Lecture Notes for Algorithm Analysis and Design

Sandeep Sen1

March 16, 2008

1 Department of Computer Science and Engineering, IIT Delhi, New Delhi 110016, India.
E-mail:ssen@cse.iitd.ernet.in
Contents

1 Model and Analysis 6

1.1 Computing Fibonacci numbers . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Fast Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Model of Computation . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Other models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.1 External memory model . . . . . . . . . . . . . . . . . . . . . 11
1.4.2 Parallel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Warm up problems 14
2.1 Euclid’s algorithm for GCD . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Extended Euclid’s algorithm . . . . . . . . . . . . . . . . . . . 15
2.2 Finding the k-th element . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Choosing a random splitter . . . . . . . . . . . . . . . . . . . 17
2.2.2 Median of medians . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Sorting words . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Mergeable heaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.1 Merging Binomial Heaps . . . . . . . . . . . . . . . . . . . . . 21

3 Optimization I :
Brute force and Greedy strategy 22
3.1 Heuristic search approaches . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.1 Game Trees * . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 A framework for Greedy Algorithms . . . . . . . . . . . . . . . . . . . 26
3.2.1 Maximal Spanning Tree . . . . . . . . . . . . . . . . . . . . . 28
3.2.2 A Scheduling Problem . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Efficient data structures for MST algorithms . . . . . . . . . . . . . . 29
3.3.1 A simple data structure for union-find . . . . . . . . . . . . . 30
3.3.2 A faster scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.3 The slowest growing function ? . . . . . . . . . . . . . . . . . 32
3.3.4 Putting things together . . . . . . . . . . . . . . . . . . . . . . 33

1
 3.3.5 Path compression only . . . . . . . . . . . . . . . . . . . . . . 34

4 Optimization II :
Dynamic Programming 35
4.1 A generic dynamic programming formulation . . . . . . . . . . . . . . 36
4.2 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.1 Context Free Parsing . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.2 Longest monotonic subsequence . . . . . . . . . . . . . . . . . 37
4.2.3 Function approximation . . . . . . . . . . . . . . . . . . . . . 38
4.2.4 Viterbi’s algorithm for Expectation Maximization . . . . . . . 39

5 Searching 41
5.1 Skip Lists - a simple dictionary . . . . . . . . . . . . . . . . . . . . . 41
5.1.1 Construction of Skip-lists . . . . . . . . . . . . . . . . . . . . . 41
5.1.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 Treaps : Randomized Search Trees . . . . . . . . . . . . . . . . . . . 44
5.3 Universal Hashing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.3.1 Example of a Universal Hash function . . . . . . . . . . . . . 47
5.4 Perfect Hash function . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4.1 Converting expected bound to worst case bound . . . . . . . . 49
5.5 A log log N priority queue . . . . . . . . . . . . . . . . . . . . . . . . 49

6 Multidimensional Searching and Geometric algorithms 51

6.1 Interval Trees and Range Trees . . . . . . . . . . . . . . . . . . . . . 51
6.1.1 Two Dimensional Range Queries . . . . . . . . . . . . . . . . 52
6.2 k-d trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.3 Priority Search Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.4 Planar Convex Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.4.1 Jarvis March . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.4.2 Graham’s Scan . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.4.3 Sorting and Convex hulls . . . . . . . . . . . . . . . . . . . . . 57
6.5 A Quickhull Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.5.1 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.5.2 Expected running time ∗ . . . . . . . . . . . . . . . . . . . . . 61
6.6 Point location using persistent data structure . . . . . . . . . . . . . 62

7 Fast Fourier Transform and Applications 65

7.1 Polynomial evaluation and interpolation . . . . . . . . . . . . . . . . 65
7.2 Cooley-Tukey algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.3 The butterfly network . . . . . . . . . . . . . . . . . . . . . . . . . . 68

2
 7.4 Schonage and Strassen’s fast multiplication . . . . . . . . . . . . . . . 69

8 String matching and finger printing 72

8.1 Rabin Karp fingerprinting . . . . . . . . . . . . . . . . . . . . . . . . 72
8.2 KMP algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8.2.1 Potential method and amortized analysis . . . . . . . . . . . . 75
8.2.2 Analysis of the KMP algorithm . . . . . . . . . . . . . . . . . 75
8.2.3 Pattern Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 76
8.3 Generalized String matching . . . . . . . . . . . . . . . . . . . . . . . 76
8.3.1 Convolution based approach . . . . . . . . . . . . . . . . . . . 76

9 Graph Algorithms 79
9.1 Applications of DFS . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
9.1.1 Strongly Connected Components (SCC) . . . . . . . . . . . . 79
9.1.2 Finding Biconnected Components (BCC) . . . . . . . . . . . 80
9.2 Path problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
9.2.1 Bellman Ford SSSP Algorithm . . . . . . . . . . . . . . . . . . 82
9.2.2 Dijkstra’s SSSP algorithm . . . . . . . . . . . . . . . . . . . . 84
9.2.3 Floyd-Warshall APSP algorithm . . . . . . . . . . . . . . . . . 85
9.3 Maximum flows in graphs . . . . . . . . . . . . . . . . . . . . . . . . 85
9.3.1 Max Flow Min Cut . . . . . . . . . . . . . . . . . . . . . . . . 87
9.3.2 Ford and Fulkerson method . . . . . . . . . . . . . . . . . . . 88
9.3.3 Edmond Karp augmentation strategy . . . . . . . . . . . . . . 88
9.3.4 Monotonicity Lemma and bounding the iterations . . . . . . . 88
9.4 Global Mincut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
9.4.1 The contraction algorithm . . . . . . . . . . . . . . . . . . . . 90
9.4.2 Probability of mincut . . . . . . . . . . . . . . . . . . . . . . . 91

10 NP Completeness and Approximation Algorithms 93

10.1 Classes and reducibility . . . . . . . . . . . . . . . . . . . . . . . . . . 94
10.2 Cook Levin theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10.3 Common NP complete problems . . . . . . . . . . . . . . . . . . . . . 97
10.3.1 Other important complexity classes . . . . . . . . . . . . . . . 97
10.4 Combating hardness with approximation . . . . . . . . . . . . . . . . 99
10.4.1 Equal partition . . . . . . . . . . . . . . . . . . . . . . . . . . 99
10.4.2 Greedy set cover . . . . . . . . . . . . . . . . . . . . . . . . . 100
10.4.3 The metric TSP problem . . . . . . . . . . . . . . . . . . . . . 101
10.4.4 Three colouring . . . . . . . . . . . . . . . . . . . . . . . . . . 102
10.4.5 Maxcut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

3
A Recurrences and generating functions 104
A.1 An iterative method - summation . . . . . . . . . . . . . . . . . . . . 104
A.2 Linear recurrence equations . . . . . . . . . . . . . . . . . . . . . . . 106
A.2.1 Homogeneous equations . . . . . . . . . . . . . . . . . . . . . 106
A.2.2 Inhomogeneous equations . . . . . . . . . . . . . . . . . . . . 107
A.3 Generating functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
A.3.1 Binomial theorem . . . . . . . . . . . . . . . . . . . . . . . . . 109
A.4 Exponential generating functions . . . . . . . . . . . . . . . . . . . . 109
A.5 Recurrences with two variables . . . . . . . . . . . . . . . . . . . . . 110

B Refresher in discrete probability and probabilistic inequalities 112

B.1 Probability generating functions . . . . . . . . . . . . . . . . . . . . . 113
B.1.1 Probabilistic inequalities . . . . . . . . . . . . . . . . . . . . . 114

4
 Preface
This write-up is a rough chronological sequence of topics that I have covered in the
past in postgraduate and undergraduate courses on Design and Analysis of Algorithms
in IIT Delhi. A quick browse will reveal that these topics are covered by many
standard textbooks in Algorithms like AHU, HS, CLRS, and more recent ones like
Kleinberg-Tardos and Dasgupta-Papadimitrou-Vazirani.
What motivated me to write these notes are
(i) As a teacher, I feel that the sequence in which the topics are exposed has a
significant impact on the appreciation and understanding of the subject.
(ii) Most textbooks have far too much material for one semester and often intimidate
an average student. Even when I pick and choose topics, the ones not covered have
a distracting effect on the reader.
(iii) Not prescribing a textbook in a course (that I have done in the past) creates
insecurity among many students who are not adept at writing down notes as well
as participating in class discussions so important for a course like algorithms. (As a
corollary, this may make it easier for some of the students to skip some lectures.)
(iv) Gives me some flexibility about asking students to fill up some of the gaps left
deliberately or inadvertently in my lectures.
(v) Gives my students some idea of the level of formalism I expect in the assignments
and exams - this is somewhat instructor dependent.
(vi) The students are not uniformly mature so that I can demand formal scribing.
I am sure that many of my colleagues have felt about one or more of the above
reasons before they took the initiative at some point as is evidenced by the availability
of many excellent notes that are accessible via internet. This is a first draft that I
am making available in the beginning of the semester and I am hoping to refine and
fill up some of the incomplete parts by the middle of this semester. The notes are
likely to contain errors, in particular, typographic. I am also relying on the power
of collective scrutiny of some of the brightest students to point these out and I will
endeavour to update this every week.
I have also promised myself that I will not get carried away in adding to the
present version beyond 1xx pages.
Sandeep Sen
July 2007

5
Chapter 1

Model and Analysis

When we make a claim like Algorithm A has running time O(n2 log n), we have an
underlying computational model where this statement is valid. It may not be true if
we change the model. Before we formalize the notion of a computational model, let
us consider the example of computing Fibonacci numbers.

1.1 Computing Fibonacci numbers

One of the most popular sequence is the Fibonacci sequence defined by

 0 i=0
Fi = 1 i=1
Fi−1 + Fi−2 otherwise for i ≥ 2


Exercise 1.1 How large is Fn , the n-th Fibonacci number - derive a closed form.

It is not difficult to argue that it grows exponentially with n. You can also prove that
n−2
X
Fn = 1 + Fi
i=0

Since the closed form solution for Fn involves the golden ratio - an irrational number,
we must find out a way to compute it efficiently without incurring numerical errors
or approximation as it is an integer.
Method 1
Simply use the recursive formula. Unfortunately, one can easily argue that the
number of operations (primarily additions) involved is proportional to the value of Fn
(just unfold the recursion tree where each internal node corresponds to an addition.

6
As we had noted earlier this leads to an exponential time algorithm and we can’t
afford it.
Method 2
Observe that we only need the last two terms of the series to compute the new
term. So by applying the idea of dynamic programming we gradually compute the
Fn starting with F0 = 0 and F1 = 1.
This takes time that is proportional to approximately n additions where each
addition involves adding (increasingly large) numbers. The size of F ⌈n/2⌉ is about
n/2 bits so the last n/2 computations are going to take Ω(n) steps 1 culminating in
an O(n2 ) algorithm.
Since the n-th Fibonacci number is at most n bits, it is reasonable to look for a
faster algorithm.
Method 3
    
Fi 1 1 Fi−1
=
Fi−1 1 0 Fi−2
By iterating the above equation we obtain
   n−1  
Fn 1 1 1
=
Fn−1 1 0 0

To compute An , where A is a square matrix we can extend the following strategy for
computing xn where n is an integer.
 2
x2k = (xk ) for even integral powers
2k+1 2k
x =x·x for odd integral powers
The number of multiplications taken by the above approach to compute xn is
bounded by 2log n (Convince yourself by writing a recurrence). However, the actual
running time depends on the time to multiply two numbers which in turn depends on
their lengths (number of digits). If we assume that M(n) is the number of (bit-wise)
steps to multiply two n bit numbers. Therefore the number of steps to implement
the above approach must take into account the lengths of numbers that are being
multiplied. The following observations will be useful.
The length of xk is bounded by k · |x| where |x| is the length of x.
Therefore, the cost of the the squaring of xk is bounded by M(k|x|). Similarly, the
cost of computing x × x2k can also be bound by M(2k|x|).The overall recurrence for
computing xn can be written as

TB (n) ≤ TB (⌊n/2⌋) + M(n|x|)

1
Adding two k bit numbers take Θ(k)

7
where TB (n) is the number of bit operations to compute the n-th power using the
previous recurrence. The solution of the above recurrence can be written as the
following summation (by unfolding)
log n
X
M(2i |x|)
i=1

If M(2i) > 2M(i), then the above summation can be bounded by O(M(n|x|), i.e. the
cost the last squaring operation.
In our case, A is a 2 × 2 matrix - each squaring operation involves 8 multiplication
and 4 additions involving entries of the matrix. Since multiplications are more ex-
pensive than additions, let us count the cost of multiplications only. Here, we have to
keep track of the lengths of the entries of the matrix. Observe that if the maximum
size of an entry is |x|, then the maximum size of an entry after squaring is at most
2|x| + 1 (Why ?).

Exercise 1.2 Show that the cost of computing An is O(M(n|x|) where A is a 2 × 2

matrix and the maximum length of any entry is |x|.

So the running time of computing Fn using Method 3 is dependent on the multipli-

cation algorithm. Well, multiplication is multiplication - what can we do about it
? Before that let us summarize what we know about it. Multiplying two n digit
numbers using the add-and-shift method takes O(n2 ) steps where each step involves
multiplying two single digits (bits in the case of binary representation), and generat-
ing and managing carries. For binary representation this takes O(n) for multiplying
with each bit and finally n shifted summands are added - the whole process takes
O(n2 ) steps.
Using such a method of multiplication implies that we cannot do better than
Ω(n2 ) steps to compute Fn . For any significant (asymptotically better) improvement,
we must find a way to multiply faster.

1.2 Fast Multiplication

Problem Given two numbers A and B in binary, we want to compute the product
A × B.
Let us assume that the numbers A and B have lengths equal to n = 2k - this will
keep our calculations simpler without affecting the asymptotic analysis.

A × B = (2n/2 · A1 + A2 ) × (2n/2 · B1 + B2 )

8
where A1 (B1 ) is the leading n/2 bits of A (B). Likewise A2 is the trailing n/2 bits
of A. We can expand the above product as

A1 × B1 · 2n/2 + (A1 × B2 + A2 × B1 ) · 2n/2 + A2 × B2

Observe that multiplication by 2k can be easily achieved in binary by adding k trailing

0’s (likewise in any radix r, multiplying by r k can be done by adding trailing zeros).
So the product of two n bit numbers can be achieved by recursively computing four
products of n/2 bit numbers.

Exercise 1.3 What is the time to multiply using the above method - write and solve
an appropriate recurrence ?

We can achieve an improvement by reducing it to three recursive calls of multiplying

n/2 bit numbers by rewriting the coefficient of 2n/2 as follows

A1 × B2 + A2 × B1 = (A1 + A2 ) × (B1 + B2 ) − (A1 × B1 ) − (A2 × B2 )

Although strictly speaking, A1 + A2 is not n/2 bits but at most n/2 + 1 bits (Why ?),
we can still view this as computing three separate products involving n/2 bit numbers
recursively and subsequently subtracting appropriate terms to get the required prod-
ucts. Subtraction and additions are identical in modulo arithmetic (2’s complement),
so the cost of subtraction can be bounded by O(n). (What is maximum size of the
numbers involved in subtraction ?). This gives us the following recurrence

TB (n) ≤ 3 · TB (n/2) + O(n)

where the last term accounts for addition, subtractions and shifts.

Exercise 1.4 With appropriate terminating condition, show that the solution to the
recurrence is O(nlog2 3 ).

The running time is roughly O(n1.7 ) which is asymptotically better than n2 and there-
fore we have succeeded in designing an algorithm to compute Fn faster than n2 .

It is possible to multiply much faster using a generalization of the above method in

O(n log n log log n) by a method of Schonage and Strassen. However it is quite in-
volved as it uses Discrete Fourier Transform computation over modulo integer rings
and has fairly large constants that neutralize the advantage of the asymptotic im-
provement unless the numbers are a few thousand bits long. It is however conceivable
that such methods will become more relevant as we may need to multiply large keys
for cryptographic/security requirements.

9
1.3 Model of Computation
Although there are a few thousand variations of the computer with different archi-
tectures and internal organization, it is best to think about them at the level of the
assembly language. Despite architectural variations, the assembly level language sup-
port is very similar - the major difference being in the number of registers and the word
length of the machine. But these parameters are also in a restricted range of a factor
of two, and hence asymptotically in the same ball park. In summary, think about any
computer as a machine that supports a basic instruction set consisting of arithmetic
and logical operations and memory accesses (including indirect addressing). We will
avoid cumbersome details of the exact instruction set and assume realistically that
any instruction of one machine can be simulated using a constant number of available
instruction of another machine. Since analysis of algorithms involves counting the
number of operations and not the exact timings (which could differ by an order of
magnitude), the above simplification is justified.
The careful reader would have noticed that during our detailed analysis of Method
3 in the previous sections, we were not simply counting the number of arithmetic
operations but actually the number of bit-level operations. Therefore the cost of
a multiplication or addition was not unity but proportional to the length of the
input. Had we only counted the number of multiplications for computing xn , that
would only be O(log n). This would indeed be the analysis in a uniform cost model
where only the number of arithmetic (also logical) operations are counted and does
not depend on the length of the operands. A very common us of this model is for
comparison-based problems like sorting, selection, merging, and many data-structure
operations. For these problems, we often count only the number of comparisons (not
even other arithmetic operations) without bothering about the length of the operands
involved. In other words, we implicitly assume O(1) cost for any comparison. This
is not considered unreasonable since the size of the numbers involved in sorting do
not increase during the course of the algorithm for majority of the commonly known
sorting problems. On the other hand consider the following problem of repeated
n
squaring n times starting with 2. The resultant is a number 22 which requires 2n
bits to be represented. It will be very unreasonable to assume that a number that is
exponentially long can be written out (or even stored) in O(n) time. Therefore the
uniform cost model will not reflect any realistic setting for this problem.
On the other extreme is the logarithmic cost model where the cost of an operation
is proportional to length of the operands. This is very consistent with the physical
world and also has close relation with the Turing Machine model which is a favorite of
complexity theorists. Our analysis in the previous sections is actually done with this
model in mind. It is not only the arithmetic operations but also the cost of memory
access is proportional to the length of the address and the operand.

10
 The most commonly used model is something in between. We assume that for
an input of size n, any operation involving operands of size log n 2 takes O(1) steps.
This is justified as follows. All microprocessor chips have specialized hardware circuits
for arithmetic operations like multiplication, addition, division etc. that take a fixed
number of clock cycles when the operands fit into a word. The reason that log n is a
natural choice for a word is that, even to address an input size n, you require log n
bits of address space. The present high end microprocessor chips have typically 2-4
GBytes of RAM and about 64 bits word size - clearly 264 exceeds 4 GBytes. We will
also use this model, popularly known as Random Access Machine (or RAM in
short) except for problems that deal with numbers as inputs like multiplication in
the previous section where we will invoke the log cost model. In the beginning, it
is desirable that for any algorithm, you get an estimate of the maximum size of the
numbers to ensure that operands do not exceed Ω(log n) so that it is safe to use the
RAM model.

1.4 Other models

There is clear trade-off between the simplicity and the fidelity achieved by an abstract
model. One of the obvious (and sometimes serious) drawback of the RAM model is
the assumption of unbounded number of registers since the memory access cost is
uniform. In reality, there is a memory hierarchy comprising of registers, several levels
of cache, main memory and finally the disks. We incur a higher access cost as we
go from registers towards the disk and for techological reason, the size of the faster
memory is limited. There could be a disparity of 105 between the fastest andthe
slowest memory which makes the RAM model somewhat suspect for larger input
sizes. This has been redressed by the external memory model.

1.4.1 External memory model

In this model, the primary concern is the number of disk accesses. Given the rather
high cost of a disk access compared to any CPU operation, this model actually ignores
all other costs and counts only the number of disk accesses. The disk is accessed as
contiguous memory locations called blocks. The blocks have a fixed size B and the
simplest model is parameterized by B and the size of the faster memory M. In this
two level model, the algorithms are only charged for transferring a block between the
internal and external memory and all other computation is
free. It turns out that in
this model, the cost of sorting n elements is O Bn logM/B Bn disk accesses and this is
also optimal.
2
We can also work with c log n bits as the asymptotic analysis does not change for a constant c.

11
 There are further refinements to this model that parameterizes multiple levels and
also accounts for internal computation. As the model becomes more complicated,
designing algorithms also becomes more challenging and often more laborious.

1.4.2 Parallel Model

The basic idea of parallel computing is extremely intuitive and a fundamental in-
tellectual pursuit. At the most intuitive level it symbolises what can be achieved
by cooperation among individuals in terms of expediting an activity. It is not in
terms of division of labor (or specialisation), but actually assuming similar capabil-
ities. Putting more labourers clearly speeds up the construction and similarly using
more than one processor is likely to speed up computation. Ideally, by using p pro-
cessors we would like to obtain a p-fold speed up over the conventional algorithms;
however the principle of decreasing marginal utility shows up. One of the intuititive
reasons for this that with more processors (as with more individuals), the commu-
nication requirements tend to dominate after a while. But more surprisingly, there
are algorithmic constraints that pose serious limitations to our objective of obtaining
proportional speed-up.
This is best demonstrated in the model called PRAM (or Parallel Random Ac-
cess Machine) which is the analogue of the RAM. Here p processors are connected to
a shared memory and the communication happens through reading and writing ina
globally shared memory. It is left to the algorithm designer to avoid read and write
conflicts. It is further assumed that all operations are synchorized globally and there
is no cost of synchronization. In this model, there is no extra overhead for communi-
cation as it charged in the same way as a local memory access. Even in this model, it
has been shown that it is not always possible to obtain ideal speed up. As an example
consider the elementary problem of finding the minimum of n elements. It has been
proved that with n processors, the time (parallel time) is at least Ω(log log n). For
certain problems, like depth first search of graphs, it is known that even if we use
any polynomial number of processors, we cannot obtain polylogarithmic time ! So,
clearly not all problems can be parallelized effectively.
A more realistic parallel model is the interconnection network model that has an
underlying communication network, usually a regular topology like a two-dimensional
mesh, hypercube etc. These can be embedded into VLSI chips and can be scaled
according to our needs. To implement any any parallel algorithm, we have to design
efficient schemes for data routing.
A very common model of parallel computation is a hardware circuit comprising of
basic logic gates. The signals are transmitted in parallel through different paths and
the output is a function of the input. The size of the circuit is the number of gates
and the (parallel) time is usually measured in terms of the maximum path length

12
from any input gate to the output gate (each gate contributes to a unit delay). Those
familiar with circuits for addition, comparison can analyse them in this framework.
The carry-save adder is a low-depth circuit that adds two n-bit numbers in about
O(log n) steps which is much faster than a sequential circuit that adds one bit at a
time taking n steps.
One of the most fascinating developments is the Quantum Model which is in-
herently parallel but it is also fundamentally different from the previous models. A
breakthrough result in recent years is a polynomial time algorithm for factorization
which forms the basis of many cryptographic protocals in the conventional model.
Biological Computing models is a very active area of research where scientists
are trying to assemble a machine out of DNA strands. It has potentially many
advantages over silicon based devices and is inherently parallel.

13
Chapter 2

Warm up problems

One of the primary challenges in algorithm design is to come up with provably optimal
algorithms. The optimality is with respect to the underlying model. In this chapter,
we look closely at some well-known algorithms for basic problems that uses basic
properties of the problem domain in conjunction with elementary analytical methods.

2.1 Euclid’s algorithm for GCD

Euclid’s algorithm for computing the greatest common divisor (gcd) of two positive
integers is allegedly the earliest known algorithm in a true sense. It is based on two
very simple observations that the gcd of numbers a, b satisfies
gcd(a, b) = gcd(a, a + b)
gcd(a, b) = b if b divides a
Exercise 2.1 Prove this rigorously.
The above also implies that gcd(a, b) = gcd(a − b, b) for b < a and repeated appli-
cations imply that gcd(a, b) = gcd(a mod b, b) where mod denotes the remainder
operation. So we have essentially derived Euclid’s algorithm, described formally as

Algorithm Euclid GCD

Input: Positive integers a, b such that b ≤ a

Output GCD of a, b

Let c = a mod b.
If c = 0 then return b else
return Euclid GCD(b, c)

14
Let us now analyze the running time of Euclid’s algorithm in the bit model. Since
it depends on integer division, which is a topic in its own right, let us address the
number of iterations of Euclid’s algorithm in the worst case.
Observation 2.1 The number a mod b ≤ a2 , i.e. the size of a mod b is strictly less
than |a|.
This is a simple case analysis based on b ≤ a2 and b > a2 . As a consequence of the above
observation, it follows that the the number of iterations of the Euclid’s algorithm is
bounded by |a|, or equivalently O(log a).
Exercise 2.2 Construct an input for which the number of iterations match this
bound.
So, by using the long division method to compute mod , the running time is bounded
by O(n2 log n) where n = |a| + |b|.

2.1.1 Extended Euclid’s algorithm

If you consider the numbers defined by the linear combinations of a, b, namely, {xa +
yb| x, yare integers it is known that

gcd(a, b) = min{xa + yb|xa + yb > 0}

Proof: Let ℓ = min{xa + yb|xa + yb > 0}. Clearly gcd(a, b) divides ℓ and hence
gcd(a, b) ≤ ℓ. We now prove that ℓ divides a (also b). Let us prove by contradiction
that a = ℓq + r where ℓ > r > 0. Now r = a − ℓq = (1 − xq)a − (yq)b contradicting
the minimality of ℓ. 2

The above result can be restated as ℓ divides a and b. For some applications,
we are interested in computing x and y corresponding to gcd(a, b). We can compute
them recursively along with the Euclid’s algorithm.
Exercise 2.3 Let (x′ , y ′ ) correspond to gcd(b, a mod b), i.e. gcd(b, a mod b) = x′ ·
b + y ′ · (a mod b). Then show that gcd(a, b) = y ′ · a + (x′ − q)b where q is the quotient
of the integer division of a by b.
One immediate application of the extended Euclid’s algorithm is computing the in-
verse in a multiplicative prime field Fp∗ where p is prime. Fq∗ = {0, 1, 2 . . . (q − 1)}
where the multiplication is performed modulo p. It is known 1 that for every number
1
since it forms a group

15
x ∈ Fq∗ − {0}, there exists y ∈ Fq∗ such that x · y = 1 which is also called the inverse
of x. To compute the inverse of a we can use extended Euclid algorithm to find s, t
such that sa + tp = 1 since a is relatively prime to p. By taking remainder modulo p,
we see that s mod p is te required inverse.

