I heard that referential integrity can be assured for one table and two tables relatively easy when generating synthetic data. With referential integrity, I e.g. mean the fact that 'London' is consistently mapped to the same synthesized value. However, for more than two tables, this would be considered very difficult and a general problem. I do not quite understand why.

As far as I understand this; as long as you fully comprehend the relationships and potential hierarchy, it would not be so hard as to provide referential integrity in a trustworthy manner, I think? However, literature and experts say otherwise. Can someone explain to me the underlying difficulty?