Functional Programming Using F#

This introduction to the principles of functional programming using F# shows how to apply
theoretical concepts to produce succinct and elegant programs. The book shows how
mainstream computer science problems can be solved using functional programming. It
introduces a model-based approach exploiting the rich type system of F#. It also demonstrates
the role of functional programming in a wide spectrum of applications including databases
and systems that engage in a dialogue with a user. Coverage also includes advanced features
in the .NET library, the imperative features of F# and topics such as sequences, computation
expressions and asynchronous computations.
With a broad spectrum of examples and exercises, the book is intended for courses in
functional programming as well as for self-study. Enhancing its use as a text is a website with
downloadable programs, lecture slides, mini-projects and links to further F# sources.

Michael R. Hansen is an Associate Professor in Computer Science at the Technical University

of Denmark. He is the author of Introduction to Programming Using SML (with Hans Rischel)
and Duration Calculus: A Formal Approach to Real-Time Systems (with Zhou Chaochen).

Hans Rischel is a former Associate Professor in Computer Science at the Technical

University of Denmark. He is the author of Introduction to Programming Using SML (with
Michael R. Hansen).
Functional Programming Using F#

MICHAEL R. HANSEN
Technical University of Denmark, Lyngby

HANS RISCHEL
Technical University of Denmark, Lyngby
 CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town,
Singapore, São Paulo, Delhi, Mexico City
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9781107684065
© Michael R. Hansen and Hans Rischel 2013
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2013
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
Hansen, Michael R., author.
Functional programming Using F# / Michael R. Hansen, Technical University of Denmark,
Lyngby, Hans Rischel, Technical University of Denmark, Lyngby.
pages cm
Includes bibliographical references and index.
ISBN 978-1-107-01902-7 (hardback) – ISBN 978-1-107-68406-5 (paperback)
1. Functional programming (Computer science) 2. F# (Computer program language) I.
Rischel, Hans, author. II. Title.
QA76.62.H37 2013
005.1′14–dc23 2012040414
ISBN 978-1-107-01902-7 Hardback
ISBN 978-1-107-68406-5 Paperback
Additional resources for this publication at http://www.cambridge.org/9781107019027
Cambridge University Press has no responsibility for the persistence or accuracy of URLs for
external or third-party Internet websites referred to in this publication and does not guarantee
that any content on such websites is, or will remain, accurate or appropriate.
 Contents

Preface
1 Getting started
1.1 Values, types, identifiers and declarations
1.2 Simple function declarations
1.3 Anonymous functions. Function expressions
1.4 Recursion
1.5 Pairs
1.6 Types and type checking
1.7 Bindings and environments
1.8 Euclid′s algorithm
1.9 Evaluations with environments
1.10 Free-standing programs
Summary
Exercises
2 Values, operators, expressions and functions
2.1 Numbers. Truth values. The unit type
2.2 Operator precedence and association
2.3 Characters and strings
2.4 If-then-else expressions
2.5 Overloaded functions and operators
2.6 Type inference
2.7 Functions are first-class citizens
2.8 Closures
2.9 Declaring prefix and infix operators
2.10 Equality and ordering
2.11 Function application operators |> and <|
2.12 Summary of the basic types
Summary
Exercises
3 Tuples, records and tagged values
 3.1 Tuples
3.2 Polymorphism
3.3 Example: Geometric vectors
3.4 Records
3.5 Example: Quadratic equations
3.6 Locally declared identifiers
3.7 Example: Rational numbers. Invariants
3.8 Tagged values. Constructors
3.9 Enumeration types
3.10 Exceptions
3.11 Partial functions. The option type
Summary
Exercises
4 Lists
4.1 The concept of a list
4.2 Construction and decomposition of lists
4.3 Typical recursions over lists
4.4 Polymorphism
4.5 The value restrictions on polymorphic expressions
4.6 Examples. A model-based approach
Summary
Exercises
5 Collections: Lists, maps and sets
5.1 Lists
5.2 Finite sets
5.3 Maps
Summary
Exercises
6 Finite trees
6.1 Chinese boxes
6.2 Symbolic differentiation
6.3 Binary trees. Parameterized types
6.4 Traversal of binary trees. Search trees
6.5 Expression trees
6.6 Trees with a variable number of sub-trees. Mutual recursion
6.7 Electrical circuits
Summary
Exercises
7 Modules
7.1 Abstractions
 7.2 Signature and implementation
7.3 Type augmentation. Operators in modules
7.4 Type extension
7.5 Classes and objects
7.6 Parameterized modules. Type variables in signatures
7.7 Customizing equality, hashing and the string function
7.8 Customizing ordering and indexing
7.9 Example: Piecewise linear plane curves
Summary
Exercises
8 Imperative features
8.1 Locations
8.2 Operators on locations
8.3 Default values
8.4 Sequential composition
8.5 Mutable record fields
8.6 References
8.7 While loops
8.8 Imperative functions on lists and other collections
8.9 Imperative tree traversal
8.10 Arrays
8.11 Imperative set and map
8.12 Functions on collections. Enumerator functions
8.13 Imperative queue
8.14 Restrictions on polymorphic expressions
Summary
Exercises
9 Efficiency
9.1 Resource measures
9.2 Memory management
9.3 Two problems
9.4 Solutions using accumulating parameters
9.5 Iterative function declarations
9.6 Tail recursion obtained using continuations
Summary
Exercises
10 Text processing programs
10.1 Keyword index example: Problem statement
10.2 Capturing data using regular expressions
10.3 Text I/O
10.4 File handling. Save and restore values in files
 10.5 Reserving, using and disposing resources
10.6 Culture-dependent information. String orderings
10.7 Conversion to textual form. Date and time
10.8 Keyword index example: The IndexGen program
10.9 Keyword index example: Analysis of a web-source
10.10 Keyword index example: Putting it all together
Summary
Exercises
11 Sequences
11.1 The sequence concept in F#
11.2 Some operations on sequences
11.3 Delays, recursion and side-effects
11.4 Example: Sieve of Eratosthenes
11.5 Limits of sequences: Newton-Raphson approximations
11.6 Sequence expressions
11.7 Specializations of sequences
11.8 Type providers and databases
Summary
Exercises
12 Computation expressions
12.1 The agenda when defining your own computations
12.2 Introducing computation expressions using sequence expressions
12.3 The basic functions: For and Yield
12.4 The technical setting when defining your own computations
12.5 Example: Expression evaluation with error handling
12.6 The basic functions: Bind, Return, ReturnFrom and Zero
12.7 Controlling the computations: Delay and Start
12.8 The basic function: Delay
12.9 The fundamental properties of For and Yield, Bind and Return
12.10 Monadic parsers
Summary
Exercises
13 Asynchronous and parallel computations
13.1 Multi-core processors, cache memories and main memory
13.2 Processes, threads and tasks
13.3 Challenges and pitfalls in concurrency
13.4 Asynchronous computations
13.5 Reactive programs
13.6 Parallel computations
Summary
Exercises
Appendix A Programs from the keyword example
A.1 Web source files
A.2 The IndexGen program
A.3 The NextLevelRefs program
Appendix B The TextProcessing library
Appendix C The dialogue program from Chapter
References
Index
 Preface

The purpose of this book is to introduce a wide range of readers – from the professional
programmer to the computer science student – to the rich world of functional programming
using the F# programming language. The book is intended as the textbook in a course on
functional programming and aims at showing the role of functional programming in a wide
spectrum of applications ranging from computer science examples over database examples to
systems that engage in a dialogue with a user.

Why functional programming using F#?

Functional programming languages have existed in academia for more than a quarter of a
century, starting with the untyped Lisp language, followed by strongly typed languages like
Haskell and Standard ML.
The penetration of functional languages to the software industry has, nevertheless, been
surprisingly slow. The reason is probably lack of support of functional languages by
commercial software development platforms, and software development managers are
reluctant to base software development on languages living in a non-commercial
environment.
This state of affairs has been changed completely by the appearance of F#, an open-source,
full-blown functional language integrated in the Visual Studio development platform and with
access to all features in the .NET program library. The language is also supported on Linux
and MAC systems using the Mono platform.

The background
The material in this book has been developed in connection with courses taught at the
Technical University of Denmark, originating from the textbook Introduction to
Programming Using SML by Hansen and Rischel (Addison-Wesley, 1999).
It has been an exciting experience for us to learn the many elegant and useful features of the
F# language, and this excitement is hopefully transferred to the reader of this book.

The chapters

Chapter 1: The basic concepts of F#, including values, types and recursive functions, are
introduced in a manner that allows readers to solve interesting problems from the start.
 Chapter 2: A thorough introduction to the basic types in F# is given, together with a
gentle introduction to the notion of higher-order functions.
Chapter 3: The simplest composite types of F#, tuples and records, are introduced. They
allow several values to be grouped together into one component. Furthermore, tagged
values are introduced.
Chapter 4: A list is a finite sequence of values with the same type. Standard recursions on
lists are studied and examples illustrating a model-based approach to functional
programming are given.
Chapter 5: The concepts of sets and maps are introduced and the powerful F# collection
libraries for lists, sets and maps are studied and applied in connection with a model-
based approach.
Chapter 6: The concept of finite tree is introduced and illustrated through a broad
selection of examples.
Chapter 7: It is shown how users can make their own libraries by means of modules
consisting of signature and implementation files. Furthermore, object-oriented features
of F# are mentioned.
Chapter 8: Imperative features of F# are introduced, including the array part of the
collection library and the imperative sets and maps from the .NET framework.
Chapter 9: The memory management concepts, stack, heap and garbage collection, are
described. Tail-recursive functions are introduced and two techniques for deriving such
functions are presented: one using accumulating parameters, the other continuations.
Their efficiency advantages are illustrated.
Chapter 10: A variety of facilities for processing text are introduced, including regular
expressions, file operations, web-based operations and culture-dependent string
ordering. The facilities are illustrated using a real-world example.
Chapter 11: A sequence is a, possibly infinite, collection of elements that are computed
on-demand only. Sequence functions are expressed using library functions or sequence
expressions that provide a step-by-step method for generating elements. Database tables
are viewed as sequences (using a type provider) and operations on databases are
expressed using query expressions.
Chapter 12: The notion of computation expression, which is based on the theory of
monads, is studied and used to hide low-level details of a computation from its
definition. Monadic parsing is used as a major example to illustrate the techniques.
Chapter 13: This last chapter describes how to construct asynchronous reactive
programs, spending most of their time awaiting a request or a response from an external
agent, and parallel programs, exploiting the multi-core processor of the computer.

The first six chapters cover a standard curriculum in functional programming, while the other
chapters cover more advanced topics.

Further material
The book contains a large number of exercises, and further material is available at the book′s
homepage. A link to this homepage is found at:
http://www.cambridge.org/9781107019027
This material includes a complete set of slides for a course in functional programming plus a
collection of problems and descriptions of topics to be used in student projects.

Acknowledgments
Special thanks go to Peter Sestoft, Don Syme and Anh-Dung Phan. The idea to make a
textbook on functional programming on the basis of F# originates from Peter, who patiently
commented on the manuscript during its production and helped with advice and suggestions.
From the very start of this project we had the support of Don. This is strongly appreciated and
so is the help, clarifications and constructive comments that we received throughout this
project. Phan helped with many comments, suggestions and insights about the platform. We
are grateful for this help, for many discussions and for careful comments on all the chapters.
Furthermore, we are grateful to Nils Andersen, Mary E. Böker, Diego Colombo and Niels
Hallenberg for reading and commenting on the complete manuscript.
Earlier versions of this manuscript have been used in connection with courses at the
Technical University of Denmark and the IT-University of Copenhagen. The comments we
received from the students in these courses are greatly appreciated.

Lyngby, July 31, 2012

Michael R. Hansen and Hans
Rischel
 1
Getting started

In this chapter we will introduce some of the main concepts of functional programming
languages. In particular we will introduce the concepts of value, expression, declaration,
recursive function and type. Furthermore, to explain the meaning of programs we will
introduce the notions: binding, environment and evaluation of expressions.
The purpose of the chapter is to acquaint the reader with these concepts, in order to address
interesting problems from the very beginning. The reader will obtain a thorough knowledge
of these concepts and skills in applying them as we elaborate on them throughout this book.
There is support of both compilation of F# programs to executable code and the execution
of programs in an interactive mode. The programs in this book are usually illustrated by the
use of the interactive mode.
The interface of the interactive F# compiler is very advanced as, for example, structured
values like tuples, lists, trees and functions can be communicated directly between the user and
the system without any conversions. Thus, it is very easy to experiment with programs and
program designs and this allows us to focus on the main structures of programs and program
designs, that is, the core of programming, as input and output of structured values can be
handled by the F# system.
 1.1 Values, types, identifiers and declarations
In this section we illustrate how to use an F# system in interactive mode.
The interactive interface allows the user to enter, for example, an arithmetic expression in a
line, followed by two semicolons and terminated by pressing the return key:
2*3 + 4;;

The answer from the system contains the value and the type of the expression:
val it : int = 10

The system will add some leading characters in the input line to make a distinction between
input from the user and output from the system. The dialogue may look as follows:
> 2*3 + 4;;
val it : int = 10
>

The leading string “>” is output whenever this particular system is awaiting input from the
user. It is called the prompt, as it “prompts” for input from the user. The input from the user is
ended by a double semicolon “;;” while the next line contains the answer from the system.
In the following we will distinguish between user input and answer from the system by the
use of different type fonts:
2*3 + 4;;
val it : int = 10

The input from the user is written in typewriter font while the answer from the system is
written in italic typewriter font.
The above answer starts with the reserved word val, which indicates that a value has been
computed, while the special identifier it is a name for the computed value, that is, 10. The
type of the result is int, denoting the subset of the integers { …, −2, −1, 0, 1, 2, … } that can
be represented using the system.
The user can give a name to a value by entering a declaration, for instance:
let price = 125;;

where the reserved word let starts the declarations. In this case the system answers:
val price : int = 125

The identifier price is now a name for the integer value 125. We also say that the identifier
price is bound to 125.
Identifiers which are bound to values can be used in expressions:
price * 20;;
val it : int = 2500

The identifier it is now bound to the integer value 2500, and this identifier can also be used
in expressions:
it / price = 20;;
val it : bool = true

The operator / is the quotient operator on integers. The expression it/price = 20 is a

question to the system and the identifier it is now bound to the answer true of type bool,
where bool is a type denoting the two-element set { true, false } of truth values. Note that the
equality sign in the input is part of an expression of type bool, whereas the equality sign in the
answer expresses a binding of the identifier it to a value.
 1.2 Simple function declarations
We now consider the declaration of functions. One can name a function, just as one can name
an integer constant. As an example, we want to compute the area of a circle with given radius
r, using the well known area function: circleArea(r) = π r2.

The constant π is found in the Library under the name System.Math.PI:

System.Math.PI;;
val it : float = 3.141592654

The type float denotes the subset of the real numbers that can be represented in the system,
and System.Math.PI is bound to a value of this type.
We choose the name circleArea for the circle area function, and the function is then
declared using a let-declaration:
let circleArea r = System.Math.PI * r * r;;
val circleArea : float -> float

The answer says that the identifier circleArea now denotes a value, as indicated by the
reserved word val occurring in the answer. This value is a function with the type float ->
float, where the symbol -> indicates a function type and the argument as well as the value of
the function has type float. Thus, the answer says that circleArea is bound to a value that is
some function of type float -> float.
The function circleArea can be applied to different arguments. These arguments must
have the type float, and the result has type float too:
circleArea 1.0;;
val it : float = 3.141592654

circleArea (2.0);;
val it : float = 12.56637061

Brackets around the argument 1.0 or (2.0) are optional, as indicated here.
The identifier System.Math.PI is a composite identifier. The identifier System denotes a
namespace where the identifier Math is defined, and System.Math denotes a namespace where
the identifier PI is defined. Furthermore, System and System.Math denote parts of the .NET
Library. We encourage the reader to use program libraries whenever appropriate. In Chapter
7 we describe how to make your own program libraries.
 Comments
A string enclosed within a matching pair (* and *) is a comment which is ignored by the F#
system. Comments can be used to make programs more readable for a human reader by
explaining the intention of the program, for example:
(* Area of circle with radius r *)
let circleArea r = System.Math.PI * r * r;;
val circleArea : float -> float

Two slash characters // can be used for one-line comments:

// Area of circle with radius r
let circleArea r = System.Math.PI * r * r;;
val circleArea : float -> float

A comment line can also begin with three slash characters ///. The tool XMLDocs can
produce program documentation from such comment, but we will not pursue this any further
in this book.
Comments can be very useful, especially in large programs, but long comments should be
avoided as they tend to make it more difficult for the reader to get an overview of the
program.
 1.3 Anonymous functions. Function expressions
A function can be created in F# without getting any name. This is done by evaluating a
function expression, that is an expression where the value is a function. This section introduces
simple function expressions and function expressions with patterns.
A nameless, anonymous function can be defined by a simple function expression, also called a
lambda expression,1 for example:
fun r -> System.Math.PI * r * r;;
val it : float -> float = <fun:clo@10-1>
it 2.0;;
val it : float = 12.56637061

The expression fun r -> System.Math.PI * r * r denotes the circle-area function and
it reads: “the function of r given by π · r2”. The reserved word fun indicates that a function is
defined, the identifier r occurring to the left of -> is a pattern for the argument of the
function, and System.Math.PI * r * r is the expression for the value of the function.
The declaration of the circle-area function could be made as follows:
let circleArea = fun r -> System.Math.PI * r * r;;
val circleArea : float -> float

but it is more natural in this case to use a let-declaration let circleArea r = … with an
argument pattern. We shall later see many uses of anonymous functions.
 Function expressions with patterns
It is often convenient to define a function in terms of a number of cases. Consider, for
example, a function giving the number of days in a month, where a month is given by its
number, that is, an integer between 1 and 12. Suppose that the year of consideration is not a
leap year. This function can thus be expressed as:
function
| 1 -> 31 // January
| 2 -> 28 // February
| 3 -> 31 // March
| 4 -> 30 // April
| 5 -> 31 // May
| 6 -> 30 // June
| 7 -> 31 // July
| 8 -> 31 // August
| 9 -> 30 // September
| 10 -> 31 // October
| 11 -> 30 // November
| 12 -> 31;;// December
function
^
stdin(17,1): warning FS0025: Incomplete pattern matches on
this expression. For example, the value ′0′ may indicate a
case not covered by the pattern(s).
val it : int -> int = <fun:clo@17-2>

The last part of the answer shows that the computed value, named by it, is a function with the
type int -> int, that is, a function from integers to integers. The answer also shows the
internal name for that function. The first part of the answer is a warning that the set of patterns
used in the function-expression is incomplete. The expression enumerates a value for every
legal number for a month (1, 2, …, 12). At this moment we do not care about other numbers.
The function can be applied to 2 to find the number of days in February:
it 2;;
val it : int = 28

This function can be expressed more compactly using a wildcard pattern “_”:
function
| 2 -> 28 // February
| 4 -> 30 // April
| 6 -> 30 // June
| 9 -> 30 // September
| 11 -> 30 // November
| _ -> 31;;// All other months

In this case, the function is defined using six clauses. The first clause 2 -> 28 consists of a
pattern 2 and a corresponding expression 28. The next four clauses have a similar
explanation, and the last clause contains a wildcard pattern. Applying the function to a value v,
the system finds the clause containing the first pattern that matches v, and returns the value of
the corresponding expression. In this example there are just two kinds of matches we should
know:

A constant, like 2, matches itself only, and

the wildcard pattern _ matches any value.

For example, applying the function to 4 gives 30, and applying it to 7 gives 31.
An even more succinct definition can be given using an or-pattern:
function
| 2 -> 28 // February
| 4|6|9|11 -> 30 // April, June, September, November
| _ -> 31 // All other months
;;

The or-pattern 4|6|9|11 matches any of the values 4,6,9,11, and no other values.
We shall make extensive use of such a case splitting in the definition of functions, also
when declaring named functions:
let daysOfMonth = function
| 2 -> 28 // February
| 4|6|9|11 -> 30 // April, June, September, November
| _ -> 31 // All other months
;;
val daysOfMonth : int -> int

daysOfMonth 3;;
val it : int = 31

daysOfMonth 9;;
val it : int = 30
 1.4 Recursion
This section introduces the concept of recursion formula and recursive declaration of
functions by an example: the factorial function n!. It is defined by:

where n is a non-negative integer. The dots … indicate that all integers from 1 to n should be
multiplied. For example:
 Recursion formula
The underbraced part of the below expression for n! is the expression for (n-1)!:

so we get the formula:

This formula is actually correct also for n = 1 as:

so we get:

This formula is called a recursion formula for the factorial function (_!) as it expresses the
value of the function for some argument n in terms of the value of the function for some
other argument (here: n-1).
 Computations
This definition has a form that can be used in the computation of values of the function. For
example:

The clauses of the definition of the factorial function are applied in a purely “mechanical”
way in the above computation of 4!. We will now take a closer look at this mechanical process
as the system will compute function values in a similar manner:
 Substitution in clauses
The first step is obtained from Clause 2, by substituting 4 for n. The condition for using the
second clause is satisfied as 4 > 0. This step can be written in more detail as:
 Computation of arguments
The new argument (4−1) of the factorial function in the expression (4−1)! is computed in the
next step:

Thus, the principles used in the first two steps of the computation of 4! are:

Substitute a value for n in Clause 2.

Compute argument.

These are the only principles used in the above computation until we arrive at the expression:

The next computation step is obtained by using Clause 1 to obtain a value of 0!:

and the multiplications are then performed in the last step:

This recursion formula for the factorial function is an example of a general pattern that
will appear over and over again throughout the book. It contains a clause for a base case “0!”,
and it contains a clause where a more general case “ n!” is reduced to an expression “ n · (n −
1)!” involving a “smaller” instance “(n − 1)!” of the function being characterized. For such
recursion formulas, the computation process will terminate, that is, the computation of n! will
terminate for all n ≥ 0.
 Recursive declaration
We name the factorial function fact, and this function is then declared as follows:
let rec fact = function
| 0 -> 1
| n -> n * fact(n-1);;
val fact : int -> int

This declaration corresponds to the recursion formula for n!. The reserved word rec
occurring in the let-declaration allows the identifier being declared (fact in this case) to
occur in the defining expression.
This declaration consists of two clauses

each initiated by a vertical bar. The pattern of the first clause is the constant 0, while the
pattern of the second clause is the identifier n.
The patterns are matched with integer arguments during the evaluation of function values
as we shall see below. The only value matching the pattern 0 is 0. On the other hand, every
value matches the pattern n, as an identifier can name any value.
 Evaluation
The system uses the declaration of fact to evaluate function values in a way that resembles
the above computation of 4!.

Substitution in clauses
To evaluate fact 4, the system searches for a clause in the declaration of fact, where 4
matches the pattern of the clause.
The system starts with the first clause of the declaration: 0 -> 1. This clause is skipped as
the value 4 does not match the pattern 0 of this clause.
Then, the second clause: n -> n * fact(n - 1) is investigated. The value 4 matches the
pattern of this clause, that is, the identifier n. The value 4 is bound to n and then substituted for
n in the right-hand side of this clause thereby obtaining the expression: 4 * fact(4 - 1).
We say that the expression fact 4 evaluates to 4 * fact(4-1) and this evaluation is
written as:

where we use the symbol for a step in the evaluation of an expression. Note that the symbol
is not part of any program, but a symbol used in explaining the evaluation of expressions.

Evaluation of arguments
The next step in the evaluation is to evaluate the argument 4 - 1 of fact:

The evaluation of the expression fact 4 proceeds until a value is reached:

The argument values 4, 3, 2 and 1 do not match the pattern 0 in the first clause of the
declaration of fact, but they match the second pattern n. Thus, the second clause is chosen for
further evaluation in the evaluation steps (1), (3), (5) and (7).
 The argument value 0 does, however, match the pattern 0, so the first clause is chosen for
further evaluation in step (9). The steps (2), (4), (6) and (8) evaluate argument values to fact,
while the last steps (10) - (13) reduce the expression built in the previous steps.
 Unsuccessful evaluations
The evaluation of fact n may not evaluate to a value, because

the system will run out of memory due to long expressions,

the evaluation may involve bigger integers than the system can handle, or
the evaluation of an expression may not terminate.2

For example, applying fact to a negative integer leads to an infinite evaluation:

 A remark on recursion formulas
The above recursive function declaration was motivated by the recursion formula:

which gives a unique characterization of the factorial function.

The factorial function may, however, be characterized by other recursion formulas, for
example:

This formula is not well-suited for computations of values, because the corresponding
function declaration based on this formula (where / denotes integer division):
let rec f = function
| 0 -> 1
| n -> f(n+1)/(n+1);;
val f : int -> int

gives an infinite evaluation of f k when k > 0. For example:

Thus, in finding a declaration of a function, one has to look for a suitable recursion formula
expressing the computation of function values. This declaration of f contains a base case “f
0”. However, the second clause does not reduce the general case “f(n)” to an instance which is
closer to the base case, and the evaluation of f(n) will not terminate when n > 0.
 14.5 Pairs
Consider the function:

where x is a real number and n is a natural number.

The under-braced part of the expression below for xn is the expression for xn− 1:

Using the convention: x0 = 1, the function can be characterized by the recursion formula:

In mathematics xn is a function of two variables x and n, but it is treated differently in F#

using the concept of a pair:

If a1 and a2 are values of types 1 and 2 then (a1,a2) is a value of type 1* 2

For example:
let a = (2.0,3);;
val a = (2.0, 3) : float * int

Furthermore, given patterns pat1 and pat2 there is a composite pattern (pat1,pat2). It
matches a pair (a1,a2) exactly when pat1 matches a1 and pat2 matches a2, for example:
let (x,y) = a;;
val y : int = 3
val x : float = 2.0

The concept of a pair is a special case of tuples that are treated in Section 3.1.
Using these concepts we represent xn as a function power with a pair (x,n) as the argument.
The following declaration is based on the above recursion formula, using composite patterns
(x,0) and (x,n):

let rec power = function

| (x,0) -> 1.0 // (1)
| (x,n) -> x * power(x,n-1);; // (2)
val power : float * int -> float

The type of power is float * int -> float. The argument of power is therefore a pair of
type float * int while the value of the function is of type float.
The power function can be applied to pairs of type float * int:
power a;;
val it : float = 8.0
power(4.0,2);;
val it : float = 16.0

A function in F# has one argument and one value. In this case the argument is a pair (u,i) of
type float * int, while the value of the function is of type float.
The system evaluates the expression power(4.0,2) as follows:
 Notes on pattern matching
Note that the order of the clauses in the declaration of power is significant. The following
declaration will not work:
let rec powerNo = function
| (x, n) -> x * powerNo(x,n-1) // This does NOT work
| (x, 0) -> 1.0
;;

The first pattern (x,n) will match any pair of form (u, i) and the second clause will
consequently never come into use. The F# compiler actually discovers this and issues a
warning:
| (x, 0) -> 1.0
—-^^^^^^
... warning FS0026: This rule will never be matched

The function can be applied to an argument (despite the warning), but that would give an
infinite evaluation since the base case (x, 0) -> 1.0 is never reached.
A similar remark on the order of clauses applies to the declaration of fact.
One should also note that a prior binding of an identifier used in a pattern has no effect on the
pattern matching.3 Hence, the following will also not work:
let zero = 0;;

let rec powerNo = function

| (x,zero) -> 1.0 // This does NOT work
| (x,n) -> x * powerNo(x,n-1)
;;

The first pattern (x,zero) will match any pair of form (u, i), binding x to u and zero to i so
the second clause will again never come into use. The F# compiler issues a warning like in
the previous example.
 1.6 Types and type checking
The examples in the previous sections show that types like float * int -> float or int
form an integral part of the responses from the system.
In fact, F# will try to infer a type for each value, expression and declaration entered. If the
system can infer a type for the input, then the input is accepted by the system. Otherwise the
system will reject the input with an error message.
For example, the expression circleArea 2.0 is accepted, because

circleArea has the type float -> float, and

2.0 has the type float.

Furthermore, the result of evaluating circleArea 2.0, that is 12.5663706144, has type
float.
On the other hand, the system will reject the expression circleArea 2 with an error
message since 2 has type int while the argument for circleArea must be of type float:
circleArea 2;;
circleArea 2;;
———–^

stdin(95,12): error FS0001: This expression was expected to

have type
float
but here has type
int

The above type consideration for function application f(e) is a special case of the general
type rule for function application:

if f has type 1 −> 2 and e has type 1 then f(e) has type 2.

Using the notation e : τ to assert that the expression e has type τ, this rule can be presented
more succinctly as follows:

if f : 1 −> 2 and e : 1 then f(e) : 2.

Consider, for example, the function power with type float * int -> float. In this case,
τ1 is float * int and τ2 is float. Furthermore, the pair (4.0,2) has type float * int
(which is τ1). According to the above rule, the expression power(4.0,2) hence has type float
(which is τ2).
 1.7 Bindings and environments
In the previous sections we have seen that identifiers can be bound to denote an integer, a
floating-point value, a pair or a function. The notions of binding and environment are used to
explain that entities are bound by identifiers.
The execution of a declaration, say let x = e, causes the identifier x to be bound to the value
of the expression e. For example, the execution of the declaration:
let a = 3;;
val a : int = 3

causes the identifier a to be bound to 3. This binding is denoted by a 3.

Execution of further declarations gives extra bindings. For example, execution of
let b = 7.0;;
val b : float = 7.0

gives a further binding b 7.0.

A collection of bindings is called an environment, and the environment env1 obtained from
execution of the above two declarations is denoted by:

Note that this notation is not part of any program. Bindings and environments are
mathematical objects used to explain the meaning of programs.
The execution of an additional declaration causes an extension of env1. For example
let c = (2, 8);;
val c : int * int = (2, 8)

let circleArea r = System.Math.PI * r * r;;

val circleArea : float -> float

adds bindings of the identifiers c and circleArea to the environment env1 giving the
environment env2:

The value of an expression is always evaluated in the actual environment, that contains the
bindings of identifiers that are valid at evaluation time. When the F# system is activated, the
actual environment is the Basis Environment that gives meanings to /, +, -, sqrt, for
example. When using environments we will usually not show bindings from the Basis
Environment. We will usually also omit bindings of identifiers like System.Math.PI from the
Library.
 1.8 Euclid’s algorithm
This section presents the famous algorithm of Euclid for computing the greatest common
divisor of two natural numbers.
For a given integer n, an integer d is called a divisor of n (written d | n) if there exists an
integer q such that n = q · d. Hence, the number 1 is a divisor of any integer. Any integer n ≠ 0
has a finite number of divisors as each divisor has absolute value ≤|n|, while 0 has infinitely
many divisors as any integer is a divisor of 0. Thus, integers m, n have at least one common
divisor (namely 1), and if either m≠ 0 or n≠ 0, then the set of common divisors of m and n is
finite.
The GCD theorem of Euclid states that for any integers m, n there exists an integer gcd(m, n)
such that gcd(m, n) ≥ 0, and such that the common divisors of m and n are precisely the divisors
of gcd(m, n).
Note that if m≠ 0 or n≠ 0 then gcd(m, n) is the greatest common divisor of m and n. For m =
0 and n = 0 we have gcd(0,0) = 0, as the common divisors for 0 and 0 are precisely the
divisors of 0, but 0 and 0 have no greatest common divisor as any number is a divisor of 0.
Euclid gave an algorithm for computing gcd(m, n) for arbitrary integers m and n and this
algorithm gives at the same time a proof of the theorem.
 Division with remainder. The / and % operators
Euclid’s algorithm is based on the concept of integer division with remainder. Let m and n be
integers with m ≠ 0. An identity with integers q and r of the form:

is then called a division with quotient q and remainder r. There are infinite many possible
remainders (corresponding to different quotients q):

It follows that there are two possibilities concerning remainders r with −|m| < r < |m|:

1. The integer 0 is a remainder and any other remainder r satisfies |r| ≥ |m|.
2. There are two remainders rneg and rpos such that − |m| < rneg < 0 < rpos < |m|.

The F# operators / and % (quotient and remainder) are defined (for m ≠ 0) such that:
(1.1)
(1.2)
(1.3)
(1.4)
so n % m = 0 when m is a divisor of n, otherwise rpos is used if n > 0 and rneg if n < 0.
Note that the corresponding operators in other programming languages may use different
conventions for negative integers.
 Euclid’s algorithm in F#
Euclid’s algorithm is now expressed in the following declaration
let rec gcd = function
| (0,n) -> n
| (m,n) -> gcd(n
val gcd : int * int -> int

For example:
gcd(12,27);;
val it : int = 3

gcd(36, 116);;
val it : int = 4
 Termination of Euclid’s algorithm
It is not obvious that the evaluation of gcd(m, n) will terminate with a result for all integers m
and n. We will now prove that the second clause in the declaration is in fact used at most |m|
times in the evaluation of gcd(m, n). It follows that the evaluation always terminates.
Consider an evaluation with at least k (> 0) steps using the second clause:

The right-hand side of the second clause gives the identities:

Using |n % m| < |m| when m ≠ 0, we get:

It follows that

because |m1|, |m2|, …, |mk| are k mutually different integers among the |m| integers |m|−1, |m|−2,
…, 1, 0.
The evaluation of gcd(m,n) will hence involve at most |m| uses of the second clause.
 Proof of Euclid’s theorem
The key to prove Euclid’s theorem is that the following holds when m ≠ 0:
The integers n % m and m have the same common divisors as the integers n and m
This follows from the identities:

which show that any common divisor of (n % m) and m is also a divisor of n, and hence a
common divisor of n and m – and conversely – any common divisor of n and m is also a
divisor of (n % m), and hence a common divisor of (n % m) and m.
Using the above integers n1, n2, … and m1, m2,… we hence get:

m1 and n1 have same common divisors as m and n

m2 and n2 have same common divisors as m1 and n1
… … …
mp and np have same common divisors as mp−1 and np−1

where the evaluation terminates with an index p where mp = 0 and np = gcd( m, n).

The common divisors of 0 and np are, however, exactly the divisors of np = gcd( m, n) as
any integer is a divisor of 0. It follows by induction that the common divisors of m and n are
exactly the divisors of gcd( m, n).
 1.9 Evaluations with environments
During the evaluation of expressions the system may create and use temporary bindings of
identifiers. This is, for example, the case for function applications like gcd(36,116) where
the function gcd is applied to the argument (36,116). We will study such bindings as it gives
insight into how recursive functions are evaluated.
The declaration:
let rec gcd = function
| (0,n) -> n
| (m,n) -> gcd(n
val gcd : int * int -> int

contains two clauses: One with pattern (0,n) and expression n and another with pattern (m,n)
and expression gcd(n % m,m). There are hence two cases in the evaluation of an expression
gcd(x,y) corresponding to the two clauses:

1. gcd(0,y): The argument (0,y) matches the pattern (0,n) in the first clause giving the
binding n y, and the system will evaluate the corresponding right-hand side expression
n using this binding:

2. gcd(x,y) with x ≠ 0: The argument (x,y) does not match the pattern (0,n) in the first
clause but it matches the pattern (m,n) in the second clause giving the bindings m x, n
y, and the system will evaluate the corresponding right-hand side expression gcd(n %
m,m) using these bindings:

Consider, for example, the expression gcd(36,116). The value (36,116) does not
match the pattern (0,n), so the first evaluation step is based on the second clause:

The expression gcd(n % m, m) will then be further evaluated using the bindings for m
and n. The next evaluation steps evaluate the argument expression (n % m, m) using the
bindings:

The evaluation continues evaluating the expression gcd(8,36) and this proceeds in the
same way, but with different values bound to m and n:
The evaluation will in the same way reduce the expression gcd(4,8) to gcd(0,4), but the
evaluation of gcd(0,4) will use the first clause in the declaration of gcd, and the
evaluation terminates with result 4:

Note that different bindings for m and n occur in this evaluation and that all these
bindings have disappeared when the result of the evaluation (that is, 4) is reached.
 1.10 Free-standing programs
A free-standing program contains a main function of type:
string[] -> int

preceded by the entry point attribute:

...
[<EntryPoint>]
let main (param: string[]) =
...

The type string[] is an array type (cf. Section 8.10) and the argument param consists of k
strings (cf. Section 2.3):

The following is a simple, free-standing “hello world” program:

open System;;
[<EntryPoint>]
let main(param: string[]) =
printf "Hello %s\n" param.[0]
0;;

It uses the printf function (cf. Section 10.7) to make some output. The zero result signals
normal termination of the program. The program source file Hello.fsx compiles to an exe-
file using the F# batch compiler:
fsc Hello.fsx -o Hello.exe

and the program can now be called from a command prompt:

>Hello Peter
Hello Peter

>Hello "Sue and Allan"

Hello Sue and Allan

Using the fsc command requires that the directory path of the F# compiler (with file name
fsc.exe or Fsc.exe) is included in the PATH environment variable.
 Summary
The main purpose of this chapter is to familiarize the reader with some of the main concepts
of F# to an extent where she/he can start experimenting with the system. To this end, we have
introduced the F# notions of values, expressions, types and declarations, including recursive
function declarations.
The main concepts needed to explain the meaning of these notions are: integers and
floating-point numbers, bindings and environments, and step by step evaluation of
expressions.
 Exercise
1.1 Declare a function g: int −> int, where g(n) = n + 4.
1.2 Declare a function h: float * float -> float, where . Hint: Use the
function System.Math.Sqrt.
1.3 Write function expressions corresponding to the functions g and h in the exercises 1.1
and 1.2.
1.4 Declare a recursive function f: int -> int, where

for n ≥ 0. (Hint: use two clauses with 0 and n as patterns.)

State the recursion formula corresponding to the declaration.
Give an evaluation for f(4).
1.5 The sequence F0, F1, F2, … of Fibonacci numbers is defined by:

Thus, the first members of the sequence are 0,1,1,2,3,5,8,13, ….

Declare an F# function to compute Fn. Use a declaration with three clauses, where the
patterns correspond to the three cases of the above definition.
Give an evaluations for F4.
1.6 Declare a recursive function sum: int * int -> int, where

for m ≥ 0 and n ≥ 0. (Hint: use two clauses with (m,0) and (m,n) as patterns.)
Give the recursion formula corresponding to the declaration.
1.7 Determine a type for each of the expressions:
(System.Math.PI, fact -1)
fact(fact 4)
power(System.Math.PI, fact 2)
(power, fact)

1.8 Consider the declarations:

let a = 5;;
let f a = a + 1;;
let g b = (f b) + a;;

Find the environment obtained from these declarations and write the evaluations of the
expressions f 3 and g 3.
___________________
1 Lambda calculus was introduced by Alonzo Church in the 1930s. In this calculus an
expression of the form λ x.e was used to denote the function of x given by the expression e.
The fun-notation in F# is a direct translation from λ-expressions.
2 Note that a text like fact n is not part of F#. It is a schema where one can obtain a program
piece by replacing the meta symbol n with a suitable F# entity. In the following we will often
use such schemas containing meta symbols in italic font.
3 Identifiers that are constructors are, however, treated in a special way (cf. Section 3.8).
 2
Values, operators, expressions and functions

The purpose of this chapter is to illustrate the use of values of basic types: numbers,
characters, truth values and strings by means of some examples. The concepts of operator
overloading and type inference are explained. Furthermore, the chapter contains a gentle
introduction to higher-order functions. It is explained how to declare operators, and the
concepts of equality and ordering in F# are introduced. After reading the chapter the reader
should be able to construct simple programs using numbers, characters, strings and truth
values.
 2.1 Numbers. Truth values. The unit type
From mathematics we know the set of natural numbers as a subset of the set of integers, which
again is a subset of the rational numbers (i.e., fractions), and so on. In F#, however, the set of
values with the type: int, for the integers, is considered to be disjoint from the set of values
with the type: float, for floating-point numbers, that is, the part of the real numbers that are
representable in the computer. The reason is that the encodings of integer and float values in
the computer are different, and that the computer has different machine instructions for
adding integer values and for adding float values, for example.
A value of type int is written as a sequence of digits possibly prefixed with the minus sign
“-”. Real numbers are written using decimal point notation or using exponential notation, or
using both:
0;;
val it : int = 0

0.0;;
val it : float = 0.0

0123;;
val it : int = 123

-7.235;;
val it : float = -7.235

-388890;;
val it : int = -388890

1.23e-17;;
val it : float = 1.23e-17

where 1.23e-17 denotes 1.23 · 10-17.

 Operators
We will use the term operator as a synonym for function and the components of the
argument of an operator will be called operands. Furthermore, a monadic operator is an
operator with one operand, while a dyadic operator has two operands. Most monadic
operators are used in prefix notation where the operator is written in front of the operand.
Examples of operators on numbers are monadic minus -, and the dyadic operators addition
+, subtraction -, multiplication * and division /. Furthermore, the relations: =, <> (denoting
inequality ≠), >, > = (denoting ≥), < and <= (denoting ≤), between numbers are considered to
be operators on numbers computing a truth value.
The symbol “−” is used for three purposes in F# as in mathematics. In number constants
like “-2” it denotes the sign of the constant, in expressions like “- 2” and “-(2+1)” it denotes
an application of the monadic minus operator, and in the expression “1-2” it denotes the
dyadic subtraction operator.
Consider, as a strange example:
2 - - -1;;
val it : int = 1

Starting from the right, -1 denotes the the integer “minus one ”, the expression - -1 denotes
monadic minus applied to minus one, and the full expression denotes the dyadic operation two
minus one.
Division is not defined on integers, but we have instead the operators / for quotient and % for
remainder as described on Page 15, for example:
13 / -5;;
val it : int = -2

13 % -5;;
val it : int = 3
 Truth values
There are two values true and false of the type bool:
true;;
val it : bool = true

false;;
val it : bool = false

Functions can have truth values as results. Consider, for example, a function even
determining whether an integer n is even (i.e., n % 2 = 0). This function can be declared as
follows:
let even n = n % 2 = 0;;
val even : int -> bool

A truth-valued function such as even is called a predicate.

Logical operators
not (unary) negation
&& logical and (conjunction)
|| logical or (disjunction)

Table 2.1 Operators on truth values

Functions on truth values are often called logical operators, and some of the main ones are
shown in Table 2.1. The negation operator not applies to truth values, and the comparison
operators = and <> are defined for truth values. For example:
not true <> false;;
val it : bool = false

Furthermore, there are expressions e1 || e2 and e1 && e2 corresponding to the disjunction

and conjunction operators of propositional logic. The expression e1 || e2 is true if either e1
or e2 (or both) are true; otherwise the expression is false. The expression e1 && e2 is true if
both e1 and e2 are true; otherwise the expression is false.
Evaluations of e1 || e2 and e1 && e2 will only evaluate the expression e2 when needed, that
is, the expression e2 in e1 || e2 is not evaluated if e1 evaluates to true, and the expression e2
in e1 && e2 is not evaluated if e1 evaluates to false. For example:
1 = 2 && fact -1 = 0;;
val it : bool = false

Thus, 1 = 2 && fact -1 = 0 evaluates to false without attempting to evaluate the

expression fact -1 = 0, which would result in a non-terminating evaluation.
 The unit type
There is only one value, written (), of type unit. It is mentioned here as it belongs to the
basic types in F#. It is used in the imperative part of F# as a “dummy” result of a computation
consisting solely of side-effects like input-output or modification of mutable data. There are
no operators on the value () of type unit.
 2.2 Operator precedence and association
The monadic operator - is written in front of the argument (like other function names), while
the dyadic operators are written in infix notation, where the operator is placed between the
operands. Table 2.2 shows the arithmetic operators.

+ unary plus
− unary minus
+ addition
− subtraction
* multiplication
/ division
% modulo (remainder)
** exponentiation

Table 2.2 Arithmetic operators

Usual rules for omitting brackets in mathematical expressions also apply to F# expressions.
These rules are governed by two concepts: operator precedence and operator association for
dyadic operators as shown in Table 2.3. The operators occurring in the same row have same
precedence, which is higher than that of operators occurring in succeeding rows. For
example, * and / have the same precedence. This precedence is higher than that of +.

Operator Association
** Associates to the right
* / % Associates to the left
+ − Associates to the left
= <> > >= < <= No association
&& Associates to the left
|| Associates to the left

Table 2.3Operator precedence and association

Furthermore, a monadic operator (including function application) has higher precedence

than any dyadic operator. The idea is that higher (larger) precedence means earlier evaluation.
For example:
and

The dyadic operators for numbers and truth values (except **) associate to the left, which
means that operators of the same precedence are applied starting from the left, so the
evaluation of an expression will proceed as if the expression was fully bracketed. For
example:
 2.3 Characters and strings
A character is a letter, a digit or a special character (i.e., a punctuation symbol like comma or
semicolon or a control character). Characters are encoded in the computer as integer values
using the Unicode alphabet, which is an international standard for encoding characters.
A character value is written as the character c enclosed in apostrophes. Examples of values
of type char are:
′a′;;
val it : char = ′a′

′ ′;;
val it : char = ′ ′

where the last one denotes the space character.

The new line, apostrophe, quote and backslash characters are written by means of the escape
sequences shown in Table 2.4. Functions on characters are found in the System.Char library.

Sequence Meaning
\′ Apostrophe
\" Quote
\\ Backslash
\b Backspace
\n Newline
\r Carriage return
\t Horizontal tab

Table 2.4 Character escape sequences

The operators ||, && and not are convenient when declaring functions with results of type
bool, like in the following declarations of the functions isLowerCaseConsonant and
isLowerCaseVowel determining whether a character is a lower-case consonant or vowel,
respectively:
let isLowerCaseVowel ch =
ch=′a′ || ch=′e′ || ch=′i′ || ch=′o′ || ch=′u′;;
val isLowerCaseVowel : char -> bool
let isLowerCaseConsonant ch =
System.Char.IsLower ch && not (isLowerCaseVowel ch);;
val isLowerCaseConsonant : char -> bool

isLowerCaseVowel ′i′ && not (isLowerCaseConsonant ′i′);;

val it : bool = true
isLowerCaseVowel ′I′ || isLowerCaseConsonant ′I′;;
val it : bool = false
not (isLowerCaseVowel ′z′) && isLowerCaseConsonant ′z′;;
val it : bool = true

where we use the function IsLower from the library System.Char to check whether ch is a
lower-case letter. This library contains predicates IsDigit, IsSeparator, and so on,
expressing properties of a character.
 Strings
A string is a sequence of characters. Strings are values of the type string. A string is written
inside enclosing quotes that are not part of the string. Quote, backslash or control characters
in a string are written by using the escape sequences. Comments cannot occur inside strings as
comment brackets ((* or *)) inside a string simply are interpreted as parts of the string.
Examples of values of type string are:
"abcd—";;
val it : string = "abcd—"

"\"1234\"";;
val it : string = "\"1234\""
"";;

val it : string = ""

The first one denotes the 7-character string “abcd—”, the second uses escape sequences to get
the 6-character string “"1234"” including the quotes, while the last denotes the empty string
containing no characters.
Strings can also be written using the verbatim string notation where the character @ is
placed in front of the first quote:
@"c0 c1 … cn−1 "

It denotes the string of characters c0 c1 … cn−1 without any conversion of escape sequences.
Hence @"\\\\" denotes a string of four backslash characters:
@"\\\\";;
val it : string = "\\\\"

while the escape sequence \\ for backslash is converted in the string "\\\\":

"\\\\";;
val it : string = "\\"

Verbatim strings are useful when making strings containing backslash characters. Note that it
is not possible to make a verbatim string containing a quote character because \" is
interpreted as a backslash character followed by the terminating quote character.
 Functions on strings
The String library contains a variety of functions on strings. In this section we will just
illustrate the use of a few of them by some examples.
The length function computes the number of characters in a string:
String.length "1234";;
val it : int = 4

String.length "\"1234\"";;
val it : int = 6

String.length "";; // size of the empty string

val it : int = 0

The concatenation function + joins two strings together forming a new string by placing the
two strings one after another. The operator + is used in infix mode:
let text = "abcd—";;
val text : string = "abcd—"

text + text;;
val it: string = "abcd—abcd—"

text + " " = text;;

val it : bool = false

text + "" = text;;

val it : bool = true

"" + text = text;;

val it : bool = true

The last two examples show that the empty string is the neutral element for concatenation
of strings just like the number 0 is the neutral element for addition of integers.
Note that the same operator symbol + is used for integer addition and string concatenation.
This overloading of operator symbols is treated in Section 2.5.
A string s with length n is given by a sequence of n characters s = “c0 c1 ... cn−1”, where the
convention in F# is that the numbering starts at 0. For any such string s there is a function,
written s.[i], to extract the i’th character in s for 0 ≤ i ≤ n−1. The integer i used in s.[i] is called
an index. For example:
"abc".[0];;
val it : char = ′a′
"abc".[2];;
val it : char = ′c′

"abc".[3];;
System.IndexOutOfRangeException: ...
Stopped due to error
where the last example shows (a part of) the error message which will occur when the index is
out of bounds.
If we want to concatenate a string and a character, we need to use the string function to
convert the character to a string, for example
"abc" + string ′d′;;
val it : string = "abcd"

as the operator + in this case denotes string concatenation, and this operator cannot
concatenate a string with a character.
Conversion of integer, real or Boolean values to their string representations are done by
using the function string, for example:
string -4;;
val it : string = "-4"
string 7.89;;
val it : string = "7.89"
string true;;
val it : string = "True"

A simple application of this conversion function is the declaration of the function nameAge:
let nameAge(name,age) =
name + " is " + (string age) + " years old";;

It converts the integer value of the age to the corresponding string of digits and builds a string
containing the string for the name and the age. For example:
nameAge("Diana",15+4);;
val it : string = "Diana is 19 years old"
nameAge("Philip",1-4);;
val it : string = "Philip is -3 years old"

The string function can actually give a string representation of every value, including
values belonging to user-defined types. We shall return to this in Section 7.7. Examples of
string representations are:
string (12, ′a′);;
val it : string = "(12, a)"
string nameAge;;
val it : string = "FSI0022+it@29-4"

where the pair (12, ′a′) has a natural string representation in contrast to that of the user-
defined nameAge function.
 2.4 If-then-else expressions
An if-then-else expression has form:
if exp1 then exp2 else exp3

where exp1 is an expression of type bool while exp2 and exp3 are expressions of the same
type. The if-then-else expression is evaluated by first evaluating exp1. If exp1 evaluates to
true then the expression exp2 is evaluated; otherwise, if exp1 evaluates to false then the
expression exp3 is evaluated. Note that at most one of the expressions exp2 and exp3 will be
evaluated (none of them will be evaluated if the evaluation of exp1 does not terminate).
An if-then-else expression is used whenever one has to express a splitting into cases that
cannot be expressed conveniently by use of patterns. As an example we may declare a function
on strings that adjusts a string to even size by putting a space character in front of the string if
the size is odd. Using the function even on Page 23 and if-then-else for the splitting into
cases gives the following declaration:
let even n = n % 2 = 0;;
val even : int -> bool
let adjString s = if even(String.length s)
then s else " " + s;;
val adjString : string -> string
adjString "123";;
val it : string = " 123"
adjString "1234";;
val it : string = "1234"

One may, of course, use an if-then-else expression instead of splitting into clauses by
pattern matching. But pattern matching is to be preferred, as illustrated by the following (less
readable) alternative declaration of the gcd function (cf. Page 16):
let rec gcd(m,n) = if m=0 then n
else gcd(n % m,m);;
val gcd : int * int -> int

One should also avoid expressions of the forms:

if e1 then true else e2
if e1 then e2 else false

for defining Boolean combinations of expressions and instead use the shorter, equivalent
forms:
e1 || e2
e1 && e2
 2.5 Overloaded functions and operators
A name or symbol for a function or operator is overloaded if it has different meanings
when applied to arguments or operands of different types. We have already seen that the plus
operator + denote addition for integers but concatenation for strings.
A (mathematical) function on real numbers is considered different from the corresponding
function on integers, as they are implemented in F# by different machine instructions. An
operator of this kind is hence overloaded: it denotes different functions depending on the
context, and it depends on the types of the operands whether, for example, the operator *
denotes multiplication on integers (of type int) or multiplication on real numbers (of type
float). The F# system tries to resolve these ambiguities in the following way:

If the type can be inferred from the context, then an overloaded operator symbol is
interpreted as denoting the function on the inferred type.
If the type cannot be inferred from the context, then an overloaded operator symbol with
a default type will default to this type. The default type is int if the operator can be
applied to integers.

For example, the obvious declaration of a squaring function yields the function on integers:
let square x = x * x;;
val square : int -> int

Declaring a squaring function on reals can be done either by specifying the type of the
argument:
let square (x:float) = x * x;;
val square : float -> float

or by specifying the type of the result:

let square x : float = x * x;;
val square : float -> float

or by specifying the type of the expression for the function value:

let square x = x * x : float;;
val square : float -> float

or by choosing any mixture of the above possibilities.

abs, acos, atan, atan2, ceil, cos, cosh, exp, floor, log
log10, pow, pown, round, sin, sinh, sqrt, tan, tanh

Table 2.5 Mathematical functions

 There are many overloaded operators in F#, in particular mathematical functions that can
be applied to integers as well as to real numbers. Some of them can be found in Table 2.5. The
function abs, for example, computes the absolute value of a number that can be of type int,
float or any of the number types in Table 2.6, for example, float32:

abs -1;;
val it : int = 1

abs -1.0;;
val it : float = 1.0

abs -3.2f;;
val it : float32 = 3.20000000f

Overloading is extensively used in the .NET library and typing of arguments is frequently
needed to resolve ambiguities. The user may declare overloaded operators and functions
inside a type declaration as explained in Section 7.3.
 2.6 Type inference
When an expression is entered, the F# system will try to determine a unique type using so-
called type inference. If this does not succeed then the expression is not accepted and an error
message is issued.
Consider once more the declaration of the function power (cf. Section 1.5):
let rec power = function
| (x, 0) -> 1.0 (* 1 *)
| (x, n) -> x * power(x,n-1) (* 2 *);;
val power : float * int -> float

The F# system deduces that power has the type: float * int -> float. We can see how F# is
able to infer this type of power by arguing as follows:

1. The keyword function indicates that the type of power is a function type τ -> τ′, for
some types τ and τ′.
2. Since power is applied to a pair (x, n) in the declaration, the type τ must have the form
τ1 * τ2 for some types τ1 and τ2.
3. We have τ2 = int, since the pattern of the first clause is (x,0), and 0 has type int.
4. We have that τ′ = float, since the expression for the function value in the first clause: 1.0
has type float.
5. We know that power(x,n-1) has the type float since τ′ = float. Thus, the overloaded
operator symbol * in x * power(x,n-1) resolves to float multiplication and x must be
of type float. We hence get τ1 = float.

The above declaration of the power function has been used for illustrating the declaration
of recursive functions and the type inference performed by the system. As described above
there is already a power operator ** in F# and this should of course be used in programs. In
general we recommend to inspect the F# and .NET libraries and use available library
functions when appropriate.
 2.7 Functions are first-class citizens
In functional languages, and F# is no exception, functions are what is called first-class
citizens. An implication of this is that a function can be argument of another function and that
the value of a function can again be a function. In this section we shall give a first, gentle
introduction to this concept, which also is known as higher-order functions.
 The value of a function can be a function
As a first example we shall consider the infix operator +. There is a version of this operator
that is not written between the operands. This non-fix version is written (+), and we shall now
study its type:
(+);;
val it : (int -> int -> int) = <fun:it@1>

The type operator “->” associates to the right, so (+) has the type:
(+) : int -> (int -> int)

This type shows that the value of the function (+) is another function with type int -> int.
Applying (+) to an integer n thus gives a function:
(+) n: int -> int

For example:
let plusThree = (+) 3;;
val plusThree : (int -> int)
plusThree 5;;
val it : int = 8
plusThree -7;;
val it : int = -4

The sum of two integers m and n can be computed as ((+) m) n. The brackets can be omitted
because function application associates to the left. For example:
(+) 1 3;;
val it : int = 4
 The argument of a function can be a function
Function composition f g is defined in mathematics by: (f g)(x) = f(g(x)). This operator on
functions is well-defined when domains and ranges of f and g match:
If f : A → B and g : C → A, then f g : C → B
For example, if f(y) = y + 3 and g(x) = x2, then (f g)(z) = z2 + 3.
We want to construe the function composition as a function, and this function will obviously
take functions as arguments. There is actually an infix operator << in F# denoting function
composition, and the above example can hence be paraphrased as follows:
let f = fun y -> y+3;; // f(y) = y+3
val f : int -> int
let g = fun x -> x*x;; // g(x) = x*x
val g : int -> int
let h = f << g;; // h = (f o g)
val h : int -> int
h 4;; // h(4) = (f o g)(4)
val it : int = 19

Using function expressions instead of named functions f, g and h, the example looks as
follows:
((fun y -> y+3) << (fun x -> x*x)) 4;;
val it : int = 19
 Declaration of higher-order functions
So far we have seen higher-order built-in functions like (+) and (<<). We shall now illustrate
ways to declare such functions by means of a simple example.
Suppose that we have a cube with side length s, containing a liquid with density ρ. The
weight of the liquid is then given by ρ · s3. If the unit of measure of ρ is kg/m3 and the unit of
measure of s is m then the unit of measure of the weight will be kg.
Consider the following declaration of the weight function:
let weight ro = fun s -> ro * s ** 3.0;;
val weight : float -> float -> float

where we use the operator ** to compute xy for floating-point numbers x and y. A function
value weight ρ is again a function as the expression on the right-hand side of the declaration
is a fun-expression. This property of the function value is also visible in the type of weight.
We can make partial evaluations of the function weight to define functions for computing
the weight of a cube of either water or methanol (having the densities 1000 kg/ m3 and 786.5
kg/ m3 respectively under “normal” pressure and temperature):
let waterWeight = weight 1000.0;;
val waterWeight : (float -> float)

waterWeight 1.0;;
val it : float = 1000.0

waterWeight 2.0;;
val it : float = 8000.0

let methanolWeight = weight 786.5;;

val methanolWeight : (float -> float)

methanolWeight 1.0;;
val it : float = 786.5

methanolWeight 2.0;;
val it : float = 6292.0

Higher-order functions may alternatively be defined by supplying the arguments as follows

in the let-declaration:
let weight ro s = ro * s ** 3.0;;
val weight : float -> float -> float

and this is normally the preferred way of defining higher-order functions.

 2.8 Closures
A closure gives the means of explaining a value that is a function. A closure is a triple:

where x is an argument identifier, exp is the expression to evaluate to get a function value,
while env is an environment (cf. Section 1.7) giving bindings to be used in such an evaluation.
Consider as an example the evaluation of weight 786.5 in the previous example. The result
is the closure:

The environment contains bindings of all identifiers in the expression ro*s**3.0 except the
argument s.
Note that a closure is a value in F# – functions are first-class citizens.
The following simple example illustrates the role of the environment in the closure:
let pi = System.Math.PI;;
let circleArea r = pi * r * r;;
val circleArea : float -> float

These declarations bind the identifier pi to a float value and circleArea to a closure:

A fresh binding of pi does not affect the meaning of circleArea that uses the binding of pi
in the closure:
let pi = 0;;
circleArea 1.0;;
val it : float = 3.141592654

This feature of F# is called static binding of identifers occurring in functions.

 2.9 Declaring prefix and infix operators
Expressions containing functions on pairs can often be given a more readable form by using
infix notation where the function symbol is written between the two components of the
argument. Infix form is used for the dyadic arithmetic operators +, -, %, /, for example. This
allows us to makepressions more readable by use of rules for omission of brackets. For
example: x - y - z means (x - y) - z and x + y * z means x + (y * z). These rules for omitting
brackets are governed by precedence and association for the operators (see Section 2.2).
Operators are written using special character strings which cannot be used as “normal”
identifiers. Infix operators are sequences of the following symbols1
! % & * + - . / < = > ? @ ^ |

while prefix operators are one of

+ - +. -. & && % %%
~ ~~ ~~~ ~~~~ (tilde characters)

The bracket notation converts from infix or prefix operator to (prefix) function:

The corresponding (prefix) function for an infix operator op is denoted by (op).

The corresponding (prefix) function for a prefix operator op is denoted by (~op).

An infix operator is declared using the bracket notation as in the following declaration of
an infix exclusive-or operator .||. on truth values:
let (.||.) p q = (p || q) && not(p && q);;
val ( .||. ) : bool -> bool -> bool
(1 > 2) .||. (2 + 3 < 5);;
val it : bool = false

The system determines the precedence and association of declared operators on the basis of
the characters in the operator. In the case of .||. the periods have no influence on this, so the
precedence and association of .||. will be the same as those of ||. Therefore,
true .||. false && true;;

is equivalent to
true .||. (false && true);;

as && has higher precedence than || and .||..

A prefix operator is declared using a leading tilde character. We may, for example, declare
a prefix operator %% to calculate the reciprocal value of a float as follows:
let ( ˜%%) x = 1.0 / x;;
val ( ˜%% ) : float -> float
%% 0.5;;
val it : float = 2.0

Remark: When defining an operator starting or ending in an asterisks “*” a space must be
inserted after “(” or before “)” to avoid a conflict with the comment convention using “(*”
and “*)”.
 2.10 Equality and ordering
The equality and inequality operators = and <> are defined on any basic type and on strings:
3.5 = 2e-3;;
val it : bool = false
"abc" <> "ab";;
val it : bool = true

It is not defined on functions (closures):

cos = sin;;
stdin(5,1): error FS0001: The type ′( a -> a) ...
does not support the ′equality′ constraint because
it is a function type

No type containing a function type can support equality as F# has no means to decide whether
two functions are equal: It is a fundamental fact of theoretical computer science that there
exists no (always terminating) algorithm to determine whether two arbitrary programs f and
g (i.e., two closures) denote the same function.

The equality function is automatically extended by F# whenever the user defines a new type –
in so far as the type does not contain function types.
The type of the function eqText declared by:
let eqText x y =
if x = y then "equal" else "not equal";;
val eqText : ′a -> ′a -> string when ′a : equality

contains a type variable ′a with the constraint: when ′a : equality.

This means that eqText will accept parameters x and y of any type τ equipped with equality:
eqText 3 4;;
val it : string = "not equal"

eqText ′ ′ (char 32);;

val it : string = "equal"
 Ordering
The ordering operators: >, > =, <, and <= are defined on values of basic types and on strings.
They correspond to the usual ordering of numbers. The ordering of characters is given by the
ordering of the Unicode values, while true > false in the ordering of truthlues. Strings are
ordered in the lexicographical ordering. That is, for two strings s1 and s2 we have that s1< s2
if s1 would occur before s2 in a lexicon. For example:
// Upper case letters precede
′A′ < ′a′;; // lower case letters
val it : bool = true

"automobile" < "car";;

val it : bool = true

"" < " ";;

val it : bool = true

Thus, the empty string precedes the string containing a space character, and the empty string
precedes any other string in the lexicographical ordering. Ordering is automatically extended
by F# whenever the user defines a new type, in so far as the type does not contain functions.
Using the comparison operators one may declare functions on values of an arbitrary type
equipped with an ordering:
let ordText x y = if x > y then "greater"
else if x = y then "equal"
else "less";;
val ordText : ′a -> ′a -> string when ′a : comparison

The type of x and y contains a type variable ′a with the constraint

when ′a : comparison

indicating that x and y can be of any type equipped with an ordering.

The library function compare is defined such that:

where the precise value of compare x y depends on the structure of the values x and y.
It may be convenient to use pattern matching with guards when declaring functions using
the compare function, for instance:
let ordText x y = match compare x y with
| t when t > 0 -> "greater"
| 0 -> "equal"
| _ -> "less";;
val ordText : ′a -> ′a -> string when ′a : comparison
The guard “when t > 0” restricts the matching, while the pattern “t” would otherwise match
any value.

Type Description Constant

bool Logical value true, false
unit Void ()

char Character ′char ′

byte 8-bit unsigned integer digits uy or 0x hexdigits uy
sbyte 8-bit signed integer {−}digits y or {−} 0x hexdigits y
int16 16-bit signed integer {−}digits s or {−} 0x hexdigits s
uint16 16-bit unsigned integer digits us or 0x hexdigits us
int
(or int32)
32-bit signed integer {−}digits or {−} 0x hexdigits
uint32 32-bit unsigned integer digits u or 0x hexdigits u
int64 64-bit signed integer {−} digits L or {−} 0x hexdigits L
uint64 64-bit unsigned integer digits UL or 0x hexdigits UL
nativeint Machines integer {−}digits n or {−} 0x hexdigits n
unativeint Machines unsigned integer digits un or 0x hexdigits un
float32 (or {−} digits.digits f or
32-bit IEEE floating-point
single) {−}digits{.digits} e {−}digits f
float {−}digits.digits or
(or double) 64-bit IEEE floating-point
{−}digits{.digits} e {−}digits
decimal High-precision decimal digits M or {−}digits.digits M
bigint Arbitrary integer {−}digits I
bignum Arbitrary rational number {−}digits N

Table 2.6 Basic Types

 2.11 Function application operators |> and <|
The operator |> means “send the value as argument to the function on the right” while <|
means “send the value as argument to the function on the left,” that is:

These operators are sometimes useful to make expressions more readable. There are two
reasons for that:

The operator |> allows you to write the argument to the left of the function.
The operators |> and <| have lower precedence than the arithmetic operators.

