I am trying to get a definitive version of what is First Normal Form. Everything I read has a slightly different spin.

Many authorities, such as Date, say that by definition a relation is always in First Normal Form, while others give a list of requirements. This means that there are from zero to many requirements for 1NF.

I suppose the difference is that between tables and relations: tables can be a complete mess, while a relationship follows certain restrictions. The fact that a relation is represented as a table in SQL thus creates some confusion.

I am focusing specifically on 1NF as it relates to SQL databases. The question is: what properties are required to ensure that a table is in the first normal form?

Many authorities suggest that if a table represents a relation, then it is already in 1NF. This pushes the definition of 1NF back to the definition of a relation.

Here are some properties of a table in 1NF:

  • Column Order is insignificant [1]
  • Rows Order is insignificant
  • All rows are the same length (ie, row data matches column headers)
  • There are no duplicate rows (this can be guaranteed using a surrogate primary key, but the PK is itself not required)
  • There are no repeating columns
  • Each column contains a single value (atomic)

[1] Technically attributes are unordered, but in a table, the row data must be in the same order as the column headers. However, the actual order is insignificant.

On multiple data:

The concept of atomic data is that an item cannot be further broken down. This concept has been qualified in that although technically everything can be broken down ad nauseum, the data in question cannot be practically broken down any further, depending on how the data will be used.

For example, a full address or a full name should normally be broken down further, but the components such as the given name or town name should probably not be broken down any further, despite the fact that as strings they can be.

As regards repeating columns, it is a poor design column to have nearly repeating columns, such as phone1, phone2 etc. In general, repeated data indicates the need for an additional related table.


There should not be any relationship between rows, other than that they conform to the same headers.

There should also be no relationship between columns, but I believe that is the subject of higher normal forms.

The question is: How much of the above is in the definition of 1NF? Does the independent rows bit also come into it?


5 Answers 5



The definition of normal form (which from the presentation of “Further Normalization of the Data Base Relational Model” in 1971 is known as first normal form) and the definition of the relational paradigm itself was published in 1970 in the scientific paper that provided a strong foundation for the practice of database administration, i.e., “A Relational Model of Data for Large Shared Data Banks” (RM for brevity) created by Dr. E. F. Codd, who is a Turing Award recipient and the authority with regard to the relational framework.

Yes, there is plenty of explanations, interpretations, expositions, deviations and opinions about Dr. Codd’s text, but I personally prefer to stick to the original source and I highly suggest that you analyze it by yourself so that you can draw your own conclusions.

I certainly do not understand the RM in its entirety, but what I do understand of it permits me appreciating its excellence, vision, intention and scope, and although decades later one can note that it has a few slight imprecisions, they do not reduce, in any way, its genius and elegance. In its field, the RM has stood the test of time in a unique way, and remains unmatched.

The act of emphasizing the aforementioned imprecisions would be —using a charitable term— unfair because, seeing it from from a considerable distance, this seminal material required a few refinements and extensions, yes, but the main body of the work was rock solid from the very conception (and, indeed, Dr. Codd made most —if not all— of such refinements and extensions himself).

I continue rereading the RM constantly in order to strengthen my comprehension of this exceptional source of knowledge (and my esteem of it keeps growing on every reread); the objective is to stand on the shoulders of giants.

Relations and tables

It is important to note that as relations are abstract resources, Dr. Codd envisioned the utility of representing them in tabular form (he initially used the term “array representation” but subsequently utilized “table” or “r-table”), so that the users, designers and administrators of a relational database (RDB) can approach them in a more familiar or concrete manner. Therefore, within the context of a RDB implementation, it is valid to use table as a shorthand for relation, as long as said table stands for an actual relation. This feature is —although obvious— quite significant because before evaluating whether or not a table represents a relation that complies with first normal form (1NF), it has to represent, precisely, a relation.

The RM naturally contains the qualities that a table must have to determine if it in fact portrays a relation, but I will offer an informal and unpretentious interpretation about them here (another one, yes!):

  • It must have a name (each particular relation in a database structure must be distinguished from the rest).
  • Each of its rows must depict exactly one tuple of the pertinent relation.
  • The order of its rows is not important at all.
  • Each of its columns must have a name that stands for the meaning of exactly one domain of the concerning relation, and said name must be different from the names of the rest of the columns of the table (a column must be uniquely differentiated and must carry a distinctive meaning and, yes, the role played by a database modeler and the business experts to define each domain of significance with accuracy is paramount)
  • The order of its columns has no significance.
  • All of its rows must have the same number of columns.
  • It must have at least one column, or one combination of columns, that uniquely identifies every one of the tuples depicted via rows; in this way, all of the rows must be different (yes, this stresses the importance of having at least one KEY declared, and when there are two or more KEYs one should be defined as PRIMARY based on pragmatic reasons, while the rest can be deemed as ALTERNATE; but yes, before making the decision, each of the KEYs was a “candidate” for a definition as PRIMARY).

