100% found this document useful (1 vote)

329 views

Elementary Algorithms

A book about algorithms

Uploaded by

Aditya Paliwal

Available Formats

Download as PDF, TXT or read online on Scribd

100% found this document useful (1 vote)

329 views

Elementary Algorithms

A book about algorithms

Uploaded by

Aditya Paliwal

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 618

Elementary Algorithms

Larry LIU Xinyu

July 25, 2014

Larry LIU Xinyu

Version: 0.6180339887498949
Email: [email protected]

Contents
I

Preface
0.1
0.2

0.3

0.4
0.5
0.6

Why? . . . . . . . . . . . . . . . . . . . . . . . . . . .
The smallest free ID problem, the power of algorithms
0.2.1 Improvement 1 . . . . . . . . . . . . . . . . . .
0.2.2 Improvement 2, Divide and Conquer . . . . . .
0.2.3 Expressiveness vs. Performance . . . . . . . . .
The number puzzle, power of data structure . . . . . .
0.3.1 The brute-force solution . . . . . . . . . . . . .
0.3.2 Improvement 1 . . . . . . . . . . . . . . . . . .
0.3.3 Improvement 2 . . . . . . . . . . . . . . . . . .
Notes and short summary . . . . . . . . . . . . . . . .
Structure of the contents . . . . . . . . . . . . . . . . .
Appendix . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

Trees

1 Binary search tree, the hello world data

1.1 Introduction . . . . . . . . . . . . . . . . .
1.2 Data Layout . . . . . . . . . . . . . . . . .
1.3 Insertion . . . . . . . . . . . . . . . . . . .
1.4 Traversing . . . . . . . . . . . . . . . . . .
1.5 Querying a binary search tree . . . . . . .
1.5.1 Looking up . . . . . . . . . . . . .
1.5.2 Minimum and maximum . . . . . .
1.5.3 Successor and predecessor . . . . .
1.6 Deletion . . . . . . . . . . . . . . . . . . .
1.7 Randomly build binary search tree . . . .
1.8 Appendix . . . . . . . . . . . . . . . . . .
2 The
2.1
2.2
2.3
2.4
2.5
2.6

7
7
8
9
10
12
12
12
15
18
18
19

evolution of insertion sort

Introduction . . . . . . . . . . . . . . . .
Insertion . . . . . . . . . . . . . . . . . .
Improvement 1 . . . . . . . . . . . . . .
Improvement 2 . . . . . . . . . . . . . .
Final improvement by binary search tree
Short summary . . . . . . . . . . . . . .
3

.
.
.
.
.
.

structure
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

25
25
27
29
30
33
33
34
34
36
40
40

.
.
.
.
.
.

43
43
44
46
47
49
50

CONTENTS

3 Red-black tree, not so complex as it was thought

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Exploit the binary search tree . . . . . . . .
3.1.2 How to ensure the balance of the tree . . .
3.1.3 Tree rotation . . . . . . . . . . . . . . . . .
3.2 Definition of red-black tree . . . . . . . . . . . . .
3.3 Insertion . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Deletion . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Imperative red-black tree algorithm ? . . . . . . .
3.6 More words . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

53
53
53
54
56
58
59
63
71
73

4 AVL tree
4.1 Introduction . . . . . . . . . . . . . . . . . . .
4.1.1 How to measure the balance of a tree?
4.2 Definition of AVL tree . . . . . . . . . . . . .
4.3 Insertion . . . . . . . . . . . . . . . . . . . . .
4.3.1 Balancing adjustment . . . . . . . . .
4.3.2 Pattern Matching . . . . . . . . . . . .
4.4 Deletion . . . . . . . . . . . . . . . . . . . . .
4.5 Imperative AVL tree algorithm ? . . . . . . .
4.6 Chapter note . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

77
77
77
77
80
82
86
87
88
92

5 Trie and Patricia

5.1 Introduction . . . . . . . . . . . . . . . . . . . .
5.2 Integer Trie . . . . . . . . . . . . . . . . . . . .
5.2.1 Definition of integer Trie . . . . . . . . .
5.2.2 Insertion . . . . . . . . . . . . . . . . . .
5.2.3 Look up . . . . . . . . . . . . . . . . . .
5.3 Integer Patricia . . . . . . . . . . . . . . . . . .
5.3.1 Definition . . . . . . . . . . . . . . . . .
5.3.2 Insertion . . . . . . . . . . . . . . . . . .
5.3.3 Look up . . . . . . . . . . . . . . . . . .
5.4 Alphabetic Trie . . . . . . . . . . . . . . . . . .
5.4.1 Definition . . . . . . . . . . . . . . . . .
5.4.2 Insertion . . . . . . . . . . . . . . . . . .
5.4.3 Look up . . . . . . . . . . . . . . . . . .
5.5 Alphabetic Patricia . . . . . . . . . . . . . . . .
5.5.1 Definition . . . . . . . . . . . . . . . . .
5.5.2 Insertion . . . . . . . . . . . . . . . . . .
5.5.3 Look up . . . . . . . . . . . . . . . . . .
5.6 Trie and Patricia applications . . . . . . . . . .
5.6.1 E-dictionary and word auto-completion
5.6.2 T9 input method . . . . . . . . . . . . .
5.7 Summary . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

95
95
96
96
97
99
100
100
102
108
109
109
111
112
113
113
114
119
121
121
125
130

6 Suffix Tree
6.1 Introduction . . . . . . . . . . . . . .
6.2 Suffix trie . . . . . . . . . . . . . . .
6.2.1 Node transfer and suffix link
6.2.2 On-line construction . . . . .

.
.
.
.

133
133
134
135
136

.
.
.
.

CONTENTS
6.3
6.4

6.5

Suffix
6.3.1
Suffix
6.4.1
6.4.2
6.4.3
6.4.4
6.4.5
Notes

5
Tree . . . . . . . . . . . . . . . . . . .
On-line construction . . . . . . . . .
tree applications . . . . . . . . . . . .
String/Pattern searching . . . . . . .
Find the longest repeated sub-string
Find the longest common sub-string
Find the longest palindrome . . . . .
Others . . . . . . . . . . . . . . . . .
and short summary . . . . . . . . . .

7 B-Trees
7.1 Introduction . . . . . . . . . . . . .
7.2 Insertion . . . . . . . . . . . . . . .
7.2.1 Splitting . . . . . . . . . . .
7.3 Deletion . . . . . . . . . . . . . . .
7.3.1 Merge before delete method
7.3.2 Delete and fix method . . .
7.4 Searching . . . . . . . . . . . . . .
7.5 Notes and short summary . . . . .

III

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

141
141
150
150
152
153
155
156
156

.
.
.
.
.
.
.
.

159
159
161
161
168
168
176
181
183

Heaps

8 Binary Heaps
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Implicit binary heap by array . . . . . . . . . . . . . .
8.2.1 Definition . . . . . . . . . . . . . . . . . . . . .
8.2.2 Heapify . . . . . . . . . . . . . . . . . . . . . .
8.2.3 Build a heap . . . . . . . . . . . . . . . . . . .
8.2.4 Basic heap operations . . . . . . . . . . . . . .
8.2.5 Heap sort . . . . . . . . . . . . . . . . . . . . .
8.3 Leftist heap and Skew heap, the explicit binary heaps
8.3.1 Definition . . . . . . . . . . . . . . . . . . . . .
8.3.2 Merge . . . . . . . . . . . . . . . . . . . . . . .
8.3.3 Basic heap operations . . . . . . . . . . . . . .
8.3.4 Heap sort by Leftist Heap . . . . . . . . . . . .
8.3.5 Skew heaps . . . . . . . . . . . . . . . . . . . .
8.4 Splay heap . . . . . . . . . . . . . . . . . . . . . . . .
8.4.1 Definition . . . . . . . . . . . . . . . . . . . . .
8.4.2 Heap sort . . . . . . . . . . . . . . . . . . . . .
8.5 Notes and short summary . . . . . . . . . . . . . . . .

187
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

9 From grape to the world cup, the evolution of selection

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2 Finding the minimum . . . . . . . . . . . . . . . . . . . .
9.2.1 Labeling . . . . . . . . . . . . . . . . . . . . . . . .
9.2.2 Grouping . . . . . . . . . . . . . . . . . . . . . . .
9.2.3 performance of the basic selection sorting . . . . .
9.3 Minor Improvement . . . . . . . . . . . . . . . . . . . . .
9.3.1 Parameterize the comparator . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

189
189
189
190
191
192
194
199
201
202
203
204
206
206
208
208
215
215

sort
. . .
. . .
. . .
. . .
. . .
. . .
. . .

.
.
.
.
.
.
.

219
219
221
222
223
224
225
225

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

CONTENTS
.
.
.
.
.
.

.
.
.
.
.
.

226
227
231
231
239
240

10 Binomial heap, Fibonacci heap, and pairing heap

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . .
10.2 Binomial Heaps . . . . . . . . . . . . . . . . . . . .
10.2.1 Definition . . . . . . . . . . . . . . . . . . .
10.2.2 Basic heap operations . . . . . . . . . . . .
10.3 Fibonacci Heaps . . . . . . . . . . . . . . . . . . .
10.3.1 Definition . . . . . . . . . . . . . . . . . . .
10.3.2 Basic heap operations . . . . . . . . . . . .
10.3.3 Running time of pop . . . . . . . . . . . . .
10.3.4 Decreasing key . . . . . . . . . . . . . . . .
10.3.5 The name of Fibonacci Heap . . . . . . . .
10.4 Pairing Heaps . . . . . . . . . . . . . . . . . . . . .
10.4.1 Definition . . . . . . . . . . . . . . . . . . .
10.4.2 Basic heap operations . . . . . . . . . . . .
10.5 Notes and short summary . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.

243
243
243
243
248
258
258
260
269
271
273
275
275
276
282

9.4

9.5

9.3.2 Trivial fine tune . . . . . .

9.3.3 Cock-tail sort . . . . . . . .
Major improvement . . . . . . . .
9.4.1 Tournament knock out . . .
9.4.2 Final improvement by using
Short summary . . . . . . . . . . .

. . .
. . .
. . .
. . .
heap
. . .

. . .
. . .
. . .
. . .
sort
. . .

.
.
.
.
.
.

Queues and Sequences

285

11 Queue, not so simple as it was thought

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2 Queue by linked-list and circular buffer . . . . . . . . . .
11.2.1 Singly linked-list solution . . . . . . . . . . . . . .
11.2.2 Circular buffer solution . . . . . . . . . . . . . . .
11.3 Purely functional solution . . . . . . . . . . . . . . . . . .
11.3.1 Paired-list queue . . . . . . . . . . . . . . . . . . .
11.3.2 Paired-array queue - a symmetric implementation
11.4 A small improvement, Balanced Queue . . . . . . . . . . .
11.5 One more step improvement, Real-time Queue . . . . . .
11.6 Lazy real-time queue . . . . . . . . . . . . . . . . . . . . .
11.7 Notes and short summary . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

287
287
288
288
291
294
294
296
298
300
307
310

12 Sequences, The last brick

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
12.2 Binary random access list . . . . . . . . . . . . . . .
12.2.1 Review of plain-array and list . . . . . . . . .
12.2.2 Represent sequence by trees . . . . . . . . . .
12.2.3 Insertion to the head of the sequence . . . . .
12.3 Numeric representation for binary random access list
12.3.1 Imperative binary access list . . . . . . . . .
12.4 Imperative paired-array list . . . . . . . . . . . . . .
12.4.1 Definition . . . . . . . . . . . . . . . . . . . .
12.4.2 Insertion and appending . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

313
313
314
314
314
316
322
324
327
327
328

.
.
.
.
.
.
.
.
.
.

CONTENTS

12.4.3 random access . . . . . . . . . . . . . . . . . . . .

12.4.4 removing and balancing . . . . . . . . . . . . . . .
12.5 Concatenate-able list . . . . . . . . . . . . . . . . . . . . .
12.6 Finger tree . . . . . . . . . . . . . . . . . . . . . . . . . .
12.6.1 Definition . . . . . . . . . . . . . . . . . . . . . . .
12.6.2 Insert element to the head of sequence . . . . . . .
12.6.3 Remove element from the head of sequence . . . .
12.6.4 Handling the ill-formed finger tree when removing
12.6.5 append element to the tail of the sequence . . . . .
12.6.6 remove element from the tail of the sequence . . .
12.6.7 concatenate . . . . . . . . . . . . . . . . . . . . . .
12.6.8 Random access of finger tree . . . . . . . . . . . .
12.7 Notes and short summary . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

Sorting and Searching

328
329
331
335
335
337
340
341
346
347
349
354
365

369

13 Divide and conquer, Quick sort vs. Merge sort

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
13.2 Quick sort . . . . . . . . . . . . . . . . . . . . . . . .
13.2.1 Basic version . . . . . . . . . . . . . . . . . .
13.2.2 Strict weak ordering . . . . . . . . . . . . . .
13.2.3 Partition . . . . . . . . . . . . . . . . . . . .
13.2.4 Minor improvement in functional partition .
13.3 Performance analysis for quick sort . . . . . . . . . .
13.3.1 Average case analysis ? . . . . . . . . . . . .
13.4 Engineering Improvement . . . . . . . . . . . . . . .
13.4.1 Engineering solution to duplicated elements .
13.5 Engineering solution to the worst case . . . . . . . .
13.6 Other engineering practice . . . . . . . . . . . . . . .
13.7 Side words . . . . . . . . . . . . . . . . . . . . . . . .
13.8 Merge sort . . . . . . . . . . . . . . . . . . . . . . . .
13.8.1 Basic version . . . . . . . . . . . . . . . . . .
13.9 In-place merge sort . . . . . . . . . . . . . . . . . . .
13.9.1 Naive in-place merge . . . . . . . . . . . . . .
13.9.2 in-place working area . . . . . . . . . . . . .
13.9.3 In-place merge sort vs. linked-list merge sort
13.10Nature merge sort . . . . . . . . . . . . . . . . . . .
13.11Bottom-up merge sort . . . . . . . . . . . . . . . . .
13.12Parallelism . . . . . . . . . . . . . . . . . . . . . . .
13.13Short summary . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

371
371
371
372
373
374
377
379
380
383
383
390
394
395
395
396
403
403
404
409
411
416
419
419

14 Searching
14.1 Introduction . . . . . . . . . . . . .
14.2 Sequence search . . . . . . . . . . .
14.2.1 Divide and conquer search .
14.2.2 Information reuse . . . . . .
14.3 Solution searching . . . . . . . . .
14.3.1 DFS and BFS . . . . . . . .
14.3.2 Search the optimal solution

.
.
.
.
.
.
.

423
423
423
424
444
471
471
507

.
.
.
.
.
.
.

CONTENTS
14.4 Short summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535

Appendix

539

Appendices
A Lists
A.1 Introduction . . . . . . . . . . . . . . . . . . .
A.2 List Definition . . . . . . . . . . . . . . . . .
A.2.1 Empty list . . . . . . . . . . . . . . . .
A.2.2 Access the element and the sub list . .
A.3 Basic list manipulation . . . . . . . . . . . . .
A.3.1 Construction . . . . . . . . . . . . . .
A.3.2 Empty testing and length calculating .
A.3.3 indexing . . . . . . . . . . . . . . . . .
A.3.4 Access the last element . . . . . . . .
A.3.5 Reverse indexing . . . . . . . . . . . .
A.3.6 Mutating . . . . . . . . . . . . . . . .
A.3.7 sum and product . . . . . . . . . . . .
A.3.8 maximum and minimum . . . . . . . .
A.4 Transformation . . . . . . . . . . . . . . . . .
A.4.1 mapping and for-each . . . . . . . . .
A.4.2 reverse . . . . . . . . . . . . . . . . . .
A.5 Extract sub-lists . . . . . . . . . . . . . . . .
A.5.1 take, drop, and split-at . . . . . . . .
A.5.2 breaking and grouping . . . . . . . . .
A.6 Folding . . . . . . . . . . . . . . . . . . . . .
A.6.1 folding from right . . . . . . . . . . . .
A.6.2 folding from left . . . . . . . . . . . .
A.6.3 folding in practice . . . . . . . . . . .
A.7 Searching and matching . . . . . . . . . . . .
A.7.1 Existence testing . . . . . . . . . . . .
A.7.2 Looking up . . . . . . . . . . . . . . .
A.7.3 finding and filtering . . . . . . . . . .
A.7.4 Matching . . . . . . . . . . . . . . . .
A.8 zipping and unzipping . . . . . . . . . . . . .
A.9 Notes and short summary . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

541
541
541
542
542
543
543
544
545
546
547
549
559
563
567
567
573
575
575
577
582
582
584
587
588
588
588
589
592
594
597

GNU Free Documentation License

1. APPLICABILITY AND DEFINITIONS . . . . . . .
2. VERBATIM COPYING . . . . . . . . . . . . . . . .
3. COPYING IN QUANTITY . . . . . . . . . . . . . .
4. MODIFICATIONS . . . . . . . . . . . . . . . . . . .
5. COMBINING DOCUMENTS . . . . . . . . . . . . .
6. COLLECTIONS OF DOCUMENTS . . . . . . . . .
7. AGGREGATION WITH INDEPENDENT WORKS
8. TRANSLATION . . . . . . . . . . . . . . . . . . . .
9. TERMINATION . . . . . . . . . . . . . . . . . . . .
10. FUTURE REVISIONS OF THIS LICENSE . . . .

.
.
.
.
.
.
.
.
.
.

601
601
603
603
604
605
606
606
606
607
607

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

CONTENTS

11. RELICENSING . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607

ADDENDUM: How to use this License for your documents . . . . . . 608

CONTENTS

Part I

Preface

Elementary Algorithms

0.1

Why?

Are algorithms useful?. Some programmers say that they seldom use any
serious data structures or algorithms in real work such as commercial application
development. Even when they need some of them, they have already been
provided by libraries. For example, the C++ standard template library (STL)
provides sort and selection algorithms as well as the vector, queue, and set data
structures. It seems that knowing about how to use the library as a tool is quite
enough.
Instead of answering this question directly, I would like to say algorithms
and data structures are critical in solving interesting problems, the usefulness
of the problem set aside.
Lets start with two problems that looks like they can be solved in a bruteforce way even by a fresh programmer.

0.2

The smallest free ID problem, the power of

algorithms

This problem is discussed in Chapter 1 of Richard Birds book [1]. Its common
that applications and systems use ID (identifier) to manage objects and entities.
At any time, some IDs are used, and some of them are available for use. When
some client tries to acquire a new ID, we want to always allocate it the smallest
available one. Suppose IDs are non-negative integers and all IDs in use are kept
in a list (or an array) which is not ordered. For example:
[18, 4, 8, 9, 16, 1, 14, 7, 19, 3, 0, 5, 2, 11, 6]
How can you find the smallest free ID, which is 10, from the list?
It seems the solution is quite easy even without any serious algorithms.
1: function Min-Free(A)
2:
x0
3:
loop
4:
if x
/ A then
5:
return x
6:
else
7:
xx+1
Where the
/ is realized like below.
1: function (x,
/
X)
2:
for i 1 to |X| do
3:
if x = X[i] then
4:
return False
5:
return True
Some languages provide handy tools which wrap this linear time process. For
example in Python, this algorithm can be directly translated as the following.
def b r u t e f o r c e ( l s t ) :
i = 0
while True :
i f i not in l s t :

Preface
return i
i = i + 1

It seems this problem is trivial. However, There will be millions of IDs in a

large system. The speed of this solution is poor in such case for it takes O(n2 )
time, where n is the length of the ID list. In my computer (2 Cores 2.10 GHz,
with 2G RAM), a C program using this solution takes an average of 5.4 seconds
to search a minimum free number among 100,000 IDs1 . And it takes more than
8 minutes to handle a million numbers.

0.2.1

Improvement 1

The key idea to improve the solution is based on a fact that for a series of n
numbers x1 , x2 , ..., xn , if there are free numbers, some of the xi are outside the
range [0, n); otherwise the list is exactly a permutation of 0, 1, ..., n 1 and n
should be returned as the minimum free number. It means that max(xi ) n1.
And we have the following fact.
minf ree(x1 , x2 , ..., xn ) n

(1)

One solution is to use an array of n + 1 flags to mark whether a number in

range [0, n] is free.
1: function Min-Free(A)
2:
F [F alse, F alse, ..., F alse] where |F | = n + 1
3:
for x A do
4:
if x < n then
5:
F [x] True
6:
for i [0, n] do
7:
if F [i] = False then
8:
return i
Line 2 initializes a flag array all of False values. This takes O(n) time. Then
the algorithm scans all numbers in A and mark the relative flag to True if the
value is less than n, This step also takes O(n) time. Finally, the algorithm
performs a linear time search to find the first flag with False value. So the total
performance of this algorithm is O(n). Note that we use n + 1 flags instead of
n flags to cover the special case that sorted(A) = [0, 1, 2, ..., n 1].
Although the algorithm only takes O(n) time, it needs extra O(n) spaces to
store the flags.
This solution is much faster than the brute force one. On my computer,
the relevant Python program takes an average of 0.02 second when dealing with
100,000 numbers.
We havent fine tuned this algorithm yet. Observe that each time we have
to allocate memory to create a n + 1 elements array of flags, and release the
memory when finished. The memory allocation and release is very expensive
thus they cost us a lot of processing time.
There are two ways in which we can improve on this solution. One is to
allocate the flags array in advance and reuse it for all the calls of our function to
find the smallest free number. The other is to use bit-wise flags instead of a flag
array. The following is the C program based on these two minor improvements.
1 All

programs can be downloaded along with this series posts.

0.2. THE SMALLEST FREE ID PROBLEM, THE POWER OF ALGORITHMS15

#define N 1000000 // 1 m i l l i o n
#define WORD LENGTH s i z e o f ( int ) 8
void s e t b i t ( unsigned int b i t s , unsigned int i ) {
b i t s [ i / WORD LENGTH] |= 1<<( i % WORD LENGTH) ;
}
int t e s t b i t ( unsigned int b i t s , unsigned int i ) {
return b i t s [ i /WORD LENGTH] & (1<<( i % WORD LENGTH) ) ;
}
unsigned int b i t s [N/WORD LENGTH+ 1 ] ;
int m i n f r e e ( int xs , int n ) {
int i , l e n = N/WORD LENGTH+1;
fo r ( i =0; i <l e n ; ++i )
b i t s [ i ]=0;
fo r ( i =0; i <n ; ++i )
i f ( xs [ i ]<n )
s e t b i t ( b i t s , xs [ i ] ) ;
fo r ( i =0; i<=n ; ++i )
i f ( ! t e s t b i t ( bits , i ))
return i ;
}
This C program can handle 1,000,000 (1 million) IDs in just 0.023 second
on my computer.
The last for-loop can be further improved as seen below but this is just minor
fine-tuning.
fo r ( i =0; ; ++i )
i f ( b i t s [ i ] !=0 )
for ( j =0; ; ++j )
i f ( ! t e s t b i t ( b i t s , i WORD LENGTH+j ) )
return i WORD LENGTH+j ;

0.2.2

Improvement 2, Divide and Conquer

Although the above improvement is much faster, it costs O(n) extra spaces to
keep a check list. if n is huge number this means a huge amount of space is
wasted.
The typical divide and conquer strategy is to break the problem into some
smaller ones, and solve these to get the final answer.
We can put all numbers xi bn/2c as a sub-list A0 and put all the others
as a second sub-list A00 . Based on formula 1 if the length of A0 is exactly bn/2c,
this means the first half of numbers are full, which indicates that the minimum
free number must be in A00 and so well need to recursively seek in the shorter
list A00 . Otherwise, it means the minimum free number is located in A0 , which
again leads to a smaller problem.
When we search the minimum free number in A00 , the conditions changes
a little bit, we are not searching the smallest free number starting from 0, but

Preface

actually from bn/2c + 1 as the lower bound. So the algorithm is something like
minf ree(A, l, u), where l is the lower bound and u is the upper bound index of
the element.
Note that there is a trivial case, that if the number list is empty, we merely
return the lower bound as the result.
This divide and conquer solution can be formally expressed as a function :
minf ree(A) = search(A, 0, |A| 1)

search(A, l, u) =

l : A=
search(A00 , m + 1, u) : |A0 | = m l + 1

search(A0 , l, m) : otherwise

where
l+u
c
2
0
A = {x A x m}
A00 = {x A x > m}

m=b

It is obvious that this algorithm doesnt need any extra space2 . Each call
performs O(|A|) comparison to build A0 and A00 . After that the problem scale
halves. So the time needed for this algorithm is T (n) = T (n/2) + O(n) which
reduce to O(n). Another way to analyze the performance is by observing that
the first call takes O(n) to build A0 and A00 and the second call takes O(n/2), and
O(n/4) for the third... The total time is O(n + n/2 + n/4 + ...) = O(2n) = O(n)
.
In functional programming languages such as Haskell, partitioning a list has
already been provided in the basic library and this algorithm can be translated
as the following.
import Data.List
minFree xs = bsearch xs 0 (length xs - 1)
bsearch xs l u | xs == [] = l
| length as == m - l + 1 = bsearch bs (m+1) u
| otherwise = bsearch as l m
where
m = (l + u) div 2
(as, bs) = partition ( m) xs

0.2.3

Expressiveness vs. Performance

Imperative language programmers may be concerned about the performance of

this kind of implementation. For instance in this minimum free ID problem, the
number of recursive calls is in O(lg n) , which means the stack size consumed
is in O(lg n). Its not free in terms of space. But if we want to avoid that , we
2 Procedural programmer may note that it actually takes O(lg n) stack spaces for bookkeeping. As well see later, this can be eliminated either by tail recursion optimization, for
instance gcc -O2. or by manually changing the recursion to iteration

0.2. THE SMALLEST FREE ID PROBLEM, THE POWER OF ALGORITHMS17

can eliminate the recursion by replacing it by an iteration
following C program.

which yields the

int min_free(int xs, int n){

int l=0;
int u=n-1;
while(n){
int m = (l + u) / 2;
int right, left = 0;
for(right = 0; right < n; ++ right)
if(xs[right] m){
swap(xs[left], xs[right]);
++left;
}
if(left == m - l + 1){
xs = xs + left;
n = n - left;
l = m+1;
}
else{
n = left;
u = m;
}
}
return l;
}

This program uses a quick-sort like approach to re-arrange the array so that
all the elements before lef t are less than or equal to m; while those between
lef t and right are greater than m. This is shown in figure 1.

left

x[i]<=m

right

x[i]>m

...?...

Figure 1: Divide the array, all x[i] m where 0 i < lef t; while all x[i] > m
where lef t i < right. The left elements are unknown.
This program is fast and it doesnt need extra stack space. However, compared to the previous Haskell program, its hard to read and the expressiveness
decreased. We have to balance performance and expressiveness.
3 This is done automatically in most functional languages since our function is in tail
recursive form which lends itself perfectly to this transformation

0.3

Preface

The number puzzle, power of data structure

If the first problem, to find the minimum free number, is a some what useful in
practice, this problem is a pure one for fun. The puzzle is to find the 1,500th
number, which only contains factor 2, 3 or 5. The first 3 numbers are of course
2, 3, and 5. Number 60 = 22 31 51 , However it is the 25th number. Number
21 = 20 31 71 , isnt a valid number because it contains a factor 7. The first 10
such numbers are list as the following.
2,3,4,5,6,8,9,10,12,15
If we consider 1 = 20 30 50 , then 1 is also a valid number and it is the first
one.

0.3.1

The brute-force solution

It seems the solution is quite easy without need any serious algorithms. We can
check all numbers from 1, then extract all factors of 2, 3 and 5 to see if the left
part is 1.
1: function Get-Number(n)
2:
x1
3:
i0
4:
loop
5:
if Valid?(x) then
6:
ii+1
7:
if i = n then
8:
return x
9:
xx+1
function Valid?(x)
while x mod 2 = 0 do
12:
x x/2
13:
while x mod 3 = 0 do
14:
x x/3
15:
while x mod 5 = 0 do
16:
x x/5
17:
if x = 1 then
18:
return T rue
19:
else
20:
return F alse
This brute-force algorithm works for most small n. However, to find the
1500th number (which is 859963392), the C program based on this algorithm
takes 40.39 seconds in my computer. I have to kill the program after 10 minutes
when I increased n to 15,000.
10:
11:

0.3.2

Improvement 1

Analysis of the above algorithm shows that modular and divide calculations
are very expensive [2]. And they executed a lot in loops. Instead of checking a
number contains only 2, 3, or 5 as factors, one alternative solution is to construct
such number by these factors.

0.3. THE NUMBER PUZZLE, POWER OF DATA STRUCTURE

We start from 1, and times it with 2, or 3, or 5 to generate rest numbers.

The problem turns to be how to generate the candidate number in order? One
handy way is to utilize the queue data structure.
A queue data structure is used to push elements at one end, and pops them
at the other end. So that the element be pushed first is also be popped out first.
This property is called FIFO (First-In-First-Out).
The idea is to push 1 as the only element to the queue, then we pop an
element, times it with 2, 3, and 5, to get 3 new elements. We then push them
back to the queue in order. Note that, the new elements may have already
existed in the queue. In such case, we just drop the element. The new element
may also smaller than the others in the queue, so we must put them to the
correct position. Figure 2 illustrates this idea.
1

1*2=2

3*2=6

1*3=3

3*3=9

1*5=5

3*5=15

2*2=4

4*2=8

2*3=6

4*3=12

2*5=10

4*5=20

Figure 2: First 4 steps of constructing numbers with a queue.

1. Queue is initialized with 1 as the only element;
2. New elements 2, 3, and 5 are pushed back;
3. New elements 4, 6, and 10, are pushed back in order;
4. New elements 9 and 15 are pushed back, element 6 already exists.

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:

This algorithm is shown as the following.

function Get-Number(n)
Q N IL
Enqueue(Q, 1)
while n > 0 do
x Dequeue(Q)
Unique-Enqueue(Q, 2x)
Unique-Enqueue(Q, 3x)
Unique-Enqueue(Q, 5x)
nn1
return x
function Unique-Enqueue(Q, x)
i0
while i < |Q| Q[i] < x do
ii+1
if i < |Q| x = Q[i] then
return
Insert(Q, i, x)

Preface

The insert function takes O(|Q|) time to find the proper position and insert
it. If the element has already existed, it just returns.
A rough estimation tells that the length of the queue increase proportion to
n, (Each time, we extract one element, and pushed 3 new, the increase ratio
2), so the total running time is O(1 + 2 + 3 + ... + n) = O(n2 ).
Figure3 shows the number of queue access time against n. It is quadratic
curve which reflect the O(n2 ) performance.

Figure 3: Queue access count v.s. n.

The C program based on this algorithm takes only 0.016[s] to get the right
answer 859963392. Which is 2500 times faster than the brute force solution.
Improvement 1 can also be considered in recursive way. Suppose X is the
infinity series for all numbers which only contain factors of 2, 3, or 5. The
following formula shows an interesting relationship.
X = {1} {2x : x X} {3x : x X} {5x : x X}

(2)

Where we can define to a special form so that all elements are stored
in order as well as unique to each other. Suppose that X = {x1 , x2 , x3 ...},
Y = {y1 , y2 , y3 , ...}, X 0 = {x2 , x3 , ...} and Y 0 = {y2 , y3 , ...}. We have

X : Y =

Y : X=

{x1 , X 0 Y } : x1 < y1
X Y =

{x1 , X 0 Y 0 } : x1 = y1

{y1 , X Y 0 } : x1 > y1
In a functional programming language such as Haskell, which supports lazy
evaluation, The above infinity series functions can be translate into the following
program.
ns = 1:merge (map (2) ns) (merge (map (3) ns) (map (5) ns))
merge [] l = l
merge l [] = l
merge (x:xs) (y:ys) | x <y = x : merge xs (y:ys)

0.3. THE NUMBER PUZZLE, POWER OF DATA STRUCTURE

| x ==y = x : merge xs ys
| otherwise = y : merge (x:xs) ys

By evaluate ns !! (n-1), we can get the 1500th number as below.

>ns !! (1500-1)
859963392

0.3.3

Improvement 2

Considering the above solution, although it is much faster than the brute-force
one, It still has some drawbacks. It produces many duplicated numbers and
they are finally dropped when examine the queue. Secondly, it does linear scan
and insertion to keep the order of all elements in the queue, which degrade the
ENQUEUE operation from O(1) to O(|Q|).
If we use three queues instead of using only one, we can improve the solution
one step ahead. Denote these queues as Q2 , Q3 , and Q5 , and we initialize them
as Q2 = {2}, Q3 = {3} and Q5 = {5}. Each time we DEQUEUEed the smallest
one from Q2 , Q3 , and Q5 as x. And do the following test:
If x comes from Q2 , we ENQUEUE 2x, 3x, and 5x back to Q2 , Q3 , and
Q5 respectively;
If x comes from Q3 , we only need ENQUEUE 3x to Q3 , and 5x to Q5 ;
We neednt ENQUEUE 2x to Q2 , because 2x have already existed in Q3 ;
If x comes from Q5 , we only need ENQUEUE 5x to Q5 ; there is no need
to ENQUEUE 3x, 5x to Q3 , Q5 because they have already been in the
queues;

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:

We repeatedly ENQUEUE the smallest one until we find the n-th element.
The algorithm based on this idea is implemented as below.
function Get-Number(n)
if n = 1 then
return 1
else
Q2 {2}
Q3 {3}
Q5 {5}
while n > 1 do
x min(Head(Q2 ), Head(Q3 ), Head(Q5 ))
if x = Head(Q2 ) then
Dequeue(Q2 )
Enqueue(Q2 , 2x)
Enqueue(Q3 , 3x)
Enqueue(Q5 , 5x)
else if x = Head(Q3 ) then
Dequeue(Q3 )
Enqueue(Q3 , 3x)
Enqueue(Q5 , 5x)
else
Dequeue(Q5 )

Preface

2*min=4

3*min=6

5*min=10

3*min=9

min=2

5*min=15

min=3

2*min=8

3*min=12

5*min=20

min=4

5*min=25

min=5

Figure 4: First 4 steps of constructing numbers with Q2 , Q3 , and Q5 .

1. Queues are initialized with 2, 3, 5 as the only element;
2. New elements 4, 6, and 10 are pushed back;
3. New elements 9, and 15, are pushed back;
4. New elements 8, 12, and 20 are pushed back;
5. New element 25 is pushed back.

0.3. THE NUMBER PUZZLE, POWER OF DATA STRUCTURE

Enqueue(Q5 , 5x)
22:
nn1
23:
return x
This algorithm loops n times, and within each loop, it extract one head
element from the three queues, which takes constant time. Then it appends
one to three new elements at the end of queues which bounds to constant time
too. So the total time of the algorithm bounds to O(n). The C++ program
translated from this algorithm shown below takes less than 1 s to produce the
1500th number, 859963392.
21:

typedef unsigned long Integer;

Integer get_number(int n){
if(n==1)
return 1;
queue<Integer> Q2, Q3, Q5;
Q2.push(2);
Q3.push(3);
Q5.push(5);
Integer x;
while(n-- > 1){
x = min(min(Q2.front(), Q3.front()), Q5.front());
if(x==Q2.front()){
Q2.pop();
Q2.push(x2);
Q3.push(x3);
Q5.push(x5);
}
else if(x==Q3.front()){
Q3.pop();
Q3.push(x3);
Q5.push(x5);
}
else{
Q5.pop();
Q5.push(x5);
}
}
return x;
}

This solution can be also implemented in Functional way. We define a function take(n), which will return the first n numbers contains only factor 2, 3, or
5.
take(n) = f (n, {1}, {2}, {3}, {5})
Where

f (n, X, Q2 , Q3 , Q5 ) =

X : n=1
f (n 1, X {x}, Q02 , Q03 , Q05 ) : otherwise

x = min(Q21 , Q31 , Q51 )

Preface

{Q22 , Q23 , ...} {2x}, Q3 {3x}, Q5 {5x}

Q2 , {Q32 , Q33 , ...} {3x}, Q5 {5x}
Q02 , Q03 , Q05 =

Q2 , Q3 , {Q52 , Q53 , ...} {5x}

: x = Q21
: x = Q31
: x = Q51

And these functional definition can be realized in Haskell as the following.

ks 1 xs _ = xs
ks n xs (q2, q3, q5) = ks (n-1) (xs++[x]) update
where
x = minimum $ map head [q2, q3, q5]
update | x == head q2 = ((tail q2)++[x2], q3++[x3], q5++[x5])
| x == head q3 = (q2, (tail q3)++[x3], q5++[x5])
| otherwise = (q2, q3, (tail q5)++[x5])
takeN n = ks n [1] ([2], [3], [5])

Invoke last takeN 1500 will generate the correct answer 859963392.

0.4

Notes and short summary

If review the 2 puzzles, we found in both cases, the brute-force solutions are so
weak. In the first problem, its quite poor in dealing with long ID list, while in
the second problem, it doesnt work at all.
The first problem shows the power of algorithms, while the second problem
tells why data structure is important. There are plenty of interesting problems,
which are hard to solve before computer was invented. With the aid of computer and programming, we are able to find the answer in a quite different way.
Compare to what we learned in mathematics course in school, we havent been
taught the method like this.
While there have been already a lot of wonderful books about algorithms,
data structures and math, however, few of them provide the comparison between
the procedural solution and the functional solution. From the above discussion,
it can be found that functional solution sometimes is very expressive and they
are close to what we are familiar in mathematics.
This series of post focus on providing both imperative and functional algorithms and data structures. Many functional data structures can be referenced
from Okasakis book[6]. While the imperative ones can be founded in classic
text books [2] or even in WIKIpedia. Multiple programming languages, including, C, C++, Python, Haskell, and Scheme/Lisp will be used. In order to make
it easy to read by programmers with different background, pseudo code and
mathematical function are the regular descriptions of each post.
The author is NOT a native English speaker, the reason why this book is
only available in English for the time being is because the contents are still
changing frequently. Any feedback, comments, or criticizes are welcome.

0.5

Structure of the contents

In the following series of post, Ill first introduce about elementary data structures before algorithms, because many algorithms need knowledge of data structures as prerequisite.

0.6. APPENDIX

The hello world data structure, binary search tree is the first topic; Then
we introduce how to solve the balance problem of binary search tree. After
that, Ill show other interesting trees. Trie, Patricia, suffix trees are useful in
text manipulation. While B-trees are commonly used in file system and data
base implementation.
The second part of data structures is about heaps. Well provide a general Heap definition and introduce about binary heaps by array and by explicit
binary trees. Then well extend to K-ary heaps including Binomial heaps, Fibonacci heaps, and pairing heaps.
Array and queues are considered among the easiest data structures typically,
However, well show how difficult to implement them in the third part.
As the elementary sort algorithms, well introduce insertion sort, quick sort,
merge sort etc in both imperative way and functional way.
The final part is about searching, besides the element searching, well also
show string matching algorithms such as KMP.
All the posts are provided under GNU FDL (Free document license), and
programs are under GNU GPL.

0.6

Appendix

All programs provided along with this book are free for downloading. download
position: http://sites.google.com/site/algoxy/home

Preface

Bibliography
[1] Richard Bird. Pearls of functional algorithm design. Cambridge University Press; 1 edition (November 1, 2010). ISBN-10: 0521513383
[2] Jon Bentley. Programming Pearls(2nd Edition). Addison-Wesley Professional; 2 edition (October 7, 1999). ISBN-13: 978-0201657883
[3] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[4] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.

BIBLIOGRAPHY

Part II

Trees

Chapter 1

Binary search tree, the

hello world data structure
1.1

Introduction

Arrays or lists are typically considered the hello world data structures. However, well see they are not actually particularly easy to implement. In some
procedural settings, arrays are the most elementary data structures, and it is
possible to implement linked lists using arrays (see section 10.3 in [2]). On the
other hand, in some functional settings, linked lists are the elementary building
blocks used to create arrays and other data structures.
Considering these factors, we start with Binary Search Trees (or BST) as the
hello world data structure using an interesting problem Jon Bentley mentioned
in Programming Pearls [2]. The problem is to count the number of times each
word occurs in a large text. One solution in C++ is below:
int main(int, char ){
map<string, int> dict;
string s;
while(cin>>s)
++dict[s];
map<string, int>::iterator it=dict.begin();
for(; it!=dict.end(); ++it)
cout<<itfirst<<": "<<itsecond<<"n";
}

And we can run it to produce the result using the following UNIX commands
1

$ g++ wordcount.cpp -o wordcount

$ cat bbe.txt | ./wordcount > wc.txt
The map provided in the standard template library is a kind of balanced
BST with augmented data. Here we use the words in the text as the keys and
the number of occurrences as the augmented data. This program is fast, and
1 This is not a UNIX unique command, in Windows OS, it can be achieved by:
type bbe.txt|wordcount.exe > wc.txt

32CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

it reflects the power of BSTs. Well introduce how to implement BSTs in this
section and show how to balance them in a later section.
Before we dive into BSTs, lets first introduce the more general binary tree.
Binary trees are recursively defined. BSTs are just one type of binary tree.
A binary tree is usually defined in the following way.
A binary tree is
either an empty node;
or a node containing 3 parts: a value, a left child which is a binary tree
and a right child which is also a binary tree.
Figure 1.1 shows this concept and an example binary tree.
k

(a) Concept of binary tree

(b) An example binary tree

Figure 1.1: Binary tree concept and an example.

A BST is a binary tree where the following applies to each node:
all the values in left child tree are less than the value of this node;
the value of this node is less than any values in its right child tree.
Figure 1.2 shows an example of binary search tree. Comparing with Figure
1.1, we can see the differences in how keys are ordered between them.

1.2. DATA LAYOUT

33
4

Figure 1.2: A Binary search tree example.

1.2

Data Layout

Based on the recursive definition of BSTs, we can draw the data layout in
procedural setting with pointers as in Figure 1.3.
The node contains a field of key, which can be augmented with satellite data;
a field contains a pointer to the left child and a field point to the right child. In
order to back-track an ancestor easily, a parent field can be provided as well.
In this post, well ignore the satellite data for simple illustration purpose.
Based on this layout, the node of binary search tree can be defined in a procedural language, such as C++ as the following.
template<class T>
struct node{
node(T x):key(x), left(0), right(0), parent(0){}
~node(){
delete left;
delete right;
}
node left;
node right;
node parent; //parent is optional, its helpful for succ/pred
T key;
};

There is another setting, for instance in Scheme/Lisp languages, the elementary data structure is linked-list. Figure 1.4 shows how a binary search tree
node can be built on top of linked-list.

34CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

key + satellite data

left
right
parent

key + satellite data

left

right

parent

...

Figure 1.3: Layout of nodes with parent field.

key

left ...

right ...

NIL

Figure 1.4: Binary search tree node layout on top of linked list. Where left...
and right ... are either empty or binary search tree node composed in the same
way.

1.3. INSERTION

Because in pure functional setting, Its hard to use pointer for back tracking
the ancestors, (and typically, there is no need to do back tracking, since we can
provide top-down solution in recursive way) there is not parent field in such
layout.
For simplified reason, well skip the detailed layout in the future, and only
focus on the logic layout of data structures. For example, below is the definition
of binary search tree node in Haskell.
data Tree a = Empty
| Node (Tree a) a (Tree a)

1.3

Insertion

To insert a key k (may be along with a value in practice) to a binary search tree
T , we can follow a quite straight forward way.
If the tree is empty, then construct a leave node with key=k;
If k is less than the key of root node, insert it to the left child;
If k is greater than the key of root, insert it to the right child;
There is an exceptional case that if k is equal to the key of root, it means it
has already existed, we can either overwrite the data, or just do nothing. For
simple reason, this case is skipped in this post.
This algorithm is described recursively. It is so simple that is why we consider
binary search tree is hello world data structure. Formally, the algorithm can
be represented with a recursive function.

insert(T, k) =

node(, k, ) : T =
node(insert(L, k), Key, R) : k < Key

node(L, Key, insert(R, k)) : otherwise

(1.1)

Where
L = lef t(T )
R = right(T )
Key = key(T )
The node function creates a new node with given left sub-tree, a key and a
right sub-tree as parameters. means NIL or Empty. function lef t, right and
key are access functions which can get the left sub-tree, right sub-tree and the
key of a node.
Translate the above functions directly to Haskell yields the following program.
insert::(Ord a) Tree a a Tree a
insert Empty k = Node Empty k Empty
insert (Node l x r) k | k < x = Node (insert l k) x r
| otherwise = Node l x (insert r k)

This program utilized the pattern matching features provided by the language. However, even in functional settings without this feature, for instance,
Scheme/Lisp, the program is still expressive.

36CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

(define (insert tree x)
(cond ((null? tree) (list () x ()))
((< x (key tree))
(make-tree (insert (left tree) x)
(key tree)
(right tree)))
((> x (key tree))
(make-tree (left tree)
(key tree)
(insert (right tree) x)))))

It is possible to turn the algorithm completely into imperative way without

recursion.
1: function Insert(T, k)
2:
root T
3:
x Create-Leaf(k)
4:
parent N IL
5:
while T 6= N IL do
6:
parent T
7:
if k < Key(T ) then
8:
T Left(T )
9:
else
10:
T Right(T )
11:
Parent(x) parent
12:
if parent = N IL then
. tree T is empty
13:
return x
14:
else if k < Key(parent) then
15:
Left(parent) x
16:
else
17:
Right(parent) x
18:
return root
function Create-Leaf(k)
x Empty-Node
21:
Key(x) k
22:
Left(x) N IL
23:
Right(x) N IL
24:
Parent(x) N IL
25:
return x
Compare with the functional algorithm, it is obviously that this one is more
complex although it is fast and can handle very deep tree. A complete C++
program and a python program are available along with this post for reference.
19:
20:

1.4

Traversing

Traversing means visiting every element one by one in a binary search tree.
There are 3 ways to traverse a binary tree, pre-order tree walk, in-order tree
walk, and post-order tree walk. The names of these traversing methods highlight
the order of when we visit the root of a binary search tree.

1.4. TRAVERSING

Since there are three parts in a tree, as left child, the root, which contains the key and satellite data, and the right child. If we denote them as
(lef t, current, right), the three traversing methods are defined as the following.
pre-order traverse, visit current, then lef t, finally right;
in-order traverse, visit lef t , then current, finally right;
post-order traverse, visit lef t, then right, finally current.
Note that each visiting operation is recursive. And we see the order of
visiting current determines the name of the traversing method.
For the binary search tree shown in figure 1.2, below are the three different
traversing results.
pre-order traverse result: 4, 3, 1, 2, 8, 7, 16, 10, 9, 14;
in-order traverse result: 1, 2, 3, 4, 7, 8, 9, 10, 14, 16;
post-order traverse result: 2, 1, 3, 7, 9, 14, 10, 16, 8, 4;
It can be found that the in-order walk of a binary search tree outputs the
elements in increase order, which is particularly helpful. The definition of binary
search tree ensures this interesting property, while the proof of this fact is left
as an exercise of this post.
In-order tree walk algorithm can be described as the following:
If the tree is empty, just return;
traverse the left child by in-order walk, then access the key, finally traverse
the right child by in-order walk.
Translate the above description yields a generic map function

: T =
map(f, T ) =
node(l0 , k 0 , r0 ) : otherwise

(1.2)

where
l0 = map(f, lef t(T ))
r0 = map(f, right(T ))
k 0 = f (key(T ))
If we only need access the key without create the transformed tree, we can
realize this algorithm in procedural way lie the below C++ program.
template<class T, class F>
void in_order_walk(node<T> t, F f){
if(t){
in_order_walk(tleft, f);
f(tvalue);
in_order_walk(tright, f);
}
}

38CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

The function takes a parameter f, it can be a real function, or a function
object, this program will apply f to the node by in-order tree walk.
We can simplified this algorithm one more step to define a function which
turns a binary search tree to a sorted list by in-order traversing.

: T =
toList(lef t(T )) {key(T )} toList(right(T )) : otherwise
(1.3)
Below is the Haskell program based on this definition.

toList(T ) =

toList::(Ord a)Tree a [a]

toList Empty = []
toList (Node l x r) = toList l ++ [x] ++ toList r

This provides us a method to sort a list of elements. We can first build

a binary search tree from the list, then output the tree by in-order traversing.
This method is called as tree sort. Lets denote the list X = {x1 , x2 , x3 , ..., xn }.
sort(X) = toList(f romList(X))

(1.4)

And we can write it in function composition form.

sort = toList.f romList
Where function f romList repeatedly insert every element to a binary search
tree.
f romList(X) = f oldL(insert, , X)

(1.5)

It can also be written in partial application form like below.

f romList = f oldL

insert

For the readers who are not familiar with folding from left, this function can
also be defined recursively as the following.

f romList(X) =

: X=
insert(f romList({x2 , x3 , ..., xn }), x1 ) : otherwise

Well intense use folding function as well as the function composition and
partial evaluation in the future, please refer to appendix of this book or [6] [7]
and [8] for more information.

Exercise 1.1
Given the in-order traverse result and pre-order traverse result, can you reconstruct the tree from these result and figure out the post-order traversing
result?
Pre-order result: 1, 2, 4, 3, 5, 6; In-order result: 4, 2, 1, 5, 3, 6; Post-order
result: ?
Write a program in your favorite language to re-construct the binary tree
from pre-order result and in-order result.

1.5. QUERYING A BINARY SEARCH TREE

Prove why in-order walk output the elements stored in a binary search
tree in increase order?
Can you analyze the performance of tree sort with big-O notation?

1.5

Querying a binary search tree

There are three types of querying for binary search tree, searching a key in the
tree, find the minimum or maximum element in the tree, and find the predecessor
or successor of an element in the tree.

1.5.1

Looking up

According to the definition of binary search tree, search a key in a tree can be
realized as the following.
If the tree is empty, the searching fails;
If the key of the root is equal to the value to be found, the search succeed.
The root is returned as the result;
If the value is less than the key of the root, search in the left child.
Else, which means that the value is greater than the key of the root, search
in the right child.
This algorithm can be described with a recursive

T :
lookup(T, x) =
lookup(lef t(T ), x) :

lookup(right(T ), x) :

function as below.
T =
key(T ) = x
x < key(T )
otherwise

(1.6)

In the real application, we may return the satellite data instead of the node
as the search result. This algorithm is simple and straightforward. Here is a
translation of Haskell program.
lookup::(Ord a) Tree a a Tree a
lookup Empty _ = Empty
lookup t@(Node l k r) x | k == x = t
| x < k = lookup l x
| otherwise = lookup r x

If the binary search tree is well balanced, which means that almost all nodes
have both non-NIL left child and right child, for N elements, the search algorithm takes O(lg N ) time to perform. This is not formal definition of balance.
Well show it in later post about red-black-tree. If the tree is poor balanced,
the worst case takes O(N ) time to search for a key. If we denote the height of
the tree as h, we can uniform the performance of the algorithm as O(h).
The search algorithm can also be realized without using recursion in a procedural manner.
1: function Search(T, x)
2:
while T 6= N IL Key(T ) 6= x do

40CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

3:
4:
5:
6:
7:

if x < Key(T ) then

T Left(T )
else
T Right(T )
return T
Below is the C++ program based on this algorithm.

template<class T>
node<T> search(node<T> t, T x){
while(t && tkey!=x){
if(x < tkey) t=tleft;
else t=tright;
}
return t;
}

1.5.2

Minimum and maximum

Minimum and maximum can be implemented from the property of binary search
tree, less keys are always in left child, and greater keys are in right.
For minimum, we can continue traverse the left sub tree until it is empty.
While for maximum, we traverse the right.

key(T ) : lef t(T ) =
min(T ) =
(1.7)
min(lef t(T )) : otherwise

key(T ) : right(T ) =
max(T ) =
(1.8)
max(right(T )) : otherwise
Both function bound to O(h) time, where h is the height of the tree. For the
balanced binary search tree, min/max are bound to O(lg N ) time, while they
are O(N ) in the worst cases.
We skip translating them to programs, Its also possible to implement them
in pure procedural way without using recursion.

1.5.3

Successor and predecessor

The last kind of querying, to find the successor or predecessor of an element

is useful when a tree is treated as a generic container and traversed by using
iterator. It will be relative easier to implement if parent of a node can be
accessed directly.
It seems that the functional solution is hard to be found, because there is no
pointer like field linking to the parent node. One solution is to left breadcrumbs
when we visit the tree, and use these information to back-track or even reconstruct the whole tree. Such data structure, that contains both the tree and
breadcrumbs is called zipper. please refer to [9] for details.
However, If we consider the original purpose of providing succ/pred function, to traverse all the binary search tree elements one by one as a generic
container, we realize that they dont make significant sense in functional settings
because we can traverse the tree in increase order by mapT function we defined
previously.

1.5. QUERYING A BINARY SEARCH TREE

Well meet many problems in this series of post that they are only valid in
imperative settings, and they are not meaningful problems in functional settings
at all. One good example is how to delete an element in red-black-tree[3].
In this section, well only present the imperative algorithm for finding the
successor and predecessor in a binary search tree.
When finding the successor of element x, which is the smallest one y that
satisfies y > x, there are two cases. If the node with value x has non-NIL right
child, the minimum element in right child is the answer; For example, in Figure
1.2, in order to find the successor of 8, we search its right sub tree for the
minimum one, which yields 9 as the result. While if node x dont have right
child, we need back-track to find the closest ancestors whose left child is also
ancestor of x. In Figure 1.2, since 2 dont have right sub tree, we go back to its
parent 1. However, node 1 dont have left child, so we go back again and reach
to node 3, the left child of 3, is also ancestor of 2, thus, 3 is the successor of
node 2.
Based on this description, the algorithm can be given as the following.
1: function Succ(x)
2:
if Right(x) 6= N IL then
3:
return Min(Right(x))
4:
else
5:
p Parent(x)
6:
while p 6= N IL and x = Right(p) do
7:
xp
8:
p Parent(p)
9:
return p
The predecessor case is quite similar to the successor algorithm, they are
symmetrical to each other.
1: function Pred(x)
2:
if Left(x) 6= N IL then
3:
return Max(Left(x))
4:
else
5:
p Parent(x)
6:
while p 6= N IL and x = Left(p) do
7:
xp
8:
p Parent(p)
9:
return p
Below are the Python programs based on these algorithms. They are changed
a bit in while loop conditions.
def succ(x):
if x.right is not None: return tree_min(x.right)
p = x.parent
while p is not None and p.left != x:
x=p
p = p.parent
return p
def pred(x):
if x.left is not None: return tree_max(x.left)
p = x.parent

42CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

while p is not None and p.right != x:
x=p
p = p.parent
return p

Exercise 1.2
Can you figure out how to iterate a tree as a generic container by using
pred()/succ()? Whats the performance of such traversing process in terms
of big-O?
A reader discussed about traversing all elements inside a range [a, b]. In
C++, the algorithm looks like the below code:
for each(m.lower bound(12), m.upper bound(26), f );
Can you provide the purely function solution for this problem?

1.6

Deletion

Deletion is another imperative only topic for binary search tree. This is because
deletion mutate the tree, while in purely functional settings, we dont modify
the tree after building it in most application.
However, One method of deleting element from binary search tree in purely
functional way is shown in this section. Its actually reconstructing the tree but
not modifying the tree.
Deletion is the most complex operation for binary search tree. this is because
we must keep the BST property, that for any node, all keys in left sub tree are
less than the key of this node, and they are all less than any keys in right sub
tree. Deleting a node can break this property.
In this post, different with the algorithm described in [2], A simpler one from
SGI STL implementation is used.[6]
To delete a node x from a tree.
If x has no child or only one child, splice x out;
Otherwise (x has two children), use minimum element of its right sub tree
to replace x, and splice the original minimum element out.
The simplicity comes from the truth that, the minimum element is stored in
a node in the right sub tree, which cant have two non-NIL children. It ends up
in the trivial case, the the node can be directly splice out from the tree.
Figure 1.5, 1.6, and 1.7 illustrate these different cases when deleting a node
from the tree.
Based on this idea, the deletion can be defined as the below function.

node(delete(L,
x),
K,
R)

node(L, K, delete(R, x))

delete(T, x) =
R

node(L, y, delete(R, y))

:
:
:
:
:
:

T =
x<K
x>K
x=K L=
x=K R=
otherwise

(1.9)

1.6. DELETION

Tree

NIL

Figure 1.5: x can be spliced out.

Tree
Tree

x
L

NIL

(a) Before delete x

(b) After delete x

x is spliced out, and replaced by its left child.

Tree
Tree

x
R

NIL

(c) Before delete x

(d) Before delete x

x is spliced out, and replaced by its right child.

Figure 1.6: Delete a node which has only one non-NIL child.

44CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

Tree

min(R)

Tree

delete(R, min(R))

(a) Before delete x

(b) After delete x

x is replaced by splicing the minimum element from its right child.

Figure 1.7: Delete a node which has both children.

Where
L = lef t(T )
R = right(T )
K = key(T )
y = min(R)
Translating the function to Haskell yields the below program.
delete::(Ord a) Tree a a Tree a
delete Empty _ = Empty
delete (Node l k r) x | x < k = (Node (delete l x) k r)
| x > k = (Node l k (delete r x))
-- x == k
| isEmpty l = r
| isEmpty r = l
| otherwise = (Node l k (delete r k))
where k = min r

Function isEmpty is used to test if a tree is empty (). Note that the
algorithm first performs search to locate the node where the element need be
deleted, after that it execute the deletion. This algorithm takes O(h) time where
h is the height of the tree.
Its also possible to pass the node but not the element to the algorithm for
deletion. Thus the searching is no more needed.
The imperative algorithm is more complex because it need set the parent
properly. The function will return the root of the result tree.
1: function Delete(T, x)
2:
root T

1.6. DELETION

x0 x
. save x
4:
parent Parent(x)
5:
if Left(x) = N IL then
6:
x Right(x)
7:
else if Right(x) = N IL then
8:
x Left(x)
9:
else
. both children are non-NIL
10:
y Min(Right(x))
11:
Key(x) Key(y)
12:
Copy other satellite data from y to x
13:
if Parent(y) 6= x then
. y hasnt left sub tree
14:
Left(Parent(y)) Right(y)
15:
else
. y is the root of right child of x
16:
Right(x) Right(y)
17:
Remove y
18:
return root
19:
if x 6= N IL then
20:
Parent(x) parent
21:
if parent = N IL then
. We are removing the root of the tree
22:
root x
23:
else
24:
if Left(parent) = x0 then
25:
Left(parent) x
26:
else
27:
Right(parent) x
28:
Remove x0
29:
return root
Here we assume the node to be deleted is not empty (otherwise we can simply
returns the original tree). In other cases, it will first record the root of the tree,
create copy pointers to x, and its parent.
If either of the children is empty, the algorithm just splice x out. If it has
two non-NIL children, we first located the minimum of right child, replace the
key of x to ys, copy the satellite data as well, then splice y out. Note that there
is a special case that y is the root node of xs left sub tree.
Finally we need reset the stored parent if the original x has only one nonNIL child. If the parent pointer we copied before is empty, it means that we are
deleting the root node, so we need return the new root. After the parent is set
properly, we finally remove the old x from memory.
The relative Python program for deleting algorithm is given as below. Because Python provides GC, we neednt explicitly remove the node from the
memory.
3:

def tree_delete(t, x):

if x is None:
return t
[root, old_x, parent] = [t, x, x.parent]
if x.left is None:
x = x.right
elif x.right is None:
x = x.left

46CHAPTER 1. BINARY SEARCH TREE, THE HELLO WORLD DATA STRUCTURE

else:
y = tree_min(x.right)
x.key = y.key
if y.parent != x:
y.parent.left = y.right
else:
x.right = y.right
return root
if x is not None:
x.parent = parent
if parent is None:
root = x
else:
if parent.left == old_x:
parent.left = x
else:
parent.right = x
return root

Because the procedure seeks minimum element, it runs in O(h) time on a

tree of height h.

Exercise 1.3
There is a symmetrical solution for deleting a node which has two non-NIL
children, to replace the element by splicing the maximum one out off the
left sub-tree. Write a program to implement this solution.

1.7

Randomly build binary search tree

It can be found that all operations given in this post bound to O(h) time for a
tree of height h. The height affects the performance a lot. For a very unbalanced
tree, h tends to be O(N ), which leads to the worst case. While for balanced
tree, h close to O(lg N ). We can gain the good performance.
How to make the binary search tree balanced will be discussed in next post.
However, there exists a simple way. Binary search tree can be randomly built as
described in [2]. Randomly building can help to avoid (decrease the possibility)
unbalanced binary trees. The idea is that before building the tree, we can call
a random process, to shuffle the elements.

Exercise 1.4
Write a randomly building process for binary search tree.

1.8

Appendix

All programs are provided along with this post. They are free for downloading.
We provided C, C++, Python, Haskell, and Scheme/Lisp programs as example.

Bibliography
[1] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. ISBN:0262032937.
The MIT Press. 2001
[2] Jon Bentley. Programming Pearls(2nd Edition). Addison-Wesley Professional; 2 edition (October 7, 1999). ISBN-13: 978-0201657883
[3] Chris Okasaki. Ten Years of Purely Functional Data Structures.
http://okasaki.blogspot.com/2008/02/ten-years-of-purely-functionaldata.html
[4] SGI.
Standard
Template
http://www.sgi.com/tech/stl/

Library

Programmers

Guide.

[5] http://en.literateprograms.org/Category:Binary search tree

[6] http://en.wikipedia.org/wiki/Foldl
[7] http://en.wikipedia.org/wiki/Function composition
[8] http://en.wikipedia.org/wiki/Partial application
[9] Miran Lipovaca. Learn You a Haskell for Great Good! A Beginners
Guide. the last chapter. No Starch Press; 1 edition April 2011, 400 pp.
ISBN: 978-1-59327-283-8

The evolution of insertion sort

Chapter 2

The evolution of insertion

sort
2.1

Introduction

In previous chapter, we introduced the hello world data structure, binary

search tree. In this chapter, we explain insertion sort, which can be think of
the hello world sorting algorithm 1 . Its straightforward, but the performance
is not as good as some divide and conqueror sorting approaches, such as quick
sort and merge sort. Thus insertion sort is seldom used as generic sorting utility
in modern software libraries. Well analyze the problems why it is slow, and
trying to improve it bit by bit till we reach the best bound of comparison based
sorting algorithms, O(n lg n), by evolution to tree sort. And we finally show the
connection between the hello world data structure and hello world sorting
algorithm.
The idea of insertion sort can be vivid illustrated by a real life poker game[2].
Suppose the cards are shuffled, and a player starts taking card one by one.
At any time, all cards in players hand are well sorted. When the player
gets a new card, he insert it in proper position according to the order of points.
Figure 2.1 shows this insertion example.
Based on this idea, the algorithm of insertion sort can be directly given as
the following.
function Sort(A)
X
for each x A do
Insert(X, x)
return X
Its easy to express this process with folding, which we mentioned in the
chapter of binary search tree.
insert = f oldL insert

(2.1)

1 Some reader may argue that Bubble sort is the easiest sort algorithm. Bubble sort isnt
covered in this book as we dont think its a valuable algorithm[1]

CHAPTER 2. THE EVOLUTION OF INSERTION SORT

Figure 2.1: Insert card 8 to proper position in a deck.

Note that in above algorithm, we store the sorted result in X, so this isnt
in-place sorting. Its easy to change it to in-place algorithm.
function Sort(A)
for i 2 to Length(A) do
insert Ai to sorted sequence {A01 , A02 , ..., A0i1 }
At any time, when we process the i-th element, all elements before i have
already been sorted. we continuously insert the current elements until consume
all the unsorted data. This idea is illustrated as in figure 9.3.

insert

... sorted elements ...

... unsorted elements ...

Figure 2.2: The left part is sorted data, continuously insert elements to sorted
part.
We can find there is recursive concept in this definition. Thus it can be
expressed as the following.

: A=
sort(A) =
(2.2)
insert(sort({A2 , A3 , ...}), A1 ) : otherwise

2.2

Insertion

We havent answered the question about how to realize insertion however. Its
a puzzle how does human locate the proper position so quickly.
For computer, its an obvious option to perform a scan. We can either scan
from left to right or vice versa. However, if the sequence is stored in plain array,
its necessary to scan from right to left.
function Sort(A)
for i 2 to Length(A) do . Insert A[i] to sorted sequence A[1...i 1]

2.2. INSERTION

x A[i]
j i1
while j > 0 x < A[j] do
A[j + 1] A[j]
j j1
A[j + 1] x
One may think scan from left to right is natural. However, it isnt as effect
as above algorithm for plain array. The reason is that, its expensive to insert an
element in arbitrary position in an array. As array stores elements continuously.
If we want to insert new element x in position i, we must shift all elements after
i, including i + 1, i + 2, ... one position to right. After that the cell at position i
is empty, and we can put x in it. This is illustrated in figure 2.3.

x
insert
A[1]

A[2]

...

A[i-1]

A[i]

A[i+1]

A[i+2]

...

A[n-1]

Figure 2.3: Insert x to array A at position i.

If the length of array is n, this indicates we need examine the first i elements,
then perform n i + 1 moves, and then insert x to the i-th cell. So insertion
from left to right need traverse the whole array anyway. While if we scan from
right to left, we totally examine the last j = n i + 1 elements, and perform
the same amount of moves. If j is small (e.g. less than n/2), there is possibility
to perform less operations than scan from left to right.
Translate the above algorithm to Python yields the following code.
def isort(xs):
n = len(xs)
for i in range(1, n):
x = xs[i]
j=i - 1
while j 0 and x < xs[j]:
xs[j+1] = xs[j]
j=j - 1
xs[j+1] = x

It can be found some other equivalent programs, for instance the following
ANSI C program. However this version isnt as effective as the pseudo code.
void isort(Key xs, int n){
int i, j;
for(i=1; i<n; ++i)
for(j=i-1; j 0 && xs[j+1] < xs[j]; --j)
swap(xs, j, j+1);
}

A[n]

empty

CHAPTER 2. THE EVOLUTION OF INSERTION SORT

This is because the swapping function, which can exchange two elements
typically uses a temporary variable like the following:
void swap(Key xs, int i, int j){
Key temp = xs[i];
xs[i] = xs[j];
xs[j] = temp;
}

So the ANSI C program presented above takes 3m times assignment, where

m is the number of inner loops. While the pseudo code as well as the Python
program use shift operation instead of swapping. There are n + 2 times assignment.
We can also provide Insert() function explicitly, and call it from the general
insertion sort algorithm in previous section. We skip the detailed realization here
and left it as an exercise.
All the insertion algorithms are bound to O(n), where n is the length of
the sequence. No matter what difference among them, such as scan from left
or from right. Thus the over all performance for insertion sort is quadratic as
O(n2 ).

Exercise 2.1
Provide explicit insertion function, and call it with general insertion sort
algorithm. Please realize it in both procedural way and functional way.

2.3

Improvement 1

Lets go back to the question, that why human being can find the proper position
for insertion so quickly. We have shown a solution based on scan. Note the fact
that at any time, all cards at hands have been well sorted, another possible
solution is to use binary search to find that location.
Well explain the search algorithms in other dedicated chapter. Binary search
is just briefly introduced for illustration purpose here.
The algorithm will be changed to call a binary search procedure.
function Sort(A)
for i 2 to Length(A) do
x A[i]
p Binary-Search(A[1...i 1], x)
for j i down to p do
A[j] A[j 1]
A[p] x
Instead of scan elements one by one, binary search utilize the information
that all elements in slice of array {A1 , ..., Ai1 } are sorted. Lets assume the
order is monotonic increase order. To find a position j that satisfies Aj1
x Aj . We can first examine the middle element, for example, Abi/2c . If x is
less than it, we need next recursively perform binary search in the first half of
the sequence; otherwise, we only need search in last half.
Every time, we halve the elements to be examined, this search process runs
O(lg n) time to locate the insertion position.
function Binary-Search(A, x)

2.4. IMPROVEMENT 2

l1
u 1+ Length(A)
while l < u do
m b l+u
2 c
if Am = x then
return m
else if Am < x then
l m+1
else
um
return l

. Find a duplicated element

The improved insertion sort algorithm is still bound to O(n2 ), compare to

previous section, which we use O(n2 ) times comparison and O(n2 ) moves, with
binary search, we just use O(n lg n) times comparison and O(n2 ) moves.
The Python program regarding to this algorithm is given below.
def isort(xs):
n = len(xs)
for i in range(1, n):
x = xs[i]
p = binary_search(xs[:i], x)
for j in range(i, p, -1):
xs[j] = xs[j-1]
xs[p] = x
def binary_search(xs, x):
l=0
u = len(xs)
while l < u:
m = (l+u)/2
if xs[m] == x:
return m
elif xs[m] < x:
l=m+1
else:
u=m
return l

Exercise 2.2
Write the binary search in recursive manner. You neednt use purely functional programming language.

2.4

Improvement 2

Although we improve the search time to O(n lg n) in previous section, the number of moves is still O(n2 ). The reason of why movement takes so long time, is
because the sequence is stored in plain array. The nature of array is continuously layout data structure, so the insertion operation is expensive. This hints
us that we can use linked-list setting to represent the sequence. It can improve

CHAPTER 2. THE EVOLUTION OF INSERTION SORT

the insertion operation from O(n) to constant time O(1).

insert(A, x) =

{x} : A =
{x} A : x < A1

{A1 } insert({A2 , A3 , ...An }, x) : otherwise

(2.3)

Translating the algorithm to Haskell yields the below program.

insert :: (Ord a) [a] a [a]
insert [] x = [x]
insert (y:ys) x = if x < y then x:y:ys else y:insert ys x

And we can complete the two versions of insertion sort program based on
the first two equations in this chapter.
isort [] = []
isort (x:xs) = insert (isort xs) x

Or we can represent the recursion with folding.

isort = foldl insert []

Linked-list setting solution can also be described imperatively. Suppose

function Key(x), returns the value of element stored in node x, and Next(x)
accesses the next node in the linked-list.
function Insert(L, x)
p N IL
HL
while L 6= N IL Key(L) < Key(x) do
pL
L Next(L)
Next(x) L
if p 6= N IL then
Hx
else
Next(p) x
return H
For example in ANSI C, the linked-list can be defined as the following.
struct node{
Key key;
struct node next;
};

Thus the insert function can be given as below.

struct node insert(struct node lst, struct node x){
struct node p, head;
p = NULL;
for(head = lst; lst && xkey > lstkey; lst = lstnext)
p = lst;
xnext = lst;
if(!p)
return x;
pnext = x;

2.5. FINAL IMPROVEMENT BY BINARY SEARCH TREE

return head;
}

Instead of using explicit linked-list such as by pointer or reference based

structure. Linked-list can also be realized by another index array. For any
array element Ai , N exti stores the index of next element follows Ai . It means
AN exti is the next element after Ai .
The insertion algorithm based on this solution is given like below.
function Insert(A, N ext, i)
j
while N extj 6= N IL AN extj < Ai do
j N extj
N exti N extj
N extj i
Here means the head of the N ext table. And the relative Python program
for this algorithm is given as the following.
def isort(xs):
n = len(xs)
next = [-1](n+1)
for i in range(n):
insert(xs, next, i)
return next
def insert(xs, next, i):
j = -1
while next[j] != -1 and xs[next[j]] < xs[i]:
j = next[j]
next[j], next[i] = i, next[j]

Although we change the insertion operation to constant time by using linkedlist. However, we have to traverse the linked-list to find the position, which
results O(n2 ) times comparison. This is because linked-list, unlike array, doesnt
support random access. It means we cant use binary search with linked-list
setting.

Exercise 2.3
Complete the insertion sort by using linked-list insertion function in your
favorate imperative programming language.
The index based linked-list return the sequence of rearranged index as
result. Write a program to re-order the original array of elements from
this result.

2.5

Final improvement by binary search tree

It seems that we drive into a corner. We must improve both the comparison
and the insertion at the same time, or we will end up with O(n2 ) performance.
We must use binary search, this is the only way to improve the comparison
time to O(lg n). On the other hand, we must change the data structure, because
we cant achieve constant time insertion at a position with plain array.

CHAPTER 2. THE EVOLUTION OF INSERTION SORT

This remind us about our hello world data structure, binary search tree. It
naturally support binary search from its definition. At the same time, We can
insert a new leaf in binary search tree in O(1) constant time if we already find
the location.
So the algorithm changes to this.
function Sort(A)
T
for each x A do
T Insert-Tree(T, x)
return To-List(T )
Where Insert-Tree() and To-List() are described in previous chapter
about binary search tree.
As we have analyzed for binary search tree, the performance of tree sort is
bound to O(n lg n), which is the lower limit of comparison based sort[3].

2.6

Short summary

In this chapter, we present the evolution process of insertion sort. Insertion sort
is well explained in most textbooks as the first sorting algorithm. It has simple
and straightforward idea, but the performance is quadratic. Some textbooks
stop here, but we want to show that there exist ways to improve it by different
point of view. We first try to save the comparison time by using binary search,
and then try to save the insertion operation by changing the data structure to
linked-list. Finally, we combine these two ideas and evolute insertion sort to
tree sort.

Bibliography
[1] http://en.wikipedia.org/wiki/Bubble sort
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. ISBN:0262032937.
The MIT Press. 2001
[3] Donald E. Knuth. The Art of Computer Programming, Volume 3: Sorting
and Searching (2nd Edition). Addison-Wesley Professional; 2 edition (May
4, 1998) ISBN-10: 0201896850 ISBN-13: 978-0201896855

Red black tree

Chapter 3

Red-black tree, not so

complex as it was thought
3.1

Introduction

3.1.1

Exploit the binary search tree

We have shown the power of binary search tree by using it to count the occurrence of every word in Bible. The idea is to use binary search tree as a dictionary
for counting.
One may come to the idea that to feed a yellow page book 1 to a binary
search tree, and use it to look up the phone number for a contact.
By modifying a bit of the program for word occurrence counting yields the
following code.
int main(int, char ){
ifstream f("yp.txt");
map<string, string> dict;
string name, phone;
while(f>>name && f>>phone)
dict[name]=phone;
for(;;){
cout<<"nname: ";
cin>>name;
if(dict.find(name)==dict.end())
cout<<"not found";
else
cout<<"phone: "<<dict[name];
}
}

This program works well. However, if you replace the STL map with the binary search tree as mentioned previously, the performance will be bad, especially
when you search some names such as Zara, Zed, Zulu.
This is because the content of yellow page is typically listed in lexicographic
order. Which means the name list is in increase order. If we try to insert a
1A

name-phone number contact list book

60CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

sequence of number 1, 2, 3, ..., n to a binary search tree. We will get a tree like
in Figure 3.1.
1

...

Figure 3.1: unbalanced tree

This is a extreme unbalanced binary search tree. The looking up performs
O(h) for a tree with height h. In balanced case, we benefit from binary search
tree by O(lg N ) search time. But in this extreme case, the search time downgraded to O(N ). Its no better than a normal link-list.

Exercise 3.1
For a very big yellow page list, one may want to speed up the dictionary
building process by two concurrent tasks (threads or processes). One task
reads the name-phone pair from the head of the list, while the other one
reads from the tail. The building terminates when these two tasks meet
at the middle of the list. What will be the binary search tree looks like
after building? What if you split the the list more than two and use more
tasks?
Can you find any more cases to exploit a binary search tree? Please
consider the unbalanced trees shown in figure 3.2.

3.1.2

How to ensure the balance of the tree

In order to avoid such case, we can shuffle the input sequence by randomized
algorithm, such as described in Section 12.4 in [2]. However, this method doesnt
always work, for example the input is fed from user interactively, and the tree
need to be built/updated after each input.
There are many solutions people have ever found to make binary search tree
balanced. Many of them rely on the rotation operations to binary search tree.
Rotation operations change the tree structure while maintain the ordering of
the elements. Thus it either improve or keep the balance property of the binary
search tree.

3.1. INTRODUCTION

n
n

n-1

n-2

n-1

...

(a)

(b)

m-1

m+1

m-2

m+2

...

(c)

Figure 3.2: Some unbalanced trees

62CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

In this chapter, well first introduce about red-black tree, which is one of
the most popular and widely used self-adjusting balanced binary search tree. In
next chapter, well introduce about AVL tree, which is another intuitive solution;
In later chapter about binary heaps, well show another interesting tree called
splay tree, which can gradually adjust the the tree to make it more and more
balanced.

3.1.3

Tree rotation

(a)

(b)

Figure 3.3: Tree rotation, rotate-left transforms the tree from left side to right
side, and rotate-right does the inverse transformation.
Tree rotation is a kind of special operation that can transform the tree
structure without changing the in-order traverse result. It based on the fact
that for a specified ordering, there are multiple binary search trees correspond
to it. Figure 3.3 shows the tree rotation. For a binary search tree on the left
side, left rotate transforms it to the tree on the right, and right rotate does the
inverse transformation.
Although tree rotation can be realized in procedural way, there exists quite
simple function description if using pattern matching.

rotateL(T ) =

node(node(a, X, b), Y, c)
T

: pattern(T ) = node(a, X, node(b, Y, c))

: otherwise
(3.1)

node(a, X, node(b, Y, c)) : pattern(T ) = node(node(a, X, b), Y, c))

T : otherwise
(3.2)
However, the pseudo code dealing imperatively has to set all fields accordingly.
1: function Left-Rotate(T, x)
2:
p Parent(x)
3:
y Right(x)
. Assume y 6= N IL
4:
a Left(x)
5:
b Left(y)
6:
c Right(y)

rotateR(T ) =

3.1. INTRODUCTION
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:

Replace(x, y)
Set-Children(x, a, b)
Set-Children(y, x, c)
if p = N IL then
T y
return T
function Right-Rotate(T, y)
p Parent(y)
x Left(y)
a Left(x)
b Right(x)
c Right(y)
Replace(y, x)
Set-Children(y, b, c)
Set-Children(x, a, y)
if p = N IL then
T x
return T

. Assume x 6= N IL

function Set-Left(x, y)
Left(x) y
if y 6= N IL then Parent(y) x
function Set-Right(x, y)
Right(x) y
if y 6= N IL then Parent(y) x
function Set-Children(x, L, R)
Set-Left(x, L)
Set-Right(x, R)
function Replace(x, y)
if Parent(x) = N IL then
if y 6= N IL then Parent(y) N IL
else if Left(Parent(x)) = x then Set-Left(Parent(x), y)
elseSet-Right(Parent(x), y)
Parent(x) N IL

Compare these pseudo codes with the pattern matching functions, the former
focus on the structure states changing, while the latter focus on the rotation
process. As the title of this chapter indicated, red-black tree neednt be so
complex as it was thought. Most traditional algorithm text books use the classic
procedural way to teach red-black tree, there are several cases need to deal and
all need carefulness to manipulate the node fields. However, by changing the
mind to functional settings, things get intuitive and simple. Although there is
some performance overhead.
Most of the content in this chapter is based on Chris Okasakis work in [2].

64CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

3.2

Definition of red-black tree

Red-black tree is a type of self-balancing binary search tree[3]. 2 By using color

changing and rotation, red-black tree provides a very simple and straightforward
way to keep the tree balanced.
For a binary search tree, we can augment the nodes with a color field, a node
can be colored either red or black. We call a binary search tree red-black tree
if it satisfies the following 5 properties[2].
1. Every node is either red or black.
2. The root is black.
3. Every leaf (NIL) is black.
4. If a node is red, then both its children are black.
5. For each node, all paths from the node to descendant leaves contain the
same number of black nodes.
Why this 5 properties can ensure the red-black tree is well balanced? Because
they have a key characteristic, the longest path from root to a leaf cant be as
2 times longer than the shortest path.
Please note the 4-th property, which means there wont be two adjacent red
nodes. so the shortest path only contains black nodes, any paths longer than
the shortest one has interval red nodes. According to property 5, all paths have
the same number of black nodes, this finally ensure there wont be any path is
2 times longer than others[3]. Figure 3.4 shows an example red-black tree.
13

Figure 3.4: An example red-black tree

2 Red-black tree is one of the equivalent form of 2-3-4 tree (see chapter B-tree about 2-3-4
tree). That is to say, for any 2-3-4 tree, there is at least one red-black tree has the same data
order.

3.3. INSERTION

All read only operations such as search, min/max are as same as in binary
search tree. While only the insertion and deletion are special.
As we have shown in word occurrence example, many implementation of
set or map container are based on red-black tree. One example is the C++
Standard library (STL)[6].
As mentioned previously, the only change in data layout is that there is color
information augmented to binary search tree. This can be represented as a data
field in imperative languages such as C++ like below.
enum Color {Red, Black};
template <class T>
struct node{
Color color;
T key;
node left;
node right;
node parent;
};

In functional settings, we can add the color information in constructors,

below is the Haskell example of red-black tree definition.
data Color = R | B
data RBTree a = Empty
| Node Color (RBTree a) a (RBTree a)

Exercise 3.2
Can you prove that a red-black tree with n nodes has height at most
2 lg(n + 1)?

3.3

Insertion

Inserting a new node as what has been mentioned in binary search tree may
cause the tree unbalanced. The red-black properties has to be maintained, so
we need do some fixing by transform the tree after insertion.
When we insert a new key, one good practice is to always insert it as a red
node. As far as the new inserted node isnt the root of the tree, we can keep all
properties except the 4-th one. that it may bring two adjacent red nodes.
Functional and procedural implementation have different fixing methods.
One is intuitive but has some overhead, the other is a bit complex but has
higher performance. Most text books about algorithm introduce the later one.
In this chapter, we focus on the former to show how easily a red-black tree
insertion algorithm can be realized. The traditional procedural method will be
given only for comparison purpose.
As described by Chris Okasaki, there are total 4 cases which violate property
4. All of them has 2 adjacent red nodes. However, they have a uniformed form
after fixing[2] as shown in figure 4.3.
Note that this transformation will move the redness one level up. So this is a
bottom-up recursive fixing, the last step will make the root node red. According

66CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

@
A

@
R
@

@
I
@

@
A

Figure 3.5: 4 cases for balancing a red-black tree after insertion

3.3. INSERTION

to property 2, root is always black. Thus we need final fixing to revert the root
color to black.
Observing that the 4 cases and the fixed result have strong pattern features,
the fixing function can be defined by using the similar method we mentioned in
tree rotation. To avoid too long formula, we abbreviate Color as C, Black as
B, and Red as R.

: match(T )
: otherwise
(3.3)
where function node() can construct a red-black tree node with 4 parameters
as color, the left child, the key and the right child. Function match() can test
if a tree is one of the 4 possible patterns as the following.
balance(T ) =

node(R, node(B, A, x, B), y, node(B, C, z, D))

pattern(T ) = node(B, node(R, node(R, A, x, B), y, C), z, D)

pattern(T ) = node(B, node(R, A, x, node(R, B, y, C), z, D))
match(T ) =
pattern(T ) = node(B, A, x, node(R, B, y, node(R, C, z, D)))
pattern(T ) = node(B, A, x, node(R, node(R, B, y, C), z, D))
With function balance() defined, we can modify the previous binary search
tree insertion functions to make it work for red-black tree.
insert(T, k) = makeBlack(ins(T, k))

(3.4)

where

ins(T, k) =

node(R, , k, ) : T =
balance(node(ins(L, k), Key, R)) : k < Key

balance(node(L, Key, ins(R, k))) : otherwise

(3.5)

L, R, Key represent the left child, right child and the key of a tree.
L = lef t(T )
R = right(T )
Key = key(T )
Function makeBlack() is defined as the following, it forces the color of a
non-empty tree to be black.
makeBlack(T ) = node(B, L, Key, R)

(3.6)

Summarize the above functions and use language supported pattern matching features, we can come to the following Haskell program.
insert::(Ord a)RBTree a a RBTree a
insert t x = makeBlack $ ins t where
ins Empty = Node R Empty x Empty
ins (Node color l k r)
| x<k
= balance color (ins l) k r
| otherwise = balance color l k (ins r) --[3]
makeBlack(Node _ l k r) = Node B l k r

68CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

balance::Color RBTree a a RBTree a RBTree a

balance B (Node R (Node R a x b) y c) z d =
Node R (Node B a x b) y (Node B c z d)
balance B (Node R a x (Node R b y c)) z d =
Node R (Node B a x b) y (Node B c z d)
balance B a x (Node R b y (Node R c z d)) =
Node R (Node B a x b) y (Node B c z d)
balance B a x (Node R (Node R b y c) z d) =
Node R (Node B a x b) y (Node B c z d)
balance color l k r = Node color l k r

Note that the balance function is changed a bit from the original definition.
Instead of passing the tree, we pass the color, the left child, the key and the
right child to it. This can save a pair of boxing and un-boxing operations.
This program doesnt handle the case of inserting a duplicated key. However,
it is possible to handle it either by overwriting, or skipping. Another option is
to augment the data with a linked list[2].
Figure 3.6 shows two red-black trees built from feeding list 11, 2, 14, 1, 7,
15, 5, 8, 4 and 1, 2, ..., 8.
7

Figure 3.6: insert results generated by the Haskell algorithm

This algorithm shows great simplicity by summarizing the uniform feature
from the four different unbalanced cases. It is expressive over the traditional tree
rotation approach, that even in programming languages which dont support
pattern matching, the algorithm can still be implemented by manually check
the pattern. A Scheme/Lisp program is available along with this book can be
referenced as an example.
The insertion algorithm takes O(lg N ) time to insert a key to a red-black
tree which has N nodes.

Exercise 3.3

Write a program in an imperative language, such as C, C++ or python

to realize the same algorithm in this section. Note that, because there is
no language supported pattern matching, you need to test the 4 different
cases manually.

3.4. DELETION

3.4

Deletion

Remind the deletion section in binary search tree. Deletion is imperative only
for red-black tree as well. In typically practice, it often builds the tree just one
time, and performs looking up frequently after that. Okasaki explained why
he didnt provide red-black tree deletion in his work in [3]. One reason is that
deletions are much messier than insertions.
The purpose of this section is just to show that red-black tree deletion is
possible in purely functional settings, although it actually rebuilds the tree
because trees are read only in terms of purely functional data structure. In real
world, its up to the user (or actually the programmer) to adopt the proper
solution. One option is to mark the node be deleted with a flag, and perform
a tree rebuilding when the number of deleted nodes exceeds 50% of the total
number of nodes.
Not only in functional settings, even in imperative settings, deletion is more
complex than insertion. We face more cases to fix. Deletion may also violate the
red black tree properties, so we need fix it after the normal deletion as described
in binary search tree.
The deletion algorithm in this book are based on top of a handout in [5].
The problem only happens if you try to delete a black node, because it will
violate the 4-th property of red-black tree, which means the number of black
node in the path may decreased so that it is not uniformed black-height any
more.
When delete a black node, we can resume red-black property number 4 by
introducing a doubly-black concept[2]. It means that the although the node
is deleted, the blackness is kept by storing it in the parent node. If the parent
node is red, it turns to black, However, if it has been already black, it turns to
doubly-black.
In order to express the doubly-black node, The definition need some modification accordingly.
data Color = R | B | BB -- BB: doubly black for deletion
data RBTree a = Empty | BBEmpty -- doubly black empty
| Node Color (RBTree a) a (RBTree a)

When deleting a node, we first perform the same deleting algorithm in binary search tree mentioned in previous chapter. After that, if the node to be
sliced out is black, we need fix the tree to keep the red-black properties. Lets
denote the empty tree as , and for non-empty tree, it can be decomposed to
node(Color, L, Key, R) for its color, left sub-tree, key and the right sub-tree.
The delete function is defined as the following.

delete(T, k) = blackenRoot(del(T, k))

(3.7)

70CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

where

f
ixBlack
(node(C,
del(L,
k),
Key,
R))

f ixBlack

(node(C, L, Key, del(R, k)))

mkBlk(R) : C = B
del(T, k) =
=
R : otherwise

mkBlk(L)
: C=B

L
: otherwise

f ixBlack2 (node(C, L, k 0 , del(R, k 0 )))

: T =
: k < Key
: k > Key
: L=

(3.8)

: R=
: otherwise

The real deleting happens inside function del. For the trivial case, that the
tree is empty, the deletion result is ; If the key to be deleted is less than the
key of the current node, we recursively perform deletion on its left sub-tree; if
it is bigger than the key of the current node, then we recursively delete the key
from the right sub-tree; Because it may bring doubly-blackness, so we need fix
it.
If the key to be deleted is equal to the key of the current node, we need
splice it out. If one of its children is empty, we just replace the node by the
other one and reserve the blackness of this node. otherwise we cut and past the
minimum element k 0 = min(R) from the right sub-tree.
Function delete just forces the result tree of del to have a black root. This
is realized by function blackenRoot.

: T =
blackenRoot(T ) =
(3.9)
node(B, L, Key, R) : otherwise
Compare with the makeBlack function, which is defined in red-black tree
insertion section, they are almost same, except for the case of empty tree. This
is only valid in deletion, because insertion cant result an empty tree, while
deletion may.
Function mkBlk is defined to reserved the blackness of a node. If the node
to be sliced isnt black, this function wont be applied, otherwise, it turns a red
node to black and turns a black node to doubly-black. This function also marks
an empty tree to doubly-black empty.

: T =

node(B, L, Key, R) : C = R
mkBlk(T ) =
(3.10)
node(B 2 , L, Key, R) : C = B

T : otherwise
where means doubly-black empty node and B2 is the doubly-black color.
Summarizing the above functions yields the following Haskell program.
delete::(Ord a)RBTree a a RBTree a
delete t x = blackenRoot(del t x) where
del Empty _ = Empty
del (Node color l k r) x
| x < k = fixDB color (del l x) k r
| x > k = fixDB color l k (del r x)
-- x == k, delete this node
| isEmpty l = if color==B then makeBlack r else r

3.4. DELETION

| isEmpty r = if color==B then makeBlack l else l

| otherwise = fixDB color l k (del r k) where k= min r
blackenRoot (Node _ l k r) = Node B l k r
blackenRoot _ = Empty
makeBlack::RBTree
makeBlack (Node B
makeBlack (Node _
makeBlack Empty =
makeBlack t = t

a RBTree a
l k r) = Node BB l k r -- doubly black
l k r) = Node B l k r
BBEmpty

The final attack to the red-black tree deletion algorithm is to realize the
f ixBlack2 function. The purpose of this function is to eliminate the doublyblack colored node by rotation and color changing.
Lets solve the doubly-black empty node first. For any node, If one of its
child is doubly-black empty, and the other child is non-empty, we can safely
replace the doubly-black empty with a normal empty node.
Like figure 3.7, if we are going to delete the node 4 from the tree (Instead
show the whole tree, only part of the tree is shown), the program will use a
doubly-black empty node to replace node 4. In the figure, the doubly-black
node is shown in black circle with 2 edges. It means that for node 5, it has a
doubly-black empty left child and has a right non-empty child (a leaf node with
key 6). In such case we can safely change the doubly-black empty to normal
empty node. which wont violate any red-black properties.
3

(a) Delete 4 from the tree.

3
3
2

5
2

NIL

6
1

(b) After 4 is sliced off, it is doubly-black empty.

(c) We can safely change it to normal NIL.

Figure 3.7: One child is doubly-black empty node, the other child is non-empty
On the other hand, if a node has a doubly-black empty node and the other
child is empty, we have to push the doubly-blackness up one level. For example

72CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

in figure 3.8, if we want to delete node 1 from the tree, the program will use a
doubly-black empty node to replace 1. Then node 2 has a doubly-black empty
node and has an empty right node. In such case we must mark node 2 as
doubly-black after change its left child back to empty.

(a) Delete 1 from the tree.

3
3

5
2

NIL

6
4

(b) After 1 is sliced off, it is doubly-black empty. (c) We must push the doubly-blackness up
to node 2.

Figure 3.8: One child is doubly-black empty node, the other child is empty.

Based on above analysis, in order to fix the doubly-black empty node, we

define the function partially like the following.

node(B2 , , Key, )

node(C, , Key, R)
f ixBlack2 (T ) =
node(C, L, Key, )

...

:
:
:
:

(L = R = ) (L = R = )
L = R 6=
R = L 6=
...
(3.11)

After dealing with doubly-black empty node, we need to fix the case that
the sibling of the doubly-black node is black and it has one red child. In this
situation, we can fix the doubly-blackness with one rotation. Actually there are
4 different sub-cases, all of them can be transformed to one uniformed pattern.
They are shown in the figure 3.9. These cases are described in [2] as case 3 and
case 4.

3.4. DELETION

Figure 3.9: Fix the doubly black by rotation, the sibling of the doubly-black
node is black, and it has one red child.

74CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

The handling of these 4 sub-cases can be defined on top of formula 3.11.

... : ...
node(C, node(B, mkBlk(A), x, B), y, node(B, C, z, D)) : p1.1
f ixBlack (T ) =
node(C, node(B, A, x, B), y, node(B, C, z, mkBlk(D))) : p1.2

... : ...
(3.12)
where p1.1 and p1.2 each represent 2 patterns as the following.
2

node(C, A, x, node(B, node(R, B, y, C), z, D)) Color(A) = B 2

p1.1 =

node(C, A, x, node(B, B, y, node(R, C, z, D))) Color(A) = B2

node(C, node(B, A, x, node(R, B, y, C)), z, D) Color(D) = B2

p1.2 =

node(C, node(B, node(R, A, x, B), y, C), z, D) Color(D) = B2

Besides the above cases, there is another one that not only the sibling of the
doubly-black node is black, but also its two children are black. We can change
the color of the sibling node to red; resume the doubly-black node to black and
propagate the doubly-blackness one level up to the parent node as shown in
figure 3.10. Note that there are two symmetric sub-cases. This case is described
in [2] as case 2.
We go on adding this fixing after formula 3.12.

...
mkBlk(node(C, mkBlk(A), x, node(R, B, y, C)))
f ixBlack (T ) =
mkBlk(node(C, node(R, A, x, B), y, mkBlk(C)))

...
2

:
:
:
:

...
p2.1
p2.2
...
(3.13)

where p2.1 and p2.2 are two patterns as below.

node(C, A, x, node(B, B, y, C))
p2.1 =
Color(A) = B2 Color(B) = Color(C) = B

node(C, node(B, A, x, B), y, C)
p2.2 =
Color(C) = B2 Color(A) = Color(B) = B
There is a final case left, that the sibling of the doubly-black node is red.
We can do a rotation to change this case to pattern p1.1 or p1.2. Figure 3.11
shows about it.
We can finish formula 3.13 with 3.14.

...
f ixBlack2 (B, f ixBlack2 (node(R, A, x, B), y, C)
f ixBlack (T ) =
f ixBlack2 (B, A, x, f ixBlack2 (node(R, B, y, C))

T
2

:
:
:
:

...
p3.1
p3.2
otherwise
(3.14)

3.4. DELETION

=
b

(a) Color of x can be either black or red. (b) If x was red, then it becomes black, otherwise, it becomes doubly-black.
y

=
a

Figure 3.10: propagate the blackness up.

Figure 3.11: The sibling of the doubly-black node is red.

76CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

where p3.1 and p3.2 are two patterns as the following.
p3.1 = {Color(T ) = B Color(L) = B 2 Color(R) = R}

p3.2 = {Color(T ) = B Color(L) = R Color(R) = B 2 }

This two cases are described in [2] as case 1.
Fixing the doubly-black node with all above different cases is a recursive
function. There are two termination conditions. One contains pattern p1.1 and
p1.2, The doubly-black node was eliminated. The other cases may continuously
propagate the doubly-blackness from bottom to top till the root. Finally the
algorithm marks the root node as black anyway. The doubly-blackness will be
removed.
Put formula 3.11, 3.12, 3.13, and 3.14 together, we can write the final Haskell
program.
fixDB::Color RBTree a a RBTree a RBTree a
fixDB color BBEmpty k Empty = Node BB Empty k Empty
fixDB color BBEmpty k r = Node color Empty k r
fixDB color Empty k BBEmpty = Node BB Empty k Empty
fixDB color l k BBEmpty = Node color l k Empty
-- the sibling is black, and it has one red child
fixDB color a@(Node BB _ _ _) x (Node B (Node R b y c) z d) =
Node color (Node B (makeBlack a) x b) y (Node B c z d)
fixDB color a@(Node BB _ _ _) x (Node B b y (Node R c z d)) =
Node color (Node B (makeBlack a) x b) y (Node B c z d)
fixDB color (Node B a x (Node R b y c)) z d@(Node BB _ _ _) =
Node color (Node B a x b) y (Node B c z (makeBlack d))
fixDB color (Node B (Node R a x b) y c) z d@(Node BB _ _ _) =
Node color (Node B a x b) y (Node B c z (makeBlack d))
-- the sibling and its 2 children are all black, propagate the blackness up
fixDB color a@(Node BB _ _ _) x (Node B b@(Node B _ _ _) y c@(Node B _ _ _))
= makeBlack (Node color (makeBlack a) x (Node R b y c))
fixDB color (Node B a@(Node B _ _ _) x b@(Node B _ _ _)) y c@(Node BB _ _ _)
= makeBlack (Node color (Node R a x b) y (makeBlack c))
-- the sibling is red
fixDB B a@(Node BB _ _ _) x (Node R b y c) = fixDB B (fixDB R a x b) y c
fixDB B (Node R a x b) y c@(Node BB _ _ _) = fixDB B a x (fixDB R b y c)
-- otherwise
fixDB color l k r = Node color l k r

The deletion algorithm takes O(lg N ) time to delete a key from a red-black
tree with N nodes.

Exercise 3.4

As we mentioned in this section, deletion can be implemented by just

marking the node as deleted without actually removing it. Once the number of marked nodes exceeds 50% of the total node number, a tree re-build
is performed. Try to implement this method in your favorite programming
language.

3.5. IMPERATIVE RED-BLACK TREE ALGORITHM ?

3.5

Imperative red-black tree algorithm ?

We almost finished all the content in this chapter. By induction the patterns, we
can implement the red-black tree in a simple way compare to the imperative tree
rotation solution. However, we should show the comparator for completeness.
For insertion, the basic idea is to use the similar algorithm as described in
binary search tree. And then fix the balance problem by rotation and return
the final result.
1: function Insert(T, k)
2:
root T
3:
x Create-Leaf(k)
4:
Color(x) RED
5:
parent N IL
6:
while T 6= N IL do
7:
parent T
8:
if k < Key(T ) then
9:
T Left(T )
10:
else
11:
T Right(T )
12:
Parent(x) parent
13:
if parent = N IL then
. tree T is empty
14:
return x
15:
else if k < Key(parent) then
16:
Left(parent) x
17:
else
18:
Right(parent) x
19:
return Insert-Fix(root, x)
The only difference from the binary search tree insertion algorithm is that
we set the color of the new node as red, and perform fixing before return. It is
easy to translate the pseudo code to real imperative programming language, for
instance Python 3 .
def rb_insert(t, key):
root = t
x = Node(key)
parent = None
while(t):
parent = t
if(key < t.key):
t = t.left
else:
t = t.right
if parent is None: #tree is empty
root = x
elif key < parent.key:
parent.set_left(x)
else:
parent.set_right(x)
return rb_insert_fix(root, x)
3 C,

and C++ source codes are available along with this book

78CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

There are 3 base cases for fixing, and if we take the left-right symmetric
into consideration. there are total 6 cases. Among them two cases can be
merged together, because they all have uncle node in red color, we can toggle
the parent color and uncle color to black and set grand parent color to red.
With this merging, the fixing algorithm can be realized as the following.
1: function Insert-Fix(T, x)
2:
while Parent(x) 6= N IL and Color(Parent(x)) = RED do
3:
if Color(Uncle(x)) = RED then
. Case 1, xs uncle is red
4:
Color(Parent(x)) BLACK
5:
Color(Grand-Parent(x)) RED
6:
Color(Uncle(x)) BLACK
7:
x Grandparent(x)
8:
else
. xs uncle is black
9:
if Parent(x) = Left(Grand-Parent(x)) then
10:
if x = Right(Parent(x)) then . Case 2, x is a right child
11:
x Parent(x)
12:
T Left-Rotate(T, x)
. Case 3, x is a left child
13:
Color(Parent(x)) BLACK
14:
Color(Grand-Parent(x)) RED
15:
T Right-Rotate(T , Grand-Parent(x))
16:
else
17:
if x = Left(Parent(x)) then
. Case 2, Symmetric
18:
x Parent(x)
19:
T Right-Rotate(T, x)
. Case 3, Symmetric
20:
Color(Parent(x)) BLACK
21:
Color(Grand-Parent(x)) RED
22:
T Left-Rotate(T , Grand-Parent(x))
23:
Color(T ) BLACK
24:
return T
This program takes O(lg N ) time to insert a new key to the red-black tree.
Compare this pseudo code and the balance function we defined in previous
section, we can see the difference. They differ not only in terms of simplicity,
but also in logic. Even if we feed the same series of keys to the two algorithms,
they may build different red-black trees. There is a bit performance overhead
in the pattern matching algorithm. Okasaki discussed about the difference in
detail in his paper[2].
Translate the above algorithm to Python yields the below program.
# Fix the redred violation
def rb_insert_fix(t, x):
while(x.parent and x.parent.color==RED):
if x.uncle().color == RED:
#case 1: ((a:R x:R b) y:B c:R) = ((a:R x:B b) y:R c:B)
set_color([x.parent, x.grandparent(), x.uncle()],
[BLACK, RED, BLACK])
x = x.grandparent()
else:
if x.parent == x.grandparent().left:
if x == x.parent.right:

3.6. MORE WORDS

#case 2: ((a x:R b:R) y:B c) = case 3

x = x.parent
t=left_rotate(t, x)
# case 3: ((a:R x:R b) y:B c) = (a:R x:B (b y:R c))
set_color([x.parent, x.grandparent()], [BLACK, RED])
t=right_rotate(t, x.grandparent())
else:
if x == x.parent.left:
#case 2: (a x:B (b:R y:R c)) = case 3
x = x.parent
t = right_rotate(t, x)
# case 3: (a x:B (b y:R c:R)) = ((a x:R b) y:B c:R)
set_color([x.parent, x.grandparent()], [BLACK, RED])
t=left_rotate(t, x.grandparent())
t.color = BLACK
return t

Figure 3.12 shows the results of feeding same series of keys to the above
python insertion program. Compare them with figure 3.6, one can tell the
difference clearly.
11
5

14
2

(a)

(b)

Figure 3.12: Red-black trees created by imperative algorithm.

We skip the red-black tree delete algorithm in imperative settings, because
it is even more complex than the insertion. The implementation of deleting is
left as an exercise of this chapter.

Exercise 3.5
Implement the red-black tree deleting algorithm in your favorite imperative programming language. you can refer to [2] for algorithm details.

3.6

More words

Red-black tree is the most popular implementation of balanced binary search

tree. Another one is the AVL tree, which well introduce in next chapter. Redblack tree can be a good start point for more data structures. If we extend the

80CHAPTER 3. RED-BLACK TREE, NOT SO COMPLEX AS IT WAS THOUGHT

number of children from 2 to K, and keep the balance as well, it leads to Btree, If we store the data along with edge but not inside node, it leads to Tries.
However, the multiple cases handling and the long program tends to make new
comers think red-black tree is complex.
Okasakis work helps making the red-black tree much easily understand.
There are many implementation in other programming languages in that manner
[7]. Its also inspired me to find the pattern matching solution for Splay tree
and AVL tree etc.

Bibliography
[1] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. ISBN:0262032937.
The MIT Press. 2001
[2] Chris Okasaki. FUNCTIONAL PEARLS Red-Black Trees in a Functional
Setting. J. Functional Programming. 1998
[3] Chris Okasaki. Ten Years of Purely Functional Data Structures.
http://okasaki.blogspot.com/2008/02/ten-years-of-purely-functionaldata.html
[4] Wikipedia. Red-black tree. http://en.wikipedia.org/wiki/Red-black tree
[5] Lyn Turbak. Red-Black Trees. cs.wellesley.edu/ cs231/fall01/redblack.pdf Nov. 2, 2001.
[6] SGI STL. http://www.sgi.com/tech/stl/
[7] Pattern matching. http://rosettacode.org/wiki/Pattern matching

AVL tree

Chapter 4

AVL tree
4.1
4.1.1

Introduction
How to measure the balance of a tree?

Besides red-black tree, are there any other intuitive solutions of self-balancing
binary search tree? In order to measure how balancing a binary search tree,
one idea is to compare the height of the left sub-tree and right sub-tree. If
they differs a lot, the tree isnt well balanced. Lets denote the difference height
between two children as below
(T ) = |L| |R|

(4.1)

Where |T | means the height of tree T , and L, R denotes the left sub-tree
and right sub-tree.
If (T ) = 0, The tree is definitely balanced. For example, a complete binary
tree has N = 2h 1 nodes for height h. There is no empty branches unless the
leafs. Another trivial case is empty tree. () = 0. The less absolute value of
(T ) the more balancing the tree is.
We define (T ) as the balance factor of a binary search tree.

4.2

Definition of AVL tree

An AVL tree is a special binary search tree, that all sub-trees satisfying the
following criteria.
|(T )| 1

(4.2)

The absolute value of balance factor is less than or equal to 1, which means
there are only three valid values, -1, 0 and 1. Figure 4.1 shows an example AVL
tree.
Why AVL tree can keep the tree balanced? In other words, Can this definition ensure the height of the tree as O(lg N ) where N is the number of the
nodes in the tree? Lets prove this fact.
For an AVL tree of height h, The number of nodes varies. It can have at
most 2h 1 nodes for a complete binary tree. We are interesting about how
83

CHAPTER 4. AVL TREE

Figure 4.1: An example AVL tree

many nodes there are at least. Lets denote the minimum number of nodes for
height h AVL tree as N (h). Its obvious for the trivial cases as below.
For empty tree, h = 0, N (0) = 0;
For a singleton root, h = 1, N (1) = 1;
Whats the situation for common case N (h)? Figure 4.2 shows an AVL tree
T of height h. It contains three part, the root node, and two sub trees A, B.
We have the following fact.
h = max(height(L), height(R)) + 1

(4.3)

We immediately know that, there must be one child has height h 1. Lets
say height(A) = h 1. According to the definition of AVL tree, we have.
|height(A) height(B)| 1. This leads to the fact that the height of other
tree B cant be lower than h 2, So the total number of the nodes of T is the
number of nodes in tree A, and B plus 1 (for the root node). We exclaim that.
N (h) = N (h 1) + N (h 2) + 1

(4.4)

This recursion reminds us the famous Fibonacci series. Actually we can

transform it to Fibonacci series by defining N 0 (h) = N (h) + 1. So equation 4.4
changes to.
N 0 (h) = N 0 (h 1) + N 0 (h 2)

(4.5)

Lemma 4.2.1. Let N (h) be the minimum number of nodes for an AVL tree
with height h. and N 0 (h) = N (h) + 1, then
N 0 (h) h
Where =

5+1
2

is the golden ratio.

Proof. For the trivial case, we have

(4.6)

4.2. DEFINITION OF AVL TREE

85
k

h-1

h-2

Figure 4.2: An AVL tree with height h, one of the sub-tree with height h 1,
the other is h 2
h = 0, N 0 (0) = 1 0 = 1
h = 1, N 0 (1) = 2 1 = 1.618...
For the induction case, suppose N 0 (h) h .
N 0 (h + 1)

= N 0 (h) + N 0 (h 1) {F ibonacci}
h + h1
= h1 ( + 1)
{ + 1 = 2 =
h+1
=

5+3
2 }

From Lemma 4.2.1, we immediately get

h log (N + 1) = log (2) lg(N + 1) 1.44 lg(N + 1)

(4.7)

It tells that the height of AVL tree is proportion to O(lg N ), which means
that AVL tree is balanced.
During the basic mutable tree operations such as insertion and deletion, if
the balance factor changes to any invalid value, some fixing has to be performed
to resume || within 1. Most implementations utilize tree rotations. In this
chapter, well show the pattern matching solution which is inspired by Okasakis
red-black tree solution[2]. Because of this modify-fixing approach, AVL tree is
also a kind of self-balancing binary search tree. For comparison purpose, well
also show the procedural algorithms.
Of course we can compute the value recursively, another option is to store
the balance factor inside each nodes, and update them when we modify the tree.
The latter one avoid computing the same value every time.
Based on this idea, we can add one data field to the original binary search
tree as the following C++ code example 1 .
template <class T>
struct node{
int delta;
T key;
node left;
1 Some

implementations store the height of a tree instead of as in [5]

CHAPTER 4. AVL TREE

node right;
node parent;
};

In purely functional setting, some implementation use different constructor

to store the information. for example in [1], there are 4 constructors, E, N,
P, Z defined. E for empty tree, N for tree with negative 1 balance factor, P for
tree with positive 1 balance factor and Z for zero case.
In this chapter, well explicitly store the balance factor inside the node.
data AVLTree a = Empty
| Br (AVLTree a) a (AVLTree a) Int

The immutable operations, including looking up, finding the maximum and
minimum elements are all same as the binary search tree. Well skip them and
focus on the mutable operations.

4.3

Insertion

Insert a new element to an AVL tree may violate the AVL tree property that the
absolute value exceeds 1. To resume it, one option is to do the tree rotation
according to the different insertion cases. Most implementation is based on this
approach
Another way is to use the similar pattern matching method mentioned by
Okasaki in his red-black tree implementation [2]. Inspired by this idea, it is
possible to provide a simple and intuitive solution.
When insert a new key to the AVL tree, the balance factor of the root may
changes in range [1, 1], and the height may increase at most by one, which
we need recursively use this information to update the value in upper level
nodes. We can define the result of the insertion algorithm as a pair of data
(T 0 , H). Where T 0 is the new tree and H is the increment of height. Lets
denote function f irst(pair) which can return the first element in a pair. We
can modify the binary search tree insertion algorithm as the following to handle
AVL tree.
insert(T, k) = f irst(ins(T, k))

(4.8)

where

ins(T, k) =

(node(, k, , 0), 1) :
(tree(ins(L, k), Key, (R, 0)), ) :

(tree((L, 0), Key, ins(R, k)), ) :

T =
k < Key
otherwise

(4.9)

L, R, Key, represent the left child, right child, the key and the balance
factor of a tree.
L = lef t(T )
R = right(T )
Key = key(T )
= (T )

4.3. INSERTION

When we insert a new key k to a AVL tree T , if the tree is empty, we just
need create a leaf node with k, set the balance factor as 0, and the height is
increased by one. This is the trivial case. Function node() is defined to build a
tree by taking a left sub-tree, a right sub-tree, a key and a balance factor.
If T isnt empty, we need compare the Key with k. If k is less than the key,
we recursively insert it to the left child, otherwise we insert it into the right
child.
As we defined above, the result of the recursive insertion is a pair like
(L0 , Hl ), we need do balancing adjustment as well as updating the increment
of height. Function tree() is defined to dealing with this task. It takes 4 parameters as (L0 , Hl ), Key, (R0 , Hr ), and . The result of this function is
defined as (T 0 , H), where T 0 is the new tree after adjustment, and H is the
new increment of height which is defined as
H = |T 0 | |T |

(4.10)

This can be further detailed deduced in 4 cases.

= |T 0 | |T |
= 1 + max(|R0 |, |L0 |) (1 + max(|R|, |L|))
0
= max(|R
|, |L0 |) max(|R|, |L|)

Hr : 0 0 0

+ Hr : 0 0 0
=
Hl : 0 0 0

Hl : otherwise

(4.11)

To prove this equation, note the fact that the height cant increase both in
left and right with only one insertion.
These 4 cases can be explained from the definition of balance factor definition
that it equal to the difference from the right sub tree and left sub tree.
If 0 and 0 0, it means that the height of right sub tree isnt less
than the height of left sub tree both before insertion and after insertion.
In this case, the increment in height of the tree is only contributed from
the right sub tree, which is Hr .
If 0, which means the height of left sub tree isnt less than the height
of right sub tree before, and it becomes 0 0, which means that the
height of right sub tree increases due to insertion, and the left side keeps
same (|L0 | = |L|). So the increment in height is
H

= max(|R0 |, |L0 |) max(|R|, |L|)

= |R0 | |L0 |
= |R| + Hr |L|
= + Hr

{ 0 0 0}
{|L| = |L0 |}

For the case 0 0 0, Similar as the second one, we can get.

= max(|R0 |, |L0 |) max(|R|, |L|)

= |L0 | |R|
= |L| + Hl |R|
= Hl

{ 0 0 0}

CHAPTER 4. AVL TREE

For the last case, the both and 0 is no bigger than zero, which means
the height left sub tree is always greater than or equal to the right sub
tree, so the increment in height is only contributed from the right sub
tree, which is Hl .

The next problem in front of us is how to determine the new balancing factor
value 0 before performing balancing adjustment. According to the definition
of AVL tree, the balancing factor is the height of right sub tree minus the height
of right sub tree. We have the following facts.
0

= |R0 | |L0 |
= |R| + Hr (|L| + Hl )
= |R| |L| + Hr Hl
= + Hr Hl

(4.12)

With all these changes in height and balancing factor get clear, its possible
to define the tree() function mentioned in (4.9).
tree((L0 , Hl ), Key, (R0 , Hr ), ) = balance(node(L0 , Key, R0 , 0 ), H)
(4.13)
Before we moving into details of balancing adjustment, lets translate the
above equations to real programs in Haskell.
First is the insert function.
insert::(Ord a)AVLTree a a AVLTree a
insert t x = fst $ ins t where
ins Empty = (Br Empty x Empty 0, 1)
ins (Br l k r d)
| x<k
= tree (ins l) k (r, 0) d
| x == k
= (Br l k r d, 0)
| otherwise = tree (l, 0) k (ins r) d

Here we also handle the case that inserting a duplicated key (which means
the key has already existed.) as just overwriting.
tree::(AVLTree a, Int) a (AVLTree a, Int) Int (AVLTree a, Int)
tree (l, dl) k (r, dr) d = balance (Br l k r d, delta) where
d = d + dr - dl
delta = deltaH d d dl dr

And the definition of height increment is as below.

deltaH :: Int Int
deltaH d d dl dr
| d 0 && d
| d 0 && d
| d 0 && d
| otherwise =

4.3.1

Int Int Int

0 = dr
0 = d+dr
0 = dl - d
dl

Balancing adjustment

As the pattern matching approach is adopted in doing re-balancing. We need

consider what kind of patterns violate the AVL tree property.

4.3. INSERTION

Figure 4.3 shows the 4 cases which need fix. For all these 4 cases the balancing factors are either -2, or +2 which exceed the range of [1, 1]. After
balancing adjustment, this factor turns to be 0, which means the height of left
sub tree is equal to the right sub tree.

(z) = 2

(x) = 2

(y) = 1
y

@
B

(x) = 1

@
I
@

(x) = 2

@
A

(z) = 1

@
R
@
x

(z) = 2

(y) = 1

0 (y) = 0

Figure 4.3: 4 cases for balancing a AVL tree after insertion

We call these four cases left-left lean, right-right lean, right-left lean, and leftright lean cases in clock-wise direction from top-left. We denote the balancing
factor before fixing as (x), (y), and (z), while after fixing, they changes to
0 (x), 0 (y), and 0 (z) respectively.
Well next prove that, after fixing, we have (y) = 0 for all four cases, and
well provide the result values of 0 (x) and 0 (z).

CHAPTER 4. AVL TREE

Left-left lean case

As the structure of sub tree x doesnt change due to fixing, we immediately get
0 (x) = (x).
Since (y) = 1 and (z) = 2, we have
(y) = |C| |x| = 1 |C| = |x| 1
(z) = |D| |y| = 2 |D| = |y| 2

(4.14)

= |D| |C|
{F rom(4.14)}
= |y| 2 (|x| 1)
= |y| |x| 1
{x is child of y |y| |x| = 1}
=0

(4.15)

After fixing.
0 (z)

For 0 (y), we have the following fact after fixing.

0 (y) = |z| |x|
= 1 + max(|C|, |D|) |x| {By (4.15), we have|C| = |D|}
= 1 + |C| |x|
{By (4.14)}
= 1 + |x| 1 |x|
=0

(4.16)

Summarize the above results, the left-left lean case adjust the balancing
factors as the following.
0 (x) = (x)
0 (y) = 0
0 (z) = 0

(4.17)

Right-right lean case

Since right-right case is symmetric to left-left case, we can easily achieve the
result balancing factors as
0 (x) = 0
0 (y) = 0
0 (z) = (z)

(4.18)

Right-left lean case

First lets consider 0 (x). After balance fixing, we have.
0 (x) = |B| |A|

(4.19)

Before fixing, if we calculate the height of z, we can get.

|z| = 1 + max(|y|, |D|)
= 1 + |y|
= 2 + max(|B|, |C|)

{(z) = 1 |y| > |D|}

(4.20)

4.3. INSERTION

While since (x) = 2, we can deduce that.

(x) = 2 |z| |A| = 2
2 + max(|B|, |C|) |A| = 2
max(|B|, |C|) |A| = 0

{By (4.20)}
(4.21)

If (y) = 1, which means |C| |B| = 1, it means

max(|B|, |C|) = |C| = |B| + 1

(4.22)

Take this into (4.21) yields

|B| + 1 |A| = 0 |B| |A| = 1 {By (4.19) }
0 (x) = 1

(4.23)

If (y) 6= 1, it means max(|B|, |C|) = |B|, taking this into (4.21), yields.
|B| |A| = 0 {By (4.19)}
0 (x) = 0

(4.24)

Summarize these 2 cases, we get relationship of 0 (x) and (y) as the following.
0 (x) =

1 : (y) = 1
0 : otherwise

(4.25)

For 0 (z) according to definition, it is equal to.

0 (z) = |D| |C|
{(z) = 1 = |D| |y|}
= |y| |C| 1
{|y| = 1 + max(|B|, |C|)}
= max(|B|, |C|) |C|

(4.26)

If (y) = 1, then we have |C| |B| = 1, so max(|B|, |C|) = |B| = |C| + 1.

Takes this into (4.26), we get 0 (z) = 1.
If (y) 6= 1, then max(|B|, |C|) = |C|, we get 0 (z) = 0.
Combined these two cases, the relationship between 0 (z) and (y) is as
below.
0 (z) =

1 :
0 :

(y) = 1
otherwise

(4.27)

Finally, for 0 (y), we deduce it like below.

0 (y) = |z| |x|
= max(|C|, |D|) max(|A|, |B|)

(4.28)

There are three cases.

If (y) = 0, it means |B| = |C|, and according to (4.25) and (4.27), we
have 0 (x) = 0 |A| = |B|, and 0 (z) = 0 |C| = |D|, these lead to
0 (y) = 0.

CHAPTER 4. AVL TREE

If (y) = 1, From (4.27), we have 0 (z) = 0 |C| = |D|.
0 (y) = max(|C|, |D|) max(|A|, |B|)
= |C| max(|A|, |B|)
= |C| (|B| + 1)
=0

{|C| = |D|}
{From (4.25): 0 (x) = 1 |B| |A| = 1}
{(y) = 1 |C| |B| = 1}

If (y) = 1, From (4.25), we have 0 (x) = 0 |A| = |B|.

0 (y) = max(|C|, |D|) max(|A|, |B|)
= max(|C|, |D|) |B|
= |C| + 1 |B|
=0

{|A| = |B|}
{From (4.27): |D| |C| = 1}
{(y) = 1 |C| |B| = 1}

Both three cases lead to the same result that 0 (y) = 0.

Collect all the above results, we get the new balancing factors after fixing as
the following.

1 : (y) = 1
0 (x) =
0 : otherwise
0 (y) =
0
(4.29)
1 : (y) = 1
0
(z) =
0 : otherwise
Left-right lean case
Left-right lean case is symmetric to the Right-left lean case. By using the similar
deduction, we can find the new balancing factors are identical to the result in
(4.29).

4.3.2

Pattern Matching

All the problems have been solved and its time to define the final pattern
matching fixing function.

(node(node(A, x, B, (x)), y, node(C, z, D, 0), 0), 0) :

(node(node(A, x, B, 0), y, node(C, z, D, (z)), 0), 0) :
balance(T, H) =
(node(node(A,
x, B, 0 (x)), y, node(C, z, D, 0 (z)), 0), 0) :

(T, H) :
(4.30)
Where Pll (T ) means the pattern of tree T is left-left lean respectively. 0 (x)
and delta0 (z) are defined in (4.29). The four patterns are tested as below.
Pll (T ) = node(node(node(A, x, B, (x)), y, C, 1), z, D, 2)
Prr (T ) = node(A, x, node(B, y, node(C, z, D, (z)), 1), 2)
Prl (T ) = node(node(A, x, node(B, y, C, (y)), 1), z, D, 2)
Plr (T ) = node(A, x, node(node(B, y, C, (y)), z, D, 1), 2)

(4.31)

Translating the above function definition to Haskell yields a simple and intuitive program.

Pll (T )
Prr (T )
Prl (T ) Plr (T )
otherwise

4.3. INSERTION

balance :: (AVLTree a, Int) (AVLTree a, Int)

balance (Br (Br (Br a x b dx) y c (-1)) z d (-2),
(Br (Br a x b dx) y (Br c z d 0) 0, 0)
balance (Br a x (Br b y (Br c z d dz)
1)
2,
(Br (Br a x b 0) y (Br c z d dz) 0, 0)
balance (Br (Br a x (Br b y c dy)
1) z d (-2),
(Br (Br a x b dx) y (Br c z d dz) 0, 0)
dx = if dy == 1 then -1 else 0
dz = if dy == -1 then 1 else 0
balance (Br a x (Br (Br b y c dy) z d (-1))
2,
(Br (Br a x b dx) y (Br c z d dz) 0, 0)
dx = if dy == 1 then -1 else 0
dz = if dy == -1 then 1 else 0
balance (t, d) = (t, d)

_) =
_) =
_) =
where

_) =
where

The insertion algorithm takes time proportion to the height of the tree, and
according to the result we proved above, its performance is O(lg N ) where N is
the number of elements stored in the AVL tree.
Verification
One can easily create a function to verify a tree is AVL tree. Actually we need
verify two things, first, its a binary search tree; second, it satisfies AVL tree
property.
We left the first verification problem as an exercise to the reader.
In order to test if a binary tree satisfies AVL tree property, we can test
the difference in height between its two children, and recursively test that both
children conform to AVL property until we arrive at an empty leaf.

avl?(T ) =

T rue
avl?(L) avl?(R) ||R| |L|| 1

: T =
: otherwise

(4.32)

And the height of a AVL tree can also be calculate from the definition.

|T | =

0 : T =
1 + max(|R|, |L|) : otherwise

(4.33)

The corresponding Haskell program is given as the following.

isAVL :: (AVLTree a) Bool
isAVL Empty = True
isAVL (Br l _ r d) = and [isAVL l, isAVL r, abs (height r - height l) 1]
height :: (AVLTree a) Int
height Empty = 0
height (Br l _ r _) = 1 + max (height l) (height r)

Exercise 4.1
Write a program to verify a binary tree is a binary search tree in your
favorite programming language. If you choose to use an imperative language,
please consider realize this program without recursion.

CHAPTER 4. AVL TREE

4.4

Deletion

As we mentioned before, deletion doesnt make significant sense in purely functional settings. As the tree is read only, its typically performs frequently looking
up after build.
Even if we implement deletion, its actually re-building the tree as we presented in chapter of red-black tree. We left the deletion of AVL tree as an
exercise to the reader.

Exercise 4.2

Take red-black tree deletion algorithm as an example, write the AVL tree
deletion program in purely functional approach in your favorite programming language.
Write the deletion algorithm in imperative approach in your favorite programming language.

4.5

Imperative AVL tree algorithm ?

We almost finished all the content in this chapter about AVL tree. However, it
necessary to show the traditional insert-and-rotate approach as the comparator
to pattern matching algorithm.
Similar as the imperative red-black tree algorithm, the strategy is first to do
the insertion as same as for binary search tree, then fix the balance problem by
rotation and return the final result.
1: function Insert(T, k)
2:
root T
3:
x Create-Leaf(k)
4:
(x) 0
5:
parent N IL
6:
while T 6= N IL do
7:
parent T
8:
if k < Key(T ) then
9:
T Left(T )
10:
else
11:
T Right(T )
12:
Parent(x) parent
13:
if parent = N IL then
. tree T is empty
14:
return x
15:
else if k < Key(parent) then
16:
Left(parent) x
17:
else
18:
Right(parent) x
19:
return AVL-Insert-Fix(root, x)
Note that after insertion, the height of the tree may increase, so that the
balancing factor may also change, insert on right side will increase by 1,
while insert on left side will decrease it. By the end of this algorithm, we need
perform bottom-up fixing from node x towards root.