Both expressions a+b |> sin and sin <| a+b do hence mean sin(a+b). The operator |> has
precedence over <| so 2 |> (-) <| 3 means (2 |> (-)) <| 3.
Both operators associate to the left. The parentheses in 2 |> (3 |> (-)) are hence needed
to get the rightmost |> operator applied before the leftmost.
 2.12 Summary of the basic types
The F# system supports a number of basic types not addressed previously in this chapter.
Table 2.6 depicts the basic types, where the column “Constant” describes how constants are
written. The meta symbols digits, hexdigits, char, and {…} have the following meanings:
digits: One or more decimal digits: 0, 1, … , 9
hexdigits: One or more hex digits: 0, 1, … , 9, A, B, … , F, a, b, … , f
char: A character or an escape sequence denoting a character.
{…}: The part between the brackets is optional. The brackets are not part of the string.
Hence 33e−8 is a constant of type float and -0x1as is a constant of type int16 while 32f is
not accepted by F#.
Each type name denotes an overloaded conversion function converting to a value of the
type in question (in so far as this is possible).
 Summary
In this chapter we have described values and functions belonging to the basic F# types:
integers, reals, characters, truth values and strings. Furthermore, we have discussed
evaluation of infix operators with precedences, and the typing of arithmetic expressions
where some operators may be overloaded. The concept of higher-ordernctions was
introduced and the concept of a closure was used to explain the meaning of a function in F#. It
was explained how to declare operators, and finally, the concepts of equality and ordering
were explained.
 Exercise
2.1 Declare a function f: int -> bool such that f(n) = true exactly when n is divisible
by 2 or divisible by 3 but not divisible by 5. Write down the expected values of f(24),
f(27), f(29) and f(30) and compare with the result. Hint: n is divisible by q when n% q = 0.
2.2 Declare an F# function pow: string * int -> string, where:

where we use · to denote string concatenation. (The F# representation is +.)

2.3 Declare the F# function isIthChar: string * int * char -> bool where the value
of isIthChar(str,i,ch) is true if and only if ch is the i’th character in the string str
(numbering starting at zero).
2.4 Declare the F# function occFromIth: string * int * char -> int where

Hint: the value should be 0 for i ≥ size str.

2.5 Declare the F# function occInString: string * char -> int where

2.6 Declare the F# function notDivisible: int * int −> bool where

For example notDivisible(2, 5) is true, and notDivisible(3, 9) is false.

2.7

1. Declare the F# function test: int * int * int -> bool. The value of test(a, b,
c), for a ≤ b, is the truth value of:

2. Declare an F# function prime: int -> bool, where prime(n) = true, if and only if n
is a prime number.
3. Declare an F# function nextPrime: int -> int, where nextPrime(n) is the smallest
prime number > n.

2.8 The following figure gives the first part of Pascal’s triangle:
The entries of the triangle are called binomial coefficients. The k’th binomial coefficient
of the n’th row is denoted (nk), for n ≥ 0 and 0 ≤ ≤n. For example, (21) = 2 and (42) = 6. The
first and last binomial coefficients, that is, (n0) and (nn), of row n are both 1.binomial
coefficient inside a row is the sum of the two binomial coefficients immediately above it.
These properties can be expressed as follows:

and

Declare an F# function bin: int * int -> int to compute binomial coefficients.
2.9 Consider the declaration:
let rec f = function
| (0,y) -> y
| (x,y) -> f(x-1, x*y);;

1. Determine the type of f.

2. For which arguments does the evaluation of f terminate?
3. Write the evaluation steps for f(2, 3).
4. What is the mathematical meaning of f(x, y)?

2.10 Consider the following declaration:

let test(c,e) = if c then e else 0;;

1. What is the type of test?

2. What is the result of evaluating test(false, fact(−1))?
3. Compare this with the result of evaluating
if false then fact -1 else 0

2.11 Declare a function VAT: int −> float −> float such that the value VAT n x is
obtained by increasing x by n percent.
Declare a function unVAT: int −> float −> float such that

Hint: Use the conversion function float to convert an int value to a float value.
2.12 Declare a function min of type (int −> int) −> int. The value of min(f) is the
smallest natural number n where f(n) = 0 (if it exists).
2.13 The functions curry and uncurry of types
curry : (′a * ′b -> ′c) -> ′a -> ′b -> ′c
uncurry : (′a -> ′b -> ′c) -> ′a * ′b -> ′c

are defined in the following way:

 curry f is the function g where g x is the function h where h y = f(x, y).
uncurry g is the function f where f(x, y) is the value h y for the function h = g x.
Write declarations of curry and uncurry.
___________________
1 This description of legal operators in F# is incomplete. The precise rules are complicated.
 3
Tuples, records and tagged values

Tuples, records and tagged values are compound values obtained by combining values of
other types. Tuples are used in expressing “functions of several variables” where the
argument is a tuple, and in expressing functions where the result is a tuple. The components in
a record are identified by special identifiers called labels. Tagged values are used when we
group together values of different kinds to form a single set of values. Tuples, records and
tagged values are treated as “first-class citizens” in F#: They can enter into expressions and
the value of an expression can be a tuple, a record or a tagged value. Functions on tuples,
records or tagged values can be defined by use of patterns.
 3.1 Tuples
An ordered collection of n values (v1, v2, …, vn), where n > 1, is called an n-tuple.
Examples of n-tuples are:
(10, true);;
val it : int * bool = (10, true)

(("abc",1),-3);;
val it : (string * int) * int = (("abc", 1), -3)

A 2-tuple like (10,true) is also called a pair. The last example shows that a pair, for
example, (("abc",1),-3), can have a component that is again a pair ("abc",1). In general,
tuples can have arbitrary values as components. A 3-tuple is called a triple and a 4-tuple is
called a quadruple. An expression like (true) is not a tuple but just the expression true
enclosed in brackets, so there is no concept of 1-tuple. The symbol () denotes the only value
of type unit (cf. Page 23).
The n-tuple (v1, v2, …, vn) represents the graph:

The tuples (true,"abc",1,-3) and ((true,"abc"),1,-3) contain the same values true,
"abc", 1 and -3, but they are different because they have a different structure. This difference
is easily seen from the structure of the corresponding graphs in Figure 3.1, where the 4-tuple
(true,"abc",1,-3) represents the graph with four branches while the 3-tuple
((true,"abc"),1,-3) represents the graph with three branches and a sub-graph with two
branches.

Fig ure 3.1 Graphs for tuple values

 Tuple expressions
A tuple expression (expr1, expr2, …, exprn) is obtained by enclosing n expressions expr1,
expr2, …, exprn in parentheses. It has the type
τ1 * τ2 * ·s * τn when expr1, expr2, …, exprn have types τ1, τ2, …, τn. For example:
(1<2,"abc",1,1-4) has type bool * string * int * int
(true,"abc") has type bool * string
((2>1,"abc"),3-2,-3) has type (bool * string) * int * int

Remark: The tuple type τ1 * τ2 * ·s * τn corresponds to the Cartesian Product

of n sets A1, A2, …, An in mathematics. An element a of the set A is a tuple a = (a1, a2, … ,an)
of elements a1 A1, a2 A2, … , an An.
A tuple expression (expr1, expr2, …, exprn) is evaluated from left to right, that is, by first
evaluating expr1, then expr2, and so on. Tuple expressions can be used in declarations
whereby an identifier is bound to a tuple value, for example:
let tp1 = ((1<2, "abc"), 1, 1-4);;
val tp1 : (bool * string) * int * int = ((true, "abc"), 1, -3)

let tp2 = (2>1, "abc", 3-2, -3);;

val tp2 : bool * string * int * int = (true, "abc", 1, -3)

Fig ure 3.2 Graphs for tuples (true,"abc") and ((true, "abc"), 1, -3)
 Tuples are individual values
A tuple expression may contain identifiers that are already bound to tuples, for example:
let t1 = (true,"abc");;
val t1 : bool * string = (true, "abc")

let t2 = (t1,1,-3);;
val t2 : (bool * string) * int * int = ((true, "abc"), 1, -3)

The value bound to the identifier t1 is then found as a subcomponent of the value bound to t2
as shown in Figure 3.2. A fresh binding of t1 is, however, not going to affect the value of t2:
let t1 = -7 > 2;;
val t1 : bool = false

t2;;
val it : (bool * string) * int * int = ((true, "abc"), 1, -3)

The subcomponent (true,"abc") is a value in its own right and it depends in no way on
possible future bindings of t1 once the value of the expression (t1,1,-3) has been evaluated.
 Equality
Equality is defined for n-tuples of the same type, provided that equality is defined for the
components. The equality is defined componentwise, that is, (v1, v2, …, vn) is equal to (v′1, v
′2, …, v′n) if vi is equal to v′i for 1 ≤ i ≤ n. This corresponds to equality of the graphs
represented by the tuples. For example:
("abc", 2, 4, 9) = ("ABC", 2, 4, 9);;
val it : bool = false

(1, (2,true)) = (2-1, (2,2>1));;

val it : bool = true

(1, (2,true)) = (1, 2, 2>1);;

(1, (2,true)) = (1, 2, 2>1);;
—————–^^^^^^^^^
stdin(25,18): error FS0001: Type mismatch. Expecting a
int * (int * bool)
but given a
int * (int * bool) * ′a
The tuples have differing lengths of 2 and 3

An error message occurs in the last example. The pair (1, (2,true)) on the left-hand side
of the equality has type int * (int * bool) while the tuple on the right-hand side has type
int * int * bool. The system recognizes that these types are different and issues an error
message.
 Ordering
The ordering operators: >, > =, <, and <=, and the compare function are defined on n-tuples of
the same type, provided ordering is defined for the components. Tuples are ordered
lexicographically:

exactly when, for some k, where 1 ≤ k ≤ n, we have:

For example:
(1, "a") < (1, "ab");;
val it : bool = true

(2, "a") < (1, "ab");;

val it : bool = false

since “a” < “ab” holds while 2 < 1 does not.

The other comparison operators and the compare function can be defined in terms of = and
< as usual, for example:

(′a′, ("b",true), 10.0) >= (′a′, ("b",false), 0.0);;

val it : bool = true

compare ("abcd", (true, 1)) ("abcd", (false, 2));;

val it : int = 1
 Tuple patterns
A tuple pattern represents a graph. For example, the pattern (x,n) is a tuple pattern. It
represents the graph shown to the left containing the identifiers x and n:

The graph represented by the value (3,2) (shown to the right) matches the graph for the
pattern in the sense that the graph for the value is obtained from the graph for the pattern by
substituting suitable values for the identifiers in the pattern – in this case the value 3 for the
identifier x and the value 2 for the identifier n. Hence, the pattern matching gives the bindings
x 3 and n 2.
Patterns can be used on the left-hand side in a let declaration which binds the identifiers in
the pattern to the values obtained by the pattern matching, for example:
let (x,n) = (3,2);;
val x : int = 3
val n : int = 2

Patterns may contain constants like the pattern (x,0), for example, containing the constant
0. It matches any pair (v1 ,v2 ) where v2 = 0, and the binding x v1 is then obtained:
let (x,0) = ((3,"a"),0);;
val x : int * string = (3, "a")

This example also illustrates that the pattern matching may bind an identifier (here: x) to a
value which is a tuple.
The pattern (x,0) is incomplete in the sense that it just matches pairs where the second
component is 0 and there are other pairs of type τ*int that do not match the pattern. The
system gives a warning when an incomplete pattern is used:
let (x,0) = ((3,"a"),0);;
—-^^^^^
stdin(46,5): warning FS0025: Incomplete pattern matches on
this expression. For example, the value ′(_,1)′ may indicate a
case not covered by the pattern(s).

The warning can be ignored since the second component of ((3,"a"),0) is, in fact, 0.
By contrast the declaration:
let (x,0) = (3,2);;
let (x,0) = (3,2);;
—-^^^^^
stdin(49,5): warning FS0025: Incomplete pattern matches on
this expression. For example, the value ′(_,1)′ may indicate a
case not covered by the pattern(s).
Microsoft.FSharp.Core.MatchFailureException: The match cases
were incomplete at <Startup$CodeFSI_0036>.$FSI0036.main@()
Stopped due to error

generates an error message because the constant 0 in the pattern does not match the
corresponding value 2 on the right-hand side. The system cannot generate any binding in this
case.
The wildcard pattern can be used in tuple patterns. Every value matches this pattern, but the
matching provides no bindings. For example:
let ((_,x),_,z) = ((1,true), (1,2,3), false);;
val z : bool = false
val x : bool = true

A pattern cannot contain multiple occurrences of the same identifier, so (x,x), for
example, is an illegal pattern:
let (x,x) = (1,1);;
let (x,x) = (1,1);;
——-^
... error FS0038: ′x′ is bound twice in this pattern
 3.2 Polymorphism
Consider the function swap interchanging the components of a pair:
let swap (x,y) = (y,x);;
val swap : ′a * ′b -> ′b * ′a

swap (′a′,"ab");;
val it : string * char = ("ab", ′a′)

swap ((1,3),("ab",true));;
val it : (string*bool) * (int*int) = (("ab", true), (1, 3))

The examples show that the function applies to all kinds of pairs. This is reflected in the type
of the function: ′a * ′b -> ′b * ′a.
The type of swap expresses that the argument (type ′a * ′b) must be a pair, and that the
value will be a pair (type ′b * ′a) such that the first/second component of the value is of
same type as the second/first component of the argument.
The type of swap contains two type variables ′a and ′b. A type containing type variables is
called a polymorphic type and a function with polymorphic type like swap is called a
polymorphic function. Polymorphic means “of many forms”: In our case the F# compiler is
able to generate a single F# function swap working on any kind of pairs and which is hence
capable of handling data of many forms.
Polymorphism is related to overloading (cf. Section 2.5) as we in both cases can apply the
same function name or operator to arguments of different types, but an overloaded operator
denotes different F# functions for different argument types (like + denoting integer addition
when applied to int’s and floating-point addition when applied to float’s).
There are two predefined, polymorphic functions
fst: ′a * ′b -> ′a and snd: ′a * ′b -> ′b

on pairs, that select the first and second component, respectively. For example:
fst((1,"a",true), "xyz");;
val it : int * string * bool = (1, "a", true)

snd(′z′, ("abc", 3.0));;

val it : string * float = ("abc", 3.0)
 3.3 Example: Geometric vectors
A proper vector in the plane is a direction in the plane together with a non-negative length.
The null vector is any direction together with the length 0. A vector can be represented by its
set of Cartesian coordinates which is a pair of real numbers. A vector might instead be
represented by its polar coordinates, which is also a pair of real numbers for the length and
the angle. These two representations are different, and the operators on vectors (addition of
vectors, multiplication by a scalar, etc.) are expressed by different functions on the
representing pairs of numbers.
In the following we will just consider the Cartesian coordinate representation, where a
vector in the plane will be represented by a value of type float * float.
We will consider the following operators on vectors:

We cannot use the operator symbols +,-,*, and so on, to denote the operations on vectors,
as this would overwrite their initial meaning. But using +. for vector addition, -. for vector
reversal and subtraction, *. for product by a scalar and &. for dot product, we obtain
operators having a direct resemblance to the mathematical vector operators and having the
associations and precedences that we would expect.
The prefix operator for vector reversal is declared by (cf. Section 2.9):
let ( -.) (x:float,y:float) = (-x,-y);;
val ( -. ) : float * float -> float * float

and the infix operators are declared by:

let (+.) (x1, y1) (x2,y2) = (x1+x2,y1+y2): float*float;;
val ( +. ) : float * float -> float * float -> float * float

let (-.) v1 v2 = v1 +. -. v2;;

val ( -. ) : float * float -> float * float -> float * float

let ( *.) x (x1,y1) = (x*x1, x*y1): float*float;;

val ( *. ) : float -> float * float -> float * float

let (&.) (x1,y1) (x2,y2) = x1*x2 + y1*y2: float;;

val (&. ) : float * float -> float * float -> float

The norm function is declared using the sqrt function (cf. Table 2.5) by:
let norm(x1:float,y1:float) = sqrt(x1*x1+y1*y1);;
val norm : float * float -> float

These functions allow us to write vector expressions in a form resembling the mathematical
notation for vectors. For example:
let a = (1.0,-2.0);;
val a : float * float = (1.0, -2.0)

let b = (3.0,4.0);;
val b : float * float = (3.0, 4.0)

let c = 2.0 *. a -. b;;

val c : float * float = (-1.0, -8.0)

let d = c &. a;;

val d : float = 15.0

let e = norm b;;

val e : float = 5.0
 3.4 Records
A record is a generalized tuple where each component is identified by a label instead of the
position in the tuple.
The record type must be declared before a record can be made. We may for example
declare a type person as follows:
type Person = {age : int; birthday : int * int;
name : string; sex : string};;

The keyword type indicates that this is a type declaration and the braces { and } indicate a
record type. The (distinct) identifiers age, birthday, name and sex are called record labels
and they are considered part of the type.
A value of type Person is entered as follows:
let john = {name = "John"; age = 29;
sex = "M"; birthday = (2,11)};;
val john : Person = { = 29;
birthday = (2, 11);
name = "John";
sex = "M";}

This record contains the following fields: The string “John” with label name, the integer 29
with label age, the string “M” with label sex, and the integer pair (2,11) with label birthday.
The declaration creates the following binding of the identifier john:
john { age 29 , birthday (2,11), name “John” , sex “M” }
A record is hence a local environment packaged in a certain way. It contains a binding of each
record label to the corresponding value.
A field in the record denoted by john is obtained by suffixing the identifier john with the
corresponding record label:
john.birthday;;
val it : int * int = (2, 11)
john.sex;;
val it : string = "M"
 Equality and ordering
The equality of two records with the same type is defined componentwise from the equality of
values associated with the same labels, so the ordering of the components in the record is of
no importance when entering values. For example
john = {age = 29; name = "John";
sex = "M"; birthday = (2,11)};;
val it : bool = true

Hence two records are equal if they are of the same type and contain the same local bindings
of the labels.
Ordering of records is based on a lexicographical ordering using the ordering of the labels
in the record type declaration. Consider, for example:
type T1 = {a:int; b:string};;
let v1 = {a=1; b="abc"};;
let v2 = {a=2; b="ab"};;
v1<v2;;
val it : bool = true

type T2 = {b:string; a:int};;

let v1′ = {T2.a=1; b="abc"};;
let v2′ = {T2.a=2; b="ab"};;
v1′>v2′;;
val it : bool = true

The value v1 is smaller than the value v2 because the label a occurs first in the record type T1
and v1.a = 1 is smaller than v2.a = 2 – while the value v1′ is larger than the value v2′
because the label b occurs first in the record type T2 and v1′.b = "abc" is larger than v2′.b =
"ab".
The composite identifier T2.a consists of the record label a prefixed with the record type
T2. It is used in order to resolve the ambiguity created by reuse of record labels.
Note that the values v1 and v1′ cannot be compared as they are of different types.

Fig ure 3.3 Record pattern and record

 Record patterns
A record pattern is used to decompose a record into its fields. The pattern
{name = x; age = y; sex = s; birthday =(d,m)}

denotes the graph shown in Figure 3.3. It generates bindings of the identifiers x, y, s, d and m
when matched with a person record:
let sue = {name="Sue"; age = 19; sex="F";
birthday = (24,12)};;
let {name = x; age = y; sex = s; birthday = (d,m)} = sue;;
val y : int = 19
val x : string = "Sue"
val s : string = "F"
val m : int = 12
val d : int = 24

Record patterns are used when defining functions. Consider, for example, the declaration of
a function age where the argument is a record of type Person:
let age {age = a; name = _; sex=_; birthday=} = a;;
val age : Person -> int

let isYoungLady {age=a; sex=s; name=_; birthday=}

= a < 25 && s = "F";;
val isYoungLady : Person -> bool

age john;;
val it : int = 29

isYoungLady john;;
val it : bool = false

isYoungLady sue;;
val it : bool = true

The type of the above functions can be inferred from the context since name, age, and so on
are labels of the record type Person only.
 3.5 Example: Quadratic equations
In this section we consider the problem of finding solutions to quadratic equations

with real coefficients a, b, c.

The equation has no solution in the real numbers if the discriminant b2 - 4 a c is negative;
otherwise, if b2 − 4ac ≥ 0 and a ≠ 0, then the equation has the solutions x1 and x2 where:

Note that x1 = x2 if b2 - 4 a c = 0.
We may represent the equation ax2 + bx + c = 0 by the triple (a, b, c) of real numbers and
the solutions x1 and x2 by the pair (x1, x2) of real numbers. This representation is captured in
the type declarations:
type Equation = float * float * float;;
type Equation = float * float * float

type Solution = float * float;;

type Solution = float * float

A function:
solve: Equation -> Solution

for computing the solutions of the equation should then have the indicated type. Note that type
declarations like the ones above are useful in program documentation as they communicate
the intention of the program in a succinct way. The system does, however, just treat the
identifiers Equation and Solution as shorthand for the corresponding types.
 Error handling
The function solve must give an error message when b2 − 4ac < 0 or a = 0 as there is no
solution in these cases. Such an error message can be signalled by using an exception. An
exception is named by an exception declaration. We may, for example, name an exception
Solve by the declaration:

exception Solve;;
exception Solve

The declaration of the function solve is:

let solve(a,b,c) =
if b*b-4.0*a*c < 0.0 || a = 0.0 then raise Solve
else ((-b + sqrt(b*b-4.0*a*c))/(2.0*a),
(-b - sqrt(b*b-4.0*a*c))/(2.0*a));;
val solve : float * float * float -> float * float

The then branch of this declaration contains the expression: raise Solve. An evaluation
of this expression terminates with an error message. For example:
solve(1.0, 0.0, 1.0);;
FSI0015+Solve: Exception of type ′FSI0015+Solve′ was thrown.
at FSI0016.solve(Double a, Double b, Double c)
at <StartupCode$FSI0017>.$FSI0017.main@()
Stopped due to error

We say that the exception Solve is raised. Note that the use of the exception does not
influence the type of solve.
Other examples of the use of solve are:
solve(1.0, 1.0, -2.0);;
val it : float * float = (1.0, -2.0)

solve(2.0, 8.0, 8.0);;

val it : float * float = (-2.0, -2.0)

An alternative to declaring your own exception is to use the built-in function:

failwith: string -> ′a

that takes a string as argument. An application failwith s raises the exception Faiure s and
the argument string s is shown on the console. The function can be applied in any context
because the function value has a polymorphic type. For example:
let solve(a,b,c) =
if b*b-4.0*a*c < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else ((-b + sqrt(b*b-4.0*a*c))/(2.0*a),
(-b - sqrt(b*b-4.0*a*c))/(2.0*a));;

solve(0.0,1.0,2.0);;
System.Exception: discriminant is negative or a=0.0
at FSI0037.solve(Double a, Double b, Double c)
at <StartupCode$FSI0038>.$FSI0038.main@()
Stopped due to error

We shall on Page 63 study how raised exceptions can be caught.

 3.6 Locally declared identifiers
It is often convenient to use locally declared identifiers in function declarations. Consider for
example the above declaration of the function solve. The expression b*b-4.0*a*c is
evaluated three times during the evaluation of solve(1.0,1.0,-2.0) and this is not
satisfactory from an efficiency point of view. Furthermore, the readability of the declaration
suffers from the repeated occurrences of the same subexpression.
These problems are avoided using a locally declared identifier d:
let solve(a,b,c) =
let d = b*b-4.0*a*c
if d < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else ((-b + sqrt d)/(2.0*a),(-b - sqrt d)/(2.0*a));;
val solve : float * float * float -> float * float

There is room for improvement in the above declaration as the expression sqrt d is
evaluated twice during the evaluation of a function value. This leads to yet another declaration
of solve with a further locally declared identifier sqrtD:
let solve(a,b,c) =
let sqrtD =
let d = b*b-4.0*a*c
if d < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else sqrt d
((-b + sqrtD)/(2.0*a),(-b - sqrtD)/(2.0*a));;
val solve : float * float * float -> float * float

The evaluation of solve(1.0,1.0,-2.0) proceeds as follows, where env denotes the

environment:
env = [a 1.0, b 1.0, c −2.0]
obtained by binding the parameters a, b and c to the actual values 1.0, 1.0 and −2.0:

1. The binding of sqrtD can only be established when the value of the subexpression has
been evaluated.
 2. The evaluation of the subexpression starts with the declaration let b = … .
3. The expression b*b-4.0*a*c evaluates to 9.0 using the bindings in env. A binding of d to
this value is added to the environment.
4. The evaluation of if d < 0.0 … else sqrt d gives the value 3.0 using the bindings in
the environment env plus d 9.0.
5. The evaluation of the subexpression is completed and the binding of d is removed from
the environment.
6. A binding of sqrtD to the value 3.0 is added to the environment, and the expression ((-b
+ sqrtD … is evaluated in this environment.
7. The bindings of a, b, c and sqrtD are removed and the evaluation terminates with result
(1.0, −2.0).

Note the role of indentation in F#. The let-expression:

let d = b*b-4.0*a*c
if d < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else sqrt d

is terminated by the occurrence of a less indented line:

((-b + sqrtD)/(2.0*a),(-b - sqrtD)/(2.0*a));;

and this also ends the lifetime of the binding of d. One says that the let-expression
constitutes the scope of the declaration of d.
The surrounding let-expression
let sqrtD =
let d = b*b-4.0*a*c
if d < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else sqrt d
((-b + sqrtD)/(2.0*a),(-b - sqrtD)/(2.0*a))

is terminated by the double semicolon. Note that the expression ((-b + … must be on the
same indentation level as let sqrtD =.
A let-expression may contain more than one local declaration as shown in yet another
version of solve (probably the most readable):
let solve(a,b,c) =
let d = b*b-4.0*a*c
if d < 0.0 || a = 0.0
then failwith "discriminant is negative or a=0.0"
else let sqrtD = sqrt d
((-b + sqrtD)/(2.0*a),(-b - sqrtD)/(2.0*a));;
val solve : float * float * float -> float * float

The evaluation of solve(1.0,1.0,-2.0) in this version of the function will add the binding
of d to the environment env. Later the binding of sqrtD is further added without removing the
binding of b. Finally the expression in the last line is evaluated and the bindings of a, b, c, d
and sqrtD are all removed at the same time.
 3.7 Example: Rational numbers. Invariants
A rational number q is a fraction q = , where a and b are integers with b ≠ 0.
Ideas to express the operations on rational numbers by function declarations come easily
from the following well-known rules of arithmetic, where a, b, c and d are integers such that
b ≠ 0 and d ≠ 0:

(3.1)
 Representation. Invariant
We use the representation (a, b), where b > 0 and where the fraction is irreducible, that is,
gcd(a, b) = 1, to represent the rational number . Thus, a value (a, b) of type int * int
represents a rational number if b > 0 and gcd(a, b) = 1, and we name this condition the
invariant for pairs representing rational numbers. Any rational number has a unique normal
form of this kind. This leads to the type declaration:
type Qnum = int*int;; // (a,b) where b > 0 and gcd(a,b) = 1

where the invariant is stated as a comment to the declaration. (The declaration of gcd is found
on Page 15.)
 Operators
It is convenient to declare a function canc that cancels common divisors and thereby reduces
any fraction with non-zero denominator to the normal form satisfying the invariant:
let canc(p,q) =
let sign = if p*q < 0 then -1 else 1
let ap = abs p
let aq = abs q
let d = gcd(ap,aq)
(sign * (ap / d), aq / d);;

In the below declarations for the other functions, canc is applied to guarantee that the
resulting values satisfy the invariant.
When a rational number is generated from a pair of integers, we must check for division
by zero and enforce that the invariant is established for the result. The function mkQ does that
by the use of canc:
let mkQ = function
| (_,0) -> failwith "Division by zero"
| pr -> canc pr;;

The operators on rational numbers are declared below. These declarations follow the rules
(3.1) for rational numbers. We assume that the arguments are legal representations of rational
numbers, that is, they respect the invariant. Under this assumption, the result of any of the
functions must respect the invariant. This is enforced by the use of canc and mkQ:
let (.+.) (a,b) (c,d) = canc(a*d + b*c, b*d);; // Addition

let (.-.) (a,b) (c,d) = canc(a*d - b*c, b*d);; // Subtraction

let (.*.) (a,b) (c,d) = canc(a*c, b*d);; // Multiplication

let (./.) (a,b) (c,d) = (a,b) .*. mkQ(d,c);; // Division

let (.=.) (a,b) (c,d) = (a,b) = (c,d);; // Equality

Note that the definition of equality assumes the invariant. Equality should be declared by
a*d=b*c if we allow integer pairs not satisfying the invariant as there would then be many
different integer pairs representing the same rational number.
It is straightforward to convert a rational number representation to a string:
let toString(p:int,q:int) = (string p) + "/" + (string q);;

as the representation is unique. We can operate on rational numbers in a familiar manner:

let q1 = mkQ(2,-3);;
val q1 : int * int = (-2, 3)
let q2 = mkQ(5,10);;
val q2 : int * int = (1, 2)
let q3 = q1 .+. q2;;
val q3 : int * int = (-1, 6)

toString(q1 .-. q3 ./. q2);;

val it : string = "-1/3"
 3.8 Tagged values. Constructors
Tagged values are used when we group together values of different kinds to form a single
set of values.
For example, we may represent a circle by its radius r, a square by its side length a, and a
triangle by the triple (a, b, c) of its side lengths a, b and c. Circles, squares, and triangles may
then be grouped together to form a single collection of shapes if we put a tag on each
representing value. The tag should be Circle, Square, or Triangle depending on the kind of
shape. The circle with radius 1.2, the square with side length 3.4 and the triangle with side
lengths 3.0, 4.0 and 5.0 are then represented by the tagged values shown in the following
graphs:

In F#, a collection of tagged values is declared by a type declaration. For example, a type
for shapes is declared by:
type Shape = | Circle of float
| Square of float
| Triangle of float*float*float;;
type Shape =
| Circle of float
| Square of float
| Triangle of float * float * float
 Constructors and values
The response from F# indicates that Shape names a type, and that Circle, Square and
Triangle are bound to value constructors. These value constructors are functions and they
give a tagged value of type Shape when applied to an argument.
For example, Circle is a value constructor with type float -> Shape. This means that
Circle r denotes a value of type Shape, for every float r. For example, Circle 1.2 denotes
the leftmost graph in the above figure and Circle 1.2 is an example of a tagged value, where
Circle is the tag.
We can observe that Circle 1.2 is a value which is not evaluated further by F#:
Circle 1.2;;
val it : Shape = Circle 1.2

as the value in the answer is equal to the expression being evaluated, that is, Circle 1.2.
Values can be constructed using Square and Triangle in a similar way.
Since constructors are functions in F#, Circle can be applied to an expression of type
float:

Circle(8.0 - 2.0*3.4);;
val it : Shape = Circle 1.2

Thus, the declaration of Shape allows one to write tagged values like Circle 1.2, Square
3.4 and Triangle(3.0,4.0,5.0) using the constructors, and any value of type Shape has one
of the forms:
Circle r
Square a
Triangle (a, b, c)

for some float value r, float value a, or triple (a, b, c) of float values.
 Equality and ordering
Equality and ordering are defined for tagged values provided they are defined for their
components. Two tagged values are equal if they have the same constructor and their
components are equal. This corresponds to equality of the graphs represented by the tagged
values. For example:
Circle 1.2 = Circle(1.0 + 0.2);;
val it : bool = true

Circle 1.2 = Square 1.2;;

val it : bool = false

The sequence in which the tags occur in the declaration is significant for the ordering. For
example, any circle is smaller than any square, which again is smaller than any triangle due to
the order in which the corresponding tags are declared. For example:
Circle 1.2 < Circle 1.0;;
val it : bool = false

Circle 1.2 < Square 1.2;;

val it : bool = true

Triangle(1.0,1.0,1.0) > Square 4.0;;

val it : bool = true
 Constructors in patterns
Constructors can be used in patterns. For example, an area function for shapes is declared by:
let area = function
| Circle r -> System.Math.PI * r * r
| Square a -> a * a
| Triangle(a,b,c) ->
let s = (a + b + c)/2.0
sqrt(s*(s-a)*(s-b)*(s-c));;
val area : Shape -> float

The pattern matching treats constructors differently from other identifiers:

A constructor matches itself only in a pattern match while other identifiers match
any value.
For example, the value Circle 1.2 will match the pattern Circle r, but not the other
patterns in the function declaration. The matching binds the identifier r to the value 1.2, and
the expression Math.pi * r * r is evaluated using this binding:

The value Triangle(3.0,4.0,5.0) will in a similar way only match the pattern in the third
clause in the declaration, and we get bindings of a, b and c to 3.0, 4.0 and 5.0, and the let
expression is evaluated using these bindings:
 Invariant for the representation of shapes
Some values of type Shape do not represent geometric shapes. For example, Circle -1.0
does not represent a circle, as a circle cannot have a negative radius, Square -2.0 does not
represent a square, as a square cannot have a negative side length, and
Triangle(3.0, 4.0, 7.5)

does not represent a triangle, as 7.5 > 3.0 + 4.0 and, therefore, one of the triangle inequalities
is not satisfied.
Therefore, there is an invariant for this representation of shapes: the real numbers have to
be positive, and the triangle inequalities must be satisfied. This invariant can be declared as a
predicate isShape:
let isShape = function
| Circle r -> r > 0.0
| Square a -> a > 0.0
| Triangle(a,b,c) ->
a > 0.0 && b > 0.0 && c > 0.0
&& a < b + c && b < c + a && c < a + b;;
val isShape : Shape -> bool

We consider now the declaration of an area function for geometric shapes that raises an
exception when the argument of the function does not satisfy the invariant. If we try to modify
the above area function:
let area x = if not (isShape x)
then failwith "not a legal shape"
else ...

then the else-branch must have means to select the right area-expression depending on the
form of x. This is done using a match ... with ... expression:
let area x =
if not (isShape x)
then failwith "not a legal shape" raise
else match x with
| Circle r -> System.Math.PI * r * r
| Square a -> a * a
| Triangle(a,b,c) ->
let s = (a + b + c)/2.0
sqrt(s*(s-a)*(s-b)*(s-c));;
val area : Shape -> float

The modified area function computes the area of legal values of the type Shape and
terminates the evaluation raising an exception for illegal values:
area (Triangle(3.0,4.0,5.0));;
val it : float = 6.0

area (Triangle(3.0,4.0,7.5));;
System.Exception: not a legal shape
...
 3.9 Enumeration types
Value constructors need not have any argument, so we can make special type declarations
like:
type Colour = Red | Blue | Green | Yellow | Purple;;
type Colour =
| Red
| Blue
| Green
| Yellow
| Purple

Types like Colour are called enumeration types, as the declaration of Colour just enumerates
five constructors:
Red, Blue, Green, Yellow, Purple

where each constructor is a value of type Colour, for example:

Green;;
val it : Colour = Green

Functions on enumeration types may be declared by pattern matching:

let niceColour = function
| Red -> true
| Blue -> true
| _ -> false;;
val niceColour : Colour -> bool

niceColour Purple;;
val it : bool = false

The days in a month example on Page 4 can be nicely expressed using an enumeration type:
type Month = January | February | March | April
| May | June | July | August | September
| October | November | December;;

let daysOfMonth = function

| February -> 28
| April | June | September | November -> 30
| _ -> 31;;

The Boolean type is actually a predefined enumeration type:

type bool = false | true

where the order of the constructors reflects that false < true. Notice that user-defined
constructors must start with uppercase letters.
 3.10 Exceptions
Exceptions have already been used in several examples earlier in this chapter. In this section
we give a systematic account of this subject.
Raising an exception terminates the evaluation of a call of a function as we have seen for
the solve function on Page 53 that raises the exception Solve when an error situation is
encountered. In the examples presented so far the exception propagates all the way to top level
where an error message is issued.
It is possible to catch an exception using a try…with expression as in the following
solveText function:

let solveText eq =
try
string(solve eq)
with
| Solve -> "No solutions";;
val solveText : float * float * float -> string

It calls solve with a float triple eq representing a quadratic equation and returns the string
representation of the solutions of the equation:
solveText (1.0,1.0,-2.0);;
val it : string = "(1, -2)"

The string “No solutions” is returned if the equation has no solutions:

solveText (1.0, 0.0, 1.0);;
val it : string = "No solutions"

An application of the function failwith s will raise the exception Failure s and this
exception can also be caught. Application of the function mkQ (see Page 57), for example, will
call failwith in the case of a division by zero:
try
toString(mkQ(2,0))
with
| Failure s -> s;;
val it : string = "Division by zero"

A try…with expression has the general form:

try e with match
where e is an expression (possibly extending over several lines) and match is a construct of
the form:
| pat1 -> e1
| pat2 -> e2
...
| patn -> en

with patterns pat1, … patn and expressions e1, …, en.

A try e with match expression is evaluated as follows:

Evaluate the expression e. If this evaluation terminates normally with a value v then
return v as the result of evaluating the try …with … expression.
If the evaluation raises an exception Exc then evaluate match by matching Exc to the
patterns pat1, … patn. If Exc matches a pattern patk then evaluate the corresponding
expression ek. If Exc matches none of the patterns then propagate the exception as a result
of evaluating the try …with … expression.

The exception mechanism in F# and .NET is not intended for use in the “normal case” in a
program but for error handling only.
Library functions (e.g., for performing I/O) may raise exceptions that can only be captured
using a match on type (cf. Section 7.7).
 3.11 Partial functions. The option type
A function f is a partial function on a set A if the domain of f is a proper subset of A. For
example, the factorial function is a partial function on the set of integers because it is
undefined on negative integers.
In declaring a partial function, F# offers the programmer three ways of handling argument
values where the function is undefined:

1. The evaluation of the function value does not terminate.

2. The evaluation of the function value is terminated by raising an exception.
3. The evaluation of the function value gives a special result, indicating that the function is
undefined for the actual argument.

The first choice was used in the declaration of the factorial function fact, where, for
example, the evaluation of fact -1 never terminates.
The second choice was selected for the improved area function (cf. Page 61).
The third choice uses the predefined option type:
type ′a option = None | Some of ′a

where None is used as result for arguments where the function is undefined while Some v is
used when the function has value v.
The constructor Some is polymorphic and can be applied to values of any type:
Some false;;
val it : bool option = Some false

Some (1, "a");;

val it : (int * string) option = Some (1, "a")

The value None is a polymorphic value of type ′a option:

None;;
val it : ′a option = None

The library function

Option.get : ′a option -> ′a

“removes the Some”, that is, Option.get(Some n) = n. It raises an exception when applied to
None. For example:

Option.get(Some (1,"a"));;
val it : int * string = (1, "a")

Option.get(Some 1);;
val it : int = 1

Option.get None + 1;;

System.ArgumentException: The option value was None ...

We may, for instance, declare a modified factorial function optFact(n) with value Some n!
for n ≥ 0 and None for n < 0:
let optFact n = if n < 0 then None else Some(fact n);;
val optFact : int -> int option

The function application optFact n always gives a result:

optFact 5;;
val it : int option = Some 120

optFact -2;;
val it : int option = None

The declaration of optFact presumes that fact has already been declared. An independent
declaration of optFact is achieved using the Option.get function:
let rec optFact = function
| 0 -> Some 1
| n when n > 0 -> Some(n * Option.get(optFact(n-1)))
| _ -> None;;
val optFact : int -> int option

Note the use of guarded patterns in this declaration (cf. Section 2.10).
 Summary
This chapter introduces the notions of tuples and tuple types, the notions of records and
record types, and the notions of tagged values and tagged-value types. Tuples and records are
composite values, and we have introduced the notion of patterns that is used to decompose a
composite value into its parts. Tagged values are used to express disjoint unions.
An operator can be given infix mode and precedence, and this feature was exploited in
writing the operators on geometric vectors in the same way as they are written in
mathematical notation.
The notion of exceptions was introduced for handling errors and let expressions were
introduced for having locally declared identifiers.
 Exercise
3.1 A time of day can be represented as a triple (hours, minutes, f) where f is either AM or
PM – or as a record. Declare a function to test whether one time of day comes before
another. For example, (11,59,"AM") comes before (1,15,"PM"). Make solutions with
triples as well as with records. Declare the functions in infix notation.
3.2 The former British currency had 12 pence to a shilling and 20 shillings to a pound.
Declare functions to add and subtract two amounts, represented by triples (pounds,
shillings, pence) of integers, and declare the functions when a representation by records is
used. Declare the functions in infix notation with proper precedences, and use patterns to
obtain readable declarations.
3.3 The set of complex numbers is the set of pairs of real numbers. Complex numbers
behave almost like real numbers if addition and multiplication are defined by:

1. Declare suitable infix functions for addition and multiplication of complex numbers.
2. The inverse of (a, b) with regard to addition, that is, −(a, b), is (−a, −b), and the
inverse of (a, b) with regard to multiplication, that is, 1/(a, b), is (a/(a2 + b2), −b/(a2 +
b2)) (provided that a and b are not both zero). Declare infix functions for subtraction
and division of complex numbers.
3. Use let-expressions in the declaration of the division of complex numbers in order
to avoid repeated evaluation of identical subexpressions.

3.4 A straight line y = ax +b in the plane can be represented by the pair (a,b) of real
numbers.

1. Declare a type StraightLine for straight lines.

2. Declare functions to mirror straight lines around the x and y-axes.
3. Declare a function to give a string representation for the equation of a straight line.

3.5 Make a type Solution capturing the three capabilities for roots in a quadratic
equation: two roots, one root and no root (cf. Section 3.5). Declare a corresponding solve
function.
3.6 Solve Exercise 3.1 using tagged values to represent AM and PM.
3.7 Give a declaration for the area function on Page 61 using guarded patterns rather than
an if...then...else expression.
 4
Lists

Lists are at the core of functional programming. A large number of applications can be
modelled and implemented using lists. In this chapter we introduce the list concept, including
list values, patterns and basic operations, and we study a collection of recursion schemas over
lists. We end the chapter introducing a model-based approach to functional programming on
the basis of two examples. The concept of a list is a special case of a collection. In the next
chapter, when we consider collections more generally, we shall see that the F# library
comprises a rich collection of powerful functions on lists.
 4.1 The concept of a list
A list is a finite sequence of values

of the same type. For example, [2], [3; 2], and [2; 3; 2] are lists of integers. A list can
contain an arbitrary number of elements.
A list [v0; v1; …; vn−1] is either empty (when n=0), or it is a non-empty list and can be
characterized by the first element v0 called its head, and the rest [v1; …; vn−1] called its tail.

Fig ure 4.1 Graphs for two lists

Figure 4.1 shows the graphs for the lists [2; 3; 2] and [2]. The list [2; 3; 2] is hence a
tagged pair with tag :: where the first component, the head of the list, is the integer 2, while
the second component, the tail of the list, is the list [3; 2] with just two elements. This list is
again a tagged pair with tag ::, head 3 and tail [2]. Finally the head of the list [2] is the
integer 2, while the tail is the empty list [].
 List constants in F#
Lists can be entered as values:
let xs = [2;3;2];;
val xs : int list = [2; 3; 2]

let ys = ["Big"; "Mac"];;

val ys : string list = ["Big"; "Mac"]

The types int list and string list, containing the type constructor list, indicate that
the value of xs is a list of integers and that the value of ys is a list of strings.
We may have lists with any element type, so we can, for example, build lists of pairs:
[("b",2);("c",3);("e",5)];;
val it : (string * int) list = [("b", 2);("c", 3);("e", 5)]

lists of records:
type P = {name: string; age: int}
[{name = "Brown"; age = 25}; {name = "Cook"; age = 45}];;
val it : P list =
[{name = "Brown"; age = 25}; {name = "Cook"; age = 45}]

lists of functions:
[sin; cos];;
val it : (float -> float) list = [<fun:it@7>; <fun:it@7-1>]

or even lists of lists:

[[2;3];[3];[2;3;3]];;
val it : int list list = [[2; 3]; [3]; [2; 3; 3]]

Furthermore, lists can be components of other values. We can, for example, have pairs
containing lists:
("bce", [2;3;5]);;
val it : string * int list = ("bce", [2; 3; 5])
 The type constructor list
The type constructor list has higher precedence than * and -> in type expressions, so the type
string * int list means string * (int list). The type constructor list is used in
postfix notation like the factorial function _! in 3! and associates to the left, so int list list
means (int list) list. Note that int (list list) would not make sense.
All elements in a list must have the same type. For example, the following is not a legal
value in F#:
["a";1];;
—–^
stdin(8,6): error FS0001:
This expression was expected to have type
string
but here has type
int
 Equality of lists
Two lists [x0; x1; …; xm-1] and [y0; y1; …, yn−1] (of the same type) are equal when m=n
and xi = yi, for all i such that 0 ≤i < m. This corresponds to equality of the graphs represented
by the lists. Hence, the order of the elements as well as repetitions of the same value are
significant in a list.
The equality operator = of F# can be used to test equality of two lists provided that the
elements of the lists are of the same type and provided that the equality operator can be used
on values of that element type.
For example:
[2;3;2] = [2;3];;
val it : bool = false

[2;3;2] = [2;3;3];;
val it : bool = false

The differences are easily recognized from the graphs representing [2; 3; 2], [2; 3] and
[2; 3; 3].

Lists containing functions cannot be compared because F# equality is not defined for
functions.
For example:
[sin; cos] = [];;
-^^^
... The type ′(^a -> ^a ) when
^a : (static member Sin : ^a -> ^a)′ does not support
the ′equality′ constraint because it is a function type
 Ordering of lists
Lists of the same type are ordered lexicographically, provided there is an ordering defined on
the elements:

exactly when

for some k, where 0 ≤ k < min {m − 1, n − 1}.

There are two cases in this definition of xs < ys:

1. The list xs is a proper prefix of ys:

[1; 2; 3] < [1; 2; 3; 4];;
val it : bool = true

[′1′; ′2′; ′3′] < [′1′; ′2′; ′3′; ′4′];;

val it : bool = true

The examples illustrate comparisons of integer lists and character lists. Furthermore, the
empty list is smaller than any non-empty list:
[] < [1; 2; 3];;
val it : bool = true

[] < [[]; [(true,2)]];;

val it : bool = true

2. The lists agree on the first k elements and xk+1 < yk+1. For example:
[1; 2; 3; 0; 9; 10] < [1; 2; 3; 4];;
val it : bool = true

["research"; "articles"] < ["research"; "books"];;

val it : bool = true

because 0 < 4 and "articles" < "books".

The other comparison relations can be defined in terms of = and < as usual. For example:
[1; 1; 6; 10] >= [1; 2];;
val it : bool = false

The compare function is defined for lists, provided it is defined for the element type. For
example:
compare [1; 1; 6; 10] [1; 2];;
val it : int = -1
compare [1;2] [1; 1; 6; 10];;
val it : int = 1
4.2 Construction and decomposition of lists
 The cons operator
The infix operator :: (called “cons”) builds a list from its head and its tail as shown in
Figures 4.2 and 4.3 so it adds an element at the front of a list:
let x = 2::[3;4;5];;
val x : int list = [2; 3; 4; 5]

let y = ""::[];;
val y : string list = [""]

Fig ure 4.2 Graph for the list x:: xs

The operator associates to the right, so x0::x1:: xs means x0::(x1:: xs) where x0 and x1
have the same type and xs is a list with elements of that same type (cf. Figure 4.3) so we get,
for example:
let z = 2::3::[4;5];;
val z : int list = [2; 3; 4; 5]

Fig ure 4.3 Graph for the list x0 ::(x1 :: xs)

 List patterns
While the cons operator can be used to construct a list from a (head) element and a (tail) list,
it is also used in list patterns. List patterns and pattern matching for lists are used in the
subsequent sections to declare functions on lists by using the bindings of identifiers in a list
pattern obtained by matching a list to the pattern.
There is the list pattern [] for the empty list while patterns for non-empty lists are
constructed using the cons operator, that is, x::xs matches a non-empty list.

Fig ure 4.4 Graphs for list patterns

The patterns [] and x::xs denote the graphs in Figure 4.4. The pattern [] matches the
empty list only, while the pattern x::xs matches any non-empty list [x0;x1;…; xn−1]. The
latter matching gives the bindings x x0 and xs [x1;…; xn−1] of the identifiers x and xs,
as the list [x0;x1;…; xn−1] denotes the graph in Figure 4.5.

Fig ure 4.5 Graph for the list [x0 ; x1 ;…; xn−1 ]

For example, the execution of the declarations:

let x::xs = [1;2;3];;
val xs : int list = [2; 3]
val x : int = 1

will simultaneously bind x to the value 1 and xs to the value [2;3] by matching the value
[1;2;3] to the pattern x::xs.
A list pattern for a list with a fixed number of elements, for example, three, may be written
as x0::x1::x2::[] or in the shorter form [x0;x1;x2]. This pattern will match any list with
precisely three elements [x0;x1;x2], and the matching binds x0 to x0, x1 to x1, and x2 to x2.
For example:
let [x0;x1;x2] = [(1,true); (2,false); (3, false)];;
let [x0;x1;x2] = [(1,true); (2,false); (3, false)];;
—-^^^^^^^^^^
stdin(1,5): warning FS0025: Incomplete pattern matches on this
expression. For example, the value ′[_;_;_;_]′ may indicate a
case not covered by the pattern(s).
val x2 : int * bool = (3, false)
val x1 : int * bool = (2, false)
val x0 : int * bool = (1, true)

This generalizes to any fixed number of elements. (The F# compiler issues a warning because
list patterns with a fixed number of elements are in general not recommended, but the
bindings are, nevertheless, made.)
List patterns may have more structure than illustrated above. For example, we can construct
list patterns that match lists with two or more elements (e.g., x0::x1::xs), and list patterns
matching only non-empty lists of pairs (e.g., (y1,y2)::ys), and so on. For example:
let x0::x1::xs = [1.1; 2.2; 3.3; 4.4; 5.5];;
val xs : float list = [3.3; 4.4; 5.5]
val x1 : float = 2.2
val x0 : float = 1.1

let (y1, y2)::ys = [(1,[1]); (2, [2]); (3, [3]); (4,[4])];;

val ys : (int * int list) list =
[(2, [2]); (3, [3]); (4, [4])]
val y2 : int list = [1]
val y1 : int = 1

We shall see examples of more involved patterns in this chapter and throughout the book.
Note the different roles of the operator symbol :: in patterns and expressions. It denotes
decomposing a list into smaller parts when used in a pattern like x0::x1::xs, and it denotes
building a list from smaller parts in an expression like 0::[1; 2].
 Simple list expressions
In F# there are special constructs that can generate lists. In this section we will just introduce
the two simple forms of expressions called range expressions:

where b, e and s are expressions having number types.

The range expression [b .. e], where e ≥ b, generates the list of consecutive elements:

where n is chosen such that b + n ≤ e < b + n + 1. The range expression generates the empty
list when e < b.
For example, the list of integers from −3 to 5 is generated by:
[ -3 .. 5 ];;
val it : int list = [-3; -2; -1; 0; 1; 2; 3; 4; 5]

and a list of floats is, for example, generated by:

[2.4 .. 3.0 ** 1.7];;
val it : float list = [2.4; 3.4; 4.4; 5.4; 6.4]

Note that 3.0 ** 1.7 = 6.47300784.

The expression s in the range expression [b..s..e] is called the step. It can be positive or
negative, but not zero:

The generated list will be either ascending or descending depending on the sign of s. For
example, the descending list of integers from 6 to 2 is generated by:
[6 .. -1 .. 2];;
val it : int list = [6; 5; 4; 3; 2]

and the float-based representation of the list consisting of 0, π/2, π, π, 2π is generated by:
[0.0 .. System.Math.PI/2.0 .. 2.0*System.Math.PI];;
val it : float list =
[0.0; 1.570796327; 3.141592654; 4.71238898; 6.283185307]

An exception is raised if the step is 0:

> [0 .. 0 .. 0];;
System.ArgumentException: The step of a range cannot be zero.
Parameter name: step
......
Stopped due to error
 4.3 Typical recursions over lists
In this section we shall consider a collection of archetypical recursive function declarations
on lists.
 Function declarations with two clauses
Let us consider the function suml that computes the sum of a list of integers:

We get the recursion formula:

We define the value of the “empty” sum, that is, suml [], to be 0 and we arrive at a recursive
function declaration with two clauses:
let rec suml = function
| [] -> 0
| x::xs -> x + suml xs;;
val suml : int list -> int

In evaluating a function value for suml xs, F# scans the clauses and selects the first clause
where the argument matches the pattern. Hence, the evaluation of suml [1;2] proceeds as
follows:

This example shows that patterns are convenient in order to split up a function declaration
into clauses covering different forms of the argument. In this example, one clause of the
declaration gives the function value for the empty list, and the other clause reduces the
computation of the function value for a non-empty list suml(x::xs) to a simple operation
(addition) on the head x and the value of suml on the tail xs (i.e., suml xs), where the length of
the argument list has been reduced by one.
It is easy to see that an evaluation for suml [x0; …; xn−1] will terminate, as it contains
precisely n + 1 recursive calls of suml.
The above declaration is an example of a typical recursion schema for the declaration of
functions on lists.
 Function declarations with several clauses
One can have function declarations with any number (≥ 1) of clauses. Consider, for example,
the alternate sum of an integer list:

In declaring this function we consider three different forms of the argument:

1. empty list: altsum [] = 0

2. list with one element: altsum [x0] = x0
3. list with two or more elements:

Fig ure 4.6 List patterns for altsum declaration

These cases are covered by the patterns in Figure 4.6. Thus, the function can be declared by:
let rec altsum = function
| [] -> 0
| [x] -> x
| x0::x1::xs -> x0 - x1 + altsum xs;;
val altsum : int list -> int

altsum [2; -1; 3];;

val it : int = 6

It is left as an exercise to give a declaration for altsum containing only two clauses.
 Layered patterns
We want to define a function succPairs such that:

Using the pattern x0::x1::xs as in the above example we get the declaration
let rec succPairs = function
| x0 :: x1 :: xs -> (x0,x1) :: succPairs(x1::xs)
| _ -> [];;
val succPairs : ′a list -> (′a * ′a) list

This works OK, but we may get a smarter declaration avoiding the cons expression x1::xs in
the recursive call in the following way:

Fig ure 4.7 A pattern x0::(x1::_ as xs) containing a layered sub-pattern x1::_ as xs

let rec succPairs = function

| x0::(x1::_ as xs) -> (x0,x1) :: succPairs xs
| _ -> [];;
val succPairs : ′a list -> (′a * ′a) list

succPairs [1;2;3];;
val it : (int * int) list = [(1, 2); (2, 3)]

The pattern x1::_ as xs is an example of a layered pattern. It is part of the pattern shown in
Figure 4.7. A layered pattern has the general form:

with pattern pat and identifier id. A value val matches this pattern exactly when the value
matches the pattern pat. The matching binds identifiers in the pattern pat as usual with the
addition that the identifier id is bound to val. Matching the list [x0 ;x1; …] with the pattern
x0::(x1::_ as xs) will hence give the following bindings:

which is exactly what is needed in this case.

 Pattern matching on result of recursive call
The following example illustrates the use of pattern matching to split the result of a recursive
call into components. The function sumProd computes the pair consisting of the sum and the
product of the elements in a list of integers, that is:

The declaration is based on the recursion formula:

where

This gives the declaration:

let rec sumProd = function
| [] -> (0,1)
| x::rest ->
let (rSum,rProd) = sumProd rest
(x+rSum,x*rProd);;
val sumProd : int list -> int * int

sumProd [2;5];;
val it : int * int = (7, 10)

Another example is the unzip function that maps a list of pairs to a pair of lists:

The declaration of unzip looks as follows:

let rec unzip = function
| [] -> ([],[])
| (x,y)::rest ->
let (xs,ys) = unzip rest
(x::xs,y::ys);;
val unzip : (′a * ′b) list -> ′a list * ′b list

unzip [(1,"a");(2,"b")];;
val it : int list * string list = ([1; 2], ["a"; "b"])

The unzip function is found as List.unzip is the F# library.

 Pattern matching on pairs of lists
We want to declare a function mix that mixes the elements of two lists with the same length:

mix ([x0 ; x1 ; …; xn−1], [ y0 ; y1 ; …; yn−1])

= [x0; y0; x1; y1; …; xn−1; yn−1]

It is declared using pattern matching on the pair of lists:

let rec mix = function
| (x::xs,y::ys) -> x::y::(mix (xs,ys))
| ([],[]) -> []
| _ -> failwith "mix: parameter error";;
val mix : ′a list * ′a list -> ′a list

mix ([1;2;3],[4;5;6]);;
val it : int list = [1; 4; 2; 5; 3; 6]

The corresponding higher-order function is defined using a match expression:

let rec mix xlst ylst =
match (xlst,ylst) with
| (x::xs,y::ys) -> x::y::(mix xs ys)
| ([],[]) -> []
| _ -> failwith "mix: parameter error";;
val mix : ′a list -> ′a list -> ′a list

mix [1;2;3] [4;5;6];;

val it : int list = [1; 4; 2; 5; 3; 6]
 4.4 Polymorphism
In this section we will study some general kinds of polymorphism, appearing frequently in
connection with lists. We will do that on the basis of three useful list functions that all can be
declared using the same structure of recursion as shown in Section 4.3.
 List membership
The member function for lists determines whether a value x is equal to one of the elements in
a list [y0;y1;…;yn−1], that is:

Since no x can be a member of the empty list, we arrive at the declaration:

let rec isMember x = function
| y::ys -> x=y || (isMember x ys)
| [] -> false;;
val isMember : ′a -> ′a list -> bool when ′a : equality

The function isMember can be useful in certain cases, but it is not included in the F# library.
The annotation ′a : equality indicates that ′a is an equality type variable; see Section
2.10. The equality type is inferred from the expression x=y. It implies that the function
isMember will only allow an argument x where the equality operator = is defined for values of
the type of x. A type such as int * (bool * string) list * int list is an equality type,
and the function can be applied to elements of this type.
 Append and reverse. Two built-in functions
The infix operator @ (called ‘append′) joins two lists of the same type:

and the function List.rev (called “reverse”) reverses a list:

These functions are predefined in F#, but their declarations reveal important issues and are
therefore discussed here. The operator @ is actually the infix operator corresponding to the
library function List.append.
The declaration of the (infix) function @ is based on the recursion formula:

This leads to the declaration:

let rec (@) xs ys =
match xs with
| [] -> ys
| x::xs′ -> x::(xs′ @ ys);;
val (@) : ′a list -> ′a list -> ′a list

The evaluation of append decomposes the left-hand list into its elements, that are afterwards
‘cons′ed′ onto the right-hand list:

The evaluation of xs @ ys comprises m + 1 pattern matches plus m cons’es where m is the

length of xs.
The notion of polymorphism is very convenient for the programmer because one need not
write a special function for appending, for example, integer lists and another function for
appending lists of integer lists, as the polymorphic append function is capable of both:
[1;2] @ [3;4];;
val it : int list = [1; 2; 3; 4]

[[1];[2;3]] @ [[4]];;
val it : int list list = [[1]; [2; 3]; [4]]

The operators :: and @ have the same precedence (5) and both associate to the right. A
mixture of these operators also associates to the right, so [1]@2::[3], for example, is
interpreted as [1]@(2::[3]), while 1::[2]@[3] is interpreted as 1::([2]@[3]):
[1] @ 2 :: [3];;
val it : int list = [1; 2; 3]

1 :: [2] @ [3];;
val it : int list = [1; 2; 3]

For the reverse function rev, where

we have the recursion formula:

because

This leads immediately to a naive declaration of a reverse function:

let rec naiveRev xls =
match xls with
| [] -> []
| x::xs -> naiveRev xs @ [x];;
val naiveRev : ′a list -> ′a list

This declaration corresponds directly to the recursion formula for rev: the tail list xs is
reversed and the head element x is inserted at the end of the resulting list — but it may be
considered naive as it gives a very inefficient evaluation of the reversed list:

We will make a much more efficient declaration of the reverse function in a later chapter
(Page 208). The library function List.rev is, of course, implemented using an efficient
declaration.
 4.5 The value restrictions on polymorphic expressions
The type system and type inference of F# is very general and flexible. It has, however, been
necessary to make a restriction on the use of polymorphic expressions in order to ensure type
correctness in all situations.
The formulation of this restriction is based on the concept of value expressions. A value
expression is an expression that is not reduced further by an evaluation, that is, it has already
the same form as its value. The following expressions are hence value expressions:
[] Some [] (5,[]) (fun x -> [x])

while
List.rev [] [] @ []

do not qualify as a value expression as they can be further evaluated. Note that a function
expression (a closure) is considered a value expression because it is only evaluated further
when applied to an argument.
The restriction applies to the expression exp in declarations
let id = exp

At top level, polymorphic expressions are allowed only if they are value expressions.
Polymorphic expressions can be used freely for intermediate results.

Hence F# allows values of polymorphic types, such as the empty list [], the pair (5,[[]]) or
the function (fun x -> [x]):
let z = [];;
val z : ′a list

(5,[[]]);;
val it : int * ′a list list = (5, [[]]

let p = (fun x -> [x]);;

val p : ′a -> ′a list

On the other hand, the following is refused at top level:

List.rev [];;
^^^^^^^^^^
stdin(86,1): error FS0030: Value restriction.
The value ′it′ has been inferred to have generic type

The restriction on polymorphic expressions may be paraphrased as follows:

 All monomorphic expressions are OK, even non-value expressions,
all value expressions are OK, even polymorphic ones, and
at top-level, polymorphic non-value expressions are forbidden,

where the type of a monomorphic expression does not contain type variables, that is, it is a
monomorphic type.
The rationale for these restrictions will only become clear much later when imperative
features of F# are introduced in Chapter 8. In the meantime, we just have to accept the
restrictions, and they will really not do us much harm.
Remark: A list expression a0 :: a1 :: ·s :: ak :: [ak+1,…,an−1] containing values a0, a2, …, an−1 is
considered a value expression with the value [a0, a1,…,an−1].
 4.6 Examples. A model-based approach
In this section we will introduce a model-based approach to functional programming by
means of two examples. The goal is to get a program directly reflecting the problem
formulation. An important step in achieving this goal is to identify names denoting key
concepts in the problem and to associate F# types with these names. We shall return to these
two examples in the next chapter when we have a richer set of collection types with associated
library functions.
 Example: Cash register
Consider an electronic cash register that contains a data register associating the name of the
article and its price to each valid article code. A purchase comprises a sequence of items,
where each item describes the purchase of one or several pieces of a specific article.
The task is to construct a program that makes a bill of a purchase. For each item the bill
must contain the name of the article, the number of pieces, and the total price, and the bill
must also contain the grand total of the entire purchase.
Article code and article name are central concepts that are named and associated with a
type:
type ArticleCode = string;;
type ArticleName = string;;

where the choice of the string type for ArticleCode is somewhat arbitrary. An alternative
choice could be the int type.
The register associates article name and article price with each article code, and we model
a register by a list of pairs. Each pair has the form:

where ac is an article code, aname is an article name, and aprice is an article price. We choose
(non-negative) integers to represent prices (in the smallest currency unit):
type Price = int;; // pr where pr >= 0

and we get the following type for a register:

type Register = (ArticleCode * (ArticleName*Price)) list;;

The following declaration names a register:

let reg = [("a1",("cheese",25));
("a2",("herring",4));
("a3",("soft drink",5)) ];;

A purchase comprises a list of items, where each item comprises a pair:

describing a number of pieces np (that is a non-negative integer) purchased of an article with

code ac:
type NoPieces = int;; // np where np >= 0
type Item = NoPieces * ArticleCode;;
type Purchase = Item list;;

The following declaration names a purchase:

let pur = [(3,"a2"); (1,"a1")];;
 A bill comprises an information list infos for the individual items and the grand total sum,
and this composite structure is modelled by a pair:

where each element in the list infos is a triple

of the number of pieces np, the name aname, and the total price tprice of a purchased article:
type Info = NoPieces * ArticleName * Price;;
type Infoseq = Info list;;
type Bill = Infoseq * Price;;

The following value is an example of a bill:

([(3,"herring",12); (1,"cheese",25)],37)

The function makeBill computes a bill given a purchase and a register and it has the type:
makeBill: Register -> Purchase -> Bill

In this example, it is convenient to declare a auxiliary function:

findArticle: ArticleCode -> Register -> ArticleName * Price

to find the article name and price in the register for a given article code. This will make the
declaration for the function makeBill easier to comprehend. An exception is raised when no
article with the given code occurs in the register:
let rec findArticle ac = function
| (ac′,adesc)::_ when ac=ac′ -> adesc
| _::reg -> findArticle ac reg
| _ ->
failwith(ac + " is an unknown article code");;
val findArticle : string -> (string * ′a) list -> ′a

Then the bill is made by the function:

let rec makeBill reg = function
| [] -> ([],0)
| (np,ac)::pur -> let (aname,aprice) = findArticle ac reg
let tprice = np*aprice
let (billtl,sumtl) = makeBill reg pur
((np,aname,tprice)::billtl,tprice+sumtl);;
val makeBill :
(string * (′a * int)) list -> (int * string) list
-> (int * ′a * int) list * int
makeBill reg pur;;
val it : (int * string * int) list * int =
([(3, "herring", 12); (1, "cheese", 25)], 37)

The declaration of makeBill uses the pattern introduced in Section 4.3 to decompose the
value of the recursive call.
 Note that the F# system infers a more general type for the makeBill function than the type
given in our model. This is, however, no problem as the specified type is an instance of the
inferred type – makeBill has the specified type (among others).
 Example: Map colouring
We shall now consider the problem of colouring a map in a way so that neighbouring
countries get different colours. We will provide a model using named types, similar to what
we did in the previous example. But the map colouring problem is more complex than the
cash register example, and we use it to illustrate functional decomposition by devising a
collection of simple well-understood functions that can be composed to solve the problem.
This problem is a famous mathematical problem and it has been proved that any
(reasonable) map can be coloured by using at most four different colours. We will not aim
for an “optimal” solution. Neither will we consider the trivial solution where each country
always gets its own separate colour. We will assume that each country is connected. This is an
over-simplification as Alaska and Kaliningrad, for example, are not connected to other
regions of the United States and Russia, respectively.
A country is represented by its name, that is a string, and a map is represented by a
neighbour relation, that is represented by a list of pairs of countries with a common border:
type Country = string;;
type Map = (Country * Country) list;;

Consider the map in Figure 4.8 with four countries “a”, “b”, “c”, and “d”, where the country
“a” has the neighbouring countries “b” and “d”, the country “b” has the neighbouring country
“a”, and so on. The F# value for this map is given by the declaration of exMap.

