Prerequisites
Covariance matrices, quadratic minimization, spatial prediction, and unbiasedness.
With an unknown constant mean, an unbiased linear predictor w^Tz requires weights summing to one. Minimize prediction-error variance C00-2w^Tc+w^TCw subject to that constraint. The weights depend on covariance, not only distance or map appearance.
Worked example
If two sites have symmetric covariance with the target and each other, the constrained optimum can give equal weights. Asymmetric target covariance can shift the optimum even if observed values happen to be equal.
Model and Assumptions
Invented covariance model: observation covariance C=[[4,1],[1,4]], target covariances c=(2,1), target variance C00=4. Both observations and target share an unknown constant mean. Observed values are z=(10,16). Assume this covariance is known and valid.
Further inquiry
Derive the constrained solution using both substitution and a Lagrange multiplier. Compare its variance with equal weighting, then describe how trend misspecification or uncertain covariance parameters would change your assessment.
Review criteria
- Recover weights (2/3,1/3) and prediction 12.
- Compare variance 10/3 with equal-weight variance 3.5.
- Distinguish model-based error variance from realized error and omitted uncertainty.