Chapter 1

Introduction to Database
Concepts

1.1 Databases and Database Systems

1.2 The Architecture of Database Systems
1.3 A Historical Perspective of Database Systems
1.4 Bibliographical Comments

1.1 Databases and Database Systems

1.1.1 What Is a Database?
A database can be summarily described as a repository for data. This makes
clear that building databases is really a continuation of a human activity that has
existed since writing began; it can be applied to the result of any bookkeeping
or recording activity that occurred long before the advent of the computer era.
However, this description is too vague for some of our purposes, and we refine
it as we go along.
The creation of a database is required by the operation of an enterprise. We
use the term enterprise to designate a variety of endeavors that range from an
airline, a bank, or a manufacturing company to a stamp collection or keeping
track of people to whom you want to write New Year cards.
Throughout this book we use a running example that deals with the database
of a small college. The college keeps track of its students, its instructors, the
courses taught by the college, grades received by students, and the assignment
of advisors to students, as well as other aspects of the activity of the institution
that we discuss later. These data items constitute the operational data — that
is, the data that the college needs to function. Operational data are built from
various input data (application forms for students, registration forms, grade
lists, schedules) and is used for generating output data (transcripts, registration
records, administrative reports, etc.) Note that no computer is necessary for
2 Introduction to Database Concepts

using such a database; a college of the 1930’s would have kept the same database
in paper form. However, the existence of computers to store and manipulate
the data does change user expectations: we expect to store more data and make
more sophisticated use of these data.

1.1.2 Database Management Systems

A database management system (DBMS) is an aggregate of data, hardware,
software, and users that helps an enterprise manage its operational data. The
main function of a DBMS is to provide efficient and reliable methods of data
retrieval to many users. If our college has 10,000 students each year and each
student can have approximately 10 grade records per year, then over 10 years,
the college will accumulate 1,000,000 grade records. It is not easy to extract
records satisfying certain criteria from such a set, and by current standards,
this set of records is quite small! Given the current concern for “grade infla-
tion”, a typical question that we may try to answer is determining the evolution
of the grade averages in introductory programming courses over a 10-year pe-
riod. Therefore, it is clear that efficient data retrieval is an essential function of
database systems.
Most DBMSs deal with several users who try simultaneously to access several
data items and, frequently, the same data item. For instance, suppose that we
wish to introduce an automatic registration system for students. Students may
register by using terminals or workstations. Of course, we assume that the
database contains information that describes the capacity of the courses and
the number of seats currently available. Suppose that several students wish to
register for cs210 in the spring semester of 2003. Unfortunately, the capacity
of the course is limited, and not all demands can be satisfied. If, say, only one
seat remains available in that class, the database must handle these competing
demands and allow only one registration to go through.

Database System Hardware

Database management systems are, in most cases, installed on general-purpose
computers. Since the characteristics of the hardware have strongly influenced
the development of DBMSs, we discuss some of the most important of these
characteristics.
For our purposes, it is helpful to categorize computer memory into two
classes: internal memory and external memory. Although some internal mem-
ory is permanent, such as ROM,1 we are interested here only in memory that
can be changed by programs. This memory is often known as RAM.2 This
memory is volatile, and any electrical interruption causes the loss of data.
By contrast, magnetic disks and tapes are common forms of external mem-
ory. They are nonvolatile memory, and they retain their content for practically
1 ROM stands for Read Only Memory; it is memory that must be written using special

equipment or special procedures, and for our purposes is considered unchangeable.

2 RAM stands for Random Access Memory.
1.1 Databases and Database Systems 3

unlimited amounts of time. The physical characteristics of magnetic tapes force

them to be accessed sequentially, making them useful for backup purposes, but
not for quick access to specific data.
In examining the memory needs of a DBMS, we need to consider the following
issues:
• Data of a DBMS must have a persistent character; in other words, data
must remain available long after any program that is using it has com-
pleted its work. Also, data must remain intact even if the system breaks
down.
• A DBMS must access data at a relatively high rate.
• Such a large quantity of data need to be stored that the storage medium
must be low cost.
These requirements are satisfied at the present stage of technological develop-
ment only by magnetic disks.

Database System Software

Users interact with database systems through query languages. The query lan-
guage of a DBMS has two broad tasks: to define the data structures that serve
as receptacles for the data of the database, and to allow the speedy retrieval
and modification of data. Accordingly, we distinguish between two components
of a query language: the data definition component and the data manipulation
component.
The main tasks of data manipulation are data retrieval and data update.
Data retrieval entails obtaining data stored in the database that satisfies a
certain specification formulated by the user in a query. Data updates include
data modification, deletion and insertion.
Programming in query languages of DBMSs is done differently from pro-
gramming in higher-level programming languages. The typical program written
in C, Pascal, or PL/1 directly implements an algorithm for solving a problem.
A query written in a database query language merely states what the problem
is and leaves the construction of the code that solves the problem to a special
component of the DBMS software. This approach to programming is called
nonprocedural.
A central task of DBMSs is transaction management. A transaction is a
sequence of database operations (that usually consists of updates, with possible
retrievals) that must be executed in its entirety or not at all. This property of
transactions is known as atomicity. A typical example includes the transfer of
funds between two account records A and B in the database of a bank. Such a
banking operation should not modify the total amount of funds that the bank
has in its accounts, which is a clear consistency requirement for the database.
The transaction consists of the following sequence of operations:
1. Decrease the balance of account A by d dollars;
2. Increase the balance of account B by d dollars.
4 Introduction to Database Concepts

If only the first operation is executed, then d dollars will disappear from the
funds deposited with the bank. If only the second is executed, then the total
funds will increase by d dollars. In either case, the consistency of the database
will be compromised. Thus, a transaction transforms one consistent database
state into another consistent database state, a property of transactions known
as consistency.
Typically, at any given moment in time a large number of transactions co-
exist in the database system. The transaction management component ensures
that the execution of one transaction is not influenced by the execution of any
other transaction. This is the isolation property of transactions. Finally, the
effect of a transaction to the state of the database must by durable, that is, it
must persist in the database after the execution of the transaction is completed.
This is the durability property of transactions. Collectively, the four fundamen-
tal properties of transactions outlined above are known as the ACID properties,
the acronym of atomicity, consistency, isolation, and durability.
DBMS software usually contains application development tools in addition
to query languages. The role of these tools is to facilitate user interface develop-
ment. They include forms systems, procedural and nonprocedural programming
languages that integrate database querying with various user interfaces, etc.

The Users of a Database System

The community of users of a DBMS includes a variety of individuals and orga-
nizational entities. These users are classified based on their roles and interests
in accessing and managing the databases.
Once a database is created, it is the job of the database administrator to
make decisions about the nature of data to be stored in the database, the access
policies to be enforced (who is going to access certain parts of the database),
monitoring and tuning the performance of the database, etc.
At the other extremity of the user range, we have the end users. These
users have limited access rights, and they need to have only minimal technical
knowledge of the database. For instance, the end users of the database of the
reservation system of an airline are travel and sales agents. The end users of a
DBMS of a bank are bank tellers, users of the ATM machines, etc.
A particularly important category of users of DBMSs (on whom we focus
in this book) consists of application programmers. Their role is to work within
existing DBMS systems and, using a combination of the query languages and
higher-level languages, to create various reports based on the data contained in
the database. In some cases, they write more general programs that depend on
these data.

1.2 The Architecture of Database Systems

The architecture of a DBMS can be examined from several angles: the functional
architecture that identifies the main components of a DBMS, the application
1.2 The Architecture of Database Systems 5

architecture that focuses on application uses of DBMSs, and the logical archi-
tecture that describes various levels of data abstractions.
Functionally, a DBMS contains several main components shown in Fig-
ure 1.1:
• the memory manager;
• the query processor;
• the transaction manager.
The query processor converts a user query into instructions the DBMS can
process efficiently, taking into account the current structure of the database
(also referred as metadata — which means data about data).
The memory manager obtains data from the database that satisfies queries
compiled by the query processor and manages the structures that contain data,
according to the DDL directives.
Finally, the transaction manager ensures that the execution of possibly many
transactions on the DBMS satisfies the ACID properties mentioned above and,
also, provides facilities for recovery from system and media failures.
The standard application architecture of DBMSs is based on a client/server
model. The client, which can be a user or an application, generates a query that
is conveyed to the server. The server processes the query (a process that includes
parsing, generation of optimized execution code, and execution) and returns an
answer to the client. This architecture is known as two-tier architecture. In
general, the number of clients may vary over time.
In large organizations, it is often necessary to create more layers of process-
ing, with, say, a layer of software to concentrate the data activities of a branch
office and organize the communication between the branch and the main data
repository. This leads to what is called a multi-tier architecture. In this setting
data are scattered among various data sources that could be DBMSs, file sys-
tems, etc. These constitute the lowest tier of the architecture, that is, the tier
that is closest to the data. The highest tier consists of users that act through
user interfaces and applications to obtain answers to queries. The intermediate
tiers constitute the middleware, and their role is, in general, to serve as me-
diators between the highest and the lowest tiers. Middleware may be consist
of web servers, data warehouses, and may be considerably complex. Multi-tier
architecture is virtually a requirement for world wide web applications.
The logical architecture, also known as the ANSI/SPARC architecture, was
elaborated at the beginning of the 1970s. It distinguishes three layers of data
abstraction:
1. The physical layer contains specific and detailed information that describes
how data are stored: addresses of various data components, lengths in
bytes, etc. DBMSs aim to achieve data independence, which means that
the database organization at the physical level should be indifferent to
application programs.
2. The logical layer describes data in a manner that is similar to, say, defini-
tions of structures in C. This layer has a conceptual character; it shields
the user from the tedium of details contained by the physical layer, but is
6 Introduction to Database Concepts

Data definition Queries (DML)

directives (DDL)

Query
Processor

Transaction
Manager

Memory
Manager

Database
metadata

Database

Figure 1.1: Functional Architecture of DBMSs

1.3 A Historical Perspective of Database Systems 7

essential in formulating queries for the DMBS.

3. The user layer contains each user’s perspective of the content of the data-
base.

1.3 A Historical Perspective of Database Sys-

tems
The history of DBMSs begins in the late 1960s, when an IBM product named
IMS (Information Management System) was launched. Data was structured hi-
erarchically, in forests of trees of records, providing very fast access. A few years
after IMS appeared, in 1971, the CODASYL Database Task Group proposed a
new type of database models known today as the network model. The original
report considered DBMSs as extensions of the COBOL language, and struc-
tured data contained by databases as graphs of records, consisting essentially of
circular linked lists. The origins of the relational model, that is the mainstay of
contemporary databases are in E. F. Codd’s work in the early and mid 1970s.
The development of relational database began in the late 1970s and early 1980s
with an experimental relational database sytem at IBM called System R, a pre-
cursor of commercial IBM DBMSs, SQL/DS and DB2. A multitude of DBMSs
emerged in the 1980s, such as ORACLE, INGRES, Rdb, etc. Relational tech-
nology evolved further in the 1990s with the addition of ideas and techniques
inspired by object-oriented programming.

1.4 Bibliographical Comments

Codd’s foundational work in relational databases is included in several arti-
cles [Codd, 1970; Codd, 1972a; Codd, 1972b; Codd, 1974], and [Codd, 1990].
Standard references in the database literature that contain extensive bibli-
ographies are [Date, 2003; Elmasri and Navathe, 2006; Silberschatz et al., 2005]
and [Ullman, 1988a; Ullman, 1988b].
8 Introduction to Database Concepts
Chapter 2

The Entity–Relationship
Model

2.1 The Main Concepts of the E/R Model

2.2 Attributes
2.3 Keys
2.4 Participation Constraints
2.5 Weak Entities
2.6 Is-a Relationships
2.7 Exercises
2.8 Bibliographical Comments
The entity–relationship model (the E/R model) was developed by P. P. Chen
and is an important tool for database design. After an introductory section, we
define the main elements of the E/R model, and we discuss the use of the E/R
model to facilitate the database design process.

2.1 The Main Concepts of the E/R Model

The E/R model uses the notions of entity, relationship, and attribute. These
notions are quite intuitive. Informally, entities are objects that need to be
represented in the database; relationships reflect interactions between entities;
attributes are properties of entities and relationships.
For the present, the database of the college used for our running example
reflects the following information:
• Students: any student who has ever registered at the college;
• Instructors: anyone who has ever taught at the college;
• Courses: any course ever taught at the college;
• Advising: which instructor currently advises which student, and
• Grades: the grade received by each student in each course, including the
semester and the instructor.
10 The Entity–Relationship Model

STUDENTS ADVISING

GRADES INSTRUCTORS

COURSES

Figure 2.1: The E/R Diagram of the College Database

We stress that this example database is intentionally simplified; it is used

here to illustrate certain ideas. To make it fully useful, we would need to include
many more entities and relationships.
A single student is represented by an entity; a student’s grade in a course is
a single relationship between the student, the course, and the instructor. The
fact that an instructor advises a student is represented by a relationship between
them.
Individual entities and individual relationships are grouped into homoge-
neous sets of entities (STUDENTS, COURSES, and INSTRUCTORS) and homo-
geneous sets of relationships (ADVISING, GRADES). Thus, for example, STU-
DENTS represent all the student entities, and ADVISING, all the individual
advising relationships. We refer to such sets as entity sets and relationship sets,
respectively.
Definition 2.1.1 An n-ary relationship is a relationship that involves n entities
from n pairwise distinct sets of entities E1 , . . . , En .
We use the entity/relationship diagram, a graphical representation of the
E/R model, where entity sets are represented by rectangles and sets of relation-
ships by diamonds. (See Figure 2.1 for a representation of the entity/relationship
diagram of the college database.)
An E/R diagram of a database can be viewed as a graph whose vertices are
the sets of entities and the sets of relationships. An edge may exist only between
a set of relationships and a set of entities. Also, every vertex must be joined by
at least one edge to some other vertex of the graph; in other words, this graph
must be connected. This is an expression of the fact that data contained in a
database have an integrated character. This means that various parts of the
database are logically related and data redundancies are minimized. An E/R
design that results in a graph that is not connected indicates that we are dealing
2.2 Attributes 11

advisee
STUDENTS ADVISING

graded
advisor

grader
GRADES INSTRUCTORS

subject

COURSES

Figure 2.2: Roles of Entities in the College E/R Diagram

with more than one database.

