Deciding What to Keep Before the Network Reconnects: Budgeted Coincidence Triggers for Intermittent Detector Arrays

Rare-event searches in physics often depend on coincidence. A weak flash at one detector station means little alone, but the same flash seen at the same moment by another station may be a real signal. Coincidence analyses in the literature assume that every station's candidates eventually reach one central computer. Consider instead a wide-area array of autonomous stations, such as remote cosmic-ray or neutrino sensors, connected by radio links that fail for minutes at a time. Each station keeps its raw waveforms only in a short rolling memory buffer, and it can send only a small, fixed amount of raw data per hour. Such a station has to decide whether to save a weak candidate before it learns whether any other station saw something coincident. If it discards too much, real signals are lost. If it saves too much, it exceeds the array's shared bandwidth.

We propose a trigger in which each station's willingness to keep weak candidates grows with how little it currently knows about its neighbors. Stations whose links have gone quiet keep more borderline events, because they cannot yet rule out a coincidence. The array's fixed readout budget moves toward those stations through pairwise exchanges that never let the total exceed the budget, even when messages are lost. This combines three existing ideas in one place: the consensus-aware uncertainty bounds from decentralized bandits, the "anytime feasible" resource-sharing dynamics from distributed power-grid control, and the Bayesian coincidence update from multi-messenger astronomy. To our knowledge, no prior work joins them.

Core Claim

The coincidence update of Enhancing the significance of astrophysical events with multimessenger coincidences shows why weak candidates matter. That paper computes the posterior probability that a gravitational-wave trigger is astrophysical, using time, sky-position and energy agreement with IceCube neutrinos. A trigger at 7.7% rose to 86% (moderate emission model) or 97% (optimistic model) when paired with a coincident 100 TeV neutrino. The analysis is offline and centralized. It never faces the case where the partner station's report has not yet arrived.

Our claim concerns that missing-information case. A station should rank candidates by an optimistic posterior: it counts the coincidence evidence it has actually received, plus a bounded allowance for stations it has not heard from recently. A station also needs readout quota in proportion to how much this allowance inflates its candidate list. We write the station-$i$ score for candidate $e$ as

$$ U_i(e) = \log \frac{\pi}{1-\pi} + \sum_{j \in H_i(t)} \log \Lambda_{j}(e) + \sum_{j \notin H_i(t)} c\big(s_{ij}(t)\big), $$

where $\pi$ is the prior signal probability. $H_i(t)$ is the set of stations whose reports covering $e$'s time window station $i$ has received by time $t$. $\Lambda_j(e)$ is the time, position and discriminator likelihood ratio contributed by station $j$, following Veske et al. $c(s_{ij})$ is a nonnegative allowance that increases with the staleness $s_{ij}(t)$ of the last message from $j$, and it is capped at the largest likelihood ratio a single station can plausibly contribute.

This allowance plays the same role as the consensus error term that Distributed Multi-Agent Bandits Over Erdős–Rényi Random Networks adds to its confidence intervals. That term keeps agents from discarding the best arm because of delayed gossip. Here it keeps stations from discarding a true coincidence because of a dropped link.

Station $i$ saves a candidate when $U_i(e)$ exceeds a threshold, and that threshold is set by its readout quota $q_i$. Quota moves between neighbors by

$$ q_i(k+1) = q_i(k) + \gamma \sum_{j} a_{ij}(k)\,\big(g_j(k) - g_i(k)\big), $$

where $a_{ij}(k)\in\{0,1\}$ indicates that the link between $i$ and $j$ succeeded at round $k$. $g_i$ is the marginal expected number of true coincidences station $i$ would lose per unit of quota removed, estimated from its current candidate scores. $\gamma$ is a step size.

This is the primal Laplacian structure of Accelerated Distributed Allocation and Distributed Automatic Generation Control subject to Ramp-Rate-Limits. Those papers move a resource along graph edges so the total stays exactly at the budget at every iteration, under time-varying or uniformly $B$-connected graphs. Each exchange is antisymmetric, so the total quota is conserved.

Our addition handles lossy links. A transfer is two-phase: the sender subtracts quota only after an acknowledgment, and the receiver may spend it only after confirming. The constraint therefore relaxes from exact equality to $\sum_i q_i \le B$, which still holds at every round even when packets are lost mid-transfer.

Research Program

The work has three parts.

Bound on lost coincidences. We will derive how many true coincidences the staleness-aware rule loses compared with an oracle that sees all reports instantly. The bound should be stated in terms of link success probability $p$, ring-buffer retention time $T_r$, and the base graph's algebraic connectivity, mirroring how the bandit regret bound scales with $1/p$ and the Fiedler value. The expected qualitative form is simple. A coincidence is lost only if the partner report fails to arrive within $T_r$ and the allowance $c$ is too small to save the event anyway. The first probability decays roughly geometrically in $pT_r$ for an Erdős–Rényi link model.

Time-scale tension. The research file notes a conflict with the gossip estimator of Convergence Rate Analysis of Distributed Gossip (Linear Parameter) Estimation. That estimator reaches centralized efficiency only asymptotically, with consensus gains decaying more slowly than measurement gains, whereas our signals last milliseconds and buffers last minutes. We do not ask the network to reach consensus on any estimate. We use Kar and Moura's work as a reason not to wait for convergence. Coincidence evidence travels as small, discrete messages (timestamp, position, discriminator score), and only the slowly varying quota uses consensus-style dynamics. The quota only needs to track how outage patterns change, which happens over minutes, not individual events.

