KATRIN, the Karlsruhe Tritium Neutrino experiment, weighs the neutrino by looking at the very top of the tritium beta-decay electron spectrum. A filter lets through only electrons above a chosen voltage threshold, and the experiment counts how many arrive. Its central design choice is how many hours to spend at each threshold. This schedule is called the measurement time distribution. In the published analyses it comes from heuristic scan patterns and empirical sensitivity scans, not from a formal optimization. Optimal experimental design is the statistics of choosing where to measure so that unknown parameters are pinned down as tightly as possible. One of its standard tools is D-optimality: maximize the determinant of the Fisher information matrix, which is the same as shrinking the volume of the parameter uncertainty ellipsoid. We propose to treat KATRIN's voltage schedule as an optimal-design problem. The schedule would target the neutrino mass together with the systematic parameters that will dominate once statistics stop being the bottleneck. A convex relaxation would also give a certified bound on how far any schedule, including the one actually used, falls short of the best possible.
The stakes are concrete. In the second campaign the m²-uncertainty was 0.29 eV² statistical against 0.17 eV² systematic. Reaching the 0.2 eV goal needs about 1,000 days of data (Direct neutrino-mass measurement with sub-electronvolt sensitivity), and over that span background slope, non-Poissonian background, and energy-loss modelling take over the error budget (Analysis methods for the first KATRIN neutrino-mass measurement). A schedule tuned for counting statistics is the wrong schedule for that regime. Every percent of beam time saved is days of running a tritium facility.
Core Claim
We claim three things. First, the right objective for the long-run KATRIN regime is not plain D-optimality over all fit parameters, which wastes time on well-measured nuisances such as the overall rate. It is a subset criterion: the log-determinant of the information on the neutrino mass squared $m_\nu^2$ together with the poorly calibrated systematic parameters, after the remaining nuisances have been integrated out and with external calibrations entering as prior information. Second, a design that maximizes this criterion, averaged over the plausible range of $m_\nu^2$ and over competing systematic models, will reach a given sensitivity in measurably less time than the heuristic schedule. It will also show smaller bias when the assumed systematic model is wrong. Third, because the time-allocation problem is concave in its continuous form, we can certify an upper bound on achievable information. That bound turns "is KATRIN's schedule near-optimal?" into a number.
Concretely, let $w_j \ge 0$ be the fraction of time spent at candidate voltage $U_j$, with $\sum_j w_j = 1$. Let $I_j(\theta)$ be the Fisher information one unit of time at $U_j$ gives about the full parameter vector $\theta$, and let $I_0$ be the prior information from calibration measurements. The total information is $$M(w,\theta) = I_0 + T\sum_j w_j I_j(\theta),$$ where $T$ is total live time. We maximize the expectation, over a prior $\pi$ on $\theta$ and a set of systematic models, of $\log\det$ of the Schur complement of $M$ onto the parameters of interest. This is the information left on those parameters after the other nuisances are profiled out. For each fixed $\theta$ this objective is concave in $w$, so the continuous problem has a certifiable global optimum. Minimum dwell times and limits on the number of setpoints turn it into the combinatorial subset problem studied in D-Optimal Data Fusion.
Relation to Prior Work
The method comes from the optimal-design literature. D-Optimal Data Fusion: Exact and Approximation Algorithms poses exactly the "existing information $I_0$ plus selected new measurements" log-determinant problem. It proves the problem NP-hard and supplies branch-and-bound with submodular cuts, local search, and randomized rounding. However, it assumes linear models with parameter-independent information, has no nuisance marginalization, and is tested only on power-grid buses. We reuse its solvers for the discrete layer and add the Schur-complement objective and the parameter dependence. Optimal experimental design: Formulations and computations surveys Bayesian expected information gain and its estimators. We use it to justify averaging over a prior to handle the dependence of the design on the unknown parameter, and to decide when a full information-gain estimator is unnecessary. Active sequential estimation (Mukherjee et al.) supplies the idea of re-estimating the parameters between campaigns and re-optimizing. KATRIN's campaign structure (KNM1 to KNM5) makes that natural.
The physics side motivates the problem. The KATRIN analysis papers (Analysis methods for the first KATRIN neutrino-mass measurement; Direct neutrino-mass measurement based on 259 days of KATRIN data) describe the measurement time distribution and how the setup changed across campaigns, but do not optimize it formally. They also report best fits that fall in the unphysical negative region, such as $m_\nu^2 = -0.14 \pm 0.13$ eV², which hints at sensitivity to systematic modelling. The closest work that connects Fisher information and neutrino mass is the BOSS voids analysis (Neutrino Mass Constraint from an Implicit Likelihood Analysis of BOSS Voids). It uses Fisher-based compression only after the data are taken, on a fixed survey, and does not design anything.
One paper constrains the evaluation. Statistical evaluation of experimental determinations of neutrino mass hierarchy shows that Gaussian and Wilks-type reporting badly misstates coverage when the parameter space is bounded or discrete. It reports a nominal 68.27% interval that covers 95.12% of the time. Our design criterion is a local Fisher quantity, and $m_\nu^2$ sits against a physical boundary at zero. We therefore never judge a design by its Fisher value alone. Every comparison is made with full Monte Carlo upper limits and coverage.
