superquantile-opt

A semismooth-Newton-based proximal augmented Lagrangian (p-ALM) solver$^{[1]}$ for the sample average approximation (SAA) of the superquantile constrained problem

$$\begin{equation} \begin{array}{rl} \displaystyle \underset{{x\in X\subseteq \mathbb{R}^n}}{\text{minimize}}\; & \displaystyle f(x) + \mathsf{CVaR}_{\tau_0}\bigl(G^{\,0}(x;\omega^{[m]})\bigr) \\[0.1in] \mbox{subject to} & \mathsf{CVaR}_{\tau_{\ell}}\bigl(G^{\,\ell}(x;\omega^{[m]})\bigr) \leq 0,\; \ell \in\{1,2,\ldots,L\}, \end{array} \end{equation}$$

where

• $\mathsf{CVaR}_{\tau}\bigl(y\bigr)$ denotes the (empirical) superquantile of a (realization of a random) vector $y\in\mathbb{R}^m$ at confidence $\tau\in(0,1)$ with $\mathsf{CVaR}_{\tau}\bigl(y\bigr) = k(\tau)^{-1}\mathsf{T}_{k(\tau)}(y)$ where $k(\tau) := m\cdot(1-\tau)$ and where $\mathsf{T}_k(\cdot)$ denotes the top- $\!\!k$-sum operator$^{[2]}$
• $f$ is smooth and convex;
• for each $\ell\in\{0,1,\ldots,L\}$, $G^\ell(x;\omega^{[m]}) := \bigl\{g^\ell(x;\omega^j)\bigr\}_{j=1}^m$ is a collection of convex continuously differentiable scalar functions $g^\ell(\,\cdot\,; \omega)$ evaluated at $m$ SAA scenarios $\{ \omega^j \}_{j=1}^m$.

Note that the solver framework can handle different a number of scenarios for each superquantile constraint $\ell\in\{0,1,\ldots,L\}$ by fixing $g^{\ell}(\,\cdot\,;\omega)\equiv0$ for certain scenario indices. In the implementation, each $\ell$ is explicitly given $m_\ell$.

The repository contains two directories related to the paper https://arxiv.org/abs/2405.07965:

• src/: implementations of the p-ALM.
• run/: scripts for running the experiments.

---

$^{[1]}$ See https://arxiv.org/abs/2405.07965 for methodological details.
$^{[2]}$ See https://arxiv.org/abs/2310.07224 and https://github.com/jacob-roth/top-k-sum.