The paper can be found here.
1. “We describe a method for replacing certain stages of rigid body simulation with a statistically-based approximation [to significantly improve performance by simplifying collision detection].” (first sentence, abstract, and second sentence, conclusion)
2. Rigid body simulation is computationally expensive, particularly when there are large numbers of rigid bodies involved.
3. Obtain a statistical “signature” for an object based on data collected regarding changes in linear and angular momentum for collisions of the object. Use this compact representation of the object’s response to collision events to replace collision detection and collision response calculations with simpler calculations.
4.a Achieves significant improvement in the performance of rigid body simulation at the cost of some fidelity.
4.b Appropriate for any graphics application where a tradeoff of fidelity for performance is desirable, such as when an object is obstructed, sufficiently far away from the viewer, or of sufficient complexity that a viewer cannot easily discern individual events.
5. none that I could find (but here’s the authors web page)