Solving matrix equations of the form $X = AXA^T + C$

The last displayed line is wrong. The RHS is indeed $\begin{bmatrix} 18.8802 & -11.3672 \\ -11.3672 & 9.5013 \end{bmatrix}$, which is $X$. So, your solution for $X$ is correct.


Your solution for $X$ is correct.

Here is another solution (it is cheating, though, because I used Octave). On he plus side, it shows another approach to solving the equation.

$A$ is diagonalizable, and so has a basis of eigenvectors, $u_1,u_2$. Define the linear operator $L(X) = X-A X A^*$. You want to solve $L(X) = C$. Note that $u_i u_j^*$ forms a basis for the space of $2 \times 2 $ matrices, and $L(u_i u_j^*) = (1-\lambda_i \overline{\lambda_j}) u_i u_j^*$, hence $L$ is diagonalizable.

To express $C$ in terms of $u_i u_j^*$, let $C = \sum_{i,j} [\Gamma]_{ij} u_i u_j^*$, and solve for $\Gamma$. Let $U$ be the matrix whose columns are $u_1 ,u_2$. Note that $\sum_i [\Gamma]_{ij} u_i = U \Gamma e_j$, so we have $C = \sum_{j} U \Gamma e_j u_j^* = \sum_{j} U \Gamma e_j (U e_j)^* = \sum_{j} U \Gamma e_j e_j^T U^* = U \Gamma (\sum_{j} e_j e_j^T) U^* = U \Gamma U^*$. It follows that $\Gamma = U^{-1} C (U^{-1})^*$, and if we solve $L(\sum_{i,j} [\tilde{X}]_{ij} u_i u_j^*) = \sum_{i,j} [\Gamma]_{ij} u_i u_j^*$, we get $[\tilde{X}]_{ij} = \frac{1}{1-\lambda_i \overline{\lambda_j}} [\Gamma]_{ij}$.

Reversing the procedure to obtain $X$ in the standard basis, we have $X = U \tilde{X} U^*$.

Using Octave we obtain the same result as above.

a = [1.5 1 ; -0.7 0 ]
c = [1 0.5 ; 0.5 0.25 ]

[u,d] = eig(a)

l = [d(1,1) ; d(2,2) ]

phi = ones(2,2)-l*l'

gamma = inv(u)*c*inv(u')

x_tilde = (1./phi).*gamma
x = u*x_tilde*u'

# check result...
a*x*a'+c-x

Here is the output:

octave:44> a*x*a'+c-x
ans =

   0.00000 - 0.00000i   0.00000 - 0.00000i
   0.00000 - 0.00000i   0.00000 + 0.00000i

octave:45> x
x =

   18.8802 +  0.0000i  -11.3672 -  0.0000i
  -11.3672 -  0.0000i    9.5013 +  0.0000i

Given $\mathrm A, \mathrm B \in \mathbb R^{n \times n}$, we have a linear matrix equation in $\mathrm X \in \mathbb R^{n \times n}$

$$\rm X = A X A^{\top} + B$$

Vectorizing both sides, we obtain a system of $n^2$ linear equations in $n^2$ unknowns

$$\left( \mathrm I_{n^2} - \left( \mathrm A \otimes \mathrm A \right) \right) \mbox{vec} (\mathrm X) = \mbox{vec} (\mathrm B)$$


Example

Let

$$\mathrm A := \begin{bmatrix} 1.5 & 1 \\ -0.7 & 0\end{bmatrix} \qquad \qquad \qquad \mathrm B := \begin{bmatrix} 1 & 0.5 \\ 0.5 & 0.25\end{bmatrix}$$

Using NumPy, we solve the linear system and then un-vectorize the solution:

>>> import numpy as np
>>> A = np.array([[ 1.5, 1.0],
                  [-0.7, 0.0]])
>>> B = np.array([[ 1.0, 0.5],
                  [ 0.5, 0.25]])
>>> I4 = np.identity(4)
>>> solution = np.linalg.solve(I4 - np.kron(A,A), np.reshape(B, 4))
>>> X = np.reshape(solution, (2,2))
>>> X
array([[ 18.88020833, -11.3671875 ],
       [-11.3671875 ,   9.50130208]])