When a detector records only a few dozen events, the question "is there a cluster in the sky?" is limited by counting statistics rather than by the instrument. The ultra-high-energy cosmic-ray hotspot reported by Telescope Array is the textbook case. It rests on 72 events. Nineteen of them fell in a region expecting 4.49 from background. The 5.1σ local excess dropped to 3.4σ once the authors corrected for how many places and window sizes they had searched (Abbasi et al.). That search, like most in this regime, counts events inside a fixed 20° circular window. A circle cannot follow an elongated or irregular source. The obvious fix is to let the analysis choose the region's shape. The danger is that a flexible shape can fit any random clump, so the significance correction grows until the gain disappears.
This note proposes a way to test whether flexible shapes help at all. We fit a region of free shape directly to the exact event coordinates (an "unbinned" likelihood, meaning events are not first sorted into histogram bins). The fit penalizes the region's boundary length using a penalty that treats every direction equally. We then calibrate the full shape-fitting procedure against 10⁵–10⁶ simulated background-only skies, run in batches on a GPU. We also make one sharper claim. In this setting the stair-step bias of pixel grids is not cosmetic. When boundary length is counted along lattice edges, diagonal boundaries cost more than axis-aligned ones. A detector built that way should therefore find sources more easily when they happen to line up with the grid axes. We call this "metrication error": a bias in detection power that depends on the source's orientation relative to the grid. We predict it is measurable at the sample sizes of real rare-event searches, and that an isotropic continuous penalty removes it.
Core Claim
We make three claims, and each can fail on its own. First, a hotspot search built on an unbinned Poisson likelihood with a continuous boundary-length penalty has higher power, after full look-elsewhere calibration, than circular-window counting and than parametric single-blob fits. This should hold for elongated or irregular sources at N ≈ 72, while roughly tying the parametric fit on compact round sources. Second, replacing the continuous penalty with a lattice-edge penalty makes detection power depend on source orientation. Third, calibrating the entire optimizer rather than a fixed statistic requires running it on each simulated sky. That becomes practical only because the solver is a local stencil computation that parallelizes across pixels and across simulated skies.
The three input directions do not play equal roles. Statistical scarcity sets the regime, and the look-elsewhere correction is the arbiter of success. Metrication error is the mechanism under test. Parallelization is the constraint that makes honest calibration feasible, not a contribution in itself.
The Objective
Let $x_1,\dots,x_N$ be event arrival directions and $b(x)$ the expected background intensity. The background includes detector exposure; we use exposure here to mean the detector's direction-dependent sensitivity. Let $u(x)\in[0,1]$ be a relaxed indicator of the source region and $a\ge 0$ the fractional excess inside it. We maximize a Poisson log-likelihood in which each event contributes $\log b(x_i)(1+a\,u(x_i))$. The fitted $u$ is evaluated at each event's exact coordinate by interpolation, not by counting events per pixel. From that likelihood we subtract a boundary-length penalty. Equivalently, we minimize
$$ E_a(u) = a\int b(x)\,u(x)\,dx \;-\; \sum_{i=1}^{N}\log\big(1+a\,u(x_i)\big) \;+\; \alpha\,\mathrm{TV}(u), $$
where $\mathrm{TV}(u)=\int|\nabla u|\,dx$ is the total variation. Total variation measures boundary length isotropically, and here it is computed with the spherical metric. The weight $\alpha$ sets how much boundary length costs. For fixed $a$, the first term is linear in $u$, the second is convex because $-\log$ is convex, and TV is convex. Each subproblem is therefore convex. We solve it with the primal-dual projected-gradient scheme of Chambolle et al. and scan $a$ on a one-dimensional grid. The test statistic is twice the log-likelihood gain over the background-only fit.
The weight $\alpha$ plays two roles. It smooths the shape, and it controls how many effectively different shapes the search can explore. It therefore directly sets the size of the look-elsewhere penalty. We either fix $\alpha$ in advance or treat the choice of $\alpha$ as part of the search and include that choice inside the calibration.