The notion of role that we are about to introduce helps explain the signifi-
cance of entities in relationships. Roles appear as labels of the edges of the E/R
diagram.
Example 2.1.2 We consider the following roles in the college database:

Role Relationship Set Entity Set

advisee ADVISING STUDENTS
advisor ADVISING INSTRUCTORS
graded GRADES STUDENTS
grader GRADES INSTRUCTORS
subject GRADES COURSES

These role explain which entities are involved in the relationship and in
which capacity: who is graded, who is the instructor who gave the grade, and
in which course was the grade given. In Figure 2.2 we show the diagram from
Figure 2.1 with the edges marked by the roles discussed above.

2.2 Attributes
Properties of entities and relationships are described by attributes. Each at-
tribute A has an associated set of values, which we refer to as the domain of A
and denote by Dom(A). The set of attributes of a set of entities E is denoted
by Attr(E); similarly, the set of attributes of a set of relationships R is denoted
by Attr(R).
Example 2.2.1 The set of entities STUDENTS of the college database has
12 The Entity–Relationship Model

the attributes student identification number (stno), student name (name), street
address (addr), city (city), state of residence (state), zip code (zip).
The student Edwards P. David, who lives at 10 Red Rd. in Newton, MA,
02129, has been assigned ID number 1011. The value of his attributes are:
Attribute Value
stno ’1011’
name ’Edwards P. David’
addr ’10 Red Rd.’
city ’Newton’
state ’MA’
zip ’02129’

We assume that domains of attributes consist of atomic values. This means

that the elements of such domains must be “simple” values such as integers,
dates, or strings of characters. Domains may not contain such values as sets,
trees, relations, or any other complex objects. Simple values are those that are
not further decomposed in working with them.
If e is an entity and A is an attribute of that entity, then we denote by
A(e) the value of the domain of A that the attribute associates with the entity
e. Similarly, when r is a relationship, we denote the value associated by an
attribute B to r as B(r). For example, if s is a student entity, then the values
associated to s are denoted by

stno(s), name(s), addr(s), city(s), state(s), zip(s).

A DBMS must support attribute domains. Such support includes validity

checks and implementation of operations specific to the domains. For instance,
whenever an assignment A(e) = v is made, where e is an entity and A is an
attribute of e, the DBMS should verify whether v belongs to Dom(A). Oper-
ations defined on specific domains include string concatenation for strings of
characters, various computations involving dates, and arithmetic operations on
numeric domains.
Dom(name) is the set of all possible names for students. However, such a
definition is clearly impractical for a real database because it would make the
support of such a domain an untenable task. Such support would imply that
the DBMS must somehow store the list of all possible names that human beings
may adopt. Only in this way would it be possible to check the validity of an
assignment of a name. Thus, in practice, we define Dom(name) as the set of all
strings of length less or equal to a certain length n. For the sake of this example,
we adopt n = 35.
The set of all strings of characters of length k is denoted by CHAR(k).
The set of all 4-bytes integers that is implemented on our system is denoted
by INTEGERS. Similarly, we could consider the set of two-byte integers and
denote this set with SMALLINT. Thus, in Figure 2.3, we use CHAR(35) as the
domain for name, SMALLINT as domain for cr, and INTEGER for roomno.
2.2 Attributes 13

Entity Set Attribute Domain Description

STUDENTS stno CHAR(10) college-assigned student ID number
name CHAR(35) full name
addr CHAR(35) street address
city CHAR(20) home city
state CHAR(2) home state
zip CHAR(10) home zip
COURSES cno CHAR(5) college-assigned course number
cname CHAR(30) course title
cr SMALLINT number of credits
cap INTEGER maximum number of students
INSTRUCTORS empno CHAR(11) college-assigned employee ID number
name CHAR(35) full name
rank CHAR(12) academic rank
roomno INTEGER office number
telno CHAR(4) office telephone number

Figure 2.3: Attributes of Sets of Entities

Relationship Set Attribute Domain

GRADES stno CHAR(10)
empno CHAR(11)
cno CHAR(5)
sem CHAR(6)
year INTEGER
grade INTEGER
ADVISING stno CHAR(10)
empno CHAR(11)

Figure 2.4: Attributes of Sets of Relationships

The attributes of the sets of entities considered in our current example (the
college database) are summarized in Figure 2.3.
If several sets of entities that occur in the same context each have an at-
tribute A, we qualify the attribute with the name of the entity set to be able to
differentiate between these attributes. For example, because both STUDENTS
and INSTRUCTORS have the attribute name, we use the qualified attributes
STUDENTS.name and INSTRUCTORS.name.
Attributes of relationships may either be attributes of the entities they relate,
or be new attributes, specific to the relationship. For instance, a grade involves
a student, a course, and an instructor, and for these, we use attributes from the
participating entities: stno, cno, and empno, respectively. In addition, we need
to specify the semester and year when the grade was given as well as the grade
itself. For these, we use new attributes: sem, year, and grade. Therefore, the set
of relationships GRADES has the attributes stno (from STUDENTS), cno (from
COURSES), and empno (from INSTRUCTORS), and also its own attributes sem,
year, and grade. By contrast, the set of relationships ADVISING has only at-
tributes gathered from the entities it relates stno (from STUDENTS) and empno
(from INSTRUCTORS). The attributes of the sets of relationships GRADES and
ADVISING are listed in Figure 2.4. Note that in our college, grades are integers
(between 0 and 100) rather than letters.
It is a feature of the E/R model that the distinction between entities and re-
lationships is intentionally vague. This allows different views of the constituents
14 The Entity–Relationship Model

stno name addr city zip

advisee
STUDENTS ADVISING

graded
advisor
sem
grader
year GRADES INSTRUCTORS

grade
empno name rank roomno telno
subject

COURSES

cno cname cr cap

Figure 2.5: The E/R Diagram of the College Database

of the model to be adopted by different database designers. The distinction be-

tween entities and relationships is a decision of the model builder that reflects his
or her understanding of the semantics of the model. In other words, an object
is categorized as an entity or a relationship depending on a particular design
choice at a given moment; this design decision could change if circumstances
change. For instance, the E/R model of the college database regards GRADES
as a set of relationships between STUDENTS, COURSES, and INSTRUCTORS.
An alternative solution could involve regarding GRADES as a set of entities
and then introducing sets of relationships linking GRADES with STUDENTS,
COURSES, and INSTRUCTORS, etc.
Sometimes attributes are represented by circles linked to the rectangles or
diamonds by undirected edges. However, to simplify the drawings, we list the
attributes of sets of entities or relationships close to the graphical representations
of those sets as in Figure 2.6.

2.3 Keys
In order to talk about a specific student, you have to be able to identify him. A
common way to do this is to use his name, and generally, this works reasonably
well. So, you can ask something like, “Where does Roland Novak live?” In
database terminology, we are using the student’s name as a “key”, an attribute
(or set of attributes) that uniquely identifies each student. So long as no two
students have the same name, you can use the name attribute as a key.
What would happen, though, if there were two students named “Helen
2.3 Keys 15

Rivers”? Then, the question, “Where does Helen Rivers live?” could not be
answered without additional information. The name attribute would no longer
uniquely identify students, so it could not be used as a key for STUDENTS.
The college solves this problem in a common way: it assigns a unique identi-
fier (corresponding to the stno attribute) to each student when he first enrolls.
This identifier can then be used to specify a student unambiguously; i.e., it can
be used as a key. If one Helen Rivers has ID 6568 and the other has ID 4140,
then instead of talking about “Helen Rivers”, leaving your listener wondering
which one is meant, you can talk about “the student with ID number 6568”.
It’s less natural in conversation, but it makes clear which student is meant.
Avoiding ambiguity is especially important for computer programs, so having
an attribute, or a set of attributes, that uniquely identifies each entity in a
collection is generally a necessity for electronic databases.
We discuss the notion of keys for both sets of entities and sets of relation-
ships. We begin with sets of entities.
Let E be a set of entities having A1 , . . . , An as its attributes. The set
{A1 , . . . , An } is denoted by A1 . . . An . Unfortunately, this notation conflicts
with standard mathematical notation; however, it has been consecrated by its
use in databases, so we adhere to it when dealing with sets of attributes. Further,
if H and L are two sets of attributes, their union is denoted by concatenation;
namely, we write HL = A1 . . . An B1 . . . Bm for H ∪ L if H = A1 . . . An and
L = B1 . . . Bm .
Definition 2.3.1 Let E be a set of entities such that Attr(E) = A1 . . . An . A
key of E is a nonempty subset L of Attr(E) such that the following conditions
are satisfied:
1. For all entities, e, e′ in E, if A(e) = A(e′ ) for every attribute A of L, then
e = e′ (the unique identification property of keys).
2. No proper, nonempty subset of L has the unique identification property
(the minimality property of keys).

Example 2.3.2 In the college database, the value of the attribute stno is suf-
ficient to identify a student entity. Since the set stno has no proper, nonempty
subsets, it clearly satisfies the minimality condition and, therefore, it is a key
for the STUDENTS entity set. For our college, the entity set COURSES both
cno and cname are keys. Note that this reflects a “business rule”, namely that
no two courses may have the same name, even if they are offered by different
departments.
Example 2.3.3 Consider the design of the database of the customers of a town
library. We introduce the entity sets PATRONS and BOOKS and the set of
relationships LOANS between BOOKS and PATRONS. The E/R diagram of this
database is represented in Figure 2.6.
The inventory number invno is clearly a key for the set of entities BOOKS.
If the library never buys more than one copy of any title, then the ISBN
number, isbn, is another key, and so is the set of attributes author title publ
16 The Entity–Relationship Model

PATRONS LOANS BOOKS

name isbn
addr date invno
city duration title
zip authors
telno publ
date_of_birth place
year

Figure 2.6: The E/R Diagram of the Town Library Database

year. For the PATRONS set of entities, it is easy to see that the sets H =
name telno date of birth and L = name addr city date of birth are keys. Indeed,
it is consistent with the usual interpretation of these attributes to assume that
a reader can be uniquely identified by his name, his telephone number, and his
date of birth. Note that the set H satisfies the minimality property. Assume,
for example, that we drop the date of birth attribute. In this case, a father and
a son who live in the same household and are both named “John Smith” can-
not be distinguished through the values of the attributes name and telno. On
the other hand, we may not drop the attribute telno because we can have two
different readers with the same name and date of birth. Finally, we may not
drop name from H because we could not distinguish between two individuals
who live in the same household and have the same date of birth (for instance,
between twins). Similar reasoning shows that L is also a key (see Exercise 10).

Example 2.3.3 shows that it is possible to have several keys for a set of
entities. One of these keys is chosen as the primary key; the remaining keys are
alternate keys.
The primary key of a set of entities E is used by other constituents of the
E/R model to refer to the entities of E.
As we now see, the definition of keys for sets of relationships is completely
parallel to the definition of keys for sets of entities.
Definition 2.3.4 Let R be a set of relationships. A subset L of the set of
attributes of R is a key of R if it satisfies the following conditions:
1. If A(r) = A(r′ ) for every attribute A of L, then r = r′ (the unique identi-
fication property of relationships).
2. No proper subset of L has the unique identification property (the mini-
mality property of keys of relationships).

Note that the attributes that form a key of a set R of relationships are
themselves either attributes of R or keys of the entities that participate in
the relationships of R. The presence of the keys of the entities is necessary
to indicate which entities actually participate in the relationships. There is no
logical necessity that any particular key be chosen, but the reason for designating
2.4 Participation Constraints 17

one of the keys as the primary key is to make sure a single key is used to access
entities of the corresponding set.
Example 2.3.5 For instance, if we designate
H = name telno date of birth
as the primary key for PATRONS and invno as primary key for BOOKS, we
obtain the following primary key for LOANS:
K = name telno date of birth invno date
To account for the possibility that a single patron borrows the same book repeat-
edly, thereby creating several loan relationships, the date attribute is necessary
to distinguish among them.
Definition 2.3.6 A foreign key for a set of relationships is a set of attributes
that is a primary key of a set of entities that participates in the relationship set.

Example 2.3.7 The set of attributes name telno date of birth is a foreign key
for the set of relationships LOANS because it is a primary key of PATRONS.
We conclude this initial presentation of keys by stressing that the identifi-
cation of the primary key and of the alternate keys is a semantic statement: It
reflects our understanding of the role played by various attributes in the real
world. In other words, choosing the primary key from among the available keys
is a choice of the designer.

2.4 Participation Constraints

The E/R model allows us to impose constraints on the number of relationships
in which an entity is allowed to participate. Let R be a set of relationships
between the sets of entities E1 , . . . , En . The database satisfies the participa-
tion constraint (Ej , u, v, R) if every entity e in Ej participates in at least u
relationships and no more than v relationships.
Example 2.4.1 Suppose, for instance, that the college requires that a student
complete at least one course and no more than 45 courses (during the entire
duration of his or her studies). This corresponds to a participation constraint
(STUDENTS, 1, 45, GRADES).
If every student must choose an advisor, and an instructor may not advise
more than 7 students, we have the participation constraints
(STUDENTS, 1, 1, ADVISING)
and
(INSTRUCTORS, 0, 7, ADVISING)
18 The Entity–Relationship Model

stno advisee
name 1:1
addr STUDENTS ADVISING
city
zip
graded
1:45 advisor
0:7
sem
year grader
GRADES INSTRUCTORS
grade
empno name rank roomno telno

subject

COURSES

cno cname cr cap

Figure 2.7: Participation Restrictions

If (E, u, v, R) is a participation constraint we may add u : v to whatever

other labels may be on the edge joining E to R. When there is no upper limit
to the number of relationships in which an entity may participate, we write
u : +.
Figure 2.7 reflects the roles and the participation constraints mentioned in
Example 2.4.1.
Example 2.4.2 If a reader can have no more than 20 books on loan from
the town library discussed in Example 2.3.3, then we impose the participation
constraints

(PATRONS, 0, 20, LOANS) and (BOOKS, 0, 1, LOANS).

