Chapter 10 The Rosenbrock matrix and solvability of the regulator equations

  • Definition 10.1. The Rosenbrock matrix of \(\dot {x}=Ax+Bu\), \(y=Cx+Du\) is the \((n+p)\times (n+m)\) matrix indexed by \(s\in \mC \)

    \[ \bbm {sI-A&-B\\C&D}. \]

  • Proposition 10.2. Let \(s\in \mC \) not be an eigenvalue of \(A\). Then \(G(s)\) is surjective if and only if the Rosenbrock matrix is surjective for this \(s\) and \(G(s)\) is injective if and only if the Rosenbrock matrix is injective for this \(s\).

  • Proof. Assume that \(G(s)\) is surjective. Let \(q\in \mR ^n\) and \(r\in \mR ^p\) be arbitrary. We then have

    \[ \bbm {q\\r}=\bbm {sI-A&-B\\C&D}\bbm {x\\u}, \]

    where \(u\) is a solution of

    \[ G(s)u=r-C(sI-A)^{-1}q, \]

    (which exists by our surjectivity assumption) and

    \[ x=(sI-A)^{-1}q+(sI-A)^{-1}Bu, \]

    since

    \[ \bbm {sI-A&-B\\C&D}\bbm {(sI-A)^{-1}q+(sI-A)^{-1}Bu\\u} =\bbm {q+Bu-Bu\\C(sI-A)^{-1}q+C(sI-A)^{-1}Bu+Du}=\bbm {q\\r}. \]

    Assume that the Rosenbrock matrix is surjective. Let \(r\in \mR ^p\). Then there exist \(x\in \mR ^n\) and \(u\in \mR ^m\) such that

    \[ \bbm {0\\r}=\bbm {sI-A&-B\\C&D}\bbm {x\\u}. \]

    It follows that \(x=(sI-A)^{-1}Bu\) and subsequently that \(r=Cx+Du=C(sI-A)^{-1}Bu+Du=G(s)u\). Hence \(G(s)\) is surjective.

    Assume that \(G(s)\) is injective. Let \(x\in \mR ^n\) and \(u\in \mR ^m\) be such that

    \[ \bbm {sI-A&-B\\C&D}\bbm {x\\u}=\bbm {0\\0}. \]

    This means

    \[ (sI-A)x-Bu=0,\qquad Cx+Du=0. \]

    From the first of these equations we obtain \(x=(sI-A)^{-1}Bu\), which substituted in the second equation gives \(C(sI-A)^{-1}Bu+Du=0\), i.e. \(G(s)u=0\). By the injectivity assumption we then have \(u=0\). From the first of the above equations we then have \((sI-A)x=0\) which since \(s\) is not an eigenvalue of \(A\) gives \(x=0\). Therefore the Rosenbrock matrix for this \(s\) is injective.

    Assume that the Rosenbrock matrix is injective. Let \(G(s)u=0\). Then

    \[ \bbm {sI-A&-B\\C&D}\bbm {(sI-A)^{-1}Bu\\u}=\bbm {0\\0}. \]

    By the injectivity assumption it follows that \(\sbm {(sI-A)^{-1}Bu\\u}=0\) and in particular that \(u=0\). Hence \(G(s)\) is injective.

  • Remark 10.3. We can re-write the regulator equations as

    \[ \bbm {I&0\\0&0}\bbm {\Pi \\V}A_e-\bbm {A&B_2\\C_1&0}\bbm {\Pi \\V}=\bbm {B_1\\D_{11}}C_e, \]

    which shows that the regulator equations are a generalized Sylvester equation.

    Using the Kronecker product and the vectorization of a matrix, this can in turn be written as the standard linear system

    \begin{equation} \label {eq:Kronecker} \left (A_e^T\otimes \bbm {I_n&0_{n\times m_2}\\0_{p_1\times n}&0_{p_1\times m_2}} -I_{n_e}\otimes \bbm {A&B_2\\C_1&0_{p_1\times m_2}}\right ) \vecc \left (\bbm {\Pi \\V}\right ) =\vecc \left (\bbm {B_1C_e\\D_{11}C_e}\right ). \end{equation}

  • Proposition 10.4. If the Rosenbrock matrix

    \[ \bbm {sI-A&-B_2\\C_1&0}, \]

    is surjective for all eigenvalues \(s\) of \(A_e\), then the regulator equations are solvable.

  • Remark 10.5. A partial converse of Proposition 10.4 is also true in case \(A_e\) has only eigenvalues with nonnegative real part: if the regulator equations are solvable for all \(B_1\), \(D_{11}\) and \(C_e\), then the surjectivity condition on the Rosenbrock matrix \(\sbm {sI-A&-B_2\\C_1&0}\) must hold. This relates to the Kronecker form (10.1) where \(B_1\), \(D_{11}\) and \(C_e\) appear on the right-hand side.

  • Corollary 10.6. If the eigenvalues of \(A_e\) are not eigenvalues of \(A\) and the transfer function \(G(s)=C_1(sI-A)^{-1}B_2\) is surjective for all eigenvalues \(s\) of \(A_e\), then the regulator equations are solvable.

10.1 Examples

  • Example 10.7. Consider the second order scalar differential equation

    \[ \ddot {q}(t)+2\zeta \dot {q}(t)+q(t)=u(t), \]

    where \(\zeta >0\). Let \(x:=\sbm {q\\\dot {q}}\). The objective is to find a control \(u\) such that for a given \(r\in \mR \) we have \(\lim _{t\to \infty }\dot {q}(t)=r\) no matter what \(x(0)\) is.

    We initially define the performance output and exo-system

    \[ z=x_2-w,\qquad \dot {x}_e=0,\qquad w=x_e,\qquad x_e(0)=r, \]

    since then the regulation requirement becomes what we desire.

