Sequential vs. Simultaneous Entanglement Swapping under Optimal Link-Layer Control
Pith reviewed 2026-05-06 03:46 UTC · model claude-opus-4-7
The pith
Sequential entanglement swapping on a four-link chain delivers nothing once external memory coherence falls below 25 heralding ticks, while simultaneous SWAP-ASAP holds steady — and the gap closes only near 10,000 ticks.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
For a four-link chain with the link layer fixed by a trained policy, sequential ("swap-and-wait") entanglement swapping delivers no end-to-end pairs when the external memory coherence time is at most 25 per-link heralding ticks, begins recovering at 50 ticks, and matches simultaneous "wait-and-swap" SWAP-ASAP within 0.4% only once the ratio reaches roughly ten thousand. Simultaneous SWAP-ASAP, by contrast, delivers a constant rate across the entire sweep. The collapse is shown to depend only on external coherence — neither the internal link-layer coherence time nor the trained link policy moves the chain-level outcome — and is mechanistically traced to the per-pair fidelity cutoff falling be
What carries the argument
A two-layer simulation in which every elementary link runs a fixed reinforcement-learned policy that converges to the same dimensionless operating point (~0.1357 bits per heralding tick), so the network-layer protocol is the only independent variable. The mechanism is the per-pair fidelity cutoff t_cut derived from the depolarizing decay law and the six-state QKD floor: when t_cut drops below one heralding tick τ, link pairs expire faster than a sequential controller can pipeline them into a partial chain, while a single-tick swap tree never holds chain storage and is mechanically immune.
If this is right
- The figure of merit hardware roadmaps must beat to make packet-switched quantum networks viable on a single flow is the dimensionless ratio T_c^ext/τ, not absolute coherence time or link length.
- Improvements to internal communication memory or to the link-layer policy will not close the connection-less gap; only the external storage tier moves the boundary.
- Below T_c^ext/τ ≈ 25, near-term deployments should pick centrally coordinated SWAP-ASAP; above ~10⁴ the architectural advantages of sequential swapping become available at negligible rate cost.
- The per-pair cutoff falling below one heralding tick is a sharp, calculable predictor of when a sequential pipeline cannot sustain itself, giving operators a direct design rule.
- For chains longer than four hops, the threshold ratio should grow with n because sequential's chain-assembly window scales linearly while simultaneous's collapse stays single-tick.
Where Pith is reading between the lines
- Because sequential's chain-assembly window scales linearly in n while simultaneous's collapse stays a single tick, the threshold ratio almost certainly grows with chain length, so the millisecond-coherence regime declared 'near-term' here may still be far from sufficient for ten- or twenty-hop networks.
- Modeling local gates and classical communication as instantaneous flatters sequential, which needs n−1 signaling rounds versus ⌈log₂ n⌉ for simultaneous; once latency is restored, the relaxed-coherence equivalence at the upper end of the sweep should narrow rather than vanish.
- The single-flow setting cannot exhibit the contention dynamics that motivate connection-less operation in the first place, so the headline 'penalty is near-term' verdict is best read as a necessary condition for connection-less viability, not a sufficient one.
- The unmeasured T_c^ext/τ ∈ (100, 10⁴) gap covers nearly three decades; the qualitative regime story is consistent with either a sharp transition or a long tail, and these have very different implications for hardware planning.
Load-bearing premise
The conclusion that this is a near-term, memory-limited problem rests on a four-link, single-flow study with idealized gates and zero classical-communication delay, plus a three-decade gap in the coherence sweep that was not measured directly.
What would settle it
Repeat the sweep at chain length n = 4 with the same fixed link policy and fill in the gap T_c^ext/τ ∈ (100, 10⁴): if sequential's delivery rate stays well below simultaneous SWAP-ASAP across that whole window, the "near-term penalty" framing fails; if it converges by a few hundred ticks, the framing is supported. Equivalently, larger-n runs that hold the threshold ratio fixed (rather than growing with n) would refute the cutoff-vs-tick mechanism.