Having a table that in actual fact represents a relation is critical as, when it undergoes manipulation operations of a relational kind, then the result is, again, a table that represents a relation. In this manner, the behavior of said table is predictable.

Atomic domains (columns)

In the first sections of the RM, Dr. Codd presents several samples of relations in order to introduce some concepts; so, in order to comprehend the meaning of atomic domain, let us start with the following excerpt from the RM that details some pertinent points:

So far, we have discussed examples of relations which are defined on simple domains—domains whose elements are atomic (nondecomposable) values. Nonatomic values can be discussed within the relational framework. Thus, some domains may have relations as elements. These relations may, in turn, be defined on nonsimple domains, and so on.

In this way, one can say that each of the aforementioned expository relations fits in one of two kinds, say either kind A or kind B:

  • Kind A groups only relations (tables) that are structured with domains (columns) that contain exclusively simple values in every one of their tuples (rows), i.e., such domains (columns) do not contain relations (tables) as values, which in this context means that the values are atomic because they cannot be decomposed successively into new relations (tables). Hence, the relations of this class are the ones that are normalized, i.e., they comply with 1NF, their form is desirable.

  • Kind B is exclusively integrated by relations (tables) that have one or more domains (columns) that hold relations as values in each respective tuple (row), and that signifies that said values are nonatomic since they can be subsequently broken down into new relations (tables), i.e., they are decomposable. Thus, the relations of this sort are unnormalized, i.e., they infringe 1NF, they are in an undesirable form.


Dr. Codd introduces the section about normalization in the RM with the following paragraph:

A relation whose domains are all simple can be represented in storage by a two-dimensional column-homogeneous array of the kind discussed above. Some more complicated data structure is necessary for a relation with one or more nonsimple domains. For this reason (and others to be cited below) the possibility of eliminating nonsimple domains appears worth investigating! There is, in fact, a very simple elimination procedure, which we shall call normalization.

Then he goes on to show:

  1. A group of relations where one is unnormalized (it has domains that contain relations as values, i.e., they are nonatomic; i.e., they are non-simple)

  2. A group of relations that are a normalized (i.e., one that was decomposed; i.e., one which relation valued domains were broken down into simple ones which signifies that they are atomic)

And then he describes the procedure for obtaining normalized relations from unnormalized ones.

In this respect, the relations he employed to illustrate a normalization exercise and the exercise description itself are quite clear, and I recommend again that you analyze them yourself (and I as well hope this encourages some readers to engage with the text).

Sucessively, he indicates:

Further operations of a normalizing kind are possible. These are not discussed in this paper.

And said operations, i.e., second and third normal form (2NF and 3NF) are actually detailed in “Further Normalization of the Data Base Relational Model”, and as mentioned above, after the presentation (and the later printing and publication) of this paper, the original normal form became known as first normal form.

As a practitioner can observe, having unnormalized relations (tables) introduces (almost always unnecessary) convolution into RDB implementations.

A relation that satisfies 1NF, eases the definition of constraints and data manipulation operations which can be implemented by means of a data sublanguage that is less complex than that required for unnormalized relations (tables), as Dr. Codd points out in the following lines:

The adoption of a relational model of data, as described above, permits the development of a universal data sublanguage based on an applied predicate calculus. A first-order predicate calculus suffices if the collection of relations is in normal form. Such a language would provide a yardstick of linguistic power for all other proposed data languages, and would itself be a strong candidate for embedding (with appropriate syntactic modification) in a variety of host languages (programming, command- or problem- oriented). […]


The universality of the data sublanguage lies in its descriptive ability (not its computing ability).

The bewilderment

From my point of view, the bewilderment has arisen, due to (a) the aforementioned excess of interpretations, explanations, etc., about 1NF and the RM itself, and because of (b) further attempts to redefine 1NF that state that having relations with domains that hold values that are, in turn, relations comply with 1NF as long as they are one single value for each corresponding tuple.

