Jacobi iteration

The simplest method is the Jacobi iteration.

$$\begin{array}{c}A=D-L-U,M=D\\ D{x}^{(k+1)}=b+(L+U)x^{(k)}\end{array}$$

• We can check that it follows the general framework of splitting methods.
• Very easy to implement
• Fast iteration

Code

for k = 1:maxiter
    xnew = zeros(n)
    for i = 1:n
        for j = 1:n
            if j != i
                xnew[i] += A[i,j] * x[j]
            end
        end
        xnew[i] = (b[i] - xnew[i]) / A[i,i]
    end
    x = xnew
end

Convergence of Jacobi

If $A$ is strictly row diagonally dominant

$$|a_{ii}|>\sum _{j\ne i}|a_{ij}|$$

then Jacobi converges. We have a similar result if $A$ is strictly column diagonally dominant.

• Jacobi is very easy to implement on parallel computers because $D$ is diagonal (or block diagonal).
• Solution step is parallel, i.e., entries of $x^{(k+1)}$ can be computed concurrently using independent computing cores:

$$D{x}^{(k+1)}=b+(L+U)x^{(k)}$$

Proof

The iteration matrix in Jacobi is

$$G=D^{-1}(L+U)$$

Assume that $A$ is row-diagonally dominant. Then from the definition of the operator $\infty$ norm, we find that

$$\Vert G{\Vert }_{\infty }<1$$

This proves the convergence of Jacobi since

$$\Vert G^k{\Vert }_{\infty }\le \Vert G{\Vert }_{\infty }^k\to 0$$

Assume now that $A$ is column-diagonally dominant. Then:

$$G^k=D^{-1}[(L+U)D^{-1}{]}^{k}D$$

But

$$\Vert (L+U)D^{-1}{\Vert }_1<1$$

So

$$\Vert ((L+U)D^{-1}{)}^k{\Vert }_1\le \Vert (L+U)D^{-1}{\Vert }_1^k\to 0$$

and $\Vert G^k{\Vert }_1\to 0$.

There is also a proof that uses the Gershgorin Circle Theorem.

Gershgorin Circle Theorem. For any $n\times n$ complex matrix $A$, each eigenvalue of $A$ lies within at least one of the Gershgorin discs in the complex plane. The $i$-th Gershgorin disc is centered at $a_{ii}$ with radius $R_i={\sum }_{j\ne i}|a_{ij}|$. Mathematically, this is expressed as:

$$|\lambda -a_{ii}|\le R_i$$

where $\lambda$ represents an eigenvalue of $A$.

Convergence proof using Gershgorin’s Theorem. $G=D^{-1}(L+U)$ is the iteration matrix in Jacobi. For the Jacobi method to converge, the spectral radius $\rho (G)$ must be less than 1.

Applying Gershgorin’s Theorem. The diagonal elements of $G$ are zero, and the off-diagonal elements are $-\frac{a_{ij}}{a_{ii}}$. The sum of the absolute values of the off-diagonal elements in the $i$-th row is:

$$\sum _{j\ne i}\left|-\frac{a_{ij}}{a_{ii}}\right|=\frac{1}{|a_{ii}|}\sum _{j\ne i}|a_{ij}|$$

If $A$ is strictly diagonally dominant, then:

$$|a_{ii}|>\sum _{j\ne i}|a_{ij}|$$

This implies:

$$\frac{1}{|a_{ii}|}\sum _{j\ne i}|a_{ij}|<1$$

Therefore, each Gershgorin disc for $G$ is centered at zero with a radius less than 1. Consequently, all eigenvalues of $G$ lie within the unit circle.

Since all eigenvalues of $G$ have magnitudes less than 1, the spectral radius $\rho (G)<1$. This ensures that the Jacobi iteration converges for any initial guess $x^{(0)}$.

The same proof applies for column diagonally dominant matrices.