A P P E N D I X C

Advanced Relational
Database Design

In this appendix we cover advanced topics in relational database design. We first

present the theory of multivalued dependencies, including a set of sound and com-
plete inference rules for multivalued dependencies. We then present PJNF and DKNF,
two normal forms based on classes of constraints that generalize multivalued depen-
dencies.

C.1 Multivalued Dependencies

As we did for functional dependencies and 3NF and BCNF, we shall need to deter-
mine all the multivalued dependencies that are logically implied by a given set of
multivalued dependencies.

C.1.1 Theory of Multivalued Dependencies

We take the same approach here that we did earlier for functional dependencies. Let
D denote a set of functional and multivalued dependencies. The closure D+ of D is
the set of all functional and multivalued dependencies logically implied by D. As
we did for functional dependencies, we can compute D+ from D, using the formal
definitions of functional dependencies and multivalued dependencies. However, it
is usually easier to reason about sets of dependencies by using a system of inference
rules.
The following list of inference rules for functional and multivalued dependencies
is sound and complete. Recall that sound rules do not generate any dependencies that
are not logically implied by D, and complete rules allow us to generate all dependen-
cies in D+ . The first three rules are Armstrong’s axioms, which we saw earlier in
Chapter 7.

1
2 Appendix C Advanced Relational Database Design

1. Reflexivity rule. If α is a set of attributes, and β ⊆ α, then α → β holds.

2. Augmentation rule. If α → β holds, and γ is a set of attributes, then γα → γβ
holds.
3. Transitivity rule. If α → β holds, and β → γ holds, then α → γ holds.
4. Complementation rule. If α →
→ β holds, then α →
→ R − β − α holds.
5. Multivalued augmentation rule. If α →
→ β holds, and γ ⊆ R and δ ⊆ γ, then
γα →
→ δβ holds.
6. Multivalued transitivity rule. If α →
→ β holds, and β →
→ γ holds, then α →
→
γ − β holds.
→ β.
7. Replication rule. If α → β holds, then α →
8. Coalescence rule. If α →→ β holds, and γ ⊆ β, and there is a δ such that δ ⊆
R, and δ ∩ β = ∅, and δ → γ, then α → γ holds.

The bibliographical notes provide references to proofs that the preceding rules are
sound and complete. The following examples provide insight into how the formal
proofs proceed.
Let R = (A, B, C, G, H, I) be a relation schema. Suppose that A → → BC holds. The
definition of multivalued dependencies implies that, if t1 [A] = t2 [A], then there exist
tuples t3 and t4 such that

t1 [A] = t2 [A] = t3 [A] = t4 [A]

t3 [BC] = t1 [BC]
t3 [GHI] = t2 [GHI]
t4 [GHI] = t1 [GHI]
t4 [BC] = t2 [BC]

The complementation rule states that, if A → → BC , then A → → GHI . Observe that t3

and t4 satisfy the definition of A →
→ GHI if we simply change the subscripts.
We can provide similar justification for rules 5 and 6 (see Exercise C.2) using the
definition of multivalued dependencies.
Rule 7, the replication rule, involves functional and multivalued dependencies.
Suppose that A → BC holds on R. If t1 [A] = t2 [A] and t1 [BC] = t2 [BC], then t1 and t2
themselves serve as the tuples t3 and t4 required by the definition of the multivalued
dependency A → → BC.
Rule 8, the coalescence rule, is the most difficult of the eight rules to verify (see
Exercise C.4).
We can simplify the computation of the closure of D by using the following rules,
which we can prove using rules 1 to 8 (see Exercise C.5):

• Multivalued union rule. If α →

→ β holds, and α →
→ γ holds, then α →
→ βγ
holds.
• Intersection rule. If α →
→ β holds, and α →
→ γ holds, then α →
→ β ∩ γ holds.
 C.1 Multivalued Dependencies 3

• Difference rule. If α →
→ β holds, and α →
→ γ holds, then α →
→ β − γ holds
and α →→ γ − β holds.