2.2 Finding the k-th element

Problem Given a set S of n elements, and an integer k, 1 ≤ k ≤ n, find an element
x ∈ S such that the rank of x is k. The rank of an element in a set S is k if x = xk
in the sorted set x1 , x2 , . . . xn where xi ∈ S. We will denote the rank of x in S by
R(x, S).
Note that k is not unique if the value of x is not unique, but the value of the k-th
element is unique. If S is a multiset, we can (hypothetically) append log n trailing
bits equal to the input index to each element. So an element xi can be thought of as
a pair (xi , i) so that every pair is unique since the input index in unique. The case
k = 1 (k = n) corresponds to finding the minimum (maximum) element.
We can easily reduce the selection problem to sorting by first sorting S and then
reporting the k-th element of the sorted set. But this also implies that we cannot
circumvent the lower bound of Ω(n log n) for comparison based sorting. If we want a
faster algorithm, we cannot afford to sort. For instance, when k = 1 or k = n, we can
easily select the minimum (maximum) element using n − 1 comparisons. The basic
idea for a faster selection algorithm is based on the following observation.
Given an element x ∈ S, we can answer the following query in n − 1 comparisons

Is x the k-th element or is x larger than the k-th element or is x smaller

than the k-th element ?

This is easily done by comparing x with all elements in S − {x} and finding the rank
of x. Using an arbitrary element x as a filter, we can subsequently confine our search
for the k-th element to either

(i) S> = {y ∈ S − {x}|y > x} if R(x, S) < k or

(ii) S< = {y ∈ S − {x}|y < x} if R(x, S) > k

In the fortuitous situation, R(x, S) = k,x is the required element. In case 1, we must
find k ′ -th element in S> where k ′ = k − R(x, S).
Suppose T (n) is the worst case running time for selecting the k-th element for any
k, then we can write the following recurrence

T (n) ≤ max{T (|S< |), T (|S> |)} + O(n)

16
A quick inspection tells us that if we can ensure max{|S< |, |S> |} ≤ ǫn for some
1/2 ≤ ǫ < n−1 n
, (Why the bounds ?) T (n) is bounded by O( 1−ǫ 1
· n). So it could
2
vary between Ω(n) and O(n ) - where a better running time is achieved by ensuring
a smaller value of ǫ.
An element x used to divide the set is often called a splitter or a pivot. So, now
we will discuss methods to select a good splitter. From our previous discussion, we
would like to select a splitter that has a rank in the range [ǫ · n, (1 − ǫ) · n] for a fixed
fraction ǫ. Typically, ǫ will be chosen as 1/4.

2.2.1 Choosing a random splitter

Let us analyze the situation where the splitter is chosen uniformly at random from S,
i.e., any of the n elements is equally likely to be chosen as the splitter. This can be
done using standard routines for random number generation in the range (1, 2, . . . n).
A central observation is
For a randomly chosen element r ∈ S, the probability
Pr{n/4 ≤ R(r, S) ≤ 3n/4} ≥ 1/2

It is easy to verify if the rank R(r, S) falls in the above range, and if it does not, then
we choose another element independently at random. This process is repeated till we
find a splitter in the above range - let us call such a splitter a good splitter.
How many times do we need to repeat the process ?
To answer this, we have to take a slightly different view. One can argue easily
that there is no guarantee that we will terminate after some fixed number of trials,
while it is also intuitively clear that it is extremely unlikely that we need to repeat
this more than say 10 times. The probability of failing 9 consecutive times, when the
success probability of picking a good splitter is ≥ 1/2 independently is ≤ 219 . More
precisely, the expected 2 number of trials is bounded by 2. So, in (expected) two trials,
we will find a good splitter that reduces the size of the problem to at most 43 n. This
argument can be repeated for the recursive calls, namely, the expected number of
splitter selection (and verification of its rank) is 2. If ni is the size of the problem
after i recursive calls with n0 = n, then the expected number of comparisons done
after the i-th recursive call is 2ni . The total expected number of comparisons X after
t calls can be written as X0 + X1 + . . . Xt where t is sufficiently large such that the
problem size nt ≤ C for some constant C (you can choose other stopping criteria)
and Xi is the number of comparisons done at stage i. By taking expectation on both
sides
E[X] = E[X1 + X2 + . . . Xt ] = E[X1 ] + E[X2 ] + . . . E[Xt ]
2
Please refer to the Appendix for a quick recap of basic measures of discrete probability

17
From the previous discussion E[Xi ] = 2ni and moreover ni ≤ 43 ni−1 . Therefore the
expected number of comparisons is bounded by 4n.

2.2.2 Median of medians

Partition the elements into groups of 5 and choose the median of each group. Denote
the groups by Gi and their medians by mi . Now consider the median of the set {mi }
which contains about n/5 elements. Denote the median of medians by M.
How many elements are guaranteed to be smaller than M ?
Wlog, assume all elements are distinct and that implies about n/103 medians that
are smaller than M. For each such median there are 3 elements that are smaller than
M, giving a total of at least n/10 · 3 = 3n/10 elements smaller than M. Likewise, we
can argue that there are at least 3n/10 elements larger than M. Therefore we can
conclude that 3n/10 ≤ R(M, S) ≤ 7n/10 which satisfies the requirement of a good
splitter. The next question is how to find M which is the median of medians. Each
mi can be determined in O(1) time because we are dealing with groups of size 5.
However, finding the median of n/5 elements is like going back to square one ! But
it is n/5 elements instead of n and therefore, we can apply a recursive strategy. We
can write a recurrence for running time as follows
 
7n n
T (n) ≤ T +T + O(n)
10 5

where the second recursive call is to find the median of medians (for finding a good
splitter). After we find the splitter ( by recursively applying the same algorithm), we
use it to reduce the original problem size to at most 7n 10
.

Exercise 2.4 By using an appropriate terminating condition, show that T (n) ∈

O(n). Try to minimize the leading constant by adjusting the size of the group.

2.3 Sorting words

Problem Given n words w1 , w2 . . . wn of lengths l1 , l2 . . . ln respectively, arrange the
words in a lexicographic order. A word is an ordered sequence of characters from a
given alphabet Σ. P
Recall that lexicographic ordering refers to the dictionary ordering. Let N = i li ,
i.e. the cumulative length of all words. A single word may be very long and we
cannot assume that it fits into a single word of the computer. So, we cannot use
3
Strictly speaking we should be using the floor function but we are avoiding the extra symbols
and it does not affect the analysis.

18
straightforward comparison sorting. Let us recall some basic results about integer
sorting.

Claim 2.1 n integers in the range [1..m] can be sorted in O(n + m) steps.

A sorting algorithm is considered stable if the relative order of input elements having
identical values is preserved in the sorted output.

Claim 2.2 Using stable sorting, n integers in the range [1..mk ] can be sorted in
O(k(n + m) steps.

The above is easily achieved by applying integer sorting in the range [1..m] starting
from the least significant digits - note that the maximum number of digits in radix
m representation is k. If we apply the same algorithm for sorting words, then the
running time will be O(L(n + |Σ|) where L = max{l1 , l2 ..ln }. This is not satisfactory
since L · n can be much larger than N (size of input).
The reason that the above method is potentially inefficient is that many words
may be much shorter than L and hence by considering them to be length L words (by
hypothetical trailing blanks), we are increasing the input size asymptotically. When
we considered radix sort as a possible solution, the words have to be left-aligned, i.e.,
all words begin from the same position. To make radix sort efficient and to avoid
redundant comparison (of blanks), we should not consider a word until the radix sort
reaches the right boundary of the word. The radix sort will take a maximum of L
rounds and a word of length l will start participating from the L − l + 1 iteration.
This can be easily achieved. A bigger challenge is to reduce the range of sorting in
each iteration depending on which symbols of the alphabet participate.
Given a word wi = ai,1 ai,2 . . . ai,li , where ai,j ∈ Σ, we form the following pairs -
(1, ai,1 ), (2, ai,2) . . . . There are N such pairs from the n words and we can think of
them as length two strings where the first symbol is from the range [1..L] and the
second symbol is from Σ. We can sort them using radix sort in two rounds in time
proportional to O(N + L + |Σ|) which is O(N + |Σ|) since N > L. From the sorted
pairs we know exactly which symbols appear in a given position (between 1 and L)
- let there be mi words that have non-blank symbols in position i. We also have the
ordering of symbols in position i which is crucial to implement integer sort in O(mi )
steps.
Now we go back to sorting the given words using radix sort where we will use
the information available from the sorted pairs. When we are sorting position i from
the left, we apply integer sort in the range [1..mi ] where the ordered buckets are also
defined by the sorted pairs. We do not move the entire words into the buckets but
only the pointers (which could be the input index of the words) associated with the
words. For every round, we allocate an array of size mi where we place the pointers

19
according to the sorted order of the symbols involved. For same symbols, we maintain
the order of the previous round (stable sorting). We must also take care of the new
words that start participating in the radix sort - once a word participates, it will
participate in all future rounds. (Where should the new words be placed within its
symbol group ?)
The analysis of this algorithm
P can be done by looking at the cost of each radix
sort which is proportional to Li=1 O(mi ) which can be bounded by N. Therefore
overall running time of the algorithm is the sum of sorting the pairs and the radix
sort. This is given by O(N + |Σ|). If |Σ| < N, then the optimal running time is given
by O(N).

2.4 Mergeable heaps

Heaps 4 are one of the most common implementation of priority queues and are
known to support the operations min , delete-min , insert , delete in logarithmic
time. A complete binary tree (often implemented as an array) is one of the simplest
ways to represent a heap. In many situations, we are interested in an additional
operation, namely, combining two heaps into a single heap. A binary tree doesn’t
support fast (polylogarithmic) merging and is not suitable for this purpose - instead
we use binomial trees.
A binomial tree Bi of order i is recursively defined as follows
• B0 is a single node

• Combine Bi+1 consists of making the root node of one Bi a child of the other
Bi .

Exercise 2.5 Prove the following properties for Bi by induction

(i) The number of nodes in Bi equals 2i .
(ii) The height of Bk is k
(by definition B0 has height 0).
(iii) There are exactly ki nodes at depth k for k = 0, 1 . . .

A binomial heap is an ordered set of binomial trees such that for any i there is
at most one Bi .
Let us refer to the above property as the unique-order property. We actually maintain
list of the root nodes in increasing order of their degrees.
You may think of the above property as a binary representation of a number where
the i-th bit from right is 0 or 1 and in the latter case, its contribution is 2i (for LSB
i = 0). From the above analogue, a Binomial Heap on n elements has log n Binomial
4
we are assuming min heaps

20
trees. Therefore, finding the minimum element can be done in O(log n) comparisons
by finding the minimum of the log n roots.

2.4.1 Merging Binomial Heaps

Merging two Binomial Heaps amounts to merging the root lists and restoring the
unique-order property. First we merge two lists of size at most log n. Subsequently,
we walk along the list combining two trees of the same degree whenever we find them
- they must be consecutive. We know that by combining two Bi trees, we obtain a
Bi+1 tree which may have to be now combined with the next Bi+1 tree if there exists
one. In this process, if you have three consecutive Binomial trees of order i (can it
happen ?), merge the second and third instead the first and second - it simplifies the
procedure. Combining two binomial trees takes O(1) time, so the running time is
proportional to the number of times we combine.

Claim 2.3 Two Binomial heaps can be combined in O(log n) steps where the total
number of nodes in the two trees is n.

Every time we combine two trees, the number of binomial trees decreases by one, so
there can be at most 2 log n times where we combine trees.
Remark The reader may compare this with the method for summing two numbers
in binary representation.

Exercise 2.6 Show that the delete-min operation can be implemented in O(log n)
steps using merging.

Inserting a new element is easy - add a node to the root list and merge. Deletion
takes a little thought. Let us first consider an operation decrease-key. This happens
when a key value of a node x decreases. Clearly, the min-heap property of the parent
node, parent(x) may not hold. But this can be restored by exchanging the node x
with its parent. This operation may have to be repeated at the parent node. This
continues until the value of x is greater than its current parent or x doesn’t have a
parent, i.e., it is the root node. The cost is the height of a Binomial tree which is
O(log n).

Exercise 2.7 Show how to implement the insert operation in O(log n) comparisons.

21
Chapter 3

Optimization I :
Brute force and Greedy strategy

A generic definition of an optimization problem involves a set of constraints that

defines a subset in some underlying space (like the Euclidean space Rn ) called the
feasible subset and an objective function that we are trying to maximize or minimize
as the case may be over the feasible set. A very important common optimization
problem is Linear Programming where there is a finite set of linear constraints and
the objective function is also linear - the underlying space is Euclidean. A convex
function f satisfies f (λ · x + (1 − λ) · y) ≤ λf (x) + (1 − λ)f (y) where 0 < λ < 1.
A convex programming problem is one where the objective function is convex and so
is the feasible set. Convex programming problem over Euclidean real space have a
nice property that local optima equals the global optimum. Linear programming falls
under this category and there are provably efficient algorithms for linear programming.
Many important real life problems can be formulated as optimization problems
and therefore solving them efficiently is one of the most important area of algorithm
design.

3.1 Heuristic search approaches

In this section, we will use the knapsack problem as the running example to expose
some of the algorithmic ideas. The 0-1 Knapsack problem is defined as follows.
Given a knapsack of capacity C and n objects of volumes {w1 , w2 . . . wn } and
profits {p1 , p2 . . . pn }, the objective is to choose a subset of n objects that fits into the
knapsack and that maximizes the total profit.
In more formal terms, let xi be 1 if object i is present in the subset and 0 otherwise.

22
The knapsack problem can be stated as
n
X n
X
Maximize xi · pi subject to xi · wi ≤ C
i=0 i=0

Note that the constraint xi ∈ {0, 1} is not linear. A simplistic approach will be to
enumerate all subsets and select the one that satisfies the constraints and maximizes
the profits. Any solution that satisfies the capacity constraint is called a feasible
solution. The obvious problem with this strategy is the running time which is at
least 2n corresponding to the power-set of n objects.
We can imagine that the solution space is generated by a binary tree where we start
from the root with an empty set and then move left or right according to selecting
x1 . At the second level, we again associate the left and right branches with the
choice of x2 . In this way, the 2n leaf nodes correspond to each possible subset of the
power-set which corresponds to a n length 0-1 vector. For example, a vector 000 . . . 1
corresponds to the subset that only contains xn .
Any intermediate node at level j from the root corresponds to partial choice among
the objects x1 , x2 . . . xj . As we traverse the tree, we keep track of the best feasible
solution among the nodes visited - let us denote this by T . At a node v, let S(v)
denote the subtree rooted at v. If we can estimate the maximum profit P (v) among
the leaf nodes of S(v), we may be able to prune the search. Suppose L(v) and U(v)
are the lower and upperbounds of P (v), i.e. L(v) ≤ P (v) ≤ U(v). If U(v) < T , then
there is no need to explore S(v) as we cannot improve the current best solution, viz.,
T . In fact, it is enough to work with only the upper-bound of the estimates and L(v)
is essentially the current partial solution. As we traverse the tree, we also update
U(v) and if it is less than T , we do not search the subtree any further. This method
of pruning search is called branch and bound and although it is clear that there it is
advantageous to use the strategy, there may not be any provable savings in the worst
case.
Exercise 3.1 Construct an instance of a knapsack problem that visits every leaf node,
even if you use branch and bound. You can choose any well defined estimation.

Example 3.1

Let the capacity of the knapsack be 15 and the weights and profits are respectively
Profits 10 10 12 18
Volume 2 4 6 9
We will use the ratio of profit per volume as an estimation for upperbound. For
the above objects the ratios are 5, 2.5, 2 and 2. Initially, T = 0 and U = 5 × 15 = 75.
After including x1 , the residual capacity is 13 and T = 10. By proceeding this way, we

23
obtain T = 38 for {x1 , x2 , x4 }. By exploring further, we come to a stage when we have
included x1 and decided against including x2 so that L(v) = 10, and residual capacity
is 13. Should we explore the subtree regarding {x3 , x4 } ? Since profit per volume of
x3 is 2, we can obtain U(v) = 2 × 13 + 10 = 36 < L = 38. So we need not search this
subtree. By continuing in this fashion, we may be able to prune large portions of the
search tree. However, it is not possible to obtain any provable improvements.

3.1.1 Game Trees *

A minmax tree is used to represent a game between two players who alternately make
moves trying to win the game. We will focus on a special class of minmax trees called
AND-OR trees where alternate levels of trees are labelled as OR nodes starting with
the root node and the remaining are labelled AND nodes. Let 1 represent a win for
player 1 and 0 represent a loss for player 1 who makes the first move at the root -
the numbers are flipped for player 2. The leaf nodes correspond to the final state of
the game and are labelled 1 or 0 corresponding to win or loss for player 1. We want
to compute the values of the intermediate nodes using the following criteria. An OR
node has value 1 if one of the children is 1, and 0 otherwise - so it is like the boolean
function OR . An AND node behaves like a boolean AND function - it is 0 if one of
the children is 0. The interpretation is as follows - the player at the root can choose
any of the branches that leads to a win. However at the next level, he is at the mercy
of the other player - only when both branches for the other player leads to a win (for
the root), the root will win, otherwise the other player can inflict a loss.
So the evaluation of this Game Tree works as follows. Each leaf node is labelled 0
or 1 and an internal node as AND or OR - these will compute the boolean function
of the value of the two child nodes. The value of the gametree is the value available
at the root node.
First consider a single level AND tree that evaluates to 0. If we use a fixed order
to inspect the leaf nodes, in the worst case, both leaf nodes may have to be searched if
there is one 0 and it is the second branch. On the other hand, if we randomly choose
the order, for an answer 0, with probability 1/2, we will end up searching only one
node (the one labelled 0) and we do not have to evaluate the other one. Therefore
the expected number of look-ups is 3/2 for an AND node with answer 0 and the same
holds for an OR node with answer 1. For the other cases, there is no saving. However
any interesting game tree will have at least two levels, one AND and the other OR.
Then you can see that for an AND node to be 1, both the child OR nodes must be 1
which is the good case for OR.
In essence, we are applying the branch-and-bound method to this problem, and we
obtain a provable improvement in the following way. The two children are evaluated
in a random order.

24
 Consider a tree with depth 2k (i.e. 4k leaf nodes) with alternating AND and OR
nodes, each type having k nodes. We will show that the expected cost of evaluation
is 3k by induction on k.

Exercise 3.2 Show that for k = 1, the expected number of evaluations is 3. (You
must consider all cases of output and take the worst, since we are not assuming any
distribution on input or output).

Let us consider a root with label OR and its two AND children, say y and z, whose
children are OR nodes with 2(k − 1) depth. We have the two cases

output is 0 at the root Both y and z must evaluate to 0. Since these are AND
nodes, again with probability 1/2, we will end up evaluating only one of the
children (of y, z) that requires expected 21 · (1 + 2) · 3k−1 = 23 · 3k−1 steps for
y as well as z from Induction hypothesis. This adds upto a total of expected
2 · 23 · 3k−1 = 3k steps for y and z.

Output is 1 at the root At least one of the AND nodes y, z must be 1. With prob-
ability 1/2 this will be chosen first and this can be evaluated using the expected
cost of evaluating two OR nodes with output 1. By induction hypothesis this
is 2 · 3k−1.
The other possibility (also with probability 1/2) is that the first AND node (say
y) is 0 and the second AND node is 1. The expected cost of the first AND node
with 0 output is 1/2 · 3k−1 + 1/2 · (3k−1 + 3k−1 ) - the first term corresponds to
the scenario that the first child evaluates to 0 and the second term corresponds
to evaluating both children of y are evaluated. The expected cost of evaluating
y having value 0 is 3/2 · 3k−1.
The expected number of evaluation for second AND node z with output 1 is
2 · 3k−1 since both children must be evaluated.
So the total expected cost is 1/2 · 3k−1 (2 + 3/2 + 2) = 2.75 · 3k−1 < 3k .
In summary, for an OR root node, regardless of the output, the expected number
of evaluations is bounded by 3k .

Exercise 3.3 Establish a similar result for the AND root node.

If N the number of leaf nodes, then the expected number of evaluations is N log4 3 = N α
where α < 0.8.

25
3.2 A framework for Greedy Algorithms
There are very few algorithmic techniques for which the underlying theory is as precise
and clean as what we will discuss here. Let us define the framework. Let S be a set
and M be a subset 1 of 2S . A subset system (S, M) is called a matroid if it satisfies
the following property

For all subsets T ∈ M, for any T ′ ⊂ T , T ′ ∈ M

Note that the empty subset Φ ∈ M. The family of subsets M is often referred to as
independent subsets and one may think of M as the feasible subsets.

Example 3.2 For the maximal spanning tree problem on a graph G = (V, E), (E, F )
is a matroid where F is the set of all subgraphs without cycles (i.e. all the forests).

For any weight function w : S → R+ , the optimization problem is defined as finding

a subset from M for which the cumulative weight of the elements is maximum among
all choices of subsets from M. A simple way to construct a subset is the following
greedy approach.

Algorithm Gen Greedy

Let e1 , e2 . . . en be the elements of S in decreasing order of weights. Initialize T = Φ.

For i = 1 to n do

In the i-th stage

If T ∪ {ei } ∈ M, then T ← T ∪ {ei }

Output T as the solution

The running time of the algorithm is dependent mainly on the test for independence
which depends on the specific problem. M is not given explicitly as it may be very
large (even exponential2 ). Instead, a characterization of M is used to perform the
test.
What seems more important is the question - Is T the maximum weight subset ?
This is answered by the next result

Theorem 3.1 The following are equivalent

1
M is a family of subsets of S
2
The number of spanning trees of a complete graph is nn−2

26
 1. Algorithm Gen Greedy outputs the optimal subset for any choice of the weight
function.

2. exchange property
For any s1 , s2 ∈ M where |s1 | < |s2 |, then there exists e ∈ s2 − s1 such that
s1 ∪ {e} ∈ M.

3. For any A ⊂ S, all maximal subsets of A have the same cardinality. A maximal
subset T of A implies that there is no element e ∈ A − T such that T ∪ {e} ∈ M.
The obvious use of the theorem is to establish properties 2 or 3 to justify that a greedy
approach works for the problem. On the contrary, we can try to prove that one of the
properties doesn’t hold (by a suitable counterexample), then greedy cannot always
return the optimum subset.
Proof: We will prove it in the following cyclic implications - Property 1 implies Prop-
erty 2. Then Property 2 implies Property 3 and finally Property 3 implies Property
1.
Property 1 implies Property 2 We will prove it by contradiction. Suppose Prop-
erty 2 doesn’t hold for some subsets s1 and s2 . That is, we cannot add any element
from s2 −s1 to s1 and keep it independent. Further, wlog, let |s2 | = p+1 and |s1 | = p.
Let us define a weight function on the elements of S as follows

 p + 2 if e ∈ s1
w(e) = p + 1 if e ∈ s2 − s1
0 otherwise


The greedy approach will pick up all elements from s1 and then it won’t be able to
choose any element from s2 − s1 . The greedy solution has weight p + 2|s1| = (p + 2) · p.
By choosing all elements of s2 , the solution has cost (p + 1) · (p + 1) which has a higher
cost than greedy and hence it is a contradiction of Property 1 that is assumed to be
true.
Property 2 implies Property 3 If two maximal subsets of a set A have different
cardinality, it is a violation of Property 2. Since both of these sets are independent,
we should be able augment the set s1 with an element from s2 .
Property 3 implies Property 1 Again we will prove by contradiction. Let e1 e2 . . . ei . . . en
be the edges chosen by the greedy algorithm in decreasing order of their weights. Fur-
ther, let e′1 e′2 . . . e′i . . . e′m be the edges of an optimal solution in decreasing order - (
Is m = n ?). Since the weight of the greedy solution is not optimal, there must a
j ≤ n such that ej < e′j . Let A = {e ∈ S|w(e) ≥ w(e′j )}. The subset {e1 , e2 . . . ej−1 }
is maximal with respect to A (Why ?). All the elements in {e′1 , e′2 . . . e′j } form an
independent subset of A that has greater cardinality. This contradicts Property 3. 2

27
Example 3.3 Half Matching Problem Given a directed graph with non-negative
edge weights, find out the maximum weighted subset of edges such that the in-degree
of any node is at most 1.
The problem defines a subset system S is the set of edges and M is the family
of all subsets of edges such that no two incoming edges share a vertex. Let us verify
Property 2 by considering two subsets Sp and Sp+1 with p and p + 1 edges respectively.
Sp+1 must have at least p + 1 distinct vertices incident on the p + 1 incoming edges
and there must be at least one vertex is not part of Sp ’s vertex set incident to Sp ’s
incoming edges. Clearly, we can add this edge to Sp without affecting independence.

Example 3.4 Weighted Bipartite Matching

Consider a simple graph with a zig-zag. There are two maximal independent sets
(set of edges that do not share an end-point), one with cardinality 2 and the other
having only 1 edge. There Property 3 is violated.

3.2.1 Maximal Spanning Tree

Let us try to verify the exchange property. Let F1 and F2 be two forests such that
F2 has one edge less than F1 . We want to show that for some e ∈ F1 − F2 , F2 ∪ {e}
is a forest. There are two cases
Case 1 F1 has a vertex that is not present in F2 , i.e. one of the end-points of an edge,
say e′ ∈ F1 is not present in F2 . Then e′ cannot induce a cycle in F2 .
Case 2 The set of vertices in F1 is the same3 as F2 . Since there are more edges in F1 ,
the number of connected components in F2 is more than F1 . Recall that if vF , eF , cF
represent the the number of vertices, edges and connected components in a forest F
then

vF − cF = eF (Starting from a tree removal of an edge increases components by 1)

We will show that there is an edge (u′ , v ′ ) in F1 that connects such that u′ and v ′ are
in different connected components of F2 and therefore (u′ , v ′ ) ∪ F2 cannot contain a
cycle.
You can easily argue that there is some component in F1 that is split up into some
components of F2 . If you imagine coloring the vertices of F1 according to the compo-
nents in F2 , at least one edge will have its end-points colored differently.
Exercise 3.4 The matroid theory is about maximizing the total weight of a subset.
How would you extend it to finding minimum weighted subset - for example Minimal
Spanning Trees ?
3
The set of end-points of F1 may be a proper subset of F2 . Even then we can restrict our
arguments to the end-points of F2 - Why ?

