For the simulated reads, they know exactly where the reads came from, because they are able to create that link at the time paired-end reads are generated from the genome. Those simulated reads would be created through a process similar to the following:
- Pick a single position in the genome at random
- Pick another region in the same chromosome that is <fragment length>+error bases away from the first location
- Extend these positions by <read length> (one read in the forward direction, and the other read in the reverse complement direction) to generate a read pair
- Change at random some bases (or none) within the reads to simulate sequencing error
- store the read pair in a FASTA file, together with the original position
- repeat steps 1-5 a few hundred million times
The effectiveness of the algorithm would be based on how good the algorithm was at picking the correct genomic location.
For the real paired-end data, it's a classical mapping problem: there's no ground truth, so all that can be done is starting at step 5, and working back to all the possibilities for step 1. If the two sequences from a read pair are derived from the same fragment, they should map to approximately the same genomic location and [usually] in a reverse-complement orientation to each other. Such an alignment would be called "consistent", or "properly mapped".