next up previous
Next: Second order sufficient optimality Up: First and second-order optimality Previous: First and second-order optimality

First order necessary conditions

Let the control $\bar{u}$ be locally optimal for (P) with associated state $\bar{y}$, i.e.

\begin{displaymath}
J(y,u) \ge J(\bar{y},\bar{u})
\end{displaymath} (2.1)

holds for all $(y,u)$ satisfying the constraints (1.2-1.4), where $u$ belongs to a sufficiently small $L^\infty$-neighborhood of $\bar{u}$. Suppose further that $(\bar{y},\bar{u})$ is regular. Then there exist Lagrange multipliers $\bar{p}\in W(0,T) \cap C(\bar{Q})$ (the adjoint state) and $\bar{\lambda} \ge 0$ such that the adjoint equation
\begin{displaymath}
\begin{array}{rclll}
-\bar{p}_t - \bar{p}_{xx}&= &\alpha(\ba...
...l,t)\, \bar{p}(l,t)&=&a_y(t)&\mbox{ in }&(0, T),\\
\end{array}\end{displaymath} (2.2)

the variational inequality
\begin{displaymath}
\int \limits_0^T (\nu \bar{u}(t) + \bar{p}(l,t) + a_u(t))(u(t) - \bar{u}(t))\, dt \ge 0
\quad \forall u \in U_{ad},
\end{displaymath} (2.3)

and the complementary slackness condition
\begin{displaymath}
\bar{\lambda} \int \int_Q \bar{y}(x,t)\, dxdt = 0
\end{displaymath} (2.4)

are fulfilled, see [3] or [11]. We mention that (2.3) is equivalent to the well-known projection property
\begin{displaymath}
\bar{u}(t) = \Pi_{[u_a,u_b]}\{ -\frac{1}{\nu}(\bar{p}(l,t) + a_u(t))\},
\end{displaymath} (2.5)

where $\Pi_{[u_a,u_b]}: \mbox{\piz R}\rightarrow [u_a,u_b] $ denotes projection onto $[u_a,u_b]$. Moreover, we recall that these conditions can be derived by variational principles applied to the Lagrange function ${\cal L}$,

\begin{displaymath}
\begin{array}{rcl}
{\cal L}(y,u,p,\lambda) &=&J(y,u) - \int ...
... + y^2(l,t)- u(t) - e_{\Sigma }(t) ) p(l,t) \, dt .
\end{array}\end{displaymath}

Defining ${\cal L}$ in this way, we tacitly assume that the homogeneous initial and boundary conditions of $y$ are formally included in the state space. The conditions (2.2-2.3) follow from ${\cal L}_y(\bar{y},\bar{u},\bar{p},\bar{\lambda}) y = 0$ for all admissible increments $y$ and ${\cal L}_u(\bar{y},\bar{u},\bar{p},\bar{\lambda}) (u -\bar{u}) \ge 0 \quad \forall u \in
U_{ad}$. Let $\tau > 0$ be given. We define

\begin{displaymath}
\begin{array}{rcl}
A^+(\tau) &=& \{t \in (0,T) \, \vert \, \...
...\nu \bar{u}(t) + \bar{p}(l,t) + a_u(t) \ge
\tau \}.
\end{array}\end{displaymath}

It holds $\bar{u}= u_b$ on $A^+$ and $\bar{u}= u_a$ on $A^-$. These sets indicate strongly active control constraints.


next up previous
Next: Second order sufficient optimality Up: First and second-order optimality Previous: First and second-order optimality
Hans D. Mittelmann
2003-01-25