Properties of Discrete-time Algebraic Riccati Equation (DARE)


This post serves as a test of emacs org-mode and jekyll static site generator.

Preliminary

We will denote \(\mathbb S_n,\,\mathbb S_n^+,\mathbb S_n^{++}\) to be the set of all real symmetric matrices, the set of all positive semi-definite matrices and the set of all positive definite matrices respectively. For any two matrices \(A,\,B\in \mathbb S_n\), we call \(A \geq B\) (\(A > B\)) if \(A - B\) is positive semi-definite (positive definite).

Let \(A\in \mathbb R^{n\times n}\) and \(C \in \mathbb R^{m\times n}\) be two matrices. Further define \(Q \in \mathbb S_n\) and \(R\in\mathbb S_n\) to be two positive definite matrices. The DARE function \(g: \mathbb S_n^+\rightarrow\mathbb S_n^+\) is defined as

\begin{equation} \label{eq:riccati1} g(X) \triangleq \left[\left(A X A' + Q\right)^{-1} + C'R^{-1}C\right]^{-1}. \end{equation}

From Equation \eqref{eq:riccati1}, we can see that \(g\) is monotonically increasing with respect to \(X\).

Using the matrix inversion lemma, we can write \(g\) as

\begin{equation} \label{eq:riccati2} g(X) = h(X) - h(X) C' \left(C h(X) C'+R\right)^{-1}Ch(X), \end{equation}

where \(h(X) = AXA' + Q\). For simplicity, we will define

\begin{equation} \tilde g(X) = X - X C' \left(C X C'+R\right)^{-1}CX. \end{equation}

Hence, \(g = \tilde g \circ h\).

Concavity of \(g\) with respect to \(X\)

We only need to show that \(\tilde g\) is concave since \(h\) is linear with respect to \(X\). Consider the following function

\begin{align} \label{eq:linear} \phi (K,X) &\triangleq (I-KC) X (I-KC)' + KRK'\\ & = X - KCX - XC'K' + K(CXC'+R)K'. \end{align}

By completing the square, \(\phi(K,X)\) can be rewritten as

\begin{equation} \label{eq:linear2} \phi(K,X) = \tilde g(X) + (K-K_*) (CXC'+R)^{-1}(K-K_*)', \end{equation}

where \(K_* = XC'(CXC'+R)^{-1}\). Therefore,

\begin{align} \label{eq:riccatilinear} \tilde g(X) = \min_{K} \phi(K,X). \end{align}

Since \(\phi(K,X)\) is linear with respect to \(X\) for a fixed \(K\), \(\tilde g(X)\) is concave with respect to \(X\).

The fixed point \(X = g(X)\)

It is well known that if the matrix pair \((A,C)\) is detectable, then the equation \(X = g(X)\) has a unique positive definite solution, which we will denote as \(X_*\).

Let \(X_0 = 0\) and denote \(X_{k+1} = g(X_k)\). Clearly \(X_0 \leq X_*\) and \(X_0 \leq X_1\). Now apply \(g\) on both sides of the inequalities, we have \[ g(X_0) \leq g(X_*),\,g(X_0)\leq g(X_1). \] Notice that \(g(X_0) = X_1,\,g(X_1) = X_2\) and \(g(X_*) = X_*\). Therefore, we have the following chain of inequalities \[ X_0 \leq X_1 \leq X_2 \leq \dots \leq X_*. \]

Since \(\{X_k\}\) is an increasing sequence with an upper bound, \(\{X_k\}\) has a limit, which we can denote as \(X_\infty\). Now by the fact that \(g\) is continuous, \(X_\infty = g(X_\infty)\). Finally, be the fact that \(X_*\) is the unique fixed point, we have that \(X_\infty = X_*\).

Similar trick can be applied to the case where we can find an \(X_1 = g(X_0) \leq X_0\). In that case, we know that \(X_*\) must be no greater than \(X_0\).

Therefore, we can establish the equivalency of the following statements:

  1. \(X_* \leq P\)
  2. There exists an \(X_0 \leq P\), such that \(X_1 = g(X_0) \leq X_0\).

We already proved that Statement 2 implies 1. On the other hand, if Statement 1 holds, then \(X_* = g(X_*)\leq X_*\).

Now if we define \(S = X_0^{-1}\), then \(X_0 \leq P\) is equivalent to \(S \geq P^{-1}\). Moreover, \(g(X_0)\leq X_0\) is equivalent to

\begin{align} \label{eq:fixpointequivalence} (A X A' + Q)^{-1} + C'R^{-1}C \geq S \end{align}

Using matrix inversion lemma on the LFS of \eqref{eq:fixpointequivalence}, we have \[ Q^{-1} - Q^{-1}A(S + A'Q^{-1}A)^{-1}A'Q^{-1} + C'R^{-1}C\geq S, \] which is equivalent to

\begin{align} \label{eq:fixpointequivalence2} Q^{-1} - S + C'R^{-1}C \geq Q^{-1}A(S + A'Q^{-1}A)^{-1}A'Q^{-1}. \end{align}

Now using Schur's complement, \eqref{eq:fixpointequivalence2} is equivalent to the following linear matrix inequality:

\begin{align} \label{eq:lmi} \begin{bmatrix} Q^{-1} - S + C'R^{-1}C&Q^{-1}A\\ A'Q^{-1}& S+A'Q^{-1}A \end{bmatrix} \geq 0. \end{align}

As a result, the condition that \(X_* \leq P\) is equivalent to \eqref{eq:lmi} and \(S\geq P^{-1}\). Notice that the LFS of \eqref{eq:lmi} is linear with respect to \(C'R^{-1}C\). This fact can leveraged for sensor placement problems.

