Christoffel subsampling for the Ishigami functionΒΆ

Optimal sampling draws the design points from the Christoffel distribution associated to the approximation space instead of the reference measure. Here the function is the Ishigami model and the reference measure is its natural distribution, the independent uniform law on [-\pi, \pi]^3. The approximation space is spanned by tensor Legendre polynomials, orthonormal with respect to that measure, enumerated in linear (total-degree) order. At equal budget, is the Christoffel subsample more accurate than the natural sample? We answer by repeating both fits and comparing the distributions of their squared L^2 errors.

Two practical choices shape the protocol. With an inexhaustible basis the experiment works with m = n functions and samples a projection DPP sequentially; the cost grows quickly with the budget. We use 30 points (a fraction of a second per design) instead, which already separates the two methods by two orders of magnitude. Likewise the error is measured against the best approximation in the 30-dimensional space rather than the function itself: the truncation error is then identical for both arms and the comparison isolates the design quality. That projection is computed once, quasi-exactly, by weighted least squares on a tensor Gauss rule; each repetition is then scored by the coefficient distance to it, exact by orthonormality, instead of costly quadrature.

import openturns as ot
import openturns.experimental as otexp
import openturns.viewer as otv
from openturns.usecases import ishigami_function

ot.RandomGenerator.SetSeed(0)

We load the Ishigami model and build the tensor Legendre basis orthonormal with respect to its natural distribution. The experiment deduces the space dimension from the budget: with this inexhaustible basis it is the budget itself. The reference projection is the weighted least squares fit on a tensor Gauss rule, quasi-exact for this smooth model: rules of 16 and 20 points per direction agree on the 60 coefficients to 1e-12.

im = ishigami_function.IshigamiModel()
measure = im.distribution
model = im.model
enumerateFunction = ot.LinearEnumerateFunction(measure.getDimension())
basis = ot.OrthogonalBasis(
    ot.OrthogonalProductPolynomialFactory(
        [measure.getMarginal(i) for i in range(measure.getDimension())],
        enumerateFunction,
    )
)
budget = 30
experiment = otexp.ChristoffelSubsampleExperiment(basis, budget)
expansionSize = experiment.getSpaceDimension()


def computeProjection(level):
    nodes, weights = ot.GaussProductExperiment(
        measure, [level] * 3
    ).generateWithWeights()
    projection = ot.LeastSquaresExpansion(
        nodes, ot.Point(weights), model(nodes), measure, basis, expansionSize
    )
    projection.run()
    return projection.getResult().getCoefficients()


coarse = computeProjection(16)
fine = computeProjection(20)
assert max(abs(x) for x in coarse.asPoint() - fine.asPoint()) < 1e-12
reference = coarse

Each repetition fits the same truncated space twice: once on a Christoffel subsample with its density-ratio weights, once on a natural sample with uniform weights. Both fits use LeastSquaresExpansion on the first expansionSize basis functions, and each is scored by its squared coefficient distance to the reference projection, which is the squared L^2 distance by orthonormality. At m = n the square DPP design systematically misses the default frame band, so the loop runs with warnings silenced; the error distributions below show the design still wins by far.

repetitions = 60
christoffelErrors = ot.Sample(repetitions, 1)
naturalErrors = ot.Sample(repetitions, 1)


def coefficientError(expansion):
    gap = expansion.getResult().getCoefficients() - reference
    return gap.computeRawMoment(2)[0] * gap.getSize()


ot.Log.Show(ot.Log.ERROR)
for repetition in range(repetitions):
    sample, weights = experiment.generateWithWeights()
    observations = model(sample)
    expansion = ot.LeastSquaresExpansion(
        sample, ot.Point(weights), observations, measure, basis, expansionSize
    )
    expansion.run()
    christoffelErrors[repetition, 0] = coefficientError(expansion)
    naturalSample = measure.getSample(budget)
    naturalObservations = model(naturalSample)
    naturalExpansion = ot.LeastSquaresExpansion(
        naturalSample, naturalObservations, measure, basis, expansionSize
    )
    naturalExpansion.run()
    naturalErrors[repetition, 0] = coefficientError(naturalExpansion)
print(f"median Christoffel error={christoffelErrors.computeQuantile(0.5)[0]:.3e}")
print(f"median natural error={naturalErrors.computeQuantile(0.5)[0]:.3e}")
ot.Log.Show(ot.Log.DEFAULT)
median Christoffel error=7.994e+01
median natural error=6.216e+03

Squared errors live on [0, +\infty), so each density is estimated with a KernelSmoothing truncated at the lower bound 0. The two densities are drawn on the same doubly logarithmic figure: the Christoffel errors concentrate two orders of magnitude below the natural ones, and the log scale resolves both bodies and tails. The kernel estimates decay far below the plotted range in their extreme tails, so the vertical axis is clipped at 10^{-8} to keep both distributions readable.

Squared L2 error distributions

Display all figures