Analysis of Algorithms
Introduction
What is Algorithm?
a clearly specified set of simple instructions to be followed to
solve a problem
Takes a set of values, as input and
produces a value, or set of values, as output
May be specified
InEnglish
As a computer program
As a pseudo-code
Data structures
Methods of organizing data
Program = algorithms + data structures
Introduction
Why need algorithm analysis ?
writing a working program is not good enough
The program may be inefficient!
If the program is run on a large data set, then the
running time becomes an issue
Example: Selection Problem
Given a list of N numbers, determine the kth
largest, where k N.
Algorithm 1:
(1) Read N numbers into an array
(2) Sort the array in decreasing order by some
simple algorithm
(3) Return the element in position k
Example: Selection Problem…
Algorithm 2:
(1) Read the first k elements into an array and sort
them in decreasing order
(2) Each remaining element is read one by one
Ifsmaller than the kth element, then it is ignored
Otherwise, it is placed in its correct spot in the array,
bumping one element out of the array.
(3) The element in the kth position is returned as
the answer.
Example: Selection Problem…
Which algorithm is better when
N =100 and k = 100?
N =100 and k = 1?
What happens when N = 1,000,000 and k =
500,000?
There exist better algorithms
Algorithm Analysis
We only analyze correct algorithms
An algorithm is correct
If, for every input instance, it halts with the correct output
Incorrect algorithms
Might not halt at all on some input instances
Might halt with other than the desired answer
Analyzing an algorithm
Predicting the resources that the algorithm requires
Resources include
Memory
Communication bandwidth
Computational time (usually most important)
Algorithm Analysis…
Factors affecting the running time
computer
compiler
algorithm used
input to the algorithm
The content of the input affects the running time
typically, the input size (number of items in the input) is the main
consideration
E.g. sorting problem the number of items to be sorted
E.g. multiply two matrices together the total number of
elements in the two matrices
Machine model assumed
Instructions are executed one after another, with no
concurrent operations Not parallel computers
Example
N
Calculate i
i 1
3
1
1
2 2N+2
3 4N
4 1
Lines 1 and 4 count for one unit each
Line 3: executed N times, each time four units
Line 2: (1 for initialization, N+1 for all the tests, N for
all the increments) total 2N + 2
total cost: 6N + 4 O(N)
Worst- / average- / best-case
Worst-case running time of an algorithm
The longest running time for any input of size n
An upper bound on the running time for any input
guarantee that the algorithm will never take longer
Example: Sort a set of numbers in increasing order; and the
data is in decreasing order
The worst case can occur fairly often
E.g. in searching a database for a particular piece of information
Best-case running time
sort a set of numbers in increasing order; and the data is
already in increasing order
Average-case running time
May be difficult to define what “average” means
Running-time of algorithms
Bounds are for the algorithms, rather than
programs
programs are just implementations of an algorithm,
and almost always the details of the program do
not affect the bounds
Bounds are for algorithms, rather than
problems
A problem can be solved with several algorithms,
some are more efficient than others
Growth Rate
The idea is to establish a relative order among functions for large
n
c , n0 > 0 such that f(N) c g(N) when N n0
f(N) grows no faster than g(N) for “large” N