Figures
read the original abstract
Connection-less, packet-switched quantum network architectures distribute entanglement across multi-hop paths through sequential entanglement swapping, in which each node acts on purely local state information. The architectural advantages over the connection-oriented alternative -- simultaneous SWAP-ASAP -- are compelling, but sequential swapping holds partial chains in intermediate buffers between successive swaps, exposing them to memory decoherence in a way simultaneous SWAP-ASAP avoids by design. We present a proof-of-principle study at fixed chain length $n = 4$ in which each elementary link is governed by a fixed reinforcement-learning policy optimizing the secret-key rate of the six-state protocol, leaving the network-layer protocol as the sole independent variable. Sweeping the network-layer memory coherence time $T_c^{\mathrm{ext}}$ over four orders of magnitude reveals a clear regime structure governed by the dimensionless ratio $T_c^{\mathrm{ext}}/\tau$, where $\tau$ is the per-link entanglement heralding latency. Simultaneous SWAP-ASAP delivers a constant rate across the full sweep. Sequential swapping, by contrast, collapses to zero end-to-end deliveries below $T_c^{\mathrm{ext}}/\tau = 25$, and begins recovering at $T_c^{\mathrm{ext}}/\tau = 50$. It remains limited by the simultaneous rate, which it saturates only at the relaxed end of the sweep. These results suggest that the connection-less penalty is a near-term phenomenon tied to present-day memory coherence rather than a fundamental property of sequential swapping.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript presents a proof-of-principle simulation study comparing sequential ("swap-and-wait") and simultaneous SWAP-ASAP ("wait-and-swap") entanglement-distribution protocols on a fixed n=4 chain. The link layer is held fixed via a WN2M2 reinforcement-learning agent (Yau et al.) trained to optimize per-link six-state SKR, and the network-layer protocol is the sole independent variable. The authors sweep T_c^ext/τ over a stressed regime (5–100) and report a relaxed reference (≈10⁴–8×10⁴), finding (i) sequential delivers zero E2E pairs for T_c^ext/τ ≤ 25, partially recovers at 50, and matches simultaneous within 0.4% at the relaxed reference; (ii) simultaneous SWAP-ASAP is invariant in T_c^ext across the sweep; (iii) an off-diagonal T_c^int × T_c^ext sweep shows row-flat behavior, isolating T_c^ext as the controlling variable. A chain-buffer dwell-time diagnostic and a per-pair cutoff calculation (t_cut ≈ 9.2 µs < τ = 50 µs at T_c^ext = 1250 µs) support the proposed mechanism. The authors frame the connection-less penalty as a near-term phenomenon tied to present-day memory coherence.
Significance. The paper makes a useful methodological contribution by separating link-layer policy from network-layer protocol via the link-buffer interface, then empirically verifying that separation through the off-diagonal T_c^int × T_c^ext sweep (Fig. 6). The factorization is a real strength: because the RL policy is trained on a per-link SKR objective while the comparison is at the chain level, the central comparison is not circular — a point the off-diagonal flatness substantiates. The mechanism narrative (per-pair cutoff dropping below one heralding tick) is made quantitative and matches the dwell-time diagnostic. Reporting 95% confidence intervals, a closed-form cutoff expression in Appendix A, and pseudocode for both controllers makes the work reproducible. The result that T_c^int does not enter chain-level performance (within the configurations studied) is a concrete, falsifiable design implication for hardware roadmaps. Within the stated n=4 single-flow scope, the regime structure is well-supported and provides a useful empirical anchor for ongoing work on packet-switched quantum-network architectures.
major comments (4)
- [Abstract / §IV / §V] The central interpretive claim — that the connection-less penalty is 'a near-term phenomenon tied to present-day memory coherence' — extrapolates across the unmeasured range T_c^ext/τ ∈ (100, 10⁴). Section III-B sweeps only up to T_c^ext/τ = 100 in the stressed cluster, and the relaxed reference sits at T_c^ext/τ ≈ 10⁴–8×10⁴. The crossover from 'substantial gap' to 'within 0.4%' is therefore bracketed but not localized over four decades. The authors acknowledge this in §IV ('not directly measured') and §V, but the abstract and contribution list do not carry that hedge. Either localize the crossover with a few intermediate points (e.g., T_c^ext/τ ∈ {250, 1000, 4000}) or revise the abstract/§V framing to match the §IV hedge. As written, the headline claim outruns the data shown.