Linear Fractional Transformations and Riccati Equation

Consider the following functions:

\begin{align} \label{eq:linearfractional} f_1(x) = \frac{a_1x+b_1}{c_1x+d_1}, f_2(x) = \frac{a_2x+b_2}{c_2x+d_2}. \end{align}

Then

\begin{align} \label{eq:linearfractionalcomposition} f_1(f_2(x)) &= \frac{a_1 \frac{a_2x+b_2}{c_2x+d_2} + b_1}{c_1 \frac{a_2x+b_2}{c_2x+d_2} + d_1}\\ &=\frac{(a_1a_2+b_1c_2)x + a_1b_2 + b_1d_2}{(c_1a_2+d_1c_2)x + c_1b_2 + d_1d_2}. \end{align}

One important observation is that the coefficients on the numerator and the denominator can be computed via matrix multiplication:

\begin{align} \begin{bmatrix} a_1&b_1\\ c_1&d_1 \end{bmatrix}\times\begin{bmatrix} a_2&b_2\\ c_2&d_2 \end{bmatrix} = \begin{bmatrix} a_1a_2+b_1c_2&a_1b_2+b_1d_2\\ c_1a_2+d_1c_2&c_1b_2+d_1d_2 \end{bmatrix}. \end{align}

We will now move to the case where \(X\) is a matrix instead of just a scalar. Let us consider the following matrix: \[ \Phi = \begin{bmatrix} \Phi_{11}&\Phi_{12} \\ \Phi_{21}&\Phi_{22} \end{bmatrix} \in \mathbb R^{2n\times 2n}, \] where each \(\Phi_{ij}\in \mathbb R^{n\times n}\). Define the following operator

\begin{align} \label{eq:homographic} F(\Phi,X) \triangleq (\Phi_{11}X+\Phi_{12})(\Phi_{21}X+\Phi_{22})^{-1}. \end{align}

Similarly for any \(\Phi_1,\,\Phi_2 \in \mathbb R^{2n\times 2n}\), we have

\begin{align} \label{eq:homographiccomposition} F(\Phi_1,F(\Phi_2,X)) = F(\Phi_1\Phi_2,X). \end{align}

We now establish the link between linear fractional transformations and Riccati equations. We will assume that \(A\) is invertible. We know that

\begin{align} \label{eq:riccatidecompose1} AXA' &= F\left(\begin{bmatrix} A & 0\\ 0 & (A')^{-1} \end{bmatrix}, X\right),\\ \label{eq:riccatidecompose2} (X + Q)^{-1} &= F\left(\begin{bmatrix} 0 & I\\ I & Q \end{bmatrix}, X\right),\\ \label{eq:riccatidecompose3} (X + C'R^{-1}C)^{-1} &= F\left(\begin{bmatrix} 0 & I\\ I & C'R^{-1}C \end{bmatrix}, X\right). \end{align}

As a result, using \eqref{eq:homographiccomposition}, we have

\begin{align} \label{eq:riccatihomo} g(X) = F(\Phi_g,X), \end{align}

where

\begin{align} \label{eq:Phimatrixg} \Phi_g \triangleq \begin{bmatrix} A&Q(A')^{-1} \\ C^TR^{-1}CA&(I+C^TR^{-1}CQ)(A')^{-1} \end{bmatrix}. \end{align}

Hence, \(g^{(n)}(X) = F(\Phi_g^n,X)\).

Contraction Property of Riccati Equation

We will simply state several key theorems in this paper.

Consider the following function \(\delta: \mathbb S_n^{++}\times \mathbb S_n^{++}\rightarrow \mathbb R\) defined as

\begin{align} \label{eq:metric} \delta(X,Y) = \sqrt{\sum_{i=1}^n \left(\log \lambda_i\right)^2}, \end{align}

where \(\lambda_1,\dots,\lambda_n\) are the eigenvalues of \(XY^{-1}\). One can prove that it is a metric on the space of \(\mathbb S_n^{++}\). The main property of this metric is that it is invariant under conjugacy and inversion, i.e.,

\begin{align} \label{eq:invariance} \delta(X^{-1},Y^{-1}) =\delta(X,Y) = \delta (AXA',AYA'). \end{align}

Moreover, if a positive semi-definite matrix \(P \geq 0\) and the smallest eigenvalue of \(P\) is \(\beta\), then

\begin{align} \label{eq:contraction} \delta(X+P,Y+P)\leq \frac{\alpha}{\alpha+\beta}\delta(X,Y) \leq \delta(X,Y). \end{align}

where \(\alpha\) is the largest eigenvalue of \(X\) and \(Y\).

Therefore, from \eqref{eq:riccatidecompose1}, \eqref{eq:riccatidecompose2} and \eqref{eq:riccatidecompose3} we know that

\begin{align} \label{eq:riccaticontraction} \delta(g(X),g(Y))\leq \delta(X,Y). \end{align}

It can be further proved that if \((A,C)\) is observable, then there exists a \(k\) and \(0 < \rho < 1\), such that

\begin{align} \label{eq:riccaticontraction2} \delta(g^{(k)}(X),g^{(k)}(Y))\leq\rho \delta(X,Y). \end{align}