4.5. IMPERATIVE AVL TREE ALGORITHM ?

We can translate the pseudo code to real programming language, such as

Python 2 .
def avl_insert(t, key):
root = t
x = Node(key)
parent = None
while(t):
parent = t
if(key < t.key):
t = t.left
else:
t = t.right
if parent is None: #tree is empty
root = x
elif key < parent.key:
parent.set_left(x)
else:
parent.set_right(x)
return avl_insert_fix(root, x)

This is a top-down algorithm search the tree from root down to the proper
position and insert the new key as a leaf. By the end of this algorithm, it calls
fixing procedure, by passing the root and the new node inserted.
Note that we reuse the same methods of set left() and set right() as we
defined in chapter of red-black tree.
In order to resume the AVL tree balance property by fixing, we first determine if the new node is inserted on left hand or right hand. If it is on left, the
balancing factor decreases, otherwise it increases. If we denote the new value
as 0 , there are 3 cases of the relationship between and 0 .
If || = 1 and | 0 | = 0, this means adding the new node makes the tree
perfectly balanced, the height of the parent node doesnt change, the algorithm can be terminated.
If || = 0 and | 0 | = 1, it means that either the height of left sub tree or
right sub tree increases, we need go on check the upper level of the tree.
If || = 1 and | 0 | = 2, it means the AVL tree property is violated due to
the new insertion. We need perform rotation to fix it.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:

function AVL-Insert-Fix(T, x)
while Parent(x) 6= N IL do
(Parent(x))
if x = Left(Parent(x)) then
0 1
else
0 + 1
(Parent(x)) 0
P Parent(x)
L Left(x)
R Right(x)

and C++ source code are available along with this book

96
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:

CHAPTER 4. AVL TREE

if || = 1 and | 0 | = 0 then
. Height doesnt change, terminates.
return T
else if || = 0 and | 0 | = 1 then
. Go on bottom-up updating.
xP
else if || = 1 and | 0 | = 2 then
if 0 = 2 then
if (R) = 1 then
. Right-right case
(P ) 0
. By (4.18)
(R) 0
T Left-Rotate(T, P )
if (R) = 1 then
. Right-left case
y (Left(R))
. By (4.29)
if y = 1 then
(P ) 1
else
(P ) 0
(Left(R)) 0
if y = 1 then
(R) 1
else
(R) 0
T Right-Rotate(T, R)
T Left-Rotate(T, P )
if 0 = 2 then
if (L) = 1 then
. Left-left case
(P ) 0
(L) 0
Right-Rotate(T, P )
else
. Left-Right case
y (Right(L))
if y = 1 then
(L) 1
else
(L) 0
(Right(L)) 0
if y = 1 then
(P ) 1
else
(P ) 0
Left-Rotate(T, L)
Right-Rotate(T, P )
break
return T

Here we reuse the rotation algorithms mentioned in red-black tree chapter.

Rotation operation doesnt update balancing factor at all, However, since
rotation changes (actually improves) the balance situation we should update
these factors. Here we refer the results from above section. Among the four
cases, right-right case and left-left case only need one rotation, while right-left

4.5. IMPERATIVE AVL TREE ALGORITHM ?

case and left-right case need two rotations.
The relative python program is shown as the following.
def avl_insert_fix(t, x):
while x.parent is not None:
d2 = d1 = x.parent.delta
if x == x.parent.left:
d2 = d2 - 1
else:
d2 = d2 + 1
x.parent.delta = d2
(p, l, r) = (x.parent, x.parent.left, x.parent.right)
if abs(d1) == 1 and abs(d2) == 0:
return t
elif abs(d1) == 0 and abs(d2) == 1:
x = x.parent
elif abs(d1)==1 and abs(d2) == 2:
if d2 == 2:
if r.delta == 1: # Right-right case
p.delta = 0
r.delta = 0
t = left_rotate(t, p)
if r.delta == -1: # Right-Left case
dy = r.left.delta
if dy == 1:
p.delta = -1
else:
p.delta = 0
r.left.delta = 0
if dy == -1:
r.delta = 1
else:
r.delta = 0
t = right_rotate(t, r)
t = left_rotate(t, p)
if d2 == -2:
if l.delta == -1: # Left-left case
p.delta = 0
l.delta = 0
t = right_rotate(t, p)
if l.delta == 1: # Left-right case
dy = l.right.delta
if dy == 1:
l.delta = -1
else:
l.delta = 0
l.right.delta = 0
if dy == -1:
p.delta = 1
else:
p.delta = 0
t = left_rotate(t, l)
t = right_rotate(t, p)
break
return t

CHAPTER 4. AVL TREE

We skip the AVL tree deletion algorithm and left this as an exercise to the
reader.

4.6

Chapter note

AVL tree was invented in 1962 by Adelson-Velskii and Landis[3], [4]. The name
AVL tree comes from the two inventors name. Its earlier than red-black tree.
Its very common to compare AVL tree and red-black tree, both are selfbalancing binary search trees, and for all the major operations, they both consume O(lg N ) time. From the result of (4.7), AVL tree is more rigidly balanced
hence they are faster than red-black tree in looking up intensive applications
[3]. However, red-black trees could perform better in frequently insertion and
removal cases.
Many popular self-balancing binary search tree libraries are implemented on
top of red-black tree such as STL etc. However, AVL tree provides an intuitive
and effective solution to the balance problem as well.
After this chapter, well extend the tree data structure from storing data in
node to storing information on edges, which leads to Trie and Patrica, etc. If
we extend the number of children from two to more, we can get B-tree. These
data structures will be introduced next.

Bibliography
[1] Data.Tree.AVL http://hackage.haskell.org/packages/archive/AvlTree/4.2/doc/html/DataTree-AVL.html
[2] Chris Okasaki. FUNCTIONAL PEARLS Red-Black Trees in a Functional
Setting. J. Functional Programming. 1998
[3] Wikipedia. AVL tree. http://en.wikipedia.org/wiki/AVL tree
[4] Guy Cousinear, Michel Mauny. The Functional Approach to Programming. Cambridge University Press; English Ed edition (October 29, 1998).
ISBN-13: 978-0521576819
[5] Pavel Grafov. Implementation of an
http://github.com/pgrafov/python-avl-tree

AVL

tree

Python.

100

Trie and Patricia

Chapter 5

Trie and Patricia

5.1

Introduction

The binary trees introduced so far store information in nodes. Its possible to
store the information in edges. Trie and Patricia are important data structures
in information retrieving and manipulating. They were invented in 1960s. And
are widely used in compiler design[2], and bio-information area, such as DNA
pattern matching [3].

0
1

011

100

1011

Figure 5.1: Radix tree.

Figure 5.1 shows a radix tree[2]. It contains the strings of bit 1011, 10, 011,
100 and 0. When searching a key k = (b0 b1 ...bn )2 , we take the first bit b0 (MSB
101

102

CHAPTER 5. TRIE AND PATRICIA

from left), check if it is 0 or 1, if it is 0, we turn left; and turn right for 1. Then
we take the second bit and repeat this search till either meet a leaf or finish all
n bits.
The radix tree neednt store keys in node at all. The information is represented by edges. The nodes marked with keys in the above figure are only for
illustration purpose.
It is very natural to come to the idea is it possible to represent key in
integer instead of string? Because integer can be written in binary format,
such approach can save spaces. Another advantage is that the speed is fast
because we can use bit-wise manipulation in most programming environment.

5.2

Integer Trie

The data structure shown in figure 5.1 is called as binary trie. Trie is invented
by Edward Fredkin. It comes from retrieval, pronounce as /tri:/ by the
inventor, while it is pronounced /trai/ try by other authors [5]. Trie is also
called prefix tree or radix tree. A binary trie is a special binary tree in which
the placement of each key is controlled by its bits, each 0 means go left and
each 1 means go right[2].
Because integers can be represented in binary format, it is possible to store
integer keys rather than 0, 1 strings. When insert an integer as the new key to
the trie, we change it to binary form, then examine the first bit, if it is 0, we
recursively insert the rest bits to the left sub tree; otherwise if it is 1, we insert
into the right sub tree.
There is a problem when treat the key as integer. Consider a binary trie
shown in figure 5.2. If represented in 0, 1 strings, all the three keys are different.
But they are identical when turn into integers. Where should we insert decimal
3, for example, to the trie?
One approach is to treat all the prefix zero as effective bits. Suppose the
integer is represented with 32-bits, If we want to insert key 1, it ends up with
a tree of 32 levels. There are 31 nodes, each only has the left sub tree. the last
node only has the right sub tree. It is very inefficient in terms of space.
Okasaki shows a method to solve this problem in [2]. Instead of using bigendian integer, we can use the little-endian integer to represent key. Thus
decimal integer 1 is represented as binary 1. Insert it to the empty binary trie,
the result is a trie with a root and a right leaf. There is only 1 level. decimal 2
is represented as 01, and decimal 3 is (11)2 in little-endian binary format. There
is no need to add any prefix 0, the position in the trie is uniquely determined.

5.2.1

Definition of integer Trie

In order to define the little-endian binary trie, we can reuse the structure of
binary tree. A binary trie node is either empty, or a branch node. The branch
node contains a left child, a right node, and optional value as satellite data. The
left sub tree is encoded as 0 and the right sub tree is encoded as 1.
The following example Haskell code defines the trie algebraic data type.
data IntTrie a = Empty
| Branch (IntTrie a) (Maybe a) (IntTrie a)

5.2. INTEGER TRIE

103

011

0011

Figure 5.2: a big-endian trie

Below is another example definition in Python.
class IntTrie:
def __init__(self):
self.left = self.right = None
self.value = None

5.2.2

Insertion

Since the key is little-endian integer, when insert a key, we take the bit one by
one from the right most (LSB). If it is 0, we go to the left, otherwise go to the
right for 1. If the child is empty, we need create a new node, and repeat this to
the last bit (MSB) of the key.
1: function Insert(T, k, v)
2:
if T = NIL then
3:
T Empty-Node
4:
pT
5:
while k 6= 0 do
6:
if Even?(k) then
7:
if Left(p) = NIL then
8:
Left(p) Empty-Node
9:
p Left(p)
10:
else
11:
if Right(p) = NIL then
12:
Right(p) Empty-Node
13:
p Right(p)

104

CHAPTER 5. TRIE AND PATRICIA

k bk/2c
15:
Data(p) v
16:
return T
This algorithm takes 3 arguments, a Trie T , a key k, and the satellite date v.
The following Python example code implements the insertion algorithm. The
satellite data is optional, it is empty by default.
14:

def trie_insert(t, key, value = None):

if t is None:
t = IntTrie()
p=t
while key != 0:
if key & 1 == 0:
if p.left is None:
p.left = IntTrie()
p = p.left
else:
if p.right is None:
p.right = IntTrie()
p = p.right
key = key>>1
p.value = value
return t

Figure 5.2 shows a trie which is created by inserting pairs of key and value
{1 a, 4 b, 5 c, 9 d} to the empty trie.

1
1:a

4:b

1
5:c

1
9:d

Figure 5.3: An little-endian integer binary trie for the map {1 a, 4 b, 5

c, 9 d}.
Because the definition of the integer trie is recursive, its nature to define
the insertion algorithm recursively. If the LSB is 0, it means that the key to
be inserted is even, we recursively insert to the left child, we can divide the

5.2. INTEGER TRIE

105

key by 2 to get rid of the LSB. If the LSB is 1, the key is odd number, the
recursive insertion is applied to the right child. For trie T , denote the left and
right children as Tl and Tr respectively. Thus T = (Tl , d, Tr ), where d is the
optional satellite data. if T is empty, Tl , Tr and d are defined as empty as well.

insert(T, k, v) =

(Tl , v, Tr ) : k = 0
(insert(Tl , k/2, v), d, Tr ) : even(k)

(Tl , d, insert(Tr , bk/2c, v)) : otherwise

(5.1)

If the key to be inserted already exists, this algorithm just overwrites the
previous stored data. It can be replaced with other alternatives, such as storing
data as with linked-list etc.
The following Haskell example program implements the insertion algorithm.
insert t 0 x = Branch (left t) (Just x) (right t)
insert t k x | even k = Branch (insert (left t) (k div 2) x) (value t) (right t)
| otherwise = Branch (left t) (value t) (insert (right t) (k div 2) x)
left (Branch l _ _) = l
left Empty = Empty
right (Branch _ _ r) = r
right Empty = Empty
value (Branch _ v _) = v
value Empty = Nothing

For a given integer k with m bits in binary, the insertion algorithm reuses
m levels. The performance is bound to O(m) time.

5.2.3

Look up

To look up key k in the little-endian integer binary trie. We take each bit of
k from the left (LSB), then go left if this bit is 0, otherwise, we go right. The
looking up completes when all bits are consumed.
1: function Lookup(T, k)
2:
while x 6= 0 T 6=NIL do
3:
if Even?(x) then
4:
T Left(T )
5:
else
6:
T Right(T )
7:
k bk/2c
8:
if T 6= NIL then
9:
return Data(T )
10:
else
11:
return not found
Below Python example code uses bit-wise operation to implements the looking up algorithm.
def lookup(t, key):
while key != 0 and (t is not None):
if key & 1 == 0:

106

CHAPTER 5. TRIE AND PATRICIA

t = t.left
else:
t = t.right
key = key>>1
if t is not None:
return t.value
else:
return None

Looking up can also be define in recursive manner. If the tree is empty, the
looking up fails; If k = 0, the satellite data is the result to be found; If the last
bit is 0, we recursively look up the left child; otherwise, we look up the right
child.

: T =

d : k=0
lookup(T, k) =
(5.2)
lookup(T
,
k/2)
: even(k)

lookup(Tr , bk/2c) : otherwise

The following Haskell example program implements the recursive look up
algorithm.
search Empty k = Nothing
search t 0 = value t
search t k = if even k then search (left t) (k div 2)
else search (right t) (k div 2)

The looking up algorithm is bound to O(m) time, where m is the number of

bits for a given key.

5.3

Integer Patricia

Trie has some drawbacks. It wastes a lot of spaces. Note in figure 5.2, only
leafs store the real data. Typically, the integer binary trie contains many nodes
only have one child. One improvement idea is to compress the chained nodes
together. Patricia is such a data structure invented by Donald R. Morrison
in 1968. Patricia means practical algorithm to retrieve information coded in
alphanumeric[3]. It is another kind of prefix tree.
Okasaki gives implementation of integer Patricia in [2]. If we merge the
chained nodes which have only one child together in figure 5.3, We can get a
Patricia as shown in figure 5.4.
From this figure, we can find that the key for the sibling nodes is the longest
common prefix for them. They branches out at certain bit. Patricia saves a lot
of spaces compare to trie.
Different from integer trie, using the big-endian integer in Patricia doesnt
cause the padding zero problem mentioned in section 5.2. All zero bits before
MSB are omitted to save space. Okasaki lists some significant advantages of
big-endian Patricia[2].

5.3.1

Definition

Integer Patricia tree is a special kind of binary tree. It is either empty or is a

node. There are two different types of node.

5.3. INTEGER PATRICIA

107

001
4:b

1
1:a
0

01
9:d

1
5:c

Figure 5.4: Little endian Patricia for the map {1 a, 4 b, 5 c, 9 d}.

It can be a leaf contains integer key and optional satellite data;
or a branch node, contains the left and the right children. The two children
shares the longest common prefix bits for their keys. For the left child,
the next bit of the key is zero, while the next bit is one for the right child.
The following Haskell example code defines Patricia accordingly.
type Key = Int
type Prefix = Int
type Mask = Int
data IntTree a = Empty
| Leaf Key a
| Branch Prefix Mask (IntTree a) (IntTree a)

In order to tell from which bit the left and right children differ, a mask is
recorded in the branch node. Typically, a mask is power of 2, like 2n for some
non-negative integer n, all bits being lower than n dont belong to the common
prefix.
The following Python example code defines Patricia as well as some auxiliary
functions.
class IntTree:
def __init__(self, key = None, value = None):
self.key = key
self.value = value
self.prefix = self.mask = None
self.left = self.right = None
def set_children(self, l, r):
self.left = l
self.right = r
def replace_child(self, x, y):
if self.left == x:
self.left = y

108

CHAPTER 5. TRIE AND PATRICIA

else:
self.right = y
def is_leaf(self):
return self.left is None and self.right is None
def get_prefix(self):
if self.prefix is None:
return self.key
else:
return self.prefix

5.3.2

Insertion

When insert a key, if the tree is empty, we can just create a leaf node with the
given key and satellite data, as shown in figure 5.5).

NIL

Figure 5.5: Left: the empty tree; Right: After insert key 12.
If the tree is just a singleton leaf node x, we can create a new leaf y, put the
key and data into it. After that, we need create a new branch node, and set x
and y as the two children. In order to determine if the y should be the left or
right node, we need find the longest common prefix of x and y. For example
if key(x) is 12 ((1100)2 in binary), key(y) is 15 ((1111)2 in binary), then the
longest common prefix is (11oo)2 . Where o denotes the bits we dont care about.
We can use another integer to mask those bits. In this case, the mask number
is 4 (100 in binary). The next bit after the longest common prefix presents 21 .
This bit is 0 in key(x), while it is 1 in key(y). We should set x as the left child
and y as the right child. Figure 5.6 shows this example.
In case the tree is neither empty, nor a singleton leaf, we need firstly check if
the key to be inserted matches the longest common prefix recorded in the root.
Then recursively insert the key to the left or the right child according to the
next bit of the common prefix. For example, if insert key 14 ((1110)2 in binary)
to the result tree in figure 5.6, since the common prefix is (11oo)2 , and the next
bit (the bit of 21 ) is 1, we need recursively insert to the right child.
If the key to be inserted doesnt match the longest common prefix stored in
the root, we need branch a new leaf out. Figure 5.7 shows these two different
cases.
For a given key k and value v, denote (k, v) is the leaf node. For branch
node, denote it in form of (p, m, Tl , Tr ), where p is the longest common prefix,
m is the mask, Tl and Tr are the left and right children. Summarize the above

5.3. INTEGER PATRICIA

109

prefix=1100
mask=100

Figure 5.6: Left: A tree with singleton leaf 12; Right: After insert key 15.

prefix=1100
mask=100

prefix=1110
mask=10

(a) Insert key 14. It matches the longest common

prefix (1100)2 ; 14 is then recursively inserted to the
right sub tree.

prefix=1100
mask=100

prefix=0
mask=10000

prefix=1110
mask=10

(b) Insert key 5. It doesnt match the longest common prefix

(1100)2 , a new leaf is branched out.

Figure 5.7: Insert key to branch node.

110

CHAPTER 5. TRIE AND PATRICIA

cases, the insertion algorithm can be defined as the following.

= T = (k, v 0 )
= (k 0 , v 0 )
= (p, m, Tl , Tr ), match(k, p, m), zero(k, m)
insert(T, k, v) =

= (p, m, Tl , Tr ), match(k, p, m), zero(k, m)

= (p, m, Tl , Tr ), match(k, p, m)
(5.3)
The first clause deals with the edge cases, that either T is empty or it is a
leaf node with the same key. The algorithm overwrites the previous value for
the later case.
The second clause handles the case that T is a leaf node, but with different
key. Here we need branch out another new leaf. We need extract the longest
common prefix, and determine which leaf should be set as the left, and which
should be set as the right child. Function join(k1 , T1 , K2 , T2 ) does this work.
Well define it later.
The third clause deals with the case that T is a branch node, the longest
common prefix matches the key to be inserted, and the next bit to the common
prefix is zero. Here we need recursively insert to the left child.
The fourth clause handles the similar case as the third clause, except that
the next bit to the common prefix is one, but not zero. We need recursively
insert to the right child.
The last clause is for the case that the key to be inserted doesnt match the
longest common prefix stored in the branch. We need branch out a new leaf by
calling the join function.
We need define function match(k, p, m) to test if the key k, has the same
prefix p above the masked bits m. For example, suppose the prefix stored in
a branch node is (pn pn1 ...pi ...p0 )2 in binary, key k is (kn kn1 ...ki ...k0 )2 in
binary, and the mask is (100...0)2 = 2i . They match if and only if pj = kj for
all i j n.
One solution to realize match is to test if mask(k, m) = p is satisfied. Where
mask(x, m) = m 1&x, that we perform bitwise-not of m 1, then perform
bitwise-and with x.
Function zero(k, m) test the next bit of the common prefix is zero. With the
help of the mask m, we can shift m one bit to the right, then perform bitwise
and with the key.
(k, v)
join(k, (k, v), k 0 , T )
(p, m, insert(Tl , k, v), Tr )
(p, m, Tl , insert(Tr , k, v))
join(k, (k, v), p, T )

:
:
:
:
:

T
T
T
T
T

zero(k, m) = x&shif tr (m, 1)

(5.4)

If the mask m = (100..0)2 = 2i , k = (kn kn1 ...ki 1...k0 )2 , because the bit
next to ki is 1, zero(k, m) returns false value; if k = (kn kn1 ...ki 0...k0 )2 , then
the result is true.
Function join(p1 , T1 , p2 , T2 ) takes two different prefixes and trees. It extracts
the longest common prefix of p1 and p2 , create a new branch node, and set T1
and T2 as the two children.

join(p1 , T1 , p2 , T2 ) =

(p, m, T1 , T2 ) :
(p, m, T2 , T1 ) :

zero(p1, m), (p, m) = LCP (p1 , p2 )

zero(p1, m)
(5.5)

5.3. INTEGER PATRICIA

111

In order to calculate the longest common prefix of p1 and p2 , we can firstly

compute bitwise exclusive-or for them, then count the number of bits in this
result, and generate a mask m = 2|xor(p1 ,p2 )| . The longest common prefix p can
be given by masking the bits with m for either p1 or p2 .
p = mask(p1 , m)

(5.6)

The following Haskell example code implements the insertion algorithm.

import Data.Bits
insert t k x
= case t of
Empty Leaf k x
Leaf k x if k==k then Leaf k x
else join k (Leaf k x) k t -- t@(Leaf k x)
Branch p m l r
| match k p m if zero k m
then Branch p m (insert l k x) r
else Branch p m l (insert r k x)
| otherwise join k (Leaf k x) p t -- t@(Branch p m l r)
join p1 t1 p2 t2 = if zero p1 m then Branch p m t1 t2
else Branch p m t2 t1
where
(p, m) = lcp p1 p2
lcp :: Prefix Prefix (Prefix, Mask)
lcp p1 p2 = (p, m) where
m = bit (highestBit (p1 xor p2))
p = mask p1 m
highestBit x = if x == 0 then 0 else 1 + highestBit (shiftR x 1)
mask x m = (x &. complement (m-1)) -- complement means bit-wise not.
zero x m = x &. (shiftR m 1) == 0
match k p m = (mask k m) == p

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:

The insertion algorithm can also be realized imperatively.

function Insert(T, k, v)
if T = NIL then
T Create-Leaf(k, v)
return T
yT
p NIL
while y is not leaf, and Match(k, Prefix(y), Mask(y)) do
py
if Zero?(k, Mask(y)) then
y Left(y)
else
y Right(y)

112
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:

CHAPTER 5. TRIE AND PATRICIA

if y is leaf, and k = Key(y) then
Data(y) v
else
z Branch(y, Create-Leaf(k, v))
if p = NIL then
T z
else
if Left(p) = y then
Left(p) z
else
Right(p) z
return T

Function Branch(T1 , T2 ) does the similar job as what join is defined. It

creates a new branch node, extracts the longest common prefix, and sets T1 and
T2 as the two children.
1: function Branch(T1 , T2 )
2:
T Empty-Node
3:
( Prefix(T ), Mask(T ) ) LCP(Prefix(T1 ), Prefix(T2 ))
4:
if Zero?(Prefix(T1 ), Mask(T )) then
5:
Left(T ) T1
6:
Right(T ) T2
7:
else
8:
Left(T ) T2
9:
Right(T ) T1
10:
return T
The following Python example program implements the insertion algorithm.
def insert(t, key, value = None):
if t is None:
t = IntTree(key, value)
return t
node = t
parent = None
while(True):
if match(key, node):
parent = node
if zero(key, node.mask):
node = node.left
else:
node = node.right
else:
if node.is_leaf() and key == node.key:
node.value = value
else:
new_node = branch(node, IntTree(key, value))
if parent is None:
t = new_node
else:
parent.replace_child(node, new_node)
break

5.3. INTEGER PATRICIA

113

return t

The auxiliary functions, match, branch, lcp etc. are given as below.
def maskbit(x, mask):
return x & (~(mask-1))
def match(key, tree):
return (not tree.is_leaf()) and maskbit(key, tree.mask) == tree.prefix
def zero(x, mask):
return x & (mask>>1) == 0
def lcp(p1, p2):
diff = (p1 ^ p2)
mask=1
while(diff!=0):
diff>>=1
mask< 1
return (maskbit(p1, mask), mask)
def branch(t1, t2):
t = IntTree()
(t.prefix, t.mask) = lcp(t1.get_prefix(), t2.get_prefix())
if zero(t1.get_prefix(), t.mask):
t.set_children(t1, t2)
else:
t.set_children(t2, t1)
return t

Figure 5.8 shows the example Patricia created with the insertion algorithm.

prefix=0
mask=8

1:x

prefix=100
mask=2

0
4:y

1
5:z

Figure 5.8: Insert map 1 x, 4 y, 5 z into the big-endian integer Patricia

tree.

114

5.3.3

CHAPTER 5. TRIE AND PATRICIA

Look up

Consider the property of integer Patricia tree. When look up a key, if it has
common prefix with the root, then we check the next bit. If this bit is zero, we
recursively look up the left child; otherwise if the bit is one, we next look up
the right child.
When reach a leaf node, we can directly check if the key of the leaf is equal
to what we are looking up. This algorithm can be described with the following
pseudo code.
1: function Look-Up(T, k)
2:
if T = NIL then
3:
return N IL
. Not found
4:
while T is not leaf, and Match(k, Prefix(T ), Mask(T )) do
5:
if Zero?(k, Mask(T )) then
6:
T Left(T )
7:
else
8:
T Right(T )
9:
if T is leaf, and Key(T ) = k then
10:
return Data(T )
11:
else
12:
return N IL
. Not found
Below Python example program implements the looking up algorithm.
def lookup(t, key):
if t is None:
return None
while (not t.is_leaf()) and match(key, t):
if zero(key, t.mask):
t = t.left
else:
t = t.right
if t.is_leaf() and t.key == key:
return t.value
else:
return None

The looking up algorithm can also be realized in recursive approach. If the

Patricia tree T is empty, or its a singleton leaf with different key from what we
are looking up, the result is empty to indicate not found error; If the tree is a
singleton leaf, and the key of this leaf is equal to what we are looking up, we
are done. Otherwise, T is a branch node, we need check if the common prefix
matches the key to be looked up, and recursively look up the child according
to the next bit. If the common prefix doesnt match the key, it means the key
doesnt exist in the tree. We can return empty result to indicate not found
error.

lookup(T, k) =

v
lookup(Tl , k)
lookup(Tr , k)

:
:
:
:
:

T = (T = (k 0 , v), k 0 6= k)
T = (k 0 , v), k 0 = k
T = (p, m, Tl , Tr ), match(k, p, m), zero(k, m)
T = (p, m, Tl , Tr ), match(k, p, m), zero(k, m)
otherwise
(5.7)

5.4. ALPHABETIC TRIE

115

The following Haskell example program implements this recursive looking

up algorithm.
search t k
= case t of
Empty Nothing
Leaf k x if k==k then Just x else Nothing
Branch p m l r
| match k p m if zero k m then search l k
else search r k
| otherwise Nothing

5.4

Alphabetic Trie

Integer based Trie and Patricia Tree can be a good start point. Such technical
plays important role in Compiler implementation. Okasaki pointed that the
widely used Glasgow Haskell Compiler, GHC, utilized the similar implementation for several years before 1998 [2].
If we extend the key from integer to alphabetic value, Trie and Patricia tree
can be very powerful in solving textual manipulation problems.

5.4.1

Definition

Its not enough to just use the left and right children to represent alphabetic
keys. Using English for example, there are 26 letters and each can be lower or
upper case. If we dont care about the case, one solution is to limit the number
of branches (children) to 26. Some simplified ANSI C implementations of Trie
are defined by using the array of 26 letters. This can be illustrated as in Figure
5.9.
Not all branch nodes contain data. For instance, in Figure 5.9, the root
only has three non-empty branches representing letter a, b, and z. Other
branches such as for letter c, are all empty. We dont show empty branches in
the rest of this chapter.
If deal with case sensitive problems, or handle languages other than English,
there can be more letters than 26. The problem of dynamic size of sub branches
can be solved by using some collection data structures. Such as Hash table or
map.
A alphabetic trie is either empty or a node. There are two types of node.
A leaf node dont has any sub trees;
A branch node contains multiple sub trees. Each sub tree is bound to a
character.
Both leaf and branch may contain optional satellite data. The following
Haskell code shows the example definition.
data Trie a = Trie { value :: Maybe a
, children :: [(Char, Trie a)]}
empty = Trie Nothing []

116

CHAPTER 5. TRIE AND PATRICIA

nil

...

boy

zoo

l
bool

r
another

Figure 5.9: A Trie with 26 branches, containing key a, an, another, bool,
boy and zoo.

5.4. ALPHABETIC TRIE

117

Below ANSI C code defines the alphabetic trie. For illustration purpose
only, we limit the character set to lower case English letters, from a to z.
struct Trie {
struct Trie children[26];
void data;
};

5.4.2

Insertion

When insert string as key, start from the root, we pick the character one by
one from the string, examine which child represents the character. If the corresponding child is empty, a new empty node is created. After that, the next
character is used to select the proper grand child.
We repeat this process for all the characters, and finally store the optional
satellite data in the node we arrived at.
Below pseudo code describes the insertion algorithm.
1: function Insert(T, k, v)
2:
if T = NIL then
3:
T Empty-Node
4:
pT
5:
for each c in k do
6:
if Children(p)[c] = NIL then
7:
Children(p)[c] Empty-Node
8:
p Children(p)[c]
9:
Data(p) v
10:
return T
The following example ANSI C program implements the insertion algorithm.
struct Trie insert(struct Trie t, const char key, void value){
int c;
struct Trie p;
if(!t)
t = create_node();
for (p = t; key; ++key, p = pchildren[c]) {
c = key - a;
if (!pchildren[c])
pchildren[c] = create_node();
}
pdata = value;
return t;
}

Where function create node creates new empty node, with all children initialized to empty.
struct Trie create_node(){
struct Trie t = (struct Trie) malloc(sizeof(struct Trie));
int i;
for (i=0; i<26; ++i)
tchildren[i] = NULL;
tdata = NULL;
return t;

118

CHAPTER 5. TRIE AND PATRICIA

The insertion can also be realized in recursive way. Denote the key to be
inserted as K = k1 k2 ...kn , where ki is the i-th character. K 0 is the rest of
characters except k1 . v 0 is the satellite data to be inserted. The trie is in form
T = (v, C), where v is the satellite data, C = {(c1 , T1 ), (c2 , T2 ), ..., (cm , Tm )} is
the map of children. It maps from character ci to sub-tree Ti . if T is empty,
then C is also empty.

(v 0 , C) : K =
(5.8)
insert(T, K, v 0 ) =
(v, ins(C, k1 , K 0 , v 0 )) : otherwise.
If the key is empty, the previous value v is overwritten with v 0 . Otherwise,
we need check the children and perform recursive insertion. This is realized in
function ins(C, k1 , K 0 , v 0 ). It examines key-sub tree pairs in C one by one. Let
C 0 be the rest of pairs except for the first one. This function can be defined as
below.

{(k1 , insert(, K 0 , v 0 ))} : C =

{k1 , insert(T1 , K 0 , v 0 )} C 0 : k1 = c1
ins(C, k1 , K , v ) =

{(c1 , T1 )} ins(C 0 , k1 , K 0 , v 0 ) : otherwise

(5.9)

If C is empty, we create a pair, mapping from character k1 to a new empty

tree, and recursively insert the rest characters. Otherwise, the algorithm locates
the child which is mapped from k1 for further insertion.
The following Haskell example program implements the insertion algorithm.
insert t []
x=
insert t (k:ks) x =
ins [] k ks x =
ins (p:ps) k ks

5.4.3

Trie (Just x) (children t)

Trie (value t) (ins (children t) k ks x) where
[(k, (insert empty ks x))]
x = if fst p == k
then (k, insert (snd p) ks x):ps
else p:(ins ps k ks x)

Look up

To look up a key, we also extract the character from the key one by one. For each
character, we search among the children to see if there is a branch match this
character. If there is no such a child, the look up process terminates immediately
to indicate the not found error. When we reach the last character of the key,
The data stored in the current node is what we are looking up.
1: function Look-Up(T, key)
2:
if T = NIL then
3:
return not found
4:
for each c in key do
5:
if Children(T )[c] = NIL then
6:
return not found
7:
T Children(T )[c]
8:
return Data(T )
Below ANSI C program implements the look up algorithm. It returns NULL
to indicate not found error.

5.5. ALPHABETIC PATRICIA

119

void lookup(struct Trie t, const char key) {

while (key && t && tchildren[key - a])
t = tchildren[key++ - a];
return (key | | !t) ? NULL : tdata;
}

The look up algorithm can also be realized in recursive manner. When look
up a key, we start from the first character. If it is bound to some child, we
then recursively search the rest characters in that child. Denote the trie as
T = (v, C), the key being searched as K = k1 k2 ...kn if it isnt empty. The first
character in the key is k1 , and the rest characters are denoted as K 0 .

lookup(T, K) =

v : K=
: f ind(C, k1 ) =

lookup(T 0 , K 0 ) : f ind(C, k1 ) = T 0

(5.10)

Where function f ind(C, k) examine the pairs of key-child one by one to

check if any child is bound to character k. If the list of pairs C is empty,
the result is empty to indicate non-existence of such a child. Otherwise, let
C = {(k1 , T1 ), (k2 , T2 ), ..., (km , Tm )}, the first sub tree T1 is bound to k1 ; the
rest of pairs are represented as C 0 . Below equation defines the f ind function.

: C=

T1 : k1 = k
(5.11)
f ind(C, k) =

f ind(C 0 , k) : otherwise
The following Haskell example program implements the trie looking up algorithm. It uses the find function provided in standard library[5].
find t [] = value t
find t (k:ks) = case lookup k (children t) of
Nothing Nothing
Just t find t ks

Exercise 5.1
Develop imperative trie by using collection data structure to manage multiple sub trees in alphabetic trie.

5.5

Alphabetic Patricia

Similar to integer trie, alphabetic trie is not memory efficient. We can use the
same method to compress alphabetic trie to Patricia.

5.5.1

Definition

Alphabetic Patricia tree is a special prefix tree, each node contains multiple
branches. All children of a node share the longest common prefix string. As the
result, there is no node with only one child, because it conflicts with the longest
common prefix property.
If we turn the trie shown in figure 5.9 into Patricia by compressing all nodes
which have only one child. we can get a Patricia prefix tree as in figure 5.10.

120

CHAPTER 5. TRIE AND PATRICIA

zoo

n
an

zoo

ol
bool

y
boy

other
another

Figure 5.10: A Patricia prefix tree, with keys: a, an, another, bool, boy
and zoo.
We can modify the definition of alphabetic trie a bit to adapt it to Patricia.
The Patricia is either empty, or a node in form T = (v, C). Where v is the
optional satellite data; C = {(s1 , T1 ), (s2 , T2 ), ..., (sn , Tn )} is a list of pairs.
Each pair contains a string si , which is bound to a sub tree Ti .
The following Haskell example code defines Patricia accordingly.
type Key = String
data Patricia a = Patricia { value :: Maybe a
, children :: [(Key, Patricia a)]}
empty = Patricia Nothing []

Below Python code reuses the definition for trie to define Patricia.
class Patricia:
def __init__(self, value = None):
self.value = value
self.children = {}

5.5.2

Insertion

When insert a key, s, if the Patricia is empty, we create a leaf node as shown
in figure 5.11 (a). Otherwise, we need check the children. If there is some sub
tree Ti bound to the string si , and there exists common prefix between si and
s, we need branch out a new leaf Tj . The method is to create a new internal
branch node, bind it with the common prefix. Then set Ti as one child of this
branch, and Tj as the other child. Ti and Tj share the common prefix. This is
shown in figure 5.11 (b). However, there are two special cases, because s may
be the prefix of si as shown in figure 5.11 (c). And si may be the prefix of s as
in figure 5.11 (d).
The insertion algorithm can be described as below.
1: function Insert(T, k, v)
2:
if T = NIL then

5.5. ALPHABETIC PATRICIA

121

bo
NIL

boy

(a) Insert key boy into the empty Patri-(b) Insert key bool. A new branch
cia, the result is a leaf.
with common prefix bo is created.

another

p1 p2

...

other
x
p1 p2

...

(c) Insert key an with value y into x with prefix another.

another

insert

an p1

...

an p1

...

insert

other

(d) Insert another, into the node with prefix an. We recursively insert key other to the child.

Figure 5.11: Patricia insertion

122

CHAPTER 5. TRIE AND PATRICIA

3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:

}
22:
23:
24:
25:
26:
27:

T Empty-Node
pT
loop
match FALSE
for each (si , Ti ) Children(p) do
if k = si then
Value(p) v
return T
c LCP(k, si )
k1 k c
k2 si c
if c 6= NIL then
match TRUE
if k2 = NIL then
. si is prefix of k
p Ti
k k1
break
else
. Branch out a new leaf
Children(p) Children(p) { (c, Branch(k1 , v, k2 , Ti ))
Delete(Children(p), (si , Ti ))
return T
if match then
. Add a new leaf
Children(p) Children(p) { (k, Create-Leaf(v)) }
return T
return T

In the above algorithm, LCP function finds the longest common prefix of
two given strings, for example, string bool and boy has the longest common
prefix bo. The subtraction symbol - for strings gives the different part of two
strings. For example bool - bo = ol. Branch function creates a branch
node and updates keys accordingly.
The longest common prefix can be extracted by checking the characters in
the two strings one by one till there are two characters dont match.
1: function LCP(A, B)
2:
i1
3:
while i |A| i |B| A[i] = B[i] do
4:
ii+1
5:
return A[1...i 1]
There are two cases when branch out a new leaf. Branch(s1 , T1 , s2 , T2 )
takes two different keys and two trees. If s1 is empty, we are dealing the case
such as insert key an into a child bound to string another. We set T2 as the
child of T1 . Otherwise, we create a new branch node and set T1 and T2 as the
two children.
1: function Branch(s1 , T1 , s2 , T2 )
2:
if s1 = then
3:
Children(T1 ) Children(T1 ) {(s2 , T2 )}
4:
return T1
5:
T Empty-Node

5.5. ALPHABETIC PATRICIA

123

Children(T ) {(s1 , T1 ), (s2 , T2 )}

return T
The following example Python program implements the Patricia insertion
algorithm.
6:
7:

def insert(t, key, value = None):

if t is None:
t = Patricia()
node = t
while True:
match = False
for k, tr in node.children.items():
if key == k: # just overwrite
node.value = value
return t
(prefix, k1, k2) = lcp(key, k)
if prefix != "":
match = True
if k2 == "":
# example: insert "another" into "an", go on traversing
node = tr
key = k1
break
else: #branch out a new leaf
node.children[prefix] = branch(k1, Patricia(value), k2, tr)
del node.children[k]
return t
if not match: # add a new leaf
node.children[key] = Patricia(value)
return t
return t

Where the functions to find the longest common prefix, and branch out are
implemented as below.
# returns (p, s1, s2), where p is lcp, s1=s1-p, s2=s2-p
def lcp(s1, s2):
j=0
while j < len(s1) and j < len(s2) and s1[j] == s2[j]:
j += 1
return (s1[0:j], s1[j:], s2[j:])
def branch(key1, tree1, key2, tree2):
if key1 == "":
#example: insert "an" into "another"
tree1.children[key2] = tree2
return tree1
t = Patricia()
t.children[key1] = tree1
t.children[key2] = tree2
return t

The insertion can also be realized recursively. Start from the root, the program checks all the children to find if there is a node matches the key. Matching
means they have the common prefix. For duplicated keys, the program overwrites previous value. There are also alternative solution to handle duplicated

124

CHAPTER 5. TRIE AND PATRICIA

keys, such as using linked-list etc. If there is no child matches the key, the
program creates a new leaf, and add it to the children.
For Patricia T = (v, C), function insert(T, k, v 0 ) inserts key k, and value v 0
to the tree.
insert(T, k, v 0 ) = (v, ins(C, k, v 0 ))

(5.12)

This function calls another internal function ins(C, k, v 0 ). If the children C

is empty, a new leaf is created; Otherwise the children are examined one by
one. Denote C = {(k1 , T1 ), (k2 , T2 ), ..., (kn , Tn )}, C 0 holds all the prefix-sub tree
pairs except for the first one.

{(k, (v 0 , ))}
{(k, (v , CT1 ))} C 0
ins(C, k, v 0 ) =
{branch(k, v 0 , k1 , T1 )} C 0

{(k1 , T1 )} ins(C 0 , k, v 0 )
0

:
:
:
:

C=
k1 = k
match(k1 , k)
otherwise

(5.13)

The first clause deals with the edge case of empty children. A leaf node
containing v 0 which is bound to k will be returned as the only child. The second
clause overwrites the previous value with v 0 if there is some child bound to the
same key. Note the CT1 means the children of sub tree T1 . The third clause
branches out a new leaf if the first child matches the key k. The last clause goes
on checking the rest children.
We define two keys A and B matching if they have non-empty common
prefix.
match(A, B) = A 6= B 6= a1 = b1

(5.14)

Where a1 and b1 are the first characters in A and B if they are not empty.
Function branch(k1 , v, k2 , T2 ) takes tow keys, a value and a tree. Extract the
longest common prefix k = lcp(k1 , k2 ), Denote the different part as k10 = k1 k,
k20 = k2 k. The algorithm firstly handles the edge cases that either k1 is the
prefix of k2 or k2 is the prefix of k1 . For the former case, It creates a new node
containing v, bind this node to k, and set (k20 , T2 ) as the only child; For the later
case, It recursively insert k10 and v to T2 . Otherwise, the algorithm creates a
branch node, binds it to the longest common prefix k, and set two children for
it. One child is (k20 , T2 ), the other is a leaf node containing v, and being bound
to k10 .

(k, (v, {(k20 , T2 )})) : k = k1

(k, insert(T2 , k10 , v)) : k = k2
branch(k1 , v, k2 , T2 ) =

(k, (, {(k10 , (v, )), (k20 , T2 )}) : otherwise

(5.15)
And function lcp(A, B) keeps taking same characters from A and B one by
one. Denote a1 and b1 as the first characters in A and B if they are not empty.
A0 and B 0 are the rest parts except for the first characters.

lcp(A, B) =

: A = B = a1 6= b1
{a1 } lcp(A0 , B 0 ) : a1 = b1

(5.16)

5.5. ALPHABETIC PATRICIA

125

The following Haskell example program implements the Patricia insertion

algorithm.
insert t k x = Patricia (value t) (ins (children t) k x) where
ins []
k x = [(k, Patricia (Just x) [])]
ins (p:ps) k x
| (fst p) == k
= (k, Patricia (Just x) (children (snd p))):ps --overwrite
| match (fst p) k
= (branch k x (fst p) (snd p)):ps
| otherwise
= p:(ins ps k x)
match x y = x /= [] && y /= [] && head x == head y
branch k1 x k2 t2
| k1 == k
-- ex: insert "an" into "another"
= (k, Patricia (Just x) [(k2, t2)])
| k2 == k
-- ex: insert "another" into "an"
= (k, insert t2 k1 x)
| otherwise = (k, Patricia Nothing [(k1, leaf x), (k2, t2)])
where
k = lcp k1 k2
k1 = drop (length k) k1
k2 = drop (length k) k2
lcp [] _ = []
lcp _ [] = []
lcp (x:xs) (y:ys) = if x == y then x:(lcp xs ys) else []

5.5.3

Look up

When look up a key, we cant examine the characters one by one as in trie. Start
from the root, we need search among the children to see if any one is bound
to a prefix of the key. If there is such a child, we update the key by removing
the prefix part, and recursively look up the updated key in this child. If there
arent any children bound to any prefix of the key, the looking up fails.
1: function Look-Up(T, k)
2:
if T = NIL then
3:
return not found
4:
repeat
5:
match FALSE
6:
for (ki , Ti ) Children(T ) do
7:
if k = ki then
8:
return Data(Ti )
9:
if ki is prefix of k then
10:
match TRUE
11:
k k ki
12:
T Ti
13:
break

126
14:
15:

CHAPTER 5. TRIE AND PATRICIA

until match
return not found

Below Python example program implements the looking up algorithm. It

reuses the lcp(s1, s2) function defined previously to test if a string is the
prefix of the other.
def lookup(t, key):
if t is None:
return None
while True:
match = False
for k, tr in t.children.items():
if k == key:
return tr.value
(prefix, k1, k2) = lcp(key, k)
if prefix != "" and k2 == "":
match = True
key = k1
t = tr
break
if not match:
return None

This algorithm can also be realized recursively. For Patricia in form T =

(v, C), it calls another function to find among the children C.
lookup(T, k) = f ind(C, k)

(5.17)

If C is empty, the looking up fails; Otherwise, For C = {(k1 , T1 ), (k2 , T2 ), ..., (kn , Tn )},
we firstly examine if k is the prefix of k1 , if not the recursively check the rest
pairs denoted as C 0 .

vT1
f ind(C, k) =
lookup(T1 , k k1 )

f ind(C 0 , k)

:
:
:
:

C=
k = k1
k1 @ k
otherwise

(5.18)

Where A @ B means string A is prefix of B. f ind mutually calls lookup if

some child is bound to a string which is prefix of the key.
Below Haskell example program implements the looking up algorithm.
import qualified Data.List
find t k = find (children t) k where
find [] _ = Nothing
find (p:ps) k
| (fst p) == k = value (snd p)
| (fst p) Data.List.isPrefixOf k = find (snd p) (diff (fst p) k)
| otherwise = find ps k
diff k1 k2 = drop (length (lcp k1 k2)) k2

5.6. TRIE AND PATRICIA APPLICATIONS

5.6

127

Trie and Patricia applications

Trie and Patricia can be used to solving some interesting problems. Integer
based prefix tree is used in compiler implementation. Some daily used software
applications have many interesting features which can be realized with trie or
Patricia. In this section, some applications are given as examples, including,
e-dictionary, word auto-completion, T9 input method etc. The commercial implementations typically do not adopt trie or Patricia directly. The solutions we
demonstrated here are for illustration purpose only.

5.6.1

E-dictionary and word auto-completion

Figure 5.12 shows a screen shot of an English-Chinese E-dictionary. In order to

provide good user experience, the dictionary searches its word library, and lists
all candidate words and phrases similar to what user has entered.

Figure 5.12: E-dictionary. All candidates starting with what the user input are
listed.
A E-dictionary typically contains hundreds of thousands words. Its very expensive to performs a whole word search. Commercial software adopts complex
approaches, including caching, indexing etc to speed up this process.
Similar with e-dictionary, figure 5.13 shows a popular Internet search engine.
When user input something, it provides a candidate lists, with all items starting
with what the user has entered. And these candidates are shown in the order
of popularity. The more people search, the upper position it is in the list.
In both cases, the software provides a kind of word auto-completion mechanism. In some modern IDEs, the editor can even help users to auto-complete
program code.
Lets see how to implementation of the e-dictionary with trie or Patricia. To
simplify the problem, we assume the dictionary only supports English - English
information.

128

CHAPTER 5. TRIE AND PATRICIA

Figure 5.13: A search engine. All candidates starting with what user input are
listed.

A dictionary stores key-value pairs, the keys are English words or phrases,
the values are the meaning described in English sentences.
We can store all the words and their meanings in a trie, but it isnt space
effective especially when there are huge amount of items. Well use Patricia to
realize e-dictionary.
When user wants to look up word a, the dictionary does not only return
the meaning of a, but also provides a list of candidate words, which all start
with a, including abandon, about, accent, adam, ... Of course all these
words are stored in the Patricia.
If there are too many candidates, one solution is only displaying the top 10
words, and the user can browse for more.
The following algorithm reuses the looking up defined for Patricia. When it
finds a node bound to a string which is the prefix of what we are looking for, it
expands all its children until getting n candidates.
1: function Look-Up(T, k, n)
2:
if T = NIL then
3:
return
4:
pref ix NIL
5:
repeat
6:
match FALSE
7:
for (ki , Ti ) Children(T ) do
8:
if k is prefix of ki then
9:
return Expand(Ti , pref ix, n)
10:
if ki is prefix of k then
11:
match TRUE
12:
k k ki
13:
T Ti
14:
pref ix pref ix + ki

5.6. TRIE AND PATRICIA APPLICATIONS

15:
16:
17:

129

break
until match
return

Where function Expand(T, pref ix, n) picks n sub trees, which share the
same prefix in T . It is realized as BFS (Bread-First-Search) traverse. Chapter
search explains BFS in detail.
1: function Expand(T, pref ix, n)
2:
R
3:
Q {(pref ix, T )}
4:
while |R| < n |Q| > 0 do
5:
(k, T ) Pop(Q)
6:
if Data(T ) 6= NIL then
7:
R R {(k, Data(T ) )}
8:
for (ki , Ti ) Children(T ) do
9:
Push(Q, (k + ki , Ti ))
The following example Python program implements the e-dictionary application. When testing if a string is prefix of another one, it uses the find function
provided in standard string library.
import string
def patricia_lookup(t, key, n):
if t is None:
return None
prefix = ""
while True:
match = False
for k, tr in t.children.items():
if string.find(k, key) == 0: #is prefix of
return expand(prefix+k, tr, n)
if string.find(key, k) ==0:
match = True
key = key[len(k):]
t = tr
prefix += k
break
if not match:
return None
def expand(prefix, t, n):
res = []
q = [(prefix, t)]
while len(res)<n and len(q)>0:
(s, p) = q.pop(0)
if p.value is not None:
res.append((s, p.value))
for k, tr in p.children.items():
q.append((s+k, tr))
return res

This algorithm can also be implemented recursively, if the string we are looking for is empty, we expand all children until getting n candidates. Otherwise

130

CHAPTER 5. TRIE AND PATRICIA

we recursively examine the children to find one which has prefix equal to this
string.
In programming environments supporting lazy evaluation. An intuitive solution is to expand all candidates, and take the first n on demand. Denote the
Patricia prefix tree in form T = (v, C), below function enumerates all items
starts with key k.

f indAll(T, k) =

enum(C) :
{(, v)} enum(C) :

f ind(C, k) :

k = , v =
k = , v 6=
k 6=

(5.19)

The first two clauses deal with the edge cases the the key is empty. All the
children are enumerated except for those with empty values. The last clause
finds child matches k.
For non-empty children, C = {(k1 , T1 ), (k2 , T2 ), ..., (km , Tm )}, denote the
rest pairs except for the first one as C 0 . The enumeration algorithm can be
defined as below.

: C=
mapAppend(k1 , f indAll(T1 , )) enum(C 0 ) :
(5.20)
Where mapAppend(k, L) = {(k + ki , vi )|(ki , vi ) L}. It concatenate the
prefix k in front of every key-value pair in list L.
Function f ind(C, k) is defined as the following. For empty children, the
result is empty as well; Otherwise, it examines the first child T1 which is bound
to string k1 . If k1 is equal to k, it calls mapAppend to add prefix to the keys of
all the children under T1 ; If k1 is prefix of k, the algorithm recursively find all
children start with k k1 ; On the other hand, if k is prefix of k1 , all children
under T1 are valid result. Otherwise, the algorithm by-pass the first child and
goes on find the rest children.
enum(C) =

f ind(C, k) =

mapAppend(k, f indAll(T1 , ))
mapAppend(k1 , f indAll(T1 , k k1 ))
f indAll(T1 , )
f ind(C 0 , k)

:
:
:
:
:

C=
k1 = k
k1 @ k
k @ k1
otherwise

(5.21)

Below example Haskell program implements the e-dictionary application according to the above equations.
findAll :: Patricia a Key [(Key, a)]
findAll t [] =
case value t of
Nothing enum $ children t
Just x ("", x):(enum $ children t)
where
enum [] = []
enum (p:ps) = (mapAppend (fst p) (findAll (snd p) [])) ++ (enum ps)
findAll t k = find (children t) k where
find [] _ = []

5.6. TRIE AND PATRICIA APPLICATIONS

131

find (p:ps) k
| (fst p) == k
= mapAppend k (findAll (snd p) [])
| (fst p) Data.List.isPrefixOf k
= mapAppend (fst p) (findAll (snd p) (k diff (fst p)))
| k Data.List.isPrefixOf (fst p)
= findAll (snd p) []
| otherwise = find ps k
diff x y = drop (length y) x
mapAppend s lst = map (p(s++(fst p), snd p)) lst

In the lazy evaluation environment, the top n candidates can be gotten like
take(n, f indAll(T, k)). Appendix A has detailed definition of take function.

5.6.2

T9 input method

Most mobile phones around year 2000 are equipped with a key pad. Users have
quite different experience from PC when editing a short message or email. This
is because the mobile-phone key pad, or so called ITU-T key pad has much
fewer keys than PC. Figure 5.14 shows one example.

Figure 5.14: an ITU-T keypad for mobile phone.

There are typical two methods to input English word or phrases with ITU-T
key pad. For instance, if user wants to enter a word home, He can press the
key in below sequence.
Press key 4 twice to enter the letter h;
Press key 6 three times to enter the letter o;
Press key 6 twice to enter the letter m;
Press key 3 twice to enter the letter e;
Another much quicker way is to just press the following keys.
Press key 4, 6, 6, 3, word home appears on top of the candidate list;
Press key * to change a candidate word, so word good appears;

132

CHAPTER 5. TRIE AND PATRICIA

Press key * again to change another candidate word, next word gone
appears;
...

Compare these two methods, we can see the second one is much easier for
the end user. The only overhead is to store a dictionary of candidate words.
Method 2 is called as T9 input method, or predictive input method [6], [7].
The abbreviation T9 stands for textonym. It start with T with 9 characters.
T9 input can also be realized with trie or Patricia.
In order to provide candidate words to user, a dictionary must be prepared in
advance. Trie or Patricia can be used to store the dictionary. The commercial T9
implementations typically use complex indexing dictionary in both file system
and cache. The realization shown here is for illustration purpose only.
Firstly, we need define the T9 mapping, which maps from digit to candidate
characters.
MT 9 = { 2 abc, 3 def, 4 ghi,
5 jkl, 6 mno, 7 pqrs,
8 tuv, 9 wxyz}

(5.22)

With this mapping, MT 9 [i] returns the corresponding characters for digit i.
Suppose user input digits D = d1 d2 ...dn , If D isnt empty, denote the rest
digits except for d1 as D0 , below pseudo code shows how to realize T9 with trie.
1: function Look-Up-T9(T, D)
2:
Q {(, D, T )}
3:
R
4:
while Q 6= do
5:
(pref ix, D, T ) Pop(Q)
6:
for each c in MT 9 [d1 ] do
7:
if c Children(T ) then
8:
if D0 = then
9:
R R {pref ix + c}
10:
else
11:
Push(Q, (pref ix + c, D0 , Children(t)[c]))
12:
return R
Where pref ix + c means appending character c to the end of string pref ix.
Again, this algorithm performs BFS search with a queue Q. The queue is initialized with a tuple (pref ix, D, T ), containing empty prefix, the digit sequence
to be searched, and the trie. It keeps picking the tuple from the queue as far
as it isnt empty. Then get the candidate characters from the first digit to be
processed via the T9 map. For each character c, if there is a sub-tree bound to
it, we created a new tuple, update the prefix by appending c, using the rest of
digits to update D, and use that sub-tree. This new tuple is pushed back to the
queue for further searching. If all the digits are processed, it means a candidate
word is found. We put this word to the result list R.
The following example program in Python implements this T9 search with
trie.
T9MAP={2:"abc", 3:"def", 4:"ghi", 5:"jkl",
6:"mno", 7:"pqrs", 8:"tuv", 9:"wxyz"}

5.6. TRIE AND PATRICIA APPLICATIONS

133

def trie_lookup_t9(t, key):

if t is None or key == "":
return None
q = [("", key, t)]
res = []
while len(q)>0:
(prefix, k, t) = q.pop(0)
i=k[0]
if not i in T9MAP:
return None #invalid input
for c in T9MAP[i]:
if c in t.children:
if k[1:]=="":
res.append((prefix+c, t.children[c].value))
else:
q.append((prefix+c, k[1:], t.children[c]))
return res

Because trie is not space effective, we can modify the above algorithm with
Patricia solution. As far as the queue isnt empty, the algorithm pops the tuple.
This time, we examine all the prefix-sub tree pairs. For every pair (ki , Ti ), we
convert the alphabetic prefix ki back to digits sequence D0 by looking up the T9
map. If D0 exactly matches the digits of what user input, we find a candidate
word; otherwise if the digit sequence is prefix of what user inputs, the program
creates a new tuple, updates the prefix, the digits to be processed, and the
sub-tree. Then put the tuple back to the queue for further search.
1: function Look-Up-T9(T, D)
2:
Q {(, D, T )}
3:
R
4:
while Q 6= do
5:
(pref ix, D, T ) Pop(Q)
6:
for each (ki , Ti ) Children(T ) do
7:
D0 Convert-T9(ki )
8:
if D0 @ D then
. D0 is prefix of D
0
9:
if D = D then
10:
R R {pref ix + ki }
11:
else
12:
Push(Q, (pref ix + ki , D D0 , Ti ))
13:
return R
Function Convert-T9(K) converts each character in K back to digit.
1: function Convert-T9(K)
2:
D
3:
for each c K do
4:
for each (d S) MT 9 do
5:
if c S then
6:
D D {d}
7:
break
8:
return D
The following example Python program implements the T9 input method
with Patricia.

134

CHAPTER 5. TRIE AND PATRICIA

def patricia_lookup_t9(t, key):

if t is None or key == "":
return None
q = [("", key, t)]
res = []
while len(q)>0:
(prefix, key, t) = q.pop(0)
for k, tr in t.children.items():
digits = toT9(k)
if string.find(key, digits)==0: #is prefix of
if key == digits:
res.append((prefix+k, tr.value))
else:
q.append((prefix+k, key[len(k):], tr))
return res

T9 input method can also be realized recursively. Lets first define the trie
solution. The algorithm takes two arguments, a trie storing all the candidate
words, and a sequence of digits that is input by the user. If the sequence
is empty, the result is empty as well; Otherwise, it looks up C to find those
children which are bound to the first digit d1 according to T9 map.

f indT 9(T, D) =

{} : D =
f old(f, , lookupT 9(d1 , C)) : otherwise

(5.23)

Where folding is defined in Appendix A. Function f takes two arguments,

an intermediate list of candidates which is initialized empty, and a pair (c, T 0 ),
where c is candidate character, to which sub tree T 0 is bound. It append character c to all the candidate words, and concatenate this to the result list.
f (L, (c, T 0 )) = mapAppend(c, f indT 9(T 0 , D0 )) L

(5.24)

Note this mapAppend function is a bit different from the previous one defined
in e-dictionary application. The first argument is a character, but not a string.
Function lookupT 9(k, C) checks all the possible characters mapped to digit
k. If the character is bound to some child in C, it is record as one candidate.
lookupT 9(d, C) = f old(g, , MT 9 [k])

(5.25)

Where

g(L, k) =

L : f ind(C, k) =
{(k, T 0 )} L : f ind(C, k) = T 0

(5.26)

Below Haskell example program implements the T9 look up algorithm with

trie.
mapT9 = [(2, "abc"), (3, "def"), (4, "ghi"), (5, "jkl"),
(6, "mno"), (7, "pqrs"), (8, "tuv"), (9, "wxyz")]
findT9 t [] = [("", value t)]
findT9 t (k:ks) = foldl f [] (lookupT9 k (children t))
where
f lst (c, tr) = (mapAppend c (findT9 tr ks)) ++ lst

5.6. TRIE AND PATRICIA APPLICATIONS

135

lookupT9 c children = case lookup c mapT9 of

Nothing []
Just s foldl f [] s where
f lst x = case lookup x children of
Nothing lst
Just t (x, t):lst
mapAppend x lst = map (p(x:(fst p), snd p)) lst

There are few modifications when change the realization from trie to Patricia.
Firstly, the sub-tree is bound to prefix string, but not a single character.

f indT 9(T, D) =

{} : D =
f old(f, , f indP ref ixT 9(D, C)) : otherwise

(5.27)

The list for folding is given by calling function f indP ref ixT 9(D, C). And
f is also modified to reflect this change. It appends the candidate prefix D0 in
front of every result output by the recursive search, and then accumulates the
words.
f (L, (D0 , T 0 )) = mapAppend(D0 , f indT 9(T 0 , D D0 )) L

(5.28)

Function f indP ref ixT 9(D, C) examines all the children. For every pair
(ki , Ti ), if converting ki back to digits yields a prefix of D, then this pair is
selected as a candidate.
f indP ref ixT 9(D, C) = {(ki , Ti )|(ki , Ti ) C, convertT 9(ki ) @ D}

(5.29)

Function convertT 9(k) converts every alphabetic character in k back to digits according to T9 map.
convertT 9(K) = {d|c k, (d S) MT 9 c S}

(5.30)

The following example Haskell program implements the T9 input algorithm

with Patricia.
findT9 t [] = [("", value t)]
findT9 t k = foldl f [] (findPrefixT9 k (children t))
where
f lst (s, tr) = (mapAppend s (findT9 tr (k diff s))) ++ lst
diff x y = drop (length y) x
findPrefixT9 s lst = filter f lst where
f (k, _) = (toT9 k) Data.List.isPrefixOf s
toT9 = map (c head $ [ d | (d, s) mapT9, c elem s])

Exercise 5.2
For the T9 input, compare the results of the algorithms realized with trie
and Patricia, the sequences are different. Why does this happen? How to
modify the algorithm so that they output the candidates with the same
sequence?

136

5.7

CHAPTER 5. TRIE AND PATRICIA

Summary

In this chapter, we start from the integer base trie and Patricia. The map
data structure based on integer Patricia plays an important role in Compiler
implementation. Alphabetic trie and Patricia are natural extensions. They can
be used to manipulate text information. As examples, predictive e-dictionary
and T9 input method are realized with trie or Patricia. Although these examples
are different from the real implementation in commercial software. They show
simple approaches to solve some problems. Other important data structure,
such as suffix tree, has close relationship with them. Suffix tree is introduced in
the next chapter.

Bibliography
[1] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford Stein. Introduction to Algorithms, Second Edition. Problem 12-1.
ISBN:0262032937. The MIT Press. 2001
[2] Chris Okasaki and Andrew Gill. Fast Mergeable Integer Maps. Workshop
on ML, September 1998, pages 77-86, http://www.cse.ogi.edu/ andy/pub/finite.htm
[3] D.R. Morrison, PATRICIA Practical Algorithm To Retrieve Information
Coded In Alphanumeric, Journal of the ACM, 15(4), October 1968, pages
514-534.
[4] Suffix Tree, Wikipedia. http://en.wikipedia.org/wiki/Suffix tree
[5] Trie, Wikipedia. http://en.wikipedia.org/wiki/Trie
[6] T9 (predictive text), Wikipedia. http://en.wikipedia.org/wiki/T9 (predictive text)
[7] Predictive text, Wikipedia. http://en.wikipedia.org/wiki/Predictive text

137

138

Suffix Tree

Chapter 6

Suffix Tree
6.1

Introduction

Suffix Tree is an important data structure. It can be used to realize many

important string operations particularly fast[3]. It is also widely used in bioinformation area such as DNA pattern matching[4]. Weiner introduced suffix
tree in 1973[2]. The latest on-line construction algorithm was found in 1995[1].
The suffix tree for a string S is a special Patricia. Each edge is labeled with
some sub-string of S. Each suffix of S corresponds to exactly one path from the
root to a leaf. Figure 6.1 shows the suffix tree for English word banana.

anana banana nana

Figure 6.1: The suffix tree for banana

All suffixes, banana, anana, nana, ana, na, a, can be found in the
above tree. Among them the first 3 suffixes are explicitly shown; others are
implicitly represented. The reason for why ana, na, a, and are not shown
is because they are prefixes of the others. In order to show all suffixes explicitly,
we can append a special pad terminal symbol, which doesnt occur in other
places in the string. Such terminator is typically denoted as $. By this means,
there is no suffix being the prefix of the others.
Although the suffix tree for banana is simple, the suffix tree for bananas,
as shown in figure 6.2, is quite different.
To create suffix suffix tree for a given string, we can utilize the insertion
algorithm explained in previous chapter for Patricia.
1: function Suffix-Tree(S)
2:
T NIL
3:
for i 1 to |S| do
4:
T Patricia-Insert(T , Right(S, i))
5:
return T
139

140

CHAPTER 6. SUFFIX TREE

a bananas na

na s

nas s

Figure 6.2: The suffix tree for bananas

For non-empty string S = s1 s2 ...si ...sn of length n = |S|, function Right(S, i)
= si si+1...sn . It extracts the sub-string of S from the i-th character to the last
one. This straightforward algorithm can also be defined as below.
suf f ixT (S) = f old(insertP atricia , , suf f ixes(S))

(6.1)

Where function suf f ixes(S) gives all the suffixes for string S. If the string
is empty, the result is one empty string; otherwise, S itself is one suffix, the
others can be given by recursively call suf f ixes(S 0 ), where S 0 is given by drop
the first character from S.

{} : S =
suf f ixes(S) =
(6.2)
{S} suf f ixes(S 0 ) : otherwise
This solution constructs suffix tree in O(n2 ) time, for string of length n.
It totally inserts n suffixes to the tree, and each insertion takes linear time
proportion to the length of the suffix. The efficiency isnt good enough.
In this chapter, we firstly explain a fast on-line suffix trie construction solution by using suffix link concept. Because trie isnt space efficient, we next
introduce a linear time on-line suffix tree construction algorithm found by Ukkonen. and show how to solve some interesting string manipulation problems with
suffix tree.

6.2

Suffix trie

Just likes the relationship between trie and Patricia, Suffix trie has much simpler
structure than suffix tree. Figure 6.3 shows the suffix trie for banana.
Compare with figure 6.1, we can find the difference between suffix tree and
suffix trie. Instead of representing a word, every edge in suffix trie only represents a character. Thus suffix trie needs much more spaces. If we pack all nodes
which have only one child, the suffix trie is turned into a suffix tree.
We can reuse the trie definition for suffix tree. Each node is bound to a
character, and contains multiple sub trees as children. A child can be referred
from the bounded character.

6.2. SUFFIX TRIE

141

Figure 6.3: Suffix trie for banana

6.2.1

Node transfer and suffix link

For string S of length n, define Si = s1 s2 ...si . It is the prefix contains the first
i characters.
In suffix trie, each node represents a suffix string. for example in figure 6.4,
node X represents suffix a, by adding character c, node X transfers to Y
which represents suffix ac. We say node X transfer to Y with the edge of
character c[1].
Y Children(X)[c]
We also say that node X has a c-child Y . Below Python expression reflects
this concept.
y = x.children[c]

If node A in a suffix trie represents suffix si si+1 ...sn , and node B represents
suffix si+1 si+2 ...sn , we say node B represents the suffix of node A. We can
create a link from A to B. This link is defined as the suffix link of node A[1].
Suffix link is drawn in dotted style. In figure 6.4, the suffix link of node A points
to node B, and the suffix link of node B points to node C.
Suffix link is valid for all nodes except the root. We can add a suffix link
field to the trie definition. Below Python example code shows this update.
class STrie:
def __init__(self, suffix=None):
self.children = {}
self.suffix = suffix

142

CHAPTER 6. SUFFIX TREE

root

X
c

a
C

B
o

a
A
o

Figure 6.4: Suffix trie for string cacao. Node X a, node Y ac, X
transfers to Y with character c
suffix string
s1 s2 s3 ...si
s2 s3 ...si
...
si1 si
si

Table 6.1: suffixes for Si

6.2.2

On-line construction

For string S, Suppose we have constructed suffix trie for the i-th prefix Si =
s1 s2 ...si . Denote it as Suf f ixT rie(Si ). Lets consider how to obtain Suf f ixT rie(Si+1 )
from Suf f ixT rie(Si ).
If list all suffixes corresponding to Suf f ixT rie(Si ), from the longest (which
is Si ) to the shortest (which is empty), we can get table 6.1. There are total
i + 1 suffixes.
One solution is to append the character si+1 to every suffix in this table,
then add another empty string. This idea can be realized by adding a new child
for every node in the trie, and binding all these new child with edge of character
si+1 .
However, some nodes in Suf f ixT rie(Si ) may have the si+1 -child already.
For example, in figure 6.5, node X and Y are corresponding for suffix cac and
ac respectively. They dont have the a-child. But node Z, which represents
suffix c has the a-child already.

6.2. SUFFIX TRIE

143

Algorithm 1 Update Suf f ixT rie(Si ) to Suf f ixT rie(Si+1 ), initial version.
1: for T Suf f ixT rie(Si ) do
2:
Children(T )[si+1 ] Create-Empty-Node

root

a c

Z
c

Y
c

(a) Suffix trie for string cac.

(b) Suffix trie for string caca.

Figure 6.5: Suffix Trie of cac and caca

144

CHAPTER 6. SUFFIX TREE

When append si+1 to Suf f ixT rie(Si ). In this example si+1 is character a.
We need create new nodes for X and Y , but we neednt do this for Z.
If check the nodes one by one according to table 6.1, we can stop immediately
when meet a node which has the si+1 -child. This is because if node X in
Suf f ixT rie(Si ) has the si+1 -child, according to the definition of suffix link,
any suffix nodes X 0 of X in Suf f ixT rie(Si ) must also have the si+1 -child. In
other words, let c = si+1 , if wc is a sub-string of Si , then every suffix of wc is
also a sub-string of Si [1]. The only exception is the root, which represents for
empty string .
According to this fact, we can refine the algorithm 1 to the following.
Algorithm 2 Update Suf f ixT rie(Si ) to Suf f ixT rie(Si+1 ), second version.
1: for each T Suf f ixT rie(Si ) in descending order of suffix length do
2:
if Children(T )[si+1 ] = NIL then
3:
Children(T )[si+1 ] Create-Empty-Node
4:
else
5:
break
The next question is how to iterate all nodes in descending order of the suffix
length? Define the top of a suffix trie as the deepest leaf node. This definition
ensures the top represents the longest suffix. Along the suffix link from the top
to the next node, the length of the suffix decrease by one. This fact tells us
that We can traverse the suffix tree from the top to the root by using the suffix
links. And the order of such traversing is exactly what we want. Finally, there
is a special suffix trie for empty string Suf f ixT rie(N IL), We define the top
equals to the root in this case.
function Insert(top, c)
if top = NIL then
. The trie is empty
top Create-Empty-Node
T top
T 0 Create-Empty-Node
. dummy init value
while T 6= NIL Children(T )[c] = NIL do
Children(T )[c] Create-Empty-Node
Suffix-Link(T 0 ) Children(T )[c]
T 0 Children(T )[c]
T Suffix-Link(T )
if T 6= NIL then
Suffix-Link(T 0 ) Children(T )[c]
return Children(top)[c]
. returns the new top
Function Insert, updates Suf f ixT rie(Si ) to Suf f ixT rie(Si+1 ). It takes
two arguments, one is the top of Suf f ixT rie(Si ), the other is si+1 character.
If the top is NIL, it means the tree is empty, so there is no root. The algorithm
creates a root node in this case. A sentinel empty node T 0 is created. It keeps
tracking the previous created new node. In the main loop, the algorithm checks
every node one by one along the suffix link. If the node hasnt the si+1 -child, it
then creates a new node, and binds the edge to character si+1 . The algorithm
repeatedly goes up along the suffix link until either arrives at the root, or find a
node which has the si+1 -child already. After the loop, if the node isnt empty,

6.2. SUFFIX TRIE

145

it means we stop at a node which has the si+1 -child. The last suffix link then
points to that child. Finally, the new top position is returned, so that we can
further insert other characters to the suffix trie.
For a given string S, the suffix trie can be built by repeatedly calling Insert
function.
1: function Suffix-Trie(S)
2:
t NIL
3:
for i 1 to |S| do
4:
t Insert(t, si )
5:
return t
This algorithm returns the top of the suffix trie, but not the root. In order
to access the root, we can traverse along the suffix link.
1: function Root(T )
2:
while Suffix-Link(T ) 6= NIL do
3:
T Suffix-Link(T )
4:
return T
Figure 6.6 shows the steps when construct suffix trie for cacao. Only the
last layer of suffix links are shown.
For Insert algorithm, the computation time is proportion to the size of suffix
trie. In the worse case, the suffix trie is built in O(n2 ) time, where n = |S|. One
example is S = an bn , that there are n characters of a and n characters of b.
The following example Python program implements the suffix trie construction algorithm.
def suffix_trie(str):
t = None
for c in str:
t = insert(t, c)
return root(t)
def insert(top, c):
if top is None:
top=STrie()
node = top
new_node = STrie() #dummy init value
while (node is not None) and (c not in node.children):
new_node.suffix = node.children[c] = STrie(node)
new_node = node.children[c]
node = node.suffix
if node is not None:
new_node.suffix = node.children[c]
return top.children[c] #update top
def root(node):
while node.suffix is not None:
node = node.suffix
return node

146

CHAPTER 6. SUFFIX TREE

root

(a) Empty

(b) c

root

a
root

X
c

root
Y
a c

Z
c

a
A

(d) cac

(e) caca

(f) cacao

Figure 6.6: Construct suffix trie for cacao. There are 6 steps. Only the last
layer of suffix links are shown in dotted arrow.

6.3. SUFFIX TREE

6.3

147

Suffix Tree

Suffix trie isnt space efficient, and the construction time is quadratic. If dont
care about the speed, we can compress the suffix trie to suffix tree[6]. Ukkonen
found a linear time on-line suffix tree construction algorithm in 1995.

6.3.1

On-line construction

Active point and end point

The suffix trie construction algorithm shows very important fact about what
happens when Suf f ixT rie(Si ) updates to Suf f ixT rie(Si+1 ). Lets review the
last two steps in figure 6.6.
There are two different updates.
1. All leaves are appended with a new node for si+1 ;
2. Some non-leaf nodes are branched out with a new node for si+1 .
The first type of update is trivial, because for all new coming characters, we
need do this work anyway. Ukkonen defines leaf as the open node.
The second type of update is important. We need figure out which internal
nodes need branch out. We only focus on these nodes and apply the update.
Ukkonen defines the path along the suffix links from the top to the end as
boundary path. Denote the nodes in boundary path as, n1 , n2 , ..., nj , ..., nk .
These nodes start from the leaf node (the first one is the top position), suppose
that after the j-th node, they are not leaves any longer, we need repeatedly
branch out from this time point till the k-th node.
Ukkonen defines the first none-leaf node nj as active point and the last
node nk as end point. The end point can be the root.
Reference pair

a bananas na

X
na s

nas s

Y
nas s

Figure 6.7: Suffix tree of bananas. X transfer to Y with sub-string na.

Figure 6.7 shows the suffix tree of English word bananas. Node X represents suffix a. By adding sub-string na, node X transfers to node Y , which

148

CHAPTER 6. SUFFIX TREE

represents suffix ana. In other words, we can represent Y with a pair of node
and sub-string, like (X, w), where w = na. Ukkonen defines such kind of
pair as reference pair. Not only the explicit node, but also the implicit position
in suffix tree can be represented with reference pair. For example, (X, n )
represents to a position which is not an explicit node. By using reference pair,
we can represent every position in a suffix tree.
In order to save spaces, for string S, all sub-strings can be represented as
a pair of index (l, r), where l is the left index and r is the right index of the
character for the sub-string. For instance, if S = bananas, and the index
starts from 1, sub-string na can be represented with pair (3, 4). As the result,
there is only one copy of the complete string, and all positions in the suffix tree
is defined as (node, (l, r)). This is the final form of reference pair.
With reference pair, node transfer for suffix tree can be defined as the following.
Children(X)[sl ] ((l, r), Y ) Y (X, (l, r))
If character sl = c, we say that node X has a c-child, This child is Y . Y can
be transferred from X with sub string (l, r) Each node can have at most one
c-child.
canonical reference pair
Its obvious that the one position in a suffix tree may have multiple reference
pairs. For example, node Y in Figure 6.7 can be either denoted as (X, (3, 4)) or
(root, (2, 4)). If we define empty string = (i, i 1), Y can also be represented
as (Y, ).
The canonical reference pair is the one which has the closest node to the position. Specially, in case the position is an explicit node, the canonical reference
pair is (node, ), so (Y, ) is the canonical reference pair of node Y .
Below algorithm converts a reference pair (node, (l, r)) to the canonical reference pair (node0 , (l0 , r)). Note that since r doesnt change, the algorithm can
only return (node0 , l0 ) as the result.
Algorithm 3 Convert reference pair to canonical reference pair
1: function Canonize(node, (l, r))
2:
if node = NIL then
3:
if (l, r) = then
4:
return ( NIL, l)
5:
else
6:
return Canonize(root, (l + 1, r))
7:
while l r do
. (l, r)isn0 tempty
8:
((l0 , r0 ), node0 ) Children(node)[sl ]
9:
if r l r0 l0 then
10:
l l + r 0 l0 + 1
. Remove |(l0 , r0 )| chars from (l, r)
11:
node node0
12:
else
13:
break
14:
return (node, l)
If the passed in node parameter is NIL, it means a very special case. The

6.3. SUFFIX TREE

149

function is called like the following.

Canonize(Suffix-Link(root), (l, r))
Because the suffix link of root points to NIL, the result should be (root, (l +
1, r)) if (l, r) is not . Otherwise, (NIL, ) is returned to indicate a terminal
position.
We explain this special case in detail in later sections.
The algorithm
In 6.3.1, we mentioned, all updating to leaves is trivial, because we only need
append the new coming character to the leaf. With reference pair, it means,
when update Suf f ixT ree(Si ) to Suf f ixT ree(Si+1 ), all reference pairs in form
(node, (l, i)), are leaves. They will change to (node, (l, i+1)) next time. Ukkonen
defines leaf in form (node, (l, )), here means open to grow. We can skip
all leaves until the suffix tree is completely constructed. After that, we can
change all to the length of the string.
So the main algorithm only cares about positions from the active point to
the end point. However, how to find the active point and the end point?
When start suffix tree construction, there is only a root node. There arent
any branches or leaves. The active point should be (root, ), or (root, (1, 0)) (the
string index starts from 1).
For the end point, it is a position where we can finish updating Suf f ixT ree(Si ).
According to the suffix trie algorithm, we know it should be a position which has
the si+1 -child already. Because a position in suffix trie may not be an explicit
node in suffix tree, if (node, (l, r)) is the end point, there are two cases.
1. (l, r) = . It means the node itself is the end point. This node has the
si+1 -child, which means Children(node)[si+1 ] 6= NIL;
2. Otherwise, l r, the end point is an implicit position. It must satisfy si+1 = sl0 +|(l,r)| , where Children(node)[sl ]= ((l0 , r0 ), node0 ), |(l, r)|
means the length of sub-string (l, r). It equals to r l + 1. This is illustrated in figure 6.8. We can also say that (node, (l, r)) has a si+1 -child
implicitly.

Figure 6.8: Implicit end point

150

CHAPTER 6. SUFFIX TREE

Ukkonen finds a very important fact that if (node, (l, i)) is the end point of
Suf f ixT ree(Si ), then (node, (l, i + 1)) is the active point of Suf f ixT ree(Si+1 ).
This is because if (node, (l, i)) is the end point of Suf f ixT ree(Si ), it must
have a si+1 -child (either explicitly or implicitly). If this end point represents
suffix sk sk+1 ...si , it is the longest suffix in Suf f ixT ree(Si ) which satisfies
sk sk+1 ...si si+1 is a sub-string of Si . Consider Si+1 , sk sk+1 ...si si+1 must occur at least twice in Si+1 , so position (node, (l, i + 1)) is the active point of
Suf f ixT ree(Si+1 ). Figure 6.9 shows about this truth.

Figure 6.9:
End
Suf f ixT ree(Si+1 ).

point

Suf f ixT ree(Si )

and

active

point

Summarize the above facts, the algorithm of Ukkonens on-line construction

can be given as the following.
1: function Update(node, (l, i))
2:
prev Create-Empty-Node
. Initialized as sentinel
3:
loop
. Traverse along the suffix links
4:
(f inish, node0 ) End-Point-Branch?(node, (l, i 1), si )
5:
if f inish then
6:
break
7:
Children(node0 )[si ] ((i, ), Create-Empty-Node)
8:
Suffix-Link(prev) node0
9:
prev node0
10:
(node, l) = Canonize(Suffix-Link(node), (l, i 1))
11:
Suffix-Link(prev) node
12:
return (node, l)
. The end point
This algorithm takes reference pair (node, (l, i)) as arguments, note that
position (node, (l, i 1) is the active point for Suf f ixT ree(Si1 ). Then we
start a loop, this loop goes along the suffix links until the current position
(node, (l, i 1)) is the end point. Otherwise, function End-Point-Branch?

6.3. SUFFIX TREE

151

returns a position, from where the new leaf branch out.

The End-Point-Branch? algorithm is realized as below.
function End-Point-Branch?(node, (l, r), c)
if (l, r) = then
if node = NIL then
return (TRUE, root)
else
return (Children(node)[c] = NIL, node)
else
((l0 , r0 ), node0 ) Children(node)[sl ]
pos l0 + |(l, r)|
if spos = c then
return (TRUE, node)
else
p Create-Empty-Node
Children(node)[sl0 ] ((l0 , pos 1), p)
Children(p)[spos ] ((pos, r0 ), node0 )
return (FALSE, p)
If the position is (root, ), it means we have arrived at the root. Its definitely
the end point, so that we can finish this round of updating. If the position is
in form of (node, ), it means the reference pair represents an explicit node, we
can examine if this node has already the c-child, where c = si . If not, we need
branch out a leaf.
Otherwise, the position (node, (l, r)) points to an implicit node. We need
find the exact position next to it to see if there is a c-child. If yes, we meet an
end point, the updating loop can be finished; else, we turn the position to an
explicit node, and return it for further branching.
We can finalize the Ukkonens algorithm as below.
1: function Suffix-Tree(S)
2:
root Create-Empty-Node
3:
node root, l 0
4:
for i 1 to |S| do
5:
(node, l) = Update(node, (l, i))
6:
(node, l) = Canonize(node, (l, i))
7:
return root
Figure 6.10 shows the steps when constructing the suffix tree for string cacao.
Note that we neednt set suffix link for leaf nodes, only branch nodes need
suffix links.
The following example Python program implements Ukkonens algorithm.
First is the node definition.
class Node:
def __init__(self, suffix=None):
self.children = {} # c:(word, Node), where word = (l, r)
self.suffix = suffix

Because there is only one copy of the complete string, all sub-strings are
represent in (lef t, right) pairs, and the leaf are open pairs as (lef t, ). The
suffix tree is defined like below.

152

CHAPTER 6. SUFFIX TREE

root

(a) Empty

(b) c

ca a

a c cac

(d) cac

aca caca

(e) caca

cao o

(f) cacao

Figure 6.10: Construct suffix tree for cacao. There are 6 steps. Only the last
layer of suffix links are shown in dotted arrow.

class STree:
def __init__(self, s):
self.str = s
self.infinity = len(s)+1000
self.root = Node()

The infinity is defined as the length of the string plus a big number. Some
auxiliary functions are defined.
def substr(str, str_ref):
(l, r)=str_ref
return str[l:r+1]
def length(str_ref):
(l, r)=str_ref
return r-l+1

The main entry for Ukkonens algorithm is implemented as the following.

def suffix_tree(str):
t = STree(str)
node = t.root # init active point is (root, Empty)
l=0
for i in range(len(str)):
(node, l) = update(t, node, (l, i))
(node, l) = canonize(t, node, (l, i))
return t
def update(t, node, str_ref):
(l, i) = str_ref

6.3. SUFFIX TREE

153

c = t.str[i] # current char

prev = Node() # dummy init
while True:
(finish, p) = branch(t, node, (l, i-1), c)
if finish:
break
p.children[c]=((i, t.infinity), Node())
prev.suffix = p
prev = p
(node, l) = canonize(t, node.suffix, (l, i-1))
prev.suffix = node
return (node, l)
def branch(t, node, str_ref, c):
(l, r) = str_ref
if length(str_ref) 0: # (node, empty)
if node is None: #_ | _
return (True, t.root)
else:
return ((c in node.children), node)
else:
((l1, r1), node1) = node.children[t.str[l]]
pos = l1+length(str_ref)
if t.str[pos]==c:
return (True, node)
else:
branch_node = Node()
node.children[t.str[l1]]=((l1, pos-1), branch_node)
branch_node.children[t.str[pos]] = ((pos, r1), node1)
return (False, branch_node)
def canonize(t, node, str_ref):
(l, r) = str_ref
if node is None:
if length(str_ref) 0:
return (None, l)
else:
return canonize(t, t.root, (l+1, r))
while l r: # str_ref is not empty
((l1, r1), child) = node.children[t.str[l]]
if r-l r1-l1:
l += r1-l1+1
node = child
else:
break
return (node, l)

Functional suffix tree construction

Giegerich and Kurtz found Ukkonens algorithm can be transformed to McCreights algorithm[7]. The three suffix tree construction algorithms found by
Weiner, McCreight, and Ukkonen are all bound to O(n) time. Giegerich and
Kurtz conjectured any sequential suffix tree construction method doesnt base

154

CHAPTER 6. SUFFIX TREE

on suffix links, active suffixes, etc., fails to meet the O(n)-criterion.

There is implementation in PLT/Scheme[10] based on Ukkonens algorithm,
However, it updates suffix links during the processing, which is not purely functional.
A lazy suffix tree construction method is discussed in [8]. And this method
is contributed to Haskell Hackage by Bryan OSullivan. [9]. The method depends on the lazy evaluation property. The tree wont be constructed until it is
traversed. However, it cant ensure the O(n) performance if the programming
environments or languages dont support lazy evaluation.
The following Haskell program defines the suffix tree. A suffix tree is either
a leaf, or a branch containing multiple sub trees. Each sub tree is bound to a
string.
data Tr = Lf | Br [(String, Tr)] deriving (Eq)
type EdgeFunc = [String](String, [String])

The edge function extracts a common prefix from a list of strings. The prefix
returned by edge function may not be the longest one, empty string is also
allowed. The exact behavior can be customized with different edge functions.
build(edge, X)
This defines a generic radix tree building function. It takes an edge function,
and a set of strings. X can be all suffixes of a string, so that we get suffix trie
or suffix tree. Well also explain later that X can be all prefixes, which lead to
normal prefix trie or Patricia.
Suppose all the strings are built from a character set . When build the
tree, if the string is empty, X only contains one empty string as well. The result
tree is an empty leaf; Otherwise, we examine every character in , group the
strings in X with their initial characters, the apply the edge function to these
groups.

branch({({c} p,
build(edge, X) ==

leaf
build(edge, X 0 ))|
c ,
G {group(X, c)},
(p, X 0 ) {edge(G)}})

: X = {}
: otherwise

(6.3)
The algorithm categorizes all suffixes by the first letter in several groups.
It removes the first letter for each element in every group. For example, the
suffixes {acac, cac, ac, c} are categorized to groups {(a, [cac, c]),
(c, [ac, ])}.
group(X, c) = {C 0 |{c1 } C 0 X, c1 = c}

(6.4)

Function group enumerates all suffixes in X, for each one, denote the first
character as c1 , the rest characters as C 0 . If c1 is same as the given character c,
then C 0 is collected.
Below example Haskell program implements the generic radix tree building
algorithm.

6.3. SUFFIX TREE

155

alpha = [a..z]++[A..Z]
lazyTree::EdgeFunc [String] Tr
lazyTree edge = build where
build [[]] = Lf
build ss = Br [(a:prefix, build ss) |
aalpha,
xs@(x:_) [[cs | c:csss, c==a]],
(prefix, ss)[edge xs]]

Different edge functions produce different radix trees. Since edge function
extracts common prefix from a set of strings. The simplest one constantly uses
the empty string as the common prefix. This edge function builds a trie.
edgeT rie(X) = (, X)

(6.5)

We can also realize an edge function, that extracts the longest common
prefix. Such edge function builds a Patricia. Denote the strings as X =
{x1 , x2 , ..., xn }, for the each string xi , let the initial character be ci , and the
rest characters in xi as Wi . If there is only one string in X, the longest common prefix is definitely this string; If there are two strings start with different
initial characters, the longest common prefix is empty; Otherwise,it means all
the strings share the same initial character. This character definitely belongs to
the longest common prefix. We can remove it from all strings, and recursively
call the edge function.

edgeT ree(X) =

(x1 , {}) : X = {x1 }

(, X) : |X| > 1, xi X, ci 6= c1

({c1 } p, Y ) : (p, Y ) = edgeT ree({Wi |xi X})

(6.6)

Here are some examples for edgeT ree function.

edgeT ree({an, another, and}) = (an, {, other, d})
edgeT ree({bool, foo, bar}) = (, {bool, fool, bar})
The following example Haskell program implements this edge function.
edgeTree::EdgeFunc
edgeTree [s] = (s, [[]])
edgeTree awss@((a:w):ss) | null [c | c:_ss, a/=c] = (a:prefix, ss)
| otherwise
= ("", awss)
where (prefix, ss) = edgeTree (w:[u | _:uss])
edgeTree ss = ("", ss)

For any given string, we can build suffix trie and suffix tree by feeding suffixes
to these two edge functions.
suf f ixT rie(S) = build(edgeT rie, suf f ixes(S))

(6.7)

suf f ixT ree(S) = build(edgeT ree, suf f ixes(S))

(6.8)

Because the build(edge, X) is generic, it can be used to build other radix

trees, such as the normal prefix trie and Patricia.
trie(S) = build(edgeT rie, pref ixes(S))

(6.9)

156

CHAPTER 6. SUFFIX TREE

tree(S) = build(edgeT ree, pref ixes(S))

6.4

(6.10)

Suffix tree applications

Suffix tree can help to solve many string and DNA pattern manipulation problems particularly fast.

6.4.1

String/Pattern searching

There a plenty of string searching algorithms, such as the famous KMP(KnuthMorris-Pratt algorithm is introduced in the chapter of search) algorithm. Suffix
tree can perform at the same level as KMP[11]. the string searching in bound
to O(m) time, where m is the length of the sub-string to be search. However,
O(n) time is required to build the suffix tree in advance, where n is the length
of the text[12].
Not only sub-string searching, but also pattern matching, including regular
expression matching can be solved with suffix tree. Ukkonen summarizes this
kind of problems as sub-string motifs: For a string S, Suf f ixT ree(S) gives
complete occurrence counts of all sub-string motifs of S in O(n) time, although
S may have O(n2 ) sub-strings.
Find the number of sub-string occurrence
Every internal node in Suf f ixT ree(S) is corresponding to a sub-string occurs
multiple times in S. If this sub-string occurs k times in S, then there are total
k sub-trees under this node[13].
1: function Lookup-Pattern(T, s)
2:
loop
3:
match FALSE
4:
for (si , Ti ) Values(Children(T )) do
5:
if s @ si then
6:
return Max(|Children(Ti )|, 1)
7:
else if si @ s then
8:
match TRUE
9:
T Ti
10:
s s si
11:
break
12:
if match then
13:
return 0
When look up a sub-string pattern s in text w, we build the suffix tree T
from the text. Start from the root, we iterate all children. For every string
reference pair si and sub-tree Ti , we check if the s is prefix of si . If yes, the
number of sub-trees in Ti is returned as the result. There is a special case that
Ti is a leaf without any children. We need return 1 but not zero. This is why
we use the maximum function. Otherwise, if si is prefix of s, then we remove
si part from s, and recursively look up in Ti .
The following Python program implements this algorithm.

6.4. SUFFIX TREE APPLICATIONS

157

def lookup_pattern(t, s):

node = t.root
while True:
match = False
for _, (str_ref, tr) in node.children.items():
edge = substr(t, str_ref)
if string.find(edge, s)==0: #s isPrefixOf edge
return max(len(tr.children), 1)
elif string.find(s, edge)==0: #edge isPrefixOf s
match = True
node = tr
s = s[len(edge):]
break
if not match:
return 0
return 0 # not found

This algorithm can also be realized in recursive way. For the non-leaf suffix
tree T , denote the children as C = {(s1 , T1 ), (s2 , T2 ), ...}. We search the sub
string among the children.
lookuppattern (T, s) = f ind(C, s)

(6.11)

If children C is empty, it means the sub string doesnt occurs at all. Otherwise, we examine the first pair (s1 , T1 ), if s is prefix of s1 , then the number
of sub-trees in T1 is the result. If s1 is prefix of s, we remove s1 from s, and
recursively look up it in T1 ; otherwise, we go on to examine the rest children
denoted as C 0 .

0
max(1, |C1 |)
f ind(C, s) =
lookuppattern (T1 , s s1 )

f ind(C 0 , s)

:
:
:
:

C=
s @ s1
s1 @ s
otherwise

(6.12)

The following Haskell example code implements this algorithm.

lookupPattern (Br lst) ptn = find lst where
find [] = 0
find ((s, t):xs)
| ptn isPrefixOf s = numberOfBranch t
| s isPrefixOf ptn = lookupPattern t (drop (length s) ptn)
| otherwise = find xs
numberOfBranch (Br ys) = length ys
numberOfBranch _ = 1
findPattern s ptn = lookupPattern (suffixTree $ s++"$") ptn

We always append special terminator to the string (the $ in above program), so that there wont be any suffix becomes the prefix of the other[3].
Suffix tree also supports searching pattern like a**n, we skip it here. Readers can refer to [13] and [14] for details.

158

6.4.2

CHAPTER 6. SUFFIX TREE

Find the longest repeated sub-string

After adding a special terminator character to string S, The longest repeated

sub-string can be found by searching the deepest branches in suffix tree.
Consider the example suffix tree shown in figure 6.11

mississippi$ p

$ ppi$ ssi
A
ppi$ ssippi$

i$ pi$

B
ppi$ ssippi$

C
ppi$ ssippi$

Figure 6.11: The suffix tree for mississippi$

There are three branch nodes, A, B, and C with depth 3. However, A
represents the longest repeated sub-string issi. B and C represent for si,
ssi, they are shorter than A.
This example tells us that the depth of the branch node should be measured by the number of characters traversed from the root. But not the number
of explicit branch nodes.
To find the longest repeated sub-string, we can perform BFS in the suffix
tree.
1: function Longest-Repeated-Substring(T )
2:
Q (NIL, Root(T ))
3:
R NIL
4:
while Q is not empty do
5:
(s, T ) Pop(Q)
6:
for each ((l, r), T 0 ) Children(T ) do
7:
if T 0 is not leaf then
8:
s0 Concatenate(s, (l, r))
9:
Push(Q, (s0 , T 0 ))
10:
R Update(R, s0 )
11:
return R
This algorithm initializes a queue with a pair of an empty string and the
root. Then it repeatedly examine the candidate in the queue.
For each node, the algorithm examines each children one by one. If it is a
branch node, the child is pushed back to the queue for further search. And the
sub-string represented by this child will be treated as a candidate of the longest
repeated sub-string.
Function Update(R, s0 ) updates the longest repeated sub-string candidates.
If multiple candidates have the same length, they are all kept in a result list.

6.4. SUFFIX TREE APPLICATIONS

159

function Update(L, s)
if L = NIL |l1 | < |s| then
return l {s}
4:
if |l1 | = |s| then
5:
return Append(L, s)
6:
return L
The above algorithm can be implemented in Python as the following example
program.
1:
2:
3:

def lrs(t):
queue = [("", t.root)]
res = []
while len(queue)>0:
(s, node) = queue.pop(0)
for _, (str_ref, tr) in node.children.items():
if len(tr.children)>0:
s1 = s+t.substr(str_ref)
queue.append((s1, tr))
res = update_max(res, s1)
return res
def update_max(lst, x):
if lst ==[] or len(lst[0]) < len(x):
return [x]
if len(lst[0]) == len(x):
return lst + [x]
return lst

Searching the deepest branch can also be realized recursively. If the tree is
just a leaf node, empty string is returned, else the algorithm tries to find the
longest repeated sub-string from the children.

: leaf (T )
longest({si LRS(Ti )|(si , Ti ) C, leaf (Ti )}) : otherwise
(6.13)
The following Haskell example program implements the longest repeated
sub-string algorithm.
LRS(T ) =

isLeaf Lf = True
isLeaf _ = False
lrs::TrString
lrs Lf = ""
lrs (Br lst) = find $ filter (not isLeaf snd) lst where
find [] = ""
find ((s, t):xs) = maximumBy (compare on length) [s++(lrs t), find xs]

6.4.3

Find the longest common sub-string

The longest common sub-string, can also be quickly found with suffix tree. The
solution is to build a generalized suffix tree. If the two strings are denoted as
txt1 and txt2 , a generalized suffix tree is Suf f ixT ree(txt1 $1 txt2 $2 ). Where

160

CHAPTER 6. SUFFIX TREE

$1 is a special terminator character for txt1 , and $2 6= $1 is another special

terminator character for txt2 .
The longest common sub-string is indicated by the deepest branch node, with
two forks corresponding to both ...$1 ... and ...$2 (no $1 ). The definition of
the deepest node is as same as the one for the longest repeated sub-string, it is
the number of characters traversed from root.
If a node has ...$1 ... under it, the node must represent a sub-string of
txt1 , as $1 is the terminator of txt1 . On the other hand, since it also has ...$2
(without $1 ), this node must represent a sub-string of txt2 too. Because its the
deepest one satisfied this criteria, so the node represents the longest common
sub-string.
Again, we can use BFS (bread first search) to find the longest common
sub-string.
1: function Longest-Common-Substring(T )
2:
Q (NIL, Root(T ))
3:
R NIL
4:
while Q is not empty do
5:
(s, T ) POP(Q)
6:
if Match-Fork(T ) then
7:
R Update(R, s)
8:
for each ((l, r), T 0 ) Children(T ) do
9:
if T 0 is not leaf then
10:
s0 Concatenate(s, (l, r))
11:
Push(Q, (s0 , T 0 ))
12:
return R
Most part is as same as the the longest repeated sub-sting searching algorithm. The function Match-Fork checks if the children satisfy the common
sub-string criteria.
1: function Match-Fork(T )
2:
if | Children(T ) | = 2 then
3:
{(s1 , T1 ), (s2 , T2 )} Children(T )
4:
return T1 is leaf T2 is leaf Xor($1 s1 , $1 s2 ))
5:
return FALSE
In this function, it checks if the two children are both leaf. One contains $2 ,
while the other doesnt. This is because if one child is a leaf, it always contains
$1 according to the definition of suffix tree.
The following Python program implement the longest common sub-string
program.
def lcs(t):
queue = [("", t.root)]
res = []
while len(queue)>0:
(s, node) = queue.pop(0)
if match_fork(t, node):
res = update_max(res, s)
for _, (str_ref, tr) in node.children.items():
if len(tr.children)>0:
s1 = s + t.substr(str_ref)
queue.append((s1, tr))

6.4. SUFFIX TREE APPLICATIONS

161

return res
def is_leaf(node):
return node.children=={}
def match_fork(t, node):
if len(node.children)==2:
[(_, (str_ref1, tr1)), (_, (str_ref2, tr2))]=node.children.items()
return is_leaf(tr1) and is_leaf(tr2) and
(t.substr(str_ref1).find(#)!=-1) !=
(t.substr(str_ref2).find(#)!=-1)
return False

The longest common sub-string finding algorithm can also be realized recursively. If the suffix tree T is a leaf, the result is empty; Otherwise, we examine
all children in T . For those satisfy the matching criteria, the sub-string are
collected as candidates; for those dont matching, we recursively search the
common sub-string among the children. The longest candidate is selected as
the final result.

: leaf (T )
{si |(si , Ti ) C, match(Ti )}
: otherwise

{si LCS(Ti )|(si , Ti ) C, match(Ti )})

(6.14)
The following Haskell example program implements the longest common
sub-string algorithm.
LCS(T ) =

longest(

lcs Lf = []
lcs (Br lst) = find $ filter (not isLeaf snd) lst where
find [] = []
find ((s, t):xs) = maxBy (compare on length)
(if match t
then s:(find xs)
else (map (s++) (lcs t)) ++ (find xs))
match (Br [(s1, Lf), (s2, Lf)]) = ("#" isInfixOf s1) /= ("#" isInfixOf s2)
match _ = False

6.4.4

Find the longest palindrome

A palindrome is a string, S, such that S = reverse(S) For example, level,

rotator, civic are all palindrome.
The longest palindrome in a string s1 s2 ...sn can be found in O(n) time with
suffix tree. The solution can be benefit from the longest common sub-string
algorithm.
For string S, if sub-string w is a palindrome, then it must be sub-string
of reverse(S) too. for instance, issi is a palindrome, it is a sub-string of
mississippi. When reverse to ippississim, issi is also a sub-string.
Based on this fact, we can find the longest palindrome by searching the
longest common sub-string for S and reverse(S).
palindromem (S) = LCS(suf f ixT ree(S reverse(S)))

(6.15)

162

CHAPTER 6. SUFFIX TREE

The following Haskell example program finds the longest palindrome.

longestPalindromes s = lcs $ suffixTree (s++"#"++(reverse s)++"$")

6.4.5

Others

Suffix tree can also be used for data compression, such as Burrows-Wheeler
transform, LZW compression (LZSS) etc. [3]

6.5

Notes and short summary

Suffix Tree was first introduced by Weiner in 1973 [?]. In 1976, McCreight
greatly simplified the construction algorithm. McCreight constructs the suffix
tree from right to left. In 1995, Ukkonen gave the first on-line construction
algorithms from left to right. All the three algorithms are linear time (O(n)).
And some research shows the relationship among these 3 algorithms. [7]

Bibliography
[1] Esko Ukkonen. On-line construction of suffix trees. Algorithmica
14
(3):
249260.
doi:10.1007/BF01206331.
http://www.cs.helsinki.fi/u/ukkonen/SuffixT1withFigs.pdf
[2] Weiner, P. Linear pattern matching algorithms, 14th Annual
IEEE Symposium on Switching and Automata Theory, pp. 1C11,
doi:10.1109/SWAT.1973.13
[3] Suffix Tree, Wikipedia. http://en.wikipedia.org/wiki/Suffix tree
[4] Esko Ukkonen. Suffix tree and suffix array techniques for pattern analysis
in strings. http://www.cs.helsinki.fi/u/ukkonen/Erice2005.ppt
[5] Trie, Wikipedia. http://en.wikipedia.org/wiki/Trie
[6] Suffix Tree (Java). http://en.literateprograms.org/Suffix tree (Java)
[7] Robert Giegerich and Stefan Kurtz. From Ukkonen to McCreight
and Weiner: A Unifying View of Linear-Time Suffix Tree Construction. Science of Computer Programming 25(2-3):187-218, 1995.
http://citeseer.ist.psu.edu/giegerich95comparison.html
[8] Robert Giegerich and Stefan Kurtz. A Comparison of Imperative and Purely Functional Suffix Tree Constructions. Algorithmica 19 (3):
331353. doi:10.1007/PL00009177. www.zbh.unihamburg.de/pubs/pdf/GieKur1997.pdf
[9] Bryan OSullivan. suffixtree: Efficient, lazy suffix tree implementation.
http://hackage.haskell.org/package/suffixtree
[10] Danny. http://hkn.eecs.berkeley.edu/ dyoo/plt/suffixtree/
[11] Zhang
Shaojie.
Lecture
of
Suffix
http://www.cs.ucf.edu/ shzhang/Combio09/lec3.pdf

Trees.

[12] Lloyd Allison. Suffix Trees. http://www.allisons.org/ll/AlgDS/Tree/Suffix/

[13] Esko Ukkonen. Suffix tree and suffix array techniques for pattern analysis
in strings. http://www.cs.helsinki.fi/u/ukkonen/Erice2005.ppt
[14] Esko Ukkonen Approximate string-matching over suffix trees. Proc. CPM
93. Lecture Notes in Computer Science 684, pp. 228-242, Springer 1993.
http://www.cs.helsinki.fi/u/ukkonen/cpm931.ps

163

164

B-Trees

Chapter 7

B-Trees
7.1

Introduction

B-Tree is important data structure. It is widely used in modern file systems.

Some are implemented based on B+ tree, which is extended from B-tree. B-tree
is also widely used in database systems.
Some textbooks introduce B-tree with the the problem of how to access a
large block of data on magnetic disks or secondary storage devices[2]. It is
also helpful to understand B-tree as a generalization of balanced binary search
tree[2].
Refer to the Figure 7.1, It is easy to find the difference and similarity of
B-tree regarding to binary search tree.
M

Figure 7.1: Example B-Tree

Remind the definition of binary search tree. A binary search tree is

either an empty node;
or a node contains 3 parts, a value, a left child and a right child. Both
children are also binary search trees.
The binary search tree satisfies the constraint that.
all the values on the left child are not greater than the value of of this
node;
the value of this node is not greater than any values on the right child.
165

166

CHAPTER 7. B-TREES

For non-empty binary tree (L, k, R), where L, R and k are the left, right children, and the key. Function Key(T ) accesses the key of tree T . The constraint
can be represented as the following.
x L, y R Key(x) k Key(y)

(7.1)

If we extend this definition to allow multiple keys and children, we get the
B-tree definition.
A B-tree
is either empty;
or contains n keys, and n + 1 children, each child is also a B-Tree, we
denote these keys and children as k1 , k2 , ..., kn and c1 , c2 , ..., cn , cn+1 .
Figure 7.2 illustrates a B-Tree node.
C[1]

K[1]

C[2]

K[2]

...

C[n]

K[n]

C[n+1]

Figure 7.2: A B-Tree node

The keys and children in a node satisfy the following order constraints.
Keys are stored in non-decreasing order. that k1 k2 ... kn ;
for each ki , all elements stored in child ci are not greater than ki , while
ki is not greater than any values stored in child ci+1 .
The constraints can be represented as in equation (7.2) as well.
xi ci , i = 0, 1, ..., n, x1 k1 x2 k2 ... xn kn xn+1

(7.2)

Finally, after adding some constraints to make the tree balanced, we get the
complete B-tree definition.
All leaves have the same depth;
We define integral number, t, as the minimum degree of B-tree;
each node can have at most 2t 1 keys;
each node can have at least t 1 keys, except the root;
Consider a B-tree holds n keys. The minimum degree t 2. The height is
h. All the nodes have at least t 1 keys except the root. The root contains at
least 1 key. There are at least 2 nodes at depth 1, at least 2t nodes at depth 2,
at least 2t2 nodes at depth 3, ..., finally, there are at least 2th1 nodes at depth
h. Times all nodes with t 1 except for root, the total number of keys satisfies
the following inequality.
n 1 + (t 1)(2 + 2t + 2t2 + ... + 2th1 )
h1
X
tk
= 1 + 2(t 1)
k=0

th 1
= 1 + 2(t 1)
t1
= 2th 1

(7.3)

7.2. INSERTION

167

Thus we have the inequality between the height and the number of keys.
n+1
(7.4)
2
This is the reason why B-tree is balanced. The simplest B-tree is so called
2-3-4 tree, where t = 2, that every node except root contains 2 or 3 or 4 keys.
red-black tree can be mapped to 2-3-4 tree essentially.
The following Python code shows example B-tree definition. It explicitly
pass t when create a node.
h logt

class BTree:
def __init__(self, t):
self.t = t
self.keys = []
self.children = []

B-tree nodes commonly have satellite data as well. We ignore satellite data
for illustration purpose.
In this chapter, we will firstly introduce how to generate B-tree by insertion.
Two different methods will be explained. One is the classic method as in [2],
that we split the node before insertion if its full; the other is the modify-fix
approach which is quite similar to the red-black tree solution [3] [2]. We will
next explain how to delete key from B-tree and how to look up a key.

7.2

Insertion

B-tree can be created by inserting keys repeatedly. The basic idea is similar to
the binary search tree. When insert key x, from the tree root, we examine all
the keys in the node to find a position where all the keys on the left are less
than x, while all the keys on the right are greater than x. If the current node
is a leaf node, and it is not full (there are less then 2t 1 keys in this node), x
will be insert at this position. Otherwise, the position points to a child node.
We need recursively insert x to it.
Figure 7.3 shows one example. The B-tree illustrated is 2-3-4 tree. When
insert key x = 22, because its greater than the root, the right child contains
key 26, 38, 45 is examined next; Since 22 < 26, the first child contains key 21
and 25 are examined. This is a leaf node, and it is not full, key 22 is inserted
to this node.
However, if there are 2t 1 keys in the leaf, the new key x cant be inserted,
because this node is full. When try to insert key 18 to the above example
B-tree will meet this problem. There are 2 methods to solve it.

7.2.1

Splitting

Split before insertion

If the node is full, one method to solve the problem is to split to node before
insertion.
For a node with at 1 keys, it can be divided into 3 parts as shown in Figure
7.4. the left part contains the first t 1 keys and t children. The right part
contains the rest t 1 keys and t children. Both left part and right part are

168

CHAPTER 7. B-TREES
20

(a) Insert key 22 to the 2-3-4 tree. 22 > 20, go to the right child; 22 < 26 go to the first child.

(b) 21 < 22 < 25, and the leaf isnt full.

Figure 7.3: Insertion is similar to binary search tree.

valid B-tree nodes. the middle part is the t-th key. We can push it up to the
parent node (if the current node is root, then the this key, with the two children
will be the new root).
For node x, denote K(x) as keys, C(x) as children. The i-th key as ki (x),
the j-th child as cj (x). Below algorithm describes how to split the i-th child for
a given node.
1: procedure Split-Child(node, i)
2:
x ci (node)
3:
y CREATE-NODE
4:
Insert(K(node), i, kt (x))
5:
Insert(C(node), i + 1, y)
6:
K(y) {kt+1 (x), kt+2 (x), ..., k2t1 (x)}
7:
K(x) {k1 (x), k2 (x), ..., kt1 (x)}
8:
if y is not leaf then
9:
C(y) {ct+1 (x), ct+2 (x), ..., c2t (x)}
10:
C(x) {c1 (x), c2 (x), ..., ct (x)}
The following example Python program implements this child splitting algorithm.
def split_child(node, i):
t = node.t
x = node.children[i]
y = BTree(t)
node.keys.insert(i, x.keys[t-1])
node.children.insert(i+1, y)
y.keys = x.keys[t:]
x.keys = x.keys[:t-1]

7.2. INSERTION

169
K[1]

C[1]

C[2]

K[2]

...

K[t]

C[t]

...

K[2t-1]

C[t+1]

...

C[2t-1]

C[2t]

(a) Before split

...

K[1]

C[1]

K[2]

C[2]

...

K[t]

...

K[t-1]

K[t+1]

C[t]

C[t+1]

...

K[2t-1]

C[2t-1]

(b) After split

Figure 7.4: Split node

if not is_leaf(x):
y.children = x.children[t:]
x.children = x.children[:t]

Where function is leaf test if a node is leaf.

def is_leaf(t):
return t.children == []

After splitting, a key is pushed up to its parent node. It is quite possible

that the parent node has already been full. And this pushing violates the B-tree
property.
In order to solve this problem, we can check from the root along the path
of insertion traversing till the leaf. If there is any node in this path is full,
the splitting is applied. Since the parent of this node has been examined, it is
ensured that there are less than 2t 1 keys in the parent. It wont make the
parent full if pushing up one key. This approach only need one single pass down
the tree without any back-tracking.
If the root need splitting, a new node is created as the new root. There is
no keys in this new created root, and the previous root is set as the only child.
After that, splitting is performed top-down. And we can insert the new key
finally.
1: function Insert(T, k)
2:
rT
3:
if r is full then
. root is full
4:
s CREATE-NODE
5:
C(s) {r}
6:
Split-Child(s, 1)
7:
rs
8:
return Insert-Nonfull(r, k)
Where algorithm Insert-Nonfull assumes the node passed in is not full.

170

CHAPTER 7. B-TREES

If it is a leaf node, the new key is inserted to the proper position based on the
order; Otherwise, the algorithm finds a proper child node to which the new key
will be inserted. If this child is full, splitting will be performed.
1: function Insert-Nonfull(T, k)
2:
if T is leaf then
3:
i1
4:
while i |K(T )| k > ki (T ) do
5:
ii+1
6:
Insert(K(T ), i, k)
7:
else
8:
i |K(T )|
9:
while i > 1 k < ki (T ) do
10:
ii1
11:
if ci (T ) is full then
12:
Split-Child(T, i)
13:
if k > ki (T ) then
14:
ii+1
15:
Insert-Nonfull(ci (T ), k)
16:
return T
This algorithm is recursive. In B-tree, the minimum degree t is typically
relative to magnetic disk structure. Even small depth can support huge amount
of data (with t = 10, maximum to 10 billion data can be stored in a B-tree with
height of 10). The recursion can also be eliminated. This is left as exercise to
the reader.
Figure 7.5 shows the result of continuously inserting keys G, M, P, X, A, C,
D, E, J, K, N, O, R, S, T, U, V, Y, Z to the empty tree. The first result is the
2-3-4 tree (t = 2). The second result shows how it varies when t = 3.
E

(a) 2-3-4 tree.

(b) t = 3

Figure 7.5: Insertion result

Below example Python program implements this algorithm.
def insert(tr, key):

7.2. INSERTION

171

root = tr
if is_full(root):
s = BTree(root.t)
s.children.insert(0, root)
split_child(s, 0)
root = s
return insert_nonfull(root, key)

And the insertion to non-full node is implemented as the following.

def insert_nonfull(tr, key):
if is_leaf(tr):
ordered_insert(tr.keys, key)
else:
i = len(tr.keys)
while i>0 and key < tr.keys[i-1]:
i = i-1
if is_full(tr.children[i]):
split_child(tr, i)
if key>tr.keys[i]:
i = i+1
insert_nonfull(tr.children[i], key)
return tr

Where function ordered insert is used to insert an element to an ordered

list. Function is full tests if a node contains 2t 1 keys.
def ordered_insert(lst, x):
i = len(lst)
lst.append(x)
while i>0 and lst[i]<lst[i-1]:
(lst[i-1], lst[i]) = (lst[i], lst[i-1])
i=i-1
def is_full(node):
return len(node.keys) 2 node.t - 1

For the array based collection, append on the tail is much more effective
than insert in other position, because the later takes O(n) time, if the length
of the collection is n. The ordered insert program firstly appends the new
element at the end of the existing collection, then iterates from the last element
to the first one, and checks if the current two elements next to each other are
ordered. If not, these two elements will be swapped.
Insert then fixing
In functional settings, B-tree insertion can be realized in a way similar to redblack tree. When insert a key to red-black tree, it is firstly inserted as in the
normal binary search tree, then recursive fixing is performed to resume the
balance of the tree. B-tree can be viewed as extension to the binary search tree,
that each node contains multiple keys and children. We can firstly insert the
key without considering if the node is full. Then perform fixing to satisfy the
minimum degree constraint.
insert(T, k) = f ix(ins(T, k))

(7.5)

172

CHAPTER 7. B-TREES

Function ins(T, k) traverse the B-tree T from root to find a proper position
where key k can be inserted. After that, function f ix is applied to resume the
B-tree properties. Denote B-tree in a form of T = (K, C, t), where K represents
keys, C represents children, and t is the minimum degree.
Below is the Haskell definition of B-tree.
data BTree a = Node{ keys :: [a]
, children :: [BTree a]
, degree :: Int} deriving (Eq)

The insertion function can be provided based on this definition.

insert tr x = fixRoot $ ins tr x

There are two cases when realize ins(T, k) function. If the tree T is leaf, k
is inserted to the keys; Otherwise if T is the branch node, we need recursively
insert k to the proper child.
Figure 7.6 shows the branch case. The algorithm first locates the position.
for certain key ki , if the new key k to be inserted satisfy ki1 < k < ki , Then
we need recursively insert k to child ci .
This position divides the node into 3 parts, the left part, the child ci and
the right part.
k, K[i-1]<k<K[i]
insert to
K[1]

C[1]

C[2]

K[2]

...

K[i-1]

C[i-1]

K[i]

C[i]

...

K[n]

C[i+1]

...

C[n]

C[n+1]

(a) Locate the child to insert.

K[1]

K[2]

...

K[i-1]

k, K[i-1]<k<K[i]

K[i]

K[i+1]

...

K[n]

recursive insert
C[1]

C[2]

...

C[i-1]

C[i]

C[i+1]

...

C[n+1]

(b) Recursive insert.

Figure 7.6: Insert a key to a branch node

(K 0 {k} K 00 , , t) : C = , (K 0 , K 00 ) = divide(K, k)
make((K , C1 ), ins(c, k), (K 00 , C20 )) : (C1 , C2 ) = split(|K 0 |, C)
(7.6)
The first clause deals with the leaf case. Function divide(K, k) divide keys
into two parts, all keys in the first part are not greater than k, and all rest keys
are not less than k.

ins(T, k) =

7.2. INSERTION

173

K = K 0 K 00 k 0 K, k 00 K 00 k 0 k k 00
The second clause handle the branch case. Function split(n, C) splits children in two parts, C1 and C2 . C1 contains the first n children; and C2 contains
the rest. Among C2 , the first child is denoted as c, and others are represented
as C20 .
Here the key k need be recursively inserted into child c. Function make
takes 3 parameter. The first and the third are pairs of key and children; the
second parameter is a child node. It examine if a B-tree node made from these
keys and children violates the minimum degree constraint and performs fixing
if necessary.

f ixF ull((K 0 , C 0 ), c, (K 00 , C 00 )) : f ull(c)

(K 0 K 00 , C 0 {c} C 00 , t) : otherwise
(7.7)
Where function f ull(c) tests if the child c is full. Function f ixF ull splits
the the child c, and forms a new B-tree node with the pushed up key.
0

make((K , C ), c, (K , C )) =

f ixF ull((K 0 , C 0 ), c, (K 00 , C 00 )) = (K 0 {k} K 00 , C 0 {c1 , c2 } C 00 , t) (7.8)

Where (c1 , k, c2 ) = split(c). During splitting, the first t 1 keys and t
children are extract to one new child, the last t 1 keys and t children form
another child. The t-th key is pushed up.
With all the above functions defined, we can realize f ix(T ) to complete the
functional B-tree insertion algorithm. It firstly checks if the root contains too
many keys. If it exceeds the limit, splitting will be applied. The split result will
be used to make a new node, so the total height of the tree increases by one.

f ix(T ) =

c : T = (, {c}, t)
({k}, {c1 , c2 }, t) : f ull(T ), (c1 , k, c2 ) = split(T )

T : otherwise

The following Haskell example code implements the B-tree insertion.

import qualified Data.List as L
ins (Node ks [] t) x = Node (L.insert x ks) [] t
ins (Node ks cs t) x = make (ks, cs) (ins c x) (ks, cs)
where
(ks, ks) = L.partition (<x) ks
(cs, (c:cs)) = L.splitAt (length ks) cs
fixRoot (Node [] [tr] _) = tr -- shrink height
fixRoot tr = if full tr then Node [k] [c1, c2] (degree tr)
else tr
where
(c1, k, c2) = split tr
make (ks, cs) c (ks, cs)
| full c = fixFull (ks, cs) c (ks, cs)

(7.9)

174

CHAPTER 7. B-TREES
| otherwise = Node (ks++ks) (cs++[c]++cs) (degree c)

fixFull (ks, cs) c (ks, cs) = Node (ks++[k]++ks)

(cs++[c1,c2]++cs) (degree c)
where
(c1, k, c2) = split c
full tr = (length $ keys tr) > 2(degree tr)-1

Figure 7.7 shows the varies of results of building B-trees by continuously

inserting keys GMPXACDEJKNORSTUVYZ.
E

(a) Insert result of a 2-3-4 tree.

(b) Insert result of a B-tree with t = 3

Figure 7.7: Insert then fixing results

Compare to the imperative insertion result as shown in figure 7.7 we can
found that there are different. However, they are all valid because all B-tree
properties are satisfied.

7.3

Deletion

Deleting a key from B-tree may violate balance properties. Except the root, a
node shouldnt contain too few keys less than t 1, where t is the minimum
degree.
Similar to the approaches for insertion, we can either do some preparation
so that the node from where the key being deleted contains enough keys; or do
some fixing after the deletion if the node has too few keys.

7.3.1

Merge before delete method

We start from the easiest case. If the key k to be deleted can be located in
node x, and x is a leaf node, we can directly remove k from x. If x is the root
(the only node of the tree), we neednt worry about there are too few keys after
deletion. This case is named as case 1 later.

7.3. DELETION

175

In most cases, we start from the root, along a path to locate where is the
node contains k. If k can be located in the internal node x, there are three sub
cases.
Case 2a, If the child y precedes k contains enough keys (more than t). We
replace k in node x with k 0 , which is the predecessor of k in child y. And
recursively remove k 0 from y.
The predecessor of k can be easily located as the last key of child y.
This is shown in figure 7.8.

Figure 7.8: Replace and delete from predecessor.

Case 2b, If y doesnt contain enough keys, while the child z follows k
contains more than t keys. We replace k in node x with k 00 , which is the
successor of k in child z. And recursively remove k 00 from z.
The successor of k can be easily located as the first key of child z.
This sub-case is illustrated in figure 7.9.
Case 2c, Otherwise, if neither y, nor z contains enough keys, we can merge
y, k and z into one new node, so that this new node contains 2t 1 keys.
After that, we can then recursively do the removing.
Note that after merge, if the current node doesnt contain any keys, which
means k is the only key in x. y and z are the only two children of x. we
need shrink the tree height by one.
Figure 7.10 illustrates this sub-case.
the last case states that, if k cant be located in node x, the algorithm need
find a child node ci in x, so that the sub-tree ci contains k. Before the deletion
is recursively applied in ci , we need make sure that there are at least t keys in
ci . If there are not enough keys, the following adjustment is performed.

176

CHAPTER 7. B-TREES

Figure 7.9: Replace and delete from successor.

Figure 7.10: Merge and delete.

7.3. DELETION

177

Case 3a, We check the two sibling of ci , which are ci1 and ci+1 . If either
one contains enough keys (at least t keys), we move one key from x down
to ci , and move one key from the sibling up to x. Also we need move the
relative child from the sibling to ci .
This operation makes ci contains enough keys for deletion. we can next
try to delete k from ci recursively.
Figure 7.11 illustrates this case.

Figure 7.11: Borrow from the right sibling.

Case 3b, In case neither one of the two siblings contains enough keys, we
then merge ci , a key from x, and either one of the sibling into a new node.
Then do the deletion on this new node.
Figure 7.12 shows this case.
Before define the B-tree delete algorithm, we need provide some auxiliary
functions. Function Can-Del tests if a node contains enough keys for deletion.
1: function Can-Del(T )
2:
return |K(T )| t
Procedure Merge-Children(T, i) merges child ci (T ), key ki (T ), and child
ci+1 (T ) into one big node.
1: procedure Merge-Children(T, i)
. Merge ci (T ), ki (T ), and ci+1 (T )
2:
x ci (T )
3:
y ci+1 (T )
4:
K(x) K(x) {ki (T )} K(y)
5:
C(x) C(x) C(y)
6:
Remove-At(K(T ), i)

178

CHAPTER 7. B-TREES

Figure 7.12: Merge ci , k, and ci+1 to a new node.

