While training a multivariate linear regression in Accord.NET, I get an exception during training when I attempt to make the regression robust.

I am attempting to train a regression from a set of inputs M to a set of outputs N. Every time it tries to Learn, it throws a System.IndexOutOfRangeException. I've traced it to the internal JaggedSingularValueDecomposition. It seems that while solving the decomposition of the matrix, there is a hard assumption that the output value y is a single column. This obviously does not work for me, or I am doing it wrong.

As an additional note, I have many more training samples than either my set of explanatory variables or my response variables.

Is there something I'm doing wrong here? When doing non-robust regression the training works, the only issue is the least squares decomposition. I've looked at the iterative/generalized models but none of the support multiple outputs.

Simple example:

var trainingInputs = new double[][]
        {
            new double[] { 1, 2, 3 },
            new double[] { 2, 3, 4 },
            new double[] { 3, 4, 5 },
            new double[] { 4, 5, 6 },
            new double[] { 5, 6, 7 },
            new double[] { 6, 7, 8 },
        };

var trainingOutputs = new double[][]
        {
            new double[] { 3, 4 },
            new double[] { 4, 5 },
            new double[] { 5, 6 },
            new double[] { 6, 7 },
            new double[] { 7, 8 },
            new double[] { 8, 9 },
        };

var ols = new OrdinaryLeastSquares() { IsRobust = true };
ols.Learn(inputs, outputs); // Throws IndexOutOfRangeException

I've attempted a workaround by instead manually calculating the JaggedSingularValueDecomposition, then solving it for each column of output samples individually. This yields me an array of coefficients and inputs that I can use.

Is this a valid method? Are there any caveats to this?