Let us apply our rules to the following example. Let R = (A, B, C, G, H, I) with the
following set of dependencies D given:

A→→B
B→→ HI
CG → H

We list several members of D+ here:

• A →
→ CGHI: Since A → → B, the complementation rule (rule 4) implies that
A→→ R − B − A. R − B − A = CGHI , so A →
→ CGHI.
• A→ → HI: Since A →→ B and B →→ HI, the multivalued transitivity rule (rule
6) implies thatA →
→ HI − B. Since HI − B = HI , A →
→ HI.
• B → H: To show this fact, we need to apply the coalescence rule (rule 8).
B→ → HI holds. Since H ⊆ HI and CG → H and CG ∩ HI = ∅, we satisfy the
statement of the coalescence rule, with α being B, β being HI, δ being CG, and
γ being H. We conclude that B → H.
• A→ → CG: We already know that A →
→CGHI and A → → HI. By the difference
rule, A →
→ CGHI − HI . Since CGHI − HI = CG, A →
→ CG.

C.1.2 Dependency Preservation

The question of dependency preservation when we have multivalued dependencies
is not as simple as it is when we have only functional dependencies.
A decomposition of schema R into schemas R1 , R2 , . . . , Rn is a dependency-
preserving decomposition with respect to a set D of functional and multivalued de-
pendencies if, for every set of relations r1 (R1 ), r2 (R2 ), . . . , rn (Rn ) such that for all i,
ri satisfies Di (the restriction of D to Ri ), there exists a relation r(R) that satisfies D
and for which ri = ΠRi (r) for all i.
Let us apply the 4NF decomposition algorithm of Figure 7.17 to the schema R =
(A, B, C, G, H, I) with D = {A → → B, B → → HI, CG → H}. We shall then test the
resulting decomposition for dependency preservation. We first need to compute the
closure of D. The nontrivial dependencies in closure include all the dependencies in
D, and the multivalued dependency A → → HI, as we saw in Section C.1.1.
R is not in 4NF. Observe that A → → B is not trivial, yet A is not a superkey. Using
A→ → B in the first iteration of the while loop, we replace R with two schemas, (A, B)
and (A, C, G, H, I). It is easy to see that (A, B) is in 4NF since all multivalued depen-
dencies that hold on (A, B) are trivial. However, the schema (A, C, G, H, I) is not in
4NF. Applying the multivalued dependency CG→ →H (which follows from the given
functional dependency CG → H by the replication rule), we replace (A, C, G, H, I)
by the two schemas (C, G, H) and (A, C, G, I).
4 Appendix C Advanced Relational Database Design

r1 : A B
a1 b1
a2 b1

r2 : C G H
c1 g1 h1
c2 g2 h2

r3 : A I
a1 i1
a2 i2

r4 : A C G
a1 c1 g1
a2 c2 g2

Figure C.1 Projection of relation r onto a 4NF decomposition of R.

A B C G H I
a1 b1 c1 g1 h1 i1
a2 b1 c2 g2 h2 i2

Figure C.2 A relation r(R) that does not satisfy B →

→ HI.

Schema (C, G, H) is in 4NF, but schema (A, C, G, I) is not. To see that (A, C, G, I)
is not in 4NF, we note that since A → → HI is in D+ , A →→ I is in the restriction of D
to (A, C, G, I). Thus, in a third iteration of the while loop, we replace (A, C, G, I) by
two schemas (A, I) and (A, C, G). The algorithm then terminates and the resulting
4NF decomposition is {(A, B), (C, G, H), (A, I), (A, C, G)}.
This 4NF decomposition is not dependency preserving, since it fails to preserve
the multivalued dependency B → → HI. Consider Figure C.1, which shows the four
relations that may result from the projection of a relation on (A, B, C, G, H, I) onto
the four schemas of our decomposition. The restriction of D to (A, B) is A → → B and
some trivial dependencies. It is easy to see that r1 satisfies A → → B, because there is
no pair of tuples with the same A value. Observe that r2 satisfies all functional and
multivalued dependencies, since no two tuples in r2 have the same value on any at-
tribute. A similar statement can be made for r3 and r4 . Therefore, the decomposed
version of our database satisfies all the dependencies in the restriction of D. How-
ever, there is no relation r on (A, B, C, G, H, I) that satisfies D and decomposes into
r1 , r2 , r3 , and r4 . Figure C.2 shows the relation r = r1 1 r2 1 r3 1 r4 . Rela-
tion r does not satisfy B → → HI. Any relation s containing r and satisfying B → → HI
must include the tuple (a2 , b1 , c2 , g2 , h1 , i1 ). However, ΠCGH (s) includes a tuple
(c2 , g2 , h1 ) that is not in r2 . Thus, our decomposition fails to detect a violation of
B→ → HI.
 C.2 Join Dependencies 5

