Tensor Tomography Overview

• Tensor tomography is the inverse problem of reconstructing symmetric tensor fields from integrals along curves, capturing orientation-dependent data and managing gauge freedoms.
• It relies on solenoidal decompositions and energy identities to ensure unique recovery up to natural gauge, particularly on simple or negatively curved manifolds.
• Modern reconstruction employs closed-form inversions, regularized optimization, and compressed sensing techniques to tackle nonuniqueness and incomplete data challenges.

Tensor tomography is the inverse problem of reconstructing a symmetric tensor field (of arbitrary order) from its integrals along a family of curves, typically geodesics, straight lines, or more general measurement trajectories. Unlike scalar tomography, which recovers scalar functions from line integrals, tensor tomography introduces distinct theoretical and computational challenges due to gauge freedoms, kernel structure, and the need to reconstruct tensorial information from orientation-dependent data. It forms the mathematical core of several applied imaging modalities, such as X-ray scattering tensor tomography, dielectric tensor tomography, polarimetric microscopy, and stress analysis in materials.

1. Fundamental Concepts: Geometry, Transform, and Gauge

Tensor tomography generalizes the geodesic X-ray transform to symmetric tensors of order $m\geq 1$, defined on a Riemannian manifold $(M,g)$ or subsets of Euclidean space. For a symmetric $m$-tensor field $f$,

$$I_mf(\gamma) = \int f_{\gamma(t)}(\dot\gamma(t),\ldots,\dot\gamma(t))\,dt$$

assigns to each geodesic $\gamma$ the integral of $f$ contracted $m$ times with the velocity. On $(M,g)$, a key property is that any tensor $f$ admits an orthogonal decomposition,

$(M,g)$0

where $(M,g)$1 is solenoidal ($(M,g)$2), and $(M,g)$3 denotes the symmetric (covariant) derivative of an order-$(M,g)$4 field $(M,g)$5. The kernel of $(M,g)$6 always contains the potential tensors $(M,g)$7, reflecting the non-injectivity for general $(M,g)$8. Reconstruction is thus fundamentally posed up to this gauge freedom.

A “solenoidal injectivity” (s-injectivity) property holds if $(M,g)$9 and $m$0 implies $m$1—uniqueness up to the natural gauge. The analytic structure of this problem depends on the geometry (e.g., simple manifolds, negative curvature, or trapped/nontrapping settings) and the presence of boundary.

2. Theoretical Results: Injectivity, Gauge Structure, and Regularity

2.1 Surfaces and Simple Manifolds

On simple Riemannian surfaces—compact, with strictly convex boundary, and no conjugate points—s-injectivity holds for all orders $m$2 (Paternain et al., 2011, Paternain et al., 2013). The proof is based on reduction to a transport equation on the unit sphere bundle $m$3 and energy identities (Pestov identity). For $m$4 solving $m$5 with boundary conditions, positivity arguments and control over Fourier modes yield that the only tensors in the kernel are the potentials.

Beyond surfaces, for higher $m$6-dimensional manifolds, s-injectivity has been established for $m$7 under additional curvature or regularity assumptions (Paternain et al., 2014, Ilmavirta et al., 2023). In particular, simple manifolds with $m$8 metrics and non-positive curvature admit s-injectivity for all $m$9 via careful adaptations of the energy-based proof using a regularity-aware calculus (Ilmavirta et al., 2023).

2.2 Decomposition Beyond Solenoidal Gauge

Recent developments have refined the canonical representative to include transverse-traceless (“tt”) or iterated-tt decompositions, especially on asymptotically hyperbolic surfaces (Eptaminitakis et al., 5 Oct 2025). For even-order tensors,

$f$0

where each $f$1 is transverse-traceless of order $f$2, naturally diagonalizing the data space into orthogonal components. This decomposition provides a finer description of the range and kernel structure.

3. Explicit Inversion, Reconstruction Algorithms, and Stability

3.1 Closed-Form Inversion and Fan-Beam Formulations

In Euclidean domains, particularly the disk, efficient inversion formulas are known using fan-beam coordinates, filtered backprojection, and integral kernel representations (Monard, 2015, Monard, 2017). Representatives in a special “reconstructible” gauge (distinct from solenoidal) allow for orthogonal recovery of each harmonic tensor mode via explicit Cauchy-type integrals. In the presence of attenuation, explicit integrating factors and holomorphic correction operators are employed to sequentially reconstruct higher-order components.

3.2 General Settings and Regularization

For general tensor tomography (e.g., dielectric or X-ray scattering tensors), the inverse problem is ill-posed and often underdetermined due to missing data in Fourier space (“missing cone”), anisotropic measurement geometry, or the limited angular coverage. Regularization methods—such as tensorial total variation (Hugonnet et al., 2022), Tikhonov (Laplacian) regularization (Nielsen et al., 2023), or low-rank tensor-based compressed sensing (Takemoto et al., 2014)—are incorporated into the variational reconstruction: $f$3 where $f$4 is the measurement operator, $f$5 the data, $f$6 a regularization term over all spatial and tensor components, and convex optimization methods (e.g., FISTA) are used for minimization.

3.3 Adjoint Methods with Attenuation and Refraction

For time-dependent, attenuated, and refracting media, the forward operator is modeled as a boundary trace of a dynamically attenuated transport equation along geodesics defined by the refractive index profile. Reconstruction leverages two explicit adjoint formulations: an integral backprojection representation and a PDE-based dual transport equation. Empirically, the integral adjoint provides significant computational speed advantage without sacrifice of accuracy; accounting for refraction robustly reduces error in the presence of index gradients (Vierus et al., 16 Jan 2026).