The second restriction reflects the fact that a book is on loan to at most one
patron.
Let R be a set of binary relationships involving the sets of entities U and
V . We single out several types of sets of binary relationships because they are
popular in the business-oriented database literature. If every entity in U is
related to exactly one entity in V , then we say that R is a set of one-to-one
relationships. If an entity in U may be related to several entities of V , then R
is a set of one-to-many relationships from U to V . If, on the other hand, many
entities of U are related to a single entity in V , then R is a set of many-to-one
relationships from U to V . And finally, if there are no such limitations between
the entities of U and V , then R is a set of many-to-many relationships.
Example 2.4.3 The set of binary relationships LOANS between BOOKS and
PATRONS considered in Example 2.4.2 is a one-to-many set of binary relation-
ships.
2.5 Weak Entities 19

p:q m:n
U R V

Figure 2.8: Binary Relationship with Participation Restrictions

STUDENTS PREREQ

Figure 2.9: Recursive Set of Relationships

This terminology is limited to sets of binary relationships. We prefer to

redefine these relationships using the participation constraints (U, p, q, R) and
(V, m, n, R) that are imposed on the sets of entities by R (see Figure 2.8). This
both makes the definitions very precise and generalizes the previous definitions
to arbitrary relationships.
The set of relationships R from U to V is:
1. one-to-one if p = 0, q = 1 and m = 0, n = 1;
2. one-to-many if p = 0, q > 1 and m = 0, n = 1;
3. many-to-one if p = 0, q = 1 and m = 0, n > 1;
4. many-to-many if p = 0, q > 1 and m = 0, n > 1.
A recursive relationship is a binary relationship connecting a set of entities
to itself.
Example 2.4.4 Suppose that we intend to incorporate in the college database
information about prerequisites for courses. This can be accomplished by intro-
ducing the set of relationships PREREQ. If we assume that a course may have up
to three prerequisites and place the appropriate participation constraint, then
we obtain the E/R diagram shown in Figure 2.9.

2.5 Weak Entities

Suppose that we need to expand our database by adding information about
student loans. This can be done, for instance, by adding a set of entities called
LOANS. We assume that a student can have several loans (for the sake of this
example, let us assume that a student can get up to 10 different loans). The
existence of a loan entity in the E/R model of the college database is conditioned
upon the existence of a student entity corresponding to the student to whom
that loan was awarded. We refer to this type of dependency as an existence
dependency.
20 The Entity–Relationship Model

STUDENTS stno
name
addr
recipient city
1:10 zip

BORROW

award
1:1
source
LOANS amount
year

Figure 2.10: Representation of Weak Sets of Entities

The sets of entities STUDENTS and LOANS are related by the one-to-many
sets of relationships BORROW.
If a student entity is deleted, the LOANS entities that depend on the student
entity should also be removed. Note that the attributes of the LOANS entity
set (source, amount, year) are not sufficient to identify an entity in this set.
Indeed, if two students (say, the student whose student number is s1 and the
student whose student number is s2 ) both got the “CALS” loan for 1993, valued
at $1000, there is no way to distinguish between these entities using their own
attributes. In other words, the set of entities LOANS does not have a key.
Definition 2.5.1 Let E, E ′ be sets of entities and let R be a set of relationships
between E and E ′ . E is a set of weak entities if the following conditions are
satisfied:
1. The set of entities E does not have a key, and
2. the participation constraint (E, 1, k, R) is satisfied for some k ≥ 1.

The second condition of Definition 2.5.1 states that no entity can exist in E
unless it is involved in a relationship of R with an entity of E ′ . According to
Definition 2.5.1, LOANS is a set of weak entities
Weak entity sets are represented in E/R diagrams by dashed boxes (see
Figure 2.10).
Example 2.5.2 Consider a personnel database that contains a set of entities
PERSINFO that contains personal information of the employees of a software
company and a set of entities EMPHIST that contains employment history
records of the employees. A set of recursive relationships REPORTING gives
the reporting lines between employees; the set of entities EMPHIST is related
to PERSINFO through the sets of relationships BELONGS TO.
Note that the existence of an employment history entity in EMPHIST is
2.6 Is-a Relationships 21

superv
1: +

sub
PERSINFO 0:1
REPORTING

emp
1: +

BELONGS_TO

pos
1:1

LOANS

Figure 2.11: E/R Diagram of the EMPLOYEES database

conditioned upon the existence of a personal information entity in PERSINF.

The E/R diagram of this database is shown in Figure 2.11.

2.6 Is-a Relationships

We often need to work with subsets of sets of entities. Because the E/R model
deals with sets of relationships, set inclusion must be expressed in these terms.
Let S, T be two sets of entities. We say that S is-a T if every entity of S is
also an entity of T . In terms of the E/R model we have the set of relationships
Ris-a . Pictorially, this is shown in Figure 2.12(a), where the representation for
S is drawn below the one for T ; we simplify this representation by replacing
the diamond in this case by an arrow marked is-a directed from S to T as in
Figure 2.12(b).
For example, foreign students are students, so we can use the notation FOR-
EIGN STUDENTS is-a STUDENTS.
Since every entity of S is also an entity of T the attributes of T are inher-
ited by S. This property of the is-a relationships is known as the descending
inheritance property of is-a.
Example 2.6.1 Consider UNDERGRADUATES and GRADUATES the sets of
entities representing the undergraduate and the graduate students of the col-
lege. For undergraduate students we add sat as an extra attribute; for graduate
students we add the attribute gre, which refers to the score obtained in the GRE
examination. Both these sets of entities are linked to STUDENTS through the
22 The Entity–Relationship Model

T T

is-a is-a

S S

(a) (b)

Figure 2.12: Representing an is-a Set of Relationships

is-a set of relationships (see Figure 2.13).

Example 2.6.2 Teaching assistants are both students and instructors, and
therefore, the corresponding set of entities, TAs, inherits its attributes from
both STUDENTS and INSTRUCTORS.
This phenomenon described in Example 2.6.2 is called multiple inheritance,
and certain precautions must be taken when it occurs. If S is-a U and S is-a V
and both U and V have an attribute A, we must have Dom(U.A) = Dom(V.A),
because otherwise it would be impossible to have any meaning for the common
restrictions of these attribute to S.
The is-a relation between sets of entities is transitive; that is, S is-a T and
T is-a U imply S is-a U . To avoid redundancy in defining the is-a relation
between entity sets (and, consequently, to eliminate redundancies involving the
is-a relationships between entities), we assume that for no set of entities S do
we have S is-a S.
The introduction of is-a relationships can be accomplished through two dis-
tinct processes, called specialization and generalization. Specialization makes a
smaller set of entities by selecting entities from a set. Generalization makes a
single set of entities by combining several sets whose attributes are those the
original sets had in common.
Definition 2.6.3 A set of entities E ′ is derived from a set of entities through
a specialization process if E ′ consists of all entities of E that satisfy a certain
condition.
If E ′ is obtained from E through specialization, then E ′ is-a E. In this case
we may mark the arrow leading from E ′ to E by is-a(sp).
Example 2.6.4 The set TAs can be regarded as a specialization of both STU-
DENTS and INSTRUCTORS (see Figure 2.14). Therefore, entities of this set
have all attributes applicable to INSTRUCTORS and STUDENTS and, in addi-
2.6 Is-a Relationships 23

stno
name
addr STUDENTS ADVISING
city
zip

UNDERGRADUATES GRADUATES TAs stipend

sat gre

sem
year GRADES INSTRUCTORS
grade
empno name rank roomno telno

COURSES

cno cname cr cap

Figure 2.13: Representation of is-a Relationships

tion, their specific attribute stipend.

Definition 2.6.5 Let E1 , . . . , En be n set of entities such that
1. no two distinct sets of entities Ei and Ej have an entity in common, and
2. there are some attributes that all entity sets have in common.
Let H be the set of attributes that all entities have in common..
The set of entities E is obtained by generalization from E1 , . . . , En if E
consists of all entities that belong to one of the sets Ei and Attr(E) = H.
If E is obtained from E1 , . . . , En through generalization, we may mark the
is-a arrows pointing from E1 , . . . , En to E by is-a(gen).
Example 2.6.6 Suppose that the construction of the college database begins
from the existing sets of entities UNDERGRADUATES and GRADUATES. Then,
the set of entities STUDENTS could have been obtained through generalization
from the previous two sets (see Figure 2.14).
The importance of the E/R methodology is that it allows the designer to or-
ganize and record his or her conception of the database. This enforces precision
and facilitates unambiguous communication among all workers on a project.
Furthermore, it imposes discipline by requiring the designer to specify the
entities and relationships, along with their attributes, and by insisting on a clear
definition of the sets of relationships between sets of entities, rather than just
assuming that they are somehow connected.
The is-a relationship imposes hierarchy on the sets of entities. Seeing which
sets are generated by specialization and which by generalization helps expose
24 The Entity–Relationship Model

stno
name
addr STUDENTS ADVISING
city
zip
is-a(gen) is-a(sp)
is-a(gen)
UNDERGRADUATES GRADUATES TAs stipend
sat gre is-a(sp)

sem
year GRADES INSTRUCTORS
grade
empno name rank roomno telno

COURSES

cno cname cr cap

Figure 2.14: Specialization and Generalization

the underlying logic as seen by the designer.

In short, we use the E/R technique as a first step in a database project
because it allows us to think about and modify the database before we commit
to the relational design, which we discuss in the next chapter.

2.7 Exercises
1. Consider the following alternative designs for the college database:
(a) Make GRADES a set of entities, and consider binary sets of relation-
ships between GRADES and each of the sets of entities STUDENTS,
COURSES, and INSTRUCTORS.
(b) Replace the set of relationship GRADES with two binary sets of rela-
tionships: One such set should relate STUDENTS with COURSES and
reflect the results obtained by students in the courses; another one
should relate COURSES with INSTRUCTORS and reflect the teaching
assignment of the instructors.
Explain the advantages and disadvantages of these design choices.
2. Consider a database that has a set of entities CUSTOMERS that consists
of all the customers of a natural gas distribution company. Suppose that
this database also records the meter readings for each customer. Each
meter reading has the date of the reading and the number read from the
meter. Bills are generated after six consecutive readings.
(a) Can you consider the readings to be weak entities?
2.7 Exercises 25

(b) Draw the E/R diagram for this database; identify relevant participa-
tion constraints.

3. The data for our college will grow quite large. One of the techniques for
dealing with the explosion of data is to remove those items that are not
used. Describe how you would augment the college database to include
information about when something was accessed (either read or written).
To what will you attach this information? How? How detailed will the
information you attach be? Why? What would you suggest be done with
this information once it is available?
4. Design an E/R model for the patient population of a small medical of-
fice. The database must reflect patients, office visits, prescriptions, bills
and payments. Explain your choice of attributes, relationships, and con-
straints.
5. Design an E/R model for the database of a bank. The database must
reflect customers, branch offices, accounts, and tellers. If you like, you
can include other features, such as deposits, withdrawals, charges, inter-
est, transfers between accounts, etc. Explain your choice of attributes,
relationships, and constraints.
6. Design an E/R model for the database of a car-rental business. Specify
entities, relationships, attributes, keys, and cardinality constraints for this
database and explain your design choices. Be sure to include such objects
like vehicles, renters, rental locations, etc.
7. A small manufacturing company needs a database for keeping track of
its inventory of parts and supplies. Parts have part numbers, names,
type, physical characteristics, etc. Some parts used by the company are
installed during the fabrication process as components of other parts that
enter the device produced by the company. Parts are stored at several
manufacturing facilities. The database must contain information about
the vendors and must keep track of the orders placed with vendors. Use
the E/R technique to design the database.
8. Let A, B, C, and D, be four entity sets linked by is-a relationships as shown
in Figure 2.15. What is wrong with the choice of these relationships?
9. Let E1 , E2 be two sets of entities.
(a) Assume that E is a nonempty set of entities that is a specialization
of both E1 and E2 . Can you construct the generalization of the sets
E1 and E2 ?
(b) Suppose that E ′ is a generalization of E1 and E2 . Can you construct
a set of entities E ′′ that is a common specialization of E1 and E2 ?
What can you say about |E ′′ |?
10. Explain why the set L in example 2.3.3 is a key. Are there reasons for
prefering H as the primary key and L as an alternate key?
26 The Entity–Relationship Model

is-a(sp)

is-a(gen) is-a(gen)

A B

Figure 2.15: Hypothetical Use of Specialization and Generalization

2.8 Bibliographical Comments

The E/R model was introduced by P. P. Chen in his article [Chen, 1976]. Other
important references on this topic are [Teorey, 1990; Elmasri and Navathe, 2006].
Chapter 3

The Relational Model

3.1 Introduction
3.2 Tables — The Main Data Structure of the Relational Model
3.3 Transforming an E/R Design into a Relational Design
3.4 Entity and Referential Integrity
3.5 Metadata
3.6 Exercises
3.7 Bibliographical Comments

3.1 Introduction
The relational model is the mainstay of contemporary databases. This chapter
presents the fundamental ideas of this model, which focus on data organization
and retrieval capabilities.
Informally, the relational model consists of:
• A class of data structures referred to as tables.
• A collection of methods for building new tables starting from an initial
collection of tables; we refer to these methods as relational algebra opera-
tions.
• A collection of constraints imposed on the data contained in tables.

3.2 Tables — The Main Data Structure of the

Relational Model
The relational model revolves around a fundamental data structure called a
table, which is a formalization of the intuitive notion of a table. For example,
the schedule of a small college may look like:
28 The Relational Model

SCHEDULE
dow cno roomno time
’Mon’ ’cs110’ 84 5:00 p.m.
’Mon’ ’cs450’ 62 7:00 p.m.
’Wed’ ’cs110’ 65 10:00 a.m.
’Wed’ ’cs310’ 63 12:00 p.m.
’Thu’ ’cs210’ 63 2:00 p.m.
’Thu’ ’cs450’ 65 3:00 p.m.
’Thu’ ’cs240’ 84 5:00 p.m.
’Fri’ ’cs310’ 63 5:00 p.m.

