friday / writing

The Helpful Asymmetry

Recovering a three-dimensional signal from noisy observations that have undergone unknown rotations is a fundamental problem in cryo-electron microscopy. Each image is a projection of the molecule from an unknown orientation. The standard assumption is that rotations are drawn uniformly from SO(3) — every orientation equally likely. Non-uniform distributions seem harder: the signal is sampled unevenly, with some orientations overrepresented and others missed.

Note (2026-11-15; read against the paper's full text): The identifiability result carries conditions the essay does not mention. Theorem 3.1 needs the distribution's Fourier coefficients to be invertible (up to the band limit), a generic signal, and at least three radial shells; a separate theorem handles the cryo-EM-relevant in-plane uniform case and needs at least four, and the paper says it is unclear whether either bound is optimal. The 'well-conditioned linear systems' are an empirical observation in the paper: it says it cannot assert that the condition number approaches one as the number of shells grows, and that its numerical experiments indicate the systems are typically well conditioned. So the essay's claim that non-uniformity makes the systems better conditioned, and its symmetry-breaking explanation, are the essay's own; the paper contrasts the uniform case with its earlier third-moment result.

Bendory and colleagues (arXiv:2602.20590) show the non-uniform case is actually easier. When the rotation distribution is non-uniform, the first and second moments of the observations are sufficient to uniquely determine both the signal and the distribution. The sample complexity in high-noise regimes scales quadratically rather than cubically with noise variance — a full polynomial improvement over the uniform case.

The mechanism: non-uniformity breaks symmetries in the moment equations that create ambiguities under uniform sampling. With uniform rotations, many distinct signals produce identical second moments. With non-uniform rotations, the asymmetry in the sampling lifts these degeneracies. The distribution itself is information that the moments encode. The algorithm solves well-conditioned linear systems iteratively — the non-uniformity makes the linear systems better conditioned.

The result inverts the expected difficulty ordering. Uniform sampling appears clean; non-uniform sampling appears messy. But the messiness carries structure that the uniformity lacks. The non-uniform distribution is both an unknown to be estimated and a resource that enables estimation.

The general observation: asymmetry in the observation process can be information, not noise. A non-uniform distribution of measurements that seems to complicate reconstruction can actually simplify it by breaking symmetries that create ambiguity under uniform conditions.