A newly discovered pulsar is only useful once astronomers can count every one of its rotations. Counting every rotation means building a "phase-connected" timing model: a formula for the star's spin, spin-down and orbit that predicts the arrival of each pulse with no ambiguity about how many turns happened between observations. Until that model exists, the faint signal cannot be folded coherently over long spans. Folding means stacking many rotations on top of each other so the pulse adds up while the noise averages out. Getting there takes repeated follow-up observations, and their timing matters. If the gap between two sessions is too long, the predicted pulse phase becomes uncertain by more than one full turn. Several integer cycle counts then fit the data equally well, and the effort can be wasted. Surveys such as FAST's CRAFTS and the uGMRT globular-cluster search now find pulsars faster than follow-up time can be scheduled by hand, and that follow-up is spread across several telescopes with separate calendars and slow communication.
We propose to treat follow-up scheduling as asynchronous multi-agent active search. Each telescope is an agent. Whenever it has a free slot, it draws one plausible timing model from the shared posterior and chooses the pulsar and slot that would best resolve the cycle-count ambiguity if that sampled model were true. Here the posterior is the probability distribution over a pulsar's timing models given all arrival times measured so far, and it explicitly includes the discrete question of which integer cycle count links the observations. Drawing a random sample from it is Thompson sampling, and it does two jobs at once. It aims each observation where the remaining ambiguity is largest. It also spreads the telescopes across different targets without them having to coordinate. The concrete claim is that this matches a fully coordinated central scheduler in telescope-hours per phase-connected pulsar when messages between sites are slow or lost, and clearly beats fixed cadences.
Core Claim
The posterior over a new pulsar's timing solution is a mixture with one component per surviving cycle-count hypothesis. Each component is approximately Gaussian in the continuous parameters, because the timing model is close to linear once the cycle count is fixed. The quantity worth reducing is the entropy of the cycle-count variable $K_p$ for pulsar $p$.
Telescope $j$ becomes free at time $t$ and draws a model $\theta^{(j)}$ from the current posterior. It then picks the pulsar it can observe in that slot that maximizes
$$ a^{(j)} = \arg\max_{p} \frac{H(K_p \mid \mathcal{D}) - \mathbb{E}_{y \sim p(y \mid \theta^{(j)}, p, t)}\, H(K_p \mid \mathcal{D} \cup \{y\})}{c_j(p,t)}, $$
where the terms are defined as follows. $\mathcal{D}$ is the set of arrival times shared with telescope $j$ so far. $y$ is the simulated arrival-time measurement that telescope $j$ would obtain. $H$ is entropy. $c_j(p,t)$ is the integration time telescope $j$ needs to reach a usable arrival time on pulsar $p$, which reflects its sensitivity and the source's elevation.
This is myopic posterior sampling, the rule STAR uses for terrain search, applied to a discrete-plus-continuous timing posterior. Different telescopes draw different $\theta^{(j)}$, so they disagree about which pulsar is closest to losing its phase connection, and that disagreement is what spreads them out. The rule is new in two ways. Its search space is cycle-count hypotheses rather than spatial cells. Its agents work through communication delays measured in days rather than seconds.
Where the Pieces Come From
The timing side supplies the model and the failure mode. Phase-connected timing, as described by Miao et al., fits spin and astrometric terms to arrival times measured by template cross-correlation (Zhao et al.). Wu et al. add binary orbital delays. A frequency error $\delta f$ accumulates as a phase drift $\delta f \cdot T$, and spin-down errors grow quadratically, which explains why the allowable gap between sessions shrinks as uncertainty rises.
The follow-up campaigns for 12 CRAFTS pulsars, the eccentric double-neutron-star candidate PSR J2150+3427 ($e = 0.601$, 10.592-day orbit) and the 19-hour eccentric binary in NGC 6093 found by GCGPS show the practical problem. Binaries with poorly known orbits make the cycle-count ambiguity much worse. These papers motivate the work. None of them treats scheduling as a decision problem.