When contemplating a table we distinguish three main components: the

name of the table, in our case SCHEDULE, the heading of the table, with one
entry for each column, in our case dow, cno, roomno, and time and the content
of the table, i.e., the list of 8 rows specified above.
The members of the heading are referred to as attributes. In keeping with
the practice of databases, if the heading H of the table consists of the attributes
A1 , . . . , An , then we write H as a string rather than a set, H = A1 · · · An .
Each attribute A has a special set that is attached to it called the domain
of A that is denoted Dom(A). This domain comprises the set of values of the
attribute; i.e., a value may occur in a column labeled by A only if it belongs
to the set Dom(A). For example, in the table SCHEDULE considered above the
domain of the attribute dow (for “day of the week”) is the set that consists of
the strings:
’Mon’, ’Tue’, ’Wed’, ’Thu’, ’Fri’, ’Sat’, ’Sun’
We need a name for each table, so we can refer to it. In principle, names
are simply arbitrary strings of symbols chosen from some fixed alphabet (for
example the modern Roman alphabet), augmented with certain special symbols,
such as “ ”, “[”, “]”, “∪”, etc., that we introduce from time to time as needed.
However, it is always a good idea to choose names that are meaningful to the
reader. Similar comments apply to the names of attributes. In most database
systems, these names are case insensitive.
When speaking informally, we use descriptions based on the visual layout of
a table T when it is displayed. So, a tuple t of T is called a row of T . Also, the
set of values that occur under an attribute may be referred to as a column of T .
The term “relational model” reflects that fact that, from a mathematical
point of view, the content of a table is what is known in mathematics as a
relation. To introduce the notion of relation we need to define the Cartesian
product of sets (sometimes called a cross product ), a fundamental set operation.
Definition 3.2.1 Let D1 , . . . , Dn be n sets. The Cartesian product of the se-
quence of sets D1 , . . . , Dn is the set that consists of all sequences of the form
(d1 , . . . , dn ), where di ∈ Di for 1 ≤ i ≤ n.
We denote the Cartesian product of D1 , . . . , Dn by D1 × · · · × Dn .
A sequence that belongs to D1 × · · · × Dn is referred to as an n-tuple, or
simply as tuple, when n is clear from context. When n has small values we use
special terms such as pair for n = 2, triple for n = 3, etc.
3.2 Tables — The Main Data Structure of the Relational Model 29

The Cartesian product can generate rather large sets starting from sets that
have a modest size. For example if D1 , D2 , D3 are three sets having 1000 ele-
ments each, then D1 × D2 × D3 contains 1,000,000,000 elements.

Example 3.2.2 Consider the domains of the attributes dow, cno, roomno and
time:

Dom(dow) = {’Mon’, ’Tue’, ’Wed’, ’Thu’, ’Fri’, ’Sat’, ’Sun’}

Dom(cno) = {’cs110’, ’cs210’, ’cs240’, ’cs310’, ’cs450’}
Dom(roomno) = {62, 63, 65, 84}
Dom(time) = {8:00a.m., . . . , 7:00p.m.}

The Cartesian product of these sets:

Dom(dow) × Dom(cno) × Dom(roomno) × Dom(time)

consists of 7 · 5 · 4 · 12 = 1680 quadruples.

If H = A1 · · · An is a sequence of attributes, we refer to the Cartesian product

Dom(A1 ) × · · · × Dom(An ) as the set of H-tuples. We will denote this set by
tupl(H). Thus, tupl(dow cno roomno time) consists of 1680 quadruples.

Definition 3.2.3 A relation on the sets D1 , . . . , Dn is a subset ρ of the Carte-

sian product D1 × · · · × Dn .

There is no requirement that the n sets be distinct. Many common relations

are defined on D1 × D2 , where D1 = D2 = D; i.e., they are defined on D × D.
Perhaps the most common example of this is the equality relation, consisting of
all pairs (a, a) for a in D.

Example 3.2.4 Consider the set D = {1, 2, 3, 4, 5, 6} and the Cartesian prod-
uct D × D, which has 36 pairs. Certain of these pairs (a, b) have the property
that a is less than b, i.e., that they satisfy the relation a < b. With a little bit
of counting, we see that there are 15 such pairs.
One way to characterize this set is to describe it is operationally. We could
say that if a and b are in D, then a < b if there is some number k in D such
that a + k = b. This has the advantage of being concise.
However, there is another way to describe < on this set: we could list out
all 15 pairs (a, b) of D × D such that a < b. If we do this in a vertical list, we
get (in no particular order)
30 The Relational Model

(3, 4)
(1, 2)
(2, 6)
(2, 5)
(1, 6)
(2, 3)
(1, 3)
(2, 4)
(1, 5)
(5, 6)
(3, 6)
(4, 5)
(4, 6)
(3, 5)
(1, 4)
With a little reformatting, to remove all those parentheses and commas, this
same list of pairs becomes a table of two columns and 15 rows, where each row
lists two elements of D, such that the first is less than the second. Furthermore,
just as in the list above, all pairs with the first element less than the second
occur as some row in this table.

3 4
1 2
2 6
2 5
1 6
2 3
1 3
2 4
1 5
5 6
3 6
4 5
4 6
3 5
1 4
Thus, we have a table that lists precisely the pairs of D × D that comprise
the < relation. In just this same manner, we can list out all the tuples of
any relation defined on finite sets as rows of a table. It is this correspondence
between tables and relations that is at the heart of the name “relational model.”

Example 3.2.5 Using the sets

Dom(dow), Dom(cno), Dom(roomno), Dom(time)
3.2 Tables — The Main Data Structure of the Relational Model 31

we can define course schedules as relations. One possible course schedule is the
relation ρ that consists of the following 8 quadruples:
(’Mon’, ’cs110’, 84, ’5:00 p.m.’), (’Mon’, ’cs450’, 62, ’7:00 p.m.’),
(’Wed’, ’cs110’, 65, ’10:00 a.m.’), (’Wed’, ’cs310’, 63, ’12:00 p.m.’),
(’Thu’, ’cs210’, 63, ’2:00 p.m.’), (’Thu’, ’cs450’, 65, ’3:00 p.m.’),
(’Thu’, ’cs240’, 84, ’5:00 p.m.’), (’Fri’, ’cs310’, 63, ’5:00 p.m.’)
If D1 , D2 , . . . , Dn are n sets with k1 , . . . , kn elements, respectively, then there
are 2k1 k2 ···kn relations that can be defined on these sets. The number of relations
that can be defined on relatively small sets can be astronomical. For example,
if each of D1 , D2 and D3 has ten elements, then there are 21000 relations that
can be defined on D1 , D2 , D3 .
It is clear now that the content of a table T having the heading H = A1 · · · An
is a relation that consists of tuples from tupl(H), and this is the essential part
of the table.
During the life of a database, the constituent tables of the database may
change through insertions or deletions of tuples, or changes to existing tuples.
Thus, at any given moment we may see a different picture of the database, which
suggests the need of introducing the the notion of relational database instance.
Definition 3.2.6 Let H1 , . . . , Hn be n sets of attributes. An relational data-
base instance is a finite collection of tables T1 , . . . , Tn that have the headings
H1 , . . . , Hn , respectively, such that all names of the tables are distinct.
Definition 3.2.7 The tables T and S are compatible if they have the same
headings.
Implicit in the definition of tables is the fact that tables do not contain
duplicate tuples. This is not a realistic assumption, and we shall remove it later,
during the study of SQL, the standard query language for relational databases.
The same relational attribute may occur in several tables of a relational data-
base. Therefore, it is important to be able to differentiate between attributes
that originate from different tables; we accomplish this using the following no-
tion.
Definition 3.2.8 Let T be a table. A qualified attribute is an attribute of the
form T.A. For every qualified attribute of the form T.A, Dom(T.A) is the same
as Dom(A).
Example 3.2.9 The qualified attributes of the table SCHEDULE are

SCHEDULE.dow, SCHEDULE.cno, SCHEDULE.roomno, SCHEDULE.time

3.2.1 Projections
For a tuple t of a table T having the heading H we may wish to consider only
some of the attributes of t while ignoring others. If L is the set of attributes
we are interested in, then t[L] is the corresponding tuple, referred to as the
projection of t on L.
32 The Relational Model

Example 3.2.10 Let H = dow cno roomno time be the set of attributes that
is the heading of the table SCHEDULE introduced above. The tuple

t = (’Mon’, ’cs110’, 84, ’5:00 p.m.’)

can be restricted to any of the sixteen subsets of the set H. For example, the
restriction of t to the set L = dow roomno is the tuple t[L] = (’Mon’, 84), and
the restriction of t to the set K = cno room time is (’cs110’, 84, ’5:00 p.m.’)
The restriction of t to H is, of course, t itself. Also, the restriction of t to
the empty set is t[∅] = (), that is, the empty sequence.
By extension, the table T itself can be projected onto L, giving a new table
named T [L], with heading L, consisting of all tuples of the form t[L], where t
is a tuple in T ; i.e., the rows of T [L] are obtained by projecting the rows of T
on L. Projecting a table often creates duplicate rows, but within the context of
the relational model, which is based on sets, only one copy of each row appears
in the table, as shown in the second projection of Example 3.2.11.
Example 3.2.11 The projection of the table SCHEDULE on the set of at-
tributes dow cno is
SCHEDULE[dow cno]
dow cno
’Mon’ ’cs110’
’Mon’ ’cs450’
’Wed’ ’cs110’
’Wed’ ’cs310’
’Thu’ ’cs210’
’Thu’ ’cs450’
’Thu’ ’cs240’
’Fri’ ’cs310’
The projection of SCHEDULE on the attribute dow gives the following table.
SCHEDULE[dow]
dow
’Mon’
’Wed’
’Thu’
’Fri’

3.3 Transforming an E/R Design into a Rela-

tional Design
The design of a database formulated in the E/R model can be naturally trans-
lated into the relational model. We show how to translate both sets of entities
and sets of relationships into tables.
3.3 Transforming an E/R Design into a Relational Design 33

From time to time, it is necessary to assume that a set of entities or a set

of relationships has a primary key. For any that does not, we can induce a
key by arbitrarily assigning a unique identifier to each element. In “real world”
examples, this is generally accomplished by picking a sequence and assigning
the next unused element to each entity as it enters the system. This can easily
be seen to be a key that we may designate to be the primary key.
For example, whenever a new patron applies for a card at the library, the
library may assign a new, distinct number to the patron; this set of numbers
could be the primary key for the entity set PATRONS. Similarly, each time a
book is loaned out, a new loan number could be assigned, and this set of numbers
could be the primary key for the set of relationships LOANS. Note, however,
that if we introduce these identifiers we in fact have made a small change to
the original in that we have actually added a new attribute to PATRONS and
to LOANS.
Consider a set of entities named E that has the set of attributes H =
A1 . . . An . Its translation is a table named E that has the heading A1 . . . An .
For each entity e, we include in the table a tuple te by that has the components
A1 (e), . . . , An (e).
In other words, for every entity e there is a tuple in the table E consisting
of the tuple (A1 (e), . . . , An (e)). For instance, if e is an entity that represents a
student (that is e ∈ STUDENTS) and
stno(e) = ’2415’
name(e) = ’Grogan A. Mary’
addr(e) = ’8 Walnut St.’
city(e) = ’Malden’
state(e) = ’MA’
zip(e) = ’02148’,
then e is represented in the table named STUDENTS by the row:
(’2415’, ’Grogan A. Mary’, ’8 Walnut St.’, ’Malden’, ’MA’, ’02148’).
The set of entities STUDENTS is translated into the table named STUDENTS
shown in Figure 3.1.

While in the E/R model we dealt with two types of basic constituents, entity
sets and relationship sets, in the relational model, we deal only with tables, and
we use these to represent both sets of entities and sets of relationships. Thus,
it is necessary to reformulate the definition of keys in this new setting. The
conditions imposed on keys are obvious translations of the conditions formulated
in Definition 2.3.1.
Definition 3.3.1 Let T be a table that has the heading H. A set of attributes
K is a key for T if K ⊆ H and the following conditions are satisfied:
1. For all tuples u, v of the table, if u[K] = v[K], then u = v (unique identi-
fication property).
2. There is no proper subset L of K that has the unique identification prop-
erty (minimality property).
34 The Relational Model

STUDENTS
stno name addr city state zip
1011 Edwards P. David 10 Red Rd. Newton MA 02159
2415 Grogan A. Mary 8 Walnut St. Malden MA 02148
2661 Mixon Leatha 100 School St. Brookline MA 02146
2890 McLane Sandy 30 Cass Rd. Boston MA 02122
3442 Novak Roland 42 Beacon St. Nashua NH 03060
3566 Pierce Richard 70 Park St. Brookline MA 02146
4022 Prior Lorraine 8 Beacon St. Boston MA 02125
5544 Rawlings Jerry 15 Pleasant Dr. Boston MA 02115
5571 Lewis Jerry 1 Main Rd. Providence RI 02904

INSTRUCTORS
empno name rank roomno telno
019 Evans Robert Professor 82 7122
023 Exxon George Professor 90 9101
056 Sawyer Kathy Assoc. Prof. 91 5110
126 Davis William Assoc. Prof. 72 5411
234 Will Samuel Assist. Prof. 90 7024

COURSES
cno cname cr cap
cs110 Introduction to Computing 4 120
cs210 Computer Programming 4 100
cs240 Computer Architecture 3 100
cs310 Data Structures 3 60
cs350 Higher Level Languages 3 50
cs410 Software Engineering 3 40
cs460 Graphics 3 30

GRADES
stno empno cno sem year grade
1011 019 cs110 Fall 2001 40
2661 019 cs110 Fall 2001 80
3566 019 cs110 Fall 2001 95
5544 019 cs110 Fall 2001 100
1011 023 cs110 Spring 2002 75
4022 023 cs110 Spring 2002 60
3566 019 cs240 Spring 2002 100
5571 019 cs240 Spring 2002 50
2415 019 cs240 Spring 2002 100
3442 234 cs410 Spring 2002 60 ADVISING
5571 234 cs410 Spring 2002 80 stno empno
1011 019 cs210 Fall 2002 90 1011 019
2661 019 cs210 Fall 2002 70 2415 019
3566 019 cs210 Fall 2002 90 2661 023
5571 019 cs210 Spring 2003 85 2890 023
4022 019 cs210 Spring 2003 70 3442 056
5544 056 cs240 Spring 2003 70 3566 126
1011 056 cs240 Spring 2003 90 4022 234
4022 056 cs240 Spring 2003 80 5544 023
2661 234 cs310 Spring 2003 100 5571 234
4022 234 cs310 Spring 2003 75

