Domain Elastic Transform (DET)

• Domain Elastic Transform (DET) is a dual-framework where grid-free Bayesian registration aligns vector-valued functions on irregular domains and elastic ray tomography maps elastic parameters via ray integrals.
• It employs a Bayesian model with Gaussian process priors to infer smooth spatial deformations, integrating both geometric and functional likelihoods for robust registration.
• DET scales to atlas-level data using Nyström approximation and variance-guided importance sampling, while addressing challenges like low overlap and significant topological changes.

Searching arXiv for the two cited papers and closely related context. Domain Elastic Transform (DET) denotes a domain-centered elastic framework in which deformation or measurement is defined on the spatial support of scientific data rather than on a regular grid. In recent arXiv usage, the term has two technically distinct realizations. In one, DET is a grid-free Bayesian method for registering vector-valued functions on irregular point sets by inferring a smooth elastic deformation of the domain from a joint spatial–functional likelihood (Hirose et al., 22 Mar 2026). In the other, the elastic ray transform can be read as a DET for elastic wave tomography because it maps elastic parameter fields on the domain to ray-based line-integral data indexed by directions and polarizations, and at rank $2$ it corresponds to linearization of travel time of elastic waves, measured for all polarizations (Ilmavirta et al., 25 Feb 2025). The common structure is that the codomain records either correspondences or ray data induced by an elastic model acting on domain-supported fields, but the two settings address different inverse problems.

1. Scope of the term

The DET literature currently spans both nonrigid registration of irregular scientific data and tensor tomography for elastic media. The distinction is not terminological only; it reflects different mathematical objects, likelihoods, and identifiability statements.

Usage in recent literature	Domain	Output
Bayesian DET	Point sets with attached feature vectors $(y_m, f_Y(y_m))$	Deformation, correspondences, and noise parameters
Elastic-ray DET	Elastic tensor fields $f : \mathbb{R}^n \to \#2{n}{m}$	Ray integrals $\#1{m}_{v,q} f(x)$

In the registration setting, DET was introduced to resolve the split between point set registration, which aligns sparse geometries, and image registration, which aligns continuous intensity fields on regular grids. The motivating data are high-dimensional vector-valued functions defined on irregular, sparse manifolds, such as spatial transcriptomics. In the tomographic setting, the relevant transform acts on elastic $m$-tensor fields and produces measurements along rays for all admissible polarizations. A plausible implication is that the two usages share a geometric philosophy—elastic structure is imposed on transformations of the domain—but they operate at different levels: Bayesian correspondence inference versus integral-geometric recovery (Hirose et al., 22 Mar 2026).

2. Bayesian DET for functions on irregular domains

In the registration formulation, DET is a training-free Bayesian algorithm for nonrigid registration of vector-valued functions defined on irregular point sets. The central objects are two functions

$$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$

discretized as target points $\{x_n\}_{n=1}^N$ with values

$$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$

and source points $\{y_m\}_{m=1}^M$ with values

$$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$

The registration goal is to infer a deformation $(y_m, f_Y(y_m))$0 such that, for corresponding points $(y_m, f_Y(y_m))$1,

$(y_m, f_Y(y_m))$2

A defining feature of DET is that the transformation acts only on spatial coordinates. The codomain $(y_m, f_Y(y_m))$3 is not warped, rotated, or otherwise mixed into the ambient space. Instead, functional similarity constrains which spatial deformations are plausible. This directly addresses the failure mode of naive augmentation in $(y_m, f_Y(y_m))$4, where spatial and feature dimensions are treated as commensurate coordinates.

The Bayesian model introduces latent variables

$(y_m, f_Y(y_m))$5

where $(y_m, f_Y(y_m))$6 is the displacement field, $(y_m, f_Y(y_m))$7 the mixture weights, $(y_m, f_Y(y_m))$8 the outlier flags, $(y_m, f_Y(y_m))$9 the correspondence indices, $f : \mathbb{R}^n \to \#2{n}{m}$0 the global similarity parameters, $f : \mathbb{R}^n \to \#2{n}{m}$1 the spatial variance, and $f : \mathbb{R}^n \to \#2{n}{m}$2 the functional covariance. The spatial transform applied to a source point is

$f : \mathbb{R}^n \to \#2{n}{m}$3

The joint spatial–functional likelihood is the model’s core coupling: $f : \mathbb{R}^n \to \#2{n}{m}$4 where $f : \mathbb{R}^n \to \#2{n}{m}$5 balances the functional term relative to the spatial term. Consequently, a target point is likely to match a source point only when both the deformed position and the attached feature vector are compatible. This gives DET its characteristic status as a unified geometric–functional registration model (Hirose et al., 22 Mar 2026).

