As-Isometric-As-Possible Regularizations

• As-Isometric-As-Possible Regularizations are methods that approximate isometric properties by soft penalties, balancing strict isometry with practical tractability.
• They employ convex relaxations, variational formulations, and lifting approaches to manage non-convex constraints in applications like non-rigid reconstruction, shape matching, and image registration.
• Empirical results show these methods achieve near-isometric mappings with robust performance metrics, offering explicit control over deformation and volume distortion in complex geometric problems.

As-isometric-as-possible regularizations comprise a family of methodologies designed to enforce or promote isometric (length- or area-preserving) mappings in geometric inverse problems, non-rigid reconstruction, shape matching, and deformable registration. These regularizations arise in contexts where strict isometry is either ill-posed, non-convex, or physically unattainable, and so one seeks penalizations that make surfaces, maps, or deformations as close as possible to isometric within a tractable optimization framework. Central to these approaches are convex relaxations, variational formulations, and principled penalty terms that interpolate between fully unconstrained deformations and strict isometry, allowing for practical computation and robust numerical methods.

1. Mathematical Formulations of As-Isometric-As-Possible Regularizations

The isometric constraint requires that the mapping under consideration—such as a surface immersion, a deformation field, or a shape correspondence—preserve particular geometric quantities (e.g., pairwise lengths, local volumes, or the first fundamental form). Direct enforcement typically leads to non-convex or even ill-posed systems. As-isometric-as-possible (AIAP) regularizations replace strict equality by a soft penalty, allowing controlled deviations measured by appropriate norms or mask functions.

Typical AIAP regularizations take the form

$$\text{Penalty} = \sum_{k} w_k \,\phi\!\left(\mathcal{Q}_k[\mathbf{X}] - \mathcal{Q}_k^\text{ref}\right)$$

where $\mathcal{Q}_k$ encodes a geometric quantity (distance, area, curvature), $\phi$ is often an $\ell_1$ or $\ell_2$-norm, and $w_k$ are weights. The precise structure depends on application: distance preservation in non-rigid structure-from-motion (NRSfM) (Sengupta et al., 2022), commutativity with the Laplace-Beltrami operator in non-rigid matching (Ren et al., 2020), or constraints on Jacobian determinants and shear for image registration (Mang et al., 2015).

Convexification is central. Non-convex quartic or quadratic forms in raw variables are lifted (e.g., via Gram or moment matrices), allowing the constraints to become linear or affine in the lifted space, thus admitting solution via semidefinite programming (SDP) (Sengupta et al., 2022).

2. Convex SDPs and Lifting Approaches in NRSfM

In NRSfM, the isometric constraint on point clouds is

$$l_{ij}^t \leq \|\mathbf{X}_i^t - \mathbf{X}_j^t\|_2 \leq u_{ij}^t$$

for all pairs $(i,j)$ and frames $t$, with $l_{ij}^t$ and $u_{ij}^t$ determined by template geodesic distances. Squaring yields quartic forms in the 3D variables, which are non-convex. The AIAP (a.k.a. "quasi-isometric") regularization of Sengupta & Bartoli performs a Gram-matrix or moment-matrix lifting, introducing PSD matrices $\mathfrak{R}^t$ or $\mathfrak{S}^{t'}$, and relaxes the exact distance matching to an $\ell_1$-norm penalty: $$\underset{\{\mathfrak R^t\},\,\{g_{ij}\}}{\min}~ \sum_{t=1}^n\mathrm{Tr}(\mathfrak{R}^t) + \lambda_I \sum_{t,i,j} \lvert \text{dist}_{ij}^t - g_{ij} \rvert, \quad \mathfrak{R}^t \succeq 0$$ where $\text{dist}_{ij}^t$ is affine in the Gram matrix. This convex SDP admits efficient optimization and, empirically, the relaxation is tight, yielding near-isometric reconstructions as the penalty parameter $\lambda_I$ increases (Sengupta et al., 2022).

3. Variational and Elliptic Regularization of Isometric Immersion

The classical isometric immersion problem for surfaces in $\mathbb{R}^3$ is governed by a first-order, fully characteristic PDE system, which is ill-posed due to lack of ellipticity. The approach of Anderson introduces an elliptic regularization by prescribing a convex blend of intrinsic (conformal class) and extrinsic (mean curvature) data: $$D_{\varepsilon}(X) = ([g], (1-\varepsilon)\lambda + \varepsilon H)$$ where $[g]$ is the pointwise conformal class, $\lambda$ the conformal factor, and $H$ the mean curvature. For $\varepsilon > 0$, the resulting system is elliptic of Fredholm index zero and permits a robust variational formulation based on the Einstein–Hilbert action with Gibbons–Hawking–York boundary terms: $$I_\varepsilon(g) = \int_M R(g)\,dV_g + (2-\varepsilon)\int_\Sigma H\,dA_g$$ Critical points realize deformations as isometric as possible, with the parameter $\varepsilon$ controlling interpolation between strict isometry and conformal/mean-curvature-driven embeddings (Anderson, 2017).