Fig ure 4.8 Colouring problem with 4 countries

A colour on a map is represented by the set of countries having this colour, and a colouring
is described by a list of mutually disjoint colours:
type Colour = Country list;;
type Colouring = Colour list;;

The countries of the map in Figure 4.8 may hence be coloured by the colouring:
[["a";"c"]; ["b"; "d"]]

This colouring has two colours ["a";"c"] and ["b"; "d"], where the countries “a” and “c”
get one colour, while the countries “b” and “d” get another colour.
An overview of the model is shown in Figure 4.9 together with sample values. This figure
also contains meta symbols used for the various types, as this helps in achieving a consistent
naming convention throughout the program.
 Fig ure 4.9 A Data model for map colouring problem

Our task is to declare a function:

colMap: Map -> Colouring

that can generate a colouring of a given map. We will express this function as a composition
of simple functions, each with a well-understood meaning. These simple functions arise from
the algorithmic idea behind the solutions to the problem. The idea we will pursue here is the
following: We start with the empty colouring, that is, the empty list containing no colours.
Then we will gradually extend the actual colouring by adding one country at a time.

Fig ure 4.10 Algorithmic idea

We illustrate this algorithmic idea on the map in Figure 4.8, with the four countries: “a”,
“b”, “c” and “d”. The four main algorithmic steps (one for each country) are shown in Figure
4.10. We give a brief comment to each step:

1. The colouring containing no colours is the empty list.

2. The colour ["a"] cannot be extended by "b" because the countries "a" and "b" are
neighbours. Hence the colouring should be extended by a new colour ["b"].
3. The colour ["a"] can be extended by "c" because "a" and "c" are not neighbours.
4. The colour ["a","c"] can not be extended by "d" while the colour ["b"] can be
extended by "d".

The task is now to make a program where the main concepts of this algorithmic idea are
directly represented. The concepts emphasized in the above discussion are:

Test whether a colour can be extended by a country for a given map.

Test whether two countries are neighbours in a given map.
Extend a colouring by a country for a given map.

The function specification of each of the main concepts documents the algorithmic idea.
These specifications are shown in Figure 4.11. We have added the specification of a function
countries for extracting the list of countries occurring in a given map and the specification
of a function colCntrs which gives the colouring for given country list and map.

Fig ure 4.11 Functional break-down for map colouring problem

We now give a declaration for each of the functions specified in Figure 4.11.

1. First we declare a predicate (i.e., a truth-valued function) areNb to determine for a given
map whether two countries are neighbours:
let areNb m c1 c2 =
isMember (c1,c2) m || isMember (c2,c1) m;;

This declaration makes use of the isMember-function declared in Section 4.4.

2. Next we declare a predicate to determine for a given map whether a colour can be
extended by a country:
let rec canBeExtBy m col c =
match col with
| [] -> true
| c′::col′ -> not(areNb m c′ c) && canBeExtBy m col′ c;;

canBeExtBy exMap ["c"] "a";;

val it : bool = true

canBeExtBy exMap ["a"; "c"] "b";;

val it : bool = false

3. Our solution strategy is to insert the countries of a map one after the other into a
colouring, starting with the empty one. To this end we declare a function extColouring
that for a given map extends a partial colouring by a country:
let rec extColouring m cols c =
match cols with
| [] -> [[c]]
| col::cols′ -> if canBeExtBy m col c
then (c::col)::cols′
else col::extColouring m cols′ c;;

extColouring exMap [] "a";;

val it : string list list = [["a"]]
 extColouring exMap [["c"]] "a";;
val it : string list list = [["a"; "c"]]

extColouring exMap [["b"]] "a";;

val it : string list list = [["b"]; ["a"]]

Note that the first of the three examples exercises the base case of the declaration, the
second example the then-branch, and the last example the else-branch (the recursion
and the base case).
4. In order to complete our task, we declare a function to extract a list of countries without
repeated elements from a map and a function to colour a list of countries given a map:
let addElem x ys = if isMember x ys then ys else x::ys;;

let rec countries = function

| [] -> []
| (c1,c2)::m -> addElem c1 (addElem c2 (countries m));;

let rec colCntrs m = function

| [] -> []
| c::cs -> extColouring m (colCntrs m cs) c;;

The function giving a colouring for a given map is declared by combination of the
functions colCntrs and countries.
let colMap m = colCntrs m (countries m);;

colMap exMap;;
val it : string list list = [["c"; "a"]; ["b"; "d"]]
 Comments
In these two examples we have just used types introduced previously in this book, and some
comments could be made concerning the adequacy of the solutions. For example, modelling a
data register by a list of pairs does not capture that each article has a unique description in the
register, and modelling a colour by a list of countries does not capture the property that the
sequence in which countries occur in the list is irrelevant. The same applies to the property
that repeated occurrences of a country in a colour are irrelevant.
In Chapter 5 we shall introduce maps and sets and we shall give more suitable models and
solutions for the two examples above.
 Summary
In this chapter we have introduced the notions of lists and list types, and the notion of list
patterns. A selection of typical recursive functions on lists were presented, and the notions of
polymorphic types and values were studied. Furthermore, we have introduced a model-based
approach to functional programming, where important concepts are named and types are
associated with the names.
 Exercise
4.1 Declare function upto: int -> int list such that upto n = [1; 2; …; n].
4.2 Declare function downto1: int -> int list such that the value of downto1 n is the
list [n; n − 1; …; 1].
4.3 Declare function evenN: int -> int list such that evenN n generates the list of the
first n non-negative even numbers.
4.4 Give a declaration for altsum (see Page 76) containing just two clauses.
4.5 Declare an F# function rmodd removing the odd-numbered elements from a list:

4.6 Declare an F# function to remove even numbers occurring in an integer list.

4.7 Declare an F# function multiplicityx xs to find the number of times the value x
occurs in the list xs.
4.8 Declare an F# function split such that:

4.9 Declare an F# function zip such that:

The function should raise an exception if the two lists are not of equal length.
4.10 Declare an F# function prefix: ′a list -> ’a list -> bool when a :
equality. The value of the expression prefix [x0 ;x1 ; …; xm] [y0 ;y1 ; …; yn ] is true if
m ≤ n and xi = yi for 0 ≤ i ≤ m, and false otherwise.
4.11 A list of integers [x0; x1; … ; xn−1] is weakly ascending if the elements satisfy:

or if the list is empty. The problem is now to declare functions on weakly ascending lists.

1. Declare an F# function count: int list * int -> int, where count(xs, x) is the
number of occurrences of the integer x in the weakly ascending list xs.
2. Declare an F# function insert: int list * int -> int list, where the value of
insert(xs, x) is a weakly ascending list obtained by inserting the number x into the
weakly ascending list xs.
3. Declare an F# function intersect: int list * int list -> int list, where the
value of intersect(xs, xs′) is a weakly ascending list containing the common
elements of the weakly ascending lists xs and xs′. For instance:

4. Declare an F# function plus: int list * int list -> int list, where the value
of plus(xs, xs′) is a weakly ascending list, that is the union of the weakly ascending
lists xs and xs′. For instance:
 5. Declare an F# function minus: int list * int list -> int list, where the
value of minus(xs, xs′) is a weakly ascending list obtained from the weakly ascending
list xs by removing those elements, that are also found in the weakly ascending list xs
′. For instance:

4.12 Declare a function sum(p, xs) where p is a predicate of type int -> bool and xs is a
list of integers. The value of sum(p, xs) is the sum of the elements in xs satisfying the
predicate p. Test the function on different predicates (e.g., p(x) = x > 0).
4.13 Naive sort function:

1. Declare an F# function finding the smallest element in a non-empty integer list.

2. Declare an F# function delete: int * int list -> int list, where the value of
delete(a, xs) is the list obtained by deleting one occurrence of a in xs (if there is
one).
3. Declare an F# function that sorts an integer list so that the elements are placed in
weakly ascending order.

Note that there is a much more efficient sort function List.sort in the library.
4.14 Declare a function of type int list -> int option for finding the smallest
element in an integer list.
4.15 Declare an F# function revrev working on a list of lists, that maps a list to the
reversed list of the reversed elements, for example:
revrev [[1;2];[3;4;5]] = [[5;4;3];[2;1]]
4.16 Consider the declarations:
let rec f = function
| (x, []) -> []
| (x, y::ys) -> (x+y)::f(x-1, ys);;
let rec g = function
| [] -> []
| (x,y)::s -> (x,y)::(y,x)::g s;;
let rec h = function
| [] -> []
| x::xs -> x::(h xs)@[x];;

Find the types for f, g and h and explain the value of the expressions:

1. f(x, [y0 , y1 , … , yn−1 ]), n ≥ 0

2. g[(x0 , y0 ) , (x1 , y1 ) , … , (xn−1 , yn−1 )], n≥0
3. h[x0 , x1 , … , xn−1 ], n ≥ 0

4.17 Consider the declaration:

let rec p q = function
 | [] -> []
| x::xs -> let ys = p q xs
if q x then x::ys else ys@[x];;

Find the type for p and explain the value of the expression:

4.18 Consider the declaration:

let rec f g = function
| [] -> []
| x::xs -> g x :: f (fun y -> g(g y)) xs;;

Find the type for f and explain the value of the expression:

4.19 Evaluation of the expression areNb m c1 c2 may traverse the map m twice. Explain
why and give an alternative declaration for areNb which avoids this problem.
4.20 Most of the auxiliary functions for the map-colouring program just assume an
arbitrary, but fixed, map. The function canBeExtBy, for example, just passes m on to
areNb, which again passes m on to isMember. The program can therefore be simplified by
declaring (most of) the auxiliary functions locally as sketched here:
...
let colMap m =
let areNb c1 c2 = ...
let canBeExtBy col c = ...
...

Revise the program by completing this skeleton.

4.21 Revise the map-colouring program so that it can cope with countries which are
islands (such as Iceland) having no neighbours.
4.22 We represent the polynomial a0 + a1 · x + ... + an · xn with integer coefficients a0, a1,
..., an by the list [a0, a1, ..., an]. For instance, the polynomial x3 + 2 is represented by the list
[2,0,0,1].

1. Declare an F# function for multiplying a polynomial by a constant.

2. Declare an F# function for multiplying a polynomial Q(x) by x.
3. Declare infix F# operators for addition and multiplication of polynomials in the
chosen representation. The following recursion formula is useful when defining the
multiplication:

4. Declare an F# function to give a textual representation for a polynomial.

4.23 A dating bureau has a file containing name, telephone number, sex, year of birth and
themes of interest for each client. You may make a request to the bureau stating your own
sex, year of birth and themes of interest and get a response listing all matching clients, that
is, clients with different sex, a deviation in age less than 10 years and with at least one
common theme of interest. The problem is to construct a program for generating the
responses from the dating bureau.
 5
Collections: Lists, maps and sets

Functional languages make it easy to express standard recursion patterns in the form of
higher-order functions. A collection of such higher-order functions on lists, for example,
provides a powerful library where many recursive functions can be obtained directly by
application of higher-order library functions. This has two important consequences:

1. The functions in the library correspond to natural abstract concepts and conscious use of
them supports high-level program design, and
2. these functions support code reuse because you can make many functions simply by
applying library functions.

In this chapter we shall study libraries for lists, sets and maps, which are parts of the
collection library of F#. This part of the collection library is studied together since:

It constitutes the immutable part of the collection library. The list, set and map collections
are finite collections programmed in a functional style.
There are many similarities in the corresponding library functions.

This chapter is a natural extension of Chapter 4 since many of the patterns introduced in
that chapter correspond to higher-order functions for lists and since more natural program
designs can be given for the two examples in Section 4.6 using sets and maps.
We will focus on the main concepts and applications in this book, and will deliberately not
cover the complete collection library of F#. The functions of the collection library do also
apply to (mutable) arrays. We address this part in Section 8.10.
 5.1 Lists
This section describes the library functions map, various library functions using a predicate
on list elements plus the functions fold and foldBack. Each description aims to provide the
following:

1. An intuitive understanding of the objective of the function.

2. Examples of use of the function.

The actual declarations of the library functions are not considered as we want to
concentrate on how to use these functions in problem solving. Declarations of fold and
foldBack are, however, of considerable theoretical interest and are therefore studied in the
last part of the section. An overview of the List-library functions considered in this section is
found in Table 5.1.
Table 5.1 A selection of functions from the List library
 The map function
The library function
List.map: (′a -> ′b) -> ′a list -> ′b list

works as follows:

In words:

The function application List.map f is the function that applies the function f
to each element x0, x1, … , xn−1 in a list [x0; x1; ·s; xn−1 ]

It is easy to use List.map:

The function addFsExt adds the F# file extension ”.fs” to every string in a list of file
names.
The function intPairToRational converts every integer pair in a list to the string of a
rational number on the basis of the declarations in Section 3.7.
The function areaList computes the area of every shape in a list on the basis of the
declarations in Section 3.8.
let addFsExt = List.map (fun s -> s + ".fs");;
val addFsExt : (string list -> string list)

let intPairToRational = List.map (toString << mkQ);;

val intPairToRational : ((int * int) list -> string list)

let areaList = List.map area;;

val areaList : (shape list -> float list)

since

addFsExt applies the function that concatenates the suffix “.fs” to a string, to every
element in a string list,
intPairToRational applies the function that converts an integer pair to the string
representation of the corresponding rational number to every element in a list of integer
pairs, and
areaList applies the area function to every element in a shape list.

The functions work as follows:

addFsExt ["ListPrograms"; "AuxiliaryPrograms"];;
val it : string list =
["ListPrograms.fs"; "AuxiliaryPrograms.fs"]

intPairToRational [(2,6); (20,-8); (-12,-4)];;

val it : string list = ["1/3"; "-5/2"; "3/1"]
areaList [Circle 2.0; Square 2.0; Triangle(2.0, 3.0, 4.0)];;
val it : float list = [12.56637061; 4.0; 2.90473751]

Alternative ways of declaring intPairToRational using List.map are

let intPairToRational = List.map (fun p -> toString(mkQ p));;
let intPairToRational ps =
List.map (fun p -> toString(mkQ p)) ps;;

where fun p -> toString(mkQ p) is an expansion of the function composition operator in

toString << mkQ and ps is used as explicit list argument in the last declaration. Explicit list
arguments could also be used in declarations of addFsExt and areaList.
 Functions using a predicate on the list elements
The F# library contains a large number of functions using a predicate of type ′a -> bool on
elements in a list of type ′a list.
We consider some of these functions here, namely (cf. Table 5.1):
List.exists : (′a -> bool) -> ′a list -> bool
List.forall : (′a -> bool) -> ′a list -> bool
List.tryFind : (′a -> bool) -> ′a list -> ′a option
List.filter : (′a -> bool) -> ′a list -> ′a list

The value of the expression

is true, if p(xk) = true holds for some list element xk, and false otherwise.

The value of the expression

is true, if p(xk) = true holds for all list elements xk, and false otherwise.
The value of the expression

is Some xk for a list element xk with p(xk) = true, or None if no such element exists.
The value of the expression

is the list of those list elements xk where p(xk) = true.

Note that the evaluation of the expression

does not terminate if the evaluation of the expression p(xk) does not terminate for some k,
where 0≤k ≤n−1 and if p(xj) = false for all j where 1 ≤j <k. A similar remark will apply to the
other functions using a predicate on list elements.
Simple applications of the functions are:
List.exists (fun x -> x>=2) [1;3;1;4];;
val it : bool = true

List.forall (fun x -> x>=2) [1;3;1;4];;

val it : bool = false

List.tryFind (fun x -> x>3) [1;5;-2;8];;

val it : int option = Some 5

List.filter (fun x -> x>3) [1;5;-2;8];;

val it : int list = [5; 8]
The function isMember (cf. Section 4.4) can be declared using List.exists:

let isMember x xs = List.exists (fun y -> y=x) xs;;

val isMember : ′a -> ′a list -> bool when ′a : equality

isMember (2,3.0) [(2, 4.0) ; (3, 7.0)];;

val it : bool = false

isMember "abc" [""; "a"; "ab"; "abc"];;

val it : bool = true
 The functions fold and foldBack
The library functions List.fold and List.foldBack are very powerful and rather useful
in many circumstances, but they are somewhat difficult to understand at first glance. To ease
the understanding we use a rather naive, almost grotesque, example to convey the ideas
behind these functions.
We consider small cheeses and a round package to contain small cheeses:

Cheeses and packages are considered elements of type cheese and package. A package may
contain zero or more cheeses.
The function

packs an extra cheese into a package:

The function List.fold can be applied to the function packCheese, a start package and a list
of cheeses. It uses packCheese to pack the elements of the list (the cheeses) into the package
one after the other – starting with the given start package:

This is a special case of the general formula:

with

because we can identify the sub-expressions on the right-hand side of the general formula in
our special case:

and

The function of List.fold can be expressed in words as follows:

The evaluation of List.fold f e [x0; x1 ; … ; xn−1 ] accumulates the list

 elements x0, x1, … , xn−1 using the accumulation function f and the start value e

One also says that the function f is folded over the list [x0; x1 ; … ; xn−1 ] starting with the
value e.
The type of List.fold is:

with list element type ′b and accumulator type ′a.

When applying List.fold one has to look for the following entities:
List element type ′b corresponding to cheese in the example
Accumulator type ′a corresponding to package in the example
Accumulator function f corresponding to packCheese in the example
Start value e corresponding to the empty package in the example
and we have to arrange the parameters in the accumulator function to suit the type of
List.fold.

As an example we consider a list vs = [v0; …; vn−1] of geometric vectors in the plane (see
Section 3.3), where vi is a pair (xi,yi) of floats, for 0 ≤ i < n. We want to compute the sum of
the norms of the vectors in vs using the norm function declared as follows in Section 3.3:
let norm(x:float,y:float) = sqrt(x*x+y*y);;
val norm : float * float -> float

This is a case for applying List.fold with:

List element type: float * float
Accumulator type: float
Accumulator function: fun s (x,y) -> s + norm(x,y)
Start value: 0.0

This leads to the declaration:

let sumOfNorms vs =
List.fold (fun s (x,y) -> s + norm(x,y)) 0.0 vs;;
val sumOfNorms : (float * float) list -> float

let vs = [(1.0,2.0); (2.0,1.0); (2.0, 5.5)];;

val vs : (float * float) list =
[(1.0, 2.0); (2.0, 1.0); (2.0, 5.5)]

sumOfNorms vs;;
val it : float = 10.32448591

The length function on lists can be defined using List.fold with

List element type: ′a
Accumulator type: int
Accumulator function: fun e _ -> e + 1
Start value: 0
This leads to the declaration:
let length lst = List.fold (fun e _ -> e+1) 0 lst;;
val length : ′a list -> int

length [[1;2];[];[3;5;8];[-2]];;
val it : int = 4

Applying fold to the following version of “cons”:

where the parameters are interchanged, gives a declaration of the reverse function for lists:
let rev xs = List.fold (fun rs x -> x::rs) [] xs;;
val rev : ′a list -> ′a list

rev [1;2;3];;
val it : int list = [3; 2; 1]

The function List.foldBack is similar to List.fold but the list elements are accumulated in
the opposite order. The type of List.foldBack is:
List.foldBack: (′a -> ′b -> ′b) -> ′a list -> ′b -> ′b

and the general formula is:

List.foldBack g [x0;x1; . . . ;xn−1] e = g x0 (g x1 (. . . (g xn−1 e) . . .))

We may use our “cheese” example also in this case with a modified accumulation function:
cheesePack: cheese -> package -> package

where

The function List.foldBack can be applied to the function cheesePack, a list of cheeses and
a start package. It uses cheesePack to pack the elements of the list (the cheeses) taken in
reverse order into the package:

This is a special case of the general formula with:

g = cheesePack

because we can identify the sub-expressions in the right-hand side of the general formula in
our special case:
and

The function of List.foldBack can be expressed in words as follows:

The evaluation of List.foldBack g [x0; x1 ; … ; xn−1 ] e accumulates the list

elements in reverse order xn−1, … , x1, x0 using the accumulation function g and
the start value e
When applying List.foldBack one has to look for the following entities:
List element type ′a corresponding to cheese in the example
Accumulator type ′b corresponding to package in the example
Accumulator function g corresponding to cheesePack in the example
Start value e corresponding to the empty package in the example
and we have to arrange the parameters in the accumulator function to suit the type of
List.foldBack.

Using List.foldBack we may define an alternative “sum of norms” function. Element and
accumulator types can be used unchanged, but the parameters in the accumulator function
must be interchanged. This gives the following declaration:
let backSumOfNorms vs =
List.foldBack (fun (x,y) s -> s + norm(x,y)) vs 0.0;;
val backSumOfNorms : (float * float) list -> float

This function will work like the previous sumOfNorms but the norms are added in the opposite
order, starting with the norm of the last vector in the list.
Applying List.foldBack on the “cons” operator:

gives the append function:

let app ys zs = List.foldBack (fun x xs -< x::xs) ys zs;;
val app : ′a list -< ′a list -< ′a list
app [1;2;3] [4;5;6];;
val it : int list = [1; 2; 3; 4; 5; 6]

The unzip function on Page 77 can be obtained using foldBack with the following data:
List element type ′a * ′b
Accumulator type ′a list * ′b list
Accumulator function fun (x,y) (xs,ys) -> (x::xs,y::ys)
Start value ([],[])
This gives the declaration
let unzip zs = List.foldBack
(fun (x,y) (xs,ys) -> (x::xs,y::ys))
zs
([],[]);;
val unzip : (′a * ′b) list -> ′a list * ′b list

unzip [(1,"a");(2,"b")];;
al it : int list * string list = ([1; 2], ["a"; "b"])

A similar construction using List.fold gives a revUnzip function where the resulting lists
are reversed:
let revUnzip zs =
List.fold (fun (xs,ys) (x,y) -> (x::xs,y::ys)) ([],[]) zs;;
val revUnzip : (′a * ′b) list -> ′a list * ′b list

revUnzip [(1,"a");(2,"b")];;
val it : int list * string list = ([2; 1], ["b"; "a"])

The prefix version of an infix operator can be used as argument in fold and foldBack:
List.fold (+) 0 [1; 2; 3];;
val it : int = 6

List.foldBack (+) [1; 2; 3] 0;;

val it : int = 6

These expression compute ((0 + 1) + 2) + 3 and 1 + (2 + (3 + 0)), but the results are
equal because + is a commutative operator: a + b = b + a.
A difference in using fold or foldBack shows up when using a non-commutative operator,
that is:
List.fold (-) 0 [1; 2; 3];;
val it : int = -6

List.foldBack (-) [1; 2; 3] 0;;

val it : int = 2

These expressions use the functions:

fun e x -> e - x
fun x e -> x - e

and we get

The map function can be declared using foldBack:

let map f xs = List.foldBack (fun x rs -> f x :: rs) xs [];;

val map : (′a -> ′b) -> ′a list -> ′b list
map (fun x -> x+1) [0; 1; 2];;
val it : int list = [1; 2; 3]

Remark
A function declared by means of fold or foldBack will always scan the whole list. Thus, the
following declaration for the exists function
let existsF p =
List.fold (fun b -> (fun x -> p x || b)) false;;
val existsF : (′a -> bool) -> (′a list -> bool)

will not behave like the function List.exists with regard to non-termination: It will give a
non-terminating evaluation if the list contains any element where the evaluation of the
predicate p does not terminate, while the library function List.exists may terminate in this
case as it does not scan the list further when an element satisfying the predicate has been
found. So it is not considered a good idea to use fold or foldBack to declare functions like
exists or find (cf. Page 95) as these functions need not scan the whole list in all cases.
 Declarations of fold and foldBack
The list functions fold and foldBack are defined on Pages 97 and 99 by the formulas:

A recursion formula for fold is obtained by observing that:

where e′ = f e x0. The expression on the right-hand side is equal to:

and we get the recursion formula:

A recursion formula for foldBack is obtained by observing that the subexpression:

on the right-hand side of the formula for foldBack is equal to:

and we get the recursion formula:

These recursion formulas lead to the declarations:

let rec fold f e = function
| x::xs -> fold f (f e x) xs
| [] -> e;;

let rec foldBack g xlst e =

match xlst with
| x::xs -> g x (foldBack g xs e)
| [] -> e;;

The evaluation of a function value fold f e [x0; x1 ; … ; xn−1 ] proceeds as follows applying
f in each evaluation step without building any large expression:

The evaluation of foldBack g [x0; x1 ; … ; xn−1 ] e first builds a large expression:

and this expression is then evaluated “inside-out” using repeated calls of g:

The evaluation of fold is obviously much more efficient than the evaluation of foldBack, so
fold should be preferred whenever possible. The List.foldBack function in the library is
more efficient than the above foldBack but List.fold is still more efficient.
 5.2 Finite sets
In solving programming problems it is often convenient to use values that are finite sets of
form { a1, a2, …, an } with elements a1, … , an from some set A. The notion of a set provides
a useful abstraction in cases where we have an unordered collection of elements where
repetitions among the elements are of no concern.
This section introduces the set concept and operations on sets in F# on the basis of the
library Set. The focus is on the principal issues so just a small part of the available operations
will be covered. Please consult the on-line documentation (in [9]) for an overview of the
complete Set library.
 The mathematical set concept
A set (in mathematics) is a collection of elements like

where it is possible to decide whether a given value is in the set. For example, Alice is not in
the set { Bob, Bill, Ben } and 7 is in the set { 1,3,5,7,9}, also written:

The empty set containing no element is written { } or ∅.

Since the order in which elements are enumerated in a set is of no concern, and repetitions
among members of a set is of no concern either, the following expressions denote the same
set:

The above examples are all finite sets; but sets may be infinite and examples are the set of
all natural numbers = { 0, 1, 2, …} and the set of all real numbers .
A set A is a subset of a set B, written A ⊆ B, if all the elements of A are also elements of B,
for example

Furthermore, two sets A and B are equal, if they are both subsets of each other:

that is, two sets are equal if they contain exactly the same elements.
The subset of a set A that consists of those elements satisfying a predicate p can be
expressed using a set-comprehension { x A | p(x) }. For example, the set { 1, 3, 5, 7, 9}
consists of the odd natural numbers that are smaller than 11:

If it is clear from the context from which set A the elements of the set-comprehension
originate, then we use the simplified notation: { x | p(x) }.

Fig ure 5.1 Venn diagrams for (a) union, (b) intersection and (c) difference

Some of the standard operations on sets are union: A B, intersection A B and difference
A B:

that is, A B is the set of elements that are in at least one of the sets A and B, A B is the set of
elements that are in both A and B, and A B is the subset of the elements from A that are not in
B. These operations are illustrated using Venn diagrams in Figure 5.1. For example:
 Sets in F#
The Set library of F# supports finite sets of elements of a type where ordering is defined, and
provides efficient implementations for a rich collection of set operations. The implementation
is based on a balanced binary tree representation of a set and this is why an ordering of the
elements is required (but we will not consider such implementation details in this section).
Consider the following example of a set in F#:
set ["Bob"; "Bill"; "Ben"];;
val it : Set<string> = set ["Ben"; "Bill"; "Bob"]

Hence, a set can be given in a manner similar to a list using the “set-builder” function set.
The resulting value is of type Set<string>, that is, a set of strings, and we can see from the
F# answer that the elements occur according to a lexicographical ordering. A standard
number ordering is used for sets of integers, for example:
set [3; 1; 9; 5; 7; 9; 1];;
val it : Set<int> = set [1; 3; 5; 7; 9]

Equality of two sets is tested in the usual manner:

set ["Bob";"Bill";"Ben"] = set ["Bill";"Ben";"Bill";"Bob"];;
val it : bool = true

and sets are ordered on the basis of a similar kind of lexicographical ordering as used for
lists. (See Section 4.1.) For example, {Ann, Jane } is smaller than { Bob, Bill, Ben } (in
the F# representation) since Ann is smaller than every element in { Bob, Bill, Ben } using
the string representation:
compare (set ["Ann";"Jane"]) (set ["Bill";"Ben"; "Bob"]);;
val it : int = -1

Table 5.2 A selection of basic operations from the Set library

 Basic properties and operations on sets
We shall now describe the basic properties of sets and the operations on sets in F# as shown in
Table 5.2. The functions Set.ofList and Set.toList are conversion functions between lists
and sets:
let males = Set.ofList ["Bob"; "Bill"; "Ben"; "Bill"];;
val males : Set<string> = set ["Ben"; "Bill"; "Bob"]

Set.toList males;;
val it : string list = ["Ben"; "Bill"; "Bob"]

Note that the resulting list is ordered and contains no repeated elements.
An element can be inserted in a set with the function Set.add:
Set.add "Barry" males;;
val it : Set<string> = set ["Barry"; "Ben"; "Bill"; "Bob"]

and removed from a set with the function Set.remove:

Set.remove "Bill" males;;
val it : Set<string> = set ["Ben"; "Bob"]

The add and remove operations do not change the original set, that is, they have no side effect.
The same observation applies for all other operations in the Set library. For example, the add
and remove operations above did not change the value of males:
males;;
val males : Set<string> = set ["Ben"; "Bill"; "Bob"]

Containment in a set is tested using Set.contains and a subset relationship is tested using
Set.isSubset:

Set.contains "Barry" males;;

val it : bool = false

Set.isSubset males (set ["Bob"; "Bill"; "Ann"]);;

val it : bool = false

Set.isSubset males (Set.add "Ben" (set ["Bob";"Bill";"Ann"]));;

val it : bool = true

Due to the ordering required for set elements, every non-empty set has a minimal and a
maximal element:
Set.minElement (set ["Bob"; "Bill"; "Ben"]);;
val it : string = "Ben"
Set.maxElement (set ["Bob"; "Bill"; "Ben"]);;
val it : string = "Bob"

Furthermore, the cardinality of a finite set is in F# given by the function Set.count:

Set.count (set ["Bob"; "Bill"; "Ben"]);;
val it : int = 3
Set.count (Set.empty);;
val it : int = 0

which also shows that the cardinality of the empty set (denoted by Set.empty) is 0.

Fundamental operations on sets

We shall now consider the selection of fundamental operations from the Set library in F#
shown in Table 5.3.
We illustrate set operations for union, intersection and difference using an example where
males are supposed to be all the males at a golf club, and boardMembers are the members of
the board for that club:
let boardMembers = Set.ofList [ "Alice"; "Bill"; "Ann"];;
val boardMembers : Set<string> = set ["Alice"; "Ann"; "Bill"]

Set.union males boardMembers;;

val it : Set<string> = set ["Alice";"Ann";"Ben";"Bill";"Bob"]

Set.intersect males boardMembers;;

val it : Set<string> = set ["Bill"]

Set.difference males boardMembers;;

val it : Set<string> = set ["Ben"; "Bob"]

Table 5.3 A selection of operations from the Set library

where, for example, the set of males being board members is obtained using intersections and
the set of those who are not board members is obtained using difference.
A function can be applied to every member of a set using Set.map in the same manner it can
be applied to every element of a list using List.map. The following function, that transforms
a set of sets S = {s0, …, sn−1} to the set {|s0|, …, |sn−1|} containing the cardinalities of the
elements of S, is, for example, defined using Set.map in a natural manner:
let setOfCounts s = Set.map Set.count s;;
val setOfCounts: Set<Set<′a>> -> Set<int> when ′a: comparison
Consider the F# value for the set of sets { { 1,3,5}, {2,4}, {7, 8, 9} }:
let ss = set [set [1;3;5]; set [2;4]; set [7;8;9] ];;
val it : Set<Set<int>>
= set [set [1; 3; 5]; set [2; 4]; set [7; 8; 9]]

setOfCounts ss;;
val it : Set<int> = set [2; 3]

The functions: Set.exists, Set.forall and Set.filter, work in a similar manner to their
List siblings:

Set.exists (fun x -> x>=2) (set [1;3;1;4]);;

val it : bool = true

Set.forall (fun x -> x>=2) (set [1;3;1;4]);;

val it : bool = false

Set.filter (fun x -> x>3) (set [1;5;-2;8]);;

val it : Set<int> = set [5; 8]

The functions Set.fold and Set.foldBack also correspond to their list siblings. This is
illustrated in the following evaluations:
Set.fold (-) 0 (set [1;2;3]) = ((0 - 1) - 2) -3 = -6
Set.foldBack (-) (set [1;2;3]) 0 = 1- (2 - (3 - 0)) = 2

where the ordering on the set elements is exploited.

The functions sumSet and setOfCounts can be succinctly declared using foldBack:

let sumSet s = Set.foldBack (+) s 0;;

val sumSet : Set<int> -> int

let setOfCounts s = Set.foldBack

(fun se sn -> Set.add (Set.count se) sn)
s
Set.empty;;
setOfCounts : Set<Set<′a>> -> Set<int> when ′a : comparison

sumSet (set [1 .. 5]);;

val it : int = 15

setOfCounts (set [set [1;3;5]; set [2;4]; set [7;8;9] ]);;

val it : Set<int> = set [2; 3]

Declarations of these functions could also be based on Set.fold:

let sumSet s = Set.fold (+) 0 s;;

let setOfCounts s = Set.fold

(fun sn se -> Set.add (Set.count se) sn)
Set.empty
s;;

Notice that is is more natural to base a declaration of setOfCounts on Set.map as done

above, rather than basing it on one of the fold functions.

Recursive functions on sets

The functions Set.map, Set.filter, Set.fold and Set.foldBack will traverse the complete
set before they terminate, unless the evaluation is aborted by raising an exception, and this
may be undesirable in some situations. Consider, for example, the function that finds the least
element in a set satisfying a given predicate:
tryFind: (′a -> bool) -> Set<′a> -> ′a option
when ′a : comparison

This function can be declared by repeated extraction of the minimal element from a set until
an element satisfying the predicate is found:
let rec tryFind p s =
if Set.isEmpty s then None
else let minE = Set.minElement s
if p minE then Some minE
else tryFind p (Set.remove minE s);;

For example, the least three-element set from a set of sets is extracted as follows:
let ss = set [set [1;3;5]; set [2;4]; set [7;8;9] ];;

tryFind (fun s -> Set.count s = 3) ss;;

val it : Set<int> option = Some (set [1; 3; 5])

A declaration of this function that is based on Set.fold will always traverse the entire set
leading to a linear best-case running time, while the function declared above will terminate as
soon as an element satisfying the predicate is found, and the best-case execution time is
dominated by the time required for finding the minimal element in a set, and that execution
time is logarithmic in the size of the set when it is represented by a balanced binary tree. Note
however, that the worst-case execution time of traversing a set S using Set.fold or
Set.foldBack is O(|S|), that is linear in the size |S| of the set, while it is O(|S| · log (|S|)) for
a function based on a recursion schema like that for tryFind, due to the logarithmic
operations for finding and removing the minimal element of a set.
A more efficient implementation of the function tryFind using an enumerator is given on
Page 191, and the efficiency of different methods for traversal of collections is analyzed in
Exercise 9.14. Enumerators for collections (to be introduced in Section 8.12) provide a far
more efficient method than the above used recursion schema for tryFind.
 Example: Map colouring
The solution of the map-colouring problem from Section 4.6 shall now be improved using
sets. The basic algorithmic idea for the solution below using sets is basically the same as that
for using lists. But the model using sets is a more natural one. Furthermore, we shall take
advantage of the higher-order library functions.
A map is mathematically modelled as a binary relation of countries, that is, as a set of
country pairs. Furthermore, since the order in which countries occur in a colour is not
relevant and since repetition among the countries in a colour is of no concern, the natural
model of a colour is a country set. A similar observation applies to a colouring:
type Country = string;;
type Map = Set<Country*Country>;;
type Colour = Set<Country>;;
type Colouring = Set<Colour>;;

Two countries c1, c2 are neighbors in a map m, if either (c1,c2) m or (c2,c1) m. In F# this
is expressed as follows:
let areNb c1 c2 m =
Set.contains (c1,c2) m || Set.contains (c2,c1) m;;

A colour col can be extended by a country c for a given map m, if for every country c′ in
col, we have that c and c′ are not neighbours in m. This can be directly expressed using
Set.forall:

let canBeExtBy m col c =

Set.forall (fun c′ -> not (areNb c′ c m)) col;;

The function
extColouring: Map -> Colouring -> Country -> Colouring

is declared as a recursive function over the colouring:

let rec extColouring m cols c =
if Set.isEmpty cols
then Set.singleton (Set.singleton c)
else let col = Set.minElement cols
let cols′ = Set.remove col cols
if canBeExtBy m col c
then Set.add (Set.add c col) cols′
else Set.add col (extColouring m cols′ c);;

This recursive declaration is preferred to using a declaration based on either Set.fold or

Set.foldBack, since the recursive version terminates as soon as a colour that can be extended
by the country is found, whereas a declaration based on one of the fold functions always will
iterate through the entire colouring.
A set of countries is obtained from a map by the function:
countries: Map -> Set<Country>

The declaration of this function is based on repeated insertion (using Set.fold) of the
countries in the map into a set:
let countries m =
Set.fold
(fun set (c1,c2) -> Set.add c1 (Set.add c2 set))
Set.empty
m;;

The function
colCntrs: Map -> Set<Country> -> Colouring

that creates a colouring for a set of countries in a given map, can be declared by repeated
insertion of countries in colourings using the extColouring function:
let colCntrs m cs = Set.fold (extColouring m) Set.empty cs;;

The function that creates a colouring from a map is declared using function composition
and used as follows:
let colMap m = colCntrs m (countries m);;

let exMap = Set.ofList [("a","b"); ("c","d"); ("d","a")];;

colMap exMap;;
val it: Set<Set<string>> = set [set ["a";"c"]; set ["b";"d"]]

Comparing this set-based solution with the list-based one in Section 4.6 we can first
observe that the set-based model is more natural, due to the facts that a map is a binary
relation of countries and a colouring is a partitioning of the set of countries in a map. For
most of the functions there is even an efficiency advantage with the set-based functions. This
advantage is due to the following

the worst-case execution time for testing for membership of a set (represented by a
balanced binary tree) is logarithmic in the size of the set, while this operation is linear
when the set is represented by a list, and
the worst-case execution time for inserting an element into a set (represented by a
balanced binary tree) is logarithmic in the size of the set, while this operation is linear
when the set is represented by a list without duplicated elements.

The use of lists has an advantage in the case of the recursive function extColouring since
the pattern matching for lists yields a more readable declaration and since the worst-case
execution time of this list-based version is linear in the size |S| of the colouring S, while it is
O(|S| · log(|S|)) for the set-based one. (See remark on Page 110.)
An improved version is therefore based on the following type declaration:
type Country = string;;
type Map = Set<Country*Country>;;
type Colour = Set<Country>;;
type Colouring = Colour list;;

Just two functions extColouring and colCntrs are affected by this change of the type for
colouring while the remaining functions are as above. The new declarations are:

let rec extColouring m cols c =

match cols with
| [] -> [Set.singleton c]
| col::cols′ -> if canBeExtBy m col c
then (Set.add c col)::cols′
else col::(extColouring m cols′ c);;

let colCntrs m cs = Set.fold (extColouring m) [] cs;;

colMap exMap;;
val it : Set<string> list = [set ["a"; "c"]; set ["b"; "d"]]
 5.3 Maps
In the modelling and solution for many problems it is often convenient to use finite functions
to uniquely associate values with keys. Such finite functions from keys to values are called
maps. This section introduces the map concept and some of the main operations on maps in
the F# Map library. Please consult the on-line documentation in 9 for an overview of the
complete Map library.
 The mathematical concept of a map
A map from a set A to a set B is a finite subset A′ of A together with a function m defined on A′:

The set A′ is called the domain of m and we write dom m = A′.

A map m can be described in a tabular form as shown below. The left column contains the
elements a0, a1, … , an−1 of the set A′, while the right column contains the corresponding
values m(a0) = b0, m(a1) = b1, …, m(an−1) = bn−1:

An element ai in the set A′ is called a key for the map m. A pair (ai,bi) is called an entry, and bi
is called the value for the key ai. Note that the order of the entries is of no significance, as the
map only expresses an association of values to keys. Note also that any two keys ai and aj in
different entries are different, as there is only one value for each key. Thus, a map may be
represented as a finite set of its entries. We use

to denote the sets of entries of a map.

The cash register example in Chapter 4.6 comprises an article register associating name
and price to article codes, and this register can be viewed as a map. A key in the map is an
article code and the corresponding value is the pair with the name and price of the article.
A particular article register is given by the following map:

It associates the value (cheese, 25) with the key a1, the value (herring, 4) with the key a2, and
the value (soft drink, 5) with the key a3. Hence, it has the domain { a1, a2, a3 }.
Table 5.4 A selection of operations from the Map library
 Maps in F#
The Map library of F# supports maps of polymorphic types Map<′a,′b>, where ′a and ′b
are the types of the keys and values, respectively, of the map. The Map is implemented using
balanced binary trees, and requires therefore that an ordering is defined on the type ′a of
keys. Some of the functions of the Map library are specified in Table 5.4.
A map in F# can be generated from a list of its entries. For example:
let reg1 = Map.ofList [("a1",("cheese",25));
("a2",("herring",4));
("a3",("soft drink",5))];;
val reg1 : Map<string,(string * int)> =
map [("a1", ("cheese", 25)); ("a2", ("herring", 4));
("a3", ("soft drink", 5))]

is an F# map for the register reg1, where keys are strings and values are pairs of the type
string*int. If the list contains multiple entries for the same key, then the last occurring entry
is the significant one:
Map.ofList [(1,"a"); (2,"b"); (2,"c"); (1,"d")];;
val it : Map<int,string> = map [(1, "d"); (2, "c")]

The list of entries of a map is achieved using the Map.toList function:

Map.toList reg1;;
val it : (string * (string * int)) list =
[("a1", ("cheese", 25)); ("a2", ("herring", 4));
("a3", ("soft drink", 5))]

An entry can be added to a map using add while the value for a key in a map is retrieved
using either find or tryFind:
let reg2 = Map.add "a4" ("bread", 6) reg1;;
val reg2 : Map<string,(string * int)> =
map [("a1", ("cheese", 25)); ("a2", ("herring", 4));
("a3", ("soft drink", 5)); ("a4", ("bread", 6))]

Map.find "a2" reg1;;

val it : string * int = ("herring", 4)

Map.tryFind "a2" reg1;;

val it : (string * int) option = Some ("herring", 4)

Map.containsKey "a4" reg1;;

val it : bool = false

Map.find "a4" reg1;;

System.Collections.Generic.KeyNotFoundException: The given key
was not present in the dictionary.
...
Stopped due to error
Map.tryFind "a4" reg1;;
val it : (string * int) option = None

where find raises an exception if the key is not in the domain of the map and tryFind returns
None in that case.
The old entry is overridden if you add an entry for an already existing key. The entry for a
given key can be deleted using the remove function:
let reg3 = Map.add "a4" ("bread", 8) reg1;;
val reg3 : Map<string,(string * int)> =
map [("a1", ("cheese", 25)); ("a2", ("herring", 4));
("a3", ("soft drink", 5)); ("a4", ("bread", 8))]

let reg4 = Map.remove "a2" reg3;;

val reg4 : Map<string,(string * int)> =
map [("a1", ("cheese", 25)); ("a3", ("soft drink", 5));
("a4", ("bread", 8))]

The Map functions exists, forall, map, fold and foldBack are similar to their List and
Set siblings. These functions are specified with type and meaning in Table 5.4, so we just give
some illustrative examples below.
The following expression tests whether there are expensive articles, for which the price
exceeds 100, in a register:
Map.exists (fun _ (_,p) -> p > 100) reg1;;
val it : bool = false

The natural requirement that every price occurring in a register must be positive is expressed
by:
Map.forall (fun _ (_,p) -> p > 0) reg1;;
val it : bool = true

The part of a register with articles having a price smaller than 7 is extracted as follows:
Map.filter (fun _ (_,p) -> p < 7) reg3;;
val it : Map<string,(string * int)> =
map [("a2", ("herring", 4)); ("a3", ("soft drink", 5))]

A new register, where a 15% discount is given on all articles, can be computed as follows:
Map.map
(fun ac (an,p) -> (an,int(round(0.85*(float p)))))
reg3;;
val it : Map<string,(string * int)> =
map [("a1", ("cheese", 21)); ("a2", ("herring", 3));
("a3", ("soft drink", 4)); ("a4", ("bread", 7))]

We can extract the list of article codes and prices for a given register using the fold functions
for maps:
Map.foldBack (fun ac (_,p) cps -> (ac,p)::cps) reg1 [];;
val it: (string*int) list = [("a1",25); ("a2",4); ("a3",5)]
Map.fold (fun cps ac (_,p) -> (ac,p)::cps) [] reg1;;
val it: (string*int) list = [("a3",5); ("a2",4); ("a1",25)]

where these two examples show that the entries of a map are ordered according to the keys.
 Example: Cash register
We give a solution to the cash register example discussed in Section 4.6. Article codes and
names, number of pieces and prices are modelled just like in Section 4.6:
type ArticleCode = string;;
type ArticleName = string;;
type NoPieces = int;;
type Price = int;;

The natural model of a register, associating article name and price with each article code, is
using a map:
type Register = Map<ArticleCode, ArticleName*Price>;;

since an article code is a unique identification of an article.

The information concerning a bill is also modelled as in Section 4.6:
type Info = NoPieces * ArticleName * Price;;
type Infoseq = Info list;;
type Bill = Infoseq * Price;;

For the remaining parts we give three versions.

Version 1
In the first version we model a purchase just as in Section 4.6:
type Item = NoPieces * ArticleCode;;
type Purchase = Item list;;

The function makebill1: Register -> Purchase -> Bill makes the bill for a given
register and purchase and it can be defined by a recursion following the structure of a
purchase:
let rec makeBill1 reg = function
| [] -> ([],0)
| (np,ac)::pur ->
match Map.tryFind ac reg with
| Some(aname,aprice) ->
let tprice = np*aprice
let (infos,sumbill) = makeBill1 reg pur
((np,aname,tprice)::infos, tprice+sumbill)
| None ->
failwith(ac + " is an unknown article code");;

where an exception signals an undefined article code in a register. We use the function
Map.tryFind in order to detect when this exception should be raised. A simple application of
the program is:
let pur = [(3,"a2"); (1,"a1")];;
makeBill1 reg1 pur;;
val it : (int * string * int) list * int =
([(3, "herring", 12); (1, "cheese", 25)], 37)

where reg1 is declared on Page 114.

Version 2
The recursion pattern of makeBill1 is the same as that of List.foldBack, and the explicit
recursion can be replaced by application of that function. Furthermore, it may be acceptable to
use the exception from the Map library instead of using failwith. This leads to the following
declaration:
let makeBill2 reg pur =
let f (np,ac) (infos,billprice) =
let (aname, aprice) = Map.find ac reg
let tprice = np*aprice
((np,aname,tprice)::infos, tprice+billprice)
List.foldBack f pur ([],0);;

makeBill2 reg1 pur;;

val it : (int * string * int) list * int =
([(3, "herring", 12); (1, "cheese", 25)], 37)

where Map.find will raise an exception if an article code is not found in a register.

Version 3
A purchase is so far just modelled as a list of items, each item consisting of a count and an
article code. The order of appearance in the list may represent the sequence in which items are
placed on the counter in the shop. One may, however, argue that a purchase of the following
three items: three herrings, one piece of cheese, and two herrings, is the same as a purchase of
one piece of cheese and five herrings. Furthermore, the latter form is more convenient if we
have to model a discount on five herrings, as the discount applies independently of the order
in which the items are placed on the counter. Thus one could model a purchase as a map,
where article codes are keys and number of pieces are values of a map.
type Purchase = Map<ArticleCode,NoPieces>;;

With this model, the makeBill3: Register -> Purchase -> Bill function is declared
and used as follows:
let makeBill3 reg pur =
let f ac np (infos,billprice) =
let (aname, aprice) = Map.find ac reg
let tprice = np*aprice
((np,aname,tprice)::infos, tprice+billprice)
Map.foldBack f pur ([],0);;

where we use Map.foldBack to fold the function f over a purchase.

 An example showing the use of this function is:
let purMap = Map.ofList [("a2",3); ("a1",1)];;
val purMap : Map<string,int> = map [("a1", 1); ("a2", 3)]

makeBill3 reg1 purMap;;

val it : (int * string * int) list * int =
([(1, "cheese", 25); (3, "herring", 12)], 37)

We leave the generation of a map for a purchase on the basis of a list of items for Exercise
5.9. Furthermore, it is left for Exercise 5.10 to take discounts for certain articles into account.
 Summary
In this chapter we have introduced the list, set and map parts from the collection library of
F#. These three libraries are efficient implementations of such finite, immutable collections.
Notice that this chapter just covers a small part of the libraries. Furthermore, in many
applications these collections provide a natural data model and we strongly encourage to use
these libraries whenever it is appropriate.
In Chapter 11 we introduce sequences, which is another part of the collection library.
Sequences are (possibly infinite) list-like structures, where just a finite part of the sequence is
computed at any stage of a computation.
 Exercise
5.1 Give a declaration for List.filter using List.foldBack.
5.2 Solve Exercise 4.15 using List.fold or List.foldBack.
5.3 Solve Exercise 4.12 using List.fold or List.foldBack.
5.4 Declare a function downto1 such that:

Declare the factorial function by use of downto1.

Use downto1 to declare a function that builds the list [g(1), g(2), … , g(n)] for a
function g and an integer n.
5.5 Consider the map colouring example in Section 4.6. Give declarations for the
functions areNb canBeExtBy, extColouring, countries and colCntrs using higher-
order list functions. Are there cases where the old declaration from Section 4.6 is
preferable?
5.6 We define a relation from a set A to a set B as a subset of A × B. A relation r′ is said to
be smaller than r, if r′ is a subset of r, that is, if r′ ⊆ r. A relation r is called finite if it is a
finite subset of A × B. Assuming that the sets A and B are represented by F# types ′a and
′b allowing comparison we can represent a finite relation r by a value of type set<′a *
′b>.

1. The domain dom r of a relation r is the set of elements a in A where there exists an
element b in B such that (a, b) r. Write an F# declaration expressing the domain
function.
2. The range rng r of a relation r is the set of elements b in B where there exists an
element a in A such that (a, b) r. Write an F# declaration expressing the range
function.
3. If r is a finite relation from A to B and a is an element of A, then the application of r
to a, apply r a, is the set of elements b in B such that (a, b) r. Write an F# declaration
expressing the apply function.
4. A relation r from a set A to the same set is said to be symmetric if (a1,a2) r implies
(a2,a1) r for any elements a1 and a2 in A. The symmetric closure of a relation r is the
smallest symmetric relation containing r. Declare an F# function to compute the
symmetric closure.
5. The relation composition r s of a relation r from a set A to a set B and a relation s
from B to a set C is a relation from A to C. It is defined as the set of pairs (a, c) where
there exist an element b in B such that (a, b) r and (b, c) s. Declare an F# function to
compute the relational composition.
 6. A relation r from a set A to the same set A is said to be transitive if (a1,a2) r and
(a2,a3) r implies (a1,a3) r for any elements a1, a2 and a3 in A. The transitive
closure of a relation r is the smallest transitive relation containing r. If r contains n
elements, then the transitive closure can be computed as the union of the following n
relations:

Declare an F# function to compute the transitive closure.

5.7 Declare a function allSubsets such that allSubsets n k is the set of all subsets of
{1, 2, …, n} containing exactly k elements. Hint: use ideas from Exercise 2.8. For
example, (nk) is the number of subsets of {1, 2, …, n} containing exactly k elements.
5.8 Give declarations for makeBill3 using map.fold rather than map.foldBack.
5.9 Declare a function to give a purchase map (see Version 3 on Page 118) on the basis of
a list of items (from the Versions 1 and 2).
5.10 Extend the cash register example to take discounts for certain articles into account.
For example, find a suitable representation of discounts and revise the function to make a
bill accordingly.
5.11 Give a solution for Exercise 4.23 using the Set and Map libraries.
 6
Finite trees

This chapter is about trees, which are structures that may contain subcomponents of the
same type. A list is an example of a tree. The list 1::[2;3;4], for example, contains a
subcomponent [2;3;4] that is also a list. In this chapter we will introduce the concept of a tree
through a variety of examples.
In F# we use a recursive type declaration to represent a set of values which are trees. The
constructors of the type correspond to the rules for building trees, and patterns containing
constructors are used when declaring functions on trees.
We motivate the use of finite trees and recursive types by a number of examples: Chinese
boxes, symbolic differentiation, expression trees, search trees, file systems, trees with
different kinds of nodes and electrical circuits.
 6.1 Chinese boxes
A Chinese box is a coloured cube that contains a coloured cube that ... that contains a coloured
cube that contains nothing. More precisely, a Chinese box is either Nothing or a Cube
characterised by its side length, colour and the contained Chinese box. This characterization
can be considered as stating rules for generating Chinese boxes, and it is used in the
following definition of Chinese boxes as trees:
The set Cbox of Chinese boxes can be represented as the set of trees generated by the rules:
Rule 1: The tree Nothing is in Cbox.
Rule 2: If r is a float number, if c is a colour, and if cb is in Cbox, then the tree:

is also in Cbox.
Rule 3: The set Cbox contains no other values than the trees generated by repeated use of
Rule 1 and Rule 2.
The following example shows how this definition can be used to generate elements of
Cbox.
Step a: The void tree Nothing is a member of Cbox by Rule 1.
Step b: The following tree is a member of Cbox by Step a and Rule 2:

Step c: The following tree is a member of Cbox by Step b and Rule 2:

Step d: The following tree is a member of Cbox by Step c and Rule 2:

 Type declaration
Using the following type Colour:
type Colour = Red | Blue | Green | Yellow | Purple;;

we declare a type Cbox representing the set of Chinese boxes as follows:

type Cbox = | Nothing // 1.
| Cube of float * Colour * Cbox;; // 2.

The declaration is recursive, because the declared type Cbox occurs in the argument type of
the constructor Cube. The constructors Nothing and Cube correspond to the above rules 1 and
2 for generating trees, so we can redo the above steps a through d with values of type Cbox:
Step a′: The constructor Nothing is a value of type Cbox.
Step b′: The value Cube(0.5,Red,Nothing) of type Cbox represents the tree generated in
Step b.
Step c′: The value Cube(1.0,Green,Cube(0.5,Red,Nothing)) of type Cbox represents
the tree generated in Step c.
Step d′: The value:
Cube(2.0,Yellow,Cube(1.0,Green,Cube(0.5,Red,Nothing)))

of type Cbox represents the tree generated in Step d.

These examples show the relationship between trees and values of type Cbox, and we note the
following statements where the last one follows from Rule 3 for generating trees:

Different values of type Cbox represent different trees.

Any tree is represented by a value of type Cbox.

Hence, a value of type Cbox is just a way of writing a tree instead of drawing it. F# does not
draw trees when printing values of type Cbox – the interactive F# system prints the textual
form of the value:
let cb1 = Cube(0.5, Red, Nothing);;
val cb1 : Cbox = Cube (0.5,Red,Nothing)

let cb2 = Cube(1.0, Green, cb1);;

val cb2 : Cbox = Cube (1.0,Green,Cube (0.5,Red,Nothing))

let cb3 = Cube(2.0, Yellow, cb2);;

val cb3 : Cbox = Cube (2.0,Yellow,Cube (1.0,Green,
Cube (0.5,Red,Nothing)))
 Patterns
In Section 3.8 we have seen declarations containing patterns for tagged values. Constructors
for trees can occur in patterns just like constructors for tagged values. An example of a tree
pattern is Cube(r,c,cb), containing identifiers r, c and cb for the components. This pattern
denotes the tree in Figure 6.1.
This pattern will, for example, match the tree shown in Figure 6.2 corresponding to the
value Cube(1.0,Green,Cube(0.5,Red,Nothing)) with bindings

Fig ure 6.1 Tree for pattern Cube(r,c,cb)

Fig ure 6.2 Tree for value cb2

where cb is bound to a value of type cbox corresponding to the tree shown in Step b on Page
122.
The inductive definition of the trees implies that any tree will either match the empty tree
corresponding to the pattern:
Nothing

according to Rule 1 in the definition of trees, or the tree pattern for a cube in Figure 6.1
corresponding to:
Cube(r,c,cb)

according to Rule 2 in the definition of trees.

 Function declarations
We give a declaration of the function:
count: Cbox -> int

such that the value of the expression: count( cb) is the number of cubes of the Chinese box
cb:
let rec count = function
| Nothing -> 0
| Cube(r,c,cb) -> 1 + count cb;;
val count : Cbox -> int

The declaration divides into two cases, one with pattern Nothing and the other with pattern
Cube(r,c,cb). Thus, the declaration follows the inductive definition of Chinese boxes.
This function can be applied to the above values cb2 and cb3:
count cb2 + count cb3;;
val it : int = 5
 Invariant for Chinese boxes
A Chinese box must satisfy the invariant that the length of its sides is a positive floating-point
number, which is larger than the side length of any cube it contains. The above four Chinese
boxes in steps a to d satisfy this invariant, but using the generation process for trees one can
construct the tree in Figure 6.3 that violates the invariant (i.e., it does not correspond to any
Chinese box).

Fig ure 6.3 A tree violating the invariant

When declaring a function on Chinese boxes by the use of the type Cbox we must ensure
that the function respects the invariant, that is, the function will only compute values of type
Cbox satisfying the invariant when applied to values satisfying the invariant.
 Insertion function
We can declare an insertion function on Chinese boxes:
insert: float * Colour * Cbox -> Cbox

The value of the expression insert( r, c, cb) is the Chinese box obtained from cb by
inserting an extra cube with side length r and colour c at the proper place among the cubes in
the box. The function insert is a partial function, that raises an exception in case the insertion
would violate the invariant for Chinese boxes:
let rec insert(r,c,cb) =
if r <= 0.0 then failwith "ChineseBox"
else match cb with
| Nothing -> Cube(r,c,Nothing)
| Cube(r1,c1,cb1) ->
match compare r r1 with
| t when t > 0 -> Cube(r,c,cb)
| 0 -> failwith "ChineseBox"
| _ -> Cube(r1,c1,insert(r,c,cb1));;
insert(2.0,Yellow,insert(1.0,Green,Nothing));;
val it : Cbox = Cube (2.0,Yellow,Cube (1.0,Green,Nothing))

insert(1.0,Green,insert(2.0,Yellow,Nothing));;
val it : Cbox = Cube (2.0,Yellow,Cube (1.0,Green,Nothing))

insert(1.0,Green,Cube(2.0,Yellow,Cube(1.0,Green,Nothing)));;
System.Exception: ChineseBox
Stopped due to error

Note, that any legal Chinese box can be generated from the box Nothing by repeated use of
insert.
 Other F# representations of Chinese Boxes
One may argue that the type cbox is unnecessarily complicated as Chinese boxes may simply
be modelled using lists:
type Cbox = (float * Colour) list

This is, however, essentially the same as the above Cbox type of trees, as the list type is a
special case of the general concept of recursive types (cf. Section 6.3).
One may also argue that it is strange to have a constructor Nothing denoting a non-existing
Chinese box, and one might rather discard the empty box and divide the Chinese boxes into
those consisting of a single cube and those consisting of multiple cubes, as expressed in the
following declaration:
type Cbox1 = | Single of float * Colour
| Multiple of float * Colour * Cbox1;;

Using this type, we get the following declarations of the functions count and insert:
let rec count1 = function
| Single _ -> 1
| Multiple(_,_,cb) -> 1 + count1 cb;;

let rec insert1 (r1,c1,cb2) =

if r1 <= 0.0 then failwith "insert1: Chinese box"
else match cb2 with
| Single (r2,c2) ->
match compare r1 r2 with
| t when t < 0 -> Multiple(r2,c2,Single(r1,c1))
| 0 -> failwith "ChineseBox"
| _ -> Multiple(r1,c1,cb2)
| Multiple (r2,c2,cb3) ->
match compare r1 r2 with
| t when t < 0 -> Multiple(r2,c2,insert1(r1,c1,cb3))
| 0 -> failwith "ChineseBox"
| _ -> Multiple(r1,c1,cb2);;

We have now suggested several representations for Chinese boxes. The preferable choice of
representation will in general depend on which functions we have to define. The clumsy
declaration of the insert1 function contains repeated sub-expressions and this indicates that
the first model for Chinese boxes with a Nothing value is to be preferred.
 6.2 Symbolic differentiation
We want to construct a program for computing the derivative of a real function of one
variable. The program should, for example, compute the derivative f′(x) = 2x · cos(x2) of the
function f(x) = sin (x2). The concept of function in F# cannot be used for this purpose, as such
a function declaration just gives the means of computing values of the function. For example:
let f x = sin(x * x);;
val f : float -> float

f 2.0;;
val it : float = -0.7568024953

The differentiation is a manipulation of the expression denoting a function, so we need a

representation of the structure of such expressions. This can be done using expression trees.
We restrict our attention to expressions constructed from real-valued constants and the
variable x, using the arithmetic functions: addition, subtraction, multiplication and division,
and the real functions sin, cos, log and exp. We use the symbols Add, Sub, Mul and Div to
represent the arithmetic operators, and the symbols Sin, Cos, Log and Exp to represent the
special functions. The expressions sin(x· x) and (sin x)· x will then be represented by the
expression trees shown in Figure 6.4.

Fig ure 6.4 Trees for sin(x · x) and (sin x) · x

The different order of the operators in these expressions is reflected in the trees: the tree
for sin(x · x) contains a sub-tree for the sub-expression x· x, which again contains two sub-
trees for the sub-expressions x and x, while the tree for (sin x)· x contains sub-trees for the
sub-expressions sin x and x.
The set of finite expression trees Fexpr is generated inductively by the following rules:
Rule 1: For every float number r, the tree for the constant r shown in Figure 6.5 is a
member of Fexpr.

Fig ure 6.5 Tree for the constant r and trees for dyadic operators

Rule 2: The tree X is in Fexpr.

Rule 3: If fe1 and fe2 are in Fexpr, then the trees for the dyadic operators for addition,
subtraction, multiplication and division shown in Figure 6.5 are members of Fexpr.
Rule 4: If fe is in Fexpr, then the trees for the special functions shown in Figure 6.6 are in
Fexpr.

Fig ure 6.6 Trees for special functions

Rule 5: The set Fexpr contains no other values than the trees generated by rules 1. to 4.
 Type declaration
Expression trees can be represented in F# by values of a recursively defined type. We get the
following declaration of the type Fexpr:
type Fexpr = | Const of float
| X
| Add of Fexpr * Fexpr
| Sub of Fexpr * Fexpr
| Mul of Fexpr * Fexpr
| Div of Fexpr * Fexpr
| Sin of Fexpr
| Cos of Fexpr
| Log of Fexpr
| Exp of Fexpr;;

For instance, the expression trees for sin(x · x) and (sin x)· x are represented by the values
Sin(Mul(X,X)) and Mul(Sin X,X) of type Fexpr.
 Patterns
The following patterns correspond to values of type Fexpr:
Const r X
Add(fe1,fe2) Sub(fe1,fe2) Mul(fe1,fe2) Div(fe1,fe2)
Sin fe Cos fe Log fe Exp fe

These patterns can be used in function declarations with a division into clauses according to
the structure of expression trees.
Table 6.1 Differentiation rules

g′(x) = 0 when g(x) = c, for c Constant

f′(x) = 1 when f(x) = x Identity
(f(x) + g(x))′ = f′(x) + g′(x) Addition
(f(x)-g(x))′ = f′(x) - g′(x) Subtraction
(f(x) · g(x))′ = f′(x) · g(x) + f(x) · g′(x) Multiplication
(f(x) / g(x))′ = (f′(x) · g(x) - f(x) · g′(x)) / (g(x))2 Division
f(g(x))′ = f′(g((x))) · g′(x) Composition
(sin x)′ = cos x and (cos x)′ = − sin x Trigonometry
(log x)′ = 1 / x Logarithmic
(exp x)′ = exp x Exponential
 Function declaration
We are now in a position to declare a function
D: Fexpr -> Fexpr

such that D(fe) is a representation of the derivative with respect to x of the function represented
by fe. The declaration for D has a clause for each constructor generating a value of type
Fexpr, and each clause is a direct translation of the corresponding mathematical
differentiation rule (see Tables 6.1 and 6.2):
Table 6.2 Differentiation function

The following examples illustrate the use of the function:

D(Sin(Mul(X, X)));;
val it : Fexpr =
Mul (Cos (Mul (X,X)),
Add (Mul (Const 1.0,X),Mul (X,Const 1.0)))

D(Mul(Const 3.0, Exp X));;

val it : Fexpr =
Add (Mul (Const 0.0,Exp X),
Mul (Const 3.0,Mul (Exp X,Const 1.0)))

Note, that these examples show results which can be reduced. For example, the above value
of D(Mul(Const 3.0, Exp X)) could be reduced to Mul(Const 3.0, Exp X) if a product
with a zero factor was reduced to zero, and if adding zero or multiplying by one was
absorbed. It is an interesting, non-trivial, task to declare a function that reduces expressions to
a particular, simple form.
 Conversion to textual representation
The following function: toString: Fexpr -> string, will produce a textual representation
of a function expression:
let rec toString = function
| Const x -> string x
| X -> "x"
| Add(fe1,fe2) -> "(" + (toString fe1) + ")"
+ " + " + "(" + (toString fe2) + ")"
| Sub(fe1,fe2) -> "(" + (toString fe1) + ")"
+ " - " + "(" + (toString fe2) + ")"
| Mul(fe1,fe2) -> "(" + (toString fe1) + ")"
+ " * " + "(" + (toString fe2) + ")"
| Div(fe1,fe2) -> "(" + (toString fe1) + ")"
+ " / " + "(" + (toString fe2) + ")"
| Sin fe -> "sin(" + (toString fe) + ")"
| Cos fe -> "cos(" + (toString fe) + ")"
| Log fe -> "log(" + (toString fe) + ")"
| Exp fe -> "exp(" + (toString fe) + ")";;
val toString : Fexpr -> string\

toString(Mul(Cos(Mul(X, X)),
Add(Mul(Const 1.0, X), Mul(X, Const 1.0))));;

val it : string =
"(cos((x) * (x))) * (((1) * (x)) + ((x) * (1)))"

toString(Add(Mul(X, Mul(X, X)) , Mul(X, X)));;

val it : string = "((x) * ((x) * (x))) + ((x) * (x))"

The function toString puts brackets around every operand of an operator and every
argument of a function. It is possible to declare a better toString function that avoids
unnecessary brackets. See Exercise 6.3.
 6.3 Binary trees. Parameterized types
The constructors in a type declaration may have polymorphic types containing type variables.
These type variables are parameters of the type, and written in angle brackets <…> just
following the type constructor in the declaration.
An example of a type declaration with parameters is the type BinTree<′a,′b> of binary
trees with leaves containing elements of type ′a and nodes containing elements of type ′b:
type BinTree<′a,′b> =
| Leaf of ′a
| Node of BinTree<′a,′b> * ′b * BinTree<′a,′b>;;

Fig ure 6.7 A tree t1 of type BinTree<int,string>

The tree t1 in Figure 6.7 corresponds to the value

let t1 = Node(Node(Leaf 1,"cd",Leaf 2),"ab",Leaf 3);;

of type BinTree<int,string>. The top node with element "ab" is called the root of the tree t1
(trees are drawn upside-down in computer science with the root at the top and the leaves at the
bottom). The trees t2 and t3 in Figure 6.8 corresponding to the values
let t2 = Node(Leaf 1,"cd",Leaf 2);;
let t3 = Leaf 3;;

are called the left sub-tree and the right sub-tree of t1.
The type BinTree is polymorphic and allows polymorphic values like:
Node(Node(Leaf [],[],Leaf []),[],Leaf []);;

of type BinTree<′a list,′b list>.

