next up previous
Next: Conclusion Up: Solving LMI and BMI Previous: Optimization by QE and

Solving BMI by QE

Algorithm 1 (Feasible problems)   ``Find X which satisfies $M(X) >_d (or \geq_d) 0$''

1.
Transform $M(X) >_d (or \geq_d) 0$ to an equivalent formula $\varphi(X)$ by using Sylvester's theorem.

2.
Construct the first-order formula $\psi = \exists X (\varphi(X))$.

3.
Perform QE for $\psi$.

QE outputs true or false. In the case of true QE returns one feasible solution. false means there are no feasible solutions.

Example 1   ``Does BMI B(x1,x2) <d 0 have any feasible solutions in $(x_1,x_2) \in [-\frac{1}{2},2]\times [-3,7]$? [9]'' where

B(x1,x2)= $
\left [
\begin{array}{ccc}
-10 & -\frac{1}{2} & -2 \\
-\frac{1}{2} & \frac{9}...
...-\frac{1}{10} & \frac{6}{5} & -1 \\
-\frac{2}{5} & -1 & 0
\end{array}\right ]
$

$+x_2
\left [
\begin{array}{ccc}
9 & \frac{1}{2} & 0 \\
\frac{1}{2} & 0 & -3 \\...
...y}{ccc}
0 & 0 & 2 \\
0 & -\frac{11}{2} & 3 \\
2 & 3 & 0
\end{array}\right ].
$

This problem is equivalent to the following QE problem:


\begin{displaymath}\exists x_1 \exists x_2
(A_1 \land A_2 \land A_3 \land -\frac{1}{2} \leq x_1 \leq 2 \land
-3 \leq x_2 \leq 7)\end{displaymath}

where


A1 = -45x1+9x2+50 > 0 ,


A2 = (-4950x2-25)x12+(990x22+6590x2+4100)x1-217x22

-2020x2-4525 > 0,


A3 = (-11000x23+37500x22-102750x2+40375)x13+ (-700x23

-15850x22+141250x2-27500)x12+ (3680x23+12935x22

-71900x2-19625)x1-764x23-3280x22+7000x2+9000 > 0.


After performing QE for this, we have ``true''. In the case of $(x_1,x_2) \in {\bf R}^2$, after performing QE for $\exists x_1 \exists x_2 (A_1 \land A_2 \land A_3)$we also have true.

Algorithm 2 (linear objective minimization problems)   ``minimize 'cT X subject to $M(X) >_d (or \geq_d) 0$.

1.
Transform $M(X) >_d (or \geq_d) 0$ to an equivalent formula $\varphi(X)$ by using Sylvester's theorem.

2.
Assign the new variable z to the objective function; $z-c^T X \geq 0$.

3.
Construct the first-order formula $\psi = \exists X \ (\varphi(X) \ \land \ z-c^T X \geq 0 )$.

4.
Perform QE for $\psi$.

We consider some examples (which are modified one taken from [20]), and apply the above mentioned method to them in order to demonstrate the potential of QE approach to SDP problems.


$\bullet$Non-Parametric: Minimize x1+x2 subject to

 \begin{displaymath}
\left [
\begin{array}{ccc}
1 & x_1 & 0 \\
x_1 & x_2 & 0 \\
0 & 0 & x_1+1
\end{array}\right] \geq_d 0.
\end{displaymath} (8)

The semidefinite constraint (8) is reduced to a Boolean system of inequalities constraints by using Sylvester's criterion;


$\psi=
(1 \geq 0) \land
(x_2 \geq 0) \land
(x_1+1 \geq 0) \land
(x_2-x_1^2 \ge...
..._2x_1+x_2) \geq 0) \land
(x_1+1 \geq 0) \land
(-x_1^3-x_1^2+x_2x_1+x_2 \geq 0).$


Assign new slack variable z to objective function x1+x2 and let $\psi' = \psi \land (z-x_1-x_2 \geq 0).$Then the problem is formulated as a first-order formula $\varphi = \ ^\exists x_1 \ ^\exists x_2 \ ( \psi' ).$After QE we have an equivalent qf formula describing the range of objective function;

 \begin{displaymath}
4z + 1 \geq 0
\end{displaymath} (9)

$\bullet$Parametric(1): Minimize ax1+x2 subject to (8) with a parameter a.


