CS 4221: Database Design

The Relational Model

Ling Tok Wang

National University of Singapore

1
CS4221: The Relational Model
 Topics:
 Basic concepts in Relational Model
o FD, transitive dependency, key, primary key, updating anomalies, properties of FDs
 Normal Forms
o 1NF, 2NF, 3NF, BCNF; redundancy in NF relations
 Decomposition Approach
o Universal Relation Assumption, problems of decomposition approach
 Sythesizing Approach
o FD inference rules, closure of FDs, closure of attributes, FD membership test, criteria for
normalization, local/global redundancy, Bernstein’s Algorithm and its weak points
 4NF
o MVDs, MVD inference rules, properties of FDs and MVDs, decomposition approach,
MVDs and hierarchical model
 5NF and DKNF
o Will not be covered/examined due to time limit


Will show many commonly misunderstood important concepts
and errors. 2
CS4221: The Relational Model
 First Normal Form (1NF) Relation
Defn: Given sets of atomic (i.e. non-decomposable) elements D1,
D2, …, Dn (not necessarily distinct), R is a first normal
form (1NF) relation on these n sets if it is a set of ordered
n-tuples < d1, d2, …, dn > such that
di  Di  i = 1, 2, ..., n. (Note:  means “for all”)
Thus R  D1 x D2 x … x Dn
where x is the Cartesian product operator.

Note: A set has no duplicates. An n-tuple is ordered means the orders of
the n components of the tuple are important.

Defn: D1, …, Dn are called the domains of R. Each domain may be

assigned a unique role name, called an attribute of R.
Defn: For any tuple in R, the value of an attribute named B is
referred to as a B-value.
For a set of attributes X = {B1, …, Bm}, the values of the
attributes in X of any tuple in R is referred to as an X-value. 3
CS4221: The Relational Model
E.g. A relation Take which contains information on courses taken
by students. Take is a 1NF relation.

There are different ways to express the relation Take:

(1) Take  char(10) x char(6) x char(30) x char(60) x int (domains)

(2) Take  Student# x Course# x S-name x C-desc x Mark (attributes)


(3) Take (Student#, Course#, S-name, C-desc, Mark) (attributes)

Take Student# Course# S-name C-desc Mark

95001 CS1101 Tan CK Programming 75

95023 CS1101 Lee SL Programming 58

94257 CS2103 Tan CK Data Stru 64

…
4
CS4221: The Relational Model
Defn: A set of attributes Y of R is said to be functionally
dependent (FD) on a set of attributes X of R if each X-value
in R has associated with exactly one Y-value in R at any
time.
This is denoted by
XY
and is called a functional dependency of R.

Q: Why “at any time”?



Defn: A functional dependency X  Y is said to be trivial if Y  X.

Q: Why call it “trivial”?



Defn: A functional dependency X  Y of R is said to be a

full dependency of R (or Y is fully dependent on X)
if it is a non-trivial FD and there exists no proper
subset X of X such that X  Y.
5
CS4221: The Relational Model
 Defn: A set of attributes K of a relation R is said to be a
candidate key (or simply key) of R if all attributes of R
are functionally dependent on K and there exists no
proper subset K of K such that all attributes of R are
functionally dependent on K.

Defn: If there are more than one key for a relation, one of the
keys is designated as the primary key of the relation.

 Q: How do we choose the primary key of a relation?

What are the selection criteria?

Defn: An attribute of R is called a prime attribute (or

prime) if it is contained in some key of R. All other
attributes of R are called non-prime attributes of R.
6
CS4221: The Relational Model
Example 1.
Let Take be a relation with the set of attributes:
{STUDENT#, COURSE#, S-NAME, C-DESCRIPTION, MARK}

We have the following functional dependencies in Take:

STUDENT#  S-NAME
COURSE#  C-DESCRIPTION
STUDENT#, COURSE#  MARK

Q: How can we find/know these FDs?
Can we use some data mining techniques to find FDs in a RDB?
Why each student only has one name?

{STUDENT#, COURSE#} is the only key of the relation.

STUDENT# and COURSE# are primes, the rest are non-primes.

Q: Do the below FDs also hold in the relation Take?

STUDENT#, COURSE#  S-NAME
STUDENT#, COURSE#  C-DESCRIPTION
STUDENT#, S-NAME, COURSE#  MARK
7
CS4221: The Relational Model
 • Insertion anomaly – if a new course is created but no students have taken this
course, then we cannot enter the information about this course because the use
of null values or undefined values in the primary key could cause problem.
• Deletion Anomaly - similiar
• Rewriting anomaly - similiar
These three anomalies are called the updating anomalies.

Q: What causes these updating anomalies?

• One process which attempts to remove these undesirable updating anomalies

from the relation is called normalization.

• The relation Take can be decomposed into (Q: How?)

R1 (STUDENT#, S-NAME)
R2 (COURSE#, C-DESCRIPTION)
R3 (STUDENT#, COURSE#, MARK)

Notation: A contiguous underline indicates a key of the relation.
E.g. In R3, attributes STUDENT# and COURSE# form a key of the relation R3.
The above 3 relations do not have updating anomalies. Prove it!
8
CS4221: The Relational Model
 Second Normal Form (2NF) Relation
Defn: A first normal form relation is called a second normal
form (2NF) relation if and only if every non-prime
attribute of R is fully dependent on each key of R.
Note that the relation Take in Example 1 is not in 2NF.
Take (STUDENT#, COURSE#, S-NAME, C-DESCRIPTION, MARK)
For example, S-Name is a non-prime and it is not fully dependent
on the key {STUDENT#, COURSE#}. Q: Why?
The name of a student is duplicated if the student takes more than one course.
Example 2. SP (S#, Sname, P#, Pname, Price)
A supplier with supplier number (S#) and name (Sname) supplies a part with part number
(P#) and name (Pname) with a price (Price). FDs in relation SP are:
S#  Sname (A supplier only has one name)
P#  Pname (A part only has one name)
S#, P#  Price (A supplier supplies a part with one price at any one time)

{S#, P#} is the only key of the relation SP. Prove it!

