The most energetic cosmic rays arrive at Earth in tiny numbers. The Telescope Array saw only 72 events above 57 EeV in five years. Because these particles are charged, galactic magnetic fields bend their paths, so a real source shows up as a smeared patch on the sky rather than a point. Current searches handle this smearing crudely. They slide a fixed 20° circle across the sky and count events, and then pay a heavy statistical penalty for having looked in many places. This note argues that the magnetic field itself can carry statistical information, and that the most basic fact about magnetic fields, that they have no sources or sinks ($\nabla\cdot\mathbf{B}=0$, "no magnetic monopoles"), is an underused constraint in that inference.
The idea is to treat the unknown galactic field as a nuisance quantity to be fit jointly with source positions, so that events are "un-bent" rather than merely smeared. The field model is then restricted to divergence-free fields by construction. The question is quantitative: how many fewer events does a field-aware, divergence-constrained search need to reach a given significance, compared with top-hat counting and with the same search run on an unconstrained field model? If the constraint removes field configurations that can mimic or fake clusters, it should both sharpen real signals and shrink the effective look-elsewhere penalty. That would be a gain on a fixed exposure, of the kind EXO-200 obtained from better analysis rather than more hardware.
Core Claim
To first order, a nucleus of charge $Ze$ and energy $E$ crossing the Galaxy along direction $\hat n$ is deflected by
$$\boldsymbol{\delta} \approx \frac{Ze}{E}\int \hat n \times \mathbf{B}(\mathbf{x})\, dl ,$$
the line integral of the field component transverse to the path. Any vector field can be split into a divergence-free part and a gradient part $\nabla\phi$. A gradient field generally produces a non-zero transverse line integral, so it deflects particles. An inference that does not impose $\nabla\cdot\mathbf{B}=0$ therefore has extra freedom: one extra scalar degree of freedom per point, which is one of three Fourier components, the one parallel to $\mathbf{k}$. Those configurations can redirect events and are physically impossible.
The claim has three parts. First, in sparse data this extra freedom both inflates the variance of the inferred source position and lets the model "focus" isotropic noise into clusters. Second, removing it by construction raises post-trial significance at fixed event count. Third, the size of the gain can be predicted from how much of the deflection signal the gradient modes can absorb.
This claim can fail in an informative way. If the gradient modes turn out to be nearly orthogonal to what the data constrain, the constraint buys nothing, and the proposal will report that.
The proposal differs from the obvious combination listed in the research file: replacing trilinear interpolation in CRPropa grids with divergence-conforming interpolation. That fix addresses a truncation-level error that is likely small next to the uncertainty in galactic field models. The field-inference setting matters much more, because there the unconstrained degrees of freedom are fit to the data rather than fixed. We run the interpolation question only as a sanity check.
Prior Work and Its Role
Two observational papers set the problem. The Telescope Array hotspot analysis found 19 events where 4.49 were expected. That excess is 5.1σ before correcting for the many sky positions scanned, but only 3.4σ after $10^6$ Monte Carlo trials. The analysis treats magnetic deflection as a fixed angular blur, which is exactly what we replace. CRPropa 3 supplies the forward model: it propagates nuclei through galactic field models and compresses the result into precomputed "lensing" matrices that map source directions to arrival directions. It assumes one fixed field model with no constraint on discrete divergence. We reuse its propagation for simulation and as a fixed-field baseline, but we let the field vary.
Three method papers supply ways to impose the constraint. Neural Conservation Laws builds networks whose outputs are divergence-free by construction (by taking a curl) and shows that soft penalties are hard to tune and leak violations where data are sparse. It was only tested on dense, noise-free benchmarks; we put it under a Poisson likelihood. The vector-potential schemes of Christlieb et al. and the Raviart-Thomas elements of Chandrashekar and Kumar give grid-based alternatives that keep the divergence at about $10^{-12}$.
One paper constrains how we can argue. Balsara et al. show that divergence errors cause crashes and spurious forces in time-evolving MHD. That is a dynamical feedback mechanism, which is absent for a static field traced by test particles. We therefore do not borrow their results as evidence. The benefit in our setting has to come from statistics, not from stability.
EXO-200 motivates the whole program. On the same 234.1 kg·yr exposure, a better analysis improved half-life sensitivity by about 25%. Analysis-side gains can be worth years of running time.
Research Program
We build a hierarchical likelihood for unbinned arrival directions and energies. It has an isotropic background, a small number of sources with unknown positions and fluxes, and a galactic field. The field is a parameterized coherent component plus a turbulent component with a prior on its power spectrum.
The field enters through a differentiable back-tracking step, so source positions and field parameters can be fit or sampled together. We implement three field parameterizations with matched parameter counts:
- Vector potential. $\mathbf{B}=\nabla\times\mathbf{A}$, with $\mathbf{A}$ either a neural network in the style of Neural Conservation Laws or a spectral expansion. This is divergence-free exactly.
- Unconstrained. The same expansion applied directly to $\mathbf{B}$.
