Immanuel Bomze,Bo Peng,Yuzhou Qiu,E. Alper Yıldırım
Abstract:Standard quadratic optimization problems (StQPs) provide a versatile modelling tool in various applications. In this paper, we consider StQPs with a hard sparsity constraint, referred to as sparse StQPs. We focus on various tractable convex relaxations of sparse StQPs arising from a mixed-binary quadratic formulation, namely, the linear optimization relaxation given by the reformulation-linearization technique, the Shor relaxation, and the relaxation resulting from their combination. We establish several structural properties of these relaxations in relation to the corresponding relaxations of StQPs without any sparsity constraints, and pay particular attention to the rank-one feasible solutions retained by these relaxations. We then utilize these relations to establish several results about the quality of the lower bounds arising from different relaxations. We also present several conditions that ensure the exactness of each relaxation.
What problem does this paper attempt to address?
This paper attempts to solve the difficult problem of solving the standard quadratic optimization problem (StQP) with sparsity constraints. Specifically, the standard quadratic optimization problem (StQP) is the problem of minimizing a quadratic form on the standard simplex. When the sparsity constraint is introduced, that is, the number of non - zero elements in the solution is restricted, such problems become more complex and difficult to solve.
### Research Background and Motivation
1. **NP - hard Property**:
- StQP itself is NP - hard because it can be reduced to the maximum clique problem.
- After introducing the sparsity constraint, even if the Hessian matrix is positive definite, the problem still remains NP - hard.
2. **Wide Application**:
- StQP has wide applications in finance (such as Markowitz portfolio optimization), economics, machine learning, life sciences and other fields.
- Especially in some applications where the number of investment assets needs to be restricted, such as sparse portfolio optimization problems, the sparsity constraint is necessary.
3. **Limitations of Existing Methods**:
- It is very difficult to directly solve StQP with sparsity constraints, especially in high - dimensional cases.
- Therefore, researchers focus on finding effective convex relaxation methods to provide acceptable approximate solutions.
### Main Contributions of the Paper
1. **Convex Relaxation Methods**:
- The paper explores multiple convex relaxation methods, including linear programming (LP) and semidefinite programming (SDP) methods.
- Specific methods include:
- Reformulation - Linearization Technique (RLT) relaxation
- Shor relaxation
- Combination of RLT and Shor relaxation
2. **Analysis of Structural Properties**:
- Analyzes the structural properties of these relaxation methods under different sparsity levels.
- Studies the relationship between these relaxation methods and StQP without sparsity constraints, especially regarding the retention of rank - one feasible solutions.
3. **Evaluation of Lower Bound Quality**:
- Establishes the quality results of the lower bounds generated by different relaxation methods.
- Proposes conditions to ensure the accuracy of each relaxation method.
4. **Impact of Sparsity Constraints**:
- Pays special attention to the impact of sparsity constraints on problem complexity and solution quality.
- For a smaller sparsity parameter \( \rho \), proposes a more accurate relaxation method.
### Conclusion
This paper systematically studies the standard quadratic optimization problem with sparsity constraints, and proposes and analyzes multiple effective convex relaxation methods. These methods not only provide a theoretical basis for solving this type of difficult problems, but also provide new ideas and tools for efficient solution in practical applications.