28
3.2.2 A Scheduling Problem
We are given a set of jobs J1 , J2 . . . Jn , their corresponding deadlines di for completion
and the corresponding penalties pi if a job completes after deadlines. The jobs have
unit processing time on a single available machine. We want to find out the minimize
the total penalty incurred by the jobs that are not completed before their deadlines.
Stated otherwise, we want to maximize the penalty of the jobs that get completed
before their deadlines.
A set A of jobs is independent if there exists a schedule to complete all jobs in
A without incurring any penalty. We will try to verify Property 2. Let A, B be two
independent sets of jobs with |B| > |A|. We would like to show that for some job
J ∈ B, {J} ∪ A is independent. Let |A| = m < n = |B|. Start with any feasible
schedules for A and B and compress them, i.e. remove any idle time between the
jobs by transforming the schedule where there is no gap between the finish time of
a job and start time of the next. This shifting to left does not affect independence.
Let us denote the (ordered) jobs in A by A1 , A2 . . . Am and the times for scheduling
of jobs in A be d1 , d2 . . . dm respectively. Likewise, let the jobs in B be B1 , B2 . . . Bn
and their scheduling times d′1 , d′2 . . . d′n .
If Bn ∈
/ A, then we can add Bn to A and schedule it as the last job. If Bn = Aj ,
then move Aj to the same position as Bn (this doesn’t violate the deadline) creating
a gap at the j-th position in A. We can now shift-to-left the jobs in A − Aj and now
by ignoring the jobs Bn = Aj , we have one less job in A and B. We can renumber
the jobs and are in a similar position as before. By applying this strategy inductively,
either we succeed in adding a job from B to A without conflict or we are in a situation
where A is empty and B is not so that we can now add without conflict.

3.3 Efficient data structures for MST algorithms

The greedy algorithm described in the previous section is known as Kruskal’s algo-
rithm that was discovered much before the matroid theory was developed. In the
usual implementation of Kruskal’s algorithm, the edges are sorted in increasing order
of their weights.

29
 Algorithm Kruskal MST

input Graph G = V, E) and a weight function on edges.

output A minimum Spanning Tree of G
Let e1 , e2 . . . em be the elements of E in increasing order of weights. Initialize T = Φ.
For i = 1 to m do

In the i-th stage

If T ∪ {ei } doesn’t contain a cycle, then T ← T ∪ {ei }

Output T .

The key to an efficient implementation is the cycle test, i.e., how do we quickly de-
termine if adding an edge induces a cycle in T . We can view Kruskal’s algorithm as
a process that starts with a forest of singleton vertices and gradually connects the
graph by adding edges and growing the trees. The algorithm adds edges that connect
distinct trees (connected components) - an edge whose endpoints are within the same
tree may not be added since it induces cycles. Therefore we can maintain a data
structure that supports the following operations
Find For a vertex, find out which connected component it belongs to.

Union Combine two connected components.

For obvious reasons, such a data structure is called a union-find structure and can be
seen in a more general context where we have a family of subsets and for any given
element we can find the subset it belongs to and we can merge two subsets into one.
The subsets are assumed to be disjont.

3.3.1 A simple data structure for union-find

Let us try to use arrays to represent the sets, viz., in an array A, let A(i) contain
the label of vertex i. The labels are also in the range 1, 2 . . . n. Initially all the labels
are distinct. For each set (label), we also have pointers to all its elements, i.e., the
indices of the array that belong to the set.
Find Is really simple - for vertex i report A(i). This takes O(1) time.
Union To do union(Sj , Sk ), we change the labels of all elements pointed to by j
and link them with k. Thus after union, we have labelled all the elements in
the union as k. The time for this operation is proportional to the number of

30
 elements in set j. For obvious reasons, we would change labels of the smaller
subset.
Although the time for a single union operation can be quite large, in the context of
MST, we will analyze a sequence of union operations - there are at most n − 1 union
operations in Kruskal’s algorithm. Consider a fixed element x ∈ S. The key to the
analysis lies in the answer to the following question.
How many times can the label of x change ?
Every time there is a label change the size of the set containing x increases by a factor
of two (Why ?). Since the size of a set is ≤ n, this implies that the maximum number
of label changes is log n). Kruskal’s algorithm involves |E| finds and at most |V | − 1
unions; from the previous discussion this can be done in O(m + n log n) steps using
the array data-structure.

3.3.2 A faster scheme

The previous data structure gives optimal performance for m ∈ Ω(n log n) so theo-
retically we want to design better schemes for graphs with fewer edges. For this we
will explore faster schemes for union-find.
Instead of arrays, let us use trees4 to represent subsets. Initially all trees are
singleton nodes. The root of each tree is associated with a label (of the subset) and
a rank which denotes the maximum depth of any leaf node. To perform Find x), we
traverse the tree starting from the node x till we reach the root and report its label.
So the cost of a Find operation is the maximum depth of a node.
To perform Union (T1 , T2 ), we make the root of one tree the child of the root of the
other tree. To minimize the depth of a tree, we attach the root of the smaller rank
tree to the root of the larger rank tree. This strategy is known as the union by rank
heuristic. The rank of a tree increases by one only when T1 and T2 have the same
ranks. Once a root node becomes a child of another node, the rank doesn’t change
(by convention). The union operation takes O(1) steps.
Exercise 3.5 Prove that
(i) The number of nodes in tree of rank r is at least 2r if we use the union by rank
heuristic.
(ii) The maximum depth of a tree (or equivalently the value of any rank) is at most
log n.
(iii) There are at most 2nr nodes of rank r.
(iv) The ranks of nodes along any path from a node to the root are increasing mono-
tonically.
4
this tree should not be confused with the MST that we are trying to construct

31
So we are in a situation where Find takes O(log n) and Union operation takes O(1).
Seemingly, we haven’t quite gained anything so let us use the following heuristic.
Path compression
When we do a Find(x) operation, let x0 = root of x, x1 , x2 . . . x be the sequence
of nodes visited. Then we make the subtrees rooted at xi (minus the subtree rooted
at xi+1 ) the children of the root node. Clearly, the motivation is to bring more nodes
closer to the root node, so that the time for the Find operation decreases. And Path
compression does not increase the asymptotic cost of the current Find operation (it
is factor of two).
While it is intuitively clear that it should give us an advantage, we have to rigor-
ously analyze if it indeed leads to any asymptotic improvement.

3.3.3 The slowest growing function ?

Let us look at a very rapidly growing function, namely the tower of two. The tower
i looks like 2) .
2. i

22
which can defined more formally as a function
 1
 2 i=0
B(i) = 22 i=1
 B(i−1)
2 otherwise for i ≥ 2

Let 
(i) n i=0
log n= (i−1)
log(log n) for i ≥ 1
The inverse of B(i) is defined as

log∗ n = min{i ≥ 0| log(i) n ≤ 1}

In other words,
. 2)
2. n

log∗ 22 =n+1
We will use the function B() and log∗ () to analyze the effect of path compression.
We will say that two integers x and y are in the same block if log∗ x = log∗ y.
Although log∗ appears to slower than anything we can imagine, (for example
log∗ 265536 ≤ 5), there is a closely related family of function called the inverse Acker-
man function that is even slower !

32
 Ackerman’s function is defined as
A(1, j) = 2j for j ≥ 1
A(i, 1) = A(i − 1, 2) for i ≥ 2
A(i, j) = A(i − 1, A(i, j − 1)) for i, j ≥ 2

Note that A(2, j) is similar to B(j) defined earlier. The inverse-Ackerman function
is given by
m
α(m, n) = min{i ≥ 1|A(i, ⌊ ⌋) > log n}
n
To get a feel for how slowly it grows, verify that
. 2)
2. 16

α(n, n) = 4 for n = 22

3.3.4 Putting things together

Clearly the cost of Find holds key to the analysis of this Union Find data structure.
Since the rank is less than log n, we already have an upperbound of (log n) for any
individual Find operation. We will adopt the following strategy for counting the cost
of Find operations. We will associate a hypothetical counter with each node that we
will increment whenever it is visited by some Find operation. Finally, by summing
the counts of all the nodes, we can bound the cost of all the find operations. We will
refer to the incremental cost of Find operation as a charge.
We further distinguish between two kinds of charges

Block charge If the block number of the parent node is strictly greater then a node
incurs a block charge. Clearly the maximum number of block charges for a
single Find operation is O(log∗ n)

Path charge Any charge incurred by a Find operation that is not a block charge.

From our previous observation, we will focus on counting the path charges.

Observation 3.1 Once the rank of a node and its parent are in different blocks, they
continue to be in different blocks, i.e. once a node incurs block charges, it will never
incur any more path charges.

The parent of a node may change because of path compression preceded by one or
more union operations, but the new parent will have a rank higher than the previous
parent. Consequently, a node in block j can incur path charges at most B(j) − B(j −

33
1) ≤ B(j) times. Since the number of elements with rank r is at most 2nr , the number
of elements having ranks in block i is
 
n n n 1 1 1 n
+ +. . . = n + + . . . ≤ 2n =
2B(i−1)+1 2B(i−1)+2 2B(i) 2B(i−1)+1 2B(i−1)+2 2B(i−1)+1 2B(i−1)
n
Therefore the total number of path charges for elements in block i is at most 2B(i−1) ·
B(i) which is O(n). For all the log∗ n blocks the cumulative path charges is O(n log∗ n)
to which we have to add O(m log∗ n) block charges.
Remark For some technical reason, in a Find operation, the child of the root node
always incurs a block charge (Why ?)

3.3.5 Path compression only

If we only use path-compression, without the union-by-rank heuristic, we cannot
bound the rank of a node by log n. Since the union-by-rank heuristic does not change
the asymptotic bound of the union operation, it is essentially to gain a better under-
standing of role of the path compression.
Note that the ranks of the nodes from any node the root still increases monoton-
ically - without the union-by-rank heuristic, the rank of node can increase by more
than one (in fact arbitrarily) after a union operation. Let us denote parent of a node
x as p(x), parent of parent of x as p2 (x) and likewise. Let us define the level of a
node x by ℓ(x) as an integer i such that 2i−1 ≤ rank(p(x)) − rank(x) ≤ 2i . Therefore
ℓ(x) ≤ log n.
We account for the cost of a find(x) operation by charging a cost one to all the
nodes in the path from x to the root. The only exception is that for any level
i, 1 ≤ i ≤ log n, the last node (in the path to the root) in level i is not charged.
Instead the cost is charged to the find operation. Clearly the number of charges to
the find operation is O(log n). For any other node y, note that the ℓ(y) increases at
least by one from the monotonicity of ranks and the fact that it is not the last node
of its level. Therefore over the course of all the union-find operations, a node can
get charged at most log n times resulting in a total cost of O(m log n) for all the find
operations.

Exercise 3.6 Can you construct an example to show that the above analysis is tight
?

34
Chapter 4

Optimization II :
Dynamic Programming

Let us try to solve the knapsack problem using a somewhat different strategy. Let
Fi (y) denote the optimal solution for a knapsack capacity y and using only the objects
in {x1 , x2 . . . xi }. Under this notation, Fn (M) is the final solution to the knapsack
problem with n objects and capacity M. Let us further assume that all the weights
are integral as also M. We can write the following equation
Fi (y) = max{Fi−1 (y), Fi−1(y − wi ) + pi }
where the two terms correspond to inclusion or exclusion of object i in the optimal
solution. Also note that, once we decide about the choice of xi , the remaining choices
must be optimal with respect to the remaining objects and the residual capacity of
the knapsack.
We can represent the above solution in a tabular form, where the rows correspond
to the residual capacity from 1 to M and the column i represents the choice of objects
restricted to the subset {1, 2 . . . i}.
The first column corresponds to the base case of the subset containing only object
{x1 } and varying the capacity from 1 to M. Since the weight of the object is w1 ,
for all i < w1 , F1 (i) = 0 and p1 otherwise. From the recurrence, it is clear that the
i-th column can be filled up from the (i − 1)-st column and therefore after having
computed the entries of column 1, we can successively fill up all the columns (till n).
The value of Fn (M) is readily obtained from the last column.
The overall time required to fill up the table is proportional to the size of the
table multiplied by the time to compute each entry. Each entry is a function of two
previously computed terms and therefore the total running time is O(n · M).
Comment The running time (nM) should be examined carefully. M is the capac-
ity of knapsack, for which log M bits are necessary for its representation. For the

35
remaining input data about n objects, let us assume that we need b · n bits where b
is the number of bits per object. This makes the input size N = b · n + log M so if
log M = N/2 then the running time is clearly exponential (M = 2N/2 ).

4.1 A generic dynamic programming formulation

We begin with a recurrence (or an inductive) relation. In a typical recurrence, you
may find repeated subproblems as we unfold the recurrence relation. There is an
interesting property that the dynamic programming problems satisfy. The overall
optimal solution can be described in terms of optimal solution of subproblems. This
is sometimes known as optimal substructure property. This is what enables us to write
an appropriate recurrence for the optimal solution.
Following this, we describe a table that contains the solutions to the various
subproblem. Each entry of the table T must be computable using only the previously
computed entries. This sequencing is very critical to carry the computation forward.
The running time is proportional to
X
t(s) where t(s) is the time time to compute an entry s
s∈T

In the knapsack problem t(s) = O(1). The space bound is proportional to part of table
that must be retained to compute the remaining entries. This is where we can make
substantial savings by sequencing the computation cleverly. Dynamic programming
is often seen as a trade-off between space and running time, where we are reducing the
running time at the expense of extra space. By storing the solutions of the repeated
subproblems, we save the time for recomputation. For the knapsack problem, we only
need to store the previous column - so instead of M · n space, we can do with O(n)
space.

4.2 Illustrative examples

4.2.1 Context Free Parsing
Given a context free grammar G in a Chomsky Normal Form (CNF) and a string
X = x1 x2 . . . xn over some alphabet Σ, we want to determine if X can be derived
from the grammar G.
A grammar in CNF has the following production rules

A → BC A → a

36
where A, B, C are non-terminals and a is a terminal (symbol of the alphabet). All
derivations must start from a special non-terminal S which is the start symbol. We
∗
will use the notation S ⇒ α to denote that S can derive the sentence α in finite
number of steps by applying production rules of the grammar.
The basis of our algorithm is the following observation
∗ ∗
Observation 4.1 A ⇒ xi xi+1 . . . xk iff A ⇒ BC and there exists a i < j < k such
∗ ∗
that B ⇒ xi xi+1 . . . xj and C ⇒ xj+1 . . . xk .
There are k − i possible partitions of the string and we must check for all partitions
if the above condition is satisfied. More generally, for the given string x1 x2 . . . xn , we
consider all substrings Xi,k = xi xi+1 . . . xk where 1 ≤ i < k ≤ n - there are O(n2 ) such
substrings. For each substring, we try to determine the set of non-terminals A that
can derive this substring. To determine this, we use the the previous observation.
Note that both B and C derive substrings that are strictly smaller than Xi,j . For
substrings of length one, it is easy to check which non-terminals derive them, so these
serve as base cases.
We define a two dimensional table T such that the entry T (s, t) corresponds to all
non-terminals that derive the substring starting at xs of length t. For a fixed t, the
possible values of s are 1, 2, . . . n − t + 1 which makes the table triangular. Each entry
in the table
P can be filled up in O(t) time for column t. That yields a total running
time of nt=1 O((n − t) · t) which is O(n3 ). The space required is the size of the table
which is O(n2 ). This algorithm is known as CYK (Cocke-Young-Kassimi) after the
inventors.

4.2.2 Longest monotonic subsequence

Problem Given a sequence S of numbers x1 , x2 . . . xn a subsequence xi1 , xi2 . . . xik
where ij+1 > ij is monotonic if xij+1 ≥ xij . We want to find the longest (there may
be more than one) monotonic subsequence.
Exercise 4.1 For any sequence of length n prove that either the longest increasing
√
monotonic subsequence or the longest decreasing subsequence has length at least ⌈ n⌉.
This is known as the Erdos-Szekeres theorem.
The previous result is only an existential result but here we would like to find the
actual sequence. Let us define the longest monotonic subsequence in x1 , x2 . . . xi
ending at xi (i.e. xi must be the last element) as Si . Clearly we are looking for
maxni=1 Si 1 . With a little thought we can write a recurrence for Si as
Si = max{Sj + 1|xj ≤ xi }
j<i

1
We are using Si for both the length and the sequence interchangeably.

37
Computing Si takes O(i) time and therefore, we can compute the longest monotonic
subsequence in O(n2 ) steps. The space required is O(n).
Can we improve the running time ? For this, we will actually address a more gen-
eral problem, namely for each i, we will compute the longest monotonic subsequence
of length i (if it exists). For each i ≤ n, let Mi,j j ≤ i be the longest monotonic
subsequence of length j in x1 x2 . . . xi . Clearly, if Mi,j exists then Mi,j−1 exists and
the maximum length subsequence is

max{Mn,j |Mn,j exists}

j

Among all subsequences of length i, we will compute an Mi,j which has the minimum
value of the last element. For example, among the subsequences 2,4,5,9 and 1,4,5,8
(both length 4), we will choose the second one, since 8 < 9.
Let ℓi,j be the last element of Mi,j - by convention we will implicitly initialise all
ℓi,j = ∞. We can write a recurrence for ℓi,j as follows

ℓi,j−1 if ℓi,j−1 ≤ xi+1 < ℓi,j
ℓi+1,j =
ℓi,j otherwise
This follows since, Mi+1,j is either Mi,j or xi+1 must be the last element of Mi+1,j .
If we implement this idea in a straightforward fashion, the running time will be Ω(n2 ),
since there are potentially n subsequences and all these may have to be updated for
every 1 ≤ i ≤ n. To improve it further, we make an important additional observation
Observation 4.2 The ℓi,j ’s form a non-decreasing sequence in j for all i.

Exercise 4.2 Prove it rigorously.

This paves way for updating and maintaining the information about ℓi,j in a compact
manner. We will maintain a sorted sequence of ℓi,j such that when we scan xi+1 ,
we can quickly identify ℓi,k such that ℓi,k−1 ≤ xi+1 < ℓi,k . Moreover, Si , the longest
monotonic subsequence ending at xi+1 is of length k. We can reconstruct Si by
keeping track of the predecessor. We can easily maintain the ℓi,j ’s in a dynamic
dictionary data structure (like AVL tree) O(log n) time. Therefore, the total running
time reduces to O(n log n).

4.2.3 Function approximation

Consider an integer valued function h(i) on integers {1, 2 . . . n}. We want to define
another function g(i) with a maximum of k steps k ≤ n such that the difference
between g(i) and h(i) is minimized according to some measure. One of the most
common measures is the sum of the squares of the differences of the two functions.

38
 Let gj∗ (i) denote the optimal j-step function for this problem restricted to the
points 1, . . . i - we are interested in computing gk∗(n).

Exercise 4.3 Show that g1∗ (j) = n1 ji=1 h(i), i.e., it is a constant function equal to
P
the mean.

We can now write a recurrence for the gi∗ (ℓ) as follows

gi∗ (ℓ) = min gi−1

∗
(j) + Dj,ℓ
j<ℓ

where Dj,ℓ denotes the sum of squares of deviation of h() from the mean value of h
in the interval [j, ℓ].
The recurrence captures the property that an optimal k step approximation can be
expressed as an optimal k − 1 step approximation till an intermediate point followed
by the best 1 step approximation of the remaining interval (which is the mean value
in this interval from our previous observation). Assuming that Dj,ℓ are precomputed
for all 1 ≤ j < ℓ ≤ n, we can compute the gi∗ (j) for all 1 ≤ i ≤ k and 1 ≤ j ≤ n in a
table of size kn. The entries can be computed in increasing order of i and thereafter
in increasing order of j’s. The base case of i = 1 can be computed directly from the
result of the previous exercise. Each entry can be computed from j − 1 previously
computed entries yielding a total time of
i=k X
X n
O(j) = O(k · n2 )
i=1 j=1

The space required is proportional to the previous column (i.e. we need to keep track
of the previous value of i), given that D(j, ℓ) can be stored/computed quickly.

Exercise 4.4 Complete the analysis of the above algorithm considering the compu-
tation of the D(i, j).

4.2.4 Viterbi’s algorithm for Expectation Maximization

In this problem we have a weighted directed graph G = (V, E) where the weights are
actually probabilities and the sum of the probabilities on outgoing edges from any
given vertex is 1. Further, the edges are labelled with symbols from an alphabet Σ -
note that more than one edge can share the same label. Given a string σ = x1 x2 . . . xn
over Σ, find the most probable path in the graph starting at vo with label equal to σ.
The label of a path is the concatenation of labels associated with the edges. To find
the most probable path, we can actually find the path that achieves the maximum
weight with label σ. Here the weights are logarithms of the probability - by taking

39
log, we have changed the product of the probabilities to sum of log (note that log of
probabilities are negative numbers).
We can write a recurrence based on the following observation.
The optimal least-weight path x1 , x2 . . . xn starting at vertex x1 with label σ1 σ2 . . . σn
is such that the path x2 x3 . . . xn is optimal with respect to the label σ2 , σ3 . . . σn . For
paths of lengths one, it is easy to find the optimal labelled path. Let Pi,j (v) denote
the optimal labelled path for the labels σi σi+1 . . . σj starting at vertex v. We are
interested in P1,n (vo ).

Pi,j (v) = min{Pi+1,j (w)|label of (v, w) = σi }

w

Starting from the base case of length one paths, we build length 2 paths from each
vertex and so on. Note that the length i + 1 paths from a vertex v can be built from
length i paths from w (computed for all vertices w ∈ V ). The paths that we compute
are of the form Pi,n for all 1 ≤ i ≤ n. Therefore we can compute the entries of the
table starting from i = n − 1. From the previous recurrence, we can now compute
the entries of the Pn−2,n etc. by comparing at most |V | entries (more specifically the
outdegree ) for each starting vertex v 2 . Given that the size of table is n · |V |, the total
time required to compute all the entries is O(n·|V |2 ). However, the space requirement
can be reduced to O(|V |) from the observation that only the (i − 1) length paths are
required to compute the optimal i length paths.

2
You can argue that each iteration takes O(|E|) steps where |E| is the number of edges.

40
Chapter 5

Searching

5.1 Skip Lists - a simple dictionary

Skip-list is a data structure introduced by Pugh1 as an alternative to balanced bi-
nary search trees for handling dictionary operations on ordered lists. The underlying
idea is to substitute complex book-keeping information used for maintaining balance
conditions for binary trees by random sampling techniques. It has been shown that,
given access to random bits, the expected search time in a skip-list of n elements is
O(log n) 2 which compares very favourably with balanced binary trees. Moreover, the
procedures for insertion and deletion are very simple which makes this data-structure
a very attractive alternative to the balanced binary trees.
Since the search time is a stochastic variable (because of the use of randomization),
it is of considerable interest to determine the bounds on the tails of its distribution.
Often, it is crucial to know the behavior for any individual access rather than a chain
of operations since it is more closely related to the real-time response.

5.1.1 Construction of Skip-lists

This data-structure is maintained as a hierarchy of sorted linked-lists. The bottom-
most level is the entire set of keys S. We denote the linked list at level i from
the bottom as Li and let |Li | = Ni . By definition L0 = S and |L0 | = n. For all
0 ≤ i, Li ⊂ Li−1 and the topmost level, say level k has constant number of elements.
Moreover, correspondences are maintained between common elements of lists Li and
Li−1 . For a key with value E, for each level i, we denote by Ti a tuple (li , ri ) such
that li ≤ E ≤ ri and li , ri ∈ Li . We call this tuple straddling pair (of E) in level i.
1
William Pugh. Skip list a probablistic alternative to balanced trees. CACM June 1990 Vol33
Num6 668-676, 1990
2
Note that all logarithms are to base 2 unless otherwise mentioned.

41
 The search begins from the topmost level Lk where Tk can be determined in
constant time. If lk = E or rk = T E then the search is successful else we recursively
search among the elements [lk , rk ] L0 . Here [lk , rk ] denotes the closed interval bound
by lk and rk . This is done by searching the elements of Lk−1 which are bounded by
lk and rk . Since both lk , rk ∈ Lk−1 , the descendence from level k to k − 1 is easily
achieved in O(1) time. In general, at any level i we determine the tuple Ti by walking
through a portion of the list Li . If li or ri equals E then we are done else we repeat
this procedure by descending to level i − 1.
In other words, we refine the search progressively until we find an element in S
equal to E or we terminate when we have determined (l0 , r0 ). This procedure can
also be viewed as searching in a tree that has variable degree (not necessarily two as
in binary tree).
Of course, to be able to analyze this algorithm, one has to specify how the lists Li
are constructed and how they are dynamically maintained under deletions and addi-
tions. Very roughly, the idea is to have elements in i-th level point to approximately
2i nodes ahead (in S) so that the number of levels T is approximately O(log n). The
time spent at each level i depends on [li+1 , ri+1 ] Li and hence the objective is to
keep this small. To achieve these conditions on-line, we use the following intuitive
method. The nodes from the bottom-most layer (level 0) are chosen with probability
p (for the purpose of our discussion we shall assume p = 0.5) to be in the first level.
Subsequently at any level i, the nodes of level i are chosen to be in level i + 1 inde-
pendently with probability p and at any level we maintain a simple linked list where
the elements are in sorted order. If p = 0.5, then it is not difficult to verify that for a
list of size n, the expected number of elements in level i is approximately n/2i and are
spaced about 2i elements apart. The expected number of levels is clearly O(log n),
(when we have just a trivial length list) and the expected space requirement is O(n).
To insert an element, we first locate its position using the search strategy described
previously. Note that a byproduct of the search algorithm are all the Ti ’s. At level
0, we choose it with probability p to be in level L1 . If it is selected, we insert it in
the proper position (which can be trivially done from the knowledge of T1 ), update
the pointers and repeat this process from the present level. Deletion is very similar
and it can be readily verified that deletion and insertion have the same asymptotic
run time as the search operation. So we shall focus on this operation.

5.1.2 Analysis
To analyze the run-time of the search procedure, we look at it backwards, i.e., retrace
the path from level 0. The search time is clearly the length of the path (number of
links) traversed over all the levels. So one can count the number of links one traverses
before climbing up a level. In other words the expected search time can be expressed

42
in the following recurrence

C(k) = (1 − p)(1 + C(k)) + p(1 + C(k − 1))

where C(k) is the expected cost for climbing k levels. From the boundary condition
C(0) = 0, one readily obtains C(k) = k/p. For k = O(log n), this is O(log n). The
recurrence captures the crux of the method in the following manner. At any node of
a given level, we climb up if this node has been chosen to be in the next level or else
we add one to the cost of the present level. The probability of this event (climbing up
a level) is p which we consider to be a success event. Now the entire search procedure
can be viewed in the following alternate manner. We are tossing a coin which turns up
heads with probability p - how many times should we toss to come up with O(log n)
heads ? Each head corresponds to the event of climbing up one level in the data
structure and the total number of tosses is the cost of the search algorithm. We are
done when we have climbed up O(log n) levels (there is some technicality about the
number of levels being O(log n) but that will be addressed later). The number of
heads obtained by tossing a coin N times is given by a Binomial random variable X
with parameters N and p. Using Chernoff bounds (see Appendix, equation B.1.5),
for N = 15 log n and p = 0.5, Pr[X ≤ 1.5 log n] ≤ 1/n2 (using ǫ = 9/10 in equation
1). Using appropriate constants, we can get rapidly decreasing probabilities of the
form Pr[X ≤ c log n] ≤ 1/nα for c, α > 0 and α increases with c. These constants can
be fine tuned although we shall not bother with such an exercise here.
We thus state the following lemma.

