Figure 1.

The number of conjugate gradient iterations grows slightly worse than $\sqrt{\kappa (P^{(1)})}$. The stopping criterion is a tolerance of 10⁻⁶ for the norm of the relative residual. All matrices are rank 12, with smaller eigenvalues as κ(P⁽¹⁾) increases, but identical eigenvectors. The same right-hand side is solved for all cases.