Tom,

I'm well aware that the relational model and TTM require relations that consist of atomic attributes. What I see as a problem is that your article uses another criterion for 1NF:

You define 1NF based on intended usage. This doesn't make any sense to me. Suppose I create a relation variable using my RDBMS and call it T. T happens to have a string attribute that encodes multiple telephone numbers. That's OK because a string is by definition an atomic value. According to your definition T is in 1NF as long as "the database doesn't know anything about that encoding". I don't have any problem with that so far because I don't ever intend to split out the phone numbers in the database. T is in 1NF.

Later on however, I change my mind. I want to split out the phone numbers using some code in my database. As you put it: "There will have to be code in queries to decode that complex string". Now you say that T is not in 1NF. But it's the same relation variable as before! The atomicity of the string and the structure of the relation has not changed. Just because I intend to use it differently T cannot suddenly stop being a relation. So your suggested criteria is an additional requirement for 1NF over and above the requirement for T to be a relation - a relation being a thing which is defined by structure alone and never by intended usage. According to you, the relation variable T can change from moment to moment from being in 1NF to not being in 1NF depending on what code exists in my database.

The relations which according to your definition aren't in 1NF could also still satisfy higher normal forms like BCNF or 5NF even though you say they violate 1NF. If I've misunderstood your criterion for deciding 1NF then I think you need to clarify this point. To me it is fundamental that 1NF means nothing more than the definition of a relation, which is simply a data structure in a database. 1NF does not depend in any way on the intended usage of the data but purely on its structure.

However Codd says that Null marks are not values and Date and pretty much everyone else agrees. Codd never used the the domain theory definition of a domain in any work of his that I'm aware of. In fact he specifically requires that null marks be independent of domain - by which I understand him to mean they are orthogonal to the concept of a domain. If you are putting forward a different definition of nulls and null behaviour (as you have done in another SSC thread) then you ought to make it clear that your 1NF definition isn't generally applicable to other systems using different things called "nulls".