We have seen that, if we are given a set of multivalued and functional dependen-
cies, it is advantageous to find a database design that meets the three criteria of

1. 4NF
2. Dependency preservation
3. Lossless join

If all we have are functional dependencies, then the first criterion is just BCNF.
We have seen also that it is not always possible to meet all three of these criteria.
We succeeded in finding such a decomposition for the bank example, but failed for
the example of schema R = (A, B, C, G, H, I).
When we cannot achieve our three goals, we have to compromise on one of 4NF
or dependency preservation.

C.2 Join Dependencies

We have seen that the lossless-join property is one of several properties of a good
database design. Indeed, this property is essential: Without it, information is lost.
When we restrict the set of legal relations to those satisfying a set of functional and
multivalued dependencies, we are able to use these dependencies to show that cer-
tain decompositions are lossless-join decompositions.
Because of the importance of the concept of lossless join, it is useful to be able
to constrain the set of legal relations over a schema R to those relations for which a
given decomposition is a lossless-join decomposition. In this section, we define such a
constraint, called a join dependency. Just as types of dependency led to other normal
forms, join dependencies will lead to a normal form called project-join normal form
(PJNF).

C.2.1 Definition of Join Dependencies

Let R be a relation schema and R1 , R2 , . . . , Rn be a decomposition of R. The join
dependency *(R1 , R2 , . . . , Rn ) is used to restrict the set of legal relations to those for
which R1 , R2 , . . . , Rn is a lossless-join decomposition of R. Formally, if R = R1 ∪
R2 ∪ . . . ∪ Rn , we say that a relation r(R) satisfies the join dependency *(R1 , R2 , . . . , Rn )
if
r = ΠR1 (r) 1 ΠR2 (r) 1 · · · 1 ΠRn (r)
A join dependency is trivial if one of the Ri is R itself.
Consider the join dependency *(R1 , R2 ) on schema R. This dependency requires
that, for all legal r(R),
r = ΠR1 (r) 1 ΠR2 (r)
Let r contain the two tuples t1 and t2 , defined as follows:

t1 [R1 − R2 ] = (a1 , a2 , . . . , ai ) t2 [R1 − R2 ] = (b1 , b2 , . . . , bi )

t1 [R1 ∩ R2 ] = (ai + 1 , . . . , aj ) t2 [R1 ∩ R2 ] = (ai + 1 , . . . , aj )
t1 [R2 − R1 ] = (aj + 1 , . . . , an ) t2 [R2 − R1 ] = (bj + 1 , . . . , bn )
6 Appendix C Advanced Relational Database Design

R1 – R2 R1 ∩ R2
ΠR1 (t1) a1 . . . ai ai + 1 . . . aj
ΠR1 (t2) b1 . . . bi ai + 1 . . . aj

R1 ∩ R2 R2 – R1
ΠR2 (t1) ai + 1 . . . aj aj + 1 . . . an
ΠR2 (t2) ai + 1 . . . aj bj + 1 . . . bn

Figure C.3 ΠR1 (r) and ΠR2 (r).

