>From the perspective of laying out the scatterplot, the point cloud could just as well be replaced by a lines-view or even a histogram-view or a density-view. Similarly, some other kind of view - perhaps a boxplot-view - could be used in place of the axes in the margins. In fact, we have defined a type of view - called a plot - of which the scatterplot is a special case, whose task it is to position various views in the margins around an interior view. The plot format encompasses many different kinds of plot, since there is no restriction on the type of view that may be placed at each location. Typically, only the left and bottom margin views are present, and these are axes. More generally, plots allow for multiple interior views which are overlaid on the interior region, similarly for left, bottom, top and right views. Of these, the ``layers'' of interior view are by far the most useful and interesting: one can superimpose smooths or a least-squares fit on a point cloud, or overlay a density-view on a histogram-view for instance.
For the scatterplot to be meaningful there must be some connection or ``glue'' between its constituent point cloud and axes. For instance, a point symbol at x=0 should line up with the x-axis tic mark at 0. More precisely, the transformation from abstract to screen coordinates for x locations should be the same for the point cloud and the x-axis, (similarly for y locations and the y-axis.) In our system, this occurs because (i) the scatterplot assigns locations to its subviews so that they line up in the appropriate way and (ii) a constraint system ensures that the horizontal extent of the interior view is the same as the horizontal extent of the bottom and top views; similarly for vertical extents.