Radial basis functions (RBFs) can be used to approximate derivatives and
solve differential equations in several ways. Here, we compare one important scheme
to ordinary finite differences by a mixture of numerical experiments and theoretical
Fourier analysis, that is, by deriving and discussing analytical formulas for the error in differentiating exp(ikx) for arbitrary k. "Truncated RBF differences" are derived
from the same strategy as Fourier and Chebyshev pseudospectral methods: Differentiation of the Fourier, Chebyshev or RBF interpolant generates a differentiation matrix that maps the grid point values or samples of a function u(x) into the values of
its derivative on the grid. For Fourier and Chebyshev interpolants, the action of the
differentiation matrix can be computed indirectly but efficiently by the Fast Fourier
Transform (FFT). For RBF functions, alas, the FFT is inapplicable and direct use of the
dense differentiation matrix on a grid of N points is prohibitively expensive (O(N2))
unless N is tiny. However, for Gaussian RBFs, which are exponentially localized, there
is another option, which is to truncate the dense matrix to a banded matrix, yielding
“truncated RBF differences”. The resulting formulas are identical in form to finite differences except for the difference weights. On a grid of spacing h with the RBF as
φ(x)=exp(−α2(x/h)2),
where without approximation wm=(−1)m+12α2/sinh(mα2). We derive explicit formula for the differentiation of the linear function, f(X)≡X, and the errors therein. We
show that Gaussian radial basis functions (GARBF), when truncated to give differentiation formulas of stencil width (2M+1), are significantly less accurate than (2M)-th order finite differences of the same stencil width. The error of the infinite series
(M=∞) decreases exponentially as α→0. However, truncated GARBF series have a
second error (truncation error) that grows exponentially as α → 0. Even for α ∼ O(1) where the sum of these two errors is minimized, it is shown that the finite difference
formulas are always superior. We explain, less rigorously, why these arguments extend to more general species of RBFs and to an irregular grid. There are, however, a
variety of alternative differentiation strategies which will be analyzed in future work,
so it is far too soon to dismiss RBFs as a tool for solving differential equations.