In one dimension, polynomial interpolation is straightforward: any $N+1$ distinct points uniquely determine a polynomial of degree $N$. This guaranteed uniqueness—known as unisolvence—holds because a non-zero 1D polynomial of degree $N$ cannot have more than $N$ roots. Consequently, the underlying Vandermonde matrix is always non-singular.
Move to multiple dimensions or curved surfaces, however, and this guarantee vanishes. In a multivariate setting, distinctness is no longer enough. For example, three distinct points in a plane uniquely define a linear polynomial (a plane) unless they happen to lie on a single straight line. If they are collinear, an infinite number of planes can pass through them, rendering the interpolation matrix singular.
When interpolating data on grids or manifolds, avoiding these singular configurations is a profound mathematical challenge. A modern thread of research addresses this by developing constructive frameworks that guarantee unisolvence, ensuring that interpolation matrices remain invertible and numerically stable.
The Geometry of Failure: Poisedness and Unisolvence
To understand why multivariate interpolation fails, we must look at the spatial arrangement of the data points. A set of points is called poised if it uniquely determines a polynomial from a given subspace. As Zhang et al. emphasize, the core requirement for unisolvence is that the chosen interpolation points must not all lie on a single algebraic hypersurface (the zero-set of a multivariate polynomial).
If the points do fall on such a hypersurface, any polynomial evaluating to zero on that surface can be added to an existing solution, destroying uniqueness. Identifying valid, poised sets of points becomes exceptionally difficult in higher dimensions or on irregular geometries. Randomly selected points are statistically likely to be poised, but as zu Castell et al. point out, proving this analytically for specific configurations is notoriously hard. This difficulty motivates the creation of structured, constructive methods to guarantee poisedness.
Poised Configuration Non-Poised Configuration
(Uniquely determines plane) (Infinitely many planes possible)
o o-------o-------o
/ \ Collinear points lie on a
o---o shared algebraic hypersurface
Constructive Point Sets on Manifolds
One way to guarantee uniqueness is to carefully constrain where the points are placed. This is particularly vital when interpolating functions on curved manifolds, such as spheres or closed surfaces, which are common in geographic and physical simulations.
For the unit sphere, zu Castell et al. developed a factorization method to systematically construct poised point sets. By analyzing the conditions under which spherical polynomials vanish, the authors established specific configurations (such as even numbers of points distributed across latitudes) that form Chebyshev systems. These systems provide an analytical guarantee that the interpolation problem possesses a unique solution, bypassing the need to check matrix determinants numerically.
On more complex, arbitrary closed surfaces, Thekke Veettil et al. leverage unisolvent interpolation techniques to represent entire shapes as Global Polynomial Level Sets (GPLS). By utilizing stable polynomial bases like Newton or Lagrange, they ensure that a discrete point cloud uniquely locks down a global algebraic variety. This mathematical rigidity allows researchers to compute high-order geometric properties—like surface curvature—with near machine precision.
Adapting the Polynomial Subspace
Instead of forcing points into rigid geometric arrangements, an alternative strategy alters the polynomial space itself to fit the points.
When solving partial differential equations (PDEs) on curved surfaces, algorithms often use local clusters of points. Narayan et al. introduced a method called Least Orthogonal Interpolation (LOI) to handle these clusters. Instead of projecting points onto a local flat tangent plane—which introduces geometric distortions—LOI operates directly in the ambient Cartesian embedding space.
To guarantee unique solvability, LOI constructs a custom, minimum-degree polynomial subspace tailored to the specific local point cloud. By generating an orthogonal basis that matches the intrinsic dimensionality of the data points, the method ensures the interpolation matrix is always invertible, eliminating the stagnation errors that typically plague manifold-based PDE solvers.
Verifying Non-Singularity Algebraically
Ultimately, proving that an interpolation rule is unique requires demonstrating that the determinant of its coefficient matrix is non-zero. This becomes increasingly intricate when the interpolation rule incorporates not just function values, but also derivatives (Hermite interpolation).
Kim addressed this by using a linear functional approach to construct high-degree interpolation rules on equally spaced nodes. By restructuring the problem, the author broke down a massive, potentially ill-conditioned interpolation matrix into two smaller, manageable linear systems. This algebraic factorization made it possible to prove that the underlying determinants are strictly non-zero for distinct nodes, providing an ironclad guarantee of uniqueness regardless of how many derivative orders are included.
Conclusion
Whether by algorithmically searching for poised lattices (as in the work of Zhang et al.), constructing explicit Chebyshev systems on spheres, or dynamically tailoring the polynomial basis to local data, ensuring unisolvence is the bedrock of stable multivariate approximation. By understanding the geometric interplay between point placement and polynomial degrees, these techniques transform fragile numerical approximations into robust, mathematically guaranteed solvers.
Key Foundational Work
Polynomial interpolation on the unit sphere II (2007) — contributed 8 concept elements to this direction.
HIGH-DEGREE INTERPOLATION RULES GENERATED BY A LINEAR FUNCTIONAL (2007) — contributed 3 concept elements to this direction.
RBF-LOI: Augmenting Radial Basis Functions (RBFs) with Least Orthogonal Interpolation (LOI) for Solving PDEs on Surfaces (2018) — contributed 2 concept elements to this direction.