Procrustes Post-Processing: Techniques and Applications

• Procrustes post-processing is a technique that aligns data by removing translation, rotation, and, in some cases, scaling to correct non-identifiability.
• It is applied to improve interoperability in embedding models, sensor calibration, and morphometric comparisons while preserving internal geometry.
• Methods leverage SVD-based closed-form solutions and robust convex relaxations to address challenges in high-dimensional and contaminated data settings.

Procrustes post-processing is the downstream alignment of one set of observations, embeddings, landmarks, or factor matrices to another by minimizing a discrepancy after translation removal, orthogonal transformation, and, in some variants, global scaling. In contemporary usage, it functions as a registration layer that removes rotational non-identifiability in singular or eigen decompositions, makes separately trained embedding models interoperable, converts segmentation outputs into pose estimates, stabilizes morphometric comparisons, and can even enforce exact Pearson correlation in synthetic data through a left-orthogonal transform (Jasa et al., 5 Oct 2025, Maystre et al., 15 Oct 2025, Sterzentsenko et al., 2020, Garcia-Perez et al., 2017, Ounissi et al., 2 Oct 2025). Across these settings, the central idea is unchanged: preserve the internal geometry that should remain invariant, while choosing the transformation that minimizes a task-relevant mismatch.

1. Scope and motivation

In random matrix and network settings, Procrustes post-processing removes the inherent rotational non-identifiability of singular or eigen decompositions and latent-position models. If $\hat{X},\hat{Y}\in\mathbb{R}^{n\times d}$ are adjacency spectral embeddings of random dot product graphs, latent positions are identifiable only up to an orthogonal transform, so comparing $\hat{X}$ and $\hat{Y}$ requires Procrustes alignment (Jasa et al., 5 Oct 2025).

In embedding systems, the same issue appears because many training losses depend only on dot products or distances. Two models trained separately on similar data can therefore encode essentially the same geometry in different coordinate frames. Orthogonal Procrustes post-processing estimates a single orthogonal transformation that aligns one embedding space to the other while preserving norms, angles, and pairwise distances inside the source space exactly (Maystre et al., 15 Oct 2025).

In volumetric sensor calibration, Procrustes post-processing is used differently but with the same mathematical core. A segmentation network produces soft class probabilities, these probabilities define soft 3D correspondences, and a small rigid alignment solved by SVD converts those correspondences into an initial sensor-to-structure pose. At inference, the same computation acts as a lightweight post-processing step before dense multi-view ICP refinement (Sterzentsenko et al., 2020).

In geometric morphometrics and machine-learning pipelines built on landmarks, Procrustes alignment is both indispensable and potentially problematic. It is the standard mechanism for removing translation, rotation, and scale before shape analysis, yet generalized Procrustes analysis performed on the full dataset before train–test splitting can induce cross-sample dependency. A leakage-free alternative aligns test specimens only to the training consensus shape (Courtenay, 26 Jan 2026).

2. Core formulations

The classical orthogonal Procrustes problem seeks

$\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$

with $A,B\in\mathbb{R}^{d\times n}$ or $\mathbb{R}^{n\times d}$ depending on convention. Setting $M=AB^\top$ and computing the SVD $M=U\Sigma V^\top$ gives the closed-form minimizer

$Q_F^*=UV^\top.$

If one restricts to $Q\in SO(d)$, the Umeyama sign correction is

$\hat{X}$0

For rigid 3D registration with corresponded points, the Kabsch formulation uses centered point sets, covariance $\hat{X}$1, and

$\hat{X}$2

while the similarity variant adds

$\hat{X}$3

(Jasa et al., 5 Oct 2025, Sterzentsenko et al., 2020).

Weighted and masked versions replace ordinary centroids and covariances by weighted analogues. In the soft Procrustes formulation for sensor alignment, class weights are $\hat{X}$4, soft 3D centroids are

$\hat{X}$5

and the weighted covariance is

$\hat{X}$6

before the same sign-corrected SVD solve (Sterzentsenko et al., 2020).

Several important post-processing variants do not admit such a closed form. The spectral-norm problem,

$\hat{X}$7

is convex in its argument but nonsmooth when the top singular value has multiplicity greater than $\hat{X}$8. The robust mixed $\hat{X}$9 problem,

$\hat{Y}$0

encourages alignment of most columns while tolerating a small number of mismatched or outlier columns, but is nonsmooth even away from multiplicities (Jasa et al., 5 Oct 2025).

A distinct left-orthogonal formulation appears in synthetic tabular data. There, one solves

$\hat{Y}$1

so that the transform acts on samples rather than features. Because $\hat{Y}$2, the column Gram matrix is preserved, and after centering and standardization this yields exact preservation of the Pearson correlation matrix (Ounissi et al., 2 Oct 2025).