4. Extensions: Scattering Tensor Tomography, Dielectric and Optical Methods

4.1 X-ray and Small-Angle Scattering

In X-ray scattering tensor tomography, the exit intensity is modeled as Gaussian blurring by a local 2x2 (projected) scattering tensor, with the forward operator mapping the full 3x3 symmetric scattering tensor through sequence of measured projections at various rotation/tilt angles. The adjoint operator rotates and backprojects each measured 2x2 tensor into the 3D field, with solution iteratively updated. Validation demonstrates robust support for various wavefront modulator geometries and experimental settings (Lautizi et al., 2024).

Advanced small-angle scattering tensor tomography reconstructs voxel-wise reciprocal space maps through a global linear inversion for the spherical harmonic coefficients representing the texture, enabling model-independent recovery of nanostructural anisotropy and phase content (Nielsen et al., 2023).

4.2 Dielectric Tensor Tomography and Polarimetric Microscopy

Dielectric tensor tomography (DTT) reconstructs spatially varying 3x3 dielectric tensors from polarization-resolved transmission measurements using frameworks such as 3D vectorial diffraction (Born/Rytov approximation) and experimental multiplexing for efficient acquisition. Regularization with tensor total variation yields resolutions and orientation accuracy suitable for imaging complex organic and liquid crystal systems (Hugonnet et al., 2022, Lee et al., 2022). Tensorial ptychography and differential phase contrast approaches merge multi-angle, polarization, and phase information to map the full $f$7 or $f$8 permittivity tensor volumes for biological and advanced material samples (Xu et al., 2023, Xu et al., 2022).

5. Nonuniqueness, Kernel Characterizations, and Geometry Dependence

Non-simple geometries (e.g., periodic slabs, manifolds with trapped geodesics, or nonconvex boundaries) admit nontrivial kernels beyond classical potential tensors. In slabs $f$9, the kernel additionally contains “depth-only” functions and gauge artifacts due to the system’s symmetry (Ilmavirta et al., 2017). On asymptotically hyperbolic or noncompact surfaces, refined decompositions allow range characterizations of the X-ray transform via explicit orthogonal projections, spectral moment conditions, and one-sided Fourier distributions (Eptaminitakis et al., 5 Oct 2025). For circular and V-line tomography, explicit inversion and kernel descriptions are given in terms of moment families and potential decompositions (Mishra et al., 2024).

The presence or absence of injectivity and the structure of the kernel depend delicately on global geometric and topological properties (e.g., simplicity, negative curvature, convexity, or the existence of invariant distributions), which are characterized through dynamical and microlocal analytic techniques, such as the analysis of the normal operator, Beurling transforms, and Pestov-type estimates (Paternain et al., 2014, Ilmavirta et al., 2023).

6. Numerical Methods and Applications

Numerical implementations of tensor tomography utilize both direct inversion (e.g., closed-form formulas for low-order tensors in disks), iterative regularized optimization (for large-scale inverse problems with incomplete data), and compressed sensing for high-dimensional or sparse fields (Monard, 2015, Hugonnet et al., 2022, Takemoto et al., 2014). Key applications include:

• X-ray scattering analysis of fiber and bone microstructure, anisotropic texture and orientation mapping (Lautizi et al., 2024, Nielsen et al., 2023)
• Dielectric tensor reconstruction in liquid crystals, biological tissue polarization microscopy, and stress tensor imaging (Hugonnet et al., 2022, Xu et al., 2023, Xu et al., 2022)
• Wireless tomography (RF shadowing) and quantum state tomography (via tensor network representations) (Takemoto et al., 2014, Lidiak et al., 2022)
• Stress analysis in photoelastic materials, neural implicit reconstructions from polarization (Dave et al., 2024)
• Dynamic imaging of time-dependent or refracting tensor fields utilizing adjoint-accelerated solvers (Vierus et al., 16 Jan 2026)

7. Open Problems and Future Directions

• s-injectivity in higher dimensions ($$I_mf(\gamma) = \int f_{\gamma(t)}(\dot\gamma(t),\ldots,\dot\gamma(t))\,dt$$0, $$I_mf(\gamma) = \int f_{\gamma(t)}(\dot\gamma(t),\ldots,\dot\gamma(t))\,dt$$1) for general simple or Anosov metrics remains unresolved, with most progress in 2D, negative curvature, or analytic/generic metrics (Paternain et al., 2013, Paternain et al., 2014).
• Development of explicit, stable algorithmic inversion and kernel descriptions for arbitrary geometries and measurement configurations, including partial data and non-unitary connections, is ongoing.
• Integration of advanced regularization (low-rank, plug-and-play, or model-based priors) and compressed sensing for efficient reconstruction from highly redundant or underdetermined data (Hugonnet et al., 2022, Takemoto et al., 2014).
• Extension to nonlinear, multiple-scattering, and non-Born regimes for strongly scattering and thick samples, especially in optical tensor tomography (Xu et al., 2023, Xu et al., 2022).
• Exploitation of tensor tomography for in vivo biological anisotropy imaging, stress tensor mapping in microengineered devices, and tensorial quantum state or process tomography, all requiring scalable measurement and computation (Lidiak et al., 2022, Lautizi et al., 2024).

Tensor tomography thus combines deep geometric inverse theory with diverse practical methodologies, spanning from analytic reconstruction on simple domains to advanced computational imaging under real-world experimental constraints. The subject remains an active area of research with ongoing theoretical and technological developments.