Chapter 5

Algorithm Analysis and Asymptotic

Notation

5.1 Correctness, running time of programs

So far we have been proving statements about databases, mathematics and arithmetic, or sequences of
numbers. Though these types of statements are common in computer science, you'll probably encounter
algorithms most of the time. Often we want to reason about algorithms and even prove things about them.
Wouldn't it be nice to be able to prove that your program is correct? Especially if you're programming a
heart monitor or a NASA spacecraft?
In this chapter we'll introduce a number of tools for dealing with computer algorithms, formalizing their
expression, and techniques for analyzing properties of algorithms, so that we can prove correctness or prove
bounds on the resources that are required.

5.2 Binary (base 2) notation

Let's rst think about numbers. In our everyday life, we write numbers in decimal (base 10) notation
(although I heard of one kid who learned to use the ngers of her left hand to count from 0 to 31 in base
2). In decimal, the sequence of digits 20395 represents (parsing from the right):
5 + 9(10) + 3(100) + 0(1000) + 2(10000) =
5(100 ) + 9(101 ) + 3(102 ) + 0(103 ) + 2(104 )
Each position represents a power of 10, and 10 is called the base. Each position has a digit from [0; 9]
representing how many of that power to add. Why do we use 10? Perhaps due to having 10 ngers
(however, humans at various times have used base 60, base 20, and mixed base 20,18 (Mayans)). In the last
case there were (105)20 18 days in the year. Any integer with absolute value greater than 1 will work (so
;

experiment with base 2).

Consider using 2 as the base for our notation. What digits should we use?1 We don't need digits 2 or
higher, since they are expressed by choosing a dierent position for our digits (just as in base 10, where
there is no single digit for numbers 10 and greater).
Here are some examples of binary numbers:
(10011)2
represents
1(20 ) + 1(21 ) + 0(22 ) + 0(23 ) + 1(24 ) = (19)10

49
Course notes for csc 165 h

We can extend the idea, and imitate the decimal point (with a \binary point"?) from base 10:
(1011:101)2 = 19 58
How did we do that?2 Here are some questions:
 How do you multiply two base 10 numbers?3 Work out 37  43.
 How do you multiply two binary numbers?4
 What does \right shifting" (eliminating the right-most digit) do in base 10?5
 What does \right shifting" do in binary?6
 What does the rightmost digit tell us in base 10? In binary?
Convert some numbers from decimal to binary notation. Try 57. We'd like to represent 57 by adding either
0 or 1 of each power of 2 that is no greater than 57. So 57 = 32 + 16 + 8 + 1 = (111001)2 . We can also ll
in the binary digits, systematically, from the bottom up, using the % operator from Python (the remainder
after division operator, at least for positive arguments):
57 % 2 = 1 so (?????1)2
(57 1)=2 = 28 % 2 = 0 so (????01)2
28=2 = 14 % 2 = 0 so (???001)2
14=2 = 7 % 2 = 1 so (??1001)2
(7 1)=2 = 3 % 2 = 1 so (?11001)2
(3 1)=2 = 1 % 2 = 1 so (111001)2
Addition in binary is the same as (only dierent from. . . ) addition in decimal. Just remember that (1)2 +
(1)2 = (10)2 . If we add two binary numbers, this tells us when to \carry" 1:
1011
+ 1011
10110
5.3 log2
How many 5-digit binary numbers are there (including those with leading 0s)? These numbers run from
(00000) through (11111) , or 0 through 31 in decimal | 32 numbers. Another way to count them is to
2 2
consider that there are two choices for each digit, hence 2 strings of digits. If we add one more digit we get
5

twice as many numbers. Every digit doubles the range of numbers, so there are two 1-digit binary numbers
(0 and 1), four 2-digit binary numbers (0 through 3), 8 3-digit binary numbers (0 through 7), and so on.
Reverse the question: how many digits are required to represent a given number. In other words, what
is the smallest integer power of 2 needed to exceed a given number? log x is the power of 2 that gives
2
2 2 = x. You can think of it as how many times you must multiply 1 by 2 to get x, or roughly the number
log x

of digits in the binary representation of x. (The precise number of digits needed is dlog (x + 1)e | which
2
happens to be equal to b(log x) + 1c for all positive values of x).
2

5.4 Loop invariant for base 2 multiplication

Integers are naturally represented on a computer in binary, since a gate can be in either an on or o (1 or
0) position. It is very easy to multiply or divide by 2, since all we need to do is perform a left or right shift
(an easy hardware operation). Similarly, it is also very easy to determine whether an integer is even or odd.
Putting these together, we can write a multiplication algorithm that uses these fast operations:

50
 Chapter 5. Algorithm Analysis and Asymptotic Notation

def mult(m,n):
""" Multiply integers m and n. """
# Precondition: m >= 0
x = m
y = n
z = 0

# loop invariant: z = mn - xy
while x != 0:
if x % 2 == 1: z = z + y # x is odd
x = x >> 1 # x = x / 2 (right shift)
y = y << 1 # y = y * 2 (left shift)