3. Geometry, optimization, and surrogate solutions

The feasible set for orthogonal alignment is the orthogonal group $\hat{Y}$3, a Riemannian manifold with two components, $\hat{Y}$4, whose $\hat{Y}$5 component is $\hat{Y}$6. Optimization on this space uses tangent spaces, retractions, the exponential map, and vector transports. For spectral and robust objectives, one derivative-free manifold method used in recent work is Lower-Triangular Mesh Adaptive Direct Search (LTMADS), with poll and search steps defined in tangent space and step-size control constrained by the injectivity radius (Jasa et al., 5 Oct 2025).

A practical theme in current Procrustes post-processing is that the Frobenius solution is often a strong surrogate even when the target norm is spectral or robust. If $\hat{Y}$7 for $\hat{Y}$8 and $\hat{Y}$9 minimizes $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$0, then the surrogate statistic

$\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$1

satisfies

$\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$2

for $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$3. In network testing, this implies conservative surrogate critical values, $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$4, and therefore potentially reduced power; empirically, however, the gap can be very small, with one reported 2D example giving a difference of approximately $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$5 between $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$6 and $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$7 (Jasa et al., 5 Oct 2025).

For robustness to outliers in rigid matching, a different route replaces nonsmooth manifold search by convex relaxation. The symmetrized robust Procrustes formulation introduces convex SOCP relaxations for the power-$\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$8 objective and, after rounding by the nearest orthogonal projection, yields constant-factor guarantees. For $\min_{Q\in \mathcal{O}(d)} \|A-QB\|_F,$9, the paper proves a $A,B\in\mathbb{R}^{d\times n}$0-factor approximation, and under linear or affine dominance-of-inliers conditions it exactly recovers the true orthogonal or rigid motion (Amir et al., 2022).

When the feasible set includes orthogonality, obliqueness, projection-type constraints, positivity, block structure, or partially observed targets, a conic reformulation becomes useful. Rank-constrained semidefinite programs can encode Frobenius, $A,B\in\mathbb{R}^{d\times n}$1, $A,B\in\mathbb{R}^{d\times n}$2, and spectral objectives together with quadratic feasibility constraints through PSD blocks and rank conditions, providing a unified treatment of constrained Procrustes problems beyond the classical SVD regime (Fulová et al., 2023).

4. Statistical, geometric, and machine-learning applications

In random matrix and network inference, norm choice changes operating characteristics. Simulations reported for random dot product graph testing show that diffuse isotropic perturbations favor Frobenius statistics, salt-and-pepper contamination favors the robust $A,B\in\mathbb{R}^{d\times n}$3 statistic, and rank-one alternatives can shift the advantage between Frobenius and robust norms depending on the perturbation angle. Across these regimes, using $A,B\in\mathbb{R}^{d\times n}$4 inside spectral or robust objectives often tracks the optimal statistics closely in power curves (Jasa et al., 5 Oct 2025).

In interoperable embedding systems, Procrustes post-processing is justified by explicit bounds. If

$A,B\in\mathbb{R}^{d\times n}$5

then

$A,B\in\mathbb{R}^{d\times n}$6

A corollary bounds average alignment error by $A,B\in\mathbb{R}^{d\times n}$7 when average squared dot-product deviation is at most $A,B\in\mathbb{R}^{d\times n}$8. Empirically, the same work reports successful use in compatibility across retrainings on MovieLens-25M, cross-model text retrieval on MMTEB tasks, and mixed-modality search on MixBench, with orthogonal alignment consistently preserving source geometry better than unconstrained linear maps (Maystre et al., 15 Oct 2025).

In markerless volumetric sensor alignment, Procrustes operates as both a training regularizer and an inference post-processor. The total loss is

$A,B\in\mathbb{R}^{d\times n}$9

with

$\mathbb{R}^{n\times d}$0

The in-loop geometric term improves robustness in enlarged pose spaces, while post-processing converts segmentation outputs into initial extrinsics for any number of sensors and arbitrary placements, after which ICP refines the poses jointly (Sterzentsenko et al., 2020).

In synthetic tabular data, Procrustes post-processing is used not to resolve non-identifiability but to impose exact second-order structure. After centering and standardization, a left-orthogonal transform of the real-data surrogate matrix preserves the target Gram matrix, so the post-processed synthetic dataset has exactly the same Pearson correlation matrix as the target data while remaining as close as possible in Frobenius norm to the original synthetic matrix. An empirical illustration on the Madeira Island SustData energy-consumption dataset demonstrates exact correlation recovery up to numerical precision (Ounissi et al., 2 Oct 2025).