Thus, t1 [R1 ∩ R2 ] = t2 [R1 ∩ R2 ], but t1 and t2 have different values on all other
attributes. Let us compute ΠR1 (r) 1 ΠR2 (r). Figure C.3 shows ΠR1 (r) and ΠR2 (r).
When we compute the join, we get two tuples in addition to t1 and t2 , shown by t3
and t4 in Figure C.4.
If *(R1 , R2 ) holds, then, whenever we have tuples t1 and t2 , we must also have
t3 and t4 . Thus, Figure C.4 shows a tabular representation of the join dependency
*(R1, R2 ). Compare Figure C.4 with Figure 7.14, in which we gave a tabular repre-
sentation of α → → β. If we let α = R1 ∩ R2 and β = R1 , then we can see that
the two tabular representations in these figures are the same. Indeed, *(R1 , R2 ) is
just another way of stating R1 ∩ R2 → → R1 . Using the complementation and aug-
mentation rules for multivalued dependencies, we can show that R1 ∩ R2 → → R1
implies R1 ∩ R2 → → R2 . Thus, *(R1 , R2 ) is equivalent to R1 ∩ R2 → → R2 . This
observation is not surprising in light of the fact we noted earlier that R1 and R2 form
a lossless-join decomposition of R if and only if R1 ∩ R2 → → R2 or R1 ∩ R2 → → R1 .
Every join dependency of the form *(R1 , R2 ) is therefore equivalent to a multival-
ued dependency. However, there are join dependencies that are not equivalent to any
multivalued dependency. The simplest example of such a dependency is on schema
R = (A, B, C). The join dependency

*((A, B), (B, C), (A, C))

is not equivalent to any collection of multivalued dependencies. Figure C.5 shows

a tabular representation of this join dependency. To see that no set of multivalued
dependencies logically implies *((A, B), (B, C), (A, C)), we consider Figure C.5
as a relation r (A, B, C), as in Figure C.6. Relation r satisfies the join dependency

R1 – R2 R1 ∩ R2 R2 – R1
t1 a1 ... ai ai + 1 ... aj aj + 1 . . . an
t2 b1 ... bi ai + 1 ... aj bj + 1 . . . bn
t3 a1 ... ai ai + 1 ... aj bj + 1 . . . bn
t4 b1 ... bi ai + 1 ... aj aj + 1 . . . an

Figure C.4 Tabular representation of *(R1 , R2 ).

 C.2 Join Dependencies 7

A B C
a1 b1 c2
a2 b1 c1
a1 b2 c1
a1 b1 c1

Figure C.5 Tabular representation of *((A, B), (B, C), (A, C)).

*((A, B), (B, C), (A, C)), as we can verify by computing

ΠAB (r) 1 ΠBC (r) 1 ΠAC (r)
and by showing that the result is exactly r. However, r does not satisfy any nontrivial
multivalued dependency. To see that it does not, we verify that r fails to satisfy any
of A → → B, A →→ C, B → → A, B →
→ C, C → → A, or C →→ B.
Just as a multivalued dependency is a way of stating the independence of a pair
of relationships, a join dependency is a way of stating that the members of a set of
relationships are all independent. This notion of independence of relationships is a
natural consequence of the way that we generally define a relation. Consider

Loan info schema = (branch name, customer name, loan number, amount)

from our banking example. We can define a relation loan info (Loan info schema) as the
set of all tuples on Loan info schema such that

• The loan represented by loan number is made by the branch named branch
name.
• The loan represented by loan number is made to the customer named customer
name.
• The loan represented by loan number is in the amount given by amount.

The preceding definition of the loan info relation is a conjunction of three predicates:
one on loan number and branch name, one on loan number and customer name, and one
on loan number and amount. Surprisingly, it can be shown that the preceding intu-
itive definition of loan info logically implies the join dependency *((loan number, branch
name), (loan number, customer name), (loan number, amount)).

A B C
a1 b1 c2
a2 b1 c1
a1 b2 c1
a1 b1 c1

Figure C.6 Relation r (A, B, C).

8 Appendix C Advanced Relational Database Design

Thus, join dependencies have an intuitive appeal and correspond to one of our
three criteria for a good database design.
For functional and multivalued dependencies, we were able to give a system of
inference rules that are sound and complete. Unfortunately, no such set of rules is
known for join dependencies. It appears that we must consider more general classes
of dependencies than join dependencies to construct a sound and complete set of
inference rules. The bibliographical notes contain references to research in this area.