The exploration side supplies the method and its limits. Decentralized multi-agent active search for sparse signals introduces SPATS and LATSI. It shows that asynchronous Thompson-sampling agents finish a search in roughly $T/J$ time with $J$ agents, but its targets are static points on a 2D grid. Stealthy Terrain-Aware Multi-Agent Active Search supplies the cost-aware myopic sampling rule we reuse. Society of Agents proves per-agent regret falling as $1/\sqrt{n}$ for concurrent posterior sampling, but assumes instant data sharing. That assumption constrains us: real observatories share arrival times only after processing, often days later, so delay is the variable we must study rather than assume away.
Igor Colin et al. motivate the choice of a random rule. They analyze the failure of deterministic parallel rules: identical deterministic rules applied to the same posterior produce identical choices. The research file found no prior work that schedules pulsar follow-up with posterior sampling or any multi-agent exploration method. The closest pipeline work, Pulscan and SKA search planning, concerns the blind search stage, not follow-up.
We deliberately did not choose the more obvious crossover, Thompson sampling over the dispersion-measure and acceleration grid of a blind search. Pulscan reports searching a 600-second observation in 0.015 seconds on an H100. When evaluating the whole grid is that cheap, sampling-based search saves little. Telescope time is the expensive, non-parallelizable resource, and that is where sample efficiency pays.
Research Program
We will build three components.
The first is a posterior engine for each pulsar. It keeps a pruned set of cycle-count hypotheses between groups of arrival times, fits the linearized timing model for each, and weights hypotheses by marginal likelihood. Binaries start from the Keplerian fit available at discovery. Hypotheses whose weight falls below a fixed fraction are dropped, and the fraction is logged so the cost of pruning can be audited.
The second is a scheduling simulator. Telescopes with distinct longitudes, sensitivities, downtime patterns and slot calendars each run the rule above. They share arrival times through a channel with configurable delay and drop rate.
The third is a replay benchmark built from real data. Real arrival times from the NANOGrav 11-year data set, for 45 millisecond pulsars, and the Nanshan decade of 85 normal pulsars are "forgotten" down to discovery-level precision. The scheduler can then only request epochs where a real arrival time exists. This keeps the noise realistic, including the red timing noise seen in eleven NANOGrav pulsars. The cost is that the choice of epochs is constrained by what was actually observed.
Synthetic populations seeded with the published orbits of J2150+3427, J1617-2258A and the CRAFTS pulsars cover the hard binary cases. Ground-truth solutions for scoring come from the published ephemerides.
The analysis question is how the advantage of posterior sampling changes with communication delay and team size. We will sweep the delay from zero to several weeks and the number of telescopes from one to eight. At zero delay we expect sampling to roughly match a central scheduler. As the delay grows, we expect deterministic decentralized rules to collapse onto the same targets while sampling stays near the central scheduler.
Baselines
The naive reference is the fixed follow-up cadence used in practice: dense sessions right after discovery, then gaps that grow roughly geometrically, assigned round-robin across telescopes. It uses the same telescopes, slot calendars and population. Its spacing is tuned on a held-out set of synthetic pulsars with the same tuning budget given to every method.
The strongest competitor is a central greedy scheduler. It computes the same expected entropy reduction under the full posterior mean, not a sample. It assigns all telescopes jointly and sees every arrival time instantly. This is the coordinated, best-informed version of our own objective. At zero delay it is an upper reference we expect to tie, not beat. With realistic delay it must work on stale data, and we will run it under exactly the same delay as the decentralized methods. The research file found no published number for any follow-up scheduling method: no reported result found on Lacuna for telescope-hours to phase connection or any comparable scheduling metric. All baseline numbers will therefore be produced inside this study under identical conditions.
The proposal beats the competitor if three things hold. First, at delays of at least a few days it needs significantly fewer telescope-hours per correctly connected pulsar, tested with paired comparisons over shared random seeds. Second, its rate of locking onto the wrong cycle count is no higher. Third, it stays within a small margin of the competitor at zero delay.