Lemma 5.1 The probability that access time for a fixed element in a skip-list data
structure of length n exceeds c log n steps is less than O(1/n2 ) for an appropriate
constant c > 1.

Proof We compute the probability of obtaining fewer than k (the number of levels
in the data-structure) heads when we toss a fair coin (p = 1/2) c log n times for some
fixed constant c > 1. That is, we compute the probability that our search procedure
exceeds c log n steps. Recall that each head is equivalent to climbing up one level
and we are done when we have climbed k levels. To bound the number of levels, it
is easy to see that the probability that any element of S appears in level i is at most
1/2i , i.e. it has turned up i consecutive heads. So the probability that any fixed
element appears in level 3 log n is at most 1/n3 . The probability that k > 3 log n is
the probability that at least one element of S appears in L3 log n . This is clearly at
most n times the probability that any fixed element survives and hence probability
of k exceeding 3 log n is less than 1/n2 .
Given that k ≤ 3 log n we choose a value of c, say c0 (to be plugged into equation
B.1.6 of Chernoff bounds) such that the probability of obtaining fewer than 3 log n

43
heads in c0 log n tosses is less than 1/n2 . The search algorithm for a fixed key exceeds
c0 log n steps if one of the above events fail; either the number of levels exceeds 3 log n
or we get fewer than 3 log n heads from c0 log n tosses. This is clearly the summation
of the failure probabilities of the individual events which is O(1/n2). 2.
Theorem 5.1 The probability that the access time for any arbitrary element in skip-
list exceeds O(log n) is less than 1/nα for any fixed α > 0.
Proof: A list of n elements induces n + 1 intervals. From the previous lemma,
the probability P that the search time for a fixed element exceeding c log n is less
than 1/n2 . Note that all elements in a fixed interval [l0 , r0 ] follow the same path in
the data-structure. It follows that for any interval the probability of the access time
exceeding O(log n) is n times P . As mentioned before, the constants can be chosen
appropriately to achieve this. 2
It is possible to obtain even tighter bounds on the space requirement for a skip
list of n elements. We can show that the expected space is O(n) since the expected
number of times a node survives is 2.
Exercise 5.1 Prove the following stronger bound on space using Chernoff bounds -
For any constant α > 0, the probability of the space exceeding 2n + α · n, is less than
2
expΩ(−α n) .

5.2 Treaps : Randomized Search Trees

The class of binary (dynamic) search trees is perhaps the first introduction to non-
trivial data-structure in computer science. However, the update operations, although
asymptotically very fast are not the easiest to remember. The rules for rotations
and the double-rotations of the AVL trees, the splitting/joining in B-trees and the
color-changes of red-black trees are often complex, as well as their correctness proofs.
The Randomized Search Trees (also known as randomized treaps) provide a practical
alternative to the Balanced BST. We still rely on rotations, but no explicit balancing
rules are used. Instead we rely on the magical properties of random numbers.
The Randomized Search Tree (RST) is a binary tree that has the keys in an in-
order ordering. In addition, each element is assigned a priority (Wlog, the priorities
are unique) and the nodes of the tree are heap-ordered based on the priorities. Si-
multaneously, the key values follow in-order numbering. It is not difficult to see that
for a given assignment of priorities, there is exactly one tree. If the priorities are
assigned randomly in the range [1, N] for N nodes, the expected height of the tree is
small. This is the crux of the following analysis of the performance of the RSTs.
Let us first look at the search time using a technique known as backward analysis.
For that we (hypothetically) insert the N elements in a decreasing order of their

44
priorities and then count the number of elements that Q can see during the course
of their insertions. This method (of assigning the random numbers on-line) makes
arguments easier and the reader must convince himself that it doesn’t affect the final
results. Q can see an element Ni if there are no previously inserted elements in
between.

Lemma 5.2 The tree constructed by inserting the nodes in order of their priorities
(highest priority is the root) is the same as the tree constructed on-line.

Lemma 5.3 The number of nodes Q sees is exactly the number of comparisons per-
formed for searching Q. In fact, the order in which it sees corresponds to the search
path of Q.

Theorem 5.2 The expected length of search path in RST is O(HN ) where HN is the
N-th harmonic number.

In the spirit of backward analysis, we pretend that the tree-construction is being

reversed, i.e. nodes are being deleted starting from the last node. In the forward
direction, we would count the expected number of nodes that Q sees. In the reverse
direction, it is the number of times Q’s visibility changes (convince yourself that
these notions are identical). Let Xi be a Bernoulli rv that is 1 if Q sees Ni (in
the forward direction) or conversely Q’s visibility changes when Ni is deleted in the
reverse sequence. Let X be the length of the search path.
X
X= Xi
X X
E[X] = E[ Xi ] = E[Xi ]
We claim that E[Xi ] = 2i . Note that the expectation of a Bernoulli variable is the
probability that it is 1. We are computing this probability over all permutations of
N being equally likely. In other words, if we consider a prefix of length i, all subsets
of size i are equally likely. Let us find the probability that Xi = 1, conditioned on a
fixed subset N i ⊂ N consisting of i elements. Unconditioning is easy if the probability
that Xi = 1 does not depend on N i itself. Since, for a fixed N i , all N i−1 are equally
likely, the probability that Xi = 1 is the probability that one of the (maximum two)
neighboring elements was removed in the reverse direction. The probability of that is
less than 2i which is independent of any specific N i . So, the unconditional probability
is
Pthe same asPconditional probability - hence E[Xi ] = 2i . The theorem follows as
1
i E[Xi ] = 2 i i = O(HN ).
The Xi ’s defined in the previous proof are independent but not identical. We can
still apply Chernoff-Hoeffding bounds to obtain strong tail-estimates for deviation
from the expected bound.

45
Theorem 5.3 The probability that the search time exceeds 2 log n comparisons in a
randomized triep is less than O(1/n).

A similar technique can be used for counting the number of rotations required for
RST during insertion and deletions. Backward analysis is a very elegant technique
for analyzing randomized algorithms, in particular in a paradigm called randomized
incremental construction.

5.3 Universal Hashing

Hashing is often used as a technique to achieve O(1) search time by fixing the location
where a key is assigned. For the simple reason that the number of possible key values
is much larger than the table size, it is inevitable that more than one key is mapped
to the same location. The number of conflicts increase the search time. If the keys
are randomly chosen, then it is known that the expected number of conflicts is O(1).
However this may be an unrealistic assumption, so we must design a scheme to handle
any arbitrary subset of keys. We begin with some useful notations

• Universe : U, Let the elements be 0, 1, 2, . . . N − 1

• Set of elements : S also |S| = n

• Hash locations : {0, 1, ..., m − 1} usually, n ≥ m

Collision If x, y ∈ U are mapped to the same location by a hash function h.


1 : h(x) = h(y), x 6= y
δh (x, y) =
0 : otherwise
X
δh (x, S) = δh (x, y)
y∈S

Hash by chaining: The more the collision the worse the performance. Look at a
sequence O1 (x2 ), O2 (x2 ), ..., On (xn ) where Oi ∈ {Insert, Delete, Search} and xi ∈ U
Let us make the following assumptions

1. |h−1 (i)| = |h−1 (i′ )| where i, i′ ∈ {0, 1, ..., m − 1}

2. In the sequence, xi can be any element of U with equal probability.

46
 n
Claim: Total expected cost = O((1 + β)n) where β = m (load factor).
Proof: Expected cost of (k + 1)th operation = expected number of elements in
location ≤ 1 + k( m1 ) assuming all the previous operations were Insert.
So total expected cost ≤ nk=1 1 + m k
= n + n(n+1) = (1 + β2 )n. This is worst case
P
2m
over operations but not over elements. 2

Universal Hash Functions

Definition 5.1 A collection H ⊂ {h|h : [0...N − 1] → [0...m − 1]} is c-universal if

for all x, y ∈ [0...N − 1] x 6= y,

|H|
|{h|h ∈ H and h(x) = h(y)}| ≤ c
m

≤ c |H|
P
for some small constant c. Roughly h δh (x, y) m

Claim:
1 X n
1 + δh (x, S) ≤ 1 + c
|H| h∈H m

where |S| = n.
Proof: Working from the LHS, we obtain
1 X 1 XX
= 1+ δh (x, y)
|H| h∈H |H| h y∈S
1 XX
= 1+ δh (x, y)
|H| y
h
1 X |H|
≤ 1+ c
|H| y m
c
= 1+ n
m

ci
P
So expected cost of n operation = (1 + m
) ≤ (1 + cβ)n 2

5.3.1 Example of a Universal Hash function

H ′ : ha,b ; hab (x) → ((ax + b) mod N) mod m where a, b ∈ 0...N − 1 (N is prime).

47
 If hab (x) = hab (y) then for some q ∈ [0...m − 1]and r, s ∈ [0... Nm−1 ]

ax + b = (q + rm) mod N
ay + b = (q + sm) mod N
2
This is a unique solution for a, b once q, r, s are fixed. So there are a total of m( Nm )
2
solutions = Nm . Also, since |H ′ | = N 2 , therefore H ′ is ”1” universal.

5.4 Perfect Hash function

Universal hashing is very useful method but may not be acceptable in a situation,
where we don’t want any conflicts. Open addressing is method that achieves this at
the expense of increased search time. In case of conflicts, we define a sequence of
probes that is guaranteed to find an empty location (if there exists one).
We will extend the scheme of universal hashing to one where there is no collision
without increasing the expected search time. Recall that the probability that an
element x collides with another element y is less than mc for some constant c. There-
fore, the
expected number of collisions in a subset of size n by considering all pairs
is f = n2 · mc . By Markov inequality, the probability that the number of collisions
exceeds 2f is less than 1/2. For c = 2 and m ≥ 4n2 , the value of 2f is less than 21 ,
i.e. there are no collisions.
We use a two level hashing scheme. In the first level, we hash it to locations
1, 2 . . . m. If there are ni keys that get mapped to location i, we subsequently map
them to 4n2i locations. From our previous discussion, we know that we can avoid
collisions with probablity at least 1/2. So we may have to repeat the second level
hashing a number of times (expected value is 2) before we achieve zero collision for
the ni keys. So the searchPtime is O(1) total expected for both levels.
The space bound is 4 i n2i . We can write
X
n2i = 2 δ(x, y) + ni .
x,y|h(x)=h(y)=i

X X X
n2i = 2 (δ(x, y) + ni )
i i x,y|h(x)=h(y)=i
X X X
= 2 ni + 2 δ(x, y)
i i x,y|h(x)=h(y)=i
X
= 2n + δ(x, y)
x,y|h(x)=h(y)

48
Taking expectation
P on both sides (with respect to choice of a random hash function),
the R.H.S. is 2E[ x,y∈S δ(x, y)]+n. This equals 2 n2 · m2 since E[δ(x, y)] = Pr[h(x) =

h(y)] ≤ mc . Therefore the total expected space required is only O(n).

5.4.1 Converting expected bound to worst case bound

We can convert the expected space bound to worst case space bound inPthe following
manner. In the first level, we repeatedly choose a hash function until i n2i is O(n).
We need to repeat this twice in the expected sense. Subsequently at the second stage,
for each i, we repeat it till there are no collisions in mapping ni elements in O(n2i )
locations. Again, the expected number of trials for each i is two that takes overall
O(n) time for n keys. Note that this method makes the space worst case O(n) at the
expense of making the time expected O(n). But once the hash table is created, for
any future query the time is worst case O(1).
For practical implementation, the n keys will be stored in a single array of size
O(n) where the first level table locations will contain the starting positions of keys
with value i and the hash function used in level 2 hash table.

5.5 A log log N priority queue

Searching in bounded universe is faster by use of hashing. Can we achieve similar
improvements for other data structures ? Here we consider maintaining a priority
queue for elements drawn from universe U and let |U| = N. The operations supported
are insert , minimum and delete.
Imagine a complete binary tree on N leaf nodes that correspond to the N integers
of the universe - this tree has depth log N. Let us colour the leaf nodes of the tree
black if the corresponding integer is present in the set S ⊂ U where |S| = n. Let us
also imagine that if a leaf node is coloured then its half-ancestor (halfway from the
node to the root) is also coloured and is labelled with the smallest and the largest
integer in its subtree. Denote the set of the minimum elements by T OP and we
recursively build a data structure on the elements of T OP which has size at most
√
N . We will denote the immediate predecessor of an element x by P RED(x) and
the successor of an element by SUCC(x). The reason we are interested in P RED
and SUCC is that when the smallest element is deleted, we must find its immediate
successor in set S. Likewise, when we insert an element, we must know its immediate
predecessor. Henceforth we will focus on the operations P RED and SUCC as these
will be used to support the priority queue operations.
For a given element x, we will check if its ancestor at depth log N/2 (halfway
up the tree) is colored. If so, then we search P RED(x) within the subtree of size

49
√
N . Otherwise, we search for P RED(x) among the elements of T OP . Note that
either we search within the subtree or in the set T OP but not both. Suitable termi-
nating conditions can be defined. The search time can be captured by the following
recurrence √
T (N) = T ( N) + O(1)
which yields T (N) = O(log log N). The T OP data structure is built on keys of length
log N/2 higher order bits and the search structure is built on the lower log N/2 bits.
The space complexity of the data structure satisfies the recurrence
√ √ √
S(m) = ( m + 1)S( m) + O( m)

which yields S(N) = O(N log log N). The additive term is the storage of the elements
of T OP .

Exercise 5.2 Propose a method to decrease the space bound to O(N) - you may want
to prune the lower levels of the tree.

For the actual implementation, the tree structure is not explicitly built - only the
relevant locations are stored. To insert an element x in this data-structure, we first
find its predecessor and successor. If it is the first element in the subtree then the
ancestor at level log N/2 is appropriately initialized. SUCC(P RED(x)) ← x where
P RED(x) is in a different subtree. Notice that we do not have to process within the
subtree any further. Otherwise, we find its P RED within the subtree. It may appear
that the time for insertion is (log log2 N) since it satisfies the recurrence
√
I(N) = I( N ) + O(log log N) or I(N) ∈ log log2 N

To match the O(log log N) time bound for searching, we will actually do the search
for P RED simultaneously with the insert operation.

Exercise 5.3 Show how to implement delete operation in O(log log N) steps.

One of the most serious drawbacks of this data structure is the high space require-
ment, which is proportional to the size of the universe. Note that, as long as
log log N ∈ o(log n), this is faster than the conventional heap. For example, when
log n/ log log n
N ≤ 22 , this holds an advantage, but the space is exponential. However, we
can reduce the space to O(n) by using the techniques mentioned in this chapter to
store only those nodes that are coloured black.

50
Chapter 6

Multidimensional Searching and

Geometric algorithms

Searching in a dictionary is one of the most primitive kind of search problem and it is
relatively simple because of the property that the elements can be ordered. Instead,
suppose that the points are from d dimensional space Rd - for searching we can build
a data-structure based on lexicographic ordering. If we denote a d dimensional point
by (x0 , x1 . . . xd ) then show the following.

Exercise 6.1 The immediate predecessor of a query point can be determined in O(d+
log n) comparisons in a set of n d-dimensional points.

Queries can be far more sophisticated than just point-queries but we will address the
more restricted kind.

6.1 Interval Trees and Range Trees

One Dimensional Range Searching
Given a a set S of n points on a line (wlog, say the x-axis), we have to build a
data-structure to report the points inside a given query interval interval [xℓ : xu ].
The counting version of range query only reports the number of points in this interval
instead of the points themselves.
Let S = {p1 , p2 , . . . , pn } be the given set of points on the real line. We can solve
the one- dimensional range searching problem using a balanced binary search tree T
in a straightforward manner. The leaves of T store the points of S and the internal
nodes of T store splitters to guide the search. If the splitter-value at node v is xv ,
the left subtree L(v) of a node v contains all the points smaller than or equal to xv

51
and right subtree R(v) contains all the points strictly greater than xv . It is easy to
see that we can build a balanced tree using median as the splitter value.
To report the points in the range query [xℓ : xu ] we search with xℓ and xu in the
tree T . Let ℓ1 and ℓ2 be the leaves where the searches end. Then the points in the
interval [xℓ : xu ] are the points stored between the leaves ℓ1 and ℓ2 . Another way to
view the set of points is the union of the leaves of some subtrees of T .

Exercise 6.2 Show that for any query interval, the points belong to at most 2 log n
subtrees.

Proof Sketch If you examine the search path of xℓ and xu , they share a common
path from root to some vertex (may be the root itself), where the paths fork to the
left and right - let us call this the fork node. The leaf nodes correspond to the union
of the right subtrees of the left path and the left subtrees of the right path.
Complexity The tree takes O(n) space and O(n log n) time in preprocessing. Each
query takes O(log n + k) time, where k is the number of points in the interval, i.e.,
the output size. The counting query takes O(log n) time. This is clearly the best we
can hope for.
You can hypothetically associate a half-open interval with each internal node x,
(l(x) : r(x)] where l(x) (r(x)) is the value of the leftmost (rightmost) leaf node in the
subtree rooted at x.

Exercise 6.3 Show how to use threading to solve the range query in a BST without
having leaf based storage.

6.1.1 Two Dimensional Range Queries

Each point has two attributes: its x coordinate and its y coordinate - the two dimen-
sional range query is a Cartesian product of two one-dimensional intervals. Given
a query [xℓ : xu ] × [yℓ : yu ] (a two dimensional rectangle), we want to build a data
structure to report the points inside the rectangular region (or alternately, count the
points in the region.)
We extend the previous one dimensional solution by first considering the vertical
slab [xℓ : xu ] 1 . Let us build the one-dimensional range tree identical to the previous
scheme (by ignoring the y coordinates of points). Therefore we can obtain the answer
to the slab-query. As we had observed every internal node represents an interval in
the one dimensional case and analogously the corresponding vertical slab in the two
dimensional case. The answer to the original query [xℓ : xu ] × [yℓ : yu ] is a subset of
[xℓ : xu ] × [−∞ : +∞]. Since our objective is to obtain a time bound proportional
1
You can think about the [yℓ : yu ] as [−∞ : +∞]

52
to the final output, we cannot afford to list out all the points of the vertical slab.
However, if we had the one dimensional data structure available for this slab, we can
quickly find out the final points by doing a range query with [yℓ : yu ]. A naive scheme
will build the data structure for all possible vertical slabs but we can do much better
using the following observation.

Observation 6.1 Each vertical slab is a union of 2log n canonical slabs.

It follows from the same argument as any interval. Each canonical slab corresponds
to the vertical slab (the corresponding [xℓ : xu ]) spanned by an internal node. We
can therefore build a one-dimensional range tree for all the points spanned by corre-
sponding vertical slab - this time in the y-direction. So the final answer to the two
dimensional range query is P the union of at most 2log n one-dimensional range query,
giving a total query time of Pti=1 O(log n+ki ) where ki is the number of output points
in slab i among t slabs and i ki = k. This results in a query time of O(t log n + k)
where t is bounded by 2log n.
The space is bounded by O(n log n) since in a given level of the tree T , a point is
stored exactly once.
The natural extension of this scheme leads us to d-dimensional range search trees
with the following performance parameters.

Q(d) ≤ 2 log n·Q(d−1) Q(1) = O(log n+k) Q(d) is the query time in d dimensions for n points

which yields Q(d) = O(2d · logd n) 2 .

Exercise 6.4 What is the space bound for d dimensional range trees ?

Exercise 6.5 How would you modify the data structure for counting range queries ?

6.2 k-d trees

A serious drawback of range trees is that both the space and the query time increases
exponentially with dimensions. Even for two dimensions, the space is super-linear.
For many applications, we cannot afford to have such a large blow-up in space (for a
million records log n = 20).
We do a divide-and-conquer on the set of points - we partition the space into
regions that contain a subset of the given set of points. The input rectangle is tested
against all the regions of the partition. If it doesn’t intersect a region U then we
do not search further. If U is completely contained within the rectangle then we
2
Strictly speaking Q() is a variable of two parameters n and d

53
report all the points associated with U otherwise we search recursively in U. We may
have to search in more than one region - we define a search tree where each region is
associated with a node of the tree. The leaf nodes correspond to the original point
set. In general, this strategy will work for other (than rectangular) kinds of regions
also.
For the rectangular query, we split on x-coordinate and next on y-coordinate,
then alternately on each coordinate. We partition with a vertical line at nodes whose
depth is even and we split with a horizontal line at nodes whose depth is odd. The
time to build the 2-D tree is as follows.
The region R(v) corresponding to a node v is a rectangle which is bounded by
horizontal and vertical lines and it is a subset of the parent node. The root of a tree
is associated with a (bounded) rectangle that contains all the n points. We search
a subtree rooted at v iff the query rectangle intersects the associated with node v.
This involves testing if two rectangles (the query rectangle and R(v)) overlap that
can be done in O(1) time. We traverse the 2-D tree, but visit only nodes whose
region is intersected by the query rectangle. When a region is fully contained in the
query rectangle, we can report all the points stored in its subtree. When the traversal
reaches a leaf, we have to check whether the point stored at the leaf is contained in
the query region and, if so, report it.

Search(Q, v)

If Q ⊂ R(v), then report all points in R(v)

Else

Let R(u) and R(w) be rectangles associated with the children u, w.

If Q ∩ R(u) is non-empty Search(Q, u)
If R ∩ R(w) is non-empty, Search (Q, w )

Since a point is stored exactly once and the description of a region corresponding to
a node takes O(1) space, the total space taken up the search tree is O(n).
Query Time Let Q(i) be the number of nodes at distance i from the root that
are visited in the worst case by a rectangular query. Since a vertical segment of Q
intersects only horizontal partitioning edges, we can write a recurrence for Q(i) by
observing that the number of nodes can increase by a factor 2 by descending 2 levels.
Hence Q(i) satisfies the recurrence
Q(i + 2) ≤ 2Q(i)
which one can √ verify to be Q(i) ∈ O(2⌊i/2⌋ ) or total number of nodes visited in the
last level is O( n).

54
6.3 Priority Search Trees
The combination of BST with heap property resulted in a simple strategy for main-
taining balanced search trees called treaps. The heap property was useful to keep a
check on the expected height of the tree within O(log n). What if we want to main-
tain a heap explicitly on set of parameters (say the y coordinates) along with a total
ordering required for binary search on the x coordinates. Such a data structure would
be useful to support a three sided range query in linear space.
A three sided query is a rectangle [xℓ : xu ] × yℓ : ∞], i.e. a half-infinite vertical slab.
If we had a data structure that is a BST on x coordinates, we can first locate the
two points xℓ and xu to determine (at most) 2log n subtrees whose union contains
the points in the interval [xℓ : xu ]. Say, these are T1 , T2 . . . Tk . Within each such tree
Ti , we want to find the points whose y coordinates are larger than yl . If Ti forms a
max-heap on the y coordinates then we can output the points as follows -

Let vy denote the y coordinate associated with a node v.

If vy < yℓ or if v is a leaf node, then terminate search else

Output the point associated with v. Search u where u is the

left child of v.
Search w where w is the right child of v.

Since v is a root of max-heap, if vy < yℓ , then all the descendents of v and therefore we
do not need to search any further. This establishes correctness of the search procedure.
Let us mark all the nodes that are visited by the procedure in the second phase. When
we visit a node in the second phase, we either output a point or terminate the search.
For the nodes that are output, we can charge it to the output size. For the nodes that
are not output, let us add a charge to its parent - the maximum charge to a node is
two because of its two children. The first phase takes O(log n) time to determine the
canonical sub-intervals and so the total search time is O(log n + k) where k is number
of output points3 .
Until now, we assmed that such a dual-purpose data structure exists. How do we
construct one ?
First we can build a leaf based BST on the x coordinates. Next, we promote the
points according to the heap ordering. If a node is empty, we inspect its two children
and the pull up the larger value. We terminate when no value moves up.

Exercise 6.6 Work out the details of the construction of the second phase and also
analyse the running time.
3
This kind of analysis where we are amortizing the cost on the output points is called filtering
search.

55
This combo data structure is known as priority search trees that takes only O(n) space
and supports O(log n) time three sided query.

6.4 Planar Convex Hull

Problem Given a set P of n points in the plane,we want to compute the smallest
convex polygon containing the points.
A polygon is convex if for any two given points a,b inside the polygon,the line
segment a, b is completely inside the polygon.
A planar hull is usually represented by an ordering of the extreme points - a point
is extreme iff it cannot be expressed as a convex linear combination4 of three other
points in the convex hull. We make a few observations about the convex hull CH(P ).

Observation 6.2 CH(P ) can be described by an ordered subset x1 , x2 . . . of P , such

that it is the intersection of the half-planes supported by (xi , xi+1 ).

We know that the entire segment (xi , xi+1 ) should be inside the hull, so if all the points
of P (other than xi , xi+1 lie to one side, then CH(P ) ⊂ half-plane supported by xi , xi+1 .
Therefore CH(P ) is a sequence of extreme points and the edges joining those points
and clearly there cannot be a smaller convex set containing P since any point in the
intersection must belong to the convex hull.
For building the hull, we divide the points by a diagonal joining the leftmost and
the rightmost point - the points above the diagonal form the upper hull and the points
below form the lower hull. We also rotate the hull so that the diagonal is parallel to
x-axis. We will describe algorithms to compute the upper hull - computing the lower
hull is analogous.
The planar convex hull is a two dimensional problem and it cannot be done using
a simple comparison model. While building the hull, we will need to test whether
three points (x0 , y0 ), (x1 , y1 ), and (x2 , y2) are clockwise (counter-clockwise) oriented.
Since the x-coordinates of all the points are ordered, all we need to do is test whether
the middle point is above or below the line segment formed by the other two. A triple
of points (p0 , p1 , p2 ) is said to form a right turn iff the determinant
 
 x0 y0 1 
 
 x1 y1 1  < 0
 
 x2 y2 1 
where (x1 , y1 ) are the co-ordinates of p1 . If the determinant is positive, then the triple
points form a left turn. If the determinant is 0, the points are collinear.
4
inside the triangle

56
6.4.1 Jarvis March
A very intuitive algorithm for computing convex hulls which simply simulates (or gift
wrapping). It starts with any extreme point and repeatedly finds the successive points
in clockwise direction by choosing the point with the least polar angle with respect to
the positive horizontal ray from the first vertex. The algorithm runs in O(nh) time
where h is the number of extreme points in CH(P ). Note that we actually never
compute angles; instead we rely on the determinant method to compare the angle
between two points, to see which is smaller. To the extent possible, we only rely on
algebraic functions when we are solving problems in Rd . Computing angles require
inverse trigonometric functions that we avoid. Jarvis march starts by computing the
leftmost point ℓ, i.e., the point whose x-coordinate is smallest which takes linear time.
When h is o(log n), Jarvis march is asymptotically faster than Graham’s scan.