Remove-At(C(T ), i + 1)

Procedure Merge-Children merges the i-th child, the i-th key, and i + 1th child of node T into a new child, and remove the i-th key and i + 1-th child
from T after merging.
With these functions defined, the B-tree deletion algorithm can be given by
realizing the above 3 cases.
1: function Delete(T, k)
2:
i1
3:
while i |K(T )| do
4:
if k = ki (T ) then
5:
if T is leaf then
. case 1
6:
Remove(K(T ), k)
7:
else
. case 2
8:
if Can-Del(ci (T )) then
. case 2a
9:
ki (T ) Last-Key(ci (T ))
10:
Delete(ci (T ), ki (T ))
11:
else if Can-Del(ci+1 (T )) then
. case 2b
12:
ki (T ) First-Key(ci+1 (T ))
13:
Delete(ci+1 (T ), ki (T ))
14:
else
. case 2c
15:
Merge-Children(T, i)
16:
Delete(ci (T ), k)
17:
if K(T ) = N IL then
18:
T ci (T )
. Shrinks height

7.3. DELETION
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:

179

return T
else if k < ki (T ) then
Break
else
ii+1
if T is leaf then
return T
. k doesnt exist in T .
if Can-Del(ci (T )) then
. case 3
if i > 1 Can-Del(ci1 (T )) then
. case 3a: left sibling
Insert(K(ci (T )), ki1 (T ))
ki1 (T ) Pop-Back(K(ci1 (T )))
if ci (T ) isnt leaf then
c Pop-Back(C(ci1 (T )))
Insert(C(ci (T )), c)
else if i |C(T )| Can-Del(ci1 (T )) then . case 3a: right sibling
Append(K(ci (T )), ki (T ))
ki (T ) Pop-Front(K(ci+1 (T )))
if ci (T ) isnt leaf then
c Pop-Front(C(ci+1 (T )))
Append(C(ci (T )), c)
else
. case 3b
if i > 1 then
Merge-Children(T, i 1)
else
Merge-Children(T, i)
Delete(ci (T ), k)
. recursive delete
if K(T ) = N IL then
. Shrinks height
T c1 (T )
return T

Figure 7.13, 7.14, and 7.15 show the deleting process step by step. The nodes
modified are shaded.
The following example Python program implements the B-tree deletion algorithm.
def can_remove(tr):
return len(tr.keys) tr.t
def replace_key(tr, i, k):
tr.keys[i] = k
return k
def merge_children(tr, i):
tr.children[i].keys += [tr.keys[i]] + tr.children[i+1].keys
tr.children[i].children += tr.children[i+1].children
tr.keys.pop(i)
tr.children.pop(i+1)
def B_tree_delete(tr, key):
i = len(tr.keys)

180

CHAPTER 7. B-TREES

(a) A B-tree before deleting.

(b) After delete key F, case 1.

Figure 7.13: Result of B-tree deleting (1).

(a) After delete key M, case 2a.

(b) After delete key G, case 2c.

Figure 7.14: Result of B-tree deleting program (2)

7.3. DELETION

181
C

(a) After delete key D, case 3b, and height is shrunk.

(b) After delete key B, case 3a, borrow from right sibling.

Figure 7.15: Result of B-tree deleting program (3)

while i>0:
if key == tr.keys[i-1]:
if tr.leaf: # case 1 in CLRS
tr.keys.remove(key)
else: # case 2 in CLRS
if tr.children[i-1].can_remove(): # case 2a
key = tr.replace_key(i-1, tr.children[i-1].keys[-1])
B_tree_delete(tr.children[i-1], key)
elif tr.children[i].can_remove(): # case 2b
key = tr.replace_key(i-1, tr.children[i].keys[0])
B_tree_delete(tr.children[i], key)
else: # case 2c
tr.merge_children(i-1)
B_tree_delete(tr.children[i-1], key)
if tr.keys==[]: # tree shrinks in height
tr = tr.children[i-1]
return tr
elif key > tr.keys[i-1]:
break
else:
i = i-1
# case 3
if tr.leaf:
return tr #key doesnt exist at all
if not tr.children[i].can_remove():
if i>0 and tr.children[i-1].can_remove(): #left sibling
tr.children[i].keys.insert(0, tr.keys[i-1])
tr.keys[i-1] = tr.children[i-1].keys.pop()

182

CHAPTER 7. B-TREES
if not tr.children[i].leaf:
tr.children[i].children.insert(0, tr.children[i-1].children.pop())
elif i<len(tr.children) and tr.children[i+1].can_remove(): #right sibling
tr.children[i].keys.append(tr.keys[i])
tr.keys[i]=tr.children[i+1].keys.pop(0)
if not tr.children[i].leaf:
tr.children[i].children.append(tr.children[i+1].children.pop(0))
else: # case 3b
if i>0:
tr.merge_children(i-1)
else:
tr.merge_children(i)
B_tree_delete(tr.children[i], key)
if tr.keys==[]: # tree shrinks in height
tr = tr.children[0]
return tr

7.3.2

Delete and fix method

The merge and delete algorithm is a bit complex. There are several cases, and
in each case, there are sub cases to deal.
Another approach to design the deleting algorithm is to perform fixing after
deletion. It is similar to the insert-then-fix strategy.
delete(T, k) = f ix(del(T, k))

(7.10)

When delete a key from B-tree, we firstly locate which node this key is
contained. We traverse from the root to the leaves till find this key in some
node.
If this node is a leaf, we can remove the key, and then examine if the deletion
makes the node contains too few keys to satisfy the B-tree balance properties.
If it is a branch node, removing the key breaks the node into two parts. We
need merge them together. The merging is a recursive process which is shown
in figure 7.16.
When do merging, if the two nodes are not leaves, we merge the keys together, and recursively merge the last child of the left part and the first child
of the right part to one new node. Otherwise, if they are leaves, we merely put
all keys together.
Till now, the deleting in performed in straightforward way. However, deleting decreases the number of keys of a node, and it may result in violating the
B-tree balance properties. The solution is to perform fixing along the path
traversed from root.
During the recursive deletion, the branch node is broken into 3 parts. The
left part contains all keys less than k, includes k1 , k2 , ..., ki1 , and children
c1 , c2 , ..., ci1 , the right part contains all keys greater than k, say ki , ki+1 , ..., kn+1 ,
and children ci+1 , ci+2 , ..., cn+1 . Then key k is recursively deleted from child ci .
Denote the result becomes c0i after that. We need make a new node from these
3 parts, as shown in figure 7.17.
At this time point, we need examine if c0i contains enough keys. If there
are to less keys (less than t 1, but not t in contrast to the merge-and-delete
approach), we can either borrow a key-child pair from the left or the right part,

7.3. DELETION

183

Figure 7.16: Delete a key from a branch node. Removing ki breaks the node
into 2 parts. Merging these 2 parts is a recursive process. When the two parts
are leaves, the merging terminates.

Figure 7.17: After delete key k from node ci , denote the result as c0i . The fixing
makes a new node from the left part, c0i and the right part.

184

CHAPTER 7. B-TREES

and do inverse operation of splitting. Figure 7.18 shows example of borrowing

from the left part.

Figure 7.18: Borrow a key-child pair from left part and un-split to a new child.
If both left part and right part are empty, we can simply push c0i up.
Denote the B-tree as T = (K, C, t), where K and C are keys and children.
The del(T, k) function deletes key k from the tree.

del(T, k) =

(delete(K, k), , t) : C =
merge((K1 , C1 , t), (K2 , C2 , t)) : ki = k

/K
make((K10 , C10 ), del(c, k), (K20 , C20 )) : k

(7.11)

If children C = is empty, T is leaf. k is deleted from keys directly.

Otherwise, T is internal node. If k K, removing it separates the keys and
children in two parts (K1 , C1 ) and (K2 , C2 ). They will be recursively merged.
K1 = {k1 , k2 , ..., ki1 }
K2 = {ki+1 , ki+2 , ..., km }
C1 = {c1 , c2 , ..., ci }
C2 = {ci+1 , ci+2 , ..., cm+1 }
If k
/ K, we need locate a child c, and further delete k from it.
(K10 , K20 ) = ({k 0 |k 0 K, k 0 < k}, {k 0 |k 0 K, k < k 0 })
(C10 , {c} C20 ) = splitAt(|K10 |, C)
The recursive merge function is defined as the following. When merge two
trees T1 = (K1 , C1 , t) and T2 = (K2 , C2 , t), if both are leaves, we create a new
leave by concatenating the keys. Otherwise, the last child in C1 , and the first
child in C2 are recursively merged. And we call make function to form the new
tree. When C1 and C2 are not empty, denote the last child of C1 as c1,m , the

7.3. DELETION

185

rest as C10 ; the first child of C2 as C2,1 , the rest as C20 . Below equation defines
the merge function.

(K1 K2 , , t) : C1 = C2 =
make((K1 , C10 ), merge(c1,m , c2,1 ), (K2 , C20 )) : otherwise
(7.12)
The make function defined above only handles the case that a node contains
too many keys due to insertion. When delete key, it may cause a node contains
too few keys. We need test and fix this situation as well.

merge(T1 , T2 ) =

f ixF ull((K 0 , C 0 ), c, (K 00 , C 00 )) : f ull(c)

0
0
00
00
f ixLow((K 0 , C 0 ), c, (K 00 , C 00 )) : low(c)
make((K , C ), c, (K , C )) =

(K 0 K 00 , C 0 {c} C 00 , t) : otherwise
(7.13)
Where low(T ) checks if there are too few keys less than t 1. Function
f ixLow(Pl , c, Pr ) takes three arguments, the left pair of keys and children, a
child node, and the right pair of keys and children. If the left part isnt empty, we
borrow a pair of key-child, and do un-splitting to make the child contain enough
keys, then recursively call make; If the right part isnt empty, we borrow a pair
from the right; and if both sides are empty, we return the child node as result.
In this case, the height of the tree shrinks.
Denote the left part Pl = (Kl , Cl ). If Kl isnt empty, the last key and child
are represented as kl,m and cl,m respectively. The rest keys and children become
Kl0 and Cl0 ; Similarly, the right part is denoted as Pr = (Kr , Cr ). If Kr isnt
empty, the first key and child are represented as kr,1 , and cr,1 . The rest keys
and children are Kr0 and Cr0 . Below equation gives the definition of f ixLow.

make((Kl0 , Cl0 ), unsplit(cl,m , kl,m , c), (Kr , Cr )) : Kl 6=

6
make((Kr , Cr ), unsplit(c, kr,1 , cr,1 ), (Kr0 , Cr0 )) : Kr =
f ixLow(Pl , c, Pr ) =

c : otherwise
(7.14)
Function unsplit(T1 , k, T2 ) is the inverse operation to splitting. It forms a
new B-tree nodes from two small nodes and a key.
unsplit(T1 , k, T2 ) = (K1 {k} K2 , C1 C2 , t)

(7.15)

The following example Haskell program implements the B-tree deletion algorithm.
import qualified Data.List as L
delete tr x = fixRoot $ del tr x
del:: (Ord a) BTree a a BTree a
del (Node ks [] t) x = Node (L.delete x ks) [] t
del (Node ks cs t) x =
case L.elemIndex x ks of
Just i merge (Node (take i ks) (take (i+1) cs) t)
(Node (drop (i+1) ks) (drop (i+1) cs) t)
Nothing make (ks, cs) (del c x) (ks, cs)

186

CHAPTER 7. B-TREES
where
(ks, ks) = L.partition (<x) ks
(cs, (c:cs)) = L.splitAt (length ks) cs

merge (Node ks [] t) (Node ks [] _) = Node (ks++ks) [] t

merge (Node ks cs t) (Node ks cs _) = make (ks, init cs)
(merge (last cs) (head cs))
(ks, tail cs)
make (ks, cs) c (ks, cs)
| full c = fixFull (ks, cs) c (ks, cs)
| low c = fixLow (ks, cs) c (ks, cs)
| otherwise = Node (ks++ks) (cs++[c]++cs) (degree c)
low tr = (length $ keys tr) < (degree tr)-1
fixLow (ks@(_:_), cs) c (ks, cs) = make (init ks, init cs)
(unsplit (last cs) (last ks) c)
(ks, cs)
fixLow (ks, cs) c (ks@(_:_), cs) = make (ks, cs)
(unsplit c (head ks) (head cs))
(tail ks, tail cs)
fixLow _ c _ = c
unsplit c1 k c2 = Node ((keys c1)++[k]++(keys c2))
((children c1)++(children c2)) (degree c1)

When delete the same keys from the B-tree as in merge and fixing approach,
the results are different. However, both satisfy the B-tree properties, so they
are all valid.
M

(a) B-tree before deleting

(b) After delete key E.

Figure 7.19: Result of delete-then-fixing (1)

7.4. SEARCHING

187
M

(a) After delete key G,

(b) After delete key A.

Figure 7.20: Result of delete-then-fixing (2)

(a) After delete key M.

(b) After delete key U.

Figure 7.21: Result of delete-then-fixing (3)

7.4

Searching

Searching in B-tree can be considered as the generalized tree search extended

from binary search tree.
When searching in the binary tree, there are only 2 different directions, the
left and the right. However, there are multiple directions in B-tree.
1: function Search(T, k)
2:
loop
3:
i1
4:
while i |K(T )| k > ki (T ) do
5:
ii+1
6:
if i |K(T )| k = ki (T ) then

188
7:
8:
9:
10:
11:

CHAPTER 7. B-TREES
return (T, i)
if T is leaf then
return N IL
else
T ci (T )

. k doesnt exist

Starts from the root, this program examines each key one by one from the
smallest to the biggest. In case it finds the matched key, it returns the current
node and the index of this key. Otherwise, if it finds the position i that ki <
k < ki+1 , the program will next search the child node ci+1 for the key. If it
traverses to some leaf node, and fails to find the key, the empty value is returned
to indicate that this key doesnt exist in the tree.
The following example Python program implements the search algorithm.
def B_tree_search(tr, key):
while True:
for i in range(len(tr.keys)):
if key tr.keys[i]:
break
if key == tr.keys[i]:
return (tr, i)
if tr.leaf:
return None
else:
if key > tr.keys[-1]:
i=i+1
tr = tr.children[i]

The search algorithm can also be realized by recursion. When search key k
in B-tree T = (K, C, t), we partition the keys with k.
K1 = {k 0 |k 0 < k}
K2 = {k 0 |k k 0 }
Thus K1 contains all the keys less than k, and K2 holds the rest. If the first
element in K2 is equal to k, we find the key. Otherwise, we recursively search
the key in child c|K1 |+1 .

search(T, k) =

(T, |K1 | + 1) :
:

search(c|K1 |+1 , k) :

k K2
C=
otherwise

Below example Haskell program implements this algorithm.

search :: (Ord a) BTree a a Maybe (BTree a, Int)
search tr@(Node ks cs _) k
| matchFirst k $ drop len ks = Just (tr, len)
| otherwise = if null cs then Nothing
else search (cs !! len) k
where
matchFirst x (y:_) = x==y
matchFirst x _ = False
len = length $ filter (<k) ks

(7.16)

7.5. NOTES AND SHORT SUMMARY

7.5

189

Notes and short summary

In this chapter, we explained the B-tree data structure as a kind of extension

from binary search tree. The background knowledge of magnetic disk access is
skipped, user can refer to [2] for detail. For the three main operations, insertion,
deletion, and searching, both imperative and functional algorithms are given.
They traverse from the root to the leaf. All the three operations perform in
time proportion to the height of the tree. Because B-tree always maintains the
balance properties. The performance is ensured to bound to O(lg n) time, where
n is the number of the keys in B-tree.

Exercise 7.1
Eliminate the recursion in imperative B-tree insertion algorithm.

190

CHAPTER 7. B-TREES

Bibliography
[1] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.
[2] B-tree, Wikipedia. http://en.wikipedia.org/wiki/B-tree
[3] Chris Okasaki. FUNCTIONAL PEARLS Red-Black Trees in a Functional
Setting. J. Functional Programming. 1998

191

192

BIBLIOGRAPHY

Part III

Heaps

193

Chapter 8

Binary Heaps
8.1

Introduction

Heap is one of the elementary data structure. It is widely used to solve some
practical problems, such as sorting, prioritized scheduling, and graph algorithms[2].
Most popular implementations of heap are using a kind of implicit binary
heap by array, which is described in [2]. Examples include C++/STL heap
and Python heapq. The most efficient heap sort algorithm is also realized with
binary heap as proposed by R. W. Floyd [3] [5].
However, heaps can be general and realized with varies of other data structures besides array. In this chapter, explicit binary tree is used. It leads to
Leftist heaps, Skew heaps, and Splay heaps, which are suitable for purely functional implementation as shown by Okasaki[6].
A heap is a data structure that satisfies the following heap property.
Top operation always returns the minimum (maximum) element;
Pop operation removes the top element from the heap while the heap
property should be kept, so that the new top element is still the minimum
(maximum) one;
Insert a new element to heap should keep the heap property. That the
new top is still the minimum (maximum) element;
Other operations including merge etc should all keep the heap property.
This is a kind of recursive definition, while it doesnt limit the under ground
data structure.
We call the heap with the minimum element on top as min-heap, while if
the top keeps the maximum element, we call it max-heap.

8.2

Implicit binary heap by array

Considering the heap definition in previous section, one option to implement

heap is by using trees. A straightforward solution is to store the minimum
(maximum) element in the root of the tree, so for top operation, we simply
195

196

CHAPTER 8. BINARY HEAPS

return the root as the result. And for pop operation, we can remove the root
and rebuild the tree from the children.
If binary tree is used to implement heap the heap, we can call it binary heap.
This chapter explains three different realizations for binary heap.

8.2.1

Definition

The first one is implicit binary tree. Consider the problem how to represent
a complete binary tree with array. (For example, try to represent a complete
binary tree in the programming language doesnt support structure or record
data type. Only array can be used). One solution is to pack all elements from
top level (root) down to bottom level (leaves).
Figure 8.1 shows a complete binary tree and its corresponding array representation.
16

Figure 8.1: Mapping between a complete binary tree and array

This mapping between tree and array can be defined as the following equations (The array index starts from 1).
1: function PARENT(i)
2:
return b 2i c
3:
4:

function LEFT(i)
return 2i

function LEFT(i)
return 2i + 1

For a given tree node which is represented as the i-th element of the array,
since the tree is complete, we can easily find its parent node as the bi/2c-th

8.2. IMPLICIT BINARY HEAP BY ARRAY

197

element; Its left child with index of 2i and right child of 2i + 1. If the index of
the child exceeds the length of the array, it means this node does not have such
a child (leaf for example).
In real implementation, this mapping can be calculated fast with bit-wise
operation like the following example ANSI C code. Note that, the array index
starts from zero in C like languages.
#define PARENT(i) ((((i) + 1) >> 1) - 1)
#define LEFT(i) (((i) << 1) + 1)
#define RIGHT(i) (((i) + 1) << 1)

8.2.2

Heapify

The most important thing for heap algorithm is to maintain the heap property,
that the top element should be the minimum (maximum) one.
For the implicit binary heap by array, it means for a given node, which is
represented as the i-th index, we need develop a method to check if both its two
children conform to this property. In case there is violation, we need swap the
parent and child recursively [2] to fix the problem.
Below algorithm shows the iterative solution to enforce the min-heap property from the given index of the array.
1: function Heapify(A, i)
2:
n |A|
3:
loop
4:
l Left(i)
5:
r Right(i)
6:
smallest i
7:
if l < n A[l] < A[i] then
8:
smallest l
9:
if r < n A[r] < A[smallest] then
10:
smallest r
11:
if smallest 6= i then
12:
Exchange A[i] A[smallest]
13:
i smallest
14:
else
15:
return
For array A and the given index i, None its children should be bigger than
A[i], in case there is violation, we pick the smallest element as A[i], and swap
the previous A[i] to child. The algorithm traverses the tree top-down to fix the
heap property until either reach a leaf or there is no heap property violation.
The Heapify algorithm takes O(lg n) time, where n is the number of elements. This is because the loop time is proportion to the height of the binary
tree.
When implement this algorithm, the comparison method can be passed as
a parameter, so that both min-heap and max-heap can be supported. The
following ANSI C example code uses this approach.
typedef int (Less)(Key, Key);
int less(Key x, Key y) { return x < y; }

198

CHAPTER 8. BINARY HEAPS

int notless(Key x, Key y) { return !less(x, y); }

void heapify(Key a, int i, int n, Less lt) {
int l, r, m;
while (1) {
l = LEFT(i);
r = RIGHT(i);
m = i;
if (l < n && lt(a[l], a[i]))
m = l;
if (r < n && lt(a[r], a[m]))
m = r;
if (m != i) {
swap(a, i, m);
i = m;
}
else
break;
}
}

Figure 8.2 illustrates the steps when Heapify processing the array {16, 4, 10, 14, 7, 9, 3, 2, 8, 1}.
The array changes to {16, 14, 10, 8, 7, 9, 3, 2, 4, 1}.

8.2.3

Build a heap

With Heapify algorithm defined, it is easy to build a heap from the arbitrary
array. Observe that the numbers of nodes in a complete binary tree for each
level is a list like below:
1, 2, 4, 8, ..., 2i , ....
The only exception is the last level. Since the tree may not full (note that
complete binary tree doesnt mean full binary tree), the last level contains at
most 2p1 nodes, where 2p n and n is the length of the array.
The Heapify algorithm doesnt have any effect on leave node. We can skip
applying Heapify for all leaves. In other words, all leaf nodes have already
satisfied the heap property. We only need start checking and maintain the heap
property from the last branch node. the index of the last branch node is no
greater than bn/2c.
Based on this fact, we can build a heap with the following algorithm. (Assume the heap is min-heap).
1: function Build-Heap(A)
2:
n |A|
3:
for i bn/2c down to 1 do
4:
Heapify(A, i)
Although the complexity of Heapify is O(lg n), the running time of BuildHeap is not bound to O(n lg n) but O(n). This is a linear time algorithm. [2]
provides the detailed proof.
Below ANSI C example program implements this heap building function.
void build_heap(Key a, int n, Less lt) {
int i;
for (i = (n-1) >> 1; i 0; --i)

8.2. IMPLICIT BINARY HEAP BY ARRAY

199

a. Step 1, 14 is the biggest element among 4, 14, and 7. Swap 4 with the left
child;
16

b. Step 2, 8 is the biggest element among 2, 4, and 8. Swap 4 with the right
child;
16

c. 4 is the leaf node. It hasnt any children. Process terminates.

Figure 8.2: Heapify example, a max-heap case.

200

CHAPTER 8. BINARY HEAPS

heapify(a, i, n, lt);

Figure 8.3 and 8.4 show the steps when building a heap from array {4, 1, 3, 2, 16, 9, 10, 14, 8, 7}.
The node in black color is the one Heapify being applied, the nodes in gray
color are swapped during to keep the heap property.
4

a. An array in arbitrary order before heap building process;

b. Step 1, The array is mapped to binary tree. The first branch node, which is
16 is examined;
4

c. Step 2, 16 is the largest element in current sub tree, next is to check node
with value 2;
Figure 8.3: Build a heap from the arbitrary array. Gray nodes are changed in
each step, black node will be processed next step.

8.2.4

Basic heap operations

From the generic definition of heap (not necessarily the binary heap), Its essential to provides basic operations so that user can access the data and modify
it.
The most important operations include accessing the top element (find the
minimum or maximum one), popping the top element from the heap, finding
the top n elements, decreasing a key (this is for min-heap, and it is increasing
a key for max-heap), and insertion.
For the binary tree, most of operations are bound to O(lg n) in worst-case,
some of them, such as top is O(1) constant time.

8.2. IMPLICIT BINARY HEAP BY ARRAY

201

d. Step 3, 14 is the largest value in the sub-tree, swap 14 and 2; next is to

check node with value 3;
4

e. Step 4, 10 is the largest value in the sub-tree, swap 10 and 3; next is to

check node with value 1;
4

f. Step 5, 16 is the largest value in current node, swap 16 and 1 first; then
similarly, swap 1 and 7; next is to check the root node with value 4;
16

g. Step 6, Swap 4 and 16, then swap 4 and 14, and then swap 4 and 8; And
the whole build process finish.
Figure 8.4: Build a heap from the arbitrary array. Gray nodes are changed in
each step, black node will be processed next step.

202

CHAPTER 8. BINARY HEAPS

Access the top element (the minimum)

We need provide a way to access the top element efficiently. For the binary tree
realization, it is the root stores the minimum (maximum) value. This is the first
element in the array.
1: function TOP(A)
2:
return A[1]
This operation is trivial. It takes O(1) time. Here we skip the error handling
for empty case. If the heap is empty, one option is to raise an error.
Heap Pop
Pop operation is more complex than accessing the top, because the heap property has to be maintained after the top element is removed.
The solution is to apply Heapify algorithm to the next element after the
root is removed.
One simple but slow method based on this idea looks like the following.
1: function Pop-Slow(A)
2:
x Top(A)
3:
Remove(A, 1)
4:
if A is not empty then
5:
Heapify(A, 1)
6:
return x
This algorithm firstly records the top element in x, then it removes the first
element from the array, the size of this array is reduced by one. After that if the
array isnt empty, Heapify will applied to the new array from the first element
(It was previously the second one).
Removing the first element from array takes O(n) time, where n is the length
of the array. This is because we need shift all the rest elements one by one. This
bottle neck slows the whole algorithm to linear time.
In order to solve this problem, one alternative is to swap the first element
with the last one in the array, then shrink the array size by one.
1: function Pop(A)
2:
x Top(A)
3:
n Heap-Size(A)
4:
Exchange A[1] A[n]
5:
Remove(A, n)
6:
if A is not empty then
7:
Heapify(A, 1)
8:
return x
Removing the last element from the array takes only constant O(1) time, and
Heapify is bound to O(lg n). Thus the whole algorithm performs in O(lg n)
time. The following example ANSI C program implements this algorithm1 .
Key pop(Key a, int n, Less lt) {
swap(a, 0, --n);
heapify(a, 0, n, lt);
1 This program does not actually remove the last element, it reuse the last cell to store the
popped result

8.2. IMPLICIT BINARY HEAP BY ARRAY

203

return a[n];
}

Since the top element is removed from the array, instead swapping, this
program overwrites it with the last element before applying Heapify.
Find the top k elements
With pop defined, it is easy to find the top k elements from array. we can build
a max-heap from the array, then perform pop operation k times.
1: function Top-k(A, k)
2:
R
3:
Build-Heap(A)
4:
for i 1 to Min(k, Length(A)) do
5:
Append(R, Pop(A))
6:
return R
If k is greater than the length of the array, we need return the whole array
as the result. Thats why it calls the Min function to determine the number of
loops.
Below example Python program implements the top-k algorithm.
def top_k(x, k, less_p = MIN_HEAP):
build_heap(x, less_p)
return [heap_pop(x, less_p) for _ in range(min(k, len(x)))]

Decrease key
Heap can be used to implement priority queue. It is important to support key
modification in heap. One typical operation is to increase the priority of a tasks
so that it can be performed earlier.
Here we present the decrease key operation for a min-heap. The corresponding operation is increase key for max-heap. Figure 8.5 illustrate such a case for
a max-heap. The key of the 9-th node is increased from 4 to 15.
Once a key is decreased in a min-heap, it may make the node conflict with
the heap property, that the key may be less than some ancestor. In order to
maintain the invariant, the following auxiliary algorithm is defined to resume
the heap property.
1: function Heap-Fix(A, i)
2:
while i > 1 A[i] < A[ Parent(i) ] do
3:
Exchange A[i] A[ Parent(i) ]
4:
i Parent(i)
This algorithm repeatedly compares the keys of parent node and current
node. It swap the nodes if the parent contains the smaller key. This process
is performed from the current node towards the root node till it finds that the
parent node holds the smaller key.
With this auxiliary algorithm, decrease key can be realized as below.
1: function Decrease-Key(A, i, k)
2:
if k < A[i] then
3:
A[i] k
4:
Heap-Fix(A, i)

204

CHAPTER 8. BINARY HEAPS

a. The 9-th node with key 4 will be modified;

b. The key is modified to 15, which is greater than its parent;

c. According the max-heap property, 8 and 15 are swapped.

d. Since 15 is greater than its parent 14, they are swapped. After that,
because 15 is less than 16, the process terminates.
Figure 8.5: Example process when increase a key in a max-heap.

8.2. IMPLICIT BINARY HEAP BY ARRAY

205

This algorithm is only triggered when the new key is less than the original
key. The performance is bound to O(lg n). Below example ANSI C program
implements the algorithm.
void heap_fix(Key a, int i, Less lt) {
while (i > 0 && lt(a[i], a[PARENT(i)])) {
swap(a, i, PARENT(i));
i = PARENT(i);
}
}
void decrease_key(Key a, int i, Key k, Less lt) {
if (lt(k, a[i])) {
a[i] = k;
heap_fix(a, i, lt);
}
}

Insertion
In some materials like [2], insertion is implemented by using Decrease-Key.
A new node with as key is created. According to the min-heap property, it
should be the last element in the under ground array. After that, the key is
decreased to the value to be inserted, and Decrease-Key is called to fix any
violation to the heap property.
Alternatively, we can reuse Heap-Fix to implement insertion. The new key
is directly appended at the end of the array, and the Heap-Fix is applied to
this new node.
1: function Heap-Push(A, k)
2:
Append(A, k)
3:
Heap-Fix(A, |A|)
The following example Python program implements the heap insertion algorithm.
def heap_insert(x, key, less_p = MIN_HEAP):
i = len(x)
x.append(key)
heap_fix(x, i, less_p)

8.2.5

Heap sort

Heap sort is interesting application of heap. According to the heap property,

the min(max) element can be easily accessed by from the top of the heap. A
straightforward way to sort a list of values is to build a heap from them, then
continuously pop the smallest element till the heap is empty.
The algorithm based on this idea can be defined like below.
1: function Heap-Sort(A)
2:
R
3:
Build-Heap(A)
4:
while A 6= do
5:
Append(R, Heap-Pop(A))

206
6:

CHAPTER 8. BINARY HEAPS

return R
The following Python example program implements this definition.

def heap_sort(x, less_p = MIN_HEAP):

res = []
build_heap(x, less_p)
while x!=[]:
res.append(heap_pop(x, less_p))
return res

When sort n elements, the Build-Heap is bound to O(n). Since pop is

O(lg n), and it is called n times, so the overall sorting takes O(n lg n) time to
run. Because we use another list to hold the result, the space requirement is
O(n).
Robert. W. Floyd found a fast implementation of heap sort. The idea is to
build a max-heap instead of min-heap, so the first element is the biggest one.
Then the biggest element is swapped with the last element in the array, so that
it is in the right position after sorting. As the last element becomes the new
top, it may violate the heap property. We can shrink the heap size by one and
perform Heapify to resume the heap property. This process is repeated till
there is only one element left in the heap.
1: function Heap-Sort(A)
2:
Build-Max-Heap(A)
3:
while |A| > 1 do
4:
Exchange A[1] A[n]
5:
|A| |A| 1
6:
Heapify(A, 1)
This is in-place algorithm, it neednt any extra spaces to hold the result.
The following. The following ANSI C example code implements this algorithm.
void heap_sort(Key a, int n) {
build_heap(a, n, notless);
while(n > 1) {
swap(a, 0, --n);
heapify(a, 0, n, notless);
}
}

Exercise 8.1
Somebody considers one alternative to realize in-place heap sort. Take
sorting the array in ascending order as example, the first step is to build the
array as a minimum heap A, but not the maximum heap like the Floyds
method. After that the first element a1 is in the correct place. Next, treat
the rest {a2 , a3 , ..., an } as a new heap, and perform Heapify to them from
a2 for these n 1 elements. Repeating this advance and Heapify step
from left to right would sort the array. The following example ANSI C
code illustrates this idea. Is this solution correct? If yes, prove it; if not,
why?
void heap_sort(Key a, int n) {
build_heap(a, n, less);
while(--n)

8.3. LEFTIST HEAP AND SKEW HEAP, THE EXPLICIT BINARY HEAPS207
heapify(++a, 0, n, less);
}

Because of the same reason, can we perform Heapify from left to right k
times to realize in-place top-k algorithm like below ANSI C code?
int tops(int k, Key a, int n, Less lt) {
build_heap(a, n, lt);
for (k = MIN(k, n) - 1; k; --k)
heapify(++a, 0, --n, lt);
return k;
}

8.3

Leftist heap and Skew heap, the explicit binary heaps

Instead of using implicit binary tree by array, it is natural to consider why we

cant use explicit binary tree to realize heap?
There are some problems must be solved if we turn into explicit binary tree
as the under ground data structure.
The first problem is about the Heap-Pop or Delete-Min operation. Consider the binary tree is represented in form of left, key, and right as (L, k, R),
which is shown in figure 8.6

Figure 8.6: A binary tree, all elements in children are smaller than k.
If k is the top element, all elements in left and right children are less than k.
After k is popped, only left and right children are left. They have to be merged
to a new tree. Since heap property should be maintained after merge, the new
root is still the smallest element.
Because both left and right children are binary trees conforming heap property, the two trivial cases can be defined immediately.

H 2 : H1 =
H 1 : H2 =
merge(H1 , H2 ) =

? : otherwise

208

CHAPTER 8. BINARY HEAPS

Where means empty heap.

If neither left nor right child is empty, because they all fit heap property, the
top elements of them are all the minimum respectively. We can compare these
two roots, and select the smaller as the new root of the merged heap.
For instance, let L = (A, x, B) and R = (A0 , y, B 0 ), where A, A0 , B, and B 0
are all sub trees. If x < y, x will be the new root. We can either keep A, and
recursively merge B and R; or keep B, and merge A and R, so the new heap
can be one of the following.
(merge(A, R), x, B)
(A, x, merge(B, R))
Both are correct. One simplified solution is to only merge the right sub tree.
Leftist tree provides a systematically approach based on this idea.

8.3.1

Definition

The heap implemented by Leftist tree is called Leftist heap. Leftist tree is first
introduced by C. A. Crane in 1972[6].
Rank (S-value)
In Leftist tree, a rank value (or S value) is defined for each node. Rank is the
distance to the nearest external node. Where external node is the NIL concept
extended from the leaf node.
For example, in figure 8.7, the rank of NIL is defined 0, consider the root
node 4, The nearest leaf node is the child of node 8. So the rank of root node
4 is 2. Because node 6 and node 8 both only contain NIL, so their rank values
are 1. Although node 5 has non-NIL left child, However, since the right child is
NIL, so the rank value, which is the minimum distance to leaf is still 1.
4

NIL

Figure 8.7: rank(4) = 2, rank(6) = rank(8) = rank(5) = 1.

NIL

8.3. LEFTIST HEAP AND SKEW HEAP, THE EXPLICIT BINARY HEAPS209
Leftist property
With rank defined, we can create a strategy when merging.
Every time when merging, we always merge to right child; Denote the
rank of the new right sub tree as rr ;
Compare the ranks of the left and right children, if the rank of left sub
tree is rl and rl < rr , we swap the left and the right children.
We call this Leftist property. In general, a Leftist tree always has the
shortest path to some external node on the right.
Leftist tree tends to be very unbalanced, However, it ensures important
property as specified in the following theorem.
Theorem 8.3.1. If a Leftist tree T contains n internal nodes, the path from
root to the rightmost external node contains at most blog(n + 1)c nodes.
We skip the proof here, readers can refer to [7] and [1] for more information.
With this theorem, algorithms operate along this path are all bound to O(lg n).
We can reuse the binary tree definition, and augment with a rank field to
define the Leftist tree. For example in form of (r, k, L, R) for non-empty case.
The below Haskell defines Leftist tree.
data LHeap a = E -- Empty
| Node Int a (LHeap a) (LHeap a) -- rank, element, left, right

For empty tree, the rank is defined as zero. Otherwise, its the value of the
augmented field. A rank(H) function can be given to cover both cases.

0 : H=
rank(H) =
(8.1)
r : otherwise, H = (r, k, L, R)
Here is the example Haskell rank function.
rank E = 0
rank (Node r _ _ _) = r

In the rest of this section, we denote rank(H) as rH

8.3.2

Merge

In order to realize merge, we need develop the auxiliary algorithm to compare

the ranks and swap the children if necessary.

(rA + 1, k, B, A) : rA < rB
mk(k, A, B) =
(8.2)
(rB + 1, k, A, B) : otherwise
This function takes three arguments, a key and two sub trees A, and B. if
the rank of A is smaller, it builds a bigger tree with B as the left child, and A
as the right child. It increment the rank of A by 1 as the rank of the new tree;
Otherwise if B holds the smaller rank, then A is set as the left child, and B
becomes the right. The resulting rank is rb + 1.
The reason why rank need be increased by one is because there is a new key
added on top of the tree. It causes the rank increasing.

210

CHAPTER 8. BINARY HEAPS

Denote the key, the left and right children for H1 and H2 as k1 , L1 , R1 , and
k2 , L2 , R2 respectively. The merge(H1 , H2 ) function can be completed by using
this auxiliary tool as below

H2 : H 1 =
H1 : H 2 =
merge(H1 , H2 ) =
mk(k
,
L
,
merge(R
,
H

1
1
1
2 )) : k1 < k2

mk(k2 , L2 , merge(H1 , R2 )) : otherwise

(8.3)

The merge function is always recursively called on the right side, and the
Leftist property is maintained. These facts ensure the performance being bound
to O(lg n).
The following Haskell example code implements the merge program.
merge E h = h
merge h E = h
merge h1@(Node _ x l r) h2@(Node _ y l r) =
if x < y then makeNode x l (merge r h2)
else makeNode y l (merge h1 r)
makeNode x a b = if rank a < rank b then Node (rank a + 1) x b a
else Node (rank b + 1) x a b

Merge operation in implicit binary heap by array

Implicit binary heap by array performs very fast in most cases, and it fits modern
computer with cache technology well. However, merge is the algorithm bounds
to O(n) time. The typical realization is to concatenate two arrays together and
make a heap for this array [13].
1: function Merge-Heap(A, B)
2:
C Concat(A, B)
3:
Build-Heap(C)

8.3.3

Basic heap operations

Most of the basic heap operations can be implemented with merge algorithm
defined above.
Top and pop
Because the smallest element is always held in root, its trivial to find the
minimum value. Its constant O(1) operation. Below equation extracts the root
from non-empty heap H = (r, k, L, R). The error handling for empty case is
skipped here.
top(H) = k

(8.4)

For pop operation, firstly, the top element is removed, then left and right
children are merged to a new heap.
pop(H) = merge(L, R)

(8.5)

8.3. LEFTIST HEAP AND SKEW HEAP, THE EXPLICIT BINARY HEAPS211
Because it calls merge directly, the pop operation on Leftist heap is bound
to O(lg n).
Insertion
To insert a new element, one solution is to create a single leaf node with the
element, and then merge this leaf node to the existing Leftist tree.
insert(H, k) = merge(H, (1, k, , ))

(8.6)

It is O(lg n) algorithm since insertion also calls merge directly.

There is a convenient way to build the Leftist heap from a list. We can
continuously insert the elements one by one to the empty heap. This can be
realized by folding.
build(L) = f old(insert, , L)

(8.7)

Figure 8.8 shows one example Leftist tree built in this way.
1

Figure 8.8: A Leftist tree built from list {9, 4, 16, 7, 10, 2, 14, 3, 8, 1}.

The following example Haskell code gives reference implementation for the
Leftist tree operations.
insert h x = merge (Node 1 x E E) h
findMin (Node _ x _ _) = x
deleteMin (Node _ _ l r) = merge l r
fromList = foldl insert E

212

CHAPTER 8. BINARY HEAPS

8.3.4

Heap sort by Leftist Heap

With all the basic operations defined, its straightforward to implement heap
sort. We can firstly turn the list into a Leftist heap, then continuously extract
the minimum element from it.
sort(L) = heapSort(build(L))

heapSort(H) =

(8.8)

: H=
{top(H)} heapSort(pop(H)) : otherwise

(8.9)

Because pop is logarithm operation, and it is recursively called n times, this

algorithm takes O(n lg n) time in total. The following Haskell example program
implements heap sort with Leftist tree.
heapSort = hsort fromList where
hsort E = []
hsort h = (findMin h):(hsort $ deleteMin h)

8.3.5

Skew heaps

Leftist heap performs poor in some cases. Figure 8.9 shows one example. The
Leftist tree is built by folding on list {16, 14, 10, 8, 7, 9, 3, 2, 4, 1}.
1

Figure 8.9:
A very
{16, 14, 10, 8, 7, 9, 3, 2, 4, 1}.

unbalanced

Leftist

tree

build

from

list

The binary tree almost turns to a linked-list. The worst case happens when
feed the ordered list to build Leftist tree. Because the tree downgrades to
linked-list, the performance drops from O(lg n) to O(n).
Skew heap (or self-adjusting heap) simplifies Leftist heap realization and can
solve the performance issue[9] [10].

8.3. LEFTIST HEAP AND SKEW HEAP, THE EXPLICIT BINARY HEAPS213
When construct the Leftist heap, we swap the left and right children during
merge if the rank on left side is less than the right side. This comparison-andswap strategy doesnt work when either sub tree has only one child. Because
in such case, the rank of the sub tree is always 1 no matter how big it is. A
Brute-force approach is to swap the left and right children every time when
merge. This idea leads to Skew heap.
Definition of Skew heap
Skew heap is the heap realized with Skew tree. Skew tree is a special binary
tree. The minimum element is stored in root. Every sub tree is also a skew tree.
It neednt keep the rank (or S-value) field. We can reuse the binary tree
definition for Skew heap. The tree is either empty, or in a pre-order form
(k, L, R). Below Haskell code defines Skew heap like this.
data SHeap a = E -- Empty
| Node a (SHeap a) (SHeap a) -- element, left, right

Merge
The merge algorithm tends to be very simple. When merge two non-empty Skew
trees, we compare the roots, and pick the smaller one as the new root, then the
other tree contains the bigger element is merged onto one sub tree, finally, the
tow children are swapped. Denote H1 = (k1 , L1 , R1 ) and H2 = (k2 , L2 , R2 ) if
they are not empty. if k1 < k2 for instance, select k1 as the new root. We
can either merge H2 to L1 , or merge H2 to R1 . Without loss of generality,
lets merge to R1 . And after swapping the two children, the final result is
(k1 , merge(R1 , H2 ), L1 ). Take account of edge cases, the merge algorithm is
defined as the following.

H1
H2
merge(H1 , H2 ) =
(k
,
merge(R
,
H
),
L

1
1
2
1)

(k2 , merge(H1 , R2 ), L2 )

:
:
:
:

H2 =
H1 =
k1 < k2
otherwise

(8.10)

All the rest operations, including insert, top and pop are all realized as same
as the Leftist heap by using merge, except that we neednt the rank any more.
Translating the above algorithm into Haskell yields the following example
program.
merge E h = h
merge h E = h
merge h1@(Node x l r) h2@(Node y l r) =
if x < y then Node x (merge r h2) l
else Node y (merge h1 r) l
insert h x = merge (Node x E E) h
findMin (Node x _ _) = x
deleteMin (Node _ l r) = merge l r

214

CHAPTER 8. BINARY HEAPS

Different from the Leftist heap, if we feed ordered list to Skew heap, it can
build a fairly balanced binary tree as illustrated in figure 8.10.
1

Figure 8.10: Skew tree is still balanced even the input is an ordered list
{1, 2, ..., 10}.

8.4

Splay heap

The Leftist heap and Skew heap show the fact that its quite possible to realize
heap data structure with explicit binary tree. Skew heap gives one method to
solve the tree balance problem. Splay heap on the other hand, use another
method to keep the tree balanced.
The binary trees used in Leftist heap and Skew heap are not Binary Search
tree (BST). If we turn the underground data structure to binary search tree,
the minimum(or maximum) element is not root any more. It takes O(lg n) time
to find the minimum(or maximum) element.
Binary search tree becomes inefficient if it isnt well balanced. Most operations degrade to O(n) in the worst case. Although red-black tree can be
used to realize binary heap, its overkill. Splay tree provides a light weight
implementation with acceptable dynamic balancing result.

8.4.1

Definition

Splay tree uses cache-like approach. It keeps rotating the current access node
close to the top, so that the node can be accessed fast next time. It defines
such kinds of operation as Splay. For the unbalanced binary search tree, after
several splay operations, the tree tends to be more and more balanced. Most
basic operations of Splay tree perform in amortized O(lg n) time. Splay tree

8.4. SPLAY HEAP

215

was invented by Daniel Dominic Sleator and Robert Endre Tarjan in 1985[11]
[12].

Splaying
There are two methods to do splaying. The first one need deal with many
different cases, but can be implemented fairly easy with pattern matching. The
second one has a uniformed form, but the implementation is complex.
Denote the node currently being accessed as X, the parent node as P , and
the grand parent node as G (If there are). There are 3 steps for splaying. Each
step contains 2 symmetric cases. For illustration purpose, only one case is shown
for each step.
Zig-zig step. As shown in figure 8.11, in this case, X and P are children
on the same side, either both on the left or on the right. By rotating 2
times, X becomes the new root.

(a) X and P are either left or right children.

(b) X becomes new root after rotating 2 times.

Figure 8.11: Zig-zig case.

Zig-zag step. As shown in figure 8.12, in this case, X and P are children
on different sides. X is on the left, P is on the right. Or X is on the right,
P is on the left. After rotation, X becomes the new root, P and G are
siblings.
Zig step. As shown in figure 8.13, in this case, P is the root, we rotate the
tree, so that X becomes new root. This is the last step in splay operation.

Although there are 6 different cases, they can be handled in the environments
support pattern matching. Denote the binary tree in form (L, k, R). When

216

CHAPTER 8. BINARY HEAPS

(a) X and P are children on different sides.

(b) X becomes new root. P and G are siblings.

Figure 8.12: Zig-zag case.

(a) P is the root.

(b) Rotate the tree to make X be new root.

Figure 8.13: Zig case.

8.4. SPLAY HEAP

217

access key Y in tree T , the splay operation can be defined as below.

(a, X, (b, P, (c, G, d)))

(((a, G, b), P, c), X, d)

((a, P, b), X, (c, G, d))

((a, G, b), X, (c, P, d))
splay(T, X) =

(a, X, (b, P, c))

((a, P, b), X, c)

:
:
:
:
:
:
:

T = (((a, X, b), P, c), G, d), X = Y

T = (a, G, (b, P, (c, X, d))), X = Y
T = (a, P, (b, X, c), G, d), X = Y
T = (a, G, ((b, X, c), P, d)), X = Y
T = ((a, X, b), P, c), X = Y
T = (a, P, (b, X, c)), X = Y
otherwise
(8.11)
The first two clauses handle the zig-zig cases; the next two clauses handle
the zig-zag cases; the last two clauses handle the zig cases. The tree arent
changed for all other situations.
The following Haskell program implements this splay function.

data STree a = E -- Empty

| Node (STree a) a (STree a) -- left, key, right
-- zig-zig
splay t@(Node
if x == y
splay t@(Node
if x == y
-- zig-zag
splay t@(Node
if x == y
splay t@(Node
if x == y
-- zig
splay t@(Node
splay t@(Node
-- otherwise
splay t _ = t

(Node (Node a x b) p c)
then Node a x (Node b p
a g (Node b p (Node c x
then Node (Node (Node a
(Node a p
then Node
a g (Node
then Node

(Node
(Node
(Node
(Node

b
a
b
a

x
p
x
g

g d) y =
(Node c g d)) else t
d))) y =
g b) p c) x d else t

c)) g d) y
b) x (Node
c) p d)) y
b) x (Node

=
c g d) else t
=
c p d) else t

(Node a x b) p c) y = if x == y then Node a x (Node b p c) else t

a p (Node b x c)) y = if x == y then Node (Node a p b) x c else t

With splay operation defined, every time when insert a new key, we call
the splay function to adjust the tree. If the tree is empty, the result is a leaf;
otherwise we compare this key with the root, if it is less than the root, we
recursively insert it into the left child, and perform splaying after that; else the
key is inserted into the right child.

(, x, ) : T =
splay((insert(L, x), k, R), x) : T = (L, k, R), x < k

splay(L, k, insert(R, x)) : otherwise

(8.12)
The following Haskell program implements this insertion algorithm.

insert(T, x) =

insert E y = Node E y E
insert (Node l x r) y
| x>y
= splay (Node (insert l y) x r) y
| otherwise = splay (Node l x (insert r y)) y

Figure 8.14 shows the result of using this function. It inserts the ordered
elements {1, 2, ..., 10} one by one to the empty tree. This would build a very

218

CHAPTER 8. BINARY HEAPS

poor result which downgrade to linked-list with normal binary search tree. The
splay method creates more balanced result.

Figure 8.14: Splaying helps improving the balance.

Okasaki found a simple rule for Splaying [6]. Whenever we follow two left
branches, or two right branches continuously, we rotate the two nodes.

Based on this rule, splaying can be realized in such a way. When we access
node for a key x (can be during the process of inserting a node, or looking up a
node, or deleting a node), if we traverse two left branches or two right branches,
we partition the tree in two parts L and R, where L contains all nodes smaller
than x, and R contains all the rest. We can then create a new tree (for instance
in insertion), with x as the root, L as the left child, and R being the right child.

8.4. SPLAY HEAP

219

The partition process is recursive, because it will splay its children as well.

(, )

(T, )

((L, k, L0 ), k 0 , A, B)

((L, k, A), (B, k 0 , R0 ))

partition(T, p) =

: T =
: T = (L, k, R) R =
:

T = (L, k, (L0 , k 0 , R0 ))
k < p, k 0 < p
(A, B) = partition(R0 , p)

T = (L, K, (L0 , k 0 , R0 ))
k < p k0
(A, B) = partition(L0 , p)

(, T ) : T = (L, k, R) L =

T = ((L0 , k 0 , R0 ), k, R)

(A, (L0 , k 0 , (R0 , k, R)) :

p k, p k 0

(A, B) = partition(L0 , p)

T = ((L0 , k 0 , R0 ), k, R)

((L0 , k 0 , A), (B, k, R)) :

k0 p k

(A, B) = partition(R0 , p)
(8.13)
Function partition(T, p) takes a tree T , and a pivot p as arguments. The
first clause is edge case. The partition result for empty is a pair of empty left
and right trees. Otherwise, denote the tree as (L, k, R). we need compare the
pivot p and the root k. If k < p, there are two sub-cases. one is trivial case that
R is empty. According to the property of binary search tree, All elements are
less than p, so the result pair is (T, ); For the other case, R = (L0 , k 0 , R0 ), we
need further compare k 0 with the pivot p. If k 0 < p is also true, we recursively
partition R0 with the pivot, all the elements less than p in R0 is held in tree A,
and the rest is in tree B. The result pair can be composed with two trees, one is
((L, k, L0 ), k 0 , A); the other is B. If the key of the right sub tree is not less than
the pivot. we recursively partition L0 with the pivot to give the intermediate
pair (A, B), the final pair trees can be composed with (L, k, A) and (B, k 0 , R0 ).
There is a symmetric cases for p k. They are handled in the last three clauses.
Translating the above algorithm into Haskell yields the following partition
program.
partition E _ = (E, E)
partition t@(Node l x r) y
| x<y=
case r of
E (t, E)
Node l x r
if x < y then
let (small, big) = partition r y in
(Node (Node l x l) x small, big)
else
let (small, big) = partition l y in
(Node l x small, Node big x r)

220

CHAPTER 8. BINARY HEAPS

| otherwise =
case l of
E (E, t)
Node l x r
if y < x then
let (small, big) =
(small, Node l x
else
let (small, big) =
(Node l x small,

partition l y in
(Node r x r))
partition r y in
Node big x r)

Alternatively, insertion can be realized with partition algorithm. When

insert a new element k into the splay heap T , we can first partition the heap
into two trees, L and R. Where L contains all nodes smaller than k, and R
contains the rest. We then construct a new node, with k as the root and L, R
as the children.
insert(T, k) = (L, k, R), (L, R) = partition(T, k)

(8.14)

The corresponding Haskell example program is as the following.

insert t x = Node small x big where (small, big) = partition t x

Top and pop

Since splay tree is just a special binary search tree, the minimum element is
stored in the left most node. We need keep traversing the left child to realize the
top operation. Denote the none empty tree T = (L, k, R), the top(T ) function
can be defined as below.

k : L=
top(T ) =
(8.15)
top(L) : otherwise
This is exactly the min(T ) algorithm for binary search tree.
For pop operation, the algorithm need remove the minimum element from
the tree. Whenever there are two left nodes traversed, the splaying operation
should be performed.

pop(T ) =

R : T = (, k, R)
(R0 , k, R) : T = ((, k 0 , R0 )k, R)

(pop(L0 ), k 0 , (R0 , k, R)) : T = ((L0 , k 0 , R0 ), k, R)

(8.16)

Note that the third clause performs splaying without explicitly call the
partition function. It utilizes the property of binary search tree directly.
Both the top and pop algorithms are bound to O(lg n) time because the
splay tree is balanced.
The following Haskell example programs implement the top and pop operations.
findMin (Node E x _) = x
findMin (Node l x _) = findMin l
deleteMin (Node E x r) = r
deleteMin (Node (Node E x r) x r) = Node r x r
deleteMin (Node (Node l x r) x r) = Node (deleteMin l) x (Node r x r)

8.5. NOTES AND SHORT SUMMARY

221

Merge
Merge is another basic operation for heaps as it is widely used in Graph algorithms. By using the partition algorithm, merge can be realized in O(lg n)
time.
When merging two splay trees, for non-trivial case, we can take the root of
the first tree as the new root, then partition the second tree with this new root
as the pivot. After that we recursively merge the children of the first tree to
the partition result. This algorithm is defined as the following.

T2 : T1 =
(merge(L, A), k, merge(R, B)) : T1 = (L, k, R), (A, B) = partition(T2 , k)
(8.17)
If the first heap is empty, the result is definitely the second heap. Otherwise,
denote the first splay heap as (L, k, R), we partition T2 with k as the pivot to
yield (A, B), where A contains all the elements in T2 which are less than k, and
B holds the rest. We next recursively merge A with L; and merge B with R as
the new children for T1 .
Translating the definition to Haskell gives the following example program.

merge(T1 , T2 ) =

merge E t = t
merge (Node l x r) t = Node (merge l l) x (merge r r)
where (l, r) = partition t x

8.4.2

Heap sort

Since the internal implementation of the Splay heap is completely transparent

to the heap interface, the heap sort algorithm can be reused. It means that the
heap sort algorithm is generic no matter what the underground data structure
is.

8.5

Notes and short summary

In this chapter, we define binary heap more general so that as long as the heap
property is maintained, all binary representation of data structures can be used
to implement binary heap.
This definition doesnt limit to the popular array based binary heap, but
also extends to the explicit binary heaps including Leftist heap, Skew heap and
Splay heap. The array based binary heap is particularly convenient for the
imperative implementation because it intensely uses random index access which
can be mapped to a completely binary tree. Its hard to find directly functional
counterpart in this way.
However, by using explicit binary tree, functional implementation can be
achieved, most of them have O(lg n) worst case performance, and some of them
even reach O(1) amortize time. Okasaki in [6] shows detailed analysis of these
data structures.
In this chapter, only purely functional realization for Leftist heap, Skew heap,
and Splay heap are explained, they can all be realized in imperative approaches.

222

CHAPTER 8. BINARY HEAPS

Its very natural to extend the concept from binary tree to k-ary (k-way)
tree, which leads to other useful heaps such as Binomial heap, Fibonacci heap
and pairing heap. They are introduced in the next chapter.

Exercise 8.2
Realize the imperative Leftist heap, Skew heap, and Splay heap.

Bibliography
[1] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.
[2] Heap (data structure), Wikipedia. http://en.wikipedia.org/wiki/Heap (data structure)
[3] Heapsort, Wikipedia. http://en.wikipedia.org/wiki/Heapsort
[4] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[5] Sorting
algorithms/Heapsort.
Rosetta
http://rosettacode.org/wiki/Sorting algorithms/Heapsort

Code.

[6] Leftist Tree, Wikipedia. http://en.wikipedia.org/wiki/Leftist tree

[7] Bruno R. Preiss. Data Structures and Algorithms with Object-Oriented Design Patterns in Java. http://www.brpreiss.com/books/opus5/index.html
[8] Donald E. Knuth. The Art of Computer Programming. Volume 3: Sorting
and Searching.. Addison-Wesley Professional; 2nd Edition (October 15,
1998). ISBN-13: 978-0201485417. Section 5.2.3 and 6.2.3
[9] Skew heap, Wikipedia. http://en.wikipedia.org/wiki/Skew heap
[10] Sleator, Daniel Dominic; Jarjan, Robert Endre. Self-adjusting heaps
SIAM Journal on Computing 15(1):52-69. doi:10.1137/0215004 ISSN
00975397 (1986)
[11] Splay tree, Wikipedia. http://en.wikipedia.org/wiki/Splay tree
[12] Sleator, Daniel D.; Tarjan, Robert E. (1985), Self-Adjusting Binary Search
Trees, Journal of the ACM 32(3):652 - 686, doi: 10.1145/3828.3835
[13] NIST, binary heap. http://xw2k.nist.gov/dads//HTML/binaryheap.html

223

224

The evolution of selection sort

Chapter 9

From grape to the world

cup, the evolution of
selection sort
9.1

Introduction

We have introduced the hello world sorting algorithm, insertion sort. In this
short chapter, we explain another straightforward sorting method, selection sort.
The basic version of selection sort doesnt perform as good as the divide and
conqueror methods, e.g. quick sort and merge sort. Well use the same approaches in the chapter of insertion sort, to analyze why its slow, and try to
improve it by varies of attempts till reach the best bound of comparison based
sorting, O(N lg N ), by evolving to heap sort.
The idea of selection sort can be illustrated by a real life story. Consider
a kid eating a bunch of grapes. There are two types of children according to
my observation. One is optimistic type, that the kid always eats the biggest
grape he/she can ever find; the other is pessimistic, that he/she always eats the
smallest one.
The first type of kids actually eat the grape in an order that the size decreases
monotonically; while the other eat in a increase order. The kid sorts the grapes
in order of size in fact, and the method used here is selection sort.
Based on this idea, the algorithm of selection sort can be directly described
as the following.
In order to sort a series of elements:
The trivial case, if the series is empty, then we are done, the result is also
empty;
Otherwise, we find the smallest element, and append it to the tail of the
result;
Note that this algorithm sorts the elements in increase order; Its easy to
sort in decrease order by picking the biggest element instead; Well introduce
about passing a comparator as a parameter later on.
225

226CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT

Figure 9.1: Always picking the smallest grape.

This description can be formalized to a equation.

: A=
sort(A) =
{m} sort(A0 ) : otherwise

(9.1)

Where m is the minimum element among collection A, and A0 is all the rest
elements except m:
m = min(A)
A0 = A {m}
We dont limit the data structure of the collection here. Typically, A is an
array in imperative environment, and a list (singly linked-list particularly) in
functional environment, and it can even be other data struture which will be
introduced later.
The algorithm can also be given in imperative manner.
function Sort(A)
X
while A 6= do
x Min(A)
A Del(A, x)
X Append(X, x)
return X
Figure 9.2 depicts the process of this algorithm.

pick
... sorted elements ...

min

... unsorted elements ...

Figure 9.2: The left part is sorted data, continuously pick the minimum element
in the rest and append it to the result.
We just translate the very original idea of eating grapes line by line without
considering any expense of time and space. This realization stores the result in

9.2. FINDING THE MINIMUM

227

X, and when an selected element is appended to X, we delete the same element

from A. This indicates that we can change it to in-place sorting to reuse the
spaces in A.
The idea is to store the minimum element in the first cell in A (we use
term cell if A is an array, and say node if A is a list); then store the second
minimum element in the next cell, then the third cell, ...
One solution to realize this sorting strategy is swapping. When we select
the i-th minimum element, we swap it with the element in the i-th cell:
function Sort(A)
for i 1 to |A| do
m Min(A[i...])
Exchange A[i] m
Denote A = {a1 , a2 , ..., aN }. At any time, when we process the i-th element,
all elements before i, as {a1 , a2 , ..., ai1 } have already been sorted. We locate
the minimum element among the {ai , ai+1 , ..., aN }, and exchange it with ai , so
that the i-th cell contains the right value. The process is repeatedly executed
until we arrived at the last element.
This idea can be illustrated by figure 9.3.

insert

... sorted elements ...

... unsorted elements ...

Figure 9.3: The left part is sorted data, continuously pick the minimum element
in the rest and put it to the right position.

9.2

Finding the minimum

We havent completely realized the selection sort, because we take the operation
of finding the minimum (or the maximum) element as a black box. Its a puzzle
how does a kid locate the biggest or the smallest grape. And this is an interesting
topic for computer algorithms.
The easiest but not so fast way to find the minimum in a collection is to
perform a scan. There are several ways to interpret this scan process. Consider
that we want to pick the biggest grape. We start from any grape, compare
it with another one, and pick the bigger one; then we take a next grape and
compare it with the one we selected so far, pick the bigger one and go on the
take-and-compare process, until there are not any grapes we havent compared.
Its easy to get loss in real practice if we dont mark which grape has been
compared. There are two ways to to solve this problem, which are suitable for
different data-structures respectively.

228CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT

9.2.1

Labeling

Method 1 is to label each grape with a number: {1, 2, ..., N }, and we systematically perform the comparison in the order of this sequence of labels. That we
first compare grape number 1 and grape number 2, pick the bigger one; then we
take grape number 3, and do the comparison, ... We repeat this process until
arrive at grape number N . This is quite suitable for elements stored in an array.
function Min(A)
m A[1]
for i 2 to |A| do
if A[i] < m then
m A[i]
return m
With Min defined, we can complete the basic version of selection sort (or
naive version without any optimization in terms of time and space).
However, this algorithm returns the value of the minimum element instead
of its location (or the label of the grape), which needs a bit tweaking for the
in-place version. Some languages such as ISO C++, support returning the
reference as result, so that the swap can be achieved directly as below.
template<typename T>
T& min(T from, T to) {
T m;
for (m = from++; from != to; ++from)
if (from < m)
m = from;
return m;
}
template<typename T>
void ssort(T xs, int n) {
int i;
for (i = 0; i < n; ++i)
std::swap(xs[i], min(xs+i, xs+n));
}

In environments without reference semantics, the solution is to return the

location of the minimum element instead of the value:
function Min-At(A)
m First-Index(A)
for i m + 1 to |A| do
if A[i] < A[m] then
mi
return m
Note that since we pass A[i...] to Min-At as the argument, we assume the
first element A[i] as the smallest one, and examine all elements A[i + 1], A[i +
2], ... one by one. Function First-Index() is used to retrieve i from the input
parameter.
The following Python example program, for example, completes the basic
in-place selection sort algorithm based on this idea. It explicitly passes the
range information to the function of finding the minimum location.

9.2. FINDING THE MINIMUM

229

def ssort(xs):
n = len(xs)
for i in range(n):
m = min_at(xs, i, n)
(xs[i], xs[m]) = (xs[m], xs[i])
return xs
def min_at(xs, i, n):
m = i;
for j in range(i+1, n):
if xs[j] < xs[m]:
m=j
return m

9.2.2

Grouping

Another method is to group all grapes in two parts: the group we have examined,
and the rest we havent. We denote these two groups as A and B; All the
elements (grapes) as L. At the beginning, we havent examine any grapes at
all, thus A is empty (), and B contains all grapes. We can select arbitrary two
grapes from B, compare them, and put the loser (the smaller one for example) to
A. After that, we repeat this process by continuously picking arbitrary grapes
from B, and compare with the winner of the previous time until B becomes
empty. At this time being, the final winner is the minimum element. And A
turns to be L{min(L)}, which can be used for the next time minimum finding.
There is an invariant of this method, that at any time, we have L = A
{m} B, where m is the winner so far we hold.
This approach doesnt need the collection of grapes being indexed (as being
labeled in method 1). Its suitable for any traversable data structures, including
linked-list etc. Suppose b1 is an arbitrary element in B if B isnt empty, and B 0
is the rest of elements with b1 being removed, this method can be formalized as
the below auxiliary function.

(m, A) : B =

min0 (A {m}, b1 , B 0 ) : b1 < m

min0 (A, m, B) =
(9.2)

min0 (A {b1 }, m, B 0 ) : otherwise

In order to pick the minimum element, we call this auxiliary function by
passing an empty A, and use an arbitrary element (for instance, the first one)
to initialize m:
extractM in(L) = min0 (, l1 , L0 )

(9.3)

Where L0 is all elements in L except for the first one l1 . The algorithm
extractM in) doesnt not only find the minimum element, but also returns the
updated collection which doesnt contain this minimum. Summarize this minimum extracting algorithm up to the basic selection sort definition, we can create
a complete functional sorting program, for example as this Haskell code snippet.
sort [] = []
sort xs = x : sort xs where
(x, xs) = extractMin xs

230CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT

extractMin (x:xs) = min [] x xs where

min ys m [] = (m, ys)
min ys m (x:xs) = if m < x then min (x:ys) m xs else min (m:ys) x xs

The first line handles the trivial edge case that the sorting result for empty
list is obvious empty; The second clause ensures that, there is at least one
element, thats why the extractMin function neednt other pattern-matching.
One may think the second clause of min function should be written like
below:
min ys m (x:xs) = if m < x then min ys ++ [x] m xs
else min ys ++ [m] x xs

Or it will produce the updated list in reverse order. Actually, its necessary
to use cons instead of appending here. This is because appending is linear
operation which is proportion to the length of part A, while cons is constant
O(1) time operation. In fact, we neednt keep the relative order of the list to
be sorted, as it will be re-arranged anyway during sorting.
Its quite possible to keep the relative order during sorting, while ensure the
performance of finding the minimum element not degrade to quadratic. The
following equation defines a solution.

(l1 , ) : |L| = 1
(l1 , L0 ) : l1 < m, (m, L00 ) = extractM in(L0 )

(m, l1 L00 ) : otherwise

(9.4)
If L is a singleton, the minimum is the only element it contains. Otherwise,
denote l1 as the first element in L, and L0 contains the rest elements except for
l1 , that L0 = {l2 , l3 , ...}. The algorithm recursively finding the minimum element
in L0 , which yields the intermediate result as (m, L00 ), that m is the minimum
element in L0 , and L00 contains all rest elements except for m. Comparing l1
with m, we can determine which of them is the final minimum result.
The following Haskell program implements this version of selection sort.
extractM in(L) =

sort [] = []
sort xs = x : sort xs where
(x, xs) = extractMin xs
extractMin [x] = (x, [])
extractMin (x:xs) = if x < m then (x, xs) else (m, x:xs) where
(m, xs) = extractMin xs

Note that only cons operation is used, we neednt appending at all because
the algorithm actually examines the list from right to left. However, its not
free, as this program need book-keeping the context (via call stack typically).
The relative order is ensured by the nature of recursion. Please refer to the
appendix about tail recursion call for detailed discussion.

9.2.3

performance of the basic selection sorting

Both the labeling method, and the grouping method need examine all the elements to pick the minimum in every round; and we totally pick up the minimum

9.3. MINOR IMPROVEMENT

231

element N times. Thus the performance is around N + (N 1) + (N 2) + ... + 1

which is N (N2+1) . Selection sort is a quadratic algorithm bound to O(N 2 ) time.
Compare to the insertion sort, which we introduced previously, selection sort
performs same in its best case, worst case and average case. While insertion
sort performs well in best case (that the list has been reverse ordered, and it is
stored in linked-list) as O(N ), and the worst performance is O(N 2 ).
In the next sections, well examine, why selection sort performs poor, and
try to improve it step by step.

Exercise 9.1
Implement the basic imperative selection sort algorithm (the none in-place
version) in your favorite programming language. Compare it with the inplace version, and analyze the time and space effectiveness.

9.3
9.3.1

Minor Improvement
Parameterize the comparator

Before any improvement in terms of performance, lets make the selection sort
algorithm general enough to handle different sorting criteria.
Weve seen two opposite examples so far, that one may need sort the elements
in ascending order or descending order. For the former case, we need repeatedly
finding the minimum, while for the later, we need find the maximum instead.
They are just two special cases. In real world practice, one may want to sort
things in varies criteria, e.g. in terms of size, weight, age, ...
One solution to handle them all is to passing the criteria as a compare
function to the basic selection sort algorithms. For example:

sort(c, L) =

: L=
m sort(c, L00 ) : otherwise, (m, L00 ) = extract(c, L0 )
(9.5)

And the algorithm extract(c, L) is defined as below.

extract(c, L) =

(l1 , ) : |L| = 1
(l1 , L0 ) : c(l1 , m), (m, L00 ) = extract(c, L0 )

(m, {l1 } L00 ) : c(l1 , m)

(9.6)
Where c is a comparator function, it takes two elements, compare them and
returns the result of which one is preceding of the other. Passing less than
operator (<) turns this algorithm to be the version we introduced in previous
section.
Some environments require to pass the total ordering comparator, which
returns result among less than, equal, and greater than. We neednt such
strong condition here, that c only tests if less than is satisfied. However, as the
minimum requirement, the comparator should meet the strict weak ordering as
following [16]:
Irreflexivity, for all x, its not the case that x < x;

232CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
Asymmetric, For all x and y, if x < y, then its not the case y < x;
Transitivity, For all x, y, and z, if x < y, and y < z, then x < z;
The following Scheme/Lisp program translates this generic selection sorting
algorithm. The reason why we choose Scheme/Lisp here is because the lexical
scope can simplify the needs to pass the less than comparator for every function
calls.
(define (sel-sort-by ltp? lst)
(define (ssort lst)
(if (null? lst)
lst
(let ((p (extract-min lst)))
(cons (car p) (ssort (cdr p))))))
(define (extract-min lst)
(if (null? (cdr lst))
lst
(let ((p (extract-min (cdr lst))))
(if (ltp? (car lst) (car p))
lst
(cons (car p) (cons (car lst) (cdr p)))))))
(ssort lst))

Note that, both ssort and extract-min are inner functions, so that the
less than comparator ltp? is available to them. Passing < to this function
yields the normal sorting in ascending order:
(sel-sort-by < (3 1 2 4 5 10 9))
;Value 16: (1 2 3 4 5 9 10)

Its possible to pass varies of comparator to imperative selection sort as well.

This is left as an exercise to the reader.
For the sake of brevity, we only consider sorting elements in ascending order
in the rest of this chapter. And well not pass comparator as a parameter unless
its necessary.

9.3.2

Trivial fine tune

The basic in-place imperative selection sorting algorithm iterates all elements,
and picking the minimum by traversing as well. It can be written in a compact
way, that we inline the minimum finding part as an inner loop.
procedure Sort(A)
for i 1 to |A| do
mi
for j i + 1 to |A| do
if A[i] < A[m] then
mi
Exchange A[i] A[m]
Observe that, when we are sorting N elements, after the first N 1 minimum
ones are selected, the left only one, is definitely the N -th big element, so that
we need NOT find the minimum if there is only one element in the list. This
indicates that the outer loop can iterate to N 1 instead of N .

9.3. MINOR IMPROVEMENT

233

Another place we can fine tune, is that we neednt swap the elements if the
i-th minimum one is just A[i]. The algorithm can be modified accordingly as
below:
procedure Sort(A)
for i 1 to |A| 1 do
mi
for j i + 1 to |A| do
if A[i] < A[m] then
mi
if m 6= i then
Exchange A[i] A[m]
Definitely, these modifications wont affects the performance in terms of bigO.

9.3.3

Cock-tail sort

Knuth gave an alternative realization of selection sort in [1]. Instead of selecting

the minimum each time, we can select the maximum element, and put it to the
last position. This method can be illustrated by the following algorithm.
procedure Sort(A)
for i |A| down-to 2 do
mi
for j 1 to i 1 do
if A[m] < A[i] then
mi
Exchange A[i] A[m]
As shown in figure 13.1, at any time, the elements on right most side are
sorted. The algorithm scans all unsorted ones, and locate the maximum. Then,
put it to the tail of the unsorted range by swapping.

swap
...

max

...

... sorted elements ...

Figure 9.4: Select the maximum every time and put it to the end.
This version reveals the fact that, selecting the maximum element can sort
the element in ascending order as well. Whats more, we can find both the
minimum and the maximum elements in one pass of traversing, putting the
minimum at the first location, while putting the maximum at the last position.
This approach can speed up the sorting slightly (halve the times of the outer
loop).
procedure Sort(A)
for i 1 to b |A|
2 c do
min i
max |A| + 1 i

234CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
if A[max] < A[min] then
Exchange A[min] A[max]
for j i + 1 to |A| i do
if A[j] < A[min] then
min j
if A[max] < A[j] then
max j
Exchange A[i] A[min]
Exchange A[|A| + 1 i] A[max]
This algorithm can be illustrated as in figure 9.5, at any time, the left most
and right most parts contain sorted elements so far. That the smaller sorted ones
are on the left, while the bigger sorted ones are on the right. The algorithm scans
the unsorted ranges, located both the minimum and the maximum positions,
then put them to the head and the tail position of the unsorted ranges by
swapping.
swap
... sorted small ones ...

x ...

max

...

min

...

... sorted big ones ...

Figure 9.5: Select both the minimum and maximum in one pass, and put them
to the proper positions.
Note that its necessary to swap the left most and right most elements before
the inner loop if they are not in correct order. This is because we scan the range
excluding these two elements. Another method is to initialize the first element of
the unsorted range as both the maximum and minimum before the inner loop.
However, since we need two swapping operations after the scan, its possible
that the first swapping moves the maximum or the minimum from the position
we just found, which leads the second swapping malfunctioned. How to solve
this problem is left as exercise to the reader.
The following Python example program implements this cock-tail sort algorithm.
def cocktail_sort(xs):
n = len(xs)
for i in range(n / 2):
(mi, ma) = (i, n - 1 -i)
if xs[ma] < xs[mi]:
(xs[mi], xs[ma]) = (xs[ma], xs[mi])
for j in range(i+1, n - 1 - i):
if xs[j] < xs[mi]:
mi = j
if xs[ma] < xs[j]:
ma = j
(xs[i], xs[mi]) = (xs[mi], xs[i])
(xs[n - 1 - i], xs[ma]) = (xs[ma], xs[n - 1 - i])
return xs

9.3. MINOR IMPROVEMENT

235

Its possible to realize cock-tail sort in functional approach as well. An

intuitive recursive description can be given like this:
Trivial edge case: If the list is empty, or there is only one element in the
list, the sorted result is obviously the origin list;
Otherwise, we select the minimum and the maximum, put them in the
head and tail positions, then recursively sort the rest elements.
This algorithm description can be formalized by the following equation.

L : |L| 1
sort(L) =
(9.7)
{lmin } sort(L00 ) {lmax } : otherwise
Where the minimum and the maximum are extracted from L by a function
select(L).
(lmin , L00 , lmax ) = select(L)
Note that, the minimum is actually linked to the front of the recursive sort
result. Its semantic is a constant O(1) time cons (refer to the appendix of this
book for detail). While the maximum is appending to the tail. This is typically
a linear O(N ) time expensive operation. Well optimize it later.
Function select(L) scans the whole list to find both the minimum and the
maximum. It can be defined as below:

(min(l1 , l2 ), max(l1 , l2 )) : L = {l1 , l2 }

(l1 , {lmin } L00 , lmax ) : l1 < lmin

select(L) =
(9.8)
(lmin , {lmax } L00 , l1 ) : lmax < l1

(lmin , {l1 } L00 , lmax ) : otherwise

Where (lmin , L00 , lmax ) = select(L0 ) and L0 is the rest of the list except for
the first element l1 . If there are only two elements in the list, we pick the
smaller as the minimum, and the bigger as the maximum. After extract them,
the list becomes empty. This is the trivial edge case; Otherwise, we take the first
element l1 out, then recursively perform selection on the rest of the list. After
that, we compare if l1 is less then the minimum or greater than the maximum
candidates, so that we can finalize the result.
Note that for all the cases, there is no appending operation to form the result.
However, since selection must scan all the element to determine the minimum
and the maximum, it is bound to O(N ) linear time.
The complete example Haskell program is given as the following.
csort [] = []
csort [x] = [x]
csort xs = mi : csort xs ++ [ma] where
(mi, xs, ma) = extractMinMax xs
extractMinMax [x, y] = (min x y, [], max x y)
extractMinMax (x:xs) | x < mi = (x, mi:xs, ma)
| ma < x = (mi, ma:xs, x)
| otherwise = (mi, x:xs, ma)
where (mi, xs, ma) = extractMinMax xs

236CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
We mentioned that the appending operation is expensive in this intuitive
version. It can be improved. This can be achieved in two steps. The first step is
to convert the cock-tail sort into tail-recursive call. Denote the sorted small ones
as A, and sorted big ones as B in figure 9.5. We use A and B as accumulators.
The new cock-tail sort is defined as the following.

: L = |L| = 1
: otherwise
(9.9)
Where lmin , lmax and L00 are defined as same as before. And we start sorting
by passing empty A and B: sort(L) = sort0 (, L, ).
Besides the edge case, observing that the appending operation only happens
on A {lmin }; while lmax is only linked to the head of B. This appending
occurs in every recursive call. To eliminate it, we can store A in reverse order

as A , so that lmax can be cons to the head instead of appending. Denote

cons(x, L) = {x} L and append(L, x) = L {x}, we have the below equation.
sort0 (A, L, B) =

ALB
sort0 (A {lmin }, L00 , {lmax } B)

append(L, x)

= reverse(cons(x, reverse(L)))
(9.10)

= reverse(cons(x, L ))

Finally, we perform a reverse to turn A back to A. Based on this idea, the

algorithm can be improved one more step as the following.

sort0 (A, L, B) =

reverse(A) B : L =
reverse({l1 } A) B : |L| = 1

sort0 ({lmin } A, L00 , {lmax } B) :

(9.11)

This algorithm can be implemented by Haskell as below.

csort xs = cocktail [] xs [] where
cocktail as [] bs = reverse as ++ bs
cocktail as [x] bs = reverse (x:as) ++ bs
cocktail as xs bs = let (mi, xs, ma) = extractMinMax xs
in cocktail (mi:as) xs (ma:bs)

Exercise 9.2
Realize the imperative basic selection sort algorithm, which can take a
comparator as a parameter. Please try both dynamic typed language and
static typed language. How to annotate the type of the comparator as
general as possible in a static typed language?
Implement Knuths version of selection sort in your favorite programming
language.
An alternative to realize cock-tail sort is to assume the i-th element both
the minimum and the maximum, after the inner loop, the minimum and
maximum are found, then we can swap the the minimum to the i-th
position, and the maximum to position |A|+1i. Implement this solution
in your favorite imperative language. Please note that there are several
special edge cases should be handled correctly:

9.4. MAJOR IMPROVEMENT

237

A = {max, min, ...};

A = {..., max, min};
A = {max, ..., min}.
Please dont refer to the example source code along with this chapter
before you try to solve this problem.

9.4

Major improvement

Although cock-tail sort halves the numbers of loop, the performance is still
bound to quadratic time. It means that, the method we developed so far handles
big data poorly compare to other divide and conquer sorting solutions.
To improve selection based sort essentially, we must analyze where is the
bottle-neck. In order to sort the elements by comparison, we must examine all
the elements for ordering. Thus the outer loop of selection sort is necessary.
However, must it scan all the elements every time to select the minimum? Note
that when we pick the smallest one at the first time, we actually traverse the
whole collection, so that we know which ones are relative big, and which ones
are relative small partially.
The problem is that, when we select the further minimum elements, instead
of re-using the ordering information we obtained previously, we drop them all,
and blindly start a new traverse.
So the key point to improve selection based sort is to re-use the previous
result. There are several approaches, well adopt an intuitive idea inspired by
football match in this chapter.

9.4.1

Tournament knock out

The football world cup is held every four years. There are 32 teams from
different continent play the final games. Before 1982, there were 16 teams
compete for the tournament finals[4].
For simplification purpose, lets go back to 1978 and imagine a way to determine the champion: In the first round, the teams are grouped into 8 pairs
to play the game; After that, there will be 8 winner, and 8 teams will be out.
Then in the second round, these 8 teams are grouped into 4 pairs. This time
there will be 4 winners after the second round of games; Then the top 4 teams
are divided into 2 pairs, so that there will be only two teams left for the final
game.
The champion is determined after the total 4 rounds of games. And there
are actually 8 + 4 + 2 + 1 = 16 games. Now we have the world cup champion,
however, the world cup game wont finish at this stage, we need to determine
which is the silver medal team.
Readers may argue that isnt the team beaten by the champion at the final game the second best? This is true according to the real world cup rule.
However, it isnt fair enough in some sense.
We often heard about the so called group of death, Lets suppose that
Brazil team is grouped with Deutch team at the very beginning. Although both
teams are quite strong, one of them must be knocked out. Its quite possible

238CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
that even the team loss that game can beat all the other teams except for the
champion. Figure 9.6 illustrates such case.
16

Figure 9.6: The element 15 is knocked out in the first round.

Imagine that every team has a number. The bigger the number, the stronger
the team. Suppose that the stronger team always beats the team with smaller
number, although this is not true in real world. But this simplification is fair
enough for us to develop the tournament knock out solution. This maximum
number which represents the champion is 16. Definitely, team with number 14
isnt the second best according to our rules. It should be 15, which is knocked
out at the first round of comparison.
The key question here is to find an effective way to locate the second maximum number in this tournament tree. After that, what we need is to apply
the same method to select the third, the fourth, ..., to accomplish the selection
based sort.
One idea is to assign the champion a very small number (for instance, ),
so that it wont be selected next time, and the second best one, becomes the
new champion. However, suppose there are 2m teams for some natural number
m, it still takes 2m1 + 2m2 + ... + 2 + 1 = 2m times of comparison to determine
the new champion. Which is as slow as the first time.
Actually, we neednt perform a bottom-up comparison at all since the tournament tree stores plenty of ordering information. Observe that, the second
best team must be beaten by the champion at sometime, or it will be the final
winner. So we can track the path from the root of the tournament tree to the
leaf of the champion, examine all the teams along with this path to find the
second best team.
In figure 9.6, this path is marked in gray color, the elements to be examined
are {14, 13, 7, 15}. Based on this idea, we refine the algorithm like below.
1. Build a tournament tree from the elements to be sorted, so that the champion (the maximum) becomes the root;
2. Extract the root from the tree, perform a top-down pass and replace the
maximum with ;
3. Perform a bottom-up back-track along the path, determine the new champion and make it as the new root;
4. Repeat step 2 until all elements have been extracted.
Figure 9.7, 9.8, and 9.9 show the steps of applying this strategy.

9.4. MAJOR IMPROVEMENT

239

-INF

Figure 9.7: Extract 16, replace it with , 15 sifts up to root.

-INF

Figure 9.8: Extract 15, replace it with , 14 sifts up to root.

-INF

Figure 9.9: Extract 14, replace it with , 13 sifts up to root.

-INF

240CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
We can reuse the binary tree definition given in the first chapter of this
book to represent tournament tree. In order to back-track from leaf to the root,
every node should hold a reference to its parent (concept of pointer in some
environment such as ANSI C):
struct Node {
Key key;
struct Node left, right, parent;
};

To build a tournament tree from a list of elements (suppose the number of

elements are 2m for some m), we can first wrap each element as a leaf, so that
we obtain a list of binary trees. We take every two trees from this list, compare
their keys, and form a new binary tree with the bigger key as the root; the two
trees are set as the left and right children of this new binary tree. Repeat this
operation to build a new list of trees. The height of each tree is increased by 1.
Note that the size of the tree list halves after such a pass, so that we can keep
reducing the list until there is only one tree left. And this tree is the finally
built tournament tree.
function Build-Tree(A)
T
for each x A do
t Create-Node
Key(t) x
Append(T, t)
while |T | > 1 do
T0
for every t1 , t2 T do
t Create-Node
Key(t) Max(Key(t1 ), Key(t2 ))
Left(t) t1
Right(t) t2
Parent(t1 ) t
Parent(t2 ) t
Append(T 0 , t)
T T0
return T [1]
Suppose the length of the list A is N , this algorithm firstly traverses the list
to build tree, which is linear to N time. Then it repeatedly compares pairs,
which loops proportion to N + N2 + N4 + ... + 2 = 2N . So the total performance
is bound to O(N ) time.
The following ANSI C program implements this tournament tree building
algorithm.
struct Node build(const Key xs, int n) {
int i;
struct Node t, ts = (struct Node) malloc(sizeof(struct Node) n);
for (i = 0; i < n; ++i)
ts[i] = leaf(xs[i]);
for (; n > 1; n /= 2)
for (i = 0; i < n; i += 2)

9.4. MAJOR IMPROVEMENT

241

ts[i/2] = branch(max(ts[i]key, ts[i+1]key), ts[i], ts[i+1]);

t = ts[0];
free(ts);
return t;
}

The type of key can be defined somewhere, for example:

typedef int Key;

Function leaf(x) creats a leaf node, with value x as key, and sets all its
fields, left, right and parent to NIL. While function branch(key, left, right)
creates a branch node, and links the new created node as parent of its two
children if they are not empty. For the sake of brevity, we skip the detail of
them. They are left as exercise to the reader, and the complete program can be
downloaded along with this book.
Some programming environments, such as Python provides tool to iterate
every two elements at a time, for example:
for x, y in zip([iter(ts)]2):

We skip such language specific feature, readers can refer to the Python example program along with this book for details.
When the maximum element is extracted from the tournament tree, we
replace it with , and repeatedly replace all these values from the root to the
leaf. Next, we back-track to root through the parent field, and determine the
new maximum element.
function Extract-Max(T )
m Key(T )
Key(T )
while Leaf?(T ) do
. The top down pass
if Key(Left(T )) = m then
T Left(T )
else
T Right(T )
Key(T )
while Parent(T ) 6= do
. The bottom up pass
T Parent(T )
Key(T ) Max(Key(Left(T )), Key(Right(T )))
return m
This algorithm returns the extracted maximum element, and modifies the
tournament tree in-place. Because we cant represent in real program by
limited length of word, one approach is to define a relative negative big number,
which is less than all the elements in the tournament tree, for example, suppose
all the elements are greater than -65535, we can define negative infinity as below:
#define N_INF -65535

We can implements this algorithm as the following ANSI C example program.

Key pop(struct Node t) {
Key x = tkey;
tkey = N_INF;
while (!isleaf(t)) {

242CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
t = tleftkey == x ? tleft : tright;
tkey = N_INF;
}
while (tparent) {
t = tparent;
tkey = max(tleftkey, trightkey);
}
return x;
}

The behavior of Extract-Max is quite similar to the pop operation for

some data structures, such as queue, and heap, thus we name it as pop in this
code snippet.
Algorithm Extract-Max process the tree in tow passes, one is top-down,
then a bottom-up along the path that the champion team wins the world cup.
Because the tournament tree is well balanced, the length of this path, which
is the height of the tree, is bound to O(lg N ), where N is the number of the
elements to be sorted (which are equal to the number of leaves). Thus the
performance of this algorithm is O(lg N ).
Its possible to realize the tournament knock out sort now. We build a
tournament tree from the elements to be sorted, then continuously extract the
maximum. If we want to sort in monotonically increase order, we put the first
extracted one to the right most, then insert the further extracted elements one
by one to left; Otherwise if we want to sort in decrease order, we can just append
the extracted elements to the result. Below is the algorithm sorts elements in
ascending order.
procedure Sort(A)
T Build-Tree(A)
for i |A| down to 1 do
A[i] Extract-Max(T )
Translating it to ANSI C example program is straightforward.
void tsort(Key xs, int n) {
struct Node t = build(xs, n);
while(n)
xs[--n] = pop(t);
release(t);
}

This algorithm firstly takes O(N ) time to build the tournament tree, then
performs N pops to select the maximum elements so far left in the tree. Since
each pop operation is bound to O(lg N ), thus the total performance of tournament knock out sorting is O(N lg N ).
Refine the tournament knock out
Its possible to design the tournament knock out algorithm in purely functional
approach. And well see that the two passes (first top-down replace the champion with , then bottom-up determine the new champion) in pop operation
can be combined in recursive manner, so that we neednt the parent field any
more. We can re-use the functional binary tree definition as the following example Haskell code.

9.4. MAJOR IMPROVEMENT

243

data Tr a = Empty | Br (Tr a) a (Tr a)

Thus a binary tree is either empty or a branch node contains a key, a left
sub tree and a right sub tree. Both children are again binary trees.
Weve use hard coded big negative number to represents . However, this
solution is ad-hoc, and it forces all elements to be sorted are greater than this
pre-defined magic number. Some programming environments support algebraic
type, so that we can define negative infinity explicitly. For instance, the below
Haskell program setups the concept of infinity 1 .
data Infinite a = NegInf | Only a | Inf deriving (Eq, Ord)

From now on, we switch back to use the min() function to determine the
winner, so that the tournament selects the minimum instead of the maximum
as the champion.
Denote function key(T ) returns the key of the tree rooted at T . Function
wrap(x) wraps the element x into a leaf node. Function tree(l, k, r) creates a
branch node, with k as the key, l and r as the two children respectively.
The knock out process, can be represented as comparing two trees, picking
the smaller key as the new key, and setting these two trees as children:
branch(T1 , T2 ) = tree(T1 , min(key(T1 ), key(T2 )), T2 )

(9.12)

This can be implemented in Haskell word by word:

branch t1 t2 = Br t1 (min (key t1) (key t2)) t2

There is limitation in our tournament sorting algorithm so far. It only

accepts collection of elements with size of 2m , or we cant build a complete
binary tree. This can be actually solved in the tree building process. Remind
that we pick two trees every time, compare and pick the winner. This is perfect
if there are always even number of trees. Considering a case in football match,
that one team is absent for some reason (sever flight delay or whatever), so that
there left one team without a challenger. One option is to make this team the
winner, so that it will attend the further games. Actually, we can use the similar
approach.
To build the tournament tree from a list of elements, we wrap every element
into a leaf, then start the building process.
build(L) = build0 ({wrap(x)|x L})

(9.13)

The build0 (T) function terminates when there is only one tree left in T, which
is the champion. This is the trivial edge case. Otherwise, it groups every two
trees in a pair to determine the winners. When there are odd numbers of trees,
it just makes the last tree as the winner to attend the next level of tournament
and recursively repeats the building process.

T : |T| 1
build0 (T) =
(9.14)
build0 (pair(T)) : otherwise
1 The order of the definition of NegInf, regular number, and Inf is significant if we want
to derive the default, correct comparing behavior of Ord. Anyway, its possible to specify the
detailed order by make it as an instance of Ord. However, this is Language specific feature
which is out of the scope of this book. Please refer to other textbook about Haskell.

244CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
Note that this algorithm actually handles another special cases, that the list
to be sort is empty. The result is obviously empty.
Denote T = {T1 , T2 , ...} if there are at least two trees, and T0 represents the
left trees by removing the first two. Function pair(T) is defined as the following.

pair(T) =

{branch(T1 , T2 )} pair(T0 ) : |T| 2

T : otherwise

(9.15)

The complete tournament tree building algorithm can be implemented as

the below example Haskell program.
fromList :: (Ord a) [a] Tr (Infinite a)
fromList = build (map wrap) where
build [] = Empty
build [t] = t
build ts = build $ pair ts
pair (t1:t2:ts) = (branch t1 t2):pair ts
pair ts = ts

When extracting the champion (the minimum) from the tournament tree,
we need examine either the left child sub-tree or the right one has the same key,
and recursively extract on that tree until arrive at the leaf node. Denote the
left sub-tree of T as L, right sub-tree as R, and K as its key. We can define this
popping algorithm as the following.

tree(, , ) : L = R =
tree(L0 , min(key(L0 ), key(R)), R) : K = key(L), L0 = pop(L)

tree(L, min(key(L), key(R0 )), R0 ) : K = key(R), R0 = pop(R)

(9.16)
Its straightforward to translate this algorithm into example Haskell code.

pop(T ) =

pop (Br Empty _ Empty) = Br Empty Inf Empty

pop (Br l k r) | k == key l = let l = pop l in Br l (min (key l) (key r)) r
| k == key r = let r = pop r in Br l (min (key l) (key r)) r

Note that this algorithm only removes the current champion without returning it. So its necessary to define a function to get the champion at the root
node.
top(T ) = key(T )

(9.17)

With these functions defined, tournament knock out sorting can be formalized by using them.
sort(L) = sort0 (build(L))

(9.18)

Where sort (T ) continuously pops the minimum element to form a result list

sort (T ) =

:
{top(T )} sort0 (pop(T )) :

T = key(T ) =
otherwise

(9.19)

The rest of the Haskell code is given below to complete the implementation.

9.4. MAJOR IMPROVEMENT

245

top = only key

tsort :: (Ord a) [a] [a]
tsort = sort fromList where
sort Empty = []
sort (Br _ Inf _) = []
sort t = (top t) : (sort $ pop t)

And the auxiliary function only, key, wrap accomplished with explicit infinity support are list as the following.
only (Only x) = x
key (Br _ k _ ) = k
wrap x = Br Empty (Only x) Empty

Exercise 9.3
Implement the helper function leaf(), branch, max() lsleaf(), and
release() to complete the imperative tournament tree program.
Implement the imperative tournament tree in a programming language
support GC (garbage collection).
Why can our tournament tree knock out sort algorithm handle duplicated
elements (elements with same value)? We say a sorting algorithm stable, if
it keeps the original order of elements with same value. Is the tournament
tree knock out sorting stable?
Design an imperative tournament tree knock out sort algorithm, which
satisfies the following:
Can handle arbitrary number of elements;
Without using hard coded negative infinity, so that it can take elements with any value.
Compare the tournament tree knock out sort algorithm and binary tree
sort algorithm, analyze efficiency both in time and space.
Compare the heap sort algorithm and binary tree sort algorithm, and do
same analysis for them.

9.4.2

Final improvement by using heap sort

We manage improving the performance of selection based sorting to O(N lg N )

by using tournament knock out. This is the limit of comparison based sort
according to [1]. However, there are still rooms for improvement. After sorting,
there lefts a complete binary tree with all leaves and branches hold useless
infinite values. This isnt space efficient at all. Can we release the nodes when
popping?
Another observation is that if there are N elements to be sorted, we actually
allocate about 2N tree nodes. N for leaves and N for branches. Is there any
better way to halve the space usage?

246CHAPTER 9. FROM GRAPE TO THE WORLD CUP, THE EVOLUTION OF SELECTION SORT
The final sorting structure described in equation 9.19 can be easily uniformed
to a more general one if we treat the case that the tree is empty if its root holds
infinity as key:

: T =
sort0 (T ) =
(9.20)
{top(T )} sort0 (pop(T )) : otherwise
This is exactly as same as the one of heap sort we gave in previous chapter.
Heap always keeps the minimum (or the maximum) on the top, and provides
fast pop operation. The binary heap by implicit array encodes the tree structure
in array index, so there arent any extra spaces allocated except for the N array
cells. The functional heaps, such as leftist heap and splay heap allocate N nodes
as well. Well introduce more heaps in next chapter which perform well in many
aspects.

9.5

Short summary

In this chapter, we present the evolution process of selection based sort. selection
sort is easy and commonly used as example to teach students about embedded
looping. It has simple and straightforward structure, but the performance is
quadratic. In this chapter, we do see that there exists ways to improve it not
only by some fine tuning, but also fundamentally change the data structure,
which leads to tournament knock out and heap sort.

247

248

Binomial heap, Fibonacci heap, and pairing heap

Chapter 10

Binomial heap, Fibonacci

heap, and pairing heap
10.1

Introduction

In previous chapter, we mentioned that heaps can be generalized and implemented with varies of data structures. However, only binary heaps are focused
so far no matter by explicit binary trees or implicit array.
Its quite natural to extend the binary tree to K-ary [1] tree. In this chapter,
we first show Binomial heaps which is actually consist of forest of K-ary trees.
Binomial heaps gain the performance for all operations to O(lg N ), as well as
keeping the finding minimum element to O(1) time.
If we delay some operations in Binomial heaps by using lazy strategy, it
turns to be Fibonacci heap.
All binary heaps we have shown perform no less than O(lg N ) time for merging, well show its possible to improve it to O(1) with Fibonacci heap, which
is quite helpful to graph algorithms. Actually, Fibonacci heap achieves almost
all operations to good amortized time bound as O(1), and left the heap pop to
O(lg N ).
Finally, well introduce about the pairing heaps. It has the best performance
in practice although the proof of it is still a conjecture for the time being.

10.2

Binomial Heaps

10.2.1

Definition

Binomial heap is more complex than most of the binary heaps. However, it has
excellent merge performance which bound to O(lg N ) time. A binomial heap is
consist of a list of binomial trees.
Binomial tree
In order to explain why the name of the tree is binomial, lets review the
famous Pascals triangle (Also know as the Jia Xians triangle to memorize the
Chinese methematican Jia Xian (1010-1070).) [4].
249

250CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

1
1 1
1 2 1
1 3 3 1
1 4 6 4 1
...
In each row, the numbers are all binomial coefficients. There are many
ways to gain a series of binomial coefficient numbers. One of them is by using
recursive composition. Binomial trees, as well, can be defined in this way as the
following.
A binomial tree of rank 0 has only a node as the root;
A binomial tree of rank N is consist of two rank N 1 binomial trees,
Among these 2 sub trees, the one has the bigger root element is linked as
the leftmost child of the other.
We denote a binomial tree of rank 0 as B0 , and the binomial tree of rank n
as Bn .
Figure 10.1 shows a B0 tree and how to link 2 Bn1 trees to a Bn tree.

(a) A B0 tree.

rank=n-1

...

(b) Linking 2 Bn1 trees yields a Bn tree.

Figure 10.1: Recursive definition of binomial trees

With this recursive definition, it easy to draw the form of binomial trees of
rank 0, 1, 2, ..., as shown in figure 10.2
Observing the binomial trees reveals some interesting properties. For each
rank N binomial tree, if counting the number of nodes in each row, it can be
found that it is the binomial number.
For instance for rank 4 binomial tree, there is 1 node as the root; and in the
second level next to root, there are 4 nodes; and in 3rd level, there are 6 nodes;

10.2. BINOMIAL HEAPS

251

(a) B0 tree;

(b) B1 tree;

(d) B3 tree;
4

...
(e) B4 tree;

Figure 10.2: Forms of binomial trees with rank = 0, 1, 2, 3, 4, ...

252CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

and in 4-th level, there are 4 nodes; and the 5-th level, there is 1 node. They
are exactly 1, 4, 6, 4, 1 which is the 5th row in Pascals triangle. Thats why
we call it binomial tree.
Another interesting property is that the total number of node for a binomial
tree with rank N is 2N . This can be proved either by binomial theory or the
recursive definition directly.
Binomial heap
With binomial tree defined, we can introduce the definition of binomial heap.
A binomial heap is a set of binomial trees (or a forest of binomial trees) that
satisfied the following properties.
Each binomial tree in the heap conforms to heap property, that the key
of a node is equal or greater than the key of its parent. Here the heap is
actually min-heap, for max-heap, it changes to equal or less than. In this
chapter, we only discuss about min-heap, and max-heap can be equally
applied by changing the comparison condition.
There is at most one binomial tree which has the rank r. In other words,
there are no two binomial trees have the same rank.
This definition leads to an important result that for a binomial heap contains
N elements, and if convert N to binary format yields a0 , a1 , a2 , ..., am , where a0
is the LSB and am is the MSB, then for each 0 i m, if ai = 0, there is no
binomial tree of rank i and if ai = 1, there must be a binomial tree of rank i.
For example, if a binomial heap contains 5 element, as 5 is (LSB)101(MSB),
then there are 2 binomial trees in this heap, one tree has rank 0, the other has
rank 2.
Figure 10.3 shows a binomial heap which have 19 nodes, as 19 is (LSB)11001(MSB)
in binary format, so there is a B0 tree, a B1 tree and a B4 tree.
18

Figure 10.3: A binomial heap with 19 elements

10.2. BINOMIAL HEAPS

253

Data layout
There are two ways to define K-ary trees imperatively. One is by using leftchild, right-sibling approach[2]. It is compatible with the typical binary tree
structure. For each node, it has two fields, left field and right field. We use the
left field to point to the first child of this node, and use the right field to point to
the sibling node of this node. All siblings are represented as a single directional
linked list. Figure 10.4 shows an example tree represented in this way.
R

...

NIL

...

Figure 10.4: Example tree represented in left-child, right-sibling way. R is the

root node, it has no sibling, so it right side is pointed to N IL. C1 , C2 , ..., Cn
are children of R. C1 is linked from the left side of R, other siblings of C1 are
0
are children of
linked one next to each other on the right side of C1 . C20 , ..., Cm
C1 .
The other way is to use the library defined collection container, such as array
or list to represent all children of a node.
Since the rank of a tree plays very important role, we also defined it as a
field.
For left-child, right-sibling method, we defined the binomial tree as the
following.1
class BinomialTree:
def __init__(self, x = None):
self.rank = 0
self.key = x
self.parent = None
self.child = None
self.sibling = None

When initialize a tree with a key, we create a leaf node, set its rank as zero
and all other fields are set as NIL.
It quite nature to utilize pre-defined list to represent multiple children as
below.
class BinomialTree:
def __init__(self, x = None):
self.rank = 0
self.key = x
self.parent = None
self.children = []
1C

programs are also provided along with this book.

254CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

For purely functional settings, such as in Haskell language, binomial tree are
defined as the following.
data BiTree a = Node { rank :: Int
, root :: a
, children :: [BiTree a]}

While binomial heap are defined as a list of binomial trees (a forest) with
ranks in monotonically increase order. And as another implicit constraint, there
are no two binomial trees have the same rank.
type BiHeap a = [BiTree a]

10.2.2

Basic heap operations

Linking trees
Before dive into the basic heap operations such as pop and insert, Well first
realize how to link two binomial trees with same rank into a bigger one. According to the definition of binomial tree and heap property that the root always
contains the minimum key, we firstly compare the two root values, select the
smaller one as the new root, and insert the other tree as the first child in front
of all other children. Suppose function Key(T ), Children(T ), and Rank(T )
access the key, children and rank of a binomial tree respectively.

link(T1 , T2 ) =

node(r + 1, x, {T2 } C1 ) : x < y

node(r + 1, y, {T1 } C2 ) : otherwise

(10.1)

Where
x = Key(T1 )
y = Key(T2 )
r = Rank(T1 ) = Rank(T2 )
C1 = Children(T1 )
C2 = Children(T2 )

...

Figure 10.5: Suppose x < y, insert y as the first child of x.

Note that the link operation is bound to O(1) time if the is a constant
time operation. Its easy to translate the link function to Haskell program as
the following.

10.2. BINOMIAL HEAPS

255

link :: (Ord a) BiTree a BiTree a BiTree a

link t1@(Node r x c1) t2@(Node _ y c2) =
if x<y then Node (r+1) x (t2:c1)
else Node (r+1) y (t1:c2)

Its possible to realize the link operation in imperative way. If we use left
child, right sibling approach, we just link the tree which has the bigger key to
the left side of the others key, and link the children of it to the right side as
sibling. Figure 10.6 shows the result of one case.
1: function Link(T1 , T2 )
2:
if Key(T2 ) < Key(T1 ) then
3:
Exchange T1 T2
4:
Sibling(T2 ) Child(T1 )
5:
Child(T1 ) T2
6:
Parent(T2 ) T1
7:
Rank(T1 ) Rank(T1 ) + 1
8:
return T1
x

...

Figure 10.6: Suppose x < y, link y to the left side of x and link the original
children of x to the right side of y.
And if we use a container to manage all children of a node, the algorithm is
like below.
1: function Link(T1 , T2 )
2:
if Key(T2 ) < Key(T1 ) then
3:
Exchange T1 T2
4:
Parent(T2 ) T1
5:
Insert-Before(Children(T1 ), T2 )
6:
Rank(T1 ) Rank(T1 ) + 1
7:
return T1
Its easy to translate both algorithms to real program. Here we only show
the Python program of Link for illustration purpose 2 .
def link(t1, t2):
if t2.key < t1.key:
(t1, t2) = (t2, t1)
t2.parent = t1
t1.children.insert(0, t2)
t1.rank = t1.rank + 1
return t1
2 The

C and C++ programs are also available along with this book

256CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

Exercise 10.1
Implement the tree-linking program in your favorite language with left-child,
right-sibling method.

We mentioned linking is a constant time algorithm and it is true when using

left-child, right-sibling approach, However, if we use container to manage the
children, the performance depends on the concrete implementation of the container. If it is plain array, the linking time will be proportion to the number of
children. In this chapter, we assume the time is constant. This is true if the
container is implemented in linked-list.
Insert a new element to the heap (push)
As the rank of binomial trees in a forest is monotonically increasing, by using
the link function defined above, its possible to define an auxiliary function, so
that we can insert a new tree, with rank no bigger than the smallest one, to the
heap which is a forest actually.
Denote the non-empty heap as H = {T1 , T2 , ..., Tn }, we define

insertT (H, T ) =

{T } : H =
{T } H : Rank(T ) < Rank(T1 )

insertT (H 0 , link(T, T1 )) : otherwise

(10.2)

where
H 0 = {T2 , T3 , ..., Tn }
The idea is that for the empty heap, we set the new tree as the only element
to create a singleton forest; otherwise, we compare the ranks of the new tree
and the first tree in the forest, if they are same, we link them together, and
recursively insert the linked result (a tree with rank increased by one) to the
rest of the forest; If they are not same, since the pre-condition constraints the
rank of the new tree, it must be the smallest, we put this new tree in front of
all the other trees in the forest.
From the binomial properties mentioned above, there are at most O(lg N )
binomial trees in the forest, where N is the total number of nodes. Thus function
insertT performs at most O(lg N ) times linking, which are all constant time
operation. So the performance of insertT is O(lg N ). 3
The relative Haskell program is given as below.
insertTree :: (Ord a) BiHeap a BiTree a BiHeap a
insertTree [] t = [t]
insertTree ts@(t:ts) t = if rank t < rank t then t:ts
else insertTree ts (link t t)
3 There is interesting observation by comparing this operation with adding two binary
numbers. Which will lead to topic of numeric representation[6].

10.2. BINOMIAL HEAPS

257

With this auxiliary function, its easy to realize the insertion. We can wrap
the new element to be inserted as the only leaf of a tree, then insert this tree to
the binomial heap.
insert(H, x) = insertT (H, node(0, x, ))

(10.3)

And we can continuously build a heap from a series of elements by folding.

For example the following Haskell define a helper function fromList.
fromList = foldl insert []

Since wrapping an element as a singleton tree takes O(1) time, the real work
is done in insertT , the performance of binomial heap insertion is bound to
O(lg N ).
The insertion algorithm can also be realized with imperative approach.
Algorithm 4 Insert a tree with left-child-right-sibling method.
1:
2:
3:
4:
5:
6:

function Insert-Tree(H, T )
while H 6= Rank(Head(H)) = Rank(T ) do
(T1 , H) Extract-Head(H)
T Link(T, T1 )
Sibling(T ) H
return T

Algorithm 4 continuously linking the first tree in a heap with the new tree
to be inserted if they have the same rank. After that, it puts the linked-list of
the rest trees as the sibling, and returns the new linked-list.
If using a container to manage the children of a node, the algorithm can be
given in Algorithm 5.
Algorithm 5 Insert a tree with children managed by a container.
1:
2:
3:
4:
5:
6:

function Insert-Tree(H, T )
while H 6= Rank(H[0]) = Rank(T ) do
T1 Pop(H)
T Link(T, T1 )
Head-Insert(H, T )
return H

In this algorithm, function Pop removes the first tree T1 = H[0] from the
forest. And function Head-Insert, insert a new tree before any other trees in
the heap, so that it becomes the first element in the forest.
With either Insert-Tree or Insert-Tree defined. Realize the binomial
heap insertion is trivial.
Algorithm 6 Imperative insert algorithm
1:
2:

function Insert(H, x)
return Insert-Tree(H, Node(0, x, ))

The following python program implement the insert algorithm by using a

container to manage sub-trees. the left-child, right-sibling program is left as
an exercise.

258CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

def insert_tree(ts, t):
while ts !=[] and t.rank == ts[0].rank:
t = link(t, ts.pop(0))
ts.insert(0, t)
return ts
def insert(h, x):
return insert_tree(h, BinomialTree(x))

Exercise 10.2
Write the insertion program in your favorite imperative programming language by using the left-child, right-sibling approach.
Merge two heaps
When merge two binomial heaps, we actually try to merge two forests of binomial trees. According to the definition, there cant be two trees with the same
rank and the ranks are in monotonically increasing order. Our strategy is very
similar to merge sort. That in every iteration, we take the first tree from each
forest, compare their ranks, and pick the smaller one to the result heap; if the
ranks are equal, we then perform linking to get a new tree, and recursively insert
this new tree to the result of merging the rest trees.
Figure 10.7 illustrates the idea of this algorithm. This method is different
from the one given in [2].
We can formalize this idea with a function. For non-empty cases, we denote
the two heaps as H1 = {T1 , T2 , ...} and H2 = {T10 , T20 , ...}. Let H10 = {T2 , T3 , ...}
and H20 = {T20 , T30 , ...}.

H1
H2
{T1 } merge(H10 , H2 )
{T10 } merge(H1 , H20 )
insertT (merge(H10 , H20 ), link(T1 , T10 ))

:
:
:
:
:

H2 =
H1 =
Rank(T1 ) < Rank(T10 )
merge(H1 , H2 ) =

Rank(T1 ) > Rank(T10 )

otherwise
(10.4)
To analysis the performance of merge, suppose there are m1 trees in H1 ,
and m2 trees in H2 . There are at most m1 + m2 trees in the merged result.
If there are no two trees have the same rank, the merge operation is bound to
O(m1 + m2 ). While if there need linking for the trees with same rank, insertT
performs at most O(m1 + m2 ) time. Consider the fact that m1 = 1 + blg N1 c
and m2 = 1 + blg N2 c, where N1 , N2 are the numbers of nodes in each heap,
and blg N1 c + blg N2 c 2blg N c, where N = N1 + N2 , is the total number of
nodes. the final performance of merging is O(lg N ).
Translating this algorithm to Haskell yields the following program.
merge:: (Ord a) BiHeap a BiHeap a BiHeap a
merge ts1 [] = ts1
merge [] ts2 = ts2
merge ts1@(t1:ts1) ts2@(t2:ts2)
| rank t1 < rank t2 = t1:(merge ts1 ts2)
| rank t1 > rank t2 = t2:(merge ts1 ts2)
| otherwise = insertTree (merge ts1 ts2) (link t1 t2)

10.2. BINOMIAL HEAPS

259

...

Rank(t1)<Rank(t2)?
the smaller
T1

...

(a) Pick the tree with smaller rank to the

result.

...

Rank(t1)=Rank(t2)
link(t1, t2)
T1

...

insert

merge rest

(b) If two trees have same rank, link them to a new tree, and recursively insert it to the
merge result of the rest.

Figure 10.7: Merge two heaps.

260CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

Merge algorithm can also be described in imperative way as shown in algorithm 7.
Algorithm 7 imperative merge two binomial heaps
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:

function Merge(H1 , H2 )
if H1 = then
return H2
if H2 = then
return H1
H
while H1 6= H2 6= do
T
if Rank(H1 ) < Rank(H2 ) then
(T, H1 ) Extract-Head(H1 )
else if Rank(H2 ) < Rank(H1 ) then
(T, H2 ) Extract-Head(H2 )
else
(T1 , H1 ) Extract-Head(H1 )
(T2 , H2 ) Extract-Head(H2 )
T Link(T1 , T2 )
Append-Tree(H, T )
if H1 6= then
Append-Trees(H, H1 )
if H2 6= then
Append-Trees(H, H2 )
return H

. Equal rank

Since both heaps contain binomial trees with rank in monotonically increasing order. Each iteration, we pick the tree with smallest rank and append it to
the result heap. If both trees have same rank we perform linking first. Consider
the Append-Tree algorithm, The rank of the new tree to be appended, cant
be less than any other trees in the result heap according to our merge strategy,
however, it might be equal to the rank of the last tree in the result heap. This
can happen if the last tree appended are the result of linking, which will increase
the rank by one. In this case, we must link the new tree to be inserted with the
last tree. In below algorithm, suppose function Last(H) refers to the last tree
in a heap, and Append(H, T ) just appends a new tree at the end of a forest.
1: function Append-Tree(H, T )
2:
if H 6= Rank(T ) = Rank(Last(H)) then
3:
Last(H) Link(T , Last(H))
4:
else
5:
Append(H, T )
Function Append-Trees repeatedly call this function, so that it can append
all trees in a heap to the other heap.
1: function Append-Trees(H1 , H2 )
2:
for each T H2 do
3:
H1 Append-Tree(H1 , T )

10.2. BINOMIAL HEAPS

261

The following Python program translates the merge algorithm.

def append_tree(ts, t):
if ts != [] and ts[-1].rank == t.rank:
ts[-1] = link(ts[-1], t)
else:
ts.append(t)
return ts
def append_trees(ts1, ts2):
return reduce(append_tree, ts2, ts1)
def merge(ts1, ts2):
if ts1 == []:
return ts2
if ts2 == []:
return ts1
ts = []
while ts1 != [] and ts2 != []:
t = None
if ts1[0].rank < ts2[0].rank:
t = ts1.pop(0)
elif ts2[0].rank < ts1[0].rank:
t = ts2.pop(0)
else:
t = link(ts1.pop(0), ts2.pop(0))
ts = append_tree(ts, t)
ts = append_trees(ts, ts1)
ts = append_trees(ts, ts2)
return ts

Exercise 10.3
The program given above uses a container to manage sub-trees. Implement
the merge algorithm in your favorite imperative programming language with
left-child, right-sibling approach.
Pop
Among the forest which forms the binomial heap, each binomial tree conforms
to heap property that the root contains the minimum element in that tree.
However, the order relationship of these roots can be arbitrary. To find the
minimum element in the heap, we can select the smallest root of these trees.
Since there are lg N binomial trees, this approach takes O(lg N ) time.
However, after we locate the minimum element (which is also know as the
top element of a heap), we need remove it from the heap and keep the binomial
property to accomplish heap-pop operation. Suppose the forest forms the binomial heap consists trees of Bi , Bj , ..., Bp , ..., Bm , where Bk is a binomial tree of
rank k, and the minimum element is the root of Bp . If we delete it, there will
be p children left, which are all binomial trees with ranks p 1, p 2, ..., 0.
One tool at hand is that we have defined O(lg N ) merge function. A possible
approach is to reverse the p children, so that their ranks change to monotonically
increasing order, and forms a binomial heap Hp . The rest of trees is still a

262CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

binomial heap, we represent it as H 0 = H Bp . Merging Hp and H 0 given the
final result of pop. Figure 10.8 illustrates this idea.

Figure 10.8: Pop the minimum element from a binomial heap.

In order to realize this algorithm, we first need to define an auxiliary function,
which can extract the tree contains the minimum element at root from the forest.

extractM in(H) =

(T, ) : H is a singleton as {T }
(T1 , H 0 ) : Root(T1 ) < Root(T 0 )

(T 0 , {T1 } H 00 ) : otherwise

(10.5)

where
H = {T1 , T2 , ...}
H 0 = {T2 , T3 , ...}
(T 0 , H 00 ) = extractM in(H 0 )

for the non-empty forest case;

is the forest without the first tree;

The result of this function is a tuple. The first part is the tree which has the
minimum element at root, the second part is the rest of the trees after remove
the first part from the forest.
This function examine each of the trees in the forest thus is bound to O(lg N )
time.
The relative Haskell program can be give respectively.
extractMin :: (Ord a) BiHeap a (BiTree a, BiHeap a)

10.2. BINOMIAL HEAPS

263

extractMin [t] = (t, [])

extractMin (t:ts) = if root t < root t then (t, ts)
else (t, t:ts)
where
(t, ts) = extractMin ts

With this function defined, to return the minimum element is trivial.

findMin :: (Ord a) BiHeap a a
findMin = root fst. extractMin

Of course, its possible to just traverse forest and pick the minimum root
without remove the tree for this purpose. Below imperative algorithm describes
it with left child, right sibling approach.
1: function Find-Minimum(H)
2:
T Head(H)
3:
min
4:
while T 6= do
5:
if Key(T )< min then
6:
min Key(T )
7:
T Sibling(T )
8:
return min
While if we manage the children with collection containers, the link list
traversing is abstracted as to find the minimum element among the list. The
following Python program shows about this situation.
def find_min(ts):
min_t = min(ts, key=lambda t: t.key)
return min_t.key

Next we define the function to delete the minimum element from the heap
by using extractM in.
delteM in(H) = merge(reverse(Children(T )), H 0 )

(10.6)

where
(T, H 0 ) = extractM in(H)
Translate the formula to Haskell program is trivial and well skip it.
To realize the algorithm in procedural way takes extra efforts including list
reversing etc. We left these details as exercise to the reader. The following
pseudo code illustrate the imperative pop algorithm
1: function Extract-Min(H)
2:
(Tmin , H) Extract-Min-Tree(H)
3:
H Merge(H, Reverse(Children(Tmin )))
4:
return (Key(Tmin ), H)
With pop operation defined, we can realize heap sort by creating a binomial
heap from a series of numbers, than keep popping the smallest number from the
heap till it becomes empty.
sort(xs) = heapSort(f romList(xs))

(10.7)

264CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

And the real work is done in function heapSort.

: H=
{f indM in(H)} heapSort(deleteM in(H)) : otherwise
(10.8)
Translate to Haskell yields the following program.

heapSort(H) =

heapSort :: (Ord a) [a] [a]

heapSort = hsort fromList where
hsort [] = []
hsort h = (findMin h):(hsort $ deleteMin h)

Function fromList can be defined by folding. Heap sort can also be expressed
in procedural way respectively. Please refer to previous chapter about binary
heap for detail.

Exercise 10.4
Write the program to return the minimum element from a binomial heap
in your favorite imperative programming language with left-child, rightsibling approach.
Realize the Extract-Min-Tree() Algorithm.
For left-child, right-sibling approach, reversing all children of a tree is
actually reversing a single-direct linked-list. Write a program to reverse
such linked-list in your favorite imperative programming language.
More words about binomial heap
As what we have shown that insertion and merge are bound to O(lg N ) time.
The results are all ensure for the worst case. The amortized performance are
O(1). We skip the proof for this fact.

10.3

Fibonacci Heaps

Its interesting that why the name is given as Fibonacci heap. In fact, there is
no direct connection from the structure design to Fibonacci series. The inventors
of Fibonacci heap, Michael L. Fredman and Robert E. Tarjan, utilized the
property of Fibonacci series to prove the performance time bound, so they
decided to use Fibonacci to name this data structure.[2]

10.3.1

Definition

Fibonacci heap is essentially a lazy evaluated binomial heap. Note that, it

doesnt mean implementing binomial heap in lazy evaluation settings, for instance Haskell, brings Fibonacci heap automatically. However, lazy evaluation
setting does help in realization. For example in [5], presents a elegant implementation.
Fibonacci heap has excellent performance theoretically. All operations except for pop are bound to amortized O(1) time. In this section, well give an

10.3. FIBONACCI HEAPS

265

algorithm different from some popular textbook[2]. Most of the ideas present
here are based on Okasakis work[6].
Lets review and compare the performance of binomial heap and Fibonacci
heap (more precisely, the performance goal of Fibonacci heap).
operation Binomial heap
Fibonacci heap
insertion
O(lg N )
O(1)
merge
O(lg N )
O(1)
top
O(lg N )
O(1)
pop
O(lg N )
amortized O(lg N )
Consider where is the bottleneck of inserting a new element x to binomial
heap. We actually wrap x as a singleton leaf and insert this tree into the heap
which is actually a forest.
During this operation, we inserted the tree in monotonically increasing order
of rank, and once the rank is equal, recursively linking and inserting will happen,
which lead to the O(lg N ) time.
As the lazy strategy, we can postpone the ordered-rank insertion and merging
operations. On the contrary, we just put the singleton leaf to the forest. The
problem is that when we try to find the minimum element, for example the top
operation, the performance will be bad, because we need check all trees in the
forest, and there arent only O(lg N ) trees.
In order to locate the top element in constant time, we must remember where
is the tree contains the minimum element as root.
Based on this idea, we can reuse the definition of binomial tree and give the
definition of Fibonacci heap as the following Haskell program for example.
data BiTree a = Node { rank :: Int
, root :: a
, children :: [BiTree a]}

The Fibonacci heap is either empty or a forest of binomial trees with the
minimum element stored in a special one explicitly.
data FibHeap a = E | FH { size :: Int
, minTree :: BiTree a
, trees :: [BiTree a]}

For convenient purpose, we also add a size field to record how many elements
are there in a heap.
The data layout can also be defined in imperative way as the following ANSI
C code.
struct node{
Key key;
struct node next, prev, parent, children;
int degree; / As known as rank /
int mark;
};
struct FibHeap{
struct node roots;
struct node minTr;
int n; / number of nodes /
};

266CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

For generality, Key can be a customized type, we use integer for illustration
purpose.
typedef int Key;

In this chapter, we use the circular doubly linked-list for imperative settings
to realize the Fibonacci Heap as described in [2]. It makes many operations
easy and fast. Note that, there are two extra fields added. A degree also known
as rank for a node is the number of children of this node; Flag mark is used
only in decreasing key operation. It will be explained in detail in later section.

10.3.2

Basic heap operations

As we mentioned that Fibonacci heap is essentially binomial heap implemented

in a lazy evaluation strategy, well reuse many algorithms defined for binomial
heap.
Insert a new element to the heap
Recall the insertion algorithm of binomial tree. It can be treated as a special
case of merge operation, that one heap contains only a singleton tree. So that
the inserting algorithm can be defined by means of merging.
insert(H, x) = merge(H, singleton(x))

(10.9)

where singleton is an auxiliary function to wrap an element to a one-leaf-tree.

singleton(x) = F ibHeap(1, node(1, x, ), )
Note that function F ibHeap() accepts three parameters, a size value, which
is 1 for this one-leaf-tree, a special tree which contains the minimum element as
root, and a list of other binomial trees in the forest. The meaning of function
node() is as same as before, that it creates a binomial tree from a rank, an
element, and a list of children.
Insertion can also be realized directly by appending the new node to the
forest and updated the record of the tree which contains the minimum element.
1: function Insert(H, k)
2:
x Singleton(k)
. Wrap x to a node
3:
append x to root list of H
4:
if Tmin (H) = N IL k < Key(Tmin (H)) then
5:
Tmin (H) x
6:
n(H) n(H)+1
Where function Tmin () returns the tree which contains the minimum element
at root.
The following C source snippet is a translation for this algorithm.
struct FibHeap insert_node(struct FibHeap h, struct node x){
h = add_tree(h, x);
if(hminTr == NULL | | xkey < hminTrkey)
hminTr = x;
hn++;
return h;
}

10.3. FIBONACCI HEAPS

267

Exercise 10.5
Implement the insert algorithm in your favorite imperative programming
language completely. This is also an exercise to circular doubly linked list manipulation.
Merge two heaps
Different with the merging algorithm of binomial heap, we post-pone the linking
operations later. The idea is to just put all binomial trees from each heap
together, and choose one special tree which record the minimum element for
the result heap.

