Preconditioned CG

The conjugate gradient method is one of the most powerful methods for the solutions of linear systems $Ax_\star=b$, where $A$ is a symmetric and positive definite matrix. It is particularly convenient for large and sparse matrices as, at each step, the only operation involving $A$ we are required to do is the product of $A$ times a vector. Recall that the convergence rate of the method can be estimated by looking at the energy norm of the error $E_{k} = x_k -x_\star$. The following upper bound holds \begin{equation}\label{eq:CG-conv-rate} \frac{\|E_{k}\|_A}{\|E_0\|_A}\leq 2\left(\frac{\sqrt{\kappa(A)}-1}{\sqrt{\kappa(A)}+1}\right)^k, , \end{equation} being $\kappa(A)=\lambda_{\max}(A)/\lambda_{\min}(A)$ the $\ell^2$ condition number of $A$.

Given a matrix $A$, in this lab you are asked to:

• Fix a solution vector $x_\star$ (it can be random or any fixed vector you like) and define the vector $b= Ax_\star$
• Design the Conjugate Gradient routine to address the solution of the system $Ax=b$
• Build from $A$ the Jacobi preconditioner $P$
• Solve the linear system $Ax=b$ with the Preconditioned Conjugate Gradient method, using the Jacobi and the Incomplete Cholesky preconditioners
• Generate convergence plots in semi-log scale of the conjugate gradient iteration for all the test problems. The plots should display the Log10 of the relative norm of the residual $r_k$ vs. the iteration number $k$, and should include different choices of preconditioners, including the unpreconditioned case $P=I$
• Generate plots that compare the convergence behavior of the PCG residual in comparison with the real residual $r_k=b-Ax_k$ and the theoretical bound \eqref{eq:CG-conv-rate}.

The matrix $A$ you should consider to work with are:

1. Diagonal $100\times 100$ matrix with eigenvalues $\lambda_1=5$, $\lambda_2=4$, $\lambda_3 = 1.5$, $\lambda_4=1.4$, $\lambda_5=1.3$ and $\lambda_6=\cdots=\lambda_{100}=1$.

2. $$ \begin{cases} A_{ii} = \sqrt{i} & \newline A_{ij} = 1, & \text{if } |i-j| = k\newline A_{ij} = 0 & \text{otherwise} \end{cases} $$

3.

• nos4, $100\times 100$ symmetric positive definite matrix, estimated condition number 1.5e+03, download it from its page on Suite Sparse Matrix Collection
• nos6, $675\times 675$ symmetric positive definite matrix, estimated condition number 7.65e+06, download it from its page on Suite Sparse Matrix Collection
• bcsstk14, $1806\times 1806$ symmetric positive definite matrix, estimated condition number 1.19e+10, download it from its page on Suite Sparse Matrix Collection

Pseudocode of the PCG algorithm

function x = my_pcg(A,P,b,x0,maxiter,tol)
%  Input: A, SPD matrix
%         P, preconditioner
%         b, rhs vector
%         x0, starting vector
%         maxiter, max number of iterations allowed
%         tol, convergence tolerance
% Output: x_approx, approximation of the solution to Ax=b

M = P^{-1};
r0 = b - A*x0;
p1 = M*r0;
for k = 1 : maxiter
    a_k = (r_{k-1}, M*r_{k-1}) /  (p_k,A*p_k);
    x_k = x_{k-1} + a_k * p_k;
    r_k = r_{k-1} - a_k * A * p_k;
    b_{k+1} = (r_k, M*r_k) / (r_{k-1}, M*r_{k-1});
    p_{k+1} = P^{-1}*r_k + b_{k+1}*p_k;

    if tol has been reached
        x_approx = x_k;
        break;
    end
end

end

Implementation notes

Note that the method can (has to) be implemented so that only four vectors are allocated in the memory, overall. Moreover, note that the most costly operation per step is $A\times vector$, which requires $O(nnz)$ flops, being $nnz$ the number of non-zero entries of $A$.

You can plot the norm of the residual $\|r_m\|$ and the norm of the true residual $\|Ax_m-b\|$ and compare them. You should observe that their behavior is different due to machine precision. Also, you should observe that the upper-bound \eqref{eq:CG-conv-rate} is quite loose, but still, when $\kappa(A)\gg 1$ the method can be very slow.