 Fig ure 6.8 Left and right sub-tree t2 and t3 of the tree t1

Fig ure 6.9 Trees for patterns Node(t1,x,t2) and Leaf x

The constructors Node and Leaf can be used in patterns like

Leaf x
Node(t1,x,t2)

corresponding to the pattern trees in Figure 6.9. Using this we may, for example, declare a
function depth computing the depth of a binary tree:
let rec depth = function
| Leaf _ -> 0
| Node(t1,_,t2) -> 1 + max (depth t1) (depth t2);;
val depth : BinTree<′a,′b> -> int

depth t1;;
val it : int = 2

Fig ure 6.10 Simplified drawing of the tree t1 in Figure 6.7

In the following we will often use simplified drawings of trees where the constructors have
been left out and replaced by the value attached to the node. Such a simplified drawing of the
tree t1 in Figure 6.7 is shown in Figure 6.10.
 6.4 Traversal of binary trees. Search trees
A traversal of a binary tree is a function visiting the nodes of the tree in a certain order. We
may hence assume that the leafs do not carry any information corresponding to a simplified
BinTree type:

type BinTree<′a> = | Leaf

| Node of BinTree<′a> * ′a * BinTree<′a>;;

A traversal of a tree of this type is then described by a function of type:

BinTree<′a> -> ′a list

We consider three kinds of traversals:

Pre-order traversal: First visit the root node, then traverse the left sub-tree in pre-order
and finally traverse the right sub-tree in pre-order.
In-order traversal: First traverse the left sub-tree in in-order, then visit the root node and
finally traverse the right sub-tree in in-order.
Post-order traversal: First traverse the left sub-tree in post-order, then traverse the right
sub-tree in post-order and finally visit the root node.
The declarations look as follows:
let rec preOrder = function
| Leaf -> []
| Node(tl,x,tr) -> x :: (preOrder tl) @ (preOrder tr);;
val preOrder : BinTree<′a> -> ′a list

let rec inOrder = function

| Leaf -> []
| Node(tl,x,tr) -> (inOrder tl) @ [x] @ (inOrder tr);;
val inOrder : BinTree<′a> -> ′a list

let rec postOrder = function

| Leaf -> []
| Node(tl,x,tr) -> (postOrder tl) @ (postOrder tr) @ [x];;
val postOrder : BinTree<′a> -> ′a list

We define values t3 and t4 of type BinTree<int> by:

let t3 = Node(Node(Leaf, -3, Leaf), 0, Node(Leaf, 2, Leaf));;
let t4 = Node(t3, 5, Node(Leaf, 7, Leaf));;

Fig ure 6.11 Trees corresponding to values t3 and t4

Simplified drawings of the corresponding trees without constructors are shown in Figure
6.11. We will use such simplified drawings in the following.
Applying the traversal functions to t4 we get:
preOrder t4;;
val it : int list = [5; 0; -3; 2; 7]

inOrder t4;;
val it : int list = [-3; 0; 2; 5; 7]

postOrder t4;;
val it : int list = [-3; 2; 0; 7; 5]

The reader should compare these lists with the figure and the oral descriptions of the traversal
functions.
Traversal of binary trees can more generally be described by fold and foldBack functions
defined such that the following holds:

preFold fet= List.fold f e (preOrder t)

preFoldBack fte= List.foldBack f (preOrder t) e

and similar for in-order and post-order traversals. These functions should be declared to
accumulate the values in the nodes while traversing the tree – without actually building the list.
We show one of the declarations:
let rec postFoldBack f t e =
match t with
| Leaf -> e
| Node(tl,x,tr) ->
let ex = f x e
let er = postFoldBack f tr ex
postFoldBack f tl er;;
val postFoldBack : (′a -> ′b -> ′b) -> BinTree<′a> -> ′b -> ′b

postFoldBack (fun x xs -> x::xs) t4 [];;

val it : int list = [-3; 2; 0; 7; 5]

The other declarations are left as exercises.

The type system of F# allows a great variety of tree types and there is no standard Tree
library with a standard tree type. Functions like inOrder or inFold are hence defined
individually according to need in each program using a tree type.
There are also imperative tree traversal functions where an imperative function is called
whenever an element in the tree is visited, see Section 8.9.
 Search trees
We restrict the type variable ′a in our BinTree type to types with an ordering:
type BinTree<′a when ′a : comparison> =
| Leaf
| Node of BinTree<′a> * ′a * BinTree<′a>;;

A value of type BinTree<′a> is then called a search tree if it satisfies the following condition:
Every node Node(tleft, a, tright) satisfies:
a′ < a for every value a′ occurring in tleft and
a″ > a for every value a” occurring in tright.
This condition is called the search tree invariant. The trees t3 and t4 defined above and
shown in Figure 6.11 satisfy this invariant and are hence search trees.
A search tree can be used to represent a finite set {a0, a1, …, an−1}. This representation is
particularly efficient when the tree is balanced (see discussion on Page 136).
A function add for adding a value to a search tree can be defined as follows:
let rec add x t =
match t with
| Leaf -> Node(Leaf,x,Leaf)
| Node(tl,a,tr) when x<a -> Node(add x tl,a,tr)
| Node(tl,a,tr) when x>a -> Node(tl,a,add x tr)
| _ -> t;;
val add: ′a -> BinTree<′a> -> BinTree<′a> when ′a: comparison

It builds a single-node tree when adding a value x to an empty tree. When adding to a non-
empty tree with root a the value is added to the left sub-tree if x < a and to the right sub-tree if
x > a. The tree is left unchanged if x = a because the value x is then already member of the
represented set.
Adding the value 4 to the search tree t4
let t5 = add 4 t4;;
val t5 : BinTree<int> =
Node
(Node(Node(Leaf,-3,Leaf),0,Node(Leaf,2,Node(Leaf,4,Leaf))),
5,Node(Leaf,7,Leaf))

gives the tree in Figure 6.12.

Fig ure 6.12 Search trees corresponding to the value t5

 It follows by an inductive argument that an in-order traversal of a search tree will visit the
elements in ascending order because the elements in the left sub-tree are smaller than the root
element while the elements in the right sub-tree are larger – and this applies inductively to any
sub-tree. We get for instance:
inOrder t5;;
val it : int list = [-3; 0; 2; 4; 5; 7]

An in-order traversal of a search tree will hence give a list where the elements in the nodes
occur in ascending order.
A function contains for testing set membership can be declared by:
let rec contains x = function
| Leaf -> false
| Node(tl,a,_) when x<a -> contains x tl
| Node(_,a,tr) when x>a -> contains x tr
| _ -> true;;
val contains : ′a -> BinTree<′a> -> bool when ′a : comparison

contains 4 t5;;
val it : bool = true

It uses the search tree property in only testing the left sub-tree if x < the root node value and
only the right sub-tree if x > the root node value. The number of comparisons made when
evaluating a function value: contains x t is hence less or equal to the depth of the tree t. It
follows that the tree t5 in Figure 6.12 is not an optimal representation of the set, because the
set can be represented by the tree of depth 2 in Figure 6.13. The tree t5 was created by the
above add function, and it would hence require a more sophisticated add function to get the
“balanced” tree in Figure 6.13 instead.

Fig ure 6.13 Search tree of depth 2 representing same set as t5

The number of nodes in a balanced tree with depth k is approximately 2k and the depth of a
balanced tree with n nodes is hence approximately log2 n. The Set and Map collections in the
F# library use balanced search trees to get efficient implementations. A function like
Set.contains will hence require circa log2 n comparisons when used on a set with n
elements. Searching a value (e.g., using List.exists) in a list of length n may require up to n
comparisons. That makes a big difference for large n (e.g., log2 n ≈ 20 when n = 1000000).
 6.5 Expression trees
Tree representation of expressions is a common technique in compiler technology. This
section gives a bit of the flavour of this technique. The subject is related to the function
expression trees presented in Section 6.2.
We consider integer expressions of the form:

integer constant
identifier
-expression
expression + expression
expression - expression
expression * expression
let identifier = expression in expression
( expression )
They are represented by expression trees of the following type:
type ExprTree = | Const of int
| Ident of string
| Minus of ExprTree
| Sum of ExprTree * ExprTree
| Diff of ExprTree * ExprTree
| Prod of ExprTree * ExprTree
| Let of string * ExprTree * ExprTree;;

such that, for example, the expression:

a * (-3 + (let x = 5 in x + a))

is represented by the value:

let et =
Prod(Ident "a",
Sum(Minus (Const 3),
Let("x", Const 5, Sum(Ident "x", Ident "a"))));;

An expression is evaluated in an environment containing bindings of identifiers to values. An

environment is represented by a value env of type map<string,int> containing entries with
identifier and corresponding value. A let tree

is evaluated as follows in an environment env:

 1. Evaluate t1 to value v1
2. Evaluate t2 in the environment env extended with the binding of str to v.

An evaluation function
eval: ExprTree -> map<string,int> -> int

can now be defined recursively by dividing into cases according to the structure of the tree:
let rec eval t env =
match t with
| Const n n ->
| Ident s ->
Map.find s env
| Minus t ->
- (eval t env)
| Sum(t1,t2) ->
eval t1 env + eval t2 env
| Diff(t1,t2) ->
eval t1 env - eval t2 env
| Prod(t1,t2) ->
eval t1 env * eval t2 env
| Let(s,t1,t2) ->
let v1 = eval t1 env
let env1 = Map.add s v1 env
eval t2 env1;;
val eval : ExprTree -> Map<string,int> -> int

We may, for example, evaluate the above representation et of an expression in the

environment env where the identifier "a" is bound to the value -7:
let env = Map.add "a" -7 Map.empty;;

eval et env;;
val it : int = 35
 6.6 Trees with a variable number of sub-trees. Mutual
recursion
Trees with a variable number of sub-trees are obtained by using a type where each node
contains a (possibly empty) list of sub-trees. An example is the type:
type ListTree<′a> = Node of ′a * (ListTree<′a> list);;

Values of ListTree type represent trees where:

Node( x,[]) represents a leaf tree containing the value x

represents a tree with value x in the root and with n
Node( x,[t0 ;…; tn−1 ])
sub-trees represented by the values t0, …, tn−1

Such a tree is shown in Figure 6.14. It is represented by the value t1 where

let t7 = Node(7,[]);; let t6 = Node(6,[]);;
let t5 = Node(5,[]);; let t3 = Node(3,[]);;
let t2 = Node(2,[t5]);; let t4 = Node(4,[t6; t7]);;
let t1 = Node(1,[t2; t3; t4]);;

Fig ure 6.14 Tree represented by the value t1

 Traversal of list trees
We consider two kinds of traversal of list-trees: depth-first and breadth-first traversal. These
traversals correspond to the following order of the elements of the tree in Figure 6.14:

Depth-first: 1, 2, 5, 3, 4, 6, 7
Breadth-first: 1, 2, 3, 4, 5, 6, 7
In both cases we define a function to generate the list of nodes as well as fold and foldBack
functions. The declarations involve functions on lists because the sub-trees of a node are
organized as a list.
In the depth-first order we first visit the root node and then the nodes in each element of the
list of immediate sub-trees: The function depthFirst generating a list of elements is declared
using the library function List.collect (cf. Table 5.1) to apply depthFirst to each sub-tree
in the list and afterwards collect the obtained lists into one list:
let rec depthFirst (Node(x,ts)) =
x :: (List.collect depthFirst ts);;
val depthFirst : ListTree<′a> -> ′a list

The function depthFirstFold f is declared using List.fold to apply the function to each
tree in the list of sub-trees and to transfer the accumulated value to the call on the next sub-
tree:
let rec depthFirstFold f e (Node(x,ts)) =
List.fold (depthFirstFold f) (f e x) ts;;
val depthFirstFold: (′a->′b->′a) -> ′a -> ListTree<′b> -> ′a

depthFirstFold (fun a x -> x::a) [] t1;;

val it : int list = [7; 6; 4; 3; 5; 2; 1]

The reader should appreciate this short and elegant combination of library functions.
The declaration of depthFirstFoldBack is left as an exercise to the reader (cf. Exercise
6.12).
In the breadth-first order we should first visit the root and then the roots of the immediate
sub-trees and so on. This view of the problem does, unfortunately, not lead to any useful
recursion because the remaining part becomes organized in an inconvenient list of lists of
sub-trees.
A nice recursive pattern is instead obtained by constantly keeping track of the list rest of
sub-trees where the nodes still remain to be visited. Using this idea on the tree in Figure 6.14
we get:

Visit rest
[t2; t3;
1
t4]
 [t3; t4;
2
t5]
3 [t4; t5]
[t5; t6;
4
t7]
… …
Each step in this scheme will do the following:

1. Remove the head element t of the list rest.

2. Get a new rest list by appending the list of immediate sub-trees of t.
3. Visit the root of t.

The traversal finishes when the list rest becomes empty.

This pattern is used in the following declarations where the argument of the auxiliary
function breadthFirstList is a list corresponding to the rest list of sub-trees:
let rec breadthFirstList = function
| [] -> []
| (Node(x,ts)) :: rest ->
x :: breadthFirstList(rest@ts);;
val breadthFirstList : ListTree<′a> list -> ′a list

let breadthFirst t = breadthFirstList [t];;

val breadthFirst : ListTree<′a> -> ′a list

breadthFirst t1;;
val it : int list = [1; 2; 3; 4; 5; 6; 7]

The declaration of breadthFirstFoldBack follows the same pattern:

let rec breadthFirstFoldBackList f ts e =
match ts with
| [] -> e
| (Node(x,ts))::rest ->
f x (breadthFirstFoldBackList f (rest@ts) e);;
val breadthFirstFoldBackList :
(′a -> ′b -> ′b) -> ListTree<′a> list -> ′b -> ′b

let breadthFirstFoldBack f t e =
breadthFirstFoldBackList f [t] e;;
val breadthFirstFoldBack :
(′a -> ′b -> ′b) -> ListTree<′a> -> ′b -> ′b

breadthFirstFoldBack (fun x a -> x::a) t1 [];;

val it : int list = [1; 2; 3; 4; 5; 6; 7]

The declaration of breadthFirstFold is left as an exercise to the reader (cf. Exercise 6.12).
There are also imperative versions of these tree traversals where an imperative function is
called whenever a node is visited, see Section 8.9 for depth-first traversal and Section 8.13 for
breadth-first traversal. The breadth-first traversal uses an imperative queue.
 Example of list trees: File system
A file system is a list of named files and named directories where each directory contains
another file system. Figure 6.15 shows a directory named d1 with its associated file system.
The directory d1 contains two files a1 and a4 and two directories d2 and d3. The directory d2
contains a file a2 and a directory d3, and so on. Note that the same name may occur in
different directories. This structure is an example of a tree with variable number of sub-trees.

Fig ure 6.15 Directory with file system

Discarding the contents of files we represent a file system and its contents by two
declarations:
type FileSys = Element list
and Element = | File of string
| Dir of string * FileSys;;

The first declaration refers to a type Element which is declared in the second declaration.
This “forward” reference to the type Element is allowed by the F# system because Element is
declared in the second declaration using the keyword and. These two declarations constitute
an example of mutually recursive type declarations, as the type Element occurs in the
declaration of FileSys and the type FileSys occurs in the declaration of Element.
The directory shown in Figure 6.15 is represented by the value:
let d1 =
Dir("d1",[File "a1";
Dir("d2", [File "a2"; Dir("d3", [File "a3"])]);
File "a4";
Dir("d3", [File "a5"])
]);;

The declarations below yield functions namesFileSys and namesElement extracting a list of
names of all files (including files in subdirectories) for file systems and elements,
respectively:
let rec namesFileSys = function
| [] -> []
| e::es -> (namesElement e) @ (namesFileSys es)
and namesElement = function
| File s -> [s]
| Dir(s,fs) -> s :: (namesFileSys fs);;
val namesFileSys : Element list -> string list
val namesElement : Element -> string list

The above function declarations are mutually recursive as the identifier namesElement
occurs in the declaration of namesFileSys while the identifier namesFileSys occurs in the
declaration of namesElement. Mutually recursive functions are declared using the keyword
and to combine the individual function declarations.
The names of file and directories in the directory d1 may now be extracted:
namesElement d1;;
val it : string list = ["d1"; "a1"; "d2"; "a2";
"d3"; "a3"; "a4"; "d3"; "a5"]
 6.7 Electrical circuits
We consider electrical circuits built from components by serial or parallel composition. We
represent a circuit by a value of the following type:
type Circuit<′a> = | Comp of ′a
| Ser of Circuit<′a> * Circuit<′a>
| Par of Circuit<′a> * Circuit<′a>;;

In Figure 6.16 we show a circuit containing three components with attached values 0.25, 1.0,
and 1.5 together with the tree representing the circuit:

Fig ure 6.16 Circuit and corresponding tree

In F# the value is written Ser(Par(Comp 0.25, Comp 1.0), Comp 1.5):

let cmp = Ser(Par(Comp 0.25, Comp 1.0), Comp 1.5);;

val cmp : Circuit<float>
= Ser(Par(Comp 0.25,Comp 1.0),Comp 1.5)

Using this representation of circuits we can define a function count for computing the
number of components in a circuit:
let rec count = function
| Comp _ -> 1
| Ser(c1,c2) -> count c1 + count c2
| Par(c1,c2) -> count c1 + count c2;;
val count : Circuit<′a> -> int

For example:
count cmp;;
val it : int = 3

We consider now circuits consisting of resistances where the attached values are the
resistances of the individual components. Suppose c1 and c2 are two circuits with resistances
r1 and r2, respectively. The resistance of a serial combination of c1 and c2 is r1 + r2, and the
resistance of a parallel combination of c1 and c2 is given by the formula:
Thus, a function resistance computing the resistance of a circuit can be declared by:
let rec resistance = function
| Comp r -> r
| Ser(c1,c2) -> resistance c1 + resistance c2
| Par(c1,c2) ->
1.0 / (1.0/resistance c1 + 1.0/resistance c2);;
val resistance : Circuit<float> -> float

For example:
resistance cmp;;
val it : float = 1.7
 Tree recursion
The functions count and resistance on circuits can be expressed using a generic higher-
order function circRec for traversing circuits. This function must be parameterized with
three functions c, s and p, where

c: ′a -> ′b The value for a single component.

s: ′b -> ′b -> ′b The combined value for two circuits connected in series.
p: ′b -> ′b -> ′b The combined value for two circuits connected in parallel.
Note that s and p have the type ′b -> ′b -> ′b because they operate on the values for two
circuits. Thus, a general higher-order recursion function for circuits will have the type:
(′a -> ′b) * (′b -> ′b -> ′b) * (′b -> ′b -> ′b)
-> Circuit<′a> -> ′b

and the function is declared by:

let rec circRec (c,s,p) = function
| Comp x -> c x
| Ser(c1,c2) ->
s (circRec (c,s,p) c1) (circRec (c,s,p) c2)
| Par(c1,c2) ->
p (circRec (c,s,p) c1) (circRec (c,s,p) c2);;

The function circRec can, for example, be used to compute the number of components in a
circuit by use of the following functions c, s, and p:

let count circ = circRec((fun _ -> 1), (+), (+)) circ : int;;
val count : Circuit<′a> -> int

The type int is required to resolve the overloaded plus operators.

count(Ser(Par(Comp 0.25, Comp 1.0), Comp 1.5));;
val it : int = 3

Suppose again that the value attached to every component in a circuit is the resistance of the
component. Then the function circRec can be used to compute the resistance of a circuit by
use of the following functions c, s, and p:

c is fun r -> r

The attached value is the resistance.

s is (+)

The resistance of a serial composition is the sum of the resistances.

 p is fun r1 r2 -> 1.0/(1.0/r1+1.0/r2)

The resistance of a parallel composition is computed by this formula.

Using these functions for c, s and p we get:

let resistance =
circRec(
(fun r -> r),
(+),
(fun r1 r2 -> 1.0/(1.0/r1 + 1.0/r2)));;
val resistance : (Circuit<float> -> float)

resistance(Ser(Par(Comp 0.25, Comp 1.0), Comp 1.5));;

val it : float = 1.7
 Summary
We have introduced the notion of finite trees and motivated this concept through a variety of
examples. In F# a recursive type declaration is used to represent a set of values which are
trees. The constructors of the type correspond to the rules for building trees, and patterns
containing constructors are used when declaring functions on trees. We have also introduced
the notions of parameterized types, and mutually recursive type and function declarations.
 Exercise
6.1 Declare a function red of type Fexpr -> Fexpr to reduce expressions generated from
the differentiation program in Section 6.2. For example, sub-expressions of form Const
1.0 * e can be reduced to e. (A solution is satisfactory if the expression becomes “nicer”.
It is difficult to design a reduce function so that all trivial sub-expressions are eliminated.)
6.2 Postfix form is a particular representation of arithmetic expressions where each
operator is preceded by its operand(s), for example:
(x + 7.0) has postfix form x 7.0 +
(x + 7.0)*(x - 5.0) has postfix form x 7.0 + x 5.0 - *

Declare an F# function with type Fexpr -> string computing the textual, postfix form of
expression trees from Section 6.2.
6.3 Make a refined version of the toString function on Page 130 using the following
conventions: A subtrahend, factor or dividend must be in brackets if it is an addition or
subtraction. A divisor must be in brackets if it is an addition, subtraction, multiplication or
division. The argument of a function must be in brackets unless it is a constant or the
variable x. (Hint: use a set of mutually recursive declarations.)
6.4 Consider binary trees of type BinTree<′a,′b> as defined in Section 6.3. Declare
functions

1. leafVals: BinTree<′a,′b> -> Set<′a> such that leafVals t is the set of values
occurring the leaves of t,
2. nodeVals: BinTree<′a,′b> -> Set<′b> such that nodeVals t is the set of values
occurring the nodes of t, and
3. vals: BinTree<′a,′b> -> Set<′a>*Set<′b> such that vals t = (ls,ns), where ls is
the set of values occurring the leaves of t and ns is the set of values occurring the
nodes of t

6.5 An ancestor tree contains the name of a person and of some of the ancestors of this
person. We define the type AncTree by:
type AncTree = | Unspec
| Info of AncTree * string * AncTree;;

The left sub-tree is the ancestor tree of the farther while the right sub-tree is the ancestor
tree of the mother. Write a value of type ancTree with at least 5 nodes and make a drawing
of the corresponding tree.
Declare functions maleAnc and femaleAnc to compute the list of names of male and
female ancestors of a person in an ancestor tree.
6.6 Consider search trees of type BinTree<′a> as defined in Section 6.4. Declare an F#
function that can delete an element in such a tree. Hint: Make use of an auxiliary function
that deletes the smallest element in a non-empty search tree (and returns that value).
6.7

1. Define a type to represent formulas in propositional logic. A proposition is either an

atom given by its name which is a string, or a composite proposition built from
atoms using the operators for negation ( ), conjunction ( ), and disjunction ( ).
2. A proposition is in negation normal form if the negation operator only appears as
applied directly to atoms. Write an F# function transforming a proposition into an
equivalent proposition in negation normal form, using the de Morgan laws:

and the law: ( p) ⇔ p.

3. A literal is an atom or its negation. A proposition is in conjunctive normal form if it
is a conjunction of propositions, where each conjunct (i.e., proposition in the
conjunction) is a disjunction of literals. Write an F# function that transforms a
proposition into an equivalent proposition in conjunctive normal form, using the
above result and the laws:

4. A proposition is a tautology if it has truth value true for any assignment of truth
values to the atoms. A disjunction of literals is a tautology exactly when it contains
the atom as well as the negated atom for some name occurring in the disjunction. A
conjunction is a tautology precisely when each conjunct is a tautology. Write a
tautology checker in F#, that is, an F# function which determines whether a
proposition is a tautology or not.

6.8 We consider a simple calculator with instructions for addition, subtraction,

multiplication and division of floats, and the functions: sin, cos, log and exp.
The instruction set of the calculator is modelled by the following F# type:

type Instruction = | ADD | SUB | MULT | DIV | SIN

| COS | LOG | EXP | PUSH of float

The calculator is a stack machine, where a stack is a list of floats.

The execution of an instruction maps a stack to a new stack:
The execution of ADD with stack yields a new stack: , where the top two
elements a and b on the stack have been replaced by the single element (b+a). Similarly
with regard to the instructions, SUB, MULT and DIV, which all work on the top two elements
of the stack.
The execution of one of the instructions SIN, COS, LOG and EXP applies the corresponding
function to the top element of the stack. For example, the execution of LOG with stack
 yields the new stack: .
The execution of PUSH r with the stack pushes r on top of the stack, that is, the new
stack is: .

1. Declare a type Stack for representing the stack, and declare an F# function to
interpret the execution of a single instruction:
intpInstr: Stack -> Instruction -> Stack

2. A program for the calculator is a list of instructions [i1, i2, …, in]. A program is
executed by executing the instructions i1, i2, …, in one after the other, in that order,
starting with an empty stack. The result of the execution is the top value of the stack
when all instructions have been executed.
Declare an F# function to interpret the execution of a program:
intpProg: Instruction list -> float

3. Declare an F# function

where Fexpr is the type for expression trees declared in Section 6.2. The value of the
expression trans(fe, x) is a program prg such that intpProg(prg) gives the float
value of fe when X has the value x. Hint: The instruction list can be obtained from the
postfix form of the expression. (See Exercise 6.2.)

6.9 A company consists of departments with sub-departments, which again can have sub-
departments, and so on. (The company can also be considered as a department.)

1. Assume that each department has a name and a (possibly empty) list of sub-
departments. Declare an F# type Department.
2. Extend this type so that each department has its own gross income.
3. Declare a function to extract a list of pairs (department name, gross income), for all
departments.
4. Declare a function to extract the total income for a given department by adding up its
gross income, including the income of its sub-departments.
5. Declare a function to extract a list of pairs (department name, total income) for all
departments.
6. Declare a function format of type Department -> string, which can be used to get
a textual form of a department such that names of sub-departments will occur suitably
indented (e.g., with four spaces) on separate lines. (Use printf to print out the result.
Do not use printf in the declaration of format.)

6.10 Consider expression trees of type ExprTree declared in Section 6.5. Extend the type
with an if-then-else expression of the form: if b then e1 else e2, where b is a boolean
expression and e1 and e2 are expressions. An example could be:
 if a*3>b+c && a>0 then c+d else e

Furthermore, extend the declaration of the eval function accordingly. Hint: make use of a
mutually recursive type and function declarations.
6.11 Write the steps of the evaluation of the expression:
depthFirstFold (fun a x -> x::a) [] t1

See Page 139.

6.12 Declare the functions depthFirstFold and breadthFirstFoldBack on list trees, cf.
Section 6.6.
 7
Modules

Throughout the book we have used programs from the F# core library and from the .NET
library, and we have seen that programs from these libraries are reused in many different
applications. In this chapter we show how the user can make own libraries by means of
modules consisting of signature and implementation files. The implementation file contains
the declarations of the entities in the library while the signature file specifies the user′s
interface to the library.
Overloaded operators are defined by adding augmentations to type definitions. Type
augmentations can also be used to customize the equality and comparison operators and the
string conversion function. Libraries with polymorphic types are obtained by using
signatures containing type variables.
These features of the module system are illustrated by small examples: plane geometric
vectors and queues of values with arbitrary type. The last part of the chapter illustrates the
module system by a larger example of piecewise linear plane curves. The curve library is
used to describe the recursively defined family of Hilbert curves, and these curves are shown
in a window using the .NET library. The theme of an exercise is a picture library used to
describe families of recursively defined pictures like Escher′s fishes.
 7.1 Abstractions
A key concept in designing a good program library is abstraction: the library must provide a
service where a user can get a general understanding of what a library function is doing
without being forced to learn details about how this function is implemented. The interface to
any standard library, like, for example, the Set library, is based on useful abstractions, in this
case the (mathematical) concept of a set. Based on a general understanding of this concept you
may use the Set library while still being able to focus your main attention on other aspects of
your program.
Modules are the technical means of dividing a programming problem into smaller parts,
but this process must be guided by useful abstractions. Creation of useful abstractions is the
basis for obtaining a successful modular program design.
Abstractions are described on the semantical level by:

Description of a collection of entities to be represented by an F# type.

Description of specific values and computations to be implemented by F# values and
functions.

The above example of the Set library comes with a natural language explanation of the
concept of sets and of operations on sets plus certain special sets like the empty set and
singleton sets (cf. Section 5.2). Throughout the book we follow this style in describing the
semantics, as seen from a user, of the library in two parts:

A general, narrative description of the conceptual framework.

A precise, but short description of the intended working of each type, value and function.

Parts of this description may be included as comments in the signature file.

Abstractions are described on the syntactical level by specifications. The main forms are:
type TypeName ...
val Name : type

They are part of an F# program and found in the signature file of a module. They specify
entities to be represented and implemented in the library.
A type specification without type expression
type TypeName
hides the structure of the type from the user and this structure is then only found in the
implementation of the library. The user ’s access to the library is restricted to use values and
functions according to the specifications in the signature and the user cannot look into details
of the representation of values of this type or make such values “by hand.” This feature of the
interface to a library is called data hiding. It gives the means for protecting the integrity of
representations of values such that, for example, invariants are not violated.
 7.2 Signature and implementation
An F# module consists of F# source files that are compiled together to make a library. This
library can then later be used by other programs. The first file in the compilation is a
signature file containing specifications of the user ’s interface to the resulting program library
while the other consists of F# declarations to define the implementation of this interface.
Signature and implementation files are text files edited by the programmer while the library
file is produced by the F# compiler when signature and implementation are compiled
together.
The compilation of a pair of signature and implementation requires that the implementation
matches the signature: each specification in the signature should be implemented by a
corresponding declaration in the implementation file. A declaration in the implementation
may have a more general type than specified in the signature but the type is then specialized to
the type in the signature. Declarations of entities that do not appear in the signature are local
to the implementation and not visible to a user of the library.
 Example: vectors in the plane
Consider the example of vectors in the plane in Section 3.3. We want to make a vector library
with hidden Vector type. Besides the functions in the earlier version of this example we have
to provide a function make to make values of type Vector and a function coord to inspect a
value of this type because the user would otherwise not be able to make or inspect any such
value.
The signature should hence contain a specification of a (hidden) type Vector without type
expression:
type Vector

together with specifications of the functions. This gives the following signature of a module
with module name Vector:
module Vector // Vector signature
type Vector
val ( ~-. ) : Vector -> Vector // Vector sign change
val ( +. ) : Vector -> Vector -> Vector // Vector sum
val ( -. ) : Vector -> Vector -> Vector // Vector difference
val ( *. ) : float -> Vector -> Vector // Product wth number
val ( &. ) : Vector -> Vector -> float // Dot product
val norm : Vector -> float // Length of vector
val make : float * float -> Vector // Make vector
val coord : Vector -> float * float // Get coordinates

The implementation must contain a definition of the type Vector and declarations of all
values specified in the signature. The type Vector specified as hidden in the signature must be
a tagged type or a record type, so we have to add a tag, say V, in the type definition:
module Vector // Vector implementation
type Vector = V of float * float
let (~-.) (V(x,y)) = V(-x,-y)
let (+.) (V(x1,y1)) (V(x2,y2)) = V(x1+x2,y1+y2)
let (-.) v1 v2 = v1 +. -. v2
let ( *.) a (V(x1,y1)) = V(a*x1,a*y1)
let (&.) (V(x1,y1)) (V(x2,y2)) = x1*x2 + y1*y2
let norm (V(x1,y1)) = sqrt(x1*x1+y1*y1)
let make (x,y) = V(x,y)
let coord (V(x,y)) = (x,y)

The resulting library can be used as follows:

open Vector;;
let a = make(1.0,-2.0);;
val a : Vector

let b = make(3.0,4.0);;
val b : Vector

let c = 2.0 *. a -. b;;

val c : Vector
coord c;;
val it : float * float = (-1.0, -8.0)

let d = c &. a;;

val d : float = 15.0

let e = norm b;;

val e : float = 5.0

Note that the response from the system displays only the type name and no value when an
expression of type Vector is entered: the structure of the type is hidden from the user.
Remark: The above method of defining infix operators in a library is not recommended. One
should instead use the syntax described in Section 7.3. The reason is that the operators in this
example are only available in a program using the library when the library has been opened,
in our case by “open Vector”. One would normally prefer not to open the library in order to
limit the “pollution of the namespace”. Entities in the library are then accessed using
composite names like Vector.make, and the operators will then only be available in the
inconvenient prefix form like Vector.(+.) and the elegance of the infix operators is not
available. It is even worse if we, for instance, declare an operator + on vectors using the
above syntax. This declaration will then override the existing declaration of + when the
module is opened, and the + operator on numbers will no longer be available.
The module-declaration in signature and implementation can use a composite module name,
for example:
module MyLibrary.Vector

A program may then access the library in any of the following ways:
open MyLibrary.Vector;;
let a = make(1.0,-2.0);;

or
open MyLibrary;;
let a = Vector.make(1.0,-2.0);;

or
let a = Mylibrary.Vector.make(1.0,-2.0);;
 Files and compilation
The F# module system uses the file types fsi, fs and dll as follows:
FileName.fsi F# signature file Text file edited by programmer
FileName.fs F# implementation file Text file edited by programmer
FileName.dll Library file Binary output file from F# compiler
The separate compilation of a module is supported by any of the development platforms for
F#. Using the batch compiler fsc (cf. Section 1.10) the compilation of a library with signature
file SignatureName.fsi and implementation file LibraryName.fs is made by the command:
fsc -a Signature.fsi Library.fs
This compilation produces a library file:
Library.dll
The library gets the same file name as the implementation file – with a different file type –
while the signature can have a different file name. These file names should not be confused
with the module name introduced by the module-declarations in signature and implementation
files as Vector in the above example.
Compilation of a program using the library Library.dll is made by a command like:
fsc ... -r Library.dll ...
Compilation of library and program should use the same version of F# and .NET.
A library Library.dll may be used as follows from an interactive environment:

1. Move the library file Library.dll to a special directory, for example:

c: \Documents and Settings\FsharpBook\lib
2. Refer to the library from the interactive environment using an #r directive:
#r @'c:\Documents and Settings\FsharpBook\lib\Library.dll';;
 7.3 Type augmentation. Operators in modules
A type augmentation adds declarations to the definition of a tagged type or a record type
and it allows declaration of (overloaded) operators. The type augmentation uses an OO-
flavoured syntax. This is illustrated in the signature and implementation files in Tables 7.1 and
7.2. The operators +, - and * on numbers are overloaded to denote also operations on values
of type Vector and the operator * is even overloaded to denote two different operations on
vectors. Resolving the overloading is made using the types of the operands, so the two
versions of the operator * could not have identical operand types.

Table 7.1 Signature file with type augmentation

Table 7.2 Implementation file with type augmentation

The member declarations cannot be intermixed with let declarations (but local let
declarations are, of course, allowed inside the expressions in the declarations).
The functions make, coord and norm are specified and implemented as usual F# functions –
the OO-features should only be used to obtain an effect (here: operators) that cannot be
obtained using normal F# style.
Note that the implementation file in Table 7.2 compiles without signature file if the module
declaration is commented out. This is often convenient during the implementation and test of
a module. Furthermore, the output from the interactive F# compiler in such a compilation can
be useful in getting details in the signature correct.
Note the following:
 The attribute [<Sealed>] is mandatory when a type augmentation is used. It protects the
library against unintentional use in the OO-world.
The heading “|” in the type and the declarations in the augmentation must be at the same
indentation level.
The “member” specification and declaration of an infix operator (e.g. +) corresponds to a
type of form type1 * type2 -> type3, while the earlier val specification and let
declaration (of e.g. (+.)) on Page 151 use a higher-order type of form type1 -> type2 ->
type3. The indicated type is required to get the overloading. The resulting prefix function
(like e.g. (+)) will, nevertheless, get the usual higher-order type.
The operators +, - and * are available on vectors without opening the library. The
operators can still be used on numbers.
The functions make, coord and norm could have been declared as “static members” using
the OO notation. The usual F# form is more succinct and should be used whenever
possible.

The following are examples of use of the Vector library specified in Table 7.1:
let a = Vector.make(1.0,-2.0);;
val a : Vector.Vector

let b = Vector.make(3.0,4.0);;
val b : Vector.Vector

let c = 2.0 * a - b;;

val c : Vector.Vector

Vector.coord c;;
val it : float * float = (-1.0, -8.0)

let d = c * a;;
val d : float = 15.0

let e = Vector.norm b;;

val e : float = 5.0

let g = (+) a b;;

val g : Vector.Vector

Vector.coord g;;
val it : float * float = (4.0, 2.0)
 7.4 Type extension
The implementation in Table 7.2 can instead be made using a type extension;
type ... with ...

as shown in Table 7.3. This implementation compiles with the signature in Table 7.1 and has
the same effect as the implementation in Table 7.2, but it offers the possibility of inserting
usual function declarations between the type definition and the member declarations like make
and coord in Table 7.3. Such functions can be used in the member declarations and that may
sometime allow simplifications. This possibility is used later in the example of plane curves
in Section 7.9.

Table 7.3 Implementation module with type extension

 7.5 Classes and objects
There are full features for Object-oriented (OO) programming in F#, but this is not a major
theme of this book. We just give a brief introduction to cover the topics that are needed when
using the .NET library in an F# program and when making computational expressions as
described in a later chapter. The OO-features in F# are only used on a larger scale when
implementing applications that can be used from programs made in another .NET language.
A class definition looks syntactically like an augmented type definition where the type
expression has been removed and replaced by declarations of constructor functions. A class in
F# determines a type and a value of such a type is called an object. An object is obtained by
calling a constructor of the class. The call of a constructor is often preceded by the keyword
new. We illustrate classes and objects by an OO-version of our vector example:

type ObjVector(X: float, Y: float) =

member v.x = X
member v.y = Y
member v.coord() = (v.x, v.y)
member v.norm() = sqrt(v.x * v.x + v.y * v.y)
static member ( -) (v: ObjVector) = ObjVector(- v.x,- v.y)
static member (+) (v1: ObjVector, v2:ObjVector)
= ObjVector(v1.x + v2.x, v1.y + v2.y)
static member (-) (v1: ObjVector, v2:ObjVector)
= ObjVector(v1.x - v2.x, v1.y - v2.y)
static member (*) (a,v:ObjVector) = ObjVector(a*v.x,a*v.y)
static member (*) (v1: ObjVector, v2:ObjVector)
= v1.x * v2.x + v1.y * v2.y

The constructor ObjVector initializes the members x and y using the parameter values X and
Y. The following show some uses of the class:

let a = ObjVector(1.0,-2.0);;
val a : CbjVector

let b = ObjVector(Y = 4.0, X = 3.0);; // Named arguments

val b : ObjVector

b.coord();;
val it : float * float = (3.0, 4.0)

let c = 2.0 * a - b;;

val c : ObjVector

c.coord();;
val it : float * float = (-1.0, -8.0)

b.x;;
val it : float = 3.0

let d = c * a;;
val d : float = 15.0

let e = b.norm();;
val e : float = 5.0

let g = (+) a b;;

val g : ObjVector

g.coord();;
val it : float * float = (4.0,2.0)

The above examples

let b = ObjVector (Y=4.0, X=3.0);;

illustrates the use of named arguments where arguments in a function call are identified by
name instead of position in the argument list. Named arguments can make call of functions
from the .NET library more readable as the meaning of each argument is visible from the
context, while the meaning of an argument can otherwise only be found by studying the
documentation of the function in question.
The example of plane curves uses a similar feature called optional property setting (cf.
Section 7.9).
Note that members coord, norm and x are written as a suffix to the values c and b. They are
in the OO-world considered as belonging to the values c and b. Using fun-expressions they
determine functions
fun v -> v.coord()
fun v -> v.norm()
fun v -> v.x

where, for example:

Using OO-style constructs is daily life for the F# programmer as the .NET library is 100
percent OO, and the OO-features of F# give a quite streamlined access to this library. An
object member is used as argument of a higher-order function by packaging it into a fun-
expression as shown above.
 7.6 Parameterized modules. Type variables in signatures
A module in F# can be parameterized by type variables and may thereby implement
polymorphic types, values and functions. This is illustrated by the example of a queue: A
queue is a row of values of the same type. The put function inserts a new value at the rear end
of the queue while the get function gets the front element. An exception should be raised if
get is attempted on an empty queue. This idea is specified in the signature in Table 7.4.
The implementation uses an interesting data representation due to L.C. Paulson (cf. [10],
Chapter 7) where a queue is represented by two lists, a front list containing the first queue
elements in the order of insertion and a rear list containing the remaining queue elements in
the reverse order of insertion. The representation of a queue containing values 1, 2, 3 may
hence look as follows:
front [1]
rear [3; 2]

Using put to insert a value, say 4, will simply “cons” the value onto the rear list:
front [1]
rear [4; 3; 2]

while get removes the heading element 1 from the front list:
front []
rear [4; 3; 2]

A call of get in this situation with empty front list will reverse the rear list to get the list [2;
3; 4] with the queue elements in the order of insertion. This list is then used as front list
while the rear list becomes empty.
front [3; 4] (the front element 2 has been removed by get)
rear []

The implementation module in Table 7.5 uses this idea and represents a Queue value as a
record {front:′a list; rear:′a list} containing the two lists. Note that the
representation of a queue is not unique because different pairs of front and rear lists may
represent the same queue.

Table 7.4 Signature of parameterized Queue module

 Table 7.5 Implementation of parameterized Queue module

The Queue library can be used as follows:

let q0 = Queue.empty: Queue.Queue<int>;;
val q0 : Queue.Queue<int>

let q1 = Queue.put 1 q0;;

val q1 : Queue.Queue<int>

let q2 = Queue.put 2 q1;;

val q2 : Queue.Queue<int>

let (x,q3) = Queue.get q2;;

val x : int = 1
val q3 : Queue.Queue<int>

let q4 = Queue.put 4 q3;;

val q4 : Queue.Queue<int>

let (x2,q5) = Queue.get q4;;

val x2 : int = 2
val q5 : Queue.Queue<int>

let (x3,q6) = Queue.get q5;;

val x3 : int = 4
val q6 : Queue.Queue<int>
 7.7 Customizing equality, hashing and the string function
The F# compiler will automatically generate a default equality operator for the above type
Queue<′a> whenever the type variable ′a is instantiated with an equality type. This default
equality operator is, however, not the wanted operator because it distinguishes values that we
want to consider equal. We may for instance get a queue containing the single integer 2 in two
ways: as the above queue value q3 where the integers 1 and 2 are put into the empty queue q0
followed by a get to remove the integer 1, or as the below queue value qnew where we just
put the integer 2 into the empty queue q0. These two values are considered different by the
default equality operator:
let qnew = Queue.put 2 q0 ;;
val qnew : Queue.Queue<int>
qnew = q3;;
val it : bool = false

The reason is that the queues qnew and q3 are represented by different values of type Queue:
The value of qnew is represented by {front = []; rear = [2]}
The value of q3 is represented by {front = [2]; rear = []}

and the default equality operator is based on structural equality of the representing values.
Hence, qnew and q3 are considered different by this operator.

Table 7.6 Type definition with augmentation for equality, hashing and string

It is possible to override the default equality operator using a type augmentation as shown
in Table 7.6. The signature in Table 7.4 needs an equality constraint on the type variable ′a of
queue elements as the Equals function uses equality for ′a list values:
type Queue<′a when ′a : equality>

The signature can otherwise be used unchanged.

The Equals function contains the clause:
:? Queue<′a> as q2 -> ...

It expresses a match on type. The value of qobj matches the pattern if the type of qobj matches
the type Queue<′a> in the pattern, that is, if the type of qobj is an instance of this type.The
identifier q2 is then bound to the value of qobj.
Note the following:

The customized equality compares single lists containing all queue elements. This list
q.list() is obtained from the used representation {front=xs; rear=ys} of a queue as the
front list q.front with the reversed of the rear list q.rear appended.
The overriding cannot be given in a separate type extension. There are hence no
possibility of declaring a local function to be used in the member-declarations. The
frequently used expression q.front @ (List.rev q.rear) is therefore defined as a
member function q.list().
The compiler gives a warning if the hash function is not customized because values
considered equal should have same hash code. This condition becomes critical if the
imperative collections HashSet or Directory (cf. Section 8.11) are used with elements
of type Queue.
Overriding ToString gives a reasonable conversion of a queue to a string by using
string on q.list().

Applying the new Queue module with customized comparison and string function to the
example in Section 7.6 with declarations of q0, q1,..., q6 and s we now get:
qnew = q3;;
val it : bool = true

string q2;;
val it : string = "[1; 2]"
 7.8 Customizing ordering and indexing
Using a suitable type augmentation one may also customize the ordering: q1 < q2 and
indexing: q.[n] on values of a defined type. The corresponding type augmentation in the
queue example is shown in Table 7.7.
The ordering is declared by overriding the CompareTo method in the IComparable
interface. The implemented comparison uses compare on the lists of the queue elements in
insertion order. The signature must tell that this interface is used:
interface System.IComparable

The indexing is expressed by the get part of an Item member function. The implementation
uses list indexing in the list of queue elements in insertion order. The signature must contain
the corresponding specification:
member Item : int -> ′a with get

Table 7.7 Type augmentation for ordering and indexing in queue module

The following illustrates uses of ordering and indexing:

let q0 = Queue.empty;;
let q1 = Queue.put 1 q0;;
let q2 = Queue.put 2 q1;;

q2 > q1 ;;
val it : bool = true
q2.[1] ;;
val it : int = 2
 7.9 Example: Piecewise linear plane curves
In this example we consider piecewise linear curves in the plane following an idea due to
Fokkinga (cf. [4]). Such a curve consists of a point P0 and a (possible empty) sequence of line
segments P0P1, P1P2, ... ,Pn-2Pn−1 where P0, P1, ..., Pn−1 are points in the plane. The point P0 is
called the start point of the curve while Pn−1 is called the end point. We use usual rectangular,
Cartesian coordinates in the plane, so points and vectors in the plane correspond to
coordinates that are pairs of float numbers.
We want represent a curve by a F# value and to implement the operations on curves shown
in Table 7.8. A corresponding signature is shown in Table 7.9. Note that the user of the library
can understand and use the functions while thinking purely in geometrical terms, so we have
obtained the wanted abstraction.

Table 7.8 Operations on curves

Use of the infix operators |^ for the rotate function is overloaded to also allow integer
angle values. The infix operators allow Curve expressions to be written using a minimum of
parentheses.
 Table 7.9 Signature of Curve library

We present an application of the Curve library before presenting its implementation.

 Example: Hilbert curves
The Hilbert curves h0, h1, h2, ... form a system of curves, where h0 consists of the point with
coordinates (0,0) while each curve hn+1 is obtained by joining four curves c1, c2, c3, c4
obtained from hn by transformations composed of reflections, rotations and translations. The
Hilbert curves h0, h1, h2 and h3 are shown in Figure 7.1. All Hilbert curves start in the origin
and the connecting segments (dotted lines in the figure) are of length 1.
We want to declare a function
hilbert: Curve.Curve -> Curve.Curve

such that

Fig ure 7.1 Hilbert curves

Studying Figure 7.1 we note that c2 and c3 can be obtained from hn by parallel translations
while c1 and c4 must be obtained from a mirror image of hn. The following figure shows the
curve c0 obtained by horizontal reflection of hn in the vertical line through the start point
(0.0,0.0) and the curves obtained from c0 by rotations through −90° and 90°:

Using the width and height of hn:

w = Curve . width hn
h = Curve . height hn

we can express the coordinates of the start points P1, P2, P3, P4 of c1, c2, c3, c4:
P1 : (0.0, 0.0)
P2 : (0.0, w + 1.0)
P3 : (h + 1.0, w + 1.0)
P4 : (h + h + 1.0, w)

Note that the height and width of c1 and c4 are the width and height, respectively, of hn. The
height of hn is actually equal to its width.
These considerations leads to the wanted declaration:
let h0 = Curve.point (0.0,0.0);;
val h0 : Curve.Curve

let hilbert hn =
let w = Curve.width hn
let h = Curve.height hn
let c0 = hn >< 0.0
let c1 = c0 |^ -90
let c2 = hn --> (0.0, w + 1.0)
let c3 = hn --> (h + 1.0, w + 1.0)
let c4 = (c0 |^ 90) --> (h + h + 1.0, w)
c1 + c2 + c3 + c4;;
val hilbert : Curve.Curve -> Curve.Curve

Note that the programming of the hilbert function has been done using geometric concepts
only. We do not need any knowledge about the implementation of the Curve library.
 Displaying curves
We want to make a function to display a curve in a window using the .NET library. Before
getting to the programming we have to make some geometric considerations.
The display is made in a panel belonging to a window. The panel uses Cartesian
coordinates where the y-axis points downwards and the upper left corner of the panel has
panel coordinates (0,0). The situation is depicted in Figure 7.2. The thick box is the panel with
width pw and height ph. The picture shows that a curve point with coordinates (x, y) has panel
coordinates:

The program uses two libraries:

System.Windows.Forms containing facilities to set up a window with scroll-bars and
underlying panel to contain the drawing
System.Drawing containing facilities to draw the curve in the panel

Fig ure 7.2 Panel coordinates

 Table 7.10 The display function

The function display declared in Table 7.10 has two parameters: the title to be written on
top of the window and a triple comprising the Curve to be displayed plus width pw and height
ph of the panel. The function consists of five parts:

1. The function f converts a set of coordinates (x, y) to a Point object containing the
corresponding panel coordinates. The panel coordinates are integers and the conversion
from float to int consists of a round followed by an int conversion. The formula (*)
on Page 165 is used in converting to panel coordinates.
The list clst of coordinates of points on the curve is extracted and the function f is
applied to each element to get the corresponding list Ptlst of Point objects. Finally the
corresponding array pArr of Point objects is made. It is ready to be used by the
Graphics member function DrawLines.
2. A Pen object pen is created and a function draw drawing the curve on a Graphics object
is declared. It calls DrawLines using pen and the array pArr of curve point coordinates.
3. The Panel object is created and configured to fill all of the window (DockStyle). The
draw function is added to the panel’s collection of Paint objects
4. The window (Form) is created using the specified title. The size is set and scrolling is
enabled. The value of AutoScrollMinSize is set to allow the window to scroll to any
part of the panel, and scrolling is activated. Finally, the panel is added to the collection of
Controls of the window.
5. The window is shown (win.Show()).
A window is a live object that handles a number of events. The actual window has only events
corresponding to manipulation of the window like: resizing of window, use of scroll-bars,
window comes in to foreground. The part of the panel inside the window is then redrawn
using the function in the Paint collection of the panel. The parameter e is actually an Event
object.
Some of the above calls of constructors use optional property setting like the argument
Dock=DockStyle.Fill in the argument list of constructor Panel. This constructor has
actually no Dock argument. The specified value DockStyle.Fill is instead used as initial
value of the Dock property of the created Panel object.
A curve requires some adjustment before the display function can be used: the curve must be
suitable scaled to get a proper size of the details of the curve, and the curve must be moved
away from the boundary of the panel as boundary points are invisible. This job is done by the
function adjust. It multiplies the curve c by the factor a and makes a parallel translation of
the curve to leave a blank band of 10 pixels in the panel around the curve:
let adjust(c:Curve.Curve, a: float) =
let c1 = a * c --> (10.0, 10.0)
let (_,(maxX,maxY)) = Curve.boundingBox c1
let pw = int(round maxX) + 20
let ph = int(round maxY) + 20
(c1,pw,ph);;

The value of adjust can be used directly as second parameter in the display function.
 Displaying Hilbert Curves
The display function can be used to display Hilbert curves (using the above declarations of
value h0 and function hilbert):
let h1 = hilbert h0;;
let h2 = hilbert h1;;
let h3 = hilbert h2;;
let h4 = hilbert h3;;

let h5 = hilbert h4;;

let h6 = hilbert h5;;
display("Hilbert Curve 6", adjust(h6, 10.0));;

The displayed curve has been scaled by a factor 10.0 to get a reasonable drawing.
 Implementation of the Curve library
The F# implementation of the Curve library is given in Tables 7.11 and 7.12.
A curve is represented by a value of tagged type
C of (float * float) * ((float * float) list)

containing the coordinates of the start point of the curve plus the (possibly empty) list of
coordinates of the remaining points. Any value of this type represents a curve (there is no
invariant) and the functions can hence be implemented without any error case where an
exception should be raised.

Table 7.11 First part of the implementation of Curve library

 Table 7.12 Last part of the implementation of Curve library

A simplification is obtained by introducing two higher-order local functions. The first

function map applies a function f to the coordinates of each curve point including the start
point. It is used in declaring functions like parallel translation --> and reflection >< where the
same transformation is applied to all curve points.
The second function mapP leaves the start point p0 unchanged and applies a partially
evaluated function g p0 to the coordinates of the remaining curve points. It is used in
declaring functions like multiplication * or rotation |^.
The combined curve c1+c2 is obtained as the start point of c1 together with the list of
remaining points of c1 with all points of c2 appended.
The multiplication with factor a from the point P0 : (x0, y0) maps a point P : (x, y) to the
point P′ : (x′, y′) where , that is;

and the function multA is declared accordingly.

The declaration of rot is based on the fact that the rotation with angle v around P0 maps a
point P in to the point P′ where is obtained from by a rotation with angle v.
The function minmax extends a (bounding) box with lower left corner (minX,minY) and
upper right corner (maxX,maxY) to contain also the point (x,y). The bounding box of the curve
is then obtained starting with the one-point box containing the start point and folding the
minmax function over the remaining points of the curve.
 Summary
We have introduced the notions of module, signature and implementation – concepts that are
needed when a programmer makes his own libraries. Moreover, we have introduced the
notion of type augmentation and shown how it can be used to declare overloaded operators
and to customize the equality and comparison operations and the string conversion.
 Exercise
7.1 Make an implementation file of the vector example in this section using a record type:
type Vector = {x: float; y: float}

while using the same signature file.

7.2 Make signature and implementation files for a library of complex numbers with
overloaded arithmetic operators (cf. Exercise 3.3).
7.3 Make signature and implementation files for a library of multi-sets of integers
represented by weakly ascending lists (cf. Exercise 4.11).
7.4 Make signature and implementation files for a library of polynomials with integer
coefficients (cf. Exercise 4.22).
7.5 Customize the string function in the library of polynomials inExercise 7.4.
7.6 Make an indexing in the library of multi-sets of integers in Exercise 7.3 such that the
value of s.[n] is the number of occurences of n in the multi-set s.
7.7 Make an indexing in the library of polynomials in Exercise 7.4 such that p.[n] is the
coffecient to xn in the polynomium p.
7.8 The Sierpinski curves s0, s1, s2, ... are a system of curves, where the curve sn+1 is
obtained by joining four curves which are obtained from the curve sn by transformations
composed of reflections, rotations and translations.

The figure shows the Sierpinski curves s0, s1 and s2 and how each of the curves s1 and s2
is obtained by joining four curves. All vertical and horizontally segments in a Sierpinski
curve have length 1 and all curves s0, s1, ... start in the origin. Use the Curve library to
declare the function sierpinski that computes the curve sn+1 from the curve sn for any n
= 0, 1, ... . Use this function to display the curve s4 in a window.
7.9 The Peano curves p0, p1, p2, ... are a system of curves, where the curve pn+1 is obtained
by joining 9 curves which are obtained from the curve pn by transformations composed of
reflections, rotations and translations.
The figure shows the Peano curves p0, p1 og p2 and how each of the curves p1 and p2 is
obtained by joining 9 curves. All Peano curves start in the origin and the joining segments
(thin lines in the Figure) are of length 1. Use the Curve library to declare the function
peano that computes the curve pn+1 from the curve pn for any n = 0, 1, ... (in getting from
pn to pn+1 it may be convenient to group the 9 curves into 3 groups each consisting of 3
curves and first build the curve for each of these 3 groups). Use the peano function to
make a program display the curve p4 in a window.

7.10 Add a minus operator of type Curve -> Curve to the Curve library. It should
compute the reversed curve, that is, - c should contain the same point as c but taken in the
opposite order.
7.11 Make a library for manipulation of pictures (following ideas due to Henderson, cf.
[16]). A picture is a set of segments together with a rectangular, upright bounding box in
the plane. The bounding box is not shown when drawing a picture but it is used when
defining operations on pictures. We use usual rectangular, Cartesian coordinates in the
plane, so points in the plane are represented by coordinates which are pairs (x, y) of float
numbers. The point with coordinates (0.0,0.0) is called the origin of the plane. A picture
is normally placed in the coordinate system such that the bounding box is situated in the
lower left corner of the first quadrant.
If c is a float number with c > 0.0 then a picture can be scaled by factor c by mapping
each point (x, y) to the point (c*x, c*y). The scaled picture will have width c*a and height
c*b where a and b are width and height of the original picture. Scaling is used in some of
the below operations in order to adjust the width or the height of a picture.
The library should contain the following functions on pictures:
Grid: Computes a picture directly from width and height of the bounding box and the
coordinates of the pairs of end-points of the segments in the picture. The function should
be declared such that all the numbers in the input are integers (the function must convert to
float numbers as used in the value representing a picture).
Rotate: Computes the picture p′ obtained from the picture p by first rotating 90° in the
positive (counter-clockwise) direction around the origin and then translating the resulting
picture to the right to get its lower left-hand corner into the origin. The height of p′ will be
the width of p and the width of p′ will be the height of p.
Flip: Computes the picture obtained from a picture by horizontal reflection around the
vertical line through the middle of the bounding box.
 Beside: Computes the picture obtained from two pictures p1 and p2 by uniting p1 with a
version of p2 that has been placed to the right of p1 and scaled to the same height.

Above: Computes the picture obtained from two pictures p1 and p2 by uniting p2 with a
version of p1 that has been placed on top of p2 and scaled to the same width.
Row: Computes the picture obtained by placing n copies of a picture p beside each other.
Column: Computes the picture obtained by placing n copies of a picture p on top of each
other.
Coordinates: Computes the pair ((width, height), segmentList) where width and height are
width and height of a picture while segmentList is a list of coordinate pairs ((x, y), (x′, y′))
of end-points of the segments in the picture.
You should chose your own names of the functions and use operators whenever
appropriate. Furthermore, you should implement a function to display a picture in a
window. (Hint: DrawLine(Pen,Point1,Point2) draws a segment.)
The library should be used to construct pictures of persons and Escher ’s fishes – as
described in the following.

Persons

Fig ure 7.3 The man picture and derived pictures

The starting point is the picture man shown in Figure 7.3. It has width 14 and heigh 20.
Using the functions on pictures you should now make programs to construct the pictures
couple and crowd shown in Figure 7.3.

Escher’s fishes
The starting point of Escher ’s fishes is the four (16 × 16) pictures p, q, r, and s shown in
Figure 7.4. By combining these four pictures we get the picture t in Figure 7.5, while the
picture a is obtained by combining suitably rotated copies of q. Finally the picture b1 is
obtained by combining two suitably rotated copies of t.
The Escher fish pictures e0, e1 and e2 are now obtained by combining the pictures in
Figure 7.5 as shown in Figure 7.6. The pictures b2, b3 and b4 are obtained from b1 by
successive rotations. The transition from an Escher picture to the next adds a border around
the picture consisting of a picture a in each corner, a row of b1’s at the top, a column of b2’s
at the left, a row of b3’s at the bottom, and a column of b4’s at the right. In this border there
will be one b1 on top of an a and two b1’s on top of a b1, one b2 to the left of an a and two
b2’s to the left of a b2, etc.
You should make a program to generate the Escher fish pictures e0, e1 and e2.

Fig ure 7.4 Basic fish pictures

Fig ure 7.5 Escher fish building blocks

Fig ure 7.6 Building Escher fish pictures

 8
Imperative features

Imperative features are language constructs to express commands using and modifying the
state of mutable objects. The working of these features are explained using a new component,
the store, beside the environment. A store consists of a set of locations containing values,
together with operators to access and change the store. We explain why the restriction on
polymorphically typed expressions is needed because of the imperative features. The F#
concept of a mutable record field gives the means of assigning values to object members. The
while loop is introduced and we study its relationship with iterative functions. The F# and
.NET libraries offer imperative collections: arrays, imperative sets and imperative maps.
 8.1 Locations
A (mutable) location is a part of the computer memory where the F# program may store
different values at different points in time. A location is obtained by executing a let mutable
declaration, for example:
let mutable x = 1;;
val mutable x : int = 1

The keyword mutable requests F# to create a location, and the answer tells that x is bound to
a location of type int, currently containing the value 1. The situation obtained is illustrated as
follows:
x loc1 loc1: 1

Locations may be of any type:

let mutable y = (3,(5,8));;
val mutable y : int * (int * int) = (3, (5, 8))

let mutable z = [1;2;3];;

val mutable z : int list = [1; 2; 3]

let mutable w = Some (3,(5,8));;

val mutable w : (int * (int * int)) option
= Some (3, (5, 8))

let mutable f = sin ;;

val mutable f : (float -> float)

After these declarations we have the following situation:

Environment Store
x loc1 loc1: 1
y loc2 loc2: (3,(5,8))
z loc3 loc3: [1;2;3]
w loc4 loc4: Some (3,(5,8))
f loc5 loc5: “the sine function”

A set of locations with contents is called a store.

The F# concept of location should be interpreted in an “abstract” sense: An F# location of
basic type (like int) corresponds to a physical memory location in the computer containing
the stored value. An F# location of, for example, a list type corresponds on the other hand to
a physical memory location containing a link to the list that is stored elsewhere in memory.
This difference in management of physical memory is, however, not visible from the F#
program.
We shall study the memory management of the system in Chapter 9.
 8.2 Operators on locations
There are two operations on a location:

An assignment expression exp1 <− exp2 is evaluated as follows:

1. Evaluate exp1 to get a location loc.

2. Evaluate exp2 to get a value v.
3. Store the value v in the location loc.
4. Deliver the value () of type unit as the result.

We may, for example, assign the value 7 to x:

x <- 7;;
val it : unit = ()

The evaluation proceeds as follows:

This evaluation has a side effect: the content of the location loc1 is changed. It is called a side-
effect because it is not visible in the result which is just the uninteresting value () of type
unit. Note that the binding of x is unaffected by the assignment: x remains bound to the same
location.
We may do more complex assignments:
f <- cos;;
val it : unit = ()

y <- (-2, (3,1));;

val it : unit = ()

but the value assigned to a location must have the same type as the location, so the following
attempt fails:
x <- (2,3);;
x <- (2,3);;
——^^^
...: error FS0001: This expression was expected to
have type
int
but here has type
 ′a * ′b

A contentsOf expression exp is evaluated as follows:

1. Evaluate exp to get a location loc.

2. Deliver the contents v of the location loc as the result.

The contentsOf operator has no visible operator symbol and the operator is automatically
inserted by the F# compiler according to the following coercion rule:

A contentsOf operator is automatically inserted by the F# compiler whenever a location

occurs in a context where a value is required. The right-hand side of a let-declaration is
such a context.