H1
H2
merge(H1 , H2 ) =
F
ibHeap(s
+
s
,
T
,
{T
}

1
2
1min
2min
1
2)

F ibHeap(s1 + s2 , T2min , {T1min } T1 T2 )

:
:
:
:

H2 =
H1 =
root(T1min ) < root(T2min )
otherwise
(10.10)
where s1 and s2 are the size of H1 and H2 ; T1min and T2min are the special trees with minimum element as root in H1 and H2 respectively; T1 =
{T11 , T12 , ...} is a forest contains all other binomial trees in H1 ; while T2 has
the same meaning as T1 except that it represents the forest in H2 . Function
root(T ) return the root element of a binomial tree.
Note that as long as the operation takes constant time, these merge algorithm is bound to O(1). The following Haskell program is the translation of
this algorithm.
merge:: (Ord a) FibHeap a FibHeap a FibHeap a
merge h E = h
merge E h = h
merge h1@(FH sz1 minTr1 ts1) h2@(FH sz2 minTr2 ts2)
| root minTr1 < root minTr2 = FH (sz1+sz2) minTr1 (minTr2:ts2++ts1)
| otherwise = FH (sz1+sz2) minTr2 (minTr1:ts1++ts2)

Merge algorithm can also be realized imperatively by concatenating the root

lists of the two heaps.
1: function Merge(H1 , H2 )
2:
H
3:
Root(H) Concat(Root(H1 ), Root(H2 ))
4:
if Key(Tmin (H1 )) < Key(Tmin (H2 )) then
5:
Tmin (H) Tmin (H1 )
6:
else
7:
Tmin (H) Tmin (H2 )
n(H) = n(H1 ) + n(H2 )
8:
release H1 and H2
9:
return H
This function assumes neither H1 , nor H2 is empty. And its easy to add
handling to these special cases as the following ANSI C program.
struct FibHeap merge(struct FibHeap h1, struct FibHeap h2){
struct FibHeap h;
if(is_empty(h1))

268CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

return h2;
if(is_empty(h2))
return h1;
h = empty();
hroots = concat(h1roots, h2roots);
if(h1minTrkey < h2minTrkey)
hminTr = h1minTr;
else
hminTr = h2minTr;
hn = h1n + h2n;
free(h1);
free(h2);
return h;
}

With merge function defined, the O(1) insertion algorithm is realized as

well. And we can also give the O(1) time top function as below.
top(H) = root(Tmin )

(10.11)

Exercise 10.6
Implement the circular doubly linked list concatenation function in your
favorite imperative programming language.
Extract the minimum element from the heap (pop)
The pop (delete the minimum element) operation is the most complex one in
Fibonacci heap. Since we postpone the tree consolidation in merge algorithm.
We have to compensate it somewhere. Pop is the only place left as we have
defined, insert, merge, top already.
There is an elegant procedural algorithm to do the tree consolidation by
using an auxiliary array[2]. Well show it later in imperative approach section.
In order to realize the purely functional consolidation algorithm, lets first
consider a similar number puzzle.
Given a list of numbers, such as {2, 1, 1, 4, 8, 1, 1, 2, 4}, we want to add any
two values if they are same. And repeat this procedure till all numbers are
unique. The result of the example list should be {8, 16} for instance.
One solution to this problem will as the following.
consolidate(L) = f old(meld, , L)

(10.12)

Where f old() function is defined to iterate all elements from a list, applying
a specified function to the intermediate result and each element. it is sometimes
called as reducing. Please refer to the chapter of binary search tree for it.
L = {x1 , x2 , ..., xn }, denotes a list of numbers; and well use L0 = {x2 , x3 , ..., xn }
to represent the rest of the list with the first element removed. Function meld()
is defined as below.

{x} : L =

meld(L0 , x + x1 ) : x = x1
meld(L, x) =
(10.13)
{x} L : x < x1

0
{x1 } meld(L , x) : otherwise

10.3. FIBONACCI HEAPS

269

Table 10.1: Steps of consolidate numbers

number
2
1
1
4
8
1
1
2
4

intermediate result
2
1, 2
(1+1), 2
(4+4)
(8+8)
1, 16
(1+1), 16
(2+2), 16
(4+4), 16

result
2
1, 2
4
8
16
1, 16
2, 16
4, 16
8, 16

The consolidate() function works as the follows. It maintains an ordered

result list L, contains only unique numbers, which is initialized from an empty
list . Each time it process an element x, it firstly check if the first element in L
is equal to x, if so, it will add them together (which yields 2x), and repeatedly
check if 2x is equal to the next element in L. This process wont stop until either
the element to be melt is not equal to the head element in the rest of the list, or
the list becomes empty. Table 10.1 illustrates the process of consolidating number sequence {2, 1, 1, 4, 8, 1, 1, 2, 4}. Column one lists the number scanned one
by one; Column two shows the intermediate result, typically the new scanned
number is compared with the first number in result list. If they are equal, they
are enclosed in a pair of parentheses; The last column is the result of meld, and
it will be used as the input to next step processing.
The Haskell program can be give accordingly.
consolidate = foldl meld [] where
meld [] x = [x]
meld (x:xs) x | x == x = meld xs (x+x)
| x < x = x:x:xs
| otherwise = x: meld xs x

Well analyze the performance of consolidation as a part of pop operation in

later section.
The tree consolidation is very similar to this algorithm except it performs
based on rank. The only thing we need to do is to modify meld() function a
bit, so that it compare on ranks and do linking instead of adding.

{x}
meld(L0 , link(x, x1 ))
meld(L, x) =
{x} L

{x1 } meld(L0 , x)

:
:
:
:

L=
rank(x) = rank(x1 )
rank(x) < rank(x1 )
otherwise

The final consolidate Haskell program changes to the below version.

consolidate :: (Ord a) [BiTree a] [BiTree a]
consolidate = foldl meld [] where
meld [] t = [t]
meld (t:ts) t | rank t == rank t = meld ts (link t t)
| rank t < rank t = t:t:ts
| otherwise = t : meld ts t

(10.14)

270CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

Figure 10.9 and 10.10 show the steps of consolidation when processing a
Fibonacci Heap contains different ranks of trees. Comparing with table 10.1
reveals the similarity.
a

(a) Before consolidation

(b) Step 1, 2 (c) Step 3, d is firstly linked to c,

then repeatedly linked to a.

(d) Step 4

Figure 10.9: Steps of consolidation

After we merge all binomial trees, including the special tree record for the
minimum element in root, in a Fibonacci heap, the heap becomes a Binomial
heap. And we lost the special tree, which gives us the ability to return the top
element in O(1) time.
Its necessary to perform a O(lg N ) time search to resume the special tree.
We can reuse the function extractM in() defined for Binomial heap.
Its time to give the final pop function for Fibonacci heap as all the sub
problems have been solved. Let Tmin denote the special tree in the heap to
record the minimum element in root; T denote the forest contains all the other
trees except for the special tree, s represents the size of a heap, and function
children() returns all sub trees except the root of a binomial tree.

deleteM in(H) =
Where

: T = children(Tmin ) =
0
F ibHeap(s 1, Tmin
, T0 ) : otherwise
(10.15)

10.3. FIBONACCI HEAPS

271

(b) Step 6

(a) Step 5

(c) Step 7, 8, r is firstly linked to q, then s is linked to q.

Figure 10.10: Steps of consolidation

272CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

0
(Tmin
, T0 ) = extractM in(consolidate(children(Tmin ) T))

Translate to Haskell yields the below program.

deleteMin :: (Ord a) FibHeap a FibHeap a
deleteMin (FH _ (Node _ x []) []) = E
deleteMin h@(FH sz minTr ts) = FH (sz-1) minTr ts where
(minTr, ts) = extractMin $ consolidate (children minTr ++ ts)

The main part of the imperative realization is similar. We cut all children
of Tmin and append them to root list, then perform consolidation to merge all
trees with the same rank until all trees are unique in term of rank.
1: function Delete-Min(H)
2:
x Tmin (H)
3:
if x 6= N IL then
4:
for each y Children(x) do
5:
append y to root list of H
6:
Parent(y) N IL
7:
remove x from root list of H
8:
n(H) n(H) - 1
9:
Consolidate(H)
10:
return x
Algorithm Consolidate utilizes an auxiliary array A to do the merge job.
Array A[i] is defined to store the tree with rank (degree) i. During the traverse
of root list, if we meet another tree of rank i, we link them together to get a
new tree of rank i + 1. Next we clean A[i], and check if A[i + 1] is empty and
perform further linking if necessary. After we finish traversing all roots, array
A stores all result trees and we can re-construct the heap from it.
1: function Consolidate(H)
2:
D Max-Degree(n(H))
3:
for i 0 to D do
4:
A[i] N IL
5:
for each x root list of H do
6:
remove x from root list of H
7:
d Degree(x)
8:
while A[d] 6= N IL do
9:
y A[d]
10:
x Link(x, y)
11:
A[d] N IL
12:
dd+1
13:
A[d] x
14:
Tmin (H) N IL
. root list is NIL at the time
15:
for i 0 to D do
16:
if A[i] 6= N IL then
17:
append A[i] to root list of H.
18:
if Tmin = N IL Key(A[i]) < Key(Tmin (H)) then
19:
Tmin (H) A[i]
The only unclear sub algorithm is Max-Degree, which can determine the
upper bound of the degree of any node in a Fibonacci Heap. Well delay the

10.3. FIBONACCI HEAPS

273

realization of it to the last sub section.

Feed a Fibonacci Heap shown in Figure 10.9 to the above algorithm, Figure
10.11, 10.12 and 10.13 show the result trees stored in auxiliary array A in every
steps.

A[0]

A[1]

A[2]

A[3]

(a) Step 1, 2

A[1]

A[2]

A[4]

A[3]

A[2]

A[4]

A[1]

A[3]

(b) Step 3, Since A0 6= N IL,

d is firstly linked to c, and
clear A0 to N IL. Again, as
A1 6= N IL, c is linked to a
and the new tree is stored in A2 .

A[4]

Figure 10.11: Steps of consolidation

Translate the above algorithm to ANSI C yields the below program.
void consolidate(struct FibHeap h){
if(!hroots)
return;
int D = max_degree(hn)+1;
struct node x, y;
struct node a = (struct node)malloc(sizeof(struct node)(D+1));
int i, d;
for(i=0; i D; ++i)
a[i] = NULL;
while(hroots){
x = hroots;
hroots = remove_node(hroots, x);
d= xdegree;
while(a[d]){
y = a[d]; / Another node has the same degree as x /
x = link(x, y);
a[d++] = NULL;
}
a[d] = x;
}
hminTr = hroots = NULL;
for(i=0; i D; ++i)
if(a[i]){

274CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

A[0]

A[1]

A[2]

A[3]

A[4]

(a) Step 5

A[0]

A[1]

A[2]

A[3]

A[4]

(b) Step 6

Figure 10.12: Steps of consolidation

10.3. FIBONACCI HEAPS

A[0]

A[1]

A[2]

275

A[3]

A[4]

(a) Step 7, 8, Since A0 6= N IL, r is firstly linked to q, and the new

tree is stored in A1 (A0 is cleared); then s is linked to q, and stored
in A2 (A1 is cleared).

Figure 10.13: Steps of consolidation

hroots = append(hroots, a[i]);

if(hminTr == NULL | | a[i]key < hminTrkey)
hminTr = a[i];
}
free(a);
}

Exercise 10.7
Implement the remove function for circular doubly linked list in your favorite
imperative programming language.

10.3.3

Running time of pop

In order to analyze the amortize performance of pop, we adopt potential method.

Reader can refer to [2] for a formal definition. In this chapter, we only give a
intuitive illustration.
Remind the gravity potential energy, which is defined as
E =M gh
Suppose there is a complex process, which moves the object with mass M up
and down, and finally the object stop at height h0 . And if there exists friction
resistance Wf , We say the process works the following power.
W = M g (h0 h) + Wf
Figure 10.14 illustrated this concept.

276CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

Figure 10.14: Gravity potential energy.

We treat the Fibonacci heap pop operation in a similar way, in order to
evaluate the cost, we firstly define the potential (H) before extract the minimum element. This potential is accumulated by insertion and merge operations
executed so far. And after tree consolidation and we get the result H 0 , we then
calculate the new potential (H 0 ). The difference between (H 0 ) and (H) plus
the contribution of consolidate algorithm indicates the amortized performance
of pop.
For pop operation analysis, the potential can be defined as
(H) = t(H)

(10.16)

Where t(H) is the number of trees in Fibonacci heap forest. We have t(H) =
1 + length(T) for any non-empty heap.
For the N -nodes Fibonacci heap, suppose there is an upper bound of ranks
for all trees as D(N ). After consolidation, it ensures that the number of trees
in the heap forest is at most D(N ) + 1.
Before consolidation, we actually did another important thing, which also
contribute to running time, we removed the root of the minimum tree, and
concatenate all children left to the forest. So consolidate operation at most
processes D(N ) + t(H) 1 trees.
Summarize all the above factors, we deduce the amortized cost as below.
T

= Tconsolidation + (H 0 ) (H)
= O(D(N ) + t(H) 1) + (D(N ) + 1) t(H)
= O(D(N ))

(10.17)

If only insertion, merge, and pop function are applied to Fibonacci heap.
We ensure that all trees are binomial trees. It is easy to estimate the upper
limit D(N ) if O(lg N ). (Suppose the extreme case, that all nodes are in only
one Binomial tree).
However, well show in next sub section that, there is operation can violate
the binomial tree assumption.

10.3. FIBONACCI HEAPS

10.3.4

277

Decreasing key

There is a special heap operation left. It only makes sense for imperative settings. Its about decreasing key of a certain node. Decreasing key plays important role in some Graphic algorithms such as Minimum Spanning tree algorithm
and Dijkstras algorithm [2]. In that case we hope the decreasing key takes O(1)
amortized time.
However, we cant define a function like Decrease(H, k, k 0 ), which first locates a node with key k, then decrease k to k 0 by replacement, and then resume
the heap properties. This is because the time for locating phase is bound to
O(N ) time, since we dont have a pointer to the target node.
In imperative setting, we can define the algorithm as Decrease-Key(H, x, k).
Here x is a node in heap H, which we want to decrease its key to k. We neednt
perform a search, as we have x at hand. Its possible to give an amortized O(1)
solution.
When we decreased the key of a node, if its not a root, this operation may
violate the property Binomial tree that the key of parent is less than all keys of
children. So we need to compare the decreased key with the parent node, and if
this case happens, we can cut this node and append it to the root list. (Remind
the recursive swapping solution for binary heap which leads to O(lg N ))

...

@
@
x

...

Figure 10.15: x < y, cut tree x from its parent, and add x to root list.
Figure 10.15 illustrates this situation. After decreasing key of node x, it is
less than y, we cut x off its parent y, and past the whole tree rooted at x to
root list.
Although we recover the property of that parent is less than all children, the
tree isnt any longer a Binomial tree after it losses some sub tree. If a tree losses
too many of its children because of cutting, we cant ensure the performance of
merge-able heap operations. Fibonacci Heap adds another constraints to avoid
such problem:
If a node losses its second child, it is immediately cut from parent, and added
to root list

278CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

1:
2:
3:
4:
5:
6:
7:
8:

The final Decrease-Key algorithm is given as below.

function Decrease-Key(H, x, k)
Key(x) k
p Parent(x)
if p 6= N IL k < Key(p) then
Cut(H, x)
Cascading-Cut(H, p)
if k < Key(Tmin (H)) then
Tmin (H) x

Where function Cascading-Cut uses the mark to determine if the node is

losing the second child. the node is marked after it losses the first child. And
the mark is cleared in Cut function.
1: function Cut(H, x)
2:
p Parent(x)
3:
remove x from p
4:
Degree(p) Degree(p) - 1
5:
add x to root list of H
6:
Parent(x) N IL
7:
Mark(x) F ALSE
During cascading cut process, if x is marked, which means it has already
lost one child. We recursively performs cut and cascading cut on its parent till
reach to root.
1: function Cascading-Cut(H, x)
2:
p Parent(x)
3:
if p 6= N IL then
4:
if Mark(x) = F ALSE then
5:
Mark(x) T RU E
6:
else
7:
Cut(H, x)
8:
Cascading-Cut(H, p)
The relevant ANSI C decreasing key program is given as the following.
void decrease_key(struct FibHeap h, struct node x, Key k){
struct node p = xparent;
xkey = k;
if(p && k < pkey){
cut(h, x);
cascading_cut(h, p);
}
if(k < hminTrkey)
hminTr = x;
}
void cut(struct FibHeap h, struct node x){
struct node p = xparent;
pchildren = remove_node(pchildren, x);
pdegree--;
hroots = append(hroots, x);
xparent = NULL;
xmark = 0;

10.3. FIBONACCI HEAPS

279

}
void cascading_cut(struct FibHeap h, struct node x){
struct node p = xparent;
if(p){
if(!xmark)
xmark = 1;
else{
cut(h, x);
cascading_cut(h, p);
}
}
}

Exercise 10.8
Prove that Decrease-Key algorithm is amortized O(1) time.

10.3.5

The name of Fibonacci Heap

Its time to reveal the reason why the data structure is named as Fibonacci
Heap.
There is only one undefined algorithm so far, Max-Degree(N ). Which can
determine the upper bound of degree for any node in a N nodes Fibonacci Heap.
Well give the proof by using Fibonacci series and finally realize Max-Degree
algorithm.
Lemma 10.3.1. For any node x in a Fibonacci Heap, denote k = degree(x),
and |x| = size(x), then
|x| Fk+2
(10.18)
Where Fk is Fibonacci series defined as the following.

0 : k=0

1 : k=1
Fk =

Fk1 + Fk2 : k 2
Proof. Consider all k children of node x, we denote them as y1 , y2 , ..., yk in the
order of time when they were linked to x. Where y1 is the oldest, and yk is the
youngest.
Obviously, yi 0. When we link yi to x, children y1 , y2 , ..., yi1 have already
been there. And algorithm LINK only links nodes with the same degree. Which
indicates at that time, we have
degree(yi ) = degree(x) = i 1
After that, node yi can at most lost 1 child, (due to the decreasing key
operation) otherwise, if it will be immediately cut off and append to root list
after the second child loss. Thus we conclude
degree(yi ) i 2
For any i = 2, 3, ..., k.
Let sk be the minimum possible size of node x, where degree(x) = k. For
trivial cases, s0 = 1, s1 = 2, and we have

280CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

|x|

= 2+

k
X

sdegree(yi )

i=2

k
X

si2

i=2

We next show that sk > Fk+2 . This can be proved by induction. For trivial
cases, we have s0 = 1 F2 = 1, and s1 = 2 F3 = 2. For induction case k 2.
We have
|x|

k
X

si2

i=2

k
X

i=2

k
X

i=0

At this point, we need prove that

Fk+2 = 1 +

k
X

(10.19)

This can also be proved by using induction:

Trivial case, F2 = 1 + F0 = 2
Induction case,
Fk+2

= Fk+1 + Fk
= 1+

k1
X

Fi + Fk

i=0

k
X

i=0

Summarize all above we have the final result.

N |x| Fk + 2

(10.20)

10.4. PAIRING HEAPS

281

Recall the result of AVL tree, that Fk k , where = 1+2 5 is the golden
ratio. We also proved that pop operation is amortized O(lg N ) algorithm.
Based on this result. We can define Function M axDegree as the following.
M axDegree(N ) = 1 + blog N c

(10.21)

The imperative Max-Degree algorithm can also be realized by using Fibonacci sequences.
1: function Max-Degree(N )
2:
F0 0
3:
F1 1
4:
k2
5:
repeat
6:
Fk Fk1 + Fk2
7:
k k+1
8:
until Fk < N
9:
return k 2
Translate the algorithm to ANSI C given the following program.
int max_degree(int n){
int k, F;
int F2 = 0;
int F1 = 1;
for(F=F1+F2, k=2; F<n; ++k){
F2 = F1;
F1 = F;
F = F1 + F2;
}
return k-2;
}

10.4

Pairing Heaps

Although Fibonacci Heaps provide excellent performance theoretically, it is complex to realize. People find that the constant behind the big-O is big. Actually,
Fibonacci Heap is more significant in theory than in practice.
In this section, well introduce another solution, Pairing heap, which is one of
the best heaps ever known in terms of performance. Most operations including
insertion, finding minimum element (top), merging are all bounds to O(1) time,
while deleting minimum element (pop) is conjectured to amortized O(lg N ) time
[7] [6]. Note that this is still a conjecture for 15 years by the time I write this
chapter. Nobody has been proven it although there are much experimental data
support the O(lg N ) amortized result.
Besides that, pairing heap is simple. There exist both elegant imperative
and functional implementations.

10.4.1

Definition

Both Binomial Heaps and Fibonacci Heaps are realized with forest. While a
pairing heaps is essentially a K-ary tree. The minimum element is stored at
root. All other elements are stored in sub trees.

282CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

The following Haskell program defines pairing heap.
data PHeap a = E | Node a [PHeap a]

This is a recursive definition, that a pairing heap is either empty or a K-ary

tree, which is consist of a root node, and a list of sub trees.
Pairing heap can also be defined in procedural languages, for example ANSI
C as below. For illustration purpose, all heaps we mentioned later are minimumheap, and we assume the type of key is integer 4 . We use same linked-list based
left-child, right-sibling approach (aka, binary tree representation[2]).
typedef int Key;
struct node{
Key key;
struct node next, children, parent;
};

Note that the parent field does only make sense for decreasing key operation,
which will be explained later on. we can omit it for the time being.

10.4.2

Basic heap operations

In this section, we first give the merging operation for pairing heap, which can
be used to realize the insertion. Merging, insertion, and finding the minimum
element are relative trivial compare to the extracting minimum element operation.
Merge, insert, and find the minimum element (top)
The idea of merging is similar to the linking algorithm we shown previously for
Binomial heap. When we merge two pairing heaps, there are two cases.
Trivial case, one heap is empty, we simply return the other heap as the
result;
Otherwise, we compare the root element of the two heaps, make the heap
with bigger root element as a new children of the other.
Let H1 , and H2 denote the two heaps, x and y be the root element of H1
and H2 respectively. Function Children() returns the children of a K-ary tree.
Function N ode() can construct a K-ary tree from a root element and a list of
children.

H1 : H 2 =
H2 : H 1 =
merge(H1 , H2 ) =
N
ode(x,
{H
}

Children(H

2
1 )) : x < y

N ode(y, {H1 } Children(H2 )) : otherwise

(10.22)

Where
x = Root(H1 )
y = Root(H2 )
4 We can parametrize the key type with C++ template, but this is beyond our scope, please
refer to the example programs along with this book

10.4. PAIRING HEAPS

283

Its obviously that merging algorithm is bound to O(1) time 5 . The merge
equation can be translated to the following Haskell program.
merge :: (Ord a) PHeap a PHeap a PHeap a
merge h E = h
merge E h = h
merge h1@(Node x hs1) h2@(Node y hs2) =
if x < y then Node x (h2:hs1) else Node y (h1:hs2)

Merge can also be realized imperatively. With left-child, right sibling approach, we can just link the heap, which is in fact a K-ary tree, with larger key
as the first new child of the other. This is constant time operation as described
below.
1: function Merge(H1 , H2 )
2:
if H1 = NIL then
3:
return H2
4:
if H2 = NIL then
5:
return H1
6:
if Key(H2 ) < Key(H1 ) then
7:
Exchange(H1 H2 )
8:
Insert H2 in front of Children(H1 )
9:
Parent(H2 ) H1
10:
return H1
Note that we also update the parent field accordingly. The ANSI C example
program is given as the following.
struct node merge(struct node h1, struct node h2){
if(h1 == NULL)
return h2;
if(h2 == NULL)
return h1;
if(h2key < h1key)
swap(&h1, &h2);
h2next = h1children;
h1children = h2;
h2parent = h1;
h1next = NULL; /Break previous link if any/
return h1;
}

Where function swap() is defined in a similar way as Fibonacci Heap.

With merge defined, insertion can be realized as same as Fibonacci Heap in
Equation 10.9. Definitely its O(1) time operation. As the minimum element is
always stored in root, finding it is trivial.
top(H) = Root(H)

(10.23)

Same as the other two above operations, its bound to O(1) time.

Exercise 10.9
Implement the insertion and top operation in your favorite programming
language.
5 Assume is constant time operation, this is true for linked-list settings, including cons
like operation in functional programming languages.

284CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

Decrease key of a node
There is another trivial operation, to decrease key of a given node, which only
makes sense in imperative settings as we explained in Fibonacci Heap section.
The solution is simple, that we can cut the node with the new smaller key
from its parent along with all its children. Then merge it again to the heap.
The only special case is that if the given node is the root, then we can directly
set the new key without doing anything else.
The following algorithm describes this procedure for a given node x, with
new key k.
1: function Decrease-Key(H, x, k)
2:
Key(x) k
3:
if Parent(x) 6= NIL then
4:
Remove x from Children(Parent(x))
Parent(x) NIL
5:
return Merge(H, x)
The following ANSI C program translates this algorithm.
struct node decrease_key(struct node h, struct node x, Key key){
xkey = key; / Assume key xkey /
if(xparent)
xparentchildren = remove_node(xparentchildren, x);
xparent = NULL;
return merge(h, x);
}

Exercise 10.10
Implement the program of removing a node from the children of its parent
in your favorite imperative programming language. Consider how can we ensure
the overall performance of decreasing key is O(1) time? Is left-child, right sibling
approach enough?
Delete the minimum element from the heap (pop)
Since the minimum element is always stored at root, after delete it during popping, the rest things left are all sub-trees. These trees can be merged to one big
tree.
pop(H) = mergeP airs(Children(H))

(10.24)

Pairing Heap uses a special approach that it merges every two sub-trees
from left to right in pair. Then merge these paired results from right to left
which forms a final result tree. The name of Pairing Heap comes from the
characteristic of this pair-merging.
Figure 10.16 and 10.17 illustrate the procedure of pair-merging.
The recursive pair-merging solution is quite similar to the bottom up merge
sort[6]. Denote the children of a pairing heap as A, which is a list of trees of
{T1 , T2 , T3 , ..., Tm } for example. The mergeP airs() function can be given as

10.4. PAIRING HEAPS

285

(a) A pairing heap before pop.

(b) After root element 2 being removed, there are 9 sub-trees left.

Figure 10.16: Remove the root element, and merge children in pairs.

286CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

(a) Merge tree with 9, and tree with root 6.

(b) Merge tree with root 7 to the result.

(c) Merge tree with root 3 to the result.

(d) Merge tree with root 4 to the result.

Figure 10.17: Steps of merge from right to left.

10.4. PAIRING HEAPS

287

below.

mergeP airs(A) =

: A=
T1 : A = {T1 }

merge(merge(T1 , T2 ), mergeP airs(A0 )) : otherwise

(10.25)

where
A0 = {T3 , T4 , ..., Tm }
is the rest of the children without the first two trees.
The relative Haskell program of popping is given as the following.
deleteMin :: (Ord a) PHeap a PHeap a
deleteMin (Node _ hs) = mergePairs hs where
mergePairs [] = E
mergePairs [h] = h
mergePairs (h1:h2:hs) = merge (merge h1 h2) (mergePairs hs)

The popping operation can also be explained in the following procedural

algorithm.
1: function Pop(H)
2:
L N IL
3:
for every 2 trees Tx , Ty Children(H) from left to right do
4:
Extract x, and y from Children(H)
5:
T Merge(Tx , Ty )
6:
Insert T at the beginning of L
7:
H Children(H)
. H is either N IL or one tree.
8:
for T L from left to right do
9:
H Merge(H, T )
10:
return H
Note that L is initialized as an empty linked-list, then the algorithm iterates
every two trees in pair in the children of the K-ary tree, from left to right, and
performs merging, the result is inserted at the beginning of L. Because we insert
to front end, so when we traverse L later on, we actually process from right to
left. There may be odd number of sub-trees in H, in that case, it will leave one
tree after pair-merging. We handle it by start the right to left merging from
this left tree.
Below is the ANSI C program to this algorithm.
struct node pop(struct node h){
struct node x, y, lst = NULL;
while((x = hchildren) != NULL){
if((hchildren = y = xnext) != NULL)
hchildren = hchildrennext;
lst = push_front(lst, merge(x, y));
}
x = NULL;
while((y = lst) != NULL){
lst = lstnext;
x = merge(x, y);
}

288CHAPTER 10. BINOMIAL HEAP, FIBONACCI HEAP, AND PAIRING HEAP

free(h);
return x;
}

The pairing heap pop operation is conjectured to be amortized O(lg N ) time

[7].

Exercise 10.11
Write a program to insert a tree at the beginning of a linked-list in your
favorite imperative programming language.
Delete a node
We didnt mention delete in Binomial heap or Fibonacci Heap. Deletion can be
realized by first decreasing key to minus infinity (), then performing pop.
In this section, we present another solution for delete node.
The algorithm is to define the function delete(H, x), where x is a node in a
pairing heap H 6 .
If x is root, we can just perform a pop operation. Otherwise, we can cut x
from H, perform a pop on x, and then merge the pop result back to H. This
can be described as the following.

delete(H, x) =

pop(H) : x is root of
merge(cut(H, x), pop(x)) : otherwise

(10.26)

As delete algorithm uses pop, the performance is conjectured to be amortized

O(1) time.

Exercise 10.12
Write procedural pseudo code for delete algorithm.
Write the delete operation in your favorite imperative programming language
Consider how to realize delete in purely functional setting.

10.5

Notes and short summary

In this chapter, we extend the heap implementation from binary tree to more
generic approach. Binomial heap and Fibonacci heap use Forest of K-ary trees
as under ground data structure, while Pairing heap use a K-ary tree to represent
heap. Its a good point to post pone some expensive operation, so that the over
all amortized performance is ensured. Although Fibonacci Heap gives good
performance in theory, the implementation is a bit complex. It was removed in
some latest textbooks. We also present pairing heap, which is easy to realize
and have good performance in practice.
The elementary tree based data structures are all introduced in this book.
There are still many tree based data structures which we cant covers them all
and skip here. We encourage the reader to refer to other textbooks about them.
From next chapter, well introduce generic sequence data structures, array and
queue.
6 Here

the semantic of x is a reference to a node.

Bibliography
[1] K-ary tree, Wikipedia. http://en.wikipedia.org/wiki/K-ary tree
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.
[3] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[4] Wikipedia, Pascals triangle. http://en.wikipedia.org/wiki/Pascals triangle
[5] Hackage. An alternate implementation of a priority queue based on
a Fibonacci heap., http://hackage.haskell.org/packages/archive/pqueuemtl/1.0.7/doc/html/src/Data-Queue-FibQueue.html
[6] Chris Okasaki. Fibonacci Heaps. http://darcs.haskell.org/nofib/gc/fibheaps/orig
[7] Michael L. Fredman, Robert Sedgewick, Daniel D. Sleator, and Robert E.
Tarjan. The Pairing Heap: A New Form of Self-Adjusting Heap Algorithmica (1986) 1: 111-129.

289

290

BIBLIOGRAPHY

Part IV

Queues and Sequences

291

Chapter 11

Queue, not so simple as it

was thought
11.1

Introduction

It seems that queues are relative simple. A queue provides FIFO (first-in, firstout) data manipulation support. There are many options to realize queue includes singly linked-list, doubly linked-list, circular buffer etc. However, well
show that its not so easy to realize queue in purely functional settings if it must
satisfy abstract queue properties.
In this chapter, well present several different approaches to implement queue.
And in next chapter, well explain how to realize sequence.
A queue is a FIFO data structure satisfies the following performance constraints.
Element can be added to the tail of the queue in O(1) constant time;
Element can be removed from the head of the queue in O(1) constant
time.
These two properties must be satisfied. And its common to add some extra
goals, such as dynamic memory allocation etc.
Of course such abstract queue interface can be implemented with doublylinked list trivially. But this is a overkill solution. We can even implement
imperative queue with singly linked-list or plain array. However, our main
question here is about how to realize a purely functional queue as well?
Well first review the typical queue solution which is realized by singly linkedlist and circular buffer in first section; Then we give a simple and straightforward
functional solution in the second section. While the performance is ensured in
terms of amortized constant time, we need find real-time solution (or worst-case
solution) for some special case. Such solution will be described in the third
and the fourth section. Finally, well show a very simple real-time queue which
depends on lazy evaluation.
Most of the functional contents are based on Chris, Okasakis great work in
[6]. There are more than 16 different types of purely functional queue given in
that material.
293

294

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

11.2

Queue by linked-list and circular buffer

11.2.1

Singly linked-list solution

Queue can be implemented with singly linked-list. Its easy to add and remove
element at the front end of a linked-list in O(1) time. However, in order to
keep the FIFO order, if we execute one operation on head, we must perform the
inverse operation on tail.
For plain singly linked-list, we must traverse the whole list before adding or
removing. Traversing is bound to O(N ) time, where N is the length of the list.
This doesnt match the abstract queue properties.
The solution is to use an extra record to store the tail of the linked-list. A
sentinel is often used to simplify the boundary handling. The following ANSI
C 1 code defines a queue realized by singly linked-list.
typedef int Key;
struct Node{
Key key;
struct Node next;
};
struct Queue{
struct Node head, tail;
};

Figure 11.1 illustrates an empty list. Both head and tail point to the sentinel
NIL node.

head

tail

S
Figure 11.1: The empty queue, both head and tail point to sentinel node.
We summarize the abstract queue interface as the following.
function Empty
. Create an empty queue
function Empty?(Q)
. Test if Q is empty
function Enqueue(Q, x)
. Add a new element x to queue Q
function Dequeue(Q)
. Remove element from queue Q
function Head(Q)
. get the next element in queue Q in FIFO order
1 Its possible to parameterize the type of the key with C++ template. ANSI C is used
here for illustration purpose.

11.2. QUEUE BY LINKED-LIST AND CIRCULAR BUFFER

295

Note the difference between Dequeue and Head. Head only retrieve next
element in FIFO order without removing it, while Dequeue performs removing.
In some programming languages, such as Haskell, and most object-oriented
languages, the above abstract queue interface can be ensured by some definition.
For example, the following Haskell code specifies the abstract queue.
class Queue q where
empty :: q a
isEmpty :: q a Bool
push :: q a a q a -- aka snoc or append, or push_back
pop :: q a q a -- aka tail or pop_front
front :: q a a -- aka head

To ensure the constant time Enqueue and Dequeue, we add new element
to head and remove element from tail.2
function Enqueue(Q, x)
p Create-New-Node
Key(p) x
Next(p) N IL
Next(Tail(Q)) p
Tail(Q) p
Note that, as we use the sentinel node, there are at least one node, the
sentinel in the queue. Thats why we neednt check the validation of of the tail
before we append the new created node p to it.
function Dequeue(Q)
x Head(Q)
Next(Head(Q)) Next(x)
if x = Tail(Q) then
. Q gets empty
Tail(Q) Head(Q)
return Key(x)
As we always put the sentinel node in front of all the other nodes, function
Head actually returns the next node to the sentinel.
Figure 11.2 illustrates Enqueue and Dequeue process with sentinel node.
Translating the pseudo code to ANSI C program yields the below code.
struct Queue enqueue(struct Queue q, Key x){
struct Node p = (struct Node)malloc(sizeof(struct Node));
pkey = x;
pnext = NULL;
qtailnext = p;
qtail = p;
return q;
}
Key dequeue(struct Queue q){
struct Node p = head(q); /gets the node next to sentinel/
Key x = key(p);
qheadnext = pnext;
if(qtail == p)
2 Its possible to add new element to the tail, while remove element from head, but the
operations are more complex than this approach.

296

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

head

tail

NIL

Enqueue
Sentinel

...

NIL

(a) Before Enqueue x to queue

head

tail

Sentinel

...

NIL

(b) After Enqueue x to queue

head

tail

Sentinel

...

b
Dequeue

(c) Before Dequeue x to queue

head

Sentinel

tail

...

NIL

(d) After Dequeue x to queue

Figure 11.2: Enqueue and Dequeue to linked-list queue.

NIL

11.2. QUEUE BY LINKED-LIST AND CIRCULAR BUFFER

297

qtail = qhead;
free(p);
return x;
}

This solution is simple and robust. Its easy to extend this solution even to
the concurrent environment (e.g. multicores). We can assign a lock to the head
and use another lock to the tail. The sentinel helps us from being dead-locked
due to the empty case [1] [2].

Exercise 11.1
Realize the Empty? and Head algorithms for linked-list queue.
Implement the singly linked-list queue in your favorite imperative programming language. Note that you need provide functions to initialize
and destroy the queue.

11.2.2

Circular buffer solution

Another typical solution to realize queue is to use plain array as a circular buffer
(also known as ring buffer). Oppose to linked-list, array support appending to
the tail in constant O(1) time if there are still spaces. Of course we need reallocate spaces if the array is fully occupied. However, Array performs poor
in O(N ) time when removing element from head and packing the space. This
is because we need shift all rest elements one cell ahead. The idea of circular
buffer is to reuse the free cells before the first valid element after we remove
elements from head.
The idea of circular buffer can be described in figure 11.3 and 11.4.
If we set a maximum size of the buffer instead of dynamically allocate memories, the queue can be defined with the below ANSI C code.
struct Queue{
Key buf;
int head, tail, size;
};

When initialize the queue, we are explicitly asked to provide the maximum
size as argument.
struct Queue createQ(int max){
struct Queue q = (struct Queue)malloc(sizeof(struct Queue));
qbuf = (Key)malloc(sizeof(Key)max);
qsize = max;
qhead = qtail = 0;
return q;
}

To test if a queue is empty is trivial.

function Empty?(Q)
return Head(Q) = Tail(Q)
One brute-force implementation for Enqueue and Dequeue is to calculate
the modular of index blindly as the following.
function Enqueue(Q, x)

298

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

head

tail

a[0]

a[1]

...

boundery

a[i]

...

(a) Continuously add some elements.

head

...

a[j]

tail

...

boundery

a[i]

...

(b) After remove some elements from head, there are free
cells.

head

...

a[j]

tail

...

boundery

a[i]

(c) Go on adding elements till the boundary of the array.

tail

a[0]

head

...

a[j]

boundery

...

(d) The next element is added to the first free cell

on head.

tail

a[0]

a[1]

...

a[j-1]

head

a[j]

boundery

...

(e) All cells are occupied. The queue is full.

Figure 11.3: A queue is realized with ring buffer.

11.2. QUEUE BY LINKED-LIST AND CIRCULAR BUFFER

299

Figure 11.4: The circular buffer.

if Full?(Q) then
Tail(Q) (Tail(Q) + 1)
Buffer(Q)[Tail(Q)] x

mod Size(Q)

300

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

Exercise 11.2
As the circular buffer is allocated with a maximum size parameter, please
write a function to test if a queue is full to avoid overflow. Note there are two
cases, one is that the head is in front of the tail, the other is on the contrary.

11.3

Purely functional solution

11.3.1

Paired-list queue

We cant just use a list to implement queue, or we cant satisfy abstract queue
properties. This is because singly linked-list, which is the back-end data structure in most functional settings, performs well on head in constant O(1) time,
while it performs in linear O(N ) time on tail, where N is the length of the list.
Either dequeue or enqueue will perform proportion to the number of elements
stored in the list as shown in figure 11.5.
EnQueue O(1)

x[N]

x[N-1]

...

x[2]

x[1]

NIL

DeQueue O(N)

x[1]

NIL

DeQueue O(1)

(a) DeQueue performs poorly.

EnQueue O(N)

x[N]

x[N-1]

...

x[2]

(b) EnQueue performs poorly.

Figure 11.5: DeQueue and EnQueue cant perform both in constant O(1)
time with a list.
We neither can add a pointer to record the tail position of the list as what
we have done in the imperative settings like in the ANSI C program, because
of the nature of purely functional.
Chris Okasaki mentioned a simple and straightforward functional solution in
[6]. The idea is to maintain two linked-lists as a queue, and concatenate these
two lists in a tail-to-tail manner. The shape of the queue looks like a horseshoe
magnet as shown in figure 11.6.
With this setup, we push new element to the head of the rear list, which is
ensure to be O(1) constant time; on the other hand, we pop element from the
head of the front list, which is also O(1) constant time. So that the abstract
queue properties can be satisfied.
The definition of such paired-list queue can be expressed in the following
Haskell code.
type Queue a = ([a], [a])
empty = ([], [])

Suppose function f ront(Q) and rear(Q) return the front and rear list in
such setup, and Queue(F, R) create a paired-list queue from two lists F and R.
The EnQueue (push) and DeQueue (pop) operations can be easily realized
based on this setup.
push(Q, x) = Queue(f ront(Q), {x} rear(Q))

(11.1)

11.3. PURELY FUNCTIONAL SOLUTION

301

(a) a horseshoe magnet.

front

DeQueue O(1)

x[N]

x[N-1]

...

x[2]

x[1]

NIL

EnQueue O(1)

y[M]

y[M-1]

...

y[2]

y[1]

NIL

rear

(b) concatenate two lists tail-to-tail.

Figure 11.6: A queue with front and rear list shapes like a horseshoe magnet.

302

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

pop(Q) = Queue(tail(f ront(Q)), rear(Q))

(11.2)

where if a list X = {x1 , x2 , ..., xn }, function tail(X) = {x2 , x3 , ..., xn } returns

the rest of the list without the first element.
However, we must next solve the problem that after several pop operations,
the front list becomes empty, while there are still elements in rear list. One
method is to rebuild the queue by reversing the rear list, and use it to replace
front list.
Hence a balance operation will be execute after popping. Lets denote the
front and rear list of a queue Q as F = f ront(Q), and R = f ear(Q).

balance(F, R) =

Queue(reverse(R), ) : F =
Q : otherwise

(11.3)

Thus if front list isnt empty, we do nothing, while when the front list becomes empty, we use the reversed rear list as the new front list, and the new
rear list is empty.
The new enqueue and dequeue algorithms are updated as below.
push(Q, x) = balance(F, {x} R)

(11.4)

pop(Q) = balance(tail(F ), R)

(11.5)

Sum up the above algorithms and translate them to Haskell yields the following program.
balance :: Queue a Queue a
balance ([], r) = (reverse r, [])
balance q = q
push :: Queue a a Queue a
push (f, r) x = balance (f, x:r)
pop :: Queue a Queue a
pop ([], _) = error "Empty"
pop (_:f, r) = balance (f, r)

However, although we only touch the heads of front list and rear list, the
overall performance cant be kept always as O(1). Actually, the performance of
this algorithm is amortized O(1). This is because the reverse operation takes
time proportion to the length of the rear list. its bound O(N ) time, where
N = |R|. We left the prove of amortized performance as an exercise to the
reader.

11.3.2

Paired-array queue - a symmetric implementation

There is an interesting implementation which is symmetric to the paired-list

queue. In some old programming languages, such as legacy version of BASIC,
There is array supported, but there is no pointers, nor records to represent
linked-list. Although we can use another array to store indexes so that we

11.3. PURELY FUNCTIONAL SOLUTION

303

can represent linked-list with implicit array, there is another option to realized
amortized O(1) queue.
Compare the performance of array and linked-list. Below table reveals some
facts (Suppose both contain N elements).
operation
Array Linked-list
insert on head
O(N )
O(1)
insert on tail
O(1)
O(N )
O(1)
remove on head O(N )
remove on tail
O(1)
O(N )
Note that linked-list performs in constant time on head, but in linear time
on tail; while array performs in constant time on tail (suppose there is enough
memory spaces, and omit the memory reallocation for simplification), but in
linear time on head. This is because we need do shifting when prepare or
eliminate an empty cell in array. (see chapter the evolution of insertion sort
for detail.)
The above table shows an interesting characteristic, that we can exploit it
and provide a solution mimic to the paired-list queue: We concatenate two
arrays, head-to-head, to make a horseshoe shape queue like in figure 11.7.

front array

x[1]

x[2]

...

x[N-1]

x[N]

EnQueue O(1)

y[1]

y[2]

...

y[M-1]

y[M]

DeQueue O(1)

rear array

(a) a horseshoe magnet.

(b) concatenate two arrays head-to-head.

Figure 11.7: A queue with front and rear arrays shapes like a horseshoe magnet.
We can define such paired-array queue like the following Python code

class Queue:
def __init__(self):
self.front = []
self.rear = []
def is_empty(q):
return q.front == [] and q.rear == []

The relative Push() and Pop() algorithm only manipulate on the tail of the
arrays.
function Push(Q, x)
3 Legacy Basic code is not presented here. And we actually use list but not array in Python
to illustrate the idea. ANSI C and ISO C++ programs are provides along with this chapter,
they show more in a purely array manner.

304

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

Append(Rear(Q), x)

Here we assume that the Append() algorithm append element x to the end
of the array, and handle the necessary memory allocation etc. Actually, there
are multiple memory handling approaches. For example, besides the dynamic
re-allocation, we can initialize the array with enough space, and just report error
if its full.
function Pop(Q)
if Front(Q) = then
Front(Q) Reverse(Rear(Q))
Rear(Q)
N Length(Front(Q))
x Front(Q)[N]
Length(Front(Q)) N 1
return x
For simplification and pure illustration purpose, the array isnt shrunk explicitly after elements removed. So test if front array is empty () can be
realized as check if the length of the array is zero. We omit all these details
here.
The enqueue and dequeue algorithms can be translated to Python programs
straightforwardly.
def push(q, x):
q.rear.append(x)
def pop(q):
if q.front == []:
q.rear.reverse()
(q.front, q.rear) = (q.rear, [])
return q.front.pop()

Similar to the paired-list queue, the performance is amortized O(1) because

the reverse procedure takes linear time.

Exercise 11.3
Prove that the amortized performance of paired-list queue is O(1).
Prove that the amortized performance of paired-array queue is O(1).

11.4

A small improvement, Balanced Queue

Although paired-list queue is amortized O(1) for popping and pushing, the solution we proposed in previous section performs poor in the worst case. For
example, there is one element in the front list, and we push N elements continuously to the queue, here N is a big number. After that executing a pop
operation will cause the worst case.
According to the strategy we used so far, all the N elements are added to
the rear list. The front list turns to be empty after a pop operation. So the
algorithm starts to reverse the rear list. This reversing procedure is bound to
O(N ) time, which is proportion to the length of the rear list. Sometimes, it
cant be acceptable for a very big N .

11.4. A SMALL IMPROVEMENT, BALANCED QUEUE

305

The reason why this worst case happens is because the front and rear lists are
extremely unbalanced. We can improve our paired-list queue design by making
them more balanced. One option is to add a balancing constraint.
|R| |F |

(11.6)

Where R = Rear(Q), F = F ront(Q), and |L| is the length of list L. This

constraint ensure the length of the rear list is less than the length of the front
list. So that the reverse procedure will be executed once the rear list grows
longer than the front list.
Here we need frequently access the length information of a list. However,
calculate the length takes linear time for singly linked-list. We can record the
length to a variable and update it as adding and removing elements. This
approach enables us to get the length information in constant time.
Below example shows the modified paired-list queue definition which is augmented with length fields.
data BalanceQueue a = BQ [a] Int [a] Int

As we keep the invariant as specified in (11.6), we can easily tell if a queue

is empty by testing the length of the front list.
F = |F | = 0

(11.7)

In the rest part of this section, we suppose the length of a list L, can be
retrieved as |L| in constant time.
Push and pop are almost as same as before except that we check the balance
invariant by passing length information and performs reversing accordingly.
push(Q, x) = balance(F, |F |, {x} R, |R| + 1)

(11.8)

pop(Q) = balance(tail(F ), |F | 1, R, |R|)

(11.9)

Where function balance() is defined as the following.

Queue(F, |F |, R, |R|) : |R| |F |

Queue(F reverse(R), |F | + |R|, , 0) : otherwise
(11.10)
Note that the function Queue() takes four parameters, the front list along
with its length (recorded), and the rear list along with its length, and forms a
paired-list queue augmented with length fields.
We can easily translate the equations to Haskell program. And we can
enforce the abstract queue interface by making the implementation an instance
of the Queue type class.
balance(F, |F |, R, |R|) =

instance Queue BalanceQueue where

empty = BQ [] 0 [] 0
isEmpty (BQ _ lenf _ _) = lenf == 0
-- Amortized O(1) time push
push (BQ f lenf r lenr) x = balance f lenf (x:r) (lenr + 1)

306

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

-- Amortized O(1) time pop

pop (BQ (_:f) lenf r lenr) = balance f (lenf - 1) r lenr
front (BQ (x:_) _ _ _) = x
balance f lenf r lenr
| lenr lenf = BQ f lenf r lenr
| otherwise = BQ (f ++ (reverse r)) (lenf + lenr) [] 0

Exercise 11.4
Write the symmetric balance improvement solution for paired-array queue
in your favorite imperative programming language.

11.5

One more step improvement, Real-time Queue

Although the extremely worst case can be avoided by improving the balancing
as what has been presented in previous section, the performance of reversing
rear list is still bound to O(N ), where N = |R|. So if the rear list is very long,
the instant performance is still unacceptable poor even if the amortized time is
O(1). It is particularly important in some real-time system to ensure the worst
case performance.
As we have analyzed, the bottleneck is the computation of F reverse(R).
This happens when |R| > |F |. Considering that |F | and |R| are all integers, so
this computation happens when
|R| = |F | + 1

(11.11)

Both F and the result of reverse(R) are singly linked-list, It takes O(|F |)
time to concatenate them together, and it takes extra O(|R|) time to reverse
the rear list, so the total computation is bound to O(|N |), where N = |F | + |R|.
Which is proportion to the total number of elements in the queue.
In order to realize a real-time queue, we cant computing F reverse(R)
monolithic. Our strategy is to distribute this expensive computation to every
pop and push operations. Thus although each pop and push get a bit slow, we
may avoid the extremely slow worst pop or push case.
Incremental reverse
Lets examine how functional reverse algorithm is implemented typically.

: X=
reverse(X) =
(11.12)
reverse(X 0 ) {x1 } : otherwise
Where X 0 = tail(X) = {x2 , x3 , ...}.
This is a typical recursive algorithm, that if the list to be reversed is empty,
the result is just an empty list. This is the edge case; otherwise, we take the first
element x1 from the list, reverse the rest {x2 , x3 , ..., xn }, to {xn , xn1 , .., x3 , x2 }
and append x1 after it.
However, this algorithm performs poor, as appending an element to the end
of a list is proportion to the length of the list. So its O(N 2 ), but not a linear
time reverse algorithm.

11.5. ONE MORE STEP IMPROVEMENT, REAL-TIME QUEUE

307

There exists another implementation which utilizes an accumulator A, like

below.
reverse(X) = reverse0 (X, )

(11.13)

Where

reverse (X, A) =

A : X=
reverse0 (X 0 , {x1 } A) : otherwise

(11.14)

We call A as accumulator because it accumulates intermediate reverse result

at any time. Every time we call reverse0 (X, A), list X contains the rest of
elements wait to be reversed, and A holds all the reversed elements so far. For
instance when we call reverse0 () at i-th time, X and A contains the following
elements:
X = {xi , xi+1 , ..., xn } A = {xi1 , xi2 , ...x1 }
In every non-trivial case, we takes the first element from X in O(1) time;
then put it in front of the accumulator A, which is again O(1) constant time.
We repeat it N times, so this is a linear time (O(N )) algorithm.
The latter version of reverse is obviously a tail-recursion algorithm, see [5]
and [6] for detail. Such characteristic is easy to change from monolithic algorithm to incremental manner.
The solution is state transferring. We can use a state machine contains two
types of stat: reversing state Sr to indicate that the reverse is still on-going (not
finished), and finish state Sf to indicate the reverse has been done (finished).
In Haskell programming language, it can be defined as a type.
data State a = | Reverse [a] [a]
| Done [a]

And we can schedule (slow-down) the above reverse0 (X, A) function with
these two types of state.

step(S, X, A) =

(Sf , A) :
(Sr , X 0 , {x1 } A) :

S = Sr X =
S = Sr X 6=

(11.15)

Each step, we examine the state type first, if the current state is Sr (ongoing), and the rest elements to be reversed in X is empty, we can turn the
algorithm to finish state Sf ; otherwise, we take the first element from X, put it
in front of A just as same as above, but we do NOT perform recursion, instead,
we just finish this step. We can store the current state as well as the resulted
X and A, the reverse can be continued at any time when we call next step
function in the future with the stored state, X and A passed in.
Here is an example of this step-by-step reverse algorithm.
step(Sr , hello, )
step(Sr , ello, h)
...
step(Sr , o, lleh)
step(Sr , , olleh)

= (Sr , ello, h)
= (Sr , llo, eh)
= (Sr , , olleh)
= (Sf , olleh)

And in Haskell code manner, the example is like the following.

308
step
step
...
step
step

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

$ Reverse "hello" [] = Reverse "ello" "h"
$ Reverse "ello" "h" = Reverse "llo" "eh"
$ Reverse "o" "lleh" = Reverse [] "olleh"
$ Reverse [] "olleh" = Done "olleh"

Now we can distribute the reverse into steps in every pop and push operations. However, the problem is just half solved. We want to break down
F reverse(R), and we have broken reverse(R) into steps, we next need to
schedule(slow-down) the list concatenation part F ..., which is bound to O(|F |),
into incremental manner so that we can distribute it to pop and push operations.
Incremental concatenate
Its a bit more challenge to implement incremental list concatenation than list
reversing. However, its possible to re-use the result we gained from increment

reverse by a small trick: In order to realize X Y , we can first reverse X to X ,

then take elements one by one from X and put them in front of Y just as what
we have done in reverse0 .
X Y

reverse(reverse(X)) Y
reverse0 (reverse(X), ) Y
reverse0 (reverse(X), Y )

reverse0 ( X , Y )

(11.16)

This fact indicates us that we can use an extra state to instruct the step()

function to continuously concatenating F after R is reversed.

The strategy is to do the total work in two phases:

1. Reverse both F and R in parallel to get F = reverse(F ), and R =

reverse(R) incrementally;

2. Incrementally take elements from F and put them in front of R .

So we define three types of state: Sr represents reversing; Sc represents
concatenating; and Sf represents finish.
In Haskell, these types of state are defined as the following.
data State a = Reverse [a] [a] [a] [a]
| Concat [a] [a]
| Done [a]

Because we reverse F and R simultaneously, so reversing state takes two

pairs of lists and accumulators.
The state transfering is defined according to the two phases strategy described previously. Denotes that F = {f1 , f2 , ...}, F 0 = tail(F ) = {f2 , f3 , ...},
R = {r1 , r2 , ...}, R0 = tail(R) = {r2 , r3 , ...}. A state S, contains its type S,
which has the value among Sr , Sc , and Sf . Note that S also contains necessary

parameters such as F , F , X, A etc as intermediate results. These parameters

11.5. ONE MORE STEP IMPROVEMENT, REAL-TIME QUEUE

309

vary according to the different states.

(Sr , F 0 , {f1 } F , R0 , {r1 } R ) : S = Sr F 6= R 6=

(Sc , F , {r1 } R ) : S = Sr F = R = {r1 }

next(S) =

(Sf , A) : S = Sc X =

(Sc , X 0 , {x1 } A) : S = Sc X 6=
(11.17)
The relative Haskell program is list as below.
next
next
next
next

(Reverse (x:f) f (y:r) r) = Reverse f (x:f) r (y:r)

(Reverse [] f [y] r) = Concat f (y:r)
(Concat 0 _ acc) = Done acc
(Concat (x:f) acc) = Concat f (x:acc)

All left to us is to distribute these incremental steps into every pop and push
operations to implement a real-time O(1) purely functional queue.
Sum up
Before we dive into the final real-time queue implementation. Lets analyze
how many incremental steps are taken to achieve the result of F reverse(R).
According to the balance variant we used previously, |R| = |F |+1, Lets denotes
M = |F |.
Once the queue gets unbalanced due to some push or pop operation, we start
this incremental F reverse(R). It needs M + 1 steps to reverse R, and at the
same time, we finish reversing the list F within these steps. After that, we need
extra M + 1 steps to execute the concatenation. So there are 2M + 2 steps.
It seems that distribute one step inside one pop or push operation is the
natural solution, However, there is a critical question must be answered: Is it
possible that before we finish these 2M + 2 steps, the queue gets unbalanced
again due to a series push and pop?
There are two facts about this question, one is good news and the other is
bad news.
Lets first show the good news, that luckily, continuously pushing cant make
the queue unbalanced again before we finish these 2M + 2 steps to achieve
F reverse(R). This is because once we start re-balancing, we can get a new
front list F 0 = F reverse(R) after 2M +2 steps. While the next time unbalance
is triggered when
|R0 | = |F 0 | + 1
= |F | + |R| + 1
= 2M + 2

(11.18)

That is to say, even we continuously pushing as mush elements as possible

after the last unbalanced time, when the queue gets unbalanced again, the
2M + 2 steps exactly get finished at that time point. Which means the new
front list F 0 is calculated OK. We can safely go on to compute F 0 reverse(R0 ).
Thanks to the balance invariant which is designed in previous section.
But, the bad news is that, pop operation can happen at anytime before these
2M + 2 steps finish. The situation is that once we want to extract element from
front list, the new front list F 0 = F reverse(R) hasnt been ready yet. We
dont have a valid front list at hand.

310

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

front copy
{fi , fi+1 , ..., fM }
first i 1 elements popped

on-going computation
...)
(Sr , F , ..., R,

intermediate result F and R

new rear
{...}
new elements pushed

Table 11.1: Intermediate state of a queue before first M steps finish.

One solution to solve this problem is to keep a copy of original front list
F , during the time we are calculating reverse(F ) which is described in phase
1 of our incremental computing strategy. So that we are still safe even if user
continuously performs first M pop operations. So the queue looks like in table
11.1 at some time after we start the incremental computation and before phase
1 (reverse F and R simultaneously) ending4 .
After these M pop operations, the copy of F is exhausted. And we just start
incremental concatenation phase at that time. What if user goes on popping?
The fact is that since F is exhausted (becomes ), we neednt do concate

nation at all. Since F R = R = R .

It indicates us, when doing concatenation, we only need to concatenate those
elements havent been popped, which are still left in F . As user pops elements
one by one continuously from the head of front list F , one method is to use a
counter, record how many elements there are still in F . The counter is initialized
as 0 when we start computing F reverse(R), its increased by one when we
reverse one element in F , which means we need concatenate thie element in the
future; and its decreased by one every time when pop is performed, which means
we can concatenate one element less; of course we need decrease this counter as
well in every steps of concatenation. If and only if this counter becomes zero,
we neednt do concatenations any more.
We can give the realization of purely functional real-time queue according
to the above analysis.
We first add an idle state S0 to simplify some state transfering. Below
Haskell program is an example of this modified state definition.
data State a =
|
|
|

Empty
Reverse Int [a] [a] [a] [a] -- n, f, acc_f r, acc_r
Append Int [a] [a]
-- n, rev_f, acc
Done [a] -- result: f ++ reverse r

And the data structure is defined with three parts, the front list (augmented
with length); the on-going state of computing F reverse(R); and the rear list
(augmented with length).
Here is the Haskell definition of real-time queue.
data RealtimeQueue a = RTQ [a] Int (State a) [a] Int

The empty queue is composed with empty front and rear list together with
idle state S0 as Queue(, 0, S0 , , 0). And we can test if a queue is empty by
4 One may wonder that copying a list takes linear time to the length of the list. If so
the whole solution would make no sense. Actually, this linear time copying wont happen
at all. This is because the purely functional nature, the front list wont be mutated either
by popping or by reversing. However, if trying to realize a symmetric solution with pairedarray and mutate the array in-place, this issue should be stated, and we can perform a lazy
copying, that the real copying work wont execute immediately, instead, it copies one element
every step we do incremental reversing. The detailed implementation is left as an exercise.

11.5. ONE MORE STEP IMPROVEMENT, REAL-TIME QUEUE

311

checking if |F | = 0 according to the balance invariant defined before. Push and

pop are changed accordingly.
push(Q, x) = balance(F, |F |, S, {x} R, |R| + 1)

(11.19)

pop(Q) = balance(F 0 , |F | 1, abort(S), R, |R|)

(11.20)

The major difference is abort() function. Based on our above analysis, when
there is popping, we need decrease the counter, so that we can concatenate
one element less. We define this as aborting. The details will be given after
balance() function.
The relative Haskell code for push and pop are listed like this.
push (RTQ f lenf s r lenr) x = balance f lenf s (x:r) (lenr + 1)
pop (RTQ (_:f) lenf s r lenr) = balance f (lenf - 1) (abort s) r lenr

The balance() function first check the balance invariant, if its violated, we
need start re-balance it by starting compute F reverse(R) incrementally;
otherwise we just execute one step of the unfinished incremental computation.

step(F, |F |, S, R, |R|) : |R| |F |

step(F, |F | + |R|, (Sr , 0, F, , R, ), 0) : otherwise
(11.21)
The relative Haskell code is given like below.

balance(F, |F |, S, R, |R|) =

balance f lenf s r lenr

| lenr lenf = step f lenf s r lenr
| otherwise = step f (lenf + lenr) (Reverse 0 f [] r []) [] 0

The step() function typically transfer the state machine one state ahead, and
it will turn the state to idle (S0 ) when the incremental computation finishes.

step(F, |F |, S, R, |R|) =

Queue(F 0 , |F |, S0 , R, |R|) : S 0 = Sf
Queue(F, |F |, S 0 , R, |R|) : otherwise

(11.22)

Where S 0 = next(S) is the next state transferred; F 0 = F reverse(R), is

the final new front list result from the incremental computing. The real state
transferring is implemented in next() function as the following. Its different
from previous version by adding the counter field n to record how many elements
left we need to concatenate.

(Sr , n + 1, F 0 , {f1 } F , R0 , {r1 } R ) : S = Sr F 6=

(Sc , n, F , {r1 } R ) : S = Sr F =
next(S) =
(Sf , A) : S = Sc n = 0

(S
,
n

1,
X
,
{x
} A) : S = Sc n 6= 0

c
1

S : otherwise
(11.23)
And the corresponding Haskell code is like this.

312

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

next
next
next
next
next

(Reverse n (x:f)
(Reverse n [] f
(Concat 0 _ acc)
(Concat n (x:f)
s=s

f (y:r) r) = Reverse (n+1) f (x:f) r (y:r)

[y] r) = Concat n f (y:r)
= Done acc
acc) = Concat (n-1) f (x:acc)

Function abort() is used to tell the state machine, we can concatenate one
element less since it is popped.

(Sf , A0 ) :
(Sc , n 1, X 0 A) :
abort(S) =

(Sr , n 1, F, F , R, R ) :

S :

S = Sc n = 0
S = Sc n 6= 0
S = Sr
otherwise

(11.24)

Note that when n = 0 we actually rollback one concatenated element by

return A0 as the result but not A. (Why? this is left as an exercise.)
The Haskell code for abort function is like the following.
abort
abort
abort
abort

(Concat 0 _ (_:acc)) = Done acc -- Note! we rollback 1 elem

(Concat n f acc) = Concat (n-1) f acc
(Reverse n f f r r) = Reverse (n-1) f f r r
s=s

It seems that weve done, however, there is still one tricky issue hidden
behind us. If we push an element x to an empty queue, the result queue will
be:
Queue(, 1, (Sc , 0, , {x}), , 0)
If we perform pop immediately, well get an error! We found that the front
list is empty although the previous computation of F reverse(R) has been
finished. This is because it takes one more extra step to transfer from the state
(Sc , 0, , A) to (Sf , A). Its necessary to refine the S 0 in step() function a bit.

next(next(S)) : F =
S0 =
(11.25)
next(S) : otherwise
The modification reflects to the below Haskell code:
step f lenf s r lenr =
case s of
Done f RTQ f lenf Empty r lenr
s RTQ f lenf s r lenr
where s = if null f then next $ next s else next s

Note that this algorithm differs from the one given by Chris Okasaki in
[6]. Okasakis algorithm executes two steps per pop and push, while the one
presents in this chapter executes only one per pop and push, which leads to
more distributed performance.

Exercise 11.5
Why need we rollback one element when n = 0 in abort() function?
Realize the real-time queue with symmetric paired-array queue solution
in your favorite imperative programming language.

11.6. LAZY REAL-TIME QUEUE

313

In the footnote, we mentioned that when we start incremental reversing

with in-place paired-array solution, copying the array cant be done monolithic or it will lead to linear time operation. Implement the lazy copying
so that we copy one element per step along with the reversing.

11.6

Lazy real-time queue

The key to realize a real-time queue is to break down the expensive F

reverse(R) to avoid monolithic computation. Lazy evaluation is particularly
helpful in such case. In this section, well explore if there is some more elegant
solution by exploit laziness.
Suppose that there exits a function rotate(), which can compute F reverse(R)
incrementally. thats to say, with some accumulator A, the following two functions are equivalent.
rotate(X, Y, A) X reverse(Y ) A

(11.26)

Where we initialized X as the front list F , Y as the rear list R, and the
accumulator A is initialized as empty .
The trigger of rotation is still as same as before when |F | + 1 = |R|. Lets
keep this constraint as an invariant during the whole rotation process, that
|X| + 1 = |Y | always holds.
Its obvious to deduce to the trivial case:
rotate(, {y1 }, A) = {y1 } A

(11.27)

Denote X = {x1 , x2 , ...}, Y = {y1 , y2 , ...}, and X 0 = {x2 , x3 , ...}, Y 0 =

{y2 , y3 , ...} are the rest of the lists without the first element for X and Y respectively. The recursion case is ruled out as the following.
rotate(X, Y, A)

X reverse(Y ) A
{x1 } (X 0 reverse(Y ) A)
{x1 } (X 0 reverse(Y 0 ) ({y1 } A))
{x1 } rotate(X 0 , Y 0 , {y1 } A)

Definition of (11.26)
Associative of
Nature of reverse and associative of
Definition of (11.26)
(11.28)
Summarize the above two cases, yields the final incremental rotate algorithm.

rotate(X, Y, A) =

{y1 } A : X =
{x1 } rotate(X 0 , Y 0 , {y1 } A) : otherwise

(11.29)

If we execute lazily instead of strictly, that is, execute once pop or push
operation is performed, the computation of rotate can be distribute to push and
pop naturally.
Based on this idea, we modify the paired-list queue definition to change the
front list to a lazy list, and augment it with a computation stream. [5]. When
the queue triggers re-balance constraint by some pop/push, that |F | + 1 = |R|,
The algorithm creates a lazy rotation computation, then use this lazy rotation
as the new front list F 0 ; the new rear list becomes , and a copy of F 0 is
maintained as a stream.

314

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

After that, when we performs every push and pop; we consume the stream
by forcing a operation. This results us advancing one step along the stream,
{x} F 00 , where F 0 = tail(F 0 ). We can discard x, and replace the stream F 0
with F 00 .
Once all of the stream is exhausted, we can start another rotation.
In order to illustrate this idea clearly, we turns to Scheme/Lisp programming
language to show example codes, because it gives us explicit control of laziness.
In Scheme/Lisp, we have the following three tools to deal with lazy stream.
(define (cons-stream a b) (cons a (delay b)))
(define stream-car car)
(define (stream-cdr s) (cdr (force s)))

So cons-stream constructs a lazy list from an element x and an existing list

L without really evaluating the value of L; The evaluation is actually delayed
to stream-cdr, where the computation is forced. delaying can be realized by
lambda calculus, please refer to [5] for detail.
The lazy paired-list queue is defined as the following.
(define (make-queue f r s)
(list f r s))
;; Auxiliary functions
(define (front-lst q) (car q))
(define (rear-lst q) (cadr q))
(define (rots q) (caddr q))

A queue is consist of three parts, a front list, a rear list, and a stream which
represents the computation of F reverse(R). Create an empty queue is trivial
as making all these three parts null.
(define empty (make-queue () () ()))

Note that the front-list is also lazy stream actually, so we need use stream
related functions to manipulate it. For example, the following function test if
the queue is empty by checking the front lazy list stream.
(define (empty? q) (stream-null? (front-lst q)))

The push function is almost as same as the one given in previous section.
That we put the new element in front of the rear list; and then examine the
balance invariant and do necessary balancing works.
push(Q, x) = balance(F, {x} R, Rs )

(11.30)

Where R represents the lazy stream of front list; Rs is the stream of rotation
computation. The relative Scheme/Lisp code is give below.
(define (push q x)
(balance (front-lst q) (cons x (rear q)) (rots q)))

11.6. LAZY REAL-TIME QUEUE

315

While pop is a bit different, because the front list is actually lazy stream,
we need force an evaluation. All the others are as same as before.
pop(Q) = balance(F 0 , R, Rs )

(11.31)

Here F 0 , force one evaluation to F, the Scheme/Lisp code regarding to this

equation is as the following.
(define (pop q)
(balance (stream-cdr (front-lst q)) (rear q) (rots q)))

For illustration purpose, we skip the error handling (such as pop from an
empty queue etc) here.
And one can access the top element in the queue by extract from the front
list stream.
(define (front q) (stream-car (front-lst q)))

The balance function first checks if the computation stream is completely

exhausted, and starts new rotation accordingly; otherwise, it just consumes one
evaluation by enforcing the lazy stream.

Queue(F 0 , , F 0 ) : Rs =
balance(Q) =
(11.32)
Queue(F, R, R0s ) : otherwise
Here F 0 is defined to start a new rotation.
F 0 = rotate(F, R, )

(11.33)

The relative Scheme/Lisp program is listed accordingly.

(define (balance f r s)
(if (stream-null? s)
(let ((newf (rotate f r ())))
(make-queue newf () newf))
(make-queue f r (stream-cdr s))))

The implementation of incremental rotate function is just as same as what

we analyzed above.
(define (rotate xs ys acc)
(if (stream-null? xs)
(cons-stream (car ys) acc)
(cons-stream (stream-car xs)
(rotate (stream-cdr xs) (cdr ys)
(cons-stream (car ys) acc)))))

We used explicit lazy evaluation in Scheme/Lisp. Actually, this program can

be very short by using lazy programming languages, for example, Haskell.
data LazyRTQueue a = LQ [a] [a] [a] -- front, rear, f ++ reverse r
instance Queue LazyRTQueue where
empty = LQ [] [] []
isEmpty (LQ f _ _) = null f
-- O(1) time push

316

CHAPTER 11. QUEUE, NOT SO SIMPLE AS IT WAS THOUGHT

push (LQ f r rot) x = balance f (x:r) rot
-- O(1) time pop
pop (LQ (_:f) r rot) = balance f r rot
front (LQ (x:_) _ _) = x

balance f r [] = let f = rotate f r [] in LQ f [] f

balance f r (_:rot) = LQ f r rot
rotate [] [y] acc = y:acc
rotate (x:xs) (y:ys) acc = x : rotate xs ys (y:acc)

11.7

Notes and short summary

Just as mentioned in the beginning of this book in the first chapter, queue
isnt so simple as it was thought. Weve tries to explain algorithms and data
structures both in imperative and in function approaches; Sometimes, it gives
impression that functional way is simpler and more expressive in most time.
However, there are still plenty of areas, that more studies and works are needed
to give equivalent functional solution. Queue is such an important topic, that
it links to many fundamental purely functional data structures.
Thats why Chris Okasaki made intensively study and took a great amount of
discussions in [6]. With purely functional queue solved, we can easily implement
dequeue with the similar approach revealed in this chapter. As we can handle
elements effectively in both head and tail, we can advance one step ahead to
realize sequence data structures, which support fast concatenate, and finally we
can realize random access data structures to mimic array in imperative settings.
The details will be explained in later chapters.
Note that, although we havent mentioned priority queue, its quite possible
to realized it with heaps. We have covered topic of heaps in several previous
chapters.

Exercise 11.6
Realize dequeue, wich support adding and removing elements on both
sides in constant O(1) time in purely functional way.
Realize dequeue in a symmetric solution only with array in your favorite
imperative programming language.

Bibliography
[1] Maged M. Michael and Michael L. Scott. Simple, Fast, and Practical Non-Blocking and Blocking Concurrent Queue Algorithms.
http://www.cs.rochester.edu/research/synchronization/pseudocode/queues.html
[2] Herb Sutter. Writing a Generalized Concurrent Queue. Dr. Dobbs Oct
29, 2008. http://drdobbs.com/cpp/211601363?pgno=1
[3] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.
[4] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[5] Wikipedia. Tail-call. http://en.wikipedia.org/wiki/Tail call
[6] Wikipedia.
Recursion
(computer
science).
http://en.wikipedia.org/wiki/Recursion (computer science)#Tailrecursive functions
[7] Harold Abelson, Gerald Jay Sussman, Julie Sussman. Structure and Interpretation of Computer Programs, 2nd Edition. MIT Press, 1996, ISBN
0-262-51087-1

317

318

Sequences

Chapter 12

Sequences, The last brick

12.1

Introduction

In the first chapter of this book, which introduced binary search tree as the hello
world data structure, we mentioned that neither queue nor array is simple if
realized not only in imperative way, but also in functional approach. In previous
chapter, we explained functional queue, which achieves the similar performance
as its imperative counterpart. In this chapter, well dive into the topic of arraylike data structures.
We have introduced several data structures in this book so far, and it seems
that functional approaches typically bring more expressive and elegant solution. However, there are some areas, people havent found competitive purely
functional solutions which can match the imperative ones. For instance, the
Ukkonen linear time suffix tree construction algorithm. another examples is
Hashing table. Array is also among them.
Array is trivial in imperative settings, it enables randomly accessing any
elements with index in constant O(1) time. However, this performance target
cant be achieved directly in purely functional settings as there is only list can
be used.
In this chapter, we are going to abstract the concept of array to sequences.
Which support the following features
Element can be inserted to or removed from the head of the sequence
quickly in O(1) time;
Element can be inserted to or removed from the head of the sequence
quickly in O(1) time;
Support concatenate two sequences quickly (faster than linear time);
Support randomly access and update any element quickly;
Support split at any position quickly;
We call these features abstract sequence properties, and it easy to see the
fact that even array (here means plain-array) in imperative settings cant meet
them all at the same time.
319

320

CHAPTER 12. SEQUENCES, THE LAST BRICK

Well provide three solutions in this chapter. Firstly, well introduce a solution based on binary tree forest and numeric representation; Secondly, well
show a concatenate-able list solution; Finally, well give the finger tree solution.
Most of the results are based on Chris, Okasakis work in [6].

12.2

Binary random access list

12.2.1

Review of plain-array and list

Lets review the performance of plain-array and singly linked-list so that we

know how they perform in different cases.
operation
Array
Linked-list
operation on head
O(N )
O(1)
operation on tail
O(1)
O(N )
access at random position
O(1)
average O(N )
average O(N )
O(1)
remove at given position
concatenate
O(N2 )
O(N1 )
Because we hold the head of linked list, operations on head such as insert and
remove perform in constant time; while we need traverse to the end to perform
removing or appending on tail; Given a position i, it need traverse i elements
to access it. Once we are at that position, removing element from there is just
bound to constant time by modifying some pointers. In order to concatenate
two linked-lists, we need traverse to the end of the first one, and link it to the
second one, which is bound to the length of the first linked-list;
On the other hand, for array, we must prepare free cell for inserting a new
element to the head of it, and we need release the first cell after the first element
being removed, all these two operations are achieved by shifting all the rest
elements forward or backward, which costs linear time. While the operations
on the tail of array are trivial constant time. Array also support accessing
random position i by nature; However, removing the element at that position
causes shifting all elements after it one position ahead. In order to concatenate
two arrays, we need copy all elements from the second one to the end of the
first one (ignore the memory re-allocation details), which is proportion to the
length of the second array.
In the chapter about binomial heaps, we have explained the idea of using
forest, which is a list of trees. It brings us the merit that, for any given number
N , by representing it in binary number, we know how many binomial trees need
to hold them. That each bit of 1 represents a binomial tree of that rank of bit.
We can go one step ahead, if we have a N nodes binomial heap, for any given
index 1 < i < N , we can quickly know which binomial tree in the heap holds
the i-th node.

12.2.2

Represent sequence by trees

One solution to realize a random-access sequence is to manage the sequence

with a forest of complete binary trees. Figure 12.1 shows how we attach such
trees to a sequence of numbers.
Here two trees t1 and t2 are used to represent sequence {x1 , x2 , x3 , x4 , x5 , x6 }.
The size of binary tree t1 is 2. The first two elements {x1 , x2 } are leaves of t1 ;

12.2. BINARY RANDOM ACCESS LIST

321

Figure 12.1: A sequence of 6 elements can be represented in a forest.

the size of binary tree t2 is 4. The next four elements {x3 , x4 , x5 , x6 } are leaves
of t2 .
For a complete binary tree, we define the depth as 0 if the tree has only a
leaf. The tree is denoted as as ti if its depth is i + 1. Its obvious that there are
2i leaves in ti .
For any sequence contains N elements, it can be turned to a forest of complete binary trees in this manner. First we represent N in binary number like
below.
N = 20 e0 + 21 e1 + ... + 2M eM

(12.1)

Where ei is either 1 or 0, so N = (eM eM 1 ...e1 e0 )2 . If ei 6= 0, we then need

a complete binary tree with size 2i , For example in figure 12.1, as the length of
sequence is 6, which is (110)2 in binary. The lowest bit is 0, so we neednt a
tree of size 1; the second bit is 1, so we need a tree of size 2, which has depth of
2; the highest bit is also 1, thus we need a tree of size 4, which has depth of 3.
This method represents the sequence {x1 , x2 , ..., xN } to a list of trees {t0 , t1 , ..., tM }
where ti is either empty if ei = 0 or a complete binary tree if ei = 1. We call
this representation as Binary Random Access List [6].
We can reused the definition of binary tree. For example, the following
Haskell program defines the tree and the binary random access list.
data Tree a = Leaf a
| Node Int (Tree a) (Tree a)

-- size, left, right

type BRAList a = [Tree a]

The only difference from the typical binary tree is that we augment the size
information to the tree. This enable us to get the size without calculation at
every time. For instance.

322

CHAPTER 12. SEQUENCES, THE LAST BRICK

size (Leaf _) = 1
size (Node sz _ _) = sz

12.2.3

Insertion to the head of the sequence

The new forest representation of sequence enables many operation effectively.

For example, the operation of inserting a new element y in front of sequence
can be realized as the following.
1. Create a tree t0 , with y as the only one leaf;
2. Examine the first tree in the forest, compare its size with t0 , if its size is
greater than t0 , we just let t0 be the new head of the forest, since the forest
is a linked-list of tree, insert t0 to its head is trivial operation, which is
bound to constant O(1) time;
3. Otherwise, if the size of first tree in the forest is equal to t0 , lets denote
this tree in the forest as ti , we can construct a new binary tree t0i+1 by
linking ti and t0 as its left and right children. After that, we recursively
try to insert t0i+1 to the forest.
Figure 12.2 and 12.3 illustrate the steps of inserting element x1 , x2 , ..., x6 to
an empty tree.

(a) A singleton leaf of x1

(b) Insert x2 . It causes linking, results a tree of height

(c) Insert x3 . the result is two trees, t1 and t2

(d) Insert x4 . It first causes linking

two leafs to a binary tree, then it
performs linking again, which results a final tree of height 2.

Figure 12.2: Steps of inserting elements to an empty list, 1

As there are at most M trees in the forest, and M is bound to O(lg N ), so
the insertion to head algorithm is ensured to perform in O(lg N ) even in worst
case. Well prove the amortized performance is O(1) later.

12.2. BINARY RANDOM ACCESS LIST

323

(a) Insert x5 . The forest is a leaf (t0 ) and

t2 .

(b) Insert x6 . It links two leaf to t1 .

Figure 12.3: Steps of inserting elements to an empty list, 2

Lets formalize the algorithm to equations. we define the function of inserting

an element in front of a sequence as insert(S, x).
insert(S, x) = insertT ree(S, leaf (x))

(12.2)

This function just wrap element x to a singleton tree with a leaf, and call
insertT ree to insert this tree to the forest. Suppose the forest F = {t1 , t2 , ...}
if its not empty, and F 0 = {t2 , t3 , ...} is the rest of trees without the first one.

{t} : F =
{t} F : size(t) < size(t1 )

insertT ree(F 0 , link(t, t1 )) : otherwise

(12.3)
Where function link(t1 , t2 ) create a new tree from two small trees with same
size. Suppose function tree(s, t1 , t2 ) create a tree, set its size as s, makes t1 as
the left child, and t2 as the right child, linking can be realized as below.
insertT ree(F, t) =

link(t1 , t2 ) = tree(size(t1 ) + size(t2 ), t1 , t2 )

(12.4)

The relative Haskell programs can be given by translating these equations.

cons :: a BRAList a BRAList a
cons x ts = insertTree ts (Leaf x)
insertTree :: BRAList a Tree a BRAList a
insertTree [] t = [t]
insertTree (t:ts) t = if size t < size t then t:t:ts
else insertTree ts (link t t)
-- Precondition: rank t1 = rank t2
link :: Tree a Tree a Tree a
link t1 t2 = Node (size t1 + size t2) t1 t2

Here we use the Lisp tradition to name the function that insert an element
before a list as cons.

324

CHAPTER 12. SEQUENCES, THE LAST BRICK

Remove the element from the head of the sequence

Its not complex to realize the inverse operation of cons, which can remove
element from the head of the sequence.
If the first tree in the forest is a singleton leaf, remove this tree from the
forest;
otherwise, we can halve the first tree by unlinking its two children, so the
first tree in the forest becomes two trees, we recursively halve the first tree
until it turns to be a leaf.
Figure 12.4 illustrates the steps of removing elements from the head of the
sequence.

(a) A sequence of 5 elements

(b) Result of removing x5 , the leaf

is removed.

(c) Result of removing x4 , As there is not leaf tree, the tree

is firstly divided into two sub trees of size 2. The first tree is
next divided again into two leafs, after that, the first leaf, which
contains x4 is removed. What left in the forest is a leaf tree of
x3 , and a tree of size 2 with elements x2 , x1 .

Figure 12.4: Steps of removing elements from head

If we assume the sequence isnt empty, so that we can skip the error handling such as trying to remove an element from an empty sequence, this can be
expressed with the following equation. We denote the forest F = {t1 , t2 , ...} and
the trees without the first one as F 0 = {t2 , t3 , ...}

extractT ree(F ) =

(t1 , F 0 ) : t1 is leaf
extractT ree({tl , tr } F 0 ) : otherwise

where {tl , tr } = unlink(t1 ) are the two children of t1 .

It can be translated to Haskell programs like below.

(12.5)

12.2. BINARY RANDOM ACCESS LIST

325

extractTree (t@(Leaf x):ts) = (t, ts)

extractTree (t@(Node _ t1 t2):ts) = extractTree (t1:t2:ts)

With this function defined, its convenient to give head() and tail() functions,
the former returns the first element in the sequence, the latter return the rest.
head(S) = key(f irst(extractT ree(S)))

(12.6)

tail(S) = second(extractT ree(S))

(12.7)

Where function f irst() returns the first element in a paired-value (as known
as tuple); second() returns the second element respectively. function key() is
used to access the element inside a leaf. Below are Haskell programs corresponding to these two equations.
head ts = x where (Leaf x, _) = extractTree ts
tail = snd extractTree

Note that as head and tail functions have already been defined in Haskell
standard library, we given them apostrophes to make them distinct. (another
option is to hide the standard ones by importing. We skip the details as they
are language specific).
Random access the element in binary random access list
As trees in the forest help managing the elements in blocks, giving an arbitrary
index, its easy to locate which tree this element is stored, after that performing
a search in the tree yields the result. As all trees are binary (more accurate,
complete binary tree), the search is essentially binary search, which is bound to
the logarithm of the tree size. This brings us a faster random access capability
than linear search in linked-list setting.
Given an index i, and a sequence S, which is actually a forest of trees, the
algorithm is executed as the following 1 .
1. Compare i with the size of the first tree T1 in the forest, if i is less than
or equal to the size, the element exists in T1 , perform looking up in T1 ;
2. Otherwise, decrease i by the size of T1 , and repeat the previous step in
the rest of the trees in the forest.
This algorithm can be represented as the below equation.

lookupT ree(T1 , i) : i |T1 |
get(S, i) =
get(S 0 , i |T1 |) : otherwise

(12.8)

Where |T | = size(T ), and S 0 = {T2 , T3 , ...} is the rest of trees without the
first one in the forest. Note that we dont handle out of bound error case, this
is left as an exercise to the reader.
Function lookupT ree() is just a binary search algorithm, if the index i is 1,
we just return the root of the tree, otherwise, we halve the tree by unlinking, if
1 We follow the tradition that the index i starts from 1 in algorithm description; while it
starts from 0 in most programming languages

326

CHAPTER 12. SEQUENCES, THE LAST BRICK

i is less than or equal to the size of the halved tree, we recursively look up the
left tree, otherwise, we look up the right tree.

root(T ) : i = 1
lookupT ree(lef t(T )) : i b |T2 | c

lookupT ree(right(T )) : otherwise

lookupT ree(T, i) =

(12.9)

Where function lef t() returns the left tree Tl of T , while right() returns Tr .
The corresponding Haskell program is given as below.
getAt (t:ts) i = if i < size t then lookupTree t i
else getAt ts (i - size t)
lookupTree (Leaf x) 0 = x
lookupTree (Node sz t1 t2) i = if i < sz div 2 then lookupTree t1 i
else lookupTree t2 (i - sz div 2)

Figure 12.5 illustrates the steps of looking up the 4-th element in a sequence
of size 6. It first examine the first tree, since the size is 2 which is smaller than
4, so it goes on looking up for the second tree with the updated index i0 = 4 2,
which is the 2nd element in the rest of the forest. As the size of the next tree is
4, which is greater than 2, so the element to be searched should be located in
this tree. It then examines the left sub tree since the new index 2 is not greater
than the half size 4/2=2; The process next visits the right grand-child, and the
final result is returned.
t2

(a) getAt(S, 4)), 4 > size(t1 ) = 2

(b) getAt(S 0 , 4 2) lookupT ree(t2 , 2)

left(t2)

(c) 2 bsize(t2 )/2c lookupT ree(lef t(t2 ), 2)

(d) lookupT ree(right(lef t(t2 )), 1), x3 is returned.

Figure 12.5: Steps of locating the 4-th element in a sequence.

By using the similar idea, we can update element at any arbitrary position
i. We first compare the size of the first tree T1 in the forest with i, if it is less

12.2. BINARY RANDOM ACCESS LIST

327

than i, it means the element to be updated doesnt exist in the first tree. We
recursively examine the next tree in the forest, comparing it with i |T1 |, where
|T1 | represents the size of the first tree. Otherwise if this size is greater than or
equal to i, the element is in the tree, we halve the tree recursively until to get
a leaf, at this stage, we can replace the element of this leaf with a new one.

set(S, i, x) =

{updateT ree(T1 , i, x)} S 0 : i < |T1 |

{T1 } set(S 0 , i |T1 |, x) : otherwise

(12.10)

Where S 0 = {T2 , T3 , ...} is the rest of the trees in the forest without the first
one.
Function setT ree(T, i, x) performs a tree search and replace the i-th element
with the given value x.

leaf (x) : i = 0 |T | = 1
tree(|T |, setT ree(Tl , i, x), Tr ) : i < b |T2 | c

tree(|T |, Tl , setT ree(Tr , i b |T2 | c, x)) : otherwise

(12.11)
Where Tl and Tr are left and right sub tree of T respectively. The following
Haskell program translates the equation accordingly.
setT ree(T, i, x) =

setAt :: BRAList a Int a BRAList a

setAt (t:ts) i x = if i < size t then (updateTree t i x):ts
else t:setAt ts (i-size t) x
updateTree :: Tree a Int a Tree a
updateTree (Leaf _) 0 x = Leaf x
updateTree (Node sz t1 t2) i x =
if i < sz div 2 then Node sz (updateTree t1 i x) t2
else Node sz t1 (updateTree t2 (i - sz div 2) x)

As the nature of complete binary search tree, for a sequence with N elements,
which is represented by binary random access list, the number of trees in the
forest is bound to O(lg N ). Thus it takes O(lg N ) time to locate the tree for
arbitrary index i, that contains the element. the followed tree search is bound
to the heights of the tree, which is O(lg N ) as well. So the total performance of
random access is O(lg N ).

Exercise 12.1
1. The random access algorithm given in this section doesnt handle the error
such as out of bound index at all. Modify the algorithm to handle this
case, and implement it in your favorite programming language.
2. Its quite possible to realize the binary random access list in imperative
settings, which is benefited with fast operation on the head of the sequence. the random access can be realized in two steps: firstly locate
the tree, secondly use the capability of constant random access of array.
Write a program to implement it in your favorite imperative programming
language.

328

12.3

CHAPTER 12. SEQUENCES, THE LAST BRICK

Numeric representation for binary random

access list

In previous section, we mentioned that for any sequence with N elements, we

can represent N in binary format so that N = 20 e0 + 21 e1 + ... + 2M eM . Where
ei is the i-th bit, which can be either 0 or 1. If ei 6= 0 it means that there is a
complete binary tree with size 2i .
This fact indicates us that there is an explicit relationship between the binary
form of N and the forest. Insertion a new element on the head can be simulated
by increasing the binary number by one; while remove an element from the head
mimics the decreasing of the corresponding binary number by one. This is as
known as numeric representation [6].
In order to represent the binary access list with binary number, we can define
two states for a bit. That Zero means there is no such a tree with size which is
corresponding to the bit, while One, means such tree exists in the forest. And
we can attach the tree with the state if it is One.
The following Haskell program for instance defines such states.
data Digit a = Zero
| One (Tree a)
type RAList a = [Digit a]

Here we reuse the definition of complete binary tree and attach it to the
state One. Note that we cache the size information in the tree as well.
With digit defined, forest can be treated as a list of digits. Lets see how
inserting a new element can be realized as binary number increasing. Suppose
function one(t) creates a One state and attaches tree t to it. And function
getT ree(s) get the tree which is attached to the One state s. The sequence S
is a list of digits of states that S = {s1 , s2 , ...}, and S 0 is the rest of digits with
the first one removed.

{one(t)}
{one(t)} S 0

{Zero} insertT ree(S 0 , link(t, getT ree(s1 )))

: S=
: s1 = Zero
: otherwise
(12.12)
When we insert a new tree t to a forest S of binary digits, If the forest is
empty, we just create a One state, attach the tree to it, and make this state the
only digit of the binary number. This is just like 0 + 1 = 1;
Otherwise if the forest isnt empty, we need examine the first digit of the
binary number. If the first digit is Zero, we just create a One state, attach the
tree, and replace the Zero state with the new created One state. This is just like
(...digits...0)2 +1 = (...digits...1)2 . For example 6+1 = (110)2 +1 = (111)2 = 7.
The last case is that the first digit is One, here we make assumption that the
tree t to be inserted has the same size with the tree attached to this One state at
this stage. This can be ensured by calling this function from inserting a leaf, so
that the size of the tree to be inserted grows in a series of 1, 2, 4, ..., 2i , .... In such
case, we need link these two trees (one is t, the other is the tree attached to the
One state), and recursively insert the linked result to the rest of the digits. Note
that the previous One state has to be replaced with a Zero state. This is just

insertT ree(S, t) =

12.3. NUMERIC REPRESENTATION FOR BINARY RANDOM ACCESS LIST329

like (...digits...1)2 + 1 = (...digits0 ...0)2 , where (...digits0 ...)2 = (...digits...)2 + 1.
For example 7 + 1 = (111)2 + 1 = (1000)2 = 8
Translating this algorithm to Haskell yields the following program.
insertTree
insertTree
insertTree
insertTree

:: RAList a Tree a RAList a

[] t = [One t]
(Zero:ts) t = One t : ts
(One t :ts) t = Zero : insertTree ts (link t t)

All the other functions, including link(), cons() etc. are as same as before.
Next lets see how removing an element from a sequence can be represented
as binary number deduction. If the sequence is a singleton One state attached
with a leaf. After removal, it becomes empty. This is just like 1 1 = 0;
Otherwise, we examine the first digit, if it is One state, it will be replaced
with a Zero state to indicate that this tree will be no longer exist in the forest
as it being removed. This is just like (...digits...1)2 1 = (...digits...0)2 . For
example 7 1 = (111)2 1 = (110)2 = 6;
If the first digit in the sequence is a Zero state, we have to borrow from the
further digits for removal. We recursively extract a tree from the rest digits,
and halve the extracted tree to its two children. Then the Zero state will be
replaced with a One state attached with the right children, and the left children
is removed. This is something like (...digits...0)2 1 = (...digits0 ...1)2 , where
(...digits0 ...)2 = (...digits)2 1. For example 4 1 = (100)2 1 = (11)2 = 3.The
following equation illustrated this algorithm.

extractT ree(S) =

(t, ) : S = {one(t)}
(t, S 0 ) : s1 = one(t)

(tl , {one(tr )} S 00 : otherwise

(12.13)

Where (t0 , S 00 ) = extractT ree(S 0 ), tl and tr are left and right sub-trees of t0 .
All other functions, including head(), tail() are as same as before.
Numeric representation doesnt change the performance of binary random
access list, readers can refer to [2] for detailed discussion. Lets take for example,
analyze the average performance (or amortized) of insertion on head algorithm
by using aggregation analysis.
Considering the process of inserting N = 2m elements to an empty binary
random access list. The numeric representation of the forest can be listed as
the following.
i
forest (MSB ... LSB)
0
0, 0, ..., 0, 0
0, 0, ..., 0, 1
1
2
0, 0, ..., 1, 0
0, 0, ..., 1, 1
3
...
...
1, 1, ..., 1, 1
2m 1
2m
1, 0, 0, ..., 0, 0
bits changed
1, 1, 2, ... 2m1 . 2m
The LSB of the forest changed every time when there is a new element
inserted, it costs 2m units of computation; The next bit changes every two
times due to a linking operation, so it costs 2m1 units; the bit next to MSB of
the forest changed only one time which links all previous trees to a big tree as

330

CHAPTER 12. SEQUENCES, THE LAST BRICK

the only one in the forest. This happens at the half time of the total insertion
process, and after the last element is inserted, the MSB flips to 1.
Sum these costs up yield to the total cost T = 1 + 1 + 2 + 3 + ... + 2m1 + 2m =
m+1
2
So the average cost for one insertion is
O(T /N ) = O(

2m+1
) = O(1)
2m

(12.14)

Which proves that the insertion algorithm performs in amortized O(1) constant time. The proof for deletion are left as an exercise to the reader.

12.3.1

Imperative binary access list

Its trivial to implement the binary access list by using binary trees, and the
recursion can be eliminated by updating the focused tree in loops. This is left as
an exercise to the reader. In this section, well show some different imperative
implementation by using the properties of numeric representation.
Remind the chapter about binary heap. Binary heap can be represented by
implicit array. We can use similar approach that use an array of 1 element to
represent the leaf; use an array of 2 elements to represent a binary tree of height
1; and use an array of 2m to represent a complete binary tree of height m.
This brings us the capability of accessing any element with index directly
instead of divide and conquer tree search. However, the tree linking operation
has to be implemented as array copying as the expense.
The following ANSI C code defines such a forest.
#define M sizeof(int) 8
typedef int Key;
struct List {
int n;
Key tree[M];
};

Where n is the number of the elements stored in this forest. Of course we can
avoid limiting the max number of trees by using dynamic arrays, for example
as the following ISO C++ code.
template<typename Key>
struct List {
int n;
vector<vector<key> > tree;
};

For illustration purpose only, we use ANSI C here, readers can find the
complete ISO C++ example programs along with this book.
Lets review the insertion process, if the first tree is empty (a Zero digit), we
simply set the first tree as a leaf of the new element to be inserted; otherwise, the
insertion will cause tree linking anyway, and such linking may be recursive until
it reach a position (digit) that the corresponding tree is empty. The numeric
representation reveals an important fact that if the first, second, ..., (i 1)-th
trees all exist, and the i-th tree is empty, the result is creating a tree of size 2i ,
and all the elements together with the new element to be inserted are stored in

12.3. NUMERIC REPRESENTATION FOR BINARY RANDOM ACCESS LIST331

this new created tree. Whats more, all trees after position i are kept as same
as before.
Is there any good methods to locate this i position? As we can use binary
number to represent the forest of N element, after a new element is inserted, N
increases to N + 1. Compare the binary form of N and N + 1, we find that all
bits before i change from 1 to 0, the i-th bit flip from 0 to 1, and all the bits
after i keep unchanged. So we can use bit-wise exclusive or () to detect this
bit. Here is the algorithm.
function Number-Of-Bits(N )
i0
while b N2 c 6= 0 do
N b N2 c
ii+1
return i
i Number-Of-Bits(N (N + 1))
And it can be easily implemented with bit shifting, for example the below
ANSI C code.
int nbits(int n) {
int i=0;
while(n >>= 1)
++i;
return i;
}

So the imperative insertion algorithm can be realized by first locating the

bit which flip from 0 to 1, then creating a new array of size 2i to represent
a complete binary tree, and moving content of all trees before this bit to this
array as well as the new element to be inserted.
function Insert(L, x)
i Number-Of-Bits(N (N + 1))
Tree(L)[i + 1] Create-Array(2i )
l1
Tree(L)[i + 1][l] x
for j [1, i] do
for k [1, 2j ] do
l l+1
Tree(L)[i + 1][l] Tree(L)[j][k]
Tree(L)[j] NIL
Size(L) Size(L) + 1
return L
The corresponding ANSI C program is given as the following.
struct List insert(struct List a, Key x) {
int i, j, sz;
Key xs;
i = nbits( (a.n+1) ^ a.n );
xs = a.tree[i] = (Key)malloc(sizeof(Key)(1<<i));
for(j=0, xs++ = x, sz = 1; j<i; ++j, sz < 1) {
memcpy((void)xs, (void)a.tree[j], sizeof(Key)(sz));

332

CHAPTER 12. SEQUENCES, THE LAST BRICK

xs += sz;
free(a.tree[j]);
a.tree[j] = NULL;

}
++a.n;
return a;
}

However, the performance in theory isnt as good as before. This is because the linking operation downgrade from O(1) constant time to linear array
copying.
We can again calculate the average (amortized) performance by using aggregation analysis. When insert N = 2m elements to an empty list which is
represented by implicit binary trees in arrays, the numeric presentation of the
forest of arrays are as same as before except for the cost of bit flipping.
i
forest (MSB ... LSB)
0
0, 0, ..., 0, 0
1
0, 0, ..., 0, 1
2
0, 0, ..., 1, 0
3
0, 0, ..., 1, 1
...
...
2m 1
1, 1, ..., 1, 1
2m
1, 0, 0, ..., 0, 0
bit change cost 1 2m , 1 2m1 , 2 2m2 , ... 2m2 2, 2m1 1
The LSB of the forest changed every time when there is a new element
inserted, however, it creates leaf tree and performs copying only it changes from
0 to 1, so the cost is half of N unit, which is 2m1 ; The next bit flips as half as
the LSB. Each time the bit gets flipped to 1, it copies the first tree as well as
the new element to the second tree. the the cost of flipping a bit to 1 in this bit
is 2 units, but not 1; For the MSB, it only flips to 1 at the last time, but the
cost of flipping this bit, is copying all the previous trees to fill the array of size
2m .
Summing all to cost and distributing them to the N times of insertion yields
the amortized performance as below.
O(T /N ) = O( 12
= O(1 +
= O(m)

+12m1 +22m2 +...+2m1 1

)
2m
m
)
2

(12.15)

As m = O(lg N ), so the amortized performance downgrade from constant

time to logarithm, although it is still faster than the normal array insertion
which is O(N ) in average.
The random accessing gets a bit faster because we can use array indexing
instead of tree search.
function Get(L, i)
for each t Trees(L) do
if t 6= NIL then
if i Size(t) then
return t[i]
else
i i Size(t)

12.4. IMPERATIVE PAIRED-ARRAY LIST

333

Here we skip the error handling such as out of bound indexing etc. The
ANSI C program of this algorithm is like the following.
Key get(struct List a, int i) {
int j, sz;
for(j = 0, sz = 1; j < M; ++j, sz < 1)
if(a.tree[j]) {
if(i < sz)
break;
i -= sz;
}
return a.tree[j][i];
}

The imperative removal and random mutating algorithms are left as exercises
to the reader.

Exercise 12.2

1. Please implement the random access algorithms, including looking up and

updating, for binary random access list with numeric representation in
your favorite programming language.
2. Prove that the amortized performance of deletion is O(1) constant time
by using aggregation analysis.
3. Design and implement the binary random access list by implicit array in
your favorite imperative programming language.

12.4

Imperative paired-array list

12.4.1

Definition

In previous chapter about queue, a symmetric solution of paired-array is presented. It is capable to operate on both ends of the list. Because the nature
that array supports fast random access. It can be also used to realize a fast
random access sequence in imperative setting.

x[n]

...

x[2]

x[1]

y[1]

y[2]

...

y[m]

Figure 12.6: A paired-array list, which is consist of 2 arrays linking in head-head

manner.
Figure 12.6 shows the design of paired-array list. Tow arrays are linked in
head-head manner. To insert a new element on the head of the sequence, the
element is appended at the end of front list; To append a new element on the
tail of the sequence, the element is appended at the end of rear list;
Here is a ISO C++ code snippet to define the this data structure.

334

CHAPTER 12. SEQUENCES, THE LAST BRICK

template<typename Key>
struct List {
int n, m;
vector<Key> front;
vector<Key> rear;
List() : n(0), m(0) {}
int size() { return n + m; }
};

Here we use vector provides in standard library to cover the dynamic memory
management issues, so that we can concentrate on the algorithm design.

12.4.2

Insertion and appending

Suppose function Front(L) returns the front array, while Rear(L) returns the
rear array. For illustration purpose, we assume the arrays are dynamic allocated.
inserting and appending can be realized as the following.
function Insert(L, x)
F Front(L)
Size(F ) Size(F ) + 1
F [Size(F )] x
function Append(L, x)
R Rear(L)
Size(R) Size(R) + 1
R[Size(R)] x
As all the above operations manipulate the front and rear array on tail, they
are all constant O(1) time. And the following are the corresponding ISO C++
programs.
template<typename Key>
void insert(List<Key>& xs, Key x) {
++xs.n;
xs.front.push_back(x);
}
template<typename Key>
void append(List<Key>& xs, Key x) {
++xs.m;
xs.rear.push_back(x);
}

12.4.3

random access

As the inner data structure is array (dynamic array as vector), which supports
random access by nature, its trivial to implement constant time indexing algorithm.
function Get(L, i)
F Front(L)
N Size(F )

12.4. IMPERATIVE PAIRED-ARRAY LIST

335

if i N then
return F [N i + 1]
else
Rear(L)[i N ]
Here the index i [1, |L|] starts from 1. If it is not greater than the size of
front array, the element is stored in front. However, as front and rear arrays are
connect head-to-head, so the elements in front array are in reverse order. We
need locate the element by subtracting the size of front array by i; If the index i
is greater than the size of front array, the element is stored in rear array. Since
elements are stored in normal order in rear, we just need subtract the index i
by an offset which is the size of front array.
Here is the ISO C++ program implements this algorithm.
template<typename Key>
Key get(List<Key>& xs, int i) {
if( i < xs.n )
return xs.front[xs.n-i-1];
else
return xs.rear[i-xs.n];
}

The random mutating algorithm is left as an exercise to the reader.

12.4.4

removing and balancing

Removing isnt as simple as insertion and appending. This is because we must

handle the condition that one array (either front or rear) becomes empty due to
removal, while the other still contains elements. In extreme case, the list turns
to be quite unbalanced. So we must fix it to resume the balance.
One idea is to trigger this fixing when either front or rear array becomes
empty. We just cut the other array in half, and reverse the first half to form
the new pair. The algorithm is described as the following.
function Balance(L)
F Front(L), R Rear(L)
N Size(F ), M Size(R)
if F = then
F Reverse(R[1 ... b M
2 c])
M
R R[b 2 c + 1...M ]
else if R = then
R Reverse(F [1 ... b N2 c])
F F [b N2 c + 1...N ]
Actually, the operations are symmetric for the case that front is empty and
the case that rear is empty. Another approach is to swap the front and rear for
one symmetric case and recursive resumes the balance, then swap the front and
rear back. For example below ISO C++ program uses this method.
template<typename Key>
void balance(List<Key>& xs) {
if(xs.n == 0) {
back_insert_iterator<vector<Key> > i(xs.front);
reverse_copy(xs.rear.begin(), xs.rear.begin() + xs.m/2, i);

336

CHAPTER 12. SEQUENCES, THE LAST BRICK

xs.rear.erase(xs.rear.begin(), xs.rear.begin() +xs.m/2);
xs.n = xs.m/2;
xs.m -= xs.n;

}
else if(xs.m == 0) {
swap(xs.front, xs.rear);
swap(xs.n, xs.m);
balance(xs);
swap(xs.front, xs.rear);
swap(xs.n, xs.m);
}
}

With Balance algorithm defined, its trivial to implement remove algorithm

both on head and on tail.
function Remove-Head(L)
Balance(L)
F Front(L)
if F = then
Remove-Tail(L)
else
Size(F ) Size(F ) - 1
function Remove-Tail(L)
Balance(L)
R Rear(L)
if R = then
Remove-Head(L)
else
Size(R) Size(R) - 1
There is an edge case for each, that is even after balancing, the array targeted
to perform removal is still empty. This happens that there is only one element
stored in the paired-array list. The solution is just remove this singleton left
element, and the overall list results empty. Below is the ISO C++ program
implements this algorithm.
template<typename Key>
void remove_head(List<Key>& xs) {
balance(xs);
if(xs.front.empty())
remove_tail(xs); //remove the singleton elem in rear
else {
xs.front.pop_back();
--xs.n;
}
}
template<typename Key>
void remove_tail(List<Key>& xs) {
balance(xs);
if(xs.rear.empty())
remove_head(xs); //remove the singleton elem in front

12.5. CONCATENATE-ABLE LIST

337

else {
xs.rear.pop_back();
--xs.m;
}
}

Its obvious that the worst case performance is O(N ) where N is the number
of elements stored in paired-array list. This happens when balancing is triggered,
and both reverse and shifting are linear operation. However, the amortized
performance of removal is still O(1), the proof is left as exercise to the reader.

Exercise 12.3
1. Implement the random mutating algorithm in your favorite imperative
programming language.
2. We utilized vector provided in standard library to manage memory dynamically, try to realize a version using plain array and manage the memory
allocation manually. Compare this version and consider how does this
affect the performance?
3. Prove that the amortized performance of removal is O(1) for paired-array
list.

12.5

Concatenate-able list

By using binary random access list, we realized sequence data structure which
supports O(lg N ) time insertion and removal on head, as well as random accessing element with a given index.
However, its not so easy to concatenate two lists. As both lists are forests of
complete binary trees, we cant merely merge them (Since forests are essentially
list of trees, and for any size, there is at most one tree of that size. Even
concatenate forests directly is not fast). One solution is to push the element
from the first sequence one by one to a stack and then pop those elements and
insert them to the head of the second one by using cons function. Of course
the stack can be implicitly used in recursion manner, for instance:

s2 : s1 =
cons(head(s1 ), concat(tail(s1 ), s2 )) : otherwise
(12.16)
Where function cons(), head() and tail() are defined in previous section.
If the length of the two sequence is N , and M , this method takes O(N lg N )
time repeatedly push all elements from the first sequence to stacks, and then
takes (N lg(N + M )) to insert the elements in front of the second sequence.
Note that means the upper limit, There is detailed definition for it in [2].
We have already implemented the real-time queue in previous chapter. It
supports O(1) time pop and push. If we can turn the sequence concatenation
to a kind of pushing operation to queue, the performance will be improved to
O(1) as well. Okasaki gave such realization in [6], which can concatenate lists
in constant time.
concat(s1 , s2 ) =

338

CHAPTER 12. SEQUENCES, THE LAST BRICK

To represent a concatenate-able list, the data structure designed by Okasaki

is essentially a K-ary tree. The root of the tree stores the first element in the
list. So that we can access it in constant O(1) time. The sub-trees or children
are all small concatenate-able lists, which are managed by real-time queues.
Concatenating another list to the end is just adding it as the last child, which
is in turn a queue pushing operation. Appending a new element can be realized
as that, first wrapping the element to a singleton tree, which is a leaf with no
children. Then, concatenate this singleton to finalize appending.
Figure 12.7 illustrates this data structure.

x[1]

c[1]

x[2]...x[i]

c[2]

...

c[n]

x[i+1]...x[j]

x[k]...x[n]

(a) The data structure for list {x1 , x2 , ..., xn }

x[1]

c[1]

x[2]...x[i]

x[i+1]...x[j]

c[2]

...

c[n]

c[n+1]

x[k]...x[n]

(b) The result after concatenated with list {y1 , y2 , ..., ym }

Figure 12.7: Data structure for concatenate-able list

Such recursively designed data structure can be defined in the following

Haskell code.
data CList a = Empty | CList a (Queue (CList a)) deriving (Show, Eq)

It means that a concatenate-able list is either empty or a K-ary tree, which

again consists of a queue of concatenate-able sub-lists and a root element. Here
we reuse the realization of real-time queue mentioned in previous chapter.
Suppose function clist(x, Q) constructs a concatenate-able list from an element x, and a queue of sub-lists Q. While function root(s) returns the root
element of such K-ary tree implemented list. and function queue(s) returns the
queue of sub-lists respectively. We can implement the algorithm to concatenate

y[1]...y[m]

12.5. CONCATENATE-ABLE LIST

339

two lists like this.

concat(s1 , s2 ) =

s1 : s2 =
s2 : s1 =

clist(x, push(Q, s2 )) : otherwise

(12.17)

Where x = root(s1 ) and Q = queue(s1 ). The idea of concatenation is that

if either one of the list to be concatenated is empty, the result is just the other
list; otherwise, we push the second list as the last child to the queue of the first
list.
Since the push operation is O(1) constant time for a well realized real-time
queue, the performance of concatenation is bound to O(1).
The concat() function can be translated to the below Haskell program.
concat x Empty = x
concat Empty y = y
concat (CList x q) y = CList x (push q y)

Besides the good performance of concatenation, this design also brings satisfied features for adding element both on head and tail.
cons(x, s) = concat(clist(x, ), s)

(12.18)

append(s, x) = concat(s, clist(x, ))

(12.19)

Its a bit complex to realize the algorithm that removes the first element
from a concatenate-able list. This is because after the root, which is the first
element in the sequence got removed, we have to re-construct the rest things, a
queue of sub-lists, to a K-ary tree.
Before diving into the re-construction, lets solve the trivial part first. Getting the first element is just returning the root of the K-ary tree.
head(s) = root(s)

(12.20)

As we mentioned above, after root being removed, there left all children of
the K-ary tree. Note that all of them are also concatenate-able list, so that one
natural solution is to concatenate them all together to a big list.

: Q=
concat(f ront(Q), concatAll(pop(Q))) : otherwise
(12.21)
Where function f ront() just returns the first element from a queue without
removing it, while pop() does the removing work.
If the queue is empty, it means that there is no children at all, so the result is
also an empty list; Otherwise, we pop the first child, which is a concatenate-able
list, from the queue, and recursively concatenate all the rest children to a list;
finally, we concatenate this list behind the already popped first children.
With concatAll() defined, we can then implement the algorithm of removing
the first element from a list as below.
concatAll(Q) =

tail(s) = linkAll(queue(s))
The corresponding Haskell program is given like the following.

(12.22)

340

CHAPTER 12. SEQUENCES, THE LAST BRICK

head (CList x _) = x
tail (CList _ q) = linkAll q
linkAll q | isEmptyQ q = Empty
| otherwise = link (front q) (linkAll (pop q))

Function isEmptyQ is used to test a queue is empty, it is trivial and we omit

its definition. Readers can refer to the source code along with this book.
linkAll() algorithm actually traverses the queue data structure, and reduces
to a final result. This remind us of folding mentioned in the chapter of binary
search tree. readers can refer to the appendix of this book for the detailed
description of folding. Its quite possible to define a folding algorithm for queue
instead of list2 [8].

f oldQ(f, e, Q) =

e : Q=
f (f ront(Q), f oldQ(f, e, pop(Q))) : otherwise

(12.23)

Function f oldQ() takes three parameters, a function f , which is used for

reducing, an initial value e, and the queue Q to be traversed.
Here are some examples to illustrate folding on queue. Suppose a queue Q
contains elements {1, 2, 3, 4, 5} from head to tail.
f oldQ(+, 0, Q) = 1 + (2 + (3 + (4 + (5 + 0)))) = 15
f oldQ(, 1, Q) = 1 (2 (3 (4 (5 1)))) = 120
f oldQ(, 0, Q) = 1 (2 (3 (4 (5 0)))) = 0
Function linkAll can be changed by using f oldQ accordingly.
linkAll(Q) = f oldQ(link, , Q)

(12.24)

The Haskell program can be modified as well.

linkAll = foldQ link Empty
foldQ :: (a b b) b Queue a b
foldQ f z q | isEmptyQ q = z
| otherwise = (front q) f foldQ f z (pop q)

However, the performance of removing cant be ensured in all cases. The

worst case is that, user keeps appending N elements to a empty list, and then
immediately performs removing. At this time, the K-ary tree has the first
element stored in root. There are N 1 children, all are leaves. So linkAll()
algorithm downgrades to O(N ) which is linear time.
The average case is amortized O(1), if the add, append, concatenate and
removing operations are randomly performed. The proof is left as en exercise
to the reader.

Exercise 12.4

1. Can you figure out a solution to append an element to the end of a binary
random access list?
2 Some functional programming language, such as Haskell, defined type class, which is a
concept of monoid so that its easy to support folding on a customized data structure.

12.6. FINGER TREE

341

2. Prove that the amortized performance of removal operation is O(1). Hint:

using the bankers method.
3. Implement the concatenate-able list in your favorite imperative language.

12.6

Finger tree

We havent been able to meet all the performance targets listed at the beginning
of this chapter.
Binary random access list enables to insert, remove element on the head of
sequence, and random access elements fast. However, it performs poor when
concatenates lists. There is no good way to append element at the end of binary
access list.
Concatenate-able list is capable to concatenates multiple lists in a fly, and
it performs well for adding new element both on head and tail. However, it
doesnt support randomly access element with a given index.
These two examples bring us some ideas:
In order to support fast manipulation both on head and tail of the sequence, there must be some way to easily access the head and tail position;
Tree like data structure helps to turn the random access into divide and
conquer search, if the tree is well balance, the search can be ensured to be
logarithm time.

12.6.1

Definition

Finger tree[6], which was first invented in 1977, can help to realize efficient
sequence. And it is also well implemented in purely functional settings[5].
As we mentioned that the balance of the tree is critical to ensure the performance for search. One option is to use balanced tree as the under ground data
structure for finger tree. For example the 2-3 tree, which is a special B-tree.
(readers can refer to the chapter of B-tree of this book).
A 2-3 tree either contains 2 children or 3. It can be defined as below in
Haskell.
data Node a = Br2 a a | Br3 a a a

In imperative settings, node can be defined with a list of sub nodes, which
contains at most 3 children. For instance the following ANSI C code defines
node.
union Node {
Key keys;
union Node children;
};

Note in this definition, a node can either contain 2 3 keys, or 2 3 sub

nodes. Where key is the type of elements stored in leaf node.
We mark the left-most none-leaf node as the front finger and the right-most
none-leaf node as the rear finger. Since both fingers are essentially 2-3 trees
with all leafs as children, they can be directly represented as list of 2 or 3 leafs.
Of course a finger tree can be empty or contain only one element as leaf.

342

CHAPTER 12. SEQUENCES, THE LAST BRICK

So the definition of a finger tree is specified like this.
A finger tree is either empty;
or a singleton leaf;
or contains three parts: a left finger which is a list contains at most 3
elements; a sub finger tree; and a right finger which is also a list contains
at most 3 elements.

Note that this definition is recursive, so its quite possible to be translated

to functional settings. The following Haskell definition summaries these cases
for example.
data Tree a = Empty
| Lf a
| Tr [a] (Tree (Node a)) [a]

In imperative settings, we can define the finger tree in a similar manner.

Whats more, we can add a parent field, so that its possible to back-track to
root from any tree node. Below ANSI C code defines finger tree accordingly.
struct Tree {
union Node front;
union Node rear;
Tree mid;
Tree parent;
};

We can use NIL pointer to represent an empty tree; and a leaf tree contains
only one element in its front finger, both its rear finger and middle part are
empty;
Figure 12.8 and 12.9 show some examples of figure tree.

NIL

(a) An empty tree

(b) A singleton leaf

NIL

(c) Front finger and rear

finger contain one element
for each, the middle part is
empty

Figure 12.8: Examples of finger tree, 1

The first example is an empty finger tree; the second one shows the result
after inserting one element to empty, it becomes a leaf of one node; the third

12.6. FINGER TREE

343

NIL

(a) After inserting extra 3 elements to front finger, it exceeds

the 2-3 tree constraint, which isnt
balanced any more

(b) The tree resumes balancing. There are 2 elements in front finger;
The middle part is a
leaf, which contains a 3branches 2-3 tree.

Figure 12.9: Examples of finger tree, 2

example shows a finger tree contains 2 elements, one is in front finger, the other
is in rear;
If we continuously insert new elements, to the tree, those elements will be
put in the front finger one by one, until it exceeds the limit of 2-3 tree. The 4-th
example shows such condition, that there are 4 elements in front finger, which
isnt balanced any more.
The last example shows that the finger tree gets fixed so that it resumes
balancing. There are two elements in the front finger. Note that the middle
part is not empty any longer. Its a leaf of a 2-3 tree. The content of the leaf is
a tree with 3 branches, each contains an element.
We can express these 5 examples as the following Haskell expression.
Empty
Lf a
[b] Empty [a]
[e, d, c, b] Empty [a]
[f, e] Lf (Br3 d c b) [a]

As we mentioned that the definition of finger tree is recursive. The middle

part besides the front and rear finger is a deeper finger tree, which is defined as
T ree(N ode(a)). Every time we go deeper, the N ode() is embedded one more
level. if the element type of the first level tree is a, the element type for the
second level tree is N ode(a), the third level is N ode(N ode(a)), ..., the n-th level
is N ode(N ode(N ode(...(a))...)) = N oden (a), where n indicates the N ode() is
applied n times.

12.6.2

Insert element to the head of sequence

The examples list above actually reveal the typical process that the elements are
inserted one by one to a finger tree. Its possible to summarize these examples
to some cases for insertion on head algorithm.
When we insert an element x to a finger tree T ,

344

CHAPTER 12. SEQUENCES, THE LAST BRICK

If the tree is empty, the result is a leaf which contains the singleton element
x;
If the tree is a singleton leaf of element y, the result is a new finger tree.
The front finger contains the new element x, the rear finger contains the
previous element y; the middle part is a empty finger tree;
If the number of elements stored in front finger isnt bigger than the upper
limit of 2-3 tree, which is 3, the new element is just inserted to the head
of front finger;
otherwise, it means that the number of elements stored in front finger
exceeds the upper limit of 2-3 tree. the last 3 elements in front finger is
wrapped in a 2-3 tree and recursively inserted to the middle part. the new
element x is inserted in front of the rest elements in front finger.

Suppose that function leaf (x) creates a leaf of element x, function tree(F, T 0 , R)
creates a finger tree from three part: F is the front finger, which is a list contains
several elements. Similarity, R is the rear finger, which is also a list. T 0 is the
middle part which is a deeper finger tree. Function tr3(a, b, c) creates a 2-3 tree
from 3 elements a, b, c; while tr2(a, b) creates a 2-3 tree from 2 elements a and
b.

leaf (x) : T =
tree({x}, , {y}) : T = leaf (y)
insertT (x, T ) =
tree({x, x1 }, insertT (tr3(x2 , x3 , x4 ), T 0 ), R) : T = tree({x1 , x2 , x3 , x4 }, T 0 , R)

tree({x} F, T 0 , R) : otherwise
(12.25)
The performance of this algorithm is dominated by the recursive case. All
the other cases are constant O(1) time. The recursion depth is proportion to
the height of the tree, so the algorithm is bound to O(h) time, where h is the
height. As we use 2-3 tree to ensure that the tree is well balanced, h = O(lg N ),
where N is the number of elements stored in the finger tree.
More analysis reveal that the amortized performance of insertT () is O(1)
because we can amortize the expensive recursion case to other trivial cases.
Please refer to [6] and [5] for the detailed proof.
Translating the algorithm yields the below Haskell program.
cons
cons
cons
cons
cons

:: a Tree a Tree a
a Empty = Lf a
a (Lf b) = Tr [a] Empty [b]
a (Tr [b, c, d, e] m r) = Tr [a, b] (cons (Br3 c d e) m) r
a (Tr f m r) = Tr (a:f) m r

Here we use the LISP naming convention to illustrate inserting a new element
to a list.
The insertion algorithm can also be implemented in imperative approach.
Suppose function Tree() creates an empty tree, that all fields, including front
and rear finger, the middle part inner tree and parent are empty. Function
Node() creates an empty node.
function Prepend-Node(n, T )
r Tree()

12.6. FINGER TREE

pr
Connect-Mid(p, T )
while Full?(Front(T )) do
F Front(T )
Front(T ) {n, F [1]}
n Node()
Children(n) F [2..]
pT
T Mid(T )
if T = NIL then
T Tree()
Front(T ) {n}
else if | Front(T ) | = 1 Rear(T ) = then
Rear(T ) Front(T )
Front(T ) {n}
else
Front(T ) {n} Front(T )
Connect-Mid(p, T ) T
return Flat(r)

345

. F = {n1 , n2 , n3 , ...}
. F [1] = n1
. F [2..] = {n2 , n3 , ...}

Where the notation L[i..] means a sub list of L with the first i 1 elements
removed, that if L = {a1 , a2 , ..., an }, L[i..] = {ai , ai+1 , ..., an }.
Functions Front(), Rear(), Mid(), and Parent() are used to access the
front finger, the rear finger, the middle part inner tree and the parent tree
respectively; Function Children() accesses the children of a node.
Function Connect-Mid(T1 , T2 ), connect T2 as the inner middle part tree
of T1 , and set the parent of T2 as T1 if T2 isnt empty.
In this algorithm, we performs a one pass top-down traverse along the middle
part inner tree if the front finger is full that it cant afford to store any more.
The criteria for full for a 2-3 tree is that the finger contains 3 elements already.
In such case, we extract all the elements except the first one off, wrap them to a
new node (one level deeper node), and continuously insert this new node to its
middle inner tree. The first element is left in the front finger, and the element
to be inserted is put in front of it, so that this element becomes the new first
one in the front finger.
After this traversal, the algorithm either reach an empty tree, or the tree
still has room to hold more element in its front finger. We create a new leaf
for the former case, and perform a trivial list insert to the front finger for the
latter.
During the traversal, we use p to record the parent of the current tree we
are processing. So any new created tree are connected as the middle part inner
tree to p.
Finally, we return the root of the tree r. The last trick of this algorithm is
the Flat() function. In order to simplify the logic, we create an empty ground
tree and set it as the parent of the root. We need eliminate this extra ground
level before return the root. This flatten algorithm is realized as the following.
function Flat(T )
while T 6= NIL T is empty do
T Mid(T )

346

CHAPTER 12. SEQUENCES, THE LAST BRICK

if T 6= then
Parent(T )
return T
The while loop test if T is trivial empty, that its not NIL(= ), while both
its front and rear fingers are empty.
Below Python code implements the insertion algorithm for finger tree.
def insert(x, t):
return prepend_node(wrap(x), t)
def prepend_node(n, t):
root = prev = Tree()
prev.set_mid(t)
while frontFull(t):
f = t.front
t.front = [n] + f[:1]
n = wraps(f[1:])
prev = t
t = t.mid
if t is None:
t = leaf(n)
elif len(t.front)==1 and t.rear == []:
t = Tree([n], None, t.front)
else:
t = Tree([n]+t.front, t.mid, t.rear)
prev.set_mid(t)
return flat(root)
def flat(t):
while t is not None and t.empty():
t = t.mid
if t is not None:
t.parent = None
return t

The implementation of function set_mid(), frontFull(), wrap(), wraps(),

empty(), and tree constructor are trivial enough, that we skip the detail of
them here. Readers can take these as exercises.

12.6.3

Remove element from the head of sequence

Its easy to implement the reverse operation that remove the first element from
the list by reversing the insertT () algorithm line by line.
Lets denote F = {f1 , f2 , ...} is the front finger list, M is the middle part
inner finger tree. R = {r1 , r2 , ...} is the rear finger list of a finger tree, and
R0 = {r2 , r3 , ...} is the rest of element with the first one removed from R.

extractT (T ) =

(x, )
(x, leaf (y))
(x, tree({r1 }, , R0 ))
(x, tree(toList(F 0 ), M 0 , R))
(f1 , tree({f2 , f3 , ...}, M, R))

:
:
:
:
:

T = leaf (x)
T = tree({x}, , {y})
T = tree({x}, , R)
T = tree({x}, M, R), (F 0 , M 0 ) = extractT (M )
otherwise
(12.26)

12.6. FINGER TREE

347

Where function toList(T ) converts a 2-3 tree to plain list as the following.

