Immanuel M. Bomze,Daniel de Vicente
Abstract:The standard quadratic optimization problem (StQP) consists of minimizing a quadratic form over the standard simplex. Without convexity or concavity of the quadratic form, the StQP is NP-hard. This problem has many relevant real-life applications ranging portfolio optimization to pairwise clustering and replicator dynamics. Sometimes, the data matrix is uncertain. We investigate models where the distribution of the data matrix is known but where both the StQP after realization of the data matrix and the here-and-now problem are indefinite. We test the performance of a chance-constrained epigraphic StQP to the uncertain StQP.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to deal with the uncertain standard quadratic optimization problem (StQP) under distribution assumptions. Specifically, the author studies how to effectively solve the standard quadratic optimization problem when the data matrix \(Q\) is uncertain. This uncertainty means that the specific value of the data matrix \(Q\) is random, but its distribution is known.
### Main problem description
1. **Standard quadratic optimization problem (StQP)**:
The standard quadratic optimization problem refers to minimizing a quadratic form on a standard simplex:
\[
\ell(Q) := \min_{x \in \Delta} x^\top Q x
\]
where \(Q \in \mathbb{R}^{n \times n}\) is a symmetric matrix, \(\Delta=\{x \in \mathbb{R}^n : e^\top x = 1, x\geq0\}\) is the standard simplex, and \(e\) is the all - ones vector.
2. **Uncertainty**:
In many practical applications, the data matrix \(Q\) is uncertain. That is, \(Q\) may be a random matrix \(\tilde{Q}\) with a known distribution. In this case, traditional optimization methods may no longer be applicable because the optimal solution depends on the specific value of \(Q\).
3. **Chance constraints**:
To deal with this uncertainty, the author introduces chance constraints, that is, the solution \(x\) is required to satisfy the constraints under a certain probability condition:
\[
P(x^\top\tilde{Q}x\leq t)\geq\alpha
\]
where \(\alpha\in(0, 1)\) is the confidence level provided by the decision - maker.
### Main contributions of the paper
1. **StQP of representative variables with chance constraints**:
The author proposes a new model - StQP of representative variables with chance constraints (CCE - StQP). By introducing a representative variable \(t\), the original problem is transformed into an equivalent deterministic StQP:
\[
\ell_{\text{cce}}(P,\alpha):=\min_{x,t}t\quad\text{s.t.}\quad P(x^\top\tilde{Q}x\leq t)\geq\alpha,\quad x\in\Delta,\quad t\in\mathbb{R}
\]
2. **Deterministic equivalent form**:
Under reasonable distribution assumptions, CCE - StQP can be transformed into a deterministic StQP. For example, under the Gaussian orthogonal ensemble (GOE) perturbation model, CCE - StQP is equivalent to:
\[
\ell_{\text{cce}}(\alpha):=\min_{x\in\Delta}x^\top Q_{\text{cce}}x
\]
where
\[
Q_{\text{cce}} := Q_{\text{nom}}+\sqrt{2\beta}\Phi^{-1}(\alpha)I
\]
where \(\Phi\) is the cumulative distribution function of the standard normal distribution.
3. **Numerical experiments**:
The author verifies the effectiveness of the proposed method through numerical experiments and compares it with the robust optimization method. The results show that CCE - StQP can be less conservative than the robust method in some cases, especially when the confidence level is not particularly high.
### Summary
This paper aims to solve the standard quadratic optimization problem under uncertain data. By introducing chance constraints and representative variables, a new optimization model CCE - StQP is proposed, and its deterministic equivalent form under specific distribution assumptions is proved. In addition, the paper also verifies it through numerical experiments.