Relation SP is not in 2NF as Sname is not fully dependent on the key. Q: Why?
There are redundant information on Sname and Pname in SP. 9
CS4221: The Relational Model
 Third Normal Form (3NF) Relation
Defn: Let A and B be two distinct sets of attributes (i.e. not
identical) of a relation R, and d be an attribute of R
which does not belong to A or B such that
 d
A B B d B / A
A B
/
Then we say that d is transitively dependent on A under R,
and A d is a transitive dependency.
Intuitive meaning: A transitive dependency can be derived from other FDs, so it

is redundant and can be removed.

Notation: B / A means A is not functionally dependent on B.


Q: What if we have B A instead?
Defn: A relation is in Codd third normal form (3NF) if and only
if it is in 2NF and each non-prime attribute of R is not
transitively dependent on each key of R. 10
CS4221: The Relational Model
 Note: All the three relations:
R1 (STUDENT#, S-NAME)
R2 (COURSE#, C-DESCRIPTION)
R3 (STUDENT#, COURSE#, MARK)
in Example 1 are in 3NF. Prove it!
Example 3. R (Prof, Dept, Faculty)
We have the below FDs: (Q: How to find them?)
Prof  Dept, Faculty
Dept  Faculty
 Q: Why Prof  Dept ? Is it true in any university?
Faculty Note that R is in 2NF but not in 3NF because
Prof  Faculty
Prof Dept is a transitive dependency.
/
We decompose this relation into
R1 (Prof, Dept)
R2 (Dept, Faculty)
They are both in 3NF. 11
CS4221: The Relational Model
 Boyce-Codd normal form (BCNF) Relation
Defn: A relation R is in Boyce-Codd normal form (BCNF) if
and only if it is in 1NF and for every attribute set A of R,
if any attribute of R not in A is functionally dependent on
A, then all attributes in R are functionally dependent on A.

Q: Are the below 3 relations in BCNF?

R1 (STUDENT#, S-NAME)
R2 (COURSE#, C-DESCRIPTION)
R3 (STUDENT#, COURSE#, MARK)

Q: Are the below 2 relations in BCNF?

R1 (Prof, Dept)
R2 (Dept, Faculty)

 Q: Are there updating anomalies in a BCNF relation?

The answer is still yes but in fewer cases. Q: Why?
12
CS4221: The Relational Model
 Example 4.
Consider the relation STJ with the below FDs:
STJ (STUDENT, TEACHER, SUBJECT)
Assume that we have the below constraints:
1. For each subject, each student of that subject is
taught by only one teacher.
STUDENT, SUBJECT  TEACHER

2. Each teacher teaches only one subject.

TEACHER  SUBJECT

3. Some subjects are taught by more than one teacher

SUBJECT / TEACHER

Q: What are the keys of the relation SPJ? Primes ?

Q: Is it in 3NF?
Q: Is it in BCNF?
 Q: If a relation is not in BCNF, can we always normalize it to a set of
BCNF relations? Ans: Not always. 13
CS4221: The Relational Model
Example 5. R (A, B, C, D, F)
with AB  CDF, A  C, DF

C  R is not in 2NF since C is not fully dependent

on the key AB.
AB A
/
Decompose it, we get:

R1 (A, C) and R 2 (A, B, D, F)

 R2 is not in 3NF since AB  F is a

F transitive dependency. Decompose it, we get
R1 (A, C), R21 (A, B, D), R22 (D, F)
AB D
/
 All are in 3NF. Q: Are they also in BCNF?

14
CS4221: The Relational Model
E.g. R (A, B, C, D) with AB  CD and D  B

R is in 3NF but not in BCNF since D  B but D / C

Q: What are the keys ? Hint: There are 2 keys.

E.g. Enrol (S#, C#, Sname, Mark)

where S#, C#  Sname is a transitive dependency

and the relation Enrol is not in 3NF.

In fact, it is not in 2NF also. Q: Why?

15
CS4221: The Relational Model
 Decomposition & Synthesizing Method
- for Relational Database Design
 Three common methods for relational database schema design are the
decomposition method, the synthesizing method, and the Entity-Relationship
Approach.
 The decomposition method is based on the assumption that a database can
be represented by a universal relation which contains all the attributes of the
database (this is called the universal relation assumption) and this relation
is then decomposed into smaller relations in order to remove redundant data.
 The synthesizing method is based on the assumption that a database can be
described by a given set of attributes and a given set of functional
dependencies, and 3NF or BCNF relations are then synthesized based on the
given set of dependencies.
Note: Synthesizing method assumes universal relation assumption also.

 We will discuss the Entity-Relationship Approach later.

 Examples 3 & 5 use the decomposition method.

16
CS4221: The Relational Model
Properties of Universal Relation Assumption
• Decomposition method and synthesizing method do not
change any attribute name and do not delete any attribute or
add new attributes to the database.
• Two attributes with the same name from 2 relations are
referred to some same attribute in the universal relation, i.e.
they are from the same attribute and of the same semantics
(same meaning).
• Two attributes with different names from 2 different
relations or from a relation are referred to two different
attributes in the universal relation, and they have different
semantics.
Example: A database SP has the below 3 relations:
Supplier (Code, Sname),
Part (Code, Pname, Color)
Supply (Supplier, Part, Price)

This database SP does not satisfy the universal relation assumption.

Q: Why? Bad design on attribute names.
17
CS4221: The Relational Model
Some properties of normal form relations:
1. BCNF  3NF  2NF  1NF (prove them!)

