FastTransforms.jl

FastTransforms.jl allows the user to conveniently work with orthogonal polynomials with degrees well into the millions.

This package provides a Julia wrapper for the C library of the same name. Additionally, all three types of nonuniform fast Fourier transforms are available, as well as the Padua transform.

Installation

Installation, which uses BinaryBuilder for all of Julia's supported platforms (in particular Sandybridge Intel processors and beyond), may be as straightforward as:

pkg> add FastTransforms

julia> using FastTransforms, LinearAlgebra

Fast orthogonal polynomial transforms

The orthogonal polynomial transforms are listed in FastTransforms.Transforms or FastTransforms.kind2string.(instances(FastTransforms.Transforms)). Univariate transforms may be planned with the standard normalization or with orthonormalization. For multivariate transforms, the standard normalization may be too severe for floating-point computations, so it is omitted. Here are two examples:

The Chebyshev--Legendre transform

julia> c = rand(8192);

julia> leg2cheb(c);

julia> cheb2leg(c);

julia> norm(cheb2leg(leg2cheb(c; normcheb=true); normcheb=true)-c)/norm(c)
1.1866591414786334e-14

The implementation separates pre-computation into an FTPlan. This type is constructed with either plan_leg2cheb or plan_cheb2leg. Let's see how much faster it is if we pre-compute.

julia> p1 = plan_leg2cheb(c);

julia> p2 = plan_cheb2leg(c);

julia> @time leg2cheb(c);
  0.433938 seconds (9 allocations: 64.641 KiB)

julia> @time p1*c;
  0.005713 seconds (8 allocations: 64.594 KiB)

julia> @time cheb2leg(c);
  0.423865 seconds (9 allocations: 64.641 KiB)

julia> @time p2*c;
  0.005829 seconds (8 allocations: 64.594 KiB)

Furthermore, for orthogonal polynomial connection problems that are degree-preserving, we should expect to be able to apply the transforms in-place:

julia> lmul!(p1, c);

julia> lmul!(p2, c);

julia> ldiv!(p1, c);

julia> ldiv!(p2, c);

The spherical harmonic transform

Let F be an array of spherical harmonic expansion coefficients with columns arranged by increasing order in absolute value, alternating between negative and positive orders. Then sph2fourier converts the representation into a bivariate Fourier series, and fourier2sph converts it back. Once in a bivariate Fourier series on the sphere, plan_sph_synthesis converts the coefficients to function samples on an equiangular grid that does not sample the poles, and plan_sph_analysis converts them back.

julia> F = sphrandn(Float64, 1024, 2047); # convenience method

julia> P = plan_sph2fourier(F);

julia> PS = plan_sph_synthesis(F);

julia> PA = plan_sph_analysis(F);

julia> @time G = PS*(P*F);
  0.090767 seconds (12 allocations: 31.985 MiB, 1.46% gc time)

julia> @time H = P\(PA*G);
  0.092726 seconds (12 allocations: 31.985 MiB, 1.63% gc time)

julia> norm(F-H)/norm(F)
2.1541073345177038e-15

Due to the structure of the spherical harmonic connection problem, these transforms may also be performed in-place with lmul! and ldiv!.

See also FastSphericalHarmonics.jl for a simpler interface to the spherical harmonic transforms defined in this package.

Nonuniform fast Fourier transforms

The NUFFTs are implemented thanks to Alex Townsend:

• nufft1 assumes uniform samples and noninteger frequencies;
• nufft2 assumes nonuniform samples and integer frequencies;
• nufft3 ( = nufft) assumes nonuniform samples and noninteger frequencies;
• inufft1 inverts an nufft1; and,
• inufft2 inverts an nufft2.

Here is an example:

julia> n = 10^4;

julia> c = complex(rand(n));

julia> ω = collect(0:n-1) + rand(n);

julia> nufft1(c, ω, eps());

julia> p1 = plan_nufft1(ω, eps());

julia> @time p1*c;
  0.002383 seconds (6 allocations: 156.484 KiB)

julia> x = (collect(0:n-1) + 3rand(n))/n;

julia> nufft2(c, x, eps());

julia> p2 = plan_nufft2(x, eps());

julia> @time p2*c;
  0.001478 seconds (6 allocations: 156.484 KiB)

julia> nufft3(c, x, ω, eps());

julia> p3 = plan_nufft3(x, ω, eps());

julia> @time p3*c;
  0.058999 seconds (6 allocations: 156.484 KiB)

The Padua Transform

The Padua transform and its inverse are implemented thanks to Michael Clarke. These are optimized methods designed for computing the bivariate Chebyshev coefficients by interpolating a bivariate function at the Padua points on [-1,1]^2.

julia> n = 200;

julia> N = div((n+1)*(n+2), 2);

julia> v = rand(N); # The length of v is the number of Padua points

julia> @time norm(ipaduatransform(paduatransform(v)) - v)/norm(v)
  0.007373 seconds (543 allocations: 1.733 MiB)
3.925164683252905e-16

References

[1] D. Ruiz—Antolín and A. Townsend, A nonuniform fast Fourier transform based on low rank approximation, SIAM J. Sci. Comput., 40:A529–A547, 2018.

[2] T. S. Gutleb, S. Olver and R. M. Slevinsky, Polynomial and rational measure modifications of orthogonal polynomials via infinite-dimensional banded matrix factorizations, arXiv:2302.08448, 2023.

[3] S. Olver, R. M. Slevinsky, and A. Townsend, Fast algorithms using orthogonal polynomials, Acta Numerica, 29:573—699, 2020.

[4] R. M. Slevinsky, Fast and backward stable transforms between spherical harmonic expansions and bivariate Fourier series, Appl. Comput. Harmon. Anal., 47:585—606, 2019.

[5] R. M. Slevinsky, Conquering the pre-computation in two-dimensional harmonic polynomial transforms, arXiv:1711.07866, 2017.