Anytime-Valid Li & Ma: KL Confidence Bounds for Scheduling Telescopes in the 20–50 GeV Gap

Ground-based gamma-ray telescopes can now see down to about 20 GeV. That closes much of the energy range between what satellites can collect and what ground arrays could previously detect. But at these energies every faint feature is starved for photons. The single Large-Sized Telescope prototype of the Cherenkov Telescope Array Observatory (LST-1) detected the second pulse peak of the Geminga pulsar at 12.2σ in about 60 hours. In the same data the first peak reached only 2.6σ, and the authors estimate that a 5σ detection of that first peak would need more than 200 hours (Detection of the Geminga pulsar at energies down to 20 GeV with the LST-1 of CTAO). Telescope time at these energies is the scarce resource. Today it is spent through fixed, pre-planned campaigns, and significance is judged only once the data are in hand.

This proposal starts from one observation. The standard significance formula in gamma-ray astronomy, the Li & Ma statistic, is already a Kullback-Leibler (KL) divergence. KL divergence measures how distinguishable two probability distributions are. Li & Ma is the profile likelihood ratio between "background only" and "background plus source" for two Poisson counts, one from the source region ("on") and one from a background region ("off"). Bandit algorithms build their confidence intervals from exactly this kind of quantity, but they add a threshold that grows with time, so the intervals remain valid however often the data are inspected. We propose to give astronomers the time-uniform version of their own statistic. That means a confidence bound and stopping rule that may be checked after every observing run without inflating false discoveries, followed by an allocation policy that moves hours between targets once a question has been settled.

Core Claim

Recomputing Li & Ma after each night and stopping at 5σ is optional stopping. It raises the false-discovery rate above the nominal level, and nobody in the cited work quantifies how much. A KL confidence bound for the on/off Poisson model, with the unknown background removed by profiling it out of the likelihood and a threshold that grows like $\log(1/\delta) + O(\log\log t)$, gives a statistic that is valid at any stopping time. It costs only a modest number of extra hours over the fixed-sample test.

Concretely, let $N_{\rm on}, N_{\rm off}$ be the counts accumulated after exposure $t$, let $\alpha$ be the on/off exposure ratio, and let $\Lambda_t(s)$ be the profile log-likelihood ratio comparing the best-fit signal rate with signal rate $s$, with the background re-fit at each $s$. The confidence set is

$$ C_t = \{\, s \ge 0 : \Lambda_t(s) \le \beta(t,\delta) \,\}, $$

where $\beta(t,\delta)$ is a time-uniform threshold that holds simultaneously over all $t$ with probability $1-\delta$. Li & Ma significance is $\sqrt{2\Lambda_t(0)}$. A fixed-sample 5σ claim therefore corresponds to $\Lambda_t(0) = 12.5$. The anytime-valid claim requires instead that $0 \notin C_t$, which demands somewhat more evidence than that.

Once each target has such a bound, scheduling becomes a pure-exploration bandit problem, the setting where an agent spends samples only to reach correct decisions with high confidence. Each target, together with its observing condition, is an arm. The goal is to classify every scientific component (a pulse peak, the bridge emission between peaks, a spectral bin) as detected, ruled out above a flux limit, or still open, using as few telescope hours as possible.

There is a physical caveat that the scheduling problem must respect. Pointing at Geminga observes both of its peaks at once, because phase is assigned after the photons arrive. Time therefore cannot be moved from P2 to P1 on the same pulsar. A confirmed P2 frees hours only for other targets. An arm consequently returns a vector of statistics, and the policy must decide which unresolved component justifies pointing at that target at all.

Research Program

The method side has two parts.

The first part is the confidence bound itself. A Finite-Time Analysis of Multi-armed Bandits Problems with Kullback-Leibler Divergences proves finite-time KL regret bounds for one-parameter and bounded families and lists models beyond those families as open. The on/off model has two parameters, the signal and a nuisance background. We will derive a time-uniform threshold for the profiled ratio, check its calibration in simulation down to counts of a few per bin, and provide a fast root-finder for the bound. Root-finding cost is the computational objection that same paper raises, but here it is one-dimensional and negligible on the scale of an observing night.

The second part is allocation. We will adapt KL-LUCB from Information Complexity in Bandit Subset Selection to a thresholding objective. KL-LUCB repeatedly samples the arm whose interval is most ambiguous with respect to the decision boundary, and here the decision is whether a component lies above or below a detection or flux threshold. That paper shows the relevant sample complexity is governed by Chernoff information, the best achievable error exponent for telling two distributions apart. It also shows the advantage of KL bounds over additive ones grows when the means approach zero, which is the faint-signal case. Our version changes three things: Poisson on/off arms in place of Bernoulli arms, vector-valued arms with shared exposure, and hard visibility windows (zenith angle below 50°) that restrict which arms can be pulled on a given night.

From Fast and Regret Optimal Best Arm Identification we take the asymmetry it argues for: explore using upper bounds, commit using lower bounds. We use upper bounds to decide where to point and lower bounds to declare a component detected. We do not borrow its regret objective, because a telescope schedule is judged by decisions per hour, not by accumulated reward.

