From my initial look through boost::optional's implementation, it seems as though it's doing nothing to optimize its space usage in some trivial cases.
The first one I was surprised by was optional<T&> - still uses sizeof(struct { T& t; }) + sizeof(bool) space, rather than simply using a pointer to represent both the reference and the unspecified state (using null, since a reference doesn't use that state). Is there a reason for this (perhaps I'm missing something & a reference is smaller than a pointer in certain implementations). Similarly - would it be worth considering optimizing optional<bool> (though, as per the documentation, this is perhaps a little "tricky" at best, owing to the implicit bool of optional itself) to use one byte instead of two? [long term - I'm wondering if a generic "efficient optionality" could be implemented by arbitrary types - privately exposing an extra unused state from their representation for use by optional only. Then, potentially, making versions of (boost/std)::shared_ptr/unique_ptr that would be never-null (only constructible via make_shared like machinery) & optional<shared_non_null_ptr<T>> would then be an explicit representation of the optional situation, but without any overhead - shared_non_null_ptr<T> would simply expose its null state only to optional<shared_non_null_ptr<T>>]