Robust LP question using box uncertainty model

110 Views Asked by At

I am trying to solve this robust LP problem by writing it as a QP

$$\min_x x^TSx : \mu \leq r^T x , Ax \leq b$$

Under Box uncertainty model: $$R = \{r : \| r - \hat{r}\|_\infty \leq \rho\}$$

Here is the solution outlined by the instructor

First we note that

$$\mu \leq x^T r \:\:\text{ for every } r \in R$$ Can be written as $$\mu \leq \min_{r \in R} r^T x$$

The above makes sense

The rest of this lost me

$$\min_{x \in R} r^T x = \hat{r}^T x + \rho . \:\:\:\: \min_{u : \|u\|_\infty \leq 1}u^Tx = \hat{r}^T x - \rho \|x\|_1$$

$$\min_x x^T S x : \mu + \rho \|x\|_1 \leq \hat{r}^Tx, Ax \leq b$$

can be expressed as a QP $$\min_x x^T S x : \mu + \rho y^T \textbf{1} \leq \hat{r}^T , Ax \leq b , -y \leq x \leq y$$

Specifically, I don't understand how

$$\min_{x \in R} r^T x = \hat{r}^T x + \rho . \:\:\:\: \min_{u : \|u\|_\infty \leq 1}u^Tx = \hat{r}^T x - \rho \|x\|_1$$

is true with the given condition

I thought

$$\min_{x \in R} r^T x =\hat{r}^T x - |\rho x| $$

As

$$R = \{r : \| r - \hat{r}\|_\infty \leq \rho\}$$

$$ \hat{r_i} - \rho \leq r_i \leq \hat{r}_i + \rho$$

so if $x_i < 0 \Rightarrow r = \hat{r}_i + \rho$ and vice versa resulting in

$$\min_{x \in R} r^T x = \hat{r}^T x - |\rho x| $$

1

There are 1 best solutions below

0
On

Specifically, I don't understand how

$$\min_{x \in R} r^T x = \hat{r}^T x + \rho . \:\:\:\: \min_{u : > \|u\|_\infty \leq 1}u^Tx = \hat{r}^T x - \rho \|x\|_1$$

is true with the given condition

Note that the minimization is with respect to $r$. Lets do a change of variables. If we set $u = \frac{r - \hat{r}}{\rho}$, we get:

$$\min_{r \in R} r^T x = \min_{u : \|u\|_\infty \leq 1} \rho u^Tx + \hat{r}^Tx.$$

Further,

$$\min_{u : \|u\|_\infty \leq 1} \rho u^Tx + \hat{r}^Tx = -\max_{u : \|u\|_\infty \leq 1} \rho u^Tx + \hat{r}^Tx.$$

Then, you can use an identity known as Holder's inequality to the first term:

$$\max_{u : \|u\|_\infty \leq 1} \rho u^Tx = \rho\|x\|_1.$$

Thus,

$$\min_{r \in R} r^T x = -\rho\|x\|_1 + \hat{r}^Tx.$$

Substituting the above in our original program gives the desired program that you have further simplified using another set of variables $y$.