6.4.2 Graham’s Scan

Using this algorithm each point in the set P is first sorted using their x-coordinate
in O(n log n) time and then inductively constructs a convex chain of extreme points.
For the upper hull it can be seen easily that a convex chain is formed by successive
right-turns as we proceed in the clockwise direction from the left-most point. When
we consider the next point (in increasing x-coordinates), we test if the last three
points form a a convex sub-chain, i.e. they make a right turn. If so, we push it into
the stack. Otherwise the middle point among the triple is discarded (Why ?) and
the last three points (as stored in the stack) are tested for right-turn. It stops when
the convex sub-chain property is satisfied.
Time bound : Since each point can be inserted and deleted at most once, the
running time is linear after the initial sorting.

6.4.3 Sorting and Convex hulls

There is a very close relationship between sorting and convex hulls and we can reduce
sorting to convex hull in the following manner. Suppose all the input points are on a
parabola (i.e. all are extreme points). Then the ordered output of extreme points is
a sorted output.
So it is hardly surprising that almost all sorting algorithms have a counterpart in
the world of convex hull algorithms.
An algorithm based on divide-and-conquer paradigm which works by arbitrary
partitioning is called merge hull. After creating arbitrary partition, we construct
convex hulls of each partition. Finally, merge the two partitions in O(n) steps.
n
T (n) = 2T ( ) + O(n)
2
57
 The key step here is to merge the two upper hulls. Note that the two upper
hulls are not necessarily separated by a vertical line L. The merge step computes the
common tangent, called bridge over line L, of these two upper hulls.

6.5 A Quickhull Algorithm

Let S be a set of n points whose convex hull has to be constructed. We compute
the convex hull of S by constructing the upper and the lower hull of S. Let pl and
pr be the extreme points of S in the x direction. Let Sa ( Sb ) be the subset of S
which lie above ( below ) of the line through pl and pr . As we had noted previously,
Sa ∪ {pl , pr } and Sb ∪ {pl , pr } determine the upper and the lower convex hulls. We will
describe the algorithm Quickhull. to determine the upper hull using Sa ∪ {pl , pr }. As
the name suggests, this algorithm shares many features with its namesake quicksort.
The slope of the line joining p and q is denoted by slope(pq). The predicate
left-turn(x, y, z) is true if the sequence x, y, z has a counter-clockwise orientation, or
equivalently the area of the triangle has a positive sign.

58
 Algorithm Quickhull(Sa , pl , pr )
Input: Given Sa = {p1 , p2 , . . . , pn } and the leftmost extreme point pl and
the rightmost extreme point pr . All points of Sa lie above the line pl pr .
Output: Extreme points of the upper hull of Sa ∪ {pl , pr } in clockwise order.
Step 1. If Sa = {p}, then return the extreme point {p}.
Step 2. Select randomly a pair {p2i−1 , p2i } from the the pairs
{p2j−1, p2j }, j = 1, 2, . . . , ⌊ n2 ⌋.
Step 3. Select the point pm of Sa which supports a line with slope(p2i−1 p2i ).
(If there are two or more points on this line then choose a pm that is
distinct from pl and pr ). Assign Sa (l) = Sa (r) = ∅).
Step 4. For each pair {p2j−1 , p2j }, j = 1, 2, . . . , ⌊ n2 ⌋ do the following
( assuming x[p2j−1 ] < x[p2j ] )
Case 1: x[p2j ] < x[pm ]
if left-turn (pm , p2j , p2j−1 ) then Sa (l) = Sa (l) ∪ {p2j−1, p2j }
else Sa (l) = Sa (l) ∪ {p2j−1 }.
Case 2: x[pm ] < x[p2j−1 ]
if left-turn (pm , p2j−1, p2j ) then Sa (r) = Sa (r) ∪ {p2j }
else Sa (r) = Sb (r) ∪ {p2j−1 , p2j }.
Case 3: x[p2j−1 ] < x[pm ] < x[p2j ]
Sa (l) = Sa (l) ∪ {p2j−1 };
Sa (r) = Sa (r) ∪ {p2j }.
Step 5. (i) Eliminate points from Sa (l) which lie below the line joining pl and pm .
(ii) Eliminate points from Sa (r) which lie below the line joining pm and pr .
Step 6. If Sa (l) 6= ∅ then Quickhull(Sa (l), pl , pm ).
Output pm .
If Sa (r) 6= ∅ then Quickhull(Sa (r), pm , pr ).

Exercise 6.7 In step 3, show that if the pair {p2i−1 , p2i } satisfies the property that
the line containing p2i−1 p2i does not intersect the line segment pl pr , then it guarantees
that p2i−1 or p2i does not lie inside the triangle △pl p2i pr or △pl p2i−1 pr respectively.
This could improve the algorithm in practice.

6.5.1 Analysis
To get a feel for the convergence of the algorithm Quickhull we must argue that in
each recursive call, some progress is achieved. This is complicated by the possibility
that one of the end-points can be repeatedly chosen as pm . However, if pm is pl ,
then at least one point is eliminated from the pairs whose slopes are larger than the

59
 L

p p
m 2j-1

p
2j

p p
r
l

x
m

Figure 6.1: Left-turn(pm , p2j−1, p2j ) is true but slope(p2j−1 p2j ) is less than the median
slope given by L.

supporting line L through pl . If L has the largest slope, then there are no other points
on the line supporting pm (Step 3 of algorithm). Then for the pair (p2j−1 , p2j ), whose
slope equals that of L, left-turn (pm , p2j , p2j−1 ) is true, so p2j−1 will be eliminated.
Hence it follows that the number of recursive calls is O(n + h), ,since at each call,
either with an output vertex or it leads to elimination of at least one point.
Let N represent the set of slopes(p2j−1 p2j ), j = 1, 2, . . . ⌊ n2 ⌋. Let k be the rank
of the slope(p2i−1 p2i ), selected uniformly at random from N. Let nl and nr be the
sizes of the subproblems determined by the extreme point supporting the line with
slope(p2i−1 , p2i ). We can show that
Observation 6.3 max(nl , nr ) ≤ n − min(⌊ n2 ⌋ − k, k).
Without loss of generality, let us bound the size of the right sub-problem. There are
⌊ n2 ⌋ − k pairs with slopes greater than or equal to slope(p2i−1 p2i ). At most one point
out of every such pair can be an output point to the right of pm .
If we choose the median slope, i.e., k = n4 , then nl , nr ≤ 43 n. Let h be the number
of extreme points of the convex hull and hl (hr ) be the extreme points of the left
(right) subproblem. We can write the following recurrence for the running time.
T (n, h) ≤ T (nl , hl ) + T (nr , hr ) + O(n)
where nl + nr ≤ n, hl + hr ≤ h − 1.
Exercise 6.8 Show that T (n, h) is O(n log h).
Therefore this achieves the right balance between Jarvis march and Graham scan as
it scales with the output size at least as well as Jarvis march and is O(n log n) in the
worst case.

60
 ∗
6.5.2 Expected running time
Let T (n, h) be the expected running time of the algorithm randomized Quickhull to
compute h extreme upper hull vertices of a set of n points, given the extreme points
pl and pr . So the h points are in addition to pl and pr , which can be identified using
3
2
· n comparisons initially. Let p(nl , nr ) be the probability that the algorithm recurses
on two smaller size problems of sizes nl and nr containing hl and hr extreme vertices
respectively. Therefore we can write
X
T (n, h) ≤ p(nl , nr )(T (nl , hl ) + T (nr , hr )) + bn (6.5.1)
∀nl ,nr ≥0

where nl , nr ≤ n − 1 and nl + nr ≤ n, and hl , hr ≤ h − 1 and hl + hr ≤ h and

b > 0 is a constant. Here we are assuming that the extreme point pm is not pl or
pr . Although, in the Quickhull algorithm, we have not explicitly used any safeguards
against such a possibility, we can analyze the algorithm without any loss of efficiency.

Lemma 6.1 T (n, h) ∈ O(n log h).

Proof: We will use the inductive hypothesis that for h′ < h and for all n′ , there
is a fixed constant c, such that T (n′ , h′ ) ≤ cn′ log h′ . For the case that pm is not pl or
pr , from Eq. 6.5.1 we get
P
T (n, h) ≤ ∀nl ,nr ≥0 p(nl , nr )(cnl log hl + cnr log hr ) + bn.

Since nl + nr ≤ n and hl , hr ≤ h − 1,

nl log hl + nr log hr ≤ n log(h − 1) (6.5.2)

Let E denote the event that max(nl , nr ) ≤ 87 n and p denote the probability of E. Note
that p ≥ 21 .
From the law of conditional expectation, we have

T (n, h) ≤ p · [T (nl , hl |E) + T (nr , hr )|E] + (1 − p) · [T (nl , hl + |Ē) + T (nr , hr |Ē)] + bn

where Ē represents the complement of E.

When max(nl , nr ) ≤ 87 n, and hl ≥ hr ,

7 1
nl log hl + nr log hr ≤ n log hl + n log hr (6.5.3)
8 8
The right hand side of 6.5.3 is maximized when hl = 87 (h − 1) and hr = 81 (h − 1).
Therefore,

61
 nl log hl + nr log hr ≤ n log(h − 1) − tn
where t = log 8 − 87 log 7 ≥ 0.55. We get the same bounds when max(nl , nr ) ≤ 78 n
and hr ≥ hl . Therefore
T (n, h) ≤ p(cn log(h − 1) − tcn) + (1 − p)cn log(h − 1) + bn
= pcn log(h − 1) − ptcn + (1 − p)cn log(h − 1) + bn
≤ cn log h − ptcn + bn
Therefore from induction, T (n, h) ≤ cn log h for c ≥ tpb .
In case pm is an extreme point (say pl ), then we cannot apply Eq. 6.5.1 directly,
but some points will still be eliminated according to Observation 6.3. This can happen
a number of times, say r ≥ 1, at which point, Eq. 6.5.1 can be applied. We will show
that this is actually a better situation, that is, the expected time bound will be less
and hence the previous case dominates the solution of the recurrence.
The rank k of slope(p2i−1 p2i ) is uniformly distributed in [1, n2 ], so the number of
points eliminated is also uniformly distributed in the range [1, n2 ] from Observation
6.3. (We are ignoring the floor in n2 to avoid special cases for odd values of n - the
same bounds can be derived even without this simplification). Let n1 , n2 . . . nr be the
r random variables that represent the sizes of subproblems in the r consecutive
Pr times
that pm is an extreme point. It can be verified by induction, that E[ i=1 ni ] ≤ 4n
and E[nr ] P ≤ (3/4)r n where E[·] represents the expectation of a random variable.
Note that ri=1 b · ni is the expected work done in the r divide steps. Since cn log h ≥
4nb + c(3/4)r · n log h for r ≥ 1 (and log h ≥ 4), the previous case dominates. 2

6.6 Point location using persistent data structure

The point location problem involves an input planar partition (a planar graph with
an embedding on the plane), for which we build a data structure, such that given a
point, we want to report the region containing the point. This fundamental problem
has numerous applications including cartography, GIS, Computer Vision etc.
The one dimensional variant of the problem has a natural solution based on binary
search - in O(log n) time, we can find the interval containing the query point. In two
dimensions, we can also consider a closely related problem called ray shooting, in
which we shoot a vertical ray in the horizontal direction and report the first segment
that it hits. Observe that every segment borders two regions and we can report the
region below the segment as a solution to the point location problem.
Exercise 6.9 Design an efficient solution to the ray shooting problem using by ex-
tending the interval trees.
Consider a vertical slab which is cris-crossed by n line segments such that no pair
of segments intersect within the slab. Given a query point, we can easily solve the

62
binary search to answer a ray shooting query in O(log n) primitives of the following
kind - Is the point below/above a line segment. This strategy works since the line
segments are totally ordered within the slab (they mat intersect outside).
For the planar partition, imagine a vertical line V being swept from left to right
and let V (x) represent the intersection of V with the planar partition at an X-
coordinate value x. For simplicity let us assume that no segment is vertical. Further
let us order the line-segments according to V (x) and denote it by S(x).

Observation 6.4 Between two consecutive (in X direction) end-points of the planar
partition, S(x) remains unchanged.

Moreover the region between two consecutive end-points is a situation is similar to

vertical slab discussed before. So once we determine which vertical slab contains the
query point, in an additional O(log n) steps, we can solve the ray shooting prob-
lem. Finding the vertical slab is a one dimensional problem and can be answered in
O(log n) steps involving a binary search. Therefore the total query time is O(log n)
but the space bound is not nearly as desirable. If we treat the (2n − 1) vertical slabs
corresponding to the 2n end-points, we are required to build Ω(n) data structures,
each of which involves Ω(n) segments.

Exercise 6.10 Construct an example that depicts the worst case.

A crucial observation is that the two consecutive vertical slabs have almost all the
segments in common except for the one whose end-points separate the region.
Can we exploit the similarity between two ordered lists of segments and
support binary search on both list efficiently ? In particular, can we avoid
storing the duplicate segments and still support log n steps binary searches.
Here is the intuitive idea. Wlog, let as assume that an element is inserted and we
would like to maintain both versions of the tree (before and after insertion). Let us
also assume that the storage is leaf based.
path copying strategy If a node changes then make a new copy of its parent and also
copy the pointers to its children.
Once a parent is copied, it will lead to copying its parent, etc, until the entire
root-leaf path is copied. At the root, create a label for the new root. Once we know
which root node to start the binary search, we only follow pointers and the search
proceeds in the normal way that is completely oblivious to fact that there are actually
two implicit search trees. The search time also remains unchanged at O(log n). The
same strategy works for any number of versions except that to start searching at
the correct root node, we may require an additional data structure. In the context of
planar point location, we can build a binary search tree that supports one dimensional
search.

63
 The space required is path − length · #versions + n which is O(n log n). This
is much smaller than the O(n2 ) scheme tat stores each tree explicitly. With some
additional ideas, it is possible to improve it to O(n) space which will be optimal.

64
Chapter 7

Fast Fourier Transform and

Applications

7.1 Polynomial evaluation and interpolation

A polynomial P(x) of degree n−1 in indeterminate x is a power series with maximum
degree n − 1 and has the general form an−1 xn−1 + an−2 xn−2 + . . . a1 x + a0 , where ai
are coefficients over some field, typically the complex numbers C. Some of the most
common problems involving polynomials are
evaluation
Pn−1 Given a value for the indeterminate x, say x′ , we want to compute
′i
i=0 ai · x .
By Horner’s rule, the most efficient way to evaluate a polynomial is given by
the formula
(((an−1 x′ + an−2 )x′ + an−3 )x′ + . . . a0
We are interested in the more general problem of evaluating a polynomial at
multiple (distinct) points, say x0 , x1 . . . xn−1 . If we apply Horner’s rule then it
will take Ω(n2 ) operations, but we will be able to do it much faster.
interpolation Given n values (not necessarily distinct), say y0 , y1 . . . yn−1 , there is a
unique polynomial of degree n − 1 such that P(xi ) = yi xi are distinct.
This follows from the fundamental theorem of algebra which states that a poly-
nomial of degree d has at most d roots. Note that a polynomial is characterized
by its coefficients ai 0 ≤ i ≤ n − 1. A popular method for interpolation is the
Lagrange’s formula.
n−1 Q
j6=i (x − xj )
X
P(x) = yk · Q
k=0 j6=k (xk − xj )

65
 Exercise 7.1 Show that Lagrange’s formula can be used to compute the coeffi-
cients ai ’s in O(n2 ) operations.
One of the consequences of the interpolation is an alternate representation of
polynomials as {(x0 , y0), (x1 , y1) . . . (xn−1 , yn−1)} from where the coefficients can
be computed. We will call this representation as the point-value representation.
multiplication The product of two polynomials can be easily computed in O(n2 )
steps by clubbing the coefficients of the powers of x. This is assuming that the
polynomials are described by their coefficients. If the polynomials are given by
their point-value, then the problem is considerably simpler since
P (x) = P1 (x) · P2 (x) where P is the product of P1 and P2
A closely related problem is that
P of convolution where we have to perform
computations of the kind ci = l+p=i al · bp for 1 ≤ i ≤ n.
The efficiency of many polynomial related problems depends on how quickly we can
perform transformations between the two representations.

7.2 Cooley-Tukey algorithm

We will solve a restricted version of the evaluation problem where we will carefully
choose the points x0 , x1 . . . xn−1 to reduce the total number of computations, Let n
be a power of 2 and let us choose xn/2 = −x0 , xn/2+1 = −x1 , . . . xn−1 = −xn/2−1 . You
can verify that P(x) = PE (x2 ) + xPo (x2 ) where
PE = a0 + a2 x + . . . an−2 xn/2−1
P0 = a1 + a3 x + . . . an−1 xn/2−1
corresponding to the even and odd coefficients and PE , PO are polynomials of degree
n/2 − 1.
P(xn/2 ) = PE (x2n/2 ) + xn/2 PO (x2n/2 ) = PE (x20 ) − x0 PO (x20 )
since xn/2 = −x0 . More generally
P(xn/2+i ) = PE (x2n/2+i ) + xn/2+i PO (x2n/2+i ) = PE (x2i ) − xi PO (x2i ), 0 ≤ 0 ≤ n/2 − 1
since xn/2+i = −xi . Therefore we have reduced the problem of evaluating a degree
n − 1 polynomial in n points to that of evaluating two degree n/2 − 1 polynomials at
n/2 points x20 , x21 . . . x2n/2−1 . This will also involve O(n) multiplications and additions
to compute the values at the original points. To continue this reduction, √ we have
2 2
to choose points such that x0 = −xn/4 or equivalently xn/2 = −1 = ω. This
involves complex numbers if we started with coefficients in R1 . If we continue with
1
Depending on our choice of the field F , we can define ω such that ω 2 = −1.

66
this strategy of choosing points, at the j-th level of recursion, we require
j−1 j−1 n
x2i = −x2nj +i 0 ≤ i ≤ −1
2 2j
log n−1 log n−1
This yields x21 = −x02 , i.e., if we choose ω n/2 = −1 then xi = ωxi−1 . By
setting x0 = 1, the points of evaluation work out to be 1, ω, ω 2 . . . ω n/2 . . . ω n−1 which
are usually referred to as the principal n-th roots of unity.
Analysis
Let P(x)za10,z,a21...z
...an−1 denote the evaluation of P(x) with coefficients a0 , a1 . . . an−1 at
i

points z1 , z2 . . . zi . Then we can write the recurrence

2 n−1 n/2−1 n/2−1
P(x)a1,ω,ω ...ω
0 ,a1 ...an−1
= P(x)a1,w...ω
0 ,a2 ...an/2−2
+P(x)a1,w...ω
1 ,a3 ...an/2−1
+O(n) multiplications and additions

This immediately yields O(n log n) operations for the FFT computation.
For the inverse problem, i.e., interpolation of polynomials given the values at
1, ω, ω 2 . . . ω n−1, let us view the process of evaluation as a matrix vector product.
     
1 1 1 ... 1 a0 y0
 1 ω2 ω4 ... ω 2(n−1)    a1   y1 
   

 1 ω3 6 3(n−1)
ω ... ω  ·  a2  =  y2 
    
  ..   .. 

 ..
 .   .   . 
n−1 2(n−1) (n−1)(n−1) an−1 yn−1
1 ω ω ... ω

Let us denote this by the matrix equation A · a = y. In this setting, the interpolation
problem can be viewed as computing the a = A−1 · y. Even if we had A−1 available,
we still have to compute the product which could take Ω(n2 ) steps. However the good
news is that the inverse of A−1 is
 
1 1 1 ... 1
 1 1 1 1
ω2 ω4
... ω 2(n−1)

1 1

1 1
... 1


ω3 ω6 ω 3(n−1)
n  ..
 

 . 
1 1 1
1 ω n−1 ω 2(n−1)
... ω (n−1)(n−1)

which can be verified by multiplication with A. Recall that

1 + ω i + ω 2i + ω 3i + . . . w i(n−1) = 0
in −1
(Use the identity j ω ji = ωωi −1 = 0 for ω i 6= 1.)
P

Moreover ω −1 , ω −2, . . . w −(n−1) also satisfy the properties of n-th roots of unity. This
enables us to use the same algorithm as FFT itself that runs in O(n log n) operations.

67
7.3 The butterfly network

1 stage 2 stage 3 stage

a0 A(ω0 )
000 000

100 a4 A(ω1 ) 001

ω4
010 a2 A(ω2 ) 010
ω4
ω2
110 a6 A(ω3 ) 011
ω4 ω6
001 a1 A(ω4 ) 100

101 a5 A(ω5 ) 101

ω4
011 a3 A(ω6 ) 110

111 a7 A(ω7 ) 111

ω4

Figure 7.1: Computing an eight point FFT using a butterfly network

If you unroll the recursion of an 8 point FFT, then it looks like the Figure 7.1.
Let us work through some successive recursive calls.

P0,1,..7 (ω0 ) = P0,2,4,6 (ω02) + ω0 P1,3,5,7 (ω02 )

P0,1,..7(ω4 ) = P0,2,4,6 (ω02 ) − ω0 P1,3,5,7 (ω02 )

Subsequently, P0,2,4,6 (ω02 ) = P0,4 (ω04) + w02P2,6 (ω04 ) and

P0,2,4,6 (ω22 ) = P0,4 (ω04) − w02 P2,6 (ω04)

To calculate P0,4 (ω04) and P0,4 (ω14 ) we compute P0,4 (ω04 ) = P0 (ω08) + ω04 P4 (ω08 ) and

P0,4 (ω14 ) = P0 (ω08 ) − ω04 P4 (ω08)

Since Pi denotes ai , we do not recurse any further. Notice that in the above figure
a0 and a4 are the multipliers on the left-hand side. Note that the indices of the ai
on the input side correspond to the mirror image of the binary representation of i. A
butterfly operation corresponds to the gadget ⊲⊳ that corresponds to a pair of recursive
calls. The black circles correspond to ”+” and ”-” operations and the appropriate
multipliers are indicated on the edges (to avoid cluttering only a couple of them are
indicated).

68
 One advantage of using a network is that, the computation in each stage can
be carried out in parallel, leading to a total of log n parallel stages. Thus FFT is
inherently parallel and the butterfly network manages to capture the parallelism ina
natural manner.

7.4 Schonage and Strassen’s fast multiplication

In our analysis of the FFT algorithm, we obtained a time bound with respect to
multiplication and additions in the appropriate field - implicitly we assumed C, the
complex field. This is not consistent with the boolean model of computation and we
should be more careful in specifying the precision used in our computation. This is
a topic in itself and somewhat out of the scope of the discussion here. In reality, the
FFT computations are done using limited precision and operations like rounding that
inherently result in numerical errors.
In other kinds of applications, like integer multiplication, we choose an appro-
priate field where we can do exact arithmetic. However, we must ensure that the
field contains n-th roots of unity. Modular arithmetic, where computations are done
modulo a prime number is consistent with the arithmetic done in hardware.

Observation 7.1 In Zm where m = 2tn/2 + 1 and n is a power of 2, we can use

ω = 2t .

Since n and m are relatively prime, n has a unique inverse in Zm (recall extended
Euclid’s algorithm). Also
n/2 n/2
ω n = ω n/2·ω n/2 = (2t ) ·(2t ) ≡ (m−1)·(m−1) mod m ≡ (−1)·(−1) mod m ≡ 1 mod m

Claim 7.1 If the maximum size of a coefficient is b bits, the FFT and its inverse can
be computed in time proportional to O(bn log n).

Note that addition of two b bit numbers take O(b) steps and the multiplications
with powers of ω are multiplications by powers of two which can also be done in
O(b) steps. The basic idea of the algorithm is to extend the idea of polynomial
multiplication. Recall, that in Chapter 2 , we had divided each number into two parts
and subsequently recursively computed by computing product of smaller numbers. By
extending this strategy, we divide the numbers a and b into k parts ak−1 , ak−2 , . . . a0
and bk−1 , bk−2 , . . . b0 .

a × b = ak−1 · xk−1 + ak−2 · xk−2 + . . . a0 × bk−1 · xk−1 + bk−2 · xk−2 + . . . b0

69
where x = 2n/k - for simplicity assume n is divisible by k. By multiplying the RHS,
and clubbing the coefficients of xi , we obtain

a × b = ak−1 bk−1 x2(k−1) + (ak−2 b1 + bk−2 a1 )x2k−3 + . . . a0 b0

Although in the final product, x = 2n/k , we can compute the coefficients using any
method and perform the necessary multiplcations by an appropriate power of two
(which is just adding trailing 0’s). This is polynomial multiplication and each term
is a convolution, so we can invoke FFT-based methods to compute the coefficients.
The following recurrence captures the running time

T (n) ≤ P (k, n/k) + O(n)

where P (k, n/k) is the time for polynomial multiplication of two degree k − 1 polyno-
mials involving coefficients of size n/k. (In a model where the coefficients are not too
large, we could have used O(k log k) as the complexity of polynomial multiplication.)
We will have to do exact computations for the FFT and for that we can use modular
arithmetic. The modulo value must be chosen carefully so that
(i) It must be larger than the maximum value of the numbers involved, so that there
is no loss of information
(ii) Should not be too large, otherwise, operations will be expensive.
Moreover, the polynomial multiplication itself consists of three distinct phases
(i) Forward FFT transform. This takes O(bk log k) using b bits.
(ii) Paiwise product of the values of the polynomials at the roots of unity.
This will be done recursively with cost 2k · T (b) where b ≥ n/k.
The factor two accounts for the number of coefficients of the product of
two polynomials of degree k − 1.
(iii) Reverse FFT, to extract the actual coefficients. This step also takes
O(bk log k) where b is the number of bits in each operand.
So the previous recurrence can be expanded to

T (n) ≤ r · T (b) + O(bk log k)

where r · b ≥ n and we must choose an appropriate value of b. For coefficients of

size s, we can argue that the maximum size of numbers during the FFT computation
is 2s + log r bits (sum of r numbers
p of pairwise multiplication
√ of s bit numbers).
If we choose r to be roughly n/ log n, then b = n log n and we can rewrite the
recurrence as
n
r p
T (n) ≤ 2 · T (2 n log n + log n) + O(n log n) (7.4.1)
log n

70
Exercise 7.2 With appropriate teminating condition, say the O(nlog2 3 ) time multi-
plication algorithm, verify that T (n) ∈ O(n log2 n log log n).