Relation to Prior Work
Several papers supply the method. Chambolle et al. introduced continuous-TV relaxations of minimal-partition problems. They showed these avoid blocky, axis-locked boundaries and reported a 30x GPU speedup. Their data terms, however, were dense images such as natural photos, MRI and stereo, with no Poisson sparsity and no false-positive control. Koch et al. used continuous max-flow to remove grid-sticking in 3D label propagation, again on dense volumes. We reuse this solver family but swap in an event-level Poisson data term and add null calibration.
Another group of papers motivates the proposal. Ahmed et al. and Zhao et al. show that unbinned likelihoods keep information that counting boxes and histograms throw away. Their variables are one- or two-dimensional kinematic quantities fitted with fixed parametric shapes; none involves a free spatial boundary. Abbasi et al. supplies both the regime and the warning. Even a fixed circle loses 1.7σ to the trial correction.
The following papers constrain the design. Shekhovtsov et al. show that discrete higher-order boundary priors keep a residual "Manhattan effect" and are slow, taking seconds per iteration on a 158×128 image. That is why we do not build the shape prior as a discrete Markov random field. Kenmochi et al. show that digital boundaries change length and connectivity under small rotations. This is exactly the orientation dependence we predict will leak into detection power. Sheikholeslami et al. rely on quantizing space into cells, which is poorly suited to skies where most cells are empty. Finally, the Xinanjiang GPU study shows that a million likelihood evaluations can run in 13.4 s on a GPU cluster, a 251.37x speedup over serial. That makes large null ensembles plausible, although their model has no spatial structure.
Research Program
We will build a batched GPU solver that holds thousands of independent simulated skies in memory at once, each on the same sky grid. The pixel-local primal-dual updates run across pixels and skies at the same time. Only a per-sky reduction is needed to evaluate the likelihood. Event interpolation is gathered per sky, which is cheap because each sky contains only about 72 events.
Background skies are drawn from the isotropic distribution weighted by exposure, the same null used by Abbasi et al.. The analysis runs the complete optimizer on each of them, including any scan over $a$ and $\alpha$. It then converts the observed test statistic into a post-trial p-value from the resulting empirical distribution.
The orientation study reuses the identical objective with TV replaced by an unweighted 4-, 8- and 16-neighbor graph length. In this way metrication is the only thing that changes between the two versions.
Baselines
The naive reference point is the published circular top-hat search: Li–Ma significance in a 20° sliding window, calibrated against isotropic simulations. On the Telescope Array data it gives a pre-trial 5.1σ, a post-trial 3.4σ and a chance probability of 3.7×10⁻⁴ (Abbasi et al.). We will reimplement it and first confirm that we reproduce these figures on our null ensemble. That check validates our exposure model before any comparison is made.
The strongest competitor is a parametric unbinned fit. It uses one symmetric Gaussian-like blob with free center, width and amplitude. We extend it to an elliptical blob with free orientation. This is the natural expert choice. The published 10.3°±1.9° Gaussian spread shows that practitioners already think in these terms. For compact sources it is close to optimal. For the fit's significance on the Telescope Array sky there is no reported result found on Lacuna, so we will compute it ourselves under the same calibration.
The third baseline is a multi-radius, multi-ellipse window scan. It gives shape flexibility at low cost, and it is our most direct test of whether continuous boundaries add anything beyond a richer menu of fixed shapes.
The key ablation is the same TV objective with lattice-edge length. It isolates isotropy as the only varying ingredient. We run this ablation at matched $\alpha$ and at matched null false-positive rate.
Fairness rules apply to all methods. Every method sees the same simulated and real skies, the same exposure map and the same null ensemble size. Every free choice (radius set, ellipse grid, $\alpha$ set) is included in that method's own trial calibration, and we match wall-clock budget per null sky wherever the method permits. Beating a baseline means higher detection probability at a fixed calibrated false-positive rate of 3.7×10⁻⁴ or lower. The gap must exceed the Monte Carlo error bars on the injected-source suites.