# post condition: z = mn
return z

After reading this algorithm, there is no reason you should believe it actually multiplies two integers: we'll
need to prove it to you. Let's consider the precondition rst. So long as m is a positive natural number, and
n is an integer, the program claims to work. The postcondition states that z , the value that is returned, is
equal to the product of m and n (that would be nice, but we're not convinced).
Let's look at the stated loop invariant. A loop invariant is a relationship between the variables that is
always true at the start and at the end of a loop iteration (we'll need to prove this). It's sucient to verify
that the invariant is true at the start of the rst iteration, and verify that if the invariant is true at the start
of any iteration, it must be true at the end of the iteration. Before we start the loop, we set x = m, y = n
and z = 0, so it is clear that z = mn xy = mn mn = 0. Now we need to show that if z = mn xy
before executing the body of the loop, and x 6= 0, then after executing the loop body, z = mn xy is still
true (can you write this statement formally?). Here's a sketch of a proof:
Assume x ; y ; z ; x +1 ; y +1 ; z +1 ; m; n 2 Z, where x represents the value of variable x at the beginning
i i i i i i i

of the ith iteration of the loop, and similarly for the other variables and subscripts. (Note that there
is no need to subscript m; n, since they aren't changed by the loop.)
Assume z = mn x y .
i i i

Case 1: Assume x odd. i

Then z +1 = z + y , x +1 = (x 1)=2, and y +1 = 2y .

i i i i i i i

So
mn x +1 y +1 = mn (x 1)=2
2y i i (since x is odd)
i i i

= mn x y + y i i i

=z +y i i

= z +1 i

Case 2: Assume x even. i

Then z +1 = z , x +1 = x =2, and y +1 = 2y .

i i i i i i

So
mn x +1 y +1 = mn x =2
2y i i i i

= mn x y i i

=z i

= z +1 i

Since x is either even or odd, in all cases mn x +1 y +1 = z +1

i i i i

51
Course notes for csc 165 h

Thus mn x y = z ) mn x +1 y +1 = z +1 .
i i i i i i

Since x ; x +1 ; y ; y +1 ; z ; z +1 ; m; n are arbitrary elements,

i i i i i i

8x ; x +1 ; y ; y +1 ; z ; z +1 ; m; n 2 Z; mn x y = z ) mn x +1 y +1 = z +1 .
i i i i i i i i i i i i

We should probably verify the postcondition to fully convince ourselves of the correctness of this algorithm.
We've shown the loop invariant holds, so let's see what we can conclude when the loop terminates (i.e.,
when x = 0). By the loop invariant, z = mn xy = mn 0 = mn, so we know we must get the right
answer (assuming the loop eventually terminates).
We should now be fairly convinced that this algorithm is in fact correct. One might now wonder, how
many iterations of the loop are completed before the answer is returned?
Also, why is it necessary for m > 0? What happens if it isn't?

5.5 Running time of programs

For any program P and any input x, let t (x) denote the number of \steps" P takes on input x. We need to
P

specify what we mean by a \step." A \step" typically corresponds to machine instructions being executed,
or some indication of time or resources expended.
Consider the following (somewhat arbitrary) accounting for common program steps:
method call: 1 step + steps to evaluate each argument + steps to execute the method.

return statement: 1 step + steps to evaluate return value.

if statement: 1 step + steps to evaluate condition.

assignment statement: 1 step + steps to evaluate each side.

arithmetic, comparison, boolean operators: 1 step + steps to evaluate each operand.

array access: 1 step + steps to evaluate index.

member access: 2 steps.

constant, variable evaluation: 1 step.

Notice that none of these \steps" (except for method calls) depend on the size of the input (sometimes
denoted with the symbol n). The smallest and largest steps above dier from each other by a constant of
about 5, so we can make the additional simplifying assumption that they all have the same cost | 1.

5.6 Linear search

Let's use linear search as an example.
def LS(A,x):
""" Return an index i such that x == L[i]. Otherwise, return -1. """
i = 0 # (line 1)
while i < len(A): # (line 2)
if A[i] == x: # (line 3)
return i # (line 4)
i = i + 1 # (line 5)
return -1 # (line 6)

Let's trace a function call, LS([2,4,6,8],4):

Line 1: 1 step (i=0)

52
 Chapter 5. Algorithm Analysis and Asymptotic Notation

Line 2: 1 step (0 < 4)

Line 3: 1 step (A[0] == 4)
Line 5: 1 step (i = 1)
Line 2: 1 step (1 < 4)
Line 3: 1 step (A[1] == 4)
Line 4: 1 (return 1)
So t ([2; 4; 6; 8]; 4) = 7. Notice that if the rst index where x is found is j , then t (A; x) will count lines
LS LS

2, 3, and 5 once for each index from 0 to j 1 (j indices), and then count lines 2, 3, 4 for index j , and so
t (A; x) will be 1 + 3j + 3.
LS

If x does not appear in A, then t (A; x) = 1 + 3 len(A) + 2, because line 1 executes once, lines 2, 3, and
LS

5 executes once for each index from 0 to len(A) 1, and then lines 2 and 6 execute.
We want a measure that depends on the size of the input, not the particular input. There are three
standard ways. Let P be a program, and let I be the set of all inputs for P . Then:
Average-case complexity: the weighted average over all possible inputs of size n.