4. AIAP Regularization in Functional Map Matching

For non-rigid shape correspondence, AIAP regularization appears as spectral commutativity constraints within the functional map framework. Classic commutativity penalties with the Laplace–Beltrami operator diverge in the smooth/infinite-dimensional limit, motivating the introduction of a resolvent-based regularizer: $$E_\mathrm{res}^{(\gamma)}(C) = \|C\,R(\Delta_1^\gamma;\mu) - R(\Delta_2^\gamma;\mu)\,C \|_F^2$$ where $C$ is the functional map, $R(\Delta^\gamma;\mu)$ the resolvent, and $\gamma$ controls the bias toward isometry. The parameter $\gamma$ allows the penalty to interpolate between minimal bias and strict isometry, with larger $\gamma$ enforcing a stronger alignment of spectra. This yields bounded, well-posed regularizers even in the continuum and enables precise tuning to expected distortion levels. Empirical results show significant error reduction and enhanced stability relative to purely Laplacian commutativity penalties (Ren et al., 2020).

5. PDE-Constrained and Shear-Controlling Regularization for Image Registration

Diffeomorphic image registration involves constructing a deformation $y$ that maps a template image to a reference. AIAP constraints are enforced by augmenting classical $H^1$ or $H^2$ Tikhonov regularization of the velocity field with additional penalties on the divergence (to control compressibility, i.e., the determinant of the deformation gradient), as well as explicit shear penalization: $$\min_{v,w} \frac{1}{2}\|m_R - m(\cdot,1)\|_{L^2}^2 + \frac{\beta_v}{2}\|v\|_V^q + \frac{\beta_w}{2}\|w\|_W^2$$ subject to $\partial_t m + v\cdot\nabla m = 0$, $\nabla\cdot v = w$, and additional nonlinear terms controlling shear via strain-rate tensors. The framework permits precise control over local volume distortion and shear, with parameter tuning yielding registrations as volume-preserving—and thus as-isometric—as possible, without compromising data fidelity. Empirical studies demonstrate min/max Jacobian determinant tightly clustered around unity and shear profiles controlled independently of match error (Mang et al., 2015).

6. Algorithmic Implementations and Optimization

Algorithmic approaches depend on the specific context:

• Semidefinite programming (SDP) for the convex-lifted AIAP regularization in NRSfM, solved by interior-point methods (e.g., MOSEK/CVX) and shown empirically to yield nearly rank-1 Gram or moment matrices (Sengupta et al., 2022).
• Newton–Krylov methods for elliptic regularizations of isometric immersion, leveraging Newton solvers for the boundary-value problem with mesh-based discretization (Anderson, 2017).
• Least-squares solvers for resolvent-regularized functional maps, allowing direct integration into standard pipelines and supporting subsequent refinement by ICP or BCICP (Ren et al., 2020).
• Gauss–Newton–Krylov strategies for PDE-constrained registration, with preconditioners and matrix-free iterative methods scaling to large problem sizes (Mang et al., 2015).

All methods provide explicit handles—in the form of penalty parameters, exponents (e.g., $\gamma$ in resolvent masks), or convex combination weights—for interpolating between weak and strong isometry enforcement, ensuring practical tunability.

7. Empirical Performance, Limitations, and Theoretical Guarantees

Across diverse geometric tasks, AIAP regularizations exhibit the following:

• Tightness of Relaxation: In SDP-lifted approaches, the trace minimization reliably enforces near rank-1 matrix solutions, implying tightness of the relaxation (Sengupta et al., 2022).
• Robustness to Discretization: The resolvent regularizer is always bounded and remains well-posed even as the Laplacian basis size increases, in contrast to untruncated commutativity penalties (Ren et al., 2020).
• Quantitative Metrics: Performance is assessed by isometry error (e.g., deviation of pairwise lengths, $\det D\varphi$ in registration, per-vertex geodesic error in mapping), and the regularization demonstrably narrows the range of local volume changes and reduces off-diagonal distortions in function maps (Sengupta et al., 2022, Ren et al., 2020, Mang et al., 2015).
• Algorithmic Stability: Ellipticity and Fredholm index zero in regularized PDE formulations guarantee local uniqueness and stability of solutions for all positive regularization parameters (Anderson, 2017).

Limitations include the persistence of ambiguities for highly extensible or equiareal models, and a breakdown of convexity or well-posedness as regularization parameters approach their extremal values (e.g., $\varepsilon \to 0$ in PDE regularization).

---

In summary, as-isometric-as-possible regularizations constitute a principled, unifying framework across geometric inverse problems, enabling robust and flexible enforcement of near-isometry via convex relaxations, variational principles, and spectral or PDE-based penalties. These strategies have established theoretical foundations and empirical effectiveness, underpinning state-of-the-art methods in NRSfM, shape mapping, immersion geometry, and deformation field registration (Sengupta et al., 2022, Ren et al., 2020, Anderson, 2017, Mang et al., 2015).