Julia library for 1-d and 2-d splines
This is a Julia wrapper for the dierckx Fortran library, the same library underlying the spline classes in scipy.interpolate. Some of the functionality here overlaps with Grid.jl, a pure-Julia interpolation package. This package is intended to complement Grid.jl and to serve as a benchmark in cases of overlapping functionality.
- Implements B-splines (basis splines).
- Splines from first order to fifth order; default is third order (cubic).
- Fit and evaluate 1-d and 2-d splines on irregular grids.
- Fit and evaluate 2-d splines at unstructured points.
- Fit "smooth" (non-interpolating) splines with adjustable smoothing factor s.
- Derivatives of 1-d splines.
julia> Pkg.add("Dierckx")
The Fortran library source code is distributed with the package, so
you need a Fortran compiler on OSX or Linux. On Ubuntu,
sudo apt-get install gfortran
will do it.
On Windows, a compiled dll will be downloaded.
using Dierckx
Fit a 1-d spline to some input data (points can be unevenly spaced):
x = [0., 1., 2., 3., 4.]
y = [0., 1., 8., 27., 64.]
spl = Spline1D(x, y)
Evaluate the spline at some new points:
evaluate(spl, [1.5, 2.5]) # result = [3.375, 15.625]
evaluate(spl, 1.5) # result = 3.375
Evaluate derivative:
derivative(spl, 1.5) # result = 6.75
Fit a 2-d spline to data on a (possibly irregular) grid:
x = [0.5, 2., 3., 4., 5.5, 8.]
y = [0.5, 2., 3., 4.]
z = [1. 2. 1. 2.; # size is (length(x), length(y))
1. 2. 1. 2.;
1. 2. 3. 2.;
1. 2. 2. 2.;
1. 2. 1. 2.;
1. 2. 3. 1.]
spline = Spline2D(x, y, z)
Evaluate at element-wise points:
xi = [1., 1.5, 2.3, 4.5, 3.3, 3.2, 3.]
yi = [1., 2.3, 5.3, 0.5, 3.3, 1.2, 3.]
zi = evaluate(spline, xi, yi) # 1-d array of length 7
Evaluate at grid spanned by input arrays:
xi = [1., 1.5, 2.3, 4.5]
yi = [1., 2.3, 5.3]
zi = evalgrid(spline, xi, yi) # 2-d array of size (4, 3)
Spline1D(x, y; w=ones(length(x)), k=3, bc="nearest", s=0.0)
Spline1D(x, y, xknots; w=ones(length(x)), k=3, bc="nearest")
-
Create a spline of degree
k
(1 = linear, 2 = quadratic, 3 = cubic, up to 5) from 1-d arraysx
andy
. The parameterbc
specifies the behavior when evaluating the spline outside the support domain, which is(minimum(x), maximum(x))
. The allowed values are"nearest"
,"zero"
,"extrapolate"
,"error"
.In the first form, the number and positions of knots are chosen automatically. The smoothness of the spline is then achieved by minimalizing the discontinuity jumps of the
k
th derivative of the spline at the knots. The amount of smoothness is determined by the condition thatsum((w[i]*(y[i]-spline(x[i])))**2) <= s
, withs
a given non-negative constant, called the smoothing factor. The number of knots is increased until the condition is satisfied. By means of this parameter, the user can control the tradeoff between closeness of fit and smoothness of fit of the approximation. ifs
is too large, the spline will be too smooth and signal will be lost ; ifs
is too small the spline will pick up too much noise. in the extreme cases the program will return an interpolating spline ifs=0.0
and the weighted least-squares polynomial of degreek
ifs
is very large.In the second form, the knots are supplied by the user. There is no smoothing parameter in this form. The program simply minimizes the discontinuity jumps of the
k
th derivative of the spline at the given knots.
evaluate(spl, x)
- Evalute the 1-d spline
spl
at points given inx
, which can be a 1-d array or scalar. If a 1-d array, the values must be monotonically increasing.
derivative(spl, x; nu=1)
- Evaluate the
nu
-th derivative of the spline at points inx
.
Spline2D(x, y, z; w=ones(length(x)), kx=3, ky=3, s=0.0)
Spline2D(x, y, z; kx=3, ky=3, s=0.0)
-
Fit a 2-d spline to the input data.
x
andy
must be 1-d arrays.If
z
is also a 1-d array, the inputs are assumed to represent unstructured data, withz[i]
being the function value at point(x[i], y[i])
. In this case, the lengths of all inputs must match.If
z
is a 2-d array, the data are assumed to be gridded:z[i, j]
is the function value at(x[i], y[j])
. In this case, it is required thatsize(z) == (length(x), length(y))
.
evaluate(spl, x, y)
- Evalute the 2-d spline
spl
at points(x[i], y[i])
. Inputs can be Vectors or scalars. Points outside the domain of the spline are set to the values at the boundary.
evalgrid(spl, x, y)
- Evaluate the 2-d spline
spl
at the grid points spanned by the coordinate arraysx
andy
. The input arrays must be monotonically increasing.
The Spline
classes in scipy.interpolate are also thin wrappers
for the Dierckx Fortran library. The performance of Dierckx.jl should
be similar or better than the scipy.interpolate classes. (Better for
small arrays where Python overhead is more significant.) The
equivalent of a specific classes in scipy.interpolate:
scipy.interpolate class | Dierckx.jl constructor method |
---|---|
UnivariateSpline | Spline1D(x, y; s=length(x)) |
InterpolatedUnivariateSpline | Spline1D(x, y; s=0.0) |
LSQUnivariateSpline | Spline1D(x, y, xknots) |
SmoothBivariateSpline | Spline2D(x, y, z; s=length(x)) |
LSQBivariateSpline | |
RectBivariateSpline | Spline2D(x, y, z; s=0.0) (z = 2-d array) |
SmoothSphereBivariateSpline | |
LSQSphereBivariateSpline | |
RectSphereBivariateSpline |
The primary benefit of Dierckx is speed. Here are some benchmark results relative to Grid.jl v0.3.6 (Julia v0.3.3).
benchmark | Grid.jl | Dierckx.jl | ratio |
---|---|---|---|
eval 1-d grid n=10 | 710.46 ns | 399.69 ns | 1.78 |
eval 1-d grid n=1000 | 62.86 us | 26.02 us | 2.42 |
eval 1-d grid n=100000 | 6.20 ms | 2.60 ms | 2.38 |
eval 2-d grid 10x10 | 11.09 us | 2.27 us | 4.89 |
eval 2-d grid 100x100 | 1.07 ms | 132.88 us | 8.08 |
eval 2-d grid 1000x1000 | 110.78 ms | 12.87 ms | 8.61 |
Notes: This is for second order (quadratic) splines. CoordInterpGrid
is used in Grid.jl, which doesn't allow arbitrary knot positions as in
Dierckx.jl. Results obtained by running benchmark.jl
script on a
Core i7-4500U CPU @ 1.80GHz.
Dierckx.jl is distributed under a 3-clause BSD license. See LICENSE.md for details. The real*8 version of the Dierckx Fortran library as well as some test cases and error messages are copied from the scipy package, which is distributed under this license.