Abstract:We investigate optimality conditions for the sensor placement problem within optimal experimental design, wherein one must decide on the optimal manner in which a fixed number of sensors can be arranged over a large number of candidate locations. By a subgradient argument, we obtain sufficient and necessary conditions for global optimality of the relaxed problem, and demonstrate how one can take advantage of this optimality criterion to approximate optimal binary designs, i.e.~designs where no fractions of sensors are placed. To demonstrate our optimality criteria-based results, we derive a trace-free, efficient formulation of the gradient of the A-optimal objective for finite element-discretised function space settings, and study globally optimal designs for a Helmholtz-type source problem and extensions towards optimal binary designs.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to optimally place sensors in experimental design, that is, given a certain number of sensors and a large number of candidate positions, to determine the best layout of these sensors. Specifically, the paper focuses on how to solve the sensor placement problem through global optimality conditions under the Optimal Experimental Design (OED) framework and extends it to Binary A - Optimal Experimental Designs.
### Main Problems and Objectives
1. **Sensor Placement Problem**:
- Determine how to optimally arrange the positions of these sensors among a limited number of sensors and a large number of candidate positions.
- Through sub - gradient arguments, derive the global optimality conditions of the relaxed problem, which can be used for approximately optimal binary designs (i.e., designs that do not use some sensors).
2. **A - Optimal Experimental Design**:
- Research A - optimal design under specific conditions (such as linear Bayesian inverse problems), that is, minimizing the trace of the posterior covariance matrix.
- Derive an efficient traceless form of the gradient in the finite - element discretization setting and study the global optimal design and its application in Helmholtz - type source problems.
### Key Contributions
1. **Non - smooth Convex First - order Optimality Conditions**:
- By studying the 1 - relaxed best sensors placement problem, derive non - smooth convex first - order optimality conditions, so as to be able to efficiently and accurately characterize the global optimal solution.
- Provide explicitly verifiable necessary and sufficient conditions to judge whether a given design is a global optimal solution.
2. **Continuous - type Algorithm**:
- Drawing on the method in reference [11], propose a continuous - type algorithm. By gradually solving problems with different relaxation parameters \( p \), finally find the fully binary optimal design.
- Utilize low - rank decomposition techniques to efficiently evaluate the A - optimal objective function and its derivatives, thereby simplifying the optimization process.
### Methods and Techniques
1. **Low - Rank Decomposition**:
- Significantly reduce the computational complexity by approximating the prior - preprocessed misfit Hessian matrix \( \mathbf{bF}^T \mathbf{M}_w \mathbf{bF} \) with a low - rank matrix.
- Use the QR decomposition method instead of the SVD method to reduce unnecessary computational burden.
2. **Numerical Implementation**:
- Verify the effectiveness of the proposed theories and methods through specific numerical experiments, and show how to use these methods to find high - performance fully binary experimental designs.
### Conclusion
Through rigorous mathematical analysis and numerical experiments, the paper provides a set of systematic methods to solve the sensor placement problem and extends it to binary A - optimal experimental designs. These methods are not only of great significance in theory but also demonstrate their high efficiency and feasibility in practical applications.