Worst-Case-Optimal Calibration Campaigns for Delayed-Coincidence Detectors

Large neutrino and rare-decay detectors find their signals by looking for pairs of events. A quick "prompt" flash is followed microseconds later by a "delayed" flash in roughly the same place. In a reactor antineutrino detector, the prompt flash comes from a positron and the delayed flash from a captured neutron. In a radon measurement, they come from a bismuth beta decay followed by a polonium alpha decay. Searching for pairs throws away most background. It also creates a calibration problem. Physicists must know how efficiently the detector catches such pairs at every position in its volume, and they learn this by lowering radioactive sources to chosen positions for chosen durations. That source time is taken away from physics running. Today, deployment positions and durations are chosen by operational habit, usually a regular grid.

This note proposes treating the calibration campaign as a G-optimal experimental design. G-optimal means choosing where and how long to measure so that the largest remaining uncertainty anywhere in the detector is as small as possible. The proposal uses the fact that, for pair-based measurements, the noise at each candidate position has a known physical form. That noise comes mostly from random "accidental" pairings of unrelated events, which grow as the square of the local single-event rate. That rate can be read off the ordinary data stream almost for free. The claim is that a design built on this noise model reaches a target worst-case precision on the pair-efficiency map with substantially less source time than a grid or an ordinary noise-blind design. It also predicts in advance where the gains appear: near walls, sensor arrays and surfaces, where single-event rates are high.

Core Claim

The research file finds no Lacuna work joining G-optimal design with delayed-coincidence detection. The intersection is open, but most of the obvious bridges are weak.

The most tempting bridge is to treat the choice of coincidence windows (the allowed time gap Δt and distance Δr between the two flashes) as a design problem. It is really a cut-optimization problem. Windows can be re-chosen after the fact on recorded data, and per-region signal-to-background optimization already handles them. Windows change the estimator, not what data are collected.

The calibration campaign is different. It is a genuine data-collection decision. Its goal, a response map that must be trustworthy everywhere in a fiducial volume, matches the G-criterion's worst-case objective. Its noise is heteroscedastic, meaning its size varies by position, in a way the physics fixes.

Concretely, consider a source of activity $A$ deployed at position $x$ for time $T$. It produces correlated pairs detected with efficiency $\varepsilon(x)$. The coincidence count is Poisson with mean $(A\varepsilon(x) + R(x)^2\Delta t)\,T$, where $R(x)$ is the local singles rate. The accidental term $R^2\Delta t$ is the standard accidental-pair scaling. It is subtracted using a long off-time window, so the efficiency estimate has variance

$$v(x, A) \approx \frac{A\varepsilon(x) + R(x)^2\Delta t}{A^2 T}, \qquad R(x) = R_{\text{bg}}(x) + A\,s(x).$$

Here $R_{\text{bg}}$ is the ambient singles rate and $s(x)$ is the source's own singles yield per unit activity. Because the source feeds its own accidentals, there is an optimal source strength at each position rather than "stronger is better."

We model the efficiency map as linear in $d$ smooth basis functions $\phi(x)$ of position. A design $\pi$ spreads calibration time over position–strength pairs. The heteroscedastic G-objective is then

$$\min_{\pi}\ \max_{z \in \mathcal{F}}\ \phi(z)^\top \Big(\sum_{(x,A)} \pi(x,A)\, \frac{\phi(x)\phi(x)^\top}{v(x,A)}\Big)^{-1} \phi(z),$$

where $\mathcal{F}$ is the fiducial volume. This objective is the worst-case variance of the predicted efficiency over the fiducial volume.

Prior work supplying the method. Learning with Good Feature Representations in Bandits and in RL with a Generative Model analyzes the Kiefer–Wolfowitz theorem. In the homoscedastic case, that theorem guarantees a sparse optimal support (at most $d(d+1)/2$ points) with worst-case variance equal to $d$. This supplies both the solver and a certificate of optimality. SPEED: Experimental Design for Policy Evaluation in Linear Heteroscedastic Bandits is the closest algorithmic relative. It must spend a forced-exploration phase of about $\sqrt{n}$ rounds to learn an unknown quadratic noise structure before it can design. Our setting differs in that the noise form is given by physics, and its main unknown, $R(x)$, is measured from abundant singles at negligible cost. Whether that removes SPEED's exploration overhead in practice is an empirical question we test.

Prior work constraining the design. Two physics papers expose model misspecification. Signatures of muonic activation in the Majorana Demonstrator reports that Geant4 v4.9.6 over-predicts isotope production relative to v4.10.5 and FLUKA. Measurement of ambient radon progeny decay rates and energy spectra in liquid argon using the MicroBooNE detector finds that ion plate-out makes ambient activity about 2.3 times bulk estimates, a strongly position-dependent effect.

A design that concentrates time on a few anchor points is fragile when the linear basis is wrong. We therefore borrow the protection from Robust Sampling for Active Statistical Inference. That paper interpolates along a budget-preserving path between active and uniform sampling, with a burn-in-tuned mixing weight, so the result is asymptotically never worse than uniform. We apply the same idea as a mixture of the G-optimal design with a uniform grid.