    We have \(n=2\), \(m_1=m_2=p_1=1\) and

    \[ x=\bbm {q\\\dot {q}},\quad A=\bbm {0&1\\-1&-2\zeta },\quad B_1=\bbm {0\\0},\quad B_2=\bbm {0\\1},\quad C_1=\bbm {0&1},\quad D_{11}=-1. \]

    We note that the transfer function \(C_1(sI-A)^{-1}B_2\) is (by Chapter 5)

    \[ G(s)=\frac {s}{s^2+2\zeta s+1}, \]

    which has a zero at \(s=0\) which is an eigenvalue of \(A_e\). Therefore by Corollary 10.6 we expect that there might be an issue with solvability of the problem.

    The regulator equations are

    \[ \bbm {0&1\\-1&-2\zeta }\bbm {\Pi _1\\\Pi _2} +\bbm {0\\1}V =0,\qquad \bbm {0&1}\bbm {\Pi _1\\\Pi _2}-1=0. \]

    It follows from the second equation that \(\Pi _2=1\) so that the first equation is

    \[ \bbm {1\\-\Pi _1-2\zeta +V}=\bbm {0\\0}. \]

    We see that the first component of this gives the contradiction \(1=0\). Therefore the regulator equations do not have a solution (as we expected might be the case).

    We instead consider

    \[ n_e=2,\qquad A_e=\bbm {0&1\\0&0},\qquad C_e=\bbm {0&1}. \]

    The regulator equations are

    \begin{gather*} \bbm {0&1\\-1&-2\zeta }\bbm {\Pi _{11}&\Pi _{12}\\\Pi _{21}&\Pi _{22}} +\bbm {0\\1}\bbm {V_1&V_2} =\bbm {\Pi _{11}&\Pi _{12}\\\Pi _{21}&\Pi _{22}}\bbm {0&1\\0&0}, \\ \bbm {0&1}\bbm {\Pi _{11}&\Pi _{12}\\\Pi _{21}&\Pi _{22}}-\bbm {0&1}=0. \end{gather*} The second of these equations give \(\Pi _{21}=0\), \(\Pi _{22}=1\). The first equation then is

    \[ \bbm {0&1\\-1&-2\zeta }\bbm {\Pi _{11}&\Pi _{12}\\0&1} +\bbm {0&0\\V_1&V_2} =\bbm {0&\Pi _{11}\\0&0}, \]

    which is

    \[ \bbm {0&1\\-\Pi _{11}&-\Pi _{12}-2\zeta } +\bbm {0&0\\V_1&V_2} =\bbm {0&\Pi _{11}\\0&0}, \]

    from which we deduce \(\Pi _{11}=1\), \(V_1=1\) and \(\Pi _{12}=V_2-2\zeta \). We see that now the regulator equations have a solution, but that this solution is not unique (as \(V_2\in \mR \) is arbitrary). We have

    \[ \Pi =\bbm {1&V_2-2\zeta \\0&1},\qquad V=\bbm {1&V_2}. \]

    Since \(A\) is asymptotically stable, we can take \(F_1=0\) and we obtain the control

    \[ u=x_{e,1}+V_2x_{2,e}. \]

    With the initial condition \(x_e^0=\sbm {0\\r}\) we obtain \(x_e(t)=\sbm {rt\\r}\) and \(w(t)=r\), so that the above control is

    \[ u(t)=rt+V_2r, \]

    where \(V_2\in \mR \) is an arbitrary constant. We note that this arbitrary constant arises because the transfer function from \(u\) to \(z\) has a zero at \(s=0\), so that constant inputs result in a zero output (therefore the arbitrary term \(V_2r\) in \(u\) does not affect the output and hence for the purposes of output regulation is indeed irrelevant).

  • Example 10.8. Consider

    \[ \dot {x}=u. \]

    The objective is to find a control \(u\) such that for a given \(r\in \mR \) we have \(\lim _{t\to \infty }x(t)=r\) no matter what \(x(0)\) is. We have \(z=x-w\)

    \[ n=1,\quad A=0,\quad B_1=0,\quad B_2=1,\quad C_1=1,\quad D_{11}=-1. \]

    We note that zero is an eigenvalue of \(A\) which will also be an eigenvalue of \(A_e\). Therefore Corollary 10.6 will not be applicable, but Proposition 10.4 is.

    We choose the exo-system

    \[ n_e=1,\quad A_e=0,\qquad C_e=1. \]

    The Rosenbrock matrix is

    \[ \bbm {sI-A&-B_2\\C_1&0}=\bbm {s&-1\\1&0}, \]

    which has determinant 1 and therefore is invertible and therefore surjective for all \(s\in \mC \) (in particular: for all eigenvalues of \(A_e\)). Therefore by Proposition 10.4 we have solvability of the regulator equations (and therefore of the output regulation problem).

    The regulator equations are

    \[ V=0,\quad \Pi -1=0, \]

    which gives \(\Pi =1\) and \(V=0\). Since \(A\) is not asymptotically stable, we have to choose a nonzero \(F_1\); we have \(A+B_2F_1=F_1\) so any \(F_1<0\) will work. We choose \(F_1=-1\). The feedback then is

    \[ u=-x+x_e. \]

    We check directly that this indeed solves the problem. We have

    \[ \dot {x}=-x+x_e,\quad \dot {x}_e=0,\quad x(0)=r,\quad w=x_e. \]

    This gives \(w=x_e=r\) and therefore

    \[ \dot {x}=-x+r. \]

    The solution of this with \(x(0)=x^0\) is \(x(t)=r+(x^0-r)\e ^{-t}\), which indeed satisfies \(\lim _{t\to \infty }x(t)=r\) no matter what \(x(0)\) is.