In general X
A (n) =
P t (x)
p(x)
P

x of size n

where p(x) is the probability that input x is encountered.

Assuming all the inputs are equally likely, this \simplies" to
P
t (x)
A ( n) = of size
x n P
P
number of inputs of size n
(Dicult to compute.)
Best-case complexity: min(t (x)), where x is an input of size n.
P

In other words, B (n) = minft (x) j x 2 I ^ size(x) = ng.

P P

(Mostly useless.)
Worst-case complexity: max(t (x)), where x is an input of size n.
P

In other words, W (n) = maxft (x) j x 2 I ^ size(x) = ng.

P P

(Relatively easy to compute, and gives a performance guarantee.)

What is meant by \input size"? This depends on the algorithm. For linear search, the number of elements
in the array is a reasonable parameter. Technically (in csc 463 h, for example), the size is the number of
bits required to represent the input in binary. In practice we use the number of elements of input (length of
array, number of nodes in a tree, etc.)

5.7 Run time and constant factors

When calculating the running time of a program, we may know how many basic \steps" it takes as a function
of input size, but we may not know how long each step takes on a particular computer. We would like to
estimate the overall running time of an algorithm while ignoring constant factors (like how fast the CPU
is). So, for example, if we have 3 machines, where operations take 3s, 8s and 0.5s, the three functions
measuring the amount of time required, t(n) = 3n , t(n) = 8n , and t(n) = n =2 are considered the same,
2 2 2

53
Course notes for csc 165 h

ignoring (\to within") constant factors (the time required always grows according to a quadratic function
in terms of the size of the input n).
To view this another way, think back to the linear search example in the previous section. The worst-case
running time for that algorithm was given by the function
W (n) = 3n + 3

But what exactly does the constant \3" in front of the \n" represent? Or the additive \+3"? Neither value
corresponds to any intrinsic property of the algorithm itself; rather, the values are consequences of some
arbitrary choices on our part, namely, how many \steps" to count for certain Python statements. Someone
counting dierently (e.g., counting more than 1 step for statements that access list elements by index) would
arrive at a dierent expression for the worst-case running time. Would their answer be \more" or \less"
correct than ours? Neither: both answers are just as imprecise as one another! This is why we want to come
up with a tool that allows us to work with functions while ignoring constant multipliers.
The nice thing is that this means that lower order terms can be ignored as well! So f (n) = 3n2 and
g (n) = 3n2 + 2 are considered \the same," as are h(n) = 3n2 + 2n and j (n) = 5n2 . Notice that

8n 2 N; n > 1 ) f (n) 6 g(n) 6 h(n) 6 j (n)

but there's always a constant factor that can reverse any of these inequalities.
Really what we want to measure is the growth rate of functions (and in computer science, the growth
rate of functions that bound the running time of algorithms). You might be familiar with binary search and
linear search (two algorithms for searching for a value in a sorted array). Suppose one computer runs binary
search and one computer runs linear search. Which computer will give an answer rst, assuming the two
computers run at roughly the same CPU speed? What if one computer is much faster (in terms of CPU
speed) than the other, does it aect your answer? What if the array is really, really big?

How large is \sufficiently large?"

Is binary search a better algorithm than linear search?7 It depends on the size of the input. For example,
suppose you established that linear search has complexity L(n) = 3n and binary search has complexity
B (n) = 9 log2 n. For the rst few n, L(n) is smaller than B (n). However, certainly for n > 10, B (n) is
smaller, indicating less \work" for binary search.
When we say \large enough" n, we mean we are discussing the asymptotic behaviour of the complexity
function (i.e., the behaviour as n grows toward innity), and we are prepared to ignore the behaviour near
the origin.

5.8 Asymptotic notation: Making Big-O precise

We dene R>0 as the set of nonnegative real numbers, and dene R+ as the set of positive real numbers.
Here is a precise denition of \The set of functions that are eventually no more than f , to within a constant
factor":
Definition: For any function f : N ! R
>0 (i.e., any function mapping naturals to nonnegative reals), let

O(f ) = g : N ! R>0 j 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) g (n) 6 cf (n) :

Saying g 2 O(f ) says that \g grows no faster than f " (or equivalently, \f is an upper bound for g"), so
long as we modify our understanding of \growing no faster" and being an \upper bound" with the practice
of ignoring constant factors. Now we can prove some theorems.

