The first proposed criteria to measure the topology preservation of the SOM was the topographic product [1]. It tests the similarity of neighborhood on the map and in the input space. However, it has been pointed out that the topographic product does not work except for linear submanifolds because it cannot distinguish between a correct folding due to a folded data manifold from a folding due to a topological mismatch between the data manifold and the SOM lattice. In its stead Villmann et al. [42] proposed topographic function which measures the number of map units having adjacent receptive fields in the input space, but a distance greater than k on the grid:
where is the location of unit i on the grid and is the connectivity matrix; a binary matrix with element (i,j) equal to one only if map units i and j are two closest BMUs of any input vector from the training set. A problem with this measure is how to compare two different topographic functions. A simpler method was proposed by Kiviluoto [20]. He calculated topographic error as the proportion of sample vectors for which two best matching weight vectors are not in adjacent units:
where if the first and second BMUs of are not adjacent units, otherwise zero. The advantage of this measure is that the results are directly comparable between different mappings and even mappings of different data sets. It is also very easy to interpret.
Another branch of methods are those that do not use a data set at all. These methods do not measure the topology preservation property of the map but rather its local smoothness. One such method was proposed by Hämäläinen [9]. His measure of disorder requires that the distance between each unit i and its neighbors j is smaller than the distance between unit i and the neighbors k of the respective unit j:
where is the 1-neighborhood of unit i, is its cardinality, N is the total number of units on the map and is the step function.