated from counts), not full quantum states.

Research Program

The first component is a trace-replay switch. We record time-stamped detection and error events from physical point-to-point links at several fiber lengths. The intended source is the Clavis3/TECNALIA-style setup or an equivalent commercial QKD or heralded-entanglement link; collecting or obtaining these traces is a real dependency and not yet in hand. Each trace becomes one client link of a simulated star switch. A "link success" in a time slot is defined by the recorded events. The switch, scheduler, and memory logic are taken unchanged from the Panigrahy, Valls, and MEW papers. This is not a multi-user hardware experiment, and we will say so. What it adds is that the arrival processes are real, so the i.i.d. assumption is tested rather than imposed. As a fallback, if traces cannot be obtained, we fit non-stationary processes (Poisson with drifting rate, bursty on–off) to the reported marginal rates. Results under the fallback would be labeled synthetic.

The second component is the event-budget analysis. For each published claim (the 99.8% PDGA figure, MEW's 99%/120% boundary, the Purify-and-Swap versus Swap-and-Purify capacity gap), we compute the number of events and the wall-clock time at measured rates needed to confirm it at $2\sigma$ and $5\sigma$. Load-sweep tests are corrected for multiple comparisons. We then test the relationship $N \gtrsim (z/\delta)^2$ empirically by subsampling traces.

The third component is an estimation-aware scheduler. MEW and Chehimi-style policies weight decisions by expected link success and fidelity. We feed them parameters estimated from a sliding window of counts. We compare a plain plug-in estimate with a lower-confidence-bound estimate, and sweep the window length to expose the trade-off between estimation noise (short windows) and staleness under drift (long windows).

Baselines

The naive reference is the theory's own setting: the same switch and scheduler driven by stationary Bernoulli link generation, with rates matched to each trace's long-run mean. This reproduces how the cited papers validate. If trace replay gives the same stability boundaries as this baseline within the event-budget error bars, the i.i.d. assumption is vindicated for these links, and that would itself be a useful result.

The strongest existing methods are PDGA, with its reported stability to 99.8% of capacity, and MEW, with its reported stability at 70% and 99% load and instability at 120% in the 7-client, 4-memory configuration. For delay, the strongest method is the Chehimi hardware-aware policy, reported at 30% delay reduction against physics-agnostic models. We run each with its published hyperparameters and, separately, with an equal tuning budget per condition, on identical traces and seeds. The comparison is always between the same policy on synthetic versus replayed traces, so compute and tuning are matched. "Beating" a baseline means something specific here. It means showing, with corrected significance of at least $3\sigma$, that a reported stability boundary or gain moves under real traces, or that the estimation-aware scheduler recovers a stated fraction of the gap between oracle and plug-in performance. For multi-user hardware switch performance there is no reported result found on Lacuna.

Two ablations isolate the key ingredient. The first replays each trace with its time order shuffled. This keeps the marginal rate and destroys burstiness and drift, so any boundary shift that disappears under shuffling is attributable to non-stationarity rather than rate mismatch. The second gives the scheduler oracle parameters, the true long-run values, instead of estimates. This separates the cost of scarce counts from the cost of model mismatch.

Why not just…?

Why not just run longer Monte Carlo simulations with fixed depolarizing noise? Longer synthetic runs shrink simulation error, but they cannot reveal an effect that the generator never produces. Ben Rached et al. already show that fixed-error models are over-optimistic for multi-hop chains. The shuffled-trace ablation tests this directly: if shuffled and ordered traces give the same boundaries, this alternative is adequate, and we will report that.

Why not just summarize each link with one randomized-benchmarking-style error rate? Twirling noise down to a single depolarizing number is the right tool for gate characterization, per Blume-Kohout et al. It is also what Chehimi et al. do. That single number discards exactly the temporal correlation we suspect matters. It is a real competitor for the fidelity side of the problem. The oracle-versus-estimated ablation, run with a single-number summary, will show whether the extra structure changes any decision.

Why not just wait for a real multi-node testbed? A testbed would be decisive. But it faces the same event budget. Our analysis tells such a testbed how long each load point must run and which published claims it can hope to resolve, so this work prepares that experiment rather than substituting for it. We acknowledge that trace replay cannot capture crosstalk or shared-control-plane effects between links.

Why not just apply standard sequential tests from classical queueing? We do use them. The contribution is not new statistics. It is applying event-budget and trial-factor discipline to specific quantum switch claims with measured rates, which the cited switch papers do not do.

Decisive Checks

The project is a success if one of three things happens. The first possibility is that trace replay moves a published stability boundary by more than its event-budget error bar, and the shift vanishes under shuffling. The second is that the event budget for a headline claim such as the 99.8% figure exceeds the drift timescale measured in the traces, which would mean that claim cannot be validated as stated. The third is that the Chehimi-style 30% delay gain shrinks measurably under estimated parameters and is partly recovered by confidence-bound weighting. A clean null on all three, meaning i.i.d. holds, budgets are affordable, and gains survive, would also be reported. It would give these theories their first hardware-grounded support.

References

  1. On the Capacity of the Quantum Switch With and Without Entanglement Decoherence
  2. High-Fidelity Coherent-One-Way QKD Simulation Framework for 6G Networks: Bridging Theory and Reality
  3. On the Capacity Region of a Quantum Switch with Entanglement Purification
  4. Full Exploitation of Limited Memory in Quantum Entanglement Switching
  5. Entanglement Distribution Delay Optimization in Quantum Networks With Distillation
  6. Accurate and Scalable Simulation of Cavity-Based Networks in Modular Quantum Architectures
  7. Improved Measurement of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="script">B</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>μ</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>ν</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>and the Pseudoscalar Decay Constant<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>f</mml:mi><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:msub></mml:math>
  8. Precision measurement of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="script">B</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>μ</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>ν</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>and the pseudoscalar decay constant<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>f</mml:mi><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:msub></mml:math>
  9. INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
  10. Quantum Characterization, Verification, and Validation