The evaluation is a simulator calibrated to the published LST-1 results. Assuming significance grows as $\sqrt{t}$, as it does for a background-dominated Poisson excess, Geminga P2 accrues about $12.2/\sqrt{60.1} \approx 1.6\sigma$ per $\sqrt{\rm h}$ and P1 about $0.34\sigma$ per $\sqrt{\rm h}$. This extrapolation reproduces the authors' estimate of more than 200 hours for P1 at 5σ, about 220 hours, which is a useful sanity check on the calibration. On the same scaling, P2 alone would cross a fixed 5σ in roughly 10 hours. The Crab pulsar contributes a strong target (15.2σ pulsed in about 103 hours) and a faint one, the bridge emission at 5.7σ (A detailed study of the very high-energy Crab pulsar emission with the LST-1). Background rates, on/off ratios and zenith dependence will be set from these papers and from the MAGIC stereo performance study (Performance of the MAGIC stereo system obtained with Crab Nebula data), which also supplies the 15–17% energy-scale systematic that we inject as a model misspecification. Null targets, with zero true signal, are mixed in to measure false discoveries.

Baselines

The naive reference point is the static campaign: hours fixed in advance in proportion to the actual published allocations, with Li & Ma computed once at the end. It defines the hours spent per resolved component today. There is no adaptive scheduling result on Lacuna to compare against, since the file reports none for any Cherenkov telescope, so this static schedule is the honest baseline. It is run on the same simulated sky, visibility calendar and total hours as every other method.

The strongest realistic competitor is sequential Li & Ma with greedy forecasting. Each night it points at the visible target whose open component has the largest projected significance gain per hour, and it stops a component at a fixed 5σ. This is what an observatory would plausibly do on its own, needs no new theory, and will usually spend fewer hours than our method because its threshold is lower. To beat it we must show one of two things. Either it overshoots the nominal false-discovery rate on null targets by a meaningful factor while our method holds its rate. Or, once its threshold is raised enough to match our false-discovery rate, our KL-LUCB allocation resolves more components within the same hours. We will also run a variant of this baseline with a Bonferroni-style corrected threshold, so that it is not handicapped artificially.

Hoeffding/sub-Gaussian LUCB, with the same allocation logic and additive intervals, isolates the information-geometric ingredient. Information Complexity in Bandit Subset Selection reports that KL-LUCB "required significantly fewer samples than its Hoeffding-based predecessor" on the near-zero instance $B_2$, but gives no per-arm count. We will report our own counts. For the commit step, Fast and Regret Optimal Best Arm Identification reports more than 20% lower regret than vanilla UCB in synthetic bandits. That is the only quantitative adaptive-stopping figure in the file, and it is not a telescope result.

Two ablations complete the comparison. The first pairs the anytime-valid stopping rule with greedy allocation, which separates the value of the bound from the value of the allocation policy. The second keeps the allocation policy but replaces the profiled background with a plug-in estimate, which tests whether profiling the background is what keeps the method calibrated.

Why not just…?

Why not just recompute Li & Ma every night and stop at 5σ? This is the real threat. If its inflation of the false-discovery rate turns out small at realistic run cadences, our contribution shrinks to a calibration correction. The null-target experiment settles the question directly. We will report the false-discovery rate of naive sequential Li & Ma as a function of how often the data are inspected, and the extra hours our threshold costs. If the inflation is negligible at weekly inspection, we will say so and recommend the simpler rule at that cadence.

Why not just use Hoeffding or Gaussian confidence intervals? Excess counts in the 20–50 GeV range are small and background-dominated. Additive intervals do not narrow near zero, and Information Complexity in Bandit Subset Selection shows that this inflates sample cost precisely when the means shrink toward zero. The Hoeffding-LUCB baseline measures the size of that effect on Poisson arms.

Why not just use Thompson sampling? Thompson sampling is simple and may allocate time well, but it gives no frequentist stopping guarantee. For subset selection, efficient Thompson analogs are noted as open (Information Complexity in Bandit Subset Selection). We will still include Gamma-Poisson Thompson allocation paired with our stopping rule. If it matches KL-LUCB, we will keep the bound and recommend the simpler allocator.

Why not keep schedules set by the time allocation committee? Committees can still set target lists and total budgets. Our method only reorders hours within a granted pool and frees time once a question is settled. The comparison against the static schedule quantifies how much time this returns.

Decisive Checks

The work succeeds if three things hold. First, in simulation, the profiled KL bound holds its nominal false-discovery rate on null targets under nightly inspection, while naive sequential Li & Ma exceeds it. Second, the anytime threshold costs no more than a modest fraction of extra hours relative to the fixed-sample 5σ crossing; on the Geminga-calibrated P2 arm, that crossing is about 10 hours by $\sqrt{t}$ extrapolation. Third, within a fixed total budget comparable to the roughly 160 combined Geminga and Crab hours, KL-LUCB allocation resolves more components than the corrected greedy baseline, most visibly faint ones such as the Crab bridge emission and the Geminga P1 peak. The bound passes on its own if the first two hold, and the KL-LUCB allocation claim stands only if the third does as well. If corrected greedy matches our allocation, we report that result and retain the bound as the contribution.