That's not storing a collection in a table. That's approximating the storage of a collection in a table by using standard RDBMS features.
If you think these are the same somehow, I encourage you to think about the number of disk seeks required for recovering the full collection in both cases.
If all you need is the indexed values, then most mature RDBMs's supports using covering indexes in queries and satisfying the query entirely from the index.
Since the point of B-tree's is to minimize disk seeks and packing values with the same key close together, the number of disk seeks can often be made just as small as if the collection had been stored in the "owning" record.
Secondly, whether or not it is is an implementation detail. There's nothing inherently preventing RDBMSs from supporting "inlining" of collections from related tables. But in practice most RDBMSs choose to support that exactly through the much more generic optimization of supporting queries using covering indexes...
In other words: It only takes a lot of disk seeks if you use a RDBMS that is lacking in features and/or haven't added the proper indexes.
I didn't say they were the same, but instead was replying to the statement "RDBMSes typically can't do that, but document stores can.", but of course an RDBMS can store an "array of cities you've ever lived in".
I encourage you to think about the number of disk seeks required for recovering the full collection in both cases.
Ignoring XML, array types, or materialized views, for the RDBMS it is generally zero on virtually any modern platform, as the common case now is a database that is 100% in memory. A $79/m OVH server has 32GB. A low-end Dell server can have 256GB for $3000 (there are extremely few databases where the hot data surpasses even a GB). And the best thing about normalization is that it, like LZ, is a form of compression, yielding a more likely scenario that your database fits in memory.
That's all a side topic, however, and is neither here nor there on the long running NoSQL/SQL debate.
If you think these are the same somehow, I encourage you to think about the number of disk seeks required for recovering the full collection in both cases.