Weighted Procrustes Method

• Weighted Procrustes method is a technique for aligning datasets or operators using weighted transformations under orthogonal, affine, or linear constraints.
• The approach provides closed-form solutions, iterative algorithms, and convex relaxation techniques that enhance accuracy in point cloud registration and statistical shape analysis.
• Recent developments include fully differentiable solvers for deep learning and probabilistic relaxations that robustly handle large-scale, noisy data.

The weighted Procrustes method is a family of techniques for optimal alignment of vector sets, operators, or matrices under orthogonal, affine, or more general linear transformations, subject to weighting of individual correspondences or rows. Originating from classical shape analysis and statistical alignment, the weighted Procrustes problem now encompasses operator theory in Hilbert spaces, robust computer vision pose estimation, and modern conic optimization frameworks. Recent developments include fully differentiable solvers for deep learning pipelines and probabilistic relaxations for large-scale geometric estimation.

1. General Formulation of the Weighted Procrustes Problem

The weighted Procrustes problem seeks a transformation (often rigid, affine, or orthogonal) that best aligns two data sets, matrices, or operators, assigning a possibly non-uniform weight to each correspondence or data element.

• Finite-dimensional, matrix setting: Given $A \in \mathbb{R}^{p \times m}$, $B \in \mathbb{R}^{p \times n}$, and $W \in \mathbb{R}^{p \times p}$ (usually diagonal, $W \succeq 0$), the weighted Frobenius-norm Procrustes formulation is:

$$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$

subject to structural constraint $X \in C$, e.g., orthogonality or obliqueness (Fulová et al., 2023).

• Operator-theoretic (Hilbert-space) formulation: Given $A \in L(\mathcal{H})$ with closed range, $B \in L(\mathcal{H})$, and positive $W \in L(\mathcal{H})^{+}$ with $W^{1/2} \in S_p$ ($B \in \mathbb{R}^{p \times n}$0-Schatten class), the problem is:

$B \in \mathbb{R}^{p \times n}$1

(Contino et al., 2016).

• Registration/alignment version: Given paired points or keypoints $B \in \mathbb{R}^{p \times n}$2 (possibly in $B \in \mathbb{R}^{p \times n}$3), with weights $B \in \mathbb{R}^{p \times n}$4, the rigid weighted Procrustes objective is:

$B \in \mathbb{R}^{p \times n}$5

(Wei et al., 27 Feb 2025, Cheng et al., 24 Jul 2025).

2. Mathematical Solutions and Characterization

2.1. Closed-form Solution (Rigid/Orthogonal Transform, Euclidean Case)

For point clouds or keypoint alignment with non-negative weights:

• Compute weighted centroids:

$B \in \mathbb{R}^{p \times n}$6

with $B \in \mathbb{R}^{p \times n}$7.

• Compute centered, weighted cross-covariance:

$B \in \mathbb{R}^{p \times n}$8

• Obtain the optimal rotation via SVD:

$B \in \mathbb{R}^{p \times n}$9

The reflection-correction via the determinant enforces $W \in \mathbb{R}^{p \times p}$0 (Wei et al., 27 Feb 2025, Cheng et al., 24 Jul 2025).

• The translation is $W \in \mathbb{R}^{p \times p}$1.

2.2. Weighted Least-Squares and Operator-theoretic Characterization

When $W \in \mathbb{R}^{p \times p}$2 and $W \in \mathbb{R}^{p \times p}$3 are bounded operators and $W \in \mathbb{R}^{p \times p}$4 a positive operator:

• The minimum exists if and only if $W \in \mathbb{R}^{p \times p}$5, with $W \in \mathbb{R}^{p \times p}$6.
• The solution is characterized as the $W \in \mathbb{R}^{p \times p}$7-inverse of $W \in \mathbb{R}^{p \times p}$8 in $W \in \mathbb{R}^{p \times p}$9 and satisfies the normal equation:

$W \succeq 0$0

The minimum-norm solution (when $W \succeq 0$1 invertible) is $W \succeq 0$2 (Contino et al., 2016).

2.3. Numerical/Optimization Algorithms

Approach	Setting	Complexity/Remarks
SVD (closed form)	Rigid, d=2,3	$W \succeq 0$3 keypoints, $W \succeq 0$4 SVD per instance
Normal equations	General least-squares	Solve $W \succeq 0$5; Cholesky/QR direct solvers
Conic SDP reform.	Structured $W \succeq 0$6 constraints	$W \succeq 0$7 per SDP iteration; practical for moderate $W \succeq 0$8
Iterative/EM variant	Probabilistic coupling	Alternates E-step (weights) and M-step (transf.)

For deep learning or large-scale cases, differentiability with respect to weights and locations (enabled by autodiff through SVD) is essential (Wei et al., 27 Feb 2025). For outlier-robust cases, iterative reweighting or soft assignment (EM-type) algorithms are effective (Cheng et al., 24 Jul 2025).

