I think it's bug here

[henrykey]

Hi, I tried your PCA and crashed at wcenter of you method. I trace your code and find an error when the matrix is not squared. following is the code. In my case, m is 30 and n is 3000, that's m > n. Matrix S is a Transposed one. It becomes a n x m matrix, but your j is within m, the large one, So, S[i][j] will be out of range to crash. I think it's should be S[j][i], or exchange n and m in first loop like second, right?

I want to know when I assign a matrix to PCA, should row for record or column for records? the one I try is 30 records with 3000 features.

thanks.

Henry

/**
 * Center and weight Sample variance for grouped data
 *
 * s^2 = SUM(Mi - xbar)^2 / (n-1)
 *
 * @return
 */
public Matrix wcenter() {
    // center the data
    Matrix s = center().transpose();
    double[][] S = s.getArray();

    // calculate the sample variance 
    double[] sigma = new double[n];
    for (int j = 0; j < n; j++) {
        double ssum = 0;
        for (int i = 0; i < m; i++) {
            ssum += (S[i][j] * S[i][j]);
        }
        sigma[j] = Math.sqrt(ssum / (m - 1));
    }

    // weigh the data
    for (int i = 0; i < m; i++) {
        for (int j = 0; j < n; j++) {
            S[i][j] = S[i][j] / sigma[j];
        }
    }

    return s;
}

[umeding]

Henry,
Thanks for taking a look, however, I am not entirely clear how you run the code. The time series data starts off as a simple double vector (see Main.java), which is then converted into a (symmetrical) Toeplitz matrix representation (see PCAHandler.java and ToeplitzMatrix.java). The Toeplitz matrix is N x N (square), so the values for 'n' and 'm' are the same (must be).

The function of the 'wcenter' method is to relocate and apply a weight to the data so that it ends up equally distributed around (0,0) in a Cartesian coordinate system - this simplifies subsequent calculations.

Hope this helps,
Uwe

[henrykey]

Got it. thanks