An underlying assumption in writing the recurrence is that all the expressions are
integral.
√ This can actually be ensured by choosing n = 2ℓ and carefully choosing
n for even and odd values of ℓ. Using the technique of wrapped convolution, one
can save a factor of two in the degree of the polynomial, yielding the best known
O(n log n log log n) algorithm for multiplication.

71
Chapter 8

String matching and finger printing

8.1 Rabin Karp fingerprinting

Notations If Y is a string of length m, Yj represents the jth character and Y (j)
represents the substring of n symbols beginning at Yj .
In terms of the above notation, we define the string matching problem as:
Given X (the pattern), find the first index i such that X = Y (i)
The obvious way of doing this is brute-force comparision of each Y (i) with X
that could result in Ω(n2 ) comparisons. Alternatively, consider the following Idea :
Compare F (X) with F (Y (1)), F (Y (2)) etc. for some function F () that maps strings
of lengths n to relatively shorter strings. The claim is if F (X) 6= F (Y (i)) for any i
then X 6= Y (i) else there is a chance that X = Y (i) whenever F (X) = F (Y (i)).
The function F is known as the fingerprinting function1 and may be defined ac-
cording to the application. In this case let

F (X) = x mod p

. Here X is assumed to be a binary pattern (of 0 and 1) and x is its integer value.
1
It is called a hash function.

72
Theorem 8.1 (Chinese Remaindering Theorem) For k numbers n1 , n2 , ..., nk , rela-
tively prime to each other,

x ≡ y mod ni for all i ⇔ x ≡ y mod n1 n2 n3 ...nk = M

Moreover,
R
X
y≡ ci d i y i
i=1
Q
where ci di ≡ 1 mod ni , di = n1 , n2 . . . ni−1 ni+1 , . . . nk and yi = x mod ni

Let k be such that 2m < M = 2 × 3 × ... × pk i.e. the first k primes. From CRT, if
X 6= Y (i) then for some p in {2, 3, . . . , pk },

Fp (X) 6= Fp (Y (i))

Enlarge the set somewhat, i.e. {2, 3, . . . , p2k }

1
P [Fp (X) = Fp (Y (i))|X 6= Y (i)] ≤
2
Otherwise it would violate CRT. So, if we take {2, 3, ..., pt2 k } then the probability
of false match at fixed position Y (i) ≤ t12 . So for any i ∈ {1, ..., t} ≤ t t12 = 1t . Thus
k = n will suffice.

Size of numbers involved: The product of n primes 2.3....pn = M > 2n . Also

from prime number density, lnUU ≤ π(U) ≤ 1.26 lnUU , where π(x) represent number of
primes less than or equal to x.
U
> t2 n ⇒ U = O(t2 n log(t2 n))
ln U
So |p| ≤ 2 log t + log n, where |p| is the bit length.

Updating Fingerprints: Y (i+1) = 2(Y (i)−Yi ·2n−1 )+Yi+n. So Y (i+1) mod p =

[2(Y (i) mod p) − 2n mod p.Yi + Yi+n mod p] mod p. This is all modulo p. So as long
as |p| = Ω(log t) = wordsize, it is constant update time. Therefore, the expected cost
of a step = O(1) + tn2 . The O(1) is due to the cost to update fingerprint function and
the term tn2 is the probability of false-match multiplied by the cost of verification.

73
 So expected cost =
m
X n n·m
O(1) + = O(m) +
1
t2 t2

By choosing t ≥ m, it is O(m).

8.2 KMP algorithm

String matching can be done in linear time with a custom-made DFA (Deterministic
Finite Automaton) for the pattern that we are trying to find. At any stage the state
of the DFA corresponds to the extent of partial match - it is in state i, if the previous
i symbols of the text has matched the first i symbols of the pattern. It reaches the
final stage iff it has found a match. Given this DFA, we can find all occurrences of
an n symbol pattern in an m symbol text in O(m) steps, where there is a transition
for every input symbol of the text. The size of the DFA is O(n|Σ|) where Σ is the
alphabet which is optimal if Σ is of constant size.
With some additional ideas, the previous method can be made to run in O(n +
m) steps without dependence on the alphabet size. Let us introduce dome useful
notations.
Let X(i) denote the first i symbols of the pattern, i.e. a prefix of length i.
Let α < β denote α is a suffix of β.
We mimic the DFA in the sense that in case of mismatch of the input alphabet
and the pattern, we want to find the largest overlap of the pattern and the part of
the text scanned. If we have matched upto i symbols before a mismatch, we want
to find the largest j |j < i such that X(j) < X(i). Following this, we try matching
Xj+1 with the next element of the text Y .
The failure function of a string is defined as

f (i) = max{X(j) < X(i)} otherwise 0, if no such X(j) exists

j

With this definition, the strategy is as follows.

Let Yk denote the k-th symbol of the text. Suppose we have partial a match upto i
symbols of the pattern, we then try to match Xi+1 with Xk+1 -st position of the text.
In case of a match, we increase the partial match and if it is n then we have found a
match.
Otherwise (if Xi+1 doesn’t match Yk+1 ), we try to match Xf (i)+1 with Yk+1 and again
if there no match, we try Xf (f (i)+1 with Yk+1 and so on till the partial match becomes
0.
74
Let us postpone the method for computing the failure function and assume that we
have the failure function available. The analysis requires us to look at a situation,
where the pattern string keeps sliding to the right (till it cannot). We can analyze it
in many ways - here we will introduce the technique of potential function.

8.2.1 Potential method and amortized analysis

To accurately analyse the performance of an algorithm, let us denote by Φ() as a
function that captures the state of an algorithm or its associated data structure at
any stage i. We define amortized work done at step i of an algorithm as wi + ∆i
where wi is actual number of steps2 ∆i = Φ(i) − Φ(i − 1) which is referred to as the
difference
Pi=tin potential. Note that the total work done by an algorithm over t steps is
W = i=1 wi . On the other hand, the total amortized work is
t
X
(wi + ∆i ) = W + Φ(t) − Φ(0)
i=1

If Φ(t) − Φ(0) ≥ 0, amortized work is an upperbound on the actual work.

Example 8.1 A stack supports push, pop and empty-stack operations. Define Φ() as
the number of elements in the stack. If we begin from an empty stack, Φ(0) = 0. For
a sequence of push, pop and empty-stack operations, we can analyze the amortized
cost. Amortized cost of push is 2, for pop it is 0 and for empty stack it is negative.
Therefore the bound on amortized cost is O(1) and therefore the cost of n operations
is O(n). Note that the worst-case cost of an empty-stack operation can be very high.

8.2.2 Analysis of the KMP algorithm

During the algorithm, we may be comparing any given element of the text, a number
of times, depending on the failure function. Let us define the potential function as
the extent of partial match. Then

Case: match The amortised cost of a match is 2 (actual cost is one and the increase
in potential is also one).

Case mismatch The amortised cost is ≤ 0, since the potential is decreasing.

So the total amortized cost is O(m).

2
this may be hard to analyze

75
8.2.3 Pattern Analysis
The preprocessing of the pattern involves constructing the failure function f (i).

Observation 8.1 If the failure function f (i) = j, j < i, it must be true that X(j −
1) < X(i − 1) and pi = pj .

This shows that the computation of the failure function is very similar to the KMP
algorithm itself and we compute the f (i) incrementally with increasing values of i.

Exercise 8.1 Using the potential function method, show that the failure function can
be computed in O(|X|) steps.

Therefore the total running time of KMP algorithm is O(|X| + |Y |).

8.3 Generalized String matching

Very often, we encounter string matching problems where the strings are not repre-
sented explicitly. This feature lends versatility to many applications. It gives us a
way of compactly representing a set of strings and also deal with situations when we
do not have complete information about strings3 . One of the fundamental applica-
tions is parsing, where we have a compact representation of (possibly infinite) strings
in form of a grammar and given a query string, we would like to know if the string
belongs to the set described by the grammar.
In one of the simpler cases, we have to deal with wild-card symbols. For example,
there is a match between the strings acb∗ d and a∗ bed by setting the first wild card to
e and the second one as c. Here a wild-card is a placeholder for exactly one symbol.
In other applications, the wild-card may represent an entire substring of arbitrary
length. Unfortunately the none of the previous string matching algorithms are able
to handle wild-cards.

Exercise 8.2 Give a example to argue why KMP algorithm cannot handle wild-cards.
You may want to extend the definition of failure function to handle wild-cards.

8.3.1 Convolution based approach

For a start, assume that we are only dealing with binary strings. Given a pattern
A = a0 a1 a2 . . . an−1 and a text B = b0 b1 b2 . . . bm−1 where xi , yi ∈ {0, 1}, let us view
them as coefficients of polynomials. More specifically, let
3
Typical situation in many biological experiments dealing with genetic sequences

76
PA (x) = a0 xn−1 + a1 xn−2 + a2 xn−3 + . . . an−1 and PB (x) = b0 + b1 x + b2 x2 + . . . xm−1 .
Pm+n−2
The product of PA and PB can be written as i=0 ci xi Note that

cn−1+j = a0 · bj + a1 · b1+j + a2 · b2+j + . . . an−1 · bn−1+j 0 ≤ j ≤ m + n − 2

which can be interpreted as the dot product of X and Y (j) 0 ≤ j ≤ n − 1.

If we replace the {0, 1} with {−1, +1}, then we can make the following claim.

Observation 8.2 There is a match in position j iff cn−1+j = n.

Exercise 8.3 Prove it rigorously.

The above convolution can be easily done using FFT computation in O(m log m)
steps.4 When wildcard characters are present in the pattern, we can assign them the
value 0. If there are w such characters then we can modify the previous observation
by looking for terms that have value n − w (Why). However, the same may not work
if we have wildcards in the text also - try to construct a counterexample.
Wildcard in Pattern and Text
Assume that the alphabet is {1, 2 . . . s} (zero is not included). We will reserve
zero for the wildcard. For every position i of the pattern that (assume there are
no wildcards), we will associate a random number ri from Pthe set {1, 2, . . . N} for
a sufficiently large N that we will choose later. Let t = i ri Xi . Here ri s are are
random multipliers such that

Observation 8.3 For any vector P v1 , v2 . . . vn , suppose there exists some i for which
Xi 6= vi . Then the probability that i vi · ri = t is less than N1 .

We can imagine that the random numbers are chosen sequentially, so that after fixing
the first n − 1 numbers, there is only one choice for which the equation is satisfied 5 .
By choosing N ≥ n2 , the probability of a false match in any of the possible positions
is n · 1/N ≤ 1/n. P
Clearly, if the vector v1 , v2 . . . vn is the same as X, then i vi ·ri = t. So this forms
the basis for a randomized string matching algorithm. In the presence of wildcards
in the pattern X, we assign ri = 0 iff Xi =* (instead of a random non-zero number)
and the same result holds for positions that do not correspond to wildcards (these are
precisely the positions that are blanked out by 0). The number t acts like a fingerprint
or a hash function for the pattern.
4
The number involved are small enough so that we can do exact computation using O(log n) bit
integers.
5
We do the arithmetic modulo N

77
 When the text has wildcard, then the fingerprint cannot be fixed and will vary
according to the wildcards in the text. The fingerprint tk at position k of the text
can be defined as n
X
tk = δj+k−1 · rj · Xj
j=1

where δi = 0 if Xi = * and 1 otherwise. Recall that rj = 0 if Xj = *.

Now we can replace * with 0 in the text and run the previous algorithm. The
probability for error (false match) is calculated along similar lines. To calculate tj ,
we perform another convolution using FFT. To summarize, string matching in the
presence of wildcards can be done in O(m log m) operations.
Remark The use of FFT for pattern matching is due to Fisher and Patterson. How-
ever, because of superlinear running time, it is not the preferred method for simple
string matching for which KMP and Karp-Rabin are more efficient.

78
Chapter 9

Graph Algorithms

A number of graph problems use Depth First Search as the starting point. Since it
runs in linear time, it is efficient as well.

9.1 Applications of DFS

Directed Graphs
Consider a Directed Acyclic Graph (DAG), G = (V, E). We want to define a a
function f : V ⇒ {1, 2, . . . n} such that for any directed edge (u, v), f (v) > f (u).

Observation 9.1 Every DAG has at least one vertex with indegree 0 (source) and a
vertex with outdegree 0 (sink).

This is also called a topological sorting of the vertices. Consider a DFS numbering
pre(v), v ∈ V and also a post-order numbering post(v).

Observation 9.2 If there is a path from u ; v, then post(u) > post(v).

If the DFS reaches u before v, then clearly it is true. If v is reached before u, then
the DFS of v is completed before it starts at u since there is no path from v to u.

Claim 9.1 The vertices in reverse order of the post-order numbering gives a topolog-
ical sorting.

9.1.1 Strongly Connected Components (SCC)

In a directed graph G = (V, E), two vertices u, v ∈ V are in the same SCC iff u ; v
and v ; u. It is easy to verify that this is an equivalence relation on the vertices and
the equivalence classes correspond to the SCCs of the given graph. Let us define a

79
graph G = (V ′ , E ′ ) as follows - V ′ corresponds to the SCCs of G and (c1 , c2 ) ∈ E ′ if
there is an edge from some vertex in c1 to some vertex of c2 . Here c1 , c2 ∈ V ′ , i.e.,
they are SCCs of G.

Exercise 9.1 Prove that G is a DAG.

To determine the SCCs of G, notice that if we start a DFS from a vertex of a sink
component c′ of G then precisely all the vertices of c′ can be reached. Since G is not
explicitly available, we will use the following strategy to determine a sink component
of G. First, reverse the edges of G - call it GR . The SCCs of GR is the same as G
but the sink components and source components of G are interchanged.

Exercise 9.2 If we do a DFS in GR , then show that the vertex with the largest
postorder numbering is in a sink component of G.

This enables us to output the SCC corresponding to the sink component of G using
a simple DFS. Once this component is deleted (delete the vertices and the induced
edges), we can apply the same strategy to the remaining graph.

Exercise 9.3 Show that the SCCs can be determined using two DFS - one in G and
the other in GR .

9.1.2 Finding Biconnected Components (BCC)

Biconnected graphs (undirected) can be defined on the basis of vertex connectivity as
well as equivalence classes on edges. Graphs that cannot be disconnected by removing
one vertex1 (along with the incident edges) are biconnected. This implies that between
any pair of vertices, there are at least two vertex disjoint paths. Since the two vertex
disjoint paths form a simple cycle, the biconnected components also contain a lot
of information on the cyclic structure. This also leads to the alternate definition of
BCC.

Definition 9.1 Two edges belong to the same BCC iff they belong to a common
(simple) cycle.

Exercise 9.4 Show that the above relation defines an equivalence relation on edges.
Moreover, the equivalence classes are precisely the BCC (as defined by the vertex
connectivity).
1
If the removal of a vertex disconnects a graph then such a vertex is called an articulation point.

80
 The DFS on an undirected graph G = (V, E) partitions the edges into T (tree
edges) and B (back edges). Based on the DFS numbering (pre-order numbering) of
the vertices, we can direct the edges of T from a lower to a higher number and the
edges in B from a higher to a lower number. Let us denote the DFS numbering by
a function d(v) v ∈ V . Analogous to the notion of component tree in the context of
SCC, we can also define a component tree on the BCC. Here the graph G has the
biconnected components (denoted by B) and the articulation points (denoted by A)
as the set of vertices. We have an edge between a ∈ A and b ∈ B if a ∈ B.

Exercise 9.5 Show that G is a tree, i.e., it cannot contain cycles.

The basic idea behind the BCC algorithm is to detect articulation points. If there
are no articulation points then the graph is biconnected. Simultaneously, we also
determine the BCC. The DFS numbering d(v) helps us in this objective based on the
following observation.

Observation 9.3 If there are no back-edges out of some subtree of the DFS tree Tu
rooted at a vertex u that leads to a vertex w with d(w) < d(u), then u is an articulation
point.

This implies that all paths from the subtree to the remaining graph must pass through
u making u an articulation point. To detect this condition, we define an additional
numbering of the vertices based on DFS numbering. Let h(v) denote the minimum
of the d(u) where (v, u) is a back edge. Then

LOW (v) = min {LOW (w), h(v)}

w|(v,w)∈T

Exercise 9.6 How would you compute the LOW (v) v ∈ V along with the DFS num-
bering ?

The computation of LOW numbers results in an efficient algorithm for testing

biconnectivity but it does not yield the biconnected components. For this, let us
consider the component graph G. The biconnected component that corresponds to
a leaf node of G should be output as we back-up from a subtree w of v such that
LOW (w) is not smaller than d(v) (v is an articulation point). After deleting this
component from G, we consider the next leaf-component. The edges of a BCC can
be kept in stack starting from (v, w) that will be popped out till we reach the edge
(v, w).

Exercise 9.7 Formalize the above argument into an efficient algorithm that runs in
O(|V | + |E|) steps.

81
9.2 Path problems
We are given a directed graph G = (V, E) and a weight function w : E → R (may
have negative weights also). The natural versions of the shortest path problem are
as follows

distance between a pair Given vertices x, y ∈ V , find the least weighted path
starting at x and ending at y.

Single source shortest path (SSSP) Given a vertex s ∈ V , find the least weighted
path from s to all vertices in V − {s}.

All pairs shortest paths (APSP) For every pair of vertices x, y ∈ V , find least
weighted paths from x to y.

Although the first problem often arises in practice, there is no specialized algo-
rithm for it. The first problem easily reduces to the SSSP problem. Intuitively, to
find the shortest path from x to y, it is difficult to avoid any vertex z since there
may be a shorter path from z to y. Indeed, one of the most basic operations used by
shortest path algorithms is the relaxation step. It is defined as follows -

Relax(u, v) : (u, v) ∈ E,
if ∆(v) > ∆(u) + w(u, v) then ∆(v) = ∆(v) + w(u, v)

For any vertex v, ∆(v) is an upperbound on the shortest path. Initially it is set to
∞ but gradually its value decreases till it becomes equal to δ(v) which is the actual
shortest path distance (from a designated source vertex).
The other property that is exploited by all algorithms is

Observation 9.4 subpath optimality

Let s = v0 , v1 , v2 . . . vi . . . vj . . . vℓ be a shortest path from v0 tovℓ . Then for any inter-
mediate vertices, vi , vj , the subpath vi , vi+2 . . . vj is also a shortest path between vi and
vj .

This follows by a simple argument by contradiction, that otherwise the original path
is not the shortest path.

9.2.1 Bellman Ford SSSP Algorithm

The Bellman Ford algorithm is essentially based on the following recurrence

δ(v) = min {δ(u) + w(u, v)}

u∈In(v)

82
where In(v) denotes the set of vertices u ∈ V such that (u, v) ∈ E. The shortest
path to v must have one of the incoming edges into v as the last edge. The algorithm
actually maintains upperbounds ∆(v) on the distance from the source vertex s -
initially ∆(v) = ∞ for all v ∈ V − {s} and ∆(s) = 0 = δ(s). The previous recurrence
is recast in terms of ∆
∆(v) = min {∆(u) + w(u, v)}
u∈In(v)

that follows from a similar reasoning. Note that if D(u) = δ(u) for any u ∈ In(v),
then after applying relax(u, v), ∆(v) = δ(v). The underlying technique is dynamic
programming as many vertices may have common predecessors in the shortest path
recurrence.

Bellman Ford SSSP

Initialize ∆(s) = 0, ∆(v) = ∞ v ∈ V − {s}.

Repeat n − 1 times

relax(e) for all e ∈ E

Output δ(v) = ∆(v) for all v ∈ V .

The correctness of the algorithm follows from the previous discussion and the following
critical observation.
Observation 9.5 After i iterations, all vertices whose shortest paths consist of i
edges, have ∆(v) = δ(v).
It follows from a straightforward induction on the number of edges in the path with
the base case δ(s) = 0 and the definition of relax step.
So, the algorithm finds all shortest paths consisting of at most n − 1 edges with
n − 1 iterations. However, if there is a negative cycle in the graph, then you may
require more iterations and in fact, the problem is not well defined any more. We can
specify that we will output simple paths (without repeated vertices) but this version
is not easy to handle. 2
Exercise 9.8 Describe an efficient algorithm to detect negative cycle in graph.
Since each iteration involves O(|E|) relax operations - one for every edge, the total
running time is bounded by O(|V | · |E|).
To actually compute the shortest path, we keep track of the predecessor of a vertex
which is determined by the relaxation step. The shortest path can be constructed by
following the predecessor links.
2
This is equivalent to the longest path problem which is known to be intractable

83
Exercise 9.9 If there is no negative cycle, show that the predecessors form a tree
(which is called the shortest-path tree).

9.2.2 Dijkstra’s SSSP algorithm

If the graph doesn’t have negative weight edges, then we can exploit this feature to
design a faster algorithm. When we have only non-negative weights, we can actually
determine which vertex has the its ∆(v) = δ(v). In the case of Bellman Ford algo-
rithm, at every iteration, at least one vertex had its shortest path computed but we
couldn’t identify them. We maintain a partition U and V − U such s ∈ U where U is
the set of vertices v ∈ V for which ∆(v) = δ(v). One the other hand, for non-negative
weights, we can make the following claim.

Observation 9.6 The vertices v ∈ V − U for which ∆(v) is minimum, satisfies the
property that ∆(v) = δ(v).

Suppose for some vertex v that has the minimum label after some iteration, ∆(v) >
δ(v). Consider a shortest path s ; x → y ; v, where y ∈ / U and all the earlier
vertices in the path s ; x are in U. Since x ∈ U, ∆(y) ≤ δ(x) + w(x, y) = δ(y).
Since all edge weights are non-negative, δ(y) ≤ δ(v) < ∆(v) and therefore ∆(y) = δ(y)
is strictly less than ∆(v) which contradicts the minimality of ∆(v).
A crucial property exploited by Dijkstra’s algorithm is that along any shortest
path s ; u, the shortest path-lengths are non-decreasing because of non-negative
edge weights. Along similar lines, we can also assert the following

Observation 9.7 Starting with s, the vertices are inserted into U is non-decreasing
order of their shortest path lengths.

We can prove it by induction starting with s - δ(s) = 0. Suppose it is true upto

iteration i, i.e., all vertices v ∈ U are such that δ(v) ≤ δ(x), x ∈ V − U. Let
∆(u) ∈ V − U be minimum, then we claim that ∆(u) = δ(u) (from the previous
observation) and ∆(u) ≤ ∆(x), x ∈ V − U. Suppose δ(x) < δ(u), then by an
extension of the previous argument , let y be the earliest vertex in s ; x that is not
in U. Then ∆(y) = δ(y) ≤ δ(x) < δ(u) ≤ ∆(u), thereby violating the minimality of
∆(u).
To implement this algorithm efficiently, we maintain a priority queue on the values
of ∆(v) for v ∈ V −U, so that we can choose the one with the smallest value in O(log n)
steps. Each edge is relaxed exactly once since only the edges incident on vertices in U
are relaxed - however because of relax operation the ∆() of some vertices may change.
This yields a running time of ((|V | + |E|) log |V |). + Any path from s to a vertex in
V − U passes through a vertex in U, i.e., Consider the vertices in

84
9.2.3 Floyd-Warshall APSP algorithm
Consider the adjacency matrix AG of the given graph G = (V, E) where the entry
AG (u, v) contains the weight w(u, v). Let us define the matrix product of AG · AG
in a way where the multiplication and addition operators are replaced with addition
and max operator.

Claim 9.2 Define the k-th iterate of AG , namely AkG for n−1 ≥ k ≥ 2 as min{Ak−1 k−1
G , AG ·
AG } where the min is the entrywise minimum. Then AkG (u, v) contains the shortest
path distance between u and v consisting of at most k edges.

We can prove it by induction on k. Note that a path between u and v of at most k

edges consists of a path of at most k − 1 edges to one neighbours of v followed by the
last edge.

9.3 Maximum flows in graphs

Given a directed graph G = (V, E) and a capacity function C : E → R+ , and two
designated vertices s and t, we want to compute a flow function f : E → R+ such
that

1. Capacity constraint
f (e) ≤ C(e) ∀e ∈ E

2. Flow conservation
X X
∀v ∈ V − {s, t}, f (e) = f (e)
e∈in(v) e∈out(v)

where in(v) are the edges directed into vertex v and out(v) are the edges directed
out of v.

The vertices s and t are often called the source and the sink and the flow is directed
out of s and into t. P
The outflow of a vertex v is defined as e∈out(v) f (e) and the inflow into v is given
P P
by e∈in(v) f (e). The net flow is defined as outflow minus inflow = e∈out(v) f (e) −
P
e∈in(v) f (e). From the property of flow conservation, net flow is zero for all vertices
except s, t. For vertex s which is the source, the net flow is positive and for t, the net
flow is negative.

Observation 9.8 The net flow at s and the net flow at t are equal in magnitude.

85
From the flow conservation constraint
 
X X X
 f (e) − f (e) = 0
v∈V −{s,t} e∈out(v) e∈in(v)

Let E ′ be edges that are not incident on s, t (either incoming or outgoing). Then
   
X X X X X
= (f (e) − f (e)) +  f (e) − f (e) +  f (e) − f (e) = 0
e∈E ′ e∈out(s) e∈in(s) e∈out(t) e∈in(t)

For an an edge e ∈ E ′ , f (e) is counted once as incoming and once as outgoing which
cancel eachother. So
X X X X
f (e) − f (e) = f (e) − f (e)
e∈out(s) e∈in(s) e∈in(t) e∈out(t)

So the outflow at s equals the inflow at t.

Let us denote the net flow at s by F and the maximum flow f ∗ from s to t is the
maximum value of F .
Computing maxflow is one of the classic problems in combinatorial optimization
with numerous applications. Therefore designing efficient algorithms for maxflow has
been pursued by many researchers for many years. Since the constraints and the
objective function are linear, we can pose it as a linear program (LP) and use some
of the efficient (polynomial time) algorithms for LP. However the algorithms for LP
are not known to be strongly polynomial, we will explore more efficient algorithms.
bf Flow augmentation
Consider any s − t path in the graph ignoring the edge direction. If the forward
edges are not fully saturated and all the backward edges have non-zero flows, we can
increase the flow in the following manner. Let ∆ be the minimum of the residual
capacity of the forward edges and the flows of the backward edges. The edge(s) that
determines ∆ is called a bottleneck edge of the augmenting path.
By increasing the flow in the forward edges by ∆ and decreasing the backward
edges by ∆, we can preserve both the capacity constraints and the flow conservation
constraints. In this process the is now increased by ∆. Such a path is called a
augmentation path. It is clear that the flow is not maximum if there is an augmenting
path. The converse is not immediate, i.e., if there is no augmenting path then the
flow is maximum. We will establish this as a consequence of a more general result.

