Multifidelity sensor placement in Bayesian state estimation problems

We study optimal sensor placement for Bayesian state estimation problems in which sensors vary in cost and fidelity, resulting in a budget-constrained multifidelity optimal experimental design problem. Sensor placement optimality is quantified using the D-optimality criterion, and the problem is approached by leveraging connections with the column subset selection problem in numerical linear algebra. We implement a greedy approach for this problem, whose computational efficiency we improve using rank-one updates via the Sherman-Morrison formula. We additionally present an iterative algorithm that, for each feasible allocation of sensors, greedily optimizes over each sensor fidelity subject to previous sensor choices, repeating this process until a termination criterion is satisfied. To the best of our knowledge, these algorithms are novel in the context of cost constrained multifidelity sensor placement. We evaluate our methods on several benchmark state estimation problems, including reconstructions of sea surface temperature and flow around a cylinder, and empirically demonstrate improved performance over random designs.

💡 Research Summary

The paper tackles the problem of placing sensors of varying cost and fidelity under a global budget constraint in Bayesian state‑estimation tasks. The authors formulate the design objective using the Bayesian D‑optimality criterion, which seeks to maximize the log‑determinant of the posterior covariance (equivalently, to minimize the volume of the uncertainty ellipsoid). While the single‑fidelity D‑optimal sensor placement problem is well‑studied and benefits from submodularity, the introduction of multiple sensor fidelities destroys submodularity, making exact optimization NP‑hard and precluding the classic (1‑1/e) approximation guarantee.

To obtain tractable solutions, the authors propose a greedy algorithm that selects, at each step, the sensor (including its fidelity choice) that yields the largest increase in the D‑optimality objective per unit cost. The key computational insight is that adding a sensor corresponds to a rank‑one update of the information matrix. By applying the Sherman‑Morrison formula, the inverse of the posterior information matrix can be updated in O(ℓ²) time (ℓ is the reduced‑order model dimension), and the determinant change can be computed via the Matrix Determinant Lemma. This reduces the overall complexity of the greedy procedure from O(k·M·ℓ³) (as in naïve implementations) to O(k·M·ℓ²), where k is the number of selected sensors and M the number of candidate locations. The authors provide a detailed pseudo‑code and a complexity analysis, showing that the method scales to thousands of candidate sites and modest reduced‑order dimensions.

Recognizing that a single greedy pass may be sub‑optimal when costs and fidelities interact, the paper introduces an iterative refinement scheme. Starting from any feasible budget‑respecting placement, the algorithm repeatedly applies the greedy selection separately for each fidelity class while keeping the other class fixed, then merges the results and re‑evaluates the budget. This cycle continues until the marginal gain per cost falls below a prescribed threshold. The iterative method re‑uses the Sherman‑Morrison updates, so its additional computational burden is modest. Empirically, the iterative scheme consistently outperforms the one‑shot greedy approach in terms of the D‑optimality score.

Theoretical contributions include proofs of monotonicity of the D‑optimal objective under sensor addition, a Taylor‑series based analysis of marginal gains, and a discussion of the relationship between the multi‑fidelity design problem and the knapsack problem. Although submodularity is absent, the authors argue that the cost‑normalized marginal gain heuristic remains a reasonable surrogate for the true marginal benefit.

Experimental validation is performed on two benchmark problems: (1) reconstruction of sea‑surface temperature fields from satellite‑derived data, using a proper orthogonal decomposition (POD) basis of dimension ℓ≈30 and M≈5000 candidate locations; (2) flow reconstruction around a cylinder, with ℓ≈20 and M≈2000. For each case, several budget levels are tested. The proposed greedy and iterative algorithms are compared against random sensor selections. Results show that the greedy method improves the log‑determinant by roughly 15–20 % relative to random designs, while the iterative refinement adds another 5–10 % improvement. Correspondingly, the root‑mean‑square reconstruction error is reduced by a comparable margin. Moreover, the rank‑one update implementation yields speed‑ups of an order of magnitude compared with a naïve implementation.

The paper concludes that cost‑constrained multi‑fidelity sensor placement can be addressed effectively with a greedy‑based framework augmented by efficient linear‑algebraic updates, and that iterative refinement further enhances performance. Limitations include the need for a priori specification of sensor noise levels and costs, lack of formal approximation guarantees for the multi‑fidelity case, and validation on only two problem domains. Future work is suggested on learning cost‑fidelity parameters from data, extending the approach to non‑Gaussian or nonlinear Bayesian models, and developing adaptive, real‑time sensor placement strategies.