54
 Chapter 5. Algorithm Analysis and Asymptotic Notation

Suppose g(n) = 3n2 + 2 and f (n) = n2 . Then g 2 O(f ). To be more precise, we need to prove the
statement 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) 3n2 + 2 6 cn2 . It's enough to nd some c and B that \work"
in order to prove the theorem.
Finding c means nding a factor that will scale n2 up to the size of 3n2 + 2. Setting c = 3 almost works,
but there's that annoying additional term 2. Certainly 3n2 +2 < 4n2 so long as n > 2, since n > 2) n2 > 2.
So pick c = 4 and B = 2 (other values also work, but we like the ones we thought of rst). Now concoct a
proof of
9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) 3n2 + 2 6 cn2 :
Let c0 = 4 and B 0 = 2.
Then c0 2 R+ and B 0 2 N.
Assume n 2 N and n > B 0 . # direct proof for an arbitrary natural number
Then n2 > B 0 2 = 4. # squaring is monotonic on natural numbers
Then n2 > 2.
Then 3n2 + n2 > 3n2 + 2. # adding 3n2 to both sides of the inequality
Then 3n2 + 2 6 4n2 = c0 n2 # re-write
Then 8n 2 N; n > B 0 ) 3n2 + 2 6 c0 n2 # introduce 8 and )
Then 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) 3n2 + 2 6 cn2 . # introduce 9 (twice)
So, by denition, g 2 O(f ).

A more complex example

Let's prove that 2n3 5n4 + 7n6 is in O(n2 4n5 + 6n8 ). We begin with:
Let c0 = . Then c0 2 R+ .
Let B 0 = . Then B 0 2 N.
Assume n 2 N and n > B 0 . # arbitrary natural number and antecedent
Then 2n3 5n4 + 7n6 6 : : : 6 c0 (n2 4n5 + 6n8 ).
Then 8n 2 N; n > Bi0 ) 2n3 5n4 + 7n6 6 c0 (n2 4n5 + 6n8 ). # introduce ) and 8
Hence, 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) 2n3 5n4 + 7n6 6 c(n2 4n5 + 6n8 ). # introduce 9
To ll in the


we try to form a chain of inequalities, working from both ends, simplifying the expressions:
2n3 5n4 + 7n6 6 2n3 + 7n6 (drop 5n4 because it doesn't help us in an important way)
6 2n6 + 7n6 (increase n3 to n6 because we have to handle n6 anyway)
= 9n6
6 9n8 (simpler to compare)
= 2(9=2)n8 (get as close to form of the simplied end result: now choose c0 = 9=2)
= 2cn8
= c0 ( 4n8 + 6n8 ) (reading bottom up: decrease 4n5 to 4n8 because we have to
handle n8 anyway)
6 c0 ( 4n5 + 6n8 ) (reading bottom up: drop n2 because it doesn't help us in an
important way)
0
6 c (n 4n + 6n )
2 5 8

We never needed to restrict n in any way beyond n 2 N (which includes n > 0), so we can ll in c0 = 9=2,
B 0 = 0, and complete the proof.
Let's use this approach to prove n4 2= O(3n2 ). More precisely, we have to prove the negation of the
statement 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) n4 6 c3n2 .

55
Course notes for csc 165 h

Assume c 2 R+ and B 2 N. # arbitrary positive real number and natural number

Let n0 = .
..
.
So n0 2 N.
..
.
So n0 > B .
..
.
So n40 > c3n20 .
Then 8c 2 R+ ; 8B 2 N; 9n 2 N; n > B ^ n4 > c3n2 .
.
Here's our chain of inequalities (the third ..):
And n40 > n30 (don't need full power of n40 )
= n0
n20 (make form as close as possible)
> c
3n20 (if we make n0 > 3c and n0 > 0)
Now pick n0 = max(B; d3c + 1e).
.
The rst .. is:
Since c > 0, 3c + 1 > 0, so d3c + 1e 2 N.
Since B 2 N, max(B; d3c + 1e) 2 N.
.
The second .. is:
max(B; d3c + 1e) > B .
We also note just before the chain of inequalities:
n0 = max(B; d3c + 1e) > d3c + 1e > 3c + 1 > 3c.
Some points to note are:
 Don't \solve" for n until you've made the form of the two sides as close as possible.
 You're not exactly solving for n: you are nding a condition of the form n > that makes the desired
inequality true. You might nd yourself using the \max" function a lot.

Other bounds
In analogy with O(f ), consider two other denitions:
Definition: For any function f : N ! R
> , let
0



(f ) = g : N ! R> j 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) g(n) > cf (n) :
0 +

To say \g 2
(f )" expresses the concept that \g grows at least as fast as f " (f is a lower bound on g).
Definition: For any function f : N ! R
>0 , let

(f ) = g : N ! R>0 j 9c1 2 R+ ; 9c2 2 R+ ; 9B 2 N; 8n 2 N; n > B ) c1 f (n) 6 g(n) 6 c2 f (n) :
To say \g 2 (f )" expresses the concept that \g grows at the same rate as f " (f is a tight bound for g, or
f is both an upper bound and a lower bound on g ).