86
9.3.1 Max Flow Min Cut
P as a partition of V such that s ∈ S and t ∈ T . The size of
An (S, T ) cut is defined
a cut is defined as u∈S, C(u, v). Note that only the capacities of the forward edges
v∈T
are counted.

Theorem 9.1 (maxflow-mincut) The value of the s − t maxflow = s − t mincut.

Consider a flow f such that there is no augmenting path. Let S ∗ be the set of vertices
such that there is an augmenting path from s to u ∈ S ∗ . By definition, s ∈ S ∗ and
t∈/ S ∗ and T ∗ = V − S ∗ .

Observation 9.9 The forward edges from S ∗ to T ∗ are saturated and the backward
arcs from T ∗ to S ∗ are full.

Otherwise it will contradict our definition of S ∗ and T ∗ . The net flow from S ∗ to T ∗
is X X
f (e) = C(e) = C(S ∗ , T ∗ )
e∈out(S ∗ ) e∈out(S ∗ )
e∈in(T ∗ ) e∈in(T ∗ )

This implies that

f ∗ ≥ f = C(S ∗ , T ∗) mincut (9.3.1)
For any cut (S, T ) and any flow f , consider
 
X X X
 f (e) − f (e)
v∈S e∈out(v) e∈in(v)

P P
= e∈out(s) f (e)− e∈in(s) f (e) since the flow conservation holds at every other vertex
in S - this is the net flow out of s. By rewriting the the first summation over two
sets of edges E and E ′ corresponding to the cases when both endpoints of e are in S
or exactly one end-point is in S (the other is in T ), we obtain the following
X X X
(f (e) − f (e)) + f (e) − f (e)
e∈E e∈out(S) e∈in(S)

The first term is 0 and the second term equals C(S, T ). Since the third term is
negative, we can upperbound the expression by C(S, T ). Therefore f ≤ C(S, T ) for
any cut and in particular the mincut, i.e., the maxflow is less than or equal to the
mincut. As f is any flow, it implies

maxflow ≤ mincut ≤ C(S ∗ , T ∗ )

87
 In conjunction with Equation 9.3.1 , we find that f ∗ = C(S ∗ , T ∗ ) = mincut.
Since the maxflow corresponds to the situation where no augmenting path is
possible, we have proved that
The flow is maximum iff there is no augmenting path.

9.3.2 Ford and Fulkerson method

The Ford and Fulkerson strategy for maxflow is directly based on the above result,
i.e., we successively find augmenting paths till we can’t find any such path.
How do we find an augmenting path ? A residual graph G(f ) corresponding to a
flow f has the same set of nodes as the original graph and the edges are defined as
follows. For an edge e with capacity C(e) and flow f (e) such that f (e) < C(e) gen-
erates two edges e′ in forward direction and e′′ in backward direction with capacities
C(e) − f (e) and f (e) respectively. If f (e) = C(e) then only the backward edge is
present with capacity C(e). Any s-t path in G(f ) is an augmenting path that can be
found using a BFS or DFS. If t is disconnected from s, then there is no augmenting
path. Note that the capacities do not play any role in the path except that zero
capacity edges are not present.
Although the Ford Fulkerson method converges since the flow increases monoton-
ically, we do not have a bound on the maximum number of iterations that it takes
to converge to the maxflow. Bad examples (taking exponential time) can be easily
constructed and actually for irrational capacities, it converge only in the limit !

9.3.3 Edmond Karp augmentation strategy

It turns out that if we augment flow along the shortest path (in the unweighted
residual network using BFS) between s and t, we can prove much superior bounds.
The basic property that that enables us to obtain a reasonable bound is the following
result.
Claim 9.3 A fixed edge can become bottleneck in at most n/2 iterations.
We will prove the claim shortly. The claim implies that the total number of iterations
is m · n/2 or O(|V | · |E|) which is polynomial in the input size. Each iteration involves
a BFS, yielding an overall running time of O(n · m2 ).

9.3.4 Monotonicity Lemma and bounding the iterations

Let ski and tki denote the minimum number of edges in a shortest path from s to
vertex i after k iterations. We claim that
sk+1
i ≥ ski and tk+1
i ≥ ti .

88
 We will prove it by contradiction. Suppose sk+1
i < ski for some k and among all
such vertices let s ; v have minimum path length (after k + 1 iterations). Consider
the last edge in the path, say (u, v). Then

sk+1
v = sk+1
u +1 (9.3.2)

since u is on the shortest path. By assumption on minimality of violation,

sk+1
u ≥ sku (9.3.3)

From 9.3.2 , it follows that

sk+1
v ≥ sku + 1 (9.3.4)
Consider the flow f (u, v) after k iterations.

Case 1 : f (u, v) < C(u, v) Then there is a forward edge u → v and hence skv ≤ sku +1
From Equation 9.3.4 sk+1v ≥ skv . that contradicts our assumption.

Case 2: f (u, v) = C(u, v) Then u ← v is a backward arc. After k + 1 iterations, we

have a forward edge u → v, so the shortest augmenting path must have passed
through the edge v → u (after k iteration) implying that

sku = skv + 1

Combining with inequality 9.3.4 , we obtain sk+1

v = skv + 2 that contradicts our
assumption.

Let us now bound the number of times edge (u, v) can be a bottleneck for augmen-
tations passing through the edge (u, v) in either direction. If (u, v) is critical after
k-iteration in the forward direction then skv = sku + 1. From monotonicity property
sℓv ≥ skv , so
sℓv ≥ sku + 1 (9.3.5)
Let ℓ(≥ k + 1) be the next iteration when an augmenting path passes through (u, v)3.
Then (u, v) must be a backward edge and therefore

sℓu = sℓv + 1 ≥ sku + 1 + 1 = sku + 2

using inequality 9.3.5 . Therefore we can conclude that distance from u to s increases
by at least 2 every time (u, v) becomes bottleneck and hence it can become bottleneck
for at most |V |/2 augmentations.
3
it may not be a bottleneck edge.

89
9.4 Global Mincut
A cut of a given (connected) graph G = (V, E) is set of edges which when removed
disconnects the graph. An s − t cut must have the property that the designated
vertices s and t should be in separate components. A mincut is the minimum number
of edges that disconnects a graph and is sometimes referred to as global mincut to
distinguish is from s − t mincut. The weighted version of the mincut problem is the
natural analogue when the edges have non-negative associated weights. A cut can
also be represented by a set of vertices S where the cut-edges are the edges connecting
S and V − S.
It was believed for a long time that the mincut is a harder problem to solve than
the s − t mincut - in fact the earlier algorithms for mincuts determined the s − t
mincuts for all pairs s, t ∈ V . The s − t mincut can be determined from the s − t
maxflow flow algorithms and over the years, there have been improved reductions of
the global mincut problem to the s − t flow problem, such that it can now be solved
in one computation of s − t flow.
In a remarkable departure from this line of work, first Karger, followed by Karger
and Stein developed faster algorithms (than maxflow) to compute the mincut with
high probability. The algorithms produce a cut that is very likely the mincut.

9.4.1 The contraction algorithm

The basis of the algorithm is the procedure contraction described below. The funda-
mental operation contract(v1 , v2 ) replaces vertices v1 and v2 by a new vertex v and
assigns the set of edges incident on v by the union of the edges incident on v1 and v2 .
We do not merge edges from v1 and v2 with the same end-point but retain them as
multiple edges. Notice that by definition, the edges between v1 and v2 disappear.

Procedure Partition (t)

Input: A multigraph G = (V, E)
Output: A t partition of V

Repeat until t vertices remain

choose an edge (v1 , v2 ) at random

contract(v1, v2 )

contract(u, v): Merge vertices u and v into w such that

all neighbours of u and v are now neighbours of w.

90
 Procedure P artition(2) produces a 2-partition of V which defines a cut. If it is a
mincut then we are done. There are two issues that must be examined carefully.

1. How likely is it that the cut is a mincut ?

2. How do we know that it is a mincut ?

The second question addresses a more general question, namely, how does one verify
the correctness of a Monte Carlo randomized algorithm ? In most cases there are no
efficient verification procedures and we can only claim the correctness in a probabilis-
tic sense. In our context, we will show that the contraction algorithm will produce
a mincut with probability p, so that, if we run the algorithm p1 times we expect to
see the mincut at least once. Among all the cuts that are output in O( p1 ) runs of the
algorithm, we choose the one with the minimum cut value. If the minimum cut had
been produced in any of the independent runs, we will obtain the mincut.

9.4.2 Probability of mincut

Using the observation that, in an n-vertex graph with a mincut value k, the minimum
degree of a vertex is k, the probability that one of the mincut edge is contracted is
k
≤ kn/2 = n2 4 .
Given a specific mincut C, we estimate the probability that C is not outputted.
If C is output, then it means that none of the edges of C has ever been contracted.
Let A(i) denote the event that an edge of C is contracted in the ith iteration and let
E(i) denote the event that no edge of C is contracted in any of the first i iterations.
If ni is the number of vertices after i iterations (initially n0 = n) we have ni = n − i.
We have seen that Pr[Ā(1)] ≥ 1 − 2/n and similarly, Pr[Ā(i)|E(i − 1)] ≥ 1 − 2/ni−1 .
Then, using the property of conditional probability

Pr[E(i)] = Pr[Ā(i) ∩ E(i − 1)] = Pr[Ā(i)|E(i − 1)] · Pr[E(i − 1)]

where Ā denotes the complement of event A. We can use the above equation induc-
tively obtain Q
n−t
Pr[E(i)] ≥ i=1 (1 − 2/ni−1
)
Qn−t 2
= i=1 1 − n−i+1
t(t−1)
≥ n(n−1)

Claim 9.4 The probability that a specific mincut C survives at the end of Partition(t)
t(t−1)
is at least n(n−1) .
4
We will prove it only for the unweighted version but the proof can be extended using multiset
arguments.

91
Therefore Partition (2) produces a mincut with probability Ω( n12 ). Repeating the
above algorithm O(n2 ) times would ensure that the min cut is expected to be the
output at least once. If each contraction can be performed in t(n) time then the
expected running time is O(t(n) · n · n2 ).

Exercise 9.10 By using an adjacency matrix representation, show that the contrac-
tion operation can be performed in O(n) steps.

We now address the problem of choosing a random edge using the above data struc-
ture.

Claim 9.5 An edge E can chosen uniformly at random at any stage of the algorithm
in O(n) steps.

We first present a method to pick an edge randomly in O(n) time. The selection
works as follows.

• Select a vertex v at random with probability = P deg(v) = deg(v)

u∈V deg(u) 2|E|

• Select an edge (v, w) at random with probability = P #E(v,w) = #E(v,w)

z∈N(v) #E(v,z) deg(v)

where #E(u, v) denotes the number of edges between u and v and N(v) is the set of
neighbours of v.
Hence, the probability of choosing any edge (v, w) is given by

#E(v, w) deg(v) #E(w, v) deg(w)

= · +
deg(v) 2|E| deg(w) 2|E|
#E(v, w)
=
|E|

Thus, the above method picks edges with probability that is proportional to the
number of edges between v and w. When there are no multiple edges, all edges are
equally likely to be picked. For the case of integer weights, the above derivation works
directly for weighted sampling. By using an adjacency matrix M for storing the graph
where Mv,w denotes the number of edges between v and w allows us to merge vertices
v and w in O(n) time.

Exercise 9.11 Describe a method to implement Partition(2) in O(m log n) steps.

This will be faster for sparse graphs.

Hint: Can you use union-find ?

92
Chapter 10

NP Completeness and
Approximation Algorithms

Let C() be a class of problems defined by some property. We are interested in charac-
terizing the hardest problems in the class, so that if we can find an efficient algorithm
for these, it would imply fast algorithms for all the problems in C. The class that is
of great interest to computer scientists is the class P that is the set of problems for
which we can design polynomial time algorithms. A related class is N P, the class of
problems for which non-deterministic1 polynomial time algorithms can be designed.
More formally,
P = ∪i≥1 C(T P (ni ))
where C(T P (ni )) denotes problems for which O(ni ) time algorithms can be designed.

N P = ∪i≥1 C(T N P (ni ))

where T N P () represent non-deterministic time. Below we formalize the notion of

hardest problems and what is known about the hardest problems. It may be noted
that the theory developed in the context of P and N P is mostly confined to decision
problems, i.e., those that have a Yes/No answer. So we can think about a problem
P as a subset of integers as all inputs can be mapped to integers and hence we are
solving the membership problem for a given set.

Exercise 10.1 Prove the following

(i) If P ∈ P then complement of P is also in P.
(ii) If P1 , P2 P then P1 ∪ P2 ∈ P and P1 ∩ P2 ∈ P.
1
We will define it more formally later. These algorithms have a choice of more than one possible
transitions at any step that does not depend on any deterministic factor.

93
10.1 Classes and reducibility
The intuitive notion of reducibility between two problems is that if we can solve one
we can also solve the other. Reducibility is actually an asymmetric relation and also
entails some details about the cost of reduction. We will use the notation P1 ≤R P2
to denote that problem P1 is reducible to P2 using resource (time or space as the case
may be) to problem P2 . Note that it is not necessary that P2 ≤R P1 .
In the context of decision problems, a problem P1 is many-one reducible
to P2 if there is a many-to-one function g() that maps an instance I1 ∈ P1
to an instance I2 ∈ P2 such that the answer to I2 is YES iff the answer to
I1 is YES.
In other words, the many-to-one reducibility maps YES instances to YES instances
and NO instances to NO instances. Note that the mapping need not be 1-1 and
therefore reducibility is not a symmetric relation.
Further, if the mapping function g() can be computed in polynomial time
then we say that P1 is polynomial-time reducible to P2 and is denoted by
P1 ≤poly P2 .
The other important kind of reduction is logspace reduction and is denoted by

P1 ≤log P2 .

Claim 10.1 If P1 ≤log P2 then P1 ≤poly P2 .

This follows from a more general result that any finite computational process that
uses space S has a running time bounded by 2S . A rigorous proof is not difficult but
beyond the scope of this discussion.
Claim 10.2 The relation ≤poly is transitive, i.e., if P1 ≤poly P2 and P2 ≤poly P3 then
P1 ≤poly P3 .
From the first assertion there must exist polynomial time computable reduction func-
tions, say g() and g ′ () corresponding to the first and second reductions. So we can
define a function g ′ (g)) which is a composition of the two functions and we claim that
it satisfies the property of a polynomial time reduction function from P1 to P3 . Let
x be an input to P1 , then g(x) ∈ P2 2 iff x ∈ P1 . Similarly g ′ (g(x)) ∈ P3 iff g(x) ∈ P2
implying g ′(g(x)) ∈ P3 iff x ∈ P1 . Moreover the composition of two polynomials is a
polynomial, so g ′(g(x)) is polynomial time computable.
A similar result on transitivity also holds for log-space reduction, although the
proof is more subtle.
2
This is a short form of saying that g(x) is an YES instance.

94
Claim 10.3 If Π1 ≤poly Π2 then
(i) If there is a polynomial time algorithm for Π2 then there is a polynomial time
algorithm for Π1 .
(ii) If there is no polynomial time algorithm for Π1 , then there cannot be a polynomial
time algorithm for Π2 .

Part (ii) is easily proved by contradiction. For part (i), if p1 (n) is the running time
of Π1 and p2 is the time of the reduction function, then there is an algorithm for P i1
that takes p1 (p2 (n)) steps where n is the input length for Π1 .
A problem Π is called NP-hard under polynomial reduction if for any problem
Π ∈ N P, Π′ ≤poly Π.
′

A problem is Π is NP-complete (NPC) if it is NP-hard and Π ∈ N P.

Therefore these are problems that are hardest within the class N P.
Exercise 10.2 If problems A and B are NPC, then A ≤poly B and B ≤poly A.
From the previous exercise, these problems form a kind of equivalent class with re-
spect to polynomial time reductions. However, a crucial question that emerges at this
juncture is : Do NPC problems actually exist ?. A positive answer to this question
led to the development of one of the most fascinating areas of Theoretical Computer
Science and will be addressed in the next section.
So far, we have only discussed many-one reducibility that hinges on the existence
of a many-one polynomial time reduction function. There is another very useful
and perhaps more intuitive notion of reducibility, namely, Turing reducibility. The
many-to-one reduction may be thought of as using one subroutine call of P2 to solve
P1 (when P1 ≤poly P2 ) in polynomial time, if P2 has a polynomial time algorithm.
Clearly, we can afford a polynomial number of subroutine calls to the algorithm for
P2 and still get a polynomial time algorithms for P1 . In other words, we say that P1
is Turing-reducible to P2 if a polynomial time algorithm for P2 implies a polynomial
time algorithm for P1 . Moreover, we do not require that P1 , P2 be decision problems.
Although, this may seem to be the more natural notion of reducibility, we will rely
on the more restrictive definition to derive the results.

10.2 Cook Levin theorem

Given a boolean formula in boolean variables, the satisfiability problem is an assign-
ment of the truth values of the boolean variables that can make the formula evaluate
to TRUE (if it is possible). If the formula is in a conjunctive normal form (CNF) 3 ,
3
A formula, that looks like (x1 ∨ x2 ..) ∧ (xi ∨ xj ∨ ..) ∧ . . . (xℓ ∨ . . . xn )

95
then the problem is known as CNF Satisfiability. Further, if we restrict the number
of variables in each clause to be exactly k then it is known as the k-CNF Satisfiability
problem. A remarkable result attributed to Cook and Levin says the following
Theorem 10.1 The CNF Satisfiability problem is NP Complete under polynomial
time reductions.
To appreciate this result, you must realize that there are potentially infinite num-
ber of problems in the class N P, so we cannot explicitly design a reduction function.
Other than the definition of N P we have very little to rely on for a proof of the above
result. A detailed technical proof requires that we define the computing model very
precisely - it is beyond the scope of this discussion. Instead we sketch an intuition
behind the proof.
Given an arbitrary problem Π ∈ N P, we want to show that Π ≤poly CNF −SAT .
In other words, given any instance of Π, say IΠ , we would like to define a boolean
formula B(IΠ ) which has a satisfiable assignment iff IΠ is a YES instance. Moreover
the length of B(IΠ ) should be polynomial time constructable (as a function of the
length of IΠ ).
A computing machine is a transition system where we have
(i) An initial configuration
(ii) A final configuration that indicates whether or not the input is a YES
or a NO instance
(iii) A sequence of intermediate configuration Si where Si+1 follows from Si
using a valid transition. In a non-deterministic system, there can be more
than one possible transition from a configuration. A non-deterministic
machine accepts a given input iff there is some valid sequence of configu-
rations that verifies that the input is a YES instance.
All the above properties can be expressed in propositional logic, i.e., by an unquan-
tified boolean formula in a CNF. Using the fact that the number of transitions is
polynomial, we can bound the size of this formula by a polynomial. The details can
be quite messy and the interested reader can consult a formal proof in the context
of Turing Machine model. Just to give the reader a glimpse of the kind of formalism
used, consider a situation where we want to write a propositional formula to assert
that a machine is in exactly one of the k states at any given time 1 ≤ i ≤ T . Let
us use boolean variables x1,i , x2,i . . . xk,i where xj,i = 1 iff the machine is in state j at
time i. We must write a formula that will be a conjunction of two two conditions
(i) At least one variable is true at any time i:

(x1,i ∨ x2,i . . . xk,i )

96
 (ii) At most one variable is true :

(x1,i ⇒ x̄2,i ∧x̄3,i . . .∧x̄k,i )∧(x2,i ⇒ x̄1,i ∧x̄3,i . . .∧x̄k,i ) . . .∧(xk,i ⇒ x̄1,i ∧x̄2,i . . .∧x̄k−1,i )

where the implication a ⇒ b is equivalent to ā ∨ b.

A conjunction of the above formula over all 1 ≤ i ≤ T has a satisfiable assignment

of xj,i iff the machine is in exactly one state (not necessarily the same state) at each
of the time instances. The other condition should capture which states can succeed
a given state.
We have argued that CNF − SAT is NP-hard. Since we can guess an assignment
and verify the truth value of the Boolean formula, in linear time, we can claim that
CNF − SAT is in N P.

10.3 Common NP complete problems

To prove that a given problem P is NPC, the standard procedure is to establish that

(i) P ∈ N P : This is usually the easier part.

(ii) CNF − SAT ≤poly P . We already know that any P ′ ∈ N P, P ′ ≤poly
CNF − SAT . So by transitivity, P ′ ≤poly P and therefore P is NPC.

The second step can be served by reducing any known NPC to P . Some of the earliest
problems that were proved NPC include (besides CNF-SAT)

• 3D Matching

• Three colouring of graphs

• Equal partition of integers

• Maximum Clique /Independent Set

• Hamilton cycle problem

• Minimum set cover

10.3.1 Other important complexity classes

While the classes P and N P hogs the maximum limelight in complexity theory, there
are many other related classes in their own right.

97
• co − N P A problem whose complement is in N P belongs to this class. If the
problem is in P, then the complement of the problem is also in P and hence in
N P. In general we can’t say much about the relation between P and co − N P.
In general, we can’t even design an NP algorithm for a problem in co − N P, i.e.
these problems are not efficiently verifiable. For instance how would you verify
that a boolean formula is unsatisfiable (all assignments make it false) ?

Exercise 10.3 Show that the complement of an NPC problem is complete for
the class co − N P under polynomial time reduction.

Exercise 10.4 What would it imply if an NPC problem and its complement
are polynomial time reducible to eachother ?

• PSPACE The problems that run in polynomial space (but not necessarily poly-
nomial time). The satisfiability of Quantified Boolean Formula (QBF) is a com-
plete problem for this class.

• Randomized classes Depending on the type of randomized algorithms (mainly

Las Vegas or Monte Carlo) , we have the following important classes

– RP : Randomized Polynomial class of problems are characterized by

(Monte Carlo) randomized algorithms A such that
If x ∈ L ⇒ Pr[A accepts x] ≥ 1/2
If x ∈
/ L ⇒ Pr[A accepts x] = 0
These algorithms can err on one side.
– BPP When a randomized algorithm is allowed to err on both sides
If x ∈ L ⇒ Pr[A accepts x] ≥ 1/2 + ε
If x ∈
/ L ⇒ Pr[A accepts x] ≤ 1/2 − ε
where ε is a fixed non zero constant.
– ZPP Zero Error Probabilistic Polynomial Time : These are the Las Vegas
kind that do not have any errors in the answer but the running time is
expected polynomial time.

One of the celebrated problems, involving randomized algorithms is

BPP ⊂ N P?

98
10.4 Combating hardness with approximation
Since the discovery of NPC problems in early 70’s , algorithm designers have been
wary of spending efforts on designing algorithms for these problems as it is considered
to be a rather hopeless situation without a definite resolution of the P = N P question.
Unfortunately, a large number of interesting problems fall under this category and
so ignoring these problems is also not an acceptable attitude. Many researchers have
pursued non-exact methods based on heuristics to tackle these problems based on
heuristics and empirical results 4 . Some of the well known heuristics are simulated
annealing, neural network based learning methods , genetic algorithms. You will have
to be an optimist to use these techniques for any critical application.
The accepted paradigm over the last decade has been to design polynomial time
algorithms that guarantee near-optimal solution to an optimization problem. For a
maximization problem, we would like to obtain a solution that is at least f · OP T
where OP T is the value of the optimal solution and f ≤ 1 is the approximation factor
for the worst case input. Likewise, for minimization problem we would like a solution
no more than a factor f ≥ 1 larger than OP T . Clearly the closer f is to 1, the better
is the algorithm. Such algorithm are referred to as Approximation algorithms and
there exists a complexity theory of approximation. It is mainly about the extent of
approximation attainable for a certain problem.
For example, if f = 1 + ε where ε is any user defined constant, then we way
that the problem has a Polynomial Time Approximable Scheme (PTAS). Further,
if the algorithm is polynomial in 1/ε then it is called FPTAS (Fully PTAS). The
theory of hardness of approximation has yielded lower bounds (for minimization and
upper bounds for maximization problems) on the approximations factors for many
important optimization problems. A typical kind of result is that Unless P = N P
we cannot approximate the set cover problem better than log n in polynomial time.
In this section, we give several illustrative approximation algorithms. One of the
main challenges in the analysis is that even without the explicit knowledge of the
optimum solutions, we can still prove guarantees about the quality of the solution of
the algorithm.

10.4.1 Equal partition

Given nPintegers = {z1 , z2 .
. . zn }, we want to find a partition S1 , S − S1 , such

S P
that | x∈S1 x − x∈S−S1 x | is minimized. A partition is balanced if the above
difference is zero.
4
The reader must realize that our inability to compute the actual solutions makes it difficult to
evaluate these methods in in a general situation.

99
 P
Let B = i zi and consider the following a generalization of the problem, namely,
the subset sum problem. For a given integer K ≤ B, is there a subset R ⊂ S such
that the elements in R sum up to K.
Let S(j, r) denote a subset of {z1 , z2 . . . zj } that sums to r - if no such subset exists
then we define it as φ (empty subset). We can write the following recurrence

S(j, r) = S(j−1, r−zj )∪zj if zj is included or S(j−1, r) if zj is not included or φ not possible

Using the above dynamic programming formulation we can compute S(j, r) for
1 ≤ j ≤ n and r ≤ B. You can easily argue that the running time is O(n · B) which
may not be polynomial as B can be very large.
Suppose, we are given an approximation factor ε and let A = ⌈ nε ⌉ so that A1 ≤ ε/n.
zi
Then we define a new scaled problem with the integers scaled as zi′ = ⌊ z/A ⌋ and let
′ r
r = ⌊ z/A ⌋ where z is the maximum value of an integer that can participate in the
solution 5 .
Let us solve the problem for {z1′ , z2′ . . . zn′ } and r ′ using the previous dynamic
programming strategy and let So′ denote the optimal solution for the scaled problem
and let So be the solution for the original problem. Further let C and C ′ denote the
cost function for the original and the scaled problems respectively. The running time
of the algorithm is O(n · r ′ ) which is O( 1ε n2 ). We would like to show that the cost of
C(So′ ) is ≥ (1 − ε)C(So ). For any S ′′ ⊂ S
n n
C(S ′′ ) · ≥ C ′ (S ′′ ) ≥ C(S ′′ ) · − |S ′′ |
εz εz
So
εz εz  n  εz εz
C(So′ ) ≥ C ′ (So′ ) ≥ C ′ (So ) ≥ C(So ) − |So | = C(So ) − |So |
n n εz n n
The first and the third inequality follows from the previous bound and the second
inequality follows from the optimality of So′ wrt C ′ . Since C(So ) ≥ z|Sno | and so
C(So′ ) ≥ (1 − ε)C(So )

10.4.2 Greedy set cover

Given a ground set S = {x1 , x2 . . . xn } and a family of subsets S1 , S2 . . . Sm Si ⊂ S,
we want to find a minimum number of subsets from the family that covers all elements
of S. If Si have associated weights C(), then we try to minimize the total weight of
the set-cover.
5
It is a lower bound to the optimal solution - for the balanced partition, it is the maximum
integer less than B/2.

