3 ms·

Is there an analog of SVD where you optimize the L1 / sum of absolute errors, rather than the squares?