56
 Chapter 5. Algorithm Analysis and Asymptotic Notation

Induction interlude
Suppose P (n) is some predicate of the natural numbers, and:
() P (0) ^ (8n 2 N; P (n) ) P (n + 1)):

You should certainly be able to show that () implies P (0), P (1), P (2), in fact P (n) where n is any natural
number you have the patience to follow the chain of results to obtain. In fact, we feel that we can \turn the
crank" enough times to show that () implies P (n) for any natural number n. This is called the Principle
of Simple Induction (PSI). It isn't proved, it is an axiom that we assume to be true.
Here's an application of the PSI that will be useful for some big-Oh problems.
P (n): 2 > 2n.
n

I'd like to prove that 8n; P (n), using the PSI. Here's what I do:
Prove P (0): P (0) states that 2 = 1 > 2(0) = 0, which is true.
0

Prove 8n 2 N; P (n) ) P (n + 1):

Assume n 2 N. # arbitrary natural number
Assume P (n), that is 2 > 2n. # antecedent
n

Then n = 0 _ n > 0. # natural numbers are non-negative

Case 1 (assume n = 0): Then 2 = 21 = 2 > 2(n + 1) = 2.
+1 n

Case 2 (assume n > 0): Then n > 1. # n is an integer greater than 0

Then 2 > 2. # since n > 1, and 2 is monotone increasing
n n

Then 2 +1 = 2 + 2 > 2n + 2 = 2(n + 1). # by previous line and IH P (n)

n n n

Then 2 +1 > 2(n + 1), which is P (n + 1). # true in both possible cases
n

Then P (n) ) P (n + 1). # introduce )

Then 8n 2 N; P (n) ) P (n + 1). # introduce 8
I now conclude, by the PSI, 8n 2 N; P (n), that is 2 > 2n.n

Here's a big-Oh problem where we can use P (n). Let g(n) = 2 and f (n) = n. I want to show that
n

g 62 O(f ).
Assume c 2 R+ , assume B 2 N. # arbitrary values
Let k = max(0; dlog2 (c)e) + 1 + B , and let n0 = 2k.
Then n0 2 N. # dxe; 1; 2; B 2 N, N closed under max; +; 
Then n0 > B . # at least twice B
Then 2 > c. # choice of k, 2 is increasing function
k x

Then
g (n0 ) = 2
n0
= 2  2 # by choice of k
k k

> 2  2k # by P (2k)
k

= 2  n0 > cn0 # since n0 = 2k and 2

k k
>c
= cf (n0 ):
Then n0 > B ^ g(n0 ) > cf (n0 ). # introduce ^
Then 9n 2 N; n > B ^ g(n) > cf (n). # introduce 9
Then 8c 2 R; 8B 2 N; 9n 2 N; n > B ^ g(n) > cf (n). # introduce 8
So, I can conclude that g 62 O(f ).

57
Course notes for csc 165 h

What happens to induction for predicates that are true for all natural numbers after a certain point, but
untrue for the rst few natural numbers? For example, 2 grows much more quickly than n2 , but 23 is not
n

larger than 32 . Choose n big enough, though, and it is true that:

P (n) : 2 > n2 :
n

You can't prove this for all n, when it is false for n = 2; n = 3, and n = 4, so you'll need to restrict the
domain and prove that for all natural numbers greater than 4, P (n) is true. We don't have a slick way to
restrict domains in our symbolic notation. Let's consider three ways to restrict the natural numbers to just
those greater than 4, and then use induction.
Restrict by set difference: One way to restrict the domain is by set dierence:

8n 2 N n f0; 1; 2; 3; 4g; P (n)

Again, we'll need to prove P (5), and then that 8n 2 N n f0; 1; 2; 3; 4g; P (n) ) P (n + 1).
Restrict by translation: We can also restrict the domain by translating our predicate, by letting Q(n) =

P (n + 5), that is:

Q(n) : 2 +5 > (n + 5)2
n

Now our task is to prove Q(0) is true and that for all n 2 N, Q(n)) Q(n +1). This is simple induction.
Restrict using implication: Another method of restriction uses implication to restrict the domain where

we claim P (n) is true | in the same way as for sentences:

8n 2 N; n > 5 ) P (n):
The expanded predicate Q(n) : n > 5 ) P (n) now ts our pattern for simple induction, and all we
need to do is prove:
1. Q(0) is true (it is vacuously true, since 0 > 5 is false).
2. 8n 2 N; Q(n) ) Q(n + 1). This breaks into cases.
 If n < 4, then Q(n) and Q(n +1) are both vacuously true (the antecedents of the implication
are false, since n and n +1 are not greater than, nor equal to, 5), so there is nothing to prove.
 If n = 4, then Q(n) is vacuously true, but Q(n +1) has a true antecedent (5 > 5), so we need
to prove Q(5) directly: 25 > 52 is true, since 32 > 25.
 If n > 4, we can depend on the assumption of the consequent of Q(n 1) being true to prove
Q(n):
n
2 = 2 1 + 2 1 (denition of 2 )
n n n

> 2(n 1)2 (antecedent of Q(n 1))

= 2n2 2n + 2 = n2 + n(n 2) + 2 > n2 + 2 > n2 (since n > 4 > 2)
After all that work, it turns out that we need prove just two things:
1. P (5)
2. 8n 2 N, If n > 4, then P (n) ) P (n + 1).
This is the same as before, except now our base case is P (5) rather than P (0), and we get to use the fact
that n > 5 in our induction step (if we need it).
Whichever argument you're comfortable with, notice that simple induction is basically the same: you
prove the base case (which may now be greater than 0), and you prove the induction step.

