1 Basic concepts
1.1 Piecewise polynomial functions
A spline is defined by joining multiple polynomial pieces over different intervals of an independent variable. Each piece is typically simple—often low degree—yet the overall representation can mimic complex shapes. The definition is local: changing one interval’s polynomial affects the curve mainly near that region, especially in bases designed for locality.
1.2 Knots and knot vectors
The junction points between polynomial segments are called knots. For a one-dimensional spline in variable \(x\), knots are ordered values \[ x_0 \le x_1 \le \cdots \le x_n, \] and they partition the domain into intervals \([x_i, x_{i+1}]\). In many practical formulations, knot values are collected into a knot vector, which also dictates the number of polynomial segments and the flexibility of the spline family used.
1.3 Degree and smoothness requirements
The degree of the spline refers to the degree of each polynomial piece (for example, degree 3 for cubic splines). Higher degree can increase smoothness and approximation power, but it also raises computational and numerical complexity. Smoothness requirements specify how many derivatives must agree across knots, which strongly influences which spline construction is appropriate.
1.4 Continuity conditions (C0, C1, etc.)
Continuity conditions describe how the spline behaves at knots:
- \(C^0\) (function continuity): the spline value is continuous across each knot.
- \(C^1\) (first-derivative continuity): the first derivative is also continuous.
- \(C^2\) (second-derivative continuity): curvature-related properties become smooth.
More generally, a spline is said to be \(C^k\) if derivatives up to order \(k\) match at knot locations. For a piecewise polynomial of degree \(p\), the maximum possible smoothness across interior knots is limited by the degree and by how the spline is constructed.
2 Spline types and constructions
2.1 Polynomial splines
2.1.1 Simple examples (linear and quadratic splines)
The most basic splines are piecewise linear (degree 1) and piecewise quadratic (degree 2).
- A linear spline has segments joined so that the function is continuous; the derivative typically changes abruptly at knots.
- A quadratic spline can enforce additional smoothness by choosing coefficients so the first derivative matches at knots, depending on the imposed conditions and available degrees of freedom.
These small-degree cases illustrate how continuity constraints restrict coefficients and how knot placement influences the shape.
2.1.2 Cubic spline foundations
Cubic splines use degree \(3\) polynomial pieces and commonly enforce \(C^2\) continuity across interior knots, producing smooth curves with continuous first and second derivatives. With suitable boundary conditions, the resulting curve balances flexibility and smoothness and is widely used for interpolation and modeling.
A typical cubic spline construction treats the curve as a collection of polynomials on each interval with coefficients determined by:
- matching the function values at knots (interpolation or fitting),
- matching derivatives at knots (smoothness),
- enforcing boundary constraints.
2.2 Cardinal splines and uniform knots
Cardinal splines are associated with a specific shape-preserving parameterization where knots are equally spaced. Uniform knot grids simplify analysis and implementation, and the spline shape can be tuned through a tension parameter (in some variants) that controls how closely the curve follows the data points. While cardinal splines are not always used in advanced numerical methods, their uniform-knot structure makes them conceptually and computationally accessible.
2.3 Truncated power basis
One classical representation expresses a spline as a linear combination of truncated power functions: \[ (x-t)_+^p = \begin{cases} (x-t)^p, & x\ge t,\\ 0, & x<t, \end{cases} \] where \(t\) is a knot location and \(p\) relates to the spline degree. By including such terms for multiple knots, the function can change polynomial behavior after each knot while maintaining desired smoothness. This basis can be effective for deriving theoretical properties, though other bases (notably B-splines) are often favored for stable computation.
2.4 B-spline basis functions
2.4.1 Cox–de Boor recursion formula
B-splines provide a systematic way to build spline spaces using basis functions defined over a knot vector. The Cox–de Boor recursion defines basis functions \(N_{i,p}(x)\) of degree \(p\) from lower-degree ones:
- For degree \(0\), the basis functions are piecewise constants on knot intervals.
- Higher degrees are built by blending lower-degree basis functions with coefficients derived from knot spans.
This recursion yields a compact and widely used construction for spline evaluation, fitting, and approximation.