Assume that we have the binding of x and corresponding location loc1 as above:
x loc1 loc1: 7

The evaluation of the expression x <− x + 1 will then proceed as follows:

x <− x + 1 x loc1 loc1: 7
loc1 <− loc1 + 1 x loc1 loc1: 7
loc1 <− contentsOf(loc1) + 1 x loc1 loc1: 7
loc1 <− 7 + 1 x loc1 loc1: 7
loc1 <− 8 x loc1 loc1: 7
() x loc1 loc1: 8

The contentsOf operator is inserted by the coercion rule because the plus operator requires a
value as operand.
The coercion rule also apply when making a declaration:
let t = x;;
val t : int = 8

where the right-hand side x would otherwise evaluate to a location. The identifier t is hence
bound to the value 8:
x loc_1 loc_1: 8
t 8

An assignment to x will have no effect on t as there is no connection between the value stored
in the location loc1 and the value bound to t:
x <- 17;;
val it : unit = ()
t ;;
val it : int = 8

The coercion rule also apply if we enter the identifier x as an expression to be evaluated by
F# because this is interpreted as a declaration let it = x of the identifier it:
x;;
val it : int = 17

Hence, a location is not a value but an expression may evaluate to a location when, for
example, used as the left-hand side of an assignment.
Locations cannot occur as components in tuples or tagged values, as elements in lists or as
values of functions. The situation for records is different as described in Section 8.5.
 A remark on the mutable declaration
The “mutable” keyword in the declaration let mutable x = 1 describes a property of the
entity to be bound to the identifier x and not a property of x. A syntax like
let x = mutable 1 // NOT legal F#

might hence have been more appropriate as the declaration creates a mutable location to be
bound to the identifier x. The actual F# syntax has the advantage that the coercion rule
automatically applies when the right-hand side evaluates to a location like in the above
example let t = x where x is bound to a location but t becomes bound to a value.
 8.3 Default values
A mutable declaration requires an initial value to be stored in the location. This value is
obtained by evaluating the expression on the right-hand side of the declaration. It is
sometimes awkward or even impossible for the programmer to make a proper initial value at
the time of declaration. One may then use the default value of the type in question:
Unchecked.defaultof<type>

on the right-hand side of the declaration. Such a value is available for any type. It may serve
as a placeholder in the location until replaced by a proper value. Default values should be
used only for this purpose.
 8.4 Sequential composition
The semicolon symbol “;” denotes the sequential composition operator (while the double
semicolon “;;” is a terminator symbol). This operator combines two expressions exp1 and
exp2 to form a new expression:
exp1; exp2

The expression exp1 ; exp2 is evaluated as follows:

1. Evaluate exp1 and discard the result.

2. Evaluate exp2 and supply the result as the result of evaluating exp1; exp2.

Hence, if exp2 has type τ then exp1 ; exp2 has type τ as well.
The F# compiler issues a warning if exp1 is of type different from unit as the result of the
evaluation might be of some use. This warning is avoided by using the ignore function:
ignore(exp1 ) ; exp2 or exp1|> ignore ; exp2

where ignore a = () for any value a.

We may combine two assignments using sequential composition:
let mutable x = 5;;
let mutable y = 7;;
x <- y + 1 ; y <- x + 2;;
(x,y);;
val it : int * int = (8 ,10)

Note that the second assignment uses the new value stored in the location denoted by x.
The operator “;” may be omitted if the expressions are written on separate lines, that is,
exp1
exp2

means (exp1 ) ; exp2.

 8.5 Mutable record fields
A mutable record field is obtained by prefixing the label in the record type declaration by the
keyword mutable, for example:
type intRec = { mutable count : int };;

Executing a declaration of a value of this type

let r1 = { count = 0 };;
val r1 : intRec = {count = 0;}

creates the following entities:

1. A value (record) of type intRec,

2. a location containing the value 0 (due to keyword mutable in the record type),
3. a local binding inside the record of the record label count to the location, and
4. a binding of the identifier r1 to the record.

So the following is added to environment and store:

Environment Store
r1 { count loc2 } loc2: 0

One may assign a value to the count field in r1:

r1.count <- 5;;
val it : unit = ()

This assignment changes the contents of the associated location:

Environment Store
r1 { count loc2 loc2: 5

One may declare a function incrementing the counter value of an intRec record and
delivering the new counter value as the result:
let incr (x: intRec) =
x.count <- x.count + 1
x.count;;
val incr : intRec -> int

incr r1;;
val it : int = 6

incr r1;;
val it : int = 7

We may even declare a function returning a closure with an internal counter:

let makeCounter() =
let counter = { count = 0 }
fun () -> incr counter;;
val makeCounter : unit -> (unit -> int)

let clock = makeCounter();;

val clock : (unit -> int)

clock();;
val it : int = 1

clock();;
val it : int = 2

Equality of records with mutable fields is defined as for records without such fields. Consider
the declarations:
let x = { count = 0 };;
val x : intRec = {count = 0;}
let y = { count = 0 };;
val x : intRec = {count = 0;}
let z = y;;
val z : intRec = {count = 0;}

The values bound to x, y and z are considered equal:

x = y;;
val it : bool = true
y = z;;
val it : bool = true

An assignment to the count field in the record bound to y has interesting consequences:
y.count <- 1;;
val it : unit = ()
x = y;;
val it : bool = false
y = z;;
val it : bool = true
z;;
val it : intRec = {count = 1;}

The assignment to the count field of y has hence not only changed y but also z. Environment
and store give the explanation: the declarations create the following environment and store
(prior to the assignment of y.count) where x=y and y=z:
Environment Store
x { count loc3 } loc3: 0
y { count loc4 } loc4: 0
z { count loc4 }

The assignment changes the store but leaves the environment unchanged:
Environment Store
x { count loc3 } loc3: 0
y { count loc4 } loc4: 1
z { count loc4 }

The crucial point is that the records bound to y and z share the location loc4. One says that z is
an alias of y. Sharing and aliasing can have unexpected and unpleasant effects in imperative
programming. These phenomena do not exist in pure functional programming where a value
is immutable – not to be changed.
It should be remembered that a record is a value in F# – assignment to a record is not
possible. If required, one may declare a location containing a record:
let mutable t = x;;
val mutable t : intRec = {count = 0;}

This gives the following environment and store:

Environment Store
x { count loc3 } loc3: 0
y { count loc4 } loc4: 1
z { count loc4 }
t loc5 loc5: { count loc3 }

and one may, for example, assign the value of y to t:

t <- y;;
val it : unit = ()
t;;
val it : intRec = {count = 1;}

This assignment changes the contents of loc5 to the value {count loc4}.
The above examples illustrate some of the (pleasant and unpleasant) features of records
with mutable fields. The real importance is, however, their key role in handling objects from
F#. An assignable member of an object appears in F# as a mutable record field that can be
assigned using the <- operator, for example:
open System.Globalization;;
open System.Threading;;
Thread.CurrentThread.CurrentCulture <- CultureInfo "en-US";;

modifying the value of the CurrentCulture member of the CurrentThread object.

 8.6 References
The F# compiler does not accept use of locally declared mutables in locally declared
functions.1 The above clock example could hence not be made without using records.
The ref type provides a shorthand for a record type containing a single mutable field, and
the ref function provides a shorthand for a value of this type. They appear 2 as defined as
follows:
type ′a ref = { mutable contents: ′a }
let ref v = { contents = v }
let ( !) r = r.contents
let (:=) r v = r.contents <- v

The declaration
let x = ref [1; 2];;
val x : int list ref = {contents = [1;2];}

will hence give the following extension of environment and store:

Environment Store
x { contents loc loc: [1;2]
with a location loc of type int list containing the value [1;2], and a binding of x to the
record {contents loc}.
The operators ! and := work as follows:
!x;;
val it : int list = [1; 2]]
x := [3;4];;
val it : unit = ()
!x;;
val it : int list = [3; 4]

The makeCounter example in the previous section can be made using references:
let incr r = (r := !r + 1 ; !r);;
let makeCounter() =
let counter = ref 0
fun () -> incr counter;;
 8.7 While loops
If b denotes an expression of type bool and e denotes an expression of any type, then
while b do e

will be an expression of type unit. This expression is evaluated as follows:

1. Evaluate the expression b.

2. If the result is true, then evaluate the expression e and repeat the evaluation of the
expression while b do e. If the result is false, then terminate the evaluation of while b do
e and return the result () of type unit.

These rules are expressed in the following evaluation steps for a while loop:

where we have used the sequential composition operator “;” (cf. Section 8.4).
A while loop is only useful when evaluated in a context where some identifiers are bound
to mutables. The expression e should contain assignment to some of these mutables and b
should comprises tests of some of their values.
The while loop:
while b do e
has the same evaluation steps as the expression:
wh()

where the function wh is declared by:

let rec wh() = if (b) then ((e) ; wh()) else ();;

The evaluation of wh() will evaluate the expression e repeatedly until b becomes false and that
is exactly what is done by the evaluation of the while loop (we assume that the identifier wh
does not occur in b or e). Thus, any while loop can be expressed by a recursive function
declaration.
It should be noted that the F# compiler generates essentially the same binary code for the
while-loop and the function wh (the recursive call wh() is compiled to a branch instruction).
There is hence no performance advantage in using the loop instead of the recursive
declaration. See also Section 9.5, especially the examples on Page 211.
 8.8 Imperative functions on lists and other collections
The F# library contains functions iter and iteri on lists:

These functions are used to iterate the effect of an imperative function over the elements of a
list. Let

be a list with elements of type ′a and let

be an imperative function. The evaluation of the expression

List.iter f lst

will then successively apply the function f to the elements v0, v1, …, vn−1 of the list. The result
(of type unit) of the evaluation is of no interest and the interesting part of the evaluation is
the side-effect. The following is a (not very interesting) application of List.iter:
let mutable sum = 0;;
let f x = sum <- sum + x;;
List.iter f [1; 2; -3; 5];;
val it : unit = ()
sum;;
val it : int = 5

The function iteri includes the index k of the element vk in the computations. Let f be a
function of type
f: int -> ′a -> unit

The evaluation of the expression

List.iteri f lst

will then successively evaluate the function calls:

The interesting part of the evaluation is again the side-effect. The following is another (not
very interesting) application of List.iteri:
let mutable t = 0;;
let f k x = t <- t + k * x;;
List.iteri f [1; 2; -3; 5];;
t;;
val it : int = 11

The evaluation of List.iteri f [1; 2; -3; 5] accumulates the value:

0 + 0 * 1 + 1 * 2 + 2 * (-3) + 3 * 5 = 11

in the variable t.
The functions iter and iteri on other collections like Seq, Set and Map work in a similar
way.
We refer to Exercise 9.14 for an analysis of the run time of the function List.iter and
Set.iter.
 8.9 Imperative tree traversal
The tree traversal functions introduced in Section 6.3 and Section 6.6 have imperative
counterparts where an imperative function is called whenever an element in a tree is visited.
Such functions are useful in many applications and give for instance a convenient way of
producing output while traversing a tree. In solving a programming problem you will often
make problem-specific imperative traversal functions corresponding to a problem-specific
tree type. The declarations in this section and in Section 8.13 could then serve as models for
such declarations.
The imperative traversal functions on a binary tree are obtained directly from the definition
of the traversal in Section 6.3 using the same binTree<′a> type:
let rec preIter f = function
| Leaf -> ()
| Node(tl,x,tr) -> f x ; preIter f tl ; preIter f tr;;
val preIter : (′a -> unit) -> BinTree<′a> -> unit
let rec inIter f = function
| Leaf -> ()
| Node(tl,x,tr) -> inIter f tl ; f x ; inIter f tr;;
val inIter : (′a -> unit) -> BinTree<′a> -> unit

and similar for postIter. Applying, for example, preIter to the tree t4 in Section 6.3 gives:
preIter (fun x -> printf " %d" x) t4;;
5 0 -3 2 7

We may in a similar way define a function for imperative depth-first traversal of list trees as
described in Section 6.6:
let rec depthFirstIter f (Node(x,ts)) =
f x ; List.iter (depthFirstIter f) ts;;
val depthFirstIter : (′a -> unit) -> ListTree<′a> -> unit

Applying this function to the tree t1 in Section 6.6 we get:

depthFirstIter (fun x -> printf " %d" x) t1;;
1 2 5 3 4 6 7

The breadthFirstIter function is declared in Section 8.13.

 8.10 Arrays
The addresses in the physical memory of the computer are integers. Consider a sequence of n
equally sized contiguous memory locations loc0, loc1, …, locn−1 as shown in Figure 8.1.

Fig ure 8.1 Memory layout of an array

The physical address physAdrk of the k′th location lock can in this situation be computed by the
formula:

where s denotes the size of one location. The machine code computation of physAdrk requires
hence only two arithmetic operations.
This addressing scheme is used to implement arrays. An array of length n consists of n
locations loc0, loc1, …, locn−1 of the same type. The numbers 0, 1, …, n-1 are called the
indices of the elements. The type of the array is written
τ []

where τ is the type of the elements.

Arrays have the advantage over lists that any array location can be accessed and modified
in a constant (short) time, that is, in a small number of computations which is independent of
the size of the array. On the other hand, an array is a mutable object — the old value is lost
when a location is modified. Furthermore, an array cannot be extended by more elements in a
simple way as the adjacent physical memory (after the last element in the array) might be
occupied for other use. A selection of operations on arrays is shown in Table 8.1.
An array can be entered using the “[|…|]” notation, for example:
let a = [|4;5;6;7|];;
val a : int [] = [|4; 5; 6; 7|]

let b = [|′a′ .. ′f′|];;

val b : char [] = [|′a′; ′b′; ′c′; ′d′; ′e′; ′f′|]

Individual array locations are assigned using indexing:

b.[2] <- ′z′;;
val it : unit = ()
b;;
val it : char [] = [|′a′; ′b′; ′z′; ′d′; ′e′; ′f′|]
Table 8.1 Operations on arrays
 Example: Histogram
Arrays are very convenient when counting frequencies and making a histogram. The
following small program reeds a text file given by its directory path and count the frequency
of each character ′A′ to ′Z′ (not distinguishing small and capital letters) and prints the
resulting histogram. The reader may consult Section 10.3 about the used text I/O functions and
Section 10.7 about printf formats.
open System;;
open System.IO;;

let Acode = int ′A′

let histogram path =

let charCount = [| for n in ′A′..′Z′ -> 0 |]
let file = File.OpenText path
while (not file.EndOfStream) do
let ch = char( file.Read() )
if (Char.IsLetter ch) then
let n = int (Char.ToUpper ch) - Acode
charCount.[n] <- charCount.[n] + 1
else ()
file.Close()
let printOne n c = printf "%c: %4d\n" (char(n + Acode)) c
Array.iteri printOne charCount;;

Calling the function histogram on the path of the source file histogram.fsx will, for
example, give the output:
A: 20
B: 0
C: 22
...
X: 1
Y: 3
Z: 1

where we have only shown some of the lines.

 8.11 Imperative set and map
The .NET library System.Collections.Generic contains classes implementing imperative
sets and maps:

The classes SortedSet<′a> and SortedDictionary<′a,′b> are imperative versions of set<

′a> and map<′a,′b> where the member functions modify the collection in a “destructive”
update without retaining the old value. They should only be used in algorithms using and
maintaining a single “current” collection without ever referring to “old” values.
The HashSet<′a> and Dictionary<′a,′b> are implemented using hash-key technique: The
basic data structure is an array (say of length N) where an element a (entry (a,b)) is stored in
the array location with index:

where hash is the hash function of the equality type ′a. This is a rather efficient scheme, but it
runs into problems when multiple elements (entries) have the same index and hence should be
stored in the same array location. This collision problem is solved by storing the colliding
elements (entries) in a linked structure that can be accessed via the index value. The HashSet<
′a> and Dictionary<′a,′b> collections have the following characteristics:

The operations are very efficient.

They are strictly imperative with destructive updating.
They do not offer a traversal of the elements (entries) in sorted order.

A selections of operations on the imperative set and map classes are shown in Tables 8.2 and
8.3.
Indexing in a Directory by a key can be used to update a value by assignment:
map.[key] <- newValue
and this construction may also be used to add a new entry to the directory.
Table 8.2 A selection of operations on SortedSet and HashSet

Table 8.3 A selection of operations on SortedDictionary and Dictionary

 8.12 Functions on collections. Enumerator functions
Operations on collections should preferably be done using standard library functions, but this
is not always feasible. This section presents means for defining functions on collections in a
way that resembles the definition of functions on lists using list patterns as described in
Chapter 4.
 Enumerator functions
The System.Collections.Generic library contains imperative features for element-wise
traversal of any of the collections – including the F# collections list, set, map, etc. The
enumerator function of the book (to be declared on Page 193) makes these features available
in a functional setting. Applying enumerator to a collection yields a function:

where elementType is determined as follows:

An element entry of type KeyValuePair<′a,′b> corresponds to an entry in the map or

dictionary, and it has components:
entry.Key of type ′a
entry.Value of type ′b
Applying enumerator to a set creates an imperative enumerator function where successive
calls yield the elements in the set:
let f = enumerator (Set.ofList [3 ; 1; 5]);;
val f : (unit -> int option)

f();;
val it : int option = Some 1

f();;
val it : int option = Some 3

f();;
val it : int option = Some 5

f();;
val it : int option = None

Applying the function enumerator to a dictionary creates a similar enumerator function but
the entries are obtained as values of the corresponding KeyValuePair type:
let d = SortedDictionary<string,int>();;
d.Add("cd",3) ; d.Add("ab",5);;
let g = enumerator d;;
val g : (unit -> KeyValuePair<string,int> option)
g();;
val it : KeyValuePair<string,int> option =
Some [ab, 5] {Key = "ab"; Value = 5;}
g();;
val it : KeyValuePair<string,int> option =
Some [cd, 3] {Key = "cd"; Value = 3;}
g();;
val it : KeyValuePair<string,int> option = None

The elements in a set<′a> or SortedSet<′a> collection are traversed in the order given by
the ordering in the type ′a while the elements in a HashSet<′a> collection are traversed in
some order depending on the hashing function and the order in which elements were added.
The entries in a map<′a,′b> or SortedDictionary<′a,′b> collection are traversed in the
order given by the ordering in the type ′a while the elements in a Dictionary<′a,′b>
collection are traversed in some order depending on hashing function and the creation of the
collection.
The enumerator may, for example, be used to define the tryFind function on sets:
let tryFind p (s: Set<′a>) =
let f = enumerator s
let rec tFnd () =
match f() with
| None -> None
| Some x ->
if (p x) then Some x else tFnd()
tFnd();;
val tryFind : (′a -> bool) -> Set<′a> -> ′a option
when ′a : comparison

let s = Set.ofList [1;3;4;5];;

val s : Set<int> = set [1; 3; 4; 5]

tryFind (fun n -> n%2 = 0) s;;

val it : int option = Some 4

We refer to Exercise 9.13 for an analysis of the run time of this function and the version
declared on Page 109 .
 Declaring the enumerator function
The declarations of the enumerator function is based on the concept of an enumerator object
for a collection. An enumerator object is a mutable data structure that is able to describe a
traversal of the collection by pointing to each element one after the other. The implementation
of the enumerator object depends on the data representation of the associated collection but all
enumerator objects are of the same polymorphic type IEnumerator<′c>. This polymorphism
has been obtained by letting any enumerator implement the interface:
type IEnumerator<′c> =
abstract Current : ′c
abstract MoveNext : unit -> bool;;

for some type ′c. An enumerator object enum points to the element enum.Current and it is
forwarded to the next element by evaluating enum.MoveNext(). An initial call of MoveNext is
required to get a fresh enumerator to point to the first element, and the value of MoveNext
becomes false when the enumerator gets beyond the last element in the collection.
Each collection has its own GetEnumerator member to create an enumerator object. This
object gets the following type:

The implementation is made such that the GetEnumerator member for any specific collection
can be considered an instance of a polymorphic GetEnumerator member working on any
collection. This polymorphism has been obtained by letting each collection implement the
interface:
type IEnumerable<′c> =
abstract GetEnumerator : unit -> IEnumerator<′c>

Any collection is hence construed as an object of type IEnumerable<′c> where ′c is the type
parameter of the enumerator object as described above.
The enumerator function refers to the collection using the IEnumerable type and may
hence be applied to any collection. It creates a reference e to an enumerator object and this
reference is then used inside a local function f that is returned as the result. A reference is
required because of the restriction on the use of mutable in closures:
open System.Collections.Generic;;

let enumerator (m: IEnumerable<′c>) =

let e = ref (m.GetEnumerator())
let f () =
match (!e).MoveNext() with
| false -> None
| _ -> Some ((!e).Current)
f;;
val enumerator : IEnumerable<′c> -> (unit -> ′c option)
 Enumerators versus list patterns
The elegant and efficient use of list patterns as described in Chapter 4 depends on the fact that
the tail of a list is represented directly by a sub-component of the data structure representing
the list. Matching a list with the pattern x::xs is hence a very fast operation, and the same
applies to matching with other list patterns.
The situation is different for set and map collections. Consider, for example, a set s
represented by a balanced search tree. Finding the first (least) element x of s is a very fast
operation, but the remaining part of the search tree consists of two separate trees and is not
represented by a sub-tree of the search tree. The time for computing Set.remove x s is in fact
proportional to the depth of the search tree and hence proportional to the logarithm of the
number of elements in the set. The computation time for making a complete recursion over a
set with successive computations Set.remove x s of trees representing sub-sets is hence
proportional to n log n where n is the number of elements in the set.
The enumerator function of a set (or map) represented by a search tree is based on the idea
of in-order traversal of the tree. The enumerator uses an imperative data structure to represent
a stage in this traversal, where a specific element (entry) is reached. Each call of the
enumerator function steps this data structure forward to the next element (entry) by a small
modification with constant computation time. The computation time for making a complete
recursion over a search-tree based set (or map) using the enumerator function is hence
proportional to the number of elements in the set (entries in the map), and the time
performance equals the time performance of pattern-matching for lists.
The enumerator functions for the hash-key based HashSet and Dictionary collections
work in a different way because we have a different data representation using an array as
described in Section 8.11. The traversal scans forward through the array (and through each
linked structure of colliding elements). The time used by a complete recursion over a hashed
collection using the enumerator function is hence proportional to the number of elements
(entries) in the set (map), and this corresponds to the time used by a complete recursion over
a list using pattern-matching.
 8.13 Imperative queue
The System.Collections.Generic library contains an imperative queue implementation in
the form of a Queue<′a> class with members:

The queue is implemented using an array where the front queue element is stored in an array
location with index frontIndex while the rest of the queue is stored in the succeeding locations
with a wraparound to the beginning of the array if the queue extends beyond the end of the
array.

The Dequeue operation returns the array element with index frontIndex and advances
frontIndex to the next array position (with a possible wraparound) while the Enqueue
operation stores the enqueued value in the first free array location. An Enqueue operation
with a filled array causes an array replacement where a new, larger array is reserved and all
queue elements are moved to the new array upon which the old array is abandoned.
A queue can be used to make the following elegant and interesting implementation of the
breadth-first traversal of list trees in Section 6.6.
type ListTree<′a> = Node of ′a * (ListTree<′a> list);;

let breadthFirstIter f ltr =

let remains = Queue<ListTree<′a> > ()
remains.Enqueue ltr
while (remains.Count <> 0) do
let (Node (x,tl)) = remains.Dequeue()
List.iter (remains.Enqueue) tl
f x;;
val breadthFirstIter : (′a -> unit) -> listTree<′a> -> unit

The idea is to let the queue remains contain those list trees where the nodes remain to be
visited, initially the tree ltr. A list tree Node (x,tl) is dequeued, the elements of the list tl
of sub-trees are enqueued one-by-one using List.iter – and the root node x is visited. This
procedure is repeated until the queue becomes empty.
 8.14 Restrictions on polymorphic expressions
The purpose of the restriction on polymorphic expressions in Section 4.5 is to ensure that the
use of mutables is type safe. The problem is illustrated by the following hypothetical example:
let mutable a = [];; // NOT allowed !!
val mutable a = []

let f x = a <- (x :: a);;

val f : ′a -> unit

f(1);;
it : unit = ()
f("ab");;
it : unit = ()
a;;
it : ? list = ["ab"; 1] *** Oops! type error !

The point is that F# would be forced to infer a type of f prior to any use of the function. This
would result in the type ′a -> unit because apparently values of any type can be cons′ed onto
the empty list. Hence each of the applications f(1) and f("ab") would type check because int
as well as string are instances of the polymorphic type ′a. The type check would hence fail
to discover the illegal expression "ab"::[1] emerging during the evaluation of f("ab").
The declaration
let mutable a = [];;

is construed as binding a to the value of the polymorphic expression “mutable []” and this
expression is not considered a value expression. The declaration is hence rejected by the
restriction on the use of polymorphic expressions.
 Summary
The chapter provides a semantical framework, the store, for understanding the imperative
features of F# that operate on and modify the state of mutable objects. A store consists of a set
of locations containing values, together with operators to access and change the store. The
main imperative constructs of F# is introduced together with extracts of .NET libraries for
imperative collections, including arrays, sets and maps. We explain why the restriction on
polymorphically typed expressions is needed because of the imperative features.
 Exercise
8.1 Make a drawing of the environment and store obtained by the following declarations
and assignments:
let mutable x = 1;;
let mutable y = (x,2);;
let z = y;;
x <- 7;;

8.2 The sequence of declarations:

let mutable a = []
let f x = a <- (x :: a)
f(1);;

are accepted by F#. Explain why.

8.3 Make a drawing of the environment and store obtained by the following declarations
and assignments:
type t1 = { mutable a : int };;
type t2 = { mutable b : int ; c : t1 };;
let x = { a = 1 };;
let y = { b = x.a ; c = x };;
x.a <- 3;;

8.4 Declare null to denote the default value of the record type:
type t = { mutable link : t ; data : int };;

Declare some other values of type t and use assignment to build chains and circles of
values of type t. Declare a function to insert an element in the front of a chain of values of
type t.
8.5 Give a declaration of the gcd function using a while loop instead of recursion (cf.
Section 1.8).
8.6 Declare a function for computing Fibonacci numbers Fn (see Exercise 1.5) using a
while loop. Hint: introduce variables to contain the two previously computed Fibonacci
numbers.
8.7 Use a HashSet traversal for loop to declare a function
HashSetFold: (′b -> ′a -> ′b) -> ′b -> HashSet<′a> -> ′b

such that
f b set = f (… (f (f b a0) a1) … ) an−1

where a0,…, an−1 are the elements of the HashSet set.

 8.8 Declare a DictionaryFold function. The type should correspond to the type of
Map.fold.
8.9 Make declarations of breadthFirst and breadthFirstFold for list trees using an
imperative queue.
Hint: unfold the while-loop in the declaration of breadthFirstIter to a local recursive
function and use argument and value of this function to build the result.
___________________
1 The restriction is related to the memory management where problems might arise if a
function was allowed to return a closure using a locally defined mutable.
2 The symbol ! cannot be used as a user-defined prefix operator.
 9
Efficiency

The efficiency of a program is measured in terms of its memory requirements and its running
time. In this chapter we shall introduce the concepts stack and heap because a basic
understanding of these concepts is necessary in order to understand the memory management
of the system, including the garbage collection.
Furthermore, we shall study techniques that in many cases can be used to improve the
efficiency of a given function, where the idea is to search for a more general function, whose
declaration has a certain form called iterative or tail recursive. Two techniques for deriving
tail-recursive functions will be presented: One is based on using accumulating parameters and
the other is based on the concept of a continuation, that represents the rest of the computation.
The continuation-based technique is generally applicable. The technique using accumulating
parameters applies in certain cases only, but when applicable it usually gives the best results.
We give examples showing the usefulness of these programming techniques.
We relate the notion of iterative function to while loops and provide examples showing
that tail-recursive programs are in fact running faster than the corresponding programs using
while loops.
The techniques for deriving tail-recursive functions are useful programming techniques
that often can be used to obtain performance gains. The techniques do not replace a conscious
choice of good algorithms and data structures. For a systematic study of efficient algorithms,
we refer to textbooks on “Algorithms and Data Structures.”
 9.1 Resource measures
The performance of an algorithm given by a function declaration in F# is expressed by
figures for the resources used in the evaluation of a function value:

Use of computer memory: The maximum size of computer memory needed to represent
expressions and bindings during the evaluation.
Computation time: The number of individual computation steps.

The important issue is to estimate how these figures depend on the “size‘” of the argument
for “large” arguments, for example, number of digits of integer argument, length of list
argument, depth (i.e. number of levels) of tree argument, etc. These performance figures are
essentially language independent, so implementations of the same algorithm in another
programming language will show a similar behaviour.
Efficiency in performance is not the only important issue in programming. Correctness and
readability are often more important because the program should be understandable to the
readers (including the programmer herself). The choice of function declaration should
therefore be based on a trade-off between performance and readability (that is, simplicity),
using the simplest declaration for any particular function in a program – unless, there is a risk
that it becomes a performance bottleneck for the overall program.
 9.2 Memory management
The memory used by an F# program is spilt into a stack and a heap, where primitive values,
such as numbers and truth values are allocated on the stack, while composite values such as
lists and trees, closures and (most) objects are allocated on the heap. A basic understanding of
the stack and the heap is necessary to understand the memory resources required by a
program.
Consider the following declaration at the outermost level:
let xs = [5;6;7];;
val xs : int list = [5; 6; 7]

let ys = 3::4::xs;;
val ys : int list = [3; 4; 5; 6; 7]

let zs = xs @ ys;;
val zs : int list = [5; 6; 7; 3; 4; 5; 6; 7]

let n = 27;;
val n : int = 27

The stack and the heap corresponding to these declarations are shown in Figure 9.1.

Fig ure 9.1 Memory: Stack and Heap for top-level declarations

The stack contains an entry for each binding. The entry for the integer n contains the
integer value 27, while the entries for the lists xs, ys and zs contain links (i.e. memory
pointers) pointing at the implementations of these lists. A list [x0; …; xn-1] is implemented by
a linked data structure, where each list element xi is implemented by a cons cell containing the
value xi and a link to the cons cell implementing the next element in the list:

The entry for xs in the stack contains a link to the cons cell for its first element 5 in the
heap.
The entry for ys in the stack contains a link to the cons cell for its first element 3. This
cons cell contains a link to the cons cell for the next element 4 and that cons cell contains
in turn a link to the first cons cell of xs.
 The entry for zs in the stack contains a link to the first cons cell of a copy of the linked
list for xs (the first argument of @ in xs @ ys). The last cons cell of that copied linked
list contains a link to the start of the linked list for ys.

Since a list is a functional (immutable) data structure, we have that:

1. The linked lists for ys is not copied when building a linked list for y::ys.
2. Fresh cons cells are made for the elements of xs when building a linked list for xs @ ys,
as the last cons cell in the new linked list for xs must refer to the first cons cell of the
linked list for ys. The running time of @ is, therefore, linear in the length of its first
argument. This running time is in agreement with the declaration of append in Section
4.4 and with the linked-list based implementation used by the built-in append function.

These two properties will be exploited later in this section.

 Basic operations on Stack and Heap
The consecutive piece of stack memory corresponding to bindings at the same level is called
a stack frame. During the evaluation of an expression a new stack frame is added whenever
new bindings arise, for example, due to local declarations and expressions or because a
function is called. This is illustrated using the following declarations:
let zs = let xs = [1;2]
let ys = [3;4]
xs@ys;;

The evaluation of the outermost declaration will start with an empty heap and a stack frame
sf0 containing a (so far undefined) entry for zs:

Pushing a stack frame

The start of the evaluation of the local declarations will push an new stack frame on top of sf0.
This stack frame has entries for the locally declared variables xs and ys and some extra
entries including one for the result of the local expression xs @ ys:
Notice that a copy of the list xs is made in the heap during the evaluation of xs @ ys.

Popping a stack frame

When the result of the local expression xs @ ys has been computed, the stack frame sf1 is
popped, that is, removed from the stack, and the reference to the first cons cell of xs @ ys is
copied to the stack entry for zs:

The resulting heap after the evaluation of the declaration for zs contains two cons cells
marked with ’†’. These cells are obsolete because they cannot be reached from any binding,
and they are therefore later removed from the heap by the garbage collector that manages the
heap behind the scene.
The management of the stack follows the evaluation of declarations and function calls in a
simple manner, and the used part of the stack is always a consecutive sequence of the relevant
stack frames. We illustrate this by a simple example. Consider the following declarations:
let rec f n =
match n with
| 0 -> 0
| n -> f(n-1) + n;;

let x = f 3;;

The first part of the evaluation of f 3 makes repeated bindings of n corresponding to the
recursive function calls:

These bindings are implemented by four stack frames sf1,…, sf4 pushed on top of the initial
stack frame sf0 corresponding to f and x. Each of the stack frames sf1, …, sf4 corresponds to
an uncompleted evaluation of a function call:

The next evaluation step marks the completion of the “innermost” functions call f 0

and the binding n 0 is hence no longer needed. The implementation releases the memory
used to implement this binding by popping the frame sf4 off the stack:
When the evaluation terminates the stack frames sf3, sf2 and sf1 are all popped and the initial
stack frame sf0 contains a binding of x to 6.
The stack management using the push and pop operations is very simple because the stack is
maintained as a contiguous sequence of the relevant stack frames. The stack memory will
hence never be fragmented.

Garbage and garbage collection

We shall now study garbage collection closer using the declarations:
let xs = [1;2];;

let rec g = function

| 0 -> xs
| n -> let ys = n::g(n-1)
List.rev ys;;
val g : int -> int list

g 2;;
val it : int list = [1; 1; 2; 2]

Application of this function will produce garbage due to the local declaration of a list and due
to the use of List.rev. The stack and the heap upon the termination of g 2 is shown in Figure
9.2. The stack contains just one stack frame corresponding to the top-level declarations.
 Fig ure 9.2 Memory: Stack and Heap upon termination of evaluation of g 2

The heap contains five cons cells marked with ’†’, that are obsolete because they cannot be
reached from any binding, and they are removed from the heap by the garbage collector. It is
left for Exercise 9.1 to produced this stack and heap for an evaluation of g 2. The amount of
garbage produced using g grows with the size of the argument, and it is easy to measure how
much garbage the system has to collect.
 Measuring running time and garbage collection
The directive #time, that works as a toggle, can be used in the interactive F# environment to
extract information about running time and garbage collection of an operation:
#time;;

–> Timing now on

g 10000;;
Real: 00:00:01.315, CPU: 00:00:01.326,
GC gen0: 356, gen1: 24, gen2: 0
val it : int list = [9999; 9997; 9995; 9993; 9991; 9989;
9987;...]

The measurement includes two times: The Real time is the clock time elapsed during the
execution of the operation, in this case 1.315 second. The CPU time is the total time spent by
the operation on all CPUs (or cores) on your computer. If you are not exploiting the
parallelism of multiple cores, then these two times should approximately be the same.
The garbage collector manages the heap as partitioned into three groups or generations:
gen0, gen1 and gen2, according to their age. The objects in gen0 are the youngest while the
objects in gen2 are the oldest. The typical situation is that objects die young, that is, garbage
typically occurs among young objects, and the garbage collector is designed for that
situation. During the above evaluation of g 10000, the garbage collector reclaimed (collected)
356 objects among the youngest ones from group gen0 and 24 objects from gen1.
 The limits of the stack and the heap
The stack and heap sizes are resources that we must be aware of. The following examples
illustrate maximal stack and heap sizes and shows that the maximal heap size is order of
magnitudes larger than the maximal stack size.
Consider first the following function that can generate a list:
let rec bigList n = if n=0 then [] else 1::bigList(n-1);;
val bigList : int -> int list

bigList 120000;;
val it : int list = [1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1;...]

bigList 130000;;
Process is terminated due to StackOverflowException.

A call bigList n will generate n consecutive stack frames each with a binding of n and the
examples show that 120000 such stack frames are manageable while 130000 are not.
Another declaration of a function that can generate the same lists as the above one is given
below. This function can generate lists that are about 100 times longer than those generated
above, and when memory problems arise it is because the heap is exhausted:
let rec bigListA n xs = if n=0 then xs
else bigListA (n-1) (1::xs);;
val bigListA : int -> int list -> int list

let xsVeryBig = bigListA 12000000 [];;

val xsVeryBig : int list = [1; 1; 1; 1; 1; 1; 1; 1; 1;...]

let xsTooBig = bigListA 13000000 [];;

System.OutOfMemoryException:
Exception of type ′System.OutOfMemoryException′ was thrown.
at FSI_0002.bigListA(Int32 n, FSharpList‘1 xs)
at <StartupCode$FSI_0014>. $FSI_0014.main@()
Stopped due to error

In the next sections we study techniques that can be used to minimize the memory usage.
 9.3 Two problems
In this section we reconsider the declarations of the factorial function fact (Page 6) and the
reverse function for lists naiveRev (Page 80). We shall see that evaluation of a function value
for fact uses more computer memory than necessary, and that the evaluation of a function
value for naiveRev requires more evaluation steps than necessary. More efficient
implementations for these functions are given in the next section.
 The factorial function
The factorial function has previously been declared by:
let rec fact = function
| 0 -> 1
| n -> n * fact(n-1);;
val fact : int -> int

We have seen that the evaluation of the expression fact (N) proceeds through a number of
evaluation steps building an expression with a size proportional to the argument N upon
which the expression is evaluated:

The maximal size of the memory needed during this evaluation is proportional to N,
because the F# system must remember (in the heap) all N factors of the expression: N * ((N−1)
* ((N−2) * ( ··· (4 * (3 * (2 * 1))) ··· ))) during the evaluation. Furthermore, during the
evaluation the stack will grow until it has N+1 stack frame corresponding to the nested calls
of fact.
 The reverse function
The naive declaration for the reverse function is as follows:
let rec naiveRev = function
| [] -> []
| x::xs -> naiveRev xs @ [x];;
val naiveRev : ′a list -> ′a list

A part of the evaluation of the expression naiveRev [x1, x2, …, xn] is:

There are n+1 evaluation steps above and heap space of size proportional to n is required
by the F# system to represent the last expression. These figures are to be expected for
reversing a list of size n.
However, the further evaluation

requires a number of evaluation steps that is proportional to n2.

To see this, observe first that m+1 evaluation steps are needed to evaluate the expression
[y1 , …, ym] @ zs as y1 :: (y2 :: … :: (ym :: zs) … ).

Thus,

Hence, the evaluation of (( ··· ( ( [ ] @ [xn]) @ [xn−1]) @ ··· @ [x2] ) @ [x1] ) requires

steps, which is proportional to n2.

 9.4 Solutions using accumulating parameters
In this section we will show how to obtain much improved implementations of the above
functions by considering more general functions, where the argument has been extended by
an extra component (“ m” and “ys”):

Note, that n! = factA (n,1) and rev [x1, …, xn] = revA ([x1, …, xn],[ ]). So good
implementations for the above functions will provide good implementations for the factorial
and the reverse functions also.
 The factorial function
The function factA is declared by:
let rec factA = function
| (0,m) -> m
| (n,m) -> factA(n-1,n*m);;
val factA : int * int -> int

Consider the following evaluation:

This evaluation of factA(5,1) has the properties we are looking for:

1. It does not build large expressions.

2. The number of steps needed to evaluate factA(n,m) is proportional to n.

The argument pattern m in the above declaration is called an accumulating parameter, since
the result is gradually built in this parameter during the evaluation.
The main part of the above evaluation of factA(5,1) is the gradual evaluation of
arguments in the recursive calls of the function:
(5, 1), (4, 5), (3, 20), (2, 60), (1, 120), (0, 120)

Each of these values is obtained from the previous one by applying the function:
fun (n,m) -> (n-1, n*m)

so the evaluation of the arguments can be viewed as repeated (or iterated) applications of this
function.
The use of factA gives a clear improvement to the use of fact. Consider the following
example measuring the time of 1000000 computations of 16! using these two function:
let xs16 = List.init 1000000 (fun i -> 16);;
val xs16 : int list = [16; 16; 16; 16; 16; 16; 16; 16; ...]

#time;;

for i in xs16 do let _ = fact i in ();;

Real: 00:00:00.051, CPU: 00:00:00.046,
GC gen0: 0, gen1: 0, gen2: 0
val it : unit = ()

for i in xs16 do let _ = factA(i,1) in ();;

Real: 00:00:00.024, CPU: 00:00:00.031,
GC gen0: 0, gen1: 0, gen2: 0
val it : unit = ()

The performance gain of using factA is actually much better than the factor 2 indicated by the
above examples becomes the run time of the for construct alone is about 12 ms:
for i in xs16 do let _ = () in ();;
Real: 00:00:00.012, CPU: 00:00:00.015,
GC gen0: 0, gen1: 0, gen2: 0
val it : unit = ()
 The reverse function
The function revA is declared by:
let rec revA = function
| ([], ys) -> ys
| (x::xs, ys) -> revA(xs, x::ys);;

Consider the following evaluation (where the bindings are omitted):

This evaluation of revA([1, 2, 3],[]) again has the properties we are looking for:

1. It does not build large expressions.

2. The number of steps needed to evaluate revA(xs, ys) is proportional to the length of xs.

It makes a big difference for lists with large length n whether the number of evaluation steps
is proportional to n or to n2.
The argument pattern ys in the above declaration is the accumulating parameter in this
example since the result list is gradually built in this parameter during the evaluation.
Note, that each argument in the recursive calls of revA is obtained from the argument in the
previous call by applying the function:
fun (x::xs, ys) -> (xs, x::ys)

The use of revA gives a dramatically improvement to the use of naiveRev. Consider the
following example measuring the time used for reversing the list of elements from 1 to
20000:
let xs20000 = [1 .. 20000];;

naiveRev xs20000;;
Real: 00:00:07.624, CPU: 00:00:07.597,
GC gen0: 825, gen1: 253, gen2: 0
val it : int list = [20000; 19999; 19998; 19997; 19996;...]
revA(xs20000,[]);;
Real: 00:00:00.001, CPU: 00:00:00.000,
GC gen0: 0, gen1: 0, gen2: 0
val it : int list = [20000; 19999; 19998; 19997; 19996; ...]

The naive version takes 7.624 seconds while the iterative version takes just 1 ms. One way to
consider the transition from the naive version to the iterative version is that the use of append
(@) has been reduced to a use of cons (::) and this has a dramatic effect of the garbage
collection. No object is reclaimed by the garbage collector when revA is used, whereas
825+253 obsolete objects were reclaimed using the naive version and this extra memory
management takes time.
Returning to the list-generating functions on Page 204, the function bigListA is a more
general function than bigList, where the argument xs is the accumulating parameter.
 9.5 Iterative function declarations
The above declarations for factA, revA and bigListA have a certain form that we will study
in this section.
A declaration of a function g: τ -> τ′ is said to be an iteration of a function f: τ −> τ if it is an
instance of the schema:
let rec g z = if p z then g (f z) else h z;;

for suitable predicate p : τ −> bool and function h: τ -> τ′.

A function declaration following the above schema is called an iterative declaration. It is tail-
recursive in the sense that every recursive call of the function is a tail call, that is, the last
operation that is evaluated in the body of the declaration. For convenience we only study tail-
recursive declarations of the above form in this subsection.
 The function factA
The function factA is an iterative function because it can be declared as:
let rec factA(n,m) = if n<>0 then factA(n-1,n*m) else m;;

which is an instance of the above schema with:

let f(n,m) = (n-1, n*m)

let p(n,m) = n<>0

let h(n,m) = m;;

 The function revA
The function revA is also an iterative function:
let rec revA(xs,ys) =
if (not (List.isEmpty xs))
then revA(List.tail xs, (List.head xs)::ys)
else ys;;

which is an instance of the above schema with:

let f(xs,ys) = (List.tail xs, (List.head xs)::ys)

let p(xs,ys) = not (List.isEmpty xs)

let h(xs,ys) = ys

When a declaration of a function in an obvious way can be transformed into the above
form, we will call it an iterative function without further argument.
 The fold function on lists
The fold function on lists as declared in Section 5.1:
let rec fold f e = function
| x::xs -> fold f (f e x) xs
| [] -> e ;;

is an iterative function. The declaration can be written as:

let rec fold f e xs =
if not (List.isEmpty xs)
then fold f (f e (List.head xs)) (List.tail xs)
else e;;

which is an instance of the above schema. The above function revA is actually an application
of this iterative function:
let revA(xs,ys) = fold (fun e x -> x::e) ys xs;;
 Evaluation of iterative functions
The evaluation for an arbitrary iterative function:
let rec g z = if p z then g(f z) else h z;;

proceeds in the same manner as the evaluations of factA and revA:

We define the n’th iteration f n x, for n ≥ 0, of a function f : τ -> τ as follows:

Thus,

Suppose that

Then, the evaluation of the expression g x proceeds as follows:

This evaluation has three desirable properties:

1. It does not build large expressions, as the argument f z of g (f z) is evaluated at each step
due to the eager evaluation strategy of F#,
2. there are n recursive calls of g, and
3. there is only one environment used at each stage of this evaluation.

The first property implies that heap allocation of long expressions with pending operations
can be avoided, the second property implies a linear unfolding of the recursive function g,
and the last property implies that just one stack frame is needed during an evaluation of g x
(ignoring stack frames needed due to calls of other functions).
Since bigListA is a tail-recursive function, the stack will not grow during the evaluation of
bigListA n xs and the heap is hence the limiting memory resource when using this function
as we learned in connection with the examples on Page 204.
 Iterations as loops
We observed in Section 8.7 that every while loop can be expressed as an iteration. It is also the
case that every iterative function g:
let rec g z = if p z then g (f z) else h z;;

can be expressed as a while loop:

let rec g z =
let zi = ref z
while p !zi do zi := f !zi
h(!zi);;

Using this translation scheme for the iterative version factA of the factorial function, we
arrive at the declaration:
let factW n =
let ni = ref n
let r = ref 1
while !ni>0 do
r := !r * !ni ; ni := !ni-1
!r;;

where it is taken into account that the argument z in the translation scheme in this case is a
pair (n,r).
There is no efficiency gain in transforming an iteration to a while-loop. Consider for
example 1000000 computations of 16! using factA(16,1) and factW 16:
#time;;

for i in 1 .. 1000000 do let _ = factA(16,1) in ();;

Real: 00:00:00.024, CPU: 00:00:00.031,
GC gen0: 0, gen1: 0, gen2: 0
val it : unit = ()

for i in 1 .. 1000000 do let _ = factW 16 in ();;

Real: 00:00:00.048, CPU: 00:00:00.046,
GC gen0: 9, gen1: 0, gen2: 0
val it : unit = ()

which shows that the tail-recursive function actually is faster than the imperative while-loop
based version.
 9.6 Tail recursion obtained using continuations
A tail-recursive version of a function can in some cases be obtained by introducing an
accumulating parameter as we have seen in the above examples, but this technique is
insufficient in the general case. However, there is a general technique that can transform an
arbitrary declaration of a recursive function f : τ1 -> τ2 into a tail-recursive one. The
technique adds an extra argument c that is a function. At present we assume that each branch in
the recursive declaration of f contains at most one recursive call of f. The tail recursive
version fC of f is then of type τ1->(τ2->τ2)->τ2 with parameters v and c of types v: τ1 and c: τ2-
>τ2 .

The evaluation of a function value f(v) comprises recursive calls of f with arguments v0, v1,
…, vn where v0 = v and where vn corresponds to a base case in the declaration of f. The
corresponding evaluation of fC:

contains functions c0, c1, …, cn with the crucial property:

This property expresses that the function ck contains the rest of the computation once you
have computed f(vk). It is therefore called a continuation. The evaluation of fC starts with c0 =
id where id is the pre-defined identity function satisfying id a = a for any a. The effects of
the recursive calls of f are gradually accumulated in the continuations ck during the evaluation
of fC v id, and the evaluation ends by applying the continuation cn to the value f(vn) in a base
case.
The notion of a continuation has a much wider scope than achieving tail-recursive
functions (the focus in this chapter) and we refer to [12] for an in-depth study of this concept.
Consider, for example, the simple declaration of bigList from Section 9.2:
let rec bigList n = if n=0 then [] else 1::bigList(n-1);;
val bigList : int -> int list

that was used to illustrate the stack limit problems due to the fact that it is not a tail-recursive
function. The continuation-based version bigListC n c has a extra argument
c: int list -> int list

that is a continuation. The declaration of bigListC is:

let rec bigListC n c =
if n=0 then c []
else bigListC (n-1) (fun res -> c(1::res));;
val bigListC : int -> (int list -> ′a) -> ′a
The base case of bigListC is obtained from the the base case of bigList by feeding that
result into the continuation c. For the recursive case, let res denote the value of the recursive
call of bigList(n-1). The rest of the computation of bigList n is then 1::res. Hence, the
continuation of bigListC(n-1) is
fun res -> c(1::res)

because c is the continuation of bigListC n.

The function is called using the pre-defined identity function id as continuation:
bigListC 3 id;;
val it : int list = [1; 1; 1]

The important observations are:

bigListC is a tail-recursive function, and

the calls of c are tail calls in the base case of bigListC as well as in the continuation: fun
res -> c(1::res).

The stack will hence neither grow due to the evaluation of recursive calls of bigListC nor
due to calls of the continuations that have been built in the heap.
Consider the examples:
bigListA 12000000 [];;
Real: 00:00:01.142, CPU: 00:00:01.138,
GC gen0: 34, gen1: 22, gen2: 0
val it : int list = [1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1;
...]

bigListC 12000000 id;;

Real: 00:00:05.742, CPU: 00:00:05.538,
GC gen0: 60, gen1: 48, gen2: 3
val it : int list = [1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1;
...]

bigListC 16000000 id;;

Real: 00:00:08.586, CPU: 00:00:08.314,
GC gen0: 80, gen1: 60, gen2: 3
val it : int list = [1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1; 1;
...]

bigListC 17000000 id;;

System.OutOfMemoryException: Exception of type
′System.OutOfMemoryException′ was thrown. Stopped due to error

These examples show:

1. The version using an accumulating parameter is much faster (about five times) than that
using continuations.
2. The version using continuations can handle about 30% longer lists.
The run-time disadvantage of a continuation-based declaration is even more clear with an
iterative function like factA where no data structure is required to be built in the heap, when
compared to bigListC. See Exercise 9.6.

More general recursions

We shall now study the use of continuations in a more general setting with several recursive
calls of the function. This situation occurs, for example, in connection with binary trees.
Consider a function count, that counts the number of nodes in binary tree on the basis of
the type BinTree<′a> defined in Section 6.4:
type BinTree<′a> = | Leaf
| Node of BinTree<′a> * ′a * BinTree<′a>;;

let rec count = function

| Leaf -> 0
| Node(tl,n,tr) -> count tl + count tr + 1;;

A counting function: countA: int -> BinTree<′a> -> int using an accumulating
parameter will not be tail-recursive due to the expression containing recursive calls on the left
as well as the right sub-trees of a node (try, for example, Exercise 9.8). A tail-recursive
version can, however, be developed for a continuation-based version:
countC: BinTree<′a> -> (int -> ′b) -> ′b

The base case countC Leaf c returns c 0. The continuation of countC tl in the case: countC
(Node(n,tl,tr)) c is the function that takes the result vl for the left subtree and calls
countC tr. The continuation of countC tr must take the result vr for the right subtree and
feed vl+vr+1 into the continuation c:
let rec countC t c =
match t with
| Leaf -> c 0
| Node(tl,n,tr) ->
countC tl (fun vl -> countC tr (fun vr -> c(vl+vr+1)));;
val countC : BinTree<′a> -> (int -> ′b) -> ′b

countC (Node(Node(Leaf,1,Leaf),2,Node(Leaf,3,Leaf))) id;;

val it : int = 3

Note that both calls of countC are tail calls, and so are the calls of the continuation c, and the
stack will therefore not grow due to the evaluation of countC and the associated continuations.
The comparison of count and countC shows similar figures as the comparison of bigList
and bigListC: The continuation-based version can handle much larger trees since the stack
space will not be exhausted (try Exercise 9.11); but it is about 4 times slower than count when
counting a balanced tree with 20000000 nodes:
let rec genTree xs =
match xs with
| [| |] -> Leaf
 | [| x |] -> Node(Leaf,x,Leaf)
| _ -> let m = xs.Length / 2
let xsl = xs.[0..m-1]
let xm = xs.[m]
let xsr = xs.[m+1 ..]
Node(genTree xsl, xm, genTree xsr);;
val genTree : ′a [] -> BinTree<′a>

let t n = genTree [| 1..n |];;

let t20000000 = t 20000000;;

count t20000000;;
Real: 00:00:00.453, CPU: 00:00:00.889,
GC gen0: 0, gen1: 0, gen2: 0
val it : int = 20000000

countC t20000000 id;;

Real: 00:00:01.733, CPU: 00:00:01.716,
GC gen0: 305, gen1: 1, gen2: 0
val it : int = 20000000

It is possible to replace one of the continuations in the recursive case of the declaration of
countC by a simple accumulator and arrive at a tail-recursive function with the type

countAC : BinTree<′a> -> int -> (int -> ′b) -> ′b

such that count t = countAC t 0 id. The declaration and analysis of this function is left for
Exercise 9.9.
 Summary
We have introduced the concepts stack and heap that are needed in order to get a basic
understanding of the memory management in the system.
Furthermore, we have introduced the concept of tail-recursive functions and two techniques
for deriving a tail-recursive version of a given function, where one is based on accumulating
parameters and the other on the notion of a continuation. The stack will not grow during the
evaluation of tail-recursive functions (ignoring the calls of other recursive functions), and
using these techniques will in many typical cases give good performance gains.
A transformation from tail-recursive functions to loops was shown, together with
experiments showing that the tail-recursive functions run faster than the corresponding
imperative while-loop based versions.
 Exercise
9.1 Consider the function g declared on Page 202 and the stack and heap after the
evaluation of g 2 shown in Figure 9.2. Reproduce this resulting stack and heap by a
systematic application of push and pop operations on the stack, and heap allocations that
follow the step by step evaluation of g 2.
9.2 Show that the gcd function on Page 16 is iterative.
9.3 Declare an iterative solution to exercise 1.6.
9.4 Give iterative declarations of the list function List.length.
9.5 Express the function List.fold in terms of an iterative function itfold iterating a
function of type ′a list * ′b -> ′a list * ′b.
9.6 Declare a continuation-based version of the factorial function and compare the run
time with the results in Section 9.4.
9.7 Develop the following three versions of functions computing Fibonacci numbers Fn
(see Exercise 1.5):

1. A version fibA: int -> int -> int -> int with two accumulating parameters n1
and n2, where fibA n n1 n2 = Fn, when n1 = Fn−1 and n2 = Fn−2. Hint: consider suitable
definitions of F−1 and F−2.
2. A continuation-based version fibC: int -> (int -> int) -> int that is based on
the definition of Fn given in Exercise 1.5.

Compare these two functions using the directive #time, and compare this with the while-
loop based solution of Exercise 8.6.
9.8 Develop a version of the counting function for binary trees
countA: int -> BinTree<′a> -> int

that makes use of an accumulating parameter. Observe that this function is not tail
recursive.
9.9 Declare a tail-recursive functions with the type
countAC : BinTree<′a> -> int -> (int -> ′b) -> ′b

such that count t = countAC t 0 id. The intuition with countAC t a c is that a is the number
of nodes being counted so far and c is the continuation.
9.10 Consider the following list-generating function:
let rec bigListK n k =
if n=0 then k []
else bigListK (n-1) (fun res -> 1::k(res));;
The call bigListK 130000 id causes a stack overflow. Analyze this problem.
9.11 Declare tail-recursive functions leftTree and rightTree. By use of leftTree it
should be possible to generate a big unbalanced tree to the left containing n+1 values in
the nodes so that n is the value in the root, n−1 is the value in the root of the left subtree,
and so on. All subtree to the right are leaves. Similarly, using rightTree it should be
possible to generate a big unbalanced tree to the right.

1. Use these functions to show the stack limit when using count and countA from
Exercise 9.8.
2. Use these functions to test the performance of countC and countAC from Exercise
9.9.

9.12 Develop a continuation-based version of the function preOrder from Section 6.4, and
compare the performance of the two functions.
9.13 Compare the run times of the two versions of the function tryFind that are declared
on Page 109 and on Page 191.
9.14 Comparison of the efficiency of iteration functions for list and sets.
In this exercise you should declare functions
iterCollM: (′a -> unit) -> Coll<′a> -> unit

so that iterCollM f col performs f v0; f v1; …; f vn when col has v0, v1, …, vn as the
elements, and M is the method of traversal that can be based on a tail-recursive function or
using an enumerator.

1. Declare a tail-recursive function to iterate a function over the elements of a list.

2. Declare a enumerator-based version. See Page 192.
3. Declare a tail-recursive version that iterate over the elements of a set on the basis of
the recursion scheme that repeatedly removes the minimal elements from the set. (See
e.g. the declaration of tryFind on Page 109.)
4. Compare the run times of the above iteration functions and the library functions
List.iter and Set.iter. Use, for example, sets and lists containing the integers
from 0 to 10000000 and the function ignore.
 10
Text processing programs

Processing text files containing structured data is a common problem in programming –

you may just think of analysing any kind of textual data generated by electronic equipment or
retrieved data from the web.
In this chapter we show how such programs can be made in a systematic and elegant way
using F# and the .NET library. Data are extracted from text files using functions from the
RegularExpressions library. The data processing of the extracted data is done with a
systematic use of F# collections types list<′a>, Map<′a,′b> and Set<′a>. Easy access from
F# programs to the extensive text processing features of the .NET library is given in a special
TextProcessing library that can be copied from the home page of the book. The chapter
centers on a real-world example illustrating the techniques.
Time performance of programs is always a problem, even with todays very fast computers.
Poor performance of text processing programs is often caused by operations on very long
strings. The method in this chapter uses three strategies to avoid using very long strings:

1. Text input is in most cases read and processed in small pieces (one or a few lines).
2. Text is generated and written in small pieces.
3. Large amounts of internal program data are stored in many small pieces in F#
collections like list, set or map.

The main focus is on methods for handling textual data both as input and output, but we
also illustrate other topics: how to save binary data on the disk to be restored later by another
program, and how to read and analyse source files of web-pages. The techniques are
illustrated using an example: the generation of a web-page containing a keyword index of the
F# and .NET library documentation.
 10.1 Keyword index example: Problem statement
Our running example is the generation of a keyword index for the F# and .NET library
documentation. The result of this programming effort should be a web-page containing an
alphabetically sorted list of keywords, such as the one shown in Figure 10.1. Whenever a user
viewing this web-page makes a double click on a keyword, for example, observer, a
corresponding web-page in the library documentation should automatically be selected by the
internet browser.

Fig ure 10.1 A browser’s view of index.html

...
"Control.Observable Module (F#)" observer event~observer
"Control.WebExtensions Module (F#)" async~web~operation
"Microsoft.FSharp.Core Namespace (F#)"
"Core.ExtraTopLevelOperators Module (F#)" top~level~operators
"Core.LanguagePrimitives Module (F#)" language~primitives
"Core.NumericLiterals Module (F#)" numeric~literal
...

Table 10.1 An extract from keywords.txt

The source data to generate the keyword index are found in the keyword.txt file that is
edited manually by the programmer generating the index (cf. Table 10.1). Each line in this file
contains the title of a library documentation web-page together with the keywords that should
refer to this particular web-page. Space characters inside keywords are written using a tilde
character such that spaces can be used to separate keywords. The line:
"Control.Observable Module (F#)" observer event observer
contains the keywords:
observer and event observer

(the second containing a space character) with links to the library documentation web-page
with title:
Control.Observable Module (F#)

The programs generating the keyword index from these (and other) data are described in
Section 10.8.
 10.2 Capturing data using regular expressions
A basic problem in processing textual data is to capture the relevant information. In the
keyword program we may, for example, input a text line containing a title and two keywords:
"Control.Observable Module (F#)" observer event~observer

and we want to capture the value

("Control.Observable Module (F#)",
["observer"; "event observer"])

of type string * (string list) containing the title and the list of keywords.
This section presents a systematic technique of constructing functions performing such
captures. Using the technique involves three steps

1. An (informal) understanding of the syntactical structure of the input.

2. Formalizing this understanding using regular expressions.
3. Constructing the function capturing the data.

The difficult part is describing the syntactical structure in terms of regular expressions.
 Informal syntactical structure of strings
The syntactic structure of an input line in the keyword index example is illustrated in the
following picture using the above line as an example:

We want to capture the parts labelled (3) and (ii) in the figure.
This syntactical structure of an input line can be described by stating that the line should
consists of the following parts:

1. Zero or more blank characters.

2. A quote character.
3. One or more non-quote characters (title to be captured).
4. A quote character.
5. Zero or more occurrences of a sequence of.
i. One or more blank characters.
ii. One or more non-blank characters (keyword to be captured).
6. Zero or more blank characters.

Table 10.2 Selected parts of regular expression syntax

Construct Legend
Matched by the character char. Character char must be different from
char
.$ ^ { [ ( | ) ] } * + ?
\specialChar Matched by specialChar in above list (e.g. matches)
\ddd Matched by character with octal value ddd
\S Matched by any non-blank character
\s Matched by any blank character
\w Matched by any letter or digit
\d
Matched by any decimal digit
[charSet]
Matched by any character in charSet
[^charSet] Matched by any character not in charSet

Matched by the concatenation of a string matching regExpr1 and a

regExpr1 string matching regExpr2
regExpr2 Matched by the concatenation of zero or more strings each matching
regExpr * regExpr
regExpr + Matched by the concatenation of one or more strings each matching
regExpr ? regExpr
regExpr1 | Matched by the empty string or a string matching regExpr
regExpr2 Matched by a string matching regExpr1 or regExpr2
(?: regExpr)
( regExpr ) Weird notation for usual bracketing of an expression
Capturing group
The matching must start at the beginning of the string or the specified
\G
sub-string (\G is not matched to any character)
$ The matching must terminate at end of string ($ is not matched to any
character)
charSet = Sequence of chars, char matches and char ranges: char1-char2
The documentation of the System.Text.RegularExpressions library contains a
link to a regular expression manual.
The F# Power Pack uses another syntax for regular expressions.
 Regular expressions
Regular expressions formalize the above informal ideas. A regular expression works as a
pattern for strings. Some strings will match a regular expression, others will not. We will pay
special attention to two kinds of elements in the above informal description:

1. Classes of characters like “a quote character,” “a non-blank character.”

2. Constructs like “sequence of,” “one or more,” “zero or more.”

They are formalized in the regular expression notation as:

1. Single character expressions matched by single characters.

2. Operators for building composite expressions.

Selected parts of the regular expression notation in the .NET library is described in Table
10.2. The upper part of this table contains single character expressions:

The regular expression \S is matched, for example, by the character P,

the regular expression \d is matched, for example by the character 5, and
the regular expression \042 is just matched by the character ".

The single character expressions […] and [^ …] are matched by any single character in a set
of characters:

The expression [ab ] is matched by any single character among a, b or space, and
the expression [^cd] is matched by any single character except c and d.

Brackets are used in any algebraic notation whenever an operator is applied to a composite
expression like in the expression (a + b)2. In regular expressions we need a further kind of
brackets to mark the parts corresponding to data to be captured. There are hence two kinds of
brackets in regular expressions:

1. Usual brackets to enclose a sub-expression when used as operand of an operator.

2. Capturing brackets enclosing sub-expressions describing data to be captured.

The designers of the notation have for some mysterious reason decided to use the normal
parentheses (…) as capturing brackets while the strange notation (?:…) is used to denote
usual brackets. You just have to accept that the weird symbol (?: is the way of writing a usual
left bracket in this notation.
Using the notation in Table 10.2 we get the wanted formalization of our description of the
syntactical structure of lines in the keyword file in form of the regular expression:
\G\s*\042([ˆ\042]+)\042(?:\s+([ˆ\s]+))*\s*$

The details in this regular expression can be explained using a picture similar to the previous
picture explaining the structure of the string:

The first and last symbols \G and $ labelled (a) are anchors used to restrain the matching to
all of a string. They are not matched to any characters. The other parts work as follows:
Expression Matched by
1 \s* Zero or more blank characters
2 \042 A quote character
3 ([^\042]+) Capturing group of one or more non-quote characters
4 \042 A quote character
5 (?:…)* Zero or more occurrences of:
i \s+ One or more blank characters
ii ([^s]+) Capturing group of one or more non-blank chars
6 \s* Zero or more blank characters

Table 10.3 The Regex class with Match and Replace functions

Name Type Legend

string ->
Regex
Regex Creates regex object for regular expression
string -> Searches a match to regular expression in
regex.Match Match string
match.Length int Length of matched string
match.Success bool Result of matching
string * regex.Replace(string1,string2) replaces each
regex.Replace string ->
string match of regex in string1 by string2
 String matching and data capturing
The System.Text.RegularExpressions library contains the types and functions for the .NET
regular expressions. A regular expression as the above string is imbedded in a Regex object
using the Regex constructor (cf. Table 10.3):
open System.Text.RegularExpressions;;
let reg =
Regex @"\G\s*\042([^\042]+)\042(?:\s+([^\s]+))*\s*$";;

We use verbatim string constants @"…" (cf. Section 2.3) to suppress any conversion of
escape sequences because we want to keep all backslash characters unchanged in the string.
The expression:
regex.Match string
evaluates to an object match expressing the outcome of matching the string with the regular
expression. A true value of match.Success signals a successful matching. The matching will
in general try any sub-string of the string but it can be restrained using anchors G and $ as in
the above regular expression where matching only applies to the full string.
The matching of a string to a regular expression determines a mapping from the characters
of the string to the single-character expressions of the regular expression. In our example we
get the mapping shown in the following picture:

The capturing groups in the regular expression are numbered 1, 2, … according to the
order of the opening brackets. In our case we have two capturing groups ([^ \042]+) and
([^ s]+). The first is not in the scope of any operator and will hence capture exactly once
while the second is in the scope of a * operator and may hence capture zero or more times.
The picture shows that the second capturing group will capture twice in this case.
The functions captureSimple and captureList in the TextProcessing library of the
book (cf. Table 10.4) give a convenient way of extracting the captured data from the Match
object. The matching and data capture will then proceed as follows:
Table 10.4 Functions from the TextProcessing library of the book. See also
Appendix B

captureSingle : Match -> int -> string

captureSingle m n returns the first captured string of group n
of the Match m. Raises an exception if no such captured data.
captureList : Match -> int -> string list
captureList m n returns the list of captured strings of group n
of the Match m. Raises an exception if the match was unsuccessful.
 captureCount : Match -> int -> int
captureCount m n returns the number of captures of group n
captureCountList : Match -> int list
captureCountList m = [cnt0 ; cnt1 ;. . . ;cntk ] where cntn is the
number of captures of group n (and cnt0 is some integer).

open TextProcessing;;
let m = reg.Match
"\"Control.Observable Module (F#)\" observer event~observer";;

m.Success;;
val it : bool = true

captureSingle m 1;;
val it : string = "Control.Observable Module (F#)"

captureList m 2;;
val it : string list = ["observer"; "event observer"]
 Conversion of captured data
The data captured as described above are strings. In case of a string of digits (possibly with
decimal point) you usually want a conversion to values of other types like int or float. Such
conversions can be made using the conversion functions named by the type name like int or
float. Another example of conversion is the capture of a textual representation of a date-
time. Such a string may be converted to a DateTime value using the DateTime.Parse function.
Our example contains a captured title in its final form while the captured keywords need
some conversion as any tilde character in the captured string should be replaced by a space
character. This conversion is done using the Replace function (cf. Table 10.3) in the Regex
library:
let tildeReg = Regex @"~";;
let tildeReplace str = tildeReg.Replace(str," ");;
val tildeReplace : string -> string

tildeReplace "event~observer";;
val it : string = "event observer"
 Nested data
The Match object and function are less elegant in case of nested data like:
John 35 2 Sophie 27 Richard 17 89 3

where we want to capture the data in a form using nested lists:

[("John", [35; 2]); ("Sophie", [27]);
("Richard", [17; 89; 3])]

The nested syntactic structure is faithfully described in the regular expression

let regNest =
Regex @"\G(\s*([a-zA-Z]+)(?:\s+(\d+))*)*\s*$";;

with anchors \G and $ enclosing

The data groups captured by Match:

group 1: " John 35 2", " Sophie 27", " Richard 17 89 3"
group 2: "John", "Sophie", "Richard"
group 3: "35", "2", "27", "17", "89", "3"

can, however, not be used directly to get the above nested list structure – the data captured by
group 3 do not reflect the nesting.
A systematic method to capture such data using grammars and parsers is presented in
Section 12.10. At present we show two ad hoc ideas to capture the nested data:

Capture in two steps.