3. Generalizations: Constraints, Weighting Schemes, and Conic Optimization

Weighted Procrustes problems admit a wide range of generalizations:

• Structured constraints: The transformation $W \succeq 0$9 or $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$0 may be restricted to lie on a Stiefel manifold (orthogonality), be oblique (diagonal of $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$1 is 1), or subject to other algebraic constraints (Fulová et al., 2023).
• Conic SDP relaxation: Many structured versions can be reformulated as rank-constrained semidefinite programs, enabling the encoding of additional linear, quadratic, or semidefinite side-constraints on $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$2. The original objective (weighted least squares) is encoded as a linear trace over an auxiliary PSD block. Convex relaxations (dropping the rank constraint) provide lower bounds, while log-det or convex iteration heuristics yield near-optimal, feasible solutions. Practical computation is typically limited to moderate dimensions by cubic cost per SDP iteration (Fulová et al., 2023).
• Operator-norm and alternative norms: While the Frobenius norm is classical, generalization to Schatten $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$3-norms, $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$4 or $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$5 matrix norms, and operator norms is available within the same framework (Contino et al., 2016, Fulová et al., 2023).

4. Differentiable and Probabilistic Variants in Modern Applications

4.1. Differentiable Weighted Procrustes Solver

For integration with deep learning models, all components of the weighted Procrustes pipeline are designed to be fully differentiable, enabling end-to-end gradient-based training:

• Weighted centroids, covariance, SVD, and the final transformation mapping are all differentiable when supported by modern autodiff frameworks (Wei et al., 27 Feb 2025).
• Example: In monocular visual odometry (BEV-DWPVO), keypoints are extracted and weighted by learned confidences; matching and pose estimation is achieved via a single differentiable weighted Procrustes call per frame pair, providing stable, interpretable, and end-to-end scale-anchored odometry without explicit depth supervision.

4.2. Probabilistic Weighted Procrustes (EM/Softweight)

For large-scale or noisy correspondence problems, probabilistic formulations augment the weighted Procrustes objective by allowing the correspondence weights themselves to adapt:

• Weights $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$6 are introduced for each pair, with an entropic regularizer $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$7 and a "dustbin" parameter $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$8 absorbing outlier mass.
• The algorithm alternates between computing soft assignments (E-step) and updating the transformation (M-step) via weighted Procrustes minimization.
• Gradients of the weighted objective with respect to transformation parameters can be computed analytically for efficient optimization (Cheng et al., 24 Jul 2025).

This approach enables robust alignment of point clouds and poses in high-noise or partially ambiguous data, as in joint optimization of 3D Gaussian Splatting (3DGS) and global reconstructions from unposed image sets.

5. Operator-theoretic and Infinite-dimensional Perspectives

In the abstract Hilbert-space setting, the weighted Procrustes problem generalizes classical least-squares approximations:

• The existence and uniqueness of a minimizer are characterized by compatibility conditions between the weight (positive operator $$\text{minimize}_{X \in C} \;\| W^{1/2}(A X - B) \|_F^2$$9) and the image/range of $X \in C$0 and $X \in C$1.
• The solution set is described by $X \in C$2-inverses and is equivalent to the solution of the normal equation $X \in C$3 when the latter exists and is unique on the relevant subspace (Contino et al., 2016).
• When $X \in C$4 is singular or semi-definite, compatibility conditions are stricter, and the Kreĭn–Anderson–Trapp shorted operator is used to express the minimum (Contino et al., 2016).

This abstract framework supports extensions to weighted pseudoinverses, structured low-rank approximation, and $X \in C$5-Schatten class regularizations.

6. Applications and Empirical Evidence

Weighted Procrustes methodology appears extensively in statistical shape analysis, multidimensional scaling, factor analysis, computer vision, geometric point set registration, system identification, 3D scene reconstruction, and data fusion:

• Monocular visual odometry: Differentiable weighted Procrustes ensures stable and scale-anchored metric pose estimation, demonstrated across challenging driving datasets (NCLT, Oxford, KITTI) with state-of-the-art accuracy (Wei et al., 27 Feb 2025).
• 3D Gaussian Splatting and global point cloud registration: Robust probabilistic weighted Procrustes enables accurate global alignment and pose estimation across hundreds of unposed images, handling tens of millions of points and outlier rejection efficiently (Cheng et al., 24 Jul 2025).
• Constrained alignment for multivariate data: Conic optimization methods permit principled simultaneous handling of orthogonality, obliqueness, and other constraints, with documented success in moderate-scale numerical experiments and side-constraint satisfaction within tight residual bounds (Fulová et al., 2023).

Weighted Procrustes algorithms thus form a core component in robust, scalable, and interpretable data alignment and registration across a wide range of computational and applied domains.