Non-uniform B-splines: Theory & Applications

• Non-uniform B-splines are piecewise polynomial functions defined on non-decreasing, irregular knot vectors that provide local control and adaptability.
• They enable adaptive modeling in computer-aided design, robotics, and mesh-free finite element methods by supporting refined spatial and temporal resolution.
• Efficient algorithms using recursive formulas or transition functions ensure numerical stability and real-time performance in dynamic applications.

Non-uniform B-splines are piecewise polynomial or, more generally, piecewise Chebyshevian functions defined over knot vectors with unequally spaced, non-decreasing knots. They constitute the foundation of adaptive spline-based modeling, computer-aided geometric design, mesh-free finite element approximations, and modern estimation methods in robotics and statistics. Their defining feature is the local control and support properties inherited from the Cox–de Boor recursion, but generalized for arbitrary knot placement. This adaptivity enables both spatially and temporally refined representations—vital for handling anisotropies, data clustering, and dynamic motion.

1. Precise Definition and Algebraic Characterization

A univariate non-uniform B-spline of degree $p$ is associated with a non-decreasing knot sequence $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$. The B-spline basis functions $N_{i,p}(x)$ are defined recursively: $$N_{i,0}(x) = \begin{cases} 1, & t_i \leq x < t_{i+1} \ 0, & \text{otherwise} \end{cases}$$

$$N_{i,p}(x) = \frac{x-t_i}{t_{i+p}-t_i}N_{i,p-1}(x) + \frac{t_{i+p+1}-x}{t_{i+p+1}-t_{i+1}}N_{i+1,p-1}(x)$$

with $0/0$ defined as 0. These properties characterize B-splines algebraically: (i) local support on $[t_i, t_{i+p+1}]$, (ii) a two-term, affine refinement relation under knot insertion, and (iii) translation/dilation invariance. Any family of piecewise polynomials obeying these, including their derivatives, must be proportional to the classical B-spline mapping, as shown by Kamont–Passenbrunner (Kamont et al., 2021).

The general non-uniform B-spline curve is $C(x) = \sum_{i=0}^n N_{i,p}(x)P_i$ for control points $P_i \in \mathbb{R}^d$. Local support and recursive refinement make these ideal for hierarchical and adaptive algorithms.

2. Numerical Construction and Transition Functions

Algorithmic approaches for computing non-uniform B-splines on arbitrary knot vectors are either recursive or based on transition functions (Beccari et al., 2016). In the latter, basis functions are constructed as differences of cumulative transition functions $f_i(x) = \sum_{j=i}^{n} N_{j,p}(x)$ with $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$0. Each $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$1 is evaluated by Hermite interpolation, enforcing vanishing moments at the left endpoint and matching unity at the right, together with continuity at interior knots. This approach is robust with respect to high degrees and highly non-uniform knots, exhibiting better numerical stability and local algorithmic complexity compared to classical recurrences.

Knot insertion and degree elevation are handled by local updates to the transition functions and associated control points, preserving curve shape and partition of unity.

3. Multivariate and Domain-Adaptive Constructions

Tensor-product constructions extend non-uniform B-splines to higher dimensions. For general domains $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$2, the approach includes condensation of the knot vector outside a working interval to cap the maximal span (keeping local supports small) and diversification by partitioning tensor-product B-splines according to the connected components of their support within $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$3 (Sissouno, 2016). The resulting "condensed diversified B-splines" (cdB-splines) retain local partition of unity and polynomial reproduction properties. Their approximation spaces are equipped with anisotropic $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$4-error estimates independent of mesh aspect ratio and domain geometry. Quasi-interpolants built from these spaces reproduce polynomials and admit bounded, locality-preserving projection operators.

4. Adaptivity, Clamping, and Real-Time Deployment

Non-uniform knot sequences are inherently suited to adaptive representations. In continuous-time trajectory estimation, such as Coco-LIC for LiDAR-Inertial-Camera odometry, knot (control point) density is adapted online based on motion dynamics (e.g., average angular rate and specific force from IMU data) (Lang et al., 2023). By subdividing windows proportionally to motion complexity, the spline representation provides spatial/temporal detail on demand while maintaining real-time constraints.