{x, y} : T = tr2(x, y)
toList(T ) =
(12.27)
{x, y, z} : T = tr3(x, y, z)
Here we skip the error handling such as trying to remove element from empty
tree etc. If the finger tree is a leaf, the result after removal is an empty tree; If
the finger tree contains two elements, one in the front rear, the other in rear, we
return the element stored in front rear as the first element, and the resulted tree
after removal is a leaf; If there is only one element in front finger, the middle
part inner tree is empty, and the rear finger isnt empty, we return the only
element in front finger, and borrow one element from the rear finger to front;
If there is only one element in front finger, however, the middle part inner tree
isnt empty, we can recursively remove a node from the inner tree, and flatten it
to a plain list to replace the front finger, and remove the original only element
in front finger; The last case says that if the front finger contains more than one
element, we can just remove the first element from front finger and keep all the
other part unchanged.
Figure 12.10 shows the steps of removing two elements from the head of
a sequence. There are 10 elements stored in the finger tree. When the first
element is removed, there is still one element left in the front finger. However,
when the next element is removed, the front finger is empty. So we borrow one
tree node from the middle part inner tree. This is a 2-3 tree. it is converted
to a list of 3 elements, and the list is used as the new finger. the middle part
inner tree change from three parts to a singleton leaf, which contains only one
2-3 tree node. There are three elements stored in this tree node.
Below is the corresponding Haskell program for uncons.
uncons
uncons
uncons
uncons
uncons
uncons

:: Tree a (a, Tree a)

(Lf a) = (a, Empty)
(Tr [a] Empty [b]) = (a, Lf b)
(Tr [a] Empty (r:rs)) = (a, Tr [r] Empty rs)
(Tr [a] m r) = (a, Tr (nodeToList f) m r) where (f, m) = uncons m
(Tr f m r) = (head f, Tr (tail f) m r)

And the function nodeT oList is defined like this.

nodeToList :: Node a [a]
nodeToList (Br2 a b) = [a, b]
nodeToList (Br3 a b c) = [a, b, c]

Similar as above, we can define head and tail function from uncons.
head = fst uncons
tail = snd uncons

12.6.4

Handling the ill-formed finger tree when removing

The strategy used so far to remove element from finger tree is a kind of removing
and borrowing. If the front finger becomes empty after removing, we borrows
more nodes from the middle part inner tree. However there exists cases that the
tree is ill-formed, for example, both the front fingers of the tree and its middle
part inner tree are empty. Such ill-formed tree can result from imperatively
splitting, which well introduce later.

348

CHAPTER 12. SEQUENCES, THE LAST BRICK

x[10]

x[9]

x[2]

x[1]

NIL

x[8]

x[7]

x[6]

x[5]

x[4]

x[3]

(a) A sequence of 10 elements represented as a finger tree

x[9]

x[2]

x[1]

NIL

x[8]

x[7]

x[6]

x[5]

x[4]

x[3]

(b) The first element is removed. There is one element left in front finger.

x[8]

x[7]

x[6]

x[5]

x[2]

x[4]

x[1]

x[3]

(c) Another element is remove from head. We borrowed one node from
the middle part inner tree, change the node, which is a 2-3 tree to a
list, and use it as the new front finger. the middle part inner tree
becomes a leaf of one 2-3 tree node.

Figure 12.10: Examples of remove 2 elements from the head of a sequence

12.6. FINGER TREE

349
1

[]

r[1][1]

r[1][2]

...

[]

r[2][1]

r[2][2]

...

n[i][1]

n[i][2]

...

r[i][1]

r[i][2]

...

Figure 12.11: Example of an ill-formed tree. The front finger of the i-th level
sub tree isnt empty.

Here we developed an imperative algorithm which can remove the first element from finger tree even it is ill-formed. The idea is first perform a top-down
traverse to find a sub tree which either has a non-empty front finger or both
its front finger and middle part inner tree are empty. For the former case, we
can safely extract the first element which is a node from the front finger; For
the latter case, since only the rear finger isnt empty, we can swap it with the
empty front finger, and change it to the former case.
After that, we need examine the node we extracted from the front finger is
leaf node (How to do that? this is left as an exercise to the reader). If not, we
need go on extracting the first sub node from the children of this node, and left
the rest of other children as the new front finger to the parent of the current
tree. We need repeatedly go up along with the parent field till the node we
extracted is a leaf. At that time point, we arrive at the root of the tree. Figure
12.12 illustrates this process.
Based on this idea, the following algorithm realizes the removal operation
on head. The algorithm assumes that the tree passed in isnt empty.
function Extract-Head(T )
r Tree()
Connect-Mid(r, T )
while Front(T ) = Mid(T ) 6= NIL do
T Mid(T )
if Front(T ) = Rear(T ) 6= then
Exchange Front(T ) Rear(T )
n Node()
Children(n) Front(T )
repeat
L Children(n)
. L = {n1 , n2 , n3 , ...}
n L[1]
. n n1

350

CHAPTER 12. SEQUENCES, THE LAST BRICK

children of n[i][1]=

n[i-1][1]

n[i-1][2]

x[1] is extracted

[]

r[1][1]

r[1][2]

...

[]

r[2][1]

r[2][2]

...

x[2]

n[2][2]

x[3]

...

r[1][1]

r[1][2]

...

n[2][3]

...

r[2][1]

r[2][2]

...

r[i-1][1]

r[i-1][2]

...

n[i-1][2]

n[i-1][3]

...

i+1

n[i][2]

...

r[i-1][1]

r[i-1][2]

...

i+1

r[i][1]

r[i][2]

...

n[i][2]

...

r[i][1]

r[i][2]

...

(a) Extract the first element n[i][1] and put its children to the(b) Repeat this process i times, and finally x[1]
front finger of upper level tree.
is extracted.

Figure 12.12: Traverse bottom-up till a leaf is extracted.

12.6. FINGER TREE

Front(T ) L[2..]
T Parent(T )
if Mid(T ) becomes empty then
Mid(T ) NIL
until n is a leaf
return (Elem(n), Flat(r))

351
. L[2..] = {n2 , n3 , ...}

Note that function Elem(n) returns the only element stored inside leaf node
n. Similar as imperative insertion algorithm, a stub ground tree is used as the
parent of the root, which can simplify the logic a bit. Thats why we need flatten
the tree finally.
Below Python program translates the algorithm.
def extract_head(t):
root = Tree()
root.set_mid(t)
while t.front == [] and t.mid is not None:
t = t.mid
if t.front == [] and t.rear != []:
(t.front, t.rear) = (t.rear, t.front)
n = wraps(t.front)
while True: # a repeat-until loop
ns = n.children
n = ns[0]
t.front = ns[1:]
t = t.parent
if t.mid.empty():
t.mid.parent = None
t.mid = None
if n.leaf:
break
return (elem(n), flat(root))

Member function Tree.empty() returns true if all the three parts - the front
finger, the rear finger and the middle part inner tree - are empty. We put a flag
Node.leaf to mark if a node is a leaf or compound node. The exercise of this
section asks the reader to consider some alternatives.
As the ill-formed tree is allowed, the algorithms to access the first and last
element of the finger tree must be modified, so that they dont blindly return the
first or last child of the finger as the finger can be empty if the tree is ill-formed.
The idea is quite similar to the Extract-Head, that in case the finger is
empty while the middle part inner tree isnt, we need traverse along with the
inner tree till a point that either the finger becomes non-empty or all the nodes
are stored in the other finger. For instance, the following algorithm can return
the first leaf node even the tree is ill-formed.
function First-Lf(T )
while Front(T ) = Mid(T ) 6= NIL do
T Mid(T )
if Front(T ) = Rear(T ) 6= then
n Rear(T )[1]
else
n Front(T )[1]

352

CHAPTER 12. SEQUENCES, THE LAST BRICK

while n is NOT leaf do

n Children(n)[1]
return n
Note the second loop in this algorithm that it keeps traversing on the first
sub-node if current node isnt a leaf. So we always get a leaf node and its trivial
to get the element inside it.
function First(T )
return Elem(First-Lf(T ))
The following Python code translates the algorithm to real program.
def first(t):
return elem(first_leaf(t))
def first_leaf(t):
while t.front == [] and t.mid is not None:
t = t.mid
if t.front == [] and t.rear != []:
n = t.rear[0]
else:
n = t.front[0]
while not n.leaf:
n = n.children[0]
return n

To access the last element is quite similar, and we left it as an exercise to

the reader.

12.6.5

append element to the tail of the sequence

Because finger tree is symmetric, we can give the realization of appending element on tail by referencing to insertT () algorithm.

leaf (x)
tree({y}, , {x})
appendT (T, x) =
tree(F, appendT (M, tr3(x1 , x2 , x3 )), {x4 , x})

tree(F, M, R {x})

T =
T = leaf (y)
T = tree(F, M, {x1 , x2 , x3 , x4 })
otherwise
(12.28)
Generally speaking, if the rear finger is still valid 2-3 tree, that the number
of elements is not greater than 4, the new elements is directly appended to rear
finger. Otherwise, we break the rear finger, take the first 3 elements in rear
finger to create a new 2-3 tree, and recursively append it to the middle part
inner tree. If the finger tree is empty or a singleton leaf, it will be handled in
the first two cases.
Translating the equation to Haskell yields the below program.

snoc
snoc
snoc
snoc
snoc

:: Tree a a
Empty a = Lf a
(Lf a) b = Tr [a]
(Tr f m [a, b, c,
(Tr f m r) a = Tr

:
:
:
:

Tree a
Empty [b]
d]) e = Tr f (snoc m (Br3 a b c)) [d, e]
f m (r++[a])

12.6. FINGER TREE

353

Function name snoc is mirror of cons, which indicates the symmetric

relationship.
Appending new element to the end imperatively is quite similar. The following algorithm realizes appending.
function Append-Node(T, n)
r Tree()
pr
Connect-Mid(p, T )
while Full?(Rear(T )) do
R Rear(T )
. R = {n1 , n2 , ..., , nm1 , nm }
Rear(T ) {n, Last(R) }
. last element nm
n Node()
Children(n) R[1...m 1]
. {n1, n2, ..., nm1 }
pT
T Mid(T )
if T = NIL then
T Tree()
Front(T ) {n}
else if | Rear(T ) | = 1 Front(T ) = then
Front(T ) Rear(T )
Rear(T ) {n}
else
Rear(T ) Rear(T ) {n}
Connect-Mid(p, T ) T
return Flat(r)
And the corresponding Python programs is given as below.
def append_node(t, n):
root = prev = Tree()
prev.set_mid(t)
while rearFull(t):
r = t.rear
t.rear = r[-1:] + [n]
n = wraps(r[:-1])
prev = t
t = t.mid
if t is None:
t = leaf(n)
elif len(t.rear) == 1 and t.front == []:
t = Tree(t.rear, None, [n])
else:
t = Tree(t.front, t.mid, t.rear + [n])
prev.set_mid(t)
return flat(root)

12.6.6

remove element from the tail of the sequence

Similar to appendT (), we can realize the algorithm which remove the last element from finger tree in symmetric manner of extractT ().

354

CHAPTER 12. SEQUENCES, THE LAST BRICK

We denote the non-empty, non-leaf finger tree as tree(F, M, R), where F is

the front finger, M is the middle part inner tree, and R is the rear finger.

T = leaf (x)
T = tree({y}, , {x})
T = tree(F, , {x}) F 6=
removeT (T ) =

T = tree(F, M, {x}), (M 0 , R0 ) = removeT (M )

otherwise
(12.29)
Function toList(T ) is used to flatten a 2-3 tree to plain list, which is defined
previously. Function init(L) returns all elements except for the last one in list
L, that if L = {a1 , a2 , ..., an1 , an }, init(L) = {a1 , a2 , ..., an1 }. And Function
last(L) returns the last element, so that last(L) = an . Please refer to the
appendix of this book for their implementation.
Algorithm removeT () can be translated to the following Haskell program,
we name it as unsnoc to indicate its the reverse function of snoc.
unsnoc
unsnoc
unsnoc
unsnoc
unsnoc
unsnoc

(, x)
(leaf (y), x)
(tree(init(F ), , last(F )), x)
(tree(F, M 0 , toList(R0 )), x)
(tree(F, M, init(R)), last(R))

:
:
:
:
:

:: Tree a (Tree a, a)
(Lf a) = (Empty, a)
(Tr [a] Empty [b]) = (Lf a, b)
(Tr f@(_:_) Empty [a]) = (Tr (init f) Empty [last f], a)
(Tr f m [a]) = (Tr f m (nodeToList r), a) where (m, r) = unsnoc m
(Tr f m r) = (Tr f m (init r), last r)

And we can define a special function last and init for finger tree which
is similar to their counterpart for list.
last = snd unsnoc
init = fst unsnoc

Imperatively removing the element from the end is almost as same as removing on the head. Although there seems to be a special case, that as we always
store the only element (or sub node) in the front finger while the rear finger
and middle part inner tree are empty (e.g. T ree({n}, N IL, )), it might get
nothing if always try to fetch the last element from rear finger.
This can be solved by swapping the front and the rear finger if the rear is
empty as in the following algorithm.
function Extract-Tail(T )
r Tree()
Connect-Mid(r, T )
while Rear(T ) = Mid(T ) 6= NIL do
T Mid(T )
if Rear(T ) = Front(T ) 6= then
Exchange Front(T ) Rear(T )
n Node()
Children(n) Rear(T )
repeat
L Children(n)
. L = {n1 , n2 , ..., nm1 , nm }
n Last(L)
. n nm
Rear(T ) L[1...m 1]
. {n1 , n2 , ..., nm1 }
T Parent(T )

12.6. FINGER TREE

355

if Mid(T ) becomes empty then

Mid(T ) NIL
until n is a leaf
return (Elem(n), Flat(r))
How to access the last element as well as implement this algorithm to working
program are left as exercises.

12.6.7

concatenate

Consider the none-trivial case that concatenate two finger trees T1 = tree(F1 , M1 , R1 )
and T2 = tree(F2 , M2 , R2 ). One natural idea is to use F1 as the new front finger
for the concatenated result, and keep R2 being the new rear finger. The rest of
work is to merge M1 , R1 , F2 and M2 to a new middle part inner tree.
Note that both R1 and F2 are plain lists of node, so the sub-problem is to
realize a algorithm like this.
merge(M1 , R1 F2 , M2 ) =?
More observation reveals that both M1 and M2 are also finger trees, except
that they are one level deeper than T1 and T2 in terms of N ode(a), where a is
the type of element stored in the tree. We can recursively use the strategy that
keep the front finger of M1 and the rear finger of M2 , then merge the middle
part inner tree of M1 , M2 , as well as the rear finger of M1 and front finger of
M2 .
If we denote function f ront(T ) returns the front finger, rear(T ) returns
the rear finger, mid(T ) returns the middle part inner tree. the above merge()
algorithm can be expressed for non-trivial case as the following.
merge(M1 , R1 F2 , M2 ) = tree(f ront(M1 ), S, rear(M2 ))
S = merge(mid(M1 ), rear(M1 ) R1 F2 f ront(M2 ), mid(M2 ))

(12.30)

If we look back to the original concatenate solution, it can be expressed as

below.
concat(T1 , T2 ) = tree(F1 , merge(M1 , R1 R2 , M2 ), R2 )

(12.31)

And compare it with equation 12.30, its easy to note the fact that concatenating is essentially merging. So we have the final algorithm like this.
concat(T1 , T2 ) = merge(T1 , , T2 )

(12.32)

By adding edge cases, the merge() algorithm can be completed as below.

merge(T1 , S, T2 ) =

f oldR(insertT, T2 , S) : T1 =
f oldL(appendT, T1 , S) : T2 =
merge(, {x} S, T2 ) : T1 = leaf (x)
merge(T1 , S {x}, ) : T2 = leaf (x)
tree(F1 , merge(M1 , nodes(R1 S F2 ), M 2), R2 ) : otherwise
(12.33)

356

CHAPTER 12. SEQUENCES, THE LAST BRICK

Most of these cases are straightforward. If any one of T1 or T2 is empty,

the algorithm repeatedly insert/append all elements in S to the other tree;
Function f oldL and f oldR are kinds of for-each process in imperative settings.
The difference is that f oldL processes the list S from left to right while f oldR
processes from right to left.
Here are their definition. Suppose list L = {a1 , a2 , ..., an1 , an }, L0 =
{a2 , a3 , ..., an1 , an } is the rest of elements except for the first one.

f oldL(f, e, L) =

f oldR(f, e, L) =

e : L=
f oldL(f, f (e, a1 ), L0 ) : otherwise

(12.34)

e : L=
f (a1 , f oldR(f, e, L0 )) : otherwise

(12.35)

They are detailed explained in the appendix of this book.

If either one of the tree is a leaf, we can insert or append the element of this
leaf to S, so that it becomes the trivial case of concatenating one empty tree
with another.
Function nodes() is used to wrap a list of elements to a list of 2-3 trees.
This is because the contents of middle part inner tree, compare to the contents
of finger, are one level deeper in terms of N ode(). Consider the time point
that transforms from recursive case to edge case. Lets suppose M1 is empty at
that time, we then need repeatedly insert all elements from R1 S F2 to M2 .
However, we cant directly do the insertion. If the element type is a, we can
only insert N ode(a) which is 2-3 tree to M2 . This is just like what we did in
the insertT () algorithm, take out the last 3 elements, wrap them in a 2-3 tree,
and recursive perform insertT (). Here is the definition of nodes().

{tr2(x1 , x2 )}
{tr3(x1 , x2 , x3 )}
nodes(L) =
{tr2(x1 , x2 ), tr2(x3 , x4 )}

{tr3(x1 , x2 , x3 )} nodes({x4 , x5 , ...})

L = {x1 , x2 }
L = {x1 , x2 , x3 }
L = {x1 , x2 , x3 , x4 }
otherwise
(12.36)
Function nodes() follows the constraint of 2-3 tree, that if there are only 2
or 3 elements in the list, it just wrap them in singleton list contains a 2-3 tree;
If there are 4 elements in the lists, it split them into two trees each is consist
of 2 branches; Otherwise, if there are more elements than 4, it wraps the first
three in to one tree with 3 branches, and recursively call nodes() to process the
rest.
The performance of concatenation is determined by merging. Analyze the
recursive case of merging reveals that the depth of recursion is proportion to the
smaller height of the two trees. As the tree is ensured to be balanced by using
2-3 tree. its height is bound to O(lg N 0 ) where N 0 is the number of elements.
The edge case of merging performs as same as insertion, (It calls insertT () at
most 8 times) which is amortized O(1) time, and O(lg M ) at worst case, where
M is the difference in height of the two trees. So the overall performance is
bound to O(lg N ), where N is the total number of elements contains in two
finger trees.
The following Haskell program implements the concatenation algorithm.
:
:
:
:

12.6. FINGER TREE

357

concat :: Tree a Tree a Tree a

concat t1 t2 = merge t1 [] t2

Note that there is concat function defined in prelude standard library, so

we need distinct them either by hiding import or take a different name.
merge :: Tree a [a] Tree a Tree a
merge Empty ts t2 = foldr cons t2 ts
merge t1 ts Empty = foldl snoc t1 ts
merge (Lf a) ts t2 = merge Empty (a:ts) t2
merge t1 ts (Lf a) = merge t1 (ts++[a]) Empty
merge (Tr f1 m1 r1) ts (Tr f2 m2 r2) = Tr f1 (merge m1 (nodes (r1 ++ ts ++
f2)) m2) r2

And the implementation of nodes() is as below.

nodes
nodes
nodes
nodes
nodes

:: [a] [Node a]
[a, b] = [Br2 a b]
[a, b, c] = [Br3 a b c]
[a, b, c, d] = [Br2 a b, Br2 c d]
(a:b:c:xs) = Br3 a b c:nodes xs

To concatenate two finger trees T1 and T2 in imperative approach, we can

traverse the two trees along with the middle part inner tree till either tree turns
to be empty. In every iteration, we create a new tree T , choose the front finger
of T1 as the front finger of T ; and choose the rear finger of T2 as the rear finger
of T . The other two fingers (rear finger of T1 and front finger of T2 ) are put
together as a list, and this list is then balanced grouped to several 2-3 tree nodes
as N . Note that N grows along with traversing not only in terms of length, the
depth of its elements increases by one in each iteration. We attach this new tree
as the middle part inner tree of the upper level result tree to end this iteration.
Once either tree becomes empty, we stop traversing, and repeatedly insert
the 2-3 tree nodes in N to the other non-empty tree, and set it as the new
middle part inner tree of the upper level result.
Below algorithm describes this process in detail.
function Concat(T1 , T2 )
return Merge(T1 , , T2 )
function Merge(T1 , N, T2 )
r Tree()
pr
while T1 6= NIL T2 6= NIL do
T Tree()
Front(T ) Front(T1 )
Rear(T ) Rear(T2 )
Connect-Mid(p, T )
pT
N Nodes(Rear(T1 ) N Front(T2 ))
T1 Mid(T1 )
T2 Mid(T2 )
if T1 = NIL then
T T2
for each n Reverse(N ) do

358

CHAPTER 12. SEQUENCES, THE LAST BRICK

T Prepend-Node(n, T )
else if T2 = NIL then
T T1
for each n N do
T Append-Node(T, n)
Connect-Mid(p, T )
return Flat(r)

Note that the for-each loops in the algorithm can also be replaced by folding
from left and right respectively. Translating this algorithm to Python program
yields the below code.
def concat(t1, t2):
return merge(t1, [], t2)
def merge(t1, ns, t2):
root = prev = Tree() #sentinel dummy tree
while t1 is not None and t2 is not None:
t = Tree(t1.size + t2.size + sizeNs(ns), t1.front, None, t2.rear)
prev.set_mid(t)
prev = t
ns = nodes(t1.rear + ns + t2.front)
t1 = t1.mid
t2 = t2.mid
if t1 is None:
prev.set_mid(foldR(prepend_node, ns, t2))
elif t2 is None:
prev.set_mid(reduce(append_node, ns, t1))
return flat(root)

Because Python only provides folding function from left as reduce(), a

folding function from right is given like what we shown in pseudo code, that it
repeatedly applies function in reverse order of the list.
def foldR(f, xs, z):
for x in reversed(xs):
z = f(x, z)
return z

The only function in question is how to balanced-group nodes to bigger 2-3

trees. As a 2-3 tree can hold at most 3 sub trees, we can firstly take 3 nodes
and wrap them to a ternary tree if there are more than 4 nodes in the list and
continuously deal with the rest. If there are just 4 nodes, they can be wrapped
to two binary trees. For other cases (there are 3 trees, 2 trees, 1 tree), we simply
wrap them all to a tree.
Denote node list L = {n1 , n2 , ...}, The following algorithm realizes this process.
function Nodes(L)
N =
while |L| > 4 do
n Node()
Children(n) L[1..3]
. {n1 , n2 , n3 }
N N {n}
L L[4...]
. {n4 , n5 , ...}

12.6. FINGER TREE

359

if |L| = 4 then
x Node()
Children(x) {L[1], L[2]}
y Node()
Children(y) {L[3], L[4]}
N N {x, y}
else if L 6= then
n Node()
Children(n) L
N N {n}
return N
Its straight forward to translate the algorithm to below Python program.
Where function wraps() helps to create an empty node, then set a list as the
children of this node.
def nodes(xs):
res = []
while len(xs) > 4:
res.append(wraps(xs[:3]))
xs = xs[3:]
if len(xs) == 4:
res.append(wraps(xs[:2]))
res.append(wraps(xs[2:]))
elif xs != []:
res.append(wraps(xs))
return res

Exercise 12.5
1. Implement the complete finger tree insertion program in your favorite
imperative programming language. Dont check the example programs
along with this chapter before having a try.
2. How to determine a node is a leaf? Does it contain only a raw element
inside or a compound node, which contains sub nodes as children? Note
that we cant distinguish it by testing the size, as there is case that node
contains a singleton leaf, such as node(1, {node(1, {x}}). Try to solve this
problem in both dynamic typed language (e.g. Python, lisp etc) and in
strong static typed language (e.g. C++).
3. Implement the Extract-Tail algorithm in your favorite imperative programming language.
4. Realize algorithm to return the last element of a finger tree in both functional and imperative approach. The later one should be able to handle
ill-formed tree.
5. Try to implement concatenation algorithm without using folding. You can
either use recursive methods, or use imperative for-each method.

360

CHAPTER 12. SEQUENCES, THE LAST BRICK

12.6.8

Random access of finger tree

size augmentation
The strategy to provide fast random access, is to turn the looking up into treesearch. In order to avoid calculating the size of tree many times, we augment
an extra field to tree and node. The definition should be modified accordingly,
for example the following Haskell definition adds size field in its constructor.
data Tree a = Empty
| Lf a
| Tr Int [a] (Tree (Node a)) [a]

And the previous ANSI C structure is augmented with size as well.

struct Tree {
union Node front;
union Node rear;
Tree mid;
Tree parent;
int size;
};

Suppose the function tree(s, F, M, R) creates a finger tree from size s, front
finger F , rear finger R, and middle part inner tree M . When the size of the tree
is needed, we can call a size(T ) function. It will be something like this.

0 : T =
? : T = leaf (x)
size(T ) =

s : T = tree(s, F, M, R)
If the tree is empty, the size is definitely zero; and if it can be expressed as
tree(s, F, M, R), the size is s; however, what if the tree is a singleton leaf? is
it 1? No, it can be 1 only if T = leaf (a) and a isnt a tree node, but a raw
element stored in finger tree. In most cases, the size is not 1, because a can be
again a tree node. Thats why we put a ? in above equation.
The correct way is to call some size function on the tree node as the following.

0 : T =

size0 (x) : T = leaf (x)

(12.37)
size(T ) =

s : T = tree(s, F, M, R)
Note that this isnt a recursive definition since size 6= size0 , the argument
to size0 is either a tree node, which is a 2-3 tree, or a plain element stored in
the finger tree. To uniform these two cases, we can anyway wrap the single
plain element to a tree node of only one element. So that we can express all
the situation as a tree node augmented with a size field. The following Haskell
program modifies the definition of tree node.
data Node a = Br Int [a]

The ANSI C node definition is modified accordingly.

struct Node {
Key key;
struct Node children;
int size;
};

12.6. FINGER TREE

361

We change it from union to structure. Although there is a overhead field

key if the node isnt a leaf.
Suppose function tr(s, L), creates such a node (either one element being
wrapped or a 2-3 tree) from a size information s, and a list L. Here are some
example.
tr(1, {x})
a tree contains only one element
tr(2, {x, y})
a 2-3 tree contains two elements
tr(3, {x, y, z}) a 2-3 tree contains three elements
So the function size0 can be implemented as returning the size information
of a tree node. We have size0 (tr(s, L)) = s.
Wrapping an element x is just calling tr(1, {x}). We can define auxiliary
functions wrap and unwrap, for instance.
wrap(x) = tr(1, {x})
unwrap(n) = x : n = tr(1, {x})

(12.38)

As both front finger and rear finger are lists of tree nodes, in order to calculate the total size of finger, we can provide a size00 (L) function, which sums up
size of all nodes stored in the list. Denote L = {a1 , a2 , ...} and L0 = {a2 , a3 , ...}.

0 : L=
00
size (L) =
(12.39)
size0 (a1 ) + size00 (L0 ) : otherwise
Its quite OK to define size00 (L) by using some high order functions. For
example.
size00 (L) = sum(map(size0 , L))

(12.40)

And we can turn a list of tree nodes into one deeper 2-3 tree and vice-versa.
wraps(L) = tr(size00 (L), L)
unwraps(n) = L : n = tr(s, L)

(12.41)

These helper functions are translated to the following Haskell code.

size (Br s _) = s
sizeL = sum (map size)
sizeT Empty = 0
sizeT (Lf a) = size a
sizeT (Tr s _ _ _) = s

Here are the wrap and unwrap auxiliary functions.

wrap x = Br 1 [x]
unwrap (Br 1 [x]) = x
wraps xs = Br (sizeL xs) xs
unwraps (Br _ xs) = xs

We omitted their type definitions for illustration purpose.

In imperative settings, the size information for node and tree can be accessed
through the size field. And the size of a list of nodes can be summed up for this
field as the below algorithm.

362

CHAPTER 12. SEQUENCES, THE LAST BRICK

function Size-Nodes(L)
s0
for n L do
s s+ Size(n)
return s
The following Python code, for example, translates this algorithm by using
standard sum() and map() functions provided in library.
def sizeNs(xs):
return sum(map(lambda x: x.size, xs))

As NIL is typically used to represent empty tree in imperative settings, its

convenient to provide a auxiliary size function to uniformed calculate the size
of tree no matter it is NIL.
function Size-Tr(T )
if T = NIL then
return 0
else
return Size(T )
The algorithm is trivial and we skip its implementation example program.
Modification due to the augmented size
The algorithms have been presented so far need to be modified to accomplish
with the augmented size. For example the insertT () function now inserts a tree
node instead of a plain element.
insertT (x, T ) = insertT 0 (wrap(x), T )

(12.42)

The corresponding Haskell program is changed as below.

cons a t = cons (wrap a) t

After being wrapped, x is augmented with size information of 1. In the

implementation of previous insertion algorithm, function tree(F, M, R) is used
to create a finger tree from a front finger, a middle part inner tree and a rear
finger. This function should also be modified to add size information of these
three arguments.

M =R=
M =F =
F = , (F 0 , M 0 ) = extractT 0 (M
tree0 (F, M, R) =

R = , (M 0 , R0 ) = removeT 0 (M

otherwise
(12.43)
Where f romL() helps to turn a list of nodes to a finger tree by repeatedly
inserting all the element one by one to an empty tree.
f romL(F )
f romL(R)
tree0 (unwraps(F 0 ), M 0 , R)
tree0 (F, M 0 , unwraps(R0 ))
00
tree(size (F ) + size(M ) + size00 (R), F, M, R)

:
:
:
:
:

f romL(L) = f oldR(insertT 0 , , L)
Of course it can be implemented in pure recursive manner without using
folding as well.

12.6. FINGER TREE

363

The last case is the most straightforward one. If none of F , M , and R is

empty, it adds the size of these three part and construct the tree along with
this size information by calling tree(s, F, M, R) function. If both the middle
part inner tree and one of the finger is empty, the algorithm repeatedly insert
all elements stored in the other finger to an empty tree, so that the result is
constructed from a list of tree nodes. If the middle part inner tree isnt empty,
and one of the finger is empty, the algorithm borrows one tree node from the
middle part, either by extracting from head if front finger is empty or removing
from tail if rear finger is empty. Then the algorithm unwraps the borrowed
tree node to a list, and recursively call tree0 () function to construct the result.
This algorithm can be translated to the following Haskell code for example.
tree
tree
tree
tree
tree

f Empty [] = foldr cons Empty f

[] Empty r = foldr cons Empty r
[] m r = let (f, m) = uncons m in tree (unwraps f) m r
f m [] = let (m, r) = unsnoc m in tree f m (unwraps r)
f m r = Tr (sizeL f + sizeT m + sizeL r) f m r

Function tree0 () helps to minimize the modification. insertT 0 () can be realized by using it like the following.

leaf (x) : T =
0
tree
({x},
, {y}) : T = leaf (x)
insertT 0 (x, T ) =
0
0
tree
({x,
x
},
insertT
(wraps({x
,
x
,
x
}),
M ), R) : T = tree(s, {x1 , x2 , x3 , x4 }, M, R)

1
2
3
4

tree0 ({x} F, M, R) : otherwise

(12.44)
And its corresponding Haskell code is a line by line translation.
cons
cons
cons
cons

a
a
a
a

Empty = Lf a
(Lf b) = tree [a] Empty [b]
(Tr _ [b, c, d, e] m r) = tree [a, b] (cons (wraps [c, d, e]) m) r
(Tr _ f m r) = tree (a:f) m r

The similar modification for augment size should also be tuned for imperative
algorithms, for example, when a new node is prepend to the head of the finger
tree, we should update size when traverse the tree.
function Prepend-Node(n, T )
r Tree()
pr
Connect-Mid(p, T )
while Full?(Front(T )) do
F Front(T )
Front(T ) {n, F [1]}
Size(T ) Size(T ) + Size(n)
. update size
n Node()
Children(n) F [2..]
pT
T Mid(T )
if T = NIL then
T Tree()
Front(T ) {n}
else if | Front(T ) | = 1 Rear(T ) = then

364

CHAPTER 12. SEQUENCES, THE LAST BRICK

Rear(T ) Front(T )
Front(T ) {n}
else
Front(T ) {n} Front(T )
Size(T ) Size(T ) + Size(n)
Connect-Mid(p, T ) T
return Flat(r)

. update size

The corresponding Python code are modified accordingly as below.

def prepend_node(n, t):
root = prev = Tree()
prev.set_mid(t)
while frontFull(t):
f = t.front
t.front = [n] + f[:1]
t.size = t.size + n.size
n = wraps(f[1:])
prev = t
t = t.mid
if t is None:
t = leaf(n)
elif len(t.front)==1 and t.rear == []:
t = Tree(n.size + t.size, [n], None, t.front)
else:
t = Tree(n.size + t.size, [n]+t.front, t.mid, t.rear)
prev.set_mid(t)
return flat(root)

Note that the tree constructor is also modified to take a size argument as
the first parameter. And the leaf() helper function does not only construct
the tree from a node, but also set the size of the tree with the same size of the
node inside it.
For simplification purpose, we skip the detailed description of what are modified in extractT ()0 , appendT (), removeT (), and concat() algorithms. They are
left as exercises to the reader.

Split a finger tree at a given position

With size information augmented, its easy to locate a node at given position
by performing a tree search. Whats more, as the finger tree is constructed from
three part F , M , and R; and its nature of recursive, its also possible to split
it into three sub parts with a given position i: the left, the node at i, and the
right part.
The idea is straight forward. Since we have the size information for F ,
M , and R. Denote these three sizes as Sf , Sm , and Sr . if the given position
i Sf , the node must be stored in F , we can go on seeking the node inside F ; if
Sf < i Sf + Sm , the node must be stored in M , we need recursively perform
search in M ; otherwise, the node should be in R, we need search inside R.
If we skip the error handling of trying to split an empty tree, there is only
one edge case as below.

12.6. FINGER TREE

365

splitAt(i, T ) =

(, x, ) : T = leaf (x)
... : otherwise

Splitting a leaf results both the left and right parts empty, the node stored
in leaf is the resulting node.
The recursive case handles the three sub cases by comparing i with the
sizes. Suppose function splitAtL(i, L) splits a list of nodes at given position i
into three parts: (A, x, B) = splitAtL(i, L), where x is the i-th node in L, A is
a sub list contains all nodes before position i, and B is a sub list contains all
rest nodes after i.

(, x, )
(f romL(A), x, tree0 (B, M, R)
splitAt(i, T ) =
(tree0 (F, Ml , A), x, tree0 (B, Mr , R)

(tree0 (F, M, A), x, f romL(B))

:
:
:
:

T = leaf (x)
i Sf , (A, x, B) = splitAtL(i, F )
Sf < i Sf + Sm
otherwise, (A, x, B) = splitAtL(i Sf Sm , R)
(12.45)
Where Ml , x, Mr , A, B in the thrid case are calculated as the following.
(Ml , t, Mr ) = splitAt(i Sf , M )
(A, x, B) = splitAtL(i Sf size(Ml ), unwraps(t))

And the function splitAtL() is just a linear traverse, since the length of list is
limited not to exceed the constraint of 2-3 tree, the performance is still ensured
to be constant O(1) time. Denote L = {x1 , x2 , ...} and L0 = {x2 , x3 , ...}.

(, x1 , ) : i = 0 L = {x1 }
(, x1 , L0 ) : i < size0 (x1 )

({x1 } A, x, B) : otherwise, (A, x, B) = splitAtL(i size0 (x1 ), L0 )

(12.46)
The solution of splitting is a typical divide and conquer strategy. The performance of this algorithm is determined by the recursive case of searching in
middle part inner tree. Other cases are all constant time as weve analyzed. The
depth of recursion is proportion to the height of the tree h, so the algorithm is
bound to O(h). Because the tree is well balanced (by using 2-3 tree, and all
the insertion/removal algorithms keep the tree balanced), so h = O(lg N ) where
N is the number of elements stored in finger tree. The overall performance of
splitting is O(lg N ).
Lets first give the Haskell program for splitAtL() function

splitAtL(i, L) =

splitNodesAt 0 [x] = ([], x, [])

splitNodesAt i (x:xs) | i < size x = ([], x, xs)
| otherwise = let (xs, y, ys) = splitNodesAt (i-size x) xs
in (x:xs, y, ys)

Then the program for splitAt(), as there is already function defined in standard library with this name, we slightly change the name by adding a apostrophe.
splitAt _ (Lf x) = (Empty, x, Empty)
splitAt i (Tr _ f m r)
| i < szf = let (xs, y, ys) = splitNodesAt i f

366

CHAPTER 12. SEQUENCES, THE LAST BRICK

in ((foldr cons Empty xs), y, tree ys m r)
| i < szf + szm = let (m1, t, m2) = splitAt (i-szf) m
(xs, y, ys) = splitNodesAt (i-szf - sizeT m1) (unwraps t)
in (tree f m1 xs, y, tree ys m2 r)
| otherwise = let (xs, y, ys) = splitNodesAt (i-szf -szm) r
in (tree f m xs, y, foldr cons Empty ys)
where
szf = sizeL f
szm = sizeT m

Random access
With the help of splitting at any arbitrary position, its trivial to realize random
access in O(lg N ) time. Denote function mid(x) returns the 2-nd element of a
tuple, lef t(x), and right(x) return the first element and the 3-rd element of the
tuple respectively.
getAt(S, i) = unwrap(mid(splitAt(i, S)))

(12.47)

It first splits the sequence at position i, then unwraps the node to get the element stored inside it. When mutate the i-th element of sequence S represented
by finger tree, we first split it at i, then we replace the middle to what we want
to mutate, and re-construct them to one finger tree by using concatenation.
setAt(S, i, x) = concat(L, insertT (x, R))

(12.48)

where
(L, y, R) = splitAt(i, S)
Whats more, we can also realize a removeAt(S, i) function, which can remove the i-th element from sequence S. The idea is first to split at i, unwrap
and return the element of the i-th node; then concatenate the left and right to
a new finger tree.
removeAt(S, i) = (unwrap(y), concat(L, R))

(12.49)

These handy algorithms can be translated to the following Haskell program.

getAt t i = unwrap x where (_, x, _) = splitAt i t
setAt t i x = let (l, _, r) = splitAt i t in concat l (cons x r)
removeAt t i = let (l, x, r) = splitAt i t in (unwrap x, concat l r)

Imperative random access

As we can directly mutate the tree in imperative settings, its possible to realize
Get-At(T, i) and Set-At(T, i, x) without using splitting. The idea is firstly
implement a algorithm which can apply some operation to a given position. The
following algorithm takes three arguments, a finger tree T , a position index at
i which ranges from zero to the number of elements stored in the tree, and a
function f , which will be applied to the element at i.
function Apply-At(T, i, f )
while Size(T ) > 1 do
Sf Size-Nodes(Front(T ))

12.6. FINGER TREE

367

Sm Size-Tr(Mid(T ))
if i < Sf then
return Lookup-Nodes(Front(T ), i, f )
else if i < Sf + Sm then
T Mid(T )
i i Sf
else
return Lookup-Nodes(Rear(T ), i Sf Sm , f )
n First-Lf(T )
x Elem(n)
Elem(n) f (x)
return x
This algorithm is essentially a divide and conquer tree search. It repeatedly
examine the current tree till reach a tree with size of 1 (can it be determined
as a leaf? please consider the ill-formed case and refer to the exercise later).
Every time, it checks the position to be located with the size information of
front finger and middle part inner tree.
If the index i is less than the size of front finger, the location is at some
node in it. The algorithm call a sub procedure to look-up among front finger;
If the index is between the size of front finger and the total size till middle part
inner tree, it means that the location is at some node inside the middle, the
algorithm goes on traverse along the middle part inner tree with an updated
index reduced by the size of front finger; Otherwise it means the location is at
some node in rear finger, the similar looking up procedure is called accordingly.
After this loop, weve got a node, (can be a compound node) with what we
are looking for at the first leaf inside this node. We can extract the element
out, and apply the function f on it and store the new value back.
The algorithm returns the previous element before applying f as the final
result.
What hasnt been factored is the algorithm Lookup-Nodes(L, i, f ). It
takes a list of nodes, a position index, and a function to be applied. This
algorithm can be implemented by checking every node in the list. If the node is
a leaf, and the index is zero, we are at the right position to be looked up. The
function can be applied on the element stored in this leaf, and the previous value
is returned; Otherwise, we need compare the size of this node and the index to
determine if the position is inside this node and search inside the children of the
node if necessary.
function Lookup-Nodes(L, i, f )
loop
for n L do
if n is leaf i = 0 then
x Elem(n)
Elem(n) f (x)
return x
if i < Size(n) then
L Children(n)
break
i i Size(n)
The following are the corresponding Python code implements the algorithms.

368

CHAPTER 12. SEQUENCES, THE LAST BRICK

def applyAt(t, i, f):

while t.size > 1:
szf = sizeNs(t.front)
szm = sizeT(t.mid)
if i < szf:
return lookupNs(t.front, i, f)
elif i < szf + szm:
t = t.mid
i = i - szf
else:
return lookupNs(t.rear, i - szf - szm, f)
n = first_leaf(t)
x = elem(n)
n.children[0] = f(x)
return x
def lookupNs(ns, i, f):
while True:
for n in ns:
if n.leaf and i == 0:
x = elem(n)
n.children[0] = f(x)
return x
if i < n.size:
ns = n.children
break
i = i - n.size

With auxiliary algorithm that can apply function at a given position, its
trivial to implement the Get-At() and Set-At() by passing special function
for applying.
function Get-At(T, i)
return Apply-At(T, i, x .x)
function Set-At(T, i, x)
return Apply-At(T, i, y .x)
That is we pass id function to implement getting element at a position,
which doesnt change anything at all; and pass constant function to implement
setting, which set the element to new value by ignoring its previous value.
Imperative splitting
Its not enough to just realizing Apply-At algorithm in imperative settings,
this is because removing element at arbitrary position is also a typical case.
Almost all the imperative finger tree algorithms so far are kind of one-pass
top-down manner. Although we sometimes need to book keeping the root. It
means that we can even realize all of them without using the parent field.
Splitting operation, however, can be easily implemented by using parent
field. We can first perform a top-down traverse along with the middle part
inner tree as long as the splitting position doesnt located in front or rear finger.
After that, we need a bottom-up traverse along with the parent field of the two
split trees to fill out the necessary fields.

12.6. FINGER TREE

369

function Split-At(T, i)
T1 Tree()
T2 Tree()
while Sf i < Sf + Sm do
. Top-down pass
T10 Tree()
T20 Tree()
Front(T10 ) Front(T )
Rear(T20 ) Rear(T )
Connect-Mid(T1 , T10 )
Connect-Mid(T2 , T20 )
T1 T10
T2 T20
i i Sf
T Mid(T )
if i < Sf then
(X, n, Y ) Split-Nodes(Front(T ), i)
T10 From-Nodes(X)
T20 T
Size(T20 ) Size(T ) - Size-Nodes(X) - Size(n)
Front(T20 ) Y
else if Sf + Sm i then
(X, n, Y ) Split-Nodes(Rear(T ), i Sf Sm )
T20 From-Nodes(Y )
T10 T
Size(T10 ) Size(T ) - Size-Nodes(Y ) - Size(n)
Rear(T10 ) X
Connect-Mid(T1 , T10 )
Connect-Mid(T2 , T20 )
i i Size-Tr(T10 )
while n is NOT leaf do
. Bottom-up pass
(X, n, Y ) Split-Nodes(Children(n), i)
i i Size-Nodes(X)
Rear(T1 ) X
Front(T2 ) Y
Size(T1 ) Sum-Sizes(T1 )
Size(T2 ) Sum-Sizes(T2 )
T1 Parent(T1 )
T2 Parent(T2 )
return (Flat(T1 ), Elem(n), Flat(T2 ))
The algorithm first creates two trees T1 and T2 to hold the split results. Note
that they are created as ground trees which are parents of the roots. The first
pass is a top-down pass. Suppose Sf , and Sm retrieve the size of the front finger
and the size of middle part inner tree respectively. If the position at which the
tree to be split is located at middle part inner tree, we reuse the front finger of
T for new created T10 , and reuse rear finger of T for T20 . At this time point, we
cant fill the other fields for T10 and T20 , they are left empty, and well finish filling
them in the future. After that, we connect T1 and T10 so the latter becomes the
middle part inner tree of the former. The similar connection is done for T2 and
T20 as well. Finally, we update the position by deducing it by the size of front

370

CHAPTER 12. SEQUENCES, THE LAST BRICK

finger, and go on traversing along with the middle part inner tree.
When the first pass finishes, we are at a position that either the splitting
should be performed in front finger, or in rear finger. Splitting the nodes in
finger results a tuple, that the first part and the third part are lists before and
after the splitting point, while the second part is a node contains the element at
the original position to be split. As both fingers hold at most 3 nodes because
they are actually 2-3 trees, the nodes splitting algorithm can be performed by
a linear search.
function Split-Nodes(L, i)
for j [1, Length(L) ] do
if i < Size(L[j]) then
return (L[1...j 1], L[j], L[j + 1... Length(L) ])
i i Size(L[j])
We next create two new result trees T10 and T20 from this tuple, and connected
them as the final middle part inner tree of T1 and T2 .
Next we need perform a bottom-up traverse along with the result trees to
fill out all the empty information we skipped in the first pass.
We loop on the second part of the tuple, the node, till it becomes a leaf. In
each iteration, we repeatedly splitting the children of the node with an updated
position i. The first list of nodes returned from splitting is used to fill the rear
finger of T1 ; and the other list of nodes is used to fill the front finger of T2 .
After that, since all the three parts of a finger tree the front and rear finger,
and the middle part inner tree are filled, we can then calculate the size of the
tree by summing these three parts up.
function Sum-Sizes(T )
return Size-Nodes(Front(T )) + Size-Tr(Mid(T )) + Size-Nodes(Rear(T ))
Next, the iteration goes on along with the parent fields of T1 and T2 . The
last black-box algorithm is From-Nodes(L), which can create a finger tree
from a list of nodes. It can be easily realized by repeatedly perform insertion
on an empty tree. The implementation is left as an exercise to the reader.
The example Python code for splitting is given as below.
def splitAt(t, i):
(t1, t2) = (Tree(), Tree())
while szf(t) i and i < szf(t) + szm(t):
fst = Tree(0, t.front, None, [])
snd = Tree(0, [], None, t.rear)
t1.set_mid(fst)
t2.set_mid(snd)
(t1, t2) = (fst, snd)
i = i - szf(t)
t = t.mid
if i < szf(t):
(xs, n, ys) = splitNs(t.front, i)
sz = t.size - sizeNs(xs) - n.size
(fst, snd) = (fromNodes(xs), Tree(sz, ys, t.mid, t.rear))
elif szf(t) + szm(t) i:
(xs, n, ys) = splitNs(t.rear, i - szf(t) - szm(t))
sz = t.size - sizeNs(ys) - n.size

12.7. NOTES AND SHORT SUMMARY

371

(fst, snd) = (Tree(sz, t.front, t.mid, xs), fromNodes(ys))

t1.set_mid(fst)
t2.set_mid(snd)
i = i - sizeT(fst)
while not n.leaf:
(xs, n, ys) = splitNs(n.children, i)
i = i - sizeNs(xs)
(t1.rear, t2.front) = (xs, ys)
t1.size = sizeNs(t1.front) + sizeT(t1.mid) + sizeNs(t1.rear)
t2.size = sizeNs(t2.front) + sizeT(t2.mid) + sizeNs(t2.rear)
(t1, t2) = (t1.parent, t2.parent)
return (flat(t1), elem(n), flat(t2))

The program to split a list of nodes at a given position is listed like this.
def splitNs(ns, i):
for j in range(len(ns)):
if i < ns[j].size:
return (ns[:j], ns[j], ns[j+1:])
i = i - ns[j].size

With splitting defined, removing an element at arbitrary position can be

realized trivially by first performing a splitting, then concatenating the two
result tree to one big tree and return the element at that position.
function Remove-At(T, i)
(T1 , x, T2 ) Split-At(T, i)
return (x, Concat(T1 , T2 ) )

Exercise 12.6
1. Another way to realize insertT 0 () is to force increasing the size field by
one, so that we neednt write function tree0 (). Try to realize the algorithm
by using this idea.
2. Try to handle the augment size information as well as in insertT 0 () algorithm for the following algorithms (both functional and imperative):
extractT ()0 , appendT (), removeT (), and concat(). The head, tail, init
and last functions should be kept unchanged. Dont refer to the downloadable programs along with this book before you take a try.
3. In the imperative Apply-At() algorithm, it tests if the size of the current
tree is greater than one. Why dont we test if the current tree is a leaf?
Tell the difference between these two approaches.
4. Implement the From-Nodes(L) in your favorite imperative programming
language. You can either use looping or create a folding-from-right sub
algorithm.

12.7

Notes and short summary

Although we havent been able to give a purely functional realization to match

the O(1) constant time random access as arrays in imperative settings. The

372

CHAPTER 12. SEQUENCES, THE LAST BRICK

result finger tree data structure achieves an overall well performed sequence. It
manipulates fast in amortized O(1) time both on head an on tail, it can also
concatenates two sequence in logarithmic time as well as break one sequence into
two sub sequences at any position. While neither arrays in imperative settings
nor linked-list in functional settings satisfies all these goals. Some functional
programming languages adopt this sequence realization in its standard library
[7].
Just as the title of this chapter, weve presented the last corner stone of elementary data structures in both functional and imperative settings. We neednt
concern about being lack of elementary data structures when solve problems
with some typical algorithms.
For example, when writing a MTF (move-to-front) encoding algorithm[8],
with the help of the sequence data structure explained in this chapter. We can
implement it quite straightforward.
mtf (S, i) = {x} S 0
where (x, S 0 ) = removeAt(S, i).
In the next following chapters, well first explains some typical divide and
conquer sorting methods, including quick sort, merge sort and their variants;
then some elementary searching algorithms, and string matching algorithms will
be covered.

Bibliography
[1] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[2] Chris Okasaki. Purely Functional Random-Access Lists. Functional Programming Languages and Computer Architecture, June 1995, pages 86-95.
[3] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. The MIT Press, 2001.
ISBN: 0262032937.
[4] Miran Lipovaca. Learn You a Haskell for Great Good! A Beginners
Guide. No Starch Press; 1 edition April 2011, 400 pp. ISBN: 978-1-59327283-8
[5] Ralf Hinze and Ross Paterson. Finger Trees: A Simple General-purpose
Data Structure. in Journal of Functional Programming16:2 (2006), pages
197-217. http://www.soi.city.ac.uk/ ross/papers/FingerTree.html
[6] Guibas, L. J., McCreight, E. M., Plass, M. F., Roberts, J. R. (1977), A
new representation for linear lists. Conference Record of the Ninth Annual
ACM Symposium on Theory of Computing, pp. 49C60.
[7] Generic finger-tree structure. http://hackage.haskell.org/packages/archive/fingertree/0.0/doc/html/DataFingerTree.html
[8] Wikipedia. Move-to-front transform. http://en.wikipedia.org/wiki/Moveto-front transform

373

374

BIBLIOGRAPHY

Part V

Sorting and Searching

375

Chapter 13

Divide and conquer, Quick

sort vs. Merge sort
13.1

Introduction

Its proved that the best approximate performance of comparison based sorting
is O(n lg n) [1]. In this chapter, two divide and conquer sorting algorithms are
introduced. Both of them perform in O(n lg n) time. One is quick sort. It is
the most popular sorting algorithm. Quick sort has been well studied, many
programming libraries provide sorting tools based on quick sort.
In this chapter, well first introduce the idea of quick sort, which demonstrates the power of divide and conquer strategy well. Several variants will be
explained, and well see when quick sort performs poor in some special cases.
That the algorithm is not able to partition the sequence in balance.
In order to solve the unbalanced partition problem, well next introduce
about merge sort, which ensure the sequence to be well partitioned in all the
cases. Some variants of merge sort, including nature merge sort, bottom-up
merge sort are shown as well.
Same as other chapters, all the algorithm will be realized in both imperative
and functional approaches.

13.2

Quick sort

Consider a teacher arranges a group of kids in kindergarten to stand in a line

for some game. The kids need stand in order of their heights, that the shortest
one stands on the left most, while the tallest stands on the right most. How can
the teacher instruct these kids, so that they can stand in a line by themselves?
There are many strategies, and the quick sort approach can be applied here:
1. The first kid raises his/her hand. The kids who are shorter than him/her
stands to the left to this child; the kids who are taller than him/her stands
to the right of this child;
2. All the kids move to the left, if there are, repeat the above step; all the
kids move to the right repeat the same step as well.
377

378CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT

Figure 13.1: Instruct kids to stand in a line

Suppose a group of kids with their heights as {102, 100, 98, 95, 96, 99, 101, 97}
with [cm] as the unit. The following table illustrate how they stand in order of
height by following this method.
102 100 98 95 96 99 101 97
100 98 95 96 99 101 97 102
98
95 96 99 97 100 101 102
95
96 97 98 99 100 101 102
95
96 97 98 99 100 101 102
95
96 97 98 99 100 101 102
95
96 97 98 99 100 101 102
At the beginning, the first child with height 102 cm raises his/her hand. We
call this kid the pivot and mark this height in bold. It happens that this is
the tallest kid. So all others stands to the left side, which is represented in the
second row in the above table. Note that the child with height 102 cm is in the
final ordered position, thus we mark it italic. Next the kid with height 100 cm
raise hand, so the children of heights 98, 95, 96 and 99 cm stand to his/her left,
and there is only 1 child of height 101 cm who is taller than this pivot kid. So he
stands to the right hand. The 3rd row in the table shows this stage accordingly.
After that, the child of 98 cm high is selected as pivot on left hand; while the
child of 101 cm high on the right is selected as pivot. Since there are no other
kids in the unsorted group with 101 cm as pivot, this small group is ordered
already and the kid of height 101 cm is in the final proper position. The same
method is applied to the group of kids which havent been in correct order until
all of them are stands in the final position.

13.2.1

Basic version

Summarize the above instruction leads to the recursive description of quick sort.
In order to sort a sequence of elements L.
If L is empty, the result is obviously empty; This is the trivial edge case;
Otherwise, select an arbitrary element in L as a pivot, recursively sort all
elements not greater than L, put the result on the left hand of the pivot,
and recursively sort all elements which are greater than L, put the result
on the right hand of the pivot.

13.2. QUICK SORT

379

Note that the emphasized word and, we dont use then here, which indicates
its quite OK that the recursive sort on the left and right can be done in parallel.
Well return this parallelism topic soon.
Quick sort was first developed by C. A. R. Hoare in 1960 [1], [15]. What
we describe here is a basic version. Note that it doesnt state how to select the
pivot. Well see soon that the pivot selection affects the performance of quick
sort dramatically.
The most simple method to select the pivot is always choose the first one so
that quick sort can be formalized as the following.

: L=
sort({x|x L0 , x l1 }) {l1 } sort({x|x L0 , l1 < x}) : otherwise
(13.1)
Where l1 is the first element of the non-empty list L, and L0 contains the
rest elements {l2 , l3 , ...}. Note that we use Zermelo Frankel expression (ZF expression for short), which is also known as list comprehension. A ZF expression
{a|a S, p1 (a), p2 (a), ...} means taking all element in set S, if it satisfies all
the predication p1 , p2 , .... ZF expression is originally used for representing set,
we extend it to express list for the sake of brevity. There can be duplicated
elements, and different permutations represent for different list. Please refer to
the appendix about list in this book for detail.
Its quite straightforward to translate this equation to real code if list comprehension is supported. The following Haskell code is given for example:

sort(L) =

sort [] = []
sort (x:xs) = sort [y | yxs, y x] ++ [x] ++ sort [y | yxs, x < y]

This might be the shortest quick sort program in the world at the time when
this book is written. Even a verbose version is still very expressive:
sort [] = []
sort (x:xs) = as ++ [x] ++ bs where
as = sort [ a | a xs, a x]
bs = sort [ b | b xs, x < b]

There are some variants of this basic quick sort program, such as using
explicit filtering instead of list comprehension. The following Python program
demonstrates this for example:
def sort(xs):
if xs == []:
return []
pivot = xs[0]
as = sort(filter(lambda x : x pivot, xs[1:]))
bs = sort(filter(lambda x : pivot < x, xs[1:]))
return as + [pivot] + bs

13.2.2

Strict weak ordering

We assume the elements are sorted in monotonic none decreasing order so far.
Its quite possible to customize the algorithm, so that it can sort the elements
in other ordering criteria. This is necessary in practice because users may sort
numbers, strings, or other complex objects (even list of lists for example).

380CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
The typical generic solution is to abstract the comparison as a parameter as
we mentioned in chapters about insertion sort and selection sort. Although it
neednt the total ordering, the comparison must satisfy strict weak ordering at
least [17] [16].
For the sake of brevity, we only considering sort the elements by using less
than or equal (equivalent to not greater than) in the rest of the chapter.

13.2.3

Partition

Observing that the basic version actually takes two passes to find all elements
which are greater than the pivot as well as to find the others which are not
respectively. Such partition can be accomplished by only one pass. We explicitly
define the partition as below.

(, ) : L =
({l1 } A, B) : p(l1 ), (A, B) = partition(p, L0 )

(A, {l1 } B) : p(l1 )

(13.2)
Note that the operation {x} L is just a cons operation, which only takes
constant time. The quick sort can be modified accordingly.
partition(p, L) =

: L=
sort(A) {l1 } sort(B) : otherwise, (A, B) = partition(x x l1 , L0 )
(13.3)
Translating this new algorithm into Haskell yields the below code.

sort(L) =

sort [] = []
sort (x:xs) = sort as ++ [x] ++ sort bs where
(as, bs) = partition ( x) xs
partition _ [] = ([], [])
partition p (x:xs) = let (as, bs) = partition p xs in
if p x then (x:as, bs) else (as, x:bs)

The concept of partition is very critical to quick sort. Partition is also very
important to many other sort algorithms. Well explain how it generally affects
the sorting methodology by the end of this chapter. Before further discussion
about fine tuning of quick sort specific partition, lets see how to realize it inplace imperatively.
There are many partition methods. The one given by Nico Lomuto [4] [2] will
be used here as its easy to understand. Well show other partition algorithms
soon and see how partitioning affects the performance.
Figure 13.2 shows the idea of this one-pass partition method. The array is
processed from left to right. At any time, the array consists of the following
parts as shown in figure 13.2 (a):
The left most cell contains the pivot; By the end of the partition process,
the pivot will be moved to the final proper position;
A segment contains all elements which are not greater than the pivot. The
right boundary of this segment is marked as left;

13.2. QUICK SORT

381

pivot

x[l]

left

...not greater than ...

right

... greater than ...

...?...x[u]

(a) Partition invariant

pivot

left

x[l]

right

x[l+1]

...?...x[u]

(b) Start

pivot

x[l]

left

...not greater than ...

x[left]

right

... greater than ...x[u]

swap

Figure 13.2: Partition a range of array by using the left most element as pivot.
A segment contains all elements which are greater than the pivot. The
right boundary of this segment is marked as right; It means that elements
between left and right marks are greater than the pivot;
The rest of elements after right mark havent been processed yet. They
may be greater than the pivot or not.
At the beginning of partition, the left mark points to the pivot and the
right mark points to the the second element next to the pivot in the array as
in Figure 13.2 (b); Then the algorithm repeatedly advances the right mark one
element after the other till passes the end of the array.
In every iteration, the element pointed by the right mark is compared with
the pivot. If it is greater than the pivot, it should be among the segment between
the left and right marks, so that the algorithm goes on to advance the right
mark and examine the next element; Otherwise, since the element pointed by
right mark is less than or equal to the pivot (not greater than), it should be
put before the left mark. In order to achieve this, the left mark needs be
advanced by one, then exchange the elements pointed by the left and right
marks.
Once the right mark passes the last element, it means that all the elements
have been processed. The elements which are greater than the pivot have been
moved to the right hand of left mark while the others are to the left hand of this
mark. Note that the pivot should move between the two segments. An extra
exchanging between the pivot and the element pointed by left mark makes this
final one to the correct location. This is shown by the swap bi-directional arrow

382CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
in figure 13.2 (c).
The left mark (which points the pivot finally) partitions the whole array
into two parts, it is returned as the result. We typically increase the left
mark by one, so that it points to the first element greater than the pivot for
convenient. Note that the array is modified in-place.
The partition algorithm can be described as the following. It takes three
arguments, the array A, the lower and the upper bound to be partitioned 1 .
1: function Partition(A, l, u)
2:
p A[l]
. the pivot
3:
Ll
. the left mark
4:
for R [l + 1, u] do
. iterate on the right mark
5:
if (p < A[R]) then . negate of < is enough for strict weak order
6:
LL+1
7:
Exchange A[L] A[R]
8:
Exchange A[L] p
9:
return L + 1
. The partition position

Below table shows the steps of partitioning the array {3, 2, 5, 4, 0, 1, 6, 7}.
(l) 3
(r) 2
5
4
0
1
6
7
initialize, pivot = 3, l = 1, r = 2
3
(l)(r) 2
5
4
0
1
6
7
2 < 3, advances l, (r = l)
3
(l) 2
(r) 5
4
0
1
6
7
5 > 3, moves on
3
(l) 2
5
(r) 4
0
1
6
7
4 > 3, moves on
3
(l) 2
5
4
(r) 0
1
6
7
0<3
3
2
(l) 0
4
(r) 5
1
6
7
Advances l, then swap with r
3
2
(l) 0
4
5
(r) 1
6
7
1<3
3
2
0
(l) 1
5
(r) 4
6
7
Advances l, then swap with r
3
2
0
(l) 1
5
4
(r) 6
7
6 > 3, moves on
3
2
0
(l) 1
5
4
6
(r) 7 7 > 3, moves on
1
2
0
3
(l+1) 5
4
6
7
r passes the end, swap pivot and l
This version of partition algorithm can be implemented in ANSI C as the
following.
int partition(Key xs, int l, int u) {
int pivot, r;
for (pivot = l, r = l + 1; r < u; ++r)
if (!(xs[pivot] < xs[r])) {
++l;
swap(xs[l], xs[r]);
}
swap(xs[pivot], xs[l]);
return l + 1;
}

Where swap(a, b) can either be defined as function or a macro. In ISO

C++, swap(a, b) is provided as a function template. the type of the elements
can be defined somewhere or abstracted as a template parameter in ISO C++.
We omit these language specific details here.
With the in-place partition realized, the imperative in-place quick sort can
be accomplished by using it.
1 The partition algorithm used here is slightly different from the one in [2]. The latter uses
the last element in the slice as the pivot.

13.2. QUICK SORT

1:
2:
3:
4:
5:

383

procedure Quick-Sort(A, l, u)
if l < u then
m Partition(A, l, u)
Quick-Sort(A, l, m 1)
Quick-Sort(A, m, u)

When sort an array, this procedure is called by passing the whole range as
the lower and upper bounds. Quick-Sort(A, 1, |A|). Note that when l u it
means the array slice is either empty, or just contains only one element, both
can be treated as ordered, so the algorithm does nothing in such cases.
Below ANSI C example program completes the basic in-place quick sort.
void quicksort(Key xs, int l, int u) {
int m;
if (l < u) {
m = partition(xs, l, u);
quicksort(xs, l, m - 1);
quicksort(xs, m, u);
}
}

13.2.4

Minor improvement in functional partition

Before exploring how to improve the partition for basic version quick sort, its
obviously that the one presented so far can be defined by using folding. Please
refer to the appendix A of this book for definition of folding.
partition(p, L) = f old(f (p), (, ), L)

(13.4)

Where function f compares the element to the pivot with predicate p (which
is passed to f as a parameter, so that f is in curried form, see appendix A for
detail. Alternatively, f can be a lexical closure which is in the scope of partition,
so that it can access the predicate in this scope.), and update the result pair
accordingly.

({x} A, B) : p(x)
f (p, x, (A, B)) =
(13.5)
(A, {x} B) : otherwise(p(x))
Note we actually use pattern-matching style definition. In environment without pattern-matching support, the pair (A, B) should be represented by a variable, for example P , and use access functions to extract its first and second
parts.
The example Haskell program needs to be modified accordingly.
sort [] = []
sort (x:xs) = sort small ++ [x] ++ sort big where
(small, big) = foldr f ([], []) xs
f a (as, bs) = if a x then (a:as, bs) else (as, a:bs)

Accumulated partition
The partition algorithm by using folding actually accumulates to the result pair
of lists (A, B). That if the element is not greater than the pivot, its accumulated

384CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
to A, otherwise to B. We can explicitly express it which save spaces and is
friendly for tail-recursive call optimization (refer to the appendix A of this book
for detail).

partition(p, L, A, B) =

(A, B) :
partition(p, L0 , {l1 } A, B) :

partition(p, L0 , A, {l1 } B) :

L=
p(l1 )
otherwise

(13.6)

Where l1 is the first element in L if L isnt empty, and L0 contains the

rest elements except for l1 , that L0 = {l2 , l3 , ...} for example. The quick sort
algorithm then uses this accumulated partition function by passing the x x
pivot as the partition predicate.

: L=
sort(L) =
(13.7)
sort(A) {l1 } sort(B) : otherwise
Where A, B are computed by the accumulated partition function defined
above.
(A, B) = partition(x x l1 , L0 , , )
Accumulated quick sort
Observe the recursive case in the last quick sort definition. the list concatenation
operations sort(A) {l1 } sort(B) actually are proportion to the length of the
list to be concatenated. Of course we can use some general solutions introduced
in the appendix A of this book to improve it. Another way is to change the sort
algorithm to accumulated manner. Something like below:

S : L=
sort0 (L, S) =
... : otherwise
Where S is the accumulator, and we call this version by passing empty list
as the accumulator to start sorting: sort(L) = sort0 (L, ). The key intuitive is
that after the partition finishes, the two sub lists need to be recursively sorted.
We can first recursively sort the list contains the elements which are greater
than the pivot, then link the pivot in front of it and use it as an accumulator
for next step sorting.
Based on this idea, the ... part in above definition can be realized as the
following.

S : L=
sort0 (L, S) =
sort(A, {l1 } sort(B, ?)) : otherwise
The problem is whats the accumulator when sorting B. There is an important invariant actually, that at every time, the accumulator S holds the elements
have been sorted so far. So that we should sort B by accumulating to S.
sort0 (L, S) =

S : L=
sort(A, {l1 } sort(B, S)) : otherwise

(13.8)

The following Haskell example program implements the accumulated quick

sort algorithm.

13.3. PERFORMANCE ANALYSIS FOR QUICK SORT

385

asort xs = asort xs []
asort
asort
(as,
part
part

[] acc = acc
(x:xs) acc = asort as (x:asort bs acc) where
bs) = part xs [] []
[] as bs = (as, bs)
(y:ys) as bs | y x = part ys (y:as) bs
| otherwise = part ys as (y:bs)

Exercise 13.1
Implement the recursive basic quick sort algorithm in your favorite imperative programming language.
Same as the imperative algorithm, one minor improvement is that besides
the empty case, we neednt sort the singleton list, implement this idea in
the functional algorithm as well.
The accumulated quick sort algorithm developed in this section uses intermediate variable A, B. They can be eliminated by defining the partition
function to mutually recursive call the sort function. Implement this idea
in your favorite functional programming language. Please dont refer to
the downloadable example program along with this book before you try
it.

13.3

Performance analysis for quick sort

Quick sort performs well in practice, however, its not easy to give theoretical
analysis. It needs the tool of probability to prove the average case performance.
Nevertheless, its intuitive to calculate the best case and worst case performance. Its obviously that the best case happens when every partition divides
the sequence into two slices with equal size. Thus it takes O(lg n) recursive calls
as shown in figure 13.3.
There are total O(lg n) levels of recursion. In the first level, it executes one
partition, which processes n elements; In the second level, it executes partition
two times, each processes n/2 elements, so the total time in the second level
bounds to 2O(n/2) = O(n) as well. In the third level, it executes partition four
times, each processes n/4 elements. The total time in the third level is also
bound to O(n); ... In the last level, there are n small slices each contains a
single element, the time is bound to O(n). Summing all the time in each level
gives the total performance of quick sort in best case as O(n lg n).
However, in the worst case, the partition process unluckily divides the sequence to two slices with unbalanced lengths in most time. That one slices with
length O(1), the other is O(n). Thus the recursive time degrades to O(n). If
we draw a similar figure, unlike in the best case, which forms a balanced binary
tree, the worst case degrades into a very unbalanced tree that every node has
only one child, while the other is empty. The binary tree turns to be a linked
list with O(n) length. And in every level, all the elements are processed, so the
total performance in worst case is O(n2 ), which is as same poor as insertion sort
and selection sort.

386CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT

n /4

n / 2

n /4

...n...

...lg(n)...

Figure 13.3: In the best case, quick sort divides the sequence into two slices
with same length.

Lets consider when the worst case will happen. One special case is that
all the elements (or most of the elements) are same. Nico Lomutos partition
method deals with such sequence poor. Well see how to solve this problem by
introducing other partition algorithm in the next section.
The other two obvious cases which lead to worst case happen when the
sequence has already in ascending or descending order. Partition the ascending
sequence makes an empty sub list before the pivot, while the list after the
pivot contains all the rest elements. Partition the descending sequence gives an
opponent result.
There are other cases which lead quick sort performs poor. There is no
completely satisfied solution which can avoid the worst case. Well see some
engineering practice in next section which can make it very seldom to meet the
worst case.

13.3.1

Average case analysis ?

In average case, quick sort performs well. There is a vivid example that even
the partition divides the list every time to two lists with length 1 to 9. The
performance is still bound to O(n lg n) as shown in [2].
This subsection need some mathematic background, reader can safely skip
to next part.
There are two methods to proof the average case performance, one uses
an important fact that the performance is proportion to the total comparing
operations during quick sort [2]. Different with the selections sort that every two
elements have been compared. Quick sort avoid many unnecessary comparisons.
For example suppose a partition operation on list {a1 , a2 , a3 , ..., an }. Select a1
as the pivot, the partition builds two sub lists A = {x1 , x2 , ..., xk } and B =
{y1 , y2 , ..., ynk1 }. In the rest time of quick sort, The element in A will never

13.3. PERFORMANCE ANALYSIS FOR QUICK SORT

387

be compared with any elements in B.

Denote the final sorted result as {a1 , a2 , ..., an }, this indicates that if element
ai < aj , they will not be compared any longer if and only if some element ak
where ai < ak < aj has ever been selected as pivot before ai or aj being selected
as the pivot.
That is to say, the only chance that ai and aj being compared is either ai
is chosen as pivot or aj is chosen as pivot before any other elements in ordered
range ai+1 < ai+2 < ... < aj1 are selected.
Let P (i, ) represent the probability that ai and aj being compared. We have:
P (i, j) =

2
ji+1

(13.9)

Since the total number of compare operation can be given as:

n1
X

C(n) =

n
X

P (i, j)

(13.10)

i=1 j=i+1

Note the fact that if we compared ai and aj , we wont compare aj and ai

again in the quick sort algorithm, and we never compare ai onto itself. Thats
why we set the upper bound of i to n 1; and lower bound of j to i + 1.
Substitute the probability, it yields:

C(n)

=
=

n1
X

n
X

2
j

i+1
i=1 j=i+1

n1
ni
XX
i=1 k=1

(13.11)

2
k+1

Using the harmonic series [18]

Hn = 1 +

1 1
+ + .... = ln n + + n
2 3

C(n) =

n1
X

O(lg n) = O(n lg n)

(13.12)

i=1

The other method to prove the average performance is to use the recursive
fact that when sorting list of length n, the partition splits the list into two sub
lists with length i and ni1. The partition process itself takes cn time because
it examine every element with the pivot. So we have the following equation.
T (n) = T (i) + T (n i 1) + cn

(13.13)

Where T (n) is the total time when perform quick sort on list of length n.
Since i is equally like to be any of 0, 1, ..., n 1, taking math expectation to the

388CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
equation gives:
T (n)

= E(T (i)) + E(T (n i 1)) + cn

n1
n1
1X
1X
=
T (i) +
T (n i 1) + cn
n i=0
n i=0
n1
n1
1X
1X
=
T (i) +
T (j) + cn
n i=0
n j=0

(13.14)

2X
T (i) + cn
n i=0
b1

Multiply by n to both sides, the equation changes to:

nT (n) = 2

n1
X

T (i) + cn2

(13.15)

i=0

Substitute n to n 1 gives another equation:

(n 1)T (n 1) = 2

n2
X

T (i) + c(n 1)2

(13.16)

i=0

Subtract equation (13.15) and (13.16) can eliminate all the T (i) for 0 i <
n 1.
nT (n) = (n + 1)T (n 1) + 2cn c

(13.17)

As we can drop the constant time c in computing performance. The equation

can be one more step changed like below.
T (n)
T (n 1)
2c
=
+
n+1
n
n+1

(13.18)

Next we assign n to n 1, n 2, ..., which gives us n 1 equations.

T (n 1)
T (n 2) 2c
=
+
n
n1
n
T (n 2)
T (n 3)
2c
=
+
n1
n2
n1
...
T (1) 2c
T (2)
=
+
3
2
3
Sum all them up, and eliminate the same components in both sides, we can
deduce to a function of n.
X1
T (n)
T (1)
=
+ 2c
n+1
2
k
n+1

k=3

(13.19)

13.4. ENGINEERING IMPROVEMENT

389

Using the harmonic series mentioned above, the final result is:
O(

T (n)
T (1)
) = O(
+ 2c ln n + + n ) = O(lg n)
n+1
2

(13.20)

Thus
O(T (n)) = O(n lg n)

(13.21)

Exercise 13.2
Why Lomutos methods performs poor when there are many duplicated
elements?

13.4

Engineering Improvement

Quick sort performs well in most cases as mentioned in previous section. However, there does exist the worst cases which downgrade the performance to
quadratic. If the data is randomly prepared, such case is rare, however, there
are some particular sequences which lead to the worst case and these kinds of
sequences are very common in practice.
In this section, some engineering practices are introduces which either help
to avoid poor performance in handling some special input data with improved
partition algorithm, or try to uniform the possibilities among cases.

13.4.1

Engineering solution to duplicated elements

As presented in the exercise of above section, N. Lomutos partition method isnt

good at handling sequence with many duplicated elements. Consider a sequence
with n equal elements like: {x, x, ..., x}. There are actually two methods to sort
it.
1. The normal basic quick sort: That we select an arbitrary element, which
is x as the pivot, partition it to two sub sequences, one is {x, x, ..., x},
which contains n 1 elements, the other is empty. then recursively sort
the first one; this is obviously quadratic O(n2 ) solution.
2. The other way is to only pick those elements strictly smaller than x, and
strictly greater than x. Such partition results two empty sub sequences,
and n elements equal to the pivot. Next we recursively sort the sub sequences contains the smaller and the bigger elements, since both of them
are empty, the recursive call returns immediately; The only thing left is
to concatenate the sort results in front of and after the list of elements
which are equal to the pivot.
The latter one performs in O(n) time if all elements are equal. This indicates
an important improvement for partition. That instead of binary partition (split
to two sub lists and a pivot), ternary partition (split to three sub lists) handles
duplicated elements better.

390CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
We can define the ternary quick sort as the following.

sort(L) =

: L=
sort(S) sort(E) sort(G) : otherwise

(13.22)

Where S, E, G are sub lists contains all elements which are less than, equal
to, and greater than the pivot respectively.
S = {x|x L, x < l1 }
E = {x|x L, x = l1 }
G = {x|x L, l1 < x}
The basic ternary quick sort can be implemented in Haskell as the following
example code.
sort [] = []
sort (x:xs) = sort [a | axs, a<x] ++
x:[b | bxs, b==x] ++ sort [c | cxs, c>x]

Note that the comparison between elements must support abstract lessthan and equal-to operations. The basic version of ternary sort takes linear
O(n) time to concatenate the three sub lists. It can be improved by using the
standard techniques of accumulator.
Suppose function sort0 (L, A) is the accumulated ternary quick sort definition, that L is the sequence to be sorted, and the accumulator A contains the
intermediate sorted result so far. We initialize the sorting with an empty accumulator: sort(L) = sort0 (L, ).
Its easy to give the trivial edge cases like below.

A : L=
0
sort (L, A) =
... : otherwise
For the recursive case, as the ternary partition splits to three sub lists S, E, G,
only S and G need recursive sort, E contains all elements equal to the pivot,
which is in correct order thus neednt to be sorted any more. The idea is to
sort G with accumulator A, then concatenate it behind E, then use this result
as the new accumulator, and start to sort S:
0

sort (L, A) =

A : L=
sort(S, E sort(G, A)) : otherwise

(13.23)

The partition can also be realized with accumulators. It is similar to what

has been developed for the basic version of quick sort. Note that we cant just
pass only one predication for pivot comparison. It actually needs two, one for
less-than, the other for equality testing. For the sake of brevity, we pass the
pivot element instead.

(S, E, G)
partition(p, L0 , {l1 } S, E, G)
partition(p, L, S, E, G) =
partition(p, L0 , S, {l1 } E, G)

partition(p, L0 , S, E, {l1 } G)

:
:
:
:

L=
l1 < p
l1 = p
p < l1
(13.24)

13.4. ENGINEERING IMPROVEMENT

391

Where l1 is the first element in L if L isnt empty, and L0 contains all rest
elements except for l1 . Below Haskell program implements this algorithm. It
starts the recursive sorting immediately in the edge case of parition.
sort xs = sort xs []
sort []
r=r
sort (x:xs) r = part xs [] [x] [] r where
part [] as bs cs r = sort as (bs ++ sort cs
part (x:xs) as bs cs r | x < x = part xs
| x == x = part xs
| x > x = part xs

r)
(x:as) bs cs r
as (x:bs) cs r
as bs (x:cs) r

Richard Bird developed another version in [1], that instead of concatenating

the recursively sorted results, it uses a list of sorted sub lists, and performs
concatenation finally.
sort xs = concat $ pass xs []
pass [] xss = xss
pass (x:xs) xss = step xs [] [x] [] xss where
step [] as bs cs xss = pass as (bs:pass cs xss)
step (x:xs) as bs cs xss | x < x = step xs (x:as) bs cs xss
| x == x = step xs as (x:bs) cs xss
| x > x = step xs as bs (x:cs) xss

2-way partition
The cases with many duplicated elements can also be handled imperatively.
Robert Sedgewick presented a partition method [3], [4] which holds two pointers.
One moves from left to right, the other moves from right to left. The two pointers
are initialized as the left and right boundaries of the array.
When start partition, the left most element is selected as the pivot. Then
the left pointer i keeps advancing to right until it meets any element which is
not less than the pivot; On the other hand2 , The right pointer j repeatedly
scans to left until it meets any element which is not greater than the pivot.
At this time, all elements before the left pointer i are strictly less than the
pivot, while all elements after the right pointer j are greater than the pivot. i
points to an element which is either greater than or equal to the pivot; while j
points to an element which is either less than or equal to the pivot, the situation
at this stage is illustrated in figure 13.4 (a).
In order to partition all elements less than or equal to the pivot to the left,
and the others to the right, we can exchange the two elements pointed by i, and
j. After that the scan can be resumed until either i meets j, or they overlap.
At any time point during partition. There is invariant that all elements
before i (including the one pointed by i) are not greater than the pivot; while
all elements after j (including the one pointed by j) are not less than the pivot.
The elements between i and j havent been examined yet. This invariant is
shown in figure 13.4 (b).
After the left pointer i meets the right pointer j, or they overlap each other,
we need one extra exchanging to move the pivot located at the first position to
2 We

dont use then because its quite OK to perform the two scans in parallel.

392CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
pivot

x[l]

>=pivot

... less than ...

x[i]

<=pivot

...?...

x[j]

... greater than ...

(a) When pointer i, and j stop

pivot

x[l]

... not greater than ...

...?...

... not less than ...

(b) Partition invariant

Figure 13.4: Partition a range of array by using the left most element as the
pivot.

the correct place which is pointed by j. Next, the elements between the lower
bound and j as well as the sub slice between i and the upper bound of the array
are recursively sorted.
This algorithm can be described as the following.
1: procedure Sort(A, l, u)
. sort range [l, u)
2:
if u l > 1 then
. More than 1 element for non-trivial case
3:
i l, j u
4:
pivot A[l]
5:
loop
6:
repeat
7:
ii+1
8:
until A[i] pivot . Need handle error case that i u in fact.
9:
repeat
10:
j j1
11:
until A[j] pivot . Need handle error case that j < l in fact.
12:
if j < i then
13:
break
14:
Exchange A[i] A[j]
15:
Exchange A[l] A[j]
. Move the pivot
16:
Sort(A, l, j)
17:
Sort(A, i, u)
Consider the extreme case that all elements are equal, this in-place quick sort
will partition the list to two equal length sub lists although it takes n2 unnecessary swaps. As the partition is balanced, the overall performance is O(n lg n),
which avoid downgrading to quadratic. The following ANSI C example program
implements this algorithm.
void qsort(Key xs, int l, int u) {
int i, j, pivot;
if (l < u - 1) {
pivot = i = l; j = u;
while (1) {

13.4. ENGINEERING IMPROVEMENT

393

while (i < u && xs[++i] < xs[pivot]);

while (j l && xs[pivot] < xs[--j]);
if (j < i) break;
swap(xs[i], xs[j]);
}
swap(xs[pivot], xs[j]);
qsort(xs, l, j);
qsort(xs, i, u);
}
}

Comparing this algorithm with the basic version based on N. Lumotos partition method, we can find that it swaps fewer elements, because it skips those
have already in proper sides of the pivot.
3-way partition
Its obviously that, we should avoid those unnecessary swapping for the duplicated elements. Whats more, the algorithm can be developed with the idea
of ternary sort (as known as 3-way partition in some materials), that all the
elements which are strictly less than the pivot are put to the left sub slice, while
those are greater than the pivot are put to the right. The middle part holds all
the elements which are equal to the pivot. With such ternary partition, we need
only recursively sort the ones which differ from the pivot. Thus in the above
extreme case, there arent any elements need further sorting. So the overall
performance is linear O(n).
The difficulty is how to do the 3-way partition. Jon Bentley and Douglas
McIlroy developed a solution which keeps those elements equal to the pivot at
the left most and right most sides as shown in figure 13.5 (a) [5] [6].
pivot

x[l]

... equal ...

... less than...

...?...

... greater than ...

... equal ...

(a) Invariant of 3-way partition

... less than...

... equal ...

pivot

... greater than ...

(b) Swapping the equal parts to the middle

Figure 13.5: 3-way partition.

The majority part of scan process is as same as the one developed by Robert
Sedgewick, that i and j keep advancing toward each other until they meet any
element which is greater then or equal to the pivot for i, or less than or equal
to the pivot for j respectively. At this time, if i and j dont meet each other or
overlap, they are not only exchanged, but also examined if the elements pointed

394CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
by them are identical to the pivot. Then necessary exchanging happens between
i and p, as well as j and q.
By the end of the partition process, the elements equal to the pivot need
to be swapped to the middle part from the left and right ends. The number of
such extra exchanging operations are proportion to the number of duplicated
elements. Its zero operation if elements are unique which there is no overhead
in the case. The final partition result is shown in figure 13.5 (b). After that we
only need recursively sort the less-than and greater-than sub slices.
This algorithm can be given by modifying the 2-way partition as below.
1: procedure Sort(A, l, u)
2:
if u l > 1 then
3:
i l, j u
4:
p l, q u
. points to the boundaries for equal elements
5:
pivot A[l]
6:
loop
7:
repeat
8:
ii+1
9:
until A[i] pivot
. Skip the error handling for i u
10:
repeat
11:
j j1
12:
until A[j] pivot
. Skip the error handling for j < l
13:
if j i then
14:
break
. Note the difference form the above algorithm
15:
Exchange A[i] A[j]
16:
if A[i] = pivot then
. Handle the equal elements
17:
pp+1
18:
Exchange A[p] A[i]
19:
if A[j] = pivot then
20:
q q1
21:
Exchange A[q] A[j]
22:
if i = j A[i] = pivot then
. A special case
23:
j j 1, i i + 1
24:
for k from l to p do . Swap the equal elements to the middle part
25:
Exchange A[k] A[j]
26:
j j1
27:
for k from u 1 down-to q do
28:
Exchange A[k] A[i]
29:
ii+1
30:
Sort(A, l, j + 1)
31:
Sort(A, i, u)
This algorithm can be translated to the following ANSI C example program.
void qsort2(Key xs, int l, int u) {
int i, j, k, p, q, pivot;
if (l < u - 1) {
i = p = l; j = q = u; pivot = xs[l];
while (1) {
while (i < u && xs[++i] < pivot);
while (j l && pivot < xs[--j]);

13.4. ENGINEERING IMPROVEMENT

395

if (j i) break;
swap(xs[i], xs[j]);
if (xs[i] == pivot) { ++p; swap(xs[p], xs[i]); }
if (xs[j] == pivot) { --q; swap(xs[q], xs[j]); }
}
if (i == j &&
for (k = l; k
for (k = u-1;
qsort2(xs, l,
qsort2(xs, i,

xs[i] == pivot) { --j, ++i; }

p; ++k, --j) swap(xs[k], xs[j]);
k q; --k, ++i) swap(xs[k], xs[i]);
j + 1);
u);

}
}

It can be seen that the the algorithm turns to be a bit complex when it
evolves to 3-way partition. There are some tricky edge cases should be handled
with caution. Actually, we just need a ternary partition algorithm. This remind
us the N. Lumotos method, which is straightforward enough to be a start point.
The idea is to change the invariant a bit. We still select the first element as
the pivot, as shown in figure 13.6, at any time, the left most section contains
elements which are strictly less than the pivot; the next section contains the
elements equal to the pivot; the right most section holds all the elements which
are strictly greater than the pivot. The boundaries of three sections are marked
as i, k, and j respectively. The rest part, which is between k and j are elements
havent been scanned yet.
At the beginning of this algorithm, the less-than section is empty; the
equal-to section contains only one element, which is the pivot; so that i is
initialized to the lower bound of the array, and k points to the element next
to i. The greater-than section is also initialized as empty, thus j is set to the
upper bound.
i

... less than...

... equal ...

...?...

... greater than ...

Figure 13.6: 3-way partition based on N. Lumotos method.

When the partition process starts, the elements pointed by k is examined.
If its equal to the pivot, k just advances to the next one; If its greater than
the pivot, we swap it with the last element in the unknown area, so that the
length of greater-than section increases by one. Its boundary j moves to the
left. Since we dont know if the elements swapped to k is still greater than the
pivot, it should be examined again repeatedly. Otherwise, if the element is less
than the pivot, we can exchange it with the first one in the equal-to section to
resume the invariant. The partition algorithm stops when k meets j.
1: procedure Sort(A, l, u)
2:
if u l > 1 then
3:
i l, j u, k l + 1
4:
pivot A[i]
5:
while k < j do

396CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
while pivot < A[k] do
j j1
Exchange A[k] A[j]
if A[k] < pivot then
Exchange A[k] A[i]
ii+1
k k+1
Sort(A, l, i)
Sort(A, j, u)

6:
7:
8:
9:
10:
11:
12:
13:
14:

Compare this one with the previous 3-way partition quick sort algorithm, its
more simple at the cost of more swapping operations. Below ANSI C program
implements this algorithm.
void qsort(Key xs, int l, int u) {
int i, j, k; Key pivot;
if (l < u - 1) {
i = l; j = u; pivot = xs[l];
for (k = l + 1; k < j; ++k) {
while (pivot < xs[k]) { --j; swap(xs[j], xs[k]); }
if (xs[k] < pivot) { swap(xs[i], xs[k]); ++i; }
}
qsort(xs, l, i);
qsort(xs, j, u);
}
}

Exercise 13.3
All the quick sort imperative algorithms use the first element as the
pivot, another method is to choose the last one as the pivot. Realize
the quick sort algorithms, including the basic version, Sedgewick version,
and ternary (3-way partition) version by using this approach.

13.5

Engineering solution to the worst case

Although the ternary quick sort (3-way partition) solves the issue for duplicated
elements, it cant handle some typical worst cases. For example if many of the
elements in the sequence are ordered, no matter its in ascending or descending
order, the partition result will be two unbalanced sub sequences, one with few
elements, the other contains all the rest.
Consider the two extreme cases, {x1 < x2 < ... < xn } and {y1 > y2 > ... >
yn }. The partition results are shown in figure 13.7.
Its easy to give some more worst cases, for example, {xm , xm1 , ..., x2 , x1 , xm+1 , xm+2 , ...xn }
where {x1 < x2 < ... < xn }; Another one is {xn , x1 , xn1 , x2 , ...}. Their partition result trees are shown in figure 13.8.
Observing that the bad partition happens easily when blindly choose the
first element as the pivot, there is a popular work around suggested by Robert
Sedgwick in [3]. Instead of selecting the fixed position in the sequence, a small
sampling helps to find a pivot which has lower possibility to cause a bad partition. One option is to examine the first element, the middle, and the last one,

13.5. ENGINEERING SOLUTION TO THE WORST CASE

...

(a) The partition tree for {x1 < x2 < ... < xn }, There arent any elements less
than or equal to the pivot (the first element) in every partition.

n-1

n-2

...

(b) The partition tree for {y1 > y2 > ... > yn }, There
arent any elements greater than or equal to the pivot
(the first element) in every partition.

Figure 13.7: The two worst cases.

397

398CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT

m-1

m+1

m-2

m+2

...

(a) Except for the first partition, all the others are unbalanced.

n-1

...

(b) A zig-zag partition tree.

Figure 13.8: Another two worst cases.

13.5. ENGINEERING SOLUTION TO THE WORST CASE

399

then choose the median of these three element. In the worst case, it can ensure
that there is at least one element in the shorter partitioned sub list.
Note that there is one tricky in real-world implementation. Since the index
is typically represented in limited length words. It may cause overflow when
calculating the middle index by the naive expression (l + u) / 2. In order to
avoid this issue, it can be accessed as l + (u - l)/2. There are two methods
to find the median, one needs at most three comparisons [5]; the other is to
move the minimum value to the first location, the maximum value to the last
location, and the median value to the meddle location by swapping. After that
we can select the middle as the pivot. Below algorithm illustrated the second
idea before calling the partition procedure.
1: procedure Sort(A, l, u)
2:
if u l > 1 then
3:
m b l+u
. Need handle overflow error in practice
2 c
4:
if A[m] < A[l] then
. Ensure A[l] A[m]
5:
Exchange A[l] A[r]
6:
if A[u 1] < A[l] then
. Ensure A[l] A[u 1]
7:
Exchange A[l] A[u 1]
8:
if A[u 1] < A[m] then
. Ensure A[m] A[u 1]
9:
Exchange A[m] A[u]
10:
Exchange A[l] A[m]
11:
(i, j) Partition(A, l, u)
12:
Sort(A, l, i)
13:
Sort(A, j, u)
Its obviously that this algorithm performs well in the 4 special worst cases
given above. The imperative implementation of median-of-three is left as exercise to the reader.
However, in purely functional settings, its expensive to randomly access the
middle and the last element. We cant directly translate the imperative median
selection algorithm. The idea of taking a small sampling and then finding the
median element as pivot can be realized alternatively by taking the first 3. For
example, in the following Haskell program.
qsort [] = []
qsort [x] = [x]
qsort [x, y] = [min x y, max x y]
qsort (x:y:z:rest) = qsort (filter (< m) (s:rest)) ++ [m] ++ qsort (filter ( m) (l:rest)) where
xs = [x, y, z]
[s, m, l] = [minimum xs, median xs, maximum xs]

Unfortunately, none of the above 4 worst cases can be well handled by this
program, this is because the sampling is not good. We need telescope, but not
microscope to profile the whole list to be partitioned. Well see the functional
way to solve the partition problem later.
Except for the median-of-three, there is another popular engineering practice
to get good partition result. instead of always taking the first element or the
last one as the pivot. One alternative is to randomly select one. For example
as the following modification.
1: procedure Sort(A, l, u)
2:
if u l > 1 then

400CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
Exchange A[l] A[ Random(l, u) ]
(i, j) Partition(A, l, u)
Sort(A, l, i)
Sort(A, j, u)

3:
4:
5:
6:

The function Random(l, u) returns a random integer i between l and u,

that l i < u. The element at this position is exchanged with the first one, so
that it is selected as the pivot for the further partition. This algorithm is called
random quick sort [2].
Theoretically, neither median-of-three nor random quick sort can avoid the
worst case completely. If the sequence to be sorted is randomly distributed, no
matter choosing the first one as the pivot, or the any other arbitrary one are
equally in effect. Considering the underlying data structure of the sequence is
singly linked-list in functional setting, its expensive to strictly apply the idea
of random quick sort in purely functional approach.
Even with this bad news, the engineering improvement still makes sense in
real world programming.

13.6

Other engineering practice

There is some other engineering practice which doesnt focus on solving the bad
partition issue. Robert Sedgewick observed that when the list to be sorted is
short, the overhead introduced by quick sort is relative expense, on the other
hand, the insertion sort performs better in such case [4], [5]. Sedgewick, Bentley
and McIlroy tried different threshold, as known as Cut-Off, that when there
are lesson than Cut-Off elements, the sort algorithm falls back to insertion
sort.
1: procedure Sort(A, l, u)
2:
if u l > Cut-Off then
3:
Quick-Sort(A, l, u)
4:
else
5:
Insertion-Sort(A, l, u)
The implementation of this improvement is left as exercise to the reader.

Exercise 13.4
Can you figure out more quick sort worst cases besides the four given in
this section?
Implement median-of-three method in your favorite imperative programming language.
Implement random quick sort in your favorite imperative programming
language.
Implement the algorithm which falls back to insertion sort when the length
of list is small in both imperative and functional approach.

13.7. SIDE WORDS

13.7

401

Side words

Its sometimes called true quick sort if the implementation equipped with most
of the engineering practice we introduced, including insertion sort fall-back with
cut-off, in-place exchanging, choose the pivot by median-of-three method, 3-waypartition.
The purely functional one, which express the idea of quick sort perfect cant
take all of them. Thus someone think the functional quick sort is essentially
tree sort.
Actually, quick sort does have close relationship with tree sort. Richard Bird
shows how to derive quick sort from binary tree sort by deforestation [7].
Consider a binary search tree creation algorithm called unf old. Which turns
a list of elements into a binary search tree.

: L=
unf old(L) =
(13.25)
tree(Tl , l1 , Tr ) : otherwise
Where
Tl = unf old({a|a L0 , a l1 })
Tr = unf old({a|a L0 , l1 < a})

(13.26)

The interesting point is that, this algorithm creates tree in a different way
as we introduced in the chapter of binary search tree. If the list to be unfold
is empty, the result is obviously an empty tree. This is the trivial edge case;
Otherwise, the algorithm set the first element l1 in the list as the key of the
node, and recursively creates its left and right children. Where the elements
used to form the left child is those which are less than or equal to the key in
L0 , while the rest elements which are greater than the key are used to form the
right child.
Remind the algorithm which turns a binary search tree to a list by in-order
traversing:

T =
otherwise
(13.27)
We can define quick sort algorithm by composing these two functions.

toList(T ) =

:
toList(lef t(T )) {key(T )} toList(right(T )) :

quickSort = toList unf old

(13.28)

The binary search tree built in the first step of applying unf old is the intermediate result. This result is consumed by toList and dropped after the second
step. Its quite possible to eliminate this intermediate result, which leads to the
basic version of quick sort.
The elimination of the intermediate binary search tree is called deforestation.
This concept is based on Burstle-Darlingtons work [9].

13.8

Merge sort

Although quick sort performs perfectly in average cases, it cant avoid the worst
case no matter what engineering practice is applied. Merge sort, on the other

402CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
kind, ensure the performance is bound to O(n lg n) in all the cases. Its particularly useful in theoretical algorithm design and analysis. Another feature is
that merge sort is friendly for linked-space settings, which is suitable for sorting
nonconsecutive stored sequences. Some functional programming and dynamic
programming environments adopt merge sort as the standard library sorting
solution, such as Haskel, Python and Java (later than Java 7).
In this section, well first brief the intuitive idea of merge sort, provide a
basic version. After that, some variants of merge sort will be given including
nature merge sort, and bottom-up merge sort.

13.8.1

Basic version

Same as quick sort, the essential idea behind merge sort is also divide and conquer. Different from quick sort, merge sort enforces the divide to be strictly
balanced, that it always splits the sequence to be sorted at the middle point.
After that, it recursively sort the sub sequences and merge the sorted two sequences to the final result. The algorithm can be described as the following.
In order to sort a sequence L,
Trivial edge case: If the sequence to be sorted is empty, the result is
obvious empty;
Otherwise, split the sequence at the middle position, recursively sort the
two sub sequences and merge the result.
The basic merge sort algorithm can be formalized with the following equation.

sort(L) =

:
merge(sort(L1 ), sort(L2 )) :

L=
otherwise, (L1 , L2 ) = splitAt(b |L|
2 c, L)
(13.29)

Merge
There are two black-boxes in the above merge sort definition, one is the splitAt
function, which splits a list at a given position; the other is the merge function,
which can merge two sorted lists into one.
As presented in the appendix of this book, its trivial to realize splitAt in
imperative settings by using random access. However, in functional settings,
its typically realized as a linear algorithm:

(, L) : n = 0
({l1 } A, B) : otherwise, (A, B) = splitAt(n 1, L0 )
(13.30)
Where l1 is the first element of L, and L0 represents the rest elements except
of l1 if L isnt empty.
The idea of merge can be illustrated as in figure 13.9. Consider two lines of
kids. The kids have already stood in order of their heights. that the shortest
one stands at the first, then a taller one, the tallest one stands at the end of the
line.
splitAt(n, L) =

13.8. MERGE SORT

403

Figure 13.9: Two lines of kids pass a door.

Now lets ask the kids to pass a door one by one, every time there can be at
most one kid pass the door. The kids must pass this door in the order of their
height. The one cant pass the door before all the kids who are shorter than
him/her.
Since the two lines of kids have already been sorted, the solution is to ask
the first two kids, one from each line, compare their height, and let the shorter
kid pass the door; Then they repeat this step until one line is empty, after that,
all the rest kids can pass the door one by one.
This idea can be formalized in the following equation.

A
B
merge(A, B) =
{a1 } merge(A0 , B)

{b1 } merge(A, B 0 )

:
:
:
:

B=
A=
a1 b1
otherwise

(13.31)

Where a1 and b1 are the first elements in list A and B; A0 and B 0 are the
rest elements except for the first ones respectively. The first two cases are trivial
edge cases. That merge one sorted list with an empty list results the same sorted
list; Otherwise, if both lists are non-empty, we take the first elements from the
two lists, compare them, and use the minimum as the first one of the result,
then recursively merge the rest.
With merge defined, the basic version of merge sort can be implemented
like the following Haskell example code.
msort [] = []
msort [x] = [x]
msort xs = merge (msort as) (msort bs) where
(as, bs) = splitAt (length xs div 2) xs
merge xs [] = xs

404CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
merge [] ys = ys
merge (x:xs) (y:ys) | x y = x : merge xs (y:ys)
| x > y = y : merge (x:xs) ys

Note that, the implementation differs from the algorithm definition that it
treats the singleton list as trivial edge case as well.
Merge sort can also be realized imperatively. The basic version can be developed as the below algorithm.
1: procedure Sort(A)
2:
if |A| > 1 then
3:
m b |A|
2 c
4:
X Copy-Array(A[1...m])
5:
Y Copy-Array(A[m + 1...|A|])
6:
Sort(X)
7:
Sort(Y )
8:
Merge(A, X, Y )
When the array to be sorted contains at least two elements, the non-trivial
sorting process starts. It first copy the first half to a new created array A, and
the second half to a second new array B. Recursively sort them; and finally
merge the sorted result back to A.
This version uses the same amount of extra spaces of A. This is because the
Merge algorithm isnt in-place at the moment. Well introduce the imperative
in-place merge sort in later section.
The merge process almost does the same thing as the functional definition.
There is a verbose version and a simplified version by using sentinel.
The verbose merge algorithm continuously checks the element from the two
input arrays, picks the smaller one and puts it back to the result array A, it then
advances along the arrays respectively until either one input array is exhausted.
After that, the algorithm appends the rest of the elements in the other input
array to A.
1: procedure Merge(A, X, Y )
2:
i 1, j 1, k 1
3:
m |X|, n |Y |
4:
while i m j n do
5:
if X[i] < Y [j] then
6:
A[k] X[i]
7:
ii+1
8:
else
9:
A[k] Y [j]
10:
j j+1
11:
k k+1
12:
while i m do
13:
A[k] X[i]
14:
k k+1
15:
ii+1
16:
while j n do
17:
A[k] Y [j]
18:
k k+1
19:
j j+1

13.8. MERGE SORT

405

Although this algorithm is a bit verbose, it can be short in some programming environment with enough tools to manipulate array. The following Python
program is an example.
def msort(xs):
n = len(xs)
if n > 1:
ys = [x for x in xs[:n/2]]
zs = [x for x in xs[n/2:]]
ys = msort(ys)
zs = msort(zs)
xs = merge(xs, ys, zs)
return xs
def merge(xs, ys, zs):
i=0
while ys != [] and zs != []:
xs[i] = ys.pop(0) if ys[0] < zs[0] else zs.pop(0)
i=i+1
xs[i:] = ys if ys !=[] else zs
return xs

Performance
Before dive into the improvement of this basic version, lets analyze the performance of merge sort. The algorithm contains two steps, divide step, and merge
step. In divide step, the sequence to be sorted is always divided into two sub
sequences with the same length. If we draw a similar partition tree as what
we did for quick sort, it can be found this tree is a perfectly balanced binary
tree as shown in figure 13.3. Thus the height of this tree is O(lg n). It means
the recursion depth of merge sort is bound to O(lg n). Merge happens in every
level. Its intuitive to analyze the merge algorithm, that it compare elements
from two input sequences in pairs, after one sequence is fully examined the rest
one is copied one by one to the result, thus its a linear algorithm proportion to
the length of the sequence. Based on this facts, denote T (n) the time for sorting
the sequence with length n, we can write the recursive time cost as below.
n
n
T (n) = T ( ) + T ( ) + cn
2
2
n
= 2T ( ) + cn
2

(13.32)

It states that the cost consists of three parts: merge sort the first half takes
T ( n2 ), merge sort the second half takes also T ( n2 ), merge the two results takes
cn, where c is some constant. Solve this equation gives the result as O(n lg n).
Note that, this performance doesnt vary in all cases, as merge sort always
uniformly divides the input.
Another significant performance indicator is space occupation. However, it
varies a lot in different merge sort implementation. The detail space bounds
analysis will be explained in every detailed variants later.
For the basic imperative merge sort, observe that it demands same amount
of spaces as the input array in every recursion, copies the original elements

406CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
to them for recursive sort, and these spaces can be released after this level of
recursion. So the peak space requirement happens when the recursion enters to
the deepest level, which is O(n lg n).
The functional merge sort consume much less than this amount, because the
underlying data structure of the sequence is linked-list. Thus it neednt extra
spaces for merge3 . The only spaces requirement is for book-keeping the stack
for recursive calls. This can be seen in the later explanation of even-odd split
algorithm.
Minor improvement
Well next improve the basic merge sort bit by bit for both the functional and
imperative realizations. The first observation is that the imperative merge algorithm is a bit verbose. [2] presents an elegant simplification by using positive
as the sentinel. That we append as the last element to the both ordered
arrays for merging4 . Thus we neednt test which array is not exhausted. Figure
13.10 illustrates this idea.

x[1]

x[2]

...

a[i]

...

a[n]

INF

b[j]

...

b[m]

INF

x[k]

Figure 13.10: Merge with as sentinels.

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:

procedure Merge(A, X, Y )
Append(X, )
Append(Y, )
i 1, j 1
for k from 1 to |A| do
if X[i] < Y [j] then
A[k] X[i]
ii+1
else
A[k] Y [j]
j j+1

The following ANSI C program imlements this idea. It embeds the merge inside. INF is defined as a big constant number with the same type of Key. Where
3 The
4 For

complex effects caused by lazy evaluation is ignored here, please refer to [7] for detail
sorting in monotonic non-increasing order, can be used instead

13.8. MERGE SORT

407

the type can either be defined elsewhere or we can abstract the type information by passing the comparator as parameter. We skip these implementation
and language details here.
void msort(Key xs, int l, int u) {
int i, j, m;
Key as, bs;
if (u - l > 1) {
m = l + (u - l) / 2; / avoid int overflow /
msort(xs, l, m);
msort(xs, m, u);
as = (Key) malloc(sizeof(Key) (m - l + 1));
bs = (Key) malloc(sizeof(Key) (u - m + 1));
memcpy((void)as, (void)(xs + l), sizeof(Key) (m - l));
memcpy((void)bs, (void)(xs + m), sizeof(Key) (u - m));
as[m - l] = bs[u - m] = INF;
for (i = j = 0; l < u; ++l)
xs[l] = as[i] < bs[j] ? as[i++] : bs[j++];
free(as);
free(bs);
}
}

Running this program takes much more time than the quick sort. Besides
the major reason well explain later, one problem is that this version frequently
allocates and releases memories for merging. While memory allocation is one of
the well known bottle-neck in real world as mentioned by Bentley in [4]. One
solution to address this issue is to allocate another array with the same size to
the original one as the working area. The recursive sort for the first and second
halves neednt allocate any more extra spaces, but use the working area when
merging. Finally, the algorithm copies the merged result back.
This idea can be expressed as the following modified algorithm.
1: procedure Sort(A)
2:
B Create-Array(|A|)
3:
Sort(A, B, 1, |A|)
4:
5:
6:
7:
8:
9:

procedure Sort(A, B, l, u)
if u l > 0 then
m b l+u
2 c
Sort(A, B, l, m)
Sort(A, B, m + 1, u)
Merge(A, B, l, m, u)

This algorithm duplicates another array, and pass it along with the original
array to be sorted to Sort algorithm. In real implementation, this working
area should be released either manually, or by some automatic tool such as GC
(Garbage collection). The modified algorithm Merge also accepts a working
area as parameter.
1: procedure Merge(A, B, l, m, u)
2:
i l, j m + 1, k l
3:
while i m j u do
4:
if A[i] < A[j] then

408CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:

B[k] A[i]
ii+1
else
B[k] A[j]
j j+1
k k+1
while i m do
B[k] A[i]
k k+1
ii+1
while j u do
B[k] A[j]
k k+1
j j+1
for i from l to u do
A[i] B[i]

. Copy back

By using this minor improvement, the space requirement reduced to O(n)

from O(n lg n). The following ANSI C program implements this minor improvement. For illustration purpose, we manually copy the merged result back to the
original array in a loop. This can also be realized by using standard library
provided tool, such as memcpy.
void merge(Key xs, Key ys, int l, int m, int u) {
int i, j, k;
i = k = l; j = m;
while (i < m && j < u)
ys[k++] = xs[i] < xs[j] ? xs[i++] : xs[j++];
while (i < m)
ys[k++] = xs[i++];
while (j < u)
ys[k++] = xs[j++];
for(; l < u; ++l)
xs[l] = ys[l];
}
void msort(Key xs, Key ys, int l, int u) {
int m;
if (u - l > 1) {
m = l + (u - l) / 2;
msort(xs, ys, l, m);
msort(xs, ys, m, u);
merge(xs, ys, l, m, u);
}
}
void sort(Key xs, int l, int u) {
Key ys = (Key) malloc(sizeof(Key) (u - l));
kmsort(xs, ys, l, u);
free(ys);
}

13.9. IN-PLACE MERGE SORT

409

This new version runs faster than the previous one. In my test machine, it
speeds up about 20% to 25% when sorting 100,000 randomly generated numbers.
The basic functional merge sort can also be fine tuned. Observe that, it
splits the list at the middle point. However, as the underlying data structure to
represent list is singly linked-list, random access at a given position is a linear
operation (refer to appendix A for detail). Alternatively, one can split the list
in an even-odd manner. That all the elements in even position are collected
in one sub list, while all the odd elements are collected in another. As for any
lists, there are either same amount of elements in even and odd positions, or
they differ by one. So this divide strategy always leads to well splitting, thus
the performance can be ensured to be O(n lg n) in all cases.
The even-odd splitting algorithm can be defined as below.

(, ) : L =
({l1 }, ) : |L| = 1

({l1 } A, {l2 } B) : otherwise, (A, B) = split(L00 )

(13.33)
When the list is empty, the split result are two empty lists; If there is only
one element in the list, we put this single element, which is at position 1, to the
odd sub list, the even sub list is empty; Otherwise, it means there are at least
two elements in the list, We pick the first one to the odd sub list, the second
one to the even sub list, and recursively split the rest elements.
All the other functions are kept same, the modified Haskell program is given
as the following.
split(L) =

split [] = ([], [])

split [x] = ([x], [])
split (x:y:xs) = (x:xs, y:ys) where (xs, ys) = split xs

13.9

In-place merge sort

One drawback for the imperative merge sort is that it requires extra spaces for
merging, the basic version without any optimization needs O(n lg n) in peak
time, and the one by allocating a working area needs O(n).
Its nature for people to seek the in-place version merge sort, which can
reuse the original array without allocating any extra spaces. In this section,
well introduce some solutions to realize imperative in-place merge sort.

13.9.1

Naive in-place merge

The first idea is straightforward. As illustrated in figure 13.11, sub list A, and B
are sorted, when performs in-place merge, the variant ensures that all elements
before i are merged, so that they are in non-decreasing order; every time we
compare the i-th and the j-th elements. If the i-th is less than the j-th, the
marker i just advances one step to the next. This is the easy case. Otherwise,
it means that the j-th element is the next merge result, which should be put in
front of i. In order to achieve this, all elements between i and j, including the
i-th should be shift to the end by one cell. We repeat this process till all the
elements in A and B are put to the correct positions.

410CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT

shift if not xs[i] < xs[j]

merged

xs[i]

...sorted sub list A...

xs[j]

...sorted sub list B...

Figure 13.11: Naive in-place merge

1:
2:
3:
4:
5:
6:
7:
8:
9:

procedure Merge(A, l, m, u)
while l m m u do
if A[l] < A[m] then
l l+1
else
x A[m]
for i m down-to l + 1 do
A[i] A[i 1]
A[l] x

. Shift

However, this naive solution downgrades merge sort overall performance to

quadratic O(n2 )! This is because that array shifting is a linear operation. It is
proportion to the length of elements in the first sorted sub array which havent
been compared so far.
The following ANSI C program based on this algorithm runs very slow, that
it takes about 12 times slower than the previous version when sorting 10,000
random numbers.
void naive_merge(Key xs, int l, int m, int u) {
int i; Key y;
for(; l < m && m < u; ++l)
if (!(xs[l] < xs[m])) {
y = xs[m++];
for (i = m - 1; i > l; --i) / shift /
xs[i] = xs[i-1];
xs[l] = y;
}
}
void msort3(Key xs, int l, int u) {
int m;
if (u - l > 1) {
m = l + (u - l) / 2;
msort3(xs, l, m);
msort3(xs, m, u);
naive_merge(xs, l, m, u);
}
}

13.9.2

in-place working area

In order to implement the in-place merge sort in O(n lg n) time, when sorting a
sub array, the rest part of the array must be reused as working area for merging.

13.9. IN-PLACE MERGE SORT

411

As the elements stored in the working area, will be sorted later, they cant be
overwritten. We can modify the previous algorithm, which duplicates extra
spaces for merging, a bit to achieve this. The idea is that, every time when
we compare the first elements in the two sorted sub arrays, if we want to put
the less element to the target position in the working area, we in-turn exchange
what sored in the working area with this element. Thus after merging the two
sub arrays store what the working area previously contains. This idea can be
illustrated in figure 13.12.

compare
... reuse ...

A[i]

...

... reuse ...

B[j]

...

swap(A[i], C[k]) if A[i] < B[j]

... merged ...

C[k]

...

Figure 13.12: Merge without overwriting working area.

In our algorithm, both the two sorted sub arrays, and the working area for
merging are parts of the original array to be sorted. we need supply the following
arguments when merging: the start points and end points of the sorted sub
arrays, which can be represented as ranges; and the start point of the working
area. The following algorithm for example, uses [a, b) to indicate the range
include a, exclude b. It merges sorted range [i, m) and range [j, n) to the working
area starts from k.
1: procedure Merge(A, [i, m), [j, n), k)
2:
while i < m j < n do
3:
if A[i] < A[j] then
4:
Exchange A[k] A[i]
5:
ii+1
6:
else
7:
Exchange A[k] A[j]
8:
j j+1
9:
k k+1
10:
while i < m do
11:
Exchange A[k] A[i]
12:
ii+1
13:
k k+1
14:
while j < m do
15:
Exchange A[k] A[j]
16:
j j+1
17:
k k+1
Note that, the following two constraints must be satisfied when merging:
1. The working area should be within the bounds of the array. In other

412CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
words, it should be big enough to hold elements exchanged in without
causing any out-of-bound error;
2. The working area can be overlapped with either of the two sorted arrays,
however, it should be ensured that there are not any unmerged elements
being overwritten;
This algorithm can be implemented in ANSI C as the following example.
void wmerge(Key xs, int i, int m, int j, int n, int w) {
while (i < m && j < n)
swap(xs, w++, xs[i] < xs[j] ? i++ : j++);
while (i < m)
swap(xs, w++, i++);
while (j < n)
swap(xs, w++, j++);
}

With this merging algorithm defined, its easy to imagine a solution, which
can sort half of the array; The next question is, how to deal with the rest of the
unsorted part stored in the working area as shown in figure 13.13?

...unsorted...

... sorted ...

Figure 13.13: Half of the array is sorted.

One intuitive idea is to recursively sort another half of the working area, thus
there are only 14 elements havent been sorted yet. Which is shown in figure
13.14. The key point at this stage is that we must merge the sorted 14 elements
B with the sorted 12 elements A sooner or later.

unsorted 1/4

sorted B 1/4

... ... sorted A 1/2 ... ...

Figure 13.14: A and B must be merged at sometime.

Is the working area left, which only holds 14 elements, big enough for merging
A and B? Unfortunately, it isnt in the settings shown in figure 13.14.
However, the second constraint mentioned before gives us a hint, that we
can exploit it by arranging the working area to overlap with either sub array
if we can ensure the unmerged elements wont be overwritten under some well
designed merging schema.
Actually, instead of sorting the second half of the working area, we can sort
the first half, and put the working area between the two sorted arrays as shown
in figure 13.15 (a). This setup effects arranging the working area to overlap
with the sub array A. This idea is proposed in [10].
Lets consider two extreme cases:
1. All the elements in B are less than any element in A. In this case, the
merge algorithm finally moves the whole contents of B to the working

13.9. IN-PLACE MERGE SORT

sorted B 1/4

413

work area

... ... sorted A 1/2 ... ...

(a)

work area 1/4

... ... ... ... merged 3/4 ... ... ... ...
(b)

Figure 13.15: Merge A and B with the working area.

area; the cells of B holds what previously stored in the working area; As
the size of area is as same as B, its OK to exchange their contents;
2. All the elements in A are less than any element in B. In this case,
the merge algorithm continuously exchanges elements between A and the
working area. After all the previous 14 cells in the working area are filled
with elements from A, the algorithm starts to overwrite the first half of
A. Fortunately, the contents being overwritten are not those unmerged
elements. The working area is in effect advances toward the end of the
array, and finally moves to the right side; From this time point, the merge
algorithm starts exchanging contents in B with the working area. The
result is that the working area moves to the left most side which is shown
in figure 13.15 (b).
We can repeat this step, that always sort the second half of the unsorted
part, and exchange the sorted sub array to the first half as working area. Thus
we keep reducing the working area from 12 of the array, 14 of the array, 18 of
the array, ... The scale of the merge problem keeps reducing. When there is
only one element left in the working area, we neednt sort it any more since the
singleton array is sorted by nature. Merging a singleton array to the other is
equivalent to insert the element. In practice, the algorithm can finalize the last
few elements by switching to insertion sort.
The whole algorithm can be described as the following.
1: procedure Sort(A, l, u)
2:
if u l > 0 then
3:
m b l+u
2 c
4:
w l+um
5:
Sort(A, l, m, w)
. The second half contains sorted elements
6:
while w l > 1 do
7:
u0 w
0
. Ensure the working area is big enough
8:
w d l+u
2 e
9:
Sort(A, w, u0 , l)
. The first half holds the sorted elements
10:
Merge(A, [l, l + u0 w], [u0 , u], w)
11:
for i w down-to l do
. Switch to insertion sort
12:
ji
13:
while j u A[j] < A[j 1] do
14:
Exchange A[j] A[j 1]

414CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
15:

j j+1

Note that in order to satisfy the first constraint, we must ensure the working
area is big enough to hold all exchanged in elements, thats way we round it by
ceiling when sort the second half of the working area. Note that we actually
pass the ranges including the end points to the algorithm Merge.
Next, we develop a Sort algorithm, which mutually recursive call Sort and
exchange the result to the working area.
1: procedure Sort(A, l, u, w)
2:
if u l > 0 then
3:
m b l+u
2 c
4:
Sort(A, l, m)
5:
Sort(A, m + 1, u)
6:
Merge(A, [l, m], [m + 1, u], w)
7:
else
. Exchange all elements to the working area
8:
while l u do
9:
Exchange A[l] A[w]
10:
l l+1
11:
w w+1
Different from the naive in-place sort, this algorithm doesnt shift the array
during merging. The main algorithm reduces the unsorted part in sequence of
n n n
2 , 4 , 8 , ..., it takes O(lg n) steps to complete sorting. In every step, It recursively
sorts half of the rest elements, and performs linear time merging.
Denote the time cost of sorting n elements as T (n), we have the following
equation.
n
n
n
3n
n
7n
T (n) = T ( ) + c + T ( ) + c
+ T( ) + c
+ ...
2
2
4
4
8
8

(13.34)

Solving this equation by using telescope method, gets the result O(n lg n).
The detailed process is left as exercise to the reader.
The following ANSI C code completes the implementation by using the example wmerge program given above.
void imsort(Key xs, int l, int u);
void wsort(Key xs, int l, int u, int w) {
int m;
if (u - l > 1) {
m = l + (u - l) / 2;
imsort(xs, l, m);
imsort(xs, m, u);
wmerge(xs, l, m, m, u, w);
}
else
while (l < u)
swap(xs, l++, w++);
}
void imsort(Key xs, int l, int u) {
int m, n, w;
if (u - l > 1) {
m = l + (u - l) / 2;

13.9. IN-PLACE MERGE SORT

415

w = l + u - m;
wsort(xs, l, m, w); / the last half contains sorted elements /
while (w - l > 2) {
n = w;
w = l + (n - l + 1) / 2; / ceiling /
wsort(xs, w, n, l); / the first half contains sorted elements /
wmerge(xs, l, l + n - w, n, u, w);
}
for (n = w; n > l; --n) /switch to insertion sort/
for (m = n; m < u && xs[m] < xs[m-1]; ++m)
swap(xs, m, m - 1);
}
}

However, this program doesnt run faster than the version we developed in
previous section, which doubles the array in advance as working area. In my
machine, it is about 60% slower when sorting 100,000 random numbers due to
many swap operations.

13.9.3

In-place merge sort vs. linked-list merge sort

The in-place merge sort is still a live area for research. In order to save the
extra spaces for merging, some overhead has be introduced, which increases
the complexity of the merge sort algorithm. However, if the underlying data
structure isnt array, but linked-list, merge can be achieved without any extra
spaces as shown in the even-odd functional merge sort algorithm presented in
previous section.
In order to make it clearer, we can develop a purely imperative linked-list
merge sort solution. The linked-list can be defined as a record type as shown in
appendix A like below.
struct Node {
Key key;
struct Node next;
};

We can define an auxiliary function for node linking. Assume the list to be
linked isnt empty, it can be implemented as the following.
struct Node link(struct Node xs, struct Node ys) {
xsnext = ys;
return xs;
}

One method to realize the imperative even-odd splitting, is to initialize two

empty sub lists. Then iterate the list to be split. Every time, we link the
current node in front of the first sub list, then exchange the two sub lists. So
that, the second sub list will be linked at the next time iteration. This idea can
be illustrated as below.
1: function Split(L)
2:
(A, B) (, )
3:
while L 6= do
4:
pL
5:
L Next(L)

416CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
6:
7:
8:

A Link(p, A)
Exchange A B
return (A, B)

The following example ANSI C program implements this splitting algorithm

embedded.
struct Node msort(struct Node xs) {
struct Node p, as, bs;
if (!xs | | !xsnext) return xs;
as = bs = NULL;
while(xs) {
p = xs;
xs = xsnext;
as = link(p, as);
swap(as, bs);
}
as = msort(as);
bs = msort(bs);
return merge(as, bs);
}

The only thing left is to develop the imperative merging algorithm for linkedlist. The idea is quite similar to the array merging version. As long as neither of
the sub lists is exhausted, we pick the less one, and append it to the result list.
After that, it just need link the non-empty one to the tail the result, but not
a looping for copying. It needs some carefulness to initialize the result list, as
its head node is the less one among the two sub lists. One simple method is to
use a dummy sentinel head, and drop it before returning. This implementation
detail can be given as the following.
struct Node merge(struct Node as, struct Node bs) {
struct Node s, p;
p = &s;
while (as && bs) {
if (askey < bskey) {
link(p, as);
as = asnext;
}
else {
link(p, bs);
bs = bsnext;
}
p = pnext;
}
if (as)
link(p, as);
if (bs)
link(p, bs);
return s.next;
}

Exercise 13.5
Proof the performance of in-place merge sort is bound to O(n lg n).

13.10. NATURE MERGE SORT

13.10

417

Nature merge sort

Knuth gives another way to interpret the idea of divide and conquer merge sort.
It just likes burn a candle in both ends [1]. This leads to the nature merge sort
algorithm.

Figure 13.16: Burn a candle from both ends

For any given sequence, we can always find a non-decreasing sub sequence
starts at any position. One particular case is that we can find such a sub
sequence from the left-most position. The following table list some examples,
the non-decreasing sub sequences are in bold font.
15 , 0, 4, 3, 5, 2, 7, 1, 12, 14, 13, 8, 9, 6, 10, 11
8, 12, 14 , 0, 1, 4, 11, 2, 3, 5, 9, 13, 10, 6, 15, 7
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
The first row in the table illustrates the worst case, that the second element
is less than the first one, so the non-decreasing sub sequence is a singleton list,
which only contains the first element; The last row shows the best case, the the
sequence is ordered, and the non-decreasing list is the whole; The second row
shows the average case.
Symmetrically, we can always find a non-decreasing sub sequence from the
end of the sequence to the left. This indicates us that we can merge the two nondecreasing sub sequences, one from the beginning, the other form the ending
to a longer sorted sequence. The advantage of this idea is that, we utilize the
nature ordered sub sequences, so that we neednt recursive sorting at all.

8, 12, 14

0, 1, 4, 11

2, 3, 5

merge
7, 8, 12, 14, 15

13, 10, 6

15, 7

merge
... free cells ...