2. A set of 3NF relations always exists for a given set of functional
dependencies, but it is not true for Boyce-Codd norm form relation set.
E.g. The relation R (s, j, t)
with functional dependencies
s j  t, tj
is in 3NF but has no BCNF relation set which
covers the given functional dependencies

3. Even BCNF relations can suffer from the updating anomalies
E.g. Let R = { R1(a,b,c,g,h), R2(a,b,e), R3(b,c,f), R4(e,f,g) }
with the set of full dependencies:
G = { abc  g, abc  h, ab  e, bc  f, ef  g }
 Note: All the relations in R are in BCNF.
However, there are two different ways to find the g-value of
any given {a,b,c}-value via different relations. So, there are
redundancies and R has updating anomalies. In fact, g in R1 is
superfluous and can be removed.
18
CS4221: The Relational Model
 Properties of FDs (inference rules)
Defn: Given a relation R having a set of attributes A and a given set of
functional dependencies F, the closure of F, denoted by F+, is
defined as follows:
(1) F  F+
(2) Projectivity:  X, Y  A Note:  means “for all”
If Y  X then X  Y  F+
(3) Transitivity:  X, Y, Z  A
If X  Y, Y  Z  F+
Then X  Z  F+
(4) Union (or Additivity):  X, Y, Z  A
If X  Y, X  Z  F+
Then X  Y  Z  F+
(5) No other functional dependencies are in F+.
Result: F+ is sound and complete. Q: What are their meanings?
Q: What is the meaning of “closure”? 19
CS4221: The Relational Model
Another definition for the closure of F (Armstrong’s Axioms):

(1) F  F+
(2) Reflexivity: X  X  F+  X  A
(3) Augmentation:  X, Y, Z  A
If X  Z  F+ then X  Y  Z  F+
(4) Pseudo-transitivity:  X, Y, Z, W  A
if X  Y  F+ , Y  Z  W  F+
then X  Z  W  F+
(5) No other FDs are in F+

Result: The above 2 definitions for the closure of F are equivalent.

Note: We usually simply write X  Y as “X, Y” or {X, Y}.

20
CS4221: The Relational Model
 Defn: Two sets of attributes A and B of a relation are
said to be functionally equivalent if and only if
A  B  F+ and B  A  F+
A and B are said to be properly functionally
equivalent if and only if A and B are functionally
equivalent and A1  A and B1  B such that
A1  B  F or B1  A  F
+ +

Note: Ǝ means there exists, and Ǝ

/ means there does not exist
Result: A relation R is in 3NF if and only if each non-
prime attribute is not transitivity dependent on
an arbitrarily chosen key of R. (Prove it!)

 Q: What is the use of this result?

21
CS4221: The Relational Model
 E.g. Let A = {A, B, C}, F = {A  B, B  C}
F+ = { A  A, B  B, C  C,
AB  A, AB  B, AB  AB,
BC  B, BC  C, BC  BC,
AC  A, AC  C, AC  AC,
ABC  A, ABC  B, ABC  C,
ABC  AB, ABC  AC, ABC  BC,
ABC  ABC, /* all the above FDs are trivial
A  B, B  C, A C, A  BC,
A  AC, A  AB, A  ABC, B  BC,
/* all the below FDs are non full dependencies
AC  B, AC  BC, AC  AB, AC  ABC,
AB  C, AB  BC, AB  AC, AB  ABC }
Q: Do we need find the closure of a set of FDs during normalization?
Note: There are too many FDs in the closure. We don’t really need to find the closure.
However it is important test whether a FD is in a closure or not.
Q: What is the intuitive meaning of “a FD is in the closure of a set of FDs”?

FD Membership Problem:
Given a set of FDs F defined on A, X  A and y  A, is X  y  F+ ?
i.e. can X  y be derived from F ?
22
CS4221: The Relational Model
Example:
1 2 3 4
Let G = { AB  C, C  D, DE  F, A  E}

Show AB  F  G+.
Note: The numbers are used to identify the FDs.

Solution:
1 2
AB  ABC  ABCD
4
ABCDE 
3
ABCDEF
F
 AB  F  G+

Q: How to prove each step using the FD inference rules?

23
CS4221: The Relational Model
 Detailed steps for proving AB  F  G+
1
(1) Prove AB  ABC
Since AB  AB (by projectivity)
AB  C (given)
so AB  ABC (by additivity)
(2) Prove ABC  2
ABCD
Since C  D (given)
ABC  C (by projectivity)
so ABC  D (by transitivity)
Also ABC  ABC (by projectivity)
so ABC  ABCD (by additivity)
(3) Prove AB  1,2
ABCD
From (1) we have AB  ABC
From (2) we have ABC  ABCD
so AB  ABCD (by transitivity)
(4) …
Note: The proof is too long. Any better way?
24
CS4221: The Relational Model
Alternative Solution to prove X  Y in G+
Given a set of attributes X, the closure of X relative to

Defn:
G is defined as:
X+ = { y  A | X  y  G+ }

Q: What is the intuitive meaning of the closure of X?

Q: How to construct X+ relative to a given set of FDs G?

Alternative Solution: To test X  Y in G+, we can just test
whether Y is in X+, the closure of X relative to G.
1 2 3 4
E.g. Let G = { AB  C, C  D, DE  F, A  E}

25
CS4221: The Relational Model
Three Criteria for Normalization
(1) Reconstructibility (or losslessness).
If an original relation R is split into n relations R1, R2, …, Rn,
then Ri = R[Ai] (where [ ] is the projection operator)

and R1 R2 … Rn = R
where Ai is the attribute set of Ri  i = 1, 2, …, n
and is the join operator

Note: The join operator is also denoted by *.

Defn: Two sets of FDs, F and G are equivalent if and only if

F+ = G+.
If F and G are equivalent, we say F covers G,
G covers F, F is a cover of G, or G is a cover of F.
26
CS4221: The Relational Model

(2) Covering.
F+ = (F1  F2 …  Fn)+
where F is the set FDs for the original relation R and Fi is
the set of FDs in relation Ri  i = 1, 2, …, n.
(3) Each relation is free of redundant attributes (i.e. no local
redundancy – no redundancy within each relation).
Q: Is it ture that (F  G)+ = F+  G+ for any two sets of FDs F and G?


Note: In fact, free of local redundant attributes is not enough, global
redundancy (i.e. redundancy among relations) may still exist. (see LTK
normal form)
Ref: Tok Wang Ling, Frank W Tompa, Tiko Kameda, An Improved Third Normal Form for
Relational Databases, ACM TODS, vol 6, no 2, pp329-346, 1981.

