Volumetric Functional Maps

• Volumetric functional maps are a spectral framework that encodes 3D solid correspondences using bases of low-frequency volumetric Laplace eigenfunctions.
• They project both discrete and continuous signals onto compact eigenbases, enabling noise suppression, compression, and effective segmentation transfer in applications like medical imaging.
• The approach minimizes descriptor, spectral commutativity, and orthogonality errors to refine correspondences, supporting accurate mesh deformation and reduced geodesic error.

Volumetric functional maps are a spectral framework for computing correspondences, transferring signals, and structuring data between 3D solids. Extending the classical functional maps method from 2D surfaces to volumetric domains, they encode volumetric shape correspondences as compact linear operators between bases of volumetric Laplace eigenfunctions. The resulting machinery supports robust, efficient transfer of discrete and continuous volumetric information across solids, with demonstrated applications in medical imaging, geometry processing, and signal analysis (Maggioli et al., 16 Jun 2025).

1. Theoretical Foundations

Volumetric functional maps generalize the functional maps paradigm from triangulated surfaces to three-dimensional solids. The foundational object is a tetrahedral mesh $M=(V_M,T_M)$ discretizing a solid domain $\Omega\subset\mathbb{R}^3$ with boundary $\partial\Omega$. The volumetric Laplace operator $\Delta_\Omega$ is defined on scalar functions $\phi:\Omega\to\mathbb{R}$ with Neumann boundary conditions: $$\Delta_\Omega \phi_i = \lambda_i \phi_i\quad{\rm in}\;\Omega,\qquad (\partial\phi_i/\partial n) = 0\;{\rm on}\;\partial\Omega,$$ where $n$ is the outward boundary normal. In the finite-element setting, piecewise-linear basis functions $\{\phi_j\}$ yield the stiffness matrix $K$ and mass matrix $M$: $$K_{ij} = \sum_{t\in T_M} \int_t \nabla\phi_i\cdot\nabla\phi_j\,dV,\quad M_{ij} = \sum_{t\in T_M} \int_t \phi_i\phi_j\,dV.$$ The generalized eigenproblem

$K\Phi = M\Phi\Lambda$

produces a basis $\Phi\in\mathbb{R}^{n\times k}$ of the first $k$ low-frequency volumetric eigenfunctions, orthonormal under the discrete $L^2$ inner product $\langle f, g\rangle = f^\top M g$ (Maggioli et al., 16 Jun 2025).

2. Signal Representation and Projection

A key property of $\Phi$ is that any scalar signal $f\in\mathbb{R}^n$—such as a segmentation label, coordinate function, or texture—admits a compact projection: $$\alpha = \Phi^\top M f \in \mathbb{R}^k,\qquad f \approx \Phi\alpha.$$ Truncating to $k\ll n$ yields a smooth, noise-suppressed signal representation, facilitating both compression and regularization. This spectral approach is robust to perturbations and local irregularities in the input signal.

3. Construction and Optimization of Volumetric Functional Maps

Given two tetrahedral meshes $M$ and $N$ with bases $\Phi^M,\Phi^N$ and eigenvalues $\Lambda^M,\Lambda^N$, a volumetric functional map is a matrix $C\in\mathbb{R}^{k\times k}$ such that

$\Phi^M C \approx T(\Phi^N),$

where $T$ is the (unknown) linear operator induced by the (unknown) pointwise correspondence $\pi : M \to N$.

For $r$ pairs of descriptor functions (e.g., HKS, WKS, coordinates) $D^M, D^N$, the corresponding spectral coefficients $A = (\Phi^M)^\top M^M D^M$, $B = (\Phi^N)^\top M^N D^N$ are computed. The functional map $C$ is then obtained by minimizing the regularized objective: $$E(C) = \|C B - A\|_F^2 + \mu \|C \Lambda^N - \Lambda^M C\|_F^2 + \beta \|C^\top C - I\|_F^2,$$ where the terms respectively enforce descriptor preservation, spectral commutativity (isometry), and orthogonality. The closed-form solution for the first term is $C = A B^+$ with $B^+$ the pseudoinverse. When using all terms, general-purpose solvers or alternating least squares are employed over $k^2$ unknowns (Maggioli et al., 16 Jun 2025).

Several refinements—such as ZoomOut (increasing $k$ iteratively), orthogonalized eigenproducts, and coupled bases—carry over directly from the surface functional maps literature, further improving correspondence accuracy and stability.