- [§V (limitations) and contributions list] The n-scaling argument is load-bearing for the 'near-term' framing but is only stated qualitatively ('threshold T_c^ext/τ should therefore grow with chain length'). Because sequential's chain-assembly window scales linearly in n while simultaneous's collapse remains single-tick, modest increases in n could shift the crossover by an order of magnitude and substantially weaken the near-term framing precisely in the regime relevant for deployment. At minimum, please provide either (a) one additional data point at n=6 or 8 at a single T_c^ext/τ to demonstrate the scaling sign and magnitude, or (b) a concrete analytic estimate of how the collapse threshold scales with n under the present cutoff model, so that readers can judge how far the n=4 result extends.
- [§II-A and §II-C] The assumptions of instantaneous, noiseless local gates and zero classical-communication latency are stated once but are non-trivial for the central comparison. Sequential requires n−1 sequential signaling rounds whereas simultaneous requires ⌈log₂ n⌉ — under any nonzero classical-communication latency δ_cc, sequential's effective chain-assembly window grows by ~(n−1)δ_cc while simultaneous's grows by ~⌈log₂ n⌉δ_cc, which directly affects the cutoff inequality t_cut < τ that drives the collapse mechanism. Please add a brief sensitivity discussion (or one figure) showing whether the qualitative regime structure is robust to a representative δ_cc (e.g., one heralding tick), or state explicitly that the comparison is in the δ_cc → 0 limit and that sequential is favored relative to a finite-δ_cc deployment.
- [§II-C and Appendix A (per-pair cutoff)] The mechanism argument hinges on the cutoff formula t_cut = −(T_c/2) log[(F_req(m)−0.25)/(F−0.25)] with link buffers using m=n and chain buffers using m=1. Please justify the asymmetry more carefully: link-buffer entries are required to absorb the full n-link budget while chain-buffer entries are only required to remain above F_min at the buffer tier. This choice plausibly tightens link-buffer cutoffs and is part of why sequential collapses; an ablation in which both buffer tiers use a common m, or a brief argument for why the chosen asymmetry is the natural one, would strengthen the claim that the collapse is intrinsic to sequential rather than to a particular cutoff bookkeeping convention.
minor comments (9)
- [Abstract] The phrase 'saturates only at the relaxed end of the sweep' should specify the dimensionless ratio range (≈10⁴–8×10⁴) at first mention, since the reader does not yet know the relaxed cluster sits four decades above the stressed one.
- [§I, paragraph beginning 'A natural concern...'] The argument that a reserved-route simulation is statistically equivalent to a connection-less protocol relies on three assumptions (fixed path length, homogeneous links, continuous opportunistic generation). The bottleneck-position experiments in Fig. 4 use heterogeneous links; please clarify whether the equivalence claim is intended to hold for those configurations as well, or only for the symmetric topologies.
- [Fig. 3 / Fig. 4] Hatched-vs-filled bar coding is hard to read where sequential delivers 0 bps; consider plotting on a symlog axis or annotating the 0-bps bins explicitly. Similarly, the 95% CIs mentioned in §II-D do not appear visible in Figs. 3–6; please add them or note their suppression.
- [§II-D, Eq. (3)] u_SKR uses the unclamped 1−H(F̄_E2E) inside a max{0,·}. Please confirm whether F̄ is computed before or after the clamp at the SKR-positive threshold, and whether averaging fidelity before applying H(·) (rather than averaging the per-event SKR) introduces a Jensen bias relative to the Yau et al. definition.
