Laplacian-Regularized B-Spline Solver

• The paper introduces a closed-form Laplacian-regularized B-spline solver that robustly fits curves while penalizing curvature to ensure smoothness.
• It employs a discrete Laplacian operator to enforce regularization, minimizing overshoot and adapting to non-uniform or underconstrained data.
• The method enables efficient, scalable compression and denoising, making it valuable for 3D motion modeling and dynamic simulation applications.

A Laplacian-regularized B-spline solver is a closed-form computational method for fitting B-spline curves to data, subject to a discrete Laplacian (second-difference) smoothness prior on the control polygon. This approach enables robust, high-fidelity functional approximations of potentially non-uniform or underconstrained data, by minimizing a combined objective consisting of a data-fitting term and a regularization term that penalizes curvature in the control points. The Laplacian-regularized B-spline solver provides efficient and scalable compression and denoising in settings such as 3D vertex trajectory modeling, scientific simulation, and generative modeling of complex dynamic phenomena (Wang et al., 21 Feb 2026, Lenz et al., 2023).

1. B-Spline Curve Formalism

A B-spline curve of degree $d$ is defined via a nondecreasing knot vector $u = [u_0, \ldots, u_{m-1}]$, with $m = k + d + 1$ and $k$ control points $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$. The curve is evaluated at parameter $t$ as

$C(t) = \sum_{i=0}^{k-1} N_{i,d}(t) p_i$

where $N_{i,d}(t)$ are recursive Cox–de Boor B-spline basis functions. For time-sampled or parameterized data $\{t_1, \ldots, t_T\}$, one constructs the basis matrix $B \in \mathbb{R}^{T \times k}$ with $u = [u_0, \ldots, u_{m-1}]$0 so that the sampled curve becomes $u = [u_0, \ldots, u_{m-1}]$1, with $u = [u_0, \ldots, u_{m-1}]$2 stacking control points and $u = [u_0, \ldots, u_{m-1}]$3 the data or trajectory samples (Wang et al., 21 Feb 2026).

Clamped knot vectors guarantee endpoint interpolation ($u = [u_0, \ldots, u_{m-1}]$4, $u = [u_0, \ldots, u_{m-1}]$5), and uniform interior knot spacing yields $u = [u_0, \ldots, u_{m-1}]$6 continuity.

2. Laplacian Regularization

To enforce smoothness and suppress spurious curvature, a discrete Laplacian (second-difference) operator $u = [u_0, \ldots, u_{m-1}]$7 acts on the control points, with $u = [u_0, \ldots, u_{m-1}]$8 for $u = [u_0, \ldots, u_{m-1}]$9. The Laplacian-regularized energy is

$m = k + d + 1$0

where $m = k + d + 1$1 controls the curvature penalty. This discourages overshooting and fits robustly even in noisy or data-scarce regimes (Wang et al., 21 Feb 2026).

Generalizations to tensor-product spline settings yield penalty matrices built from discrete evaluations of second-derivatives at locations of maximal basis support, leading to a curvature penalty $m = k + d + 1$2 where $m = k + d + 1$3 is a second-derivative collocation matrix (Lenz et al., 2023).

3. Fitting Objective and Closed-Form Solution

The solver minimizes the sum of a data-fitting loss and Laplacian regularizer: $m = k + d + 1$4 Taking derivatives and equating to zero yields the normal equations: $m = k + d + 1$5 which is a small, positive-definite system for the control points. The closed-form minimizer is

$m = k + d + 1$6

This system admits efficient batched solutions via Cholesky factorization and matrix–matrix products, particularly with small $m = k + d + 1$7 (number of control points), making it practical for large-scale, high-dimensional data such as framewise vertex displacements or volumetric simulation fields (Wang et al., 21 Feb 2026, Lenz et al., 2023).

4. Adaptive and Spatially Varying Regularization

Uniform regularization may oversmooth high-density regions or undersmooth poorly-constrained regions. Adaptive Laplacian-regularized B-spline solvers introduce spatially varying weights $m = k + d + 1$8 per control point, based on data availability and local second-derivative magnitude. Let $m = k + d + 1$9 be the sum of basis function samples for the $k$0-th control point, and $k$1 the sum of second-derivative collocation magnitudes. Regularization is modulated as

$k$2

where $k$3 is a user-chosen “regularization threshold.” Only control points with inadequate sample coverage are smoothed, preserving features in data-rich areas. Additional, optional first-derivative penalties further suppress low-frequency drift in unconstrained regions. The result is an SPD system with adaptively assembled diagonal weighting matrices. The regularization threshold $k$4 controls the tradeoff: low $k$5 fits sharp features, high $k$6 approaches uniform smoothing. Empirically, $k$7 yields artifact-free yet accurate splines in complex data (Lenz et al., 2023).

5. Algorithmic Implementation and Efficiency

For large-scale systems, the Laplacian-regularized B-spline solver proceeds as follows:

1. Build the collocation matrix $k$8 (or $k$9 for tensor-product splines).
2. Assemble the discrete Laplacian $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$0 (or penalty matrix $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$1).
3. (If adaptive) Compute $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$2, $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$3 and diagonal regularization weights.
4. Form the SPD normal system $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$4 or its spatially-weighted generalization.
5. Compute right-hand side $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$5.
6. Solve via direct (e.g., Cholesky, if $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$6 is small) or iterative (e.g., Preconditioned Conjugate Gradient) linear solvers.

Because all channels and points of a data manifold (e.g., mesh vertices) can be processed with the same, pre-factorized system per data length, the method scales efficiently to millions of trajectories or high spatial resolutions. Reported runtimes for 200 frames and 50,000 vertices are under one second on a commodity CPU, due to the favorable scaling and reuse of factorizations (Wang et al., 21 Feb 2026, Lenz et al., 2023).

6. Integration in Motion Synthesis and Modeling Pipelines

The solver has practical impact in dynamic shape and motion modeling. In the BiMotion generative pipeline for text-guided 3D character animation, vertex-wise displacement trajectories $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$7 are compressed from variable-length sequences into fixed-size sets of B-spline control points $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$8 via the Laplacian-regularized solver. This compact, smooth encoding enables expressive variational modeling (e.g., VAE latent embeddings) over long, complex motion episodes and is robust to variable temporal granularity. The control points, along with initial vertex positions and normals, are hierarchically embedded, subsampled, and cross-attended during both training and inference. The inverse process reconstructs full-length, temporally smooth motions from generated control points for arbitrary output durations. Thus, the solver is integral to bridging raw motion data and deep generative representations, enabling semantically faithful long-horizon animation synthesis (Wang et al., 21 Feb 2026).

7. Applications and Limitations

Laplacian-regularized B-spline solvers are employed in scientific data modeling, geometric processing, and generative modeling where noise-robust, smooth, and compact functional representations are required. Feature-preserving regularization—such as spatially adaptive Laplacian weighting—enables high-accuracy fits even with non-uniform or scarce data (Lenz et al., 2023). A plausible implication is that such solvers are well-suited to inverse problems, shape completion, and signal denoising tasks on complex domains.

Limitations include possible oversmoothing in regions where $\{p_0, \ldots, p_{k-1}\} \subset \mathbb{R}^D$9 is too low to capture high-frequency content or $t$0 is mis-tuned. The solver assumes appropriate design of knot vectors and basis support, and the adaptive scheme needs careful parameter calibration for optimal feature preservation.

---

Key references:

• (Wang et al., 21 Feb 2026) “BiMotion: B-spline Motion for Text-guided Dynamic 3D Character Generation”
• (Lenz et al., 2023) “Customizable Adaptive Regularization Techniques for B-Spline Modeling”