Clamped non-uniform B-splines (C-NUBS) are constructed by repeating the endpoint knots $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$5 times, which guarantees interpolation at the boundaries and eliminates the query-time delay inherent to unclamped splines (Lu et al., 25 Feb 2026). This is critical in multi-robot or sensor-fusion scenarios with asynchronous data: clamping ensures that output (e.g., pose) is available up to the most recent measurement, regardless of the spline degree or dynamic knot insertion. Closed-form extension and shrinkage formulas for clamped splines enable rapid online knot management, supporting knot-keyknot strategies that maintain only a small number of "informative" knots for efficient sliding-window optimization.

The table summarizes adaptivity and clamping features in recent robotic estimation frameworks:

Method	Knot Adaptivity	Clamping	Real-Time Query	Reference
Coco-LIC	IMU-driven density	Optional	Yes	(Lang et al., 2023)
CT-RIO (C-NUBS)	Keyknot strategy	Yes	Yes	(Lu et al., 25 Feb 2026)
Uniform B-spline	Fixed, global	Yes	Up to last pre-set knot	--

5. Penalized Estimation and Statistical Smoothing

Non-uniform B-splines underpin general P-spline estimators, which extend the standard Eilers–Marx framework to arbitrary knot sequences (Li et al., 2022). The core innovation is a weighted difference penalty that mimics integrated squared $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$6-derivatives over irregular grids: $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$7 where $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$8 is the $$\mathbf{T} = \{t_0 \leq t_1 \leq \dots \leq t_{n+p+1}\}$$9-th order finite difference vector, and weights $N_{i,p}(x)$0 adjust for local knot spacing. This preserves the nullspace (polynomials of degree $N_{i,p}(x)$1) and provides a continuous analog to derivative-based O-spline penalties. Empirically, general P-splines on quantile knots match or outperform both standard P-splines (uniform knots) and O-splines in mean squared error. Implementation is available in R packages gps and gps.mgcv.

6. Finite Element Methods with Non-Uniform B-Splines

Non-uniform B-splines enable mesh-free, adaptive, conforming finite element spaces. The non-uniform weighted extended B-splines (NU-WEB-splines) method extends standard tensor-product B-spline bases by incorporating outer basis extensions and distance-type weight functions to handle Dirichlet boundaries (Kumar et al., 2018). The resulting space

$N_{i,p}(x)$2

supports variable support sizes tuned to knot-spacing, facilitating boundary layer and singularity refinement. Applied to nonlinear elliptic PDEs such as the $N_{i,p}(x)$3-Laplacian or quasi-Newtonian flows, these methods admit standard monotonicity/coercivity analysis, unique solutions, and a priori error bounds, with convergence rates depending on mesh width and regularity.

Implementation is meshless up to cell-based quadrature and supports boundary adaptation via the weight function. Local refinement is achieved simply by narrowing knot intervals, controlling support sizes without topological changes.

7. Applications and Comparative Properties

Non-uniform B-splines are the preferred model for data settings and representations where irregular sampling, motion complexity, domain geometry, or spatially inhomogeneous features require variable resolution. Key attributes include:

• Locality: Control points only affect the curve or surface inside their local support.
• Refinement: The basis is closed under knot insertion and degree elevation without global recomputation.
• Adaptivity: Knot placement adjusts to spatial gradients, sensor rates, or modelled phenomena (e.g., aggressive vs. smooth robotic motion).
• Versatility: Applicable across geometric modeling, meshless PDE solvers, adaptive statistical smoothing, sensor fusion, and continuous-time estimation.

Comparisons with uniform and (un)clamped splines reveal that non-uniform variants offer better tradeoffs between accuracy and computational cost in contexts where the underlying phenomena are themselves nonuniform or multiscale, as extensively demonstrated in SLAM, statistical estimation, and PDE approximation literature (Lang et al., 2023, Lu et al., 25 Feb 2026, Li et al., 2022, Kumar et al., 2018, Sissouno, 2016).

---

References:

• (Lang et al., 2023): "Coco-LIC: Continuous-Time Tightly-Coupled LiDAR-Inertial-Camera Odometry using Non-Uniform B-spline"
• (Lu et al., 25 Feb 2026): "Parallel Continuous-Time Relative Localization with Augmented Clamped Non-Uniform B-Splines"
• (Kamont et al., 2021): "An algebraic characterization of B-splines"
• (Beccari et al., 2016): "A practical method for computing with piecewise Chebyshevian splines"
• (Sissouno, 2016): "Anisotropic spline approximation with non-uniform B-splines"
• (Kumar et al., 2018): "Non uniform weighted extended B-Spline finite element analysis of non linear elliptic partial differential equations"
• (Li et al., 2022): "General P-Splines for Non-Uniform B-Splines"