13, 11, 10, 6, 4, 1, 0

Figure 13.17: Nature merge sort

Figure 13.17 illustrates this idea. We starts the algorithm by scanning from
both ends, finding the longest non-decreasing sub sequences respectively. After

418CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
that, these two sub sequences are merged to the working area. The merged
result starts from beginning. Next we repeat this step, which goes on scanning
toward the center of the original sequence. This time we merge the two ordered
sub sequences to the right hand of the working area toward the left. Such setup
is easy for the next round of scanning. When all the elements in the original
sequence have been scanned and merged to the target, we switch to use the
elements stored in the working area for sorting, and use the previous sequence
as new working area. Such switching happens repeatedly in each round. Finally,
we copy all elements from the working area to the original array if necessary.
The only question left is when this algorithm stops. The answer is that when
we start a new round of scanning, and find that the longest non-decreasing sub
list spans to the end, which means the whole list is ordered, the sorting is done.
Because this kind of merge sort proceeds the target sequence in two ways,
and uses the nature ordering of sub sequences, its named nature two-way merge
sort. In order to realize it, some carefulness must be paid. Figure 13.18 shows
the invariant during the nature merge sort. At anytime, all elements before
marker a and after marker d have been already scanned and merged. We are
trying to span the non-decreasing sub sequence [a, b) as long as possible, at the
same time, we span the sub sequence from right to left to span [c, d) as long as
possible as well. The invariant for the working area is shown in the second row.
All elements before f and after r have already been sorted. (Note that they
may contain several ordered sub sequences), For the odd times (1, 3, 5, ...), we
merge [a, b) and [c, d) from f toword right; while for the even times (2, 4, 6, ...),
we merge the two sorted sub sequences after r toward left.

... scanned ...

... span [a, b) ...

... ? ...

... span [c, d) ...

... merged ...

... scanned ...

... unused free cells ...

... merged ...

Figure 13.18: Invariant during nature merge sort

For imperative realization, the sequence is represented by array. Before
sorting starts, we duplicate the array to create a working area. The pointers
a, b are initialized to point the left most position, while c, d point to the right
most position. Pointer f starts by pointing to the front of the working area,

13.10. NATURE MERGE SORT

419

and r points to the rear position.

1: function Sort(A)
2:
if |A| > 1 then
3:
n |A|
4:
B Create-Array(n)
. Create the working area
5:
loop
6:
[a, b) [1, 1)
7:
[c, d) [n + 1, n + 1)
8:
f 1, r n
. front and rear pointers to the working area
9:
t False
. merge to front or rear
10:
while b < c do
. There are still elements for scan
11:
repeat
. Span [a, b)
12:
bb+1
13:
until b c A[b] < A[b 1]
14:
repeat
. Span [c, d)
15:
cc1
16:
until c b A[c 1] < A[c]
17:
if c < b then
. Avoid overlap
18:
cb
19:
if b a n then . Done if [a, b) spans to the whole array
20:
return A
21:
if t then
. merge to front
22:
f Merge(A, [a, b), [c, d), B, f, 1)
23:
else
. merge to rear
24:
r Merge(A, [a, b), [c, d), B, r, 1)
25:
a b, d c
26:
t t
. Switch the merge direction
27:
Exchange A B
. Switch working area
28:
return A
The merge algorithm is almost as same as before except that we need pass
a parameter to indicate the direction for merging.
1: function Merge(A, [a, b), [c, d), B, w, )
2:
while a < b c < d do
3:
if A[a] < A[d 1] then
4:
B[w] A[a]
5:
aa+1
6:
else
7:
B[w] A[d 1]
8:
dd1
9:
w w+
10:
while a < b do
11:
B[w] A[a]
12:
aa+1
13:
w w+
14:
while c < d do
15:
B[w] A[d 1]
16:
dd1
17:
w w+

420CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
return w
The following ANSI C program implements this two-way nature merge sort
algorithm. Note that it doesnt release the allocated working area explictly.
18:

int merge(Key xs, int a, int b, int c, int d, Key ys, int k, int delta) {
for(; a < b && c < d; k += delta )
ys[k] = xs[a] < xs[d-1] ? xs[a++] : xs[--d];
for(; a < b; k += delta)
ys[k] = xs[a++];
for(; c < d; k += delta)
ys[k] = xs[--d];
return k;
}
Key sort(Key xs, Key ys, int n) {
int a, b, c, d, f, r, t;
if(n < 2)
return xs;
for(;;) {
a = b = 0;
c = d = n;
f = 0;
r = n-1;
t = 1;
while(b < c) {
do {
/ span [a, b) as much as possible /
++b;
} while( b < c && xs[b-1] xs[b] );
do{
/ span [c, d) as much as possible /
--c;
} while( b < c && xs[c] xs[c-1] );
if( c < b )
c = b;
/ eliminate overlap if any /
if( b - a n)
return xs;
/ sorted /
if( t )
f = merge(xs, a, b, c, d, ys, f, 1);
else
r = merge(xs, a, b, c, d, ys, r, -1);
a = b;
d = c;
t = !t;
}
swap(&xs, &ys);
}
return xs; /cant be here/
}

The performance of nature merge sort depends on the actual ordering of

the sub arrays. However, it in fact performs well even in the worst case. Suppose that we are unlucky when scanning the array, that the length of the nondecreasing sub arrays are always 1 during the first round scan. This leads to
the result working area with merged ordered sub arrays of length 2. Suppose
that we are unlucky again in the second round of scan, however, the previous

13.10. NATURE MERGE SORT

421

results ensure that the non-decreasing sub arrays in this round are no shorter
than 2, this time, the working area will be filled with merged ordered sub arrays
of length 4, ... Repeat this we get the length of the non-decreasing sub arrays
doubled in every round, so there are at most O(lg n) rounds, and in every round
we scanned all the elements. The overall performance for this worst case is
bound to O(n lg n). Well go back to this interesting phenomena in the next
section about bottom-up merge sort.
In purely functional settings however, its not sensible to scan list from both
ends since the underlying data structure is singly linked-list. The nature merge
sort can be realized in another approach.
Observe that the list to be sorted is consist of several non-decreasing sub
lists, that we can pick every two of such sub lists and merge them to a bigger
one. We repeatedly pick and merge, so that the number of the non-decreasing
sub lists halves continuously and finally there is only one such list, which is the
sorted result. This idea can be formalized in the following equation.
sort(L) = sort0 (group(L))

(13.35)

Where function group(L) groups the list into non-decreasing sub lists. This
function can be described like below, the first two are trivial edge cases.
If the list is empty, the result is a list contains an empty list;
If there is only one element in the list, the result is a list contains a
singleton list;
Otherwise, The first two elements are compared, if the first one is less
than or equal to the second, it is linked in front of the first sub list of the
recursive grouping result; or a singleton list contains the first element is
set as the first sub list before the recursive result.

{L} : |L| 1
{{l1 } L1 , L2 , ...} : l1 l2 , {L1 , L2 , ...} = group(L0 )

{{l1 }, L1 , L2 , ...} : otherwise

(13.36)
Its quite possible to abstract the grouping criteria as a parameter to develop
a generic grouping function, for instance, as the following Haskell code 5 .
group(L) =

groupBy
groupBy
groupBy
groupBy

:: (aaBool) [a] [[a]]

_ [] = [[]]
_ [x] = [[x]]
f (x:xs@(x:_)) | f x x = (x:ys):yss
| otherwise = [x]:r

where
r@(ys:yss) = groupBy f xs
5 There

is a groupBy function provided in the Haskell standard library Data.List. However, it doesnt fit here, because it accepts an equality testing function as parameter, which
must satisfy the properties of reflexive, transitive, and symmetric. but what we use here, the
less-than or equal to operation doesnt conform to transitive. Refer to appendix A of this
book for detail.

422CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
Different from the sort function, which sorts a list of elements, function sort0
accepts a list of sub lists which is the result of grouping.

: L=

L1 : L = {L1 }
(13.37)
sort0 (L) =

sort0 (mergeP airs(L)) : otherwise

The first two are the trivial edge cases. If the list to be sorted is empty, the
result is obviously empty; If it contains only one sub list, then we are done. We
need just extract this single sub list as result; For the recursive case, we call a
function mergeP airs to merge every two sub lists, then recursively call sort0 .
The next undefined function is mergeP airs, as the name indicates, it repeatedly merges pairs of non-decreasing sub lists into bigger ones.

L : |L| 1
{merge(L1 , L2 )} mergeP airs(L00 ) : otherwise
(13.38)
When there are less than two sub lists in the list, we are done; otherwise, we
merge the first two sub lists L1 and L2 , and recursively merge the rest of pairs
in L00 . The type of the result of mergeP airs is list of lists, however, it will be
flattened by sort0 function finally.
The merge function is as same as before. The complete example Haskell
program is given as below.
mergeP airs(L) =

mergesort = sort groupBy ( )

sort [] = []
sort [xs] = xs
sort xss = sort (mergePairs xss) where
mergePairs (xs:ys:xss) = merge xs ys : mergePairs xss
mergePairs xss = xss

Alternatively, observing that we can first pick two sub lists, merge them to
an intermediate result, then repeatedly pick next sub list, and merge to this
ordered result weve gotten so far until all the rest sub lists are merged. This is
a typical folding algorithm as introduced in appendix A.
sort(L) = f old(merge, , group(L))

(13.39)

Translate this version to Haskell yields the folding version.

mergesort = foldl merge [] groupBy ( )

Exercise 13.6
Is the nature merge sort algorithm realized by folding is equivalent with
the one by using mergeP airs in terms of performance? If yes, prove it; If
not, which one is faster?

13.11

Bottom-up merge sort

The worst case analysis for nature merge sort raises an interesting topic, instead
of realizing merge sort in top-down manner, we can develop a bottom-up version.

13.11. BOTTOM-UP MERGE SORT

423

The great advantage is that, we neednt do book keeping any more, so the
algorithm is quite friendly for purely iterative implementation.
The idea of bottom-up merge sort is to turn the sequence to be sorted into n
small sub sequences each contains only one element. Then we merge every two
of such small sub sequences, so that we get n2 ordered sub sequences each with
length 2; If n is odd number, we left the last singleton sequence untouched. We
repeatedly merge these pairs, and finally we get the sorted result. Knuth names
this variant as straight two-way merge sort [1]. The bottom-up merge sort is
illustrated in figure 13.19
...

...

Figure 13.19: Bottom-up merge sort

Different with the basic version and even-odd version, we neednt explicitly
split the list to be sorted in every recursion. The whole list is split into n
singletons at the very beginning, and we merge these sub lists in the rest of the
algorithm.

wraps(L) =

sort(L) = sort0 (wraps(L))

(13.40)

: L=
{{l1 }} wraps(L0 ) : otherwise

(13.41)

Of course wraps can be implemented by using mapping as introduced in

appendix A.
sort(L) = sort0 (map(x {x}, L))
0

(13.42)

We reuse the function sort and mergeP airs which are defined in section of
nature merge sort. They repeatedly merge pairs of sub lists until there is only
one.
Implement this version in Haskell gives the following example code.
sort = sort map (x[x])

424CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
This version is based on what Okasaki presented in [6]. It is quite similar to
the nature merge sort only differs in the way of grouping. Actually, it can be
deduced as a special case (the worst case) of nature merge sort by the following
equation.
sort(L) = sort0 (groupBy(x,y F alse, L))

(13.43)

That instead of spanning the non-decreasing sub list as long as possible, the
predicate always evaluates to false, so the sub list spans only one element.
Similar with nature merge sort, bottom-up merge sort can also be defined
by folding. The detailed implementation is left as exercise to the reader.
Observing the bottom-up sort, we can find its in tail-recursion call manner,
thus its quite easy to translate into purely iterative algorithm without any
recursion.
1: function Sort(A)
2:
B
3:
for a A do
4:
B Append({a})
5:
N |B|
6:
while N > 1 do
7:
for i from 1 to b N2 c do
8:
B[i] Merge(B[2i 1], B[2i])
9:
if Odd(N ) then
10:
B[d N2 e] B[N ]
11:
12:
13:
14:

N d N2 e
if B = then
return
return B[1]

The following example Python program implements the purely iterative

bottom-up merge sort.
def mergesort(xs):
ys = [[x] for x in xs]
while len(ys) > 1:
ys.append(merge(ys.pop(0), ys.pop(0)))
return [] if ys == [] else ys.pop()
def merge(xs, ys):
zs = []
while xs != [] and ys !=[]:
zs.append(xs.pop(0) if xs[0] < ys[0] else ys.pop(0))
return zs + (xs if xs !=[] else ys)

The Python implementation exploit the fact that instead of starting next
round of merging after all pairs have been merged, we can combine these rounds
of merging by consuming the pair of lists on the head, and appending the merged
result to the tail. This greatly simply the logic of handling odd sub lists case as
shown in the above pseudo code.

Exercise 13.7
Implement the functional bottom-up merge sort by using folding.

13.12. PARALLELISM

425

Implement the iterative bottom-up merge sort only with array indexing.
Dont use any library supported tools, such as list, vector etc.

13.12

Parallelism

We mentioned in the basic version of quick sort, that the two sub sequences
can be sorted in parallel after the divide phase finished. This strategy is also
applicable for merge sort. Actually, the parallel version quick sort and morege
sort, do not only distribute the recursive sub sequences sorting into two parallel
processes, but divide the sequences into p sub sequences, where p is the number
of processors. Idealy, if we can achieve sorting in T 0 time with parallelism,
which satisifies O(n lg n) = pT 0 . We say it is linear speed up, and the algorithm
is parallel optimal.
However, a straightforward parallel extension to the sequential quick sort
algorithm which samples several pivots, divides p sub sequences, and independantly sorts them in parallel, isnt optimal. The bottleneck exists in the divide
phase, which we can only achive O(n) time in average case.
The straightforward parallel extention to merge sort, on the other hand,
block at the merge phase. Both parallel merge sort and quick sort in practice
need good designes in order to achieve the optimal speed up. Actually, the
divide and conqure nature makes merge sort and quick sort relative easy for
parallelisim. Richard Cole found the O(lg n) parallel merge sort algorithm with
n processors in 1986 in [13].
Parallelism is a big and complex topic which is out of the scope of this
elementary book. Readers can refer to [13] and [14] for details.

13.13

Short summary

In this chapter, two popular divide and conquer sorting methods, quick sort
and merge sort are introduced. Both of them meet the upper performance
limit of the comparison based sorting algorithms O(n lg n). Sedgewick said
that quick sort is the greatest algorithm invented in the 20th century. Almost
all programming environments adopt quick sort as the default sorting tool. As
time goes on, some environments, especially those manipulate abstract sequence
which is dynamic and not based on pure array switch to merge sort as the general
purpose sorting tool6 .
The reason for this interesting phenomena can be partly explained by the
treatment in this chapter. That quick sort performs perfectly in most cases,
it needs fewer swapping than most other algorithms. However, the quick sort
algorithm is based on swapping, in purely functional settings, swapping isnt
the most efficient way due to the underlying data structure is singly linked-list,
but not vectorized array. Merge sort, on the other hand, is friendly in such
environment, as it costs constant spaces, and the performance can be ensured
even in the worst case of quick sort, while the latter downgrade to quadratic
time. However, merge sort doesnt performs as well as quick sort in purely imperative settings with arrays. It either needs extra spaces for merging, which is
6 Actually, most of them are kind of hybrid sort, balanced with insertion sort to achieve
good performance when the sequence is short

426CHAPTER 13. DIVIDE AND CONQUER, QUICK SORT VS. MERGE SORT
sometimes unreasonable, for example in embedded system with limited memory,
or causes many overhead swaps by in-place workaround. In-place merging is till
an active research area.
Although the title of this chapter is quick sort vs. merge sort, its not
the case that one algorithm has nothing to do with the other. Quick sort can
be viewed as the optimized version of tree sort as explained in this chapter.
Similarly, merge sort can also be deduced from tree sort as shown in [12].
There are many ways to categorize sorting algorithms, such as in [1]. One
way is to from the point of view of easy/hard partition, and easy/hard merge
[7].
Quick sort, for example, is quite easy for merging, because all the elements
in the sub sequence before the pivot are no greater than any one after the pivot.
The merging for quick sort is actually trivial sequence concatenation.
Merge sort, on the other hand, is more complex in merging than quick sort.
However, its quite easy to divide no matter what concrete divide method is
taken: simple divide at the middle point, even-odd splitting, nature splitting,
or bottom-up straight splitting. Compare to merge sort, its more difficult for
quick sort to achieve a perfect dividing. We show that in theory, the worst
case cant be completely avoided, no matter what engineering practice is taken,
median-of-three, random quick sort, 3-way partition etc.
Weve shown some elementary sorting algorithms in this book till this chapter, including insertion sort, tree sort, selection sort, heap sort, quick sort and
merge sort. Sorting is still a hot research area in computer science. At the
time when I this chapter is written, people are challenged by the buzz word big
data, that the traditional convenient method cant handle more and more huge
data within reasonable time and resources. Sorting a sequence of hundreds of
Gigabytes becomes a routine in some fields.

Exercise 13.8
Design an algorithm to create binary search tree by using merge sort
strategy.

Bibliography
[1] Donald E. Knuth. The Art of Computer Programming, Volume 3: Sorting
and Searching (2nd Edition). Addison-Wesley Professional; 2 edition (May
4, 1998) ISBN-10: 0201896850 ISBN-13: 978-0201896855
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford
Stein. Introduction to Algorithms, Second Edition. ISBN:0262032937.
The MIT Press. 2001
[3] Robert Sedgewick. Implementing quick sort programs. Communication
of ACM. Volume 21, Number 10. 1978. pp.847 - 857.
[4] Jon Bentley. Programming pearls, Second Edition. Addison-Wesley Professional; 1999. ISBN-13: 978-0201657883
[5] Jon Bentley, Douglas McIlroy. Engineering a sort function. Software
Practice and experience VOL. 23(11), 1249-1265 1993.
[6] Robert
Sedgewick,
Jon
Bentley.
Quicksort
is
optimal.
http://www.cs.princeton.edu/ rs/talks/QuicksortIsOptimal.pdf
[7] Richard Bird. Pearls of functional algorithm design. Cambridge University Press. 2010. ISBN, 1139490605, 9781139490603
[8] Fethi Rabhi, Guy Lapalme. Algorithms: a functional programming approach. Second edition. Addison-Wesley, 1999. ISBN: 0201-59604-0
[9] Simon Peyton Jones. The Implementation of functional programming languages. Prentice-Hall International, 1987. ISBN: 0-13-453333-X
[10] Jyrki Katajainen, Tomi Pasanen, Jukka Teuhola. Practical in-place mergesort. Nordic Journal of Computing, 1996.
[11] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[12] Jos`e Bacelar Almeida and Jorge Sousa Pinto. Deriving Sorting Algorithms. Technical report, Data structures and Algorithms. 2008.
[13] Cole, Richard (August 1988). Parallel merge sort. SIAM J. Comput. 17
(4): 770C785. doi:10.1137/0217049. (August 1988)
[14] Powers, David M. W. Parallelized Quicksort and Radixsort with Optimal
Speedup, Proceedings of International Conference on Parallel Computing
Technologies. Novosibirsk. 1991.
427

428

Searching

[15] Wikipedia. Quicksort. http://en.wikipedia.org/wiki/Quicksort

[16] Wikipedia. Strict weak order. http://en.wikipedia.org/wiki/Strict weak order
[17] Wikipedia. Total order. http://en.wokipedia.org/wiki/Total order
[18] Wikipedia.
Harmonic
series
(mathematics).
http://en.wikipedia.org/wiki/Harmonic series (mathematics)

Chapter 14

Searching
14.1

Introduction

Searching is quite a big and important area. Computer makes many hard searching problems realistic. They are almost impossible for human beings. A modern
industry robot can even search and pick the correct gadget from the pipeline for
assembly; A GPS car navigator can search among the map, for the best route
to a specific place. The modern mobile phone is not only equipped with such
map navigator, but it can also search for the best price for Internet shopping.
This chapter just scratches the surface of elementary searching. One good
thing that computer offers is the brute-force scanning for a certain result in a
large sequence. The divide and conquer search strategy will be briefed with two
problems, one is to find the k-th big one among a list of unsorted elements; the
other is the popular binary search among a list of sorted elements. Well also
introduce the extension of binary search for multiple-dimension data.
Text matching is also very important in our daily life, two well-known searching algorithms, Knuth-Morris-Pratt (KMP) and Boyer-Moore algorithms will be
introduced. They set good examples for another searching strategy: information
reusing.
Besides sequence search, some elementary methods for searching solution for
some interesting problems will be introduced. They were mostly well studied
in the early phase of AI (artificial intelligence), including the basic DFS (Depth
first search), and BFS (Breadth first search).
Finally, Dynamic programming will be briefed for searching optimal solutions, and well also introduce about greedy algorithm which is applicable for
some special cases.
All algorithms will be realized in both imperative and functional approaches.

14.2

Sequence search

Although modern computer offers fast speed for brute-force searching, and even
if the Moores law could be strictly followed, the grows of huge data is too fast
to be handled well in this way. Weve seen a vivid example in the introduction
chapter of this book. Its why people study the computer search algorithms.
429

430

14.2.1

CHAPTER 14. SEARCHING

Divide and conquer search

One solution is to use divide and conquer approach. That if we can repeatedly
scale down the search domain, the data being dropped neednt be examined at
all. This will definitely speed up the search.
k-selection problem
Consider a problem of finding the k-th smallest one among n elements. The
most straightforward idea is to find the minimum first, then drop it and find
the second minimum element among the rest. Repeat this minimum finding and
dropping k steps will give the k-th smallest one. Finding the minimum among
n elements costs linear O(n) time. Thus this method performs O(kn) time, if k
is much smaller than n.
Another method is to use the heap data structure weve introduced. No
matter what concrete heap is used, e.g. binary heap with implicit array, Fibonacci heap or others, Accessing the top element followed by popping is typically bound O(lg n) time. Thus this method, as formalized in equation (14.1)
and (14.2) performs in O(k lg n) time, if k is much smaller than n.
top(k, L) = f ind(k, heapif y(L))

f ind(k, H) =

top(H) : k = 0
f ind(k 1, pop(H)) : otherwise

(14.1)
(14.2)

However, heap adds some complexity to the solution. Is there any simple,
fast method to find the k-th element?
The divide and conquer strategy can help us. If we can divide all the elements
into two sub lists A and B, and ensure all the elements in A is not greater than
any elements in B, we can scale down the problem by following this method1 :
1. Compare the length of sub list A and k;
2. If k < |A|, the k-th smallest one must be contained in A, we can drop B
and further search in A;
3. If |A| < k, the k-th smallest one must be contained in B, we can drop A
and further search the (k |A|)-th smallest one in B.
Note that the italic font emphasizes the fact of recursion. The ideal case
always divides the list into two equally big sub lists A and B, so that we can
halve the problem each time. Such ideal case leads to a performance of O(n)
linear time.
Thus the key problem is how to realize dividing, which collects the first m
smallest elements in one sub list, and put the rest in another.
This reminds us the partition algorithm in quick sort, which moves all the
elements smaller than the pivot in front of it, and moves those greater than
the pivot behind it. Based on this idea, we can develop a divide and conquer
k-selection algorithm, which is called quick selection algorithm.
1 This actually demands a more accurate definition of the k-th smallest in L: Its equal to
the k-the element of L0 , where L0 is a permutation of L, and L0 is in monotonic non-decreasing
order.

14.2. SEQUENCE SEARCH

431

1. Randomly select an element (the first for instance) as the pivot;

2. Moves all elements which arent greater than the pivot in a sub list A; and
moves the rest to sub list B;
3. Compare the length of A with k, if |A| = k 1, then the pivot is the k-th
smallest one;
4. If |A| > k 1, recursively find the k-th smallest one among A;
5. Otherwise, recursively find the (k |A|)-th smallest one among B;
This algorithm can be formalized in below equation. Suppose 0 < k |L|
, where L is a non-empty list of elements. Denote l1 as the first element in L.
It is chosen as the pivot; L0 contains the rest elements except for l1 . (A, B) =
partition(x x l1 , L0 ). It partitions L0 by using the same algorithm defined
in the chapter of quick sort.

l1 : |A| = k 1

top(k 1 |A|, B) : |A| < k 1

top(k, L) =
(14.3)

top(k, A) : otherwise

(, ) : L =
({l1 } A, B) : p(l1 ), (A, B) = partition(p, L0 )

(A, {l1 } B) : p(l1 )

(14.4)
The following Haskell example program implements this algorithm.
partition(p, L) =

top n (x:xs) | len == n - 1 = x

| len < n - 1 = top (n - len - 1) bs
| otherwise = top n as
where
(as, bs) = partition ( x) xs
len = length as

The partition function is provided in Haskell standard library, the detailed

implementation can be referred to previous chapter about quick sort.
The lucky case is that, the k-th smallest element is selected as the pivot at
the very beginning. The partition function examines the whole list, and finds
that there are k 1 elements not greater than the pivot, we are done in just
O(n) time. The worst case is that either the maximum or the minimum element
is selected as the pivot every time. The partition always produces an empty sub
list, that either A or B is empty. If we always pick the minimum as the pivot,
the performance is bound to O(kn). If we always pick the maximum as the
pivot, the performance is O((n k)n). If k is much less than n, it downgrades
to quadratic O(n2 ) time.
The best case (not the lucky case), is that the pivot always partition the
list perfectly. The length of A is nearly as same as the length of B. The list
is halved every time. It needs about O(lg n) partitions, each partition takes
linear time proportion to the length of the halved list. This can be expressed
as O(n + n2 + n4 + ... + 2nm ), where m is the smallest number satisfies 2nm < k.
Summing the series leads to the result of O(n).

432

CHAPTER 14. SEARCHING

The average case analysis needs tool of mathematical expectation. Its quite
similar to the proof given in previous chapter of quick sort. Its left as an exercise
to the reader.
Similar as quick sort, this divide and conquer selection algorithm performs
well most time in practice. We can take the same engineering practice such as
media-of-three, or randomly select the pivot as we did for quick sort. Below is
the imperative realization for example.
1: function Top(k, A, l, u)
2:
Exchange A[l] A[ Random(l, u) ]
. Randomly select in [l, u]
3:
p Partition(A, l, u)
4:
if p l + 1 = k then
5:
return A[p]
6:
if k < p l + 1 then
7:
return Top(k, A, l, p 1)
8:
return Top(k p + l 1, A, p + 1, u)
This algorithm searches the k-th smallest element in range of [l, u] for array
A. The boundaries are included. It first randomly selects a position, and swaps
it with the first one. Then this element is chosen as the pivot for partitioning.
The partition algorithm in-place moves elements and returns the position where
the pivot being moved. If the pivot is just located at position k, then we are
done; if there are more than k 1 elements not greater than the pivot, the
algorithm recursively searches the k-th smallest one in range [l, p1]; otherwise,
k is deduced by the number of elements before the pivot, and recursively searches
the range after the pivot [p + 1, u].
There are many methods to realize the partition algorithm, below one is
based on N. Lumotos method. Other realizations are left as exercises to the
reader.
1: function Partition(A, l, u)
2:
p A[l]
3:
Ll
4:
for R l + 1 to u do
5:
if (p < A[R]) then
6:
LL+1
7:
Exchange A[L] A[R]
8:
Exchange A[L] p
9:
return L
Below ANSI C example program implements this algorithm. Note that it
handles the special case that either the array is empty, or k is out of the boundaries of the array. It returns -1 to indicate the search failure.
int partition(Key xs, int l, int u) {
int r, p = l;
for (r = l + 1; r < u; ++r)
if (!(xs[p] < xs[r]))
swap(xs, ++l, r);
swap(xs, p, l);
return l;
}
/ The result is stored in xs[k], returns k if u-l k, otherwise -1 /

14.2. SEQUENCE SEARCH

433

int top(int k, Key xs, int l, int u) {

int p;
if (l < u) {
swap(xs, l, rand() % (u - l) + l);
p = partition(xs, l, u);
if (p - l + 1 == k)
return p;
return (k < p - l + 1) ? top(k, xs, l, p) :
top(k- p + l - 1, xs, p + 1, u);
}
return -1;
}

There is a method proposed by Blum, Floyd, Pratt, Rivest and Tarjan in

1973, which ensures the worst case performance being bound to O(n) [2], [3].
It divides the list into small groups. Each group contains no more than 5
elements. The median of each group among these 5 elements are identified
quickly. Then there are n5 median elements selected. We repeat this step, and
divide them again into groups of 5, and recursively select the median of median.
Its obviously that the final true median can be found in O(lg n) time. This is
the best pivot for partitioning the list. Next, we halve the list by this pivot and
recursively search for the k-th smallest one. The performance can be calculated
as the following.
n
T (n) = c1 lgn + c2 n + T ( )
2

(14.5)

Where c1 and c2 are constant factors for the median of median and partition
computation respectively. Solving this equation with telescope method or the
master theory in [2] gives the linear O(n) performance. The detailed algorithm
realization is left as exercise to the reader.
In case we just want to pick the top k smallest elements, but dont care
about the order of them, the algorithm can be adjusted a little bit to fit.

:
A :
tops(k, L) =
A {l1 } tops(k |A| 1, B) :

tops(k, A) :

k =0L=
|A| = k
|A| < k
otherwise

(14.6)

Where A, B have the same meaning as before that, (A, B) = partition(x

x l1 , L0 ) if L isnt empty. The relative example program in Haskell is given
as below.
tops _ [] = []
tops 0 _ = []
tops n (x:xs) | len ==n = as
| len < n = as ++ [x] ++ tops (n-len-1) bs
| otherwise = tops n as
where
(as, bs) = partition ( x) xs
len = length as

434

CHAPTER 14. SEARCHING

binary search
Another popular divide and conquer algorithm is binary search. Weve shown
it in the chapter about insertion sort. When I was in school, the teacher who
taught math played a magic to me, He asked me to consider a natural number
less than 1000. Then he asked me some questions, I only replied yes or no, and
finally he guessed my number. He typically asked questions like the following:
Is it an even number?
Is it a prime number?
Are all digits same?
Can it be divided by 3?
...
Most of the time he guessed the number within 10 questions. My classmates
and I all thought its unbelievable.
This game will not be so interesting if it downgrades to a popular TV program, that the price of a product is hidden, and you must figure out the exact
price in 30 seconds. The host of the program tells you if your guess is higher
or lower to the fact. If you win, the product is yours. The best strategy is to
use similar divide and conquer approach to perform a binary search. So its
common to find such conversation between the player and the host:
P: 1000;
H: High;
P: 500;
H: Low;
P: 750;
H: Low;
P: 890;
H: Low;
P: 990;
H: Bingo.
My math teacher told us that, because the number we considered is within
1000, if he can halve the numbers every time by designing good questions, the
number will be found in 10 questions. This is because 210 = 1024 > 1000.
However, it would be boring to just ask it is higher than 500, is lower than 250,
... Actually, the question is it even is very good, because it always halve the
numbers.
Come back to the binary search algorithm. It is only applicable to a sequence
of ordered number. Ive seen programmers tried to apply it to unsorted array,
and took several hours to figure out why it doesnt work. The idea is quite

14.2. SEQUENCE SEARCH

435

straightforward, in order to find a number x in an ordered sequence A, we

firstly check middle point number, compare it with x, if they are same, then
we are done; If x is smaller, as A is ordered, we need only recursively search it
among the first half; otherwise we search it among the second half. Once A gets
empty and we havent found x yet, it means x doesnt exist.
Before formalizing this algorithm, there is a surprising fact need to be noted.
Donald Knuth stated that Although the basic idea of binary search is comparatively straightforward, the details can be surprisingly tricky. Jon Bentley
pointed out that most binary search implementation contains errors, and even
the one given by him in the first version of Programming pearls contains an
error undetected over twenty years [4].
There are two kinds of realization, one is recursive, the other is iterative.
The recursive solution is as same as what we described. Suppose the lower and
upper boundaries of the array are l and u inclusive.
1: function Binary-Search(x, A, l, u)
2:
if u < l then
3:
Not found error
4:
else
5:
m l + b ul
. avoid overflow of b l+u
2 c
2 c
6:
if A[m] = x then
7:
return m
8:
if x < A[m] then
9:
return Binary-Search(x, A, l, m - 1)
10:
else
11:
return Binary-Search(x, A, m + 1, u)
As the comment highlights, if the integer is represented with limited words,
we cant merely use b l+u
2 c because it may cause overflow if l and u are big.
Binary search can also be realized in iterative manner, that we keep updating
the boundaries according to the middle point comparison result.
1: function Binary-Search(x, A, l, u)
2:
while l < u do
3:
m l + b ul
2 c
4:
if A[m] = x then
5:
return m
6:
if x < A[m] then
7:
um1
8:
else
9:
l m+1
return NIL
The implementation is very good exercise, we left it to the reader. Please
try all kinds of methods to verify your program.
Since the array is halved every time, the performance of binary search is
bound to O(lg n) time.
In purely functional settings, the list is represented with singly linked-list.
Its linear time to randomly access the element for a given position. Binary
search doesnt make sense in such case. However, it good to analyze what the
performance will downgrade to. Consider the following equation.

436

CHAPTER 14. SEARCHING

Err
b1
bsearch(x, L) =

bsearch(x,
A)

bsearch(x, B 0 )

:
:
:
:

L=
x = b1 , (A, B) = splitAt(b |L|
2 c, L)
B = x < b1
otherwise

Where b1 is the first element if B isnt empty, and B 0 holds the rest except
for b1 . The splitAt function takes O(n) time to divide the list into two subs A
and B (see the appendix A, and the chapter about merge sort for detail). If B
isnt empty and x is equal to b1 , the search returns; Otherwise if it is less than
b1 , as the list is sorted, we need recursively search in A, otherwise, we search in
B. If the list is empty, we raise error to indicate search failure.
As we always split the list in the middle point, the number of elements halves
in each recursion. In every recursive call, we takes linear time for splitting. The
splitting function only traverses the first half of the linked-list, Thus the total
time can be expressed as.
n
n
n
+ c + c + ...
2
4
8
This results O(n) time, which is as same as the brute force search from head
to tail:

Err : L =

l1 : x = l 1
search(x, L) =

search(x, L0 ) : otherwise
T (n) = c

As we mentioned in the chapter about insertion sort, the functional approach

of binary search is through binary search tree. That the ordered sequence is
represented in a tree ( self balanced tree if necessary), which offers logarithm
time searching 2 .
Although it doesnt make sense to apply divide and conquer binary sort on
linked-list, binary search can still be very useful in purely functional settings.
Consider solving an equation ax = y, for given natural numbers a and y, where
a y. We want to find the integer solution for x if there is. Of course bruteforce naive searching can solve it. We can examine all numbers one by one
from 0 for a0 , a1 , a2 , ..., stops if ai = y or report that there is no solution if
ai < y < ai+1 for some i. We initialize the solution domain as X = {0, 1, 2, ...},
and call the below exhausted searching function solve(a, y, X).

x1 : ax1 = y

solve(a, y, X 0 ) : ax1 < y

solve(a, y, X) =

Err : otherwise
This function examines the solution domain in monotonic increasing order.
It takes the first candidate element x1 from X, compare ax1 and y, if they are
equal, then x1 is the solution and we are done; if it is less than y, then x1 is
dropped, and we search among the rest elements represented as X 0 ; Otherwise,
since f (x) = ax is non-decreasing function when a is natural number, so the rest
2 Some readers may argue that array should be used instead of linked-list, for example in
Haskell. This book only deals with purely functional sequences in finger-tree. Different from
the Haskell array, it cant support constant time random accessing

14.2. SEQUENCE SEARCH

437

elements will only make f (x) bigger and bigger. There is no integer solution for
this equation. The function returns error to indicate no solution.
The computation of ax is expensive for big a and x if precession must be
kept3 . Can it be improved so that we can compute as less as possible? The
divide and conquer binary search can help. Actually, we can estimate the upper
limit of the solution domain. As ay y, We can search in range {0, 1, ..., y}. As
the function f (x) = ax is non-decreasing against its argument x, we can firstly
xm
check the middle point candidate xm = b 0+y
= y, the solution is found;
2 c, if a
if it is less than y, we can drop all candidate solutions before xm ; otherwise we
drop all candidate solutions after it; Both halve the solution domain. We repeat
this approach until either the solution is found or the solution domain becomes
empty, which indicates there is no integer solution.
The binary search method can be formalized as the following equation. The
non-decreasing function is abstracted as a parameter. To solve our problem, we
can just call it as bsearch(f, y, 0, y), where f (x) = ax .

Err
m
bsearch(f, y, l, u) =
bsearch(f,
y,
l,
m

bsearch(f, y, m + 1, u)

:
:
:
:

u<l
f (m) = y, m = b l+u
2 c
f (m) > y
f (m) < y
(14.7)
As we halve the solution domain in every recursion, this method computes
f (x) in O(log y) times. It is much faster than the brute-force searching.
2 dimensions search
Its quite natural to think that the idea of binary search can be extended to 2
dimensions or even more general multiple-dimensions domain. However, it is
not so easy.
Consider the example of a m n matrix M . The elements in each row and
each column are in strict increasing order. Figure 14.1 illustrates such a matrix
for example.

1
2

3
4

2
4
5
6

3
5
7
8

4
6
8
9

...
...

...
......

Figure 14.1: A matrix in strict increasing order for each row and column.
Given a value x, how to locate all elements equal to x in the matrix quickly?
We need develop an algorithm, which returns a list of locations (i, j) so that
Mi,j = x.
Richard Bird in [1] mentioned that he used this problem to interview candidates for entry to Oxford. The interesting story was that, those who had some
3 One alternative is to reuse the result of an when compute an+1 = aan . Here we consider
for general form monotonic function f (n)

438

CHAPTER 14. SEARCHING

computer background at school tended to use binary search. But its easy to
get stuck.
The usual way follows binary search idea is to examine element at M m2 , n2 .
If it is less than x, we can only drop the elements in the top-left area; If it
is greater than x, only the bottom-right area can be dropped. Both cases are
illustrated in figure 14.2, the gray areas indicate elements can be dropped.

Figure 14.2: Left: the middle point element is smaller than x. All elements in
the gray area are less than x; Right: the middle point element is greater than
x. All elements in the gray area are greater than x.
The problem is that the solution domain changes from a rectangle to a L
shape in both cases. We cant just recursively apply search on it. In order to
solve this problem systematically, we define the problem more generally, using
brute-force search as a start point, and keep improving it bit by bit.
Consider a function f (x, y), which is strict increasing for its arguments, for
instance f (x, y) = ax + by , where a and b are natural numbers. Given a value
z, which is a natural number too, we want to solve the equation f (x, y) = z by
finding all candidate pairs (x, y).
With this definition, the matrix search problem can be specialized by below
function.

f (x, y) =

Mx,y : 1 x m, 1 y n
1 : otherwise

Brute-force 2D search
As all solutions should be found for f (x, y). One can immediately give the brute
force solution by embedded looping.
1: function Solve(f, z)
2:
A
3:
for x {0, 1, 2, ..., z} do
4:
for y {0, 1, 2, ..., z} do
5:
if f (x, y) = z then
6:
A A {(x, y)}
7:
return A

14.2. SEQUENCE SEARCH

439

This definitely calculates f for (z + 1)2 times. It can be formalized as in

(14.8).
solve(f, z) = {(x, y)|x {0, 1, ..., z}, y {0, 1, ..., z}, f (x, y) = z}

(14.8)

Saddleback search
We havent utilize the fact that f (x, y) is strict increasing yet. Dijkstra pointed
out in [6], instead of searching from bottom-left corner, starting from the topleft leads to one effective solution. As illustrated in figure 14.3, the search starts
from (0, z), for every point (p, q), we compare f (p, q) with z:
If f (p, q) < z, since f is strict increasing, for all 0 y < q, we have
f (p, y) < z. We can drop all points in the vertical line section (in red
color);
If f (p, q) > z, then f (x, q) > z for all p < x z. We can drop all points
in the horizontal line section (in blue color);
Otherwise if f (p, q) = z, we mark (p, q) as one solution, then both line
sections can be dropped.
This is a systematical way to scale down the solution domain rectangle. We
keep dropping a row, or a column, or both.

Figure 14.3: Search from top-left.

This method can be formalized as a function search(f, z, p, q), which searches
solutions for equation f (x, y) = z in rectangle with top-left corner (p, q), and
bottom-right corner (z, 0). We start the searching by initializing (p, q) = (0, z)
as solve(f, z) = search(f, z, 0, z)

search(f, z, p + 1, q)
search(f, z, p, q) =
search(f, z, p, q 1)

{(p, q)} search(f, z, p + 1, q 1)

:
:
:
:

p>zq <0
f (p, q) < z
f (p, q) > z
otherwise
(14.9)

440

CHAPTER 14. SEARCHING

The first clause is the edge case, there is no solution if (p, q) isnt top-left to
(z, 0). The following example Haskell program implements this algorithm.
solve f z = search 0 z where
search p q | p > z | | q < 0 = []
| z < z = search (p + 1) q
| z > z = search p (q - 1)
| otherwise = (p, q) : search (p + 1) (q - 1)
where z = f p q

Considering the calculation of f may be expensive, this program stores the

result of f (p, q) to variable z 0 . This algorithm can also be implemented in
iterative manner, that the boundaries of solution domain keeps being updated
in a loop.
1: function Solve(f, z)
2:
p 0, q z
3:
S
4:
while p z q 0 do
5:
z 0 f (p, q)
6:
if z 0 < z then
7:
pp+1
8:
else if z 0 > z then
9:
q q1
10:
else
11:
S S {(p, q)}
12:
p p + 1, q q 1
13:
return S
Its intuitive to translate this imperative algorithm to real program, as the
following example Python code.
def solve(f, z):
(p, q) = (0, z)
res = []
while p z and q 0:
z1 = f(p, q)
if z1 < z:
p=p+1
elif z1 > z:
q=q - 1
else:
res.append((p, q))
(p, q) = (p + 1, q - 1)
return res

It is clear that in every iteration, At least one of p and q advances to the

bottom-right corner by one. Thus it takes at most 2(z + 1) steps to complete
searching. This is the worst case. There are three best cases. The first one
happens that in every iteration, both p and q advance by one, so that it needs
only z + 1 steps; The second case keeps advancing horizontally to right and ends
when p exceeds z; The last case is similar, that it keeps moving down vertically
to the bottom until q becomes negative.
Figure 14.4 illustrates the best cases and the worst cases respectively. Figure
14.4 (a) is the case that every point (x, z x) in diagonal satisfies f (x, z x) = z,

14.2. SEQUENCE SEARCH

441

it uses z + 1 steps to arrive at (z, 0); (b) is the case that every point (x, z) along
the top horizontal line gives the result f (x, z) < z, the algorithm takes z + 1
steps to finish; (c) is the case that every point (0, x) along the left vertical line
gives the result f (0, x) > z, thus the algorithm takes z + 1 steps to finish; (d) is
the worst case. If we project all the horizontal sections along the search path to
x axis, and all the vertical sections to y axis, it gives the total steps of 2(z + 1).

Figure 14.4: The best cases and the worst cases.

Compare to the quadratic brute-force method (O(z 2 )), we improve to a linear
algorithm bound to O(z).
Bird imagined that the name saddleback is because the 3D plot of f with
the smallest bottom-left and the latest top-right and two wings looks like a
saddle as shown in figure 14.5

Figure 14.5: Plot of f (x, y) = x2 + y 2 .

Improved saddleback search

We havent utilized the binary search tool so far, even the problem extends to
2-dimension domain. The basic saddleback search starts from the top-left corner

442

CHAPTER 14. SEARCHING

(0, z) to the bottom-right corner (z, 0). This is actually over-general domain.
we can constraint it a bit more accurate.
Since f is strict increasing, we can find the biggest number m, that 0 m
z, along the y axis which satisfies f (0, m) z; Similarly, we can find the biggest
n, that 0 n z, along the x axis, which satisfies f (n, 0) z; And the solution
domain shrinks from (0, z) (z, 0) to (0, m) (n, 0) as shown in figure 14.6.

Figure 14.6: A more accurate search domain shown in gray color.

Of course m and n can be found by brute-force like below.
m = max({y|0 y z, f (0, y) z})
n = max({x|0 x z, f (x, 0) z})

(14.10)

When searching m, the x variable of f is bound to 0. It turns to be one

dimension search problem for a strict increasing function (or in functional term,
a Curried function f (0, y)). Binary search works in such case. However, we
need a bit modification for equation (14.7). Different from searching a solution
l x u, so that f (x) = y for a given y; we need search for a solution l x u
so that f (x) y < f (x + 1).

ul
f (m) y < f (m + 1), m = b l+u
2 c
f (m) y
otherwise
(14.11)
The first clause handles the edge case of empty range. The lower boundary
is returned in such case; If the middle point produces a value less than or equal
to the target, while the next one evaluates to a bigger value, then the middle
point is what we are looking for; Otherwise if the point next to the middle also
evaluates to a value not greater than the target, the lower bound is increased
by one, and we perform recursively binary search; In the last case, the middle
point evaluates to a value greater than the target, upper bound is updated as
the point proceeds to the middle for further recursive searching. The following
Haskell example code implements this modified binary search.
l
m
bsearch(f, y, l, u) =
bsearch(f, y, m + 1, u)

bsearch(f, y, l, m 1)

bsearch f y (l, u) | u l = l

:
:
:
:

14.2. SEQUENCE SEARCH

443

| f m y = if f (m + 1) y
then bsearch f y (m + 1, u) else m
| otherwise = bsearch f y (l, m-1)
where m = (l + u) div 2

Then m and n can be found with this binary search function.

m = bsearch(y f (0, y), z, 0, z)
n = bsearch(x f (x, 0), z, 0, z)

(14.12)

And the improved saddleback search shrinks to this new search domain
solve(f, z) = search(f, z, 0, m):

p>nq <0
f (p, q) < z
f (p, q) > z
otherwise
(14.13)
Its almost as same as the basic saddleback version, except that it stops
if p exceeds n, but not z. In real implementation, the result of f (p, q) can
be calculated once, and stored in a variable as shown in the following Haskell
example.

search(f, z, p + 1, q)
search(f, z, p, q) =
search(f, z, p, q 1)

{(p, q)} search(f, z, p + 1, q 1)

:
:
:
:

solve f z = search 0 m where

This improved saddleback search firstly performs binary search two rounds
to find the proper m, and n. Each round is bound to O(lg z) times of calculation
for f ; After that, it takes O(m + n) time in the worst case; and O(min(m, n))
time in the best case. The overall performance is given in the following table.
times of evaluation f
worst case 2 log z + m + n
best case
2 log z + min(m, n)
For some function f (x, y) = ax + by , for positive integers a and b, m and n
will be relative small, that the performance is close to O(lg z).
This algorithm can also be realized in imperative approach. Firstly, the
binary search should be modified.
1: function Binary-Search(f, y, (l, u))
2:
while l < u do
3:
m b l+u
2 c
4:
if f (m) y then
5:
if y < f (m + 1) then
6:
return m
7:
l m+1
8:
else
9:
um

444

CHAPTER 14. SEARCHING

return l
Utilize this algorithm, the boundaries m and n can be found before performing the saddleback search.
1: function Solve(f, z)
2:
m Binary-Search(y f (0, y), z, (0, z))
3:
n Binary-Search(x f (x, 0), z, (0, z))
4:
p 0, q m
5:
S
6:
while p n q 0 do
7:
z 0 f (p, q)
8:
if z 0 < z then
9:
pp+1
10:
else if z 0 > z then
11:
q q1
12:
else
13:
S S {(p, q)}
14:
p p + 1, q q 1
15:
return S
The implementation is left as exercise to the reader.
10:

More improvement to saddleback search

In figure 14.2, two cases are shown for comparing the value of the middle point
in a matrix with the given value. One case is the center value is smaller than
the given value, the other is bigger. In both cases, we can only throw away 14
candidates, and left a L shape for further searching.
Actually, one important case is missing. We can extend the observation to
any point inside the rectangle searching area. As shown in the figure 14.7.
Suppose we are searching in a rectangle from the lower-left corner (a, b) to
the upper-right corner (c, d). If the (p, q) isnt the middle point, and f (p, q) 6= z.
We cant ensure the area to be dropped is always 1/4. However, if f (p, q) = z,
as f is strict increasing, we are not only sure both the lower-left and the upperright sub areas can be thrown, but also all the other points in the column p and
row q. The problem can be scaled down fast, because only 1/2 area is left.
This indicates us, instead of jumping to the middle point to start searching.
A more efficient way is to find a point which evaluates to the target value. One
straightforward way to find such a point, is to perform binary search along the
center horizontal line or the center vertical line of the rectangle.
The performance of binary search along a line is logarithmic to the length
of that line. A good idea is to always pick the shorter center line as shown in
figure 14.8. That if the height of the rectangle is longer than the width, we
perform binary search along the horizontal center line; otherwise we choose the
vertical center line.
However, what if we cant find a point (p, q) in the center line, that satisfies
f (p, q) = z? Lets take the center horizontal line for example. even in such case,
we can still find a point that f (p, q) < z < f (p + 1, q). The only difference is
that we cant drop the points in row p and q completely.
Combine this conditions, the binary search along the horizontally line is
to find a p, satisfies f (p, q) z < f (p + 1, q); While the vertical line search

14.2. SEQUENCE SEARCH

(a) If f (p, q) 6= z, only lower-left or upper-right sub area

(in gray color) can be thrown. Both left a L shape.

(b) If f (p, q) = z, both sub areas can be thrown, the scale

of the problem is halved.

Figure 14.7: The efficiency of scaling down the search domain.

Figure 14.8: Binary search along the shorter center line.

445

446

CHAPTER 14. SEARCHING

condition is f (p, q) z < f (p, q + 1).

The modified binary search ensures that, if all points in the line segment
give f (p, q) < z, the upper bound will be found; and the lower bound will be
found if they all greater than z. We can drop the whole area on one side of the
center line in such case.
Sum up all the ideas, we can develop the efficient improved saddleback search
as the following.
1. Perform binary search along the y axis and x axis to find the tight boundaries from (0, m) to (n, 0);
2. Denote the candidate rectangle as (a, b) (c, d), if the candidate rectangle
is empty, the solution is empty;
3. If the height of the rectangle is longer than the width, perform binary
search along the center horizontal line; otherwise, perform binary search
along the center vertical line; denote the search result as (p, q);
4. If f (p, q) = z, record (p, q) as a solution, and recursively search two sub
rectangles (a, b) (p 1, q + 1) and (p + 1, q 1) (c, d);
5. Otherwise, f (p, q) 6= z, recursively search the same two sub rectangles
plus a line section. The line section is either (p, q + 1) (p, b) as shown in
figure 14.9 (a); or (p + 1, q) (c, q) as shown in figure 14.9 (b).

Figure 14.9: Recursively search the gray areas, the bold line should be included
if f (p, q) 6= z.
This algorithm can be formalized as the following. The equation (14.11),
and (14.12) are as same as before. A new search function should be defined.
Define Search(a,b),(c,d) as a function for searching rectangle with top-left
corner (a, b), and bottom-right corner (c, d).

: c<ad<b

csearch : c a < b d
search(a,b),(c,d) =
(14.14)

rsearch : otherwise
Function csearch performs binary search in the center horizontal line to
find a point (p, q) that f (p, q) z < f (p + 1, q). This is shown in figure 14.9

14.2. SEQUENCE SEARCH

447

(a). There is a special edge case, that all points in the lines evaluate to values
greater than z. The general binary search will return the lower bound as result,
so that (p, q) = (a, b b+d
2 c). The whole upper side includes the center line can
be dropped as shown in figure 14.10 (a).

Figure 14.10: Edge cases when performing binary search in the center line.

csearch =

search(p,q1),(c,d)
search(a,b),(p1,q+1) {(p, q)} search(p+1,q1),(c,d)

search(a,b),(p,q+1) search(p+1,q1),(c,d)

: z < f (p, q)
: f (p, q) = z
: otherwise
(14.15)

Where
q = b b+d
2 c)
p = bsearch(x f (x, q), z, (a, c))
Function rsearch is quite similar except that it searches in the center horizontal line.

rsearch =

search(a,b),(p1,q)
search(a,b),(p1,q+1) {(p, q)} search(p+1,q1),(c,d)

search(a,b),(p1,q+1) search(p+1,q),(c,d)

: z < f (p, q)
: f (p, q) = z
: otherwise
(14.16)

Where
p = b a+c
2 c)
q = bsearch(y f (p, y), z, (d, b))
The following Haskell program implements this algorithm.
search f z (a, b) (c, d) | c < a | | b < d = []
| c - a < b - d = let q = (b + d) div 2 in
csearch (bsearch (x f x q) z (a, c), q)
| otherwise = let p = (a + c) div 2 in
rsearch (p, bsearch (f p) z (d, b))
where
csearch (p, q) | z < f p q = search f z (p, q - 1) (c, d)

448

CHAPTER 14. SEARCHING

| f p q == z = search f z (a, b) (p - 1, q + 1) ++
(p, q) : search f z (p + 1, q - 1) (c, d)
| otherwise = search f z (a, b) (p, q + 1) ++
search f z (p + 1, q - 1) (c, d)
rsearch (p, q) | z < f p q = search f z (a, b) (p - 1, q)
| f p q == z = search f z (a, b) (p - 1, q + 1) ++
(p, q) : search f z (p + 1, q - 1) (c, d)
| otherwise = search f z (a, b) (p - 1, q + 1) ++
search f z (p + 1, q) (c, d)

And the main program calls this function after performing binary search in
X and Y axes.
solve f z = search f z (0, m) (n, 0) where
m = bsearch (f 0) z (0, z)
n = bsearch (x f x 0) z (0, z)

Since we drop half areas in every recursion, it takes O(log(mn)) rounds of

search. However, in order to locate the point (p, q), which halves the problem,
we must perform binary search along the center line. which will call f about
O(log(min(m, n))) times. Denote the time of searching a m n rectangle as
T (m, n), the recursion relationship can be represented as the following.
T (m, n) = log(min(m, n)) + 2T (

m n
, )
2 2

(14.17)

Suppose m > n, using telescope method, for m = 2i , and n = 2j . We have:

T (2i , 2j ) = j + 2T (2i1 , 2j1 )
i1
X
=
2k (j k)
k=0

(14.18)

= O(2 (j i))
= O(m log(n/m))
i

Richard Bird proved that this is asymptotically optimal by a lower bound

of searching a given value in m n rectangle [1].
The imperative algorithm is almost as same as the functional version. We
skip it for the sake of brevity.

Exercise 14.1
Prove that the average case for the divide and conquer solution to kselection problem is O(n). Please refer to previous chapter about quick
sort.
Implement the imperative k-selection problem with 2-way partition, and
median-of-three pivot selection.
Implement the imperative k-selection problem to handle duplicated elements effectively.
Realize the median-of-median k-selection algorithm and implement it in
your favorite programming language.

14.2. SEQUENCE SEARCH

449

The tops(k, L) algorithm uses list concatenation likes A {l1 } tops(k

|A| 1, B). It is linear operation which is proportion to the length of the
list to be concatenated. Modify the algorithm so that the sub lists are
concatenated by one pass.
The author considered another divide and conquer solution for the kselection problem. It finds the maximum of the first k elements and the
minimum of the rest. Denote them as x, and y. If x is smaller than y, it
means that all the first k elements are smaller than the rest, so that they
are exactly the top k smallest; Otherwise, There are some elements in the
first k should be swapped.
1: procedure Tops(k, A)
2:
l1
3:
u |A|
4:
loop
5:
i Max-At(A[l..k])
6:
j Min-At(A[k + 1..u])
7:
if A[i] < A[j] then
8:
break
9:
Exchange A[l] A[j]
10:
Exchange A[k + 1] A[i]
11:
l Partition(A, l, k)
12:
u Partition(A, k + 1, u)
Explain why this algorithm works? Whats the performance of it?
Implement the binary search algorithm in both recursive and iterative
manner, and try to verify your version automatically. You can either
generate randomized data, test your program with the binary search invariant, or compare with the built-in binary search tool in your standard
library.
Implement the improved saddleback search by firstly performing binary
search to find a more accurate solution domain in your favorite imperative
programming language.
Realize the improved 2D search, by performing binary search along the
shorter center line, in your favorite imperative programming language.
Someone considers that the 2D search can be designed as the following.
When search a rectangle, as the minimum value is at bottom-left, and
the maximum at to-right. If the target value is less than the minimum
or greater than the maximum, then there is no solution; otherwise, the
rectangle is divided into 4 sub rectangles at the center point, then perform
recursively searching.
1: procedure Search(f, z, a, b, c, d)
. (a, b): bottom-left (c, d):
top-right
2:
if z f (a, b) f (c, d) z then
3:
if z = f (a, b) then
4:
record (a, b) as a solution
5:
if z = f (c, d) then

450

CHAPTER 14. SEARCHING

6:
7:
8:
9:
10:
11:
12:
13:

record (c, d) as a solution

return
p b a+c
2 c
q b b+d
2 c
Search(f, z, a, q, p, d)
Search(f, z, p, q, c, d)
Search(f, z, a, b, p, q)
Search(f, z, p, b, c, q)

Whats the performance of this algorithm?

14.2.2

Information reuse

One interesting behavior that is that people learning while searching. We do not
only remember lessons which cause search fails, but also learn patterns which
lead to success. This is a kind of information reusing, no matter the information
is positive or negative. However, Its not easy to determine what information
should be kept. Too little information isnt enough to help effective searching,
while keeping too much is expensive in term of spaces.
In this section, well first introduce two interesting problems, Boyer-Moore
majority number problem and the maximum sum of sub vector problem. Both
reuse information as little as possible. After that, two popular string matching
algorithms, Knuth-Morris-Pratt algorithm and Boyer-Moore algorithm will be
introduced.
Boyer-Moore majority number
Voting is quite critical to people. We use voting to choose the leader, make
decision or reject a proposal. In the months when I was writing this chapter,
there are three countries in the world voted their presidents. All of the three
voting activities utilized computer to calculate the result.
Suppose there is a country in a small island wants a new president. According
to the constitution, only if the candidate wins more than half of the votes can
be selected as the president. Given a serious of votes, such as A, B, A, C, B,
B, D, ..., can we develop a program tells who is the new president if there is, or
indicate nobody wins more than half of the votes?
Of course this problem can be solved with brute-force by using a map. As
what we did in the chapter of binary search tree4 .
template<typename T>
T majority(const T xs, int n, T fail) {
map<T, int> m;
int i, max = 0;
T r;
for (i = 0; i < n; ++i)
++m[xs[i]];
for (typename map<T, int>::iterator it = m.begin(); it != m.end(); ++it)
if (itsecond > max) {
max = itsecond;
4 There is a probabilistic sub-linear space counting algorithm published in 2004, named as
Count-min sketch[8].

14.2. SEQUENCE SEARCH

451

r = itfirst;
}
return max 2 > n ? r : fail;
}

This program first scan the votes, and accumulates the number of votes for
each individual with a map. After that, it traverse the map to find the one with
the most of votes. If the number is bigger than the half, the winner is found
otherwise, it returns a value to indicate fail.

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:

The following pseudo code describes this algorithm.

function Majority(A)
M empty map
for a A do
Put(M , a, 1+ Get(M, a))
max 0, m N IL
for (k, v) M do
if max < v then
max v, m k
if max > |A|50% then
return m
else
fail

For m individuals and n votes, this program firstly takes about O(n log m)
time to build the map if the map is implemented in self balanced tree (red-black
tree for instance); or about O(n) time if the map is hash table based. However,
the hash table needs more space. Next the program takes O(m) time to traverse
the map, and find the majority vote. The following table lists the time and space
performance for different maps.
map
self-balanced tree
hashing

time
O(n log m)
O(n)

space
O(m)
O(m) at least

Boyer and Moore invented a cleaver algorithm in 1980, which can pick the
majority element with only one scan if there is. Their algorithm only needs
O(1) space [7].
The idea is to record the first candidate as the winner so far, and mark
him with 1 vote. During the scan process, if the winner being selected gets
another vote, we just increase the vote counter; otherwise, it means somebody
vote against this candidate, so the vote counter should be decreased by one. If
the vote counter becomes zero, it means this candidate is voted out; We select
the next candidate as the new winner and repeat the above scanning process.
Suppose there is a series of votes: A, B, C, B, B, C, A, B, A, B, B, D, B.
Below table illustrates the steps of this processing.

452

CHAPTER 14. SEARCHING

winner count scan position

A
1
A, B, C, B, B, C, A, B, A, B, B, D, B
A
0
A, B, C, B, B, C, A, B, A, B, B, D, B
C
1
A, B, C, B, B, C, A, B, A, B, B, D, B
C
0
A, B, C, B, B, C, A, B, A, B, B, D, B
B
1
A, B, C, B, B, C, A, B, A, B, B, D, B
B
0
A, B, C, B, B, C, A, B, A, B, B, D, B
A
1
A, B, C, B, B, C, A, B, A, B, B, D, B
A
0
A, B, C, B, B, C, A, B, A, B, B, D, B
A
1
A, B, C, B, B, C, A, B, A, B, B, D, B
A
0
A, B, C, B, B, C, A, B, A, B, B, D, B
B
1
A, B, C, B, B, C, A, B, A, B, B, D, B
B
0
A, B, C, B, B, C, A, B, A, B, B, D, B
B
1
A, B, C, B, B, C, A, B, A, B, B, D, B
The key point is that, if there exits the majority greater than 50%, it cant
be voted out by all the others. However, if there are not any candidates win
more than half of the votes, the recorded winner is invalid. Thus it is necessary
to perform a second round scan for verification.
The following pseudo code illustrates this algorithm.
1: function Majority(A)
2:
c0
3:
for i 1 to |A| do
4:
if c = 0 then
5:
x A[i]
6:
if A[i] = x then
7:
cc+1
8:
else
9:
cc1
10:
return x
If there is the majority element, this algorithm takes one pass to scan the
votes. In every iteration, it either increases or decreases the counter according
to the vote is support or against the current selection. If the counter becomes
zero, it means the current selection is voted out. So the new one is selected as
the updated candidate for further scan.
The process is linear O(n) time, and the spaces needed are just two variables.
One for recording the selected candidate so far, the other is for vote counting.
Although this algorithm can find the majority element if there is. it still
picks an element even there isnt. The following modified algorithm verifies the
final result with another round of scan.
1: function Majority(A)
2:
c0
3:
for i 1 to |A| do
4:
if c = 0 then
5:
x A[i]
6:
if A[i] = x then
7:
cc+1
8:
else
9:
cc1

14.2. SEQUENCE SEARCH

453

c0
for i 1 to |A| do
if A[i] = x then
13:
cc+1
14:
if c > %50|A| then
15:
return x
16:
else
17:
fail
Even with this verification process, the algorithm is still bound to O(n) time,
and the space needed is constant. The following ISO C++ program implements
this algorithm 5 .
10:
11:
12:

template<typename T>
T majority(const T xs, int n, T fail) {
T m;
int i, c;
for (i = 0, c = 0; i < n; ++i) {
if (!c)
m = xs[i];
c += xs[i] == m ? 1 : -1;
}
for (i = 0, c = 0; i < n; ++i, c += xs[i] == m);
return c 2 > n ? m : fail;
}

Boyer-Moore majority algorithm can also be realized in purely functional

approach. Different from the imperative settings, which use variables to record
and update information, accumulators are used to define the core algorithm.
Define function maj(c, n, L), which takes a list of votes L, a selected candidate
c so far, and a counter n. For non empty list L, we initialize c as the first vote
l1 , and set the counter as 1 to start the algorithm: maj(l1 , 1, L0 ), where L0 is
the rest votes except for l1 . Below are the definition of this function.

c : L=

maj(c, n + 1, L0 ) : l1 = c
maj(c, n, L) =
(14.19)
maj(l1 , 1, L0 ) : n = 0 l1 6= c

0
maj(c, n 1, L ) : otherwise
We also need to define a function, which can verify the result. The idea is
that, if the list of votes is empty, the final result is a failure; otherwise, we start
the Boyer-Moore algorithm to find a candidate c, then we scan the list again to
count the total votes c wins, and verify if this number is not less than the half.

f ail : L =
c : c = maj(l1 , 1, L0 ), |{x|x L, x = c}| > %50|L|
majority(L) =

f ail : otherwise
(14.20)
Below Haskell example code implements this algorithm.
majority :: (Eq a) [a] Maybe a
5 We actually uses the ANSI C style. The C++ template is only used to generalize the
type of the element

454

CHAPTER 14. SEARCHING

majority [] = Nothing
majority (x:xs) = let m = maj x 1 xs in verify m (x:xs)
maj c n [] = c
maj c n (x:xs) | c == x = maj c (n+1) xs
| n == 0 = maj x 1 xs
| otherwise = maj c (n-1) xs
verify m xs = if 2 (length $ filter (==m) xs) > length xs
then Just m else Nothing

Maximum sum of sub vector

Jon Bentley presents another interesting puzzle which can be solved by using
quite similar idea in [4]. The problem is to find the maximum sum of sub vector.
For example in the following array, The sub vector {19, -12, 1, 9, 18} yields the
biggest sum 35.
3 -13 19 -12 1 9 18 -16 15 -15
Note that it is only required to output the value of the maximum sum. If
all the numbers are positive, the answer is definitely the sum of all. Another
special case is that all numbers are negative. We define the maximum sum is 0
for an empty sub vector.
Of course we can find the answer with brute-force, by calculating all sums of
sub vectors and picking the maximum. Such naive method is typical quadratic.
1: function Max-Sum(A)
2:
m0
3:
for i 1 to |A| do
4:
s0
5:
for j i to |A| do
6:
s s + A[j]
7:
m Max(m, s)
8:
return m
The brute force algorithm does not reuse any information in previous search.
Similar with Boyer-Moore majority vote algorithm, we can record the maximum
sum end to the position where we are scanning. Of course we also need record
the biggest sum found so far. The following figure illustrates this idea and the
invariant during scan.
i

...

max

...

max end at i

...

Figure 14.11: Invariant during scan.

At any time when we scan to the i-th position, the max sum found so far is
recorded as A. At the same time, we also record the biggest sum end at i as B.
Note that A and B may not be the same, in fact, we always maintain B A. and
when B becomes greater than A by adding with the next element, we update A

14.2. SEQUENCE SEARCH

455

with this new value. When B becomes negative, this happens when the next element is a negative number, we reset it to 0. The following tables illustrated the
steps when we scan the example vector {3, 13, 19, 12, 1, 9, 18, 16, 15, 15}.
max sum max end at i
list to be scan
0
0
{3, 13, 19, 12, 1, 9, 18, 16, 15, 15}
3
3
{13, 19, 12, 1, 9, 18, 16, 15, 15}
3
0
{19, 12, 1, 9, 18, 16, 15, 15}
19
19
{12, 1, 9, 18, 16, 15, 15}
19
7
{1, 9, 18, 16, 15, 15}
19
8
{9, 18, 16, 15, 15}
19
17
{18, 16, 15, 15}
35
35
{16, 15, 15}
35
19
{15, 15}
35
34
{15}
35
19
{}
This algorithm can be described as below.
1: function Max-Sum(V )
2:
A 0, B 0
3:
for i 1 to |V | do
4:
B Max(B + V [i], 0)
5:
A Max(A, B)
It is trivial to implement this linear time algorithm, that we skip the details
here.
This algorithm can also be defined in functional approach. Instead of mutating variables, we use accumulator to record A and B. In order to search the
maximum sum of list L, we call the below function with maxsum (0, 0, L).

maxsum (A, B, L) =
Where

A : L=
maxsum (A0 , B 0 , L0 ) : otherwise

(14.21)

B 0 = max(l1 + B, 0)
A0 = max(A, B 0 )

Below Haskell example code implements this algorithm.

maxsum = msum 0 0 where
msum a _ [] = a
msum a b (x:xs) = let b = max (x+b) 0
a = max a b
in msum a b xs

KMP
String matching is another important type of searching. Almost all the software
editors are equipped with tools to find string in the text. In chapters about Trie,
Patricia, and suffix tree, we have introduced some powerful data structures
which can help to search string. In this section, we introduce another two string
matching algorithms all based on information reusing.
Some programming environments provide built-in string search tools, however, most of them are brute-force solution including strstr function in ANSI

456

CHAPTER 14. SEARCHING

C standard library, find in C++ standard template library, indexOf in Java

Development Kit etc. Figure 14.12 illustrate how such character-by-character
comparison process works.
a

(a) The offset s = 4, after matching q = 4 characters, the 5th mismatch.

(b) Move s = 4 + 2 = 6, directly.

Figure 14.12: Match ananym in any ananthous ananym flower.

Suppose we search a pattern P in text T , as shown in figure 14.12 (a), at
offset s = 4, the process examines every character in P and T to check if they
are same. It successfully matches the first 4 characters anan. However, the
5th character in the pattern string is y. It doesnt match the corresponding
character in the text, which is t.
At this stage, the brute-force solution terminates the attempt, increases s by
one to 5, and restart the comparison between ananym and nantho.... Actually,
we can increase s not only by one. This is because we have already known that
the first four characters anan have been matched, and the failure happens at
the 5th position. Observe the two letters prefix an of the pattern string is also
a suffix of anan that we have matched so far. A more effective way is to shift
s by two but not one, which is shown in figure 14.12 (b). By this means, we
reused the information that 4 characters have been matched. This helps us to
skip invalid positions as many as possible.
Knuth, Morris and Pratt presented this idea in [9] and developed a novel
string matching algorithm. This algorithm is later called as KMP, which is
consist of the three authors initials.
For the sake of brevity, we denote the first k characters of text T as Tk .
Which means Tk is the k-character prefix of T .
The key point to shift s effectively is to find a function of q, where q is the
number of characters matched successfully. For instance, q is 4 in figure 14.12
(a), as the 5th character doesnt match.
Consider what situation we can shift s more than 1. As shown in figure
14.13, if we can shift the pattern P ahead, there must exist k, so that the first k
characters are as same as the last k characters of Pq . In other words, the prefix
Pk is suffix of Pq .

14.2. SEQUENCE SEARCH

...

T[i]

P[1]

T[i+1]

P[2]

457
T[i+2]

...

P[j]

P[1]

... ... ... ...

P[j+1]

P[2]

...

T[i+q-1]

P[q]

P[k]

...

Figure 14.13: Pk is both prefix of Pq and suffix of Pq .

Its possible that there is no such a prefix that is the suffix at the same time.
If we treat empty string as both the prefix and the suffix of any others, there
must be at least one solution that k = 0. Its also quite possible that there are
multiple k satisfy. To avoid missing any possible matching positions, we have
to find the biggest k. We can define a prefix function (q) which tells us where
we can fallback if the (q + 1)-th character does not match [2].
(q) = max{k|k < q Pk A Pq }

(14.22)

Where A is read as is suffix of. For instance, A A B means A is suffix of

B. This function is used as the following. When we match pattern P against
text T from offset s, If it fails after matching q characters, we next look up (q)
to get a fallback q 0 , and retry to compare P [q 0 ] with the previous unmatched
character. Based on this idea, the core algorithm of KMP can be described as
the following.
1: function KMP(T, P )
2:
n |T |, m |P |
3:
build prefix function from P
4:
q0
. How many characters have been matched so far.
5:
for i 1 to n do
6:
while q > 0 P [q + 1] 6= T [i] do
7:
q (q)
8:
if P [q + 1] = T [i] then
9:
q q+1
10:
if q = m then
11:
found one solution at i m
12:
q (q)
. look for next solution
Although the definition of prefix function (q) is given in equation (14.22),
realizing it blindly by finding the longest suffix isnt effective. Actually we can
use the idea of information reusing again to build the prefix function.
The trivial edge case is that, the first character doesnt match. In this case
the longest prefix, which is also the suffix is definitely empty, so (1) = k = 0.
We record the longest prefix as Pk . In this edge case Pk = P0 is the empty
string.
After that, when we scan at the q-th character in the pattern string P , we
hold the invariant that the prefix function values (i) for i in {1, 2, ..., q 1}
have already been recorded, and Pk is the longest prefix which is also the suffix
of Pq1 . As shown in figure 14.14, if P [q] = P [k + 1], A bigger k than before

458

CHAPTER 14. SEARCHING

is found, we can increase the maximum of k by one; otherwise, if they are not
same, we can use (k) to fallback to a shorter prefix Pk0 where k 0 = (k), and
check if the next character after this new prefix is same as the q-th character.
We need repeat this step until either k becomes zero (which means only empty
string satisfies), or the q-th character matches.
P[1]

P[2]

...

P[k]

P[k+1]

...

P[q-1]

P[q]

...

?
P[1]

P[2]

...

P[k]

P[k+1]

...

Figure 14.14: Pk is suffix of Pq1 , P [q] and P [k + 1] are compared.

Realizing this idea gives the KMP prefix building algorithm.
function Build-Prefix-Function(P )
2:
m |P |, k 0
3:
(1) 0
4:
for q 2 to m do
5:
while k > 0 P [q] 6= P [k + 1] do
6:
k (k)
7:
if P [q] = P [k + 1] then
8:
k k+1
9:
(q) k
10:
return
The following table lists the steps of building prefix function for pattern
string ananym. Note that the k in the table actually means the maximum k
satisfies equation (14.22).
q
Pq k Pk
1
a 0
2
an 0
3
ana 1 a
4
anan 2 an
5
anany 0
6 ananym 0
Translating the KMP algorithm to Python gives the below example code.
1:

def kmp_match(w, p):

n = len(w)
m = len(p)
fallback = fprefix(p)
k = 0 # how many elements have been matched so far.
res = []
for i in range(n):
while k > 0 and p[k] != w[i]:
k = fallback[k] #fall back
if p[k] == w[i]:
k=k+1
if k == m:
res.append(i+1-m)

14.2. SEQUENCE SEARCH

459

k = fallback[k-1] # look for next

return res
def fprefix(p):
m = len(p)
t = [0]m # fallback table
k=0
for i in range(2, m):
while k>0 and p[i-1] != p[k]:
k = t[k-1] #fallback
if p[i-1] == p[k]:
k=k+1
t[i] = k
return t

The KMP algorithm builds the prefix function for the pattern string as a
kind of pre-processing before the search. Because of this, it can reuse as much
information of the previous matching as possible.
The amortized performance of building the prefix function is O(m). This
can be proved by using potential method as in [2]. Using the similar method, it
can be proved that the matching algorithm itself is also linear. Thus the total
performance is O(m + n) at the expense of the O(m) space to record the prefix
function table.
It seems that varies pattern string would affect the performance of KMP.
Considering the case that we are finding pattern string aaa...a of length m in a
string aaa...a of length n. All the characters are same, when the last character
in the pattern is examined, we can only fallback by 1, and this 1 character
fallback repeats until it falls back to zero. Even in this extreme case, KMP
algorithm still holds its linear performance (why?). Please try to consider more
cases such as P = aaaa...b, T = aaaa...a and so on.
Purely functional KMP algorithm
It is not easy to realize KMP matching algorithm in purely functional manner. The imperative algorithm represented so far intensely uses array to record
prefix function values. Although it is possible to utilize sequence like structure in purely functional settings, such sequence is typically implemented with
finger tree. Unlike native arrays, finger tree needs logarithm time for random
accessing6 .
Richard Bird presents a formal program deduction to KMP algorithm by
using fold fusion law in chapter 17 of [1]. In this section, we show how to
develop purely functional KMP algorithm step by step from a brute-force prefix
function creation method.
Both text string and pattern are represented as singly linked-list in purely
functional settings. During the scan process, these two lists are further partitioned, every one is broken into two parts. As shown in figure 14.15, The first j
characters in the pattern string have been matched. T [i + 1] and P [j + 1] will be
compared next. If they are same, we need append the character to the matched
6 Again, we dont use native array, even it is supported in some functional programming
environments like Haskell.

460

CHAPTER 14. SEARCHING

part. However, since strings are essentially singly linked list, such appending is
proportion to j.
T[1]

T[2]

... ...

T[i-1]

T[i]

T[i+1]

T[i+2]

... ...

P[j+2]

...

T[n-1]

T[n]

?
s

P[1]

P[2]

...

P[j]

P[j+1]

P[m]

Figure 14.15: Pk is suffix of Pq1 , P [q] and P [k + 1] are compared.

Denote the first i characters as Tp , which means the prefix of T , the rest
characters as Ts for suffix; Similarly, the first j characters as Pp , and the rest
as Ps ; Denote the first character of Ts as t, the first character of Ps as p. We
have the following cons relationship.
Ts = cons(t, Ts0 )
Ps = cons(p, Ps0 )
If t = p, note the following updating process is bound to linear time.
Tp0 = Tp {t}
Pp0 = Pp {p}
Weve introduced a method in the chapter about purely functional queue,
which can solve this problem. By using a pair of front and rear list, we can
turn the linear time appending to constant time linking. The key point is to
represent the prefix part in reverse order.

T = Tp Ts = reverse(reverse(Tp )) Ts = reverse(Tp ) Ts

P = Pp Ps = reverse(reverse(Pp )) Ps = reverse(Pp ) Ps

(14.23)

The idea is to using pair (Tp , Ts ) and (Pp , Ps ) instead. With this change,
the if t = p, we can update the prefix part fast in constant time.

Tp0 = cons(t, Tp )

Pp = cons(p, Pp )

(14.24)

The KMP matching algorithm starts by initializing the success prefix parts
to empty strings as the following.
search(P, T ) = kmp(, (, P )(, T ))

(14.25)

Where is the prefix function we explained before. The core part of KMP

14.2. SEQUENCE SEARCH

461

algorithm, except for the prefix function building, can be defined as below.

{|Tp |}

{|Tp } kmp(, (Pp , Ps ), (Tp , Ts ))

kmp(, (Pp0 , Ps0 ), (Tp0 , Ts0 ))

kmp(, (Pp , Ps ), (Tp0 , Ts0 ))

kmp(, (Pp , Ps ), (Tp , Ts ))

Ps = Ts =
Ps 6= Ts =
Ps = Ts 6=

kmp(, (Pp , Ps ), (Tp , Ts )) =

t=p

:
t
6= p Pp =

: t 6= p Pp 6=
(14.26)
The first clause states that, if the scan successfully ends to both the pattern
and text strings, we get a solution, and the algorithm terminates. Note that we
use the right position in the text string as the matching point. Its easy to use
the left position by subtracting with the length of the pattern string. For sake
of brevity, we switch to right position in functional solutions.
The second clause states that if the scan arrives at the end of text string,
while there are still rest of characters in the pattern string havent been matched,
there is no solution. And the algorithm terminates.
The third clause states that, if all the characters in the pattern string have
been successfully matched, while there are still characters in the text havent
been examined, we get a solution, and we fallback by calling prefix function
to go on searching other solutions.
The fourth clause deals with the case, that the next character in pattern
string and text are same. In such case, the algorithm advances one character
ahead, and recursively performs searching.
If the the next characters are not same and this is the first character in the
pattern string, we just need advance to next character in the text, and try again.
Otherwise if this isnt the first character in the pattern, we call prefix function
to fallback, and try again.
The brute-force way to build the prefix function is just to follow the definition
equation (14.22).

(Pp , Ps ) = (Pp0 , Ps0 )

:
:
:
:

(14.27)

where
Pp0 = longest({s|s pref ixes(Pp ), s A Pp })
Ps0 = P Pp0
Every time when calculate the fallback position, the algorithm naively enumerates all prefixes of Pp , checks if it is also the suffix of Pp , and then pick the
longest one as result. Note that we reuse the subtraction symbol here for list
differ operation.
There is a tricky case which should be avoided. Because any string itself is
both its prefix and suffix. Say Pp @ Pp and Pp A Pp . We shouldnt enumerate
Pp as a candidate prefix. One solution of such prefix enumeration can be realized
as the following.

pref ixes(L) =

{} : L = |L| = 1
cons(, map(s cons(l1 , s), pref ixes(L0 ))) : otherwise
(14.28)

462

CHAPTER 14. SEARCHING

Below Haskell example program implements this version of string matching

algorithm.
kmpSearch1 ptn text = kmpSearch next ([], ptn) ([], text)
kmpSearch _ (sp, []) (sw, []) = [length sw]
kmpSearch _ _ (_, []) = []
kmpSearch f (sp, []) (sw, ws) = length sw : kmpSearch f (f sp []) (sw, ws)
kmpSearch f (sp, (p:ps)) (sw, (w:ws))
| p == w = kmpSearch f ((p:sp), ps) ((w:sw), ws)
| otherwise = if sp ==[] then kmpSearch f (sp, (p:ps)) ((w:sw), ws)
else kmpSearch f (f sp (p:ps)) (sw, (w:ws))
next sp ps = (sp, ps) where
prev = reverse sp
prefix = longest [xs | xs inits prev, xs isSuffixOf prev]
sp = reverse prefix
ps = (prev ++ ps) \\ prefix
longest = maximumBy (compare on length)
inits [] = [[]]
inits [_] = [[]]
inits (x:xs) = [] : (map (x:) $ inits xs)

This version does not only perform poorly, but it is also complex. We can
simplify it a bit. Observing the KMP matching is a scan process from left to
the right of the text, it can be represented with folding (refer to Appendix A for
detail). Firstly, we can augment each character with an index for folding like
below.
zip(T, {1, 2, ...})

(14.29)

Zipping the text string with infinity natural numbers gives list of pairs. For
example, text string The quick brown fox jumps over the lazy dog turns into
(T, 1), (h, 2), (e, 3), ... (o, 42), (g, 43).
The initial state for folding contains two parts, one is the pair of pattern
(Pp , Ps ), with prefix starts from empty, and the suffix is the whole pattern
string (, P ). For illustration purpose only, we revert back to normal pairs

but not (Pp , Ps ) notation. It can be easily replaced with reversed form in the
finalized version. This is left as exercise to the reader. The other part is a
list of positions, where the successful matching are found. It starts from empty
list. After the folding finishes, this list contains all solutions. What we need
is to extract this list from the final state. The core KMP search algorithm is
simplified like this.

kmp(P, T ) = snd(f old(search, ((, P ), ), zip(T, {1, 2, ...})))

(14.30)

The only black box is the search function, which takes a state, and a pair
of character and index, and it returns a new state as result. Denote the first
character in Ps as p and the rest characters as Ps0 (Ps = cons(p, Ps0 )), we have

14.2. SEQUENCE SEARCH

463

the following definition.

((Pp p, Ps0 ), L {i})

((Pp p, Ps0 ), L)
search(((Pp , Ps ), L), (c, i)) =
((Pp , Ps ), L)

search(((Pp , Ps ), L), (c, i))

p = c Ps0 =
p = c Ps0 6=
Pp =
otherwise
(14.31)
If the first character in Ps matches the current character c during scan, we
need further check if all the characters in the pattern have been examined, if so,
we successfully find a solution, This position i in list L is recorded; Otherwise,
we advance one character ahead and go on. If p does not match c, we need
fallback for further retry. However, there is an edge case that we cant fallback
any more. Pp is empty in this case, and we need do nothing but keep the current
state.
The prefix-function developed so far can also be improved a bit. Since we
want to find the longest prefix of Pp , which is also suffix of it, we can scan from
right to left instead. For any non empty list L, denote the first element as l1 ,
and all the rest except for the first one as L0 , define a function init(L), which
returns all the elements except for the last one as below.

: |L| = 1
init(L) =
(14.32)
cons(l1 , init(L0 )) : otherwise
:
:
:
:

Note that this function can not handle empty list. The idea of scan from
right to left for Pp is first check if init(Pp ) A Pp , if yes, then we are done;
otherwise, we examine if init(init(Pp )) is OK, and repeat this till the left most.
Based on this idea, the prefix-function can be modified as the following.

(Pp , Ps ) =

(Pp , Ps ) :
f allback(init(Pp ), cons(last(Pp ), Ps )) :

Pp =
otherwise

(14.33)

Where

f allback(A, B) =

(A, B) :
(init(A), cons(last(A), B)) :

A A Pp
otherwise

(14.34)

Note that fallback always terminates because empty string is suffix of any
string. The last(L) function returns the last element of a list, it is also a
linear time operation (refer to Appendix A for detail). However, its constant

operation if we use Pp approach. This improved prefix-function is bound to

linear time. It is still quite slower than the imperative algorithm which can
look up prefix-function in constant O(1) time. The following Haskell example
program implements this minor improvement.
failure ([], ys) = ([], ys)
failure (xs, ys) = fallback (init xs) (last xs:ys) where
fallback as bs | as isSuffixOf xs = (as, bs)
| otherwise = fallback (init as) (last as:bs)
kmpSearch ws txt = snd $ foldl f (([], ws), []) (zip txt [1..]) where

464

CHAPTER 14. SEARCHING

f (p@(xs, (y:ys)), ns) (x, n) | x == y = if ys==[] then ((xs++[y], ys), ns++[n])
else ((xs++[y], ys), ns)
| xs == [] = (p, ns)
| otherwise = f (failure p, ns) (x, n)
f (p, ns) e = f (failure p, ns) e

The bottleneck is that we can not use native array to record prefix functions
in purely functional settings. In fact the prefix function can be understood as
a state transform function. It transfer from one state to the other according to
the matching is success or fail. We can abstract such state changing as a tree.
In environment supporting algebraic data type, Haskell for example, such state
tree can be defined like below.
data State a = E | S a (State a) (State a)

A state is either empty, or contains three parts: the current state, the new
state if match fails, and the new state if match succeeds. Such definition is quite
similar to the binary tree. We can call it left-fail, right-success tree. The state
we are using here is (Pp , Ps ).
Similar as imperative KMP algorithm, which builds the prefix function from
the pattern string, the state transforming tree can also be built from the pattern.
The idea is to build the tree from the very beginning state (, P ), with both
its children empty. We replace the left child with a new state by calling
function defined above, and replace the right child by advancing one character
ahead. There is an edge case, that when the state transfers to (P, ), we can
not advance any more in success case, such node only contains child for failure
case. The build function is defined as the following.

build((Pp , Ps ), , ) =

build((Pp , Ps ), , ) : Ps =
build((Pp , Ps ), L, R) : otherwise

(14.35)

Where
L = build((Pp , Ps ), , )
R = build((Ps {p}, Ps0 ), , ))
The meaning of p and Ps0 are as same as before, that p is the first character
in Ps , and Ps0 is the rest characters. The most interesting point is that the build
function will never stop. It endless build a infinite tree. In strict programming
environment, calling this function will freeze. However, in environments support
lazy evaluation, only the nodes have to be used will be created. For example,
both Haskell and Scheme/Lisp are capable to construct such infinite state tree.
In imperative settings, it is typically realized by using pointers which links to
ancestor of a node.
Figure 14.16 illustrates such an infinite state tree for pattern string ananym.
Note that the right most edge represents the case that the matching continuously
succeed for all characters. After that, since we cant match any more, so the
right sub-tree is empty. Base on this fact, we can define a auxiliary function to
test if a state indicates the whole pattern is successfully matched.

match((Pp , Ps ), L, R) =

T rue : Ps =
F alse : otherwise

(14.36)

14.2. SEQUENCE SEARCH

465

(, ananym)
fail
(, ananym)
fail
(, ananym)

match

(a, ananym)

match
(a, nanym)
fail
(, ananym)

match
(an, anym)
fail

...

(, ananym)

match
(ana, nym)
fail

...

(a, nanym)

match
(anan, ym)
fail

...

match

(an, anym)

(anany, m)
fail

(, ananym)

match
(ananym, )
fail

(, ananym)

Figure 14.16: The infinite state tree for pattern ananym.

empty

466

CHAPTER 14. SEARCHING

With the help of state transform tree, we can realize KMP algorithm in an
automaton manner.

kmp(P, T ) = snd(f old(search, (T r, []), zip(T, {1, 2, ...})))

(14.37)

Where the tree T r = build((, P ), , ) is the infinite state transform tree.

Function search utilizes this tree to transform the state according to match or
fail. Denote the first character in Ps as p, the rest characters as Ps0 , and the
matched positions found so far as A.

(R, A {i})
(R, A)
search((((Pp , Ps ), L, R), A), (c, i)) =
((((Pp , Ps ), L, R), A)

search((L, A), (c, i))

p = c match(R)
p = c match(R)
Pp =
otherwise
(14.38)
The following Haskell example program implements this algorithm.
:
:
:
:

data State a = E | S a (State a) (State a) -- state, ok-state, fail-state

deriving (Eq, Show)
build
build
build
l
r

:: (Eq a)State ([a], [a]) State ([a], [a])

(S s@(xs, []) E E) = S s (build (S (failure s) E E)) E
(S s@(xs, (y:ys)) E E) = S s l r where
= build (S (failure s) E E) -- fail state
= build (S (xs++[y], ys) E E)

matched (S (_, []) _ _) = True

matched _ = False
kmpSearch3 :: (Eq a) [a] [a] [Int]
kmpSearch3 ws txt = snd $ foldl f (auto, []) (zip txt [1..]) where
auto = build (S ([], ws) E E)
f (s@(S (xs, ys) l r), ns) (x, n)
| [x] isPrefixOf ys = if matched r then (r, ns++[n])
else (r, ns)
| xs == [] = (s, ns)
| otherwise = f (l, ns) (x, n)

The bottle-neck is that the state tree building function calls to fallback.
While current definition of isnt effective enough, because it enumerates all
candidates from right to the left every time.
Since the state tree is infinite, we can adopt some common treatment for
infinite structures. One good example is the Fibonacci series. The first two
Fibonacci numbers are defined as 0 and 1; the rest Fibonacci numbers can be
obtained by adding the previous two numbers.
F0 = 0
F1 = 1
Fn = Fn1 + Fn2

(14.39)

14.2. SEQUENCE SEARCH

467

Thus the Fibonacci numbers can be list one by one as the following
F0
F1
F2
F3
...

=0
=1
= F1 + F0
= F2 + F1

(14.40)

We can collect all numbers in both sides, and define F = {0, 1, F1 , F2 , ...},
Thus we have the following equation.
F = {0, 1, F1 + F0 , F2 + F1 , ...}
= {0, 1} {x + y|x {F0 , F1 , F2 , ...}, y {F1 , F2 , F3 , ...}}
= {0, 1} {x + y|x F, y F 0 }

(14.41)

Where F 0 = tail(F ) is all the Fibonacci numbers except for the first one. In
environments support lazy evaluation, like Haskell for instance, this definition
can be expressed like below.
fibs = 0 : 1 : zipWith (+) fibs (tail fibs)

The recursive definition for infinite Fibonacci series indicates an idea which
can be used to get rid of the fallback function . Denote the state transfer tree
as T , we can define the transfer function when matching a character on this tree
as the following.

trans(T, c) =

root : T =
R : T = ((Pp , Ps ), L, R), c = p

trans(L, c) : otherwise

(14.42)

If we match a character against empty node, we transfer to the root of the

tree. Well define the root later soon. Otherwise, we compare if the character c
is as same as the first character p in Ps . If they match, then we transfer to the
right sub tree for this success case; otherwise, we transfer to the left sub tree
for fail case.
With transfer function defined, we can modify the previous tree building
function accordingly. This is quite similar to the previous Fibonacci series definition.
build(T, (Pp , Ps )) = ((Pp , Ps ), T, build(trans(T, p), (Pp {p}, Ps0 )))
The right hand of this equation contains three parts. The first one is the
state that we are matching (Pp , Ps ); If the match fails, Since T itself can handle
any fail case, we use it directly as the left sub tree; otherwise we recursive build
the right sub tree for success case by advancing one character ahead, and calling
transfer function we defined above.
However, there is an edge case which has to be handled specially, that if Ps
is empty, which indicates a successful match. As defined above, there isnt right
sub tree any more. Combining these cases gives the final building function.

build(T, (Pp , Ps )) =

((Pp , Ps ), T, ) : Ps =
((Pp , Ps ), T, build(trans(T, p), (Pp {p}, Ps0 ))) : otherwise
(14.43)

468

CHAPTER 14. SEARCHING

The last brick is to define the root of the infinite state transfer tree, which
initializes the building.
root = build(, (, P ))

(14.44)

And the new KMP matching algorithm is modified with this root.
kmp(P, T ) = snd(f old(trans, (root, []), zip(T, {1, 2, ...})))

(14.45)

The following Haskell example program implements this final version.

kmpSearch ws txt = snd $ foldl tr (root, []) (zip txt [1..]) where
root = build E ([], ws)
build fails (xs, []) = S (xs, []) fails E
build fails s@(xs, (y:ys)) = S s fails succs where
succs = build (fst (tr (fails, []) (y, 0))) (xs++[y], ys)
tr (E, ns) _ = (root, ns)
tr ((S (xs, ys) fails succs), ns) (x, n)
| [x] isPrefixOf ys = if matched succs then (succs, ns++[n]) else (succs, ns)
| otherwise = tr (fails, ns) (x, n)

Figure 14.17 shows the first 4 steps when search anaym in text anal. Since
the first 3 steps all succeed, so the left sub trees of these 3 states are not actually
constructed. They are marked as ?. In the fourth step, the match fails, thus the
right sub tree neednt be built. On the other hand, we must construct the left
sub tree, which is on top of the result of trans(right(right(right(T ))), n), where
function right(T ) returns the right sub tree of T . This can be further expanded
according to the definition of building and state transforming functions till we
get the concrete state ((a, nanym), L, R). The detailed deduce process is left as
exercise to the reader.
(, ananym)
fail
?

match
(a, nanym)
fail
?

match
(an, anym)
fail
?

match
(ana, nym)
fail

(a, nanym)

match
?

Figure 14.17: On demand construct the state transform tree when searching
ananym in text anal.
This algorithm depends on the lazy evaluation critically. All the states to
be transferred are built on demand. So that the building process is amortized

14.2. SEQUENCE SEARCH

469

O(m), and the total performance is amortized O(n + m). Readers can refer to
[1] for detailed proof of it.
Its worth of comparing the final purely functional and the imperative algorithms. In many cases, we have expressive functional realization, however, for
KMP matching algorithm, the imperative approach is much simpler and more
intuitive. This is because we have to mimic the raw array by a infinite state
transfer tree.
Boyer-Moore
Boyer-Moore string matching algorithm is another effective solution invited in
1977 [10]. The idea of Boyer-Moore algorithm comes from the following observation.
The bad character heuristics
When attempt to match the pattern, even if there are several characters from
the left are same, it fails if the last one does not match, as shown in figure 14.18.
Whats more, we wouldnt find a match even if we slide the pattern down by
1, or 2. Actually, the length of the pattern ananym is 5, the last character is
m, however, the corresponding character in the text is h. It does not appear
in the pattern at all. We can directly slide the pattern down by 5.
a

Figure 14.18: Since character h doesnt appear in the pattern, we wouldnt find
a match even if we slide the pattern down less than the length of the pattern.
This leads to the bad-character rule. We can do a pre-processing for the
pattern. If the character set of the text is already known, we can find all
characters which dont appear in the pattern string. During the later scan
process, as long as we find such a bad character, we can immediately slide the
pattern down by its length. The question is what if the unmatched character
does appear in the pattern? While, in order not to miss any potential matches,
we have to slide down the pattern to check again. This is shown as in the figure
14.19
Its quite possible that the unmatched character appears in the pattern more
than one position. Denote the length of the pattern as |P |, the character appears
in positions p1 , p2 , ..., pi . In such case, we take the right most one to avoid
missing any matches.
s = |P | pi

(14.46)

Note that the shifting length is 0 for the last position in the pattern according
to the above equation. Thus we can skip it in realization. Another important

470

CHAPTER 14. SEARCHING

...

(a) The last character in the pattern e

doesnt match p. However, p appears in
the pattern.

...

(b) We have to slide the pattern down by 2

to check again.

Figure 14.19: Slide the pattern if the unmatched character appears in the pattern.

point is that since the shifting length is calculated against the position aligned
with the last character in the pattern string, (we deduce it from |P |), no matter
where the mismatching happens when we scan from right to the left, we slide
down the pattern string by looking up the bad character table with the one in
the text aligned with the last character of the pattern. This is shown in figure
14.20.

(a)

...

(b)

Figure 14.20: Even the mismatching happens in the middle, between char i
and a, we look up the shifting value with character e, which is 6 (calculated
from the first e, the second e is skipped to avoid zero shifting).
There is a good result in practice, that only using the bad-character rule leads
to a simple and fast string matching algorithm, called Boyer-Moore-Horspool
algorithm [11].
1: procedure Boyer-Moore-Horspool(T, P )
2:
for c do
3:
[c] |P |
4:
for i 1 to |P | 1 do
. Skip the last position
5:
[P [i]] |P | i
6:
s0
7:
while s + |P | |T | do

14.2. SEQUENCE SEARCH

8:
9:
10:
11:
12:
13:
14:
15:

i |P |
while i 1 P [i] = T [s + i] do
ii1
if i < 1 then
found one solution at s
ss+1
else
s s + [T [s + |P |]]

471

. scan from right

. go on finding the next

The character set is denoted as , we first initialize all the values of sliding
table as the length of the pattern string |P |. After that we process the pattern
from left to right, update the sliding value. If a character appears multiple times
in the pattern, the latter value, which is on the right hand, will overwrite the
previous value. We start the matching scan process by aligning the pattern
and the text string from the very left. However, for every alignment s, we scan
from the right to the left until either there is unmatched character or all the
characters in the pattern have been examined. The latter case indicates that
weve found a match; while for the former case, we look up to slide the pattern
down to the right.
The following example Python code implements this algorithm accordingly.
def bmh_match(w, p):
n = len(w)
m = len(p)
tab = [m for _ in range(256)] # table to hold the bad character rule.
for i in range(m-1):
tab[ord(p[i])] = m - 1 - i
res = []
offset = 0
while offset + m n:
i=m - 1
while i 0 and p[i] == w[offset+i]:
i=i - 1
if i < 0:
res.append(offset)
offset = offset + 1
else:
offset = offset + tab[ord(w[offset + m - 1])]
return res

The algorithm firstly takes about O(||+|P |) time to build the sliding table.
If the character set size is small, the performance is dominated by the pattern
and the text. There is definitely the worst case that all the characters in the
pattern and text are same, e.g. searching aa...a (m of a, denoted as am )
in text aa......a (n of a, denoted as an ). The performance in the worst case
is O(mn). This algorithm performs well if the pattern is long, and there are
constant number of matching. The result is bound to linear time. This is as
same as the best case of full Boyer-Moore algorithm which will be explained
next.

472

CHAPTER 14. SEARCHING

The good suffix heuristics

Consider searching pattern abbabab in text bbbababbabab... like figure 14.21.
By using the bad-character rule, the pattern will be slided by two.
b

...

T
b

X
a

(a)

(b)

Figure 14.21: According to the bad-character rule, the pattern is slided by 2, so

that the next b is aligned.
Actually, we can do better than this. Observing that before the unmatched
point, we have already successfully matched 6 characters bbabab from right to
the left. Since ab, which is the prefix of the pattern is also the suffix of what
we matched so far, we can directly slide the pattern to align this suffix as shown
in figure 14.22.
b

...

Figure 14.22: As the prefix ab is also the suffix of what weve matched, we can
slide down the pattern to a position so that ab are aligned.
This is quite similar to the pre-processing of KMP algorithm, However, we
cant always skip so many characters. Consider the following example as shown
in figure 14.23. We have matched characters bab when the unmatch happens.
Although the prefix ab of the pattern is also the suffix of bab, we cant
slide the pattern so far. This is because bab appears somewhere else, which
starts from the 3rd character of the pattern. In order not to miss any potential
matching, we can only slide the pattern by two.
b

...

T
b

...

X
a

(a)

(b)

Figure 14.23: Weve matched bab, which appears somewhere else in the pattern
(from the 3rd to the 5th character). We can only slide down the pattern by 2
to avoid missing any potential matching.

...

14.2. SEQUENCE SEARCH

473

The above situation forms the two cases of the good-suffix rule, as shown in
figure 14.24.

(a) Case 1, Only a part of the matching suffix (b) Case 2, The matching suffix occurs some
occurs as a prefix of the pattern.
where else in the pattern.

Figure 14.24: The light gray section in the text represents the characters have
been matched; The dark gray parts indicate the same content in the pattern.
Both cases in good suffix rule handle the situation that there are multiple
characters have been matched from right. We can slide the pattern to the right
if any of the the following happens.
Case 1 states that if a part of the matching suffix occurs as a prefix of the
pattern, and the matching suffix doesnt appear in any other places in the
pattern, we can slide the pattern to the right to make this prefix aligned;
Case 2 states that if the matching suffix occurs some where else in the pattern, we can slide the pattern to make the right most occurrence aligned.
Note that in the scan process, we should apply case 2 first whenever it is
possible, and then examine case 1 if the whole matched suffix does not appears
in the pattern. Observe that both cases of the good-suffix rule only depend on
the pattern string, a table can be built by pre-process the pattern for further
looking up.
For the sake of brevity, we denote the suffix string from the i-th character
of P as Pi . That Pi is the sub-string P [i]P [i + 1]...P [m].
For case 1, we can check every suffix of P , which includes Pm , Pm1 , Pm2 ,
..., P2 to examine if it is the prefix of P . This can be achieved by a round of
scan from right to the left.
For case 2, we can check every prefix of P includes P1 , P2 , ..., Pm1 to
examine if the longest suffix is also a suffix of P . This can be achieved by
another round of scan from left to the right.
1: function Good-Suffix(P )
2:
m |P |
3:
s {0, 0, ..., 0}
. Initialize the table of length m
4:
l0
. The last suffix which is also prefix of P
5:
for i m 1 down-to 1 do
. First loop for case 1
6:
if Pi @ P then
. @ means is prefix of
7:
li

474
8:
9:
10:
11:
12:
13:

CHAPTER 14. SEARCHING

s [i] l
for i 1 to m do
s Suffix-Length(Pi )
if s 6= 0 P [i s] 6= P [m s] then
s [m s] m i
return s

. Second loop for case 2

This algorithm builds the good-suffix heuristics table s . It first checks every
suffix of P from the shortest to the longest. If the suffix Pi is also the prefix of
P , we record this suffix, and use it for all the entries until we find another suffix
Pj , j < i, and it is also the prefix of P .
After that, the algorithm checks every prefix of P from the shortest to the
longest. It calls the function Suffix-Length(Pi ), to calculate the length of the
longest suffix of Pi , which is also suffix of P . If this length s isnt zero, which
means there exists a sub-string of s, that appears as the suffix of the pattern.
It indicates that case 2 happens. The algorithm overwrites the s-th entry from
the right of the table s . Note that to avoid finding the same occurrence of the
matched suffix, we test if P [i s] and P [m s] are same.
Function Suffix-Length is designed as the following.
1: function Suffix-Length(Pi )
2:
m |P |
3:
j0
4:
while P [m j] = P [i j] j < i do
5:
j j+1
6:
return j
The following Python example program implements the good-suffix rule.
def good_suffix(p):
m = len(p)
tab = [0 for _ in range(m)]
last = 0
# first loop for case 1
for i in range(m-1, 0, -1): # m-1, m-2, .., 1
if is_prefix(p, i):
last = i
tab[i - 1] = last
# second loop for case 2
for i in range(m):
slen = suffix_len(p, i)
if slen != 0 and p[i - slen] != p[m - 1 - slen]:
tab[m - 1 - slen] = m - 1 - i
return tab
# test if p[i..m-1] is prefix of p
def is_prefix(p, i):
for j in range(len(p) - i):
if p[j] != p [i+j]:
return False
return True
# length of the longest suffix of p[..i], which is also a suffix of p
def suffix_len(p, i):

14.2. SEQUENCE SEARCH

475

m = len(p)
j=0
while p[m - 1 - j] == p[i - j] and j < i:
j=j+1
return j

Its quite possible that both the bad-character rule and the good-suffix rule
can be applied when the unmatch happens. The Boyer-Moore algorithm compares and picks the bigger shift so that it can find the solution as quick as
possible. The bad-character rule table can be explicitly built as below
1: function Bad-Character(P )
2:
for c do
3:
b [c] |P |
4:
for i 1 to |P | 1 do
5:
b [P [i]] |P | i
6:
return b
The following Python program implements the bad-character rule accordingly.
def bad_char(p):
m = len(p)
tab = [m for _ in range(256)]
for i in range(m-1):
tab[ord(p[i])] = m - 1 - i
return tab

The final Boyer-Moore algorithm firstly builds the two rules from the pattern,
then aligns the pattern to the beginning of the text and scans from right to the
left for every alignment. If any unmatch happens, it tries both rules, and slides
the pattern with the bigger shift.
1: function Boyer-Moore(T, P )
2:
n |T |, m |P |
3:
b Bad-Character(P )
4:
s Good-Suffix(P )
5:
s0
6:
while s + m n do
7:
im
8:
while i 1 P [i] = T [s + i] do
9:
ii1
10:
if i < 1 then
11:
found one solution at s
12:
ss+1
. go on finding the next
13:
else
14:
s s + max(b [T [s + m]], s [i])
Here is the example implementation of Boyer-Moore algorithm in Python.
def bm_match(w, p):
n = len(w)
m = len(p)
tab1 = bad_char(p)
tab2 = good_suffix(p)
res = []

476

CHAPTER 14. SEARCHING

offset = 0
while offset + m n:
i=m - 1
while i 0 and p[i] == w[offset + i]:
i=i - 1
if i < 0:
res.append(offset)
offset = offset + 1
else:
offset = offset + max(tab1[ord(w[offset + m - 1])], tab2[i])
return res

The Boyer-Moore algorithm published in original paper is bound to O(n+m)

in worst case only if the pattern doesnt appear in the text [10]. Knuth, Morris,
and Pratt proved this fact in 1977 [12]. However, when the pattern appears in
the text, as we shown above, Boyer-Moore performs O(nm) in the worst case.
Richard birds shows a purely functional realization of Boyer-Moore algorithm
in chapter 16 in [1]. We skipped it in this book.

Exercise 14.2
Proof that Boyer-Moore majority vote algorithm is correct.
Given a list, find the element occurs most. Are there any divide and
conqueror solutions? Are there any divide and conqueror data structures,
such as map can be used?
Bentley presents a divide and conquer algorithm to find the maximum sum
in O(n log n) time in [4]. The idea is to split the list at the middle point.
We can recursively find the maximum sum in the first half and second half;
However, we also need to find maximum sum cross the middle point. The
method is to scan from the middle point to both ends as the following.
1: function Max-Sum(A)
2:
if A = then
3:
return 0
4:
else if |A| = 1 then
5:
return Max(0, A[1])
6:
else
7:
m b |A|
2 c
8:
a Max-From(Reverse(A[1...m]))
9:
b Max-From(A[m + 1...|A|])
10:
c Max-Sum(A[1...m])
11:
d Max-Sum(A[m + 1...|A|)
12:
return Max(a + b, c, d)
13:
14:
15:
16:
17:
18:

function Max-From(A)
sum 0, m 0
for i 1 to |A| do
sum sum + A[i]
m Max(m, sum)
return m

14.3. SOLUTION SEARCHING

477

Its easy to deduce the time performance is T (n) = 2T (n/2) + O(n).

Implement this algorithm in your favorite programming language.
Explain why KMP algorithm perform in linear time even in the seemed
worst case.
Implement the purely functional KMP algorithm by using reversed Pp to
avoid the linear time appending operation.
Deduce the state of the tree lef t(right(right(right(T )))) when searching
ananym in text anal.

14.3

Solution searching

One interesting thing that computer programming can offer is solving puzzles.
In the early phase of classic artificial intelligent, people developed many methods
to search for solutions. Different from the sequence searching and string matching, the solution doesnt obviously exist among a candidates set. It typically
need construct the solution while trying varies of attempts. Some problems are
solvable, while others are not. Among the solvable problems, not all of them
just have one unique solution. For example, a maze may have multiple ways
out. People sometimes need search for the best one.

14.3.1

DFS and BFS

DFS and BFS stand for deep-first search and breadth-first search. They are
typically introduced as graph algorithms in textbooks. Graph is a comprehensive topic which is hard to be covered in this elementary book. In this section,
well show how to use DFS and BFS to solve some real puzzles without formal
introduction about the graph concept.
Maze
Maze is a classic and popular puzzle. Maze is amazing to both kids and adults.
Figure 14.25 shows an example maze. There are also real maze gardens can be
found in parks for fun. In the late 1990s, maze-solving games were quite often
hold in robot mouse competition all over the world.

Figure 14.25: A maze

478

CHAPTER 14. SEARCHING

There are multiple methods to solve maze puzzle. Well introduce an effective, yet not the best one in this section. There are some well known sayings
about how to find the way out in maze, while not all of them are true.
For example, one method states that, wherever you have multiple ways,
always turn right. This doesnt work as shown in figure 14.26. The obvious
solution is first to go along the top horizontal line, then turn right, and keep
going ahead at the T section. However, if we always turn right, well endless
loop around the inner big block.

Figure 14.26: It leads to loop way if always turns right.

This example tells us that the decision when there are multiple choices matters the solution. Like the fairy tale we read in our childhood, we can take some
bread crumbs in a maze. When there are multiple ways, we can simply select
one, left a piece of bread crumbs to mark this attempt. If we enter a died end,
we go back to the last place where weve made a decision by back-tracking the
bread crumbs. Then we can alter to another way.
At any time, if we find there have been already bread crumbs left, it means
we have entered a loop, we must go back and try different ways. Repeat these
try-and-check steps, we can either find the way out, or give the no solution
fact. In the later case, we back-track to the start point.
One easy way to describe a maze, is by a m n matrix, each element is
either 0 or 1, which indicates if there is a way at this cell. The maze illustrated
in figure 14.26 can be defined as the following matrix.
0
0
0
0
0
0
1

0
1
1
1
1
0
1

0
0
0
0
0
0
0

Given a start point s = (i, j), and a goal e = (p, q), we need find all solutions,
that are the paths from s to e.
There is an obviously recursive exhaustive search method. That in order
to find all paths from s to e, we can check all connected points to s, for every
such point k, we recursively find all paths from k to e. This method can be
illustrated as the following.
Trivial case, if the start point s is as same as the target point e, we are
done;

14.3. SOLUTION SEARCHING

479

Otherwise, for every connected point k to s, recursively find the paths

from k to e; If e can be reached via k, put section s-k in front of each path
between k and e.
However, we have to left bread crumbs to avoid repeatedly trying the same
attempts. This is because otherwise in the recursive case, we start from s, find
a connected point k, then we further try to find paths from k to e. Since s is
connected to k as well, so in the next recursion, well try to find paths from s
to e again. It turns to be the very same origin problem, and we are trapped in
infinite recursions.
Our solution is to initialize an empty list, use it to record all the points weve
visited so far. For every connected point, we look up the list to examine if it
has already been visited. We skip all the visited candidates and only try those
new ones. The corresponding algorithm can be defined like this.
solveM aze(m, s, e) = solve(s, {})

(14.47)

Where m is the matrix which defines a maze, s is the start point, and e is
the end point. Function solve is defined in the context of solveM aze, so that
the maze and the end point can be accessed. It can be realized recursively like
what we described above7 .

solve(s, P ) =

concat({

{{s} p|p P }
solve(s0 , {{s} p|p P })|
s0 adj(s), visited(s0 )})

: s=e
: otherwise

(14.48)
Note that P also serves as an accumulator. Every connected point is recorded
in all the possible paths to the current position. But they are stored in reversed
order, that is the newly visited point is put to the head of all the lists, and
the starting point is the last one. This is because the appending operation is
linear (O(n), where n is the number of elements stored in a list), while linking
to the head is just constant time. We can output the result in correct order by
reversing all possible solutions in equation (14.47)8 :
solveM aze(m, s, e) = map(reverse, solve(s, {}))

(14.49)

We need define functions adj(p) and visited(p), which finds all the connected
points to p, and tests if point p has been visited respectively. Two points are
connected if and only if they are next cells horizontally or vertically in the maze
matrix, and both have zero value.
adj((x, y)) = {(x0 , y 0 )| (x0 , y 0 ) {(x 1, y), (x + 1, y), (x, y 1), (x, y + 1)},
1 x0 M, 1 y 0 N, mx0 y0 = 0}
(14.50)
Where M and N are the widths and heights of the maze.
Function visited(p) examines if point p has been recorded in any lists in P .
visited(p, P ) = path P, p path

(14.51)

concat can flatten a list of lists. For example. concat({{a, b, c}, {x, y, z}}) =
{a, b, c, x, y, z}. Refer to appendix A for detail.
8 the detailed definition of reverse can be found in the appendix A.
7 Function

480

CHAPTER 14. SEARCHING

The following Haskell example code implements this algorithm.

solveMaze m from to = map reverse $ solve from [[]] where

solve p paths | p == to = map (p:) paths
| otherwise = concat [solve p (map (p:) paths) |
p adjacent p,
not $ visited p paths]
adjacent (x, y) = [(x, y) |
(x, y) [(x-1, y), (x+1, y), (x, y-1), (x, y+1)],
inRange (bounds m) (x, y),
m ! (x, y) == 0]
visited p paths = any (p elem) paths

For a maze defined as matrix like below example, all the solutions can be
given by this program.
mz = [[0,
[1,
[1,
[1,
[0,
[0,

0,
0,
0,
1,
0,
0,

1,
1,
0,
0,
0,
0,

0,
0,
0,
1,
0,
1,

1,
1,
0,
1,
0,
1,

1],
1],
0],
1],
0],
0]]

maze = listArray ((1,1), (6, 6)) concat

solveMaze (maze mz) (1,1) (6, 6)

As we mentioned, this is a style of exhaustive search. It recursively searches

all the connected points as candidates. In a real maze solving game, a robot
mouse competition for instance, its enough to just find a route. We can adapt
to a method close to what described at the beginning of this section. The robot
mouse always tries the first connected point, and skip the others until it gets
stuck. We need some data structure to store the bread crumbs, which help to
remember the decisions being made. As we always attempt to find the way on
top of the latest decision, it is the last-in, first-out manner. A stack can be used
to realize it.
At the very beginning, only the starting point s is stored in the stack. we
pop it out, find, for example, points a, and b, are connected to s. We push the
two possible paths: {a, s} and {b, s} to the stack. Next we pop {a, s} out, and
examine connected points to a. Then all the paths with 3 steps will be pushed
back. We repeat this process. At anytime, each element stored in the stack is
a path, from the starting point to the farthest place can arrive in the reversed
order. This can be illustrated in figure 14.27.
The stack can be realized with a list. The latest option is picked from the
head, and the new candidates are also added to the head. The maze puzzle can
be solved by using such a list of paths:
solveM aze0 (m, s, e) = reverse(solve0 ({{s}}))

(14.52)

As we are searching the first, but not all the solutions, map isnt used here.
When the stack is empty, it means that weve tried all the options and failed
to find a way out. There is no solution; otherwise, the top option is popped,
expanded with all the adjacent points which havent been visited before, and
pushed back to the stack. Denote the stack as S, if it isnt empty, the top

14.3. SOLUTION SEARCHING

481
i

[i, p, ... , s]
[p, ... , s]

[j, p, ..., s]

[q, ..., s]

[k, p, ..., s]

...

[q, ..., s]

[a, s]
...

[s]
[b, s]

...

Figure 14.27: The stack is initialized with a singleton list of the starting point
s. s is connected with point a and b. Paths {a, s} and {b, s} are pushed back.
In some step, the path ended with point p is popped. p is connected with points
i, j, and k. These 3 points are expanded as different options and pushed back
to the stack. The candidate path ended with q wont be examined unless all the
options above fail.
element is s1 , and the new stack after the top being popped as S 0 . s1 is a list
of points represents path P . Denote the first point in this path as p1 , and the
rest as P 0 . The solution can be formalized as the following.

:
s
:
1
solve0 (S) =
0
0
solve
(S
)
:

solve0 ({{p} P |p C} S) :

S=
s1 = e
C = {c|c adj(p1 ), c 6 P 0 } =
otherwise, C 6=
(14.53)
Where the adj function is defined above. This updated maze solution can
be implemented with the below example Haskell program 9 .

dfsSolve m from to = reverse $ solve [[from]] where

solve [] = []
solve (c@(p:path):cs)
| p == to = c -- stop at the first solution
| otherwise = let os = filter (notElem path) (adjacent p) in
if os == []
then solve cs
else solve ((map (:c) os) ++ cs)

Its quite easy to modify this algorithm to find all solutions. When we find
a path in the second clause, instead of returning it immediately, we record it
and go on checking the rest memorized options in the stack till until the stack
becomes empty. We left it as an exercise to the reader.
The same idea can also be realized imperatively. We maintain a stack to store
all possible paths from the starting point. In each iteration, the top option path
is popped, if the farthest position is the end point, a solution is found; otherwise,
all the adjacent, not visited yet points are appended as new paths and pushed
9 The

same code of adjacent function is skipped

482

CHAPTER 14. SEARCHING

back to the stack. This is repeated till all the candidate paths in the stacks are
checked.
We use the same notation to represent the stack S. But the paths will
be stored as arrays instead of list in imperative settings as the former is more
effective. Because of this the starting point is the first element in the path
array, while the farthest reached place is the right most element. We use pn
to represent Last(P ) for path P . The imperative algorithm can be given as
below.
1: function Solve-Maze(m, s, e)
2:
S
3:
Push(S, {s})
4:
L
. the result list
5:
while S 6= do
6:
P Pop(S)
7:
if e = pn then
8:
Add(L, P )
9:
else
10:
for p Adjacent(m, pn ) do
11:
if p
/ P then
12:
Push(S, P {p})
13:
return L
The following example Python program implements this maze solving algorithm.
def solve(m, src, dst):
stack = [[src]]
s = []
while stack != []:
path = stack.pop()
if path[-1] == dst:
s.append(path)
else:
for p in adjacent(m, path[-1]):
if not p in path:
stack.append(path + [p])
return s
def adjacent(m, p):
(x, y) = p
ds = [(0, 1), (0, -1), (1, 0), (-1, 0)]
ps = []
for (dx, dy) in ds:
x1 = x + dx
y1 = y + dy
if 0 x1 and x1 < len(m[0]) and
0 y1 and y1 < len(m) and m[y][x] == 0:
ps.append((x1, y1))
return ps

And the same maze example given above can be solved by this program like
the following.
mz = [[0, 0, 1, 0, 1, 1],

14.3. SOLUTION SEARCHING

[1,
[1,
[1,
[0,
[0,

0,
0,
1,
0,
0,

1,
0,
0,
0,
0,

0,
0,
1,
0,
1,

1,
0,
1,
0,
1,

483

1],
0],
1],
0],
0]]

solve(mz, (0, 0), (5,5))

It seems that in the worst case, there are 4 options (up, down, left, and right)
at each step, each option is pushed to the stack and eventually examined during
backtracking. Thus the complexity is bound to O(4n ). The actual time wont
be so large because we filtered out the places which have been visited before. In
the worst case, all the reachable points are visited exactly once. So the time is
bound to O(n), where n is the number of points connected in total. As a stack
is used to store candidate solutions, the space complexity is O(n2 ).
Eight queens puzzle
The eight queens puzzle is also a famous problem. Although cheese has very
long history, this puzzle was first published in 1848 by Max Bezzel[13]. Queen
in the cheese game is quite powerful. It can attack any other pieces in the same
row, column and diagonal at any distance. The puzzle is to find a solution to
put 8 queens in the board, so that none of them attack each other. Figure
14.28 (a) illustrates the places can be attacked by a queen and 14.28 (b) shows
a solution of 8 queens puzzle.

(a) A queen piece.

(b) An example solution

Figure 14.28: The eight queens puzzle.

8
Its obviously that the puzzle can be solved by brute-force, which takes P64
10
times. This number is about 4 10 . It can be easily improved by observing
that, no two queens can be in the same row, and each queen must be put on one
column between 1 to 8. Thus we can represent the arrangement as a permutation
of {1, 2, 3, 4, 5, 6, 7, 8}. For instance, the arrangement {6, 2, 7, 1, 3, 5, 8, 4} means,
we put the first queen at row 1, column 6, the second queen at row 2 column 2,
..., and the last queen at row 8, column 4. By this means, we need only examine
8! = 40320 possibilities.
We can find better solutions than this. Similar to the maze puzzle, we put
queens one by one from the first row. For the first queen, there are 8 options,
that we can put it at one of the eight columns. Then for the next queen, we

484

CHAPTER 14. SEARCHING

again examine the 8 candidate columns. Some of them are not valid because
those positions will be attacked by the first queen. We repeat this process, for
the i-th queen, we examine the 8 columns in row i, find which columns are safe.
If none column is valid, it means all the columns in this row will be attacked
by some queen weve previously arranged, we have to backtrack as what we did
in the maze puzzle. When all the 8 queens are successfully put to the board,
we find a solution. In order to find all the possible solutions, we need record
it and go on to examine other candidate columns and perform back tracking if
necessary. This process terminates when all the columns in the first row have
been examined. The below equation starts the search.
solve({}, )

(14.54)

In order to manage the candidate attempts, a stack S is used as same as

in the maze puzzle. The stack is initialized with one empty element. And a
list L is used to record all possible solutions. Denote the top element in the
stack as s1 . Its actually an intermediate state of assignment, which is a partial
permutation of 1 to 8. after pops s1 , the stack becomes S 0 . The solve function
can be defined as the following.

L : S=
0
solve(S
,
{s
}

L)
: |s1 | = 8

1
solve(S, L) =
{i} s1 | i [1, 8],

i
/ s1 ,
solve(
S 0 , L) : otherwise

saf e(i, s1 )
(14.55)
If the stack is empty, all the possible candidates have been examined, its not
possible to backtrack any more. L has been accumulated all found solutions and
returned as the result; Otherwise, if the length of the top element in the stack
is 8, a valid solution is found. We add it to L, and go on finding other solutions;
If the length is less than 8, we need try to put the next queen. Among all the
columns from 1 to 8, we pick those not already occupied by previous queens
(through the i
/ s1 clause), and must not be attacked in diagonal direction
(through the saf e predication). The valid assignments will be pushed to the
stack for the further searching.
Function saf e(x, C) detects if the assignment of a queen in position x will
be attacked by other queens in C in diagonal direction. There are 2 possible
cases, 45 and 135 directions. Since the row of this new queen is y = 1 + |C|,
where |C| is the length of C, the saf e function can be defined as the following.
saf e(x, C) = (c, r) zip(reverse(C), {1, 2, ...}), |x c| 6= |y r|

(14.56)

Where zip takes two lists, and pairs every elements in them to a new
list. Thus If C = {ci1 , ci2 , ..., c2 , c1 } represents the column of the first
i 1 queens has been assigned, the above function will check list of pairs
{(c1 , 1), (c2 , 2), ..., (ci1 , i 1)} with position (x, y) forms any diagonal lines.
Translating this algorithm into Haskell gives the below example program.
solve = dfsSolve [[]] [] where
dfsSolve [] s = s

14.3. SOLUTION SEARCHING

485

dfsSolve (c:cs) s
| length c == 8 = dfsSolve cs (c:s)
| otherwise = dfsSolve ([(x:c) | x [1..8] \\ c,
not $ attack x c] ++ cs) s
attack x cs = let y = 1 + length cs in
any ((c, r) abs(x - c) == abs(y - r)) $
zip (reverse cs) [1..]

Observing that the algorithm is tail recursive, its easy to transform it into
imperative realization. Instead of using list, we use array to represent queens
assignment. Denote the stack as S, and the possible solutions as A. The
imperative algorithm can be described as the following.
1: function Solve-Queens
2:
S {}
3:
L
. The result list
4:
while S 6= do
5:
A Pop(S)
. A is an intermediate assignment
6:
if |A| = 8 then
7:
Add(L, A)
8:
else
9:
for i 1 to 8 do
10:
if Valid(i, A) then
11:
Push(S, A {i})
12:
return L
The stack is initialized with the empty assignment. The main process repeatedly pops the top candidate from the stack. If there are still queens left, the
algorithm examines possible columns in the next row from 1 to 8. If a column
is safe, that it wont be attacked by any previous queens, this column will be
appended to the assignment, and pushed back to the stack. Different from the
functional approach, since array, but not list, is used, we neednt reverse the
solution assignment any more.
Function Valid checks if column x is safe with previous queens put in A.
It filters out the columns have already been occupied, and calculates if any
diagonal lines are formed with existing queens.
1: function Valid(x, A)
2:
y 1 + |A|
3:
for i 1 to |A| do
4:
if x = i |y i| = |x A[i]| then
5:
return False
6:
return True
The following Python example program implements this imperative algorithm.
def solve():
stack = [[]]
s = []
while stack != []:
a = stack.pop()
if len(a) == 8:
s.append(a)
else:

486

CHAPTER 14. SEARCHING

for i in range(1, 9):
if valid(i, a):
stack.append(a+[i])
return s

def valid(x, a):

y = len(a) + 1
for i in range(1, y):
if x == a[i-1] or abs(y - i) == abs(x - a[i-1]):
return False
return True

Although there are 8 optional columns for each queen, not all of them are
valid and thus further expanded. Only those columns havent been occupied by
previous queens are tried. The algorithm only examines 15720, which is far less
than 88 = 16777216, possibilities [13].
Its quite easy to extend the algorithm, so that it can solve N queens puzzle,
where N 4. However, the time cost increases fast. The backtrack algorithm
is just slightly better than the one permuting the sequence of 1 to 8 (which is
bound to o(N !)). Another extension to the algorithm is based on the fact that
the chess board is square, which is symmetric both vertically and horizontally.
Thus a solution can generate other solutions by rotating and flipping. These
aspects are left as exercises to the reader.
Peg puzzle
I once received a puzzle of the leap frogs. It said to be homework for 2nd grade
student in China. As illustrated in figure 14.29, there are 6 frogs in 7 stones.
Each frog can either hop to the next stone if it is not occupied, or leap over one
frog to another empty stone. The frogs on the left side can only move to the
right, while the ones on the right side can only move to the left. These rules are
described in figure 14.30

Figure 14.29: The leap frogs puzzle.

