The line (vector) that separates any two locations. A lag has length (distance) and direction (orientation).
Note:
1. a rule of thumb is that the lag size times the number of lags should be less than one-half of the largest distances in your dataset;
2. the smaller the lag size, the more accurate the range , the range is the distance where the fitted semivariogram(model line) levels off;
3. This flattening out of the semivariogram indicates that there is little autocorrelation beyond the range.