C.2.2 Project-Join Normal Form

Project-join normal form (PJNF) is defined in the same way as BCNF and 4NF, except
that join dependencies are used. A relation schema R is in PJNF with respect to a set
D of functional, multivalued, and join dependencies if, for all join dependencies in
D+ of the form *(R1 , R2 , . . . , Rn ), where each Ri ⊆ R and R = R1 ∪ R2 ∪ . . . ∪ Rn ,
at least one of the following holds:

• *(R1 , R2 , . . . , Rn ) is a trivial join dependency.

• Every Ri is a superkey for R.

A database design is in PJNF if each member of the set of relation schemas that con-
stitutes the design is in PJNF. PJNF is called fifth normal form (5NF) in some of the
literature on database normalization.
Consider again our banking example. Given the join dependency *((loan number,
branch name), (loan number, customer name), (loan number, amount)), Loan info schema is
not in PJNF. To put Loan info schema into PJNF, we must decompose it into the three
schemas specified by the join dependency: (loan number, branch name), (loan number,
customer name), and (loan number, amount).
Because every multivalued dependency is also a join dependency, it is easy to see
that every PJNF schema is also in 4NF. Thus, in general, we may not be able to find a
dependency-preserving decomposition into PJNF for a given schema.

C.3 Domain-Key Normal Form

The approach we have taken to normalization is to define a form of constraint (func-
tional, multivalued, or join dependency), and then to use that form of constraint to
define a normal form. Domain-key normal form (DKNF) is based on three notions.

1. Domain declaration. Let A be an attribute, and let dom be a set of values. The
domain declaration A ⊆ dom requires that the A value of all tuples be values
in dom.
2. Key declaration. Let R be a relation schema with K ⊆ R. The key declaration
key (K) requires that K be a superkey for schema R—that is, K → R. Note
that all key declarations are functional dependencies but not all functional
dependencies are key declarations.
 C.3 Domain-Key Normal Form 9

3. General constraint. A general constraint is a predicate on the set of all relations

on a given schema. The dependencies that we have studied in this chapter are
examples of general constraints. In general, a general constraint is a predicate
expressed in some agreed-on form, such as first-order logic.

We now give an example of a general constraint that is not a functional, multi-

valued, or join dependency. Suppose that all accounts whose account number begins
with the digit 9 are special high-interest accounts with a minimum balance of $2500.
Then, we include as a general constraint, “If the first digit of t[account number] is 9,
then t[balance] ≥ 2500.”
Domain declarations and key declarations are easy to test in a practical database
system. General constraints, however, may be extremely costly (in time and space)
to test. The purpose of a DKNF database design is to allow us to test the general
constraints using only domain and key constraints.
Formally, let D be a set of domain constraints and let K be a set of key constraints
for a relation schema R. Let G denote the general constraints for R. Schema R is in
DKNF if D ∪ K logically imply G.
Let us return to the general constraint that we gave on accounts. The constraint
implies that our database design is not in DKNF. To create a DKNF design, we need
two schemas in place of Account schema:

Regular acct schema = (account number, branch name, balance)

Special acct schema = (account number, branch name, balance)

We retain all the dependencies that we had on Account schema as general constraints.
The domain constraints for Special acct schema require that, for each account,

• The account number begins with 9.

• The balance is greater than 2500.

The domain constraints for Regular acct schema require that the account number does
not begin with 9. The resulting design is in DKNF, although the proof of this fact is
beyond the scope of this text.
Let us compare DKNF to the other normal forms that we have studied. Under the
other normal forms, we did not take into consideration domain constraints. We as-
sumed (implicitly) that the domain of each attribute was some infinite domain, such
as the set of all integers or the set of all character strings. We allowed key constraints
(indeed, we allowed functional dependencies). For each normal form, we allowed
a restricted form of general constraint (a set of functional, multivalued, or join de-
pendencies). Thus, we can rewrite the definitions of PJNF, 4NF, BCNF, and 3NF in a
manner that shows them to be special cases of DKNF.
We now present a DKNF-inspired rephrasing of our definition of PJNF. Let R =
(A1 , A2 , . . . , An ) be a relation schema. Let dom(Ai ) denote the domain of attribute
Ai , and let all these domains be infinite. Then all domain constraints D are of the form
Ai ⊆ dom(Ai ). Let the general constraints be a set G of functional, multivalued, or
join dependencies. If F is the set of functional dependencies in G, let the set K of key
10 Appendix C Advanced Relational Database Design

constraints be those nontrivial functional dependencies in F + of the form α → R.