Figure 3.1: An Instance of the College Database

3.3 Transforming an E/R Design into a Relational Design 35

If several keys exist for a table, one of them is designated as the primary
key of the table; the remaining keys are alternate keys. The main role of the
primary key of a table T is to serve as a reference for the tuples of T that can
be used by other tables that refer to these tuples.
Example 3.3.2 The table that results from the translation of the set of entities
PATRONS introduced in Example 2.3.3 has the keys
K = name telno date of birth
and
L = name address city date of birth.
If we consider K to be the primary key, then L is an alternate key.
As a practical matter, if K1 and K2 are both keys, where K1 has fewer
attributes than K2 , we would prefer K1 as the primary key.
Translating sets of relationships is a little more intricate than translating
tables. Let R be a set of relationships that relates the set of entities E1 , . . . , En .
Suppose that every set Ei has its own primary key Ki for 1 ≤ i ≤ n and that
no two such keys have an attribute in common. We exclude, for the moment,
the is-a relationship and the dependency relationship that relates sets of weak
entities to sets of regular entities. When translating relationships, the entities
involved are represented by their primary key values.
If the set of attributes of R itself is B1 , . . . , Bk , then a relationship r of
R relates the entities e1 , . . . , en with some values, say b1 , . . . , bk . In making
the translation of this particular relationship, each entity ei is represented by
its primary key, ei [Ki ], which may comprise several values, ei1 , . . . , eimi . To
simplify, we will write ~ei for the primary key of ei . The value of the relationship
itself is represented by the value bj of each attribute Bj . So, r can be translated
to a tuple wr = (~e1 , . . . , ~en , b1 , . . . , bk ). In other words, the translation WR of
the set of relationships R is defined on the set of all attributes that appear in
the primary keys of the entities, K1 , . . . , Kn , as well as attributes B1 , . . . , Bk ;
and the tuple wr is put together from the values of the primary keys of the
participating entities and the values of the attributes of the relationship r.
Example 3.3.3 Consider, for example, the relationship g that belongs to the
set of relationships GRADES, that relates STUDENTS, COURSES, and INSTRUC-
TORS. Further, assume that this relationship involves the student whose student
number is ’1011’, the instructor whose employee number is ’019’, and the course
whose number is ’cs110’. Further, assume that
sem(g) = ’Fall’
year(g) = ’2001’
grade(g) = 40.
Then, the relationship g will be represented by the tuple
wg = (’1011’, ’019’, ’cs110’, ’Fall’, ’2001’, 40).
36 The Relational Model

Formally, wr is given by:


A(ei ) if A is in Ki for some i, 1 ≤ i ≤ n
wr (A) =
A(r) if A is in {B1 , . . . , Bk }.

In turn, the set R is translated into a table named R whose heading contains
B1 , . . . , Bk as well as all attributes that occur in a key K1 , . . . , Kn . The content
of this table consists of all tuples of the form wr for each relationship r.
The collection of tables shown in Figure 3.1 represents an instance of the
college database obtained by the transformation of the E/R design shown in
Figure 2.5.
If E is a set of weak entities linked by a dependency relationship R to a set
of entities E ′ , then we map both the set of entities E and the set of relationships
R to a single table T defined as follows. If K is the primary key of the table
T ′ that represents the set of entities E ′ , we define H to be the set of attributes
that includes the attributes of E and the attributes of K. The content of the
table T consists of those tuples t in tupl(H) such that there exists an entity e′
in E ′ and a weak entity e in E such that

A(e) if A is an attribute of E
t(A) =
A(e′ ) if A belongs to K.

Example 3.3.4 Consider the set of weak entities LOANS dependent on the set
STUDENTS. Assuming that the primary key of STUDENTS is stno, both the
relationship GRANTS and the weak set of entities LOANS are translated into
the table named LOANS:
LOANS
stno source amount year
1011 CALS 1000 2002
1011 Stafford 1200 2003
3566 Stafford 1000 2002
3566 CALS 1200 2003
3566 Gulf Bank 2000 2003

Example 3.3.5 In Example 2.5.2 we discussed the E/R design of a personnel

database. Recall that we had the set of entities PERSINFO and the set of
weak entities EMPHIST linked to PERSINFO through the set of relationships
BELONGS TO. In addition, we had the set of relationships REPORTING. These
components of the E/R model are translated into three tables represented in
Figure 3.2.
The translation of a set of entities involved in an is-a relationship depends
on the nature of the relationship (generalization or specialization).
Suppose that a set of entities is obtained by generalization from the collection
of sets of entities E1 , . . . , En , such that no two distinct sets of entities Ei and Ej
have an entity in common. In addition we assume that there are attributes that
are shared by all sets E1 , . . . , En and we denote the set of all such attributes by
H.
3.3 Transforming an E/R Design into a Relational Design 37
PERSINFO
empno ssn name address city zip state
1000 ’340-90-5512’ ’Natalia Martins’ ’110 Beacon St.’ ’Boston’ ’02125’ ’MA’
1005 ’125-91-5172’ ’Laura Schwartz’ ’40 Tremont St.’ ’Newton’ ’02661’ ’MA’
1010 ’016-70-0033’ ’John Soriano’ ’10 Whittier Rd.’ ’Lexington’ ’02118’ ’MA’
1015 ’417-52-5751’ ’Kendall MacRae’ ’4 Maynard Dr.’ ’Cambridge’ ’02169’ ’MA’
1020 ’311-90-6688’ ’Rachel Anderson’ ’55 Columbus St.’ ’Boston’ ’02123’ ’MA’
1025 ’671-27-5577’ ’Richard Laughlin’ ’37 Greenough St.’ ’Somerville’ ’02060’ ’MA’
1030 ’508-56-7700’ ’Danielle Craig’ ’72 Dove Rd.’ ’Boston’ ’02225’ ’MA’
1035 ’870-50-5528’ ’Abby Walsh’ ’717 Park St.’ ’Roxbury’ ’02331’ ’MA’
1040 ’644-21-0887’ ’Bailey Burns’ ’35 White Pl.’ ’Cambridge’ ’02169’ ’MA’

EMPHIST
empno position dept appt date term date salary
1000 ’President’ null ’1-oct-1999’ null 150000
1005 ’Vice-President’ ’DB’ ’12-oct-1999’ null 120000
1010 ’Vice-President’ ’WWW’ ’1-jan-2000’ null 120000
1015 ’Senior Engineer’ ’DB’ ’25-oct-1999’ null 100000
1020 ’Engineer’ ’DB’ ’1-nov-1999’ null 70000
1025 ’Programmer’ ’DB’ ’10-mar-2000’ null 70000
1030 ’Senior Engineer’ ’WWW’ ’10-jan-2000’ null 90000
1035 ’Programmer’ ’WWW’ ’20-feb-2000’ null 75000
1040 ’Programmer’ ’WWW’ ’1-mar-2000’ null 70000

REPORTING
empno superv
1000 null
1005 1000
1010 1000
1015 1005
1020 1005
1025 1005
1030 1010
1035 1010
1040 1010

Figure 3.2: An Instance of the Employee Database

If Ei is translated into a table Ti , having the heading Hi for 1 ≤ i ≤ n, then

E is represented by the table T that contains every projections of every tuple
of Ti on the set H.
Example 3.3.6 Assume that we want to form a table named STUDENTS from
the tables named UNDERGRADUATES and GRADUATES in the college data-
base. If these tables have the form
UNDERGRADUATES
stno name addr city state zip major
1011 Edwards P. David 10 Red Rd. Newton MA 02159 CS
2415 Grogan A. Mary 8 Walnut St. Malden MA 02148 BIO
2661 Mixon Leatha 100 School St. Brookline MA 02146 MATH
2890 McLane Sandy 30 Cass Rd. Boston MA 02122 CS
3442 Novak Roland 42 Beacon St. Nashua NH 03060 CHEM

GRADUATES
stno name addr city state zip qualdate
3566 Pierce Richard 70 Park St. Brookline MA 02146 2/1/92
4022 Prior Lorraine 8 Beacon St. Boston MA 02125 11/5/93
5544 Rawlings Jerry 15 Pleasant Dr. Boston MA 02115 2/1/92
5571 Lewis Jerry 1 Main Rd. Providence RI 02904 11/5/93

then the table that represents the set of entities STUDENTS obtained by gen-
eralization from UNDERGRADUATES and GRADUATES is the one shown in
Figure 3.1.
If the set of entities E ′ is obtained by specialization from the set of entities
E, the heading of the table that represents E ′ must include the attributes of E
plus the extra attributes that are specific to E ′ whenever such attributes exist
(see Figure 3.3).
38 The Relational Model

A1
.. translation A1 ··· An
. E -
An
6

is-a
(sp)
A1
..
.
translation A1 ··· An B1 ··· Bℓ
An E ′ -
B1
..
.
Bℓ

Figure 3.3: Translation of Specialization

Example 3.3.7 The heading of the table that represents the set of entities TA
consists of the attributes stno, name, addr, city, state, zip, empno, rank, roomno,
telno, stipend. The extension of the table that results from the translation of TA
consists of the translation of all entities that belong to both STUDENTS and
INSTRUCTORS.

3.4 Entity and Referential Integrity

If student course registrations are recorded using the structure of this database,
a tuple must be inserted into the table GRADES. Naturally, at the beginning
of the semester there is no way to enter a numerical grade; we need a special
value to enter in the field grade of the table GRADES that indicates that the
grade component of the tuple is not yet determined. Such a value is called a
null value. We represent this value by null.
A null value can have a significant semantic content: It may indicate that
a component of a tuple is not defined yet (as is the case with the previous
example), or that a certain attribute is inapplicable to a tuple, or that the
value of the component of the tuple is unknown. Unfortunately, it is not always
possible to tell which of these three intrepretations is intended whan a null value
is encountered. This can lead to some serious problems in practical situations.
3.4 Entity and Referential Integrity 39

Example 3.4.1 Suppose that we need to expand the table STUDENTS by

adding the relational attributes SAT and GRE. The first attribute is applicable
to undergraduates, while the second can be applied only to graduate students.
Therefore, every tuple that represents an undergraduate student has a null
component for the GRE attribute, and every tuple that represents a graduate
student has a null component for the SAT attribute.
Null values cannot be allowed to occur as tuple components corresponding to
the attributes of the primary key of a table, regardless of the semantic content
of a null value. Sometimes the primary key is used to compute the memory
address of a tuple. Therefore, the presence of null components in a tuple would
jeopardize the role of the primary key in the physical placement of the tuples in
memory. Also, such null values would interfere with the role of the primary key
of “representing” the tuple in its relationships with other data in the database.
This general requirement for relational databases is known as the entity integrity
rule.
Recall that every table that represents a set of relationships R contains refer-
ences to the sets of entities involved E1 , . . . , En . These references take the form
of the primary keys of E1 , . . . , En . For instance, the table GRADES contains
the attributes stno, empno, and cno, which are primary keys for STUDENTS,
INSTRUCTORS, and COURSES, respectively. It is natural to assume that the
student number stno component of a tuple of the table GRADES refers to the
student number component of a tuple that is actually in the table STUDENTS,
which is the place where student records are kept. This requirement (and similar
requirements involving references to the tables COURSES and INSTRUCTORS)
is formalized by the notion of referential integrity.
To define the concept of referential integrity, we need to introduce the notion
of a foreign key.
Definition 3.4.2 An S-foreign key for a table T with heading H is a set of
attributes L included in H that is the primary key for some other table S of the
relational database. We omit the mention of S when it is clear from context,
and we refer to a S-foreign key of a table T simply as a foreign key.
Although a foreign key in a table T must be a primary key of some table S
in the database, it may or may not be a part of the primary key of T . Suppose,
for instance, that the college database contains a table named ROOMS that lists
all the rooms of the campus. If the primary key of this table is roomno, then
this attribute is a ROOMS-foreign key for the INSTRUCTORS table.
The relational model has the following fundamental rule.
Referential Integrity Rule: If L is an S-foreign key for a table
T , only the following two cases may occur for each tuple t of T :
1. Either all components of t[L] are null, or
2. there is a tuple s in S such that t[L] = s[L].
This rule says that if a relationship refers to a row that could be in another
table, S, then that row must be present in S.
Example 3.4.3 Since roomno is a foreign key for INSTRUCTORS, any non-
40 The Relational Model

null value that occurs under this attribute in the table INSTRUCTORS must
appear in the table ROOMS. This corresponds to the real-world constraint that
either an instructor has no office, in which case the roomno-component is null,
or the instructor’s office in one of the rooms of the college.
Of course, if an S-foreign key is a part of the primary key of a table T (as is
the case with stno for GRADES, for example), then null values are not permitted
in T under the attributes of the S-foreign key.

3.5 Metadata
Metadata is a term that refers to data that describes other data. In the context
of the relational model, metadata are data that describe the tables and their
attributes.
The relational model allows a relational database to contain tables that
describe the database itself. These tables are known as catalog tables, and they
constitute the data catalog or the data dictionary of the database.
Typically, the catalog tables of a database include a table that describes the
names, owners, and some parameters of the headings of the data tables of the
database. The owner of a table is relevant in multi-user relational database
systems, where some users are permitted only limited access to tables they do
not own.
For example, a catalog table named SYSCATALOG that describes the tables
of the college database might look like:

SYSCATALOG
owner tname dbspacename
dsim courses system
dsim students system
dsim instructors system
dsim grades system
dsim advising system
sys syscolumns system
.. .. ..
. . .

