reparameterisation

Scales and geoms interact because the geom may not be parameterised in variables that the scales know about. For example:

• bar: width, ymin = 0
• boxplot: width
• crossbar: width
• errorbar: width
• errorbarh: height
• map: uses locations of map polygons
• tile: width and height

This has impacts in two places:

• training: needs to be mediated through the geom
• mapping:
• munching: munching requires only locations

Reparameterisation is currently implemented as a data transformation after statistics are calculated and their aesthetics are remapped. This is because

For large data also want to not have to convert data into a data frame.

Position scales are special because:

• they vary across facets, so there can be multiple copies
• some statistics use them: smooth, bin, bin2d, binhex, density, abline, function
• many statistics modify them

train_position.gbar <- function(train_x, train_y)