Using successive calls of Match.

Capture in two steps

The two steps are:

1. An “outer” data capture of a list of person data strings. In the example it should capture
the list ["John 35 2"; " Sophie 27"; " Richard 17 89 3"] of strings.
2. An “inner” capture of data from each person data string. It should, for example, capture
the value ("John", [35; 2]) when applied to the string "John 35 2".
The outer data capture uses the regular expression
let regOuter = Regex @"\G(\s*[a-zA-Z]+(?:\s+\d+)*)*\s*$";;

obtained from the above expression regNest by leaving out the two innermost capturing
groups. It captures a list of person data strings:
let m = regOuter.Match
" John 35 2 Sophie 27 Richard 17 89 3 ";;
captureList m 1;;
val it : string list =
[" John 35 2"; " Sophie 27"; " Richard 17 89 3"]

The inner data capture uses the regular expression:

let regPerson1 =
Regex @"\G\s*([a-zA-Z]+)(?:\s+(\d+))*\s*$";;

It captures the person name as a letter string and each integer value as a digit string. The digit
strings need further conversions to the corresponding int values. This is done using the
List.map function to apply the conversion function int to each digit string:

let extractPersonData subStr =

let m = regPerson1.Match subStr
(captureSingle m 1, List.map int (captureList m 2));;
val extractPersonData : string -> string * int list

Combining these ideas we get the following function:

let getData1 str =
let m = regOuter.Match str
match (m.Success) with
| false -> None
| _ ->
Some (List.map extractPersonData (captureList m 1));;
val getData1 : string -> (string * int list) list option
getData1 " John 35 2 Sophie 27 Richard 17 89 3 ";;
val it : (string * int list) list option =
Some [("John", [35; 2]);
("Sophie", [27]); ("Richard", [17; 89; 3])]

Using successive calls of Match

The capture is made using successive matches where each match captures from a sub-string
containing the data of one person, for example:

We use the version of Match with a parameter specifying the start position of the sub-string to
be matched:
regex.Match (string, pos)
combined with the regular expression:
let regPerson2 =
Regex @"\G\s*([a-zA-Z]+)(?:\s+(\d+))*\s*";;

obtained from the regular expression regPerson1 by removing the trailing anchor $.
Each of the person data sub-strings will then match this regular expression when matching
from the start position of the sub-string, for instance when matching from the position ( =11)
of the character “S” in “Sophie”:
let m =
regPerson2.Match
(" John 35 2 Sophie 27 Richard 17 89 3 ", 11);;
captureSingle m 1;;
val it : string = "Sophie"
captureList m 2;;
val it : string list = ["27"]
m.Length ;;
val it : int = 10

The length of the captured sub-string is given by m.Length and the new position:

is the position of the first character “R” in the next person data sub-string.
These are combined in the function personDataList that tries to extract a list of person
data from the string str starting at position pos and terminating at position top:
let rec personDataList str pos top =
if pos >= top then Some []
else let m = regPerson2.Match(str,pos)
match m.Success with
| false -> None
| true -> let data = (captureSingle m 1,
List.map int (captureList m 2))
let newPos = pos + m.Length
match (personDataList str newPos top) with
| None -> None
| Some lst -> Some (data :: lst);;
val personDataList : string -> int -> int
-> (string * int list) list option

The function returns an empty list “Some []” if pos ≥ top. Otherwise, a match with the regular
expression regPerson2 is tried. A negative result “None” is returned if the match is
unsuccessful. Otherwise the data are captured and the new position calculated. The result now
depends on the outcome of a recursive call using the new position: A negative result is
propagated – otherwise a positive result is obtained by “cons’ing” the captured person data
onto the list found in the recursive call.
When applying personDataList to a string we start at position 0 with top position equal to
the length of the string:
let getData2 (s: string) = personDataList s 0 s.Length;;
getData2 " John 35 2 Sophie 27 Richard 17 89 3 ";;
val it : (string * int list) list option =
Some [("John", [35; 2]); ("Sophie", [27]);
("Richard", [17; 89; 3])]
 Several syntactic forms
Use of lines of different syntactic form is frequent: The textual data from a piece of electronic
equipment may for instance have the form:
Heading line
Measurement Line
...
Measurement Line
Heading Line
Measurement Line
...
where each heading line initiates a new series of measurements. In this situation one may use
one regular expression regExpr1 to describe the syntax of a header line and another
expression regExpr2 to describe the syntax of a measurement line. The program analyzing an
input line can then use the results match.Success (cf. Table 10.3) of matching the line against
regExpr1 or regExpr2 to get a division into cases.
 10.3 Text I/O
Reading and writing text files is done using File.OpenText to create a StreamReader or
using File.CreateText to create a StreamWriter as described in Table 10.5. These readers
and writers are instances of the TextReader and TextWriter classes that also comprise the
StringReader and StringWriter classes to read and write strings in memory using the same
I/O functions (an example of using a StringReader is found in Section 10.9. The reader may
consult [9] for further details).
The I/O functions insert and remove new-line characters in a proper way. An input line
obtained by calling ReadLine contains no newline character while each call of WriteLine
adds a new-line at the end of the line. The division into lines inside a text file is hence
determined by the successive calls of WriteLine plus possible new line characters occurring
in the strings written to the file. This structure is retrieved when reading the file line-by-line
using successive calls of ReadLine.
Table 10.5 Selected System.IO library functions

A StreamWriter has an internal data buffer. Part of a string sent to the writer may be
temporarily stored in the buffer for later writing to the output medium. A call of Flush
ensures that the buffer contents is written to the medium.
It is often possible to process an input text on a line-by-line basis. The program will then input
one or more lines, do some computations, input the next lines, etc. This pattern of
computation is captured in the functions of the TextProcessing library of the book described
in Table 10.6. Signature and implementation files are given in Appendix B.
A typical application of fileFold is to build a collection using a function f that captures
data from a single input line (as described in Section 10.2) and adds the data to the collection.
A typical application of fileIter is to generate a side effect for each input line such as output
of some data or updating of an imperative data structure. The applications of fileXfold and
fileXiter are similar, but involve several lines of the file.

Table 10.6 File functions of the TextProcessing library of the book. See also
 Appendix B

fileFold: (′a -> string -> ′a) -> ′a -> string -> ′a
fileFold f e path = f ( …(f (f e lin0 ) lin1 ) … ) linn-1 where
lin0, lin1, … , linn-1 are the lines in the file given by path
fileIter: (string -> unit) -> string -> unit
fileIter g path will apply g successively to each line in the file given by path
fileXfold: (′a -> StreamReader -> ′a) -> ′a -> string -> ′a
fileXfoldf e path creates a StreamReader rdr to read from the file given by
path and makes successive calls f e rdr with accumulating parameter e until end
of the file. Reading from the file is done by f using the rdr parameter.
fileXiter: (StreamReader -> unit) -> string -> unit
fileXiter g e path creates a StreamReader rdr to read from the file given by
path and makes successive calls f rdr until end of the file. Reading from the file
is done by f using the rdr parameter.
 10.4 File handling. Save and restore values in files
The File class of the System.IO library contains a large collection of functions to handle
files. A few of these functions are described in Table 10.7.
The Exists function is useful in a program if you want to ensure that an existing file is not
unintentionally overwritten because of errors in the parameters of the program call. The
Replace function is used in maintaining a file discipline with a current version and a backup
version.
The saveValue function in the TextProcessing library of the book (cf. Table 10.8) can be
used to store a value from a program in a disk file. Another program may then later restore
the value from the disk file using the restoreValue function.
Table 10.7 Some file handling function from the System.IO library

File.Delete: string -> unit

File.Delete path deletes the file given by path (if possible).
File.Exists: string -> bool
File.Exists path= true if a file as specified by path exists and false otherwise.
File.Move: string * string -> unit
File.Move(oldPath,newPath) moves file given by oldPath to newPath (if
possible).
File.Replace: string * string * string -> unit
File.Replace(tempPath,currPath,backupPath) deletes the file given by
backupPath,
renames the file given by currPath to backupPath and renames the file given by
tempPath to currPath.

Table 10.8 Save/restore functions from the books TextProcessing library. See also
Appendix B

saveValue: ′a -> string -> unit

saveValue value path saves the value val in a disk file as specified by path
restoreValue: string -> ′a
restoreValue path restores a value that has previously been saved in the file
given by path. Explicit typing of the restored value is required as the data saved
in the file do not comprise the F# type of the saved value.

The following examples show how to save two values on the disk:
open TextProcessing;;
let v1 = Map.ofList [("a", [1..3]); ("b", [4..10])];;
val v1 : Map<string,int list> =
map [("a", [1; 2; 3]); ("b", [4; 5; 6; 7; 8; 9; 10])]
saveValue v1 "v1.bin";;
val it : unit = ()

let v2 = [(fun x-> x+3); (fun x -> 2*x*x)];;

val v2 : (int -> int) list = [<fun:v2@20>; <fun:v2@20-1>]

saveValue v2 "v2.bin";;
val it : unit = ()

These values are restored as follows:

let value1:Map<string,int list> = restoreValue "v1.bin";;
val value1 : Map<string,int list> =
map [("a", [1; 2; 3]); ("b", [4; 5; 6; 7; 8; 9; 10])]

let [f;g]: (int->int) list = restoreValue "v2.bin";;

val g : (int -> int)
val f : (int -> int)

f 7;;
val it : int = 10

g 2;;
val it : int = 8

Note that arbitrary values, including functions, can be saved on the disk and retrieved again
and that the type annotations are necessary when restoring the values, because the F# system
otherwise would have no information about the types of retrieved values. Furthermore, we
have omitted the warning concerning the incomplete pattern [f;g] in the last example.
 10.5 Reserving, using and disposing resources
A function like File.OpenText reserves other resources beside allocating a piece of memory
for the stream reader object: The disk file is reserved for input to avoid any output to the file
while it is being used for input. This reservation must be released when the program has
ceased reading the file.
The release of such resources reserved by functions in the .NET and F# libraries can be
managed in a uniform way by using a use-binding instead of a let-binding when reserving
the resources, that is:
use reader = File.OpenText path

instead of
let reader = File.OpenText path

The keyword use indicates to the system that the binding of reader comprises resources that
should be released once the program is no longer using the object bound to reader. The
system will in this case release these resources when the binding of the identifier reader
cannot be accessed any longer from the program. One usually places the use declaration
inside a function such that the object is automatically released on return from the function.
This mechanism is implemented in the library functions by letting all objects that own
resources implement the IDisposable interface. This interface contains a Dispose operation
that is called when the object is released. Declaring use-bindings can only be done for such
objects.
 10.6 Culture-dependent information. String orderings
The .NET library System.Globalization offers facilities to handle culture-dependent
information like:

Alphabet with national characters. Ordering of strings, keyboard layout.

Number printing layout (period or comma). Currency and amounts.
Date-time format including names of months and days of the week.

Culture-dependent information is collected in a CultureInfo object. Such an object is created

using the Name of the culture, for example
open System.Globalization;;
let SpanishArgentina = CultureInfo "es-AR";;

The complete collection of the (more than 350) supported cultures is found in the sequence
CultureInfo.GetCultures(CultureTypes.AllCultures)

and you can get a complete (and long) list of Name and DisplayName by calling the printing
function:
let printCultures () =
Seq.iter
(fun (a:CultureInfo) ->
printf "%-12s %s\n" a.Name a.DisplayName)
(CultureInfo.GetCultures(CultureTypes.AllCultures));;

The (mutable) cultureInfo object:

System.Threading.Thread.CurrentThread.CurrentCulture

is used by default in culture-dependent formatting (cf. Section 10.7). The Name field:
System.Threading.Thread.CurrentThread.CurrentCulture.Name

gives the name of the current culture.

 Culture-dependent string orderings
The .NET library comprises culture-dependent string orderings. The orderString type in the
TextProcessing library of the book gives a convenient access to these orderings from an F#
program (cf. Table 10.9).
Applying the function
orderString: string -> string -> orderString

to a culture name:
let f = orderString cultureName

yields a function:
f : string -> orderString

to create orderString values with specified culture,

Table 10.9 The orderString type in the books TextProcessing library. See also
Appendix B

orderString: string -> (string -> orderString)

orderString cultureName
is a function to create orderString values with given culture.
string: orderString -> string
string orderString \ is the character string contained in orderString.
orderCulture: orderString -> string
orderCulture\ orderString \ is the culture name of the used string ordering.

for example:
open System.Globalization;;
open TextProcessing;;

let svString = orderString "sv-SE";;

val svString : (string -> orderString)

let dkString = orderString "da-DK";;

val dkString : (string -> orderString)

let enString = orderString "en-US";;

val enString : (string -> orderString)

The comparison operators compare, <, <=, > and >= are customized on orderString values to
the string ordering determined by the culture. We may, for example, observe that the
alphabetic order of the national letters Ø and å is different in Sweden and Denmark:
svString "Ø" < svString "å";;
val it : bool = false
dkString "Ø" < dkString "å";;
val it : bool = true

Comparing orderString values with different culture raises an exception:

dkString "a" < svString "b";;
... Exception of type ′TextProcessing+StringOrderingMismatch′ ...

The string function gives the string imbedded in an orderString value, while the function
orderCulture gives the culture:

let str = svString "abc";;

string str;;
val it : string = "abc"
orderCulture str;;
val it : string = "sv-SE"

It is easy to define an "en-US" sorting of lists of strings:

let enListSort lst =
List.map string (List.sort (List.map enString lst));;
val enListSort : string list -> string list

This function uses List.map to apply enString to each string in a list of strings. The
resulting list of orderString values is then sorted using List.sort. Finally, the strings are
recovered by applying string to each element in the sorted list using List.map.
The "en-US" ordering has interesting properties: Alphabetic order of characters overrules
upper/lower case. For example:
enListSort ["Ab" ; "ab" ; "AC" ; "ad" ] ;;
val it : string list = ["ab"; "Ab"; "AC"; "ad"]

Special characters and digits precede letters:

enListSort ["a"; "B"; "3"; "7"; "+"; ";"] ;;
val it : string list = [";"; "+"; "3"; "7"; "a"; "B"]

The string ordering corresponds to the order of the entries in a dictionary. This is almost the
lexicographical order (ignoring case) – but not quite, for example:
enListSort ["multicore";"multi-core";"multic";"multi-"];;
val it : string list
= ["multi-"; "multic"; "multicore"; "multi-core"]

The string “multicore” precedes “multi-core” because the minus character in this context is
considered a hyphen in a hyphenated word, while the string “multi-” precedes “multic”
because the minus character in this context is considered a minus sign, and this character
precedes the letter “c”.
Note the convenient use of a sorting function. One may also obtain textual output sorted
according to culture by using values of orderString type as keys in set or map collections:
the fold and iter functions will then traverse the elements of such a collection using the
described ordering. The same applies to the enumerator functions of SortedSet and
SortedDictionary collections (cf. Section 8.12).

The ordering of orderString values is defined using the String.Compare function:

String.Compare(string1,string2, cultureInfo)

The user may consult the documentation of this function in [9] for further information about
the culture-dependent orderings.
 10.7 Conversion to textual form. Date and time
The F# and .NET libraries offer two ways of converting data in a program to textual form:
the sprintf function with related functions and the String.Format function with related
functions. The latter is the most advanced, comprising culture-dependent conversion of date
or time using a CultureInfo object.
 The sprintf function
The sprintf function and its fellow functions are called as follows:
sprintf printfFormatString v0 … vn−1

printf printfFormatString v0 … vn−1

fprintf writer printfFormatString v0 … vn−1
eprintf printfFormatString v0 … vn−1
where printfFormatString is a string constant containing a text intermixed with format
placeholders fph0, fph1, …, fphn−1 specifying the formatting of the corresponding arguments.
The resulting string is obtained from the format string by replacing each format placeholder
by result of formatting the corresponding argument.
Table 10.10 Some possible format types in a format placeholder

Format boolean value as true or false

b
s String
d, i Format any basic integer type value as decimal number possibly with sign
u Format any basic integer type value as unsigned decimal number
e, E
f, F
Format floating point value with mantissa and exponent
O Format floating point value in decimal notation
Format value using conversion function string

Table 10.11 Some possible flags in a format placeholder

O Put zeroes instead of blanks in from of number

- Left-justify result (within specified width)
+ Use + sign on positive numbers

The function sprintf delivers the formatted string as the result, while the other functions
writes this string on some output media:
printf writes on Console.Out.
fprintf writer writes on StreamWriter writer
eprintf writes on Console.Error.
A format placeholder has the general form :
%{flags}{width.}{precision}formatType

where {…} means that this part is optional. Frequently used format types and flags are shown
in Table 10.10 and Table 10.11.
The integers width and precision are used in formatting numbers, where width specifies the
total number of printing positions while precision specifies the number of decimals:
sprintf "%bhood" (1=2);;
val it : string = "falsehood"
sprintf "%-6d" 67;;
val it : string = "67 "
sprintf "%+8e" 653.27;;
val it : string = "+6.532700e+002"
sprintf "a%+7.2fb" 35.62849;;
val it : string = "a +35.63b"

Further information in the documentation of the F# Core Printf Module in [9].

 Date and time
A point of time is uniquely determined by its universal time (UTC) value (defined by the
number of 100 ns ticks since New Year midnight year 1 A.C. at Greenwich). A point in time
corresponds to different date-time values at different locations depending on time zone and
local rules for daylight saving time. A standard computer configuration keeps track of local
time and UTC time. They are available at any time as the present values of the variables:
open System;;
let localNow = DateTime.Now;; // local time
let UtcNow = DateTime.UtcNow;; // Utc time

It is in general not possible to convert an arbitrary date-time object to universal time.

The System.TimeZoneInfo class contains a large array of TimeZoneInfo objects
let zones = TimeZoneInfo.GetSystemTimeZones()

A TimeZoneInfo object can be used to convert between local standard time and universal
time, but the conversion does not cater for daylight saving time. Further information can be
found in
 Conversion to textual form using String.Format
The function (static member) Format of the String class in the System library has a large
number of overloads, in particular:
String.Format(formatString, values)
String.Format(cultureInfo , formatString, values)

where
A CultureInfo object.
cultureInfo String constant consisting of fixed texts intermixed with
formatString format items.
values One or several expressions e0, e1, … , en−1 separated by
comma.
where
String not containing any brace characters {
fixed text
or }.
has one of the forms:
{ index }
format
{ index : format }
item
{ index, alignment }
{ index, alignment : format}

where
index k = 0, 1, … identifies the argument vk to be formatted.
alignment Number of printing positions including heading spaces.
format One-letter format code optionally followed by precision.

Table 10.12 Selected Numeric Format Codes

C or c Currency with national currency symbol

D or d Decimal notation
E or e Exponential notation
F or f Fixed point notation
N or n Number
X or x Hexadecimal

Table 10.13 Selected Date-time Format Codes

d Short date
D Long date
t Short time
T
Long time
 F Long date and long time
g
Short date and short time
M or m
Month and day
Y or y
Year and month

Precision is an integer between 0 and 99 specifying the number of decimals. There are two
kinds of formats: numeric formats as shown in Table 10.12 and date-time formats as shown in
Table 10.13. (Further information can be found in the .NET documentation web-pages for
Numeric Format Strings and Date Time Format Strings.)
Some examples:
open System ;;
String.Format("{,7:F2}",35.2) ;;
val it : string = " 35,20"
let dk = CultureInfo "da-DK" ;;
let en = CultureInfo "en-US" ;;
let ru = CultureInfo "ru-RU" ;;
let now = DateTime.Now ;;
String.Format(dk, "1:d...0:c", 45, now) ;;
val it : string = "17-10-2011...kr 45,00"
String.Format(ru,"{0:d}",now) ;;
val it : string = "17.10.2011"
String.Format(en,"{0:d}",now) ;;
val it : string = "10/17/2011"
String.Format(en,"{0:F}",now) ;;
val it : string = "Monday, October 17, 2011 2:57:50 PM"
let ar = CultureInfo "es-AR" ;;
String.Format(ar,"{0:F}",now) ;;
val it : string = "lunes, 17 de octubre de 2011 02:57:50 p.m."
 10.8 Keyword index example: The IndexGen program
We shall now return to the keyword index problem described in Section 10.1. A solution to
this problem will be described using the concepts introduced previously in this chapter. We
will just describe the main ingredients in the following. The complete program for the
solution appears in Appendix A.2. A system diagram for the program generating the keyword
web-page, called -IndexGen- is shown in Figure 10.2.
The program reads a keyword file keywords.txt and a binary file webCat.bin and
produces a keyword index web-page index.html. An extract from keywords.txt is given in
Table 10.1. This file is organized into lines where each line contains a title and a list of
keywords, like the line containing the title Control.Observable Module (F#) and two
keywords: observer and event observer.

Fig ure 10.2 System diagram for IndexGen program

The box webCat in the system diagram (see Figure 10.2) is a map from titles to uri’s:
webCat: Map<string,string>

Such a map is called a webCat map in the following, and could contain the entry with:
key: "Control.Observable Module (F#)"
value: "http://msdn.microsoft.com/en-us/library/ee370313"

The file webCat.bin is a binary file for a webCat map and we shall see in Section 10.9 how
this file to a large extend can be generated automatically.
The set keyWdSet in the system diagram (see Figure 10.2) has the type:
keyWdSet: Set<orderString*string>

An element in this set is called a webEntry and consists of a pair of keyword and associated
uri, where the keyword is encoded in an orderString value. The set could include the
following two elements:
("observer",
"http://msdn.microsoft.com/en-us/library/ee370313")
("event observer",
"http://msdn.microsoft.com/en-us/library/ee370313")

A set like keyWdSet is called a webEntry set in the following.

The set keyWdSet is generated from the files: keyword.txt and webCat.bin by the
function: keyWdIn. The index.html file is generated by the webOut function on the basis of
keyWdSet. Table 10.14 gives a short extract of the generated web-page and Figure 10.1 shows
how the index.html file is shown in a browser. Clicking on the keyword observer in the
browser will show the web-page with uri:
http://msdn.microsoft.com/en-us/library/ee370313

that is, the web-page containing the documentation of observer.

...
<a href="http://msdn.microsoft.com/en-us/library/ee353608">
null literal</a><br />
<a href="http://msdn.microsoft.com/en-us/library/ee353820">
numeric literal</a><br />
<br />
<a href="http://msdn.microsoft.com/en-us/library/ee370313">
observer</a><br />
<a href="http://msdn.microsoft.com/en-us/library/ee353721">
open</a><br />
<a href="http://msdn.microsoft.com/en-us/library/ee340450">
...

Table 10.14 An extract of index.html

 Data Capture
We now show how to capture the text in our keyword index file (see Section 10.1) to allow
comment lines in the keyword file. We will allow two types of comment lines, a blank line and
a line starting with two slash characters.
These two syntactic patterns are described by the regular expression comReg containing the
“or” operator |. Note that the sub-expression \G// for comment lines is without trailing
anchor “$”. A string will hence match this pattern just if the two first characters in the string
are slash characters.
let comReg = Regex @"(?:\G\s*$)|(?:\G//)";;

A normal line with keyword data matches the regular expression:

let reg =
Regex @"\G\s*\042([^\042]+)\042(?:\s+([^\s]+))*\s*$";;

The getData function should return a result of type:

type resType = | KeywData of string * string list
| Comment
| SyntError of string;;

and the replacement of tilde characters by spaces is made by the function tildeReplace:
let tildeReg = Regex @"~";;
let tildeReplace str = tildeReg.Replace(str," ");;

These ideas and the techniques from Section 10.2 are combined in the function getData:
let getData str =
let m = reg.Match str
if m.Success
then KeywData(captureSingle m 1,
List.map tildeReplace (captureList m 2))
else let m = comReg.Match str
if m.Success then Comment
else SyntError str;;
val getData : string -> resType
 The function keyWdIn
This part uses the techniques in Section 10.6 to create a webEntry set, where the keywords are
ordered using a cultural dependent ordering.
The declarations of keyWdIn is as follows:
let enString = orderString "en-US";;
val enString : (string -> orderString)

let keyWdIn() =
let webCat = restoreValue "webCat.bin"
let handleLine (keywSet: Set<orderString*string>) str =
match getData str with
| Comment -> keywSet
| SyntError str -> failwith ("SyntaxError: " + str)
| KeywData (title,keywL) ->
let uri = Map.find title webCat
let addKeywd kws kw = Set.add (enString kw, uri) kws
List.fold addKeywd keywSet keywL
let keyWdSet = Set.empty<orderString*string>
fileFold handleLine keyWdSet "keywords.txt";;
val keyWdIn : unit -> Set<orderString * string>

The idea is to build a webEntry set by folding a function handleLine over all the lines of the
keywords.txt file. The function handleLine translates the title in a line to the corresponding
uri using the webCat map (that has earlier been input from the webCat.bin file). This uri is
then paired with each keyword in the line and these pairs are inserted in the webEntry set.
 The function webOut
The function
webOut: Set<orderString * string> -> unit

that generate the file index.html on the basis of a webEntry set is given in Table 10.15. The
file index.html is encoded in the HTML (Hyper Text Mark-up Language) format. The
preamble contains the HTML string that sets up the web-page while the postamble contains
the HTML-tags that must appear at the end of the file. Please consult Appendix A.1 for a brief
introduction to HTML, and Appendix A.2 for the complete source code.
The file index.html is opened and the pre-amble is written on the file. A function outAct
to output a single keyword line is declared, and the keywords with uri’s in the webEntry set
keyWdSet are written on the file using Set.fold. The post-amble is written and the file is
closed.
The local outAct function
outAct: char -> (orderString * string) -> char

outputs one keyword with uri from keyWdSet to the web-page. An extra empty line is inserted
whenever the first character of the keyword is a letter different from the previous first
character. The argument of the function is, therefore, the first character of the previous
keyword and the value of the function is the first character of the just treated keyword.

let preamble = "<!DOCTYPE ...... /p>";;

let postamble = "</body></html>" ;;

let webOut(keyWdSet) =
use webPage = File.CreateText "index.html"
let outAct oldChar (orderKwd: orderString,uri: string) =
let keyword = string orderKwd
let newChar = keyword.[0]
if Char.ToLower newChar <> Char.ToLower oldChar
&& Char.IsLetter newChar
then webPage.WriteLine "<br />"
else ()
webPage.Write "<a href=\""
webPage.Write uri
webPage.WriteLine "\"..>"
webPage.Write (HttpUtility.HtmlEncode keyword)
webPage.WriteLine "</a><br />"
newChar

webPage.WriteLine preamble
Set.fold outAct ′a′ keyWdSet |> ignore
webPage.WriteLine postamble
webPage.Close()
 Table 10.15 The webOut function

The keyword is extracted from the orderString value. It becomes a displayed text in the
web-page and must hence be encoded in HTML encoding. This is done using the HttpUtility
function HtmlEncode from the System.Web library. The uri becomes an attribute and should
hence not be encoded.
 10.9 Keyword index example: Analysis of a web-source
This function addresses the problem of creating a suitable webCat map. A major challenge is
that the library documentation is spread over a huge number of different web-pages.
The documentation web-pages that are used in the keyword index have the property that
they are organized in two trees. Any web-page of interest can be reached by following links
starting from one of the root pages:

F# Core Library Reference

.NET Framework Class Library

These root pages have the following uri’s:

http://msdn.microsoft.com/en-us/library/ee353567.aspx
http://msdn.microsoft.com/en-us/library/gg145045.aspx

Scanning the HTML source of the web-page of the F# Core Library Reference we may for
instance find the following link to another documentation web-page:
<a data-tochassubtree="true"
href="/en-us/library/ee370255"
id="ee340636_VS.100_en-us"
title="System.Collections Namespace (F#)">
System.Collections Namespace (F#)</a>

This can be used to find the new title System.Collections Namespace (F#) with associated
path /en-us/library/ee370255. The path is relative to the current web-page but it can be
converted to an absolute uri.
 Using an XML reader
Analysis of the HTML-source of a web-page can be made using an XML reader, provided that
the web-page conforms to the XML-syntax. This is, fortunately, the case for the F# and .NET
documentation pages. The complete HTML-source of the web-page with given uri is read in
one operation and made available as a string (named doc). A stringReader is created, and an
XMLReader is then created to be used in the analysis of the HTML-source:

let baseUri = Uri uri

let webCl = new WebClient()
let doc = webCl.DownloadString baseUri
use docRd = new StringReader(doc)
let settings = XmlReaderSettings(DtdProcessing
= DtdProcessing.Ignore)
use reader = XmlReader.Create(docRd,settings)

The ignore setting of DtdProcessing is required for some security reasons. The XML reader
is a mutable data structure pointing at any time to an XML node in the HTML-source.
Successive calls of reader.Read() will step the reader through the nodes. Properties of the
current node are found as the current values of members of the reader object, where we will
use the following properties of the current node:
reader.NodeType: XmlNodeType
reader.Name: string
reader.Value: string
reader.GetAttribute: string
int
-> string
reader.Depth:
We use the following values of XmlNodeType:
XmlNodeType.Element HTML element as given by reader.Name
XmlNodeType.EndElement HTML end element as given by reader.Name
XmlNodeType.Text Text to be displayed as given by reader.Value
 The library documentation
The data capture from a Microsoft library documentation page uses a specific property that is
common for these pages, namely the navigation menu shown in the left column of the page. It
contains buttons like the following:
...
Visual F#
F# Core Library Reference
Microsoft.FSharp Collections Namespace
Collections.Array Module (F#)
Collections.Array2D Module (F#)
...
The button shown in orange (here in bold) points to the current page. It is indented one level.
The interesting part (for us) are the buttons just below pointing to the next level of
documentation. The start of this sub-menu of the navigation menu is indicated in the web-
source by a div start element with class attribute toclevel2 children:
<div … class="toclevel2 children" …>

while the end of the sub-menu is indicated by a matching div end element </div> at the same
level (that is, with the same Depth).
A button is given by
<a … href="path" … >text</a>

as described in Appendix A.1.

These observations are captured in the function nextInfo in Table 10.16 that steps an
XmlReader forward to the next “node of interest”.
The possible values returned by nextInfo indicate:
StartInfo d Found the start div element of the sub-menu at depth d.
EndDiv d Found an end div element at depth d.
RefInfo(text,path) Found a button <a . . . href="path". . .>text</a>.
EndOfFile The XmlReader has reached the end of the file.
The function nextInfo shown in Table 10.16 is the essential ingredient in constructing the
function getWebRefs
val getWEBrefs : string -> (string * string) list

that takes a uri as argument and reads through the corresponding web-page source and
extracts the list of pairs (title,uri) corresponding to buttons in the above described sub-menu
of the navigation menu. The complete program is found in Appendix A.3. The reader may pay
special attention to the following:
 The use of the HttpUtility.HtmlDecode function.
The use of an Uri object to convert a path in a link to the corresponding absolute uri.

Table 10.16 The nextInfo function

type infoType =
| StartInfo of int | EndDiv of int
| RefInfo of string * string | EndOfFile;;

let rec nextInfo(r:XmlReader) =

if not (r.Read()) then EndOfFile
else match r.NodeType with
| XmlNodeType.Element ->
match r.Name with
| "div" when (r.GetAttribute "class"
= "toclevel2 children")
-> StartInfo (r.Depth)
| "a" -> let path = r.GetAttribute "href"
ignore(r.Read())
RefInfo(r.Value,path)
| _ -> nextInfo r
| XmlNodeType.EndElement when r.Name = "div"
-> EndDiv (r.Depth)
| _ -> nextInfo r;;
val nextInfo: XmlReader -> infoType
 10.10 Keyword index example: Putting it all together
The keyword index is generated using the text files:

webCat0.txt
keywords.txt

and the programs:

NextLevelRefs
MakeWebCat
IndexGen

The text file webCat0.txt is shown in Table 10.17. It contains two pairs of lines with the title
and uri of the two root documentation pages. The text file keywords.txt contains titles of
documentation web-pages with associated keywords and an extract of the file is shown in
Table 10.1.
The keyword index is generates in three steps:

1. Generate the text files webCat1.txt and webCat2.txt using NextLevelRefs.

2. Generate the binary file webCat.bin using MakeWebCat.
3. Generate the text file index.html using IndexGen.

F# Core Library Reference

http://msdn.microsoft.com/en-us/library/ee353567.aspx
.NET Framework Class Library
http://msdn.microsoft.com/en-us/library/gg145045.aspx

Table 10.17 The file webCat0.txt

 Generating webCat1.txt and webCat2.txt
This step uses the program NextLevelRefs (as shown in Appendix A.3). It operates on
webCat text files that consists of consecutive pairs of lines containing title and uri for
documentation web-pages (like webCat0.txt). The program is called with such a file as input.
It applies the function getWebRefs to each uri in the input file and produces a webCat output
file containing all titles and uri’s found in this way.
The program is called twice: first with input file webCat0.txt to create the new webCat file
webCat1.txt containing all titles with uri’s one level down in the library documentation trees
and then with input file webCat1.txt to create the new webCat file webCat2.txt containing
all titles with uri’s two levels down in the library documentation trees.
Using a free-standing NextLevelRefs program (cf. Section 1.10 and Section 7.2) one
makes two calls in a command prompt (assuming that the webCat files and the program are
placed in the same directory):
NextLevelRefs webCat0.txt webCat1.txt
NextLevelRefs webCat1.txt webCat2.txt

Using the file NextLevelRefs.fsx in an interactive environment one makes two calls of the
main function:

main [| "webCat0.txt"; "webCat1.txt" |];;

val it : int = 0
main [| "webCat1.txt"; "webCat2.txt" |];;
val it : int = 0

where the webCat files are placed in a directory defined by the interactive environment. The
keyword index is designed to contain references to documentation web-pages two levels
down in the trees, so the files webCat0.txt, webCat1.txt and webCat2.txt contain all the
information needed to build the webCat - but in textual form.
An extract of the file webCat1.txt is shown in Table 10.18. The files webCat1.txt and
webCat2.txt has the same structure as webCat0.txt containing pairs of lines with title and
associated uri of documentation web-pages:

...
Microsoft.FSharp.Collections Namespace (F#)
http://msdn.microsoft.com/en-us/library/ee353413
Microsoft.FSharp.Control Namespace (F#)
http://msdn.microsoft.com/en-us/library/ee340440
Microsoft.FSharp.Core Namespace (F#)
http://msdn.microsoft.com/en-us/library/ee353649
...

Table 10.18 Extract of the file webCat1.txt

open System.IO;;
open TextProcessing;;

let addOneRef (e: Map<string,string>) (rd: StreamReader) =

Map.add (rd.ReadLine()) (rd.ReadLine()) e;;

let addRefsInFile (e: Map<string,string>) path =

fileXfold addOneRef e path;;

[<EntryPoint>]
let main (args: string[]) =
let webCat = Array.fold
addRefsInFile
Map.empty
args
saveValue webCat "webCat.bin"
0;;

Table 10.19 The file MakeWebCat.fsx

 Generating webCat.bin
This is done using the MakeWebCat program shown in Table 10.19. A free-standing version of
this program is called as follows in a command window:
MakeWebCat webCat0.txt webCat1.txt webCat2.txt

The program inputs the title-uri pairs in each of the specified webCat input files and collects
the corresponding webCat entries in the webCat map. This map is then saved in the file
webCat.bin using the saveValue function in the books TextProcessing library.
Using the file MakeWebCat.fsx in an interactive environment one should make the
following call of the main function:
main [| "webCat0.txt"; "webCat1.txt"; "webCat2.txt" |];;
val it : int = 0
 Generating the index.html file
The stage is now set for calling the IndexGen program described in Section 10.8. A free-
standing IndexGen program is called:
IndexGen

from a command window, assuming that the webCat.bin and keywords.txt are placed in the
same directory as the program. The program restores the webCat map from the binary file
webCat.bin and reads the text file keywords.txt. The data are captured as described in
Section 10.8 and each keyword is combined with the uri found by searching the
corresponding title in the webCat map. Upon that the index.html file is produced by the
webOut function.
Using the file IndexGen in an interactive environment one makes the following call of its
main function

main [||];;
val it : int = 0
 Summary
In this chapter we have studied a part of the text processing facilities available in F# and the
.NET library, including

regular expressions,
textual input and output from and to files,
save and retrieval of F# values on and from files,
culture-dependant ordering of strings,
retrieval of web-information, and
processing XML files.

These text-processing facilities are generally applicable, and we have illustrated their
expressive power in the construction of a program that can generate a web-page containing a
keyword index of the F# and .NET library documentation.
 Exercise
10.1 The term “word” is used in this exercise to denote a string not containing blank
characters. The blank characters of a string do hence divide the string into words. Make a
program WordCount that is called with two parameters:
WordCount inputFile outputFile

The program should read the input file and produce an output file where each line contains a
word found in the input file together with the number of occurrences, for example, “peter
3”. The words should appears in alphabetic order and the program should not distinguish
between small and capital letters.
10.2 The HTML elements <pre> … </pre> encloses a pre-formatted part of the web-page.
This part is displayed exactly as written, including spaces and line breaks, but each line
should, of course, be encoded in HTML encoding. Parts of this text can be copied using
copy-paste when the page is displayed using a web-browser. Make a program with program
call
examplePage fileName.txt

that inputs the contents of the text file fileName.txt and produces a web-page fileName.html
containing the contents of this file as preformatted text.
10.3 This exercise is a continuation of Exercise 10.1.

1. We do not consider the hyphen character “-” a proper character in a word. Make a
function to capture the list of words in a string while removing any hyphen character.
2. Make a function of type string -> (string list)*(string option) removing
hyphen characters like the previous one, but treating the last word in the line in a special
way: we get the result
([word0; … ;wordn−1],None)
if the last word in the string does not end with a hyphen character, and ([word0; …
;wordn−2 ],

Some wordn−1)
if the last word terminates with a hyphen character.

Make a new version of the WordCount program in Exercise 10.1 that in general ignores
hyphen characters but handles words that are divided from a text line to the next by means of a
hyphen character.
10.4 A position on Earth is given by geographic longitude and latitude written in the form:
14 ^ 27′35.03" E 55 ^ 13′47" N
containing degrees, minutes (1/60 degrees) and seconds (1/60 minutes) where the letters E or
W denote positive or negative sign on a longitude while N or S denote positive or negative sign
on an latitude. The seconds (here: 35.03) may have a decimal point followed by decimals or it
may just be an integer. The unit symbols ( and ′ and ") are assumed to consist of one or
several non-digit and non-letter characters. Make a program to capture a position from a
string as a value of type float*float.
10.5 Make alternative versions of the programs IndexGen and WebCatGen in the keyword
index problem, where WebCat and KeyWdSet are represented using the imperative
collections Dictionary and SortedSet (see Section 8.11). The programs should use iter
functions to build these imperative collections. The files keywords.txt, webCat0.txt,
webCat1.txt and webCat2.txt can be found on the web-page of the book.
 11
Sequences

A sequence is a possibly infinite, ordered collection of elements seq[e0; e1; …]. The
elements of a sequence are computed on demand only, as it would make no sense to actually
compute an infinite sequence. Thus, at any stage in a computation, just a finite portion of the
sequence has been computed.
The notion of a sequence provides a useful abstraction in a variety of applications where
you are dealing with elements that should be processed one after the other. Sequences are
supported by the collection library of F# and many of the library functions on lists presented
in Chapter 5 have similar sequence variants. Furthermore, sequences can be defined in F#
using sequence expressions, defining a process for generating the elements.
The type seq<′a> is a synonym for the .NET type IEnumerable<′a> and any .NET
framework type that implements this interface can be used as a sequence. One consequence of
this is, for the F# language, that lists and arrays (that are specializations of sequences) can be
used as sequence arguments for the functions in the Seq library. Another consequence is that
results from the Language-Integrated Query or LINQ component of the .NET framework can
be viewed as F# sequences. LINQ gives query support for different kinds of sources like SQL
databases and XML repositories. We shall exploit this in connection with a database for a
simple product-register application, where we shall introduce the concept type provider that
makes it possible to work with values from external data sources (like SQL databases) in a
type safe manner, and the concept of query expressions that gives support for expressing
queries on SQL databases in F#.
Sequence expressions and query expressions are special kinds of computation expressions
(also called workflows), a concept of F# that will be addressed in Chapter 12.
 11.1 The sequence concept in F#
A finite sequence can be formed like a list using the “sequence builder” seq (see also Section
11.7):
seq [10; 7; -25];;
val it : seq<int> = [10; 7; -25]

and we obtain a value of type seq<int>. This simple example works just like a list.
To study the special features of sequences, we will consider infinite sequences. An infinite
sequence can be obtained by using the library function:
Seq.initInfinite: (int -> ′a) -> seq<′a>

The value of Seq.initInfinite f, denotes the infinite sequence:

where the elements are computed on demand.

The sequence of natural numbers 0, 1, 2, … is obtained as follows:
let nat = Seq.initInfinite (fun i -> i);;
val nat : seq<int>

and the i’th element in the sequence nat, when numbering starts with 0, is obtained using the
function Seq.nth: int -> seq<’a> -> ’a. For example:
Seq.nth 5 nat;;
val it : int = 5

So far just the fifth element of nat is computed – this is the only element demanded.
To study the consequences of such on demand computation we modify the example so that the
number is printed whenever it is demanded, using the printfn function (Section 10.7).
Evaluation of the expression printfn "%d" i has the side-effect that the integer i is printed on
the console, for example:
printfn "%d" 10;;
10
val it : unit = ()

The natural number sequence with print of demanded elements is declared as follows:
let idWithPrint i = printfn "%d" i
i;;
val idWithPrint : int -> int

let natWithPrint = Seq.initInfinite idWithPrint;;

val natWithPrint : seq<int>
 The function idWithPrint is the identity function on integers that has the side-effect of
printing the returned value, for example:
idWithPrint 5;;
5
val it : int = 5

Extracting the third and fifth elements of natWithPrint will print just those elements:
Seq.nth 3 natWithPrint;;
3
val it : int = 3

Seq.nth 5 natWithPrint;;
5
val it : int = 5

In particular, the elements 0, 1, 2, and 4 are not computed at all.

Extracting the third element again will result in a reprint of that element:
Seq.nth 5 natWithPrint;;
5
val it : int = 5

Thus the n’th element of a sequence is recomputed each time it is demanded. If elements are
needed once only or if the re-computation is cheap or giving a desired side-effect, then this is
fine.
 Cached sequences
A cached sequence can be used if a re-computation of elements (such as 5 above) is
undesirable. A cached sequence remembers the initial portion of the sequence that has already
been computed. This initial portion, also called a prefix, comprises the elements e0, …, en,
when en is the demanded element with the highest index n.
We will illustrate this notion of cached sequence using the previous example. A cached
sequence of natural numbers is obtained using the library function
Seq.cache: seq<′a> -> seq<′a>

and it is used as follows:

let natWithPrintCached = Seq.cache natWithPrint;;
val natWithPrintCached : seq<int>

Demanding the third element of this sequence will lead to a computation of the prefix 0, 1,
2, 3 as we can see from the output from the system:
Seq.nth 3 natWithPrintCached;;
0
1
2
3
val it : int = 3

Demanding the fifth element will extend the prefix to 0, 1, 2, 3, 4, 5 but just two elements are
computed and hence printed:
Seq.nth 5 natWithPrintCached;;
4
5
val it : int = 5

and there is no re-computation of cached elements:

Seq.nth 5 natWithPrintCached;;
val it : int = 5
Table 11.1 Selected functions from the sequence library Seq
 11.2 Some operations on sequences
The Seq library contains a rich collection of functions. Some of those are described in Table
11.1. In the remaining part of this chapter we illustrate the use of these functions.
 Creating sequences
The empty sequence is denoted Seq.empty. A one element sequence, that is, a singleton
sequence, is created using Seq.singleton, and a finite sequence can be generated from a list
using Seq.ofList. For example:
Seq.empty;;
val it : seq<′a> = seq []

Seq.singleton "abc";;
val it : seq<string> = seq ["abc"]

Seq.ofList [(true,"a"); (false,"b"); (false,"ab")];;

val it : seq<bool * string>
= [(true, "a"); (false, "b"); (false, "ab")]

The function Seq.init is used to generate a finite sequence. The value of Seq.init n f is the
sequence

having n elements. For example:

Seq.init 3 (fun i -> 2*i);;
val it : seq<int> = seq [0; 2; 4]
 Appending sequences
The operation Seq.append appending sequences works for finite sequences like @ on lists,
except for the "on demand" property:
let s1 = Seq.append (seq [1;2;3;4]) (seq [5;6]);;
val s1 : seq<int>

Seq.toList s1;;
val it : int list = [1; 2; 3; 4; 5; 6]

where the computation of the elements of s1 is delayed until they are demanded by the
conversion of the sequence to a list using Seq.toList.
If s denotes an infinite sequence, then Seq.append s s′ is equal to s for any sequence s′. For
example:
let s2 = Seq.append nat s1;;
val s2 : seq<int>

Seq.nth 1000 s2;;

val it : int = 1000

If s = seq[e0; …; en−1] and s′ = seq[e′0; e′1; …], then

that is, s” is the sequence, where

For example:
let s3 = Seq.append s1 nat;;
(Seq.nth 2 s3 = Seq.nth 2 s1)
&& (Seq.nth 10 s3 = Seq.nth (10-6) nat);;
val it : bool = true

since s1 has six elements.

There is no function in the Seq library that directly corresponds to the function for cons′ing
an element x to a list xs, that is, x :: xs. But a cons function for sequence is easily defined
using Seq.singleton and Seq.append:
let cons x sq = Seq.append (Seq.singleton x) sq;;
val cons : 'a -> seq<'a> -> seq<'a>

cons 5 (seq [6; 7; 8]);;

val it : seq<int> = seq [5; 6; 7; 8]
 11.3 Delays, recursion and side-effects
The functions above yield sequences that are lazy, that is, elements of the resulting sequence
are not computed unless they are needed. The same is not necessarily true when sequences are
defined by recursive functions. This is illustrated by the following failing attempt to declare a
function that generates the sequence of integers starting from i:
let rec from i = cons i (from(i+1));;
val from : int -> seq<int>

Applying this function to any integer i yields a non-terminating evaluation:

The problem is that the arguments to cons are evaluated before the application of cons can
create a “lazy” sequence using Seq.append. This is caused by the eager evaluation strategy of
F#, and it results in an infinite recursive unfolding of from.
This problem can be avoided as described below by using the function
Seq.delay: (unit -> seq<′a>) -> seq<′a>

that suspends the computation of a sequence. We first motivate this function:

Suppose that e is an expression of type seq<′a>. The closure fun () -> e is then a value of
type unit -> seq<′a> and this value can be viewed as a lazy representation of the value of e.
A sequence generated by seq is for instance not lazy:
seq [idWithPrint 1];;
1
val it : seq<int> = [1]

but "packing" this expression into a closure

let sf = fun () -> seq [idWithPrint 1];;
val sf : unit -> seq<int>

gives a representation of the sequence where the element remain unevaluated. The element is
evaluated when the function is applied:
sf();;
1
val it : seq<int> = [1]

The function Seq.delay converts a closure like sf into a lazy sequence:

let s1 = Seq.delay sf;;
val s1 : seq<int>
Seq.nth 0 s1;;
1
val it : int = 1

There are two ways of using Seq.delay to modify the above recursive declaration of from
in order to avoid the problem with non-terminating evaluations. Either the recursive call of
from can be delayed:

let rec from1 i = cons i (Seq.delay (fun () -> from1(i+1)));;

val from1 : int -> seq<int>

or the computation of the function value can be delayed:

let rec from2 i = Seq.delay (fun () -> cons i (from2(i+1)));;
val from2 : int -> seq<int>

Both of these functions can be used to generate lazy sequences:

let nat10 = from1 10;;
val nat10 : seq<int>

let nat15 = from2 15;;

val nat15 : seq<int>

Seq.nth 5 nat10;;
val it : int = 15

Seq.nth 5 nat15;;
val it : int = 20

There is a difference between these results of turning from into a function that generates
lazy sequences. This difference becomes visible in the presence of the side-effect of printing
the computed numbers using the idWithPrint function:
let rec fromWithPrint1 i =
cons (idWithPrint i)
(Seq.delay (fun () -> fromWithPrint1(i+1)));;
val fromWithPrint1 : int -> seq<int>

let rec fromWithPrint2 i =

Seq.delay (fun () -> cons (idWithPrint i)
(fromWithPrint2(i+1)));;
val fromWithPrint2 : int -> seq<int>

The first element of a sequence created by using fromWithPrint1 is eagerly computed

when the sequence is created
let nat10a = fromWithPrint1 10;;
10
val nat10a : seq<int>

since the argument idWithPrint i to cons is computed, due to the eager evaluation strategy
of F#, before cons can create a lazy sequence. Therefore, 10 is printed out when the above
declaration of nat10a is elaborated and not later when an element of the sequence is
demanded:
Seq.nth 3 nat10a;;
11
12
13
val it : int = 13

On the other hand, no element of a sequence created by fromWithPrint2 is created at

declaration time, since the value of that function is delayed, and all side-effects show up when
elements are demanded:
let nat10b = fromWithPrint2 10;;
val nat10b : seq<int>

Seq.nth 3 nat10b;;
10
11
12
13
val it : int = 13
 11.4 Example: Sieve of Eratosthenes
The Greek mathematician Eratosthenes (194 – 176 BC) invented a process, called the sieve of
Eratosthenes, for generating the sequence of prime numbers. The process starts with the
sequence of natural numbers that are greater than 1. The process is described as follows:

1. Select the head element p of the current sequence as the next prime number.
2. Remove multiples of p from the current sequence, yielding a new sequence.
3. Repeat the process from 1. with the new sequence.

Suppose a is the head element of the sequence at the start of some iteration of the process. No
number between 2 and a-1 divides a since since multiples of those numbers have been
removed from the sequence by step 2. in the process.
The first three iterations of the process can be illustrated as follows:

The function that removes all multiples of a number from a sequence is called sift, and it
is declared using the filter function for sequences:
let sift a sq = Seq.filter (fun n -> n % a <> 0) sq;;
val sift : int -> seq<int> -> seq<int>

The above iterations are exemplified in F# as follows:

Seq.initInfinite (fun i -> i+2);;
val it : seq = seq<int> [2; 3; 4; 5; ...]

sift 2 it;;
val it : seq<int> = seq [3; 5; 7; 9; ...]

sift 3 it;;
val it : seq<int> = seq [5; 7; 11; 13; ...]

The “Sieve of Erastothenes”, corresponding to the process above, is obtained by a

recursive application of sieve to the sequence seq[2; 3; 4; 5; …]:
let rec sieve sq =
Seq.delay (fun () ->
let p = Seq.nth 0 sq
cons p (sieve(sift p (Seq.skip 1 sq))));;
val sieve : seq<int> -> seq<int>

let primes = sieve (Seq.initInfinite (fun n -> n+2));;

val primes : seq<int>

The head and tail of a sequence are extracted in the above declaration by the functions
Seq.nth 0 and Seq.skip 1, respectively. See also Table 11.2.
The function that finds the n′th prime number is declared as follows:
let nthPrime n = Seq.nth n primes;;
val nthPrime : int -> int

The 5’th and 100’th prime numbers are fast to compute (remember that numbering starts at
0); but it requires some seconds to compute the 700’th prime number:
nthPrime 5;;
val it : int = 13

nthPrime 100;;
val it : int = 547

nthPrime 700;;
val it : int = 5281

A re-computation of the 700’th prime number takes the same time as before, and a
computations of the 705’th prime number would take approximately the same time. A use of a
cached prime number sequence will improve on that:
let primesCached = Seq.cache primes;;

let nthPrime1 n = Seq.nth n primesCached;;

val nthPrime1 : int -> int

Computing the 700’th prime number takes the same time as before; but a subsequent
computation of the 705’th is fast since that computation is based on a prefix containing the
first 700 prime numbers.
An alternative to the above declaration of primesCached is to let the sieve function return a
cached sequence:
let rec sieve sq =
Seq.cache
(Seq.delay (fun () ->
let p = Seq.nth 0 sq
cons p (sieve(sift p (Seq.skip 1 sq)))));;
 11.5 Limits of sequences: Newton-Raphson approximations
The Newton-Raphson method for computing the square root of a number is based on the fact
that , for a ≥ 0, is the limit of a sequence:

where x0 > 0 and

(11.1)
for any i ≥ 0.
It is easy to generate such infinite sequences using F# and to use the elements to
approximate the square root of a number a within a given tolerance.
The computation of the next element in the sequence on the basis of a and the current
element x follows from (11.1):
let next a x = (a/x + x)/2.0;;
val next : float -> float -> float

This function should be iterated over a sequence and to this end two auxiliary functions are
defined:
let rec iter f x = function
| 0 -> x
| n -> iter f (f x) (n-1);;
val iter : (′a -> ′a) -> ′a -> int -> ′a

let iterate f x = Seq.initInfinite (fun i -> iter f x i);;

val iterate : (′a -> ′a) -> ′a -> seq<′a>

where iter makes an n-fold application of a function:

and iterate gives a sequence on the basis of the iteration of a function:

The “Newton-Raphson” sequence for that starts at 1.0 is:

iterate (next 2.0) 1.0;;
val it : seq = seq<float> [1.0;1.5;1.416666667;1.414215686;...]

The following function traverses trough a sequence sq using an enumerator (see Section
8.12) and returns the first element where the distance to the next is within a given tolerance
eps:

let rec inTolerance (eps:float) sq =

let f = enumerator sq
let nextVal() = Option.get(f())
 let rec loop a = let b = nextVal()
if abs(a-b) > eps then loop b else a
loop(nextVal());;

The sequence argument sq is assumed to be infinite and the enumerator f will always return
some value, which is exploited by the function nextVal. The square roots of a can be
computed according to Newton-Raphson’s method within tolerance 10−6.
let sRoot a = inTolerance 1E-6 (iterate (next a) 1.0);;
val sRoot : float -> float

sRoot 2.0;;
val it : float = 1.414213562

This example illustrates the expressiveness of infinite sequences, but the above programs
can definitely be improved, in particular when considering efficiency issues. See, for
example, Exercise 11.5 and Exercise 11.6. But notice that none of these sequence-based
solutions is as efficient as a solution that just remembers the last computed approximation and
terminates when the next approximation is within the given tolerance.
 11.6 Sequence expressions
A sequence expression is a special kind of computational expression (cf. Chapter 12) that
allows a step-by-step generation of sequences.
The corresponding general form of expressions to generate sequences is

where seqexp is a sequence expression.

Suppose that we aim at generating a sequence of type seq<’a>. Then the following
sequence expressions are fundamental for the generation:

The sequence expression yield x, with x: ′a, adds the element x to the sequence.
The sequence expression yield! sq, with sq: seq<′a>, adds the sequence sq to the
sequence.

The following examples show three ways of generating the integer sequence seq[1, 2; 3]
using yield and yield! constructs:
let s123a = seq {yield 1
yield 2
yield 3 };;

let s123b = seq {yield 1

yield! seq [2; 3] };;

let s123c = seq {yield! seq [1; 2]

yield 3 };;

These examples all show a combination of sequence expressions of the form:

The meaning of this combination is the sequence obtained by appending the sequence denoted
by seqexp1 with the sequence denoted by seqexp2.
Expressions denoting sequences may in general be obtained using functions from the Seq
library; but sequence expressions have the advantage that the generated sequences by
construction are lazy, but not cached, and elements are therefore generated by demand only.
Hence an explicit use of Seq.delay can be avoided when using sequence expressions.
This implicit delaying of a sequence is exploited in the following declaration of the
function from from Section 11.2:
let rec from i = seq {yield i
yield! from (i+1)};;
val from : int -> seq<int>

let s10 = from 10;;

val s10 : seq<int>

Seq.nth 5 s10;;
val it : int = 15

Construct Legend
yield exp generate element
yield! exp generate sequence
seqexp1
combination of two sequences by appending them
seqexp2
let pat = exp
local declaration
seqexp
for pat in exp do seqexp iteration
if exp then seqexp filter
if exp then seqexp else seqexp conditional

Table 11.2 Constructs for sequence expressions

A summary of constructs used in sequence expressions is given in Table 11.2.

 Example: Sieve of Eratosthenes
The Sieve of Eratosthenes can also be formulated using sequence expressions.
The function sift a sq makes a filtering of the elements of the sequence sq by removing
multiples of a. This is expressed using an iteration over sq and a filter:
let sift a sq = seq {for n in sq do
if n % a <> 0 then
yield n };;
val sift : int -> seq<int> -> seq<int>

The use of sequence expressions gives, in this case, a more “verbose” alternative to the
succinct formulation using Seq.filter in Section 11.4.
The use of sequence expressions in the declaration of sieve sq is attractive since the
explicit delay of the recursive call can be avoided and the combination sequence expressions
is a brief alternative to using Seq.append:
let rec sieve sq =
seq {let p = Seq.nth 0 sq
yield p
yield! sieve(sift p (Seq.skip 1 sq)) };;
val sieve : seq<int> -> seq<int>

The remaining parts of the example are unchanged, for example:

let primes = sieve(Seq.initInfinite (fun n -> n+2));;
val primes : seq<int>

let nthPrime n = Seq.nth n primes;;

val nthPrime : int -> int
 Example: Search in a directory
We shall now use sequences in connection with the search for certain files in a directory.
Although there is a finite number of files to be considered, it is natural to exploit the laziness
of sequences since a search typically succeeds before the whole directory structure is
exhausted.
The function allFiles declared below gives a sequence of all files in a directory,
including those in subdirectories. The sequence of files and directories are extracted using the
functions Directory.GetFiles and Directory.GetDirectories from the library
System.IO:

open System.IO;;

let rec allFiles dir = seq {

yield! Directory.GetFiles dir
yield! Seq.collect allFiles (Directory.GetDirectories dir)};;
val allFiles : string -> seq<string>

The function
Seq.collect: (′a -> seq<′c>) -> seq<′a> -> seq<′c>

combines the map and concatenate functionality. The value of Seq.collect f sq is obtained
(lazily) by applying f to each element ei of the sequence sq. The result is obtained by
concatenation of all the sequences f(ei), for i = 0, 1, 2,….
Hence the expression
Seq.collect allFiles (Directory.GetDirectories dir)

recursively extracts all files in sub-directories of dir and concatenates the results.
The sequence of all files in the directory C:\mrh\Forskning\Cambridge, for example, can
be extracted as follows:
// set current directory
Directory.SetCurrentDirectory @"C:\mrh\Forskning\Cambridge\";;

let files = allFiles ".";;

val files : seq<string>

Seq.nth 100 files;;

val it : string = ".\BOOK\Satisfiability.fs"

A composite file name like .\BOOK\Satisfiability.fs can be split into three parts: a path
.\BOOK\, a file name Satisfiability and an extension fs. These three parts can be extracted
from a composite file name using regular expressions (see Page 222):

The path matches a sequence of characters ending with a backslash “\”, that is, it matches
the regular expression “\S*\\”.
 The file name matches a non-empty sequence of characters not containing backslash “\”,
that is, it matches the regular expression “[^ \ \]+”.

Suppose that reExts is a regular expression matching certain file extensions. Then the
following regular expression (where the string is obtained by joining several strings)
matches composite file names having one of these file extensions:

Note that the extension is the suffix of the composite file name that starts just after the last
period (.). The regular expression has three capturing groups (enclosed in normal brackets (
and )) so that the path, the name and the extension can be extracted from a matched composite
file name. The function captureSingle from the TextProcessing library (cf. Table 10.4) is
used to extract captured strings in the following declaration of searchFiles files exts that
gives those files in the sequence files that have an extension in the list exts:
open System.Text.RegularExpressions;;

let rec searchFiles files exts =

let reExts =
List.foldBack (fun ext re -> ext+"|"+re) exts ""
let re = Regex (@"G(\S*\\)([^\\]+)\.(" + reExts + ")$")
seq for fn in files do
let m = re.Match fn
if m.Success
then let path = TextProcessing.captureSingle m 1
let name = TextProcessing.captureSingle m 2
let ext = TextProcessing.captureSingle m 3
yield (path, name, ext) };;
val searchFiles : seq<string> -> string list
-> seq<string * string * string>

The search for F# files (with extensions fs or fsi) is initiated as follows:

let funFiles =
Seq.cache (searchFiles (allFiles ".") ["fs";"fsi"]);;
val funFiles : seq<string * string * string>

In this case a cached sequence is chosen so that a search can exploit the already computed
part of the sequence from previous searches:
Seq.nth 6 funFiles;;
val it : string * string * string = (".\BOOK\","Curve","fsi")

Seq.nth 11 funFiles;;
val it : string * string * string
= (".\BOOK\", "Satisfiability", "fs")
 11.7 Specializations of sequences
The type seq<’a> is a synonym for the .NET type IEnumerable<’a> and any .NET
framework type that implements this interface can be used as a sequence. Since lists and
arrays are specializations of sequences they can be used as sequence arguments for the
functions in the Seq library.
 Specialization of Seq-library functions
Using the function Seq.append
Seq.append: seq<′a> -> seq<′a> -> seq<′a>

a sequence can, for example, be formed by appending two lists:

let sq1 = Seq.append [1; 2; 3] [4; 5; 6];;
val sq1 : seq<int>

by appending two arrays (see Section 8.10)

let sq2 = Seq.append [|1; 2; 3|] [|4; 5; 6|];;
val sq2 : seq<int>

and by appending a list and an array:

let sq3 = Seq.append [1; 2; 3] [|4; 5; 6|];;
val sq3 : seq<int>

Similarly, the range rng in a for-construct: for i in rng do seqexp can be a sequence. This
is exploited in the following example where the range is the first eleven natural numbers:
let squares = seq {for i in 0..10 do yield i*i };;
val squares : seq<int>
 Range expressions
In a construction of a sequence constant, like
seq [1; 2; 3; 4];;
val it : seq<int> = [1; 2; 3; 4]

the sequence builder seq is actually a built-in function on sequences:

seq;;
val it : (seq<′a> -> seq<′a>) = <fun:clo@16-2>

and when applying seq to the list in the above example, we are just exploiting that lists are
specializations of sequences.
Hence the range expressions [b .. e] and [b .. s .. e] for lists described in Section 4.2 can,
in a natural manner, be used to generate finite sequences:
let evenUpTo n = seq [0..2.. n];;
val evenUpTo : int -> seq<int>

Sequences constructed using this function will not be lazy:

let e20 = evenUpTo 20;;
val e20 : seq<int> = [0; 2; 4; 6; 8; 10; 12; 14; 16; 18; 20]

The answer from the system shows that the sequences constructed in this way will not be lazy,
as the answer shows all eleven elements. The reason is that the construct [0..2.. n] eagerly
will construct a list.
 11.8 Type providers and databases
The Language-Integrated Query or LINQ component of the .NET framework gives query
support for different kinds of sources like SQL databases and XML documents. We shall now
exploit that LINQ queries for SQL databases return values of type IEnumerable<T>, and
hence the result of these queries can be viewed as F# sequences.
An obstacle in doing so is that the type system of an SQL database is different from that of
F#. To overcome this obstacle a type provider for SQL is used. An F# type provider for SQL
makes it possible to work directly with the tables of the database in a type safe manner. There
are type providers for other external sources than SQL databases, and it is possible to make
user-defined type providers; but we will just illustrate the use of a type provider for SQL in
connection with a database for a simple product-register application.
 A database for a Product Register
Suppose for the moment that the product register is a database ProductRegister containing
the two tables shown in Figure 11.1.