- Soft penalty. The unconstrained form plus a term $\lambda\|\nabla\cdot\mathbf{B}\|^2$.
The analysis component works out where the gain should come from. We compute the Fisher information of simulated event sets with respect to the gradient modes. Projecting these modes out predicts a variance reduction for source position. We check that prediction against the measured gain, so the result rests on more than an empirical difference.
Significance is always computed post-trial. We simulate isotropic skies and run the entire procedure on each, including the field fitting. This captures any extra look-elsewhere cost created by a flexible field model.
Baselines
The naive reference is the Telescope Array procedure itself: a Li-Ma excess in a 20° top-hat window, scanned over the sky, with post-trial significance from Monte Carlo. Its reported result on the real 72-event set is 5.1σ pre-trial and 3.4σ post-trial (chance probability $3.7\times10^{-4}$). We re-implement it and re-run it on every simulated sky. We also give it the advantage of scanning window radius, with the scan included in its trial correction, so that it is not handicapped by a fixed radius.
The strongest existing physically informed competitor is a fixed-field search. It applies CRPropa 3 lensing with one published-style galactic field model, then runs a matched-filter scan. This baseline uses physics but does not marginalize over field uncertainty. It should win when the true field equals the assumed one and lose under misspecification. We run it both ways.
The key ablation is the unconstrained-field version of our own method. It has the same likelihood, sampler, parameter count and compute, and differs only in whether $\nabla\cdot\mathbf{B}=0$ holds. A second ablation adds the soft penalty, with $\lambda$ tuned on held-out simulations using the same tuning budget as our method's hyperparameters.
All methods see identical simulated skies. Our method beats the baselines if it reaches 3σ post-trial with at least 20% fewer events than the best baseline on the same injected sources, and also beats the unconstrained ablation by a margin outside the Monte Carlo spread. The 20% threshold is set comparable to the EXO-200 analysis gain.
Why not just…?
Why not just keep the top-hat search and optimize its window? This is a real threat. When the turbulent field dominates, deflections behave like random blur, and a well-tuned blur may be close to optimal. Field-aware methods only help when part of the deflection is coherent and can be learned from the data. The experiments sweep the ratio of coherent to turbulent field strength, so we will report the regime where the top-hat search is as good, rather than assume it does not exist.
Why not just use a fixed best-fit galactic field model? If the model were correct, it would be hard to beat. However, the CRPropa paper itself reports an unresolved mismatch with Auger composition data, and the large 10.3° spread of the hotspot shows that deflections are not well pinned down. We test fixed-field searches under deliberately misspecified fields, where we expect calibration and localization to fail.
Why not just penalize the divergence softly, or project it out cheaply? Neural Conservation Laws shows that soft penalties leak violations where data are sparse, and sparse coverage describes most of the sky here. That evidence comes from fluid benchmarks, so our soft-penalty ablation tests it directly. A cheap Fourier (Helmholtz) projection on a periodic grid imposes the same constraint as the vector potential, so we treat it as an equally valid implementation rather than a competitor. The claim is about the constraint, not about the network design.
Why not just build a bigger detector? Errors shrink only as $1/\sqrt{N}$, so quadrupling the array (TAx4) halves them. Analysis gains multiply whatever hardware delivers, and the two approaches do not compete.
Proposed Experiments
The main experiment uses CRPropa-generated skies of 50 to 500 events above 57 EeV, with mixed composition and with injected sources of varying flux and distance from the Galactic plane. These skies are propagated through ensembles of true fields that are divergence-free by construction and drawn independently of every model's prior. For each method we record three quantities:
- the event count needed to reach 3σ and 5σ post-trial;
- the angular error of source localization;
- the false-cluster rate on isotropic skies.
The decisive check concerns false clusters. If the unconstrained field model produces more significant clusters than the constrained model on isotropic skies, the "noise-focusing" mechanism is confirmed. A second check compares the gain predicted by the Fisher projection with the measured gain across the coherent-to-turbulent sweep. Agreement means we understand the mechanism. Disagreement, or a zero predicted gain, falsifies the core claim.
As a minor side test, we measure how far CRPropa trajectories shift when its trilinear field interpolation is replaced with divergence-conforming interpolation. We expect the shift to be small next to field-model uncertainty.
Finally, we apply all methods to the public 72-event hotspot configuration as described in the Telescope Array paper. We report post-trial significance and the inferred source region as an illustration only, not as a discovery claim, since 72 events cannot support one.
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
- Cosmic ray propagation with CRPropa 3
- Neural Conservation Laws: A Divergence-Free Perspective
- Finite difference weighted essentially non-oscillatory schemes with constrained transport for ideal magnetohydrodynamics
- Constraint Preserving Discontinuous Galerkin Method for Ideal Compressible MHD on 2-D Cartesian Grids
- Physical Constraint Preserving Higher Order Finite Volume Schemes for Divergence-Free Astrophysical MHD and RMHD
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset