I find it best to think of them as sets. A lot of operators treat them that way (e.g. containment), and you can use GIN indexing to accelerate access to them using those operators.