Fig ure 11.1 A Product Register Database with two tables: Part and PartsList

The table Part stores information about four parts named: “Part0”, … ,“Part3”, where
“Part0” and “Part1” are basic parts (with empty parts lists) since their IsBasic attribute is 1
(the SQL representation of true), while “Part2” and “Part3” are composite parts since their
IsBasic attribute is 0 (representing false). The PartId attribute of the Part table is a unique
identifier, that is, a key, for the description of a part.
The table PartsList contains the parts lists for all the composite parts. The attribute pair
(PartsListId, PartId) is a composite key. A row (pid, id, q) in this table, therefore,
describes that exactly q pieces of the part identified by id is required in the parts list of the
composite part identified by pid. For example, the parts list for “Part3” comprises 3 pieces of
“Part1” and 4 pieces of “Part2”.
Starting on Page 274 it is shown how this database can be created and updated. But before
that we address the issue of making queries to this database from F#.
 F# Type Providers for SQL
A type provider is a component that automatically generates types and functions for external
data sources like SQL databases and Web services. Below we shall use a built-in SQL type
provider for F#. In order to access and use this type provider, the following assemblies
should be referenced: FSharp.Data.TypeProviders.dll, System.Data.dll and
System.Data.Linq.dll, and namespaces should be opened (see Figure 11.2 for the complete
code) prior to the execution of the following declarations:
type schema =
SqlDataConnection<"Data Source=IMM-NBMRH\SQLEXPRESS;
InitialCatalog=ProductRegister;
Integrated Security=True">;;
type schema

let db = schema.GetDataContext();;
val db:
schema.ServiceTypes.SimpleDataContextTypes.ProductRegister

Fig ure 11.2 Creating a type provider for the ProductRegister database

The SQL type provider: SqlDataConnection has a connection string as argument. This
connection string has three parts: A definition of the data source, in this case an SQL server,
the initial catalog, in this case the database name, and the integrated security, which in this
case is true, meaning that the .NET credentials of the current user will be used for
authentication.
The type schema contains all the generated types that represent the database and db is an
object containing the database tables. The two database tables can be accessed as follows:
let partTable = db.Part;;
val partTable : Data.Linq.Table<schema.ServiceTypes.Part>

let partsListTable = db.PartsList;;

val partsListTable:
Data.Linq.Table<schema.ServiceTypes.PartsList>
The answers from the F# system do not reveal the F# values of these two tables.
They are in fact lazy sequences. For example:
partTable;;
val it : Data.Linq.Table<schema.ServiceTypes.Part> =
seq [Part {IsBasic = true; PartId = 0; PartName = "Part0"};;
Part {IsBasic = true; PartId = 1; PartName = "Part1"};;
Part {IsBasic = false; PartId = 2;PartName = "Part2"};;
Part {IsBasic = false; PartId = 3;PartName = "Part3"};]

where all the elements are shown in this case just because the interactive environment always
prints a short prefix of a lazy sequence. The elements of this sequence are objects belonging
to a class Part that has the attributes of the table as public fields:
let r = Seq.nth 2 partTable;;
val r : schema.ServiceTypes.Part

r.PartId;;
val it : int = 2

r.PartName;;
val it : string = "Part2"

r.IsBasic;;
val it : bool = false

Note that the SQL types bit and varchar(50) are translated to the F# types bool and string,
respectively, by the type provider.
The list of F# elements of the PartsList table is obtained as follows:
Seq.toList partsListTable;;
val it : schema.ServiceTypes.PartsList list =
[PartsList {PartId = 0; PartsListId = 2; Quantity = 5;};
PartsList {PartId = 1; PartsListId = 2; Quantity = 4;};
PartsList {PartId = 1; PartsListId = 3; Quantity = 3;};
PartsList {PartId = 2; PartsListId = 3; Quantity = 4;}]

Database queries can be expressed using the functions from the sequence library since the
tables in the database can be accessed like sequences when using the above type provider. The
names of all composite parts can, for example, be extracted as follows:
Seq.fold
(fun ns (r:schema.ServiceTypes.Part)
-> if r.IsBasic then ns else r.PartName::ns)
[]
partTable;;
val it : string list = ["Part3"; "Part2"]
 Query expressions
We shall now introduce query expressions as means for extracting information from the
database. A query expression is a computation expression (just like sequence expressions) and
it occurs in expressions of the form:
query { queryexp }
The construct select v adds the element v to the answer to the query just like yield v adds
an element to a sequence:
query {select (1, "a") };;
val it : IQueryable<int * string> = seq [(1, "a")]

The value of this query expression has type IQueryable<int * string>. The type
IQueryable<T> is a specialization of IEnumerable<T> and, therefore, values of type
IQueryable<T> can be treated as sequences.
There is a rich collection of query-expression constructs that translates to SQL queries. We
will now introduce a small part these constructs by illustrating how the following operations
of relational algebra can be expressed: projection, selection and join.

Projection
A projection operation extracts certain columns of a table and such a projection can be
expressed using an iteration.
For example, a query for the projection of the Part table with respect to PartName and
IsBasic is declared as follows:

let q1 = query {for part in db.Part do

select (part.PartName, part.IsBasic) };;

q1;;
val it : IQueryable<string * bool> =
seq [("Part0", true); ("Part1", true);
("Part2", false); ("Part3", false)]

Selection
A selection operation extracts certain rows of a table and such a selection can be expressed
using an iteration together with a where-clause and a selection.
For example, the query selecting the composite parts from the Part table is declared by:
let q2 =
query {for part in db.Part do
where (not part.IsBasic)
select (part.PartId, part.PartName, part.IsBasic)};;

q2;;
val it : IQueryable<int * string * bool> =
 seq [(2, "Part2", false); (3, "Part3", false)]

Join
A join operation combines the rows of two tables A and B. There are many different kinds of
such combinations that are supported by SQL and query expressions. We shall here just
consider what is called an equi-join, where a row a A is combined with a row b B only if
a.LA = b.LB, where LA is a given attribute of A and LB is a given attribute of B.
By an equi-join of PartsList and Part tables with PartsListId of PartsList equal to
PartId of Part we can extract tuples from PartsList where identifiers for parts list are
replaced by their names:
let q3 = query {for pl in db.PartsList do
join part in db.Part on
(pl.PartsListId = part.PartId)
select(part.PartName, pl.PartId, pl.Quantity)
};;
q3;;
val it : IQueryable<string * int * int> =
seq [("Part2", 0, 5); ("Part2", 1, 4);
("Part3", 1, 3); ("Part3", 2, 4)]

Hence “Part2” is a composite part consisting of 5 pieces of the part with PartId equal to 0
and 4 pieces of the part with PartId equal to 1. By the use of nested joins we can make a
query where these identifiers also are replaced by their names. The result of q4 cannot be used
in a join since the elements have a tuple type and not a record type. We therefore introduce a
record type:
type partListElement =
{PartName:string; PartId:int; Quantity:int}

In the following nested join, the local query qa is the variant of q3 that gives elements of
type partListElement:
let q4 =
query {let qa = query {for pl in db.PartsList do
join part in db.Part on
(pl.PartsListId = part.PartId)
select {PartName = part.PartName;
PartId = pl.PartId;
Quantity = pl.Quantity} }
for pl in qa do
join part in db.Part on
(pl.PartId = part.PartId)
select(pl.PartName, part.PartName, pl.Quantity) };;
q4;;
val it : IQueryable<string * string * int> =
seq
[("Part2", "Part0", 5); ("Part2", "Part1", 4);
("Part3", "Part1", 3); ("Part3", "Part2", 4)]
 Aggregate operations
In SQL there are so-called aggregate operations that depend on a whole table or all the values
in a column of a table, such as counting the number of elements in a table or finding the
average of the elements in a column. There are also query-expression constructs for these
functions, for example, count that counts the number of elements selected so far, exactlyOne
that returns the single element selected, and raises an exception if no element or more than
one element have been selected, and contains v that checks whether v is among the so far
selected elements.
The following function counts the number of rows in Part. Since we shall use consecutive
numbers 0, 1, …, n-1 as identifiers for existing parts, the number of rows n is the next
identifier that can be used as a key. This function is therefore named nextID:
let nextId() = query {for part in db.Part do
count };;
val nextId : unit -> int

The function getDesc extracts the description of a given identifier

let getDesc id =
query {for part in db.Part do
where (part.PartId=id)
select (part.PartName,part.IsBasic)
exactlyOne };;
val getDesc : int -> string * bool

where the description consists of the name and truth values of the Name and IsBasic attributes.
For example:
nextId();;
val it : int = 4

getDesc 3;;
val it : string * bool = ("Part3", false)

getDesc 4;;
System.InvalidOperationException: Sequence contains no elements

The predicate containsPartId checks whether a given identifier is in the Part table:
let containsPartId id = query {for part in db.Part do
select part.PartId
contains id };;
val containsPartId : int -> bool

containsPartId 3;;
val it : bool = true
containsPartId 4;;
val it : bool = false

Example: Parts Break Down

We shall now consider the problem of computing a parts list containing all the basic parts
needed to produce a given part. By a parts list we shall now understand a list of pairs: [(id1,
k1), …,(idn, kn)] where idi is the identifier of a part and ki is the quantity needed of that part.
The following function extracts the parts list for a given part:
let getPartsList id =
query {for pl in db.PartsList do
where (pl.PartsListId = id)
select (pl.PartId,pl.Quantity) };;
val getPartsList : int -> IQueryable<int * int>

getPartsList 3;;
val it : IQueryable<int * int> = seq [(1, 3); (2, 4)]

We shall need functions for adding a pair (id, k) to given parts list, for merging two parts
lists and for multiplying all quantities in a parts list by a constant. These functions are "usual"
auxiliary list functions:
let rec add pl (id,q) =
match pl with
| [] -> [(id,q)]
| (id1,q1)::pl1 when id=id1 -> (id,q+q1)::pl1
| idq::pl1 -> idq :: add pl1 (id,q);;
val add : (′a * int) list -> ′a * int -> (′a * int) list
when ′a : equality

let mergePartsList pl1 pl2 = List.fold add pl1 pl2;;

val mergePartsList :
(′a * int) list -> (′a * int) list -> (′a * int) list
when ′a : equality

let mult k pl = List.map (fun (id,q) -> (id,k*q)) pl;;

val mult : int -> (′a * int) list -> (′a * int) list

The following function partBreakDown that computes a parts list containing all basic parts
needed for producing a given part is declared in mutual recursion with the function
partsListBreakDown that computes a parts list containing all basic parts needed for
producing a given parts list. These functions access the database to extract the description and
the parts list of a given part using getDesc and getPartsList.
let rec partBreakDown id =
match getDesc id with
| (_,true) -> [(id,1)]
| _ ->
partsListBreakDown(Seq.toList(getPartsList id))
and partsListBreakDown = function
| (id,q)::pl -> let pl1 = mult q (partBreakDown id)
let pl2 = partsListBreakDown pl
mergePartsList pl1 pl2
| [] -> [];;
val partBreakDown : int -> (int * int) list
val partsListBreakDown : (int * int) list -> (int * int) list
partBreakDown 3;;
val it : (int * int) list = [(1, 19); (0, 20)]

partBreakDown 1;;
val it : (int * int) list = [(1, 1)]
 Creating a database
Executing the F# program in Figure 11.3 will setup the ProductRegister database with empty
Part and PartList tables.

Fig ure 11.3 F# program creating the tables of ProductRegister

 Updating a database
The type scheme contains service types and constructors for elements of the tables in the
database. For example
new schema.ServiceTypes.Part (PartId=id, PartName=s, IsBasic=b)

generates a new part object that can belong to the Part table.
Table objects like db.Part and db.PartsList have members InsertOnSubmit and
InsertAllOnSubmit that you can give a single row and a collection of rows, respectively, to
be inserted in the tables. These insertions are effectuated only when the function
SubmitChanges from the LINQ DataContext type has been applied.
Consider for example the following function that inserts a basic part into the Part table
given its part name:
let addBasic s =
let id = nextId()
let part = new schema.ServiceTypes.Part(PartId = id,
PartName = s,
IsBasic = true)
db.Part.InsertOnSubmit(part)
db.DataContext.SubmitChanges()
Some id;;
val addBasic : string -> int option

The function generates a key for the part and this key is returned by the function.
The insertion of a composite part into the database is based on its name s and its parts list:
[(id1, k1),… (idn, kn)]. Such an insertion is only meaningful when the identifiers idi, for 1 ≤i
≤n, are already defined in the Part table and when all the quantities ki, for 1 ≤ i ≤ n, are
positive integers. This well-formedness constraint is checked by the following function:
let isWellFormed pl =
List.forall (fun (id,k) -> containsPartId id && k>0) pl;;
val isWellFormed : (int * int) list -> bool

If this well-formedness constraint is satisfied, then the following function inserts a new
composite part into the Part table and its parts list into the PartsList table:
let addComposite s pl =
if isWellFormed pl
then
let id = nextId()
let part =
new schema.ServiceTypes.Part(PartId=id,
PartName=s,
IsBasic = false)
let partslist =
List.map
(fun (pid,k) ->
new schema.ServiceTypes.PartsList(PartsListId=id,
 PartId=pid,
Quantity=k))
pl
db.Part.InsertOnSubmit(part)
db.PartsList.InsertAllOnSubmit(partslist)
db.DataContext.SubmitChanges()
Some id
else None;;
val addComposite : string -> (int * int) list -> int option

The tables in Figure 11.1 are generated from an initial ProductRegister database with two
empty tables by evaluation of the following declarations:
let id0 = Option.get (addBasic "Part0");;
val id0 : int = 0

let id1 = Option.get (addBasic "Part1");;

val id1 : int = 1

let id2 =
Option.get (addComposite "Part2" [(id0,5);(id1,4)]);;
val id2 : int = 2

let id3 =
Option.get (addComposite "Part3" [(id1,3);(id2,4)]);;
val id3 : int = 3
 Summary
This chapter has introduced the notion of sequence, which is an ordered, possibly infinite,
collection of elements where the computation of elements is on demand only. Sequences are
convenient to use in applications where you are dealing with elements that are processed one
after they other. Functions from the sequence part of the collection library of F# have been
introduced together with cached sequences that prevents a recomputation of already computed
sequence elements. Furthermore, sequences can be defined in F# using sequence expressions
defining a step-by-step process for generating the elements.
The type seq<′a> is a synonym for the .NET type IEnumerable<′a> and any .NET
framework type that implements this interface can be used as a sequence. This has been
studied in connection with the Language-Integrated Query or LINQ component of the .NET
framework. LINQ gives query support for different kinds of data sources like SQL databases
and XML repositories. We have used LINQ in connection with a database for a simple
product-register application, where an F# type provider made it possible to work with values
from the external data sources (an SQL databases in this case) in a type safe manner. The
concept of query expressions was introduced since it gives powerful support for expressing
queries on SQL databases in F#.
 Exercise
11.1 Make a declaration for the sequence of odd numbers.
11.2 Make a declaration for the sequence of numbers 1, 1, 2, 6, …, n!, ….
11.3 Make a declaration for the sequence of seq[1; 1; 2; 6; …; n!; …], where the i+1′st
element is generated from the i′th element by multiplication with i + 1.
11.4 Declare a function that, for given i and n, selects the sublist [ai; ai+1; … ; ai+n−1] of a
sequence seq[a0; a1; …].

11.5 The declaration of the function iterate f on Page 260 has the drawback that fn x is
computed when the n′th element is demanded. Give an alternative declaration of this
function using the property that the n+1′st element of the sequence can be computed from the
n′th element by an application of f.
11.6 Have a look at the unfold function from the Seq library. Make a declaration of the
sRoot function from Section 11.5 using Seq.unfold. That declaration should be based on
the idea that the sequence generation is stopped when the desired tolerance is reached.
Measure the possible performance gains.
11.7 The exponential functions can be approximated using the Taylor′s series:
(11.2)

1. Declare a function that for a given x can generate the sequence of summands in (11.2).
Hint: Notice that the next summand can be generated from the previous one.
2. Declare a function that accumulates the elements of a sequence of floats. I.e. given a
sequence seq[x0; x1; x2; …] it generates the sequence seq[x0; x0 + x1; x0 + x1 + x2; …].
3. Declare a function to generate the sequence of approximations for the function ex on the
basis of (11.2).
4. Declare a function to approximate ex within a given tolerance.

11.8 The Madhava-Leibniz series (also called Gregory-Leibniz series) for π is:

Use this series to approximate π. (Note that there are other series for π, which converge
much faster than the above one.)
11.9 Declare a sequence denoting the following enumeration of the integers:

11.10 Use the functions in the Seq library to declare a function cartesian sqx sqy that gives
a sequence containing all pairs (x, y) where x is a member of sqx and y is a member of sqy.
Make an alternative declaration using sequence expressions.
11.11 Solve Exercise 11.3 using sequence expressions.
11.12 Solve Exercise 11.7 using sequence expressions.
11.13 Solve Exercise 11.8 using sequence expressions.
11.14 Solve Exercise 11.9 using sequence expressions.
11.15 Give a database-based solution to the cash-register example introduced in Section 4.6.
11.16 Give a database-based solution to Exercise 4.23.
 12
Computation expressions

A computation expression of F# provides the means to express a specific kind of

computations in a way where low-level details are hidden and only visible through the use of
special syntactic constructs like yield, let!, return, etc. These constructs are not part of the
normal F# syntax and are only allowed inside computation expressions. Each kind of
computation expression is defined by a computation builder object that contains the meaning
of the special constructs.
A computation expression ce belonging to the builder object comp appears in the F#
program in a construct of the form:
comp { ce }

This construct is an expression that evaluates to a value called a computation.

We have already seen examples of computation expressions in the form of sequence
expressions with builder object seq (cf. Table 11.2) and query expressions with builder object
query (cf. Section 11.8). The sequence computation expressions allow us to define
computations on sequences without having to bother about the laziness and other
implementation details, and the query expressions allow you to make database queries. In
Chapter 13 we introduce asynchronous computation expressions with builder object async
where you can define asynchronous computations without having to bother about low-level
details in the current state of the system. The seq, query and async computation expressions
are parts of F#.
This chapter describes the internals of computation expressions and how you may define
your own builder objects. This allow you to define new kinds of computations with special
state, flow of control and data management. The implementation of these special features is in
the builder object only, and the low-level implementation details are therefore invisible at the
computation-expression level.
The chapter explains and illustrates the machinery of computation expressions using a
number of examples of builder objects. The builder object mySeq explains the use and
meaning of the yield and for constructs using sequence expressions, while the maybe builder
object explains the use and meaning of the let! and return constructs by using them to
handle error cases in the evaluation of expressions represented by trees. This is followed by a
description of some fundamental properties of the For, Yield, Bind and Return builder
object methods that correspond to laws in the theory of monads – the mathematical foundation
of computation expressions. The chapter ends with a presentation of monadic parsers. They
provide a general method to construct programs capturing data from strings with a complex
syntactic structure.
 12.1 The agenda when defining your own computations
Defining a new kind of computation expressions comprises three parts:

Defining the computation type comp<′a>.

Defining the computation builder class CompBuilderClass.
Declaring the computation builder object comp.

This may look as follows:

The members comp.Bind and comp.Return provide the meaning of let! and return and
those constructs are therefore allowed in the computation expression ce in connection with
expressions of the form comp{ce}. An overview of the technical setting is given in Section
12.4, where Table 12.2 presents the main syntactical constructs and their translations to
composition of members from the associated builder object, while Table 12.3 gives the usual
types of the members.
The same name (meta symbol comp) is usually used for the type of computations as well as
for the builder object (an exception is the computation type Async<′a> with builder object
async). This double use of the name causes no conflict as the type and the builder object are
used in different contexts
By a computation we shall understand a value of type comp<′a> for some type ′a. The
pragmatics behind the development of new kinds of computations is that they are like recipes
in a cook book. A computation encapsulates pieces of programs but these pieces are only
executed if the computation is started. In the same way, a recipe is usually an operational
description of how to cook a dish; but the actual cooking is postponed until the recipe is
started (used) by a cook.
Evaluating a computation expression, for example:
let c = comp { ... }
is like editing a recipe. The special syntactic constructs in Table 12.2 build combined
computations in the same way as the pour “operator” builds a combined recipe in: “cook the
vegetables, make a béchamel sauce and pour the sauce on the vegetables.”
Each kind of computation expression has it own means for starting computations, for
instance:
seq<′a>: Operations starting computations are, for example, indexing sq.[n] that gives a
value of type ′a and the function Seq.toList that gives a value of type ′a list.
Async<′a>: An example is the function Async.Start starting the computation where the
execution in the normal case will cause some side-effects and eventually terminate with a
result of type ′a.
 12.2 Introducing computation expressions using sequence
expressions
In the introduction of sequence expressions in Section 11.6 it was shown that such
expressions are more convenient to use in certain cases than the functions of the Seq-library.
In this section we shall introduce the basic notions of computation expressions by showing
how user-defined sequence expressions can be implemented.
Consider the sequence expression:
seq {for i in seq [1 .. 3] do
for ch in seq [′a′ .. ′d′] do
yield (i,ch) };;

that expresses the sequence of pairs: (1,′a′), (1,′b′), …, (3,′d′). This sequence expression can be
considered a recipe to get this sequence since the evaluation of the declaration:
let pairRecipe = seq {for i in seq [1 .. 3] do
for ch in seq [′a′ .. ′d′] do
yield (i,ch) };;
val pairRecipe : seq<int * char>

will not cause any pair to be computed. Instead pairRecipe is bound to a computation (or a
recipe) that can be started (or cooked) at a later stage, for example, when the last element is
requested:
Seq.nth 11 pairRecipe;;
val it : int * char = (3, ′d′)

We will now show how the outer for-construct above

for i in seq {1 .. 3} do ce(i)

can be expressed using functions from the Seq library, and the obtained insight will be used in
the next section to implement our own computation expressions for sequences.
We denote the computation seq { ce(i) } corresponding to the body ce(i) by f i:
let f i = seq {for ch in seq [′a′ .. ′d′] do
yield (i,ch) };;
val f : ′a -> seq<′a * char>

and we get:
f 1;;
val it : seq<int*char> = seq [(1,′a′);(1,′b′);(1,′c′);(1,′d′)]
f 2;;
val it : seq<int*char> = seq [(2,′a′);(2,′b′);(2,′c′);(2,′d′)]
f 3;;
val it : seq<int*char> = seq [(3,′a′);(3,′b′);(3,′c′);(3,′d′)]
The sequence (1,′a′), (1,′b′), …, (3,′d′) of pairs denoted by pairRecipe is hence obtained by
appending the three sequences f 1, f 2 and f 3, and we get:
seq [(1,′a′); (...) ;(3,′d′)] = Seq.collect f (seq [1 .. 3])

where we have used the definition of Seq.collect in Table 11.1 on Page 254.
The meaning of the actual for construct can hence be expressed using Seq.Collect:
seq { for i in seq [1 .. 3] do ce(i) }
= Seq.collect f (seq [1 .. 3])

where:
f = fun i -> seq { ce(i) }
 12.3 The basic functions: For and Yield
A new kind of computation expressions can be declared in F# through the declaration of a
builder class, that implements a suitable selection of functions (cf. Section 12.1). We illustrate
this concepts by defining a builder class for sequences. This builder class contains member
functions that can perform operations on values of a parameterized type, that here is named
mySeq<′a>:

type mySeq<′a> = seq<′a>;;

To be able to make a computation expression corresponding to the pairRecipe example, the

builder class must provide implementations for the two functions:
For: mySeq<′a> * (′a-> mySeq<′b>) -> mySeq<′b>
Yield: ′a -> mySeq<′a>

where defines the meaning of the for construct and Yield defines the meaning of the
For
yield construct, in the sense of the translations shown in Table 12.1.

Construct: C Translation: T(C)

for x in e do ce For(e, fun x -> T(ce))
yield e Yield(e)

Table 12.1 Translations for for and yield

It was shown above that the for construct can be expressed using Seq.collect.
Furthermore, the function Yield, with the type ′a -> mySeq<′a> “lifts” an element to a
sequence and, therefore, Yield a returns the singleton sequence just containing a. Hence, we
arrive at the definitions:
(12.1)
(12.2)
The machinery is illustrated in Figure 12.1 on the pairRecipe example, where an ellipse
represents a sequence and a dashed box represent a sequence obtained by concatenating the
contained sequences. The figure illustrates the meaning of the for construct, where f is
applied to each element of i ∈ sq. Each application f(i) contribute with a part of the resulting
sequence where the result is obtained by concatenation of the sequences:
 Fig ure 12.1 An example illustrating the definition of For(sq, f)

The declaration of a builder class is based directly on (12.1) and (12.2):

type MySeqClass() =
member bld.Yield a: mySeq<′a> = Seq.singleton a
member bld.For(sq:mySeq<′a>, f:′a -> mySeq<′b>):mySeq<′b>
= Seq.collect f sq;;

and the builder object is obtained as follows:

let mySeq = MySeqClass();;

We can now make our own computation expressions (limited to the for and yield
constructs), for example, to declare a function that makes the Cartesian product of two
sequences sqx and sqy by constructing a sequence containing all pairs (x, y) where x is a
member of sqx and y is a member of sqy:
let cartesian sqx sqy = mySeq {for x in sqx do
for y in sqy do
yield (x,y) };;
val cartesian : mySeq<′a> -> mySeq<′b> -> mySeq<′a * ′b>

A declaration based on recursive functions or on the functions in the Seq library would not
have a comparable simplicity. Try, for example, Exercise 11.10.
Using the translation in Table 12.1, this declaration is, behind the scene, translated to:
let cartesian sqx sqy =
mySeq.For(sqx,
fun x -> mySeq.For(sqy,
fun y -> mySeq.Yield (x,y)));;

We shall not go into further details here concerning how mySeq can be extended to capture
more facilities of sequence expressions. Sequence expressions are handled in F# by a direct
translation to composition of functions from the Seq-library and not by the use of a builder
class. Further information can be found in the on-line documentation of F#.
Table 12.2 Translation of selected syntactic forms inside comp{…}
 12.4 The technical setting when defining your own
computations
Defining a builder class CompBuilderClass for a type comp<′a> of computations is like
defining a (rather simple) programming language. The syntactical constructs, like for and
yield, that are possible in expressions comp { ce } are given in advance; but the meaning of
the wanted constructs must be provided in the class CompBuilderClass of the builder object
comp. The meaning of an expression ce is given by a translation T(ce), as described in Table
12.1, to a composition of the member functions of the builder object comp.
A selection of possible syntactical constructs for computation expressions are enumerated
in Table 12.2. Meta symbols are used in this table as follows:

ce, ce1 and ce2 denote computation expressions occurring inside the brackets in an
expression comp { … }.
e denotes an expression. That expression may have the form comp′ { ce′ } and may be
subject to further translations using the builder object comp′.
pat denotes a pattern.

Usual types of the member functions are given in Table 12.3. The builder class
CompBuilderClass must contain declarations of member functions such that the translation
gives a correctly typed expression; but there are rather few restrictions beside that. One
restriction is that the builder class CompBuilderClass must contain either the members For
and Yield or the members Bind and Return (or all four member functions). These operations
are the fundamental ones and we shall study their properties in Section 12.9.

Table 12.3 Usual types for members of a builder object comp

 We shall not do any attempt to cover all these constructs, but subsequent sections will
describe Bind, Return, ReturnFrom, Zero and Delay.
The meaning of an expression comp { ce } is defined by a translation as follows:

Hence, the member function Delay provides an possibility to impose a delay from the very
start of a computation expression. We shall have a closer look at this in Section 12.8.
 12.5 Example: Expression evaluation with error handling
Consider the following type for expressions that are generated from integer constants and
variables, using operators for addition and integer division:
type Expr = | Num of int | Var of string
| Add of Expr*Expr | Div of Expr*Expr;;

An expression evaluates to a value in a given environment env: Map<string,int> that

associates values with variables. Errors may, however, occur due to a division by zero or in
the case that the environment does not give any value to a variable.
To avoid evaluations terminating by an exception we declare a function of type:
I: expr -> Map<string,int> -> option<int>

where

The function can be declared as follows:

let I e env =
let rec eval = function
| Num i -> Some i
| Var x -> Map.tryFind x env
| Add(e1,e2) -> match (eval e1, eval e2) with
| (Some v1, Some v2) -> Some(v1+v2)
| _ -> None
| Div(e1,e2) -> match (eval e1, eval e2) with
| (_ , Some 0) -> None
| (Some v1, Some v2) -> Some(v1/v2)
| _ -> None
eval e;;

Unfortunately, about half of this declaration addresses the manipulation of option values in a
rather un-elegant manner. The declaration of I is dominated by the error-handling and the
actual applications of the operators are almost put aside in a corner. We want to construct a
maybe<′a> computation expression with an associated builder class MaybeClass that takes
care of the Some/None distinguishing cases of the error handling.
 12.6 The basic functions: Bind, Return, ReturnFrom and Zero
We first present a simple version of the type maybe<′a> to introduce the main concepts
behind the members of the builder class. This first version will not allow delaying and
activation of computations, but these concepts are included in subsequent versions of the
builder class. A value of the type maybe<′a> is like a container possibly containing a value
and the type is declared as follows:
type maybe<′a> = option<′a>;;

where the value None denotes an error situation – the container is empty – and Some v denotes
that the container contains the value v.
It is not necessary to introduce maybe<′a> as a synonym for option<′a>; but in doing so
we are following the convention introduced in Section 12.1 that the type and the builder object
should have the same name.
We shall use expressions maybe { ce } where the constructs let!, return, return! and if
… then … (no else clause) are used in the computational expression ce. To do so the builder
class MaybeClass must contain declarations for the members:
Bind : maybe<′a> * (′a-> maybe<′b>) -> maybe<′b>
Return : ′a -> maybe<′a>
ReturnFrom: maybe<′a> -> maybe<′a>
Zero : unit -> maybe<′a>

where Bind defines the meaning of the let! construct, Return the meaning of the return
construct, ReturnFrom the meaning of the return! construct, and Zero defines the meaning
of the if-then construct in the sense of the translations shown in Table 12.4.

Construct: C Translation: T(C)

let! x = e
Bind(e, fun x -> T(ce))
ce
return e Return(e)
return! e ReturnFrom(e)
if e then ce if e then T(ce) else Zero()
Table 12.4 Translations for let!, return, return! and if-then

The operational readings of these constructs are as follows:

• The construct let! x = m in ce reads:
– Bind x to the value in the container m (if this value exists) and
– use this binding in ce.
• This construct is analogous to the for construct in the previous section.
 • The construct return e reads: put e in a container.
This construct is analogous to the yield construct in the previous section.
• The construct return! e reads: here is the container e.
Note that e must be a container.
• The value Zero() is used in the case of an empty else clause.
With this intuition and the fact that let! x = m in ce is translated to Bind(T(m), f) where f =
fun x -> T(ce), we arrive at the declaration:

type MaybeClass() =
member bld.Bind(m:maybe<′a>, f:′a->maybe<′b>):maybe<′b> =
match m with | None -> None
| Some a -> f a
member bld.Return a:maybe<′a> = Some a
member bld.ReturnFrom m:maybe<′a> = m
member bld.Zero():maybe<′a> = None;;
let maybe = MaybeClass();;

By use of expressions maybe { ce } the handling of error situations is managed by the Bind
function and not in the handling of dyadic operators and divisions:
let I e env =
let rec eval = function
| Num i -> maybe {return i}
| Var x -> maybe {return! Map.tryFind x env}
| Add(e1,e2) -> maybe {let! v1 = eval e1
let! v2 = eval e2
return v1+v2}
| Div(e1,e2) -> maybe {let! v2 = eval e2
if v2<>0 then
let! v1 = eval e1
return v1/v2}
eval e;;
val I : expr -> Map<string,int> -> maybe<int>

Observe that the tags None and Some are absent from this program and that the declarations
focus just on the computation of the value of an expression.
For example:
let e1 = Add(Div(Num 1, Num 0), Num 2);;
let e2 = Add(Add(Var "x", Var "y"), Num 2);;

let env = Map.ofList [("x",1);("y",2)];;

let v1 = I e1 env;;
val v1 : maybe<int> = None

let v2 = I e2 env;;
val v2 : maybe<int> = Some 5

The examples show that the maybe computation expressions eagerly evaluate e1 and e2 to
values of type maybe<int>. Hence the computation expressions for this simple version of the
class MaybeClass are not real recipes, they actually correspond to cooked dishes. This
deficiency is repaired in the next section.
 12.7 Controlling the computations: Delay and Start
The evaluation of an expression e can be delayed by “packing” it into a closure:
fun () -> e
as we already have seen in Section 11.3. For example:
let c1 = fun () -> 1+2;;
val c1 : unit -> int

The addition operation is started by a function application:

c1();;
val it : int = 3

We use another definition of the type maybe<′a>:

type maybe<′a> = unit -> option<′a>;;

in order to be able to control the delay and activation of maybe computation expressions. A
value of this type is a recipe in the shape of a closure, and the delaying of computations is
“hard coded” into the type. Such recipes must be started explicitly when the value of the
computation is asked for. The following functions for delay and start will be used:
let delay v = fun () -> v;;
val delay : ′a -> unit -> ′a

let start m = m();;

val start : (unit -> ′a) -> ′a

The builder class MaybeClass is revised to get another meaning of the let! construct. The
new definition of Bind captures that the construct
let! x = m in ce
matches the following operational reading:

1. Start the computation m.

2. Bind x to the value a of this computation if it terminates properly with a in the container.
3. Use this binding in the recipe ce.

Notice that the let! construct translates to Bind(m, f) where f is fun x -> T(ce).
These considerations lead to the first revised version of the builder class:
type MaybeClass() =
member bld.Bind(m:maybe<′a>, f:′a->maybe<′b>):maybe<′b> =
match start m with
| None -> delay None
 | Some a -> f a
member bld.Return a:maybe<′a> = delay(Some a)
member bld.ReturnFrom v:maybe<′a> = delay v
member bld.Zero():maybe<′a> = delay None;;

let maybe = MaybeClass();;

The declaration for Bind matches the operational reading of the let! construct. Delays of the
option-values are needed to lift these values to the type maybe.
Notice that the declaration of I e env can be based on this revised maybe builder without any
change. Doing so gives controlled computations in the sense of recipes:
let v1 = I e1 env;;
val v1 : maybe<int>

let v2 = I e2 env;;
val v2 : maybe<int>

The recipes v1 and v2 must be started to get values computed:

start v1;;
val it : int option = None

start v2;;
val it : int option = Some 5

Since an expression like maybe { let! x = m ... } translates to Bind(m,...) the

computation m will actually be started (check the declaration of Bind) and the values v1 and v2
contain in this sense partly cooked ingredients. This can be observed if side effects are
introduced into, for example, in the clause where addition is treated:
...
| Add(e1,e2) ->
maybe {let! v1 = eval e1
let! v2 = eval e2
return (printfn "v1: %i v2: %i" v1 v2 ; v1+v2)}
...

The result of executing the following declarations with this version of maybe:
let v2 = I e2 env;;
v1: 1 v2: 2
v1: 3 v2: 2
val v2 : maybe<int>

start v2;;
val it : int option = Some 5

shows that the computation is started and active until the outermost return or return!
statement is reached.
 12.8 The basic function: Delay
The builder class CompBuilderClass of a type comp<′a> with builder object comp may
contain a member:
Delay: (unit -> comp<′a>) -> comp<′a>

The translation of an expression comp { ce } will then use this delay function in the translation
of a computation expression ce:

This gives a possibility to enforce a delay from the very start of a computation expression:
We add the following declaration of Delay to the MaybeClass declaration in the previous
section:
type MaybeClass() =
... As above from Bind to Zero ...
member bld.Delay f:maybe<′a> = fun () -> start (f());;

The effect of this can be observed using the above “side-effect example,” where the printing
of the two lines with values to be added move from the declaration of v2 to its activation:
let v2 = I e2 env;;
val v2 : maybe<′a>

start v2;;
v1: 1 v2: 2
v1: 3 v2: 2
val it : int option = Some 5

Hence, with the introduction of the Delay member in the class declaration, the expressions of
the form maybe { ce } will denote genuine recipes. These recipes are expressed in an
operational manner by describing how to cook the dish; but the actual cooking is delayed until
the recipe is started.

Fig ure 12.2 An illustration of the law: For(Yield a, f) = f(a)

Fig ure 12.3 An illustration of the law: For(sq, Yield) = sq
 12.9 The fundamental properties of For and Yield, Bind and
Return
When declaring builder classes like MySeqClass and MaybeClass the only restriction
imposed on For and Yield and Bind and Return is that they should have the correct types. But
there are some laws that meaningful implementations should obey. These laws originate from
the theory of monads for functional programming, a theory that provides the mathematical
foundation for computation expressions.
The intuition behind these laws will be presented using the builder class for mySeq<′a> as
example, where values of this type are considered as containers for values of type ′a and
computation expressions are recipes for filling containers. But the laws are not biased
towards the mySeq computation expression builder.
 The laws for For and Yield
The first law expresses that Yield is a kind of left unit for For:
(12.3)
Hence binding f to the element of the container yielded by a is the same as applying f to a.
This is illustrated in Figure 12.2.
The second law expresses that Yield is a right unit for For:
(12.4)
Hence yielding the elements of a container sq equals that container. This is illustrated in
Figure 12.3.

Fig ure 12.4 The law: For(For(sq, f), g) = For(sq, fun a -> For(f(a), g))

The last law expresses a kind of associativity of For:

(12.5)
This law is explained in terms of the for construct. Observe first that the computation
expression for a in sq do f(a) translates as follows:

Using this technique we arrive at the following alternative formulation of the law:
 The law expresses two ways of filling a container. The left-hand-side way is by filling it
using g(b) where b is in the container obtained from for a in sq do f(a)). The right-hand-side
way is by filling it using g(bi j) where bi j is in the container obtained from f(ai), where ai is in
the container sq. This is illustrated in Figure 12.4.
 The laws for Bind and Return
We shall now have a closer look at the resemblance between Bind and Return and For and
Yield, respectively. Actually the types for For and Yield may be considered special cases of
the type of Bind and Return, and it is possible to define computation expressions for
sequences that make use of let! and return rather than for and yield.
To illustrate this, consider the following (not recommendable) declarations:
type myStrangeSeq<′a> = seq<′a>;;

type MyStrangeSeqClass() =
member bld.Return a: myStrangeSeq<′a> =
Seq.singleton a
member bld.Bind(sqs, f):myStrangeSeq<′b> =
Seq.collect f sqs;;

let myStrangeSeq = MyStrangeSeqClass();;

The pairRecipe example can now be given in the following way:

let pairRecipe = myStrangeSeq {let! i = seq {1..3}
let! ch = seq {′a′..′d′}
return (i,ch)};;
val pairRecipe : myStrangeSeq<int * char>

Seq.toList pairRecipe;;
val it : (int * char) list =
[(1, ′a′); (1, ′b′); (1, ′c′); (1, ′d′);
(2, ′a′); (2, ′b′); (2, ′c′); (2, ′d′);
(3, ′a′); (3, ′b′); (3, ′c′); (3, ′d′)]

Therefore, the fundamental laws Bind and Return are the same as those for For and Yield:
(12.5)
(12.5)
(12.8)
It is left as an exercise to justify that these properties hold for the maybe example.
 12.10 Monadic parsers
This section is a continuation of Section 10.2 where regular expressions were used in
capturing data from text lines. Functions using match on regular expressions with capturing
groups give an efficient capture of data with a simple structure. The capability of such
functions is, however, limited and we had, for example, to use ad hoc tricks in Section 10.2 to
capture data with a nested list structure. Data with a recursive structure appearing, for
example, in representing expressions to be input by the program cannot be handled in this
way.
Techniques to construct parsers to solve such a problem are well-known in compiler
technology (as described, for example, in [2]). In this section we use a technique known as
monadic parsing that has been developed by the “Haskell community” (cf. [15, 14, 7]). It gives
a simple construction of parsers that is well-suited for small-scaled parsing – and it gives an
interesting example of computation expressions. Large-scaled parsing used for instance in
compilers is made using a parser generator (cf. [12, 13]). The following presentation of
monadic parsers in F# is based on the Haskell implementation described in [7].
The first step in constructing a parser is to make a grammar describing the structure of the
input data. This comprises making regular expressions to describe the tokens, i.e. the smallest
pieces of information: names, numbers, operators and delimiters to be captured by token
parsers constructed from the regular expressions. The combination of the token parsers to get
a parser is then based directly on the grammar. We define a number of parser combinators to
be used in this construction. The definition of the parser will usually consist of a set of
mutually recursive declarations.
We use two examples to illustrate the technique: the simple example of person data from
Section 10.2 and the more complicated example of algebraic expressions. In the first example
we may, for example, get an input line of the form:
John 35 2 Sophie 27 Richard 17 89 3

and the captured value should then be:

[("John", [35;2]); ("Sophie", [27]); ("Richard", [17;89;3])]

of type
(string * (int list)) list

In the second example we have algebraic expressions like the string:

-a1 + 2 * (a2 - 3)

We want to capture this string as the value:

Add (Neg (Var "a1"), Mul (Num 2, Sub (Var "a2", Num 3)))
of type
type Expr = | Num of int | Var of string
| Neg of Expr | Add of Expr * Expr
| Sub of Expr * Expr | Mul of Expr * Expr;;

The captured value corresponds to the expression tree in Figure 12.5. This example has a
number of interesting features beside the recursion: two levels of operator precedence
(multiplication and addition operators), a precedence level with two operators (+ and -), and
use of the same operator symbol (-) with two different meanings as prefix and infix operator.

Fig ure 12.5 Expression tree for expression -a1 + 2 * (a2 - 3)

 Grammars
In the first example we have two tokens name and number with regular expressions:
open System.Text.RegularExpressions;;
let nameReg = Regex @"\G\s*([a-zA-Z][a-zA-Z0-9]*)";;
let numberReg = Regex @"\G\s*([0-9]+)";;

where the tokens name and number denote the set of strings matching the corresponding
regular expressions nameReg and numberReg.
We shall capture structured data through rules described by context-free grammars. For the
person data, the rules should capture strings with the wanted syntax:

A personData string consists of a series of zero or more person strings.

A person string consists of a name followed by a series of zero or more numbers.

The corresponding context-free grammar looks as follows:

This grammar has the following components:

Four non-terminal symbols: personData, personList, person and numberList.

Two tokens: name and number.
Four rules. Each rule has a non-terminal at the left-hand side and a definition at the right-
hand side. A definition may have one or more choices separated by | and each choice
consists of a possibly empty sequence of tokens and non-terminal symbols. The Greek
letter Λ is used to denote the empty sequence in grammars.

In order to be able to distinguish tokens from non-terminal symbols in a grammar, we write a

token in teletype font (like in name and number) while a non-terminal symbol is written in
roman font (like in person and numberList).
Each non-terminal symbol denotes a syntax class that is the set of all strings that can be
generated from that non-terminal symbol. A string is generated from a non-terminal symbol
by a derivation that repeatedly replaces a non-terminal symbol by a choice in its definition or
a token by a matching string. The derivation terminates when there are no more non-terminals
or tokens to be substituted.
For example, the string Peter 5 John belongs to the syntax class personData due to the
derivation:
While a derivation generates a string from a non-terminal symbol, we are interested in the
parsing of a string, that is, the creation of a derivation on the basis of a given string. The
technique for monadic parsing to be presented will make this derivation on the basis of
recursive definitions following the structure of the grammar, and for this approach to be
well-defined the grammar must satisfy that there is no derivation of the form:

where N is a non-terminal symbol and w is a sequence of tokens, non-terminal symbols and

strings. In particular, there should be no left-recursive rule of the form:

in the grammar. See Exercises 12.3 and 12.4.

We shall use grammars written in EBNF notation (extended Backus Naur form) allowing, for
example, use of the repetition operator * that was introduced in connection with regular
expressions (cf. Section10.2). Using the EBNF notation the above grammar gets a more
compact form:

In the second example with expressions we have tokens num, var, addOp, mulOp, sign,
leftPar, rightPar, and eos where eos denotes end of string. The corresponding regular
expressions are as follows:
let numReg = Regex @"\G\s*((?:\053|-|)\s*[0-9]+)";;
let varReg = Regex @"\G\s*([a-zA-z][a-zA-Z0-9]*)";;
let plusMinReg = Regex @"\G\s*(\053|\055)";;
let addOpReg = plusMinReg;;
let signReg = plusMinReg;;
let mulOpReg = Regex @"\G\s*(\052)";;
let leftParReg = Regex @"\G\s*(\050)";;
let rightParReg = Regex @"\G\s*(\051)";;
let eosReg = Regex @"\G(\s*)$";;

The cautious reader will observe that the addOp and sign tokens have the same regular
expressions. We will comment on this in the subsection on token parsers.
We choose syntax classes expr, term and factor to express the composite forms in
expressions and we get at the following grammar for expressions, where the rule for factor
has a number of different choices:

Building a grammar of this kind requires careful considerations of the following syntactic
issues:

Precedence levels of operators.

Left or right association of operators.

The example with expressions has two levels of precedence of operators:

1. Multiplication operator *.
2. Addition operators + and -.

while all operators associate to the left. The precedence rules of operators are captured in the
grammar:

A factor can be used as a term.

A term cannot be used as a factor – it has to be viewed as an expression and then enclosed
in parentheses in order to get a factor.

The associations of the operators are capture in the structure of the constructs expressing
repetitions, and the rule for term would, for example, have been:

if the multiplication operator should associate to the right.

The original grammars (in the Algol 60 report) used left recursion in the grammar of
expressions, like in the following:

The left recursion appears in the rule where expr can be expanded to expr addOp term and
similar for term. This grammar has the (theoretical) advantage that the steps in the derivation
of an expression correspond to the steps in building the expression tree. Such grammars can,
however, not be used in our method because the corresponding parser will enter an infinite
loop. See Exercise 12.4.
 Parsers
Each character in the input string is identified by its position that is an non-negative integer. A
parser with result type ′a scans the string searching for matches starting a specified start
position pos. A match is a pair (a,pos′) of a captured value a of type ′a and the end position
pos′ where a succeeding parsing may take over. The parser is hence “consuming” the
characters from position pos up to (but not including) the end position pos′ in producing the
result a. The collection of all possible matches starting at a specified position can hence be
represented by a list:

This corresponds to the following type of a parser with result type ′a:
type parser<′a> = string -> int -> (′a * int) list;;

Note that we allow several possible matches. This is not a complication – it is actually a key
feature in monadic parsers. An empty list indicates that the parser has failed to find any
matches. Suppose, for example, that we have a parser expr for algebraic expressions. Parsing
the input string "-a1 + 2 * (a2 - 3)" from position 0, using the expression expr "-a1 +
2 * (a2 - 3)" 0, should then give a list with three matches:

[(Neg (Var "a1"), 3);

(Add (Neg (Var "a1"),Num 2), 7);
(Add (Neg (Var "a1"),Mul (Num 2,Sub (Var "a2",Num 3))),18)]

Position 3 is just after "-a1", position 7 is just after "2" while position 18 is at the end of the
string.
 Token parsers
Tokens are parsed using token parsers. We consider two kinds of token parsers:

1. A token parser with captured data (normally to be converted).

2. A token parser without relevant captured data.

A token parser of the first kind is made using the token function. The regular expression reg
must contain a capturing group. The function conv converts the captured data:
open TextProcessing;;

let token (reg: Regex) (conv: string -> ′a) : parser<′a> =

fun str pos ->
let ma = reg.Match(str,pos)

match ma.Success with

| false -> []
| _ ->
let pos2 = pos + ma.Length
[( conv(captureSingle ma 1), pos2)];;
val token : (Regex -> (string->′a) -> parser<′a>) = <fun:...>

Token parsers without captured data are made using the emptyToken function. The regular
expression need not contain any capturing group and there are no conversion function. The
parser captures the dummy value () of type unit and its function is solely to recognize and
“consume” the data matching the regular expression:
let emptyToken (reg: Regex) : parser<unit> =
fun str pos ->
let ma = reg.Match(str,pos)
match ma.Success with
| false -> []
| _ -> let pos2 = pos + ma.Length
[( (), pos2)];;
val emptyToken : (Regex -> parser<unit>) = <fun:clo...>

Note that the function captureSingle from the TextProcessing library (cf. Table 10.4 and
Appendix B) is used in the above declarations.
 Token parsers in the examples
We declare token parsers name and number in the first example using the corresponding
regular expressions:
let name = token nameReg id;;
let number = token numberReg int;;

The conversion function is the pre-defined identity function id for the name token parser
because the captured string should be used literally “as is”. The number token parser uses the
conversion function int to convert the captured string of digits to an integer. The token
parsers num, var, sign, addOp, mulOp, leftPar and rightPar in the second example should
give values that can be used directly in building the expression tree (like the tree shown in
Figure 12.5). The token parser addOp should hence capture a “value” that can be used to join
two sub-trees, for example:
fun x y -> Add(x,y)

of type:
Expr -> Expr -> Expr

when parsing the character ’+’. The addOp token parser will hence be of type:
parser<Expr->Expr->Expr>

These token parsers use the following conversion functions:

let numFct (str: string) = Num (int str);;
let varFct = Var;;
let addOpFct = function
| "+" -> fun x y -> Add(x,y)
| _ -> fun x y -> Sub(x,y);;
let mulOpFct _ = fun x y -> Mul(x,y);;
let signFct = function
| "+" -> id
| _ -> fun x -> Neg x;;

and their declarations are as follows:

let num = token numReg numFct;;
let var = token varReg varFct;;
let addOp = token addOpReg addOpFct;;
let mulOp = token mulOpReg mulOpFct;;
let sign = token signReg signFct;;
let leftPar = emptyToken leftParReg;;
let rightPar = emptyToken rightParReg;;
let eos = emptyToken eosReg;;

The cautious reader will observe that the token parsers addOp and sign parse the same strings
– with different captures. This works in monadic parsing because the parsing is strictly top-
down: The expr parser (to be constructed later) acts according to the context and calls the
addOp token parser when an addition operator may occur – and the sign token parser when a
sign change operator may occur.
 Computation expressions for building parsers
The computation expressions aim at simplifying the construction of a parser by hiding all the
technicalities concerning the character positions in the input string. The key point in defining
the parser computation expressions is to define the Bind member in the builder class that
provides the meaning to the computation expression:
let! a = p
ce(a)
where p:parser<′a> is a parser giving parses of type ′a and the computation expression
ce(a) has type parser<′b>. The construct is translated to Bind(p, f), where f is fun a ->
T(ce(a)) using the translation T described in Section 12.4.
The operational reading of this construct is:

1. Start the parser p,

2. bind a to a result a of the parser p, and
3. use this binding in the computation expression ce(a).

The parser p is activated by a function application p str pos, where str is the input string and
pos is a position. This resembles the activation of a maybe value on Page 288. This activation
of p gives a list of pairs [ (a0, pos0) , … , (an, posn)], where ai is a captured value for the part
of the input string that starts at position pos and ends at posi - 1. Application of f to ai yields a
parser that must be activated to the input string str and the start position posi for that parser.
This resembles the definition of the for construct for sequences, see Figure 12.1. This
complete process is illustrated in Figure 12.6. Notice that the Bind function takes care of all
the data management concerning the positions.

Fig ure 12.6 Illustrating Bind(p, f) = fun str pos -> collect g f,str (p str pos)

These ideas lead to the following computation expression class and builder object:
type ParserClass() =
 member t.Bind(p: parser<′a>, f: ′a->parser<′b>):parser<′b> =
fun str pos ->
List.collect (fun (a,apos) -> f a str apos) (p str pos)
member bld.Zero() = (fun _ _ -> []): parser<′a>
member bld.Return a = (fun str pos -> [(a,pos)]): parser<′a>
member bld.ReturnFrom (p: parser<′a>) = p;;

let parser = ParserClass();;

The zero result bld.Zero() is the parser without any matches.

The “constant” parser (corresponding to the right-hand side of bld.Return):
parser { return a }

“captures” the value a without consuming any characters, that is, it gives the value a at the end
position of the previously used parser.
When a string cannot be parsed the final result is the empty list; but no informative error
report is handled by the builder class. In Exercise 12.8 you are asked make new builder class
for parsers that takes care of a simple error handling.
 Sequencing of parsers. Parsers for fixed forms
Assume that we have n parsers

and a function:

for some types ′a1 , ′a2 , … , ′an , ′b.

Any match of the sequenced parser starting from position pos is then obtained by getting a
sequence of contiguous matches (starting from position pos):

of the parser p1, p2, … , pn and applying the function F to the captured values to get a match of
the sequenced parser (starting from position pos):

Using computation expression the sequenced parser can be written:

The return expression is inserted at a place where activation of the parsers p1, p2, … , pn
have already consumed all relevant characters and where it only remains to return the value
F(a1,a2, … ,an) without consuming any further characters.
Sequencing of parsers is used when building parsers for fixed forms (containing no part
that is repeated an unspecified number of times). The simplest examples are parsers built
using the pairOf combinator:
let pairOf p1 p2 = parser {let! x1 = p1
let! x2 = p2
return (x1,x2)};;
val pairOf : (parser<′a> -> parser<′b> -> parser<′a*′b>) = ...

We may, for instance, combine the name and number token parsers using pairOf:

let nameNumber = pairOf name number;;

val nameNumber : parser<string * int>

nameNumber " abc 473 " 0;;

val it : ((string * int) * int) list = [(("abc", 473), 8)]

In building a parser in the first example we will use the pairOf parser combinator to combine
parsers of a name and of a list of numbers.
 One may define a tripleOf combinator in a similar way, but it is not of much use as most
grammars require specially built parsers for their sequencing constructs. In the expression
example we have, for instance, the form of an expression enclosed in parentheses. A
simplified version of this is a form with a variable enclosed in parentheses like:
( abc)

A parser for this form can be obtained by sequencing the token parsers leftPar, var and
rightPar:

let varInPars = parser {let! _ = leftPar

let! x = var
let! _ = rightPar
return x };;
val varInPars : parser<expr>

varInPars "( abc) " 0;;

val it : (expr * int) list = [(Var "abc", 6)]

The recursive declaration of the expr parser given later in this section comprises a
sequencing of the parsers leftPar, expr and rightPar.
 The choice combinator
The choice combinator <|> defines the choice parser:
p1 <|> p2

for two parsers of same type:

p1; parser<′a>
p2: parser<′a>

The set of matches of p1 <|> p2 is simply the union of the set of matches of p1 and the set of
matches of p2.
This is captured in the following declaration where the lists of matches of p2 is appended to
the list of matches of p1:
let (<|>) (p1: parser<′a>) (p2: parser<′a>) =
(fun str pos -> (p1 str pos) @ (p2 str pos)): parser<′a>;;
val ( <|> ) : parser<′a> -> parser<′a> -> parser<′a>

We may for instance make a parser capturing a variable or a number:

let numOrVar = num <|> var;;
val numOrVar : parser<expr>

numOrVar "ab 35" 0;;

val it : (expr * int) list = [(Var "ab", 2)]
numOrVar "ab 35" 2;;
val it : (expr * int) list = [(Num 35, 5)]
 Repetitive constructs. Combinators listOf, infixL and infixR
The listOf combinator is used when parsing lists where the parser p captures a single list
element:
let rec listOf p = parser {return []}
<|> parser {let! x = p
let! xs = listOf p
return x::xs};;
val listOf : parser<′a> -> parser<′a list>

The parser listOf p will either return an empty list:

parser {return []}

or parse one list element:

let! x = p

and put this element in front of the remaining list and return the result:
let! xs = listOf p
return x::xs

The parser listOf number will for instance parse lists of numbers:
(listOf number) " 3 5 7 " 0;;
val it : (int list * int) list =
[([], 0); ([3], 2); ([3; 5], 4); ([3; 5; 7], 6)]

The infixL combinator is used when building a parser for a syntactic form where an
arbitrary number of operands are intermixed with infix operators that are on the same
precedence level and associates to the left.
As an example we consider strings like
a - b + 5

where numbers or variables are intermixed with addition operators (+ or -). A parser for this
form can be obtained using the below defined infixL operator:
let psL = numOrVar |> infixL addOp numOrVar;;
val psL : parser<expr>

psL "a - b + 5" 0;;

val it : (expr * int) list =
[(Var "a", 1); (Sub (Var "a",Var "b"), 5);
(Add (Sub (Var "a",Var "b"),Num 5), 9)]

The three matches in the result correspond to the strings:

Note that the last expression tree:
Add(Sub(Var "a", Var "b"), Num 5)

of the full string "a - b + 5" reflects the left association of the operators: first we subtract b
from a and afterwards we add 5 to the result.
The infixL combinator is defined in a slightly more general setting with two operand
parsers p and q:
p : parser<′a>
q : parser<′b>

and an operator parser op:

op: parser<′a -> ′b -> ′a>

corresponding to strings with one operand astr matching p, n operators op1 , … , opn
matching op, and n operands bstr1, … , bstrn matching q:

The parser p |> infixL op q should have n + 1 matches on this string corresponding to a
sequence of matches of p, op and q:

The recursive pattern in these captures leads to the declaration:

let rec infixL op q =
fun p ->
p <|>
parser { let! a = p
let! f1 = op
let! b1 = q
let a1 = f1 a b1
let p1 = parser { return a1 }
return! p1 |> infixL op q } ;;
val infixL : parser<(′a -> ′b -> ′a)> -> parser<′b>
-> parser<′a> -> parser<′a>

The infixR combinator is declared as follows:

let rec infixR op q = fun p ->
q <|>
parser { let! a = p
let! f = op
let! b = p |> infixR op q
return f a b } ;;
val infixR : parser<(′a -> ′b -> ′b)> -> parser<′b>
-> parser<′a> -> parser<′b>

It is similar to infixL but it builds a tree corresponding to the evaluation of operators

associating to the right. Using infixR in the above example of addition operators would
hence give other expression trees:
let psR = numOrVar |> infixR addOp numOrVar ;;
val psR : parser<expr>

psR "a - b + 5" 0;;

val it : (expr * int) list =
[(Var "a", 1); (Sub (Var "a",Var "b"), 5);
(Sub (Var "a",Add (Var "b",Num 5)), 9)]

where Sub(Var "a",Add(Var "b",Num 5)) would define an evaluation of

a - b + 5

such that 5 is first added to b and the result then afterwards subtracted from a.
 Making parsers
The parsers in the examples are based directly on the token parsers and the EBNF grammar:

A parser is defined for each syntax class.

Each operator in a syntactic rule in the grammar is translated to a suitable parser
combinator.

One should, however, pay attention to the word “suitable”: The parser combinators should not
only correspond to the syntax but they must also give the right conversion of the textual form
to captured value. You will frequently have to write your own parsers for fixed sequence
constructs (like leftPar expr rightPar in the second example) but is it a good idea to try to
design the syntax and the structure (that is, type) of the captured value such that repetitive
constructs can be handled using the above parser combinators.
Making the parser in the first example is almost straightforward:
let person = pairOf name (listOf number);;
val person : parser<string * int list>

let personData = listOf person;;

val personData : parser<(string * int list) list>

personData "John 35 2 Sophie 27 Richard 17 89 3" 0;;

val it : ((string * int list) list * int) list =
[([], 0);
([("John", [])], 4);
([("John", [35])], 7);
...
([("John", [35; 2]); ("Sophie", [27]);
("Richard", [17; 89; 3])], 35)]

The example with expressions is more complex but the required parser combinators have
been introduced above, so the grammar can be translated directly into a monadic parser:
let rec expr = term |> infixL addOp term
and term = factor |> infixL mulOp factor
and factor = num <|> var
<|> parser {let! f = sign
let! x = factor
return (f x)}
<|> parser {let! _ = leftPar
let! x = expr
let! _ = rightPar
return x};;
val expr : parser<expr>
val term : parser<expr>
val factor : parser<expr>

The F# compiler issues a warning telling that the above system of mutually recursive
declarations may contain cyclic definitions – but the declaration is, nevertheless, accepted by
the compiler.
Applying the parser expr to the sample string gives the wanted result:
expr "-a1 + 2 * (a2 - 3)" 0;;
val it : (Expr * int) list =
[(Neg (Var "a1"), 3); (Add (Neg (Var "a1"),Num 2), 7);
(Add(Neg(Var "a1"),Mul(Num 2,Sub(Var "a2",Num 3))), 18)]
 Parsing the full string
The personData and expr parsers deliver a match for each matching sub-string. Parsers
matching only the full string are made using the eos token parser that matches end-of-string
(possibly preceded by blank characters):
let personDataString = parser {let! dt = personData
let! _ = eos
return dt };;
val personDataString : parser<(string * int list) list>
personDataString "John 35 2 Sophie 27 Richard 17 89 3" 0;;
val it : ((string * int list) list * int) list =
[([("John", [35; 2]); ("Sophie", [27]);
("Richard", [17; 89; 3])], 35)]

let exprString = parser { let! ex = expr

let! _ = eos
return ex };;
val exprString : parser<expr>

exprString "-a1 + 2 * (a2 - 3)" 0;;

val it : (Expr * int) list =
[(Add(Neg(Var "a1"),Mul(Num 2,Sub(Var "a2",Num 3))), 18)]
 Reporting errors
A simple error reporting can be obtained by letting the token parsers update a global variable
maxPos. The declarations of token and emptyToken are then preceded by

let mutable maxPos = 0

let updateMaxPos pos = if pos > maxPos then maxPos <- pos;;

and an extra line is added to the token function

let token (reg: Regex) (conv: string -> ′a) : parser<′a> =
fun str pos ->
let ma = reg.Match(str,pos)

match ma.Success with

| false -> []
| _ ->
let pos2 = pos + ma.Length
updateMaxPos pos2
[( conv(captureSingle ma 1), pos2)];;

and similarly for emptyToken.

Using this set-up we introduce the type ParseResult<′a>
type ParseResult<′a> = ParseOk of ′a | ParseError of int;;

in order to report an error when an input string cannot be parsed. In the case of such an error,
the global variable maxPos identifies the position where the error was detected and this
position is reported:
let parseString (p: parser<′a>) (s: string) =
maxPos <- 0
match p s 0 with
| (a,_)::_ -> ParseOk a
| _ -> ParseError maxPos;;
val parseString : parser<′a> -> string -> ParseResult<′a>)

parseString exprString "a - b + c";;

val it : ParseResult<expr> =
ParseOk (Add (Sub (Var "a",Var "b"),Var "c"))

parseString exprString "a - b * (1 + c" ;;

val it : ParseResult<expr> = ParseError 14

where the error in the last case was found at position 14 in the string.
In Exercise 12.8 you are asked to hide the error handling in the builder class for parsers.
 Summary
This chapter has introduced the notion of computation expressions of F#. Computation
expressions offer a possibility for using special syntactic constructs like let!, return, etc.
with a user-defined meaning through the declaration of so-called builder classes. This concept
is based on the theory of monads for functional programming introduced in connection with
the Haskell programming language.
The chapter uses sequence expressions (introduced in Chapter 11) and error handling in
connection with expression evaluation as examples to show how you may define your own
computation expressions. The last section shows how parsers can be constructed in a
convenient manner using computation expressions.
Asynchronous computations that will be introduced in Section 13.4 is an important example
of computation expressions.
 Exercise
12.1 Consider the following “alternative” to the declaration for bld.Delay on Page 290:
type MaybeClass() =
...
member bld.Delay f:maybe<′a> = delay(start (f()));;

This new declaration would not give the desired effect. Explain why.
12.2 Consider the expression evaluation on Page 287. Make a new class declaration for
computation expressions that takes care of the evaluation in the environment env and
simplify the declaration of the function I accordingly. Hint: Consider computations as
functions having the type: Map<string,′a> -> option<′a>.
12.3 The following grammar for non-empty lists of numbers uses left recursion:

A parser strictly following the structure of this grammar:

let rec numberLst = parser {let! n = number
return [n] }
<|>
parser {let! ns = numberLst
let! n = number
return ns @ [n]};;

has a problem. Analyze the parser and explain what the problem is.
12.4 Explain the problem with the grammar for expressions on Page 297 that uses left
recursion.
12.5 Consider the formulas of propositional logic introduced in Exercise 6.7. In the string
representation of such formulas conjunction ∧ can be written either as & or as and,
disjunction either as | or as or and negation either as ! or as not. For example, the
formula

has several string representations. Two of them are:

Write a parser for such formulas that takes care of:

conjunction and disjunction associates to the left,

conjunction has higher precedence that disjunction, and
negation has highest precedence.

12.6 Consider the dating bureau in Exercise 4.23. Make a string representation of the file
of the dating bureau and a parser that can convert such strings into the representation of
the file used in your solution to Exercise 4.23.
12.7 Declare a parser combinator:
pFold : (′a -> ′d -> ′a) ->
parser<′d> -> parser<′a> -> parser<′a>

such that pFold f t p captures the values

if p first captures the value a and repeated use of t afterwards captures the values d1, d2,
…, dk. Use this parse combinator to make an alternative declaration of infixL of the
form:
let infixL op q p = pFold (fun ... ) (pairOf op q) p ;;

12.8 The report of errors can be hidden in builder class for parsers and in this exercise
you shall make a new version of parsers based on the type declarations:
type ParseResult<′a> = ParseOk of ′a | ParseError of int
type parser<′a> = string -> int -> ParseResult<(′a*int) list>

The builder class should not make use of any mutable variable like maxPos, see Page 308,
and a value ParseError n should occur when an error is discovered at position n.

Make a new version of the builder class using the above type declarations.
Revise the two functions token and emptyToken for generating token parsers
accordingly.
Revise the declaration of the choice operator (<|>) and test your builder class using
the examples for parsing person data (see Page 306) and expressions (see Page 307).
 13
Asynchronous and parallel computations

This chapter is about programs where the dynamic allocation of computer resources like
processor time and memory becomes an issue. We consider two different kinds of programs
together with programming constructs to obtain the wanted management of computer
resources:

1. Asynchronous, reactive programs spending most of the wall-clock time awaiting a request
or a response from an external agent. A crucial problem for such a program is to
minimize the resource demand while the program is waiting.
2. Parallel programs exploiting the multi-core processor of the computer by performing
different parts of the computation concurrently on different cores.