Example Given R (A, B, C) with CA

R is in 3NF, but not in BCNF
If we decompose R into 2 relations
R1 (C, A) and R2 (C, B)
then we lose the FD AB  C.
This violates the covering criteria. Why? 27
CS4221: The Relational Model
Synthesizing Third Normal Form Relations
(by Philip A. Bernstein, TODS 1979)

Algorithm
1. (Eliminate extraneous attributes). Let F be the given set of FDs
where the right side of each FD is a single attribute.
Eliminate extraneous attributes from the left side of each FD in F,
producing the set G.
2. (Finding covering). Find a non-redundant covering H of G.
3. (Partition). Partition H into groups such that all of the FDs in each
group have identical left sides.
4. (Merge equivalent keys). Let J = .
For each pair of groups, say Hi and Hj with left sides X and Y resp.
If X and Y are properly equivalent, then
(a) merge Hi and Hj together
(b) add X  Y and Y  X to J
(c) if X  Z  H and Z  Y, then delete X  Z from H.
Similarly, if Y  Z  H and Z  X, then delete Y  Z from H.
28
CS4221: The Relational Model

5. (Eliminate transitive dependencies).
Find a minimal H  H such that
(H  J)+ = (H  J)+
Then add each FD of J into its corresponding group of H.

6. (Construct relations)
Each group in H forms a relation.
Each set of attributes that appears on the left side of any FD in the
group is a key of the relation formed by the group. They are called
explicit keys.
Note: There may have more than one key for some relations constructed.

Result: The relations produced by step 6 are all in 3NF.

Result: The number of relations produced is minimum.


Q: What is the difference between “minimal” and “minimum”?
29
CS4221: The Relational Model
Example 1 Given F = { A  B, A  C, B  C,
B  D, D B, A B E  F }
Step 1. (Eliminating extraneous attributes)
G = { A  B, A  C, B  C,
B  D, D B, A E  F }
(since A E  A B E  F+)
Step 2. (Find covering)
H = { A  B, B  C,
B  D, D  B, A E  F }
(since A  C  (G – {A  C})+)

Step 3 (Partition)
H1 = { A  B }
H2 = { B  C, B  D }
H3 = { D  B }
H4 = { A E  F }
30
CS4221: The Relational Model
Step 4 (Merge groups)
B and D are properly equivalent
J = { B  D, D  B }
H1 = {A  B}
H2 = H2  H3 – {B  D, D  B}
= {B  C}
H4 = {AE  F}

Step 5 (Eliminate transitive dependencies)

None! (You should verify this).

Step 6 (Construct relations)

R1 (A, B)
R2 (B, D, C)
R3 (A, E, F)

31
CS4221: The Relational Model
Example 2 (need step 5)
Given F = {X1 X2  AD, CD  X1 X2 ,
A X1  B, B X2  C, C  A}
Step 1. G=F
Step 2. H=G
Step 3 H1 ={ X1 X2  AD}
H2 = {CD  X1 X2}
H3 = {A X1  B}
H4 = {B X2  C}
H5 = {C  A}
Step 4 J = {X1 X2  CD, CD  X1 X2}
H1 = H1  H2 – J
= {X1 X2  A}
H3 = {A X1  B}
H4 = {B X2  C}
H5 = {C  A}
32
CS4221: The Relational Model

Step 5 (Eliminate TD)
We can eliminate X 1X 2  A A
since X 1 X 2  CD, C  A
and C / X 1 X2
X1 X2 C
so we get
J = {X1 X2  CD, CD  X1 X2}
H1 = 
H3 = {A X1  B}
H4 = {B X2  C}
H5 = {C  A}
Step 6 R1 (X1, X2, C,D) Note: 2 keys: {X1, X2} and {C, D}
R2 (A, X1, B)
R3 (B, X2, C)
R4 (C, A)

Note: If we omit step 5, then R1 will be
R1 (X1, X2, C,D, A)
Which is not in 3NF. Why? 33
CS4221: The Relational Model
 Some shortcomings of Bernstein’s algorithm

Shortcoming 1. Bernstein’s algorithm does not guarantee
reconstructibility (or losslessness).
Example 3. Given R (Course#, Preq#, Cname, Cdesc) with
F = {Course#, Preq#  Cname
Course#  Cname, Cdesc}
Step 1 G = {Course#  Cname, Cdesc}
Step 2 H=G
:
Step 6 R1 (Course#, Cname, Cdesc)
Note: We lose information about Preq#.

Q: How to resolve this problem?
In fact we have Course# Preq#
(Note. It is a multi-valued dependency, to be discussed later. Bernstein’s
algorithm does not handle MVDs).
We need another relation:
R2 (Course#, Preq#) 34
CS4221: The Relational Model

Shortcoming 2. Bernstein’s algorithm does not find all the keys.

Example 4. Given R (A, B, C, D)

with F = { AB  CD, C  B }
Apply the algorithm, we will get
R1 (A, B, C, D)
R2 (C, B)
 In fact, {A, C} is also a key of R1.
This is called an implicit key.

Note: R1 is not in BCNF.

Note: To find all the keys of a relation is NP-complete.
Q: What is the meaning of NP-complete? A term from complexity theory.

35
CS4221: The Relational Model

Shortcoming 3. Bernstein’s algorithm does not remove all the
superfluous attributes (i.e. redundant attributes).
Example 5. Given F = { AD  B, B  C, C  D,AB  E,
AC  F }
Step 1 G=F
Step 2 H=G=F
:
Step 6 R 1 (A, B, C, D, E, F)

R2 (B, C)
R3 (C, D)

Note: C is superfluous in R1, but R1 is in 3NF. However, D
is not superfluous. Remove C from R1 and get
R1 (A, B, D, E, F)

Note: Ling & Tompa & Kameda method removes all superfluous attributes.