58
 Chapter 5. Algorithm Analysis and Asymptotic Notation

Some theorems
Here are some general results that we now have the tools to prove.
 f 2 O(f ).
 (f 2 O(g) ^ g 2 O(h)) ) f 2 O(h).
 g 2
(f ) , f 2 O(g).
 g 2 (f ) , g 2 O(f ) ^ g 2
(f ).
Test your intuition about Big-O by doing the \scratch work" to answer the following questions:
 Are there functions f; g such that f 2 O(g) and g 2 O(f ) but f 6= g? 8

 Are there functions f; g such that f 62 O(g), and g 62 O(f )? 9

To show that (f 2 O(g) ^ g 2 O(h)) ) f 2 O(h), we need to nd a constant c 2 R and a constant B 2 N,

+

that satisfy:
8n 2 N; n > B ) f (n) 6 ch(n):
Since we have constants that scale h to g and then g to f , it seems clear that we need their product to scale
g to f . And if we take the maximum of the two starting points, we can't go wrong. Making this precise:
Theorem 5.1: For any functions f; g; h : N ! R
> , we have (f 2 O(g ) ^ g 2 O(h)) ) f 2 O(h).
0

Proof:

Assume f 2 O(g) ^ g 2 O(h).

So f 2 O(g).
So 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) f (n) 6 cg(n). # by def'n of f 2 O(g)
Let c 2 R+ ; B 2 N be such that 8n 2 N; n > B ) f (n) 6 c g(n).
g g g g

So g 2 O(h).
So 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) g(n) 6 ch(n). # by def'n of g 2 O(h)
Let c 2 R+ ; B 2 N be such that 8n 2 N; n > B ) g(n) 6 c h(n).
h h h h

Let c0 = c c . Let B 0 = max(B ; B ).

g h g h

Then, c0 2 R+ (because c ; c 2 R+ ) and B 0 2 N (because B ; B 2 N).

g h g h

Assume n 2 N and n > B 0 .

Then n > B (by denition of max), so g(n) 6 c h(n).
h h

Then n > B (by denition of max), so f (n) 6 c g(n) 6 c c h(n).

g g g h

So f (n) 6 c0 h(n).
Hence, 8n 2 N; n > B 0 ) f (n) 6 c0 h(n).
Therefore, 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) f (n) 6 ch(n).
So f 2 O(g), by denition.
So (f 2 O(g) ^ g 2 O(h)) ) f 2 O(h).
To show that g 2
(f ) , f 2 O(g), it is enough to note that the constant, c, for one direction is positive,
so its reciprocal will work for the other direction.10
Theorem 5.2: For any functions f; g : N ! R
>0 , we have g 2
(f ) , f 2 O(g ).
Proof:

g 2
(f )
() 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) g(n) > cf (n) (by denition)
() 9c0 2 R+ ; 9B 0 2 N; 8n 2 N; n > B 0 ) f (n) 6 c0 g(n) (letting c0 = 1=c and B 0 = B )
() f 2 O(g) (by denition)
To show g 2 (f ) , g 2 O(f ) ^ g 2
(f ), it's really just a matter of unwrapping the denitions.

59
Course notes for csc 165 h

Theorem 5.3: For any functions f; g : N ! R>0 , we have g 2 (f ) , g 2 O(f ) ^ g 2

(f ).
Proof:

g 2 (f )
, (by denition)
9c1 2 R+ ; 9c2 2 R+ ; 9B 2 N; 8n 2 N; n > B ) c1 f (n) 6 g(n) 6 c2 f (n).
, (combined inequality, and B = max(B1 ; B2 ))

9c1 2 R+ ; 9B1 2 N; 8n 2 N; n > B1 ) g(n) > c1 f (n)
^
9c2 2 R+ ; 9B2 2 N; 8n 2 N; n > B2 ) g(n) 6 c2 f (n)
, (by denition)
g 2
(f ) ^ g 2 O(f )
Here's an example of a corollary that recycles some of the theorems we've already proven (so we don't have to
do the grubby work). To show g 2 (f ), f 2 (g), I re-use theorems proved above and the commutativity
of ^:
Corollary: For any functions f; g : N ! R
>0 , we have g 2 (f ) , f 2 (g ).
Proof:

g 2 (f )
() g 2 O(f ) ^ g 2
(f ) (by 5.3)
() g 2 O(f ) ^ f 2 O(g) (by 5.2)
() f 2 O(g) ^ g 2 O(f ) (by commutativity of ^)
() f 2 O(g) ^ f 2
(g) (by 5.2)
() f 2 (g) (by 5.3)