Transformer compression provides yet another adaptation. In ProcrustesGPT and COMPOT, orthogonal transformations are inserted at layer interfaces or dictionary updates because RMSNorm and residual structures admit invariances under orthogonal rotations. In one case this improves compressibility of pretrained weights within structured matrix classes such as sums of Kronecker products or GS matrices; in the other, thin-SVD Procrustes updates enforce orthogonal dictionaries during calibration-optimized sparse factorization (Grishina et al., 3 Jun 2025, Makhov et al., 16 Feb 2026).

5. Robustness, contamination, and methodological cautions

A persistent misconception is that least-squares Procrustes preprocessing is automatically robust. In morphometric analysis this is explicitly false: sample means and variances make classical Procrustes sensitive to outliers, and a single aberrant landmark can shift the centroid and inflate variance. A robust pipeline replaces mean centroids by column-wise medians or trimmed means, replaces centroid-mean size by

$\mathbb{R}^{n\times d}$1

keeps the classical SVD rotation, and then performs consensus estimation and tangent-space projection. For inference under contaminated normal models, the same work derives Von Mises plus saddlepoint approximations for the Procrustes statistic’s tail probability (Garcia-Perez et al., 2017).

A second misconception is that Procrustes preprocessing is always harmless in machine-learning evaluation. In geometric morphometrics, generalized Procrustes analysis is a global operator: every specimen’s final alignment depends on the sample-wide consensus. If GPA is run on training and test specimens jointly before splitting, both sides of the split influence each other. The leakage-free alternative is fold-specific: run GPA on the training set only, compute the training consensus, and align each test specimen independently to that fixed reference by the closed-form single-specimen Procrustes solution (Courtenay, 26 Jan 2026).

A third caution concerns regularized multivariate analysis. One paper argues that the orthogonal Procrustes step commonly used in iterative regularized PCA, CCA, and OPLS pipelines is not optimal for the overall MVA objective, does not in general preserve uncorrelated extracted features, and can stall at orthogonal initialization when $\mathbb{R}^{n\times d}$2. The proposed replacement is an eigenvalue-based step derived from the stationarity conditions of the full constrained objective, with the explicit goal of preserving diagonal $\mathbb{R}^{n\times d}$3 (Muñoz-Romero et al., 2016).

These results do not imply that Procrustes post-processing is unreliable. They indicate instead that the statistical meaning of the alignment depends on the nuisance structure being removed, the loss being minimized, and the place of the alignment inside the full pipeline.

6. Limitations and open problems

Several limitations recur across the literature. For nonsmooth spectral or robust norms, manifold solvers provide stationarity guarantees but not global optimality, because the objectives may be nonconvex on $\mathbb{R}^{n\times d}$4 and need not be geodesically convex (Jasa et al., 5 Oct 2025). Analytical approximation bounds explaining why $\mathbb{R}^{n\times d}$5 is such a strong surrogate for spectral or robust alignment under random matrix models remain open (Jasa et al., 5 Oct 2025).

Orthogonal alignment also presupposes approximate preservation of pairwise dot products or distances. When two embedding models encode different semantics, exhibit substantial anisotropy, or undergo nonlinear distortion or domain drift, Procrustes alignment may underperform even though it preserves the source geometry exactly (Maystre et al., 15 Oct 2025). In synthetic-data correlation correction, the method enforces only second-order structure, and because it uses the real-data matrix or a surrogate with the same correlation, strict privacy regimes may require replacement by a protected correlation target (Ounissi et al., 2 Oct 2025).

High-dimensional settings introduce additional identifiability issues. The ProMises model addresses the non-identifiability and interpretability problems of high-dimensional Procrustes alignment by imposing a matrix von Mises–Fisher prior on the orthogonal parameter, yielding a “data-plus-prior” SVD solution and an efficient reduced algorithm for $\mathbb{R}^{n\times d}$6 regimes such as neuroimaging (Andreella et al., 2020).

Open directions recorded in the literature include global certification under structural assumptions, nonconvex robust penalties with better behavior under global structured perturbations, scalable constrained solvers for large $\mathbb{R}^{n\times d}$7 and $\mathbb{R}^{n\times d}$8, and more principled integration of Procrustes alignment with downstream statistical testing, calibration, and privacy constraints (Jasa et al., 5 Oct 2025, Fulová et al., 2023, Makhov et al., 16 Feb 2026). Taken together, these directions show that Procrustes post-processing is less a single algorithm than a family of alignment mechanisms whose effectiveness depends on how geometry, optimization, and statistical assumptions are coupled in the application at hand.