36
CS4221: The Relational Model

Shortcoming 4. The set of relations produced by the algorithm depends
on the non-redundant covering found.
Example 6. Given F = {AD  B, B  C, C  D, AB  E,
AC F, AD  F, AC  E}
Case 1 If H = {AD  B, B  C, C  D,AB  E, AC  F}
Then the set of relation is
R 1 (A , B , C, D , E, F)

R2 (B, C)
R3 (C, D)
Case 2 If H = {AD  B, B  C, C  D, AB  E, AD  F }
Then the set of relations is
R1 (A, B, D, E, F)

R2 (B, C)
R3 (C, D)
37
CS4221: The Relational Model
Case 3 If H = {AD  B, B  C, C  D,AC  F, AC  E}
Then we have
R  1 (A , C , D , B , E , F )

R2 (B, C)
R3 (C, D

Note that AB is a key but it is not found by the algorithm.

Case 4 If H = {AD  B, B  C, C  D, AC  E, AD  F }
Then we have
R 1 (A, C, D, B, E, F)
R2 (B, C)
R3 (C, D)
 Note that AB is a key but it is not found by the algorithm.

Note that Case 2 gives the best solution. What is the meaning?
38
CS4221: The Relational Model

Shortcoming 5. A BCNF relation set may contain superfluous attributes,
i.e. redundant attributes which can be removed.

Example: Given a set of relations

R1 (Model#, Serial#, Price, Color)
R2 (Model#, Name)
R3 (Serial#, Year)
R4 (Name, Year, Price)

Note: All relations are in BCNF, but R1 contains a superfluous attribute
Price, i.e. Price can be removed from R1 without losing any information.
How to prove it?

Note: 3NF and BCNF are defined for individual relations but not
the whole relational schema.
Ref: Ling, Tompa, & Kameda method takes the whole relational
schema into consideration and removes superfluous attributes.

39
CS4221: The Relational Model
 Note. Some relations generated by Step 6 may have more than
one key. We need to choose their preliminary key.
Why and how to choose?

Q: Any impact on other relations after choosing primary key for


some relation which has more than one key?

E.g. A database schema generated by Bernstein’s Algorithm has

the below relations:
Student (NRIC, S#, Name, DOB)
Course (C#, Title, Desc)
Take (NRIC, C#, Grade)
Note that Student relation has two keys, i.e. NRIC and S#. We
choose S# as its preliminary key, and we also need to change
NRIC in Take relation to S# and the relation Take becomes
Take (S#, C#, Grade)
Q: Why? 40
CS4221: The Relational Model
 Fourth Normal Form (4NF) Relation
E.g. The meaning of a given record in the below unnormalized relation
(shown on the LHS) is:

the indicated courses are taught by all of the indicated

teachers, and uses all the indicated text books.
Its normalized relation CTX is shown on the RHS.

Unnormalized relation (a nested relation) CTX - normalized relation

Course Teacher Text Course Teacher Text
Physics { Dr. Lee, {Basic Mechanics, Physics Dr. Lee Basic Mechanics
Dr. Chan} Applied Physics} Physics Dr. Lee Applied Physics
Math {Dr. Black} {Modern Algebra, Physics Dr. Chan Basic Mechanics
Geometry}
Physics Dr. Chan Applied Physics
Math Dr. Black Modern Physics
Math Dr. Black Geometry
41
CS4221: The Relational Model
Notes:
1. CTX has the following property:
if (c, t1, x1)  CTX and (c, t2, x2)  CTX
then (c, t1, x2)  CTX and (c, t2, x1)  CTX
2. A lot of redundant data in CTX.
3. CTX is in BCNF.

Defn: Given a relation R with attributes A, B, and C, the

multivalued dependency (MVD)
R.A R.B or simply A B
holds in R if and only if the set of B-values matching a given
(A-value, C-value) pair in R depends only on A-value,
i.e. if (a, b1, c1)  R, (a, b2, c2)  R
then (a, b1, c2)  R, (a, b2, c1)  R

42
CS4221: The Relational Model
 Another way to view MVD:
Defn: Let R (A, B, C) be a relation and A, B, C be sets of
attributes of R, not necessarily disjoint.
Let Ba c ={ b | (a, b, c)  R } /* a and c are some A and C values
The MVD A B is said to hold for R (A, B, C)
if and only if Ba c depends on a only,
i.e. Ba c = Ba c for all a, c, c values of attributes A and C,
whenever Ba c and Ba c are both non-empty.

• We sometime use the embedded MVD notation

A B|C
Note: Pronounce | as independent of. A multi-determines B and independent of C.

• The two definitions for MVD are equivalent.

• For the relation CTX (Course,Teacher,Text), we have
Course Teacher
Course Text
i.e. Course Teacher | Text

Q: What is the intuitive meaning? 43

CS4221: The Relational Model
 Notes: (1) X  and X Y hold for R (X, Y).
(2) X Y whenever Y  X  R for R,
there we use R to represent all attributes of relation R also.

These are called trivial multivalued dependencies.

Note:  is the symbol for the empty set.

Note: Many text books define trivial MVD using (2).

Recall: A functional dependency X  Y is said to be trivial if Y  X .

Defn. A relation R is in fourth normal form (4NF)

if and only if any non-trivial MVD X Y holds in R
implies X is a superkey of R,
i.e. X  a for all attribute a of R.
Recall: A relation R is in BCNF iff any non-trivial FD X  Y holds in R
implies X  a for all attribute a of R.
Note: A superkey is a key or a superset of a key.
44
CS4221: The Relational Model
Inference Rules for Multivalued Dependencies
Let R be a relation with attribute set A.
1. (Complementation)
If X Y then X A – X – Y (Note: “–” is the set difference)
2. (Augmentation)
If X Y and V  W
then WX VY (Note: WX means W union X, i.e. W and X together)
3. (Transitivity)
X Y and Y Z then X Z–Y
4. (Replication)
If X  Y then X Y
5. (Coalescence) If X

Y, Z  Y, and
X
Y
for some W disjoint from Y and W  Z
then X  Z holds also.
Z

W W∩Y=

Note: These 5 rules plus the 3 rules of Armstrong’s Axioms for
FDs are sound and complete for FDs and MVDs. 45
CS4221: The Relational Model
 Result: 4NF relation is also in BCNF.
Theorem. X Y holds for relation R (X, Y, Z)
if and only if R is the join of its projections
R1 (X, Y) and R2 (X, Z).
Note: We call {R1, R2} is a non-loss decomposition of R. R can be
reconstructed by joining R1 and R2.

