High-dimensional state estimation requires choosing sensor locations that maximize the information gained about critical system modes while respecting cost, communication, and physical constraints. The optimal placement depends on the estimation objective, measurement noise, and available models; however control and estimation theory offer principled criteria and scalable heuristics that perform reliably in practice.
Mathematical criteria and model reduction
Effective formulations use the observability Gramian or the Fisher information matrix as objective measures: placing sensors to maximize the log determinant or to minimize the trace or largest eigenvalue of the posterior covariance aligns directly with improved Kalman filter performance. Stephen Boyd Stanford University has shown how convex relaxations and semidefinite programming can be used to cast many sensor selection objectives into tractable forms. For very high dimensions, model reduction techniques such as proper orthogonal decomposition or balanced truncation let designers project onto dominant modes and place sensors to observe those modes first, limiting computational cost while retaining estimation quality.
Algorithms and performance guarantees
When objective functions are submodular—exhibiting diminishing returns—simple greedy algorithms achieve near-optimal guarantees. Andreas Krause ETH Zurich and collaborators developed sensor placement methods exploiting submodularity to obtain efficient, provably good selections for environmental and robotic sensing problems. Where objectives are nonconvex or non-submodular, randomized sampling, convex relaxations, and iterative refinement guided by local observability metrics are used.
Practical, cultural, and territorial considerations
Optimality on paper can conflict with real-world constraints: terrain, regulatory access, power and communication availability, and community acceptance shape feasible placements. In environmental monitoring, sensors must balance ecological impact and cultural sensitivities; in urban systems, territorial ownership and privacy impose placement limits. These constraints can change the effective objective and often require multi-objective formulations that trade pure information gain for robustness and acceptability.
Consequences of poor placement include unobservable modes, biased estimates, and degraded control or forecasting. Integrating domain expertise, reduced-order models, and principled objectives produces sensor deployments that are both computationally tractable and robust in practice, enabling reliable state estimation in large-scale systems.