Fast Volume Alignment by Frequency-Marched Newton

Abstract

We develop a fast and accurate method for 3D alignment, recovering the rotation and translation that best align a reference volume with a noisy observation. Classical matched filtering evaluates cross-correlation over a large discretized transformation space; we show that high-precision alignment can be achieved far more efficiently by treating pose estimation as a continuous optimization problem. Our starting point is a band-limited Wigner-$D$ expansion of the rotational correlation, which enables rapid evaluation and efficient closed-form gradients and Hessians. Combined with analytical control of the complexity of trigonometric-polynomial landscapes, this makes second-order optimization practical in a setting where it is often avoided due to nonconvexity and noise sensitivity. We show that Newton-type refinement is stable and effective when initialized at low angular bandwidth: a coarse low-resolution $\mathrm{SO}(3)$ search provides robust candidates, which are then refined by iterative frequency marching and Newton steps, with translations updated via FFT in an alternating scheme. We provide a deterministic convergence guarantee showing that, under verifiable spectral-decay and gap conditions, the frequency-marching scheme returns a near-optimal solution whose suboptimality is controlled by the Newton tolerance. On synthetic rotation-estimation benchmarks, the method attains sub-degree accuracy while substantially reducing runtime relative to exhaustive $\mathrm{SO}(3)$ search. Integrated into the subtomogram-averaging pipeline of RELION5, it matches the baseline reconstruction quality, reaching local resolution at the Nyquist limit, while reducing pose-refinement time by more than an order of magnitude.

Figures (8)

• Figure 1: Evolution of a 1D correlation landscape under progressive bandwidth expansion. Each curve corresponds to a different bandwidth, with lower bandwidths producing smoother objectives and typically fewer local maxima. A coarse exhaustive search (evaluated on a discretization depicted in green) at low bandwidth identifies local maxima. As the bandwidth increases, these local maxima are tracked across scales using frequency marching, while local optimization refines the solution at each stage. This strategy avoids exhaustive search on highly oscillatory high-frequency landscapes while achieving high-precision alignment. For physical interpretation of low-pass filtering the correlation, we display correspondingly low-pass filtered copies of a particle on the left.
• Figure 2: One-shot versus frequency marching on a schematic 1D correlation landscape. Left: jumping directly from $L_0$ to $L_J$, Newton's method can converge to a spurious local maximum that emerged at high bandwidth. Right: under the conditions of Theorem \ref{['thm:multiband_correct']}, marching through intermediate bandwidths keeps the tracked maximizer in the correct basin at each level; Newton refines within the basin before proceeding to the next bandwidth.
• Figure 3: Slice of a ribosome volume corrupted by Gaussian noise at the indicated SNR levels.
• Figure 4: Hyperparameter study for Matcha and SOFFT at $\mathrm{SNR}=0\,\mathrm{dB}$. In (a), dot size encodes alignment error; Matcha achieves $0.03^\circ$ accuracy in seconds, whereas SOFFT at its highest feasible oversampling ($K=18$) requires over a hundred seconds for $0.13^\circ$. In (b), increasing $L_{\max}$ improves both methods, but the effect is more pronounced for SOFFT; beyond $L_{\max}=100$, SOFFT becomes memory-prohibitive at large $K$. In (c), a single Newton step matches the accuracy that gradient ascent reaches only after several iterations, supporting the use of one Newton step per band in (a) and (b).
• Figure 5: Subtomogram averaging on the Chlamy dataset (EMPIAR-11830) using RELION5. Matcha replaces 3D auto-refine at the bin2 and bin1 stages, matching the baseline resolution ($3.8$ Å at Nyquist) while reducing cumulative alignment time by over $10\times$. Intermediate processing steps (CTF-refinement, classification, template updates) are shared between both pipelines and excluded from the reported timings.

Theorems & Definitions (14)

• Definition 3.1
• Lemma 3.1: Set-gap stability
• proof
• Definition 3.2
• Lemma 3.2: Second-order two-sided bounds from a critical point
• proof
• Lemma 3.3: Basin persistence under a $C^2$ perturbation
• proof
• Theorem 3.4: Main result
• proof