Corollary. If a relation is not in 4NF, then there is a non-loss
decomposition of R into a set of 4NF relations.
Note: However, it may not cover all the given FDs.


E.g. The relation STJ (S, J, T) with

SJ  T and TJ
STJ is not in BCNF so it is not in 4NF.
We can decompose it into two 4NF relations:
R1 (T, J) and R2 (T, S)
 R1 and R2 form a non-loss decomposition of STJ.
However they do not cover the FD: SJ  T. Bad! 46
CS4221: The Relational Model
 E.g. The relation CTX(course, teacher, text) is in BCNF but not in 4NF
since we have:
course teacher | text
i.e. course teacher
and course text

Q: How do we know the MVDs?
We can decompose the relation into 2 relations:
CT(course, teacher)
CX(course, text)
Both relations are in 4NF.
Note that the MVD
course teacher | text
does not exist in the decomposed relations CT or CX.

Intuitive meaning of the MVD: The text books of a course are
independent of who are the teachers of the course (perhaps the
textbooks of a course are decided by the curriculum committee). 47
CS4221: The Relational Model

The relation CTX (course, teacher, text) is similar to
the below hierarchical model (and XML):

This is a wrong design in hierarchical model.

Recall that the contiguous underline indicate all the attributes
form the key of the relation. It is an all key relation.


Below is a a correct design:

Note: It can be translated into 2 relations:

CT(Course, Teacher)
CX(Course, Text)
48
CS4221: The Relational Model
E.g. Let R be a relation
R(employee, child, salary, year)
A tuple < e, c, s, y > in the relation R indicates c is a child
of employee e and e got a salary s in year y.
Note that R is in BCNF but not in 4NF, and
employee child
employee {salary, year}

Q: How do we know/discover these 2 MVDs?
We can decompose R into
R1 (employee, child)
R2 (employee, salary, year)
Both relations are in 4NF.
Note that in the above relation, an employee may have more
than one salary adjustment within one year.
Q: What if an employee can only has one salary adjustment in
January? Any impact on the FDs and MVDs?
49
CS4221: The Relational Model
 3 possible hierarchical database designs (or XML) of the relation R:

Employee Employee

 Child Child Year

(wrong design)
Year Salary
(correct design)

Salary

Employee
(another correct design)

Child Year/Salary

50
CS4221: The Relational Model
 More Properties of MVDs
Result:  Y in R(Y, Z) iff R is the cartesian product
of its projection R1(Y) and R2 (Z). Prove it!
Q: What is the intuitive meaning of this MVD?

Note: If  Y in R(Y, Z) then YØ z = Yz = {y | (y, z)  R} = R[Y].

Note: A binary relation is definitely in 3NF but not necessarily in 4NF.

How about in BCNF? Yes. Prove it!

Result: If X Y and X Z
then, X YZ (multivalued union rule)
X YZ (multivalued intersection rule)
X Y–Z (multivalued difference rule)
X Z–Y
Prove them!
51
CS4221: The Relational Model
Example. Let R(A, B, C, G, H, I) with the following set of
dependencies D = { A B, B HI, CG  H}

(1) Prove A CGHI  D+

Since A B, by the complementation rule,

we have A R–B–A
i.e. A CGHI  D+
where R means all attributes of the relation R.

Q: Is A CGH  D+ ?

Q: In general, does A BC imply A B?

52
CS4221: The Relational Model
(2) Prove A HI  D+
Since A B and B HI
By the multivalued transitivity rule, we have
A HI - B
i.e. A HI  D+
(3) Prove B  H  D+
Since B HI
H  HI
CG  H
CG  HI = 
By the coalescence rule, we have
B  H  D+
(4) Prove A CG  D+
By (1) we have A CGHI  D+
By (2) we have A HI  D+
By the difference rule, we have
A CGHI – HI  D+
i.e. A CG  D+
53
CS4221: The Relational Model
4NF Decomposition Algorithm (Korth’s book page 206)

Given a relation R with a set of FDs and MVDs D

Step 1. (Initialization)
result := {R};
done := false;
Step 2. (Test for non-trivial MVD)
WHILE (not done) DO
IF (there is a relation Ri  result that is not in 4NF)
THEN BEGIN
 LET X Y be a non-trivial MVD that holds on Ri such that
X  Ri  D+; /* i.e. X is not a superkey
/* need to decompose the relation Ri into 2 smaller relations
SET result := (result – Ri)  (Ri – Y)  (Relation formed by XY)
END;
ELSE done := true;
Q: How to know relation Ri is not in 4NF? I.e. how to find such MVD

X Y that holds on Ri in Step 2?
 Note: There may have several such MVDs, can we just choose anyone of them? 54
CS4221: The Relational Model
 Example. Let R = (A, B, C, G, H, I)
D = {A B, B HI, CG  H}.
Clearly, R is not in 4NF. Why?

(1) Since A B and A is not a key of R (i.e., A  R  D+),

using 4NF decomposition algorithm we get
R1(A, B) and R2(A, C, G, H, I)
Note that R1 is in 4NF.

(2) R2 is not in 4NF (since CG  H, therefore CG H in R2 and

CG is not a key of R2)
Decompose R2 to get
R21 (C, G, H) and R22 (C, G, A, I)
Note: R21 is in 4NF.
(3) We have shown that A HI  D+ earlier.
Hence A I (prove it!) holds in R22.
Also A is not a key of R22, R22 is not in 4NF. Decompose it into:
R221 (A, I) and R222 (C, G, A)
Both are in 4NF.

Q: What happen if we first choose B HI to split the relation? Try it.
55
CS4221: The Relational Model
 Note: The 4NF decomposition algorithm is not a dependency

preserving decomposition.

