When I initially set out on my quest for a universal input method that could bring people of all languages together, I had the quad layout in mind. Now when it's put to the test it doesn't perform that well. However the tree layout that's basically the same structure came out from this.
Surprisingly the linear variant when using a dynamic prediction performs better. This is highly ironic since I wrote this as an example of the worst case performance. Keep in mind that the linear approach with a static ordering still would perform extremely bad.
The silver-lining for the rectangular approach however would be in input where it's non-trivial to gather statistical relations between successive symbols. Such as math. Applying this approach to math might be a good idea in that sense.