- [§III-A, Fig. 2] Reporting 'four-decimal precision' across ten policies is a strong invariance claim. Please report the standard deviation across policies and the per-policy 95% CI so the reader can judge whether 0.1357–0.1358 is genuine convergence or simply within Monte-Carlo noise at N_trials=200.
- [§II-B (training)] The bootstrap gradient ∂u/∂J_T = 0 below F_del ≈ 0.811 should be motivated more explicitly; in particular, whether this introduces a bias against short-episode trajectories that happen to deliver below threshold, and how training stability was monitored.
- [Appendix A, sequential pseudocode] The 're-push the original chain entry to C_{ℓ−1}' branch when the link pair is expired should clarify whether the chain's t_sw is updated to the current t (which would re-anchor decoherence) or left at its original value. The text suggests the latter; explicit pseudocode would remove ambiguity.
- [Citations / arXiv ID] The arXiv identifier 2605.04047 in the manuscript header appears to be a future-dated typo; please verify.
- [Acknowledgment] The use of an LLM for language refinement is disclosed, which is appreciated; consider also stating whether any analysis, code, or figures were produced or modified with LLM assistance, per emerging journal norms.
Simulated Author's Rebuttal
We thank the referee for a careful and constructive report that engages directly with the methodological substance of the paper. We are encouraged that the architectural factorization, the off-diagonal verification, and the cutoff-based mechanism narrative are read as the principal contributions, and that the n=4 single-flow scope is regarded as appropriate for a proof-of-principle anchor. The four major comments all target the boundary between what we measured and what we extrapolated, and we agree that the abstract and contributions list currently carry less hedging than §IV. We will revise accordingly. In summary: (i) we will align the abstract/§V framing with the §IV hedge and add a small set of intermediate T_c^ext/τ points to bracket the crossover more tightly; (ii) we will add a closed-form analytic estimate for how the collapse threshold scales with n under the present cutoff model, and clearly mark the n-scaling claim as analytic rather than measured; (iii) we will state explicitly that the comparison is in the δ_cc → 0 limit, derive the first-order correction to the cutoff inequality, and note that finite δ_cc disadvantages sequential more than simultaneous; (iv) we will justify the m=n vs. m=1 buffer-tier asymmetry on physical grounds and add an ablation in which both tiers use a common m to confirm the collapse is not a bookkeeping artifact. None of these revisions changes the central regime structure or the off-diagonal factorization result.
read point-by-point responses
-
Referee: The 'near-term phenomenon' claim extrapolates across the unmeasured range T_c^ext/τ ∈ (100, 10^4); the abstract and contribution list do not carry the §IV hedge. Either localize the crossover with intermediate points or revise the framing.
Authors: We agree. The §IV language ('bracketed but not directly localized') is the accurate description and should propagate to the abstract and the contributions list. We will (a) revise the abstract to state that the equivalence regime is established by a relaxed-coherence reference at T_c^ext/τ ≈ 10^4 rather than by a localized crossover, and replace 'near-term phenomenon tied to present-day memory coherence' with a more guarded formulation matching §IV; (b) revise contribution (2) to remove the implicit interpolation; and (c) add three intermediate measurement points at T_c^ext/τ ∈ {250, 1000, 4000} for the L=10 km symmetric topology, reported as a new sub-panel in Fig. 3 (and tabulated in Appendix B). These additions bracket the crossover to within roughly one decade rather than four. We will not claim a localized transition where one is not measured. revision: yes
-
Referee: The n-scaling argument is load-bearing for the 'near-term' framing but only stated qualitatively. Provide either an n=6 or n=8 data point or a concrete analytic estimate of how the collapse threshold scales with n.
