See first comment here:
http://weblogs.java.net/blog/campbell/archive/2009/01/effects_in_java_2.html#comments

If a node has a DropShadow (or similar Gaussian-based) effect applied to it, and it is being transformed with scale factors less than 1, it seems we should be able to scale the radius of the kernel accordingly and apply the effect to the downscaled representation. This should be more optimal than the current approach where we'd first render the node in user space, filter it to produce a shadow, and then downscale the resulting shadow image.

For example, suppose a node that's 1000x1000 in user space, with a DropShadow.radius of 10, and scaleX/scaleY=0.1. We should be able to use the vector-downscaled 100x100 representation of the node, and a DropShadow.radius of 1 (no raster scaling of the shadow image required).