Smooth versus accelerated quota dynamics. The signum-accelerated dynamics of Accelerated Distributed Allocation chatter when run in discrete time. That would make thresholds flicker and classification near the boundary inconsistent. We will therefore compare linear and smoothed nonlinear quota updates and measure threshold variance directly.

Simulator. The system will be built on DecentralizePy, which already supports time-varying topologies. We will add three components: per-station ring buffers, a two-phase quota exchange, and a coincidence likelihood module that reimplements the time-position-energy update of Veske et al.

Baselines

Naive reference: local hard cut with later upload. This is the "do nothing" design. Each station keeps only candidates whose single-station discriminator score passes a fixed high threshold, and it uploads them whenever a link appears. It defines how many coincidences are lost by ignoring cross-station evidence. Its threshold will be set so its average readout exactly equals the shared budget $B$, giving it the same data volume as every other method.

Strongest realistic competitor: gossip-informed, fixed-share trigger. Stations flood candidate metadata by randomized gossip, which Epidemic Learning showed beats static sparse topologies: up to 1.7× faster convergence and 2.2% higher top-1 accuracy on non-IID CIFAR-10 with 96 nodes. Each station computes the exact Veske-style posterior over whatever reports it has received, and keeps its top candidates within a fixed equal quota $B/n$. This is what a competent engineer would build. It uses the same coincidence statistic and the same gossip layer as our method, and differs only in the staleness allowance and quota reallocation. It receives the same message budget, the same simulated links, and a tuning sweep over gossip fan-out equal in size to our sweep over $\gamma$ and the cap on $c$.

We found no reported result on Lacuna for coincidence recovery under intermittent links with finite buffers, so no published number exists to beat. Our method counts as beating this baseline if it recovers significantly more true coincidences at equal $B$, measured by paired comparison over at least 20 random seeds per outage regime, with zero budget violations.

Dual-decomposition reference. For quota sharing, we run an asynchronous ADMM-style allocation of the same objective. Distributed Automatic Generation Control subject to Ramp-Rate-Limits reports that ADMM "significantly violated" ramp-rate constraints during transients, while the primal Laplacian method "strictly honored" a ±1 MW/s limit. We expect the analogous failure to show up as budget overshoot when links drop mid-iteration. Here the result that counts is the number of rounds in which the total quota exceeds $B$, not accuracy.

Ablations. Two ablations separate the key ingredients: the staleness allowance with fixed quotas, and quota reallocation with no allowance ($c \equiv 0$). A centralized oracle with instant, lossless reports provides the upper bound.

Why not just…?

Why not just buffer everything and decide centrally after reconnection? If ring buffers outlasted every outage, this would be the right answer and our method would be unnecessary. We do not argue around this case. We sweep the ratio of buffer retention time to median outage length. Our claim applies only below the crossover point, and we will report where that crossover falls. Above it, the store-and-forward approach should match the oracle. The concerns raised in A Framework for Communication Planning on Mobile Devices about energy and deferral costs suggest that real autonomous stations sit below the crossover, but our experiment, not that citation, will settle it.

Why not just raise every station's quota a little and skip reallocation? Uniform slack spends budget on well-connected stations that already know their candidates are background. The fixed-share-with-allowance ablation tests exactly this alternative. If it comes within noise of the full method, we will drop the reallocation layer and report the simpler trigger as the result.

Why not just use a better single-station classifier? Better discrimination helps every method equally, and it cannot recover information that exists only at another station. The 7.7%-to-86% jump in Veske et al. comes from the partner detection, not from the primary detector's own discriminator. We will check this by swapping a shallow classifier for a deep one at each station and testing whether the gap between our method and the baselines persists.

Proposed Experiments

Discriminator statistics. Per-station scores come from real classifier outputs, not invented distributions. We train the five-layer network of Searching for exotic particles in high-energy physics with deep learning on the HIGGS benchmark (11M simulated events). That paper reports 4.9σ discovery significance from low-level features, versus 2.5σ for a shallow network. We use its held-out signal and background score distributions as per-station likelihoods, and use the shallow network as the weak-discriminator condition.

Simulated array. Stations of 16 to 64 nodes receive Poisson background triggers at rates spanning several orders of magnitude above the signal rate. Signals are injected as correlated multi-station events with realistic time and position spread. Links follow time-varying Erdős–Rényi graphs over a fixed base graph, as in the bandit paper, plus a bursty two-state outage model that produces long partitions.

Metrics. We measure three quantities: the fraction of injected coincidences whose raw data are retained at all participating stations, the number of rounds violating the budget, and threshold variance as a measure of chattering.

Decisive checks. Three outcomes would decide the question. First, the method should beat the gossip-informed fixed-share trigger specifically in the regime where retention time is shorter than outages, and tie with it elsewhere; a gain everywhere would point to a simulation artifact. Second, budget violations should be exactly zero under packet loss, while the ADMM reference should show them. Third, measured coincidence loss should follow the predicted dependence on $pT_r$ and algebraic connectivity. If the advantage disappears once the bursty outage model replaces independent link drops, the staleness allowance is miscalibrated, and that negative result would be reported.