3. Elastic motion prior, correspondence model, and variational inference

The adjective “elastic” in DET refers to the deformation prior. For each coordinate component $f : \mathbb{R}^n \to \#2{n}{m}$6, the displacement field is modeled as a Gaussian process

$f : \mathbb{R}^n \to \#2{n}{m}$7

so that, for the discrete vector $f : \mathbb{R}^n \to \#2{n}{m}$8,

$f : \mathbb{R}^n \to \#2{n}{m}$9

with $\#1{m}_{v,q} f(x)$0. The stiffness parameter $\#1{m}_{v,q} f(x)$1 controls deformation magnitude: larger $\#1{m}_{v,q} f(x)$2 yields a stiffer prior. The associated quadratic form

$\#1{m}_{v,q} f(x)$3

penalizes rough or incoherent displacement fields.

For tissue-like data on manifolds, DET employs a surface coherence matrix that mixes Euclidean and geodesic distances: $\#1{m}_{v,q} f(x)$4 where $\#1{m}_{v,q} f(x)$5, $\#1{m}_{v,q} f(x)$6, and $\#1{m}_{v,q} f(x)$7. The geodesic distance is computed on a $\#1{m}_{v,q} f(x)$8-NN graph in a joint space of coordinates and weighted features. Because geodesic kernels can be indefinite, a fast PSD approximation is applied to obtain a valid covariance.

Inference proceeds by variational Bayes with structured mean-field factorization

$\#1{m}_{v,q} f(x)$9

where $m$0 is a Dirac delta at the MAP parameters. The correspondence statistics are encoded by

$m$1

with

$m$2

The posterior over mixture weights is Dirichlet, the posterior over the displacement field is Gaussian, and the global similarity parameters $m$3 are updated by a weighted Procrustes solution obtained from the SVD of $m$4.

High-dimensional features are handled through a diagonal $m$5, yielding per-feature noise variances

$m$6

This is used as automatic relevance determination (ARD): inconsistent features receive large variance and are down-weighted, whereas informative features receive small variance and contribute more strongly to alignment. DET therefore generalizes Bayesian CPD from geometry-only registration to function-aware registration while preserving the probabilistic motion-coherence structure (Hirose et al., 22 Mar 2026).

4. Scalability, empirical behavior, and practical limitations

DET was designed for atlas-scale data. Its main computational bottlenecks are the matching-probability matrix $m$7 and the kernel matrix $m$8. The implementation uses Nyström approximation of $m$9 and $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$0, k-D tree search to approximate sparse $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$1 in later iterations, and a three-stage pipeline consisting of downsampling, DET on the downsampled sets, and interpolation of the deformation back to the full data. The downsampling scheme is Variance-Guided Importance Sampling (VGIS), defined by

$$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$2

so that both functional boundaries and geometric edges are preferentially retained.

On MERFISH mouse-brain slice-to-slice alignment, the reported mean $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$3 std over $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$4 consecutive slice pairs were $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$5 Jaccard, $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$6 Topology Score, and $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$7 Smoothed PCC for DET; $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$8, $$f_X : \mathbb{R}^D \to \mathbb{R}^{D'}, \qquad f_Y : \mathbb{R}^D \to \mathbb{R}^{D'},$$9, and $\{x_n\}_{n=1}^N$0 for BCPD; and $\{x_n\}_{n=1}^N$1, $\{x_n\}_{n=1}^N$2, and $\{x_n\}_{n=1}^N$3 for PASTE. The abstract summarizes this as DET achieving 92\% topological preservation on MERFISH data where state-of-the-art optimal transport methods struggle $\{x_n\}_{n=1}^N$4. On Stereo-seq whole-embryo atlases, DET registered an E14.5 sagittal section with approximately $\{x_n\}_{n=1}^N$5k points to an E15.5 sagittal section with approximately $\{x_n\}_{n=1}^N$6k points, using the first $\{x_n\}_{n=1}^N$7 principal components of gene expression, while preserving overall tissue topology and maintaining clear organ boundaries. After acceleration, the reported complexity is approximately time $\{x_n\}_{n=1}^N$8 and memory $\{x_n\}_{n=1}^N$9; a runtime study on MacBook Air M1 reported roughly $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$0–$$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$1 s with $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$2 landmarks and up to roughly $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$3 s with $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$4 (Hirose et al., 22 Mar 2026).

Several common misconceptions are explicitly excluded by the formulation. DET is grid-free, so it is not an image-registration method in disguise; it works directly on point clouds and avoids voxelization. It is training-free and unsupervised, so it does not depend on a pretrained network. It does not explicitly enforce diffeomorphisms, and there are no explicit formal guarantees of global diffeomorphism or topology preservation; continuity and coherence arise from the GP prior, smooth kernels, and moderate deformation scales. The paper also identifies typical failure modes: insufficient initial overlap, weak or batch-biased functional signal, and severe topological changes such as missing organs or appearance/disappearance of structures. These limitations define the present boundary of DET as a practical Bayesian registration method rather than a general theorem on topology-preserving correspondence (Hirose et al., 22 Mar 2026).

