Fig. 3
Standard error computation. a Empirical runtime for the computation of standard error vector σ2 is plotted in log-scale against increasing number of parameters (also log-scale). Computation based on sparse inverse subset algorithm (orange line) achieves linear runtime in p (dotted line p), the number of parameters, compared to quadratic complexity (dotted line p2) of the “indirect” method (blue line). b Memory consumption in MByte for the computation of standard error vector σ2 is plotted against number of parameters. Though both methods achieve linear memory consumption in p, the slope of the “indirect” method (blue line) is around 4 times greater than of the sparse inverse subset algorithm (orange line) likely due to the recursive computation of the inverse instead of solving of a triangular system