[Paper Review] Maximum likelihood for high-noise group orbit estimation and single-particle cryo-EM

Zhou Fan, Roy R. Lederman+3 more|arXiv (Cornell University)|Jul 2, 2021

Advanced Electron Microscopy Techniques and Applications36 references4 citations

TL;DR

This paper develops a maximum likelihood framework for high-noise group orbit estimation in single-particle cryo-EM, analyzing the log-likelihood landscape through Fisher information eigenvalues stratified by algebraic invariants. It proves that third-order moments suffice to locally identify signals up to rotation in SO(2) and SO(3) models, confirming theoretical predictions with empirical noise-scaling validation in low-dimensional protein potential maps.

ABSTRACT

Motivated by applications to single-particle cryo-electron microscopy (cryo-EM), we study several problems of function estimation in a high noise regime, where samples are observed after random rotation and possible linear projection of the function domain. We describe a stratification of the Fisher information eigenvalues according to transcendence degrees of graded pieces of the algebra of group invariants, and we relate critical points of the log-likelihood landscape to a sequence of moment optimization problems, extending previous results for a discrete rotation group without projections. We then compute the transcendence degrees and forms of these optimization problems for several examples of function estimation under $SO(2)$ and $SO(3)$ rotations, including a simplified model of cryo-EM as introduced by Bandeira, Blum-Smith, Kileel, Perry, Weed, and Wein. We affirmatively resolve conjectures that $3^ ext{rd}$-order moments are sufficient to locally identify a generic signal up to its rotational orbit in these examples. For low-dimensional approximations of the electric potential maps of two small protein molecules, we empirically verify that the noise-scalings of the Fisher information eigenvalues conform with our theoretical predictions over a range of SNR, in a model of $SO(3)$ rotations without projections.

Motivation & Objective

• To understand the statistical and geometric structure of maximum likelihood estimation in high-noise group orbit recovery, particularly in single-particle cryo-EM.
• To characterize the Fisher information matrix in terms of transcendence degrees of group invariants, linking statistical estimation to algebraic geometry.
• To resolve conjectures about the sufficiency of third-order moments for local signal identification under SO(2) and SO(3) rotations.
• To validate theoretical predictions on noise-scaling of Fisher information eigenvalues through numerical simulations on low-dimensional protein potential maps.

Proposed method

• Stratifies Fisher information eigenvalues according to transcendence degrees of graded pieces of the algebra of group invariants under SO(2) and SO(3).
• Reformulates critical points of the log-likelihood landscape as a sequence of moment optimization problems.
• Computes transcendence degrees and moment forms for function estimation under SO(2) and SO(3), including simplified cryo-EM models.
• Uses orthogonal Procrustes alignment and continuous multi-reference alignment as analytical tools to study estimation under rotation.
• Employs numerical simulations to evaluate Fisher information eigenvalues across signal-to-noise ratios, comparing with theoretical predictions.
• Applies the Fourier-slice theorem and tomographic projection models to relate cryo-EM imaging to 3D volume reconstruction from 2D projections.

Experimental results

Research questions

• RQ1 Are third-order moments sufficient to locally identify a generic signal up to its rotational orbit in SO(2) and SO(3) group orbit models?
• RQ2 How does the Fisher information matrix structure relate to the algebraic invariants of the rotation group?
• RQ3 What is the noise-scaling behavior of Fisher information eigenvalues in cryo-EM-like models under varying signal-to-noise ratios?
• RQ4 How do theoretical predictions on Fisher information eigenvalues compare with empirical observations in low-dimensional protein potential maps?
• RQ5 Can the log-likelihood landscape be characterized via moment optimization problems in high-noise, rotationally invariant estimation models?

Key findings

• The paper affirms that third-order moments are sufficient to locally identify a generic signal up to its rotational orbit in SO(2) and SO(3) models, resolving a conjecture from prior work.
• Fisher information eigenvalues scale with noise in a manner consistent with theoretical predictions derived from transcendence degrees of group invariants, validated over a range of SNR in numerical simulations.
• The stratification of Fisher information eigenvalues by transcendence degrees of graded invariants provides a novel algebraic-geometric framework for understanding estimation difficulty in rotationally invariant models.
• For low-dimensional approximations of two small protein molecules, the noise-scaling of Fisher information eigenvalues matches theoretical predictions, confirming the model's accuracy in realistic settings.
• The critical points of the log-likelihood function are shown to correspond to solutions of a sequence of moment optimization problems, extending prior results beyond discrete groups.
• Theoretical and numerical results support the use of higher-order moments in maximum likelihood estimation for cryo-EM, even in high-noise regimes where second-order methods fail.