Spending Telescope Hours Where the Plasma Is Changing: Freshness-Optimal Monitoring of Repeating FRBs

A fast radio burst (FRB) is a millisecond radio flash from another galaxy. On its way to Earth, ionized gas delays low radio frequencies more than high ones. The size of that delay, the dispersion measure (DM), counts the free electrons along the path. In magnetized gas, the burst's polarization angle also rotates by an amount set by the rotation measure (RM). Some FRB sources repeat, and for them these numbers are not constant. Over about eleven years, FRB 20121102A showed a ~7% drop in its local DM, at a rate of −3.93 pc cm⁻³ yr⁻¹ in 2022–2023, and a 70% drop in RM to 30,755 ± 16 rad m⁻². These changes are read as an expanding young supernova remnant around the source (Decadal evolution of a repeating fast radio burst source). Yet these sources are followed with ad hoc observing campaigns, and the source often turns out to be silent when the telescope is pointed at it.

This note proposes to treat follow-up of repeating FRBs as a status-update problem from networking. A telescope with a fixed budget of hours must keep current estimates of the DM and RM of several drifting sources. It can only "receive an update" if the source happens to burst while it is looking. The cost we minimize is the Age of Incorrect Information: the time during which our best estimate is wrong by more than a tolerance. Waiting while nothing changes is not penalized. The new ingredient is an index policy that couples two beliefs per source: a belief about whether the source is currently active, and a belief about how far its plasma has drifted since the last measurement. The policy then decides which source to observe in each scheduling slot.

Core Claim

Status-update theory already shows that under a resource budget, the best way to track a changing process is a threshold rule on accumulated estimation error, not periodic sampling (The Age of Incorrect Information: A New Performance Metric for Status Updates). For repeating FRBs, this result does not transfer directly, because the "channel" that delivers an update is the source's own bursting. That bursting is clustered and has long quiet stretches. frbpoppy simulations suggest repetition statistics are better described by Weibull than Poisson processes (Synthesising the repeating FRB population using frbpoppy), and FRB 20121102A ranges from silence to 188 bursts per hour.

We claim that the right scheduling index is the expected reduction in incorrect-information age per telescope hour, multiplied by the posterior probability that the source is active. We further claim that this index reaches a target tracking accuracy and change-detection delay with substantially fewer hours than two simpler alternatives: activity-blind error thresholds and activity-only triggering. This claim is falsifiable. Our own ablations, described below, test whether the coupling between the two beliefs is what earns the gain.

Model and Policy

For source $i$, let $X_i(t)$ be the true local DM (and, separately, RM) at time $t$. Let $\hat X_i(t)$ be our current estimate, and let $\epsilon_i$ be a tolerance set by the science use. For DM, that use is host-galaxy subtraction in the DM budget. The incorrect-information age is

$$A_i(t) = t - \sup\{s \le t : |X_i(s) - \hat X_i(s)| \le \epsilon_i\},$$

which is zero while the estimate is within tolerance and grows linearly once it is not. We model the plasma as slow drift plus rare jumps. The drift is calibrated to the reported FRB 20121102A trend. The jumps represent sudden density injections or screen crossings. We model activity as a hidden two-state on/off process with Weibull-distributed waiting times.

Let $\pi_i(t)$ be the posterior probability that source $i$ is active. Let $\lambda_i$ be the burst rate when active, and $\tau$ the slot length. Then the probability of getting at least one usable burst is $q_i = \pi_i(1 - e^{-\lambda_i \tau})$. Let $\Delta_i$ be the expected drop in future $\mathbb{E}[A_i]$ if a burst is obtained. The scheduler observes, in each slot, the source with the largest

$$W_i(t) = q_i(t)\,\Delta_i(t) - \mu,$$

where $\mu$ is a Lagrange multiplier that enforces the yearly hour budget. The scheduler also has the option to idle if no $W_i$ is positive. This is a Whittle-style index for a restless bandit with an unreliable, state-dependent channel. It is close in spirit to the age-of-information bandit literature (Age-of-Information Bandits, Correlated Age-of-Information Bandits). It differs in that the channel quality (whether the source is bursting) and the content value (how stale the plasma estimate is) are separate latent processes, and both are updated by the same observation. An empty observing window lowers $\pi_i$. A detected burst both resets $A_i$ and raises $\pi_i$.

Relation to Prior Work

The FRB papers motivate the problem. The Andromeda and galaxy-cluster DM-budget analyses subtract the components of the DM as if they were fixed numbers (Evidence for a Hot Galactic Halo around the Andromeda Galaxy Using Fast Radio Bursts along Two Sightlines; Deep Synoptic Array Science: Two Fast Radio Burst Sources in Massive Galaxy Clusters). By contrast, the FRB 20121102A decadal study shows local DM moving on yearly timescales. FRB 20190520B's host contribution of about 9037 pc cm⁻³ shows how large the local term can be (A repeating fast radio burst associated with a persistent radio source). Neither line of work asks when to observe.

The status-update papers supply the method. From Age of Incorrect Information work, we take the metric and the result that threshold policies are optimal under a budget. From Optimal Status Updating for an Energy Harvesting Sensor with a Noisy Channel, which shows that uniform best-effort updating is optimal when there is no feedback, we take a warning: if bursting carried no exploitable activity state, our index should collapse to near-uniform sampling. Generalized Age of Information is proportional to change-detection delay (Information Freshness for Timely Detection of Status Changes). We use it as a secondary metric for the jump component of the model.

None of these works handles a channel that is itself a hidden, bursty astrophysical process. The research file reports no prior paper connecting DM estimation with Age of Information.

