next up previous contents
Next: Matrix Form Up: The Kriging Estimator Previous: Non-bias Condition   Contents


Minimum Estimation Variance

The estimation variance $E(Z_{V}-/hat{Z}_{V})^{2}$ can be expanded as follows:

/begin{displaymath}E(Z_{V}-/hat{Z}_{V})^{2}=EZ_{V}^{2}-2E(Z_{V}/hat{Z}_{V})+E/hat{Z}_{V}^{2}/
/ / ,/end{displaymath}

with

/begin{displaymath}EZ_{V}^{2}=/frac{1}{V^2} /int_{V} /int_{V} E[Z({/boma x})Z({/boma x}')]
d{/boma x}'d{/boma x}=/bar{C}(V,V)+m^{2}/ / / ,/end{displaymath}


/begin{displaymath}E(Z_{V}/hat{Z}_{V})= /sum_{i}/frac{/lambda_i}{Vv_i}/int_{V}/i...
...}'d{/boma x}=
/sum_{i} /lambda_{i}/bar{C}(V,v_{i})+m^{2}/ / / ,/end{displaymath}


/begin{displaymath}/begin{array}{rl}
E/hat{Z}^{2}& = /sum_i/sum_{j} /lambda_{i} ...
...ambda_{i} /lambda_{j} /bar{C}(v_{i},v_{j}) + m^{2}.
/end{array}/end{displaymath}

The terms $m^2$ are eliminated and, thus, we have

/begin{displaymath}/sigma_{E}^{2}=E(Z_{V}-/hat{Z}_{V})^{2}=/bar{C}(V,V)-2 /sum_{...
..._{i}/sum_{j}/lambda_{i} /lambda_{j}
/bar{C}(v_{i},v_{j})/ / / ./end{displaymath}

The estimation variance can, thus, be expressed as a quadratic form in $/lambda_i, /lambda_j$, and is to be minimized subject to the non-bias condition $/sum /lambda_{i}=1$. The optimal weights are obtained from standard Lagrangian techniques by setting to zero each of the $n+1$ partial derivatives (in respect to $/lambda_i$ and to the Lagrange parameter $/mu $) of

/begin{displaymath}/sigma_{E}^{2} - 2 /mu(/sum_{i} /lambda_{i} - 1)/ / ./end{displaymath}

This procedure provides a system of $n+1$ linear equations in $n+1$ unknowns (the $n$ weights $/lambda_i$, and the Lagrange parameter $/mu $) which is called the ``kriging system'':
/fbox{ $
/begin{array}{rl}
/sum_{j=1}^{n} /lambda_{j}/bar{C}(v_{i},v_{j})-/mu & ...
...r{C}(V,v_{i}), /; i=1,/ldots,n,//
/sum_{j=1}^{n} /lambda_{j}& =1.
/end{array}$}

The minimum estimation variance, or ``kriging variance'', can then be written as

/begin{displaymath}/sigma_{K}^{2}=min E(Z_{V}-/hat{Z}_{V})^{2}=/bar{C}(V,V)+/mu-
/sum_{i=1}^{n} /lambda_{i}/bar{C}(v_{i},V)/ / / ./end{displaymath}

This kriging system can also be expressed in terms of the semi-variogram function $/gamma({/boma h})$, particularly when the random function $Z({/boma x})$ is intrinsic only and the covariance function $C({/boma h})$ is not defined:

/begin{displaymath}/sum_{j} /lambda_{j} /bar{/gamma}(v_{i},v_{j})+/mu=/bar{/gamma}(v_{i},V),
/ / / i=1,/ldots,n/ / / ,/end{displaymath}


/begin{displaymath}/sum_{j} /lambda_{j}=1/end{displaymath}


/begin{displaymath}/sigma_{K}^{2}=/sum_{i=1}^{n} /lambda_{i} /bar{/gamma}(v_{i},V)+/mu- /bar{/gamma}(V,V)/ / / ./end{displaymath}


next up previous contents
Next: Matrix Form Up: The Kriging Estimator Previous: Non-bias Condition   Contents
Rudolf Dutter 2003-03-13