Capra recovered the main large-scale structures in the tested IBEX-Hi maps, including the Ribbon and a broad enhancement centered toward the map's nose. The evaluation also reported close preservation of the tested rates, intensities and reconstructed signal.
The evaluation used binned IBEX-Hi event-counting data. The input comprised 40 orbit arcs, 360 scan-angle bins per arc and six ESA energy channels; comparisons used baseline IBEX maps and THESEUS reconstructions derived from the same orbit-arc data.
It is arXiv version 1, dated 25 Aug 2026, and no journal publication is reported. It does not evaluate actual IMAP observations, so the results describe Capra's handling of the tested IBEX data rather than its performance on IMAP measurements.
How Capra builds its maps
Capra works directly on a HEALPix pixel grid for the sky maps. It produces a baseline map assigned by boresight, the observation's pointing direction, alongside an optional smoothed map.
To form its background-subtracted signal, the pipeline multiplies the background rate by exposure time to estimate background counts, then subtracts that estimate from the observed counts. The resulting signal is one of the quantities checked in the preservation tests.
All Capra maps shown in the reported results used the Gaussian smoothing kernel described in the paper, although smoothing is optional in the pipeline.
The broad pattern held across reconstructions
The Capra and THESEUS reconstructions placed the dominant large-scale structures in the same locations. Capra retained more pixel-scale variation, while the THESEUS reconstruction was smoother.
That comparison does not establish finer physical angular resolution or show that a smoother map is more accurate. Every displayed Capra map used Gaussian smoothing, even though smoothing is optional in the method itself.
Numerical checks and a working resolution
At the tested HEALPix resolution settings, labeled Nside = 64 and Nside = 16, preservation diagnostics reported rate and intensity normalization coefficients of approximately 1. Pixel-wise rate and intensity differences were typically no greater than 10^-11, and signal-reconstruction differences were below 10^-6 counts.
Convergence, meaning whether a finer map changes the result materially, was assessed with the 16-to-64 and 64-to-128 pairs. RMSDiff, a measure of average map difference, was 0.063 for the first pair and 0.033 for the second; Pearson correlation, a measure of pattern agreement, was 0.989 and 0.997, respectively. Only the 64-to-128 pair met the adopted 5% RMSDiff threshold, although both correlations exceeded 0.98.
The authors considered Nside = 64 a sufficient working resolution for intensity-map analysis of the tested IBEX-Hi dataset. The conclusion is limited to that dataset and the stated criterion: the assessment used only two map pairs, and the 5% cutoff was an operational criterion rather than a physical validation standard.
Benchmark results by kernel setting
At Nside = 64, the benchmark recorded 2,882.2 seconds with one kernel and 235.7 seconds with 16 kernels. The reported speedup at 16 kernels was 12.23, with 76.4% efficiency; at eight kernels, speedup was 7.58 and efficiency was 94.8%.
With one kernel, the runtime ratio for Nside = 64 versus Nside = 16 was approximately 15.4, consistent with O(Npix) scaling, or runtime scaling with the number of map pixels. Peak memory stayed near 0.15 GB at Nside = 16 and 0.25 GB at Nside = 64.
The paper states that the code has been made open-source. The reported benchmark is specific to the selected IBEX dataset and the hardware setup used for the test; it does not by itself establish performance across other datasets or hardware.
The test's boundaries
Beyond the smoothing and resolution choices, the evaluation does not show improved physical angular resolution beyond the instrument's sampling or point-spread characteristics. It also does not show that visually smoother maps are more accurate or validate Capra on diverse datasets.
The convergence assessment used only two map pairs, and no formal confidence intervals are reported for that comparison. Open questions include how alternative kernels, including no smoothing, would affect the maps, how uncertainty propagation and count statistics should be incorporated, and whether the convergence result remains stable across more resolution levels.
Paper data and sources
Original title: Capra: Scalable HEALPix-Native Intensity Reconstruction for High-Resolution IMAP Analyses
Authors: Nikola Bukowiecka, Daniel B. Reisenfeld, Maciej Bzowski
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text