We deliberately do not pursue the more obvious bridge: modeling CHIME's ~20-second baseband buffer as an age-of-information queue (The CHIME Fast Radio Burst Project: System Overview). That problem has hard deadlines. Earliest-deadline-first scheduling or a modestly larger buffer would likely solve it without any freshness theory, and we have no pipeline logs to show it is a real loss.

Research Program

We will build a simulator of N repeating sources. Each source has drift-plus-jump DM and RM paths and Weibull on/off activity. Burst rates span the observed range, with peaks near the reported 188 per hour. We will also model a realistic noise level for the DM and RM estimate from a single burst. frbpoppy will generate burst arrival processes consistent with CHIME/FRB Catalog 1 repeater statistics.

The simulated telescope has a yearly hour budget, slot granularity, and a visibility constraint per source. We will separately model a free, daily low-frequency DM stream, in the style of CHIME transit coverage, alongside scarce high-frequency time. High-frequency time is needed for RM at the ~10⁵ rad m⁻² level, where rotation across a narrow channel at 600 MHz depolarizes the signal.

We will derive $\Delta_i$ in closed form for the Gaussian drift case and by Monte Carlo for jumps. We will fit $\mu$ by bisection on the budget. We will then prove a simple consistency property: when the activity process is memoryless, the index reduces to the known Age of Incorrect Information threshold rule.

Baselines

The naive reference is fixed-cadence monitoring: the same hours, split evenly across sources at a monthly cadence. This is how campaigns like the FAST/GBT monitoring of FRB 20121102A are effectively run.

The second baseline is an activity-window policy. It observes only when a source is in a predicted active phase, for example the proposed 157-day cycle of FRB 20121102A, and otherwise idles. It is strong because it captures most of the "don't waste hours on silence" benefit with no state model.

The third baseline is the activity-blind Age of Incorrect Information threshold policy from The Age of Incorrect Information: A New Performance Metric for Status Updates. It is our method with $\pi_i$ fixed at its long-run average. This baseline is also the main ablation: it isolates whether coupling activity belief to plasma staleness matters.

The strongest competitor is a greedy Bayesian experimental-design policy. It observes the source with the largest expected one-slot reduction in posterior variance of current DM and RM, computed with the same activity-aware burst likelihood.

All policies share the same simulator seeds, the same hour budget, the same estimator (a Kalman smoother with jump detection), and the same tuning budget of grid search over each policy's free parameters on held-out seeds.

No reported result exists on Lacuna for telescope scheduling of repeating FRBs, or for any of these baselines in an astronomical setting. The nearest reported freshness gain is the roughly 4× lower average age of an age-based edge-computing scheduler over local processing (Age-Based Scheduling for Mobile Edge Computing: A Deep Reinforcement Learning Approach), which is context only. Beating the strongest baseline means one of two things at equal hours: a lower time-averaged incorrect-information age, or a lower median delay in detecting injected jumps, by more than seed-to-seed variation across at least 200 seeds. At equal error, the alternative criterion is a 25% or greater cut in hours.

Why not just…?

Why not just observe on a fixed cadence? Periodic sampling is provably suboptimal for tracking under a budget when the process has informative states (The Age of Incorrect Information: A New Performance Metric for Status Updates; Joint Status Sampling and Updating for Minimizing Age of Information in the Internet of Things). FRB 20121102A's silent "on" phases mean fixed slots routinely return nothing. The experiment quantifies this as wasted hours per usable burst. Feng et al.'s noisy-channel result is a real caveat: if activity carries no memory, uniform sampling is optimal. Our sweep over Weibull shape includes that limit.

Why not just observe whenever the source is predicted to be active? This is the most practical alternative and a serious threat for a single source. It ignores how stale each estimate is. With several sources active at once, it splits time poorly: it keeps re-observing a source whose DM has not moved while a jumping source goes stale. If the activity-window policy matches ours when N = 1 but loses as N grows and as jump rates differ across sources, that confirms where the index helps. If it matches everywhere, the honest conclusion is that activity prediction, not freshness, is the lever.

Why not just do greedy Bayesian experimental design? With our activity-aware likelihood, greedy variance reduction is close to our index. The two differ mainly in two ways. First, variance reduction rewards precision even when the estimate is already within tolerance. Second, it is myopic, so it does not plan around predicted quiet periods. If it ties our policy, we will report that the contribution is the joint activity-plasma model rather than the age metric, and frame the result accordingly.

Why not just use CHIME's free daily transits? For DM, daily low-frequency coverage may make scheduling unnecessary. We include it as a free observation stream, and expect the scheduler to then idle on DM. The scarce resource is high-frequency time for RM and for DM jumps hidden by scattering. The experiment with and without the free stream shows how much scheduling value remains.

Proposed Experiments

The first experiment calibrates and checks the model against reality. We fit the drift and activity parameters to the published FRB 20121102A trend and burst rates. We then confirm that the simulator reproduces a ~7% decadal DM decline and long quiescent stretches.

The second is the main comparison. We vary N from 1 to 10 sources, the hour budget, the jump rate, and the Weibull shape. We report time-averaged incorrect-information age, Generalized-Age-of-Information-based change-detection delay (Information Freshness for Timely Detection of Status Changes), and hours per usable burst for every policy.

The third is a retrospective replay. We restrict each policy to the epochs at which FRB 20121102A was actually observed, and ask which subset it would have chosen and how well that subset recovers the full DM and RM trend. This is limited to epochs that exist in the record, which we state plainly.

The decisive check is the coupling ablation. If the activity-blind threshold policy and greedy experimental design both come within noise of the full index across the sweep, the central claim fails. If the gap widens as activity becomes burstier and sources more heterogeneous, as the model predicts, the coupled index is the thing worth deploying.