Bug in Kurtosis Contrast Function

[armwal]

Hi,

there seems to be a bug in Accord.Statistics.Analysis.ContrastFunctions.Kurtosis.cs

The Evaluate function reads

public void Evaluate(double[] x, double[] output, double[] derivative)
        {
            for (int j = 0; j < x.Length; j++)
            {
                // Kurtosis contrast function and its derivative, as given
                //  in original Hyvärinen's paper. See main references for the
                //  Independent Component Analysis class for details.

                double v = x[j];

                // g(w*x)
                output[j] = v * v * v;

                // g'(w*x)
                derivative[j] = (1.0 / 3.0) * v * v;
            }
        }

However, the derivative should be:

derivative[j] = 3.0 * v * v;

This can be confirmed for example with the ComputeTest() unit test in Accord.Tests.Statistics.IndependentComponentAnalysisTest: Setting ica.Contrast = new Kurtosis(); before computing the ICA fails the test with the old derivative and succeeds with the new one.

Armin

[cesarsouza]

You are absolutely correct! Many thanks for catching (and solving) this issue!