When Biased Information Estimates Pull a Telescope Off Its Target

A fast Cherenkov telescope often has to choose between two jobs in seconds. It is in the middle of a long, quiet exposure of a dark-matter target, and an alert arrives about a fading gamma-ray burst or a gravitational-wave event somewhere in a large patch of sky. Should it slew away? A principled automated scheduler answers by estimating how much each action would teach us, its expected information gain. That quantity is a nested average: a logarithm of an average, inside another average. The standard sampling estimator of it is biased upward, and the bias grows with how uncertain the target is. We argue that this bias acts as a hidden, uncalibrated exploration bonus. It pushes an automated scheduler toward poorly localized, possibly false alerts and away from the deep integrations that dark-matter searches need.

The opportunity is timely. The CTA Large Size Telescope prototype can slew 180° in under 30 seconds and is moving from manual to automated alert handling (Carosi et al.). Competitive dark-matter limits, meanwhile, require hundreds of hours of stacked exposure (Acciari et al.). We propose to show that the bias exists and matters in a realistic scheduling simulator, to remove it with debiased gradient estimators, and to train a small "interrupt or stay" policy on the corrected objective. We will also measure honestly when cheaper, non-nested decision rules do just as well.

Core Claim

Let $a$ be an observing action, such as continuing on a dwarf galaxy, slewing to an alert, or taking one tile of a large localization region. Let $\theta$ collect the unknowns: the dark-matter cross-section and the J-factor (the line-of-sight dark-matter density that sets the expected signal), or, for an alert, whether the transient is real, where it sits, and how fast it fades. Let $y$ be the photon counts that would result from the action. The nested Monte Carlo estimator of expected information gain uses $M$ inner prior samples to approximate the marginal likelihood $p(y \mid a)$. A second-order expansion of the logarithm, which is Jensen's inequality made quantitative, gives the leading bias

$$ \mathbb{E}[\hat U_M(a)] - U(a) \approx \mathbb{E}_{y}\!\left[\frac{\mathrm{Var}_{\theta \sim p(\theta)}\,[\,p(y \mid \theta, a)\,]}{2M\,p(y \mid a)^2}\right], $$

where $U(a)$ is the true expected information gain and $\hat U_M(a)$ is its estimate.

The bias is always positive. It is large exactly when the Poisson likelihood of the counts varies strongly across the prior, which is the case for a transient localized to hundreds of square degrees against a field of view of about 4.5°. For a well-studied dwarf galaxy, the likelihood varies much less. A threshold rule such as "interrupt if $\hat U_M(\text{alert}) > \hat U_M(\text{stay})$" is therefore tilted toward exploration by an amount that depends on $M$ and has no connection to scientific value.

Dayan et al. showed that correct Bayesian updating produces principled exploration bonuses. Our claim is the mirror image: estimator bias produces a spurious one. It inflates the value of the most uncertain options, and it does not disappear as outer samples grow. Wood et al. proved that fixed-$M$ nested estimators stay biased as outer samples go to infinity, and derived the $N \propto M^2$ budget rule. Their analysis concerns static experimental designs. It does not study how bias shifts a sequential decision threshold, which is what we measure.

Research Program

We will build a scheduling simulator in three parts.

The first part is the steady-state science side. We simulate stacked dwarf-galaxy observations whose instrument responses and backgrounds come from public MAGIC data analyzed in Gammapy (MAGIC Collaboration). The likelihood follows the profile-likelihood construction of Acciari et al., with the J-factor treated as a nuisance parameter. This choice matters because J-factor uncertainty makes the dark-matter sensitivity saturate with exposure, rather than growing as the square root of time.

The second part is the transient side. We generate synthetic alert streams whose flux decays as a power law in time since the event, following Bartos et al.. Each alert carries a false-alarm probability and a localization region that ranges from point-like, as for burst alerts, to hundreds of square degrees, as for gravitational-wave and neutrino alerts.

The third part is the telescope. Slews are constrained to LST-1-like kinematics, so every repointing costs both travel time and lost flux.

The object we will study is a one-step-lookahead decision among three options: stay on the current target, slew to the alert, or take the next tile. Full Bayes-adaptive planning (Pineau et al.) is intractable at continuous pointing and sky-map scale, and practical lookahead is limited to a few steps. We accept that limitation and state it openly. We will also train an amortized policy: a small network that maps alert features and the current exposure state to action probabilities. Training uses stochastic gradients of expected information gain, since evaluating nested estimators live for every alert is wasteful.

The methodological contribution is to compare three ways of obtaining these gradients, each borrowed from existing work.

The first is naive nested Monte Carlo with fixed $M$.