Schema R is in PJNF if and only if it is in DKNF with respect to D, K, and G.
A consequence of DKNF is that all insertion and deletion anomalies are eliminated.
DKNF represents an “ultimate” normal form because it allows arbitrary constraints,
rather than dependencies, yet it allows efficient testing of these constraints. Of course,
if a schema is not in DKNF, we may be able to achieve DKNF via decomposition, but
such decompositions, as we have seen, are not always dependency-preserving de-
compositions. Thus, although DKNF is a goal of a database designer, it may have to
be sacrificed in a practical design.

C.4 Summary
In this chapter we presented the theory of multivalued dependencies, including a set
of sound and complete inference rules for multivalued dependencies.
We then presented two more normal forms based on more general classes of con-
straints. Join dependencies are a generalization of multivalued dependencies, and
lead to the definition of PJNF. DKNF is an idealized normal form that may be difficult
to achieve in practice. Yet DKNF has desirable properties that should be included to
the extent possible in a good database design.

Exercises
C.1 List all the nontrivial multivalued dependencies satisfied by the relation in Fig-
ure C.7.
C.2 Use the definition of multivalued dependency (Section 7.6.1) to argue that each
of the following axioms is sound:
a. The complementation rule
b. The multivalued augmentation rule
c. The multivalued transitivity rule
C.3 Use the definitions of functional and multivalued dependencies (Sections 7.4
and 7.6.1) to show the soundness of the replication rule.
C.4 Show that the coalescence rule is sound. (Hint: Apply the definition of α → → β
to a pair of tuples t1 and t2 such that t1 [α] = t2 [α]. Observe that since δ ∩ β =
∅, if two tuples have the same value on R − β, then they have the same value
on δ.)

A B C
a1 b1 c1
a1 b1 c2
a2 b1 c1
a2 b1 c3

Figure C.7 Relation of Exercise C.1.

 Bibliographical Notes 11

C.5 Use the axioms for functional and multivalued dependencies to show that each
of the following rules is sound:
a. The multivalued union rule
b. The intersection rule
c. The difference rule
C.6 Let R = (A, B, C, D, E), and let M be the following set of multivalued
dependencies

A →
→ BC
B →→ CD
E →→ AD

List the nontrivial dependencies in M + .

C.7 Give a lossless-join decomposition of schema R in Exercise C.6 into 4NF.
C.8 Give an example of relation schema R and a set of dependencies such that R is
in 4NF, but is not in PJNF.
C.9 Explain why PJNF is a normal form more desirable than is 4NF.
C.10 Rewrite the definitions of 4NF and BCNF using the notions of domain con-
straints and general constraints.
C.11 Explain why DKNF is a highly desirable normal form, yet is one that is difficult
to achieve in practice.

Bibliographical Notes
The notions of 4NF, PJNF, and DKNF are from Fagin [1977], Fagin [1979], and Fagin
[1981], respectively. The synthesis approach to database design is discussed in Bern-
stein [1976].
Join dependencies were introduced by Rissanen [1979]. Sciore [1982] gives a set
of axioms for a class of dependencies that properly includes the join dependencies.
In addition to their use in PJNF, join dependencies are central to the definition of
universal relation databases. Fagin et al. [1982] introduces the relationship between
join dependencies and the definition of a relation as a conjunction of predicates (see
Section C.2.1). This use of join dependencies has led to a large amount of research into
acyclic database schemas. Intuitively, a schema is acyclic if every pair of attributes is
related in a unique way. Formal treatment of acyclic schemas appears in Fagin [1983]
and in Beeri et al. [1983].
Additional dependencies are discussed in detail in Maier [1983]. Inclusion depen-
dencies are discussed by Casanova et al. [1984] and Cosmadakis et al. [1990]. Tem-
plate dependencies are covered by Sadri and Ullman [1982]. Mutual dependencies
are examined by Furtado [1978] and by Mendelzon and Maier [1979].