A supernova's light curve is how its brightness changes over time, measured through several color filters. Scientists read physics from it: the explosion energy, the type of explosion, and how far away it is. The measurements that matter most are not spread evenly. Color is the difference in brightness between two filters, so it is only useful when both filters are observed close together in time. The width of the curve is set mostly around its peak. Early points are worth far more than late ones. Now picture a constellation of small orbiting telescopes, or any fleet of observing nodes whose communication links keep appearing and disappearing as they move. Each node must decide on its own which transient to observe next and in which filter, and it often cannot reach the ground. Existing constellation schedulers treat every observation request as a fixed reward to be collected. That view cannot represent the fact that a g-band point is worth a lot if someone takes an r-band point within the hour, and almost nothing if two nodes both take g.
We propose to schedule each observation by how much it is expected to reduce uncertainty about each transient's light-curve parameters. The objective is non-additive, and it accounts for which instrument calibrated which point. Coordination then comes from nodes gossiping compact summaries of that uncertainty over the time-varying link graph, a few numbers per transient, rather than from a central controller. The question we test is concrete. With realistic gaps in ground contact and delays between satellites, does passing uncertainty summaries between nodes recover most of what a fully informed central scheduler achieves? And does it clearly beat decentralized schedulers that use hand-tuned, time-decaying rewards?
Core Claim
The closest existing system is decentralized observation scheduling for Earth-imaging constellations. Decentralized, Decomposition-Based Observation Scheduling for a Large-Scale Satellite Constellation splits the constellation into geometric neighborhoods and runs a neighborhood stochastic search. Its requests are fixed ground targets with fixed durations and additive rewards. It lands within about 3% of a centralized scheduler at a small fraction of the messaging cost. We keep that decentralized setting and change what is being maximized.
For each transient $k$ with light-curve parameters $\theta_k$, we score a set $S$ of planned observations with the following objective. Each observation is a choice of node, transient, filter and time. The parameters are amplitude, stretch, color, and time of peak, as in the SALT fits used by ZTF SN Ia DR2: Overview.
$$F(S)=\sum_k w_k\,\tfrac12\log\det\!\Big(I+\Sigma_k^{1/2}\,J_{k,S}^{\top}R_{k,S}^{-1}J_{k,S}\,\Sigma_k^{1/2}\Big)$$
The symbols are defined as follows. $\Sigma_k$ is the current posterior covariance of $\theta_k$. $J_{k,S}$ is the matrix of light-curve model derivatives at the planned filters and times. $R_{k,S}$ is the measurement covariance, and it includes a shared per-node calibration offset, so that two filters taken by the same instrument yield a cleaner color than two filters taken by different instruments. $w_k$ is a science weight, for example the probability from host-galaxy photometry that the transient is a rare class.
The formula is the expected information gain in a model linearized around the current fit. Without the calibration term, this kind of objective has diminishing returns, which is what makes greedy selection nearly optimal. With the shared offsets it is only approximately so, and we check this empirically.
Our claim has two parts. First, maximizing $F$ recovers stretch and color from sparse follow-up much better than additive rewards do, when both spend the same number of observations. Second, nodes that exchange only the matrices $\Sigma_k$, a 4×4 matrix per transient, over a link graph that changes with time can approach a central scheduler that knows everything. They should also degrade more gracefully when ground contact is lost.
Research Program
We will build the scheduler in three layers.
Science model. Each transient keeps a linearized SALT-style posterior. It is updated as simulated photometry arrives. A class prior comes from the host-galaxy estimator of Classifying Astronomical Transients Using Only Host Galaxy Photometry, which classifies from 19 host magnitudes and colors in under a minute. That paper supplies our weights $w_k$. It does not supply any scheduling.
Communication model. We use a time-expanded graph with storage links, taken from Service Function Chain Dynamic Scheduling in Space-Air-Ground Integrated Networks. In that graph a message can wait onboard until a link opens. We reuse only this representation of delay, not its learned scheduler. Posterior summaries and new-alert notices travel through the graph with realistic delays.
Planning. Each node runs local greedy maximization of $F$ over its own feasible observations, using whatever version of each $\Sigma_k$ it currently holds. It then broadcasts its planned observations and the updated $\Sigma_k$ to its neighbors. The delay model is the novel network ingredient. It decides when a stale copy of $\Sigma_k$ causes two nodes to duplicate the same filter.
The network geometry comes from real orbital elements. We use the OneWeb configuration of SkyLink: Scalable and Resilient Link Management in LEO Satellite Networks and the Starlink G6-35 shell used by the SFC-scheduling work. These paper-supplied geometries serve as stand-ins for a future telescope fleet. We place ground stations and take their visibility windows from these orbits.
Ground truth comes from real light curves. We take the ZTF DR2 subset that passes cosmology quality cuts, 2,667 SNe Ia, and treat each well-sampled gri curve as the truth. Every simulated follow-up observation is drawn from the interpolated real curve at the scheduled time and filter, with noise added to match the aperture of a small telescope.
Per-node calibration offsets are injected at the scale of a known instrument problem. ZTF's "pocket effect" shifted stretch by $\Delta x_1=-0.14$ and peak magnitudes by a few percent. For this reason, truth curves taken after November 2019 are either excluded or corrected. We draw the distribution of transient types for the rare-class experiments from the THEx host–transient pairs.