My take on your other pointss

There should not be any relationship between rows, other than that they conform to the same headers.

I am not sure if I understand the intention of that statement correctly but, apart from conforming to the same headers, there must be a connection between the (tuples) rows of a relation (table) as each of them should be an assertion about a particular occurrence of the specific entity type (defined in terms of the business context of interest) that the relation (table) is supposed to represent.

There should also be no relationship between columns, but I believe that is the subject of higher normal forms.

I do not know if I am properly interpreting the meaning of that statement either but, in fact, and in accordance with my response to the previous aspect, there must be a relationship between the domains (columns) of a relation (table) as well, which is precisely why it is a relation (the essential structure of the relational model and of a concrete RDB implementation).

To exemplify, with regard to the hypothetical relation (table)

  • Salary (PersonNumber, EffectiveDate, Amount)

the tuple (row)

  • Salary (x, y, z)

would convey the meaning

  • The Salary payed to the Person identified by PersonNumber x, on EffectiveDate y corresponds to the Amount of z

Therefore, each tuple (row) of the Salary relation (table) must fit into the structure of the assertion shown above, and the difference would be the replacement of the pertinent domain (column) values, but there must exist a relationship between (a) all the Salary domains (columns) and also between (b) all their corresponding values with respect to each tuple (row); such a relationship it is indispensable.

Higher normal forms (2NF and 3NF) are useful to get rid of functional dependencies between domains (columns) of a relation (table), they assist in avoiding undesirable connections between domains (columns), as said undesirable connections allow the introduction of update anomalies. Both 2NF and 3NF are helpful to test the soundness of the structure of the relations (tables) in a certain RDB implementation.


An illustration. Take a text string that contains a typical US address:

"123 Cornhusk Rd., South Succotash, NY 12345"

Writing a query to find all residents of Cornhusk Road or a particular resident of the town of South Succotash or the state of New York would be a daunting task. Particularly when you have the following strings in the data:

"123 Cornhusk Road, South Succotash, NY 12345"
"123 Cornhusk Rd., South Succotash, New York 12345"
"123 Cornhusk, South Succotash, NY 12345"
"123 Cornhusk Rd., SOUTH SUCCOTASH, NY 12345"

Each of these specify the same actual address (and this doesn't even consider likely misspellings like "Succatash") but writing the algorithm to successfully compare them is something I would NOT want to dedicate my last remaining years on this Earth to do.

Enter first normal form. I won't go into the details of how it's done, just consider a common form as seen in most databases:

create table Addresses(
  Street  varchar,
  City    int        references Cities( ID ),
  State   char( 2 )  references States( ID ),
  Zip     int

This is not a complete normalization. For example, you could further divide Street into StreetNum and StreetName but this is far enough for a simple illustration and it really is about as far as the normalization process is taken in real life. There is still a small problem of finding all the residents of Cornhusk Road but if you searched Street for the substring "Cornhusk" and there happened to be a town named Cornhusk somewhere, at least that wouldn't cause confusion.

The problem with the definition as provided by Date, et al, is that they fail to look inside a piece of data and consider the meaning of that data. Not that I particularly blame them, it can be quite difficult.

So we have to take a "string of characters" and turn it into an "address." But coming up with an all-inclusive definition of address is not as simple as it looks. Especially when working with addresses in more than one country.

First, we fix the context. Nothing has meaning without context. Our context here is address. Within that context, we can see parts of the data that have specific meanings, such as Street, City, etc. We assign each meaning a separate field.

The difficult part is that data, like words, can have different meanings to different people or some people can see meaning where others do not. It can be a project all to itself and require a good deal of cooperative effort. But it is a critical element in good database design.

The point of normalization is to eliminate or at least minimize anomalous data. Consider the fact that, in the table above, a Town or Zip code may be entered that does not exist in the state that was selected. Or a street entered that does not actually exist in the town that was selected. Ah, but now we are getting into second and third normal form and that is another topic.

  • While this may be an example of bad design, it has nothing to do with first normal form.
    – JacquesB
    Commented Jun 15, 2021 at 14:08

Think of 1NF as an introduction to the mathematical concept of relations, rather than a particular data structure or fixed set of rules. Mathematical structures like relations can be represented in different ways - tables are just one of the most convenient ways. When using tables, there are restrictions to ensure that they represent their intended relations clearly and that operations on tables are logically sound with respect to the represented relations.

