next up previous
Next: Numerical Verification Up: paper94 Previous: Second order sufficient optimality


The test example

We fix here the following quantities in (P):

\begin{displaymath}
\begin{array}{rcl}
T &=& 1, ~ l = \pi, ~u_a = 0,~u_b = 1, ~\...
...ght.\\
\\
a_u(t) &=&\nu + 1 - (1+ 2\nu) t,\\
\\
\end{array}\end{displaymath}


\begin{displaymath}
\begin{array}{rcl}
e_Q(t) &=&\left\{
\begin{array}{ll}
0, & ...
.../2)^4- (2t-1), & t \in (1/2,1].
\end{array} \right.
\end{array}\end{displaymath}

THEOREM 1   The quantities

\begin{displaymath}
\begin{array}{rcl}
\bar{u}&=&max\{0,2t-1\}\\
\bar{y}&=& \le...
...ght.\\
\bar{p}&=& (1-t) cosx\\
\bar{\lambda} &=&1
\end{array}\end{displaymath}

satisfy the system of first order necessary conditions.

Proof      Insert the definitions above in the state equation and adjoint equation. Then it is easy to check by elementary calculations that $\bar{u},~\bar{y}$ fulfill (1.2) and that $\bar{p}$ satisfies the adjoint equation (2.2). Moreover, the state-inequality constraint (1.4) holds as an equality, since the integral of $cosx$ over $[0,\pi]$ vanishes. Clearly, $\bar{u}$ is an admissible control. The variational inequality (2.3) is easy to verify by (2.5): We find

\begin{displaymath}
-1/\nu(\bar{p}(\pi,t) + a_u(t))= 2t-1=\left\{
\begin{array}{...
... 0, & t \in [0,1/2)\\
>0, & t \in (1/2,1].
\end{array}\right.
\end{displaymath}

Therefore,

\begin{displaymath}
\Pi_{[0,1]}\left\{ -1/\nu(\bar{p}(\pi,t) + a_u(t)) \right\} = max \{0,2t-1\} =
\bar{u}(t).
\end{displaymath}

Next we consider the second order sufficient condition (SSC) for the example analytically. What conditions must be checked to verify them? Thanks to our construction, $\bar{u}$ is strongly active on $[0,1/2)$ and $\bar{u}=0$ holds there. If $b<1/2$ is given, then

\begin{displaymath}
\nu \bar{u}(t) + \bar{p}(\pi,t) + a_u(t) = \bar{p}(\pi,t) + a_u(t) = -\nu (2t -1) > \nu
(2b-1)
\end{displaymath}

holds for $t \in [0,b]$. Therefore, $t \in A^-(\tau)$ for $\tau = \vert\nu (2b-1)\vert$. To verify the second order sufficient conditions, it suffices to confirm the coercivity condition (2.7) for all pairs $(y,u)$ coupled through the linearized equation (2.8) and satisfying $u = 0$ on $[0,b]$ ($0<b<1/2$ being arbitrary but fixed). Assume that
\begin{displaymath}
\alpha (x,t) = \left\{
\begin{array}{rl}
\alpha_0,& \, 0 \le t \le b \\
1,& \, b < t \le 1.
\end{array}\right.
\end{displaymath} (3.1)

THEOREM 2   Let $\alpha$ have the form (3.1), where $b \in [0,1/2)$. Then the second order sufficient conditions (SSC) are satisfied by $(\bar{y},\bar{u},\bar{p},\bar{\lambda})$ for arbitrary $\alpha_0 \in \mbox{\piz R}$.

Proof      Let $u$ vanish on $[0,b]$ and let $y$ solve (2.8). Then $y(x,t) = 0$ on $[0,b]$. For ${\cal L''}$ we get

\begin{displaymath}
\begin{array}{rcl}
{\cal L''}(\bar{y},\bar{u},\bar{p},\bar{\...
...^2(\pi,t)\, dt\\
&\ge& \nu \int\limits_0^1 u^2 dt.
\end{array}\end{displaymath} (3.2)

Hence the coercivity condition (2.7) is satisfied.

Notice that $\alpha_0$ was not assumed to be positive. If $\alpha_0\ge 0$, then ${\cal L''}$ is obviously coercive on the whole space $W(0,1) \times
L^2(0,1)$, and (SSC) is satisfied in a very strong sense. However, we might find negative values for $\alpha_0$ such that ${\cal L''}$ is partially indefinite.

THEOREM 3   If $\alpha_0 < 0$ is sufficiently small, then a pair $(y,u)$ exists, such that $u \ge 0$, $y$ solves the linearized equation (2.8), and
\begin{displaymath}
{\cal L''}(\bar{y},\bar{u},\bar{p},\bar{\lambda}) [y,u]^2 < 0.
\end{displaymath} (3.3)

Proof      We take an arbitrary but fixed $b<1/2$ and set

\begin{displaymath}
u(t) = \left\{
\begin{array}{rl}
1 & \mbox{ on }[0,b]\\
0& \mbox{ on }(b,1].
\end{array}\right.
\end{displaymath}

Then $\int\limits_0^\pi\int\limits_0^b y^2\, dxdt $ is positive. Hence

\begin{displaymath}
\alpha_0 \int\limits_0^\pi\int\limits_0^b y^2\, dxdt \rightarrow - \infty
\end{displaymath}

as $\alpha_0 \rightarrow - \infty$. Therefore, the expression (3.2) becomes negative for sufficiently small $\alpha_0$, if $y^2$ is substituted there for $0$ in the first integral.

For the numerical verification we need a rough estimate on how small $\alpha_0$ should be chosen. To get a negative value of ${\cal L''}(\bar{y},\bar{u},\bar{p},\bar{\lambda}) [y,u]^2$, we must have

\begin{displaymath}
\alpha_0 \int\limits_0^\pi
\int\limits_0^b y^2 \, dxdt + \in...
...ts_0^1 2(1-t)y^2(\pi,t)\, dt
+ \nu \int\limits_0^1 u^2 dt < 0,
\end{displaymath}

hence
\begin{displaymath}
\vert\alpha_0\vert > \displaystyle \frac{\int\limits_0^\pi
\...
...limits_0^b y^2 \, dxdt}
=\displaystyle \frac{I_1+I_2+I_3}{I_0}
\end{displaymath} (3.4)

must hold. Here, $b \in [0,1/2)$ can be chosen arbitrarily. We take the value $b = 1/4$. Thus, we evaluate the integrals $I_j$ for

\begin{displaymath}
u(t)= \left\{
\begin{array}{rl}
1& \, \mbox{ on }[0,1/4]\\
0& \, \mbox{ on }(1/4,1]
\end{array} \right.
\end{displaymath}

and the associated state $y$. The state $y$ solves the homogeneous heat equation subject to homogeneous initial condition, homogeneous boundary condition at $x = 0$ and

\begin{displaymath}
y_x(\pi,t)= \left\{
\begin{array}{ll}
1& \, \mbox{ on }[0,1/...
...
-2 \bar{y}(\pi,t)& \, \mbox{ on }(1/4,1].
\end{array} \right.
\end{displaymath}

To avoid tedious estimates which might be performed by means of a Fourier series representation of $y$, we have evaluated the integrals $I_j$, $j = 0,..,3$ numerically. The result is
$I_0 = .0103271,\quad I_1 = .0401844,\quad I_2 = .0708107,\quad I_3 = .001$

\begin{displaymath}
\frac{I_1+I_2+I_3}{I_0} = 10.845 .
\end{displaymath} (3.5)


next up previous
Next: Numerical Verification Up: paper94 Previous: Second order sufficient optimality
Hans D. Mittelmann
2003-01-25