In the table SYSCATALOG the attribute owner describes the creator of the
table; this coincides, in general, with the owner of that table. The attribute
tname gives the name of the table, while dbspacename indicates the memory
area (also known as the table space) where the table was placed.
Note that the above table mentions the table SYSCOLUMNS (recall that ta-
ble names are case insensitive). SYSCOLUMNS describes various attributes and
domains that occur in the user’s tables. For example, for the college database,
the table may look like:
3.6 Exercises 41

SYSCOLUMNS
owner cname tname coltype nulls length in pr key
dsim cno courses char N 5 Y
dsim cname courses char Y 20 N
dsim cr courses smallint Y 2 N
dsim cap courses integer Y 4 N
dsim stno grades char N 10 Y
dsim empno grades char N 11 N
dsim cno grades char N 5 Y
dsim sem grades char N 6 Y
dsim year grades integer N 4 Y
dsim grade grades integer Y 4 N
.. .. .. .. .. .. ..
. . . . . . .

The attributes cname and tname give the name of the column (attribute)
and the name of table where the attribute occurs. The nature of the domain
(character or numeric) is given by the attribute coltype and the size in bytes of
the values of the domain is given by the attribute length. The attribute nulls
specifies whether or not null values are allowed. Finally, the attribute in pr key
indicates whether the attribute belongs to the primary key of the table tname.
The access to and the presentation of metadata is highly dependent on the
specific database system. We examine the approach taken by ORACLE in
section 5.24.
The relational model currently dominates all database systems, and it is
likely to continue to do so for quite some time. Researchers are continually
producing enhancements to the model, adding, e.g., object-oriented and web-
centered features. Some of these features are already implemented in contem-
porary database systems, as we will see when we discuss ORACLE in detail.

3.6 Exercises
1. Convert the alternative E/R models for the college database discussed in
Exercise 1 of Chapter 2 to a relational design.
2. Convert the E/R design of the database of the customers of the natural
gas distribution company to a relational design. Specify the keys of each
relation.
3. Suppose that the set of entities E ′ is obtained by specialization from the
set of entities E and that
τ = (T, A1 . . . An , ρ),
τ ′ = (T ′ , A1 . . . An B1 . . . Bℓ , ρ′ )
are the tables that result from the translation of ρ and ρ′ , respectively.
Show that if e is an entity from E − E ′ and t is the tuple that results from
the translation of e, then t ∈ ρ − ρ′ [A1 . . . An ].
42 The Relational Model

3.7 Bibliographical Comments

The relational model was introduced by E. F. Codd in [Codd, 1970]. A revised
and extended version of the relational model is discussed in [Codd, 1990]. In-
teresting reflections on the relational model can be found in [Date and Darwen,
1993] and in [Date, 1990].
Chapter 4

Data Retrieval in the

Relational Model
4.1 Set Operations on Tables
4.2 The Basic Operations of Relational Algebra
4.4 Exercises
4.5 Bibliographical Comments

4.1 Introduction
Tables are more than simply places to store data. The real interest in tables is
in how they are used. To obtain information from a database, a user formulates
a question known as a “query.” For example, if we wanted to construct an
honor roll for the college for Fall 2002, we could examine the GRADES table and
select all students whose grades are above some threshold, say 90. Note that the
result can again be stored in a table. In this case, every tuple in the resultant
table actually appears in the original table. However, if we wanted to know the
names of the students in this table, we cannot find it out directly, as students
are represented only by their student numbers in the GRADES table. We have
to add some information from the STUDENTS table to find their names. The
result can again be stored in a table, which we can call HONOR ROLL.
In general, the method of working with relational databases is to modify and
combine tables using specific techniques. These techniques have been studied
and, of course, have names. For example, the method above that generates the
sub-table of GRADES is an example of a “selection.” This table can be thought
of as an “intermediate result” along the path of obtaining HONOR ROLL. The
method of combining this intermediate result with STUDENTS is known as
“joining.” These and various other methods are what we study under the name
“relational algebra.”
Relational algebra is thus a collection of methods for building new tables
starting from existing ones. These methods are referred to as “operations”
44 Data Retrieval in the Relational Model

on the tables. The interest in relational algebra is clear: Because a relational

database instance is a finite set of tables, and the answer to a query is again a
table, we need methods for constructing the tables corresponding to our queries.
Traditionally, relational algebra defines the minimal retrieval capabilities of
a relational database system. Thus, any system that purports to be a relational
database management system must provide retrieval capabilities that are at
least as powerful as the operations of relational algebra.
We introduce the operations of relational algebra one by one. For each, we
specify how the operation acts on the contents of the tables involved. However,
tables comprise more than just their contents, so to make the specification of
an operation complete, we must also specify the heading of the resultant table
and its name.

4.1.1 Renaming of Tables and Attributes

In building new tables, sometimes we need to create copies of existing tables.
Such a copy has the same extension (that is, contains the same tuples) as the
original table; the new copy must have a different name. In addition, for tech-
nical reasons, attributes of the new table may be different, provided each has
the same domain as the corresponding attribute of the original table.
Definition 4.1.1 Let T be a table having the heading A1 · · · An . The table
T ′ is obtained from T by renaming if T ′ 6= T , the heading of T ′ is B1 · · · Bn ,
where Dom Bi = Dom Ai , for 1 ≤ i ≤ n and the tables T and T ′ have the same
content.
We denote that T ′ was obtained from T through renaming by writing

T ′ (B1 , . . . , Bn ) := T (A1 , . . . , An ).

If B1 = A1 , . . . , Bn = An , then we may write T ′ := T . In this case we refer to

T ′ as an alias of T .
Example 4.1.2 Suppose that we need an alias of the COURSES table. If we
write

SUBJECTS := COURSES,

then we create the table:

SUBJECTS
cno cname cr cap
cs110 Introduction to Computing 4 120
cs210 Computer Programming 4 100
cs240 Computer Architecture 3 100
cs310 Data Structures 3 60
cs350 Higher Level Languages 3 50
cs410 Software Engineering 3 40
cs460 Graphics 3 30
4.1 Introduction 45

Figure 4.1: The Venn diagrams of set-theoretical operations

4.1.2 Set-Theoretical Operations

Since a table is essentially a set of tuples, it is natural to consider operations
similar to the usual set-theoretical operations.
The basic set-theoretical operations: union, intersection, and difference are
represented in Figure 4.1 using the well-known method of Venn diagrams.
In Figure 4.1 we show the intersection R ∩ S, the difference R − S, the
difference S − R, and the union R ∪ S of the sets R and S.
Unlike the set-theoretical case where the union, intersection, or difference of
any two sets exists, in relational algebra only certain tables may be involved in
these operations. The following definition introduces the necessary restriction,
that of being compatible.
Definition 4.1.3 Let T1 , T2 be two tables. The tables T1 , T2 are compatible if
they have the same headings.
46 Data Retrieval in the Relational Model

If R1 , R2 are extensions of two compatible tables T1 , T2 , respectively we say

that they are compatible relations. Otherwise, we say that the relations are
incompatible.
Example 4.1.4 The tables STUDENTS and INSTRUCTORS are incompatible
because
heading (STUDENTS) = stno name addr city state zip
heading (INSTRUCTORS) = empno name rank roomno telno.
It is not enough for the tables to have attributes in common; equality of the sets
of attributes is required for compatibility.
Now consider a table that contains courses offered by the college under a
continuing education program. Some of these courses are the same as the regular
courses; others are offered only by this program.
CED COURSES
cno cname cr cap
cs105 Computer Literacy 2 150
cs110 Introduction to Computing 4 120
cs199 Survey of Programming 3 120

The tables COURSES and CED COURSES are clearly compatible.

Definition 4.1.5 Let T1 , T2 be two compatible tables.
The union of T1 and T2 is the table (T1 ∪ T2 ) that contains the tuples that
belong to T1 or T2 .
The intersection of T1 and T2 is the table (T1 ∩ T2 ) that contains the tuples
that belong to both T1 and T2
The difference of T1 and T2 is the table (T1 − T2 ) that contains those rows
that belong to T1 but do not belong to T2 .
Note that the names of the tables that we define here have the form (T1 oper T2 ).
By this we mean that the name of the new table is a string that is the concate-
nation of the left parenthesis “(”, the name T1 , the symbol “oper”, the name
T2 , and the right parenthesis “)”, where the symbol “oper” can be “∪”, “∩”,
or “−”. Observe that these symbols must be added to the alphabet we use to
name tables. When there is no ambiguity, we simplify our notation by omitting
parentheses; e.g., we write T1 oper T2 for (T1 oper T2 ).
By an abuse of notation, renaming can also be used as an assignment to store
results obtained using relational algebra operations. Thus, we may write, for
instance, T ′ (B1 , B2 , B3 ) := T oper S, where the tables T, S have the attributes
A1 A2 and A2 A3 , respectively. This means that after performing the oper op-
eration, we rename the attributes A1 , A2 , A3 to B1 , B2 , B3 , and we rename the
resulting table T ′ .
As we introduce the operations of relational algebra, we use relational algebra
expressions informally to construct the names of the tables that we are about
to define.
Example 4.1.6 Consider the tables COURSES and CED COURSES introduced
in Example 4.1.4. If we need to determine all the courses offered by either the
4.1 Introduction 47

regular program or the continuing education division, then we compute the table
(COURSES ∪ CED COURSES):
(COURSES ∪ CED COURSES)
cno cname cr cap
cs105 Computer Literacy 2 150
cs110 Introduction to Computing 4 120
cs199 Survey of Programming 3 120
cs210 Computer Programming 4 100
cs240 Computer Architecture 3 100
cs310 Data Structures 3 60
cs350 Higher Level Languages 3 50
cs410 Software Engineering 3 40
cs460 Graphics 3 30
Courses offered under both the regular and the extension program are com-
puted in the table (COURSES ∩ CED COURSES):
(COURSES ∩ CED COURSES)
cno cname cr cap
cs110 Introduction to Computing 4 120

Finally, (COURSES − CED COURSES) contains courses offered by the regular

program but not by the continuing education division.
(COURSES − CED COURSES)
cno cname cr cap
cs210 Computer Programming 4 100
cs240 Computer Architecture 3 100
cs310 Data Structures 3 60
cs350 Higher Level Languages 3 50
cs410 Software Engineering 3 40
cs460 Graphics 3 30

Definition 4.1.7 Let T and S be two distinct tables. The product of T and
S is the table named (T × S) whose heading is T.A1 . . . T.An S.B1 . . . S.Bk and
which contains all tuples of the form

(u1 , . . . , un , v1 , . . . , vk ),

for every tuple (u1 , . . . , un ) of T and every (v1 , . . . , vk ) of S.

Example 4.1.8 The product of the tables
T
S
A B C
D E
a1 b1 c1
d1 e1
a2 b2 c4
d2 e1
a3 b1 c1
is the table
(T × S)
T.A T.B T.C S.D S.E
a1 b1 c1 d1 e1
a2 b2 c4 d1 e1
a3 b1 c1 d1 e1
a1 b1 c1 d2 e1
a2 b2 c4 d2 e1
a3 b1 c1 d2 e1
48 Data Retrieval in the Relational Model

In short, the product contains all possible combinations of the rows of the
original tables. So, we see that the product operation can create huge tables
starting from tables of modest size; for instance, the product of three tables of
1000 rows apiece yields a table with one billion tuples.
Note that the definition of the product of tables prevents us from consid-
ering the product of a table with itself. Indeed, if we were to try to con-
struct the product T × T , where the attributes of the new table would be
T.A1 , . . . , T.An , T.A1 , . . . , T.An . This contradicts the requirement that all at-
tributes of a table be distinct. To get around this restriction we create an alias
T ′ by writing T ′ := T ; then, we can compute (T ×T ′), which has T.A1 , . . . , T.An ,
T ′ .A1 , . . . , T ′ .An as its attributes.
Example 4.1.18 shows a query that requires this kind of special handling.

4.1.3 Selection
Selection is a unary operation (that is, an operation that applies to one table)
that allows us to select tuples that satisfy specified conditions. For instance,
using selection, we can extract the tuples that refer to all students who live in
Massachusetts from the STUDENTS table. To begin, we formalize the notion of
a condition.
Definition 4.1.9 Let H be a set of attributes. An atomic condition on H has
the form A oper a or A oper B, where A, B are attributes of H that have the
same domain, oper is one of =, !=, <, >, ≤, or ≥, and a is a value from the
domain of A.
As is common in query languages, we use != to represent 6=, because 6= is
not part of the ASCII character set and does not appear on most keyboards.
Example 4.1.10 Consider the table ITEMS that is a part of the database of a
department store and lists items sold by the store. We assume that the heading
consists of the following attributes:

H = itno iname dept cost retprice date.

The significance of the attributes of ITEMS is summarized below:

Attribute Meaning
itno item number
iname item name
dept store department
cost wholesale price
retprice retail price
date date when the retail
price was set
The following constructions
4.1 Introduction 49

dept = ’Sport’
cost > retprice
cost <= 1.25
are atomic conditions on the attributes of ITEMS. Note that we use quotation
marks for the value ’Sport’, because it is a part of a string domain, but there
are no quotation marks around 1.25, because this value belongs to a numerical
domain.
Starting from these atomic condtions, we can build more complicated con-
ditions using and, or, and not. So, if we want to list the sports items that
sell for under $ 1.25, we can use the condition dept = ’Sport’ and cost <= 1.25.
This method of building conditions is known as “recursive”, and we use it in
the following definition.
Definition 4.1.11 Conditions on a set of attributes H are defined recursively
as follows:
1. Every atomic condition on H is a condition on H.
2. If C1 , C2 are conditions on H, then

(C1 or C2 ), (C1 and C2 ), (not C1 )

are conditions on H.