When we say that column and row order and duplication are insignificant, this is to ensure that all significant information is recorded as values in the table and accessible to our queries, and not encoded into the presentation of the table. While few, if any, authors have explicitly banned the use of color in tables, understanding the purpose of the restrictions above would lead us to similarly exclude the use of meaningful presentation colors so that significant color values have to be explicitly recorded. Date and other authors also deprecate hidden row identifiers for the same reason.

The concept of atomicity is about ensuring that all significant structure is expressed as relations, so that we're able to analyze dependencies in all our data and prevent anomalies, and so that that all meaningful structure is uniformly accessible to relational operations. Composite values aren't invalid, but they require domain-specific operators to unpack, which increases complexity, and they can introduce redundancy. Of course, in practice some simple composite types and associated operators are convenient and help increase the compactness and expressivity of queries, but this doesn't contradict the theory.

Row dependencies like multivalued dependencies don't violate 1NF, but like dependencies among columns, is the subject of higher normal forms. Nearly repeating columns like phone1, phone2, etc don't violate 1NF either, even though they're poor design.


The definition and explanation from WikiPedia about 1NF, I think is quite good. - joanolo

Expanding on one sentence in the Wikipedia article:

Seen as telephone numbers, the text is not atomic

This may give you a handle on why there is so much confusion. If this is just a gob of text, then it is atomic. But if it's seen as telephone numbers, then it's not atomic. Date and others sidestep this issue. It has to do with what the data means. When you analyze the subject matter, you have to grapple with what the data means.

Whether there is any point to breaking it down further is relevant to the question of whether first normal form has been met. The point behind 1NF is to provide keyed access to all data. But if the thing you retrieve via keyed access has substructure, then you don't have keyed access to the substructure. - Walter Mitty

"1NF" is used to mean a bunch of different things, many of which are simultaneoulsly nonsensical and common. See my answer to "Normalization in database management system", including its links. One of those is my answer to "What is atomicity in dbms". (Both at Stack Overflow.) - philipxy

Community Wiki answer created from comments left on the question


First normal form have a straightforward definition: Eliminate domains which have relations as element.

SQL does not support columns with tables as values, so in a SQL database table are in first normal form by necessity. So outside of exam questions, most practitioners does not need to worry about 1NF.

Unfortunately a number of misunderstandings have sprung up around 1NF. Some context around the purpose of 1NF will make it clearer what exactly it means and don't mean.

In the paper "Further normalization of the data base relational model" by E.F.Codd he shows this flowchart:

     | eliminate domains which 
     | have relations as elements 
     | eliminate non-full dependence 
     | of non-prime attributes on 
     | candidate keys
     | eliminate transitive dependence 
     | of non-prime attributes on 
     | candidate keys 

(The first normal form is first defined in Codds "A relational model of data for large shared data banks.", but "Further normalization..." frames it as one of three normal forms as above.)

First thing to notice is there is a stage of unnormalized relations before normalizing to first normal form. The unnormalized relations adhere to the definition of relations, which means:

  • Rows Order is insignificant
  • All rows are the same length (ie, row data matches column headers)
  • There are no duplicate rows

Codd adds the further constraint that all attributes should have distinct names. (Strictly speaking he only requires distinct names when two attributes have the same domain, but it is easier to just require distinct names for all attributes.)

But there are no constraint on the domains of the attributes at the unnormalized stage. If domains can contain anything, they can also contain relations, which means you could have nested relations arbitrary deep.

1NF eliminates nested relations. Codd explains several advantages to this, but perhaps most importantly it simplifies a data language significantly if it doesn't have to be able to handle nested relations. Notably Codd introduces 1NF before introducing relational algebra, which is the formal foundation for languages like SQL. And this in turn is the reason SQL does not support nested tables.

Much have been made of Codds use of the phrase "atomic". But in context it just refers to types which are not relations. For example a string is not a relation and is therefore atomic with regard to the relational operators like SELECT..FROM.

Since 1NF is defined in terms of domains, it does not put any constraint on individual values. Whether you store an address in one or five attributes is irrelevant, as long as each individual attribute only allows a single string.

The thing about "repeating columns" is probably also a misunderstanding, but I'm not even sure what it is supposed to mean. Having multiple columns with the same type is explicitly allowed in 1NF - this is why there is the constraint that attributes with same domain must have unique names.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.