2.4.2 Local support and numerical advantages
A major practical strength of B-splines is local support: each basis function is nonzero only over a limited region of the domain. As a result, modifying control coefficients affects the curve only locally, improving interpretability and often computational efficiency. Additionally, B-spline formulations are numerically stable for many tasks because they avoid large dense systems inherent in some polynomial-basis approaches.
3 Interpolation and approximation
3.1 Spline interpolation
3.1.1 Interpolatory cubic splines
Interpolation uses the spline to pass through specified data points \((x_i, y_i)\). For cubic spline interpolation with ordered knots \(x_i\), each knot’s function value is matched: \[ S(x_i) = y_i. \] The remaining degrees of freedom are determined by enforcing smoothness across knots, typically \(C^2\) continuity, and by applying boundary conditions. The outcome is a smooth curve that honors the given samples exactly.
3.1.2 Boundary conditions (natural, clamped, not-a-knot)
Boundary conditions control behavior at the ends of the domain:
- Natural spline: second derivatives at endpoints are set to zero, producing a “minimal curvature” feel at boundaries.
- Clamped spline: first derivatives at endpoints are specified, aligning the curve’s slope with known endpoint behavior.
- Not-a-knot spline: the first two polynomial segments are forced to connect as if they were higher-degree at the near-end knots, reducing artificial boundary effects.
Different choices reflect different assumptions about what happens outside the sampled range.
3.2 Spline least squares approximation
3.2.1 Choosing a smoothing level
Least squares spline approximation replaces exact interpolation with a compromise that reduces sensitivity to noise. A common approach is to penalize excessive curvature through a smoothing parameter \(\lambda\), often tied to the integral of squared derivatives. As \(\lambda\) increases, the spline becomes smoother and deviates more from the data; as \(\lambda\) decreases, the fit approaches interpolation.
3.2.2 Trade-offs: bias vs. variance
Approximation introduces a classic statistical trade-off.
- More smoothing increases bias (the model may miss fine-scale features).
- Less smoothing decreases bias but can increase variance (the fit may follow noise).
Spline methods make this trade-off tunable through the smoothing level and the chosen spline space complexity (such as knot count).
3.3 Error estimation and convergence intuition
Error behavior depends on the smoothness of the underlying target function, the spline degree, and how knot placement scales with sample density. Intuitively, if the true function is sufficiently smooth, using splines of appropriate degree and refining the knot placement can reduce approximation error. Convergence is often analyzed through approximation theory and numerical analysis tools that bound the difference between the spline estimator and the ideal function.
4 Algorithms and computation
4.1 Building the spline system
4.1.1 Setting up linear equations
Many spline constructions reduce to solving a linear system for unknown coefficients. For example, interpolation and least squares in a chosen basis lead to constraints that are linear in coefficients. Smoothness conditions across knots translate into linear equations involving derivatives evaluated from each piece. The result is often a sparse or structured system, especially in B-spline or derivative-matching formulations.
4.2 Evaluating splines efficiently
4.2.1 De Boor’s algorithm for evaluation
De Boor’s algorithm evaluates B-spline curves efficiently and stably. It uses the knot vector and the degree to determine which basis functions matter at a given \(x\), then performs a sequence of local blending operations. This avoids recomputing all basis functions and scales well for repeated evaluations.
4.2.2 Complexity and stability considerations
Evaluation complexity is typically proportional to the number of relevant knot spans and the spline degree, rather than the total number of knots. Stability can be improved by algorithms that avoid subtractive cancellation and that leverage the partition-of-unity-like properties of B-spline basis functions. In practice, careful knot vector handling and floating-point considerations matter when knot spans are extremely small.
4.3 Differentiation and integration of splines
4.3.1 Derivative splines
Derivatives of a spline are again splines, generally of one lower degree. In B-spline form, derivative computation can be expressed via transformed control coefficients and an adjusted knot vector structure. This property makes it straightforward to obtain velocity-like or curvature-like information from spline models.
4.3.2 Definite integrals and antiderivatives
Definite integrals over spline segments can be computed exactly (within floating-point arithmetic) by integrating polynomial pieces or by using basis-function integral properties. For repeated queries—such as cumulative area or energy measures—precomputing segment integrals or using cumulative sums can improve efficiency.
