KernelInterpolation.jl

KernelInterpolation.jl is a Julia package that implements methods for multivariate interpolation in arbitrary dimension based on symmetric (conditionally) positive-definite kernels with a focus on radial basis functions. It can be used for classical interpolation of scattered data, as well as for generalized (Hermite-Birkhoff) interpolation by using a meshfree collocation approach. This can be used to solve partial differential equations both stationary ones and time-dependent ones by using some time integration method from OrdinaryDiffEq.jl.

Installation

If you have not yet installed Julia, then you first need to download Julia. Please follow the instructions for your operating system. KernelInterpolation.jl works with Julia v1.10 and newer. KernelInterpolation.jl is a registered Julia package. Therefore, you can install it by executing the following commands from the Julia REPL

julia> using Pkg

julia> Pkg.add("KernelInterpolation")

For visualizing the results, additionally you need to install Plots.jl, which can be done by

julia> using Pkg

julia> Pkg.add("Plots")

To create special node sets, you might also want to install QuasiMonteCarlo.jl or Meshes.jl and for solving time-dependent partial differential equations OrdinaryDiffEq.jl in a similar way as above for Plots.jl. See the documentation for more examples on how to use these packages in combination with KernelInterpolation.jl.

Usage

In the Julia REPL, first load the package KernelInterpolation.jl

julia> using KernelInterpolation

In order to interpolate discrete function values of a (potentially) multivariate function $f: \mathbb{R}^d\to \mathbb{R}$, we first need a set of nodes $X = \{x_1,\ldots,x_n\}\subset\mathbb{R}^d$, where the function values of $f$ are known. In KernelInterpolation.jl we can, e.g., construct a homogeneous grid on a hypercube in 2 dimensions by calling

julia> nodeset = homogeneous_hypercube(5, (-2, -1), (2, 1))

Here, we specified that the hypercube has 5 nodes along each of the 2 dimensions (i.e. in total we have $5^2 = 25$ nodes) and that the boundaries of the cube are given by the lower left corner located at $(-2, -1)$ and the upper right corner at $(2, 1)$. Similarly, NodeSets can be constructed by the functions random_hypercube, random_hypercube_boundary, homogeneous_hypercube_boundary, random_hypersphere or random_hypersphere_boundary or by directly passing a set of nodes to the constructor of NodeSet. Besides the nodeset, we need the function values at the nodes. Let's say, we want to reconstruct the function $f(x) = \sin(x_1\cdot x_2)$. Then, we can create the vector of function values by

julia> f(x) = sin(x[1]*x[2])
julia> ff = f.(nodeset)

Finally, we obtain the Interpolation object by calling interpolate, where we specify the kernel function that is used for the reconstruction. Here, we take a Gaussian $\phi(r) = \exp(-(\varepsilon r)^2)$ with shape parameter $\varepsilon = 1/2$ as radial-symmetric basis function:

julia> kernel = GaussKernel{dim(nodeset)}(shape_parameter = 0.5)
julia> itp = interpolate(nodeset, ff, kernel)

If the kernel is only conditionally positive definite, the interpolant will be augmented by a polynomial of the corresponding order of the kernel. Another order can also be passed explicitly with the keyword argument m of interpolate. The result itp is an object that is callable on any point $x\in\mathbb{R}^d$, e.g.,

julia> itp([-1.3, 0.26])
-0.34096946394940986

julia> f([-1.3, 0.26])
-0.33160091709280176

For more sophisticated examples also involving solving stationary or time-dependent partial differential equations, see the documentation. More examples can be found in the examples/ subdirectory.

Visualization

In order to visualize the results, you need to have Plots.jl installed and loaded

julia> using Plots

A NodeSet can simply be plotted by calling

julia> plot(nodeset)

An Interpolation object can be plotted by providing a NodeSet at which the interpolation is evaluated. Continuing the example from above, we can visualize the resulting interpolant on a finer grid

julia> nodeset_fine = homogeneous_hypercube(20, 2, (-2, -1), (2, 1))
julia> plot(nodeset_fine, itp)

To visualize the true solution f in the same plot as a surface plot we can call

julia> plot!(nodeset_fine, f, st = :surface)

KernelInterpolation.jl also supports exporting (and importing) VTK files, which can be visualized using tools such as ParaView or VisIt. See the documentation for more details.

Referencing

You can directly refer to KernelInterpolation.jl as

@misc{lampert2024kernel,
  title={{K}ernel{I}nterpolation.jl: {M}ultivariate (generalized) scattered data interpolation
         with symmetric (conditionally) positive definite kernel functions in arbitrary dimension},
  author={Lampert, Joshua},
  year={2024},
  month={06},
  howpublished={\url{https://github.com/JoshuaLampert/KernelInterpolation.jl}},
  doi={10.5281/zenodo.12599880}
}

Authors

The package is developed and maintained by Joshua Lampert (University of Hamburg).

License and contributing

KernelInterpolation.jl is published under the MIT license (see License). We are pleased to accept contributions from everyone, preferably in the form of a PR.