Authors: We agree this is the right place to do quantitative work and prefer the analytic route, since a clean simulation at larger n requires retraining the link-layer policy at a correspondingly larger F_0 (as we note in §V), which would conflate the scaling we want to isolate with a moving link-layer operating point. We will add a short subsection in Appendix A deriving the threshold T_c^ext/τ at which the link-buffer cutoff drops below one heralding tick, as a function of n, F_0, and F_min, by inverting the closed-form t_cut expression with m=n. The leading behavior is t_cut ∝ T_c^ext · [F_req(n) − 0.25]/[F − 0.25] with F_req(n) → F_min as n grows for fixed F_min, so the cutoff tightens super-linearly in n once F_0 is also scaled to preserve the end-to-end fidelity floor. We will report the predicted threshold for n ∈ {4, 6, 8, 10} and explicitly mark these as analytic, not measured. We will also weaken the 'near-term' language in §V to reflect that modest increases in n shift the crossover unfavorably for sequential. revision: yes
-
Referee: Instantaneous, noiseless local gates and zero classical-communication latency are non-trivial. Sequential needs n−1 signaling rounds vs. ⌈log₂ n⌉ for simultaneous. Add a sensitivity discussion to a representative δ_cc, or state the δ_cc → 0 limit explicitly and that sequential is favored relative to finite δ_cc.
Authors: We agree and will address this in two ways. First, we will state explicitly in §II-A and in the contributions list that the comparison is in the δ_cc → 0 limit and that this limit is favorable to sequential, not unfavorable: under finite δ_cc, sequential's chain-assembly window grows by ~(n−1)δ_cc while simultaneous's grows only by ~⌈log₂ n⌉ δ_cc, so any finite δ_cc shifts the collapse threshold to higher T_c^ext/τ for sequential while leaving simultaneous nearly invariant. Second, we will add a brief analytic sensitivity paragraph to §IV showing how the cutoff inequality t_cut < τ + δ_cc is modified, and we will plot the predicted threshold shift for δ_cc equal to one heralding tick and to ten ticks, again as an analytic overlay rather than a new simulation. The qualitative regime structure (sequential collapses below a threshold; simultaneous is invariant in T_c^ext) is robust to this correction; only the threshold location shifts, and it shifts in the direction that strengthens the conservative reading of our result. revision: yes
-
Referee: The mechanism hinges on t_cut with link buffers using m=n and chain buffers using m=1. Justify the asymmetry or run an ablation with a common m to show the collapse is intrinsic to sequential rather than a particular cutoff bookkeeping convention.
Authors: The asymmetry reflects the physical role of each tier rather than a free choice, and we will state this explicitly in §II-C and Appendix A. A link-buffer entry, at the moment it is consumed, must still survive (n−1) further swaps before reaching the application boundary, so its required floor at consumption is F_req(n) — equivalently, m=n in the inverted decoherence relation. A chain-buffer entry of length i+1, by contrast, has already absorbed the noise of its prior i swaps; the live decoherence-and-swap calculation at consumption time accounts for any subsequent loss, and the buffered entry need only remain above F_min at the buffer tier itself, which is m=1 in the same expression. Using m=n on chain buffers would double-count the budget already paid. That said, we accept the value of an ablation. We will add to Appendix B a sequential-only run of the L=10 km symmetric chain at the stressed T_c^ext sweep with both buffer tiers set to a common m=n (the conservative choice) and a common m=1 (the permissive choice). The collapse below T_c^ext/τ ≈ 25 persists in both ablations: m=n tightens it slightly, m=1 relaxes it modestly but does not remove it, confirming the effect is intrinsic to sequential's pipelined storage rather than to the bookkeeping convention. revision: yes
Circularity Check
No significant circularity: link-layer training objective is independent of the chain-level comparison metric, and T_c^int independence is empirically verified rather than assumed.