5 Practical considerations
5.1 Knot selection strategies
5.1.1 Uniform vs. non-uniform knots
Uniform knots place knot locations at regular intervals of the independent variable. Non-uniform knots allow variable resolution, allocating more flexibility where the data changes rapidly. In many applications, non-uniform placement can improve accuracy without uniformly increasing knot count.
5.1.2 Adaptive refinement concepts
Adaptive refinement increases knot density in regions where the current spline fit is insufficient, such as near steep gradients or where residual errors are large. The refinement process may iteratively add knots, recompute coefficients, and reassess error. This approach can balance computational cost and accuracy by targeting complexity where it is most needed.
5.2 Smoothing splines vs. exact interpolation
5.2.1 Controlling overfitting
Exact interpolation forces agreement with every data point and can overfit when observations contain noise. Smoothing splines mitigate this by reducing the influence of noisy variations through regularization. Overfitting control is often managed by selecting the smoothing parameter and monitoring residual behavior and generalization performance.
5.3 Handling noisy or irregular data
For noisy measurements, robust fitting often prefers smoothing or regularized least squares rather than strict interpolation. Irregular sampling—non-uniform spacing in \(x\)—can also affect numerical conditioning and local flexibility. Carefully chosen knot strategies and stable basis methods help prevent oscillatory behavior common in naive high-degree polynomial fits.
5.4 Multidimensional spline basics (curves vs. surfaces)
Spline concepts extend beyond one-dimensional functions. A spline curve uses one parameter (e.g., arc-length-like or a normalized parameter) and represents geometry in multiple spatial dimensions. Spline surfaces use two parameters and are constructed by tensor-product approaches: basis functions in each parameter direction combine to form surface patches. These extensions are fundamental in geometric modeling pipelines.
6 Applications
6.1 Data fitting and curve modeling
Splines provide flexible curve models that can capture smooth trends and local variations. In practical modeling, spline fitting can serve as a compact representation of measured relationships, offering interpretable derivatives and controllable smoothness.
6.2 Numerical methods and simulation support
In numerical simulation, splines help define smooth initial conditions, interpolate simulation outputs, or approximate unknown functions in methods requiring differentiability. They also support error-controlled refinement by enabling local adjustments to fit quality.
6.3 Computer graphics and CAD pipelines
Spline curves are central to design and rendering workflows, offering smooth paths, controllable shape editing, and efficient evaluation. CAD systems typically use spline representations to maintain geometric continuity and to support operations such as trimming, blending, and offsetting.
6.4 Signal processing and filtering perspectives
While classic signal processing often uses different bases, spline functions can act as smoothing kernels or as representations for reconstructing signals from samples. By selecting degree and knot structure, splines can model low-pass-like behavior and yield smooth interpolants suited for derivative-based analysis.
7 Special topics
7.1 Splines on irregular domains (overview)
Standard spline formulations assume a structured independent variable axis. For irregular domains—such as data on nonuniform grids or scattered samples—spline construction may require reparameterization, specialized knot strategies, or localized methods. The overarching goal remains the same: represent a smooth model while respecting constraints implied by the geometry or sampling pattern.
7.2 Periodic splines
Periodic splines enforce agreement of the function and several derivatives at the domain endpoints, treating the ends as the same point in a loop. This is useful for cyclical phenomena where boundary conditions should not create artificial discontinuities.
7.3 Rational splines (overview of NURBS concept)
Rational splines generalize polynomial splines by using weighted combinations of basis functions. In the familiar NURBS framework, weights allow representation of conic sections exactly and provide enhanced control over curve shape. The rational form is particularly important in geometric modeling where exact arcs and circles are required.
7.4 Spline constraints and shape preservation (overview)
7.4.1 Monotonicity and convexity-preserving ideas
In certain applications, it is desirable that a spline respects qualitative properties of the underlying function, such as monotonic increase/decrease or convexity. Constraint-based spline methods modify the fitting or interpolation problem to satisfy inequalities on derivatives or second derivatives. These techniques can reduce unrealistic overshoot while retaining smoothness.