It is common practice to omit parentheses when the expression is unambigu-

ous. This depends on a hierarchy of operations, where not has the higher prior-
ity, and by and and or are at the same, lower priority. Successive operations at
the same priority are associated from left-to-right. So, C1 and C2 or C3 and C4
is to be interpreted as (((C1 and C2 ) or C3 ) and C4 ).
Next, we define what it means for a tuple of a table T to satisfy a condition.
Definition 4.1.12 A tuple t satisfies an atomic condition on H, A oper a if
t[A] oper a; t satisfies the atomic condition A oper B if t[A] oper t[B].
A tuple t satisfies the condition (C1 and C2 ) if it satisfies both C1 and C2 ; t
satisfies the condition (C1 or C2 ) if it satisfies at least one of C1 and C2 . Finally,
t satisfies (not C1 ) if it fails to satisfy C1 .
To introduce the selection operation we add to the alphabet A the symbols
or, and and not; also, we add the relational attributes and the members of
their domains. Observe that a relational attribute that is written using several
letters (such as stno) is considered in this context to be a single symbol rather
than a sequence of several letters.
Definition 4.1.13 Let T be a table, and let C be a condition on H. The table
obtained by C-selection is the table (T where C) having the same heading H as
T , where the content of (T where C) consists of all tuples of T that satisfy the
condition C.
The next example shows how selection can be used to extract data from the
college database. Sometimes we show the table resulting from the operation (or
the succession of operations) that we intend to illustrate. In all such cases, we
assume that the college database is in the state shown in Figure 3.1.
50 Data Retrieval in the Relational Model

Example 4.1.14 To retrieve all students who live in Boston or in Brookline,

we write:
T1 := (STUDENTS where(city = ’Boston’ or city = ’Brookline’))
The corresponding table is:
T1
stno name addr city state zip
2661 Mixon Leatha 100 School St. Brookline MA 02146
2890 McLane Sandy 30 Cass Rd. Boston MA 02122
3566 Pierce Richard 70 Park St. Brookline MA 02146
4022 Prior Lorraine 8 Beacon St. Boston MA 02125
5544 Rawlings Jerry 15 Pleasant Dr. Boston MA 02115

Example 4.1.15 Let us find the list of grades given in CS110 during the spring
semester of 2002. This can be done by applying the following selection operation:
T := (GRADES wherecno = ’CS110’ and
sem = ’Spring’ and year = 2002).
This selection gives the table:
T
stno empno cno sem year grade
1011 023 cs110 Spring 2002 75
4022 023 cs110 Spring 2002 60

We conclude the definition of selection with the observation that selection ex-
tracts “horizontal” slices from a table. The next operation extracts vertical
slices from tables.

4.1.4 Projection
Recall that we introduced the projection of tables in Section 3.2. In this section
we re-examine this notion as a relational algebra operation.
A table may contain many attributes, but for any particular query, only
some of these may be relevant; projection allows us to chose these.
Example 4.1.16 Suppose that we wish to produce a list of instructors’ names
and the room numbers of their offices. This can be accomplished by projection:
OFFICE LIST := INSTRUCTORS[name roomno]
and we obtain the table:
OFFICE LIST
name roomno
Evans Robert 82
Exxon George 90
Sawyer Kathy 91
Davis William 72
Will Samuel 90

Example 4.1.17 Projection and selection may be combined, provided the pro-
jection does not eliminate the attributes used in the selection. Consider, for ex-
4.1 Introduction 51

ample, the task of determining the grades of the student whose student number
is 1011. The table T created by
T:=(GRADES wherestno=’1011’)[grade]
is
T
grade
40
75
90

Observe that duplicates are dropped through projection. Indeed, instead of

two grades of 90, the table shows only one. This happens because, as we pointed
out above, tables do not contain duplicate entries.
Example 4.1.18 Suppose that we need to find all pairs of instructors’ names
for instructors who share the same office. Of course, we need to compare the
office of every instructor with the office of every other instructor; we output the
names of instructors who have the same office. This query requires that we form
the product of the table INSTRUCTORS with an alias I of this table, as follows:
I := INSTRUCTORS
PROD := (INSTRUCTORS × I)
Next, we extract the pairs of instructors who have equal values for roomno. This
is accomplished using the selection:
(PROD whereINSTRUCTORS.roomno = I.roomno).
Note that this is not an entirely satisfactory solution. Indeed, we have no interest
in knowing that an instructor is in the same room as himself or herself; and,
once we know that instructor i1 is in the same room as instructor i2 it is clear
that i2 is in the same room as i1 . To eliminate this type of redundancy from
the answer we use a more restrictive selection:
(PROD where INSTRUCTORS.roomno = I.roomno
and INSTRUCTORS.empno < I.empno)
Finally, we extract the names of the instructors involved in the pairs retrieved
above:
(PROD where INSTRUCTORS.roomno = I.roomno
and INSTRUCTORS.empno < I.empno)
[INSTRUCTORS.name, I.name]

4.1.5 The Join Operation

The join operation is important for answering queries that combine data that
reside in several tables. To define the join operation between two tables, we first
introduce the join between two tuples.
Definition 4.1.19 Let T1 , T2 be two tables that have the headings

A1 · · · Am B1 · · · Bn and B1 · · · Bn C1 · · · Cp ,
52 Data Retrieval in the Relational Model

respectively. (In other words, assume that the two tables that have only the
attributes B1 , . . . , Bn in common.)
The tuples t1 in T1 and t2 in T2 are joinable if

t1 [B1 · · · Bn ] = t2 [B1 · · · Bn ].

If t1 and t2 are joinable tuples, their join is a tuple t defined on

A1 . . . Am B1 . . . Bn C1 . . . Cp

such that

t[A1 . . . Am B1 . . . Bn ] = t1 [A1 . . . Am B1 . . . Bn ],

and

t[B1 . . . Bn C1 . . . Cp ] = t2 [B1 . . . Bn C1 . . . Cp ].

The join of t1 and t2 is denoted by t1 1 t2 .

Note that in the above definition, if D is one of the attributes B1 , . . . , Bn
shared by the two tables, then t1 [D] = t2 [D] because t1 , t2 are joinable, so t[D]
can be defined correctly to be either t1 [D] or t2 [D].
Example 4.1.20 Let T1 , T2 be the tables given by:
T1
T2
A B D
B C D
t1 a2 b1 d1
u1 b1 c1 d1
t2 a1 b2 d4
u2 b2 c2 d4
t3 a3 b1 d1
u3 b3 c2 d1
t4 a3 b1 d2
u4 b2 c1 d2
t5 a1 b3 d3
The tuples t1 and u1 are joinable because t1 [BD] = u1 [BD] = (b1 d1 ); similarly,
t2 is joinable with u2 , t3 is joinable with u1 , and t4 and t5 are not joinable with
any tuple of S.
We have
t1 1 u1 = (a2 , b1 , d1 , c1 )
t2 1 u2 = (a1 , b2 , d4 , c2 )
t3 1 u 1 = (a3 , b1 , d1 , c1 )

Definition 4.1.21 Suppose that B1 , . . . , Bn are the attributes that two tables
T1 , T2 have in common.
The natural join of T1 and T2 , or simply the join, is the table named (T1 1
T2 ) having the heading A1 . . . Am B1 . . . Bn C1 . . . Cp that contains of all tuples
t1 1 t2 such that t1 is in T1 and t2 is in T2 , and t1 is joinable with t2 .
Note that if n = 0 (that is, if the tables T1 , T2 have no attributes in common),
then the joinability condition is satisfied by every tuple t1 of T1 and t2 of T2 .
In this special case, the tables T1 1 T2 and T1 × T2 are virtually identical: they
have the same rows but different names and headings.
4.1 Introduction 53

Example 4.1.22 The join T1 1 T2 of the tables considered in Example 4.1.20

is the table
T1 1 T2
A B D C
a2 b1 d1 c1
a1 b2 d4 c2
a3 b1 d1 c1

Example 4.1.23 Suppose that we need to find the names of all instructors
who have taught cs110. Initially, we extract all grade records involving cs110
using a selection operation:
T1 := (GRADES wherecno = ’cs110’).
Then, by joining T1 with the table INSTRUCTORS we extract the records of
instructors who teach this course:

T2 := (T1 1 INSTRUCTORS).

Finally, a projection on name yields the answer to the query:

ANS := T2 [name].

Example 4.1.24 To find the names of all instructors who have ever taught any
four-credit course, we can compute the join:

T1 := ((COURSES 1 GRADES) 1 INSTRUCTORS).

Then, by applying a selection we extract records corresponding to four-credit

courses:

T2 := (T1 where cr = 4).

The names of instructors are thus obtained by projection:

ANS := T2 [name].

An interesting variant of the previous example is given below:

Example 4.1.25 Let us determine the names of all instructors who have taught
any student who lives in Brookline. Observe that join cannot be used because
computing the join

((STUDENTS 1 GRADES) 1 INSTRUCTORS)

would require the name of the student to be identical with the name of the
instructor (which is, of course, not what is required by this query). Instead, we
54 Data Retrieval in the Relational Model

can use the product of tables and enforce the “limited joining” through selection:
T1 := (STUDENTS × GRADES × INSTRUCTORS)
T2 := T1 where STUDENTS.stno = GRADES.stno and
GRADES.empno = INSTRUCTORS.empno and
STUDENTS.city = ’Brookline’
Then, by projection, we extract the name of the instructors involved:

ANS := T2 [INSTRUCTORS.name].

Join can be used to express other operations. Note, for instance, that if T
and T ′ are two compatible tables, then T 1 T ′ has the same rows as T ∩ T ′ .
Indeed, since the two tables have all their attributes in common, two tuples t
in T and t′ in T ′ are joinable only if they are equal on all attributes, that is, if
they are the same.

4.1.6 Division
Definition 4.1.26 Let T1 , T2 be two tables such that the heading of T1 is
A1 . . . An B1 . . . Bk and the heading of T2 is B1 . . . Bk . The table obtained by
division of T1 by T2 is the table T1 ÷ T2 that has the heading A1 . . . An and
contains those tuples t in tupl(A1 . . . An ) such that t 1 t2 is a tuple in T1 for
every tuple t2 of T2 .
In other words, the content of the table obtained by dividing T1 by T2 ,
T1 ÷ T2 , consists of each tuple from tupl(A1 . . . An ) which, when concatenated
with every tuple of T2 , yields a tuple of T1 .
We stress that, in order for two tables T1 and T2 to be involved in a division,
the heading of T2 must be included in the heading of T1 .
Example 4.1.27 Suppose that we need to determine the courses taught by
all full professors. We can solve this query by first determining the employee
numbers (empno) for all full professors:

T1 := (INSTRUCTORS where rank = ’Professor’)[empno].

This generates the table:

T1
empno
019
023
Then, using projection, we discard all attributes from GRADES with the excep-
tion of cno and empno:

T2 := GRADES[cno, empno],

which results in
4.2 The Basic Operations of Relational Algebra 55

T2
empno cno
019 cs110
023 cs110
019 cs240
234 cs410
019 cs210
056 cs240
234 cs310

Finally, by applying division, we extract the course numbers of courses that are
taught by all full professors:

ANS := (T2 ÷ T1 ),

that is,
ANS
cno
cs110

It is essential to project GRADES on cno empno; otherwise, if we divide

GRADES by T1 , a tuple t = (s, c, m, y, g) is placed into GRADES ÷ T1 only if
the student s has taken the course c during the semester m of the year y and
has obtained the grade g from all full professors. Extracting the course number
afterwards does not help at all, since this requirement is both impossible to
satisfy and has nothing to do with our query.

4.2 The Basic Operations of Relational Algebra

So far, we have introduced nine operations: renaming, union, intersection, dif-
ference, product, selection, projection, join, and division. Now, we show that
certain operation can be expressed in terms of other operations. Our goal is to
build a list of “basic operations” that have the same computational capabilities
as the full set of operations previously introduced. In other words, for any table
created using the full set of operations, we can build a a table that has the same
content using the set of basic operations. Consequently, a relational database
systems need to implement only the basic operations and indeed, this is what
most of them do.
By convention, unary operations of relational algebra — that is, selection
and projection — have higher priority than the remaining binary operations.
Thus, for example, T1 oper T2 where C means T1 oper(T2 where C).
Let T1 and T2 be two compatible tables. It is easy to see that T1 ∩ T2 has the
same content as T1 − (T1 − T2 ). Thus intersection can be accomplished using
difference.
To see how the join operation can be expressed using the operations of
renaming, product, selection, and projection consider the following example.
Example 4.2.1 The tables T1 , T2 introduced in Example 4.1.20, have the head-
ings ABD and BCD, respectively. The table T3 := T1 × T2 is
56 Data Retrieval in the Relational Model

T3
T1 .A T1 .B T1 .D T2 .B T2 .C T2 .D
a2 b1 d1 b1 c1 d1
a1 b2 d4 b2 c2 d4
a3 b1 d1 b1 c1 d1

Then, we eliminate duplicate columns and rename the attributes in

T4 (A, B, D, C) := T3 [T1 .A, T1 .B, T2 .C, T2 .D].