Let $\psi'_{P_1} = \psi \land (z-ax_1-x_2 \geq 0).$The problem is formulated as $\varphi_{P_1} = \ ^\exists x_1 \ ^\exists x_2 \ ( \psi'_{P_1} ).$After QE we have an equivalent qf formula describing the range of objective function z with a parameter a;


$(a + z - 1 \geq 0) \lor
(z \geq 0) \lor
((a^3 - 2a^2z + a^2 + az^2 - az \leq ...
... + 2z \geq 0) \land (a - 2 \leq 0))
\lor ((a^2 + 4z \geq 0) \land (a - 2 < 0)).$


If we substitute a parameter a with 1 and simplify the result, we have same result as (9).


$\bullet$Restricted region: Minimize x1+x2 subject to (8) in restricted domains D1,D2 respectively, where (i) $D_1=
\{ x_1,x_2 \in {\bf R} \ \vert \ 0 \leq x_1 \leq 3, 0 \leq x_2 \leq 5 \}$and (ii) $D_2=
\{ x_1,x_2 \in {\bf R} \ \vert \ 0 \leq x_1 \leq 10, 5 \leq x_2 \leq 10 \}$.


(i) In this case the problem is formulated as

\begin{displaymath}\varphi_{D_1} = \ ^\exists x_1 \ ^\exists x_2 (\psi' \land
(0 \leq x_1 \leq 3 \land 0 \leq x_2 \leq 5) )\end{displaymath}

After QE we have $4z + 1 \geq 0$.


(ii) In this case the problem is formulated as

\begin{displaymath}\varphi_{D_2} = \ ^\exists x_1 \ ^\exists x_2 (\psi' \land
(0 \leq x_1 \leq 10 \land 5 \leq x_2 \leq 10) )\end{displaymath}

After QE we have


$(z - 4 \geq 0) \lor
((z^2 - 10z + 20 \leq 0) \land (z^2 - 20z + 90 \geq 0)
\la...
... 20 \leq 0) \land
(z^2 - 20z + 90 \geq 0) \land (4z + 1 \geq 0) \land (z < 0)).$


$\bullet$Parametric(2): Minimize x1+x2 subject to

 \begin{displaymath}
\left [
\begin{array}{ccc}
s & x_1 & 0 \\
x_1 & x_2 & 0 \\
0 & 0 & x_1+1
\end{array}\right] \geq_d 0.
\end{displaymath} (10)

where s is a parameter.


The semidefinite constraint (10) is reduced to


$\psi_P=
(s \geq 0) \land
(x_2 \geq 0) \land
(x_1+1 \geq 0) \land
(sx_2-x_1^2 \...
...+x_2) \geq 0) \land
(s(x_1+1) \geq 0) \land
((x_2x_1+x_2)s-x_1^3-x_1^2 \geq 0).$


Let $\psi'_{P_2} = \psi_P \land (z-x_1-x_2 \geq 0).$Then the problem is formulated as a first-order formula $\varphi_{P_2} = \ ^\exists x_1 \ ^\exists x_2 \ ( \psi'_{P_2} ).$After QE we obtain


$(sz + s - 1 \geq 0 \land s > 0 \land z + 1 \geq 0)
\lor (sz + s - 1 \geq 0 \la...
...(s > 0
\land (s^2 z + s^2 - s > 0 \lor (s^2 - 2s < 0 \land sz + s - 1 = 0))).
$


If we substitute the parameter s with 1 and simplify the result, we have same result as (9).


$\bullet$ Parameter uncertainty: Minimize x1+x2 subject to (10) within the regions of a parameter s (i) $-5 \leq s \leq -1$, (ii) $ -10 \leq s \leq 0$, respectively.


For (i),(ii), the problem is formulated as a first-order formula; $\varphi_{(i)} = \ ^\exists x_1 \ ^\exists x_2
(\psi'_{P_2} \land (-5 \leq s \leq -1))$, $\varphi_{(ii)} = \ ^\exists x_1 \ ^\exists x_2
(\psi'_{P_2} \land ( -10 \leq s \leq 0))$, respectively. After QE we obtain ``false'' for (i) and for (ii)


$(sz + s - 1 \geq 0 \land s = 0 \land z + 1 \geq 0)
\lor (sz + s \geq 0 \land s...
... \geq 0 \land (sz + s - 1 \geq 0 \lor z = 0))))))
\lor (s = 0 \land z \geq 0).$


next up previous
Next: Conclusion Up: Solving LMI and BMI Previous: Optimization by QE and
IMACS ACA'98 Electronic Proceedings