The goal of this puzzle is to arrange the frogs to jump according to the
rules, so that the positions of the 3 frogs on the left are finally exchange with
the ones on the right. If we denote the frog on the left as A, on the right as
B, and the empty stone as O. The puzzle is to find a solution to transform
from AAAOBBB to BBBOAAA.
This puzzle is just a special form of the peg puzzles. The number of pegs is
not limited to 6. it can be 8 or other bigger even numbers. Figure 14.31 shows
some variants.

14.3. SOLUTION SEARCHING

(a) Jump to the next

stone

487

(b) Jump over to the

right

(c) Jump over to the left

Figure 14.30: Moving rules.

(a) Solitaire

(b) Hop over

Figure
14.31:
Variants
of
the
http://home.comcast.net/ stegmann/jumping.htm

(c) Draught board

peg

puzzles

from

We can solve this puzzle by programing. The idea is similar to the 8 queens
puzzle. Denote the positions from the left most stone as 1, 2, ..., 7. In ideal
cases, there are 4 options to arrange the move. For example when start, the frog
on 3rd stone can hop right to the empty stone; symmetrically, the frog on the
5th stone can hop left; Alternatively, the frog on the 2nd stone can leap right,
while the frog on the 6th stone can leap left.
We can record the state and try one of these 4 options at every step. Of
course not all of them are possible at any time. If get stuck, we can backtrack
and try other options.
As we restrict the left side frogs only moving to the right, and the right frogs
only moving to the left. The moves are not reversible. There wont be any
repetition cases as what we have to deal with in the maze puzzle. However, we
still need record the steps in order to print them out finally.
In order to enforce these restriction, let A, O, B in representation AAAOBBB
be -1, 0, and 1 respectively. A state L is a list of elements, each element is one
of these 3 values. It starts from {1, 1, 1, 0, 1, 1, 1}. L[i] access the i-th element, its value indicates if the i-th stone is empty, occupied by a frog from left
side, or occupied by a frog from right side. Denote the position of the vacant
stone as p. The 4 moving options can be stated as below.
Leap left: p < 6 and L[p + 2] > 0, swap L[p] L[p + 2];
Hop left: p < 7 and L[p + 1] > 0, swap L[p] L[p + 1];
Leap right: p > 2 and L[p 2] < 0, swap L[p 2] L[p];
Hop right: p > 1 and L[p 1] < 0, swap L[p 1] L[p].
Four functions leapl (L), hopl (L), leapr (L) and hopr (L) are defined accordingly. If the state L does not satisfy the move restriction, these function return
L unchanged, otherwise, the changed state L0 is returned accordingly.

488

CHAPTER 14. SEARCHING

We can also explicitly maintain a stack S to the attempts as well as the

historic movements. The stack is initialized with a singleton list of starting
state. The solution is accumulated to a list M , which is empty at the beginning:
solve({{1, 1, 1, 0, 1, 1, 1}}, )

(14.57)

As far as the stack isnt empty, we pop one intermediate attempt. If the
latest state is equal to {1, 1, 1, 0, 1, 1, 1}, a solution is found. We append
the series of moves till this state to the result list M ; otherwise, We expand to
next possible state by trying all four possible moves, and push them back to the
stack for further search. Denote the top element in the stack S as s1 , and the
latest state in s1 as L. The algorithm can be defined as the following.

solve(S, M ) =

M : S=
solve(S 0 , {reverse(s1 )} M ) : L = {1, 1, 1, 0, 1, 1, 1}

solve(P S 0 , M ) : otherwise
(14.58)

Where P are possible moves from the latest state L:

P = {L0 |L0 {leapl (L), hopl (L), leapr (L), hopr (L)}, L 6= L0 }
Note that the starting state is stored as the last element, while the final state
is the first. That is the reason why we reverse it when adding to solution list.
Translating this algorithm to Haskell gives the following example program.
solve = dfsSolve [[[-1, -1, -1, 0, 1, 1, 1]]] [] where
dfsSolve [] s = s
dfsSolve (c:cs) s
| head c == [1, 1, 1, 0, -1, -1, -1] = dfsSolve cs (reverse c:s)
| otherwise = dfsSolve ((map (:c) $ moves $ head c) ++ cs) s
moves s = filter (/=s) [leapLeft s, hopLeft s, leapRight s, hopRight s] where
leapLeft [] = []
leapLeft (0:y:1:ys) = 1:y:0:ys
leapLeft (y:ys) = y:leapLeft ys
hopLeft [] = []
hopLeft (0:1:ys) = 1:0:ys
hopLeft (y:ys) = y:hopLeft ys
leapRight [] = []
leapRight (-1:y:0:ys) = 0:y:(-1):ys
leapRight (y:ys) = y:leapRight ys
hopRight [] = []
hopRight (-1:0:ys) = 0:(-1):ys
hopRight (y:ys) = y:hopRight ys

Running this program finds 2 symmetric solutions, each takes 15 steps. One
solution is list in the below table.

14.3. SOLUTION SEARCHING

489

step -1 -1 -1 0
1
1
1
1
-1 -1 0 -1 1
1
1
2
-1 -1 1 -1 0
1
1
3
-1 -1 1 -1 1
0
1
4
-1 -1 1
0
1 -1 1
5
-1 0
1 -1 1 -1 1
6
0 -1 1 -1 1 -1 1
7
1 -1 0 -1 1 -1 1
8
1 -1 1 -1 0 -1 1
9
1 -1 1 -1 1 -1 0
10
1 -1 1 -1 1
0 -1
11
1 -1 1
0
1 -1 -1
12
1
0
1 -1 1 -1 -1
13
1
1
0 -1 1 -1 -1
14
1
1
1 -1 0 -1 -1
15
1
1
1
0 -1 -1 -1
Observe that the algorithm is in tail recursive manner, it can also be realized
imperatively. The algorithm can be more generalized, so that it solve the puzzles
of n frogs on each side. We represent the start state {-1, -1, ..., -1, 0, 1, 1, ...,
1} as s, and the mirrored end state as e.
1: function Solve(s, e)
2:
S {{s}}
3:
M
4:
while S 6= do
5:
s1 Pop(S)
6:
if s1 [1] = e then
7:
Add(M , Reverse(s1 ))
8:
else
9:
for m Moves(s1 [1]) do
10:
Push(S, {m} s1 )
11:
return M
The possible moves can be also generalized with procedure Moves to handle arbitrary number of frogs. The following Python program implements this
solution.
def solve(start, end):
stack = [[start]]
s = []
while stack != []:
c = stack.pop()
if c[0] == end:
s.append(reversed(c))
else:
for m in moves(c[0]):
stack.append([m]+c)
return s
def moves(s):
ms = []
n = len(s)
p = s.index(0)

490

CHAPTER 14. SEARCHING

if p < n - 2 and s[p+2] > 0:
ms.append(swap(s, p, p+2))
if p < n - 1 and s[p+1] > 0:
ms.append(swap(s, p, p+1))
if p > 1 and s[p-2] < 0:
ms.append(swap(s, p, p-2))
if p > 0 and s[p-1] < 0:
ms.append(swap(s, p, p-1))
return ms

def swap(s, i, j):

a = s[:]
(a[i], a[j]) = (a[j], a[i])
return a

For 3 frogs in each side, we know that it takes 15 steps to exchange them.
Its interesting to examine the table that how many steps are needed along with
the number of frogs in each side. Our program gives the following result.
4
5 ...
number of frogs 1 2 3
number of steps 3 8 15 24 35 ...
It seems that the number of steps are all square numbers minus one. Its
natural to guess that the number of steps for n frogs in one side is (n + 1)2 1.
Actually we can prove it is true.
Compare to the final state and the start state, each frog moves ahead n + 1
stones in its opposite direction. Thus total 2n frogs move 2n(n + 1) stones.
Another important fact is that each frog on the left has to meet every one on
the right one time. And leap will happen when meets. Since the frog moves
two stones ahead by leap, and there are total n2 meets happen, so that all these
meets cause moving 2n2 stones ahead. The rest moves are not leap, but hop.
The number of hops are 2n(n + 1) 2n2 = 2n. Sum up all n2 leaps and 2n
hops, the total number of steps are n2 + 2n = (n + 1)2 1.
Summary of DFS
Observe the above three puzzles, although they vary in many aspects, their
solutions show quite similar common structures. They all have some starting
state. The maze starts from the entrance point; The 8 queens puzzle starts
from the empty board; The leap frogs start from the state of AAAOBBB. The
solution is a kind of searching, at each attempt, there are several possible ways.
For the maze puzzle, there are four different directions to try; For the 8 queens
puzzle, there are eight columns to choose; For the leap frogs puzzle, there are
four movements of leap or hop. We dont know how far we can go when make a
decision, although the final state is clear. For the maze, its the exit point; For
the 8 queens puzzle, we are done when all the 8 queens being assigned on the
board; For the leap frogs puzzle, the final state is that all frogs exchanged.
We use a common approach to solve them. We repeatedly select one possible
candidate to try, record where weve achieved; If we get stuck, we backtrack
and try other options. We are sure by using this strategy, we can either find a
solution, or tell that the problem is unsolvable.
Of course there can be some variation, that we can stop when find one
answer, or go on searching all the solutions.

14.3. SOLUTION SEARCHING

491

If we draw a tree rooted at the starting state, expand it so that every branch
stands for a different attempt, our searching process is in a manner, that it
searches deeper and deeper. We wont consider any other options in the same
depth unless the searching fails so that weve to backtrack to upper level of
the tree. Figure 14.32 illustrates the order we search a state tree. The arrow
indicates how we go down and backtrack up. The number of the nodes shows
the order we visit them.

Figure 14.32: Example of DFS search order.

This kind of search strategy is called DFS (Deep-first-search). We widely
use it unintentionally. Some programming environments, Prolog for instance,
adopt DFS as the default evaluation model. A maze is given by a set of rule
base, such as:
c(a,
c(b,
c(e,
c(f,
c(g,
c(h,

b).
c).
d),
c).
d).
f).

c(a, e).
c(b, f).
c(e, f).
c(g, h).

Where predicate c(X, Y ) means place X is connected with Y . Note that

this is a directed predicate, we can make Y to be connected with X as well by
either adding a symmetric rule, or create a undirected predicate. Figure 14.33
shows such a directed graph. Given two places X and Y , Prolog can tell if they
are connected by the following program.
go(X, X).
go(X, Y) :- c(X, Z), go(Z, Y)

This program says that, a place is connected with itself. Given two different
places X and Y , if X is connected with Z, and Z is connected with Y , then X
is connected with Y . Note that, there might be multiple choices for Z. Prolog
selects a candidate, and go on further searching. It only tries other candidates
if the recursive searching fails. In that case, Prolog backtracks and tries other
alternatives. This is exactly what DFS does.
DFS is quite straightforward when we only need a solution, but dont care
if the solution takes the fewest steps. For example, the solution it gives, may
not be the shortest path for the maze. Well see some more puzzles next. They
demands to find the solution with the minimum attempts.

492

CHAPTER 14. SEARCHING

Figure 14.33: A directed graph.

The wolf, goat, and cabbage puzzle
This puzzle says that a farmer wants to cross a river with a wolf, a goat, and a
bucket of cabbage. There is a boat. Only the farmer can drive it. But the boat
is small. it can only hold one of the wolf, the goat, and the bucket of cabbage
with the farmer at a time. The farmer has to pick them one by one to the other
side of the river. However, the wolf would eat the goat, and the goat would eat
the cabbage if the farmer is absent. The puzzle asks to find the fast solution so
that they can all safely go cross the river.

Figure 14.34: The wolf, goat, cabbage puzzle

The key point to this puzzle is that the wolf does not eat the cabbage. The
farmer can safely pick the goat to the other side. But next time, no matter if he
pick the wolf or the cabbage to cross the river, he has to take one back to avoid
the conflict. In order to find the fast the solution, at any time, if the farmer has
multiple options, we can examine all of them in parallel, so that these different
decisions compete. If we count the number of the times the farmer cross the
river without considering the direction, that crossing the river back and forth
means 2 times, we are actually checking the complete possibilities after 1 time,

14.3. SOLUTION SEARCHING

493

2 times, 3 times, ... When we find a situation, that they all arrive at the other
bank, we are done. And this solution wins the competition, which is the fast
solution.
The problem is that we cant examine all the possible solutions in parallel
ideally. Even with a super computer equipped with many CPU cores, the setup
is too expensive to solve such a simple puzzle.
Lets consider a lucky draw game. People blindly pick from a box with
colored balls. There is only one black ball, all the others are white. The one
who pick the black ball wins the game; Otherwise, he must return the ball to
the box and wait for the next chance. In order to be fair enough, we can setup
a rule that no one can try the second time before all others have tried. We can
line people to a queue. Every time the first guy pick a ball, if he does not win,
he then stands at the tail of the queue to wait for the second try. This queue
helps to ensure our rule.

Figure 14.35: A lucky-draw game, the i-th person goes from the queue, pick a
ball, then join the queue at tail if he fails to pick the black ball.
We can use the quite same idea to solve our puzzle. The two banks of the
river can be represented as two sets A and B. A contains the wolf, the goat,
the cabbage and the farmer; while B is empty. We take an element from one
set to the other each time. The two sets cant hold conflict things if the farmer
is absent. The goal is to exchange the contents of A and B with fewest steps.
We initialize a queue with state A = {w, g, c, p}, B = as the only element.
As far as the queue isnt empty, we pick the first element from the head, expand
it with all possible options, and put these new expanded candidates to the tail
of the queue. If the first element on the head is the final goal, that A = , B =
{w, g, c, p}, we are done. Figure 14.36 illustrates the idea of this search order.
Note that as all possibilities in the same level are examined, there is no need
for back-tracking.

494

CHAPTER 14. SEARCHING

Figure 14.36: Start from state 1, check all possible options 2, 3, and 4 for next
step; then all nodes in level 3, ...
There is a simple way to treat the set. A four bits binary number can be
used, each bit stands for a thing, for example, the wolf w = 1, the goat g = 2,
the cabbage c = 4, and the farmer p = 8. That 0 stands for the empty set, 15
stands for a full set. Value 3, solely means there are a wolf and a goat on the
river bank. In this case, the wolf will eat the goat. Similarly, value 6 stands for
another conflicting case. Every time, we move the highest bit (which is 8), or
together with one of the other bits (4 or 2, or 1) from one number to the other.
The possible moves can be defined as below.

{(A 8 i, B + 8 + i)|i {0, 1, 2, 4}, i = 0 Ai 6= 0} : B < 8

{(A + 8 + i, B 8 i)|i {0, 1, 2, 4}, i = 0 Bi 6= 0} : Otherwise
(14.59)
Where is the bitwise-and operation.
the solution can be given by reusing the queue defined in previous chapter.
Denote the queue as Q, which is initialed with a singleton list {(15, 0)}. If Q is
not empty, function DeQ(Q) extracts the head element M , the updated queue
becomes Q0 . M is a list of pairs, stands for a series of movements between the
river banks. The first element in m1 = (A0 , B 0 ) is the latest state. Function
EnQ0 (Q, L) is a slightly different enqueue operation. It pushes all the possible
moving sequences in L to the tail of the queue one by one and returns the
updated queue. With these notations, the solution function is defined like below.
mv(A, B) =

: Q=
0
reverse(M

) : A =0
solve(Q) =
{m} M | m mv(m1 ),

)) : otherwise
solve(EnQ0 (Q0 ,
valid(m, M )
(14.60)
Where function valid(m, M ) checks if the new moving candidate m = (A00 , B 00 )
is valid. That neither A00 nor B 00 is 3 or 6, and m hasnt been tried before in
M to avoid any repeatedly attempts.
valid(m, M ) = A00 6= 3, A00 6= 6, B 00 6= 3, B 00 6= 6, m
/M

(14.61)

The following example Haskell program implements this solution. Note that
it uses a plain list to represent the queue for illustration purpose.

14.3. SOLUTION SEARCHING

495

import Data.Bits
solve = bfsSolve [[(15, 0)]] where
bfsSolve :: [[(Int, Int)]] [(Int, Int)]
bfsSolve [] = [] -- no solution
bfsSolve (c:cs) | (fst $ head c) == 0 = reverse c
| otherwise = bfsSolve (cs ++ map (:c)
(filter (valid c) $ moves $ head c))
valid (a, b) r = not $ or [ a elem [3, 6], b elem [3, 6],
(a, b) elem r]
moves (a, b) = if b < 8 then trans a b else map swap (trans b a) where
trans x y = [(x - 8 - i, y + 8 + i)
| i [0, 1, 2, 4], i == 0 | | (x &. i) /= 0]
swap (x, y) = (y, x)

This algorithm can be easily modified to find all the possible solutions, but
not just stop after finding the first one. This is left as the exercise to the reader.
The following shows the two best solutions to this puzzle.
Solution 1:
river Right
Left
wolf, goat, cabbage, farmer
wolf, cabbage
goat, farmer
wolf, cabbage, farmer
goat
cabbage
wolf, goat, farmer
goat, cabbage, farmer
wolf
goat
wolf, cabbage, farmer
goat, farmer
wolf, cabbage
wolf, goat, cabbage, farmer
Solution 2:
Left
river Right
wolf, goat, cabbage, farmer
wolf, cabbage
goat, farmer
wolf, cabbage, farmer
goat
wolf
goat, cabbage, farmer
wolf, goat, farmer
cabbage
goat
wolf, cabbage, farmer
goat, farmer
wolf, cabbage
wolf, goat, cabbage, farmer
This algorithm can also be realized imperatively. Observing that our solution
is in tail recursive manner, we can translate it directly to a loop. We use a list
S to hold all the solutions can be found. The singleton list {(15, 0)} is pushed
to queue when initializing. As long as the queue isnt empty, we extract the
head C from the queue by calling DeQ procedure. Examine if it reaches the
final goal, if not, we expand all the possible moves and push to the tail of the
queue for further searching.
1: function Solve
2:
S
3:
Q
4:
EnQ(Q, {(15, 0)})
5:
while Q 6= do
6:
C DeQ(Q)

496

CHAPTER 14. SEARCHING

if c1 = (0, 15) then

Add(S, Reverse(C))
else
10:
for m Moves(C) do
11:
if Valid(m, C) then
12:
EnQ(Q, {m} C)
13:
return S
Where Moves, and Valid procedures are as same as before. The following
Python example program implements this imperative algorithm.
7:
8:
9:

def solve():
s = []
queue = [[(0xf, 0)]]
while queue != []:
cur = queue.pop(0)
if cur[0] == (0, 0xf):
s.append(reverse(cur))
else:
for m in moves(cur):
queue.append([m]+cur)
return s
def moves(s):
(a, b) = s[0]
return valid(s, trans(a, b) if b < 8 else swaps(trans(b, a)))
def valid(s, mv):
return [(a, b) for (a, b) in mv
if a not in [3, 6] and b not in [3, 6] and (a, b) not in s]
def trans(a, b):
masks = [ 8 | (1<<i) for i in range(4)]
return [(a ^ mask, b | mask) for mask in masks if a & mask == mask]
def swaps(s):
return [(b, a) for (a, b) in s]

There is a minor difference between the program and the pseudo code, that
the function to generate candidate moving options filters the invalid cases inside
it.
Every time, no matter the farmer drives the boat back and forth, there are
m options for him to choose, where m is the number of objects on the river bank
the farmer drives from. m is always less than 4, that the algorithm wont take
more than n4 times at step n. This estimation is far more than the actual time,
because we avoid trying all invalid cases. Our solution examines all the possible
moving in the worst case. Because we check recorded steps to avoid repeated
attempt, the algorithm takes about O(n2 ) time to search for n possible steps.
Water jugs puzzle
This is a popular puzzle in classic AI. The history of it should be very long. It
says that there are two jugs, one is 9 quarts, the other is 4 quarts. How to use
them to bring up from the river exactly 6 quarts of water?

14.3. SOLUTION SEARCHING

497

There are varies versions of this puzzle, although the volume of the jugs, and
the target volume of water differ. The solver is said to be young Blaise Pascal
when he was a child, the French mathematician, scientist in one story, and
Sim`eon Denis Poisson in another story. Later in the popular Hollywood movie
Die-Hard 3, actor Bruce Willis and Samuel L. Jackson were also confronted
with this puzzle.
P`olya gave a nice way to solve this problem backwards in [14].

Figure 14.37: Two jugs with volume of 9 and 4.

Instead of thinking from the starting state as shown in figure 14.37. P`olya
pointed out that there will be 6 quarts of water in the bigger jugs at the final
stage, which indicates the second last step, we can fill the 9 quarts jug, then
pour out 3 quarts from it. In order to achieve this, there should be 1 quart of
water left in the smaller jug as shown in figure 14.38.

Figure 14.38: The last two steps.

Its easy to see that fill the 9 quarters jug, then pour to the 4 quarters jug
twice can bring 1 quarters of water. As shown in figure 14.39. At this stage,
weve found the solution. By reversing our findings, we can give the correct
steps to bring exactly 6 quarters of water.
P`olyas methodology is general. Its still hard to solve it without concrete
algorithm. For instance, how to bring up 2 gallons of water from 899 and 1147
gallon jugs?
There are 6 ways to deal with 2 jugs in total. Denote the smaller jug as A,
the bigger jug as B.

498

CHAPTER 14. SEARCHING

Figure 14.39: Fill the bigger jugs, and pour to the smaller one twice.
Fill jug A from the river;
Fill jug B from the river;
Empty jug A;
Empty jug B;
Pour water from jug A to B;
Pour water from jug B to A.
The following sequence shows an example. Note that in this example, we
assume that a < b < 2a.
A
B
operation
0
0
start
a
0
fill A
0
a
pour A into B
a
a
fill A
2a - b
b
pour A into B
2a - b
0
empty B
0
2a - b pour A into B
a
2a - b fill A
3a - 2b b
pour A into B
...
...
...
No matter what the above operations are taken, the amount of water in each
jug can be expressed as xa + yb, where a and b are volumes of jugs, for some
integers x and y. All the amounts of water we can get are linear combination of
a and b. We can immediately tell given two jugs, if a goal g is solvable or not.
For instance, we cant bring 5 gallons of water with two jugs of volume 4
and 6 gallon. The number theory ensures that, the 2 water jugs puzzle can be
solved if and only if g can be divided by the greatest common divisor of a and
b. Written as:
gcd(a, b)|g

(14.62)

Where m|n means n can be divided by m. Whats more, if a and b are relatively prime, which means gcd(a, b) = 1, its possible to bring up any quantity
g of water.

14.3. SOLUTION SEARCHING

499

Although gcd(a, b) enables us to determine if the puzzle is solvable, it doesnt

give us the detailed pour sequence. If we can find some integer x and y, so that
g = xa + yb. We can arrange a sequence of operations (even it may not be the
best solution) to solve it. The idea is that, without loss of generality, suppose
x > 0, y < 0, we need fill jug A by x times, and empty jug B by y times in total.
Lets take a = 3, b = 5, and g = 4 for example, since 4 = 3 3 5, we can
arrange a sequence like the following.
A B operation
0 0
start
3 0
fill A
pour A into B
0 3
3 3
fill A
1 5
pour A into B
1 0
empty B
0 1
pour A into B
3 1
fill A
0 4
pour A into B
In this sequence, we fill A by 3 times, and empty B by 1 time. The procedure
can be described as the following:
Repeat x times:
1. Fill jug A;
2. Pour jug A into jug B, whenever B is full, empty it.
So the only problem left is to find the x and y. There is a powerful tool
in number theory called, Extended Euclid algorithm, which can achieve this.
Compare to the classic Euclid GCD algorithm, which can only give the greatest
common divisor, The extended Euclid algorithm can give a pair of x, y as well,
so that:
(d, x, y) = gcdext (a, b)

(14.63)

where d = gcd(a, b) and ax + by = d. Without loss of generality, suppose

a < b, there exits quotation q and remainder r that:
b = aq + r

(14.64)

Since d is the common divisor, it can divide both a and b, thus d can divide
r as well. Because r is less than a, we can scale down the problem by finding
GCD of a and r:
(d, x0 , y 0 ) = gcdext (r, a)

(14.65)

Where d = x0 r + y 0 a according to the definition of the extended Euclid

algorithm. Transform b = aq + r to r = b aq, substitute r in above equation
yields:
d

= x0 (b aq) + y 0 a
= (y 0 x0 q)a + x0 b

(14.66)

500

CHAPTER 14. SEARCHING

This is the linear combination of a and b, so that we have:
(
b
x = y 0 x0
a
y = x0

(14.67)

Note that this is a typical recursive relationship. The edge case happens
when a = 0.
gcd(0, b) = b = 0a + 1b

(14.68)

Summarize the above result, the extended Euclid algorithm can be defined
as the following:
(
(b, 0, 1) : a = 0
gcdext (a, b) =
(14.69)
b
0
(d, y x0 , x0 ) : otherwise
a
Where d, x0 , y 0 are defined in equation (14.65).
The 2 water jugs puzzle is almost solved, but there are still two detailed
problems need to be tackled. First, extended Euclid algorithm gives the linear
combination for the greatest common divisor d. While the target volume of
water g isnt necessarily equal to d. This can be easily solved by multiplying x
and y by m times, where m = g/gcd(a, b); Second, we assume x > 0, to form a
procedure to fill jug A with x times. However, the extended Euclid algorithm
doesnt ensure x to be positive. For instance gcdext (4, 9) = (1, 2, 1). Whenever
we get a negative x, since d = xa + yb, we can continuously add b to x, and
decrease y by a till a is greater than zero.
At this stage, we are able to give the complete solution to the 2 water jugs
puzzle. Below is an example Haskell program.
extGcd 0 b = (b, 0, 1)
extGcd a b = let (d, x, y) = extGcd (b mod a) a in
(d, y - x (b div a), x)
solve a b g | g mod d /= 0 = [] -- no solution
| otherwise = solve (x g div d)
where
(d, x, y) = extGcd a b
solve x | x < 0 = solve (x + b)
| otherwise = pour x [(0, 0)]
pour 0 ps = reverse ((0, g):ps)
pour x ps@((a, b):_) | a == 0 = pour (x - 1) ((a, b):ps) -- fill a
| b == b = pour x ((a, 0):ps) -- empty b
| otherwise = pour x ((max 0 (a + b - b),
min (a + b) b):ps)

Although we can solve the 2 water jugs puzzle with extended Euclid algorithm, the solution may not be the best. For instance, when we are going
to bring 4 gallons of water from 3 and 5 gallons jugs. The extended Euclid
algorithm produces the following sequence:
[(0,0),(3,0),(0,3),(3,3),(1,5),(1,0),(0,1),(3,1),
(0,4),(3,4),(2,5),(2,0),(0,2),(3,2),(0,5),(3,5),
(3,0),(0,3),(3,3),(1,5),(1,0),(0,1),(3,1),(0,4)]

14.3. SOLUTION SEARCHING

501

It takes 23 steps to achieve the goal, while the best solution only need 6
steps:
[(0,0),(0,5),(3,2),(0,2),(2,0),(2,5),(3,4)]
Observe the 23 steps, and we can find that jug B has already contained 4
gallons of water at the 8-th step. But the algorithm ignores this fact and goes
on executing the left 15 steps. The reason is that the linear combination x and y
we find with the extended Euclid algorithm are not the only numbers satisfying
g = xa + by. For all these numbers, the smaller |x| + |y|, the less steps are
needed. There is an exercise to addressing this problem in this section.
The interesting problem is how to find the best solution? We have two
approaches, one is to find x and y to minimize |x| + |y|; the other is to adopt
the quite similar idea as the wolf-goat-cabbage puzzle. We focus on the latter
in this section. Since there are at most 6 possible options: fill A, fill B, pour
A into B, pour B into A, empty A and empty B, we can try them in parallel,
and check which decision can lead to the best solution. We need record all the
states weve achieved to avoid any potential repetition. In order to realize this
parallel approach with reasonable resources, a queue can be used to arrange our
attempts. The elements stored in this queue are series of pairs (p, q), where p
and q represent the volume of waters contained in each jug. These pairs record
the sequence of our operations from the beginning to the latest. We initialize
the queue with the singleton list contains the starting state {(0, 0)}.
solve(a, b, g) = solve0 {{(0, 0)}}

(14.70)

Every time, when the queue isnt empty, we pick a sequence from the head
of the queue. If this sequence ends with a pair contains the target volume g, we
find a solution, we can print this sequence by reversing it; Otherwise, we expand
the latest pair by trying all the possible 6 options, remove any duplicated states,
and add them to the tail of the queue. Denote the queue as Q, the first sequence
stored on the head of the queue as S, the latest pair in S as (p, q), and the rest
of pairs as S 0 . After popping the head element, the queue becomes Q0 . This
algorithm can be defined like below:

Q=
p=gq =g
otherwise
(14.71)
Where function EnQ0 pushes a list of sequence to the queue one by one.
Function try(S) will try all possible 6 options to generate new pairs of water
volumes:
solve0 (Q) =

:
reverse(S) :

solve0 (EnQ0 (Q0 , {{s0 } S 0 |s0 try(S)})) :

f illA(p, q), f illB(p, q),

pourA(p, q), pourB(p, q),

try(S) = {s0 |s0
, s0
/ S0}

emptyA(p, q), emptyB(p, q)

(14.72)

Its intuitive to define the 6 options. For fill operations, the result is that the
volume of the filled jug is full; for empty operation, the result volume is empty;

502

CHAPTER 14. SEARCHING

for pour operation, we need test if the jug is big enough to hold all the water.
f illA(p, q) = (a, q)
f illB(p, q) = (p, b)
emptyA(p, q) = (0, q)
emptyB(p, q) = (p, 0)
pourA(p, q) = (max(0, p + q b), min(x + y, b))
pourB(p, q) = (min(x + y, a), max(0, x + y a))
(14.73)
The following example Haskell program implements this method:
solve a b g = bfs [[(0, 0)]] where
bfs [] = []
bfs (c:cs) | fst (head c) == g | | snd (head c) == g = reverse c
| otherwise = bfs (cs ++ map (:c) (expand c))
expand ((x, y):ps) = filter (notElem ps) $ map (f f x y)
[fillA, fillB, pourA, pourB, emptyA, emptyB]
fillA _ y = (a, y)
fillB x _ = (x, b)
emptyA _ y = (0, y)
emptyB x _ = (x, 0)
pourA x y = (max 0 (x + y - b), min (x + y) b)
pourB x y = (min (x + y) a, max 0 (x + y - a))

This method always returns the fast solution. It can also be realized in
imperative approach. Instead of storing the complete sequence of operations in
every element in the queue, we can store the unique state in a global history
list, and use links to track the operation sequence, this can save spaces.
(0, 0)
(3, 0)
(0, 5)
(0, 0)

(3, 5)
(0, 3)
(3, 2)
...

fill A

flll B

(3, 0)

(0, 5)

fill B empty A pour A

(3, 5)

(0, 0)

(0, 3)

fill A
(3, 5)

empty B
(0, 0)

pour B
(3, 2)

...

Figure 14.40: All attempted states are stored in a global list.

The idea is illustrated in figure 14.40. The initial state is (0, 0). Only fill
A and fill B are possible. They are tried and added to the record list; Next
we can try and record fill B on top of (3, 0), which yields new state (3, 5).

14.3. SOLUTION SEARCHING

503

However, when try empty A from state (3, 0), we would return to the start
state (0, 0). As this previous state has been recorded, it is ignored. All the
repeated states are in gray color in this figure.
With such settings, we neednt remember the operation sequence in each
element in the queue explicitly. We can add a parent link to each node in
figure 14.40, and use it to back-traverse to the starting point from any state.
The following example ANSI C code shows such a definition.
struct Step {
int p, q;
struct Step parent;
};
struct Step make_step(int p, int q, struct Step parent) {
struct Step s = (struct Step) malloc(sizeof(struct Step));
sp = p;
sq = q;
sparent = parent;
return s;
}

Where p, q are volumes of water in the 2 jugs. For any state s, define
functions p(s) and q(s) return these 2 values, the imperative algorithm can be
realized based on this idea as below.
1: function Solve(a, b, g)
2:
Q
3:
Push-and-record(Q, (0, 0))
4:
while Q 6= do
5:
s Pop(Q)
6:
if p(s) = g q(s) = g then
7:
return s
8:
else
9:
C Expand(s)
10:
for c C do
11:
if c 6= s Visited(c) then
12:
Push-and-record(Q, c)
13:
return NIL
Where Push-and-record does not only push an element to the queue, but
also record this element as visited, so that we can check if an element has been
visited before in the future. This can be implemented with a list. All push
operations append the new elements to the tail. For pop operation, instead of
removing the element pointed by head, the head pointer only advances to the
next one. This list contains historic data which has to be reset explicitly. The
following ANSI C code illustrates this idea.
struct Step steps[1000], head, tail = steps;
void push(struct Step s) { tail++ = s; }
struct Step pop() { return head++; }
int empty() { return head == tail; }

504

CHAPTER 14. SEARCHING

void reset() {
struct Step p;
for (p = steps; p != tail; ++p)
free(p);
head = tail = steps;
}

In order to test a state has been visited, we can traverse the list to compare
p and q.
int eq(struct Step a, struct Step b) {
return ap == bp && aq == bq;
}
int visited(struct Step s) {
struct Step p;
for (p = steps; p != tail; ++p)
if (eq(p, s)) return 1;
return 0;
}

The main program can be implemented as below:

struct Step solve(int a, int b, int g) {
int i;
struct Step cur, cs[6];
reset();
push(make_step(0, 0, NULL));
while (!empty()) {
cur = pop();
if (curp == g | | curq == g)
return cur;
else {
expand(cur, a, b, cs);
for (i = 0; i < 6; ++i)
if(!eq(cur, cs[i]) && !visited(cs[i]))
push(cs[i]);
}
}
return NULL;
}

Where function expand tries all the 6 possible options:

void expand(struct Step s, int a, int b, struct Step cs) {
int p = sp, q = sq;
cs[0] = make_step(a, q, s); /fill A/
cs[1] = make_step(p, b, s); /fill B/
cs[2] = make_step(0, q, s); /empty A/
cs[3] = make_step(p, 0, s); /empty B/
cs[4] = make_step(max(0, p + q - b), min(p + q, b), s); /pour A/
cs[5] = make_step(min(p + q, a), max(0, p + q - a), s); /pour B/
}

And the result steps is back tracked in reversed order, it can be output with
a recursive function:

14.3. SOLUTION SEARCHING

505

void print(struct Step s) {

if (s) {
print(sparent);
printf("%d, %dn", sp, sq);
}
}

Kloski
Kloski is a block sliding puzzle. It appears in many countries. There are different
sizes and layouts. Figure 14.41 illustrates a traditional Kloski game in China.

(a) Initial layout of blocks

(b) Block layout after several

movements

Figure 14.41: Huarong Dao, the traditional Kloski game in China.

In this puzzle, there are 10 blocks, each is labeled with text or icon. The
smallest block has size of 1 unit square, the biggest one is 2 2 units size. Note
there is a slot of 2 units wide at the middle-bottom of the board. The biggest
block represents a king in ancient time, while the others are enemies. The goal
is to move the biggest block to the slot, so that the king can escape. This game
is named as Huarong Dao, or Huarong Escape in China. Figure 14.42 shows
the similar Kloski puzzle in Japan. The biggest block means daughter, while
the others are her family members. This game is named as Daughter in the
box in Japan (Japanese name: hakoiri musume).

Figure 14.42: Daughter in the box, the Kloski game in Japan.

506

CHAPTER 14. SEARCHING

In this section, we want to find a solution, which can slide blocks from the
initial state to the final state with the minimum movements.
The intuitive idea to model this puzzle is to use a 5 4 matrix representing
the board. All pieces are labeled with a number. The following matrix M , for
example, shows the initial state of the puzzle.

1 10 10 2
1 10 10 2

4 5
M =

3 4
3 7
8 5
6 0
0 9
In this matrix, the cells of value i mean the i-th piece covers this cell. The
special value 0 represents a free cell. By using sequence 1, 2, ... to identify
pieces, a special layout can be further simplified as an array L. Each element
is a list of cells covered by the piece indexed with this element. For example,
L[4] = {(3, 2), (3, 3)} means the 4-th piece covers cells at position (3, 2) and
(3, 3), where (i, j) means the cell at row i and column j.
The starting layout can be written as the following Array.
{{(1, 1), (2, 1)}, {(1, 4), (2, 4)}, {(3, 1), (4, 1)}, {(3, 2), (3, 3)}, {(3, 4), (4, 4)},
{(5, 1)}, {(4, 2)}, {(4, 3)}, {(5, 4)}, {(1, 2), (1, 3), (2, 2), (2, 3)}}
When moving the Kloski blocks, we need examine all the 10 blocks, checking
each block if it can move up, down, left and right. it seems that this approach
would lead to a very huge amount of possibilities, because each step might have
10 4 options, there will be about 40n cases in the n-th step.
Actually, there wont be so much options. For example, in the first step,
there are only 4 valid moving: the 6-th piece moves right; the 7-th and 8-th
move down; and the 9-th moves left.
All others are invalid moving. Figure 14.43 shows how to test if the moving
is possible.
The left example illustrates sliding block labeled with 1 down. There are two
cells covered by this block. The upper 1 moves to the cell previously occupied
by this same block, which is also labeled with 1; The lower 1 moves to a free
cell, which is labeled with 0;
The right example, on the other hand, illustrates invalid sliding. In this case,
the upper cells could move to the cell occupied by the same block. However, the
lower cell labeled with 1 cant move to the cell occupied by other block, which
is labeled with 2.
In order to test the valid moving, we need examine all the cells a block
will cover. If they are labeled with 0 or a number as same as this block, the
moving is valid. Otherwise it conflicts with some other block. For a layout L,
the corresponding matrix is M , suppose we want to move the k-th block with
(x, y), where |x| 1, |y| 1. The following equation tells if the moving
is valid:
valid(L, k, x, y) :
(i, j) L[k] i0 = i + y, j 0 = j + x,
(1, 1) (i0 , j 0 ) (5, 4), Mi0 j 0 {k, 0}

(14.74)

14.3. SOLUTION SEARCHING

507

Figure 14.43: Left: both the upper and the lower 1 are OK; Right: the upper 1
is OK, the lower 1 conflicts with 2.

Another important point to solve Kloski puzzle, is about how to avoid repeated attempts. The obvious case is that after a series of sliding, we end up
a matrix which have been transformed from. However, it is not enough to only
avoid the same matrix. Consider the following two metrics. Although M1 6= M2 ,
we need drop options to M2 , because they are essentially the same.

M1 =

1
1
3
3
6

10
10
4
7
0

10
10
4
8
0

2
2
5
5
9

M2 =

2
2
3
3
8

10
10
4
7
0

10
10
4
6
0

1
1
5
5
9

This fact tells us, that we should compare the layout, but not merely matrix
to avoid repetition. Denote the corresponding layouts as L1 and L2 respectively,
its easy to verify that ||L1 || = ||L2 ||, where ||L|| is the normalized layout, which
is defined as below:

||L|| = sort({sort(li )|li L})

(14.75)

In other words, a normalized layout is ordered for all its elements, and every
element is also ordered. The ordering can be defined as that (a, b) (c, d)
an + b cn + d, where n is the width of the matrix.
Observing that the Kloski board is symmetric, thus a layout can be mirrored
from another one. Mirrored layout is also a kind of repeating, which should be
avoided. The following M1 and M2 show such an example.

508

CHAPTER 14. SEARCHING

M1 =

10 10 1 2
3
3
10 10 1 2

3 5 4 4
M2 = 4

7
3 5 8 9
6 7 0 0
0

1
1
4
6
0

10
10
2
2
9

10
10
5
5
8

Note that, the normalized layouts are symmetric to each other. Its easy to
get a mirrored layout like this:
mirror(L) = {{(i, n j + 1)|(i, j) l}|l L}

(14.76)

We find that the matrix representation is useful in validating the moving,

while the layout is handy to model the moving and avoid repeated attempt.
We can use the similar approach to solve the Kloski puzzle. We need a queue,
every element in the queue contains two parts: a series of moving and the latest
layout led by the moving. Each moving is in form of (k, (y, x)), which means
moving the k-th block, with y in row, and x in column in the board.
The queue contains the starting layout when initialized. Whenever this
queue isnt empty, we pick the first one from the head, checking if the biggest
block is on target, that L[10] = {(4, 2), (4, 3), (5, 2), (5, 3)}. If yes, then we are
done; otherwise, we try to move every block with 4 options: left, right, up, and
down, and store all the possible, unique new layout to the tail of the queue.
During this searching, we need record all the normalized layouts weve ever
found to avoid any duplication.
Denote the queue as Q, the historic layouts as H, the first layout on the head
of the queue as L, its corresponding matrix as M . and the moving sequence to
this layout as S. The algorithm can be defined as the following.

: Q=
reverse(S) : L[10] = {(4, 2), (4, 3), (5, 2), (5, 3)}

solve(Q0 , H 0 ) : otherwise
(14.77)
The first clause says that if the queue is empty, weve tried all the possibilities
and cant find a solution; The second clause finds a solution, it returns the
moving sequence in reversed order; These are two edge cases. Otherwise, the
algorithm expands the current layout, puts all the valid new layouts to the tail
of the queue to yield Q0 , and updates the normalized layouts to H 0 . Then it
performs recursive searching.
In order to expand a layout to valid unique new layouts, we can define a
function as below:
solve(Q, H) =

expand(L, H) = {(k, (y, x)| k {1, 2, ..., 10},

(y, x) {(0, 1), (0, 1), (1, 0), (1, 0)},
valid(L, k, x, y), unique(L0 , H)}
(14.78)
Where L0 is the the new layout by moving the k-th block with (y, x)
from L, M 0 is the corresponding matrix, and M 00 is the matrix to the mirrored
layout of L0 . Function unique is defined like this:
unique(L0 , H) = M 0
/ H M 00
/H

(14.79)

14.3. SOLUTION SEARCHING

509

Well next show some example Haskell Kloski programs. As array isnt
mutable in the purely functional settings, tree based map is used to represent
layout 10 . Some type synonyms are defined as below:
import qualified Data.Map as M
import Data.Ix
import Data.List (sort)
type Point = (Integer, Integer)
type Layout = M.Map Integer [Point]
type Move = (Integer, Point)
data Ops = Op Layout [Move]

The main program is almost as same as the sort(Q, H) function defined

above.
solve :: [Ops] [[[Point]]] [Move]
solve [] _ = [] -- no solution
solve (Op x seq : cs) visit
| M.lookup 10 x == Just [(4, 2), (4, 3), (5, 2), (5, 3)] = reverse seq
| otherwise = solve q visit
where
ops = expand x visit
visit = map (layout move x) ops ++ visit
q = cs ++ [Op (move x op) (op:seq) | op ops ]

Where function layout gives the normalized form by sorting. move returns
the updated map by sliding the i-th block with (y, x).
layout = sort map sort M.elems
move x (i, d) = M.update (Just map (flip shift d)) i x
shift (y, x) (dy, dx) = (y + dy, x + dx)

Function expand gives all the possible new options. It can be directly translated from expand(L, H).
expand :: Layout [[[Point]]] [Move]
expand x visit = [(i, d) | i [1..10],
d [(0, -1), (0, 1), (-1, 0), (1, 0)],
valid i d, unique i d] where
valid i d = all (p let p = shift p d in
inRange (bounds board) p &&
(M.keys $ M.filter (elem p) x) elem [[i], []])
(maybe [] id $ M.lookup i x)
unique i d = let mv = move x (i, d) in
all (notElem visit) (map layout [mv, mirror mv])

Note that we also filter out the mirrored layouts. The mirror function is
given as the following.
mirror = M.map (map ( (y, x) (y, 5 - x)))

This program takes several minutes to produce the best solution, which takes
116 steps. The final 3 steps are shown as below:
10 Alternatively,

finger tree based sequence shown in previous chapter can be used

510

CHAPTER 14. SEARCHING

...
[5,
[5,
[7,
[A,
[A,

3,
3,
9,
A,
A,

2,
2,
4,
6,
0,

1]
1]
4]
0]
8]

[5,
[5,
[7,
[A,
[A,

3,
3,
9,
A,
A,

2,
2,
4,
0,
0,

1]
1]
4]
6]
8]

[5,
[5,
[7,
[0,
[0,

3,
3,
9,
A,
A,

2,
2,
4,
A,
A,

1]
1]
4]
6]
8]

total 116 steps

The Kloski solution can also be realized imperatively. Note that the solve(Q, H)
is tail-recursive, its easy to transform the algorithm with looping. We can also
link one layout to its parent, so that the moving sequence can be recorded
globally. This can save some spaces, as the queue neednt store the moving information in every element. When output the result, we only need back-tracking
to the starting layout from the last one.
Suppose function Link(L0 , L) links a new layout L0 to its parent layout L.
The following algorithm takes a starting layout, and searches for best moving
sequence.
1: function Solve(L0 )
2:
H ||L0 ||
3:
Q
4:
Push(Q, Link(L0 , NIL))
5:
while Q 6= do
6:
L Pop(Q)
7:
if L[10] = {(4, 2), (4, 3), (5, 2), (5, 3)} then
8:
return L
9:
else
10:
for each L0 Expand(L, H) do
11:
Push(Q, Link(L0 , L))
12:
Append(H, ||L0 ||)
13:
return NIL
. No solution
The following example Python program implements this algorithm:
class Node:
def __init__(self, l, p = None):
self.layout = l
self.parent = p

14.3. SOLUTION SEARCHING

511

def solve(start):
visit = [normalize(start)]
queue = [Node(start)]
while queue != []:
cur = queue.pop(0)
layout = cur.layout
if layout[-1] == [(4, 2), (4, 3), (5, 2), (5, 3)]:
return cur
else:
for brd in expand(layout, visit):
queue.append(Node(brd, cur))
visit.append(normalize(brd))
return None # no solution

Where normalize and expand are implemented as below:

def normalize(layout):
return sorted([sorted(r) for r in layout])
def expand(layout, visit):
def bound(y, x):
return 1 y and y 5 and 1 x and x 4
def valid(m, i, y, x):
return m[y - 1][x - 1] in [0, i]
def unique(brd):
(m, n) = (normalize(brd), normalize(mirror(brd)))
return all(m != v and n != v for v in visit)
s = []
d = [(0, -1), (0, 1), (-1, 0), (1, 0)]
m = matrix(layout)
for i in range(1, 11):
for (dy, dx) in d:
if all(bound(y + dy, x + dx) and valid(m, i, y + dy, x + dx)
for (y, x) in layout[i - 1]):
brd = move(layout, (i, (dy, dx)))
if unique(brd):
s.append(brd)
return s

Like most programming languages, arrays are indexed from 0 but not 1 in
Python. This has to be handled properly. The rest functions including mirror,
matrix, and move are implemented as the following.
def mirror(layout):
return [[(y, 5 - x) for (y, x) in r] for r in layout]
def matrix(layout):
m = [[0]4 for _ in range(5)]
for (i, ps) in zip(range(1, 11), layout):
for (y, x) in ps:
m[y - 1][x - 1] = i
return m
def move(layout, delta):
(i, (dy, dx)) = delta

512

CHAPTER 14. SEARCHING

m = dup(layout)
m[i - 1] = [(y + dy, x + dx) for (y, x) in m[i - 1]]
return m

def dup(layout):
return [r[:] for r in layout]

Its possible to modify this Kloski algorithm, so that it does not only stop
at the fast solution, but also search all the solutions. In such case, the computation time is bound to the size of a space V , where V holds all the layouts can
be transformed from the starting layout. If all these layouts are stored globally, with a parent field point to the predecessor, the space requirement of this
algorithm is also bound to O(V ).
Summary of BFS
The above three puzzles, the wolf-goat-cabbage puzzle, the water jugs puzzle,
and the Kloski puzzle show some common solution structure. Similar to the
DFS problems, they all have the starting state and the end state. The wolfgoat-cabbage puzzle starts with the wolf, the goat, the cabbage and the farmer
all in one side, while the other side is empty. It ends up in a state that they
all moved to the other side. The water jugs puzzle starts with two empty jugs,
and ends with either jug contains a certain volume of water. The Kloski puzzle
starts from a layout and ends to another layout that the biggest block begging
slided to a given position.
All problems specify a set of rules which can transfer from one state to
another. Different form the DFS approach, we try all the possible options in
parallel. We wont search further until all the other alternatives in the same step
have been examined. This method ensures that the solution with the minimum
steps can be found before those with more steps. Review and compare the two
figures weve drawn before shows the difference between these two approaches.
For the later one, because we expand the searching horizontally, it is called as
Breadth-first search (BFS for short).

(a) Depth First Search

(b) Breadth First Search

Figure 14.44: Search orders for DFS and BFS.

As we cant perform search really in parallel, BFS realization typically utilizes a queue to store the search options. The candidate with less steps pops

14.3. SOLUTION SEARCHING

513

from the head, while the new candidate with more steps is pushed to the tail
of the queue. Note that the queue should meet constant time enqueue and dequeue requirement, which weve explained in previous chapter of queue. Strictly
speaking, the example functional programs shown above dont meet this criteria. They use list to mimic queue, which can only provide linear time pushing.
Readers can replace them with the functional queue we explained before.
BFS provides a simple method to search for optimal solutions in terms of the
number of steps. However, it cant search for more general optimal solution.
Consider another directed graph as shown in figure 14.45, the length of each
section varies. We cant use BFS to find the shortest route from one city to
another.
a
15

g
4

9
e

11
7

10
f

h
5

12
d

6
c

Figure 14.45: A weighted directed graph.

Note that the shortest route from city a to city c isnt the one with the
fewest steps a b c. The total length of this route is 22; But the route with
more steps a e f c is the best. The length of it is 20. The coming
sections introduce other algorithms to search for optimal solution.

14.3.2

Search the optimal solution

Searching for the optimal solution is quite important in many aspects. People
need the best solution to save time, space, cost, or energy. However, its not
easy to find the best solution with limited resources. There have been many
optimal problems can only be solved by brute-force. Nevertheless, weve found
that, for some of them, There exists special simplified ways to search the optimal
solution.
Grady algorithm
Huffman coding
Huffman coding is a solution to encode information with the shortest length of
code. Consider the popular ASCII code, which uses 7 bits to encode characters,
digits, and symbols. ASCII code can represent 27 = 128 different symbols. With
0, 1 bits, we need at least log2 n bits to distinguish n different symbols. For text
with only case insensitive English letters, we can define a code table like below.

514

CHAPTER 14. SEARCHING

char code
char code
A
00000 N
01101
B
00001 O
01110
C
00010 P
01111
D
00011 Q
10000
E
00100 R
10001
F
00101 S
10010
G
00110 T
10011
00111 U
10100
H
I
01000 V
10101
J
01001 W
10110
K
01010 X
10111
L
01011 Y
11000
M
01100 Z
11001
With this code table, text INTERNATIONAL is encoded to 65 bits.

00010101101100100100100011011000000110010001001110101100000011010
Observe the above code table, which actually maps the letter A to Z from
0 to 25. There are 5 bits to represent every code. Code zero is forced as 00000
but not 0 for example. Such kind of coding method, is called fixed-length
coding.
Another coding method is variable-length coding. That we can use just
one bit 0 for A, two bits 10 for C, and 5 bits 11001 for Z. Although
this approach can shorten the total length of the code for INTERNATIONAL
from 65 bits dramatically, it causes problem when decoding. When processing
a sequence of bits like 1101, we dont know if it means 1 followed by 101,
which stands for BF; or 110 followed by 1, which is GB, or 1101 which is
N.
The famous Morse code is variable-length coding system. That the most
used letter E is encoded as a dot, while Z is encoded as two dashes and two
dots. Morse code uses a special pause separator to indicate the termination of
a code, so that the above problem wont happen. There is another solution to
avoid ambiguity. Consider the following code table.
char code char code
A
110
E
1110
I
101
L
1111
N
01
O
000
R
001
T
100
Text INTERNATIONAL is encoded to 38 bits only:
10101100111000101110100101000011101111
If decode the bits against the above code table, we wont meet any ambiguity
symbols. This is because there is no code for any symbol is the prefix of another
one. Such code is called prefix-code. (You may wonder why it isnt called as
non-prefix code.) By using prefix-code, we neednt separators at all. So that
the length of the code can be shorten.
This is a very interesting problem. Can we find a prefix-code table, which
produce the shortest code for a given text? The very same problem was given
to David A. Huffman in 1951, who was still a student in MIT[15]. His professor

14.3. SOLUTION SEARCHING

515

Robert M. Fano told the class that those who could solve this problem neednt
take the final exam. Huffman almost gave up and started preparing the final
exam when he found the most efficient answer.
The idea is to create the coding table according to the frequency of the
symbol appeared in the text. The more used symbol is assigned with the shorter
code.
Its not hard to process some text, and calculate the occurrence for each
symbol. So that we have a symbol set, each one is augmented with a weight.
The weight can be the number which indicates the frequency this symbol occurs.
We can use the number of occurrence, or the probabilities for example.
Huffman discovered that a binary tree can be used to generate prefix-code.
All symbols are stored in the leaf nodes. The codes are generated by traversing
the tree from root. When go left, we add a zero; and when go right we add a
one.
Figure 14.46 illustrates a binary tree. Taking symbol N for example, starting from the root, we first go left, then right and arrive at N. Thus the code
for N is 01; While for symbol A, we can go right, right, then left. So A is
encode to 110. Note that this approach ensures none code is the prefix of the
other.
13

O, 1

N, 3

R, 1

T, 2

I, 2

A, 2

E, 1

L, 1

Figure 14.46: An encoding tree.

Note that this tree can also be used directly for decoding. When scan a
series of bits, if the bit is zero, we go left; if the bit is one, we go right. When
arrive at a leaf, we decode a symbol from that leaf. And we restart from the
root of the tree for the coming bits.
Given a list of symbols with weights, we need build such a binary tree, so
that the symbol with greater weight has shorter path from the root. Huffman
developed a bottom-up solution. When start, all symbols are put into a leaf
node. Every time, we pick two nodes, which has the smallest weight, and merge
them into a branch node. The weight of this branch is the sum of its two
children. We repeatedly pick the two smallest weighted nodes and merge till
there is only one tree left. Figure 14.47 illustrates such a building process.
We can reuse the binary tree definition to formalize Huffman coding. We

516

CHAPTER 14. SEARCHING

E, 1

L, 1

O, 1

R, 1

(a) 1.

T, 2

I, 2

(b) 2.

A, 2

E, 1

L, 1

N, 3

O, 1

R, 1

(d) 4.

(e) 5.
8

T, 2

I, 2

A, 2

E, 1

L, 1

(f) 6.
13

O, 1

N, 3

R, 1

T, 2

I, 2

A, 2

E, 1

L, 1

(g) 7.

Figure 14.47: Steps to build a Huffman tree.

14.3. SOLUTION SEARCHING

517

augment the weight information, and the symbols are only stored in leaf nodes.
The following C like definition, shows an example.
struct Node {
int w;
char c;
struct Node left, right;
};

Some limitation can be added to the definition, as empty tree isnt allowed.
A Huffman tree is either a leaf, which contains a symbol and its weight; or a
branch, which only holds total weight of all leaves. The following Haskell code,
for instance, explicitly specifies these two cases.
data HTr w a = Leaf w a | Branch w (HTr w a) (HTr w a)

When merge two Huffman trees T1 and T2 to a bigger one, These two trees
are set as its children. We can select either one as the left, and the other as
the right. the weight of the result tree T is the sum of its two children. so that
w = w1 + w2 . Define T1 < T2 if w1 < w2 , One possible Huffman tree building
algorithm can be realized as the following.

build(A) =

T1
build({merge(Ta , Tb )} A0 )

: A = {T1 }
: otherwise

(14.80)

A is a list of trees. It is initialized as leaves for all symbols and their weights.
If there is only one tree in this list, we are done, the tree is the final Huffman
tree. Otherwise, The two smallest tree Ta and Tb are extracted, and the rest
trees are hold in list A0 . Ta and Tb are merged to one bigger tree, and put back
to the tree list for further recursive building.
(Ta , Tb , A0 ) = extract(A)

(14.81)

We can scan the tree list to extract the 2 nodes with the smallest weight. Below equation shows that when the scan begins, the first 2 elements are compared
and initialized as the two minimum ones. An empty accumulator is passed as
the last argument.
extract(A) = extract0 (min(T1 , T2 ), max(T1 , T2 ), {T3 , T4 , ...}, )

(14.82)

For every tree, if its weight is less than the smallest two weve ever found,
we update the result to contain this tree. For any given tree list A, denote the
first tree in it as T1 , and the rest trees except T1 as A0 . The scan process can
be defined as the following.

(Ta , Tb , B) : A =
extract0 (Ta0 , Tb0 , A0 , {Tb } A) : T1 < Tb

extract0 (Ta , Tb , A0 , {T1 } A) : otherwise

(14.83)
Where Ta0 = min(T1 , Ta ), Tb0 = max(T1 , Ta ) are the updated two trees with
the smallest weights.
The following Haskell example program implements this Huffman tree building algorithm.
extract0 (Ta , Tb , A, B) =

518

CHAPTER 14. SEARCHING

build [x] = x
build xs = build ((merge x y) : xs) where
(x, y, xs) = extract xs
extract (x:y:xs) = min2 (min x y) (max x y) xs [] where
min2 x y [] xs = (x, y, xs)
min2 x y (z:zs) xs | z < y = min2 (min z x) (max z x) zs (y:xs)
| otherwise = min2 x y zs (z:xs)

This building solution can also be realized imperatively. Given an array of

Huffman nodes, we can use the last two cells to hold the nodes with the smallest
weights. Then we scan the rest of the array from right to left. Whenever there
is a node with the smaller weight, this node will be exchanged with the bigger
one of the last two. After all nodes have been examined, we merge the trees in
the last two cells, and drop the last cell. This shrinks the array by one. We
repeat this process till there is only one tree left.
1: function Huffman(A)
2:
while |A| > 1 do
3:
n |A|
4:
for i n 2 down to 1 do
5:
if A[i] < Max(A[n], A[n 1]) then
6:
Exchange A[i] Max(A[n], A[n 1])
7:
A[n 1] Merge(A[n], A[n 1])
8:
Drop(A[n])
9:
return A[1]
The following C++ example program implements this algorithm. Note that
this algorithm neednt the last two elements being ordered.
typedef vector<Node> Nodes;
bool lessp(Node a, Node b) { return aw < bw; }
Node max(Node a, Node b) { return lessp(a, b) ? b : a; }
void swap(Nodes& ts, int i, int j, int k) {
swap(ts[i], ts[ts[j] < ts[k] ? k : j]);
}
Node huffman(Nodes ts) {
int n;
while((n = ts.size()) > 1) {
for (int i = n - 3; i 0; --i)
if (lessp(ts[i], max(ts[n-1], ts[n-2])))
swap(ts, i, n-1, n-2);
ts[n-2] = merge(ts[n-1], ts[n-2]);
ts.pop_back();
}
return ts.front();
}

The algorithm merges all the leaves, and it need scan the list in each iteration.
Thus the performance is quadratic. This algorithm can be improved. Observe
that each time, only the two trees with the smallest weights are merged. This

14.3. SOLUTION SEARCHING

519

reminds us the heap data structure. Heap ensures to access the smallest element
fast. We can put all the leaves in a heap. For binary heap, this is typically a
linear operation. Then we extract the minimum element twice, merge them,
then put the bigger tree back to the heap. This is O(lg n) operation if binary
heap is used. So the total performance is O(n lg n), which is better than the
above algorithm. The next algorithm extracts the node from the heap, and
starts Huffman tree building.
build(H) = reduce(top(H), pop(H))

(14.84)

This algorithm stops when the heap is empty; Otherwise, it extracts another
nodes from the heap for merging.

: H=
: otherwise
(14.85)
Function build and reduce are mutually recursive. The following Haskell
example program implements this algorithm by using heap defined in previous
chapter.
reduce(T, H) =

huffman
huffman
build
reduce
reduce

T
build(insert(merge(T, top(H)), pop(H)))

:: (Num a, Ord a) [(b, a)] HTr a b

= build Heap.fromList map ((c, w) Leaf w c) where
h = reduce (Heap.findMin h) (Heap.deleteMin h)
x Heap.E = x
x h = build $ Heap.insert (Heap.deleteMin h) (merge x (Heap.findMin h))

The heap solution can also be realized imperatively. The leaves are firstly
transformed to a heap, so that the one with the minimum weight is put on the
top. As far as there are more than 1 elements in the heap, we extract the two
smallest, merge them to a bigger one, and put back to the heap. The final tree
left in the heap is the result Huffman tree.
1: function Huffman(A)
2:
Build-Heap(A)
3:
while |A| > 1 do
4:
Ta Heap-Pop(A)
5:
Tb Heap-Pop(B)
6:
Heap-Push(A, Merge(Ta , Tb ))
7:
return Heap-Pop(A)
The following example C++ code implements this heap solution. The heap
used here is provided in the standard library. Because the max-heap, but not
min-heap would be made by default, a greater predication is explicitly passed
as argument.
bool greaterp(Node a, Node b) { return bw < aw; }
Node pop(Nodes& h) {
Node m = h.front();
pop_heap(h.begin(), h.end(), greaterp);
h.pop_back();
return m;
}

520

CHAPTER 14. SEARCHING

void push(Node t, Nodes& h) {

h.push_back(t);
push_heap(h.begin(), h.end(), greaterp);
}
Node huffman1(Nodes ts) {
make_heap(ts.begin(), ts.end(), greaterp);
while (ts.size() > 1) {
Node t1 = pop(ts);
Node t2 = pop(ts);
push(merge(t1, t2), ts);
}
return ts.front();
}

When the symbol-weight list has been already sorted, there exists a linear
time method to build the Huffman tree. Observe that during the Huffman tree
building, it produces a series of merged trees with weight in ascending order.
We can use a queue to manage the merged trees. Every time, we pick the two
trees with the smallest weight from both the queue and the list, merge them
and push the result to the queue. All the trees in the list will be processed, and
there will be only one tree left in the queue. This tree is the result Huffman
tree. This process starts by passing an empty queue as below.
build0 (A) = reduce0 (extract00 (, A))

(14.86)

Suppose A is in ascending order by weight, At any time, the tree with the
smallest weight is either the header of the queue, or the first element of the list.
Denote the header of the queue is Ta , after pops it, the queue is Q0 ; The first
element in A is Tb , the rest elements are hold in A0 . Function extract00 can be
defined like the following.

(Tb , (Q, A0 )) :
00
(Ta , (Q0 , A)) :
extract (Q, A) =

(Tb , (Q, A0 )) :

Q=
A = Ta < T b
otherwise

(14.87)

Actually, the pair of queue and tree list can be viewed as a special heap.
The tree with the minimum weight is continuously extracted and merged.
0
reduce
(T, (Q, A)) =

T : Q=A=
reduce0 (extract00 (push(Q00 , merge(T, T 0 )), A00 )) : otherwise
(14.88)
Where (T 0 , (Q00 , A00 )) = extract00 (Q, A), which means extracting another
tree. The following Haskell example program shows the implementation of this
method. Note that this program explicitly sort the leaves, which isnt necessary
if the leaves are ordered. Again, the list, but not a real queue is used here for
illustration purpose. List isnt good at pushing new element, please refer to the
chapter of queue for details about it.
huffman :: (Num a, Ord a) [(b, a)] HTr a b
huffman = reduce wrap sort map ((c, w) Leaf w c) where

14.3. SOLUTION SEARCHING

521

wrap xs = delMin ([], xs)

reduce (x, ([], [])) = x
reduce (x, h) = let (y, (q, xs)) = delMin h in
reduce $ delMin (q ++ [merge x y], xs)
delMin ([], (x:xs)) = (x, ([], xs))
delMin ((q:qs), []) = (q, (qs, []))
delMin ((q:qs), (x:xs)) | q < x = (q, (qs, (x:xs)))
| otherwise = (x, ((q:qs), xs))

This algorithm can also be realized imperatively.

1: function Huffman(A)
. A is ordered by weight
2:
Q
3:
T Extract(Q, A)
4:
while Q 6= A 6= do
5:
Push(Q, Merge(T , Extract(Q, A)))
6:
T Extract(Q, A)
7:
return T
Where function Extract(Q, A) extracts the tree with the smallest weight
from the queue and the list. It mutates the queue and tree if necessary. Denote
the head of the queue is Ta , and the first element of the list as Tb .
1: function Extract(Q, A)
2:
if Q 6= (A = Ta < Tb ) then
3:
return Pop(Q)
4:
else
5:
return Detach(A)
Where procedure Detach(A), removes the first element from A, and returns
this element as result. In most imperative settings, as detaching the first element
is slow linear operation for array, we can store the trees in descending order by
weight, and remove the last element. This is a fast constant time operation.
The below C++ example code shows this idea.
Node extract(queue<Node>& q, Nodes& ts) {
Node t;
if (!q.empty() && (ts.empty() | | lessp(q.front(), ts.back()))) {
t = q.front();
q.pop();
} else {
t = ts.back();
ts.pop_back();
}
return t;
}
Node huffman2(Nodes ts) {
queue<Node> q;
sort(ts.begin(), ts.end(), greaterp);
Node t = extract(q, ts);
while (!q.empty() | | !ts.empty()) {
q.push(merge(t, extract(q, ts)));
t = extract(q, ts);
}
return t;

522

CHAPTER 14. SEARCHING

Note that the sorting isnt necessary if the trees have already been ordered.
It can be a linear time reversing in case the trees are in ascending order by
weight.
There are three different Huffman man tree building methods explained.
Although they follow the same approach developed by Huffman, the result trees
varies. Figure 14.48 shows the three different Huffman trees built with these
methods.

A, 2

N, 3

L, 1

T, 2

E, 1

I, 2

O, 1

N, 3

O, 1

R, 1

T, 2

I, 2

R, 1

A, 2

E, 1

(a) Created by scan method.

L, 1

(b) Created by heap method.

O, 1

N, 3

R, 1

A, 2

I, 2

T, 2

E, 1

L, 1

(c) Linear time building for sorted list.

Figure 14.48: Variation of Huffman trees for the same symbol list.
Although these three trees are not identical. They are all able to generate the
most efficient code. The formal proof is skipped here. The detailed information
can be referred to [15] and Section 16.3 of [2].
The Huffman tree building is the core idea of Huffman coding. Many things
can be easily achieved with the Huffman tree. For example, the code table can
be generated by traversing the tree. We start from the root with the empty
prefix p. For any branches, we append a zero to the prefix if turn left, and
append a one if turn right. When a leaf node is arrived, the symbol represented
by this node and the prefix are put to the code table. Denote the symbol of a
leaf node as c, the children for tree T as Tl and Tr respectively. The code table
association list can be built with code(T, ), which is defined as below.

code(T, p) =

{(c, p)}
code(Tl , p {0}) code(Tr , p {1})

: leaf (T )
: otherwise

(14.89)

Where function leaf (T ) tests if tree T is a leaf or a branch node. The

14.3. SOLUTION SEARCHING

523

following Haskell example program generates a map as the code table according
to this algorithm.
code tr = Map.fromList $ traverse [] tr where
traverse bits (Leaf _ c) = [(c, bits)]
traverse bits (Branch _ l r) = (traverse (bits ++ [0]) l) ++
(traverse (bits ++ [1]) r)

The imperative code table generating algorithm is left as exercise. The

encoding process can scan the text, and look up the code table to output the
bit sequence. The realization is skipped here.
The decoding process is realized by looking up the Huffman tree according
to the bit sequence. We start from the root, whenever a zero is received, we turn
left, otherwise if a one is received, we turn right. If a leaf node is arrived, the
symbol represented by this leaf is output, and we start another looking up from
the root. The decoding process ends when all the bits are consumed. Denote
the bit sequence as B = {b1 , b2 , ...}, all bits except the first one are hold in B 0 ,
below definition realizes the decoding algorithm.

{c} : B = leaf (T )
{c} decode(root(T ), B) : leaf (T )
decode(T, B) =
decode(Tl , B 0 ) : b1 = 0

decode(Tr , B 0 ) : otherwise

(14.90)

Where root(T ) returns the root of the Huffman tree. The following Haskell
example code implements this algorithm.
decode
find
find
find

tr cs =
(Leaf _
(Leaf _
(Branch

find tr cs where
c) [] = [c]
c) bs = c : find tr bs
_ l r) (b:bs) = find (if b == 0 then l else r) bs

Note that this is an on-line decoding algorithm with linear time performance.
It consumes one bit per time. This can be clearly noted from the below imperative realization, where the index keeps increasing by one.
1: function Decode(T, B)
2:
W
3:
n |B|, i 1
4:
while i < n do
5:
RT
6:
while Leaf(R) do
7:
if B[i] = 0 then
8:
R Left(R)
9:
else
10:
R Right(R)
11:
ii+1
12:
W W Symbol(R)
13:
return W
This imperative algorithm can be implemented as the following example
C++ program.
string decode(Node root, const char bits) {

524

CHAPTER 14. SEARCHING

string w;
while (bits) {
Node t = root;
while (!isleaf(t))
t = 0 == bits++ ? tleft : tright;
w += tc;
}
return w;

Huffman coding, especially the Huffman tree building shows an interesting

strategy. Each time, there are multiple options for merging. Among the trees
in the list, Huffman method always selects two trees with the smallest weight.
This is the best choice at that merge stage. However, these series of local best
options generate a global optimal prefix code.
Its not always the case that the local optimal choice also leads to the global
optimal solution. In most cases, it doesnt. Huffman coding is a special one. We
call the strategy that always choosing the local best option as greedy strategy.
Greedy method works for many problems. However, its not easy to tell
if the greedy method can be applied to get the global optimal solution. The
generic formal proof is still an active research area. Section 16.4 in [2] provides
a good treatment for Matroid tool, which covers many problems that greedy
algorithm can be applied.
Change-making problem
We often change money when visiting other countries. People tend to use credit
card more often nowadays than before, because its quite convenient to buy
things without considering much about changes. If we changed some money in
the bank, there are often some foreign money left by the end of the trip. Some
people like to change them to coins for collection. Can we find a solution, which
can change the given amount of money with the least number of coins?
Lets use USA coin system for example. There are 5 different coins: 1 cent,
5 cent, 25 cent, 50 cent, and 1 dollar. A dollar is equal to 100 cents. Using the
greedy method introduced above, we can always pick the largest coin which is
not greater than the remaining amount of money to be changed. Denote list
C = {1, 5, 25, 50, 100}, which stands for the value of coins. For any given money
X, the change coins can be generated as below.

X=0
otherwise,
{cm } change(X cm , C) :
cm = max({c C, c X})
(14.91)
If C is in descending order, cm can be found as the first one not greater
than X. If we want to change 1.42 dollar, This function produces a coin list of
{100, 25, 5, 5, 5, 1, 1}. The output coins list can be easily transformed to contain
pairs {(100, 1), (25, 1), (5, 3), (1, 2)}. That we need one dollar, a quarter, three
coins of 5 cent, and 2 coins of 1 cent to make the change. The following Haskell
example program outputs result as such.
change(X, C) =

solve x = assoc change x where

14.3. SOLUTION SEARCHING

525

change 0 _ = []
change x cs = let c = head $ filter ( x) cs in c : change (x - c) cs
assoc = (map (cs (head cs, length cs))) group

As mentioned above, this program assumes the coins are in descending order,
for instance like below.
solve 142 [100, 50, 25, 5, 1]

This algorithm is tail recursive, it can be transformed to a imperative looping.

1: function Change(X, C)
2:
R
3:
while X 6= 0 do
4:
cm = max({c C, c X})
5:
R {cm } R
6:
X X cm
7:
return R
The following example Python program implements this imperative version
and manages the result with a dictionary.
def change(x, coins):
cs = {}
while x != 0:
m = max([c for c in coins if c x])
cs[m] = 1 + cs.setdefault(m, 0)
x=x - m
return cs

For a coin system like USA, the greedy approach can find the optimal solution. The amount of coins is the minimum. Fortunately, our greedy method
works in most countries. But it is not always true. For example, suppose a
country have coins of value 1, 3, and 4 units. The best change for value 6, is
to use two coins of 3 units, however, the greedy method gives a result of three
coins: one coin of 4, two coins of 1. Which isnt the optimal result.
Summary of greedy method
As shown in the change making problem, greedy method doesnt always give the
best result. In order to find the optimal solution, we need dynamic programming
which will be introduced in the next section.
However, the result is often good enough in practice. Lets take the wordwrap problem for example. In modern software editors and browsers, text spans
to multiple lines if the length of the content is too long to be hold. With wordwrap supported, user neednt hard line breaking. Although dynamic programming can wrap with the minimum number of lines, its overkill. On the contrary,
greedy algorithm can wrap with lines approximate to the optimal result with
quite effective realization as below. Here it wraps text T , not to exceeds line
width W , with space s between each word.
1: L W
2: for w T do
3:
if |w| + s > L then

526
4:
5:
6:
7:

CHAPTER 14. SEARCHING

Insert line break
L W |w|
else
L L |w| s

For each word w in the text, it uses a greedy strategy to put as many words
in a line as possible unless it exceeds the line width. Many word processors use
a similar algorithm to do word-wrapping.
There are many cases, the strict optimal result, but not the approximate
one is necessary. Dynamic programming can help to solve such problems.
Dynamic programming
In the change-making problem, we mentioned the greedy method cant always
give the optimal solution. For any coin system, are there any way to find the
best changes?
Suppose we have find the best solution which makes X value of money.
The coins needed are contained in Cm . We can partition these coins into two
collections, C1 and C2 . They make money of X1 , and X2 respectively. Well
prove that C1 is the optimal solution for X1 , and C2 is the optimal solution for
X2 .
Proof. For X1 , Suppose there exists another solution C10 , which uses less coins
than C1 . Then changing solution C10 C2 uses less coins to make X than Cm .
This is conflict with the fact that Cm is the optimal solution to X. Similarity,
we can prove C2 is the optimal solution to X2 .
Note that it is not true in the reverse situation. If we arbitrary select a
value Y < X, divide the original problem to find the optimal solutions for sub
problems Y and X Y . Combine the two optimal solutions doesnt necessarily
yield optimal solution for X. Consider this example. There are coins with value
1, 2, and 4. The optimal solution for making value 6, is to use 2 coins of value
2, and 4; However, if we divide 6 = 3 + 3, since each 3 can be made with optimal
solution 3 = 1 + 2, the combined solution contains 4 coins (1 + 1 + 2 + 2).
If an optimal problem can be divided into several sub optimal problems, we
call it has optimal substructure. We see that the change-making problem has
optimal substructure. But the dividing has to be done based on the coins, but
not with an arbitrary value.
The optimal substructure can be expressed recursively as the following.

: X=0
least({c change(X c)|c C, c X}) : otherwise
(14.92)
For any coin system C, the changing result for zero is empty; otherwise, we
check every candidate coin c, which is not greater then value X, and recursively
find the best solution for X c; We pick the coin collection which contains the
least coins as the result.
Below Haskell example program implements this top-down recursive solution.
change(X) =

14.3. SOLUTION SEARCHING

527

change _ 0 = []
change cs x = minimumBy (compare on length)
[c:change cs (x - c) | c cs, c x]

Although this program outputs correct answer [2, 4] when evaluates change
[1, 2, 4] 6, it performs very bad when changing 1.42 dollar with USA coins
system. It failed to find the answer within 15 minutes in a computer with
2.7GHz CPU and 8G memory.
The reason why its slow is because there are a lot of duplicated computing
in the top-down recursive solution. When it computes change(142), it needs to
examine change(141), change(137), change(117), change(92), and change(42).
While change(141) next computes to smaller values by deducing with 1, 2, 25,
50 and 100 cents. it will eventually meets value 137, 117, 92, and 42 again. The
search space explodes with power of 5.
This is quite similar to compute Fibonacci numbers in a top-down recursive
way.

1 : n=1n=2
Fn =
(14.93)
Fn1 + Fn2 : otherwise
When we calculate F8 for example, we recursively calculate F7 and F6 . While
when we calculate F7 , we need calculate F6 again, and F5 , ... As shown in the
below expand forms, the calculation is doubled every time, and the same value
is calculate again and again.
F8 =
=
=
=

F7 + F6
F6 + F5 + F5 + F4
F5 + F4 + F4 + F3 + F4 + F3 + F3 + F2
...

In order to avoid duplicated computation, a table F can be maintained when

calculating the Fibonacci numbers. The first two elements are filled as 1, all
others are left blank. During the top-down recursive calculation, If need Fk ,
we first look up this table for the k-th cell, if it isnt blank, we use that value
directly. Otherwise we need further calculation. Whenever a value is calculated,
we store it in the corresponding cell for looking up in the future.
1: F {1, 1, N IL, N IL, ...}
2: function Fibonacci(n)
3:
if n > 2 F [n] = N IL then
4:
F [n] Fibonacci(n 1) + Fibonacci(n 2)
5:
return F [n]
By using the similar idea, we can develop a new top-down change-making
solution. We use a table T to maintain the best changes, it is initialized to all
empty coin list. During the top-down recursive computation, we look up this
table for smaller changing values. Whenever a intermediate value is calculated,
it is stored in the table.
1: T {, , ...}
2: function Change(X)
3:
if X > 0 T [X] = then
4:
for c C do
5:
if c X then

528
6:
7:
8:
9:

CHAPTER 14. SEARCHING

Cm {c} Change(X c)
if T [X] = |Cm | < |T [X]| then
T [X] Cm
return T [X]

The solution to change 0 money is definitely empty , otherwise, we look

up T [X] to retrieve the solution to change X money. If it is empty, we need
recursively calculate it. We examine all coins in the coin system C which is not
greater than X. This is the sub problem of making changes for money X c.
The minimum amount of coins plus one coin of c is stored in T [X] finally as the
result.
The following example Python program implements this algorithm just takes
8000 ms to give the answer of changing 1.42 dollar in US coin system.
tab = [[] for _ in range(1000)]
def change(x, cs):
if x > 0 and tab[x] == []:
for s in [[c] + change(x - c, cs) for c in cs if c x]:
if tab[x] == [] or len(s) < len(tab[x]):
tab[x] = s
return tab[x]

Another solution to calculate Fibonacci number, is to compute them in order

of F1 , F2 , F3 , ..., Fn . This is quite natural when people write down Fibonacci
series.
1: function Fibo(n)
2:
F = {1, 1, N IL, N IL, ...}
3:
for i 3 to n do
4:
F [i] F [i 1] + F [i 2]
5:
return F [n]
We can use the quite similar idea to solve the change making problem. Starts
from zero money, which can be changed by an empty list of coins, we next try
to figure out how to change money of value 1. In US coin system for example,
A cent can be used; The next values of 2, 3, and 4, can be changed by two coins
of 1 cent, three coins of 1 cent, and 4 coins of 1 cent. At this stage, the solution
table looks like below
0
1
2
3
4
{1} {1, 1} {1, 1, 1} {1, 1, 1, 1}
The interesting case happens for changing value 5. There are two options,
use another coin of 1 cent, which need 5 coins in total; The other way is to use
1 coin of 5 cent, which uses less coins than the former. So the solution table
can be extended to this.
0
1
2
3
4
5
{1} {1, 1} {1, 1, 1} {1, 1, 1, 1} {5}
For the next change value 6, since there are two types of coin, 1 cent and 5
cent, are less than this value, we need examine both of them.
If we choose the 1 cent coin, we need next make changes for 5; Since weve
already known that the best solution to change 5 is {5}, which only needs
a coin of 5 cents, by looking up the solution table, we have one candidate
solution to change 6 as {5, 1};

14.3. SOLUTION SEARCHING

529

The other option is to choose the 5 cent coin, we need next make changes
for 1; By looking up the solution table weve filled so far, the sub optimal
solution to change 1 is {1}. Thus we get another candidate solution to
change 6 as {1, 5};
It happens that, both options yield a solution of two coins, we can select
either of them as the best solution. Generally speaking, the candidate with
fewest number of coins is selected as the solution, and filled into the table.
At any iteration, when we are trying to change the i < X value of money,
we examine all the types of coin. For any coin c not greater than i, we look
up the solution table to fetch the sub solution T [i c]. The number of coins
in this sub solution plus the one coin of c are the total coins needed in this
candidate solution. The fewest candidate is then selected and updated to the
solution table.
The following algorithm realizes this bottom-up idea.
1: function Change(X)
2:
T {, , ...}
3:
for i 1 to X do
4:
for c C, c i do
5:
if T [i] = 1 + |T [i c]| < |T [i]| then
6:
T [i] {c} T [i c]
7:
return T [X]
This algorithm can be directly translated to imperative programs, like Python
for example.
def changemk(x, cs):
s = [[] for _ in range(x+1)]
for i in range(1, x+1):
for c in cs:
if c i and (s[i] == [] or 1 + len(s[i-c]) < len(s[i])):
s[i] = [c] + s[i-c]
return s[x]

Observe the solution table, its easy to find that, there are many duplicated
contents being stored.
6
7
8
9
10
...
{1, 5} {1, 1, 5} {1, 1, 1, 5} {1, 1, 1, 1, 5} {5, 5} ...
This is because the optimal sub solutions are completely copied and saved
in parent solution. In order to use less space, we can only record the delta
part from the sub optimal solution. In change-making problem, it means that
we only need to record the coin being selected for value i.
1: function Change(X)
2:
T {0, , , ...}
3:
S {N IL, N IL, ...}
4:
for i 1 to X do
5:
for c C, c i do
6:
if 1 + T [i c] < T [i] then
7:
T [i] 1 + T [i c]
8:
S[i] c
9:
while X > 0 do
10:
Print(S[X])

530
11:

CHAPTER 14. SEARCHING

X X S[X]

Instead of recording the complete solution list of coins, this new algorithm
uses two tables T and S. T holds the minimum number of coins needed for
changing value 0, 1, 2, ...; while S holds the first coin being selected for the
optimal solution. For the complete coin list to change money X, the first coin
is thus S[X], the sub optimal solution is to change money X 0 = X S[X]. We
can look up table S[X 0 ] for the next coin. The coins for sub optimal solutions
are repeatedly looked up like this till the beginning of the table. Below Python
example program implements this algorithm.
def chgmk(x, cs):
cnt = [0] + [x+1] x
s = [0]
for i in range(1, x+1):
coin = 0
for c in cs:
if c i and 1 + cnt[i-c] < cnt[i]:
cnt[i] = 1 + cnt[i-c]
coin = c
s.append(coin)
r = []
while x > 0:
r.append(s[x])
x = x - s[x]
return r

This change-making solution loops n times for given money n. It examines

at most the full coin system in each iteration. The time is bound to (nk)
where k is the number of coins for a certain coin system. The last algorithm
adds O(n) spaces to record sub optimal solutions with table T and S.
In purely functional settings, There is no means to mutate the solution
table and look up in constant time. One alternative is to use finger tree as we
mentioned in previous chapter 11 . We can store the minimum number of coins,
and the coin leads to the sub optimal solution in pairs.
The solution table, which is a finger tree, is initialized as T = {(0, 0)}. It
means change 0 money need no coin. We can fold on list {1, 2, ..., X}, start
from this table, with a binary function change(T, i). The folding will build the
solution table, and we can construct the coin list from this table by function
make(X, T ).
makeChange(X) = make(X, f old(change, {(0, 0)}, {1, 2, ..., X}))

(14.94)

In function change(T, i), all the coins not greater than i are examined to
select the one lead to the best result. The fewest number of coins, and the coin
being selected are formed to a pair. This pair is inserted to the finger tree, so
that a new solution table is returned.
change(T, i) = insert(T, f old(sel, (, 0), {c|c C, c i}))

(14.95)

11 Some purely functional programming environments, Haskell for instance, provide built-in
array; while other almost pure ones, such as ML, provide mutable array

14.3. SOLUTION SEARCHING

531

Again, folding is used to select the candidate with the minimum number of
coins. This folding starts with initial value (, 0), on all valid coins. function
sel((n, c), c0 ) accepts two arguments, one is a pair of length and coin, which
is the best solution so far; the other is a candidate coin, it examines if this
candidate can make better solution.

sel((n, c), c ) =

(1 + n0 , c0 ) : 1 + n0 < n, (n0 , c0 ) = T [i c0 ]
(n, c) : otherwise

(14.96)

After the solution table is built, the coins needed can be generated from it.

: X=0
{c} make(X c, T ) : otherwise, (n, c) = T [X]
(14.97)
The following example Haskell program uses Data.Sequence, which is the
library of finger tree, to implement change making solution.
make(X, T ) =

import Data.Sequence (Seq, singleton, index, ( |>))

changemk x cs = makeChange x $ foldl change (singleton (0, 0)) [1..x] where
change tab i = let sel c = min (1 + fst (index tab (i - c)), c)
in tab |> (foldr sel ((x + 1), 0) $ filter ( i) cs)
makeChange 0 _ = []
makeChange x tab = let c = snd $ index tab x in c : makeChange (x - c) tab

Its necessary to memorize the optimal solution to sub problems no matter

using the top-down or the bottom-up approach. This is because a sub problem is
used many times when computing the overall optimal solution. Such properties
are called overlapping sub problems.
Properties of dynamic programming
Dynamic programming was originally named by Richard Bellman in 1940s. It is
a powerful tool to search for optimal solution for problems with two properties.
Optimal sub structure. The problem can be broken down into smaller
problems, and the optimal solution can be constructed efficiently from
solutions of these sub problems;
Overlapping sub problems. The problem can be broken down into sub
problems which are reused several times in finding the overall solution.
The change-making problem, as weve explained, has both optimal sub structures, and overlapping sub problems.
Longest common subsequence problem
The longest common subsequence problem, is different with the longest common substring problem. Weve show how to solve the later in the chapter of
suffix tree. The longest common subsequence neednt be consecutive part of the
original sequence.

532

CHAPTER 14. SEARCHING

Figure 14.49: The longest common subsequence

For example, The longest common substring for text Mississippi, and
Missunderstanding is Miss, while the longest common subsequence for them
are Misssi. This is shown in figure 14.49.
If we rotate the figure vertically, and consider the two texts as two pieces of
source code, it turns to be a diff result between them. Most modern version
control tools need calculate the difference content among the versions. The
longest common subsequence problem plays a very important role.
If either one of the two strings X and Y is empty, the longest common subsequence LCS(X, Y ) is definitely empty; Otherwise, denote X = {x1 , x2 , ..., xn },
Y = {y1 , y2 , ..., ym }, if the first elements x1 and y1 are same, we can recursively
find the longest subsequence of X 0 = {x2 , x3 , ..., xn } and Y 0 = {y2 , y3 , ..., ym }.
And the final result LCS(X, Y ) can be constructed by concatenating x1 with
LCS(X 0 , Y 0 ); Otherwise if x1 6= y1 , we need recursively find the longest common subsequences of LCS(X, Y 0 ) and LCS(X 0 , Y ), and pick the longer one as
the final result. Summarize these cases gives the below definition.

X =Y =
x1 = y1
otherwise
(14.98)
Note that this algorithm shows clearly the optimal substructure, that the
longest common subsequence problem can be broken to smaller problems. The
sub problem is ensured to be at least one element shorter than the original one.
Its also clear that, there are overlapping sub-problems. The longest common
subsequences to the sub strings are used multiple times in finding the overall
optimal solution.
The existence of these two properties, the optimal substructure and the
overlapping sub-problem, indicates the dynamic programming can be used to
LCS(X, Y ) =

:
{x1 } LCS(X 0 , Y 0 ) :

longer(LCS(X, Y 0 ), LCS(X 0 , Y )) :

14.3. SOLUTION SEARCHING

533

solve this problem.

A 2-dimension table can be used to record the solutions to the sub-problems.
The rows and columns represent the substrings of X and Y respectively.
a n t e n n a
1 2 3 4 5 6 7
b 1
a 2
n 3
a 4
n 5
a 6
This table shows an example of finding the longest common subsequence for
strings antenna and banana. Their lengths are 7, and 6. The right bottom
corner of this table is looked up first, Since its empty we need compare the
7th element in antenna and the 6th in banana, they are both a, Thus we
need next recursively look up the cell at row 5, column 6; Its still empty, and
we repeated this till either get a trivial case that one substring becomes empty,
or some cell we are looking up has been filled before. Similar to the changemaking problem, whenever the optimal solution for a sub-problem is found, it is
recorded in the cell for further reusing. Note that this process is in the reversed
order comparing to the recursive equation given above, that we start from the
right most element of each string.
Considering that the longest common subsequence for any empty string is
still empty, we can extended the solution table so that the first row and column
hold the empty strings.
a n t
e n n a

b
a
n
a
n
a
Below algorithm realizes the top-down recursive dynamic programming solution with such a table.
1: T NIL
2: function LCS(X, Y )
3:
m |X|, n |Y |
4:
m0 m + 1, n0 n + 1
5:
if T = NIL then
6:
T {{, , ..., }, {, N IL, N IL, ...}, ...}
. m 0 n0
7:
if X 6= Y =
6 T [m0 ][n0 ] = NIL then
8:
if X[m] = Y [n] then
9:
T [m0 ][n0 ] Append(LCS(X[1..m 1], Y [1..n 1]), X[m])
10:
else
11:
T [m0 ][n0 ] Longer(LCS(X, Y [1..n 1]), LCS(X[1..m 1], Y ))
12:
return T [m0 ][n0 ]
The table is firstly initialized with the first row and column filled with empty

534

CHAPTER 14. SEARCHING

strings; the rest are all NIL values. Unless either string is empty, or the cell
content isnt NIL, the last two elements of the strings are compared, and recursively computes the longest common subsequence with substrings. The following
Python example program implements this algorithm.
def lcs(xs, ys):
m = len(xs)
n = len(ys)
global tab
if tab is None:
tab = [[""](n+1)] + [[""] + [None]n for _ in xrange(m)]
if m != 0 and n !=0 and tab[m][n] is None:
if xs[-1] == ys[-1]:
tab[m][n] = lcs(xs[:-1], ys[:-1]) + xs[-1]
else:
(a, b) = (lcs(xs, ys[:-1]), lcs(xs[:-1], ys))
tab[m][n] = a if len(b) < len(a) else b
return tab[m][n]

The longest common subsequence can also be found in a bottom-up manner

as what weve done with the change-making problem. Besides that, instead of
recording the whole sequences in the table, we can just store the lengths of the
longest subsequences, and later construct the subsubsequence with this table
and the two strings. This time, the table is initialized with all values set as 0.
1: function LCS(X, Y )
2:
m |X|, n |Y |
3:
T {{0, 0, ...}, {0, 0, ...}, ...}
. (m + 1) (n + 1)
4:
for i 1 to m do
5:
for j 1 to n do
6:
if X[i] = Y [j] then
7:
T [i + 1][j + 1] T [i][j] + 1
8:
else
9:
T [i + 1][j + 1] Max(T [i][j + 1], T [i + 1][j])
10:
return Get(T, X, Y, m, n)
11:
12:
13:
14:
15:
16:
17:
18:
19:

function Get(T, X, Y, i, j)
if i = 0 j = 0 then
return
else if X[i] = Y [j] then
return Append(Get(T, X, Y, i 1, j 1), X[i])
else if T [i 1][j] > T [i][j 1] then
return Get(T, X, Y, i 1, j)
else
return Get(T, X, Y, i, j 1)

In the bottom-up approach, we start from the cell at the second row and
the second column. The cell is corresponding to the first element in both X,
and Y . If they are same, the length of the longest common subsequence so far
is 1. This can be yielded by increasing the length of empty sequence, which is
stored in the top-left cell, by one; Otherwise, we pick the maximum value from
the upper cell and left cell. The table is repeatedly filled in this manner.
After that, a back-track is performed to construct the longest common sub-

14.3. SOLUTION SEARCHING

535

sequence. This time we start from the bottom-right corner of the table. If the
last elements in X and Y are same, we put this element as the last one of the
result, and go on looking up the cell along the diagonal line; Otherwise, we
compare the values in the left cell and the right cell, and go on looking up the
cell with the bigger value.
The following example Python program implements this algorithm.
def lcs(xs, ys):
m = len(xs)
n = len(ys)
c = [[0](n+1) for _ in xrange(m+1)]
for i in xrange(1, m+1):
for j in xrange(1, n+1):
if xs[i-1] == ys[j-1]:
c[i][j] = c[i-1][j-1] + 1
else:
c[i][j] = max(c[i-1][j], c[i][j-1])
return get(c, xs, ys, m, n)
def get(c, xs, ys, i, j):
if i==0 or j==0:
return []
elif xs[i-1] == ys[j-1]:
return get(c, xs, ys, i-1, j-1) + [xs[i-1]]
elif c[i-1][j] > c[i][j-1]:
return get(c, xs, ys, i-1, j)
else:
return get(c, xs, ys, i, j-1)

The bottom-up dynamic programming solution can also be defined in purely

functional way. The finger tree can be used as a table. The first row is filled
with n + 1 zero values. This table can be built by folding on sequence X. Then
the longest common subsequence is constructed from the table.
LCS(X, Y ) = construct(f old(f, {{0, 0, ..., 0}}, zip({1, 2, ...}, X)))

(14.99)

Note that, since the table need be looked up by index, X is zipped with
natural numbers. Function f creates a new row of this table by folding on
sequence Y , and records the lengths of the longest common sequence for all
possible cases so far.
f (T, (i, x)) = insert(T, f old(longest, {0}, zip({1, 2, ...}, Y )))

(14.100)

Function longest takes the intermediate filled row result, and a pair of index
and element in Y , it compares if this element is the same as the one in X. Then
fills the new cell with the length of the longest one.

longest(R, (j, y)) =

insert(R, 1 + T [i 1][j 1]) : x = y

insert(R, max(T [i 1][j], T [i][j 1])) : otherwise
(14.101)

536

CHAPTER 14. SEARCHING

After the table is built. The longest common sub sequence can be constructed recursively by looking up this table. We can pass the reversed sequences

X , and Y together with their lengths m and n for efficient building.

construct(T ) = get(( X , m), ( Y , n))

(14.102)

If the sequences are not empty, denote the first elements as x and y. The rest

elements are hold in X 0 and Y 0 respectively. The function get can be defined
as the following.

: X = Y =

get(( X 0 , i 1), ( Y 0 , j 1)) {x} : x = y

get(( X , i), ( Y , j)) =

get(( X 0 , i 1), ( Y , j)) : T [i 1][j] > T [i][j 1]

get(( X , i), ( Y , j 1)) : otherwise

(14.103)
Below Haskell example program implements this solution.
lcs xs ys = construct $ foldl f (singleton $ fromList $ replicate (n+1) 0)
(zip [1..] xs) where
(m, n) = (length xs, length ys)
f tab (i, x) = tab |> (foldl longer (singleton 0) (zip [1..] ys)) where
longer r (j, y) = r |> if x == y
then 1 + (tab index (i-1) index (j-1))
else max (tab index (i-1) index j) (r index (j-1))
construct tab = get (reverse xs, m) (reverse ys, n) where
get ([], 0) ([], 0) = []
get ((x:xs), i) ((y:ys), j)
| x == y = get (xs, i-1) (ys, j-1) ++ [x]
| (tab index (i-1) index j) > (tab index i index (j-1)) =
get (xs, i-1) ((y:ys), j)
| otherwise = get ((x:xs), i) (ys, j-1)

Subset sum problem

Dynamic programming does not limit to solve the optimization problem, but can
also solve some more general searching problems. Subset sum problem is such an
example. Given a set of integers, is there a non-empty subset sums to zero? for
example, there are two subsets of {11, 64, 82, 68, 86, 55, 88, 21, 51} both
sum to zero. One is {64, 82, 55, 88, 51}, the other is {64, 82, 68, 86}.
Of course summing to zero is a special case, because sometimes, people want
to find a subset, whose sum is a given value s. Here we are going to develop a
method to find all the candidate subsets.
There is obvious a brute-force exhausting search solution. For every element,
we can either pick it or not. So there are total 2n options for set with n elements.
Because for every selection, we need check if it sums to s. This is a linear
operation. The overall complexity is bound to O(n2n ). This is the exponential
algorithm, which takes very huge time if the set is big.
There is a recursive solution to subset sum problem. If the set is empty, there
is no solution definitely; Otherwise, let the set is X = {x1 , x2 , ...}. If x1 = s, then
subset {x1 } is a solution, we need next search for subsets X 0 = {x2 , x3 , ...} for

14.3. SOLUTION SEARCHING

537

those sum to s; Otherwise if x1 6= s, there are two different kinds of possibilities.

We need search X 0 for both sum s, and sum s x1 . For any subset sum to
s x1 , we can add x1 to it to form a new set as a solution. The following
equation defines this algorithm.

: X=
{{x1 }} solve(X 0 , s) : x1 = s

solve(X 0 , s) {{x1 } S|S solve(X 0 , s x1 )} : otherwise

(14.104)
There are clear substructures in this definition, although they are not in a
sense of optimal. And there are also overlapping sub-problems. This indicates
the problem can be solved with dynamic programming with a table to memorize
the solutions to sub-problems.
Instead of developing a solution to output all the subsets directly, lets consider how to give the existence answer firstly. That output yes if there exists
some subset sum to s, and no otherwise.
One fact is that, the upper and lower limit for all possible answer can be
calculated in one scan. If the given sum s doesnt belong to this range, there is
no solution obviously.

P
sl = P{x X, x < 0}
(14.105)
su = {x X, x > 0}

solve(X, s) =

Otherwise, if sl s su , since the values are all integers, we can use a

table, with su sl + 1 columns, each column represents a possible value in this
range, from sl to su . The value of the cell is either true or false to represents
if there exists subset sum to this value. All cells are initialized as false. Starts
from the first element x1 in X, definitely, set {x1 } can sum to x1 , so that the
cell represents this value in the first row can be filled as true.
sl sl + 1 ... x1 ... su
x1 F
F
... T ... F
With the next element x2 , There are three possible sums. Similar as the
first row, {x2 } sums to x2 ; For all possible sums in previous row, they can also
been achieved without x2 . So the cell below to x1 should also be filled as true;
By adding x2 to all possible sums so far, we can also get some new values. That
the cell represents x1 + x2 should be true.
sl sl + 1 ... x1 ... x2 ... x1 + x2 ... su
x1 F
F
... T ... F ...
F
... F
x2 F
F
... T ... T ...
T
... F
Generally speaking, when fill the i-th row, all the possible sums constructed
with {x1 , x2 , ..., xi1 so far can also be achieved with xi . So the cells previously
are true should also be true in this new row. The cell represents value xi should
also be true since the singleton set {xi } sums to it. And we can also adds xi
to all previously constructed sums to get the new results. Cells represent these
new sums should also be filled as true.
When all the elements are processed like this, a table with |X| rows is built.
Looking up the cell represents s in the last row tells if there exists subset can
sum to this value. As mentioned above, there is no solution if s < sl or su < s.
We skip handling this case for the sake of brevity.
1: function Subset-Sum(X, s)

538
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:

CHAPTER 14. SEARCHING

P
sl P{x X, x < 0}
su {x X, x > 0}
n |X|
T {{F alse, F alse, ...}, {F alse, F alse, ...}, ...}
for i 1 to n do
for j sl to su do
if X[i] = j then
T [i][j] T rue
if i > 1 then
T [i][j] T [i][j] T [i 1][j]
j 0 j X[i]
if sl j 0 su then
T [i][j] T [i][j] T [i 1][j 0 ]
return T [n][s]

. n (su sl + 1)

Note that the index to the columns of the table, doesnt range from 1 to su
sl +1, but maps directly from sl to su . Because most programming environments
dont support negative index, this can be dealt with T [i][j sl ]. The following
example Python program utilizes the property of negative indexing.
def solve(xs, s):
low = sum([x for x in xs if x < 0])
up = sum([x for x in xs if x > 0])
tab = [[False](up-low+1) for _ in xs]
for i in xrange(0, len(xs)):
for j in xrange(low, up+1):
tab[i][j] = (xs[i] == j)
j1 = j - xs[i];
tab[i][j] = tab[i][j] or tab[i-1][j] or
(low j1 and j1 up and tab[i-1][j1])
return tab[-1][s]

Note that this program doesnt use different branches for i = 0 and i =
1, 2, ..., n 1. This is because when i = 0, the row index to i 1 = 1 refers to
the last row in the table, which are all false. This simplifies the logic one more
step.
With this table built, its easy to construct all subsets sum to s. The method
is to look up the last row for cell represents s. If the last element xn = s, then
{xn } definitely is a candidate. We next look up the previous row for s, and
recursively construct all the possible subsets sum to s with {x1 , x2 , x3 , ..., xn1 }.
Finally, we look up the second last row for cell represents s xn . And for every
subset sums to this value, we add element xn to construct a new subset, which
sums to s.
1: function Get(X, s, T, n)
2:
S
3:
if X[n] = s then
4:
S S {X[n]}
5:
if n > 1 then
6:
if T [n 1][s] then
7:
S S Get(X, s, T, n 1)
8:
if T [n 1][s X[n]] then

14.3. SOLUTION SEARCHING

9:
10:

539

S S {{X[n]} S 0 |S 0 Get(X, s X[n], T, n 1) }

return S
The following Python example program translates this algorithm.

def get(xs, s, tab, n):

r = []
if xs[n] == s:
r.append([xs[n]])
if n > 0:
if tab[n-1][s]:
r = r + get(xs, s, tab, n-1)
if tab[n-1][s - xs[n]]:
r = r + [[xs[n]] + ys for ys in get(xs, s - xs[n], tab, n-1)]
return r

This dynamic programming solution to subset sum problem loops O(n(su

sl + 1)) times to build the table, and recursively uses O(n) time to construct the
final solution from this table. The space it used is also bound to O(n(su sl +1)).
Instead of using table with n rows, a vector can be used alternatively. For
every cell represents a possible sum, the list of subsets are stored. This vector
is initialized to contain all empty sets. For every element in X, we update the
vector, so that it records all the possible sums which can be built so far. When
all the elements are considered, the cell corresponding to s contains the final
result.
1: function Subset-Sum(X, s)
P
2:
sl P{x X, x < 0}
3:
su {x X, x > 0}
4:
T {, , ...}
. su sl + 1
5:
for x X do
6:
T 0 Duplicate(T )
7:
for j sl to su do
8:
j0 j x
9:
if x = j then
10:
T 0 [j] T 0 [j] {x}
11:
if sl j 0 su T [j 0 ] 6= then
12:
T 0 [j] T 0 [j] {{x} S|S T [j 0 ]}
13:
T T0
14:
return T [s]
The corresponding Python example program is given as below.
def subsetsum(xs, s):
low = sum([x for x in xs if x < 0])
up = sum([x for x in xs if x > 0])
tab = [[] for _ in xrange(low, up+1)]
for x in xs:
tab1 = tab[:]
for j in xrange(low, up+1):
if x == j:
tab1[j].append([x])
j1 = j - x
if low j1 and j1 up and tab[j1] != []:
tab1[j] = tab1[j] + [[x] + ys for ys in tab[j1]]

540

CHAPTER 14. SEARCHING

tab = tab1
return tab[s]

This imperative algorithm shows a clear structure, that the solution table is
built by looping every element. This can be realized in purely functional way
by folding. A finger tree can be used to represents the vector spans from sl to
su . It is initialized with all empty values as in the following equation.
subsetsum(X, s) = f old(build, {, , ..., }, X)[s]

(14.106)

After folding, the solution table is built, the answer is looked up at cell s12 .
For every element x X, function build folds the list {sl , sl + 1, ..., su }, with
every value j, it checks if it equals to x and appends the singleton set {x} to
the j-th cell. Not that here the cell is indexed from sl , but not 0. If the cell
corresponding to j x is not empty, the candidate solutions stored in that place
are also duplicated and add element x is added to every solution.
build(T, x) = f old(f, T, {sl , sl + 1, ..., su })

(14.107)

T 0 [j] {{x} Y |Y T [j 0 ]} : sl j 0 su T [j 0 ] 6= , j 0 = j x
T 0 : otherwise
(14.108)
Here the adjustment is applied on T 0 , which is another adjustment to T as
shown as below.

{x} T [j] : x = j
0
T =
(14.109)
T : otherwise

f (T, j) =

Note that the first clause in both equation (14.108) and (14.109) return a
new table with certain cell being updated with the given value.
The following Haskell example program implements this algorithm.
subsetsum xs s = foldl build (fromList [[] | _ [l..u]]) xs idx s where
l = sum $ filter (< 0) xs
u = sum $ filter (> 0) xs
idx t i = index t (i - l)
build tab x = foldl (t j let j = j - x in
adjustIf (l j && j u && tab idx j /= [])
(++ [(x:ys) | ys tab idx j]) j
(adjustIf (x == j) ([x]:) j t)) tab [l..u]
adjustIf pred f i seq = if pred then adjust f (i - l) seq else seq

Some materials like [16] provide common structures to abstract dynamic programming. So that problems can be solved with a generic solution by customizing the precondition, the comparison of candidate solutions for better choice,
and the merge method for sub solutions. However, the variety of problems
makes things complex in practice. Its important to study the properties of the
problem carefully.

Exercise 14.3
12 Again, here we skip the error handling to the case that s < s or s > s . There is no
u
l
solution if s is out of range.

14.4. SHORT SUMMARY

541

Realize a maze solver by using the stack approach, which can find all the
possible paths.
There are 92 distinct solutions for the 8 queens puzzle. For any one solution, rotating it 90 , 180 , 270 gives solutions too. Also flipping it vertically and horizontally also generate solutions. Some solutions are symmetric, so that rotation or flip gives the same one. There are 12 unique
solutions in this sense. Modify the program to find the 12 unique solutions. Improve the program, so that the 92 distinct solutions can be found
with fewer search.
Make the 8 queens puzzle solution generic so that it can solve n queens
puzzle.
Make the functional solution to the leap frogs puzzle generic, so that it
can solve n frogs case.
Modify the wolf, goat, and cabbage puzzle algorithm, so that it can find
all possible solutions.
Give the complete algorithm definition to solve the 2 water jugs puzzle
with extended Euclid algorithm.
We neednt the exact linear combination information x and y in fact.
After we know the puzzle is solvable by testing with GCD, we can blindly
execute the process that: fill A, pour A into B, whenever B is full, empty
it till there is expected volume in one jug. Realize this solution. Can this
one find faster solution than the original version?
Compare to the extended Euclid method, the DFS approach is a kind of
brute-force searching. Improve the extended Euclid approach by finding
the best linear combination which minimize |x| + |y|.
Realize the imperative Huffman code table generating algorithm.
One option to realize the bottom-up solution for the longest common
subsequence problem is to record the direction in the table. Thus, instead
of storing the length information, three values like N, for north, W
for west, and NW for northwest are used to indicate how to construct
the final result. We start from the bottom-right corner of the table, if
the cell value is NW, we go along the diagonal by moving to the cell
in the upper-left; if its N, we move vertically to the upper row; and
move horizontally if its W. Implement this approach in your favorite
programming language.

14.4

Short summary

This chapter introduces the elementary methods about searching. Some of them
instruct the computer to scan for interesting information among the data. They
often have some structure, that can be updated during the scan. This can be
considered as a special case for the information reusing approach. The other
commonly used strategy is divide and conquer, that the scale of the search

542

CHAPTER 14. SEARCHING

domain is kept decreasing till some obvious result. This chapter also explains
methods to search for solutions among domains. The solutions typically are not
the elements being searched. They can be a series of decisions or some operation
arrangement. If there are multiple solutions, sometimes, people want to find the
optimized one. For some spacial cases, there exist simplified approach such as
the greedy methods. And dynamic programming can be used for more wide
range of problems when they shows optimal substructures.

1985.

[7] Robert Boyer, and Strother Moore. MJRTY - A Fast Majority Vote Algorithm. Automated Reasoning: Essays in Honor of Woody Bledsoe, Automated Reasoning Series, Kluwer Academic Publishers, Dordrecht, The
Netherlands, 1991, pp. 105-117.
[8] Cormode, Graham; S. Muthukrishnan (2004). An Improved Data Stream
Summary: The Count-Min Sketch and its Applications. J. Algorithms 55:
29C38.
[9] Knuth Donald, Morris James H., jr, Pratt Vaughan. Fast pattern matching
in strings. SIAM Journal on Computing 6 (2): 323C350. 1977.
[10] Robert Boyer, Strother Moore. A Fast String Searching Algorithm.
Comm. ACM (New York, NY, USA: Association for Computing Machinery) 20 (10): 762C772. 1977
[11] R. N. Horspool. Practical fast searching in strings. Software - Practice &
Experience 10 (6): 501C506. 1980.
[12] Wikipedia.
Boyer-Moore
string
search
algorithm.
http://en.wikipedia.org/wiki/Boyer-Moore string search algorithm
[13] Wikipedia. Eight queens puzzle. http://en.wikipedia.org/wiki/Eight queens puzzle
543

544

BIBLIOGRAPHY

[14] George Polya. How to solve it: A new aspect of mathematical method.
Princeton University Press(April 25, 2004). ISBN-13: 978-0691119663
[15] Wikipedia. David A. Huffman. http://en.wikipedia.org/wiki/David A. Huffman
[16] Fethi Rabhi, Guy Lapalme Algorithms: a functional programming approach. Second edition. Addison-Wesley.

Part VI

Appendix

545

Appendix A

Lists
A.1

Introduction

This book intensely uses recursive list manipulations in purely functional settings. List can be treated as a counterpart to array in imperative settings, which
are the bricks to many algorithms and data structures.
For the readers who are not familiar with functional list manipulation, this
appendix provides a quick reference. All operations listed in this appendix
are not only described in equations, but also implemented in both functional
programming languages as well as imperative languages as examples. We also
provide a special type of implementation in C++ template meta programming
similar to [3] for interesting in next appendix.
Besides the elementary list operations, this appendix also contains explanation of some high order function concepts such as mapping, folding etc.

A.2

List Definition

Like arrays in imperative settings, lists play a critical role in functional setting1 .
Lists are built-in support in some programming languages like Lisp families and
ML families so it neednt explicitly define list in those environment.
List, or more precisely, singly linked-list is a data structure that can be
described below.
A list is either empty;
Or contains an element and a list.
Note that this definition is recursive. Figure A.1 illustrates a list with N
nodes. Each node contains two part, a key element and a sub list. The sub list
contained in the last node is empty, which is denoted as NIL.
1 Some reader may argue that lambda calculus plays the most critical role. Lambda calculus is somewhat as assembly languages to the computation world, which is worthy studying
from the essence of computation model to the practical programs. However, we dont dive
into the topic in this book. Users can refer to [4] for detail.

547

548

APPENDIX A. LISTS

key[1]

key[2]

...

key[N]

NIL

Figure A.1: A list contains N nodes

This data structure can be explicitly defined in programming languages support record (or compound type) concept. The following ISO C++ code defines
list2 .
template<typename T>
struct List {
T key;
List next;
};

A.2.1

Empty list

It is worth to mention about empty list a bit more in detail. In environment

supporting the nil concept, for example, C or java like programming languages,
empty list can have two different representations. One is the trivial NIL (or
null, or 0, which varies from languages); the other is an non-NIL empty list as
{}, the latter is typically allocated with memory but filled with nothing. In Lisp
dialects, the empty is commonly written as (). In ML families, its written as
[]. We use to denote empty list in equations and use NIL in pseudo code
sometimes to describe algorithms in this book.

A.2.2

Access the element and the sub list

Given a list L, two functions can be defined to access the element stored in
it and the sub list respectively. They are typically denoted as f irst(L), and
rest(L) or head(L) and tail(L) for the same meaning. These two functions are
named as car and cdr in Lisp for historic reason about the design of machine
2 We only use template to parameterize the type of the element in this chapter. Except
this point, all imperative source code are in ANSI C style to avoid language specific features.

A.3. BASIC LIST MANIPULATION

549

registers [5]. In languages support Pattern matching (e.g. ML families, Prolog

and Erlang etc.) These two functions are commonly realized by matching the
cons which well introduced later. for example the following Haskell program:
head (x:xs) = x
tail (x:xs) = xs

If the list is defined in record syntax like what we did above, these two
functions can be realized by accessing the record fields 3 .
template<typename T>
T first(List<T> xs) { return xskey; }
template<typename T>
List<T> rest(List<T> xs) { return xsnext; }

In this book, L0 is used to denote the rest(L) sometimes, also we uses l1

to represent f irst(L) in the context that the list is literately given in form
L = {l1 , l2 , ..., lN }.
More interesting, as far as in an environment support recursion, we can
define List. The following example define a list of integers in C++ compile
time.
struct Empty;
template<int x, typename T> struct List {
static const int first = x;
typedef T rest;
};

This line constructs a list of {1, 2, 3, 4, 5} in compile time.

typedef List<1, List<2, List<3, List<4 List<5, Empty> > > > > A;

A.3
A.3.1

Basic list manipulation

Construction

The last C++ template meta programming example actually shows literate
construction of a list. A list can be constructed from an element with a sub
list, where the sub list can be empty. We denote function cons(x, L) as the
constructor. This name is used in most Lisp dialects. In ML families, there are
cons operator defined as ::, (in Haskell its :).
We can define cons to create a record as we defined above in ISO C++, for
example4 .
template<typename T>
List<T> cons(T x, List<T> xs) {
List<T> lst = new List<T>;
lstkey = x;
lstnext = xs;
return lst;
}
3 They

can be also named as key and next or be defined as class methods.

It is often defined as a constructor method for the class template, However, we define it
as a standalone function for illustration purpose.
4

550

APPENDIX A. LISTS

A.3.2

Empty testing and length calculating

It is trivial to test if a list is empty. If the environment contains nil concept, the
testing should also handle nil case. Both Lisp dialects and ML families provide
null testing functions. Empty testing can also be realized by pattern-matching
with empty list if possible. The following Haskell program shows such example.
null [] = True
null _ = False

In this book we will either use empty(L) or L = where empty testing

happens.
With empty testing defined, its possible to calculate length for a list. In
imperative settings, Length is often implemented like the following.
function Length(L)
n0
while L 6= N IL do
nn+1
L Next(L)
This ISO C++ code translates the algorithm to real program.
template<typename T>
int length(List<T> xs) {
int n = 0;
for (; xs; ++n, xs = xsnext);
return n
}

However, in purely funcitonal setting, we cant mutate a counter variable.

the idea is that, if the list is empty, then its size is zero; otherwise, we can
recursively calculate the length of the sub list, then add it by one to get the
length of this list.

length(L) =

0 : L=
1 + length(L0 ) : otherwise

(A.1)

Here L0 = rest(L) as mentioned above, its {l2 , l3 , ..., lN } for list contains N
elements. Note that both L and L0 can be empty . In this equation, we also
use = to test if list L is empty. In order to know the length of a list, we need
traverse all the elements from the head to the end, so that this algorithm is
proportion to the number of elements stored in the list. It is a linear algorithm
bound to O(N ) time.
Below are two programs in Haskell and in Scheme/Lisp realize this recursive
algorithm.
length [] = 0
length (x:xs) = 1 + length xs
(define (length lst)
(if (null? lst) 0 (+ 1 (length (cdr lst)))))

How to testing if two lists are identical is left as exercise to the reader.

A.3. BASIC LIST MANIPULATION

A.3.3

551

indexing

One big difference between array and list (singly-linked list accurately) is that
array supports random access. Many programming languages support using
x[i] to access the i-th element stored in array in constant O(1) time. The
index typically starts from 0, but its not the all case. Some programming
languages using 1 as the first index. In this appendix, we treat index starting
from 0. However, we must traverse the list with i steps to reach the target
element. The traversing is quite similar to the length calculation. Thus its
commonly expressed as below in imperative settings.
function Get-At(L, i)
while i 6= 0 do
L Next(L)
return First(L)
Note that this algorithm doesnt handle the error case such that the index
isnt within the bound of the list. We assume that 0 i < |L|, where |L| =
length(L). The error handling is left as exercise to the reader. The following
ISO C++ code is a line-by-line translation of this algorithm.
template<typename T>
T getAt(List<T> lst, int n) {
while(n--)
lst = lstnext;
return lstkey;
}

However, in purely functional settings, we turn to recursive traversing instead of while-loop.

getAt(L, i) =

F irst(L) : i = 0
getAt(Rest(L), i 1) : otherwise

(A.2)

In order to get the i-th element, the algorithm does the following:
if i is 0, then we are done, the result is the first element in the list;
Otherwise, the result is to get the (i 1)-th element from the sub-list.
This algorithm can be translated to the following Haskell code.
getAt i (x:xs) = if i == 0 then x else getAt i-1 xs

Note that we are using pattern matching to ensure the list isnt empty, which
actually handles all out-of-bound cases with un-matched pattern error. Thus if
i > |L|, we finally arrive at a edge case that the index is i |L|, while the list is
empty; On the other hand, if i < 0, minus it by one makes it even farther away
from 0. We finally end at the same error that the index is some negative, while
the list is empty;
The indexing algorithm takes time proportion to the value of index, which is
bound to O(N ) linear time. This section only address the read semantics. How
to mutate the element at a given position is explained in later section.

552

APPENDIX A. LISTS

A.3.4

Access the last element

Although accessing the first element and the rest list L0 is trivial, the opposite
operations, that retrieving the last element and the initial sub list need linear
time without using a tail pointer. If the list isnt empty, we need traverse it till
the tail to get these two components. Below are their imperative descriptions.
function Last(L)
x NIL
while L 6= NIL do
x First(L)
L Rest(L)
return x
function Init(L)
L0 NIL
while Rest(L) 6= NIL do
L0 Append(L0 , First(L))
L Rest(L)
return L0
The algorithm assumes that the input list isnt empty, so the error handling
is skipped. Note that the Init() algorithm uses the appending algorithm which
will be defined later.
Below are the corresponding ISO C++ implementation. The optimized version by utilizing tail pointer is left as exercise.
template<typename T>
T last(List<T> xs) {
T x; / Can be set to a special value to indicate empty list err. /
for (; xs; xs = xsnext)
x = xskey;
return x;
}
template<typename T>
List<T> init(List<T> xs) {
List<T> ys = NULL;
for (; xsnext; xs = xsnext)
ys = append(ys, xskey);
return ys;
}

While these two algorithms can be implemented in purely recursive manner

as well. When we want to access the last element.
If the list contains only one element (the rest sub-list is empty), the result
is this very element;
Otherwise, the result is the last element of the rest sub-list.

last(L) =

F irst(L) : Rest(L) =
last(Rest(L)) : otherwise

(A.3)

A.3. BASIC LIST MANIPULATION

553

The similar approach can be used to get a list contains all elements except
for the last one.
The edge case: If the list contains only one element, then the result is an
empty list;
Otherwise, we can first get a list contains all elements except for the last
one from the rest sub-list, then construct the final result from the first
element and this intermediate result.

init(L) =

:
cons(l1 , init(L )) :
0

L0 =
otherwise

(A.4)

Here we denote l1 as the first element of L, and L0 is the rest sub-list. This
recursive algorithm neednt use appending, It actually construct the final result
list from right to left. Well introduce a high-level concept of such kind of
computation later in this appendix.
Below are Haskell programs implement last() and init() algorithms by using
pattern matching.
last [x] = x
last (_:xs) = last xs
init [x] = []
init (x:xs) = x : init xs

Where [x] matches the singleton list contains only one element, while (_:xs)
matches any non-empty list, and the underscore (_) is used to indicate that we
dont care about the element. For the detail of pattern matching, readers can
refer to any Haskell tutorial materials, such as [8].

A.3.5

Reverse indexing

Reverse indexing is a general case for last(), finding the i-th element in a singlylinked list with the minimized memory spaces is interesting, and this problem
is often used in technical interview in some companies. A naive implementation
takes 2 rounds of traversing, the first round is to determine the length of the
list N , then, calculate the left-hand index by N i 1. Finally a second round
of traverse is used to access the element with the left-hand index. This idea can
be give as the following equation.
getAtR(L, i) = getAt(L, length(L) i 1)
There exists better imperative solution. For illustration purpose, we omit
the error cases such as index is out-of-bound etc. The idea is to keep two
pointers p1 , p2 , with the distance of i between them, that resti (p2 ) = p1 , where
resti (p1 ) means repleatedly apply rest() function i times. It says that succeeds
i steps from p2 gets p1 . We can start p2 from the head of the list and advance
the two pointers in parallel till one of them (p1 ) arrives at the end of the list.
At that time point, pointer p2 exactly arrived at the i-th element from right.
Figure A.2 illustrates this idea.
It is straightforward to realize the imperative algorithm based on this double
pointers solution.

554

APPENDIX A. LISTS

x[1]

x[2]

...

x[i+1]

...

x[N]

(a) p2 starts from the head, which is behind p1 in i steps.

x[1]

x[2]

...

x[N-i]

...

x[N]

(b) When p1 reaches the end, p2 points to the i-th element from right.

Figure A.2: Double pointers solution to reverse indexing.

function Get-At-R(L, i)
pL
while i 6= 0 do
L Rest(L)
ii1
while Rest(L) 6= NIL do
L Rest(L)
p Rest(p)
return First(p)
The following ISO C++ code implements the double pointers right indexing
algorithm.
template<typename T>
T getAtR(List<T> xs, int i) {
List<T> p = xs;
while(i--)
xs = xsnext;
for(; xsnext; xs = xsnext, p = pnext);
return pkey;
}

The same idea can be realized recursively as well. If we want to access the
i-th element of list L, we can examine the two lists L and S = {li , li+1 , ..., lN }
simultaneously, where S is a sub-list of L without the first i elements.
The edge case: If S is a singleton list, then the i-th element from right is
the first element in L;
Otherwise, we drop the first element from L and S, and recursively exam-

A.3. BASIC LIST MANIPULATION

555

ine L0 and S 0 .
This algorithm description can be formalized as the following equations.
getAtR(L, i) = examine(L, drop(i, L))

(A.5)

Where function examine(L, S) is defined as below.

examine(L, S) =

f irst(L) : |S| = 1
examine(rest(L), rest(S)) : otherwise

(A.6)

Well explain the detail of drop() function in later section about list mutating
operations. Here it can be implemented as repeatedly call rest() with specified
times.

L : n=0
drop(n, L) =
drop(n 1, rest(L)) : otherwise
Translating the equations to Haskell yields this example program.
atR :: [a] Int a
atR xs i = get xs (drop i xs) where
get (x:_) [_] = x
get (_:xs) (_:ys) = get xs ys
drop n as@(_:as) = if n == 0 then as else drop (n-1) as

Here we use dummy variable _ as the placeholders for components we dont

care.