The second is a dual reformulation that removes the logarithm-of-average. It rewrites $\log \mathbb{E}[X]$ as a minimum over an auxiliary variable, which the network also predicts. Primal-dual methods for compositional objectives (Devraj and Chen) and for LogSumExp (Gladin et al.) use the same device. Those works prove convergence under convexity, which our policy objective lacks. We therefore test the device empirically and do not claim their guarantees.

The third is randomized multilevel estimation in the spirit of Godichon-Baggioni et al.. It trades extra variance for reduced bias.

We will also test the file's observation that discontinuous inner operators, such as the hard edge of the field of view, make the bias oscillate with $M$.

Baselines

The naive reference point is never interrupt. The telescope integrates on the dwarf-galaxy stack for the whole simulated season and ignores alerts. It anchors the dark-matter side. For calibration, Acciari et al. report 95% confidence-level upper limits on the annihilation cross-section of approximately $10^{-24}$ cm³/s at about 1 TeV from 354.3 hours on four dwarfs. Our simulator must reproduce that order of magnitude under an equivalent exposure before any scheduler comparison counts.

The operational baseline is static override, the rule used in multi-observatory neutrino follow-up (Satalecka et al.). Whenever an alert passes a significance and visibility threshold, the telescope slews, observes for a fixed window, and then resumes. We will tune the threshold and window by grid search on held-out alert streams, with the same simulation budget given to every method. No reported result for scheduler-level performance of this rule was found on Lacuna. The LST-1 commissioning report gives only the one-day manual delay for GRB 201216C and non-detections for four transients (Carosi et al.).

The strongest non-nested competitor is a greedy expected-significance or Fisher-information rule. It compares closed-form expected detection significance, or a local Fisher-information gain in the spirit of task-directed design (Wagenmaker et al.). Because it involves no nesting, it is cheap and unbiased by construction. It is the method to beat.

The ablation that isolates our key ingredient is the identical information-gain policy trained with naive fixed-$M$ gradients. We sweep $M$ over three orders of magnitude and use the same network, alert streams, and wall-clock training budget as the debiased versions.

All methods face the same alert realizations, measured by paired comparison over seeds. The debiased policy "wins" only under three conditions. It must lower the season-end expected cross-section limit and raise transient detections simultaneously, relative to the tuned override. It must beat the Fisher rule on the multimodal alert class. And it must match exact-enumeration ground truth where that ground truth exists.

Why not just…?

Why not just use static override rules? They cannot decide whether to tile, which tile to take first, or when to give up after empty pointings. Each false alarm costs clean exposure, and that exposure is the scarce resource in the dwarf-galaxy stack. The override is nonetheless a serious competitor for well-localized burst alerts, where slewing immediately is probably optimal. We expect our advantage to appear only for large-area and low-purity alerts, and we will report results stratified by localization area so that a null result there is visible.

Why not just raise $M$ until the bias is negligible? For some alert classes this will work, and we will report the $M$ at which the gap closes. The bias equation says the required $M$ scales with the relative variance of the likelihood under the prior. For sharp Poisson counts against a sky map much larger than the field of view, that variance can be enormous, and it is largest for exactly the decisions that matter. Wood et al.'s fixed-$M$ non-convergence result also means that adding outer samples does not help. We will plot the required $M$ against localization area. If it stays modest at deployment compute, the contribution shrinks to a diagnostic, and we will say so.

Why not just use Fisher information or expected significance, which avoid nesting entirely? These are local approximations. They assume a roughly Gaussian, single-peaked posterior. Whether an alert is real, and which tile contains it, produce many-peaked posteriors, and local curvature is uninformative about them. This is a genuine threat on unimodal cases such as a dwarf-galaxy flux measurement, where we expect Fisher to be nearly as good. The stratified experiments below decide the question.

Proposed Experiments

The decisive check uses small discretized scenarios, with a few tiles, a handful of decay indices, and a binary real-or-false flag. In these scenarios exact expected information gain is computable by enumeration, and Wood et al.'s rearranged estimator applies to the discrete outcomes. We will plot interrupt probability against localization area for naive estimators at each $M$, for both debiased estimators, and for the exact value.

The prediction is falsifiable. Naive estimators should over-interrupt, with an excess that grows with area and shrinks roughly as $1/M$. Debiased estimators should track the exact curve. If naive estimators show no area-dependent excess, the core claim fails.

At full scale, we run season-long simulations with realistic alert rates. The simulations are built on the Gammapy-validated MAGIC data products and the LST-1 slew and threshold figures, which are under 30 seconds for 180° and about 20 GeV. For each method we report the expected 95% cross-section limit from the dwarf stack, the number of transients detected, the exposure lost to false alarms, and total realized information gain.

A final robustness test misspecifies the J-factor and alert-purity priors. This probes the prior sensitivity that the file flags as an open problem for Bayes-adaptive methods. It tells us whether debiasing still helps when the model itself is wrong.