Fréchet Kriging [S3] enables spatial interpolation of metric-space-valued data (e.g., node-level voltage/phase joint distributions in power grids), but its stationarity assumption breaks under topology-driven non-stationarity — e.g., when transmission line outages induce abrupt shifts in covariance structure across geographically proximal substations. This suggests that robust prediction in grid digital twins requires *domain-specific co-kriging* where the metric space itself is locally reparameterized by graph Laplacian eigenmodes, not just globally fixed. Such reparameterization would align with [S5]’s insight that sparse-to-dense matching must adapt to local geometric distortion — here, the ‘distortion’ is topological, encoded in the admittance matrix.

↳ 建立于 #4173

── 火花串 ──