E.g. The relation

SJT (student, subject, teacher)
with D = {teacher  subject,
student, subject  teacher}

If we use the 4NF decomposition algorithm, we will get

R1 (teacher, subject)
R2 (teacher, student)

The resulting relations do not cover the original FD
student, subject  teacher.

56
CS4221: The Relational Model
 Another method to find 4NF relations
1. Normalize the relation R into a set of 3NF and/or BCNF
relations based on the given set of FDs.
2. For each relation, if all attributes belong to the same key
and there exists non-trivial MVDs in the relation, then
decompose the relation into 2 smaller relations.

Q: How to find such non-trivial MVDs?



Q: How about the covering criteria for normalization?

Note: MVDs are relation sensitive.
What is the meaning of “relation sensitive”?

 Note: When we normalize relations using FDs, we must maintain

(cover) the non-trivial FDs. However, when we normalize relations
to 4NF, we want to remove non-trivial MVDs.
57
CS4221: The Relational Model

MVDs are relation sensitive
Recall that we have 2 MVDs in the relation
CTX (course, teacher, text)
and CTX is not in 4NF.
However, if we add one more attribute, say percentage, to the relation and
it becomes
CTX’ (course, teacher, text, percentage)
A tuple (c,t,x,p) in the relation CTX’ means teacher t teaches course c and
p percentages of his material is from text book x. We have the FD:

course, teacher, text  percentage

 Note that now the two MVDs (in CTX):
course teacher & course text
are no longer hold in CTX’. Q: Why? Prove it.
The relation CTX’ is in 4NF.
This shows MVDs are relation sensitive.
However, we still have course teacher | text in CTX’. 58
CS4221: The Relational Model
 The Chase Algorithm
• An elegant solution for dependency membership test involving FDs
and MVDs.
• Given a set of FDs and MVDs D, does another dependency d (FD or
MVD) follow from D (i.e. d in D+)?
• FD Membership Test. If d is a FD of the form A  B, we create a
table (i.e. relation) which has all the attributes in D with 2 tuples which
have the same A-value.
Our objective is to test whether the B-values of these 2 tuples are the
same after “applying” the FDs and MVDs in D to the tuples in the
table.
If yes, then d in D+ else d is not in D+.
• MVD Membership Test. If d is a MVD of the form A  B, we
create a table which has all the attributes in D with 2 tuples which have
the same A-value.
Our objective is to test after applying the FDs and MVDs in D,
whether there are 2 new tuples in the table which have the same
attribute values of the two original tuples except their B-values are
swapped.
If yes, then d is in D+ else d is not in D+.
59
CS4221: The Relational Model
• Apply an FD in D of the form X  Y. If there are 2 tuples in the
table with same X-value, set their Y-values the same.
• Apply an MVD in D of the form X  Y. If there are 2 tuples in the
able with same X-value, we add 2 new tuples with all the same
attribute values except their Y-values are swapped.

60
CS4221: The Relational Model
Example: Prove that if A  BC and D C, then A  C.
In order to prove AC, we create 2 tuples in the relation with
the same A-value. Our objective is to prove that c1=c2.
A B C D
a b1 c1 d1
a b2 c2 d2

Since ABC, apply the MVD rule, we add 2 tuples into

the relation.
A B C D A B C D
a b1 c1 d1 a b1 c1 d1
a b2 c2 d2 a b2 c1 d2
a b2 c2 d1 a b2 c1 d1
a b1 c1 d2 a b1 c1 d2

Since D  C, and the 1st and 3rd tuples have the same D-value,
so their C-value should be set to equal, i.e. c1=c2.
So, we have proved that A  C. 61
CS4221: The Relational Model
Example: Prove that if AB and BC, then AC
in relation R(A,B,C,D).
In order to prove AC, we create 2 tuples with same A-
value in a relation and then show the 2 tuples (a, b1, c2, d1)
and (a, b2, c1, d2) are in the relation.
A B C D A B C D
Since AB
a b1 c1 d1 a b1 c1 d1
a b2 c2 d2
we add 2 tuples. a b2 c2 d2
a b2 c1 d1
Since BC, we add 2 + 2 tuples. a b1 c2 d2
A B C D
a b1 c1 d1
a b2 c2 d2
a b2 c1 d1 The 2 tuples (a, b1, c2, d1) and (a, b2, c1, d2)
a b1 c2 d2 are now in the relation. So we have proved that
a b1 c2 d1 AC
a b1 c1 d2
a b2 c1 d2
62
a b2 c2 d1 CS4221: The Relational Model
 Example (Counter example by chase).
Prove or disprove the statement:
If ABC and CD B then AB.
In order to prove or disprove AB, we create 2 tuples with same
A-value in a relation and find out whether we can conclude b1=b2.

A B C D A B C D
a b1 c1 d1
Since ABC a b1 c1 d1
a b2 c2 d2 add 2 tuples a b2 c2 d2
a b2 c2 d1
a b1 c1 d2

We cannot further apply the FD: C D B to the relation, so

the relation remains unchanged.
Since this relation satisfies the two given dependencies but it
does not satisfy AB. This relation is a counter example.
So, the above statement is not true.
63
CS4221: The Relational Model

Summary on FDs and MVDs in
Database Design
• How can we find FDs in a RDB? Can we use some data
mining techniques to find FDs in a RDB?
• How to choose the primary key of a relation? What are the
 criteria?
 • Are there updating anomalies in a BCNF relation?
• If a relation is not in BCNF, can we always normalize it to a
set of BCNF relations?
• What are the normalization criteria in database schema
design?
• Free of local redundant attributes is not enough, global
redundancy may still exist. 3NF and BCNF relations are
defined on individual relations, not the whole database, so they
may contain global redundant attributes.
• What are the main differences between decomposition vs.
synthesizing methods? What are their weak points?
64
CS4221: The Relational Model
Summary (cont.)
• How do we find non-trivial MVDs in a relation?