5. DET as the elastic ray transform in elastic wave tomography

In the tomographic usage, DET is realized by the elastic ray transform acting on elastic tensor fields. For dimensions $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$5 and integer $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$6, the paper defines elastic $$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$7-tensors by

$$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$8

that is, symmetric rank-$$F_x = \big(f_X(x_1), \dots, f_X(x_N)\big) \in \mathbb{R}^{D' \times N},$$9 tensors over the symmetric square of $\{y_m\}_{m=1}^M$0. These are not fully symmetric tensors of rank $\{y_m\}_{m=1}^M$1; they carry the elastic symmetries of elasticity theory. For $\{y_m\}_{m=1}^M$2,

$\{y_m\}_{m=1}^M$3

For a direction $\{y_m\}_{m=1}^M$4, the admissible polarizations are

$\{y_m\}_{m=1}^M$5

Thus both longitudinal $\{y_m\}_{m=1}^M$6 and shear $\{y_m\}_{m=1}^M$7 polarizations are included. The elastic ray transform of $\{y_m\}_{m=1}^M$8 is

$\{y_m\}_{m=1}^M$9

and, for $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$0,

$$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$1

Each operator extends from Schwartz space to $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$2,

$$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$3

The physical motivation is the elastic wave equation

$$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$4

with principal symbol given by the Christoffel matrix

$$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$5

In an isotropic, homogeneous background, $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$6 has one simple P-wave eigenvalue with polarization parallel to $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$7 and one shear eigenvalue of multiplicity $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$8 with polarizations orthogonal to $$F_y = \big(f_Y(y_1), \dots, f_Y(y_M)\big) \in \mathbb{R}^{D' \times M}.$$9. At rank $(y_m, f_Y(y_m))$00, the transform corresponds to a linearization of travel time for all frozen polarizations,

$(y_m, f_Y(y_m))$01

so that, conceptually, domain elastic parameters are mapped to ray-based travel-time-like data indexed by $(y_m, f_Y(y_m))$02. The paper states this explicitly: “At rank 2 it corresponds to linearization of travel time of elastic waves, measured for all polarizations” (Ilmavirta et al., 25 Feb 2025).

6. Kernel structure, identifiability, and relation to classical tensor tomography

The transform-theoretic DET has a kernel structure that differs from standard symmetric tensor tomography because the relevant potential operators are not purely first-order. For general rank $(y_m, f_Y(y_m))$03, the paper defines

$(y_m, f_Y(y_m))$04

with

$(y_m, f_Y(y_m))$05

and proves

$(y_m, f_Y(y_m))$06

At rank $(y_m, f_Y(y_m))$07,

$(y_m, f_Y(y_m))$08

At rank $(y_m, f_Y(y_m))$09, two potential operators enter: $(y_m, f_Y(y_m))$10 with formal adjoints

$(y_m, f_Y(y_m))$11

The central result is

$(y_m, f_Y(y_m))$12

together with solenoidal injectivity,

$(y_m, f_Y(y_m))$13

This yields an elastic Helmholtz-type decomposition

$(y_m, f_Y(y_m))$14

where $(y_m, f_Y(y_m))$15 is the closure of fields of the form $(y_m, f_Y(y_m))$16 and $(y_m, f_Y(y_m))$17 is the closure of fields satisfying $(y_m, f_Y(y_m))$18 and $(y_m, f_Y(y_m))$19.

The Fourier analysis is correspondingly refined. For rank $(y_m, f_Y(y_m))$20, the Fourier slice theorem states that $(y_m, f_Y(y_m))$21 is equivalent to $(y_m, f_Y(y_m))$22 and

$(y_m, f_Y(y_m))$23

for all $(y_m, f_Y(y_m))$24, $(y_m, f_Y(y_m))$25, and $(y_m, f_Y(y_m))$26. At each nonzero frequency, the symbol-level decomposition

$(y_m, f_Y(y_m))$27

separates $(y_m, f_Y(y_m))$28-generated, $(y_m, f_Y(y_m))$29-generated, and jointly solenoidal components. This is the frequency-domain mechanism underlying the kernel characterization.

The identifiability statement is therefore conditional. In general anisotropic settings, only the solenoidal component is recoverable; the potential components are gauge freedoms invisible to all ray–polarization data in the model. Under isotropic perturbations,

$(y_m, f_Y(y_m))$30

the situation simplifies. For $(y_m, f_Y(y_m))$31,

$(y_m, f_Y(y_m))$32

and for $(y_m, f_Y(y_m))$33,

$(y_m, f_Y(y_m))$34

Since $(y_m, f_Y(y_m))$35 is injective, $(y_m, f_Y(y_m))$36 and $(y_m, f_Y(y_m))$37 are uniquely determined. Relative to the classical longitudinal X-ray transform, the elastic-ray DET is therefore a hybrid longitudinal–mixed transform with an elastic-specific kernel generated by second-order potentials and the first-order operator $(y_m, f_Y(y_m))$38, rather than only by the symmetrized gradient familiar from standard tensor tomography (Ilmavirta et al., 25 Feb 2025).