I am reading these notes from Stanford called "Optimization with uncertain data".

In section 2.2, Example 2 (page 7), the author mentions the following portfolio problem $(P)$:

$$ \max \; t $$ subject to \begin{align*} \text{Prob}\left(\sum_i R_i x_i \ge t\right) &\ge 1-\epsilon \\ \sum_i x_i &= 1 \\ x_i &\ge 0 \quad \forall i \end{align*} where $R_i$ (the random return of each asset $i$) and $\epsilon$ (the value at risk) are given parameters.

Assuming $R_i \in [\mu_i -\text{u}_i,\mu_i +\text{u}_i]$ with $\text{E}(R_i)=\mu_i$, the author shows that the problem is equivalent to maximizing $$ \sum_i \mu_i x_i -\sqrt{2 \log \frac{1}{\epsilon}} \lVert\text{diag}(\text{u})x \rVert_2 $$ subject to \begin{align*} \sum_i x_i &= 1 \\ x_i &\ge 0 \quad \forall i \end{align*}

And states:

This corresponds to the uncertainty set given by the scaled $\ell_2$ ball

$$ \mathcal{U}=\{u \in \mathbb{R}^n , \lVert\text{diag}(u)^{-1/2}u \rVert_2 \le \sqrt{2 \log \frac{1}{\epsilon}}\} $$ in the robust inequality $\sum_i (\mu_i+u_i)x_i \le t$ for all $u\in \mathcal{U}$.

I don't understand the last quoted part ($\text{diag}(u)^{-1/2}u$?) and suspect there is at least one typo ($\sum_i (\mu_i+u_i)x_i \color{red}{\ge} t$?). In the end, if I want to solve the initial problem, what formulation do I use ?


1 Answer 1


I believe there is a typo in the document, and perhaps some confusion between parameters $\text{u}$ and variables $u$.

My understanding is that the initial problem $(P)$ is equivalent to:

$$ \max_x \; \sum_i \mu_i x_i -\sqrt{2 \log \frac{1}{\epsilon}} \lVert\text{diag}(\text{u})x \rVert_2 $$ subject to \begin{align*} \sum_i x_i &= 1 \\ x_i &\ge 0 \quad \forall i \end{align*}

and to: $$ \max_{x,t,u} \; t $$ subject to \begin{align*} \sum_i x_i &= 1 \\ x_i &\ge 0 \quad \forall i \\ \sum_i (\mu_i + u_i\text{u}_i)x_i &\ge t \\ \lVert u \rVert_2 &\le \sqrt{2 \log \frac{1}{\epsilon}} \end{align*}


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.