Rotation-Only Procrustes

• Rotation-only Procrustes is a method that determines the best proper rotation between two point sets by minimizing the Frobenius norm under the constraint of SO(p).
• It uses a robust SVD-based closed-form solution that ensures only proper rotations (determinant +1) are considered, avoiding reflections.
• The technique is widely applied in computer vision, neuroimaging, robotics, and molecular alignment, with extensions to quaternion, SDP, and probabilistic formulations.

The rotation-only Procrustes problem, also known as the constrained orthogonal Procrustes or special orthogonal Procrustes problem, concerns finding the optimal rotation that best aligns two datasets—typically point clouds or feature matrices—in a least-squares sense. Restricting the transformation to the special orthogonal group ($SO(d)$) ensures that only proper rotations (orthogonal matrices with determinant $+1$) are considered, excluding reflections. This constraint is central in statistical shape analysis, computer vision, robotics, neuroimaging, and 3D geometric learning, where rigid, orientation-preserving alignment is required.

1. Mathematical Formulation and Properties

Let $X, Y \in \mathbb{R}^{n \times p}$ denote two column-centered datasets. The rotation-only Procrustes problem seeks

$R^* = \arg\min_{R \in SO(p)} \| X R - Y \|_F^2$

where $$SO(p) = \{R \in \mathbb{R}^{p \times p} : R^T R=I_p,\, \det R=1\}$$. This can equivalently be recast as a trace maximization problem: $$R^* = \arg\max_{R \in SO(p)} \operatorname{tr}(R^T X^T Y)$$ due to the invariance of the Frobenius norm and orthogonality constraint (Andreella et al., 2023Lawrence et al., 2019).

The special-orthogonal constraint ensures that the solution represents a rotation without reflection. For $p=3$, $SO(3)$ is the 3D rotation group; the constraint $\det R=1$ ensures that the handedness of space is preserved.

2. Closed-Form Solution: SVD and Maximal Trace Characterization

The classical solution to the rotation-only Procrustes problem is via the Singular Value Decomposition (SVD) of the cross-covariance matrix $M = X^T Y$ (or equivalently $Y^T X$): $$M = U \Sigma V^T,\quad U, V \in O(p),\quad \Sigma=\text{diag}(\sigma_1, \dots, \sigma_p) \succeq 0$$ A maximizer of the trace objective is given by

$R_0 = V U^T$

However, $R_0$ may have $\det(R_0)=-1$ (corresponding to a reflection rather than a rotation). To enforce $\det R^* = +1$,

$$D = \mathrm{diag}(1, ..., 1, \det(VU^T)),\ \ \ R^* = V D U^T$$

This explicit formula ensures $R^* \in SO(p)$ (Andreella et al., 2023Lawrence et al., 2019Bernal et al., 2019Levinson et al., 2020).

The SVD method is robust, numerically stable, and directly generalized to weighted (“Wahba”) or high-dimensional (“Efficient ProMises”) settings (Bernal et al., 2019Andreella et al., 2023).

3. Alternative Solution Methods and Generalizations

While the SVD approach is standard, several alternative techniques exist, particularly in low dimensions:

• Quaternion-based eigenproblem (K method): Especially for $d=3$, Wahba’s problem can be reformulated as finding the unit quaternion maximizing a quadratic form $q^T K q$, where $K$ is a symmetric $4 \times 4$ matrix derived from data covariances. This reduces to an eigenproblem for $K$ (Hanson, 2018Sjogren, 2020).
• Cayley-Newton iteration: A root-finding approach suitable for generating rotations from skew-symmetric matrices via the Cayley transform, combined with Newton’s method (Bernal et al., 2019).
• Semidefinite programming (SDP) relaxations: For Procrustes variants with extra constraints (e.g., partial target, side constraints), the problem can be lifted to a rank-constrained SDP. Relaxations are often tight; the unique minimizer is attained if the cross-covariance has full rank and generic spectrum (Fulová et al., 2023Ling, 2021).
• Probabilistic or weighted alignments: Recent work admits soft correspondence weights (as in probabilistic Procrustes mapping), solved by weighted Kabsch–Umeyama algorithms via weighted SVD, supporting large-scale or noisy applications (Cheng et al., 24 Jul 2025).

