I am looking for some helps concerning an optimization problem.
I have an optimization problem defined on two sets $\mathcal{X}=\{x_i\}_{i=1}^n $ and $\mathcal{Y}=\{y_j\}_{j=1}^m $ and described as follows
$$
\begin{equation}
\begin{aligned}
& \underset{\alpha, \beta}{\text{max}}
& & \sum_{i=1}^n \alpha_i x_i^T x_i-\frac{\sum_{i=1}^n \sum_{j=1}^n \alpha_i x_i^Tx_j \alpha_j}{\sum_{i=1}^n \alpha_i}+\sum_{k=1}^m \beta_k y_k^T y_k-\frac{\sum_{k=1}^m \sum_{l=1}^m \beta_k y_k^Ty_l \beta_l}{\sum_{j=1}^m \beta_j }\\
& \text{subject to}
& & 0 \leq \alpha \leq C1 \\
&
& & 0 \leq \beta \leq C2 \\
\end{aligned}
\end{equation}
$$
where $C1$ and $C2$ are two given parameters. I'm asking if the above problem can be rewritten in the follwing form
$$
\begin{equation}
\begin{aligned}
& \underset{X}{\text{max}}
& & \frac{ X^T H X}{X}+A^TX\\ %\mathbf {1}_N
& \text{subject to}
& & 0 \leq X \leq C \\
\end{aligned}
\end{equation}
$$
If yes what would be the value of $H$,$A$ and $X$ and if not how can I optimize the first formulation by matlab knowing that the second formulation can be optimized in matlab by CVX package for convex optimization using its "quad_over_lin" function.
Thanks a lot
Asked
Active
Viewed 220 times
1
alya
- 11
- 1
-
Welcome to SciComp.Exchange. Is the $X$ a vector? If it is, how can you divide by a vector in your reformulation? – nicoguaro Dec 14 '15 at 21:17
-
1Presumably the denominator should be $B^{T}X$ where $B$ is a constant vector (in this case it would be the vector of all ones.) – Brian Borchers Dec 14 '15 at 22:06
-
Yes, as noted by Mr. Brian Borchers, the denominator take the form of $ B^T X$ where $B $is a vector of ones and $X$ depends on $\alpha$ and $\beta$. Thanks – alya Dec 15 '15 at 07:53
1 Answers
1
This question sounds a lot like a homework exercise, so I'll simply give a couple of hints.
- Break your objective into the sum of two of two terms, and then use the "epigraph trick" on each separately. So, you'd have
$\max s + t$
subject to
$s \leq -\mbox{quad_over_lin(...)}$
$t \leq -\mbox{quad_over_lin(...)}$
- Recognize that the two parts of your objective function are already almost in quad_over_lin form. How can you write
$\sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{i} x_{i}^{T}x_{j} \alpha_{j}$
as
$\| X\alpha \|_{2}^{2}$?
Brian Borchers
- 18,719
- 1
- 39
- 70
-
Thanks Mr Brian for your answers. I have used your answer to a problem presented in http://scicomp.stackexchange.com/questions/20292/minimization-of-normalized-constrained-quadratic-function – alya Dec 15 '15 at 03:50