using Schur's complement and Young's inequality to reduce matrix algebraic equation to LMI

302 Views Asked by At

This questions concerns the practical implementation of schur's complement and Young's inequality.

Consider the following

\begin{align} \begin{pmatrix} \begin{pmatrix} \mathbb{A}_{\mathbb{Z}} & \mathcal{O}\\ \mathcal{O} & \mathbb{A}_{\mathbb{P}} \end{pmatrix} + \begin{pmatrix} \mathbb{Z} & \mathcal{O}\\ \mathcal{O} & \mathcal{I}_{n} \end{pmatrix} \begin{pmatrix} \mathbb{Z} & \mathcal{O}\\ \mathcal{O} & \mathcal{I}_{n} \end{pmatrix} & \begin{pmatrix} D_{11} & D_{12}\\ D_{21} & D_{22} \end{pmatrix}\\ \begin{pmatrix} D_{11} & D_{12}\\ D_{21} & D_{22} \end{pmatrix}^T & -\mu \mathbb{I}_{q} \end{pmatrix} + \underbrace{\begin{pmatrix} \overline{B} \overline{K}\\ \mathcal{O}\\ \mathcal{O} \end{pmatrix}}_{V^T} \underbrace{\begin{pmatrix} \mathcal{O} & \tilde{B} & \mathcal{O} \end{pmatrix}}_U + \begin{pmatrix} \mathcal{O}\\ \tilde{B}^T\\ \mathcal{O} \end{pmatrix} \begin{pmatrix} \left(\overline{B} \overline{K} \right)^T & \mathcal{O} & \mathcal{O} \end{pmatrix} \leq 0 \end{align}

Now to avoid any bilinear coupling between terms, the young's inequality is used, as given by

\begin{equation} X^T \hspace{1mm} Y + Y^T \hspace{1mm} X \leq \dfrac{1}{2} \hspace{1mm} \left( X + S \hspace{1mm} Y \right)^T \hspace{1mm} S^{-1} \hspace{1mm} \left( X + S \hspace{1mm} Y \right) \end{equation}

where $S$ is a symmetric positive definite matrix. To apply the inequality, we use $S = \dfrac{1}{2} \mathbb{I}_n$ for the term $\begin{pmatrix} \mathbb{Z} & \mathcal{O}\\ \mathcal{O} & \mathcal{I}_{n} \end{pmatrix} \begin{pmatrix} \mathbb{Z} & \mathcal{O}\\ \mathcal{O} & \mathcal{I}_{n} \end{pmatrix}$ and $S = \epsilon \mathbb{Z}$ for the term $V^TU+U^TV$. Now, the intent is to use the young's inequality and convert the above algebraic equation in LMI using Schur's complement. I am unsure of the proper way and cannot reduce the equation to the requisite LMI. Given: $\mathbb{Z}$ is a $n \times n$ symmetric positive definite matrix.

Can someone walk me through the subsequent steps?

1

There are 1 best solutions below

6
On

The only term that is not linear in the unknown variables is the quadratic term in $\mathbb{Z}$, so it might be easier to replace that with another variable and add another inequality

$$ \mathbb{Z}^2 \leq \mathbb{X}, \tag{1} $$

such that the initial in equality can be written as

\begin{align} \begin{pmatrix} \begin{pmatrix} \mathbb{A}_{\mathbb{Z}} & \mathcal{O}\\ \mathcal{O} & \mathbb{A}_{\mathbb{P}} \end{pmatrix} + \begin{pmatrix} \mathbb{X} & \mathcal{O}\\ \mathcal{O} & \mathcal{I}_{n}^2 \end{pmatrix} & \begin{pmatrix} D_{11} & D_{12}\\ D_{21} & D_{22} \end{pmatrix}\\ \begin{pmatrix} D_{11} & D_{12}\\ D_{21} & D_{22} \end{pmatrix}^\top & -\mu\,\mathbb{I}_{q} \end{pmatrix} + \begin{pmatrix} \overline{B} \overline{K}\\ \mathcal{O}\\ \mathcal{O} \end{pmatrix} \begin{pmatrix} \mathcal{O} & \tilde{B} & \mathcal{O} \end{pmatrix} + \begin{pmatrix} \mathcal{O}\\ \tilde{B}^\top\\ \mathcal{O} \end{pmatrix} \begin{pmatrix} \left(\overline{B} \overline{K} \right)^\top & \mathcal{O} & \mathcal{O} \end{pmatrix} \leq 0, \tag{2} \end{align}

which is linear in all variables. One can then use Schur's complement on $(1)$ which allows us to rewrite it also as an LMI

$$ \begin{pmatrix} I & \mathbb{Z} \\ \mathbb{Z} & \mathbb{X} \end{pmatrix} \geq 0, \tag{3} $$

with $I$ an $n$ by $n$ identity matrix, $\mathbb{Z} = \mathbb{Z}^\top > 0$ and $\mathbb{X} = \mathbb{X}^\top > 0$.

So the combination of $(2)$ and $(3)$ should be equivalent to your initial inequality and are all LMI's.