A.3.6

Mutating

Strictly speaking, we cant mutate the list at all in purely functional settings.
Unlike in imperative settings, mutate is actually realized by creating new list.
Almost all functional environments support garbage collection, the original list
may either be persisted for reusing, or released (dropped) at sometime (Chapter
2 in [6]).
Appending
Function cons can be viewed as building list by insertion element always on head.
If we chains multiple cons operations, it can repeatedly construct a list from
right to the left. Appending on the other hand, is an operation adding element
to the tail. Compare to cons which is trivial constant time O(1) operation, We
must traverse the whole list to locate the appending position. It means that
appending is bound to O(N ), where N is the length of the list. In order to speed
up the appending, imperative implementation typically uses a field (variable) to
record the tail position of a list, so that the traversing can be avoided. However,
in purely functional settings we cant use such tail pointer. The appending has
to be realized in recursive manner.

append(L, x) =

{x} : L =
cons(f irst(L), append(rest(L), x)) : otherwise

That the algorithm handles two different appending cases:

(A.7)

556

APPENDIX A. LISTS
If the list is empty, the result is a singleton list contains x, which is the
element to be appended. The singleton list notion {x} = cons(x, ), is a
simplified form of cons the element with an empty list ;
Otherwise, for the none-empty list, the result can be achieved by first appending the element x to the rest sub-list, then construct the first element
of L with the recursive appending result.

For the none-trivial case, if we denote L = {l1 , l2 , ...}, and L0 = {l2 , l3 , ...}
the equation can be written as.

append(L, x) =

{x} : L =
cons(l1 , append(L0 , x)) : otherwise

(A.8)

Well use both forms in the rest of this appendix.

The following Scheme/Lisp program implements this algorithm.
(define (append lst x)
(if (null? lst)
(list x)
(cons (car lst) (append (cdr lst) x))))

Even without the tail pointer, its possible to traverse the list imperatively
and append the element at the end.
function Append(L, x)
if L = NIL then
return Cons(x, NIL)
HL
while Rest(L) 6= NIL do
L Rest(L)
Rest(L) Cons(x, NIL)
return H
The following ISO C++ programs implements this algorithm. How to utilize a tail field to speed up the appending is left as exercise to the reader for
interesting.
template<typename T>
List<T> append(List<T> xs, T x) {
List<T> tail, head;
for (head = tail = xs; xs; xs = xsnext)
tail = xs;
if (!head)
head = cons<T>(x, NULL);
else
tailnext = cons<T>(x, NULL);
return head;
}

Mutate element at a given position

Although we have defined random access algorithm getAt(L, i), we cant just
mutate the element returned by this function in a sense of purely functional

A.3. BASIC LIST MANIPULATION

557

settings. It is quite common to provide reference semantics in imperative programming languages and in some almost functional environment. Readers can
refer to [4] for detail. For example, the following ISO C++ example returns a
reference instead of a value in indexing program.
template<typename T>
T& getAt(List<T> xs, int n) {
while (n--)
xs = xsnext;
return xskey;
}

So that we can use this function to mutate the 2nd element as below.
List<int> xs = cons(1, cons(2, cons<int>(3, NULL)));
getAt(xs, 1) = 4;

In an impure functional environment, such as Scheme/Lisp, to set the ith element to a given value can be implemented by mutate the referenced cell
directly as well.
(define (set-at! lst i x)
(if (= i 0)
(set-car! lst x)
(set-at! (cdr lst) (- i 1) x)))

This program first checks if the index i is zero, if so, it mutate the first
element of the list to given value x; otherwise, it deduces the index i by one, and
tries to mutate the rest of the list at this new index with value x. This function
doesnt return meaningful value. It is for use of side-effect. For instance, the
following code mutates the 2nd element in a list.
(define lst (1 2 3 4 5))
(set-at! lst 1 4)
(display lst)
(1 4 3 4 5)

In order to realize a purely functional setAt(L, i, x) algorithm, we need avoid

directly mutating the cell, but creating a new one:
Edge case: If we want to set the value of the first element (i = 0), we
construct a new list, with the new value and the sub-list of the previous
one;
Otherwise, we construct a new list, with the previous first element, and a
new sub-list, which has the (i 1)-th element set with the new value.
This recursive description can be formalized by the following equation.

setAt(L, i, x) =

cons(x, L0 ) : i = 0
cons(l1 , setAt(L0 , i 1, x)) : otherwise

(A.9)

Comparing the below Scheme/Lisp implementation to the previous one reveals the difference from imperative mutating.

558

APPENDIX A. LISTS

(define (set-at lst i x)

(if (= i 0)
(cons x (cdr lst))
(cons (car lst) (set-at (cdr lst) (- i 1) x))))

Here we skip the error handling for out-of-bound error etc. Again, similar
to the random access algorithm, the performance is bound to linear time, as
traverse is need to locate the position to set the value.
insertion
There are two semantics about list insertion. One is to insert an element at a
given position, which can be denoted as insert(L, i, x). The algorithm is close
to setAt(L, i, x); The other is to insert an element to a sorted list, so that the
the result list is still sorted.
Lets first consider how to insert an element x at a given position i. The
obvious thing is that we need firstly traverse i elements to get to the position,
the rest of work is to construct a new sub-list with x being the head of this
sub-list. Finally, we construct the whole result by attaching this new sub-list to
the end of the first i elements.
The algorithm can be described accordingly to this idea. If we want to insert
an element x to a list L at i.
Edge case: If i is zero, then the insertion turns to be a trivial cons
operation cons(x, L);
Otherwise, we recursively insert x to the sub-list L0 at position i 1; then
construct the first element with this result.
Below equation formalizes the insertion algorithm.

insert(L, i, x) =

cons(x, L) : i = 0
cons(l1 , insert(L0 , i 1, x)) : otherwise

(A.10)

The following Haskell program implements this algorithm.

insert xs 0 y = y:xs
insert (x:xs) i y = x : insert xs (i-1) y

This algorithm doesnt handle the out-of-bound error. However, we can

interpret the case, that the position i exceeds the length of the list as appending.
Readers can considering about it in the exercise of this section.
The algorithm can also be designed imperatively: If the position is zero, just
construct the new list with the element to be inserted as the first one; Otherwise,
we record the head of the list, then start traversing the list i steps. We also need
an extra variable to memorize the previous position for the later list insertion
operation. Below is the pseudo code.
function Insert(L, i, x)
if i = 0 then
return Cons(x, L)
HL
pL

A.3. BASIC LIST MANIPULATION

559

while i 6= 0 do
pL
L Rest(L)
ii1
Rest(p) Cons(x, L)
return H
And the ISO C++ example program is given by translating this algorithm.
template<typename T>
List<T> insert(List<T> xs, int i , int x) {
List<T> head, prev;
if (i == 0)
return cons(x, xs);
for (head = xs; i; --i, xs = xsnext)
prev = xs;
prevnext = cons(x, xs);
return head;
}

If the list L is sorted, that is for any position 1 i j N , we have li lj .

We can design an algorithm which inserts a new element x to the list, so that
the result list is still sorted.

cons(x, ) : L =

cons(x, L) : x < l1
(A.11)
insert(x, L) =

cons(l1 , insert(x, L0 )) : otherwise

The idea is that, to insert an element x to a sorted list L:
If either L is empty or x is less than the first element in L, we just put x
in front of L to construct the result;
Otherwise, we recursively insert x to the sub-list L0 .
The following Haskell program implements this algorithm. Note that we use
, to determine the ordering. Actually this constraint can be loosened to the
strict less (<), that if elements can be compare in terms of <, we can design
a program to insert element so that the result list is still sorted. Readers can
refer to the chapters of sorting in this book for details about ordering.
insert y [] = [y]
insert y xs@(x:xs) = if y x then y : xs else x : insert y xs

Since the algorithm need compare the elements one by one, its also a linear
time algorithm. Note that here we use the as notion for pattern matching in
Haskell. Readers can refer to [8] and [7] for details.
This ordered insertion algorithm can be designed in imperative manner, for
example like the following pseudo code5 .
function Insert(x, L)
if L = x < First(L) then
return Cons(x, L)
5 Reader can refer to the chapter The evolution of insertion sort in this book for a minor
different one

560

APPENDIX A. LISTS

HL
while Rest(L) 6= First(Rest(L)) < x do
L Rest(L)
Rest(L) Cons(x, Rest(L))
return H
If either the list is empty, or the new element to be inserted is less than
the first element in the list, we can just put this element as the new first one;
Otherwise, we record the head, then traverse the list till a position, where x is
less than the rest of the sub-list, and put x in that position. Compare this one
to the insert at algorithm shown previously, the variable p uses to point to the
previous position during traversing is omitted by examine the sub-list instead
of current list. The following ISO C++ program implements this algorithm.
template<typename T>
List<T> insert(T x, List<T> xs) {
List<T> head;
if (!xs | | x < xskey)
return cons(x, xs);
for (head = xs; xsnext && xsnextkey < x; xs = xsnext);
xsnext = cons(x, xsnext);
return head;
}

With this linear time ordered insertion defined, its possible to implement
quadratic time insertion-sort by repeatedly inserting elements to an empty list
as formalized in this equation.

: L=
sort(L) =
(A.12)
insert(l1 , sort(L0 )) : otherwise
This equation says that if the list to be sorted is empty, the result is also
empty, otherwise, we can firstly recursively sort all elements except for the
first one, then ordered insert the first element to this intermediate result. The
corresponding Haskell program is given as below.
isort [] = []
isort (x:xs) = insert x (isort xs)

And the imperative linked-list base insertion sort is described in the following. That we initialize the result list as empty, then take the element one by
one from the list to be sorted, and ordered insert them to the result list.
function Sort(L)
L0
while L 6= do
L0 Insert(First(L), L0 )
L Rest(L)
return L0
Note that, at any time during the loop, the result list is kept sorted. There is
a major difference between the recursive algorithm (formalized by the equation)
and the procedural one (described by the pseudo code), that the former process
the list from right, while the latter from left. Well see in later section about
tail-recursion how to eliminate this difference.
The ISO C++ version of linked-list insertion sort is list like this.

A.3. BASIC LIST MANIPULATION

561

template<typename T>
List<T> isort(List<T> xs) {
List<T> ys = NULL;
for(; xs; xs = xsnext)
ys = insert(xskey, ys);
return ys;
}

There is also a dedicated chapter discusses insertion sort in this book. Please
refer to that chapter for more details including performance analysis and finetuning.
deletion
In purely functional settings, there is no deletion at all in terms of mutating, the
data is persist, what the semantic deletion means is actually to create a new
list with all the elements in previous one except for the element being deleted.
Similar to the insertion, there are also two deletion semantics. One is to
delete the element at a given position; the other is to find and delete elements
of a given value. The first can be expressed as delete(L, i), while the second is
delete(L, x).
In order to design the algorithm delete(L, i) (or delete at), we can use the
idea which is quite similar to random access and insertion, that we first traverse
the list to the specified position, then construct the result list with the elements
we have traversed, and all the others except for the next one we havent traversed
yet.
The strategy can be realized in a recursive manner that in order to delete
the i-th element from list L,
If i is zero, that we are going to delete the first element of a list, the result
is obviously the rest of the list;
If the list to be removed element is empty, the result is anyway empty;
Otherwise, we can recursively delete the (i 1)-th element from the sublist L0 , then construct the final result from the first element of L and this
intermediate result.
Note there are two edge cases, and the second case is major used for error
handling. This algorithm can be formalized with the following equation.

L0 : i = 0

: L=
delete(L, i) =
(A.13)

cons(l1 , delete(L0 , i 1)) :

Where L0 = rest(L), l1 = f irst(L). The corresponding Haskell example
program is given below.
del (_:xs) 0 = xs
del [] _ = []
del (x:xs) i = x : del xs (i-1)

This is a linear time algorithm as well, and there are also alternatives for
implementation, for example, we can first split the list at position i 1, to get
2 sub-lists L1 and L2 , then we can concatenate L1 and L02 .

562

APPENDIX A. LISTS

The delete at algorithm can also be realized imperatively, that we traverse

to the position by looping:
function Delete(L, i)
if i = 0 then
return Rest(L)
HL
pL
while i 6= 0 do
ii1
pL
L Rest(L)
Rest(p) Rest(L)
return H
Different from the recursive approach, The error handling for out-of-bound is
skipped. Besides that the algorithm also skips the handling of resource releasing
which is necessary in environments without GC (Garbage collection). Below ISO
C++ code for example, explicitly releases the node to be deleted.
template<typename T>
List<T> del(List<T> xs, int i) {
List<T> head, prev;
if (i == 0)
head = xsnext;
else {
for (head = xs; i; --i, xs = xsnext)
prev = xs;
prevnext = xsnext;
}
xsnext = NULL;
delete xs;
return head;
}

Note that the statement xs->next = NULL is neccessary if the destructor is

designed to release the whole linked-list recursively.
The find and delete semantic can further be represented in two ways, one is
just find the first occurrence of a given value, and delete this element from the
list; The other is to find ALL occurrence of this value, and delete these elements.
The later is more general case, and it can be achieved by a minor modification
of the former. We left the find all and delete algorithm as an exercise to the
reader.
The algorithm can be designed exactly as the term find and delete but
not find then delete, that the finding and deleting are processed in one pass
traversing.
If the list to be dealt with is empty, the result is obviously empty;
If the list isnt empty, we examine the first element of the list, if it is
identical to the given value, the result is the sub list;
Otherwise, we keep the first element, and recursively find and delete the
element in the sub list with the given value. The final result is a list
constructed with the kept first element, and the recursive deleting result.

A.3. BASIC LIST MANIPULATION

This algorithm can be formalized by the following equation.

: L=

L0 : l1 = x
delete(L, x) =

cons(l1 , delete(L0 , x)) : otherwise

563

(A.14)

This algorithm is bound to linear time as it traverses the list to find and
delete element. Translating this equation to Haskell program yields the below
code, note that, the first edge case is handled by pattern-matching the empty
list; while the other two cases are further processed by if-else expression.
del [] _ = []
del (x:xs) y = if x == y then xs else x : del xs y

Different from the above imperative algorithms, which skip the error handling in most cases, the imperative find and delete realization must deal with
the problem that the given value doesnt exist.
function Delete(L, x)
if L = then
. Empty list
return
if First(L) = x then
H Rest(L)
else
HL
while L 6= First(L) 6= x do
. List isnt empty
pL
L Rest(L)
if L 6= then
. Found
Rest(p) Rest(L)
return H
If the list is empty, the result is anyway empty; otherwise, the algorithm
traverses the list till either finds an element identical to the given value or to
the end of the list. If the element is found, it is removed from the list. The
following ISO C++ program implements the algorithm. Note that there are
codes release the memory explicitly.
template<typename T>
List<T> del(List<T> xs, T x) {
List<T> head, prev;
if (!xs)
return xs;
if (xskey == x)
head = xsnext;
else {
for (head = xs; xs && xskey != x; xs = xsnext)
prev = xs;
if (xs)
prevnext = xsnext;
}
if (xs) {
xsnext = NULL;
delete xs;
}

564

APPENDIX A. LISTS

return head;
}

concatenate
Concatenation can be considered as a general case for appending, that appending only adds one more extra element to the end of the list, while concatenation
adds multiple elements.
However, It will lead to quadratic algorithm if implement concatenation
naively by appending, which performs poor. Consider the following equation.

concat(L1 , L2 ) =

L1 : L2 =
concat(append(L1 , f irst(L2 )), rest(L2 )) : otherwise

Note that each appending algorithm need traverse to the end of the list,
which is proportion to the length of L1 , and we need do this linear time appending work |L2 | times, so the total performance is O(|L1 | + (|L1 | + 1) + ... +
(|L1 | + |L2 |)) = O(|L1 ||L2 | + |L2 |2 ).
The key point is that the linking operation of linked-list is fast (constant
O(1) time), we can traverse to the end of L1 only one time, and link the second
list to the tail of L1 .

: L1 =
: otherwise
(A.15)
This algorithm only traverses the first list one time to get the tail of L1 , and
then linking the second list with this tail. So the algorithm is bound to linear
O(|L1 |) time.
This algorithm is described as the following.
concat(L1 , L2 ) =

L2
cons(f irst(L1 ), concat(rest(L1 ), L2 ))

If the first list is empty, the concatenate result is the second list;
Otherwise, we concatenate the second list to the sub-list of the first one,
and construct the final result with the first element and this intermediate
result.
Most functional languages provide built-in functions or operators for list
concatenation, for example in ML families ++ is used for this purpose.
[] ++ ys = ys
xs ++ [] = xs
(x:xs) ++ ys = x : xs ++ ys

Note we add another edge case that if the second list is empty, we neednt
traverse to the end of the first one and perform linking, the result is merely the
first list.
In imperative settings, concatenation can be realized in constant O(1) time
with the augmented tail record. We skip the detailed implementation for this
method, reader can refer to the source code which can be download along with
this appendix.
The imperative algorithm without using augmented tail record can be described as below.

A.3. BASIC LIST MANIPULATION

565

function Concat(L1 , L2 )
if L1 = then
return L2
if L2 = then
return L1
H L1
while Rest(L1 ) 6= do
L1 Rest(L1 )
Rest(L1 ) L2
return H
And the corresponding ISO C++ example code is given like this.
template<typename T>
List<T> concat(List<T> xs, List<T> ys) {
List<T> head;
if (!xs)
return ys;
if (!ys)
return xs;
for (head = xs; xsnext; xs = xsnext);
xsnext = ys;
return head;
}

A.3.7

sum and product

Recursive sum and product

It is common to calculate the sum or product of a list of numbers. They are
quite similar in terms of algorithm structure. Well see how to abstract such
structure in later section.
In order to calculate the sum of a list:
If the list is empty, the result is zero;
Otherwise, the result is the first element plus the sum of the rest of the
list.
Formalize this description gives the following equation.

sum(L) =

0 : L=
l1 + sum(L0 ) : otherwise

(A.16)

However, we cant merely replace plus to times in this equation to achieve

product algorithm, because it always returns zero. We can define the product
of empty list as 1 to solve this problem.

product(L) =

1 : L=
l1 product(L0 ) : otherwise

The following Haskell program implements sum and product.

(A.17)

566

APPENDIX A. LISTS

sum [] = 0
sum (x:xs) = x + sum xs
product [] = 1
product (x:xs) = x product xs

Both algorithms traverse the whole list during calculation, so they are bound
to O(N ) linear time.
Tail call recursion
Note that both sum and product algorithms actually compute the result from
right to left. We can change them to the normal way, that calculate the accumulated result from left to right. For example with sum, the result is actually
accumulated from 0, and adds element one by one to this accumulated result
till all the list is consumed. Such approach can be described as the following.
When accumulate result of a list by summing:
If the list is empty, we are done and return the accumulated result;
Otherwise, we take the first element from the list, accumulate it to the
result by summing, and go on processing the rest of the list.
Formalize this idea to equation yields another version of sum algorithm.

A : L=
sum0 (A, L) =
(A.18)
sum0 (A + l1 , L0 ) : otherwise
And sum can be implemented by calling this function by passing start value
0 and the list as arguments.
sum(L) = sum0 (0, L)

(A.19)

The interesting point of this approach is that, besides it calculates the result
in a normal order from left to right; by observing the equation of sum0 (A, L),
we found it neednt remember any intermediate results or states when perform
recursion. All such states are either passed as arguments (A for example) or
can be dropped (previous elements of the list for example). So in a practical
implementation, such kind of recursive function can be optimized by eliminating
the recursion at all.
We call such kind of function as tail recursion (or tail call), and the optimization of removing recursion in this case as tail recursion optimization[10]
because the recursion happens as the final action in such function. The advantage of tail recursion optimization is that the performance can be greatly
improved, so that we can avoid the issue of stack overflow in deep recursion
algorithms such as sum and product.
Changing the sum and product Haskell programs to tail-recursion manner
gives the following modified programs.
sum = sum 0 where
sum acc [] = acc
sum acc (x:xs) = sum (acc + x) xs
product = product 1 where
product acc [] = acc
product acc (x:xs) = product (acc x) xs

A.3. BASIC LIST MANIPULATION

567

In previous section about insertion sort, we mentioned that the functional

version sorts the elements form right, this can also be modified to tail recursive
realization.
sort0 (A, L) =

A : L=
sort0 (insert(l1 , A), L0 ) : otherwise

(A.20)

The the sorting algorithm is just calling this function by passing empty list
as the accumulator argument.
sort(L) = sort0 (, L)

(A.21)

Implementing this tail recursive algorithm to real program is left as exercise

to the reader.
As the end of this sub-section, lets consider an interesting problem, that
how to design an algorithm to compute bN effectively? (refer to problem 1.16
in [5].)
A naive brute-force solution is to repeatedly multiply b for N times from 1,
which leads to a linear O(N ) algorithm.
function Pow(b, N )
x1
loopN times
xxb
return x
Actually, the solution can be greatly improved. Consider we are trying to
calculate b8 . By the first 2 iterations in above naive algorithm, we got x = b2 .
At this stage, we neednt multiply x with b to get b3 , we can directly calculate
x2 , which leads to b4 . And if we do this again, we get (b4 )2 = b8 . Thus we only
need looping 3 times but not 8 times.
An algorithm based on this idea to compute bN if N = 2M for some nonnegative integer M can be shown in the following equation.

pow(b, N ) =

b
pow(b, N2 )2

: N =1
: otherwise

It is easy to extend this divide and conquer algorithm so that N can be any
non-negative integer.
For the trivial case, that N is zero, the result is 1;
N

If N is even number, we can halve N , and compute b 2 first. Then calculate

the square number of this result.
Otherwise, N is odd. Since N 1 is even, we can first recursively compute
bN 1 , the multiply b one more time to this result.
Below equation formalizes this description.

1 : N =0

pow(b, N ) =
pow(b, N2 )2 : 2|N

b pow(b, N 1) : otherwise

(A.22)

568

APPENDIX A. LISTS

However, its hard to turn this algorithm to be tail-recursive mainly because

of the 2nd clause. In fact, the 2nd clause can be alternatively realized by
squaring the base number, and halve the exponent.

pow(b, N ) =

1 : N =0
pow(b2 , N2 ) : 2|N

b pow(b, N 1) : otherwise

(A.23)

With this change, its easy to get a tail-recursive algorithm as the following,
so that bN = pow0 (b, N, 1).

pow0 (b, N, A) =

A : N =0
pow0 (b2 , N2 , A) : 2|N

pow0 (b, N 1, A b) : otherwise

(A.24)

Compare to the naive brute-force algorithm, we improved the performance

to O(lg N ). Actually, this algorithm can be improved even one more step.
Observe that if we represent N in binary format N = (am am1 ...a1 a0 )2 , we
i
clear know that the computation for b2 is necessary if ai = 1. This is quite
similar to the idea of Binomial heap (reader can refer to the chapter of binomial
heap in this book). Thus we can calculate the final result by multiplying all of
them for bits with value 1.
For instance, when we compute b11 , as 11 = (1011)2 = 23 + 2 + 1, thus
3
11
b = b2 b2 b. We can get the result by the following steps.
1. calculate b1 , which is b;
2. Get b2 from previous result;
2

3. Get b2 from step 2;

4. Get b2 from step 3.

Finally, we multiply the results of step 1, 2, and 4 which yields b11 .
Summarize this idea, we can improve the algorithm as below.
0

pow (b, N, A) =

A : N =0
: 2|N
: otherwise

pow (b2 , N2 , A)
0 2
pow (b , b N2 c, A b)

(A.25)

This algorithm essentially shift N to right for 1 bit each time (by dividing N
by 2). If the LSB (Least Significant Bit, which is the lowest bit) is 0, it means
N is even. It goes on computing the square of the base, without accumulating
the final product (Just like the 3rd step in above example); If the LSB is 1, it
means N is odd. It squares the base and accumulates it to the product A; The
edge case is when N is zero, which means we exhaust all the bits in N , thus the
final result is the accumulator A. At any time, the updated base number b0 , the
shifted exponent number N 0 , and the accumulator A satisfy the invariant that
0
bN = b0N A.
This algorithm can be implemented in Haskell like the following.

A.3. BASIC LIST MANIPULATION

569

pow b n = pow b n 1 where

pow b n acc | n == 0 = acc
| even n = pow (bb) (n div 2) acc
| otherwise = pow (bb) (n div 2) (accb)

Compare to previous algorithm, which minus N by one to change it to even

when N is odd, this one halves N every time. It exactly runs m rounds, where m
is the number of bits of N . However, the performance is still bound to O(lg N ).
How to implement this algorithm imperatively is left as exercise to the reader.
Imperative sum and product
The imperative sum and product are just applying plus and times while traversing the list.
function Sum(L)
s0
while L 6= do
s s+ First(L)
L Rest(L)
return s
function Product(L)
p1
while L 6= do
p p First(L)
L Rest(L)
return p
The corresponding ISO C++ example programs are list as the following.
template<typename T>
T sum(List<T> xs) {
T s;
for (s = 0; xs; xs = xsnext)
s += xskey;
return s;
}
template<typename T>
T product(List<T> xs) {
T p;
for (p = 1; xs; xs = xsnext)
p = xskey;
return p;
}

One interesting usage of product algorithm is that we can calculate factorial

of N by calculating the product of {1, 2, ..., N } that N ! = product([1..N ]).

A.3.8

maximum and minimum

Another very useful use case is to get the minimum or maximum element of
a list. Well see that their algorithm structures are quite similar again. Well

570

APPENDIX A. LISTS

generalize this kind of feature and introduce about higher level abstraction in
later section. For both maximum and minimum algorithms, we assume that the
given list isnt empty.
In order to find the minimum element in a list.
If the list contains only one element, (a singleton list), the minimum element is this one;
Otherwise, we can firstly find the minimum element of the rest list, then
compare the first element with this intermediate result to determine the
final minimum value.
This algorithm can be formalized by the

l1 :

l1 :
min(L) =

min(L0 ) :

following equation.
L = {l1 }
l1 min(L0 )
otherwise

(A.26)

In order to get the maximum element instead of the minimum one, we can
simply replace the comparison to in the above equation.

l1 : L = {l1 }

l1 : l1 max(L0 )
max(L) =
(A.27)

max(L0 ) : otherwise
Note that both maximum and minimum actually process the list from right
to left. It remind us about tail recursion. We can modify them so that the list
is processed from left to right. Whats more, the tail recursion version brings us
on-line algorithm, that at any time, we hold the minimum or maximum result
of the list we examined so far.

a : L=

min(L0 , l1 ) : l1 < a
min0 (L, a) =
(A.28)

min(L0 , a) : otherwise

a : L=

max(L0 , l1 ) : a < l1
(A.29)
max0 (L, a) =

max(L0 , a) : otherwise
Different from the tail recursion sum and product, we cant pass constant
value to min0 , or max0 in practice, this is because we have to pass infinity
(min(L, )) or negative infinity (max(L, )) in theory, but in a real machine
neither of them can be represented since the length of word is limited.
Actually, there is workaround, we can instead pass the first element of the
list, so that the algorithms become applicable.
min(L) = min(L0 , l1 )
max(L) = max(L0 , l1 )

(A.30)

The corresponding real programs are given as the following. We skip the
none tail recursion programs, as they are intuitive enough. Reader can take
them as exercises for interesting.

A.3. BASIC LIST MANIPULATION

571

min (x:xs) = min xs x where

min [] a = a
min (x:xs) a = if x < a then min xs x else min xs a
max (x:xs) = max xs x where
max [] a = a
max (x:xs) a = if a < x then max xs x else max xs a

The tail call version can be easily translated to imperative min/max algorithms.
function Min(L)
m First(L)
L Rest(L)
while L 6= do
if First(L) < m then
m First(L)
L Rest(L)
return m
function Max(L)
m First(L)
L Rest(L)
while L 6= do
if m < First(L) then
m First(L)
L Rest(L)
return m
The corresponding ISO C++ programs are given as below.
template<typename T>
T min(List<T> xs) {
T x;
for (x = xskey; xs; xs = xsnext)
if (xskey < x)
x = xskey;
return x;
}
template<typename T>
T max(List<T> xs) {
T x;
for (x = xskey; xs; xs = xsnext)
if (x < xskey)
x = xskey;
return x;
}

Another method to achieve tail-call maximum( and minimum) algorithm is

by discarding the smaller element each time. The edge case is as same as before;
for recursion case, since there are at least two elements in the list, we can take
the first two for comparing, then drop one and go on process the rest. For a list
with more than two elements, denote L00 as rest(rest(L)) = {l3 , l4 , ...}, we have

572

APPENDIX A. LISTS

the following equation.

l1 :
max(cons(l1 , L00 )) :

max(L0 ) :

l1 :

min(cons(l1 , L00 )) :
min(L) =

min(L0 ) :

max(L) =

|L| = 1
l 2 < l1
otherwise

(A.31)

|L| = 1
l1 < l 2
otherwise

(A.32)

The relative example Haskell programs are given as below.

min [x] = x
min (x:y:xs) = if x < y then min (x:xs) else min (y:xs)
max [x] = x
max (x:y:xs) = if x < y then max (y:ys) else max (x:xs)

Exercise A.1
Given two lists L1 and L2 , design a algorithm eq(L1 , L2 ) to test if they
are equal to each other. Here equality means the lengths are same, and
at the same time, every elements in both lists are identical.
Consider varies of options to handle the out-of-bound error case when
randomly access the element in list. Realize them in both imperative
and functional programming languages. Compare the solutions based on
exception and error code.
Augment the list with a tail field, so that the appending algorithm can
be realized in constant O(1) time but not linear O(N ) time. Feel free to
choose your favorite imperative programming language. Please dont refer
to the example source code along with this book before you try it.
With tail field augmented to list, for which list operations this field must
be updated? How it affects the performance?
Handle the out-of-bound case in insertion algorithm by treating it as appending.
Write the insertion sort algorithm by only using less than (<).
Design and implement the algorithm that find all the occurrence of a given
value and delete them from the list.
Reimplenent the algorithm to calculate the length of a list in tail-call
recursion manner.
Implement the insertion sort in tail recursive manner.
Implement the O(lg N ) algorithm to calculate bN in your favorite imperative programming language. Note that we only need accumulate the
intermediate result when the bit is not zero.

A.4. TRANSFORMATION

A.4

573

Transformation

In previous section, we list some basic operations for linked-list. In this section,
we focus on the transformation algorithms for list. Some of them are corner
stones of abstraction for functional programming. Well show how to use list
transformation to solve some interesting problems.

A.4.1

mapping and for-each

It is every-day programming routine that, we need output something as readable

string. If we have a list of numbers, and we want to print the list to console
like 3 1 2 5 4. One option is to convert the numbers to strings, so that we can
feed them to the printing function. One such trivial conversion program may
like this.

: L=
(A.33)
toStr(L) =
cons(str(l1 ), toStr(L0 )) : otherwise
The other example is that we have a dictionary which is actually a list of
words grouped in their initial letters, for example: [[a, an, another, ... ], [bat,
bath, bool, bus, ...], ..., [zero, zoo, ...]]. We want to know the frequency of them
in English, so that we process some English text, for example, Hamlet or the
Bible and augment each of the word with a number of occurrence in these
texts. Now we have a list like this:
[[(a, 1041), (an, 432), (another, 802), ... ],
[(bat, 5), (bath, 34), (bool, 11), (bus, 0), ...],
...,
[(zero 12), (zoo, 0), ...]]
If we want to find which word in each initial is used most, how to write a
program to work this problem out? The output is a list of words that every one
has the most occurrence in the group, which is categorized by initial, something
like [a, but, can, ...]. We actually need a program which can transfer a list of
group of augmented words into a list of words.
Lets work it out step by step. First, we need define a function, which takes
a list of word - number pairs, and find the word has the biggest number augmented. Sorting is overkill. What we need is just a special max0 () function, Note
that the max() function developed in previous section cant be used directly.
Suppose for a pair of values p = (a, b), function f st(p) = a, and snd(p) = b are
accessors to extract the values, max0 () can be defined as the following.

l1 : |L| = 1

l1 : snd(max0 (L0 )) < snd(l1 )

max0 (L) =
(A.34)

0
max (L0 ) : otherwise
Alternatively, we can define a dedicated function to compare word-number
of occurrence pair, and generalize the max() function by passing a compare
function.
less(p1 , p2 ) = snd(p1 ) < snd(p2 )

(A.35)

574

APPENDIX A. LISTS

: |L| = 1
: cmp(l1 , maxBy(cmp, L0 ))
: otherwise
(A.36)
Then max0 () is just a special case of maxBy() with the compare function
comparing on the second value in a pair.
maxBy(cmp, L) =

l1
l1

maxBy(cmp, L0 )

max0 (L) = maxBy(less, L)

(A.37)

Here we write all functions in purely recursive way, they can be modified in
tail call manner. This is left as exercise to the reader.
With max0 () function defined, its possible to complete the solution by processing the whole list.

solve(L) =

: L=
cons(f st(max0 (l1 )), solve(L0 )) : otherwise

(A.38)

Map
Compare the solve() function in (A.38) and toStr() function in (A.33), it reveals very similar algorithm structure. although they targets on very different
problems, and one is trivial while the other is a bit complex.
The structure of toStr() applies the function str() which can turn a number
into string on every element in the list; while solve() first applies max0 () function
to every element (which is actually a list of pairs), then applies f st() function,
which essentially turns a list of pairs into a string. It is not hard to abstract
such common structure like the following equation, which is called as mapping.

: L=
map(f, L) =
(A.39)
cons(f (l1 )), map(f, L0 )) : otherwise
Because map takes a converter function f as argument, its called a kind of
high-order function. In functional programming environment such as Haskell,
mapping can be implemented just like the above equation.
map :: (ab)[a][b]
map _ [] = []
map f (x:xs) = f x : map f xs

The two concrete cases we discussed above can all be represented in high
order mapping.
toStr = map str
solve = map (f st max0 )
Where f g means function composing, that we first apply g then apply f .
For instance function h(x) = f (g(x)) can be represented as h = f g, reading
like function h is composed by f and g. Note that we use Curried form to omit
the argument L for brevity. Informally speaking, If we feed a function which
needs 2 arguments, for instance f (x, y) = z with only 1 argument, the result
turns to be a function which need 1 argument. For instance, if we feed f with

A.4. TRANSFORMATION

575

only argument x, it turns to be a new function take one argument y, defined as

g(y) = f (x, y), or g = f x. Note that x isnt a free variable any more, as it is
bound to a value. Reader can refer to any book about functional programming
for details about function composing and Currying.
Mapping can also be understood from the domain theory point of view.
Consider function y = f (x), it actually defines a mapping from domain of
variable x to the domain of value y. (x and y can have different types). If the
domains can be represented as set X, and Y , we have the following relation.
Y = {f (x)|x X}

(A.40)

This type of set definition is called Zermelo Frankel set abstraction (as known
as ZF expression) [7]. The different is that here the mapping is from a list to
another list, so there can be duplicated elements. In languages support list
comprehension, for example Haskell and Python etc (Note that the Python list
is a built-in type, but not the linked-list we discussed in this appendix), mapping
can be implemented as a special case of list comprehension.
map f xs = [ f x | x xs]

List comprehension is a powerful tool. Here is another example that realizes

the permutation algorithm in list comprehension. Many textbooks introduce
how to implement all-permutation for a list, such as [7], and [9]. It is possible
to design a more general version perm(L, r), that if the length of the list L is
N , this algorithm permutes r elements from the total N elements. We know
!
that there are PNr = (NNr)!
solutions.

: r = 0 |L| < r
: otherwise
(A.41)
In this equation, {l} P means cons(l, P ), and L {l} denotes delete(L, l),
which is defined in previous section. If we take zero element for permutation,
or there are too few elements (less than r), the result is a list contains a empty
list; Otherwise for non-trivial case, the algorithm picks one element l from the
list, and recursively permutes the rest N 1 elements by picking up r 1 ones;
then it puts all the possible l in front of all the possible r 1 permutations.
Here is the Haskell implementation of this algorithm.

perm(L, r) =

{}
{{l} P |l L, P perm(L {l}, r 1)}

perm _ 0 = [[]]
perm xs r | length xs < r = [[]]
| otherwise = [ x:ys | x xs, ys perm (delete x xs) (r-1)]

Well go back to the list comprehension later in section about filtering.

Mapping can also be realized imperatively. We can apply the function while
traversing the list, and construct the new list from left to right. Since that the
new element is appended to the result list, we can track the tail position to
achieve constant time appending, so the mapping algorithms is linear in terms
of the passed in function.
function Map(f, L)
L0
p
while L 6= do

576

APPENDIX A. LISTS

if p = then
p Cons(f ( First(L) ), )
L0 p
else
Next(p) Cons(f ( First(L) ), )
p Next(p)
L Next(L)
return L0
Because It is a bit complex to annotate the type of the passed-in function
in ISO C++, as it involves some detailed language specific features. See [11]
for detail. In fact ISO C++ provides the very same mapping concept as in
std::transform. However, it needs the reader have knowledge of function
object, iterator etc, which are out of the scope of this book. Reader can refer
to any ISO C++ STL materials for detail.
For brevity purpose, we switch to Python programming language for example
code. So that the type inference can be avoid in compile time. The definition
of a simple singly linked-list in Python is give as the following.
class List:
def __init__(self, x = None, xs = None):
self.key = x
self.next = xs
def cons(x, xs):
return List(x, xs)

The mapping program, takes a function and a linked-list, and maps the
functions to every element as described in above algorithm.
def mapL(f, xs):
ys = prev = List()
while xs is not None:
prev.next = List(f(xs.key))
prev = prev.next
xs = xs.next
return ys.next

Different from the pseudo code, this program uses a dummy node as the head
of the resulting list. So it neednt test if the variable stores the last appending
position is NIL. This small trick makes the program compact. We only need
drop the dummy node before returning the result.
For each
For the trivial task such as printing a list of elements out, its quite OK to just
print each element without converting the whole list to a list of strings. We can
actually simplify the program.
function Print(L)
while L 6= do
print First(L)
L Rest(L)
More generally, we can pass a procedure such as printing, to this list traverse,
so the procedure is performed for each element.

A.4. TRANSFORMATION

577

function For-Each(L, P )
while L 6= do
P(First(L))
L Rest(L)
For-each algorithm can be formalized in recursive approach as well.

f oreach(L, p) =

u : L=
do(p(l1 ), f oreach(L0 , p)) : otherwise

(A.42)

Here u means unit, its can be understood as doing nothing, The type of it
is similar to the void concept in C or java like programming languages. The
do() function evaluates all its arguments, discards all the results except for the
last one, and returns the last result as the final value of do(). It is equivalent to
(begin ...) in Lisp families, and do block in Haskell in some sense. For the
details about unit type, please refer to [4].
Note that the for-each algorithm is just a simplified mapping, there are only
two minor difference points:
It neednt form a result list, we care the side effect rather than the
returned value;
For each focus more on traversing, while mapping focus more on applying
function, thus the order of arguments are typically arranged as map(f, L)
and f oreach(L, p).
Some Functional programming facilities provides options for both returning
the result list or discarding it. For example Haskell Monad library provides both
mapM, mapM_ and forM, forM_. Readers can refer to language specific materials
for detail.
Examples for mapping
Well show how to use mapping by an example, which is a problem of ACM/ICPC[12].
For sake of brevity, we modified the problem description a bit. Suppose there
are N lights in a room, all of them are off. We execute the following process N
times:
1. We switch all the lights in the room, so that they are all on;
2. We switch the 2, 4, 6, ... lights, that every other light is switched, if the
light is on, it will be off, and it will be on if the previous state is off;
3. We switch every third lights, that the 3, 6, 9, ... are switched;
4. ...
And at the last round, only the last light (the N -th light) is switched.
The question is how many lights are on finally?
Before we show the best answer to this puzzle, lets first work out a naive
brute-force solution. Suppose there are N lights, which can be represented as a
list of 0, 1 numbers, where 0 means the light is off, and 1 means on. The initial
state is a list of N zeros: {0, 0, ..., 0}.

578

APPENDIX A. LISTS

We can label the light from 1 to N . A mapping can help us to turn the
above list into a labeled list6 .
map(i (i, 0), {1, 2, 3, ...N })
This mapping augments each natural number with zero, the result is a list
of pairs: L = {(1, 0), (2, 0), ..., (N, 0)}.
Next we operate this list of pairs N times from 1 to N . For every time i,
we switch the second value in this pair if the first label can be divided by i.
Consider the fact that 1 0 = 1, and 1 1 = 0, we can realize switching of 0, 1
value x by 1 x. At the i-th operation, for light (j, x), if i|j, (or j mod i = 0),
we then perform switching, otherwise, we leave the light untouched.

(j, 1 x) : j mod i = 0
switch(i, (j, x)) =
(A.43)
(j, x) : otherwise
The i-th operation on all lights can be realized as mapping again:
map(switch(i), L)

(A.44)

Note that, here we use Curried form of switch() function, which is equivalent
to
map((j,x) switch(i, (j, x)), L)
Here we need define a function proc(), which can perform the above mapping
on L over and over by N times. One option is to realize it in purely recursive
way as the following, so that we can call it like proc({1, 2, ..., N }, L)7 .

proc(I, L) =

L : I=
operate(I 0 , map(switch(i1 ), L)) : otherwise

(A.45)

Where I = cons(i1 , I 0 ) if I isnt empty.

At this stage, we can sum the second value of each pair in list L to get the
answer. The sum function has been defined in previous section, so the only
thing left is mapping.
solve(N ) = sum(map(snd, proc({1, 2, ..., N }, L)))

(A.46)

Translating this naive brute-force solution to Haskell yields below program.

solve = sum (map snd) proc where
proc n = operate [1..n] $ map (i (i, 0)) [1..n]
operate [] xs = xs
operate (i:is) xs = operate is (map (switch i) xs)

Lets see whats the answer for there are 1, 2, ..., 100 lights.
[1,1,1,2,2,2,2,2,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,5,5,5,5,5,5,5,5,5,5,5,
6,6,6,6,6,6,6,6,6,6,6,6,6,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,
8,8,8,8,8,8,8,8,8,8,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,10]
6 Readers who are familiar with functional programming, may use zipping to achieve this.
Well explain zipping in later section.
7 This can also be realized by folding, which will be explained in later section.

A.4. TRANSFORMATION

579

This result is interesting:

the first 3 answers are 1;
the 4-th to the 8-th answers are 2;
the 9-th to the 15-th answers are 3;
...
It seems that the i2 -th to the ((i + 1)2 1)-th answers are i. Actually, we
can prove this fact as the following.
Proof. Given N lights, labeled from 1 to N , consider which lights are on finally.
Since the initial states for all lights are off, we can say that, the lights which
are manipulated odd times are on. For every light i, it will be switched at the
j round if i can be divided by j (denote as j|i). So only the lights which have
odd number of factors are on at the end.
So the key point to solve this puzzle, is to find all numbers which have odd
number of factors. For any positive integer N , denote S the set of all factors of
N . S is initialized to . if p is a factor of N , there must exist a positive integer
q that N = pq, which means q is also a factor of N . So we add 2 different factors
to the set S if and only if p 6= q, which keeps |S| even all the time unless p = q.
In such case, N is a perfect square number, and we can only add 1 factor to the
set S, which leads to an odd number of factors.
At this stage, we can design a fast solution by finding the number of perfect
square numbers under N .

solve(N ) = b N c

(A.47)

The next Haskell command verifies that the answer for 1, 2, ..., 100 lights
are as same as above.
map (floor.sqrt) [1..100]
[1,1,1,2,2,2,2,2,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,5,5,5,5,5,5,5,5,5,5,5,
6,6,6,6,6,6,6,6,6,6,6,6,6,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,
8,8,8,8,8,8,8,8,8,8,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,10]

Mapping is generic concept that it doesnt only limit in linked-list, but also
can be applied to many complex data structures. The chapter about binary
search tree in this book explains how to map on trees. As long as we can
traverse a data structure in some order, and the empty data structure can be
identified, we can use the same mapping idea. Well return to this kind of
high-order concept in the section of folding later.

A.4.2

reverse

How to reverse a singly linked-list with minimum space is a popular technical interview problem in some companies. The pointer manipulation must be
arranged carefully in imperative programming languages such as ANSI C. However, well show that, there exists an easy way to write this program:
1. Firstly, write a pure recursive straightforward solution;

580

APPENDIX A. LISTS

2. Then, transform the pure recursive solution to tail-call manner;

3. Finally, translate the tail-call solution to pure imperative pointer operations.
The pure recursive solution is simple enough that we can write it out immediately. In order to reverse a list L.
If L is empty, the reversed result is empty. This is the trivial edge case;
Otherwise, we can first reverse the rest of the sub-list, then append the
first element to the end.
This idea can be formalized to the below equation.

reverse(L) =

: L=
append(reverse(L0 ), l1 ) : otherwise

(A.48)

Translating it to Haskell yields below program.

reverse [] = []
reverse (x:xs) = reverse xs ++ [x]

However, this solution doesnt perform well, as appending has to traverse

to the end of list, which leads to a quadratic time algorithm. It is not hard to
improve this program by changing it to tail-call manner. That we can use a accumulator to store the intermediate reversed result, and initialize the accumulated
result as empty. So the algorithm is formalized as reverse(L) = reverse0 (L, ).

reverse0 (L, A) =

A : L=
reverse0 (L0 , {l1 } A) : otherwise

(A.49)

Where {l1 } A means cons(l1 , A). Different from appending, its a constant
O(1) time operation. The core idea is that we repeatedly take the element one
by one from the head of the original list, and put them in front the accumulated
result. This is just like we store all the elements in a stack, them pop them out.
This is a linear time algorithm.
Below Haskell program implements this tail-call version.
reverse [] acc = acc
reverse (x:xs) acc = reverse xs (x:acc)

Since the nature of tail-recursion call neednt book-keep any context (typically by stack), most modern compilers are able to optimize it to a pure imperative loop, and reuse the current context and stack etc. Lets manually do this
optimization so that we can get a imperative algorithm.
function Reverse(L)
A
while L 6= do
A Cons(First(L), A)
L Rest(L)

A.5. EXTRACT SUB-LISTS

581

However, because we translate it directly from a functional solution, this

algorithm actually produces a new reversed list, but does not mutate the original
one. It is not hard to change it to an in-place solution by reusing L. For example,
the following ISO C++ program implements the in-place algorithm. It takes
O(1) memory space, and reverses the list in O(N ) time.
template<typename T>
List<T> reverse(List<T> xs) {
List<T> p, ys = NULL;
while (xs) {
p = xs;
xs = xsnext;
pnext = ys;
ys = p;
}
return ys;
}

Exercise A.2
Implement the algorithm to find the maximum element in a list of pair in
tail call approach in your favorite programming language.

A.5

Extract sub-lists

Different from arrays which are capable to slice a continuous segment fast and
easily, It needs more work to extract sub lists from singly linked list. Such
operations are typically linear algorithms.

A.5.1

take, drop, and split-at

Taking first N elements from a list is semantically similar to extract sub list from
the very left like sublist(L, 1, N ), where the second and the third arguments to
sublist are the positions the sub-list starts and ends. For the trivial edge case,
that either N is zero or the list is empty, the sub list is empty; Otherwise, we
can recursively take the first N 1 elements from the rest of the list, and put
the first element in front of it.

take(N, L) =

: L=N =0
cons(l1 , take(N 1, L0 )) : otherwise

(A.50)

Note that the edge cases actually handle the out-of-bound error. The following Haskell program implements this algorithm.
take _ [] = []
take 0 _ = []
take n (x:xs) = x : take (n-1) xs

Dropping on the other hand, drops the first N elements and returns the left
as result. It is equivalent to get the sub list from right like sublist(L, N + 1, |L|),

582

APPENDIX A. LISTS

where |L| is the length of the list. Dropping can be designed quite similar to
taking by discarding the first element in the recursive case.

: L=

L : N =0
(A.51)
drop(N, L) =

drop(N 1, L0 )) : otherwise
Translating the algorithm to Haskell gives the below example program.
drop _ [] = []
drop 0 L = L
drop n (x:xs) = drop (n-1) xs

The imperative taking and dropping are quite straight-forward, that they
are left as exercises to the reader.
With taking and dropping defined, extracting sub list at arbitrary position
for arbitrary length can be realized by calling them.
sublist(L, f rom, count) = take(count, drop(f rom 1, L))

(A.52)

or in another semantics by providing left and right boundaries:

sublist(L, f rom, to) = drop(f rom 1, take(to, L))

(A.53)

Note that the elements in range [f rom, to] is returned by this function, with
both ends included. All the above algorithms perform in linear time.
take-while and drop-while
Compare to taking and dropping, there is another type of operation, that we
either keep taking or dropping elements as far as a certain condition is met. The
taking and dropping algorithms can be viewed as special cases for take-while
and drop-while.
Take-while examines elements one by one as far as the condition is satisfied,
and ignore all the rest of elements even some of them satisfy the condition.
This is the different point from filtering which well explained in later section.
Take-while stops once the condition tests fail; while filtering traverses the whole
list.

takeW hile(p, L) =

:
:

cons(l1 , takeW hile(p, L0 )) :

L=
p(l1 )
otherwise

(A.54)

Take-while accepts two arguments, one is the predicate function p, which

can be applied to element in the list and returns Boolean value as result; the
other argument is the list to be processed.
It is easy to define the drop-while symmetrically.

: L=

L : p(l1 )
dropW hile(p, L) =
(A.55)

dropW hile(p, L0 ) : otherwise

The corresponding Haskell example programs are given as below.

A.5. EXTRACT SUB-LISTS

583

takeWhile _ [] = []
takeWhile p (x:xs) = if p x then x : takeWhile p xs else []
dropWhile _ [] = []
dropWhile p xs@(x:xs) = if p x then dropWhile p xs else xs

split-at
With taking and dropping defined, splitting-at can be realized trivially by calling
them.
splitAt(i, L) = (take(i, L), drop(i, L))

A.5.2

(A.56)

breaking and grouping

breaking
Breaking can be considered as a general form of splitting, instead of splitting
at a given position, breaking examines every element for a certain predicate,
and finds the longest prefix of the list for that condition. The result is a pair of
sub-lists, one is that longest prefix, the other is the rest.
There are two different breaking semantics, one is to pick elements satisfying
the predicate as long as possible; the other is to pick those dont satisfy. The
former is typically defined as span, while the later as break.
Span can be described, for example, in such recursive manner: In order to
span a list L for predicate p:
If the list is empty, the result for this edge trivial case is a pair of empty
lists (, );
Otherwise, we test the predicate against the first element l1 , if l1 satisfies
the predicate, we denote the intermediate result for spanning the rest of
list as (A, B) = span(p, L0 ), then we put l1 in front of A to get pair
({l1 } A, B), otherwise, we just return (, L) as the result.
For breaking, we just test the negate of predicate and all the others are as
same as spanning. Alternatively, one can define breaking by using span as in
the later example program.

span(p, L) =

(, ) : L =
({l1 } A, B) : p(l1 ) = T rue, (A, B) = span(p, L0 )

(, L) : otherwise
(A.57)

break(p, L) =

(, ) :
({l1 } A, B) :

(, L) :

L=
p(l1 ), (A, B) = break(p, L0 )
otherwise

(A.58)

Note that both functions only find the longest prefix, they stop immediately
when the condition is fail even if there are elements in the rest of the list meet
the predicate (or not). Translating them to Haskell gives the following example
program.

584

APPENDIX A. LISTS

span _ [] = ([], [])

span p xs@(x:xs) = if p x then let (as, bs) = span xs in (x:as, bs) else ([], xs)
break p = span (not p)

Span and break can also be realized imperatively as the following.

function Span(p, L)
A
while L 6= p(l1 ) do
A Cons(l1 , A)
L Rest(L)
return (A, L)
function Break(p, L)
return Span(p, L)
This algorithm creates a new list to hold the longest prefix, another option is
to turn it into in-place algorithm to reuse the spaces as in the following Python
example.
def span(p, xs):
ys = xs
last = None
while xs is not None and p(xs.key):
last = xs
xs = xs.next
if last is None:
return (None, xs)
last.next = None
return (ys, xs)

Note that both span and break need traverse the list to test the predicate,
thus they are linear algorithms bound to O(N ).
grouping
Grouping is a commonly used operation to solve the problems that we need
divide the list into some small groups. For example, Suppose we want to group
the string Mississippi, which is actual a list of char { M, s, s, i, s, s,
i, p, p, i}. into several small lists in sequence, that each one contains
consecutive identical characters. The grouping operation is expected to be:
group(Mississippi) = { M, i, ss, i, ss, i, pp, i}
Another example, is that we have a list of numbers:
L = {15, 9, 0, 12, 11, 7, 10, 5, 6, 13, 1, 4, 8, 3, 14, 2}
We want to divide it into several small lists, that each sub-list is ordered
descending. The grouping operation is expected to be :
group(L) = {{15, 9, 0}, {12, 11, 7}, {10, 5}, {6}, {13, 1}, {4}, {8, 3}, {14, 2}}

A.5. EXTRACT SUB-LISTS

585

Both cases play very important role in real algorithms. The string grouping
is used in creating Trie/Patricia data structure, which is a powerful tool in string
searching area; The ordered sub-list grouping can be used in nature merge sort.
There are dedicated chapters in this book explain the detail of these algorithms.
It is obvious that we need abstract the grouping condition so that we know
where to break the original list into small ones. This predicate can be passed to
the algorithm as an argument like group(p, L), where predicate p accepts two
consecutive elements and test if the condition matches.
The first idea to solve the grouping problem is traversing takes two elements
at each time, if the predicate test succeeds, put both elements into a small
group; otherwise, only put the first one into the group, and use the second one
to initialize another new group. Denote the first two elements (if there are) are
l1 , l2 , and the sub-list without the first element as L0 . The result is a list of list
G = {g1 , g2 , ...}, denoted as G = group(p, L).

{}
{{l1 }}
group(p, L) =
{{l1 } g10 , g20 , ...}

{{l1 }, g10 , g20 , ...}

:
:
:
:

L=
|L| = 1
p(l1 , l2 ), G0 = group(p, L0 ) = {g10 , g20 , ...}
otherwise
(A.59)
Note that {l1 } g10 actually means cons(l1 , g10 ), which performs in constant
time. This is a linear algorithm performs proportion to the length of the list, it
traverses the list in one pass which is bound to O(N ). Translating this program
to Haskell gives the below example code.
group _ [] = [[]]
group _ [x] = [[x]]
group p (x:xs@(x:_)) | p x x = (x:ys):yss
| otherwise = [x]:r
where
r@(ys:yss) = group p xs

It is possible to implement this algorithm in imperative approach, that we

initialize the result groups as {l1 } if L isnt empty, then we traverse the list from
the second one, and append to the last group if the two consecutive elements
satisfy the predicate; otherwise we start a new group.
function Group(p, L)
if L = then
return {}
x First(L)
L Rest(L)
g {x}
G {g}
while L 6= do
y First(L)
if p(x, y) then
g Append(g, y)
else
g {y}
G Append(G, g)

586

APPENDIX A. LISTS

xy
L Next(L)
return G
However, different from the recursive algorithm, this program performs in
quadratic time if the appending function isnt optimized by storing the tail
position. The corresponding Python program is given as below.
def group(p, xs):
if xs is None:
return List(None)
(x, xs) = (xs.key, xs.next)
g = List(x)
G = List(g)
while xs is not None:
y = xs.key
if p(x, y):
g = append(g, y)
else:
g = List(y)
G = append(G, g)
x=y
xs = xs.next
return G

With the grouping function defined, the two example cases mentioned at the
beginning of this section can be realized by passing different predictions.
group(=, {m, i, s, s, i, s, s, i, p, p, i}) = {{M }, {i}, {ss}, {i}, {ss}, {i}, {pp}, {i}}

group(, {15, 9, 0, 12, 11, 7, 10, 5, 6, 13, 1, 4, 8, 3, 14, 2})

= {{15, 9, 0}, {12, 11, 7}, {10, 5}, {6}, {13, 1}, {4}, {8, 3}, {14, 2}}
Another solution is to use the span function we have defined to realize grouping. We pass a predicate to span, which will break the list into two parts: The
first part is the longest sub-list satisfying the condition. We can repeatedly
apply the span with the same predication to the second part, till it becomes
empty.
However, the predicate function we passed to span is an unary function, that
it takes an element as argument, and test if it satisfies the condition. While in
grouping algorithm, the predicate function is a binary function. It takes two
adjacent elements for testing. The solution is that, we can use currying and pass
the first element to the binary predicate, and use it to test the rest of elements.

group(p, L) =

{} : L =
{{l1 } A} group(p, B) : otherwise

(A.60)

Where (A, B) = span(x p(l1 , x), L0 ) is the result of spanning on the rest
sub-list of L.
Although this new defined grouping function can generate correct result for
the first case as in the following Haskell code snippet.

A.5. EXTRACT SUB-LISTS

587

groupBy (==) "Mississippi"

["m","i","ss","i","ss","i","pp","i"]

However, it seems that this algorithm cant group the list of numbers into
ordered sub lists.
groupBy ( ) [15, 9, 0, 12, 11, 7, 10, 5, 6, 13, 1, 4, 8, 3, 14, 2]
[[15,9,0,12,11,7,10,5,6,13,1,4,8,3,14,2]]

The reason is because that, as the first element 15 is used as the left parameter to operator for span, while 15 is the maximum value in this list, so the
span function ends with putting all elements to A, and B is left empty. This
might seem a defect, but it is actually the correct behavior if the semantic is to
group equal elements together.
Strictly speaking, the equality predicate must satisfy three properties: reflexive, transitive, and symmetric. They are specified as the following.
Reflexive. x = x, which says that any element is equal to itself;
Transitive. x = y, y = z x = z, which says that if two elements are
equal, and one of them is equal to another, then all the tree are equal;
Symmetric. x = y y = x, which says that the order of comparing two
equal elements doesnt affect the result.
When we group character list Mississippi, the equal (=) operator is used,
which obviously conforms these three properties. So that it generates correct
grouping result. However, when passing () as equality predicate, to group a
list of numbers, it violets both reflexive and symmetric properties, that is reason
why we get wrong grouping result.
This fact means that the second algorithm we designed by using span, limits
the semantic to strictly equality, while the first one does not. It just tests the
condition for every two adjacent elements, which is much weaker than equality.

Exercise A.3
1. Implement the in-place imperative taking and dropping algorithms in your
favorite programming language, note that the out of bound cases should
be handled. Please try both languages with and without GC (Garbage
Collection) support.
2. Implement take-while and drop-while in your favorite imperative programming language. Please try both dynamic type language and static type
language (with and without type inference). How to specify the type of
predicate function as generic as possible in static type system?
3. Consider the following definition of

(, ) :

({l1 } A, B) :
span(p, L) =

(A, {l1 } B) :

span.
L=
p(l1 ) = T rue, (A, B) = span(p, L0 )
otherwise

Whats the difference between this algorithm and the the one weve shown
in this section?
4. Implement the grouping algorithm by using span in imperative way in
your favorite programming language.

588

APPENDIX A. LISTS

A.6

Folding

We are ready to introduce one of the most critical concept in high order programming, folding. It is so powerful tool that almost all the algorithms so far
in this appendix can be realized by folding. Folding is sometimes be named as
reducing (the abstracted concept is identical to the buzz term map-reduce in
cloud computing in some sense). For example, both STL and Python provide
reduce function which realizes partial form of folding.

A.6.1

folding from right

Remind the sum and product definition in previous section, they are quite similar actually.

0 : L=
sum(L) =
l1 + sum(L0 ) : otherwise

1 : L=
product(L) =
l1 product(L0 ) : otherwise
It is obvious that they have same structure. Whats more, if we list the
insertion sort definition, we can find that it also shares this structure.

: L=
sort(L) =
insert(l1 , sort(L0 )) : otherwise
This hint us that we can abstract this essential common structure, so that
we neednt repeat it again and again. Observing sum, product, and sort, there
are two different points which we can parameterize.
The result of the trivial edge case varies. It is zero for sum, 1 for product,
and empty list for sorting.
The function applied to the first element and the intermediate result varies.
It is plus for sum, multiply for product, and ordered-insertion for sorting.
If we parameterize the result of trivial edge case as initial value z (stands for
abstract zero concept), the function applied in recursive case as f (which takes
two parameters, one is the first element in the list, the other is the recursive
result for the rest of the list), this common structure can be defined as something
like the following.

z : L=
proc(f, z, L) =
f (l1 , proc(f, z, L0 )) : otherwise
Thats it, and we should name this common structure a better name instead
of the meaningless proc. Lets see the characteristic of this common structure.
For list L = {x1 , x2 , ..., xN }, we can expand the computation like the following.
proc(f, z, L) = f (x1 , proc(f, z, L0 )
= f (x1 , f (x2 , proc(f, z, L00 ))
...
= f (x1 , f (x2 , f (..., f (xN , f (f, z, ))...)
= f (x1 , f (x2 , f (..., f (xN , z))...)

A.6. FOLDING

589

Since f takes two parameters, its a binary function, thus we can write it in
infix form. The infix form is defined as below.
x f y = f (x, y)

(A.61)

The above expanded result is equivalent to the following by using infix notion.
proc(f, z, L) = x1 f (x2 f (...(xN f z))...)
Note that the parentheses are necessary, because the computation starts
from the right-most (xN f z), and repeatedly fold to left towards x1 . This is
quite similar to folding a Chinese hand-fan as illustrated in the following photos.
A Chinese hand-fan is made of bamboo and paper. Multiple bamboo frames are
stuck together with an axis at one end. The arc shape paper is fully expanded
by these frames as shown in Figure A.3 (a); The fan can be closed by folding
the paper. Figure A.3 (b) shows that some part of the fan is folded from right.
After these folding finished, the fan results a stick, as shown in Figure A.3 (c).

(a) A folding fan fully opened.

(b) The fan is partly folded on right.

(c) The fan is fully folded, closed to a stick.

Figure A.3: Folding a Chinese hand-fan

We can considered that each bamboo frame along with the paper on it as an
element, so these frames forms a list. A unit process to close the fan is to rotate

590

APPENDIX A. LISTS

a frame for a certain angle, so that it lays on top of the collapsed part. When
we start closing the fan, the initial collapsed result is the first bamboo frame.
The close process is folding from one end, and repeatedly apply the unit close
steps, till all the frames is rotated, and the folding result is a stick closed form.
Actually, the sum and product algorithms exactly do the same thing as
closing the fan.
sum({1, 2, 3, 4, 5}) = 1 + (2 + (3 + (4 + 5)))
= 1 + (2 + (3 + 9))
= 1 + (2 + 12)
= 1 + 14
= 15
product({1, 2, 3, 4, 5}) = 1 (2 (3 (4 5)))
= 1 (2 (3 20))
= 1 (2 60)
= 1 120
= 120
In functional programming, we name this process folding, and particularly,
since we execute from the most inner structure, which starts from the right-most
one. This type of folding is named folding right.

f oldr(f, z, L) =

z : L=
f (l1 , f oldr(f, z, L0 )) : otherwise

(A.62)

Lets see how to use fold-right to realize sum and product.

PN
i=1

QN
i=1

= x1 + (x2 + (x3 + ... + (xN1 + xN ))...)

= f oldr(+, 0, {x1 , x2 , ..., xN })

(A.63)

= x1 (x2 (x3 ... + (xN1 xN ))...)

= f oldr(, 1, {x1 , x2 , ..., xN })

(A.64)

The insertion-sort algorithm can also be defined by using folding right.

sort(L) = f oldr(insert, , L)

A.6.2

(A.65)

folding from left

As mentioned in section of tail recursive call. Both pure recursive sum and
product compute from right to left and they must book keep all the intermediate
results and contexts. As we abstract fold-right from the very same structure,
folding from right does the book keeping as well. This will be expensive if the
list is very long.
Since we can change the realization of sum and product to tail-recursive call
manner, it quite possible that we can provide another folding algorithm, which
processes the list from left to right in normal order, and enable the tail-call
optimization by reusing the same context.

A.6. FOLDING

591

Instead of induction from sum, product and insertion, we can directly change
the folding right to tail call. Observe that the initial value z, actually represents
the intermediate result at any time. We can use it as the accumulator.

f oldl(f, z, L) =

z : L=
f oldl(f, f (z, l1 ), L0 ) : otherwise

(A.66)

Every time when the list isnt empty, we take the first element, apply function
f on the accumulator z and it to get a new accumulator z 0 = f (z, l1 ). After
that we can repeatedly folding with the very same function f , the updated
accumulator z 0 , and list L0 .
Lets verify that this tail-call algorithm actually folding from left.
P5
i=1

i = f oldl(+, 0, {1, 2, 3, 4, 5})

= f oldl(+, 0 + 1, {2, 3, 4, 5})
= f oldl(+, (0 + 1) + 2{3, 4, 5}
= f oldl(+, ((0 + 1) + 2) + 3, {4, 5})
= f oldl(+, (((0 + 1) + 2) + 3) + 4, {5})
= f oldl(+, ((((0 + 1) + 2 + 3) + 4 + 5, )
=0+1+2+3+4+5

Note that, we actually delayed the evaluation of f (z, l1 ) in every step. (This
is the exact behavior in system support lazy-evaluation, for instance, Haskell.
However, in strict system such as standard ML, its not the case.) Actually,
they will be evaluated in sequence of {1, 3, 6, 10, 15} in each call.
Generally, folding-left can be expanded in form of
f oldl(f, z, L) = f (f (...(f (f (z, l1 ), l2 ), ..., lN )

(A.67)

Or in infix manner as
f oldl(f, z, L) = ((...(z f l1 ) f l2 ) f ...) lN

(A.68)

With folding from left defined, sum, product, and insertion-sort can be transparently implemented by calling f oldl as sum(L) = f oldl(+, 0, L), product(L) =
f oldl(+, 1, L), and sort(L) = f oldl(insert, , L). Compare with the foldingright version, they are almost same at first glares, however, the internal implementation differs.
Imperative folding and generic folding concept
The tail-call nature of folding-left algorithm is quite friendly for imperative settings, that even the compiler isnt equipped with tail-call recursive optimization,
we can anyway implement the folding in while-loop manually.
function Fold(f, z, L)
while L 6= do
z f (z, First(L) )
L Rest(L)
return z
Translating this algorithm to Python yields the following example program.

592

APPENDIX A. LISTS

def fold(f, z, xs):

for x in xs:
z = f(z, x)
return z

Actually, Python provides built-in function reduce which does the very
same thing. (in ISO C++, this is provided as reduce algorithm in STL.) Almost
no imperative environment provides folding-right function because it will cause
stack overflow problem if the list is too long. However, there still exist cases
that the folding from right semantics is necessary. For example, one defines a
container, which only provides insertion function to the head of the container,
but there is no any appending method, so that we want such a f romList tool.
f romList(L) = f oldr(insertHead, empty, L)
Calling f romList with the insertion function as well as an empty initialized
container, can turn a list into the special container. Actually the singly linkedlist is such a container, which performs well on insertion to the head, but poor
to linear time if appending on the tail. Folding from right is quite nature when
duplicate a linked-list while keeps the elements ordering. While folding from
left will generate a reversed list.
In such cases, there exists an alternative way to implement imperative folding
right by first reverse the list, and then folding the reversed one from left.
function Fold-Right(f, z, L)
return Fold(f, z, Reverse(L))
Note that, here we must use the tail-call version of reversing, or the stack
overflow issue still exists.
One may think that folding-left should be chosen in most cases over foldingright because its friendly for tail-recursion call optimization, suitable for both
functional and imperative settings, and its an online algorithm. However,
folding-right plays a critical role when the input list is infinity and the binary
function f is lazy. For example, below Haskell program wraps every element in
an infinity list to a singleton, and returns the first 10 result.
take 10 $ foldr (x xs [x]:xs) [] [1..]
[[1], [2], [3], [4], [5], [6], [7], [8], [9], [10]]

This cant be achieved by using folding left because the outer most evaluation
cant be finished until all the list being processed. The details is specific to lazy
evaluation feature, which is out of the scope of this book. Readers can refer to
[13] for details.
Although the main topic of this appendix is about singly linked-list related
algorithms, the folding concept itself is generic which doesnt only limit to list,
but also can be applied to other data structures.
We can fold a tree, a queue, or even more complicated data structures as
long as we have the following:
The empty data structure can be identified for trivial edge case; (e.g.
empty tree)
We can traverse the data structure (e.g. traverse the tree in pre-order).

A.6. FOLDING

593

Some languages provide this high-level concept support, for example, Haskell
achieve this via monoid, readers can refer to [8] for detail.
There are many chapters in this book use the widen concept of folding.

A.6.3

folding in practice

We have seen that max, min, and insertion sort all can be realized in folding.
The brute-force solution for drunk jailer puzzle shown in mapping section can
also be designed by mixed use of mapping and folding.
Remind that we create a list of pairs, each pair contains the number of the
light, and the on-off state. After that we process from 1 to N , switch the light
if the number can be divided. The whole process can be viewed as folding.
f old(step, {(1, 0), (2, 0), ..., (N, 0)}, {1, 2, ..., N })
The initial value is the very first state, that all the lights are off. The list to
be folding is the operations from 1 to N . Function step takes two arguments,
one is the light states pair list, the other is the operation time i. It then maps
on all lights and performs switching. We can then substitute the step with
mapping.
f old(L,i map(switch(i), L), {(1, 0), (2, 0), ..., (N, 0)}, {1, 2, ..., N })
Well simplify the notation, and directly write map(switch(i), l) for brevity
purpose. The result of this folding is the final states pairs, we need take the
second one of the pair for each element via mapping, then calculate the summation.
sum(map(snd, f old(map(switch(i), L), {(1, 0), (2, 0), ..., (N, 0)}, {1, 2, ..., N })))
(A.69)
There are materials provides plenty of good examples of using folding, especially in [1], folding together with fusion law are well explained.
concatenate a list of list
In previous section A.3.6 about concatenation, we explained how to concatenate two lists. Actually, concatenation of lists can be considered equivalent to
summation of numbers. Thus we can design a general algorithm, which can
concatenate multiple lists into one big list.
Whats more, we can realize this general concatenation by using folding.
As sum can be represented as sum(L) = f oldr(+, 0, L), its straightforward to
write the following equation.
concats(L) = f oldr(concat, , L)

(A.70)

Where L is a list of list, for example {{1, 2, 3}, {4, 5, 6}, {7, 8, 9}, ...}. Function concat(L1 , L2 ) is what we defined in section A.3.6.
In some environments which support lazy-evaluation, such as Haskell, this
algorithm is capable to concatenate infinite list of list, as the binary function
++ is lazy.

Exercise A.4

594

APPENDIX A. LISTS
Whats the performance of concats algorithm? is it linear or quadratic?
Design another linear time concats algorithm without using folding.
Realize mapping algorithm by using folding.

A.7

Searching and matching

Searching and matching are very important algorithms. They are not only
limited to linked list, but also applicable to a wide range of data structures. We
just scratch the surface of searching and matching in this appendix. There are
dedicated chapters explain about them in this book.

A.7.1

Existence testing

The simplest searching case is to test if a given element exists in a list. A linear
time traverse can solve this problem. In order to determine element x exists in
list L:
If the list is empty, its obvious that the element doesnt exist in L;
If the first element in the list equals to x, we know that x exists;
Otherwise, we need recursively test if x exists in the rest sub-list L0 ;
This simple description can be directly formalized to equation as the following.

F alse : L =
T rue : l1 = x
(A.71)
xL=

x L0 : otherwise
This is definitely a linear algorithm which is bound to O(N ) time. The
best case happens in the two trivial clauses that either the list is empty or the
first element is what we are finding; The worst case happens when the element
doesnt exist at all or it is the last element. In both cases, we need traverse the
whole list. If the probability is equal for all the positions, the average case takes
about N2+1 steps for traversing.
This algorithm is so trivial that we left the implementation as exercise to
the reader. If the list is ordered, one may expect to improve the algorithm
to logarithm time but not linear. However, as we discussed, since list doesnt
support constant time random accessing, binary search cant be applied here.
There is a dedicated chapter in this book discusses how to evolve the linked list
to binary tree to achieve quick searching.

A.7.2

Looking up

One extra step from existence testing is to find the interesting information stored
in the list. There are two typical methods to augment extra data to the element.
Since the linked list is chain of nodes, we can store satellite data in the node,
then provide key(n) to access the key of the node, rest(n) for the rest sub-list,
and value(n) for the augmented data. The other method, is to pair the key

A.7. SEARCHING AND MATCHING

595

and data, for example {(1, hello), (2, world), (3, f oo), ...}. Well introduce how
to form such pairing list in later section.
The algorithm is almost as same as the existence testing, that it traverses
the list, examines the key one by one. Whenever it finds a node which has the
same key as what we are looking up, it stops, and returns the augmented data.
It is obvious that this is linear strategy. If the satellite data is augmented to
the node directly, the algorithm can be defined as the following.

: L=

value(l1 ) : key(l1 ) = x
lookup(x, L) =
(A.72)

lookup(x, L0 ) : otherwise
In this algorithm, L is a list of nodes which are augmented with satellite data.
Note that the first case actually means looking up failure, so that the result is
empty. Some functional programming languages, such as Haskell, provide Maybe
type to handle the possibility of fail. This algorithm can be slightly modified to
handle the key-value pair list as well.

: L=

snd(l1 ) : f st(l1 ) = x
(A.73)
lookup(x, L) =

lookup(x, L0 ) : otherwise
Here L is a list of pairs, functions f st(p) and snd(p) access the first part and
second part of the pair respectively.
Both algorithms are in tail-call manner, they can be transformed to imperative looping easily. We left this as exercise to the reader.

A.7.3

finding and filtering

Lets take one more step ahead, looking up algorithm performs linear search by
comparing the key of an element is equal to the given value. A more general
case is to find an element matching a certain predicate. We can abstract this
matching condition as a parameter for this generic linear finding algorithm.

: L=

l1 : p(l1 )
f ind(p, L) =
(A.74)

f ind(p, L0 ) : otherwise
The algorithm traverses the list by examining if the element satisfies the
predicate p. It fails if the list is empty while there is still nothing found. This is
handled in the first trivial edge case; If the first element in the list satisfies the
condition, the algorithm returns the whole element (node), and user can further
handle it as he like (either extract the satellite data or do whatever); otherwise,
the algorithm recursively perform finding on the rest of the sub-list. Below is
the corresponding Haskell example program.
find _ [] = Nothing
find p (x:xs) = if p x then Just x else find p xs

Translating this to imperative algorithm is straightforward. Here we use

NIL to represent the fail case.
function Find(p, L)
while L 6= do

596

APPENDIX A. LISTS
if p(First(L)) then
return First(L)
L Rest(L)
return NIL
And here is the Python example of finding.

def find(p, xs):

while xs is not None:
if p(xs.key):
return xs
xs = xs.next
return None

It is quite possible that there are multiple elements in the list which satisfy
the precondition. The finding algorithm designed so far just picks the first one
it meets, and stops immediately. It can be considered as a special case of finding
all elements under a certain condition.
Another viewpoint of finding all elements with a given predicate is to treat
the finding algorithm as a black box, the input to this box is a list, while the
output is another list contains all elements satisfying the predicate. This can
be called as filtering as shown in the below figure.

input

filter p

output

Figure A.4: The input is the original list {x1 , x2 , ..., xN }, the output is a list
{x01 , x02 , ..., x0M }, that for x0i , predicate p(x0i ) is satisfied.
This figure can be formalized in another form in taste of set enumeration.
However, we actually enumerate among list instead of a set.
f ilter(p, L) = {x|x L p(x)}

(A.75)

Some environment such as Haskell (and Python for any iterable), supports
this form as list comprehension.
filter p xs = [ x | x xs, p x]

And in Python for built-in list as

def filter(p, xs):
return [x for x in xs if p(x)]

Note that the Python built-in list isnt singly-linked list as we mentioned in
this appendix.
In order to modify the finding algorithm to realize filtering, the found elements are appended to a result list. And instead of stopping the traverse, all
the rest of elements should be examined with the predicate.

: L=

cons(l1 , f ilter(p, L0 )) : p(l1 )

f ilter(p, L) =
(A.76)

f ilter(p, L0 ) : otherwise

A.7. SEARCHING AND MATCHING

597

This algorithm returns empty result if the list is empty for trivial edge case;
For non-empty list, suppose the recursive result of filtering the rest of the sublist is A, the algorithm examine if the first element satisfies the predicate, it is
put in front of A by a cons operation (O(1) time).
The corresponding Haskell program is given as below.
filter _ [] = []
filter p (x:xs) = if p x then x : filter p xs else filter p xs

Although we mentioned that the next found element is appended to the

result list, this algorithm actually constructs the result list from the right most
to the left, so that appending is avoided, which ensure the linear O(N ) performance. Compare this algorithm with the following imperative quadratic realization reveals the difference.
function Filter(p, L)
L0
while L 6= do
if p(First(L)) then
L0 Append(L0 , First(L))
. Linear operation
L Rest(L)
As the comment of appending statement, its typically proportion to the
length of the result list if the tail position isnt memorized. This fact indicates
that directly transforming the recursive filter algorithm into tail-call form will
downgrade the performance from O(N ) to O(N 2 ). As shown in the below equation, that f ilter(p, L) = f ilter0 (p, L, ) performs as poorly as the imperative
one.

f ilter0 (p, L, A) =

A : L=
f ilter0 (p, L0 , A {l1 }) : p(l1 )

f ilter0 (p, L0 , A) : otherwise

(A.77)

One solution to achieve linear time performance imperatively is to construct

the result list in reverse order, and perform the O(N ) reversion again (refer to
the above section) to get the final result. This is left as exercise to the reader.
The fact of construction the result list from right to left indicates the possibility of realizing filtering with folding-right concept. We need design some
combinator function f , so that f ilter(p, L) = f oldr(f, , L). It requires that
function f takes two arguments, one is the element iterated among the list; the
other is the intermediate result constructed from right. f (x, A) can be defined
as that it tests the predicate against x, if succeed, the result is updated to
cons(x, A), otherwise, A is kept same.

cons(x, A) : p(x)
f (x, A) =
(A.78)
A : otherwise
However, the predicate must be passed to function f as well. This can
be achieved by using currying, so f actually has the prototype f (p, x, A), and
filtering is defined as following.
f ilter(p, L) = f oldr(x,A f (p, x, A), , L)

(A.79)

598

APPENDIX A. LISTS

Which can be simplified by -conversion. For detailed definition of -conversion,

readers can refer to [2].
f ilter(p, L) = f oldr(f (p), , L)

(A.80)

The following Haskell example program implements this equation.

filter p = foldr f [] where
f x xs = if p x then x : xs else xs

Similar to mapping and folding, filtering is actually a generic concept, that

we can apply a predicate on any traversable data structures to get what we
are interesting. readers can refer to the topic about monoid in [8] for further
reading.

A.7.4

Matching

Matching generally means to find a given pattern among some data structures.
In this section, we limit the topic within list. Even this limitation will leads
to a very wide and deep topic, that there are dedicated chapters in this book
introduce matching algorithms. So we only select the algorithm to test if a given
list exists in another (typically longer) list.
Before dive into the algorithm of finding the sub-list at any position, two
special edge cases are used for warm up. They are algorithms to test if a given
list is either prefix or suffix of another.
In the section about span, we have seen how to find a prefix under a certain
condition. prefix matching can be considered as a special case in some sense.
That it compares each of the elements between the two lists from the beginning
until meets any different elements or pass the end of one list. Define P L if
P is prefix of L.

T rue : P =

F alse : p1 6= l1
(A.81)
P L=
0
P L0 : otherwise
This is obviously a linear algorithm. However, We cant use the very same
approach to test if a list is suffix of another because it isnt cheap to start from
the end of the list and keep iterating backwards. Arrays, on the other hand
which support random access can be easily traversed backwards.
As we only need the yes-no result, one solution to realize a linear suffix
testing algorithm is to reverse both lists, (which is linear time), and use prefix
testing instead. Define L P if P is suffix of L.
L P = reverse(P ) reverse(L)

(A.82)

With defined, it enables to test if a list is infix of another. The idea is

to traverse the target list, and repeatedly applying the prefix testing till any
success or arrives at the end.
function Is-Infix(P, L)
while L 6= do
if P L then
return TRUE
L Rest(L)

A.7. SEARCHING AND MATCHING

599

return FALSE
Formalize this algorithm to recursive equation leads to the below definition.

T rue : P L

F alse : L =
inf ix?(P, L) =
(A.83)

inf ix?(P, L0 ) : otherwise

Note that there is a tricky implicit constraint in this equation. If the pattern
P is empty, it is definitely the infix of any target list. This case is actually
covered by the first condition in the above equation because empty list is also
the prefix of any list. In most programming languages support pattern matching,
we cant arrange the second clause as the first edge case, or it will return false for
inf ix?(, ). (One exception is Prolog, but this is a language specific feature,
which we wont covered in this book.)
Since prefix testing is linear, and it is called while traversing the list, this
algorithm is quadratic O(N M ). where N and M are the length of the pattern
and target lists respectively. There is no trivial way to improve this position by
position scanning algorithm to linear even if the data structure changes from
linked-list to randomly accessible array.
There are chapters in this book introduce several approaches for fast matching, including suffix tree with Ukkonen algorithm, Knuth-Morris-Pratt algorithm and Boyer-Moore algorithm.
Alternatively, we can enumerate all suffixes of the target list, and check if
the pattern is prefix of any these suffixes. Which can be represented as the
following.
inf ix?(P, L) = S suf f ixes(L) P S

(A.84)

This can be represented as list comprehension, for example the below Haskell
program.
isInfixOf x y = (not null) [ s | s tails(y), x isPrefixOfs]

Where function isPrefixOf is the prefixing testing function defined according to our previous design. function tails generate all suffixes of a list. The
implementation of tails is left as an exercise to the reader.

Exercise A.5

Implement the linear existence testing in both functional and imperative

approaches in your favorite programming languages.
Implement the looking up algorithm in your favorite imperative programming language.
Realize the linear time filtering algorithm by firstly building the result
list in reverse order, and finally reverse it to resume the normal result.
Implement this algorithm in both imperative looping and functional tailrecursion call.
Implement the imperative algorithm of prefix testing in your favorite programming language.
Implement the algorithm to enumerate all suffixes of a list.

600

APPENDIX A. LISTS

A.8

zipping and unzipping

It is quite common to construct a list of paired elements. For example, in the

naive brute-force solution for Drunk jailer puzzle which is shown in section
of mapping, we need to represent the state of all lights. It is initialized as
{(1, 0), (2, 0), ..., (N, 0)}. Another example is to build a key-value list, such as
{(1, a), (2, an), (3, another), ...}.
In Drunk jailer example, the list of pairs is built like the following.
map(i (i, 0), {1, 2, ..., N })
The more general case is that, There have been already two lists prepared,
what we need is a handy zipper method.

zip(A, B) =

: A=B =
cons((a1 , b1 ), zip(A0 , B 0 )) : otherwise

(A.85)

Note that this algorithm is capable to handle the case that the two lists
being zipped have different lengths. The result list of pairs aligns with the
shorter one. And its even possible to zip an infinite list with another one with
limited length in environment support lazy evaluation. For example with this
auxiliary function defined, we can initialize the lights state as
zip({0, 0, ...}, {1, 2, ..., N }
In some languages support list enumeration, such as Haskell (Python provides similar range function, but it manipulates built-in list, which isnt linkedlist actually), this can be expressed as zip (repeat 0) [1..n]. Given a list
of words, we can also index them with consecutive numbers as
zip({1, 2, ...}, {a, an, another, ...})
Note that the zipping algorithm is linear, as it uses constant time cons operation in each recursive call. However, directly translating zip into imperative
manner would down-grade the performance to quadratic unless the linked-list
is optimized with tail position cache or we in-place modify one of the passed-in
list.
function Zip(A, B)
C
while A 6= B 6= do
C Append(C, (First(A), First(B)))
A Rest(A)
B Rest(B)
return C
Note that, the appending operation is proportion to the length of the result
list C, so it will get more and more slowly along with traversing. There are three
solutions to improve this algorithm to linear time. The first method is to use a
similar approach as we did in infix-testing, that we construct the result list of
pairs in reverse order by always insert the paired elements on head; then perform
a linear reverse operation before return the final result; The second method is

A.8. ZIPPING AND UNZIPPING

601

to modify one passed-in list, for example A, in-place while traversing. Translate
it from list of elements to list of pairs; The third method is to remember the
last appending position. Please try these solutions as exercise.
The key point of linear time zipping is that the result list is actually built
from right to left, which is similar to the infix-testing algorithm. So its quite
possible to provide a folding-right realization. This is left as exercise to the
reader.
It is natural to extend the zipper algorithm so that multiple lists can be
zipped to one list of multiple-elements. For example, Haskell standard library
provides, zip, zip3, zip4, ..., till zip7. Another typical extension to zipper
is that, sometimes, we dont want to list of pairs (or tuples more generally),
instead, we want to apply some combinator function to each pair of elements.
For example, consider the case that we have a list of unit prices for every
fruit: apple, orange, banana, ..., as {1.00, 0.80, 10.05, ...}, with same unit of
Dollar; And the cart of customer holds a list of purchased quantity, for instance
{3, 1, 0, ...}, means this customer, put 3 apples, an orange in the cart. He doesnt
take any banana, so the quantity of banana is zero. We want to generate a
list of cost for the customer, contains how much should pay for apple, orange,
banana,... respectively.
The program can be written from scratch as below.

paylist(U, Q) =

:
cons(u1 q1 , paylist(U 0 , Q0 )) :

U =Q=
otherwise

Compare this equation with the zipper algorithm. It is easy to find the
common structure of the two, and we can parameterize the combinator function
as f , so that the generic zipper algorithm can be defined as the following.

A=B =
otherwise
(A.86)
Here is an example that defines the inner-product (or dot-product)[14] by
using zipW ith.

zipW ith(f, A, B) =

:
cons(f (a1 , b1 ), zipW ith(f, A0 , B 0 )) :

A B = sum(zipW ith(, A, B))

(A.87)

It is necessary to realize the inverse operation of zipping, that converts a list

of pairs, to different lists of elements. Back to the purchasing example, It is
quite possible that the unit price information is stored in a association list like
U = {(apple, 1.00), (orange, 0.80), (banana, 10.05), ...}, so that its convenient
to look up the price with a given product name, for instance, lookup(melon, U ).
Similarly, the cart can also be represented clearly in such manner, for example,
Q = {(apple, 3), (orange, 1), (banana, 0), ...}.
Given such a product - unit price list and a product - quantity list, how
to calculate the total payment?
One straight forward idea derived from the previous solution is to extract the
unit price list and the purchased quantity list, then calculate the inner-product
of them.
pay = sum(zipW ith(, snd(unzip(P )), snd(unzip(Q))))

(A.88)

602

APPENDIX A. LISTS

Although the definition of unzip can be directly written as the inverse of

zip, here we give a realization based on folding-right.
unzip(L) = f oldr((a,b),(A,B) (cons(a, A), cons(b, B)), (, ), L)

(A.89)

The initial result is a pair of empty list. During the folding process, the
head of the list, which is a pair of elements, as well as the intermediate result
are passed to the combinator function. This combinator function is given as a
lambda expression, that it extracts the paired elements, and put them in front
of the two intermediate lists respectively. Note that we use implicit pattern
matching to extract the elements from pairs. Alternatively this can be done by
using f st, and snd functions explicitly as
p,P (cons(f st(p), f st(P )), cons(snd(p), snd(P )))
The following Haskell example code implements unzip algorithm.
unzip = foldr (a, b) (as, bs) (a:as, b:bs) ([], [])

Zip and unzip concepts can be extended more generally rather than only
limiting within linked-list. It is quite useful to zip two lists to a tree, where the
data stored in the tree are paired elements from both lists. General zip and
unzip can also be used to track the traverse path of a collection to mimic the
parent pointer in imperative implementations. Please refer to the last chapter
of [8] for a good treatment.

Exercise A.6
Design and implement iota (I) algorithm, which can enumerate a list with
some given parameters. For example:
iota(..., N ) = {1, 2, 3, ..., N };
iota(M, N ) = {M, M + 1, M + 2, ..., N }, Where M N ;
iota(M, M + a, ..., N ) = {M, M + a, M + 2a, ..., N };
iota(M, M, ...) = repeat(M ) = {M, M, M, ...};
iota(M, ...) = {M, M + 1, M + 2, ...}.
Note that the last two cases demand generate infinite list essentially. Consider how to represents infinite list? You may refer to the streaming and
lazy evaluation materials such as [5] and [8].
Design and implement a linear time imperative zipper algorithm.
Realize the zipper algorithm with folding-right approach.
For the purchase payment example, suppose the quantity association list
only contains those items with the quantity isnt zero, that instead of a
list of Q = {(apple, 3), (banana, 0), (orange, 1), ...}, it hold a list like Q =
{(apple, 3), (orange, 1), ...}. The banana information is filtered because
the customer doesnt pick any bananas. Write a program, taking the unitprice association list, and this kind of quantity list, to calculate the total
payment.

A.9. NOTES AND SHORT SUMMARY

A.9

603

Notes and short summary

In this appendix, a quick introduction about how to build, manipulate, transfer,

and searching singly linked list is briefed in both purely functional and imperative approaches. Most of the modern programming environments have been
equipped with tools to handle such elementary data structures. However, such
tools are designed for general purpose cases, Serious programming shouldnt
take them as black-boxes.
Since linked-list is so critical that it builds the corner stones for almost
all functional programming environments, just like the importance of array to
imperative settings. We take this topic as an appendix to the book. It is quite
OK that the reader starts with the first chapter about binary search tree, which
is a kind of hello world topic, and refers to this appendix when meets any
unfamiliar list operations.

604

APPENDIX A. LISTS

Bibliography
[1] Richard Bird. Pearls of Functional Algorithm Design. Cambridge University Press; 1 edition (November 1, 2010). ISBN: 978-0521513388
[2] Simon L. Peyton Jones. The Implementation of Functional Programming
Languages. Prentice-Hall International Series in Computer Since. Prentice
Hall (May 1987). ISBN: 978-0134533339
[3] Andrei Alexandrescu. Modern C++ design: Generic Programming and
Design Patterns Applied. Addison Wesley February 01, 2001, ISBN 0201-70431-5
[4] Benjamin C. Pierce. Types and Programming Languages. The MIT
Press, 2002. ISBN:0262162091
[5] Harold Abelson, Gerald Jay Sussman, Julie Sussman. Structure and Interpretation of Computer Programs, 2nd Edition. MIT Press, 1996, ISBN
0-262-51087-1
[6] Chris Okasaki. Purely Functional Data Structures. Cambridge university
press, (July 1, 1999), ISBN-13: 978-0521663502
[7] Fethi Rabhi, Guy Lapalme. Algorithms: a functional programming approach. Second edition. Addison-Wesley, 1999. ISBN: 0201-59604-0
[8] Miran Lipovaca. Learn You a Haskell for Great Good! A Beginners
Guide. No Starch Press; 1 edition April 2011, 400 pp. ISBN: 978-1-59327283-8
[9] Joe Armstrong. Programming Erlang: Software for a Concurrent World.
Pragmatic Bookshelf; 1 edition (July 18, 2007). ISBN-13: 978-1934356005
[10] Wikipedia. Tail call. https://en.wikipedia.org/wiki/Tail call
[11] SGI. transform. http://www.sgi.com/tech/stl/transform.html
[12] ACM/ICPC. The drunk jailer. Peking University judge online for
ACM/ICPC. http://poj.org/problem?id=1218.
[13] Haskell wiki. Haskell programming tips. 4.4 Choose the appropriate fold.
http://www.haskell.org/haskellwiki/Haskell programming tips
[14] Wikipedia. Dot product. http://en.wikipedia.org/wiki/Dot product

605

606

BIBLIOGRAPHY

GNU Free Documentation License

Version 1.3, 3 November 2008
c 2000, 2001, 2002, 2007, 2008 Free Software Foundation, Inc.
Copyright
http://fsf.org/
Everyone is permitted to copy and distribute verbatim copies of this license
document, but changing it is not allowed.

Preamble
The purpose of this License is to make a manual, textbook, or other functional and useful document free in the sense of freedom: to assure everyone
the effective freedom to copy and redistribute it, with or without modifying it,
either commercially or noncommercially. Secondarily, this License preserves for
the author and publisher a way to get credit for their work, while not being
considered responsible for modifications made by others.
This License is a kind of copyleft, which means that derivative works of the
document must themselves be free in the same sense. It complements the GNU
General Public License, which is a copyleft license designed for free software.
We have designed this License in order to use it for manuals for free software,
because free software needs free documentation: a free program should come
with manuals providing the same freedoms that the software does. But this
License is not limited to software manuals; it can be used for any textual work,
regardless of subject matter or whether it is published as a printed book. We
recommend this License principally for works whose purpose is instruction or
reference.

1. APPLICABILITY AND DEFINITIONS

This License applies to any manual or other work, in any medium, that
contains a notice placed by the copyright holder saying it can be distributed
under the terms of this License. Such a notice grants a world-wide, royalty-free
license, unlimited in duration, to use that work under the conditions stated
herein. The Document, below, refers to any such manual or work. Any
member of the public is a licensee, and is addressed as you. You accept the
license if you copy, modify or distribute the work in a way requiring permission
under copyright law.
A Modified Version of the Document means any work containing the
Document or a portion of it, either copied verbatim, or with modifications
and/or translated into another language.
607

608

BIBLIOGRAPHY

A Secondary Section is a named appendix or a front-matter section

of the Document that deals exclusively with the relationship of the publishers
or authors of the Document to the Documents overall subject (or to related
matters) and contains nothing that could fall directly within that overall subject.
(Thus, if the Document is in part a textbook of mathematics, a Secondary
Section may not explain any mathematics.) The relationship could be a matter
of historical connection with the subject or with related matters, or of legal,
commercial, philosophical, ethical or political position regarding them.
The Invariant Sections are certain Secondary Sections whose titles are
designated, as being those of Invariant Sections, in the notice that says that
the Document is released under this License. If a section does not fit the above
definition of Secondary then it is not allowed to be designated as Invariant.
The Document may contain zero Invariant Sections. If the Document does not
identify any Invariant Sections then there are none.
The Cover Texts are certain short passages of text that are listed, as
Front-Cover Texts or Back-Cover Texts, in the notice that says that the Document is released under this License. A Front-Cover Text may be at most 5
words, and a Back-Cover Text may be at most 25 words.
A Transparent copy of the Document means a machine-readable copy,
represented in a format whose specification is available to the general public,
that is suitable for revising the document straightforwardly with generic text
editors or (for images composed of pixels) generic paint programs or (for drawings) some widely available drawing editor, and that is suitable for input to
text formatters or for automatic translation to a variety of formats suitable for
input to text formatters. A copy made in an otherwise Transparent file format
whose markup, or absence of markup, has been arranged to thwart or discourage subsequent modification by readers is not Transparent. An image format is
not Transparent if used for any substantial amount of text. A copy that is not
Transparent is called Opaque.
Examples of suitable formats for Transparent copies include plain ASCII
without markup, Texinfo input format, LaTeX input format, SGML or XML using a publicly available DTD, and standard-conforming simple HTML, PostScript
or PDF designed for human modification. Examples of transparent image formats include PNG, XCF and JPG. Opaque formats include proprietary formats
that can be read and edited only by proprietary word processors, SGML or
XML for which the DTD and/or processing tools are not generally available,
and the machine-generated HTML, PostScript or PDF produced by some word
processors for output purposes only.
The Title Page means, for a printed book, the title page itself, plus such
following pages as are needed to hold, legibly, the material this License requires
to appear in the title page. For works in formats which do not have any title
page as such, Title Page means the text near the most prominent appearance
of the works title, preceding the beginning of the body of the text.
The publisher means any person or entity that distributes copies of the
Document to the public.
A section Entitled XYZ means a named subunit of the Document whose
title either is precisely XYZ or contains XYZ in parentheses following text
that translates XYZ in another language. (Here XYZ stands for a specific section name mentioned below, such as Acknowledgements, Dedications,
Endorsements, or History.) To Preserve the Title of such a sec-

BIBLIOGRAPHY

609

tion when you modify the Document means that it remains a section Entitled
XYZ according to this definition.
The Document may include Warranty Disclaimers next to the notice which
states that this License applies to the Document. These Warranty Disclaimers
are considered to be included by reference in this License, but only as regards
disclaiming warranties: any other implication that these Warranty Disclaimers
may have is void and has no effect on the meaning of this License.

2. VERBATIM COPYING
You may copy and distribute the Document in any medium, either commercially or noncommercially, provided that this License, the copyright notices, and
the license notice saying this License applies to the Document are reproduced
in all copies, and that you add no other conditions whatsoever to those of this
License. You may not use technical measures to obstruct or control the reading
or further copying of the copies you make or distribute. However, you may
accept compensation in exchange for copies. If you distribute a large enough
number of copies you must also follow the conditions in section 3.
You may also lend copies, under the same conditions stated above, and you
may publicly display copies.

3. COPYING IN QUANTITY
If you publish printed copies (or copies in media that commonly have printed
covers) of the Document, numbering more than 100, and the Documents license
notice requires Cover Texts, you must enclose the copies in covers that carry,
clearly and legibly, all these Cover Texts: Front-Cover Texts on the front cover,
and Back-Cover Texts on the back cover. Both covers must also clearly and
legibly identify you as the publisher of these copies. The front cover must
present the full title with all words of the title equally prominent and visible.
You may add other material on the covers in addition. Copying with changes
limited to the covers, as long as they preserve the title of the Document and
satisfy these conditions, can be treated as verbatim copying in other respects.
If the required texts for either cover are too voluminous to fit legibly, you
should put the first ones listed (as many as fit reasonably) on the actual cover,
and continue the rest onto adjacent pages.
If you publish or distribute Opaque copies of the Document numbering more
than 100, you must either include a machine-readable Transparent copy along
with each Opaque copy, or state in or with each Opaque copy a computernetwork location from which the general network-using public has access to
download using public-standard network protocols a complete Transparent copy
of the Document, free of added material. If you use the latter option, you must
take reasonably prudent steps, when you begin distribution of Opaque copies
in quantity, to ensure that this Transparent copy will remain thus accessible at
the stated location until at least one year after the last time you distribute an
Opaque copy (directly or through your agents or retailers) of that edition to the
public.
It is requested, but not required, that you contact the authors of the Document well before redistributing any large number of copies, to give them a
chance to provide you with an updated version of the Document.

610

BIBLIOGRAPHY

4. MODIFICATIONS
You may copy and distribute a Modified Version of the Document under the
conditions of sections 2 and 3 above, provided that you release the Modified
Version under precisely this License, with the Modified Version filling the role
of the Document, thus licensing distribution and modification of the Modified
Version to whoever possesses a copy of it. In addition, you must do these things
in the Modified Version:
A. Use in the Title Page (and on the covers, if any) a title distinct from that
of the Document, and from those of previous versions (which should, if
there were any, be listed in the History section of the Document). You
may use the same title as a previous version if the original publisher of
that version gives permission.
B. List on the Title Page, as authors, one or more persons or entities responsible for authorship of the modifications in the Modified Version, together
with at least five of the principal authors of the Document (all of its principal authors, if it has fewer than five), unless they release you from this
requirement.
C. State on the Title page the name of the publisher of the Modified Version,
as the publisher.
D. Preserve all the copyright notices of the Document.
E. Add an appropriate copyright notice for your modifications adjacent to
the other copyright notices.
F. Include, immediately after the copyright notices, a license notice giving
the public permission to use the Modified Version under the terms of this
License, in the form shown in the Addendum below.
G. Preserve in that license notice the full lists of Invariant Sections and required Cover Texts given in the Documents license notice.
H. Include an unaltered copy of this License.
I. Preserve the section Entitled History, Preserve its Title, and add to it
an item stating at least the title, year, new authors, and publisher of the
Modified Version as given on the Title Page. If there is no section Entitled
History in the Document, create one stating the title, year, authors, and
publisher of the Document as given on its Title Page, then add an item
describing the Modified Version as stated in the previous sentence.
J. Preserve the network location, if any, given in the Document for public
access to a Transparent copy of the Document, and likewise the network
locations given in the Document for previous versions it was based on.
These may be placed in the History section. You may omit a network
location for a work that was published at least four years before the Document itself, or if the original publisher of the version it refers to gives
permission.

BIBLIOGRAPHY

611

K. For any section Entitled Acknowledgements or Dedications, Preserve

the Title of the section, and preserve in the section all the substance and
tone of each of the contributor acknowledgements and/or dedications given
therein.
L. Preserve all the Invariant Sections of the Document, unaltered in their text
and in their titles. Section numbers or the equivalent are not considered
part of the section titles.
M. Delete any section Entitled Endorsements. Such a section may not be
included in the Modified Version.
N. Do not retitle any existing section to be Entitled Endorsements or to
conflict in title with any Invariant Section.
O. Preserve any Warranty Disclaimers.
If the Modified Version includes new front-matter sections or appendices
that qualify as Secondary Sections and contain no material copied from the
Document, you may at your option designate some or all of these sections as
invariant. To do this, add their titles to the list of Invariant Sections in the
Modified Versions license notice. These titles must be distinct from any other
section titles.
You may add a section Entitled Endorsements, provided it contains nothing but endorsements of your Modified Version by various partiesfor example,
statements of peer review or that the text has been approved by an organization
as the authoritative definition of a standard.
You may add a passage of up to five words as a Front-Cover Text, and a
passage of up to 25 words as a Back-Cover Text, to the end of the list of Cover
Texts in the Modified Version. Only one passage of Front-Cover Text and one
of Back-Cover Text may be added by (or through arrangements made by) any
one entity. If the Document already includes a cover text for the same cover,
previously added by you or by arrangement made by the same entity you are
acting on behalf of, you may not add another; but you may replace the old one,
on explicit permission from the previous publisher that added the old one.
The author(s) and publisher(s) of the Document do not by this License give
permission to use their names for publicity for or to assert or imply endorsement
of any Modified Version.

5. COMBINING DOCUMENTS
You may combine the Document with other documents released under this
License, under the terms defined in section 4 above for modified versions, provided that you include in the combination all of the Invariant Sections of all
of the original documents, unmodified, and list them all as Invariant Sections
of your combined work in its license notice, and that you preserve all their
Warranty Disclaimers.
The combined work need only contain one copy of this License, and multiple
identical Invariant Sections may be replaced with a single copy. If there are
multiple Invariant Sections with the same name but different contents, make
the title of each such section unique by adding at the end of it, in parentheses,

612

BIBLIOGRAPHY

the name of the original author or publisher of that section if known, or else a
unique number. Make the same adjustment to the section titles in the list of
Invariant Sections in the license notice of the combined work.
In the combination, you must combine any sections Entitled History in
the various original documents, forming one section Entitled History; likewise
combine any sections Entitled Acknowledgements, and any sections Entitled
Dedications. You must delete all sections Entitled Endorsements.

6. COLLECTIONS OF DOCUMENTS
You may make a collection consisting of the Document and other documents
released under this License, and replace the individual copies of this License in
the various documents with a single copy that is included in the collection,
provided that you follow the rules of this License for verbatim copying of each
of the documents in all other respects.
You may extract a single document from such a collection, and distribute it
individually under this License, provided you insert a copy of this License into
the extracted document, and follow this License in all other respects regarding
verbatim copying of that document.

7. AGGREGATION WITH INDEPENDENT

WORKS
A compilation of the Document or its derivatives with other separate and
independent documents or works, in or on a volume of a storage or distribution
medium, is called an aggregate if the copyright resulting from the compilation
is not used to limit the legal rights of the compilations users beyond what the
individual works permit. When the Document is included in an aggregate,
this License does not apply to the other works in the aggregate which are not
themselves derivative works of the Document.
If the Cover Text requirement of section 3 is applicable to these copies of the
Document, then if the Document is less than one half of the entire aggregate, the
Documents Cover Texts may be placed on covers that bracket the Document
within the aggregate, or the electronic equivalent of covers if the Document is
in electronic form. Otherwise they must appear on printed covers that bracket
the whole aggregate.

8. TRANSLATION
Translation is considered a kind of modification, so you may distribute translations of the Document under the terms of section 4. Replacing Invariant Sections with translations requires special permission from their copyright holders,
but you may include translations of some or all Invariant Sections in addition to
the original versions of these Invariant Sections. You may include a translation
of this License, and all the license notices in the Document, and any Warranty
Disclaimers, provided that you also include the original English version of this
License and the original versions of those notices and disclaimers. In case of a
disagreement between the translation and the original version of this License or
a notice or disclaimer, the original version will prevail.

BIBLIOGRAPHY

613

If a section in the Document is Entitled Acknowledgements, Dedications, or History, the requirement (section 4) to Preserve its Title (section 1)
will typically require changing the actual title.

9. TERMINATION
You may not copy, modify, sublicense, or distribute the Document except as
expressly provided under this License. Any attempt otherwise to copy, modify,
sublicense, or distribute it is void, and will automatically terminate your rights
under this License.
However, if you cease all violation of this License, then your license from
a particular copyright holder is reinstated (a) provisionally, unless and until
the copyright holder explicitly and finally terminates your license, and (b) permanently, if the copyright holder fails to notify you of the violation by some
reasonable means prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies you of the violation by some reasonable
means, this is the first time you have received notice of violation of this License
(for any work) from that copyright holder, and you cure the violation prior to
30 days after your receipt of the notice.
Termination of your rights under this section does not terminate the licenses
of parties who have received copies or rights from you under this License. If
your rights have been terminated and not permanently reinstated, receipt of a
copy of some or all of the same material does not give you any rights to use it.

10. FUTURE REVISIONS OF THIS LICENSE

The Free Software Foundation may publish new, revised versions of the
GNU Free Documentation License from time to time. Such new versions will be
similar in spirit to the present version, but may differ in detail to address new
problems or concerns. See http://www.gnu.org/copyleft/.
Each version of the License is given a distinguishing version number. If
the Document specifies that a particular numbered version of this License or
any later version applies to it, you have the option of following the terms and
conditions either of that specified version or of any later version that has been
published (not as a draft) by the Free Software Foundation. If the Document
does not specify a version number of this License, you may choose any version
ever published (not as a draft) by the Free Software Foundation. If the Document specifies that a proxy can decide which future versions of this License can
be used, that proxys public statement of acceptance of a version permanently
authorizes you to choose that version for the Document.

11. RELICENSING
Massive Multiauthor Collaboration Site (or MMC Site) means any World
Wide Web server that publishes copyrightable works and also provides prominent facilities for anybody to edit those works. A public wiki that anybody can
edit is an example of such a server. A Massive Multiauthor Collaboration
(or MMC) contained in the site means any set of copyrightable works thus
published on the MMC site.

614

BIBLIOGRAPHY

CC-BY-SA means the Creative Commons Attribution-Share Alike 3.0 license published by Creative Commons Corporation, a not-for-profit corporation
with a principal place of business in San Francisco, California, as well as future
copyleft versions of that license published by that same organization.
Incorporate means to publish or republish a Document, in whole or in
part, as part of another Document.
An MMC is eligible for relicensing if it is licensed under this License, and
if all works that were first published under this License somewhere other than
this MMC, and subsequently incorporated in whole or in part into the MMC,
(1) had no cover texts or invariant sections, and (2) were thus incorporated prior
to November 1, 2008.
The operator of an MMC Site may republish an MMC contained in the site
under CC-BY-SA on the same site at any time before August 1, 2009, provided
the MMC is eligible for relicensing.

ADDENDUM: How to use this License for your

documents
To use this License in a document you have written, include a copy of the
License in the document and put the following copyright and license notices just
after the title page:
c YEAR YOUR NAME. Permission is granted to copy,
Copyright
distribute and/or modify this document under the terms of the GNU
Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections,
no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled GNU Free Documentation
License.

If you have Invariant Sections, Front-Cover Texts and Back-Cover Texts,

replace the with . . . Texts. line with this:

with the Invariant Sections being LIST THEIR TITLES, with the
Front-Cover Texts being LIST, and with the Back-Cover Texts being
LIST.

If you have Invariant Sections without Cover Texts, or some other combination of the three, merge those two alternatives to suit the situation.
If your document contains nontrivial examples of program code, we recommend releasing these examples in parallel under your choice of free software
license, such as the GNU General Public License, to permit their use in free
software.

Index
8 queens puzzle, 477
Auto completion, 121
AVL tree, 77
balancing, 82
definition, 77
deletion, 87
imperative insertion, 88
insertion, 80
verification, 87
B-tree, 159
delete, 168
insert, 161
search, 181
split, 161
BFS, 506
Binary heap, 189
build heap, 192
decrease key, 197
heap push, 199
Heapify, 191
insertion, 199
merge, 204
pop, 196
top, 196
top-k, 197
Binary Random Access List
Definition, 314
Insertion, 316
Random access, 319
Remove from head, 318
Binary search, 428
binary search tree, 25
data layout, 27
delete, 36
insertion, 29
looking up, 33
min/max, 34
randomly build, 40
search, 33

succ/pred, 34
traverse, 30
binary tree, 26
Binomial Heap
Linking, 248
Binomial heap, 243
definition, 246
insertion, 250
pop, 255
Binomial tree, 243
merge, 252
Boyer-Moor majority number, 444
Boyer-Moore algorithm, 463
Breadth-first search, 506
Change-making problem, 518
Cock-tail sort, 227
Deep-first search, 471
DFS, 471
Dynamic programming, 520
Fibonacci Heap, 258
decrease key, 271
delete min, 262
insert, 260
merge, 261
pop, 262
Finger Tree
Imperative splitting, 362
Finger tree
Append to tail, 346
Concatenate, 349
Definition, 335
Ill-formed tree, 341
Imperative random access, 360
Insert to head, 337
Random access, 354, 360
Remove from head, 340
Remove from tail, 347
Size augmentation, 354
615

616
splitting, 358
folding, 582
Grady algorithm, 507
Heap sort, 199
Huffman coding, 507
Implicit binary heap, 189
in-order traverse, 31
Insertion sort
binary search, 46
binary search tree, 49
linked-list setting, 47
insertion sort, 43
insertion, 44
Integer Patricia, 100
insert, 102
look up, 108
Integer prefix tree, 100
Integer trie, 96
insert, 97
look up, 99
Kloski puzzle, 499
KMP, 449
Knuth-Morris-Pratt algorithm, 449
LCS, 525
left child, right sibling, 247
Leftist heap, 202
heap sort, 206
insertion, 205
merge, 203
pop, 204
rank, 202
S-value, 202
top, 204
List
append, 549
break, 577
concat, 558
concats, 587
cons, 543
Construction, 543
definition, 541
delete, 555
delete at, 555
drop, 575
drop while, 576
elem, 588

INDEX
empty, 542
empty testing, 544
existence testing, 588
Extract sub-list, 575
filter, 589
find, 589
fold from left, 584
fold from right, 582
foldl, 584
foldr, 582
for each, 570
get at, 545
group, 578
head, 542
index, 545
infix, 592
init, 546
insert, 552
insert at, 552
last, 546
length, 544
lookup, 588
map, 567, 568
matching, 592
maximum, 563
minimum, 563
mutate, 549
prefix, 592
product, 559
reverse, 573
Reverse index, 547
rindex, 547
set at, 550
span, 577
split at, 575, 577
suffix, 592
sum, 559
tail, 542
take, 575
take while, 576
Transformation, 567
unzip, 594
zip, 594
Longest common sub-string, 153
Longest common subsequence problem,
525
Longest palindrome, 155
Longest repeated sub-string, 152
Maximum sum problem, 448

INDEX
Maze problem, 471
Merge Sort, 395
Basic version, 396
Bottom-up merge sort, 416
In-place merge sort, 403
In-place working area, 404
Linked-list merge sort, 409
Merge, 396
Naive in-place merge, 403
Nature merge sort, 411
Performance analysis, 399
Work area allocation, 400
minimum free number, 7
MTF, 366
Paired-array list
Definition, 327
Insertion and appending, 328
Random access, 328
Removing and balancing, 329
Pairing heap, 275
definition, 275
delete min, 278
find min, 276
insert, 276
pop, 278
top, 276
pairing heap
decrease key, 278
delete, 282
Parallel merge sort, 419
Parallel quick sort, 419
Patricia, 113
insert, 114
look up, 119
Peg puzzle, 480
post-order traverse, 31
pre-order traverse, 31
Prefix tree, 113
Queue
Balance Queue, 298
Circular buffer, 291
Incremental concatenate, 302
Incremental reverse, 300
Lazy real-time queue, 307
Paired-array queue, 296
Paired-list queue, 294
Real-time Queue, 300
Singly linked-list, 288

617
Quick Sort
2-way partition, 385
3-way partition, 387
Accmulated partition, 377
Accumulated quick sort, 378
Average case analysis, 380
Basic version, 372
Engineering improvement, 383
Handle duplicated elements, 383
Insertion sort fall-back, 394
One pass functional partition, 377
Performance analysis, 379
Strict weak ordering, 373
Quick sort, 371
partition, 374
Radix tree, 95
range traverse, 36
red-black tree, 53, 58
deletion, 63
imperative insertion, 71
insertion, 59
red-black properties, 58
Saddelback search, 433
Selection algorithm, 424
selection sort, 219
minimum finding, 221
parameterize the comparator, 225
tail-recursive call minimum finding, 223
Sequence
Binary random access list, 314
Concatenate-able list, 331
finger tree, 335
Imperative binary access list, 324
numeric representation for binary
access list, 322
Paired-array list, 327
Skew heap, 206
insertion, 207
merge, 207
pop, 207
top, 207
Splay heap, 208
insertion, 214
merge, 215
pop, 214
splaying, 209
top, 214

618
Subset sum problem, 530
Suffix link, 135
Suffix tree, 133, 141
active point, 141
Canonical reference pair, 142
end point, 141
functional construction, 147
node transfer, 135
on-line construction, 143
reference pair, 141
string searching, 150
sub-string occurrence, 150
Suffix trie, 134
on-line construction, 136
T9, 125
Tail call, 560
Tail recursion, 560
Tail recursive call, 560
Textonym input method, 125
The wolf, goat, and cabbage puzzle,
486
Tounament knock out
explict infinity, 236
tree reconstruction, 32
tree rotation, 56
Trie, 109
insert, 111
look up, 112
Trounament knock out, 231
Ukkonens algorithm, 143
Water jugs puzzle, 490
word counter, 25

INDEX

Intrusion Detection Honeypots
From Everand
Intrusion Detection Honeypots
Chris Sanders
3/5 (2)
Marketplace: System Requirements Specification (SRS)
100% (2)
Marketplace: System Requirements Specification (SRS)
60 pages
Vijayalakshmi P. A Textbook of Data Structures and Algorithms Vol 2. 2022
No ratings yet
Vijayalakshmi P. A Textbook of Data Structures and Algorithms Vol 2. 2022
304 pages
Data Structures and Algorithms: Lecture Notes For
No ratings yet
Data Structures and Algorithms: Lecture Notes For
126 pages
Moshiri Design and Analysis of Data Structures
No ratings yet
Moshiri Design and Analysis of Data Structures
335 pages
STL Free
No ratings yet
STL Free
181 pages
Elementary Algorithms
100% (4)
Elementary Algorithms
630 pages
Elementary Algorithms
No ratings yet
Elementary Algorithms
622 pages
Elementary Algorithms
100% (1)
Elementary Algorithms
622 pages
Elementary Algorithms PDF
No ratings yet
Elementary Algorithms PDF
642 pages
AlgoXY Elementary Algorithms
No ratings yet
AlgoXY Elementary Algorithms
749 pages
Data Sructures and Algorithms
No ratings yet
Data Sructures and Algorithms
112 pages
Data Structures and Algorithms
100% (1)
Data Structures and Algorithms
111 pages
TO PRINT - DSA DOTNET SLACKERS
No ratings yet
TO PRINT - DSA DOTNET SLACKERS
56 pages
Algo Quicksheet
No ratings yet
Algo Quicksheet
70 pages
Dsa Removed
No ratings yet
Dsa Removed
124 pages
Algorithms: Selected Lecture Notes
No ratings yet
Algorithms: Selected Lecture Notes
53 pages
DSA Data Structures and Algorithms
No ratings yet
DSA Data Structures and Algorithms
126 pages
Algorithm Structure C++ - Teo OK PDF
No ratings yet
Algorithm Structure C++ - Teo OK PDF
126 pages
Dsa Book1 PDF
No ratings yet
Dsa Book1 PDF
126 pages
dsaa14
No ratings yet
dsaa14
1 page
Dsa - Barnette and Tonga - 2
No ratings yet
Dsa - Barnette and Tonga - 2
3 pages
Avl-2 0 1
No ratings yet
Avl-2 0 1
432 pages
JNTU BTECH 2-1 Data Structures NOTES
No ratings yet
JNTU BTECH 2-1 Data Structures NOTES
104 pages
Data Structures
No ratings yet
Data Structures
104 pages
Data Structures
No ratings yet
Data Structures
104 pages
Data Structures C1
No ratings yet
Data Structures C1
11 pages
Data Structures
No ratings yet
Data Structures
104 pages
Data Structures Lecture Notes
No ratings yet
Data Structures Lecture Notes
143 pages
Data Structures
No ratings yet
Data Structures
239 pages
FCE Review (Quiz3)
No ratings yet
FCE Review (Quiz3)
14 pages
AlgDs1LectureNotes-2025-02-16
No ratings yet
AlgDs1LectureNotes-2025-02-16
89 pages
Nono
No ratings yet
Nono
215 pages
An Intuitive Introduction To Data Structures Heinold
No ratings yet
An Intuitive Introduction To Data Structures Heinold
167 pages
Alg Ds 1 Lecture Notes
No ratings yet
Alg Ds 1 Lecture Notes
86 pages
The Computer Science Handbook
100% (2)
The Computer Science Handbook
271 pages
Guide
No ratings yet
Guide
160 pages
DWGX
No ratings yet
DWGX
262 pages
Advanced Data Structures
No ratings yet
Advanced Data Structures
320 pages
Hilfinger Data Structures
No ratings yet
Hilfinger Data Structures
253 pages
Data Structure PDF
No ratings yet
Data Structure PDF
233 pages
Data-Structures in Java
No ratings yet
Data-Structures in Java
233 pages
Instant Access to (Ebook) Algorithms For Competitive Programming by David Esparza Alba,J A Ruiz Leal ISBN 9788700297067, 9798700297066, 8700297062 ebook Full Chapters
100% (9)
Instant Access to (Ebook) Algorithms For Competitive Programming by David Esparza Alba,J A Ruiz Leal ISBN 9788700297067, 9798700297066, 8700297062 ebook Full Chapters
62 pages
Java Data Structures Hilfinger
No ratings yet
Java Data Structures Hilfinger
231 pages
FIT2004 (Contents)
No ratings yet
FIT2004 (Contents)
3 pages
Instant download Algorithms For Competitive Programming 1st Edition David Esparza Alba pdf all chapter
100% (2)
Instant download Algorithms For Competitive Programming 1st Edition David Esparza Alba pdf all chapter
65 pages
Java Algorithms Interview Challenger
No ratings yet
Java Algorithms Interview Challenger
194 pages
2017 Muusse
No ratings yet
2017 Muusse
48 pages
lc_1
No ratings yet
lc_1
337 pages
Mehlhorn K., Sanders P. Concise Algorithmics, The Basic Toolbox 124ñ PDF
No ratings yet
Mehlhorn K., Sanders P. Concise Algorithmics, The Basic Toolbox 124ñ PDF
124 pages
Data Structures and Algorithms With Object-Oriented Design Patterns in Python
No ratings yet
Data Structures and Algorithms With Object-Oriented Design Patterns in Python
14 pages
Algorithms For Competitive Programming 2021
No ratings yet
Algorithms For Competitive Programming 2021
371 pages
Algorithms For Competitive Programming 1st Edition David Esparza Alba - Instantly access the full ebook content in just a few seconds
100% (2)
Algorithms For Competitive Programming 1st Edition David Esparza Alba - Instantly access the full ebook content in just a few seconds
61 pages
Crash Course Coding Companion
No ratings yet
Crash Course Coding Companion
136 pages
CP Handbook PDF
No ratings yet
CP Handbook PDF
296 pages
Algorithms Everything
No ratings yet
Algorithms Everything
33 pages
Coding Interview Preparation by Kaiyu Zheng 1651769215
No ratings yet
Coding Interview Preparation by Kaiyu Zheng 1651769215
100 pages
Prep Doc Coding Algo
No ratings yet
Prep Doc Coding Algo
100 pages
Re-Cording Lives: Governing Asylum in Switzerland and the Need to Resolve
From Everand
Re-Cording Lives: Governing Asylum in Switzerland and the Need to Resolve
Ephraim Pörtner
No ratings yet
A Discourse Analysis of 1 Peter
From Everand
A Discourse Analysis of 1 Peter
Ervin Ray Starwalt
No ratings yet
Gray Hat Hacking the Ethical Hacker's
From Everand
Gray Hat Hacking the Ethical Hacker's
Çağatay Şanlı
5/5 (1)
Onefs 8.2.1 Backup and Recovery Guide
No ratings yet
Onefs 8.2.1 Backup and Recovery Guide
124 pages
Assignment 4
No ratings yet
Assignment 4
4 pages
PowerSwitch SmartFabric OS10 REST API Implementation Participant Guide
No ratings yet
PowerSwitch SmartFabric OS10 REST API Implementation Participant Guide
43 pages
F MC-16F MB90220 Series: 16-Bit Proprietary Microcontroller
No ratings yet
F MC-16F MB90220 Series: 16-Bit Proprietary Microcontroller
105 pages
3GPP TR 32.814
No ratings yet
3GPP TR 32.814
25 pages
Ubiquiti Nanostation2 Loco 8dbi AP
No ratings yet
Ubiquiti Nanostation2 Loco 8dbi AP
5 pages
Packet Tracer 4.0 CCNA1 Hands-On Final: Objective
No ratings yet
Packet Tracer 4.0 CCNA1 Hands-On Final: Objective
4 pages
Visual Studio 2022 In Depth Ockert J Du Preez - The ebook is available for quick download, easy access to content
100% (2)
Visual Studio 2022 In Depth Ockert J Du Preez - The ebook is available for quick download, easy access to content
79 pages
Cissp Sunflower Slides PDF
No ratings yet
Cissp Sunflower Slides PDF
131 pages
Terminal Paper PDF
100% (1)
Terminal Paper PDF
4 pages
DIDM Adsense & Blogging
No ratings yet
DIDM Adsense & Blogging
7 pages
CCNA Study Plan
No ratings yet
CCNA Study Plan
11 pages
Aspiring Minds Sample Papers
No ratings yet
Aspiring Minds Sample Papers
14 pages
Cloud-Based Passenger Experience Management in Bus Fare Ticketing Systems Using Random Forest Algorithm
No ratings yet
Cloud-Based Passenger Experience Management in Bus Fare Ticketing Systems Using Random Forest Algorithm
6 pages
Airport Management System
33% (3)
Airport Management System
8 pages
Speed and Pothole Detection
No ratings yet
Speed and Pothole Detection
15 pages
Log
No ratings yet
Log
296 pages
Protastructure Design Guide Ifc and DXF Import
No ratings yet
Protastructure Design Guide Ifc and DXF Import
21 pages
Azure Capacity Planning - Google Search - 2
No ratings yet
Azure Capacity Planning - Google Search - 2
2 pages
ACS
No ratings yet
ACS
14 pages
S-23 WMN model answer (2 files merged) (1)
No ratings yet
S-23 WMN model answer (2 files merged) (1)
6 pages
AOL - Oracle FNDLOAD Commands
No ratings yet
AOL - Oracle FNDLOAD Commands
8 pages
DataSheet HP470 PDF
No ratings yet
DataSheet HP470 PDF
4 pages
Extra Quadrata Brochure
No ratings yet
Extra Quadrata Brochure
15 pages
CVBN
No ratings yet
CVBN
2 pages
Deloitte - Introduction To TS&A - Cloud
100% (1)
Deloitte - Introduction To TS&A - Cloud
2 pages
3 - MT 103 - 202 Base Procedure For Receiver Review and Comments Ao
No ratings yet
3 - MT 103 - 202 Base Procedure For Receiver Review and Comments Ao
4 pages
Reviewer-MIL-2nd-Quarter-ABM-STEM
No ratings yet
Reviewer-MIL-2nd-Quarter-ABM-STEM
4 pages
IP Multicast Technical Overview
No ratings yet
IP Multicast Technical Overview
9 pages