T4
A B D C
a2 b1 d1 c1
a1 b2 d4 c2
a3 b1 d1 c1
The table T4 contains exactly the same tuples as the join T1 1 T2 .
Example 4.2.2 The query considered in Example 4.1.24 (where we use join to
find the names of instructors who have taught any four-credit course) can now
be solved using product, selection, projection, and renaming by the following
computation:

T1 = (COURSES × GRADES × INSTRUCTORS)

T2 = (T1 where COURSES.cr = 4 and
COURSES.cno = GRADES.cno and
GRADES.empno = INSTRUCTORS.empno)
T3 (name) = T2 [INSTRUCTORS.name]

The division operation can be expressed using renaming, product, selection,

projection, and difference. We illustrate how this can be accomplished by offer-
ing an alternative solution to the query in Example 4.1.27, listing the courses
taught by every full professor.
Example 4.2.3 Instead of directly finding the courses taught by all full pro-
fessors, we initially determine the courses that do not satisfy this condition. In
other words, in the first phase of the solution, we determine those courses that
are not taught by every full professor. Then, in the second phase, we eliminate
from the table GRADES[cno] (the list of courses that are actually taught) the
courses retrieved in the first phase.
We begin by forming a table T3 containing all pairs of course numbers and
employee numbers for full professors. This is accomplished by:
T1 := (INSTRUCTORS where
rank = ’Professor’)[empno]
T2 := GRADES[cno]
T3 (cno, empno) := (T2 × T1 )
4.2 The Basic Operations of Relational Algebra 57

The renaming of last step is required to replace the qualified attributes with
unqualified ones; i.e., we must consider all possible combinations (pairs) of pro-
fessors and courses, not just the courses that the various professors taught.
Next, by computing

T4 := (T3 − GRADES[cno, empno]),

we retain a pair (c, e) in T4 (where c is a course number and e is an employee

number) only if there is a full professor (whose employee number is e) who did
not teach the course that has course number c. Therefore, a course number
occurs in T5 := T4 [cno] only if there is a full professor who did not teach that
course. Consequently, courses taught by all full professors are the ones that
do not appear in T5 ; in other words, these courses can be found in the table
T6 = (COURSES[cno] − T5 ).
Starting from the database instance from Figure 3.1 we obtain the table T1
that contains all employee numbers for full professors:

T1
empno
019
023

T2 = GRADES[cno] contains all course numbers that are currently taught:

GRADES[cno]
cno
cs110
cs210
cs240
cs310
cs410

T3 gives all possible pairs of course numbers and employee numbers for full
professors:

T3
cno empno
cs110 019
cs110 023
cs210 019
cs210 023
cs240 019
cs240 023
cs310 019
cs310 023
cs410 019
cs410 023

The relation T4 := (T3 − GRADES[cno, empno]) is

58 Data Retrieval in the Relational Model

T4
cno empno
cs210 023
cs240 023
cs310 019
cs310 023
cs410 019
cs410 023
This means that the courses not taught by every full professor are:
T5
cno
cs210
cs240
cs310
cs410
Finally, the result of the computation is:
T6
cno
cs110

The same series of steps can always be used to calculate the division opera-
tion.
The arguments just presented show that only six of the nine operations of
relational algebra are required: renaming, union, difference, product, selection,
and projection.
It is natural to ask whether we can eliminate any of these remaining oper-
ations and still retain the full computational power of relational algebra. We
show, however, the set of six operations just mentioned is minimal. In other
words, if we discard any of these six operations, the remaining five are unable
to do the job of the discarded operation.
The following observation shows that the union operation cannot be dis-
carded.
Consider the one-attribute, one-tuple tables:
T1 T2
A and A
a1 a2
If we assume a1 6= a2 , then the table T1 ∪ T2 is given by
(T1 ∪ T2 )
A
a1
a2
Note that if we apply the operations of difference, product, selection, projection,
and renaming to tables that consist of at most one tuple, then the result may
contain at most one tuple; therefore, any computation that makes use only of
these operations is not capable of computing a target that contains more than
4.3 Other Relational Algebra Operations 59

one tuple, and therefore is unable to produce a table that has the same content
as T1 ∪ T2 .
The product operation cannot be eliminated from the set of basic operations.
Indeed, suppose that a database consists of two tables T and S that have the
one-attribute headings A and B, respectively. Observe that no computation
that uses renaming, union, difference, selection, and projection is capable of
computing the table (T1 × T2 ). Indeed, any table that is obtained through such
a computation may have only one attribute and, therefore, it cannot compute
T1 × T2 , which has two attributes.
Slightly more complicated examples show that the difference, selection, and
projection operations are all essential.

4.3 Other Relational Algebra Operations

We consider now three operations related to the natural join. In a join operation
tuples may be combined only if they have equal values on all columns they share
and they must have such values in all such columns. This can be awkward
in many situations; the operation we are about to introduce allows for more
flexibility.
Definition 4.3.1 Let T and T ′ be two tables such that their headings H, H ′ ,
respectively, have no common attributes.
Suppose that A1 , . . . , An are attributes of H and B1 , . . . , Bn are attributes
of H ′ such that Dom Ai = Dom Bi for 1 ≤ i ≤ n, and let θi be one of {=, ! =, <
, ≤, >, ≥} for 1 ≤ i ≤ n. Here, we use ! = to denote inequality.
If θ = (θ1 , . . . , θn ), the θ-join of τ and τ ′ is the table having the name T 1A1 θ1 B1 ,...,An θn Bn
T ′ , the heading HH ′ and the content ρA1 θ1 B1 ,...,An θn Bn , where ρA1 θ1 B1 ,...,An θn Bn
consists of those tuples u ∈ tupl(HH ′ ) for which there is t ∈ ρ and t ∈ ρ′ such
that

t[A] if A ∈ H
u[A] =
t′ [A] if A ∈ H ′ ,
and t[Ai ]θi t′ [Bi ] for 1 ≤ i ≤ n.
If θi is equality for all i, 1 ≤ i ≤ n, then we refer to the table
T 1A1 θ1 B1 ,...,An θn Bn T ′
as the equijoin of τ and τ ′ .
Example 4.3.2 Suppose we need to determine the pairs of student names and
instructor names such that the instructor is not an advisor for the student. In
order to deal with the requirement that the tables involved in a θ-join have
disjoint headings we create the tables:
ADVISING1(stno, empno1) := ADVISING,
and
INSTRUCTORS1(empno,name1) := INSTRUCTORS[empno,name].
60 Data Retrieval in the Relational Model

Since every student has one advisor it suffices to compute the θ-join:

T := (ADVISING1 1empno1!=empno INSTRUCTORS1)

Then, using natural join and projection we extract the answer:

ANS := (STUDENTS 1 T)[name,name1].

The semijoin ⋉ is another operation related to the join operation.

Definition 4.3.3 Let T1 , T2 be two tables having the headings H1 , H2 and the
contents ρ1 , ρ2 , respectively. Their semijoin is the table named T1 ⋉T2 that has
the heading H1 and the content ρ1 ⋉ρ2 , where ρ1 ⋉ρ2 = (ρ1 1 ρ2 )[H1 ].
Example 4.3.4 Let
τ1 = (T1 , ABD, ρ1 ),
τ2 = (T2 , BCD, ρ2 )
be the tables considered in Example 4.1.20. The semijoin τ1 ⋉τ2 is the table
(T1 ⋉T2 )
A B D
a2 b1 d1
a1 b2 d4
a3 b1 d1
The semijoin τ2 ⋉τ1 is:
(T2 ⋉T1 )
B D C
b1 d1 c1
b2 d4 c2

Clearly, we have in general ρ1 ⋉ρ2 6= ρ2 ⋉ρ1 .

The join operation is linked to semijoin by the identities:

ρ1 1 ρ2 = ρ1 1 (ρ2 ⋉ρ1 ) = ρ2 1 (ρ1 ⋉ρ2 ).

The semijoin of table τ1 with table τ2 computes that part of table τ1 that
consists of the tuples of τ1 that are joinable with tuples of τ2 ; in other words,
it computes the “useful” part of τ1 for the join with τ2 . This operation is
very important for distributed databases. In such databases various tables (or
even portions of tables) may reside at different computing sites, and it is often
important to minimize the amount of data traffic through the network that
connects these sites. Suppose, for example that τ1 is a very large table stored
at site S1 , τ2 is a relatively small table stored at site S2 and τ1 1 τ2 is needed
at site S2 (see Figure 4.2).
Suppose that the tuples of T1 and T2 have approximatively the same size.
Also, assume that T1 contains n1 tuples, T2 contains n2 tuples and k tuples of
T1 are joinable with the tuples of T2 . We need to compare two scenarios:
1) Ship table T1 to site S2 . The traffic cost is proportional to the size n1 of
T1 .
4.3 Other Relational Algebra Operations 61

Scenario 1:
τ1 τ2
1. Ship τ1 to S2

S1 S2

Scenario 2:
1. Ship τ2 to S1 .
2. Ship τ1 ⋉τ2 to S2 .
3. Compute τ1 1 τ2 at S2 as (τ1 ⋉τ2 ) 1 τ2

Figure 4.2: Computing a join in a two-site network

2) Ship T2 to site S1 , compute the semijoin T1 ⋉T2 at site S1 , ship the

semijoin to site S2 and compute the join T1 1 T2 using the semijoin. The cost
of the traffic is n2 +k. If this number is much smaller that n1 the second method
could be preferable.
Note that if ρ1 , ρ2 are two relations, then ρ1 − (ρ1 ⋉ρ2 ) is that part of ρ1
that consists of tuples of ρ1 that are not joinable with any tuples of ρ2 . This
observation is usful in defining the third operation that we introduce in this
section.
The tuples of a table T that is involved in a join with another table T ‘ and
are not joinable with any tuple of T ′ leave no trace in the join T 1 T ′ . By
contrast, in the operation we are about to define, all tuples, joinable or not,
participate in the final result.
Definition 4.3.5 Let T1 , T2 be two tables having th headings H1 , H2 and the
contents ρ1 , ρ2 , respectively.
The left outer join of τ1 and T2 is the table named T1 1l T2 having th
heading H1 ∪ H2 and the content ρ1 1ℓ ρ2 , where:
ρ1 1ℓ ρ2 = (ρ1 1 ρ2 ) ∪ {(a1 , . . . , an , null, . . . , null) |
(a1 , . . . , an ) ∈ ρ1 − (ρ1 ⋉ρ2 )}.
The right outer join of T1 and T2 is the table named T1 1r T2 whose heading
is H1 ∪ H2 , having the content ρ1 1r ρ2 , where
ρ1 1r ρ2 = (ρ1 1 ρ2 ) ∪ {(null, . . . , null, b1 , . . . , bp ) |
(b1 , . . . , bp ) ∈ ρ2 − (ρ2 ⋉ρ1 )}
The outer join of the tables T1 and T2 is the table named T1 1o T2 , whose
62 Data Retrieval in the Relational Model

heading is H1 ∪ H2 . The content of this table, ρ1 1o ρ2 is ρ1 1o ρ2 = (ρ1 1ℓ

ρ2 ) ∪ (ρ1 1r ρ2 ).
Example 4.3.6 Let (T1 , ABD, ρ1 ) and (T2 , BCD, ρ2 ) be the tables considered
in Example 4.3.4. The left outer join T1 1ℓ T2 is the table
(T1 1ℓ T2 )
A B D C
a2 b1 d1 c1
a1 b2 d4 c2
a3 b1 d1 c1
a3 b1 d2 null
a1 b3 d3 null
The right outer join T1 1r T2 is:
(T1 1r T2 )
A B D C
a2 b1 d1 c1
a1 b2 d4 c2
a3 b1 d1 c1
null b3 d1 c2
null b2 d2 c1
The outer join of these tables is:
(T1 1ℓ T2 )
A B D C
a2 b1 d1 c1
a1 b2 d4 c2
a3 b1 d1 c1
a3 b1 d2 null
a1 b3 d3 null
null b3 d1 c2
null b2 d2 c1

4.4 Exercises
1. Consider a database that consists of one table T :
T
A B
a b
Prove that there is no computation that uses renaming, union, product,
difference, and selection that can compute the projection T [A]. Conclude
that projection is an essential operation.

Solve the queries contained in Exercises 2–49 in relational algebra.

2. Find the names of students who live in Boston; find the names of students
who live outside Boston.
3. Find all pairs of student names and course names for grades obtained
during Fall of 2001.
4. Find the names of students who took some four-credit courses.
5. Find the names of students who took every four-credit course.
4.4 Exercises 63

6. Find the names of students who took only four-credit courses.

7. Find the names of students who took no four-credit courses.
8. Find the names of students who took a course with an instructor who is
also their advisor.
9. Find the names of students who took cs210 or cs310.
10. Find the names of students who took cs210 and cs310.
11. Find the names of all students who took neither cs210 nor cs310.
12. Find names of all students who took cs310 but never took cs210.
13. Find names of all students who have a cs210 grade higher than the higest
grade given in cs310 and did not take any course with Prof. Evans.
14. Find the names of all students whose advisor is not a full professor.
15. Find the names of students sho took cs210 or had Prof. Smith as their
advisor.
16. Find all pairs of names of students who live in the same city.
17. Find all triples of instructors’ names for instructors who taught the same
course.
18. Find instructors who taught students who are advised by another instruc-
tor who shares the same room.
19. Find course numbers for courses that enrol at least two students; solve the
same query for courses that enrol at least three students.
20. Find course numbers for courses that enrol exactly two students;
21. Find all pairs of students’ names for students who studied with the same
instructor.
22. Find the names of all students for whom no other student lives in the same
city.
23. Find the names of students who obtained the highest grade in cs210.
24. Find course numbers of courses taken by students who live in Boston and
which are taught by an associate professor.
25. Find the names of instructors who teach courses attended by students who
took a course with an instructor who is an assistant professor.
26. Find the telephone numbers of instructors who teach a course taken by
any student who lives in Boston.
27. Find the lowest grade of a student who took a course during the spring of
2003.
28. Find names of students who took every course taken by Richard Pierce.
29. Find the names for students such that if prof. Evans teaches a course,
then the student takes that course (although not necessarily with prof.
Evans).
30. Find all pairs of names of students and instructors such that the student
never took a course with the instructor.
31. Find the names of students who took only one course.
32. Find the names of students who took at least two courses.
33. Find names of courses taken by students who do not live in Massachusetts
(MA).
34. Find the names of instructors who teach no course.
35. Find course numbers of courses that have never been taught.
64 Data Retrieval in the Relational Model

36. Find course numbers of courses taken by students whose advisor is an

instructor who taught cs110.
37. Find the highest grade of a student who never took cs110.
38. Find courses that are taught by every assistant professor.
39. Find the names of the instructors who taught only one course during the
spring semester of 2001.
40. Find the names of students whose advisor did not teach them any course.
41. Find the names of students who have failed all their courses (failing is
defined as a grade less than 60).
42. Find the names of students who do not have an advisor.
43. Find course names of courses taught by more than one instructor.
44. Find the names of instructors who taught every semester when a student
from Rhode Island was enrolled.
45. Find course names of courses taken be every student advised by Prof.
Evans.
46. Find names of students who took every course taught be an instructor
who is advising at least two students.
47. Find names of instructors who teach every student they advise.
48. Find names of students who are taking every course taught by their advi-
sor.
49. Find course numbers of courses taken by every student who lives in Rhode
Island.
50. Consider a database that consists of one table T :
T
A
a1
a2

Prove that there is no computation that uses renaming, union, product,

difference, and projection that can compute the selection T where A = a1 .
Conclude that selection is an essential operation.
51. Prove that the product of tables can be expressed using renaming and
join. Conclude that there exist minimal sets of operations other than the
set of basic operations.

4.5 Bibliographical Comments

The original definition of relational algebra was given by E. F. Codd in [Codd,
1972a]. Relational algebra is presented in a rigorous manner in several sources [Fe-
jer and Simovici, 1991; Simovici and Tenney, 1995; Maier, 1983] and [Ullman,
1988a]. An excellent informal introduction can be found in [Date, 2003].