| Principal Coordinates Analysis |
| See Also |
dij = aii + ajj - 2aij
If, for example, A is a similarity matrix then aii and ajj are both equal to 1.0, and so this is equivalent to:
dij = 2 (1 - aij)
Thus similar units are placed close together and dissimilar units are further apart.
The coordinates generated can be arbitrarily located in space as this will not alter the fitted inter-point distances. By convention, the points are centred to have their mean at the origin, and rotated to principal axes, so that the first r dimensions give the best r dimensional fit.
If B is a distance matrix, so that bij gives the observed distance between the ith and jth units, then the transformation
A = - B * B / 2
will lead to the analysis generating points with inter-point squared distance
dij = aii + ajj - 2aij = 0 + 0 - 2 * (- bij * bij / 2 ) = bij2
Therefore the analysis will give points that generate the supplied distances; the first r dimensions of the solution will give the best r-dimensional fit.