full rationale
The paper's central claim is an empirical regime structure (sequential collapses below T_c^ext/τ ≈ 25, recovers at 50, equivalent at ~10⁴–8×10⁴) measured by sweeping T_c^ext while holding the link-layer policy fixed. Walking the chain: (1) Link-layer policies are trained to maximize per-link six-state SKR (a single-link objective). The compared quantity is u_SKR at the chain end-to-end level vs T_c^ext. These are not the same target — the training objective does not directly fix the chain-level comparison outcome, so there is no "fitted-input-called-prediction" reduction. (2) The "architectural factorization" claim (chain outcome depends only on T_c^ext, not on T_c^int or which policy was trained) could in principle be a self-definitional artifact of the buffer architecture, but the authors demonstrate it empirically with a 5×5 off-diagonal T_c^int × T_c^ext sweep (Fig. 6) showing rows flat to four decimals. That is an empirical test, not a definitional consequence. (3) The mechanism (per-pair cutoff falls below one tick) is derived from independent Eqs. (1)–(2) (depolarizing decay and twirled BSM) and the cutoff formula in Appendix A. The numbers (t_cut ≈ 9.2 µs at T_c^ext=1250 µs vs τ=50 µs) follow from the closed-form expression with stated F_0, F_req, not from fitting to the collapse. (4) Self-citation is light and not load-bearing: WN2M2 from Yau et al. [14] is reused as a tool, not invoked as a uniqueness theorem; the packet-switched architecture motivation cites Bacciottini et al. [9] (different group). The reserved-route ≡ connection-less equivalence is asserted with stated conditions (fixed path length, homogeneous links, continuous opportunistic generation) — this is a modeling assumption flagged as a limitation, not a circular justification. (5) The "near-term phenomenon" framing is a hypothesis explicitly bracketed by the unmeasured T_c^ext/τ ∈ (100, 10⁴) gap and the n=4 scope — this is a scope/extrapolation concern (correctness risk), not circularity. The derivation is self-contained against external benchmarks (six-state QKD threshold, depolarizing-channel decay model, fiber attenuation). The compared protocols are evaluated under a symmetric metric defined identically for both. No load-bearing step reduces to its own input by construction.
Axiom & Free-Parameter Ledger
free parameters (6)
- Delivery-fidelity target F_0 =
0.94 (n=4)
- Buffer capacity B =
20
- Coupling/loss factor K =
0.9
- Fiber attenuation length L_att =
22 km
- T_c^ext/τ sweep grid =
{5,10,25,50,100} stressed; ~10⁴–8×10⁴ relaxed
- Trial count and simulated time =
N_trials=200, T_sim=5 s
axioms (6)
- domain assumption Werner states stored for time Δt depolarize as F(Δt)=1/4+(F−1/4)exp(−2Δt/T_c)
- standard math Twirled BSM swap of Werner states obeys F_swap(F1,F2)=F1F2+(1−F1)(1−F2)/3
- ad hoc to paper Local operations (gates, measurements, memory readout) are instantaneous and noiseless
- ad hoc to paper Reserved-route simulation ≡ truly connection-less under fixed path length, homogeneous links, continuous opportunistic generation
- domain assumption Six-state SKR utility max{0,1−H(F)}/⟨Δt⟩ captures application-layer value
- ad hoc to paper Zero classical-communication latency
Lean theorems connected to this paper
-
IndisputableMonolith.Constants (octave = 8·tick)Constants.octave unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
the dimensionless ratio T_c^ext/τ, where τ is the per-link entanglement heralding latency... Sequential collapses to zero end-to-end deliveries below T_c^ext/τ = 25
-
IndisputableMonolith.Cost.FunctionalEquation (composition law)washburn_uniqueness_aczel unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
F_swap(F1,F2) = F1 F2 + (1−F1)(1−F2)/3
What do these tags mean?
- matches
- The paper's claim is directly supported by a theorem in the formal canon.
- supports
- The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
- extends
- The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
- uses
- The paper appears to rely on the theorem as machinery.
- contradicts
- The paper's claim conflicts with a theorem or certificate in the canon.
- unclear
- Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.