A very important note

Note that asymptotic notation (the Big-O, Big-
, and Big- denitions) bound the asymptotic growth rates
of functions, as n approaches innity. Often in computer science we use this asymptotic notation to bound
functions that express the running times of algorithms, perhaps in best case or in worst case. Asymptotic
notation does not express or bound the worst case or best case running time, only the functions expressing
these values.
This distinction is subtle, but crucial to understanding both running times and asymptotic notation.
Suppose U is an upper bound on the worst-case running time of some program P , denoted T (n): P

T 2 O(U )
P

() 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) T (n) 6 cU (n)

+
P

() 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) maxft (x) j x 2 I ^ size(x) = ng 6 cU (n)

+
P

() 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) 8x 2 I; size(x) = n ) t (x) 6 cU (n)

+
P

() 9c 2 R ; 9B 2 N; 8x 2 I; size(x) > B ) t (x) 6 cU (size(x))

+
P

So to show that T 2 O(U (n)), you need to nd constants c and B and show that for an arbitrary input x
P

of size n, P takes at most c
U (n) steps.

In the other direction, suppose L is a lower bound on the worst-case running time of algorithm P :
T 2
(L)
P

() 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) maxft (x) j x 2 I ^ size(x) = ng > cL(n)

+
P

() 9c 2 R ; 9B 2 N; 8n 2 N; n > B ) 9x 2 I; size(x) = n ^ t (x) > cL(n)

+
P

60
 Chapter 5. Algorithm Analysis and Asymptotic Notation

So, to prove that T 2

(L), we have to nd constants c, B and for arbitrary n, nd an input x of size n,
p

for which we can show that P takes at least cL(n) steps on input x

5.9 Insertion sort example

Here is an intuitive sorting algorithm:
11

def IS(A):
""" Sort the elements of A in non-decreasing order. """
i = 1 # (line 1)
while i < len(A): # (line 2)
t = A[i] # (line 3)
j = i # (line 4)
while j > 0 and A[j-1] > t: # (line 5)
A[j] = A[j-1] # (line 6)
j = j-1 # (line 7)
A[j] = t # (line 8)
i = i+1 # (line 9)

Let's nd an upper bound for T (n), the maximum number of steps to Insertion Sort an array of size n.
IS

We'll use the proof format to prove and nd the bound simultaneously | during the course of the proof we
can ll in the necessary values for c and B .
We show that T (n) 2 O(n2 ) (where n = len(A)):
IS

Let c0 = . Let B 0 = .
Then c0 2 R+ and B 0 2 N.
Assume n 2 N, A is an array of length n, and n > B 0 .
Then lines 5{7 execute at most n 1 times, which we can overestimate at 3n steps, plus 1 step
for the last loop test.
Then lines 2{9 take no more than n(5 + 3n) + 1 = 5n + 3n2 + 1 steps.
So 3n2 + 5n + 1 6 c0 n2 (ll in the values of c0 and B 0 that makes this so | setting c0 = 9; B 0 = 1
should do).
Since n is the length of an arbitrary array A, 8n 2 N; n > B 0 ) T (n) 6 c0 n2 (so long as B 0 > 1).
IS

Since c0 is a positive real number and B 0 is a natural number,

9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) T (n) 6 cn2 .
IS

So T 2 O(n2 ) (by denition of O(n2 )).

IS

Similarly, we prove a lower bound. Specically, T 2

(n2 ):
IS

Let c0 = . Let B 0 = .
Then c0 2 R+ and B 0 2 N.
Assume n 2 N and n > B 0 .
Let A0 = [n 1; : : : ; 1; 0] (notice that this means n > 1).
Then at any point during the outside loop, A0 [0::(i 1)] contains the same elements as before but
sorted (i.e., no element from A0 [(i + 1)::(n 1)] has been examined yet).
Then the inner while loop makes i iterations, at a cost of 3 steps per iteration, plus 1 for the nal
loop check, since the value A0 [i] is less than all the values A0 [0::(i 1)], by construction of the
array.
Then the inner loop makes strictly greater than 2i + 1, or greater than or equal to 2i + 2., steps.
Then (since the outer loop varies from i = 1 to i = n 1 and we have n 1 iterations of lines 3
and 4, plus one iteration of line 1), we have that t (n) > 1+3+5+


+(2n 1)+(2n +1) = n2
IS

(the sum of the rst n odd numbers), so long as n is at least 4.

61
Course notes for csc 165 h

So there is some array A of size n such that t (A) > c0 n2 . IS

This means T (n) > c0 n2 (setting B 0 = 4; c0 = 1 will do).

