
We define a binary spatial weight matrix as:
\[ w_{ij}(h)=\bigg\{\begin{array}{l l} 1\text{, if } d_{ij} = h\\ 0\text{, otherwise}\\ \end{array} \]
Autocovariance:
\[ C_{z}(h) = \frac{\sum_{i=1}^{n}{w_{ij}(h)(z_i^2 - \bar{z})(z_j^2 - \bar{z})}}{\sum_{i=1}^n{w_{ij}(h)}} \]
Semivariance:
\[ \hat{\gamma}_{z}(h) = \frac{\sum_{i=1}^{n}{w_{ij}(h)(z_i - z_j)^2}}{2\sum_{i=1}^n{w_{ij}(h)}} \]
Covariogram:

Semivariogram:

The autocovariance, \(C_{z}(h)\), and semivariance, \(\hat{\gamma}_{z}(h)\), are related as follows: \[ C_{z}(h) = \sigma^2 - \hat{\gamma}_{z}(h) \] where \(\sigma^2\) is the sample variance.
All kriging variants produce predictions of the form:
\[ \hat{Z}(s_0)=\sum_{i=1}^n \lambda_i Z(s_i) \]
where
The key difference from IDW or k‑nearest‑neighbor averaging is that kriging weights are determined by the spatial covariance (or semivariogram) structure, not just distance. The covariance model encodes how similarity decays with spatial separation, and kriging uses this to compute the best linear unbiased predictor (BLUP).
A spatial process with a non‑constant mean can be written as: \(Z(s_i) = f(s_i) + \epsilon(s_i)\)
where
Prediction at a new location \(s_p\) decomposes into trend + kriged residual:
\(\hat{Z}(s_p) = \underbrace{\hat{f}(s_p)}_{\text{a smooth estimator, e.g., trend surface}} + \hat{\epsilon}(s_p)\)
The residual prediction is a weighted combination of observed residuals: \(\hat{\epsilon}(s_p) = \sum_{i=1}^n {\lambda_{i}\epsilon_i}\) and \(\epsilon_i = Z(s_i) - \hat{f}(s_i)\).
To obtain the weights \(\lambda\), kriging solves:
Because unbiasedness constraint, the bias term vanishes, and the expected mean squared error reduces to the prediction variance: \[ MSE(s_p) = E[(\hat{Z}(s_p)-Z(s_p))^2] = Var(\hat{\epsilon}(s_p)-\epsilon(s_p)) \]
Once the optimal weights \(\lambda\) are obtained for location \(s_p\), the Gaussian assumption on the residual field provides a closed-form expression for the kriging prediction variance, which can be used to construct interval estimates.
2026 Zehui Yin