Xoryn II - Exact Image Recovery and Sharp Measurement Bounds for Projected Multi-Reference Alignment on Finite Grids
Open problem solved. In Section 5.2 of Projected Multi-Reference Alignment (arXiv:2605.25533), Balanov, Katz, Bendory, and Edidin ask which projections preserve low-order information about the original orbit. Xoryn II resolves a concrete finite-grid acquisition case. Given independently and uniformly translated views of a real periodic image, can a sensor retain the Fourier coordinates that are even under point inversion, add a fixed number of scalar linear measurements per view, and identify the translation orbit from population moments through degree three? Original open question Exact recovery and minimum measurement count. For every real image with full non-DC Fourier support, one fixed sine measurement per folded view suffices on any finite rectangular grid, including even side lengths and Nyquist frequencies. One is the minimum extra count within this acquisition family whenever a nonreal Fourier mode exists. The decoder uses O(N) selected cubic products for an N-pixel image and linear-time reconstruction once the moments are available. The resulting representation distinguishes point-inverted images whenever they lie in different translation orbits. Full cubic tensor and missing frequencies. Two fixed sine measurements suffice to recover the entire unprojected third-order moment tensor on finite abelian grids, including images with spectral zeros. On p × p grids for odd primes p ≥ 5, two are necessary for linear transfer of the full tensor. An integer-lattice criterion also determines the minimum separating moment order for a given Fourier support. The worst-case order on odd prime-square grids is p, and one coded sine measurement with a known mask that varies between views can recover moments through that order. An O(N log N) normalized bispectral descriptor has an inverse stability bound independent of the side lengths in two dimensions. Reproducible evidence. In noiseless tests with simulated acquisition, all 1,797 recorded 8 × 8 handwritten digits were recovered up to translation. The 1,678 full-support digits used 35 of 64 scalar readings per view, and the 119 with spectral zeros used 36. Two measurements shared across all color channels recovered 12 natural 32 × 32 RGB crops with a full-support channel mixture, using 1,544 of 3,072 readings per view. In a separate balanced test of known templates against their point inversions, at 128 views and Gaussian noise standard deviation 1 per coordinate, the one-measurement augmentation achieved 98.8% accuracy. The full-view cubic reference achieved 99.6%, and the folded-only measurements had a theoretical accuracy ceiling of 50%. The result gives a concrete design for preserving translation-invariant image information during acquisition on finite periodic grids. The companion materials include the proofs, reference implementation, frozen benchmark inputs, and numerical results.
Authors
- Maciej Nowicki
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24347473
- Primary Topic
- Mathematical Analysis and Transform Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00