IS

Since n was an arbitrary natural number, 8n 2 N; n > B 0 ) T (n) > c0 n2 . IS

Since c0 2 R+ and B 0 is a natural number, 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) T (n) > cn2 . IS

So T 2
(n2 ) (by denition of
(n2 )).
IS

From these proofs, we conclude that T 2 (n2 ). IS

Exercises for asymptotic notation

1. Prove or disprove the following claims:
(a) 7n + 11n + n 2 O(n )
3 2 3 12

(b) n + 165 2
(n )
2 4

(c) n! 2 O(n ) n

(d) n 2 O(n log n) 2

(e) 8k 2 N; k > 1 ) log n 2 (log n) k 2

(
2. Dene g(n) = np=165; n < 165 . Note that 8x 2 R; x 6 dxe < x + 1.
3

6n ; n > 165 5

Prove that g 2 O(n ). 2:5

3. Let F be the set of functions from N to R> . Prove the following theorems: 0

(a) For f; g 2 F, if g 2
(f ) then g 2
(f ). 2 2

(b) 8k 2 N; k > 1 ) 8d 2 R ; d log n 2 (log n). +

k 2
13

Notice that (b) means that all logarithms eventually grow at the same rate (up to a multiplicative
constant), so the base doesn't matter (and can be omitted inside the asymptotic notation).
4. Let F be the set of functions from N to R> . Prove or disprove the following claims:
0

(a) 8f 2 F ; 8g 2 F ; f 2 O(g) ) (f + g) 2 (g)

(b) 8f 2 F ; 8f 0 2 F ; 8g 2 F ; (f 2 O(g) ^ f 0 2 O(g)) ) (f + f 0 ) 2 O(g)
5. For each function f in the left column, choose one expression O(g) from the right column such that
f 2 O(g ). Use each expression exactly once.

(i) 3
2 2 n
(a) O( 1 )
n

(ii) 2n4 +1
n3 +2n 1
2 (b) O(1)
(iii) (n + 7)(n5 7) 2
5
(c) O(log2 n)
(iv) 2
4
n log2
2 +1
n
n n
(d) O(n)
(v) n log2
n5
2 n
(e) O(n log2 n)
(vi) 8+ 2 2 1
n (f) O(n2 )
(vii) 23 +1 2
n
(g) O(n10 )
(viii) n! 2
(h) O(2 )
n

(ix) 5 log2 ( +1)

1+ log2 3n
2 n
n
(i) O(10 )n

(x) (n 2) log2 (n3 + 4) 2 (j) O(n ) n

62
 Chapter 5. Algorithm Analysis and Asymptotic Notation

Chapter 5 Notes
1
From 0 to (2 1), if we work in analogy with base 10.
2
To parse the 0:101 part, calculate 0:101 = 1(2 1 ) + 0(2 2 ) + 1(2 3 ).
3
You should be able to look up this algorithm in an elementary school textbook.
4
Same as the previous exercise, but only write numbers that have 0's and 1's, and do binary addition.
5
Integer divide by 10.
6
Integer divide by 2.
7
Better in the sense of time complexity.
8
Sure, f = n2 , g = 3n2 + 2.
9
Sure. f and g don't need to both be monotonic, so let f (n) = n2 and
(
n; n even
g ( n) =
n3 ; n odd

So not every pair of functions from N ! R>0 can be compared using Big-O.
10
Let's try the symmetrical presentation of bi-implication.
11
but not particularly ecient. . .
12
The claim is true.
Let c0 = 8. Then c0 2 R+ .
Let B 0 = 12. Then B 0 2 N.
Assume n 2 N and n > B 0 .
Then n3 = n
n2 > 12
n2 = 11n2 + n2 > 11n2 + n. # since n > B 0 = 12
Thus c0 n3 = 8n3 = 7n3 + n3 > 7n3 + 11n2 + n.
So 8n 2 N, n > B 0 ) 7n3 + 11n2 + n 6 c0 n3 .
Since B 0 is a natural number, 9B 2 N; 8n 2 N; n > B ) 7n3 + 11n2 + n 6 c0 n3 .
Since c0 is a real positive number, 9c 2 R+ ; 9B 2 N; 8n 2 N; n > B ) 7n3 + 11n2 + n 6 cn3 .
By denition, 7n3 + 11n2 + n 2 O(n3 ).
13
Assume k 2 N and k > 1.
Assume d 2 R+ .
It suces to argue that d log n 2 (log2 n).
Let c01 = log2 . Since k > 1, log2 k 6= 0 and so c01 2 R+ .
k
d
k

Let c02 = log2 . By the same reasoning, c02 2 R+ .

d

Let B 0 = 1. Then B 0 2 N.
k

Assume n 2 N and n > B 0 .

Then c01 log2 n = log2 log2 n = d log
d
k
2
log2
= d log n 6 d log n.
n
k k k

Moreover, d log n 6 d log2 = log2 log2 n = c02 log2 n.

log2 n d

Hence, 8n 2 N; n > B 0 ) c01 log2 n 6 d log n 6 c02 log2 n.

k k k

Thus 9c1 2 R+ ; 9c2 2 R+ ; 9B 2 N; 8n 2 N; n > B ) c1 log2 n 6 d log k

n 6 c2 log2 n.
By denition, d log n 2 (log2 n).
k

63
Course notes for csc 165 h

Thus, 8d 2 R+ ; d log n 2 (log2 n).

k

Hence 8k 2 N; k > 1 ) 8d 2 R+ ; d log n 2 (log2 n).

k

64