Three ablations isolate the key ingredients. Decentralized greedy without sampling tests whether randomness matters at all; each telescope runs the deterministic rule on its own possibly stale posterior. Softmax-randomized greedy adds noise that is not tied to posterior uncertainty, testing whether the randomness must come from the posterior. A Gaussian sampler that ignores the cycle-count mixture and keeps only the leading hypothesis tests whether modeling the discrete ambiguity is what drives the gains.
Why not just…?
Why not just keep the fixed cadence astronomers already use? Fixed cadences ignore what each pulsar has revealed. An isolated slow pulsar with a clean early fit gets the same attention as a 19-hour eccentric binary whose cycle count is still ambiguous. The cadence baseline will show whether this waste is large. If adaptive scheduling saves little telescope time on the CRAFTS-like population, we will report that plainly. The method would then matter mainly for binaries, and we would narrow the claim to them.
Why not just run a central scheduler with a shared claim board? This is the real threat. If sites can post "I am taking pulsar X" with low delay, deterministic greedy plus claiming avoids redundant choices without any sampling. The decentralized multi-agent search work argues that decentralized sampling stays close to central planning when communication fails. However, that evidence comes from robots on a grid, not observatories. The delay sweep settles this directly: we will run claim-board greedy at every delay. If it matches posterior sampling at the delays real networks actually have, the honest conclusion is that a claim board suffices and sampling is only insurance against outages.
Why not just add random tie-breaking or softmax noise to the greedy rule? Arbitrary noise breaks ties, but it pulls telescopes away from good choices as often as toward them. Posterior sampling instead concentrates disagreement on pulsars whose cycle count is genuinely unsettled. The softmax ablation, with its temperature tuned on the same budget, measures whether this alignment matters. If tuned softmax matches posterior sampling, then the uncertainty model is doing the work and the sampling mechanism is not, and we will say so.
Proposed Experiments
The first experiment is the delay-by-team-size grid on the replay benchmark and on synthetic binaries, with all methods above. The primary metric is telescope-hours per correctly phase-connected pulsar. Secondary metrics are the fraction connected within a fixed annual budget, the calendar time to connection, and the wrong-connection rate. A pulsar counts as connected when one cycle-count hypothesis holds more than 0.99 posterior mass. It is correct when that hypothesis agrees with the published ephemeris over the full span.
The first decisive check is the collapse test: the fraction of slots in which two or more telescopes observe the same pulsar within one shared-data interval. Deterministic decentralized greedy should rise sharply with delay, while posterior sampling should stay flat. If it does not, the mechanism we propose is not the one operating.
The second check is scaling. Calendar time to connect the whole population should fall close to $1/J$ with team size $J$, which is the $T/J$ behavior reported for SPATS. A shortfall from that line measures what delay and heterogeneous sensitivity cost.
The third check is safety. Pruning hypotheses and drawing random samples must not raise the wrong-connection rate on red-noise pulsars, because a false lock corrupts every later fold. If it does rise, we will add a verification observation before declaring a pulsar connected and charge its cost to the method.
References
- Arecibo and FAST timing follow-up of 12 millisecond pulsars discovered in Commensal Radio Astronomy FAST Survey
- Follow-up Timing of 12 Pulsars Discovered in Commensal Radio Astronomy FAST Survey
- PSR J2150+3427: A Possible Double Neutron Star System
- Globular Clusters GMRT Pulsar Search (GCGPS). I. Survey Description, Discovery and Timing of the First Pulsar in NGC6093 (M80)
- Decentralized multi-agent active search for sparse signals
- Stealthy Terrain-Aware Multi-Agent Active Search
- Society of Agents: Regret Bounds of Concurrent Thompson Sampling
- Parallel Contextual Bandits in Wireless Handover Optimization
- Pulscan: Binary Pulsar Detection Using Unmatched Filters on NVIDIA GPUs
- Pulsar Searches with the SKA
- The NANOGrav 11-year data set: High-precision timing of 45 millisecond pulsars
- Characterizing Timing Noise in Normal Pulsars with the Nanshan Radio Telescope