Today I read Simeon and Wadler's POPL 2003 paper, The Essence of XML, which should really be titled "The Essence of XML Schema". The paper points out that XML is not a good data representation because it isn't self describing or round-trippable:

It is not always self-describing, since the internal format corresponding to an external XML description depends crucially on the XML Schema that is used for validation (for instance, to tell whether data is an integer or a string). And it is not always round-tripping, since some pathological Schemas lack this property (for instance, if there is a type union of integers and strings)

They are assuming a strongly typed world, of course. The key observation in their work is that using a named typing model instead of (more commonly used) a structural typing model leads to a model where proving theorems about validation and erasure is very easy. The model in the paper only models a subset of XML Schema, but the remaining features are accounted for in the formal semantics for XQuery and XPath 2.0.

For practitioners, the bottom line is that the subset of Schema described by the paper is roundtrippable (convert internal value to external and back -- output the value as XML and then parse/validate) except for simple types that are lists or unions. For the case of reverse-roundtripping (parse/validate then output), the only problems arise when a base type has multiple representations for the same value, like leading zeros or number bases.