Baselines
Fixed cadence (naive reference). Every visible transient is observed in g and r whenever a node can see it, with ties broken by alert age. This mirrors how surveys sample by default. It gets the same observation budget and the same simulated network. Beating it is the minimum bar. We require the median error in recovered stretch and color, measured against the full-data ZTF fit, to be clearly lower at equal budget.
Neighborhood stochastic search with decaying rewards (strongest existing method). This is the decentralized scheduler from Decentralized, Decomposition-Based Observation Scheduling for a Large-Scale Satellite Constellation. Each transient and filter becomes an additive request whose reward decays with the transient's age and is scaled by class weight. The paper reports an optimality gap of 0.409% on small instances and a satisfaction loss of about 3% relative to centralized Squeaky Wheel Optimization on large instances. No reported result was found on Lacuna for this method on light-curve parameter recovery.
For fairness, we tune its decay constants and class weights by grid search. It gets the same number of simulator evaluations that we spend tuning our weights. Beating it means lower color and stretch error at equal observations and equal message volume, across at least two constellation geometries. It also means the gap does not close as its reward is tuned further.
Centralized greedy information gain (upper reference). The same objective $F$ is maximized by a ground scheduler that uploads plans only during ground-contact windows, as a bent-pipe design would. As an idealized variant, the ground also sees everything instantly. We do not expect to beat the instant variant. The target is to recover most of its value, and to beat the contact-window variant when ground stations fail. SkyLink found that under a 3% ground-station failure, bent-pipe routing costs rose by over 67% while distributed routing costs rose by under 10%. We test whether scheduling science observations shows a similar asymmetry.
Ablations. Three ablations isolate the key ingredients. The first replaces $F$ with a per-filter additive version, which drops the coupling between filters. The second removes the calibration offset from $R_{k,S}$. The third sets the delays to zero, so planning is done as if gossip were instant, while the executed schedule still suffers the real delays.
Why not just…?
Why not just give the existing decentralized scheduler a hand-tuned priority that decays with alert age? This is the real threat. If a well-tuned scalar reward matches $F$, the information-gain machinery is unnecessary. The case against it is structural. An additive reward cannot express that an r-band point is valuable only if a g-band point is taken nearby, or that two nodes taking the same filter is waste. Whether this matters in practice is exactly what the additive-$F$ ablation and the tuned search baseline measure. If they come within noise of full $F$ on color error, we will report that the gain comes from class weighting alone, and the proposal reduces to that simpler result.
Why not just schedule centrally on the ground and uplink plans? At the scale of today's follow-up programs, central greedy scheduling is simple and near-optimal. We include it as the upper reference and do not argue it away. Our bet is narrower. Ground contact for small satellites is intermittent, and SkyLink shows that centralized designs degrade sharply when ground stations fail. The value of gossiping uncertainty summaries should therefore appear in the first hours after an alert and under outages. If the contact-window central scheduler matches ours even under failures, decentralization is not justified for this task.
Why not just wait for a full light curve before prioritizing follow-up? For roughly a million alerts per year at the Rubin Observatory's survey (LSST), waiting misses early phases. Very young events such as OPTICAL AND ULTRAVIOLET OBSERVATIONS OF THE VERY YOUNG TYPE IIP SN 2014cx IN NGC 337 show why those early phases matter. The host-galaxy prior of Kisley et al. gives usable class information at discovery, for example 48% purity for tidal disruption events. We measure how well stretch and color are recovered as a function of the time of the first scheduled observation, which quantifies what waiting costs.
Proposed Experiments
Main comparison. We sweep the observation budget per transient, the number of nodes, and the ground-station failure rate on both orbital geometries. We report median absolute error and outlier rate in stretch and color relative to the full-data ZTF fits. We also report scatter in the standardized peak magnitude. The full data give 0.15 mag of Hubble scatter, and the question is how much of that level sparse follow-up preserves. Messages exchanged and delay from alert to first observation are reported alongside.
Decisive check on the calibration term. We raise the injected offsets from zero to the pocket-effect scale. The full objective should increasingly favor taking both filters on the same node. The ablation without the calibration term should show color bias growing with offset size. If no such divergence appears, that ingredient is dropped.
Staleness experiment. We slow the gossip deliberately to find the delay at which duplicated filters erase the advantage over the tuned additive baseline. This gives a direct link between network topology and science return. Existing link managers such as SkyLink, and the service-migration controller of Graph-Aware Temporal Encoder Based Service Migration and Resource Allocation in Satellite Networks, optimize packet delivery with no notion of what the data mean. They do not provide this link.
Rare-class run. Using THEx class proportions, we check whether weighting by host-galaxy class probability shifts observations toward 91bg-like and tidal-disruption candidates without degrading common SNe Ia. Class purity at first observation is reported against the host-only estimator.
References
- Decentralized, Decomposition-Based Observation Scheduling for a Large-Scale Satellite Constellation
- ZTF SN Ia DR2: Overview
- Classifying Astronomical Transients Using Only Host Galaxy Photometry
- Service Function Chain Dynamic Scheduling in Space-Air-Ground Integrated Networks
- SkyLink: Scalable and Resilient Link Management in LEO Satellite Networks
- OPTICAL AND ULTRAVIOLET OBSERVATIONS OF THE VERY YOUNG TYPE IIP SN 2014cx IN NGC 337
- Graph-Aware Temporal Encoder-Based Service Migration and Resource Allocation in Satellite Networks