The construction of asynchronous and parallel programs is based on the hardware features in
the computer and software features in system software as described in Sections 13.1 and 13.2.
Section 13.3 addresses common challenges and pitfalls in parallel programming. Section 13.4
describes the async computation expression and illustrates its use by some simple examples.
Section 13.5 describes how asynchronous computations can be used to make reactive,
asynchronous programs with a very low resource demand. Section 13.6 describes some of the
library functions for parallel programming and their use in achieving computations executing
concurrently on several cores.
13.1 Multi-core processors, cache memories and main memory
A typical PC in today′s technology (2012) contains two multi-core processor chips, where
each processor chip corresponds to Figure 13.1. Programs and data are stored in the main
memory while the cache memories contain copies of parts of main memory. Each core gives
an independent execution of instructions, and a typical PC offers hence the possibility of four
independent executions of instructions. Instructions and data are fetched from the cache
memories whenever found there – but have otherwise to be transferred from the main
memory. Updating a memory location must always be done in main memory – and in cache if
the memory location is cached.
Typical clock frequency of processor is approx. 2 GHz while the clock frequency of main
memory is approx. 100 MHz (2012 figures), so cache memory is about 20 times faster than
main memory. Maximum speed execution of instructions is hence obtained with instructions
and data in cache while the speed may suffer a substantial degradation if there are frequent
accesses to main memory. Getting instructions and data in cache may hence give a significant
performance gain. Typical memory sizes are 4 GB main memory and 3 MB cache memory so
program and data should fit into the cache unless there is an enormous amount of data – or
some other program is using the cache.

Fig ure 13.1 A multi-core processor chip in a standard PC

The strategies used in managing the cache memories are outside the scope of this book, but
one should observe that all program activities on the computer are competing for cache.
 13.2 Processes, threads and tasks
This section gives a brief survey of the basic features in operating system and run-time
system that are used to manage the concurrent execution of several programs on the
computer.
 Processes
A process is the operating system entity to manage an instance of execution of a free-standing
program. The process contains the program and the data of the program execution. A process
may comprise multiple threads of execution that execute instructions concurrently. A double-
click on an icon on the screen will usually start a process to run the program belonging to the
icon.
A free-standing F# program comprises the Common Language Runtime System, CLR (cf.
“Interoperating with C and COM” in [13]). The Runtime System manages the memory
resources of the process using a stack for each thread and a common heap as described in
Chapter 9, and it manages the program execution using threads as described below. A
simplified drawing of the memory lay-out of such a process is shown in Figure 13.2.
The System.Diagnostics.Process library allows a program to start and manage new
processes. This topic is, however, outside the scope of the present book. The reader may
consult the Microsoft .NET documentation [9] for further information.
 Threads
A thread is the .NET vehicle for program execution on one of the cores in the computer. Each
thread has its own memory stack and separate execution of instructions. In this chapter we
consider only threads managed via a thread pool where tasks containing programs can be
enlisted as work items. Such a task will be executed when a thread and a core become
available.

Fig ure 13.2 Simplified memory lay-out of a process running an F# program

There is a simple example on Page 314 showing creation and start of threads. The reader
may consult the description of the System.Threading library in [9] for further information.
 Tasks
A task is a piece of program that can be executed by a thread. When started, a task is enlisted
as a work item in the thread pool and it is then executed when a thread becomes available.
There are two essentially different ways of executing operations like I/O where the task has to
await the completion of the operation:

Synchronous operations: The operation is started and the executing thread awaits the
completion of the operation. The thread continues executing the task when the operation
has completed. The standard library I/O functions are synchronous operations.
Asynchronous operations: The operation is started and the task becomes a wait item
awaiting the completion of the operation. The executing thread is returned to the thread
pool to be used by other tasks. The task is again enlisted as a work item when the
operation has completed. Asynchronous operations are found in the F# Async library
and in extensions to the standard I/O libraries.

The continuation of program execution after a synchronous operation is done using the stack
of the thread where the information is at hand. The mechanism is different in asynchronous
operations because there is no stack available while the task is waiting. The continuation of
the program execution is therefore saved in a special data structure when the asynchronous
operations is initiated, and this data structure is then later used to continue the task upon
completion of the operation.
A task waiting for completion of an asynchronous operation uses a small amount of
memory and no threads and a process may hence run thousands of asynchronous tasks
concurrently. The situation is quite different for synchronous tasks where the number is
limited by the number of threads available for the program.
These concepts can be illustrated using the “cook-book” metaphor of Section 12.1 where a
program is described as a recipe in a cook-book. A process is then described as a restaurant
while threads are cooks and tasks are the customer′s orders. A synchronous operation
corresponds to a cook focussing on the progress of a single order only, while asynchronous
operations corresponds to a cook switching between several customer ’s orders using kitchen
stop clocks.
Asynchronous operations are called using async computation expressions (cf. Section
13.4).
 13.3 Challenges and pitfalls in concurrency
The concurrent execution of tasks gives new challenges and pitfalls for the programmer:

Management of mutable data shared by several threads.

Deadlocks caused by threads competing for the same resources.
Debugging problems where a program failure cannot be reproduced.

Each of these problems is addressed below in a subsection.

 Shared mutable data
Updating mutable data with a complicated structure may require a number of operations
before getting to a well-defined state. The data may hence become garbled if two update
operations are intermixed by concurrent execution. This problem is solved by ensuring
mutual exclusion of updating threads such that two threads do not update the data structure at
the same time. The mutual exclusion can be obtained using a Mutex object and stipulating the
rule that a thread should acquire this object before updating the data structure and release it
after the updating.
The threads thread1 and thread2 in the following example both use the mutex object
mutex. This object is acquired using the WaitOne method and released using the
ReleaseMutex method. The example also illustrates how to create and start threads using F#.
The reader may consult the documentation of the Threading library in [9] for further details.
open System.Threading;;

let mutex = new Mutex();;

let f (n: int) () =

for k in 1..2 do
Thread.Sleep 50
mutex.WaitOne() |> ignore
printf "Thread %d gets mutex\n" n
Thread.Sleep 100
printf "Thread %d releases mutex\n" n
mutex.ReleaseMutex() |> ignore;;
val f : int -> unit -> unit

let g() =
let thread1 = Thread (f 1)
let thread2 = Thread (f 2)
thread1.Start()
thread2.Start();;
val g : unit -> unit
g();;
Thread 2 gets mutex
Thread 2 releases mutex
Thread 1 gets mutex
Thread 1 releases mutex
Thread 2 gets mutex
Thread 2 releases mutex
Thread 1 gets mutex
Thread 1 releases mutex

The System.Collections.Concurrent library provides thread-safe collections that should

be used in place of the corresponding types in the System.Collections and
System.Collections.Generic libraries whenever multiple threads may access the collection
concurrently.
These problems of shared mutable data does not occur in pure functional programming
without mutable data.
 Deadlocks
A deadlock may occur if two threads thread1 and thread2 are trying to acquire two mutex
objects mutex1 and mutex2 simultaneously as follows:

The deadlock occurs because thread1 is waiting for mutex2 that has been acquired and not yet
released by thread2 while thread2 is waiting for mutex1 that has been acquired and not yet
released by thread1 – both threads are hence stuck and will never proceed.
A problem of this kind can be solved by stipulating a fixed order of nested acquirement of
mutex objects to be used throughout the program – for instance always acquire mutex1 before
acquiring mutex2 in the example. A thread acquiring mutex1 would then always find a free
mutex object mutex2 and can proceed acquiring also mutex2. The thread should eventually
release both mutex objects – whereupon another thread may proceed acquiring these objects
(in the same order).
A program containing a potential deadlock will work most of the time and the deadlock
situation will only occur spuriously in special situations where two threads are doing the
reservations at exactly the same time. Extensive testing will most likely not reveal the problem
and you may end up with an unreliable program that occasionally stops working in stressed
situations – exactly when a correct function is most needed.
 Debugging problems. Logging facilities
A bug in an asynchronous program may have the unpleasant property that it only leads to
failure under special courses of external events and special timing conditions (like the above
described deadlock). The traditional scheme of receiving error reports from the customer
does not work for such programs. The programmer receiving the error report will most
likely not get a relevant description of the course of events and timing conditions leading to
the failure and will be unable to reproduce the failure – and hence unable to locate the bug.
The solution to this problem is to include logging facilities in the program. These facilities
should produce a disk file containing a log of events occurring while the program is running.
The contents of the log file should be part of the error report from the customer sent in case
of a program failure. The log file can then be used by the programmer to trace the course of
events leading to the failure – and hopefully to find and correct the bug.
Logging facilities should be a part of any serious real-world asynchronous system. This
theme will, however, not be pursued further in this book.
 13.4 Asynchronous computations
A value of type Async<′a> (for some type ′a) is an asynchronous computation. When started
it becomes an executing task that either continues the current task or runs as an independent
task. A terminating execution of an asynchronous task of type Async<′a> delivers a value of
type ′a.
This concept gives a very elegant notation because an asynchronous computation can be
used like any other value as argument or result of a function. Simple asynchronous
computations are build using the operations in Table 13.1 inside an async expression:
Table 13.1 Selected asynchronous operations

stream.AsyncRead: int -> Async<string>

stream.AsyncRead n = async.comp. to read n chars from stream.
stream.AsyncRead: byte [] * ?int * ?int -> Async<int>
stream.AsyncRead(buf),
stream.AsyncRead(buf,m) or
stream.AsyncRead(buf,m,n).
Async.comp to read into buf,
possibly from pos. m and possibly n chars
stream.AsyncWrite: string -> Async<unit>
stream.AsyncWrite str = async.comp. to write str to stream
stream.AsyncWrite: string * int * int -> Async<unit>
stream.AsyncWrite(str,m,n)
= asc.cmp. to write n chars of
str from pos. m
webClient.Async.DownloadString: Uri -> Async<string>
async.comp. to get WEB source determined by webClient.
webRequest.AsyncGetResponse: unit -> Async<WebResponse>
async.comp. to await response on web-request
Async.Sleep: int -> Async<unit>
Async.Sleep n = async.comp. sleeping n mS
A question mark (?) signals an optional argument.
Table 13.2 Function combining asynchronous computations

Async.Parallel: Seq<Async<′a>> -> Async<′a []>

Async.Parallel [c0 ; …; cn−1 ] = async.comp. of c0 ,…,cn−1 put in parallel

Table 13.3 Selected functions to activate or deactivate asynchronous computations

Async.RunSynchronously:
 Async<′a> * ?int * ?CancellationToken -> ′a
Activates async.comp. possibly with time-out in mS and possibly with specified
cancellation token. Awaits completion.
Async.Start: Async<unit> * ?CancellationToken -> unit
Activates async.comp. possibly with specified canc. token. Does not await
completion.
Async.StartChild: Async<′T> * ?int -> Async<Async<′T>>
Activates async.comp. and gets async.comp. to await result
Async.StartWithContinuations:
Async<′T> * (′T -> unit) * (exn -> unit)
*(OperationCanceledException -> unit) * ?CancellationToken
-> unit
Activates async.comp. with specified continuation and possibly specified
cancellation token. Does not await completion of computation.
Async.FromContinuations: ((′T -> unit) * (exn -> unit)
* (OperationCanceledException -> unit) -> unit) -> Async<′T>
Makes current asynchronous task a wait item. Argument function is called with
triple
of trigger functions as the argument and should save one or more of these
closures in
variables. The task continues as a work item when a trigger function is called.

The function Async.Parallel in Table 13.2 is used to put asynchronous computations in

parallel. The computations are started using the functions in Table 13.3. These tables do only
contain a selection of functions – further information can be found in [9].
 Simple examples of asynchronous computations
Using the asynchronous operation webClient.AsyncDownloadString we may build an
asynchronous computation to download the HTML-source of the DTU home page:
open System ;; open System.Net;; // Uri, WebClient
let downLoadDTUcomp =
async let webCl = new WebClient()
let! html =
webCl.AsyncDownloadString(Uri "http://www.dtu.dk")
return html ;;
val downLoadDTUcomp : Async<string>

This is just a value like any other – but if started, it will run a task to download the HTML-
source of the DTU home page. This task will do the following:

1. Create a WebClient object. This declaration need actually not be part of the async
expression and could hence be placed before async…
2. Initiate the download using AsyncDownloadString. This function makes the task an wait
item and returns this item in the form of an Async value comp. The asynchronous
computation comp will eventually run and terminate when the download has completed.
3. The termination of comp re-starts the rest of the computation with the identifier html
bound to the result of comp (which in this case is the result of the download).
4. The expression return html returns the value bound to html, that is, the result of the
download.

Please observe the following

The computation uses very few resources while waiting for the download – it uses for
instance no thread during this time period.
The let! construct is required to make a binding to a value that is later returned at the
termination of an asynchronous computation.
The computation expression does in most cases contain a construct like return or
return! to give a result – and will otherwise give the dummy value “()” as the result.
Using return! yields a new asynchronous computation.
 Functions computing asynchronous computations
We may generalize the above example to a function computing the asynchronous download
computation for arbitrary URL:
let downloadComp url =
let webCl = new WebClient()
async let! html = webCl.AsyncDownloadString(Uri url)
return html;;
val downloadComp : string -> Async<string>

Computations downloading the HTML sources of the DTU and Microsoft home pages may
then be obtained as function values:
let downloadDTUcomp = downloadComp "http://www.dtu.dk";;
val downloadDTUcomp : Async<string>
let downloadMScomp = downloadComp "http://www.microsoft.com";;
val downloadMScomp : Async<string>

A computation downloading the HTML-sources corresponding to an array of URL’s in

parallel can be made using Async.Parallel and Array.map:
let downlArrayComp (urlArr: string[]) =
Async.Parallel (Array.map downloadComp urlArr);;
val downlArrayComp : string [] -> Async<string []>

and we may hence download the HTML-sources of the DTU and the Microsoft home pages
concurrently and compute their lengths:
let paralDTUandMScomp =
downlArrayComp
[|"http://www.dtu.dk"; "http://www.microsoft.com"|];;
val paralDTUandMScomp : Async<string []>
Array.map (fun (s:string) -> s.Length)
(Async.RunSynchronously paralDTUandMScomp);;
val it : int [] = [|45199; 1020|]

The parallel download of HTML-sources can instead be made using the StartChild function.
This gives separated activation and waiting for completion of two child tasks:
let parallelChildrenDTUandMS =
async let! compl1 = Async.StartChild downloadDTUcomp
let! compl2 = Async.StartChild downloadMScomp
let! html1 = compl1
let! html2 = compl2
return (html1,html2);;
val parallelChildrenDTUandMS : Async<string * string>

The calls of StartChild:

let! compl1 = Async.StartChild downloadDTUcomp
let! compl2 = Async.StartChild downloadMScomp
start the downloads in two child tasks in parallel. The identifiers compl1 and compl2 are
bound to two asynchronous computations that when started will await the completion of the
child tasks. The main task is hence not blocked by the StartChild operations. It becomes
blocked when compl1 is started and awaits the completion of the corresponding child task:
let! html1 = compl1

The next let! construct: let! html2 = compl2 will in the same way afterwards await the
completion of the second child task.
 Exception and cancellation
Executing async computations includes handling premature termination caused by an
exception or a cancellation. These concepts can be informally explained using the “cook-
book recipe” metaphor of Section 12.1 by imaging a cook working in a restaurant. The cook
should not only follow the recipe when processing a customer ’s order but also handle the
following abnormal situations:

Exception: An exception has occurred (like break-down of an oven).

Cancellation: The customer has cancelled the order.

A task executing an async computation reacts in case of an exception or a cancellation by

calling the corresponding continuation. A cancellation is requested (from outside the task) by
setting the cancellation token of the execution of the computation (see example below). The
cancellation token is polled regularly by the asynchronous library functions and by the
member functions of the async computation expression. The cancellation is performed with a
proper clean-up of resources as soon as the cancellation request has been discovered.
Using the library function Async.StartWithContinuations you may supply your own
continuations when an asynchronous computation is started. This function requires three
continuations among its parameters:

Normal continuation okCon – invoked after normal termination.

Exception continuation exnCon – invoked if an exception is raised.
Cancellation continuation canCon – invoked if the computation is cancelled.

The following example executes the above function downloadComp with continuations:
open System.Threading;; // CancellationTokenSource
let okCon (s: string) = printf "Length = %dn" (s.Length);;
let exnCon _ = printf "Exception raisedn";;
let canCon _ = printf "Operation cancelledn";;

let downloadWithConts url =

use ts = new CancellationTokenSource()
Async.StartWithContinuations
((downloadComp url),okCon,exnCon,canCon,ts.Token)
ts;;
val downloadWithConts : string -> CancellationTokenSource

A computation started by a call of downloadWithConts may terminate normally:

downloadWithConts "http://www.microsoft.com" |> ignore;;
val it : unit = ()
Length = 1020

it may be terminated by an exception:

downloadWithConts "ppp" |> ignore;;
Exception raised
val it : unit = ()

or it may be cancelled:
let ts = downloadWithConts "http://www.dtu.dk";;
ts.Cancel();;
val it : unit = ()
Operation cancelled

Note the following:

The task started by Async.StartWithContinuations terminates when the selected

continuation returns the dummy value “()”. A meaningful program would hence use
continuations that initiate some other activity – for instance by sending a message to a
queue or activating another task.
Each execution of an asynchronous computation with possible cancellation should have a
fresh cancellation token source. The above function downloadWithConts ensures that by
including the allocation in the function declaration.
Requesting cancellation using Cancel on the token source can be followed by a call of
one of the other continuations if an error occurs or if the operation terminates before the
cancellation gets through.
 13.5 Reactive programs
A reactive program performs operations using asynchronous waiting and may hence be used
to perform many long lasting I/O operations simultaneously while also communicating with
the user at the same time. Such a program can be implemented using an asynchronous event
queue. It is a queue containing events of the following kinds, for example:

Mouse clicks or key-strokes or other kinds of user input.

Responses from asynchronous operations.
 Asynchronous event queue
An asynchronous event queue from the class AsyncEventQueue supports two operations:
ev.Post : ′T -> unit
ev.Receive : unit -> Async<′T>

where

ev.Post msg: inserts the element msg in the event queue ev.
ev.Receive(): Awaits the next element in the event queue ev.

The event queue class AsyncEventQueue is kindly provided by Don Syme, and its
implementation is shown in Table 13.4. It is possible that this queue will be included in the F#
standard library.
 Design of dialogue programs
We shall now consider the design of reactive programs where the system may engage in a
dialogue with a user. The systems considered here will have to react to two kinds of events:

Input from a user.

Status events from asynchronous computations.

These events are handled by use of an asynchronous event queue.

// An asynchronous event queue kindly provided by Don Syme

type AsyncEventQueue<′T>() =
let mutable cont = None
let queue = System.Collections.Generic.Queue<′T>()
let tryTrigger() =
match queue.Count, cont with
| _, None -> ()
| 0, _ -> ()
| _, Some d ->
cont <- None
d (queue.Dequeue())

let tryListen(d) =
if cont.IsSome then invalidOp "multicast not allowed"
cont <- Some d
tryTrigger()

member x.Post msg = queue.Enqueue msg; tryTrigger()

member x.Receive() =
Async.FromContinuations (fun (cont,econt,ccont) ->
tryListen cont)

Table 13.4 An implementation of AsyncEventQueue by Don Syme

Fig ure 13.3 Window of asynchronous dialogue program

Consider for example a primitive dialogue program that finds lengths of HTML-sources of
web-pages. The program shows a window as in Figure 13.3. The upper text box is used to
enter the URL while the lower text box shows the answer from the program. The buttons have
the following functions:
Start url: Starts the download of the web-page using the URL in the upper text box.
Clear: Clears the text boxes.
Cancel: Cancels a progressing download.
The program we shall construct make use of the asynchronous event queue shown in Table
13.4 and it has three parts:

open System
open System.Net
open System.Threading
open System.Windows.Forms
open System.Drawing

// The window part

let window =
new Form(Text="Web Source Length", Size=Size(525,225))
let urlBox =
new TextBox(Location=Point(50,25),Size=Size(400,25))
let ansBox = new TextBox(Location=Point(150,150),Size=Size(200,25))
let startButton =
new Button(Location=Point(50,65),MinimumSize=Size(100,50),
MaximumSize=Size(100,50),Text="START")
let clearButton =
new Button(Location=Point(200,65),MinimumSize=Size(100,50),
MaximumSize=Size(100,50),Text="CLEAR")
let cancelButton =
new Button(Location=Point(350,65),MinimumSize=Size(100,50),
MaximumSize=Size(100,50),Text="CANCEL")
let disable bs = for b in [startButton;clearButton;cancelButton] do
b.Enabled <- true
for (b:Button) in bs do
b.Enabled <- false

// The dialogue part from Table 13.7 belongs here

// Initialization
window.Controls.Add urlBox
window.Controls.Add ansBox
window.Controls.Add startButton
window.Controls.Add clearButton
window.Controls.Add cancelButton
startButton.Click.Add (fun _ -> ev.Post (Start urlBox.Text))
clearButton.Click.Add (fun _ -> ev.Post Clear)
cancelButton.Click.Add (fun _ -> ev.Post Cancel)
// Start
Async.StartImmediate (ready())
window.Show()

Table 13.5 Dialogue program: The Window, Initialization and Start parts
 The first part contains declarations corresponding to the window shown in Figure 13.3.
These declarations are shown in Table 13.5. In this part buttons and text boxes are
declared. Furthermore, a function disable is declared that controls enable/disable of the
buttons in the window. During the download of a web-page, for example, the user should
have the option to cancel the ongoing download; but the buttons for clearing the text
fields and for starting up a new download should be disabled in that situation.
The second part (see comments in Table 13.5) contains the dialogue program. We shall
focus on this part in the following.
The third part connects the buttons of the user interface to events, shows the window and
starts the dialogue program. This part is shown in the lower part of Table 13.5.

Notice that the program is an event-driven program with asynchronous operations all running
on a single thread. The complete program is found at the homepage for the book.
 Dialogue automaton
We shall design an event-driven program that reacts to user events and status events from
asynchronous operations. The user events are described above. An asynchronous download
of a web-page can result in three kinds of status events:
Web html: The event containing the html-source of a web-page.
Cancelled: The event signalling a successful cancelling of a download.
Error: The event signalling an unsuccessful download of a web-page possibly due to an
illegal URL.
The system must perform some actions in response to incoming events, for instance: the
action corresponding to a Clear event is that the text boxes are cleared, the action
corresponding to a Start url event is the start of an asynchronous download of the web-page
url, and the action corresponding to a Web html event prints the number of characters in the
string html in the lower text box.
The possible sequences of events of an reactive program are often conveniently described
by a simple automaton. An automaton is a directed graph, with a finite number of vertices also
called called states and edges also called transitions. A specific state is called the initial state.
A transition is labelled with a set of events. A path of the automaton is a sequence:

where there is a transition labelled ei from si-1 to si, for 1<i≤n. The sequence of events e1 e2 ·s
en is called a run.

Fig ure 13.4 A simple dialogue automaton

Consider the automaton in Figure 13.4 with states: ready, loading, cancelling and
finished, where ready is the initial state – it is marked with an incoming arrow, and six
events: Start url, Clear, Cancel, Web html, Cancelled and Error.
The runs starting in the initial state ready describe the allowed sequences of events. For
example, the sequence
Clear Start(url1) Web(html1) Clear Start(url2)

is allowed because it “brings the automaton” from the ready state to the loading state. Other
allowed sequences are:
Clear Start(url1) Web(html1) Clear Start(url2)
Start(url1) Cancel Cancelled Clear
Start(url1) Cancel Web(html1)

while the following sequence is forbidden:

Start(url1) Cancel Clear

because the automaton gets stuck in the cancelling state. The two first events:
Start(url1) Cancel

lead to the cancelling state and there are no outgoing transition labelled Clear from that
state.
Notice that the automaton conveys an overview of the interaction in the system. Furthermore,
the corresponding dialogue program will systematically be constructed from the dialogue
automaton in terms of four mutually recursive functions corresponding to the four states.
This leads to the program skeleton in Table 13.6.
The only parts that are missing in this program skeleton relate to the actions for the
incoming events of the states. The other parts are systematically derived from the dialogue
automaton in Figure 13.4. The function implementing a given state of the automaton, for
example, the ready state, has three parts:
Part 1: Actions corresponding to the incoming event are performed in the first part. This
part is not described in the skeleton because these details are not present in the automaton.
Part 2: Forbidden user input is disabled. By inspection of the events labelling the
transitions leaving a state, it can be observed which input the user should be able to
provide in that state. The ready state has no outgoing transition labelled Cancel and the
corresponding button is therefore disabled.
Part 3: Wait for incoming events and make a corresponding state transition. In the ready
state only Start and Clear events are allowed. A Clear event leads back to the ready state
while a Start event leads to the loading state.
The program skeleton in Table 13.6 contains a few details not present in Figure 13.4: Answer
strings are passed from the loading and cancelling states to the finished state.

let ready() =
async .... // actionReady: actions for incoming events
disable [cancelButton]
let! msg = ev.Receive()
match msg with
| Start url -> return! loading(url)
| Clear -> return! ready()
| _ -> failwith("ready: unexpected message")
 and loading(url) =
async .... // actionLoading: actions for incoming events
disable [startButton; clearButton]
let! msg = ev.Receive()
match msg with
| Web html ->
let ans = "Length = " + String.Format("0:D",html.Length)
return! finished(ans)
| Error -> return! finished("Error")
| Cancel -> ts.Cancel()
return! cancelling()
| _ -> failwith("loading: unexpected message")

and cancelling() =
async .... // actionCancelling: actions for incoming events
disable [startButton; clearButton; cancelButton]
let! msg = ev.Receive()
match msg with
| Cancelled | Error
| Web _ -> return! finished("Cancelled")
| _ -> failwith("cancelling: unexpected message")

and finished(s) =
async .... // actionFinished: actions for incoming events
disable [startButton; cancelButton]
let! msg = ev.Receive()
match msg with
| Clear -> return! ready()
| _ -> failwith("finished: unexpected message")

Table 13.6 Skeleton program for automaton in Figure 13.4

It is now straightforward to complete the whole dialogue program. The type for events (or
messages) and an event queue are declared as follows:
type Message = | Start of string | Clear | Cancel
| Web of string | Error | Cancelled;;

let ev = AsyncEventQueue();;

and the action parts missing in the skeleton program are declared as follows:

Actions for incoming event in the ready state: The two text boxes must be cleared:
urlBox.Text <- "http://"
ansBox.Text <- ""

Actions for incoming event in the loading state: The text box for the answer is set and an
asynchronous download of a web-page is started with continuations as we have seen on
Page 320:
ansBox.Text <- "Downloading"
use ts = new CancellationTokenSource()
 Async.StartWithContinuations
(async let webCl = new WebClient()
let! html = webCl.AsyncDownloadString(Uri url)
return html ,
(fun html -> ev.Post (Web html)),
(fun _ -> ev.Post Error),
(fun _ -> ev.Post Cancelled),
ts.Token)

Actions for incoming event in the cancelling state: The answer text box is set.
ansBox.Text <- "Cancelling"

Actions for incoming event in the finished state: The answer text box is set.
ansBox.Text <- s

The complete program for the dialogue automaton is found in Appendix C.

 A summary of the approach
We have considered the design of reactive systems in terms of a very simple example in order
to be able to focus on the principle elements of the approach. Many systems have a similar
form where a system is engaging in a dialogue with users and external sources like database
servers on the basis of asynchronous communication.
The type Message in the example, provide an abstract notion of the important events in the
system that abstracts away the concrete interactions with the user interface. A dialogue
automaton provides a convenient technique to define the legal sequences of events in the
system. This automaton conveys the essence of the dialogue design in a succinct manner and a
dialogue program can systematically be derived from this automaton.
The technical advantage of the approach is that the resulting asynchronous program is
executed in a single thread requiring limited computational resources.
 13.6 Parallel computations
The Task Parallel Library of the .NET platform provides a powerful framework for
exploiting multi-core parallelism. In this section we shall show that functional programming
provides an adequate platform for a programmer wanting to exploit this parallelism to
speedup the programs. Obtaining a parallel implementation of a side-effect free program
makes the correctness problem of the parallel version simple and good library support for
parallelism makes the step to the parallel version manageable.
We shall distinguish between two kinds of parallelism: data parallelism, where the same
function is applied in parallel on distributed data, and task parallelism, where a complex
problem is solved by combining solutions to simpler problems, that can be solved in parallel.
When measuring the effect of multiple cores we consider “big” problems in terms of
computation requirements. It does not pay off to parallelize small problems due to the
management overhead needed for multiple cores.
In order to experiment with parallelization, we need primitive operations that demands
some computation resources in order to make the effect of parallelization visible.
Throughout this section a prime number test on randomly generated integers will be used for
that purpose. The prime-number test is performed by the function:
let isPrime =
let rec testDiv a b c =
a>b || c%a <> 0 && testDiv (a+1) b c
function
| 0 | 1 -> false
| n -> testDiv 2 (n-1) n;;
val isPrime : int -> bool

The locally declared function testDiv a b c is true if no integer i where a ≤i ≤b divides c.

Testing whether n, with n>1, is a prime number, it suffices to test whether no integer between
2 to divides n. In order to use more computing resources isPrime is inefficiently
implemented by performing this test from 2 to n−1.
isPrime 51;;
val it : bool = false

isPrime 232012709;;
val it : bool = true

A test of a small number is fast whereas a test of a large prime number like the one in the
above example takes some observable amount of time.
We shall use randomly generated integers in our experiments. They are generated by the
following function gen, where genrange, with range > 0, generates a number that is greater
than or equal to 0 and smaller than range:
let gen = let generator = new System.Random()
 generator.Next;;
val gen : int -> int
gen 100;;
val it : int = 24

gen 100;;
val it : int = 53

The experiments in the rest of this section are conducted on a 4-core 2.67 GHz Intel I7 CPU
with 8GB shared memory.
 Data parallelism
The map function on collections is the canonical example for exploiting data parallelism,
where a function is applied in parallel to the members of a collection. Parallel
implementations of functions on arrays are found in the Array.Parallel library as shown in
Table 13.7.

choose : (′T -> ′U option) -> ′T [] -> ′U []

collect : (′T -> ′U []) -> ′T [] -> ′U []
init : int -> (int -> ′T) -> ′T []
iter : (′T -> unit) -> ′T [] -> unit
iteri : (int -> ′T -> unit) -> ′T [] -> unit
map : (′T -> ′U) -> ′T [] -> ′U []
mapi : (int -> ′T -> ′U) -> ′T [] -> ′U []
partition : (′T -> bool) -> ′T [] -> ′T [] * ′T []

Table 13.7 Functions in the library: Array.Parallel

We have studied these functions previously in the book, so we just illustrate the advantage of
using the function parallel version of the map function on an array with 5000000 numbers:
let bigArray = Array.init 5000000 (fun _ -> gen 10000);;
val bigArray : int [] = [|2436; 7975; 2647; 1590; 5959; 3951;
430; 1705; 2527; 1004; 2333; ... |]

Mapping the isPrime function on the elements of bigArray will generate a new Boolean
array, where an entry is true if and only if the corresponding entry in bigArray is a prime
number:
#time;;

Array.Parallel.map isPrime bigArray;;

Real: 00:00:05.292, CPU: 00:00:20.592,
GC gen0: 0, gen1: 0, gen2: 0
val it : bool [] = [|false; false; true; false; false; false;
false; false; false; false; true; ...|]

Array.map isPrime bigArray;;

Real: 00:00:10.220, CPU: 00:00:10.218,
GC gen0: 0, gen1: 0, gen2: 0
val it : bool [] = [|false; false; true; false; false; false;
false; false; false; false; true; ...|]

The experiment shows a speed-up of approximately 2 in real time when using the parallel
version of map. The main point is that achieving this speed-up is effortless for the
programmer. Note that the total CPU time (20.218 seconds) used on all cores is approximately
double the time needed for a non-parallel version.
In order to use the library for parallel operations on sequences you need to install the F#
Power Pack. The PSeq library in that package provides parallel versions of a rich collection
of the functions in the Seq library (see Chapter 11). These functions can also be used on lists
and arrays as we have seen in Section 11.7. We just show one experiment with the exists
function from the PSeq library:
#r @"FSHarp.PowerPack.Parallel.Seq"
open Microsoft.FSharp.Collections

let bigSequence = Seq.init 5000000 (fun _ -> gen 10000);;

Real: 00:00:00.001, CPU: 00:00:00.000,
GC gen0: 0, gen1: 0, gen2: 0
val bigSequence : seq<int>

Seq.exists (fun i -> isPrime i && i>10000) bigSequence;;

Real: 00:00:11.557, CPU: 00:00:11.528,
GC gen0: 247, gen1: 3, gen2: 1
val it : bool = false

PSeq.exists (fun i -> isPrime i && i>10000) bigSequence;;

Real: 00:00:05.985, CPU: 00:00:22.183,
GC gen0: 250, gen1: 1, gen2: 0
val it : bool = false

In the example we search for the existence of a prime number that do not exists in the
generated sequence in order to be sure that the whole sequence is traversed. The speed-up is
about 2 and the figures are similar to those for the above experiment using map.
 Task parallelism
The problem-solving strategy we have used throughout the book is to solve a complex
problem by combining solutions to simpler problems. This strategy, which also is known as
divide and conquer, fits very well with task parallelism, where a complex problem is solved
by combining solutions to simpler problems that can be solved in parallel.
We illustrate the idea on a simple example. Consider the type for binary trees given in
Section 6.4:
type BinTree<′a> = | Leaf
| Node of BinTree<′a> * ′a * BinTree<′a>;;

A function to test for the existence of an element in a binary tree satisfying a given predicate
is declared as follows:
let rec exists p t =
match t with
| Leaf -> false
| Node(_,v,_) when p v -> true
| Node(tl,_,tr) -> exists p tl || exists p tr;;
val exists: (′a -> bool) -> BinTree<′a> -> bool

The divide and conquer strategy is employed in the last clause: In order to check whether an
element in a tree satisfies the given predicate, the check is performed in the left and right
subtrees and those results are combined.
We shall generate trees using the following function:
let rec genTree n range =
if n=0 then Leaf
else let tl = genTree (n-1) range
let tr = genTree (n-1) range
Node(tl, gen range, tr);;
val genTree : int -> int -> BinTree<int>

let t = genTree 25 10000;;

The value of genTreen range is a balanced binary tree with depth n, where every element v
occurring in a node is an integer satisfying 0 ≤ v <range. The generated tree t, therefore, has
225 leaves and searching through the whole tree is a time-consuming operation:
exists (fun n -> isPrime n && n>10000) t;;
Real: 00:01:22.818, CPU: 00:01:22.727,
GC gen0: 0, gen1: 0, gen2: 0
val it : bool = false

The obvious idea for parallelization is to do the search in the left and right subtrees in
parallel and combine their results. The function Task.Factory.StartNew from the
namespace System.Threading.Tasks is used to create and start a new task:
open System.Threading.Tasks;;
let rec parExists p t =
match t with
| Leaf -> false
| Node(_,v,_) when p v -> true
| Node(tl,_,tr) ->
let b1 = Task.Factory.StartNew(fun () -> parExists p tl)
let b2 = Task.Factory.StartNew(fun () -> parExists p tr)
b1.Result||b2.Result;;
val parExists: (′a -> bool) -> BinTree<′a> -> bool

Evaluation of the declaration

let b1 = Task.Factory.StartNew(fun () -> parExists p tl)

will create and start a task object of type Task<bool> and b1 is bound to that object (similarly
for the declaration of b2). The property Result gets the result of the task upon its completion.
This parallel version does, however, not give any significant performance gain:
parExists (fun n -> isPrime n && n>10000) t;;
Real: 00:01:19.659, CPU: 00:04:43.578,
GC gen0: 2972, gen1: 10, gen2: 1
val it : bool = false

The problem with this version is that a huge amount of tasks are created and the
administration of these tasks cancels out the advantage with multiple core.
This problem is handled by the introduction of a maximal depth to which new tasks are
created:
let rec parExistsDepth p t n =
if n=0 then exists p t
else match t with
| Leaf -> false
| Node(_,v,_) when p v -> true
| Node(tl,_,tr) ->
let b1 = Task.Factory.StartNew(
fun () -> parExistsDepth p tl (n-1))
let b2 = Task.Factory.StartNew(
fun () -> parExistsDepth p tr (n-1))
b1.Result||b2.Result;;
val parExistsDepth : (′a -> bool) -> BinTree<′a> -> int -> bool

Experiments show that the best result is obtained using depth 4:

parExistsDepth (fun n -> isPrime n && n>10000) t 4;;
Real: 00:00:35.303, CPU: 00:02:18.669,
GC gen0: 0, gen1: 0, gen2: 0

The speedup is approximately 2.3. At depths starting from about 22 the degradation of
performance grows fast. This is not surprising taking the number of subtrees at such depths
into account.

Example: Quick sort

A classical algorithm that is based on the divide and conquer problem-solving technique is the
Quick sort algorithm that was developed by C.A.R. Hoare. The basic idea is very simple. The
array:

is sorted by first rearranging the elements v1 ... vn−2 vn−1 such that the resulting elements v’1 ...
v′n−2 v′n−1 can be partitioned into two sections with indices 1, … , k and k+1, …, n-1,
respectively, such that all the elements in first section are smaller than v0 and all the elements
in the second section are greater than or equal to v0:

The element v0 can now be correctly placed in its final position by swapping it with the k′s
element:

This array has the property that any element in the first section is smaller than any element in
the second section, as the elements in the first section are < v0 while the elements in the second
section are ≥ v0. The array can hence be sorted by sorting each of the sections separately. This
algorithm will have an average run time that is proportional to n· log n and a worst-case run
time proportional to n2, where n is the length of the array.
The sorting algorithms available in the libraries have a better worst-case run time
(proportional to n· log n) and they are using very efficient algorithms. So our
recommendation is to use these libraries. We just use the Quick sort algorithm here to
illustrate that the above method for parallelizing a divide and conquer algorithm applies to a
non-trivial algorithm.
The function swap exchanges two elements of an array:
let swap (a: ′a[]) i j =
let v = a.[i]
a.[i] <- a.[j]
a.[j] <- v;;
val swap : ′a [] -> int -> int -> unit

and the function partition can rearrange a section of an array:

so that the elements in the section which are smaller than a give value v comes before the
elements which are greater than or equal to v:

The value of the expression partition a v k1 k2 is K, that is, the index of the last element in
the first section containing elements smaller than v:
let rec partition (a:′a[]) v k1 k2 =
if k2=k1-1 then k2 //empty section
else if a.[k2] >= v then partition a v k1 (k2-1)
else swap a k1 k2
partition a v (k1+1) k2;;
val partition : ′a [] -> ′a -> int -> int -> int
when ′a : comparison

The basic Quick sort algorithm is declared as follows:

let rec qsort a i j =
if j-i>1 then let k = partition a a.[i] (i+1) (j-1)
swap a i k
qsort a i k
qsort a (k+1) j;;
val qsort : ′a [] -> int -> int -> unit when ′a : comparison

let sort a = qsort a 0 (Array.length a);;

val sort : ′a [] -> unit when ′a : comparison

So far we have just achieved an imperative program that can sort an array:
let a1 = [|1; -4; 0; 7; 2; 3|];;
val a1 : int [] = [|1; -4; 0; 7; 2; 3|]

sort a1;;
val it : unit = ()

a1;;
val it : int [] = [|-4; 0; 1; 2; 3; 7|]

Even though Quick sort is an imperative algorithm that changes an array, this does not cause
any problems for a parallel version since the two recursive calls of qsort work on non-
overlapping sections of the array – these two recursive call are independent of each other.
Therefore, a parallel version that creates tasks up to a certain depth only is straightforwardly
achieved using the same technique as used for the parallel search in a binary tree:
let rec pqsort a i j depth =
if j-i<= 1 then ()
else if depth=0 then qsort a i j
else let k = partition a a.[i] (i+1) (j-1)
swap a i k
let s1 = Task.Factory.StartNew
(fun () -> pqsort a i k (depth-1))
 let s2 = Task.Factory.StartNew
(fun () -> pqsort a (k+1) j (depth-1))
Task.WaitAll[|s1;s2|];;
val pqsort : ′a [] -> int -> int -> int -> unit
when ′a : comparison

let parSort a d = pqsort a 0 (Array.length a) d;;

val parSort : ′a [] -> int -> unit when ′a : comparison

Since pqsort is an imperative algorithm we need to wait for the termination of both of the
tasks s1 and s2 for the recursive calls. The function
Task.WaitAll: Task [] -> unit

is used for that purpose. It waits until all the provided tasks have completed their executions.
Experiments show a speed-up of approximately 1.7 when sorting an array of size 3200000:
let a32 = Array.init 3200000 (fun _ -> gen 1000000000);;
let a32cp = Array.copy a32;;

sort a32;;
Real: 00:00:14.090, CPU: 00:00:14.024,
GC gen0: 1009, gen1: 3, gen2: 0
val it : unit = ()

parSort a32cp 7;;

Real: 00:00:08.352, CPU: 00:00:20.030,
GC gen0: 1016, gen1: 1, gen2: 1
val it : unit = ()

It is not surprising that parSort gets a smaller speed-up than parExistsDepth. The recursive
call of parExistsDepth requires only that the disjunction || is computed on the values of the
recursive calls, and this sequential part is a very fast constant-time operation. On the other
hand, prior to the two recursive and parallel calls of parSort, a partitioning has to be made of
the section to be sorted, and this sequential component has a run time that is linear in the size (
j−i) of the section.
 Summary
In this chapter we have introduced

asynchronous, reactive programs spending most of the wall-clock awaiting a request or

a response from an external agent, and
parallel programs exploiting the multi-core processor of the computer.

The common challenges and pitfalls in parallel programming are described. The async
computation expression is introduced and it is shown how asynchronous computations can be
used to make reactive, asynchronous programs with a very low resource demand. Library
functions for parallel programming are introduced and it is show how they are used in
achieving computations executing concurrently on several cores.
 Exercise
13.1 Make program producing the deadlocked situation described on Page 315.
13.2 Make a type extension (cf. Section 7.4) of the class AsyncEventQueue<′T> with an extra
member Timer: ′T -> int -> unit such that evaluating
evnq.Timer evnt n

will start an asynchronous computation that first sleeps n milliseconds and afterwards sends
the event evnt to the queue evnq.
Hint: Apply Async.StartWithContinuations to Async.Sleep with suitable continuations.
13.3 Consider the dialogue program in Table C.1. Sometimes it is more convenient to let the
functions for the state of the automaton communicate using shared variables rather than
using function parameters. Revise the program so that loading and finished become
parameterless functions. Is this revision an improvement?
13.4 Make a quiz program where a user should guess a number by asking the following
questions:

Is the number < n?

Is the number = n?
Is the number > n?

where n is a integer. The program can give the following answers:

Yes
No
You guessed it!

The program must fix a random number between 0 and 59 to be guessed before starting the
dialogue, and each run of the program should give a new number to be guessed.
13.5 Make a geography program guessing a country in Europe. The program asks questions
to the user who answers yes or no. The program should use a binary tree with country
names in the leaves and with a question in each node, such that the left subtree is chosen in
case of answer yes and the right in case of answer no.
The program can be made to look more “intelligent” by inserting some random questions
in between the systematic questions taken from the tree. The random questions should be of
two kinds: Silly questions where the answer is not used by the program, and direct questions
guessing a specific country where the answer is used by the program in case it gets answer
yes.
 13.6 The game of Nim is played as follows. Any number of matches are arranged in heaps,
the number of heaps, and the number of matches in each heap, being arbitrary. There are
two players A and B. The first player A takes any number of matches from a heap; he may
take one only, or any number up to the whole of the heap, but he must touch one heap only.
B then makes a move conditioned similarly, and the players continue to take alternately. The
player who takes the last match wins the game.
The game has a precise mathematical theory: We define an operator xorb for non-negative
integers by forming the exclusive or of each binary digit in the binary representation of the
numbers, for example

The xorb operator in F# is written ˆ ˆ ˆ, for example:

109 ^^^ 70;;
val it : int = 43

The operator xorb is associative and commutative, and 0 is the unit element for the
operator.
Let the non-negative integers a1, …, an be the number of matches in the n heaps, and let m
denote the integer:

The following can then be proved:

1. If m ≠ 0 then there exists an index k such that ak xorb m < ak. Replacing the number ak by
ak xorb m then gives a new set of ai′s with m = 0.
2. If m = 0 and if one of the numbers ak is replaced by a smaller number, then the m-value
for the new set of ai′s will be ≠ 0.

This theory gives a strategy for playing Nim:

1. If m ≠ 0 before a move, then make a move to obtain m = 0 after the move (cf. the above
remark 1).
2. If m = 0 before a move, then remove one match from the biggest heap (hoping that the
other player will make a mistake, cf. the above remark 2).

Use this strategy to make a program playing Nim with the user.
13.7 In this exercise we shall use data and task parallelism in connection with computation
of Fibonacci numbers.

Consider your solution to Exercise 1.5. Make an array containing the integers from 0 to
40 and apply Array.map to compute the first 41 Fibonacci numbers. Measure the run time
of this operation. Make a data parallel solution using Array.Parallel.map and compare
 the solutions.
Make a task-parallel solution to Exercise 1.5 and measure the obtained speedup when
computing big Fibonacci numbers.
Compare the obtained results with a sequential solution using accumulation parameters
(Exercise 9.7). Note that the linear speedup obtained using multiple cores does not
replace the use of good algorithms.

13.8 In this exercise you shall make a list-based version of the Quick sort algorithm. Note
that you can use List.partition in your solution. Make a sequential as well as a task-
parallel version and measure the speedup obtained by using the parallel version. The
speedup for the list-based version should be smaller than that for the array-based version
due to garbage collection (that is a sequential component) and due to the sequential
operation that appends two sorted sub-lists.
 Appendix A
Programs from the keyword example

This appendix provides complete programs of the keyword program from the Chapter 10.
It consists of

a section introducing the basic HTML concepts,

a section containing the complete IndexGen program, and
a section containing the complete NextLevelRefs program.

The remaining program for the keyword example: MakeWebCat, appears in Table 10.19. The
source can also be found on the homepage of the book.
 A.1 Web source files
The source of a web-page is a file encoded in the HTML (Hyper Text Mark-up Language)
format. This section gives a brief introduction to HTML using the library documentation web-
page as an example.
An HTML file is an ordinary text file using a special syntax. Certain characters like <, >, &
and " are delimiters defining the syntactical structure. The file consists of elements of the
form <…> intermixed with text to be displayed. The following construction will for instance
make a button with a link to a web-page:
<a href="http://msdn.microsoft.com/en-us/library/ee353439.aspx">
active pattern</a>

The text:
active pattern

is displayed in the button and a click will cause the browser to select the web-page given by
the URI:
http://msdn.microsoft.com/en-us/library/ee353439.aspx

A link is hence defined by a pair of elements: a start element <a…> and an end element </a>
surrounding the text to be displayed. The construction:
href=" . . . "

defines the href attribute of the element <a…>. Attributes have special uses and are not
displayed text.
Elements in HTML appear in pairs of start and end elements, and some elements may
contain attributes. The line break element <br /> is considered a (degenerated) pair of start
and end element <br></br>.
Text to be displayed is encoded in the HTML encoding with certain characters encoded
using HTML escape sequences like &lt; and &amp; encoding < and &. The internet browser
performs the corresponding decoding when displaying a text.
The HTML notation has developed over time and web-pages around the world follow
different standards. The standard is now controlled by the World Wide Web Consortium
(W3C) and more recent standards define HTML as a specialization of the XML notation.
The HTML-source of the library keyword index page starts with the Document type
definition that is an XML <!DOCTYPE…> element:
<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01//EN"
"http://www.w3.org/TR/html4/strict.dtd">

The start of the HTML part is signalled by:

<html>

The heading starts with the title to be displayed on the boundary of the browser window:
<head>
<title>F# Program Library Documentation Keyword Index</title>

It is followed by the style section:

<style type = "text/css">
h1 {color: purple; font-size: x-large; font-family: Verdana}
p {font-family: Verdana; font-size: large; color: maroon}
a {font-family: Verdana; text-decoration: none;
font-size: medium}
</style>
</head>

This section defines the appearance of different parts of the web-page:

h1: Level 1 heading in purple with x-large Verdana font.
p: Paragraphs in large Verdana font in maroon colour
a: Links in medium-sized Verdana font without the default underlining.

The reader may consult a Cascading Style Sheet (CSS) manual for further information about
styles in HTML.
The body starts with a level 1 heading <h1>…</h1> and a paragraph <p>…</p>:
<body>
<h1>F# Program Library Documentation Keyword Index</h1>
<p>Version date: Saturday, August 27, 2011</p>

Each link is followed by a line break <br />:

<a href="http://msdn.microsoft.com/en-us/library/ee353439.aspx">
active pattern</a><br />

A empty keyword line is just an extra line break:

<br />

Document body and entire HTML document are terminated by end elements:
</body></html>

HTML-sources may contain links with an abbreviated reference. The web-page with URI:
http://msdn.microsoft.com/en-us/library/ee353439.aspx

do for instance contain a link with a path instead of a full URI:

href="/en-us/library/ee370230"

This path is interpreted relative to the base URI:

http://msdn.microsoft.com/

as pointing to the web-page with URI:

http://msdn.microsoft.com/en-us/library/ee370230

The reader may consult an HTML (or XHTML) manual for further information.
 A.2 The IndexGen program
This section contains the complete IndexGen program. For the documentation of the
program, we refer to Section 10.8. The source code is split into an input and an output parts,
that are shown in following the two tables.

Table A.1 The IndexGen program: Input part

Table A.2 The IndexGen program: Output part
 A.3 The NextLevelRefs program
This section contains the complete NextLevelRefs program. For the documentation of the
program, we refer to Section 10.9.
Table A.3 The NextLevelRefs program
 Appendix B
The TextProcessing library

This appendix contains the source code of the TextProcessing library that was introduced in
Chapter 10. It consists of a signature file TextProcessing.fsi and an implementation file
TextProcessing.fs. This library is organized into four groups:

A group on regular expressions. This group is documented on Page 224. See Table 10.4.
A group on file functions. This group is documented on Page 230. See Table 10.6.
A group on file handling. This group is documented on Page 230. See Table 10.8.
A group on culture-dependent string ordering. This group is documented in Section 10.6.
See Table 10.9.

The interface file TextProcessing.fsi is given in Table B.1. The listing of the
implementation file TextProcessing.fs is split into four tables: Table B.2 – B.5, one for
each of the above-mentioned groups. The source can also be found on the homepage of the
book.
Table B.1 The file TextProcessing.fsi

Table B.2 The file TextProcessing.fsi – Regular expression

Table B.3 The file TextProcessing.fs – File functions

Table B.4 The file TextProcessing.fs – File handling

Table B.5 The file TextProcessing.fs – Culture-dependent string ordering

 Appendix C
The dialogue program from Chapter 13

This appendix contains the complete program for the skeleton program shown in Table 13.6.
The reader should consult Section 13.5 for further information.
Table C.1 Dialogue program for automaton in Figure 13.4
 References

[1] Harold Abelson, Gerald Jay Sussman, Structure and Interpretation of Computer Programs,
second edition, The MIT Press, Cambridge, MA, USA, 1996.
[2] Alfred Aho, Monica S. Lam, Ravi Sethi, Jeffrey D. Ullman, Compilers: Principles,
Techniques and Tools, second edition, Pearson Addison-Wesley, Boston, MA, USA, 2006.
[3] Guy Cousineau, Michel Mauny, The Functional Approach to Programming, Cambridge
University Press, Cambridge, United Kingdom, 1998.
[4] Maarten M. Fokkinga, Functioneel programmeren in een vogelvlucht, INFORMATIE, vol.
27, pp. 862–873, Kluwer b.v., Deventer, The Netherlands, 1985.
[5] Michael R. Hansen, Hans Rischel, Introduction to Programming using SML, Addison-
Wesley Longman, Harlow, England, 1999.
[6] Peter Henderson, Functional Geometry, Proceedings of the 1982 ACM Symposium on LISP
and Functional Programming, pp. 179–187, ACM, Pittsburgh, PA, USA, 1982.
[7] Graham Hutton, Erik Meijer, Monadic Parsing in Haskell, Journal of Functional
Programming, vol. 8, pp. 437–444, Cambridge University Press, Cambridge, United Kingdom,
1998.
[8] Robin Milner, Mads Tofte, Robert Harper, David MacQueen, The Definition of Standard
ML, revised edition, The MIT Press, Cambridge, MA, USA, 1997.
[9] Microsoft Development Network MSDN, on the internet.
[10] L.C. Paulson, ML for the Working Programmer, second edition, Cambridge University
Press, Cambridge, United Kingdom, 1996.
[11] Peter Sestoft, Henrik I. Hansen, C# Precisely, second edition, The MIT Press, Cambridge,
MA, USA, 2012.
[12] Peter Sestoft, Programming Language Concepts for Software Developers, Springer,
London, England, 2012.
[13] Don Syme, Adam Granicz, Antonio Cisternino, Expert F# 2.0, Apress, New York, NY,
USA, 2010.
[14] Simon Thompson, Haskell. The Craft of Functional programming, third edition, Addison-
Wesley Longman, Harlow, England, 2011.
[15] Philip Wadler, Monads for functional programming, Advanced Functional Programming,
Proceedings of the Båstad Spring School, May 1995, Lecture Notes in Computer Science 925,
Springer, Berlin, Heidelberg, Germany, 1995.
The URL of [9] is found on the home page of the book (see Page x).
 Index

(), 23
(*, 3
->, 3, 32
*, 11
*), 3
.[…], 27, 186
//, 3
///, 3
::, 71, 80
:=, 182
;, 179
;;, 1
@, 79, 80
@′′…′′, 26
|>, 38
~, 35
<-, 176
<|, 38
<<, 32

accumulating parameter, 206, 209

add, 106, 114
aggregation, see database
alias, 182
and, 141
anonymous function, see function
append, 79
append, 254
argument pattern, 8
Array, 186
array, 186
assignment, 176
association, see operator
Async, 313
async, 316
Async<′a>, 316
asynchronous computation, see computation expression
asynchronous event queue, 321
asynchronous operation, 313
binary tree, 131, 133
Bind, see computation expression
binding, 14
bool, see type constructor, 62
builder class, see computation expression
builder object, see computation expression
cache, see memory
cache, 254
cached sequence, see sequence
cancellation, 319
cancellation continuation, 319
cancellation token, 319
Cartesian product, 44
Cash register, 116
cash register, 82
catch exception, see exception
char, see type constructor
character, 24
Chinese box, 121
class, 156
clause, 8, 74
closure, see function
coercion rule, 177
collect, 254
collection
iteration function, 184
Collection Library
List, 93
Map, 114
Seq, 254
Set, 104
Combine, see computation expression
comment, 3
compare, 37, 70
compare function, 36
comparison constraint, 36
compilation, see module
composite name, see module
computation
delay, 288
start, 288
computation expression, 279
Combine, 284
asynchronous, 316
Bind, 284, 286, 293
builder class, 280, 284
builder object, 280, 284
computation, 280
Delay, 284, 290
do!, 284
For, 282, 284, 291
if-then, 284
laws, 291
let!, 284
Return, 286, 293
return!, 284
 ReturnFrom, 284, 286
translation, 282, 284
try-finally, 284
TryFinally, 284
use, 284
Using, 284
While, 284
while, 284
Yield, 284, 291
yield, 284
yield!, 284
YieldFrom, 284
Zero, 284, 286
conjunction, see expression
cons, 71
constructor, 58, 59, 123
contains, 106
containsKey, 114
ContentsOf, 176
context-free grammar, 295
continuation, 212
conversion, 235
count, 106
Culture dependent information, 232
customizing equality, see override
customizing hash, see override
customizing indexing, see override
customizing ordering, see override
customizing string, see override
data capture, 221, 224
data hiding, 150
data.format, 235
database, 267
aggregation, 272
creation, 274
 join, 271
projection, 270
selection, 271
update, 274
date, 235, 236
DateTime, 236
deadlock, 315
debugging, 316
decimal point, 21
declaration, 1
exception, 53
execution, 14
infix operator, 35
local, 54
prefix operator, 35
recursive function, 8
type, 58
value, 2
default value, 179
Delay, see computation expression
dialogue, 324
dialogue design, 324
dictionary imperative, 188
difference, 108
digit, 21, 24
directory
current, 264
disjunction, see expression
division
quotient, 15, 22
remainder, 15, 22
dll file, see module 152
do!, see computation expression

empty list, 68
entry, see map
enumeration
type, 62
enumeration type, see type
enumerator function, 190, 192
environment, 14
actual, 15
Basis Environment, 15
local, 50
equality, 36, 45, 51, 69
equality constraint, 36
equality type, see type
error handling, 53
escape sequence, 25
Euclid′s algorithm, 15
evaluation, 8
environment, 17
infinite, 10
iterative function, 210
event-driven program, 324
exception, 53, 63
catch, 63
declaration, see declaration
failwith, 53
raise, 53
execution, 14
exists, 95, 108, 114
exponent, 21
expression
conjunction, 23
disjunction, 23
if-then-else, 28
let, 54
negation, 23
expression tree, 137
failwith, see exception
false, 22, 62
File, 229, 230
Close, 229
CreateText, 229
Delete, 230
EndOfStream, 229
Exists, 230
Flush, 229
Move, 230
OpenText, 229
ReadLine, 229
Replace, 230
StreamReader, 229
StreamWriter, 229
Write, 229
WriteLine, 229
file handling, 230
filter, 95, 108, 114, 254
find, 114
finite tree, 121
float, see type constructor 3, see type constructor
fold, 97, 108, 114
foldBack, 97, 108, 114
For, see computation expression
forall, 95, 108, 114
format.data, 235
fprintf, 235
free-standing program, see program
fs file, see module
fsi file, see module
function
anomymous, 4
application, 13, 14
closure, 34
composition, 32
 declaration, 2
higher order, see higher-order function
partial, 64
polymorphic, 48
function composition, see function
garbage, 200, 202
garbage collection, 200, 202, 203
gcd, 15
grammar, 293, 295
left recursion, 297
guarded pattern, 37, 65
hash set, 188
head, 68
heap, 198
limit, 204
hiding, 150
higher-order function, 31
declaration, 33
histogram, 188
html, 339
HTML-source, 243
I/O, 229
identifier, 1
IDisposable, 232
IEnumerable, 192
IEnumerable<′a>, 266
IEnumerator, 192
if-then, see computation expression
ignore function, 179
imperative features, 175
imperative map, 188
imperative set, 188
implementation file, see module
incomplete pattern, see pattern
indentation, 55
infix, 49
infix operator, 35
infix operators, 153
infix precedence, 49
init, 254
initInfinite, 254
int, see type constructor, see type constructor
integer, 21
intersect, 108
invariant, 57, 60
IQueryable<T>, 270
isSubset, 106
it, 2
iteration, 209, 211
iterative function, 197, 206, 209
join, see database
key, see map
label, 51
layered pattern, see pattern
let, see expression
let!, see computation expression
lexicographical ordering, 36, 46
library
construction of, 149
List, see Collection Library
list, 67, 93
collect, 94
equality, 69
exists, 95
expression, 73
filter, 95
fold, 97
foldBack, 97
 forall, 95
map, 94
ordering, 70
pattern, 71
tryFind, 95
list, see type constructor
list pattern, see list
list recursion, see recursion
List.iter, 184
List.iteri, 184
local declaration, see declaration
location, 175
loop, 211
main memory, see memory
Map, see Collection Library
map, 113
add, 114
containsKey, 114
entry, 113
exists, 114
filter, 114
find, 114
fold, 114
foldBack, 114
forall, 114
key, 113
map, 114
ofList, 114
toList, 114
tryFind, 114
value, 113
map, 94, 108, 114, 254
Map colouring, 110
map colouring, 85
Map.iter, 184
match on type, 160
maxElement, 106
memory
cache, 311
main, 311
memory management, 198
minElement, 106
minus sign, 21
module, 149, 150
compilation, 150, 152
composite name, 152
dll file, 152
fs file, 152
fsi file, 152
implementation file, 150
library, 150, 152
pameterized, 157
signature file, 150
monad, 291
monadic parser, 293
monomorphic expression, 82
multi-core processor, 311
mutable, 175
mutable record field, see record
mutex, 314
mutex object, 314
mutually recursive functions, 141
mutual recursion, 138
n-tuple, see tuple 43
National string ordering, 232
natural number, 21
negation, see expression
Newton-Raphson approximation, 260
None, 64
object, 156
ofList, 106, 114
operand, 22
operator, 22
association, 23, 35, 71
dyadic, 22
infix, 23, 35
monadic, 22
overloading, 29, 31
precedence, 23, 35, 71
prefix, 35
option, 64
or pattern, see pattern
ordering, 36, 46, 70
overloaded operators, 153
overloading, see operator
override
equality, 159
hash, 159
index, 161
order, 161
string, 159
pair, 11, see tuple
panel, 165
parameterized type, see type
parser, 293, 298
partial function, see function
path and uri, 339
pattern, 4, 8, 51, 123, 129
when clause, 37
composite, 11
constructor, 60
guarded, 37
illegal, 47
incomplete, 47
 layered, 76
list pattern, 74
matching, 9, 12
or, 6
wildcard, 5
piecewise linear curves, 162
polymorphic
value, 81
polymorphic function, see function
polymorphic type, see type
polymorphism, 48, 78
value polymorphism, 81, 195
value restriction, 81, 195
precedence, see operator, 80
predicate, 23
prefix operator, 35
printf, 235
priority, see operator precedence
process, 312
product register, 267
program
free standing, 19
projection, see database
property.settting, 167
quadruple, see tuple
query expression, 270
contains, 273
count, 272
exactlyOne, 272
for, 270
join, 271
select, 270
where, 271
quotient, see division
raise, see exception
raise exception, 63
range expression, 73, 266
read web-page, 243
real number, 21
record, 50
equality, 51
field, 50
label, 50
mutable field, 180
ordering, 51
pattern, 51
type declaration, 180
record field, 50
record type, see type
recursion, 6, 138
list, 74
recursive data type, 128
recursive function, 6, 8
recursive type, see type
ref-, 182
reference, 182
Regex, 224
Length, 224
Match, 224
Replace, 224
Success, 224
regular expression, 221, 222
anchor, 223
capture, 224
match, 224
Regex, 224
remainder, see division
remove, 106
reserved word, 2
resource measurement, 203
resource measures, 197
return!, see computation expression
ReturnFrom, see computation expression
rev, 79
reverse, 79
save in file, 230
scope, 55
search tree, 133, see tree
selection, see database
selector, 51
Seq, see Collection Library
seq, see type constructor
Seq.iter, 184
Seq.iteri, 184
sequence, 251
append, 254
cache, 254
caches, 253
collect, 254
combine, 262
delay, 256
filter, 254
init, 254
initInfinite, 254
map, 254
recursion, 256
side effect, 256
specialization, 266
sequence expression, 262
for, 263
if-then, 263
if-then-else, 263
let!, 263
yield, 262
 yield!, 262
sequence expressions, 282
sequential composition, 179
Set, see Collection Library
set, 104
add, 106
contains, 106
count, 106
difference, 105
difference, 108
empty, 104
exists, 108
filter, 108
fold, 108
foldBack, 108
forall, 108
intersect, 108
intersection, 105
isSubset, 106
map, 108
maxElement, 106
minElement, 106
ofList, 106
remove, 106
subset, 104
toList, 106
union, 105
union, 108
set comprehension, 104
Seq.iter, 184
set imperative, 188
shared mutable data, 314
side-effect, 176
Sieve of Eratosthenes, 258, 263
signature file, see module
Some, 64
special character, 24
sprintf, 235
SQL database, 267
stack, 198
limit, 204
static binding, 34
store, 175
string, 26
conversion, 27
function, 26
index, 27
verbatim, 26
string, see type constructor
string matching, 224
String.Format, 237
string I/O, 229
String ordering, 232
symbolic differentiation, 127
synchronous operation, 313
table
domain, 113
tagged value, 58
tagged values
equality, 59
ordering, 59
tail, 68
tail recursion, 206
tail recursive function, see iterative function
tail-recursive function, 209, 212
task, 313
work item, 313
TextProcessing, 224, 230, 233, 346
captureCount, 224, 346
captureCountList, 224, 346
 captureList, 224, 346
captureSingle, 224, 346
fileFold, 230, 346
fileIter, 230, 346
fileXfold, 230, 346
fileXiter, 230, 346
orderCulture, 233, 346
orderString, 233, 346
restoreValue, 230, 346
saveValue, 230, 346
string, 233, 346
text I/O, 229
thread, 312
create, 314
safe collections, 315
start, 314
thread pool, 313
time, 235, 236
#time, 203
token, 293
toList, 106, 114
tree
search tree, 134
variable number of sub-trees, 138, 140
tree iteration, 185
tree traversal, 133, 139, 185
bredth first, 139
depth first, 139
imperative, 185
in-order, 133
list trees, 139
post-order, 133
pre-order, 133
triple, see tuple 43
true, 22, 62
truth value, 22
try-finally, see computation expression
TryFinally, see computation expression
tryFind, 95, 114
tuple, 43
equality, 45
expression, 44
n-tuple, 43
ordering, 46
pair, 11, 43
pattern, 46
quadruple, 43
triple, 43
type, 44
tuple equality, see tuple
tuple ordering, see tuple
type, 1, 2
augmentation, 153
check, 13
declaration, 50
equality, 36
equality type, 79
equality type variable, 79
function, 3
inference, 13, 31
instance, 84
monomorphic, 82
parameterized, 131
polymorphic, 48
record, 50
recursive, 121, 123
tree, 123
variable, 48
type, 131
type constructor
 bool, 2, 22
char, 24
float, 3, 21
int, 2, 21
list, 68
seq, 251
string, 26
unit, 23
type declaration
tagged values, 58
type extension, 155
type provider, 267
SQL, 268
type variable, 36
comparison, 37
equality, 36
type variable, see type
union, 108
unit, see type constructor
URI, 339
use, see computation expression
use-binding, 232
use keyword, 232
Using, see computation expression

value, 1
value polymorphism, see polymorphism
value restriction, 81, 195
variable, 175
Venn diagram, 105
verbatim string, see string
web page format, 339
when, 37
While, see computation expression
while, 183
while loop, 183
wildcard, see pattern
window, 165
work item, see task
workflows, 279
Yield, see computation expression
yield, see computation expression
yield!, see computation expression
YieldFrom, see computation expression

Zero, see computation expression