Table 1: Summary of Solution Methods

Method	Dimension & Features	Complexity	Notable Properties
SVD/Kabsch–Umeyama	General $d$	$O(d^3)$	Numerically robust, closed-form
Quaternion Eigenproblem	$d=3$	$O(1)$	Analytical roots via quartics
Cayley–Newton	$d=3$	$O(1)$	Efficient for small matrices
SDP Relaxation	General $d$ + constraints	$\geq O(d^6)$	Supports side constraints
Probabilistic/Weighted	General $d$, large-scale, noisy problems	$O(d^3)$	Robust to outliers, scalable

4. Theoretical Properties: Optimality, Uniqueness, and Geometry

The rotation-only Procrustes solution is optimal in the following senses:

• Least-squares optimality: $R^*$ uniquely minimizes the Frobenius error for all $SO(d)$ alignments assuming the cross-covariance is full-rank and singular values are distinct (Levinson et al., 2020Lawrence et al., 2019).
• Maximum-likelihood interpretation: Under isotropic i.i.d. Gaussian noise, $R^*$ is the MLE for the unknown rotation (Levinson et al., 2020Ling, 2021).
• Geometric interpretation: The map $M \mapsto R^*$ is the orthogonal/Frobenius-projection onto $SO(d)$ (i.e., polar decomposition for $d \times d$ matrices) (Levinson et al., 2020Brégier, 2021).
• Uniqueness: The optimizer is unique except in degenerate cases (repeated singular values or zero singular values), in which a continuum of solutions is possible (Lawrence et al., 2019Levinson et al., 2020).
• Metric connection: The Procrustes residual $\|X R^* - Y\|_F$ equals the minimal achievable discrepancy under rotation.

Recent work also provides an eigenvalue-based characterization of matrices achieving maximal trace over rotations: such a matrix must be symmetric and have at most one negative eigenvalue whose magnitude does not exceed the others (Bernal et al., 2019).

5. Rotational Distance Metrics and Residual Analysis

Rotation-only Procrustes solutions yield two canonical between-matrix distances (Andreella et al., 2023):

• Rotational-based distance:

$$d_R(X_i, X_j) = \| R_i - R_j \|_F = \sqrt{2p - 2\,\mathrm{tr}(R_i^T R_j)}$$

where $R_i, R_j$ are the optimal rotations aligning $X_i, X_j$ to a common reference.

• Residual-based Procrustes distance:

$d_\mathrm{Re}(X_i, X_j) = \|X_i R_i - X_j R_j\|_F$

$d_R$ quantifies the difference of orientation prior to alignment; $d_\mathrm{Re}$ captures differences post-alignment. The choice is application-dependent: $d_R$ is preferred for comparing topographic orientation, $d_\mathrm{Re}$ for shape analysis or object coincidence (Andreella et al., 2023).

For multiple point clouds, the generalized rotation-only Procrustes reduces to joint minimization over rotations, and recent results provide tight recovery bounds and convergence guarantees for high SNR settings (Ling, 2021).

6. Computational Aspects and Applications

The SVD-based approach scales efficiently for moderate $d$ and is well supported across scientific computing platforms (e.g., LAPACK, PyTorch batch SVD). For large $d$ or when additional constraints are imposed, manifold optimization or rank-constrained SDP techniques are utilized, though computational cost increases rapidly with problem size (Fulová et al., 2023Ling, 2021).

Rotation-only Procrustes algorithms are foundational in:

• Molecular and structural biology: Protein and molecular alignment, RMSD computation (Hanson, 2018).
• Computer vision and 3D reconstruction: Multi-view registration, pose estimation, and camera calibration (Cheng et al., 24 Jul 2025).
• Neuroimaging: Alignment of fMRI or cortical data for group analysis (Andreella et al., 2023).
• Deep learning on manifolds: Differentiable mapping onto $SO(3)$ as neural network layers, with documented benefits over quaternions or axis-angle parameterizations (Brégier, 2021).

Probabilistic (weighted) variants are now deployed for scalable 3D point-cloud registration in unposed settings with outlier rejection (Cheng et al., 24 Jul 2025).

7. Extensions and Generalizations

The rotation-only Procrustes problem framework extends to:

• Weighted and probabilistic matchings: Probabilistic Procrustes with entropy-regularized soft correspondences enables robust alignment under uncertainty and outlier presence. Closed-form update steps alternate between weighted alignment (via SVD) and soft assignment (Cheng et al., 24 Jul 2025).
• Non-Euclidean geometry: In hyperbolic space, after appropriate centering (hyperbolic centroid), pure rotation alignment reduces to a maximization trace formula over $O(d)$, again solvable via SVD of a weighted cross-covariance (Tabaghi et al., 2021).
• Constrained/relaxed settings: The problem admits conic and semidefinite relaxations, permitting incorporation of additional linear, quadratic, or norm constraints, and supporting non-Frobenius norms and side-constraints (Fulová et al., 2023).

The rigid mapping, maximal trace characterization, and robust, efficient computational methods lend the rotation-only Procrustes problem its central role in high-dimensional statistical estimation, signal processing, and shape analysis.