Overdraw is simply where multiple layers of translucency stack on screen. If your translucent material had black in the alpha, that means that parts of the material were “rendered” only to be invisible in the final result. Even though those pixels were invisible, they cost just as much as the visible pixels. So you can see how if you have a HUGE polygon just to have a small translucent effect, it will be very inefficient. Making the geometry fit closely around only the visible parts of your opacity mask is something you should always do when possible.
For instance if you have a blob shadow (round), then all of the corners of a square polygon are going to be wasted overdraw. Simply turning the quad into an octagon by cutting a bit off all of the corners will make a big difference and save many wasted pixels from rendering. You could tesellate more and more to reduce overdraw but at some point you would start causing the scene to be too expensive from polygons so it is best to find some middle ground. Ie just making the shape an octagon is a good compromise because its still low poly and cuts away most of the overdraw.
For a circle or blob shadow you could calculate the savings.
A square of pixels 100x100 has an area of 10,000 pixels that all have to be rendered.
A circle with a diameter of 100 has an area of only 7,853. So by rendering a circle using square geometry you would be wasting ~1,200 pixels or 12% of the cost. This is probably not a huge deal in this case but for some shapes it can make a huge difference.
for particle systems, a feature was recently added to cut out the geomtry specifically for subUVs automatically for particle systems. Probably 4.12.