Figure 5.

Plot of the residual approximation error. Each curve reports the trajectory of the approximation residuals, computed with the algorithm's loss function. Each track is normalized separately over its maximum value, and stops at the number of iterations required to achieve the convergence criterion.