Research Program
The first component is a forward model of the integral spectrum. It covers the tritium spectrum, the filter's response, energy loss in the source, background with a slope and non-Poissonian fluctuations, and column density. We calibrate it against the public KNM1–5 release (KATRIN_data_KNM15.json, described in Direct neutrino-mass measurement based on 259 days of KATRIN data). From the model we compute the per-voltage information matrices $I_j(\theta)$ by automatic differentiation. If the release does not include the per-scan voltage and time allocation, we will reconstruct the reference schedule from the published descriptions and state that this is a reconstruction.
The second component is the design solver. The continuous concave problem is solved with Frank–Wolfe iterations, which yields the certified upper bound. A discrete layer imposes minimum dwell times and limits on setpoint count using the data-fusion branch-and-bound and local search. Robustness enters in two ways: a Bayesian average over $m_\nu^2 \in [0, 0.5^2]$ eV² and endpoint uncertainty, and a worst case over a small family of alternative systematic models. These include energy-loss function variants and background-slope shapes.
The third component is a two-stage adaptive variant. The schedule is re-optimized after each simulated campaign using that campaign's posterior.
Baselines
The naive reference point is a uniform allocation over the 40 eV analysis window. It shows how much any deliberate design matters at all.
The strongest existing method is KATRIN's own schedule. The reported figures are a 90% CL sensitivity of 0.7 eV on $m_\nu$ for KNM2 (Direct neutrino-mass measurement with sub-electronvolt sensitivity) and a total uncertainty of ±0.13 eV² on $m_\nu^2$, with an upper limit of 0.45 eV, from 259 days (Direct neutrino-mass measurement based on 259 days of KATRIN data). For a fair comparison, both schedules are run through the same calibrated forward model, at the same total live time, background level, and source activity, with the same fit and the same Monte Carlo procedure for computing limits. We first reproduce the reported sensitivity with the reference schedule to validate the model. Beating it means at least 10% less live time to reach the same median 90% CL limit in the 1,000-day projection, with equal or better coverage.
A second strong competitor is direct numerical minimization of the projected upper limit. It uses simulated data that equal the model prediction exactly (the "Asimov" dataset), at a single assumed parameter value, with known systematics. It gets the same optimizer budget as our method.
Two ablations isolate the key ingredients. Plain D-optimality over all parameters tests whether the subset criterion matters. Our criterion with the systematic priors removed, so that only $m_\nu^2$ is targeted, tests whether self-calibrating systematics is the source of any gain.
Why not just…?
Why not just optimize the projected upper limit directly, as experiments already informally do? This is the real threat. With Gaussian statistics and correctly known systematics, minimizing the projected variance of $m_\nu^2$ is close to our criterion restricted to one parameter. We therefore expect the two to agree when the systematic model is correct, and we will report it plainly if they do. The proposal's value would then rest on two things the direct approach lacks. One is the certified optimality gap. The other is behaviour under a wrong systematic model, measured as bias and coverage when we inject alternative energy-loss or background-slope shapes. If the direct design matches ours there too, the honest conclusion is that KATRIN's practice is already near-optimal, which is itself a useful certified result.
Why not just collect more data or lower the background? These are hardware levers that are complementary to scheduling. The background reduction after vacuum baking in KNM2 (Direct neutrino-mass measurement with sub-electronvolt sensitivity) changes $I_j$, and our design simply re-optimizes for it. In the regime where systematics dominate, more time cannot shrink an error that is not statistical. Only measurements that constrain the systematic parameters can, and choosing those is exactly what the design does.
Why not just use full Bayesian expected information gain with neural estimators? The model has a handful of parameters and Poisson counts. A cheap Laplace-style approximation, averaged over the prior, should be close to the full information gain. On a coarse candidate grid we will compute nested Monte Carlo estimates of the information gain (Optimal experimental design: Formulations and computations) and check whether the optimal schedules differ. If they do near the $m_\nu^2 = 0$ boundary, we switch to that estimator there.
Proposed Experiments
The validation check reproduces the KNM2 sensitivity of 0.7 eV and the 259-day uncertainty of ±0.13 eV² with the reference schedule inside our forward model. If the model cannot do this, nothing downstream is trusted.
The main experiment projects to 1,000 days. It compares, at matched live time, the median 90% CL limit and the time needed to reach 0.2 eV across all baselines and ablations, with limits computed from 10⁴ Monte Carlo pseudo-experiments per design and coverage checked against the physical boundary.
The robustness experiment fits data generated under an alternative systematic model with each design's nominal model. It measures the bias in $m_\nu^2$ and the coverage, and asks whether the robust design reduces the drift toward negative best fits.
The certification experiment reports the gap between the reference schedule's information and the concave upper bound. It also reports the gap between the discrete solution and that bound.
Finally, the adaptive experiment tests whether re-optimizing between campaigns adds anything over a fixed robust design. We expect little gain when $m_\nu^2$ is near zero, because the design depends only weakly on $m_\nu^2$ in that region.
References
- Direct neutrino-mass measurement with sub-electronvolt sensitivity
- Analysis methods for the first KATRIN neutrino-mass measurement
- D-Optimal Data Fusion: Exact and Approximation Algorithms
- Optimal experimental design: Formulations and computations
- Active Sampling of Multiple Sources for Sequential Estimation
- Direct neutrino-mass measurement based on 259 days of KATRIN data
- Neutrino Mass Constraint from an Implicit Likelihood Analysis of BOSS Voids
- Statistical evaluation of experimental determinations of neutrino mass hierarchy