4. Practical Algorithmic Pipeline

Computation of volumetric functional maps proceeds through several standard steps:

1. Mesh Preprocessing: Ensure both $M$ and $N$ are watertight and manifold; tetrahedralization is required if only surface data is available.
2. Laplace Eigen-Decomposition: Assemble and diagonalize the stiffness and mass matrices to obtain $\Phi^M$, $\Phi^N$.
3. Signal Projection: Project descriptors $D^M$, $D^N$ into the respective spectral bases.
4. Functional Map Estimation: Estimate $C$ by minimizing $E(C)$, possibly with iterative enlargement of $k$ and repeated refinement (ZoomOut).
5. Pointwise Map Recovery: For each vertex $i$ in $M$, assign the match $$\pi(i) = \arg\min_j \|\Phi^M[i,:] - (C\Phi^N[j,:])\|_2$$.

This pipeline supports both discrete and continuous signal transfer and enables extraction of approximate pointwise correspondences by nearest neighbor search in spectral space (Maggioli et al., 16 Jun 2025).

5. Applications in Geometry Processing and Medical Imaging

Volumetric functional maps facilitate a wide array of transfer and correspondence tasks:

• Segmentation Transfer: Label information encoded as one-hot vectors is mapped between volumes. The method yields noise-robust, smoothly varying labelings suitable for medical anatomical segmentation of, e.g., brain structures.
• Mesh Connectivity Transfer: With a known boundary correspondence $\pi$, one can transfer interior coordinates volumetrically or extrapolate from surface-only information. Mesh deformation is achieved with low rates of tetrahedra flips ($\leq1\%$ when using 20% of the spectrum).
• Solid Texturing: Procedural or empirical 3D textures, modeled as scalar fields on $N$, are mapped to $M$ in a manner that respects the interior geometry of the target domain.
• Improved Surface Shape Matching: Embedding surface meshes as boundaries of volumetric solids and computing volumetric functional maps can substantially reduce average geodesic error (AGE) compared to surface-only pipelines. For example, on the Su et al. 2019 dataset, volumetric ZoomOut with $k=30$ achieved AGE $\approx0.022$ versus $0.04$ for the surface-only approach, with a modest 1.6$\times$ computational overhead (Maggioli et al., 16 Jun 2025).

6. Comparison with Voxelwise and Atlas-Based Volumetric Methods

While volumetric functional maps are primarily formulated for geometric and signal transfer tasks between 3D solids, a closely related approach in brain imaging is the construction of high-resolution volumetric functional atlases, such as DiFuMo (Dadi et al., 2020). DiFuMo decomposes preprocessed 4D fMRI data $X\in\mathbb{R}^{p\times n}$ into a sparse, non-negative dictionary $D$ of $k$ spatial modes, capturing functional gradients of brain activity as high-dimensional volumetric maps. While both frameworks leverage spectral or dictionary-based bases for volumetric representation, the objectives and constraints differ: DiFuMo emphasizes soft, spatially localized, overlapping functional modes for signal reduction and statistical analysis, whereas volumetric functional maps focus on precise bijective correspondences and spectral signal transfer across meshes.

A plausible implication is that future work might integrate the regularized, sparse representations of DiFuMo with spectral volumetric mapping techniques, especially for tasks requiring both functional parcellation and accurate geometric correspondence (Dadi et al., 2020, Maggioli et al., 16 Jun 2025).

7. Limitations and Prospects

Volumetric functional maps require computation of low-frequency volumetric Laplace eigenfunctions, which can introduce modest runtime overhead (approximately 1.2–2$\times$ slower than surface-only analogues when interior vertices are $\sim$2--3$\times$ the surface count). The quality of correspondences strongly depends on tetrahedral mesh regularity, choice of descriptors, and the rank $k$ of the truncated basis.

The main advantages are robustness to non-isometric deformations, flexibility in transferring both discrete and continuous signals, and the ability to reuse spectral refinement strategies from the well-established surface functional map literature (Maggioli et al., 16 Jun 2025). Potential future directions include integration of learned or data-driven descriptors, scalable eigen-solvers, and application to domains combining functional and anatomical information, as in neuroimaging workflows.

---

Key References:

• "Volumetric Functional Maps" (Maggioli et al., 16 Jun 2025)
• "Fine-grain atlases of functional modes for fMRI analysis" (Dadi et al., 2020)