This package implements a hyperbolic cross sparse-grid method for approximating multivariate continuous functions with Chebyshev polynomials as basis functions. Flexibility is given over the resulting grid through parameters that govern the maximum number of points along each spatial dimension and the number of layers along each spatial dimension. The package also allows for isotropic and anisotropic grids.

Hyperbolic cross approximation is a sparse-grid alternative to Smolyak's method. Where existing treatments of the hyperbolic cross method are based on non-equispaced fast Fourier methods, the treatment in this code is based on Chebyshev polynomials with the coefficients in the approximating function constructed using Lagrange interpolation.

This package draws on and is related to others I have written, specifically: ChebyshevApprox.jl and SmolyakApprox.jl.

To install this package you need to type in the REPL

using Pkg
Pkg.add("HyperbolicCrossApprox")

Then the package can be used by typing

using HyperbolicCrossApprox

The nodes used to form the approximating grid can be computed using either the Chebyshev nodes (points of the first kind), Chebyshev extrema (points of the second kind), the extended Chebyshev nodes, Vertesi nodes, or Legendre nodes, with the approximating grid and the multi-index computed by

grid, multi_ind = hyperbolic_cross_grid(chebyshev_nodes,d,k,domain)

where d is the dimension of the function, k is the number of layers along each dimension, and domain is a 2d-array (2xd) containing the upper and lower bound on each variable. If domain is not provided, then it is assumed that the variables reside on the [-1,1] interval. If k is an integer, then an isotropic grid is computed whereas if k is an array of integers with length d, then an anisotropic grid is computed. For the function above, the number of grid points along each spacial dimension is given by n = 2k+1. Alternatively, the values for k and n can be provided separately (and hence not linked by n=2k+1) by

grid, multi_ind = hyperbolic_cross_grid(chebyshev_nodes,d,k,n,domain)

where n is either an interger or a 1d array of integers.

In the functions above, chebyshev_nodes can be replaced with chebyshev_extrema, chebyshev_extended, vertesi_nodes, or legendre_nodes.

With the grid and multi-index in hand, we can compute the weights, or coefficients in the approximation, according to

weights = hyperbolic_cross_weights(y,grid,multi_ind,domain)

where y is a 1d-array containing the evaluations at each grid point of the function being approximated. The weights can be computed more efficiently by computing the inverse interpolation matrix (this generally needs to be done only once, outside any loops)

inv_interp_mat = hyperbolic_cross_inverse_interpolation_matrix(grid,multi_ind,domain)

with the weights then computed through

weights = hyperbolic_cross_weights(y,inv_interp_mat)

We can evaluate the hyperbolic cross approximation of the function at any point in the domain by

y_hat = hyperbolic_cross_evaluate(weights,point,multi_ind,domain)

where point (a 1d-array) is the point in the domain where the approximation is to be evaluated.

The package can also be used to compute derivatives of the approximating function, as per

deriv = hyperbolic_cross_derivative(weights,point,multi_ind,domain,pos)

where pos is an integer reflecting the number of the variable being differentiated with respect to. The package can compute gradients according to

gradient = hyperbolic_cross_gradient(weights,point,multi_ind,domain)

and hessians according to

hess = hyperbolic_cross_hessian(weights,point,multi_ind,domain)

Lastly, the package implements a Clenshaw-Curtis integration and a Guass-Chebyshev integration scheme. To integrate a function with repsect to all variables over the approximating domain use

integral = hyperbolic_cross_integrate(f,plan,:clenshaw_curtis)
integral = hyperbolic_cross_integrate(f,plan,:gauss_chebyshev_quad)

where f is the function to be integrated and plan is a HCApproxPlan structure, described below. The :gauss_chebyshev_quad approach constructs the quadrature weights using a variant on "designed quadrature", (Keshavarzzadeh, et al, (2018, SIAM J. SCI. COMPUT)). The :clenshaw_curtis method approximates the function to be integrated with a hyperbolic cross and then numerically integrates the approximation.

There are multi-threaded functions to compute the polynomial weights and the interpolation matrix. These multi-threaded functions are accessed by adding _threaded to the end of the funtion, as per

weights = hyperbolic_cross_weights_threaded(y,inv_interp_mat)

The key structure to be aware of is the HCApproxPlan, which contains the key information needed to approximate a function.

d = 3
k = 5
domain = [2.0 2.0 2.0; -2.0 -2.0 -2.0]
grid, mi = hyperbolic_cross_grid(chebyshev_nodes,d,k,domain)
plan = HCApproxPlan(grid,mi,domain)

or

plan = hyperbolic_cross_plan(chebyshev_nodes,d,k,n,domain)
plan = hyperbolic_cross_plan(chebyshev_nodes,d,k,domain)

Once the approximation plan has been constructed it can be used to create functions to interpolate and to compute gradients and hessians.

f = hyperbolic_cross_interp(y,plan)
g = hyperbolic_cross_gradient(y,plan)
h = hyperbolic_cross_hessian(y,plan)

point = [1.0, 1.0, 1.0]

f(point)
g(point)
h(point)

There are multi-threaded versions of hyperbolic_cross_interp, hyperbolic_cross_gradient, and hyperbolic_cross_hessian; just add _threaded to the end of the function name.

• ChebyshevApprox.jl
• SmolyakApprox.jl
• PiecewiseLinearApprox.jl

Some useful references are:

Dennis, R., (2021), "Using a Hyperbolic Cross to Solve Non-linear Macroeconomic Models", CAMA working paper number 93/2021, Australian National University.

Döhler, M., Kunis, S., and D. Potts, (2010), "Nonequispaced Hyperbolic Cross Fast Fourier Transform", SIAM Journal on Numerical Analysis, 47, (6), pp. 4415-4428.

Dũng, D., Temlyakov, V., and T. Ullrich, (2018), "Hyperbolic Cross Approximation", Advanced Courses in Mathematics, CRM Barcelona, Centre de Recerca Matemàtica, Birkhäuser.

Shen, J., L-L. Wang, (2010), "Sparse Spectral Approximations of High-Dimensional Problems Based on Hyperbolic Cross" SIAM Journal on Numerical Analysis, 48, (3), pp. 1087-1109.