Notes on the precision metric used in 
D. Liang, M. Pennings, MJ Harrold, "Evaluating the Impact of 
Context-sensitivity on Andersen's Algorithm for Java Programs",
PASTE'05

Let 
1. i be an actual object instance created during program execution

2. n be the abstract object name representing that instance in the
static analysis algorithm (e.g., n might be the allocation site).  

3. C[i] be the set of callsites at which i is observed (at execution
time) as the receiver 

4. C_A[n] be the set of callsites at which algorithm A reports that
instances identified by n may be used as a receiver.

Then precision of A wrt object instance i is:

r_A[i] = |C[i]| / ( |C_A[n] intersect Cover| )

where Cover is the set of callsites that have been exercised by at least
one test during execution.

if r_A[i] == 1, then the algorithm A has perfectly identified the set
of observed executed callsites with receiver i.  if r_A[i] < 1, then
algorithm A has idenfied many call sites at which it says i can be a 
receiver, but where test data failed to observe i as a receiver.

To use this metric as a precision measure associated with algorithm A,
the authors are assuming that the static information may be spurious
if it disagrees with the dynamically observed information at
callsites.  however, this assumes that the test coverage exercises
most of the possible paths in the program.