Research Program

Simulator. We will build an open, parameterized pair-detector simulator spanning two regimes from the file. The first is liquid-scintillator inverse beta decay with hydrogen capture: 2.22 MeV delayed gamma and about 210 μs capture time, as in Initial measurement of reactor antineutrino oscillation at SNO+. The second is liquid-argon bismuth–polonium tagging: 164 μs polonium half-life and an energy threshold near 100 keV, as in the MicroBooNE radon studies (Measurement of ambient radon progeny decay rates and energy spectra in liquid argon using the MicroBooNE detector, Observation of radon mitigation in MicroBooNE by a liquid argon filtration system).

The simulator has tunable spatial contrast in the singles rate (wall and sensor-array hot zones) and a ground-truth efficiency map. It also lets us generate that truth from a different model than the one the designer assumes. This deliberate mismatch mimics the Geant4-version discrepancy and plate-out effects.

Solver. Second, we implement the heteroscedastic G-design over position and source strength with a Frank–Wolfe solver. Optionally, the solver includes a source-movement time cost. This touches the file's open problem that real instruments cannot jump instantly between anchor points.

We run it in a few batched stages. Each stage refits $\varepsilon$ and $R$ before re-designing, in the spirit of the logarithmic-stage batch scheme of Achieving Minimax Rates in Pool-Based Batch Active Learning. Calibration happens in long runs, not one sample at a time.

Analysis. Third, we derive when the heteroscedastic design differs from the homoscedastic one. The two coincide when $R(x)$ is flat. We also derive the size of the gain as a function of singles-rate contrast and of the accidental-to-signal ratio.

Baselines

Current practice. The do-nothing reference is a uniform grid of deployment positions with equal time and a fixed source strength. This is the default operational choice. We give it the same total source time and the same fitting procedure for the efficiency map.

Noise-blind G-optimal design. This is the key ablation. It is the Kiefer–Wolfowitz design computed as if noise were equal everywhere, using the theory in Learning with Good Feature Representations in Bandits and in RL with a Generative Model. It differs from our method only in dropping $v(x,A)$, so any gap between them isolates the value of the accidental-pair noise model.

Strongest existing method. This is SPEED from SPEED: Experimental Design for Policy Evaluation in Linear Heteroscedastic Bandits, adapted to learn a quadratic noise model from forced-exploration deployments and then design. It receives the same total time, with its exploration counted against the budget. SPEED reports convergence rates on synthetic, MovieLens, wine and air-quality benchmarks rather than numbers transferable here. We found no reported Lacuna result for any calibration-design method on coincidence detectors, so all baseline numbers will be produced in our simulator.

Alternative criteria. We also run A-optimal design (lowest average variance) and c-optimal design (lowest variance of the volume-integrated efficiency) with the same noise model.

Win condition. Beating the baselines means achieving a lower held-out worst-case squared error of the efficiency map at equal source time. Equivalently, it means reaching the grid's worst-case error with at least 30% less source time, consistently across random detector configurations and both physics regimes. Results are averaged over at least 50 simulated campaigns per setting.

Why not just…?

Why not just keep the uniform grid? The grid is robust and simple, and it has served Daya Bay and SNO+ well (Measurement of the Reactor Antineutrino Flux and Spectrum at Daya Bay). Its weakness is that it spends equal time where accidentals are negligible and where they dominate. The worst-case error therefore sits in the high-rate zones that matter most for background-sensitive analyses. If singles rates are nearly uniform, our method should reduce to roughly a grid-like design. That flat-rate setting is a built-in check that we do not claim gains where none should exist.

Why not just optimize cuts region by region for signal over the square root of background? Cut optimization chooses how to analyze data already taken. It does not decide where calibration data are collected, so it is complementary rather than competing. We apply the same post-hoc window optimization to every design so that it cancels out of the comparison.

Why not just use c-optimal or A-optimal design? This is a real threat. If the only downstream quantity is the total detector efficiency, c-optimal design is the right tool and will beat our worst-case design on that metric. We will report this outcome openly. The case for the G-criterion rests on analyses that need the full map: moving fiducial cuts, position-dependent backgrounds near walls, and plate-out regions. We will evaluate all criteria on both the integrated-efficiency error and the worst-case map error, so readers can see which regime they are in.

Why not just run SPEED as-is? SPEED is the closest method and may come close in long campaigns, where its $\sqrt{n}$ exploration cost becomes small. Our claim is narrower. When the noise form is known and its rate parameter is free to measure, forced exploration is wasted in short campaigns, and SPEED's generic quadratic form misses the dependence of accidentals on source strength. The budget sweep below shows whether this advantage survives beyond short campaigns.

Proposed Experiments

Main sweep. We sweep total source time across an order of magnitude and singles-rate contrast from flat to strongly wall-peaked. The headline metric is held-out worst-case error of the efficiency map. Secondary metrics are integrated-efficiency error and source time needed to reach a target precision.

Decisive check 1: gain tracks contrast. The predicted signature is that our advantage over the noise-blind ablation is near zero at flat contrast and grows with contrast. A gain that does not track contrast would falsify the mechanism.

Decisive check 2: robustness to misspecification. We generate truth from a model outside the designer's basis, mimicking the Geant4-version and plate-out discrepancies. The pure G-design may then lose to the grid. The uniform-mixture variant should never lose by more than noise, as the robust-sampling guarantee suggests.

Decisive check 3: source strength. We verify that optimizing source strength jointly with position gives measurable savings over a fixed strength. This is the most physics-specific part of the proposal. If it adds nothing, the method collapses to a reweighted Kiefer–Wolfowitz design.