• MVDs are relation sensitive.
• If a relation is not in 4NF, then there is a non-loss
decomposition of R into a set of 4NF relations.
 However, it may not cover all the given FDs.
• When we normalize relations involving onlyFDs,
we must maintain (cover) all the non-trivial FDs.
However, when we normalize relations to 4NF, we
want to remove non-trivial MVDs.
• The Chase Algorithm for FD/MVD membership test.

65
CS4221: The Relational Model
 Some other normal forms
• Fifth Normal Form (5NF) or called
Project-Join Normal Form (PJNF).
• Domain-Key Normal Form (DKNF)
• For your reading pleasure. They will not be
covered/examined.

66
CS4221: The Relational Model
Fifth Normal Form (Project-Join Normal Form)
(5NF, PJNF) (will not be covered/examined)

There exist relation that cannot be non-loss decomposed into

two relations, but can be non-loss decomposed into three or
more relations.
Example Let us consider the relation
STOCK(Agent, Company, Product)
We assume that:
1. Agents represent companies.
2. Companies make products.
3. Agents sell products
4. If an agent sells a product and he
represents the company making that
product, then he sells that product for
that company.
Note: It is an all key relation. There is no FD or MVD in the relation.
67
CS4221: The Relational Model
Relation instances:
STOCK (Agent, Company, Product)
a1 c1 p1
a1 c2 p1
a1 c1 p3
a1 c2 p4
a2 c1 p1
a2 c1 p2
a3 c2 p4

REP (Agent, Company) MAKE (Company, Product) SELL (Agent, Product)

a1 c1 c1 p1 a1 p1
a1 c2 c1 p2 a1 p3
a2 c1 c1 p3 a1 p4
a3 c2 c2 p1 a2 p1
c2 p4 a2 p2
a3 p4
68
CS4221: The Relational Model
Notes: (1) There is no FD or MVD in the relation STOCK
(2) The relation is in 4NF.
(3) There are redundant data in the relation.
(4) However, the relation can be non-loss decomposed into
3 relations, namely
REP (Agent, Company)
MAKE (Company, Product)
SELL (Agent, Product)

Q: How do you know this?

(5) REP MAKE SELL = STOCK

69
CS4221: The Relational Model
 Defn: Let R be a relation and R1, …, Rn be a decomposition of
R. We say that R satisfies the join dependency *{ R1,
R2, …, Rn} iff
n
Ri = R
i=1
( or R1 R2 … Rn = R
or R1 * R2 * … * Rn = R )
Defn: A join dependency (JD) is trivial if one of the Ri is
R itself.
Note: When n = 2, the join dependency of the form
*{R1, R2} is equivalent to a multivalued dependency.

Example. The relation STOCK(Agent, Company, product) satisfies

the join dependency:
*{R1(Agent, Company), R2(Agent, Product), R3(Company, Product)}

However, there is no MVD in the relation. 70

CS4221: The Relational Model
Defn: A relation R is in fifth normal form (5NF) or called
Project-Join normal form (PJNF) iff every non-trivial join
dependency in R is implied by the candidate keys of R.
i.e. whenever a non-trivial join dependency *{R1, R2, …, Rn}
holds in R, implies every Ri (all the attributes of Ri) is
a superkey for R.
Example: The relation STOCK(Agent, Company, Product) is not in 5NF.

Results: (1) A 5NF relation is in 4NF.


(2) Any relation can be non-loss decomposed into an
equivalent collect of 5NF relations, if covering
criteria (of FDs) is not required.

Example: The relation Stock can be non-loss decomposed into 3 relations:

REP (Agent, Company)
SELL (Agent, Product)
MAKE (Company, Product)
All are in 5NF. 71
CS4221: The Relational Model
 Domain-Key Normal Form (DKNF)
(will not be covered/examined)
Note that FDs, MVDs and JDs are some sorts of integrity constraints.
There are other types of constraints:
(1) Domain constraint - which specifies the possible values of
some attribute.
E.g. The only colors of cars are blue, white, red, grey.
E.g. The age of a person is between 0 and 150.
(2) Key constraint - which specifies keys of some relation.
Note: All key declarations are FDs but not reverse.
(3) General constraints - any other constraints which can be
expressed by the first order logic.
E.g. If the first digit of a bank account is 9, then the balance of
the account is greater than 2500.

72
CS4221: The Relational Model
Defn: Let D, K, G be the set of domain constraints, the set
of key constraints, and the set of general constraints of a
relation R.
R is said to be in domain-key normal form (DKNF) if
D  K logically implies G.
i.e. all constraints can be expressed by only domain
constraints and key constraints.

73
CS4221: The Relational Model
Example. Let Acct(acct#, balance) with acct#  balance
and a general constraint:
“ if the first digit of an account is 9,
then the balance of the account is  2500.”

• Relation Acct is not in DKNF.

• To create a DKNF design, we split the relation horizontally
into 2 relations:
Regular_Acct (acct#, balance)
Key = {acct#}
Domain constraint: the first digit of acct# is not 9.
Special_Acct (acct#, balance)
Key = {acct#}
Domain constraints:
(1) t he first digit of acct# is 9, and.
(2) balance  2500.
Both relations are in DKNF. Why?
All constraints can now be enforced as domain constraints and key constraints.
Q: How to enforce them? 74
CS4221: The Relational Model
Note: We can rewrite the definitions of PJNF, 4NF, and BCNF
in a manner which shows them to be special case of DKNF.
E.g. Let R=(A1, …, An) be a relation.
Let dom(Ai) denote the domain of attribute Ai and let all these
domains be infinite.
Then all domain constraints D are of the from
Ai  dom(Ai).
Let the general constraints be a set G of FDs and MVDs .
Let K be the set of key constraints.
R is in 4NF iff it is in DKNF with respect to D, K, G.
(i.e. every FD and MVD is implied by the domain constraints and
key constraints.)

Note: PJNF and BCNF can be rewritten similarly.

Q: How about 3NF?
75
CS4221: The Relational Model
Theorem
Let R be a relation in which dom(A) is infinite for
each attribute A.
If R is in DKNF then it is in PJNF.

Thus if all domains are infinite, then

DKNF  PJNF  4NF  BCNF  3NF

76
CS4221: The Relational Model