Why not just…?
Why not just scan more window radii and ellipse orientations? This is the most serious threat. For sources that really are elliptical it may match our method, because an ellipse scan explores fewer effective shapes and pays a smaller trial penalty. The baseline suite includes exactly this scan. The deciding evidence is the injected-source suite with non-elliptical morphologies: crescents, two-lobed shapes and curved filaments. If the ellipse scan ties us there as well, we will report that free-form shape priors are not worth their trial cost at N ≈ 72. That would be a useful negative result, and it would bound what morphology-agnostic methods can achieve in this regime.
Why not just fit a parametric source model with an unbinned likelihood? Ahmed et al. show this beats counting boxes when the source shape is known. We expect it to win or tie on round, compact injections, and we pre-register that expectation. It cannot adapt to an unknown morphology, which is the case that motivates searches along extended structures. The morphology sweep will show where the crossover lies, if there is one.
Why not just use a finer grid or more graph neighbors? Refining the grid does not change the ratio between the length of a staircase path and the true diagonal length; only adding neighbor directions does. Shekhovtsov et al. report residual axis bias even with large neighborhoods, at high computational cost. Our 4/8/16-neighbor ablation measures directly how much orientation-dependent power each option leaves behind.
Why not just run the null ensemble on a CPU cluster? With 72 events per sky, this could work. The cost is dominated by thousands of grid iterations per sky, repeated over the $a$ and $\alpha$ scans, rather than by the events themselves. We will report skies per second on multithreaded CPU versus batched GPU. If the CPU is adequate at 10⁵ skies, the parallelization contribution reduces to engineering, and we will say so. The GPU matters only for pushing direct calibration toward p-values around 10⁻⁶. The CPU-versus-GPU gap reported for the Xinanjiang workload, 3.93x–6.36x, suggests the GPU saving is real but not decisive at 10⁵ skies.
Proposed Experiments
The main suite is synthetic. We simulate skies with N = 72, drawn from the Telescope Array exposure and an isotropic background, and inject sources contributing 10–25 events each. Source shapes are round Gaussians, 4:1 ellipses, curved filaments and crescents. Each shape is rotated in 15° steps from 0° to 90° relative to the grid axes.
Four checks settle the claims, in order. The first is null validity: the empirical false-positive rate on held-out isotropic skies must match nominal levels for every method. The second is post-trial power: we plot power as a function of injected event count for each morphology. The third is orientation: we test whether power for the lattice-penalty ablation varies with rotation angle while power for isotropic TV stays flat within Monte Carlo error. A flat curve for the lattice version would falsify the claim that metrication turns into a statistical bias. The fourth is throughput: skies calibrated per second, which determines how small a p-value we can reach without extrapolating.
As a final step, we apply every method to the Telescope Array event sample described in Abbasi et al.. If the arrival directions are not accessible, we instead use simulated skies matched to its published excess. For each method we report the post-trial significance of the hotspot region and the fitted shape. We fix all analysis choices before looking and make no discovery claim. The purpose is to see whether a shape-flexible, honestly calibrated search strengthens or weakens the 3.4σ indication.
References
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- A convex relaxation approach for computing minimal partitions
- Multi-Atlas Segmentation Using Partially Annotated Data: Methods and Annotation Strategies
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Dalitz plot analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msubsup><mml:mi>D</mml:mi><mml:mi>s</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
- Curvature Prior for MRF-Based Segmentation and Shape Inpainting
- Digital shapes, digital boundaries and rigid transformations: A topological discussion
- WaveCluster: A Multi-Resolution Clustering Approach for Very Large Spatial Databases
- Massively Parallel Monte Carlo Sampling for Xinanjiang Hydrological Model Parameter Optimization Using CPU-GPU Computer Cluster