100
 In the greedy algorithm, we pick up a subset that is most cost-effective in terms
of the cost per unchosen element. The cost-effectiveness of a set U is defined by UC(U
−V
)

where V ⊂ S is the set of elements already covered. We do this repeatedly till all
elements are covered.
Let us number the elements of S in the order they were covered by the greedy
algorithm (wlog, we can renumber such that they are x1 , x2 . . .). We will apportion
C(U )
the cost of covering an element e ∈ S as w(e)
P= U −V where e is covered for the first
time by U. The total cost of the cover is = i w(xi ).

Claim 10.4
Co
w(xi ) ≤
n−i+1
where Co is the cost of an optimum cover.

In iteration i, the greedy choice is more cost effective than any left over set of the opti-
′ ′
mal cover. Suppose n the cost-effectivenessoof the best set in the optimal cover is C /U ,
C(S ) C(S ) C(S )
i.e. C ′ /U ′ = min Si −Si1 i2
, Si −S ik
. . . Si −S where Si1 , Si2 . . . Sik forms a minimum set
1 2 k
cover. It follows that
C(Si1 ) + C(Si1 ) + . . . C(Si1 ) Co
C ′ /U ′ ≤ ≤
(Si1 − S) + (Si2 − S) + . . . (Sik − S) n−i+1
Co
So w(xi ) ≤ n−i+1 .
Co
P
Thus the cost of the greedy cover is i n−i+1
which is bounded by Co · Hn . Here
1
Hn = n1 + n−1 + . . . 1.

Exercise 10.5 Formulate the Vertex cover problem as an instance of set cover prob-
lem.
Analyze the approximation factor achieved by the following algorithm. Construct a
maximal matching of the given graph and consider the union C of the end-points of
the matched edges. Prove that C is a vertex cover and the size of the optimal cover
is at least C/2. So the approximation factor achieved is better than the general set
cover.

10.4.3 The metric TSP problem

If the edges of the graph satisfies triangle inequality, i.e., for any three vertices
u, v, w C(u, v) ≤ C(u, w) + C(w, v), then we can design an approximation algo-
rithm for the TSP problem as follows.

101
Metric TSP on graphs
Input: A graph G = (V,E) with weights on edges that satisfy triangle inequality.

1. Find a Minimum Spanning Tree T of G.

2. Double every edge - call the resulting graph E ′ and construct an Euler tour T .

3. In this tour, try to take shortcuts if we have visited a vertex before.

Claim 10.5 The length of this tour no more than twice that of the optimal tour.

MST ≤ T SP , therefore 2 · MST ≤ 2 · T SP . Since shortcuts can only decrease the

tour length (because of the triangle inequality), the tour length is no more than twice
that of the optimal tour.

10.4.4 Three colouring

We will rely on the following simple observation. If a graph is three colourable, its
neighbourhood must be triangle free (or else the graph will contain a four-clique) i,e,,
it must be 2 colourable,

Observation 10.1 A graph that has maximum degree ∆ can be coloured using ∆ + 1
colours using a greedy strategy.

Use a colour that is different from its already coloured neighbours.

√Given a 3-colourable graph G = (V, E), separate out vertices that have degrees
≥ n - call this set H. Remove the set H and its incident edges and denote this
graph√by G′ = (V ′ , E ′ ). Note that G′ is 3-colourable and all vertices have degrees
√ less
than n and so from our previous observation, we can easily colour using n + 1
colours. Now, reinsert the the√ vertices of H and use an extra
√ |H| colours to complete
the colouring. Since |H| ≤ n, we have used at most 2 n colours.
It is a rather poor approximation since we have used significantly more colours
than three. However, it is known that unless P = N P, any polynomial time colouring
algorithm will use Ω(nǫ ) colours for some fixed ǫ > 0.

10.4.5 Maxcut
Problem Given a graph G = (V, E), we want to partition the vertices into sets U, V −
U such that the number of edges across U and V − U is maximized. There is a

102
corresponding weighted version for a weighted graph with a weight function w : E →
R.
We have designed a polynomial time algorithm for mincut but the maxcut is an
NP-hard problem. Let us explore a simple idea of randomly assigning the vertices to
one of the partitions. For any fixed edge (u, v) ∈ E, it either belongs to the optimal
maxcut or not depending on whether u, v belong to different partitions of the optimal
maxcut Mo . The probability that we have chosen the the right partitions is at least
half. Let Xe be a random 0-1 variable (also called indicator random variables) that
is 1 iff the algorithm choses it consistently with the maxcut. The expected size of the
cut produced by the algorithm is
X X
E[ w(e) · Xe ] = w(e) · E[Xe ] ≥ Mo /2
e e

1
Therefore we have a simple randomized algorithm that attains a 2
apprximation.

Exercise 10.6 For an unweighted graph show that a simple greedy strategy leads to
a 12 approximation algorithm.

103
Appendix A

Recurrences and generating

functions

Given a sequence a1 , a2 . . . an (i.e. a function with the domain as integers), a compact

way of representing it is an equation in terms of itself, a recurrence relation. One of
the most common examples is the Fibonacci sequence specified as an = an−1 + an−2
for n ≥ 2 and a0 = 0, a1 = 1. The values a0 , a1 are known as the boundary conditions.
Given this and the recurrence, we can compute the sequence step by step, or better
still we can write a computer program. Sometimes, we would like to find the general
term of the sequence. Very often, the running time of an algorithm is expressed as
a recurrence and we would like to know the explicit function for the running time to
make any predictions and comparisons. A typical recurrence arising from a divide-
and-conquer algorithm is
a2n = 2an + cn
which has a solution an ≤ 2cn⌈log2 n⌉. In the context of algorithm analysis, we are
often satisfied with an upper-bound. However, to the extent possible, it is desirable
to obtain an exact expression.
Unfortunately, there is no general method for solving all recurrence relations. In
this chapter, we discuss solutions to some important classes of recurrence equations.
In the second part we discuss an important technique based on generating functions
which are also important in their own right.

A.1 An iterative method - summation

As starters, some of the recurrence relations can be solved by summation or guessing
and verifying by induction.

104
Example A.1 The number of moves required to solve the Tower of Hanoi problem
with n discs can be written as
an = 2an−1 + 1
By substituting for an−1 this becomes

an = 22 an−2 + 2 + 1

By expanding this till a1 , we obtain

an = 2n−1 a1 + 2n−2 + . . . .. + 1

This gives an = 2n − 1 by using the formula for geometric series and a1 = 1.

Example A.2 For the recurrence

a2n = 2an + cn

we can use the same technique to show that a2n = i=0 log2 n2i n/2i · c + 2na1 .
P
Remark We made an assumption that n is a power of 2. In the general case, this may
present some technical complication but the nature of the answer remains unchanged.
Consider the recurrence
T (n) = 2T (⌊n/2⌋) + n
Suppose T (x) = cx log2 x for some constant c > 0 for all x < n. Then T (n) =
2c⌊n/2⌋ log2 ⌊n/2⌋+n. Then T (n) ≤ cn log2 (n/2)+n ≤ cn log2 n−(cn)+n ≤ cn log2 n
for c ≥ 1.

A very frequent recurrence equation that comes up in the context of divide-and-

conquer algorithms (like mergesort) has the form

T (n) = aT (n/b) + f (n) a, b are constants and f (n) a positive monotonic function

Theorem A.1 For the following different cases, the above recurrence has the follow-
ing solutions

• If f (n) = O(nlogb a−ǫ ) for some constant ǫ, then T (n) is Θ(nlogb a ).

• If f (n) = O(nlogb a ) then T (n) is Θ(nlogb a log n).

• If f (n) = O(nlogb a+ǫ ) for some constant ǫ, and if af (n/b) is O(f (n)) then T (n)
is Θ(f (n)).

105
Example A.3 What is the maximum number of regions induced by n lines in the
plane ? If we let Ln represent the number of regions, then we can write the following
recurrence
Ln ≤ Ln−1 + n L0 = 1
n(n+1)
Again by the method of summation, we can arrive at the answer Ln = 2
+ 1.

Example A.4 Let us try to solve the recurrence for Fibonacci, namely

Fn = Fn−1 + Fn−2 F0 = 0, F1 = 1

If we try to expand this in the way that we have done previously, it becomes unwieldy
very quickly. Instead we ”guess” the following solution
1
Fn = √ φn − φ̄n


5
√ √
where φ = (1+2 5) and φ̄ = (1−2 5) . The above solution can be verified by induction. Of
course it is far from clear how one can magically guess the right solution. We shall
address this later in the chapter.

A.2 Linear recurrence equations

A recurrence of the form

c0 ar + c1 ar−1 + c2 ar−2 . . . + ck ar−k = f (r)

where ci are constants is called a linear recurrence equation of order k. Most of

the above examples fall under this class. If f (r) = 0 then it is homogeneous linear
recurrence.

A.2.1 Homogeneous equations

We will first outline the solution for the homogeneous class and then extend it to the
general linear recurrence. Let us first determine the number of solutions. It appears
that we must know the values of a1 , a2 . . . ak to compute the values of the sequence
according to the recurrence. In absence of this there can be different solutions based
on different boundary conditions. Given the k boundary conditions, we can uniquely
determine the values of the sequence. Note that this is not true for a non-linear
recurrence like
ar 2 + ar−1 = 5 with a0 = 1

106
This observation (of unique solution) makes it somewhat easier for us to guess some
solution and verify.
Let us guess a solution of the form ar = Aαr where A is some constant. This may
be justified from the solution of Example A.1. By substituting this in the homoge-
neous linear recurrence and simplification, we obtain the following equation

c0 αk + c1 αk−1 . . . + ck = 0

This is called the characteristic equation of the recurrence relation and this degree
k equation has k roots, say α1 , α2 . . . αk . If these are all distinct then the following is
a solution to the recurrence

ar (h) = A1 α1r + A2 α2r + . . . Ak αkr

which is also called the homogeneous solution to linear recurrence. The values of
A1 , A2 . . . Ak can be determined from the k boundary conditions (by solving k simul-
taneous equations).
When the roots are not unique, i.e. some roots have multiplicity then for mul-
tiplicity m, αn , nαn , n2 αn . . . nm−1 αn are the associated solutions. This follows from
the fact that if α is a multiple root of the characteristic equation, then it is also the
root of the derivative of the equation.

A.2.2 Inhomogeneous equations

If f (n) 6= 0, then there is no general methodology. Solutions are known for some
(h)
particular cases, known as particular solutions. Let an be the solution by ignoring
(p) (h) (p)
f (n) and let an be a particular solution then it can be verified that an = an + an
is a solution to the non-homogeneous recurrence.
The following is a table of some particular solutions

d a constant B
dn B1 n + B0
dn2 B2 n2 + B1 n + B0
edn , e, d are constants Bdn

Here B, B0 , B1 , B2 are constants to be determined from initial conditions. When

f (n) = f1 (n) + f2 (n) is a sum of the above functions then we solve the equation
for f1 (n) and f2 (n) separately and then add them in the end to obtain a particular
solution for the f(n).

107
A.3 Generating functions
An alternative representation for a sequence a1 , a2 . . . ai is a polynomial function
a1 x+a2 x2 +. . . ai xi . Polynomials are very useful objects in mathematics, in particular
as ”placeholders.” For example if we know that two polynomials are equal (i.e. they
evaluate to the same value for all x), then all the corresponding coefficients must
be equal. This follows from the well known property that a degree d polynomial
has no more than d distinct roots (unless it is the zero polynomial). The issue of
convergence is not important at this stage but will be relevant when we use the
method of differentiation.
Example A.5 Consider the problem of changing a Rs 100 note using notes of the
following denomination - 50, 20, 10, 5 and 1. Suppose we have an infinite supply of
each denomination then we can represent each of these using the following polynomials
where the coefficient corresponding to xi is non-zero if we can obtain a certain sum
using the given denomination.
P1 (x) = x0 + x1 + x2 + . . .
P5 (x) = x0 + x5 + x10 + x15 + . . .
P10 (x) = x0 + x10 + x20 + x30 + . . .
P20 (x) = x0 + x20 + x40 + x60 + . . .
P50 (x) = x0 + x50 + x100 + x150 + . . .
For example, we cannot have 51 to 99 using Rs 50,so all those coefficients are zero.
By multiplying these polynomials we obtain
P (x) = E0 + E1 x + E2 x2 + . . . E100 x100 + . . . Ei xi
where Ei is the number of ways the terms of the polynomials can combine such that
the sum of the exponents is 100. Convince yourself that this is precisely what we are
looking for. However we must still obtain a formula for E100 or more generally Ei ,
which the number of ways of changing a sum of i.
Note that for the polynomials P1 , P5 . . . P50 , the following holds

Pk (1 − xk ) = 1 for k = 1, 5, ..50 giving

1
P (x) =
(1 − x)(1 − x5 )(1 − x10 )(1 − x20 )(1 − x50 )
1 5
1−x
We can now use the observations that 1−x = 1 + x2 + x3 . . . and (1−x)(1−x 5) =
2 3
1 + x + x . . .. So the corresponding coefficients are related by Bn = An + Bn−5 where
1 1
A and B are the coefficients of the polynomials 1−x and (1−x)(1−x 5 ) . Since An = 1,

this is a linear recurrence. Find the final answer by extending these observations.

108
Let us try the method of generating function on the Fibonacci sequence.
Example A.6 Let the generating function be G(z) = F0 + F1 x + F2 x2 . . . Fn xn where
Fi is the i-th Fibonacci number. Then G(z) − zG(z) − z 2 G(z) can be written as the
infinite sequence
F0 + (F1 − F2 )z + (F2 − F1 − F0 )z 2 + . . . (Fi+2 − Fi+1 − Fi )z i+2 + . . . = z
z
for F0 = 0, F1 = 1. Therefore G(z) = .
This can be worked out to be
1−z−z 2
 
1 1 1
G(z) = √ −
5 1 − φz 1 − φ̄z
1
√

where φ̄ = 1 − φ = 2
1− 5 .

A.3.1 Binomial theorem

The use of generating functions necessitates computation of the coefficients of power
series of the form (1 + x)α for |x| < 1 and any α. For that the following result is very
useful - the coefficient of xk is given by
α · (α − 1) . . . (α − k + 1)
C(α, k) =
k · (k − 1) . . . 1
This can be seen from an application of Taylor’s series. Let f (x) = (1 + x)α . Then
from Taylor’s theorem, expanding around 0 for some z,
f ′′ (0) f (k) (0)
f (z) = f (0) + zf ′ (0) + α · z + z 2 + . . . zk ...
2! k!
α(α − 1)
= f (0) + 1 + z 2 + . . . C(α, k) + . . .
2!
Therefore (1 + z)α = ∞ i
P
i=0 C(α, i)z which is known as the binomial theorem.

A.4 Exponential generating functions

If the terms of a sequence is growing too rapidly, i.e. the n-th term exceeds xn for
any 0 < x < 1, then it may not converge. It is known that a sequence converges iff
the sequence |an |1/n is bounded. Then it makes sense to divide the coefficients by a
rapidly growing function like n!. For example, if we consider the generating function
for the number of permutations of n identical objects
p1 p2 pi
G(z) = 1 + z + z 2 . . . z i
1! 2! i!
109
where pi = P (i, i). Then G(z) = ez . The number of permutations of r objects when
selected out of (an infinite supply of) n kinds of objects is given by the exponential
generating function (EGF)
∞
 p1 p2 2  n nx
X nr r
1 + z + z ... = e = x
1! 2! r=0
r!

Example A.7 Let Dn denote the number of derangements of n objects. Then it can
be shown that Dn = (n − 1)(Dn−1 + Dn−2 ). This can be rewritten as Dn − nDn−1 =
−(Dn−1 − (n − 2)Dn−2. Iterating this, we obtain Dn − nDn−1 = (−1)n−2 (D2 − 2D1 ).
Using D2 = 1, D1 = 0, we obtain

Dn − nDn−1 = (−1)n−2 = (−1)n .

xn
Multiplying both sides by n!
, and summing from n = 2 to ∞, we obtain
∞ ∞ ∞
X Dn n
X nDn−1 n
X (−1)n
x − x = xn
n=2
n! n=2
n! n=2
n!

If we let D(x) represent the exponential generating function for derangements, after
simplification, we get

D(x) − D1 x − D0 − x(D(x) − D0 ) = e−x − (1 − x)

e−x
or D(x) = 1−x
.

A.5 Recurrences with two variables

For selecting r out of n distinct objects, we can write the familiar recurrence

C(n, r) = C(n − 1, r − 1) + C(n − 1, r)

with boundary conditions C(n, 0) = 1 and C(n, 1) = n.

The general form of a linear recurrence with constant coefficients that has two
indices is

Cn,r an,r + Cn,r−1 an,r−1 + . . . Cn−k,r an−k,r . . . C0,r a0,r + . . . = f (n, r)

where Ci,j are constants. We will use the technique of generating functions to extend
the one variable method. Let

A0 (x) = a0,0 + a0,1 x + . . . a0,r xr

110
 A1 (x) = a1,0 + a1,1 x + . . . a1,r xr
An (x) = an,0 + an,1 x + . . . an,r xr
Then we can define a generating function with A0 (x), A1 (x)A3 (x) . . . as the sequence
- the new indeterminate can be chosen as y.

Ay (x) = A0 (x) + A1 (x)y + A2 (x)y 2 . . . An (x)y n

For the above example, we have

Fn (x) = C(n, 0) + C(n, 1)x + C(n, 2)x2 + . . . C(n, r)xr + . . .

∞
X ∞
X ∞
X
C(n, r)xr = C(n − 1, r − 1)xr + C(n − 1, r)xr
r=0 r=1 r=0

Fn (x) − C(n, 0) = xFn−1 (x) + Fn−1 (x) − C(n − 1, 0)

Fn (x) = (1 + x)Fn−1 (x)
or Fn (x) = (1 + x) C(0, 0) = (1 + x)n as expected.
n

111
Appendix B

Refresher in discrete probability

and probabilistic inequalities

The sample space Ω may be infinite with infinite elements that are called elementary
events. For example consider the experiment where we must toss a coin until a head
comes up for the first time. A probability space consists of a sample space with a
probability measure associated with the elementary events. The probability measure
Pr is a real valued function on events of the sample space and satisfies the following

1. For all A ⊂ Ω , 0 ≤ Pr[A] ≤ 1

2. Pr[Ω] = 1
P
3. For mutually disjoint events E1 , E2 . . . , Pr[∪i Ei ] = i Pr[Ei ]

Sometimes we are only interested in a certain collection of events (rather the entire
sample space)a, say F . If F is closed under union and complementation, then the
above properties can be modified in a way as if F = Ω.
The principle of Inclusion-Exclusion has its counterpart in the probabilistic world,
namely

Lemma B.1
X X X
Pr[∪i Ei ] = Pr[Ei ] − Pr[Ei ∩ Ej ] + Pr[Ei ∩ Ej ∩ Ek ] . . .
i i<j i<j<k

Definition B.1 A random variable (r.v.) X is a real-valued function over the sample
space, X : Ω → R. A discrete random variable is a random variable whose range is
finite or a countable finite subset of R.
The distribution function FX : R → (0, 1] for a random variable X is defined as
FX (x) ≤ Pr[X = x]. The probability density function of a discrete r.v. X, fX is

112
given by fX (x) = Pr[X = x]. P
The expectation of a r.v. X, denoted by E[X] = x x · Pr[X = x].

A very useful property of expectation, called the linearity property can be stated as
follows
Lemma B.2 If X and Y are random variables, then

E[X + Y ] = E[X] + E[Y ]

Remark Note that X and Y do not have to be independent !

Definition B.2 The conditional probability of E1 given E2 is denoted by Pr[E1 |E2 ]

and is given by
Pr[E1 ∩ E2 ]
Pr[E2 ]
assuming Pr[E2 ] > 0.

Definition B.3 A collection of events {Ei |i ∈ I} is independent if for all subsets

S⊂I
Pr[∩i∈S Ei ] = Πi∈S Pr[Ei ]

Remark E1 and E2 are independent if Pr[E1 |E2 ] = Pr[E2 ].

The conditional probability of a random variable X with respect to another random
variable Y is denoted by Pr[X = x|Y = y] is similar to the previous definition with
events E1 , E2 as X = x and Y = y respectively. The conditional expectation is defined
as X
E[X|Y = y] = Pr x · [X = x|Y = y]
x

The theorem of total expectation that can be proved easily states that
X
E[X] = E[X|Y = y]
y

B.1 Probability generating functions

The notion of generating functions have useful applications in the context of proba-
bility calculations also. Given a non-negative integer-valued discrete random variable
X with Pr[X = k] = pk , the probability generating function (PGF) of X is given by
∞
X
GX (z) = pi z i = p0 + p1 z + . . . pi z i . . .
i=0

113
This
P is also known as the z-transform of X and it is easily seen that GX (1) = 1 =
i pi . The convergence of the PGF is an important issue for some calculations
involving differentiation of the PGF. For example,

dGX (z)
E[X] = |z = 1
dz
The notion of expectation of random variable can be extended to function f (X)
of random variable X in the following way
X
E[f (X)] = pi f (X = i)
i

Therefore, PGF of X is the same as E[z X ]. A particularly useful quantity for a

number of probabilistic calculations is the Moment Generating Function (MGF)
defined as
MX (θ) = E[eXθ ]
Since
X 2 θ2 X k θk
eXθ = 1 + Xθ + + ... ...
2! k!
E[X k ]θk
MX (θ) = 1 + E[X]θ + . . . ...
k!
from which E[X k ] also known as higher moments can be calculated. There is also
a very useful theorem known for independent random variables Y1 , Y2 . . . Yt . If Y =
Y1 + Y2 + . . . Yt , then

MY (θ) = MY1 (θ) · MY2 (θ) · . . . MYt (θ)

i.e., the MGF of the sum of independent random variables is the product of the
individual MGF’s.

B.1.1 Probabilistic inequalities

In many applications, especially in the analysis of randomized algorithms, we want to
guarantee correctness or running time. Suppose we have a bound on the expectation.
Then the following inequality known as Markov’s inequality can be used.
Markov’s inequality
1
Pr[X ≥ kE[X]] ≤ (B.1.1)
k
Unfortunately there is no symmetric result.

114
 If we have knowledge of the second moment, then the following gives a stronger
result
Chebychev’s inequality

σ2
Pr[(X − E[X])2 ≥ t] ≤ (B.1.2)
t
where σ is the variance, i.e. E 2 [X] − E[X 2 ].
With
Pnknowledge of higher moments, then we have the following inequality. If
X = i xi is the sum of n mutually independent random variables where xi is
uniformly distributed in {-1 , +1 }, then for any δ > 0,
Chernoff bounds
Pr[X ≥ ∆] ≤ e−λ∆ E[eλX ] (B.1.3)
2 −λ λ
If we choose λ = ∆/n, the RHS becomes e−∆ /2n using a result that e 2+e =
2
coshh(λ) ≤ eλ /2 .
A more useful form of the above inequality is for a situation where a random
variable X is the sum of n independent
P 0-1 valued Poisson trials with a success prob-
ability of pi in each trial. If i pi = np, the following equations give us concentration
bounds of deviation of X from the expected value of np. The first equation is more
useful for large deviations whereas the other two are useful for small deviations from
a large expected value.
 np m
P rob(X ≥ m) ≤ em−np (B.1.4)
m
P rob(X ≤ (1 − ǫ)pn) ≤ exp(−ǫ2 np/2) (B.1.5)
P rob(X ≥ (1 + ǫ)np) ≤ exp(−ǫ2 np/3) (B.1.6)
for all 0 < ǫ < 1.
A special case of non-independent random variables P
Consider n 0-1 random variables y1 , y2 , . . . yn such that Pr[yi = 1] ≤ pi and pi =
np. The random variables are not known to be independent. In such a case, we will
not be able to directly invoke the previous Chernoff bounds directly but we will show
the following
P P
Lemma B.3 Let Y = i yi and let X = i xi where xi are independent Poisson
trials with xi = pi . Then

Pr[Y ≥ k] ≤ [X ≥ k]∀k, 0 ≤ k ≤ n

In this case the random variable X is known to stochastically dominate Y.

115
 Therefore we can invoke the Chernoff bounds on X to obtain a bound on Y . We
will prove the above property by induction on i (number of variables). For i = 1, (for
all k) this is true by definition. Suppose this is true for i < t (for all k ≤ i) and let
i = t. Let Xi = x1 + x2 . . . xi and Yi = y1 + y2 . . . yi. Then

Pr[Xt ≥ k] = Pr[Xt−1 ≥ k] + Pr[Xt−1 = k − 1 ∩ xt = 1]

Since xi ’s are independent, we can rewrite the above equation as

Pr[Xt ≥ k] = (pt + 1 − pt ) Pr[Xt−1 ≥ k] + Pr[Xt−1 = k − 1] · pt

= pt (Pr[Xt−1 ≥ k] + Pr[Xt−1 = k − 1]) + (1 − pt ) · Pr[Xt−1 ≥ k]

= pt Pr[Xt−1 ≥ k − 1] + (1 − pt ) · Pr[Xt−1 ≥ k]
Similarly

Pr[Yt ≥ k] ≤ pt (Pr[Yt−1 ≥ k − 1] + (1 − pt ) · Pr[Yt−1 ≥ k]

where the inequality exists because Pr[Yt−1 = k − 1 ∩ yt = 1] = Pr[yt = 1|Yt−1 =

k − 1] · Pr[Yt−1 = k − 1] ≤ pt · Pr[Yt−1 = k − 1]. By comparing the two equations term
by term and invoking induction hypothesis, the result follows.

116