boundingBox computes incorrect bounding box for transformed diagrams

[fryguybob]

(Imported from http://code.google.com/p/diagrams/issues/detail?id=87. Original issue from andreas....@gmail.com on July 25, 2012, 09:03:51 PM UTC)

What steps will reproduce the problem?

1. The expression
   boundingBox (translate (r2 (10,10)) unitSquare :: D R2) calculates a bounding box of size 10.5 times 10.5 (instead of 1 times 1).

So the result is:
BoundingBox (P (R2 {unR2 = (0.0,0.0)})) (P (R2 {unR2 = (10.5,10.5)}))

instead of the expected output:
BoundingBox (P (R2 {unR2 = (9.5,9.5)})) (P (R2 {unR2 = (10.5,10.5)}))

I tested this with version 0.5.0.1 diagrams-lib

[fryguybob]

(Imported. Original comment by byor...@gmail.com on July 25, 2012, 09:31:37 PM UTC)

Relevant IRC discussion:

(16:08) < fryguybob> abernstein: I just looked at boundingBox a little and I think I'm leaning "that's a bug" at the moment.
(16:09) :)
(16:09) fryguybob: ok
(16:32) < mgsloan> oops
(16:32) < mgsloan> what's wrong with it? I think I may have written it
(16:33) < fryguybob> mgsloan: fromPoints doesn't seem appropriate there. That will only work when the origin is inside the diagram.
(16:35) < fryguybob> mgsloan: For each unit vector and +/- only one value should be used, not all values.
(16:36) < mgsloan> ahh
(16:37) < mgsloan> but envelope returns negative numbers, right?
(16:37) < mgsloan> so it should work anyway
(16:37) < mgsloan> but you're right, that this is probably less efficient than possible
(16:38) < fryguybob> envelopeP gives a point, nothing negative going on there.
(16:45) < mgsloan> right, but it'll still work if the envelope does not contain the origin
(16:46) < byorgey> sounds like a bug in boundingBox, though I haven't looked into the details
(16:46) < byorgey> however, the fact that there is no Transformable instance for BoundingBox is NOT a bug
(16:47) < byorgey> in fact, there used to be one and I removed it. Applying affine transformations to bounding boxes does not make sense.
(16:47) < fryguybob> byorgey: I can see that.
(16:47) < byorgey> However, there IS a HasOrigin instance for BoundingBox (or at least there should be) so you can still translate them
(16:48) < byorgey> but using 'moveTo' and friends instead of 'translate'
(16:49) < mgsloan> right, if you want to apply affine transforms, you should use getAllCorners, transform those, and use fromPoints
(16:49) < byorgey> sure, that will give a conservative approximation to a transformed bounding box
(16:49) < byorgey> but it's not compositional
(16:50) i tried boundingBox (moveTo (p2 (10,10)) unitSquare :: D R2), but i still get a 10.5 times 10.5 box
(16:51) < byorgey> abernstein: and what happens if you do (moveTo (p2 (10,10)) (boundingBox (unitSquare :: D R2))) ?
(16:52) < abernstei> byorgey: that works :)
(16:53) < byorgey> oh, I see what the problem is
(16:53) < byorgey> I think
(16:54) < fryguybob> byorgey: envelopP give a point on the hyperplane, not a point on the bounding box.
(16:54) < byorgey> yes, exactly
(16:54) < fryguybob> (as it should)
(16:54) < byorgey> and in this case it happens to be giving points on each of the four axes.
(16:55) < fryguybob> Right
(16:55) < byorgey> hmm, so what's the right way to implement it
(16:56) < fryguybob> if you take the magnitudes (or use some basis function for efficiency) you can take the combinations of mins an maxs to get the box
(17:01) < fryguybob> If no one figures out something before tonight I'll come up with a fix when I'm not at work :D.
(17:01) < byorgey> ok =)
(17:01) < byorgey> yes, I guess you can use http://hackage.haskell.org/packages/archive/vector-space/0.8.2/doc/html/Data-Basis.html#v:decompose
(17:02) < byorgey> just use envelopeV then call decompose and extract the proper coordinates

[byorgey]

Fixed by 77ec072