Asymptotic notation: Big-Oh
f(N) = O(g(N))
There are positive constants c and n0 such
that
f(N) c g(N) when N n0
The growth rate of f(N) is less than or equal to
the growth rate of g(N)
g(N) is an upper bound on f(N)
Big-Oh: example
Let f(N) = 2N2. Then
f(N) = O(N4)
f(N) = O(N3)
f(N) = O(N2) (best answer, asymptotically tight)
O(N2): reads “order N-squared” or “Big-Oh N-squared”
Big Oh: more examples
N2 / 2 – 3N = O(N2)
1 + 4N = O(N)
7N2 + 10N + 3 = O(N2) = O(N3)
log10 N = log2 N / log2 10 = O(log2 N) = O(log N)
sin N = O(1); 10 = O(1), 1010 = O(1)
N
i 1
i N N O( N 2 )
i 1
N
i 2
N N 2
O ( N 3
)
log N + N = O(N)
logk N = O(N) for any constant k
N = O(2N), but 2N is not O(N)
210N is not O(2N)
Math Review: logarithmic functions
x b iff
a
log x b a
log ab log a log b
log m b
log a b
log m a
log a b log a
b
a log n n log a
log a (log a ) log a
b b b
d log e x 1
dx x
Some rules
When considering the growth rate of a function using
Big-Oh
Ignore the lower order terms and the coefficients of
the highest-order term
No need to specify the base of logarithm
Changing the base from one constant to another changes the
value of the logarithm by only a constant factor
If T1(N) = O(f(N) and T2(N) = O(g(N)), then
T1(N) + T2(N) = max(O(f(N)), O(g(N))),
T1(N) * T2(N) = O(f(N) * g(N))
Big-Omega
c , n0 > 0 such that f(N) c g(N) when N n0
f(N) grows no slower than g(N) for “large” N
Big-Omega
f(N) = (g(N))
There are positive constants c and n0 such
that
f(N) c g(N) when N n0
The growth rate of f(N) is greater than or
equal to the growth rate of g(N).
Big-Omega: examples
Let f(N) = 2N2. Then
f(N) = (N)
f(N) = (N2) (best answer)
f(N) = (g(N))
the growth rate of f(N) is the same as the growth rate
of g(N)
Big-Theta
f(N) = (g(N)) iff
f(N) = O(g(N)) and f(N) = (g(N))
The growth rate of f(N) equals the growth rate
of g(N)
Example: Let f(N)=N2 , g(N)=2N2
Since f(N) = O(g(N)) and f(N) = (g(N)),
thus f(N) = (g(N)).
Big-Theta means the bound is the tightest
possible.
Some rules
If T(N) is a polynomial of degree k, then
T(N) = (Nk).
For logarithmic functions,
T(logm N) = (log N).
Typical Growth Rates
Growth rates …
Doubling the input size
f(N) = c f(2N) = f(N) = c
f(N) = log N f(2N) = f(N) + log 2
f(N) = N f(2N) = 2 f(N)
f(N) = N2 f(2N) = 4 f(N)
f(N) = N3 f(2N) = 8 f(N)
f(N) = 2N f(2N) = f2(N)
Advantages of algorithm analysis
To eliminate bad algorithms early
pinpoints the bottlenecks, which are worth coding
carefully
Using L' Hopital's rule
L' Hopital's rule
If lim f ( N ) and lim g ( N )
n n
f (N ) f ( N )
then lim = lim
n g ( N ) n g ( N )
Determine the relative growth rates (using L' Hopital's rule if
necessary)
f (N )
compute lim
n g(N )
if 0: f(N) = O(g(N)) and f(N) is not (g(N))
if constant 0: f(N) = (g(N))
if : f(N) = (f(N)) and f(N) is not (g(N))
limit oscillates: no relation
General Rules
For loops
at most the running time of the statements inside
the for-loop (including tests) times the number of
iterations.
Nested for loops
the running time of the statement multiplied by the
product of the sizes of all the for-loops.
O(N2)
General rules (cont’d)
Consecutive statements
These just add
O(N) + O(N2) = O(N2)
If S1
Else S2
never more than the running time of the test plus the larger of
the running times of S1 and S2.
Another Example
Maximum Subsequence Sum Problem
Given (possibly negative) integers
j
A1, A2, ....,
An, find the maximum value of Ak
k i
For convenience, the maximum subsequence sum
is 0 if all the integers are negative
E.g. for input –2, 11, -4, 13, -5, -2
Answer: 20 (A2 through A4)
Algorithm 1: Simple
Exhaustively tries all possibilities (brute force)
O(N3)
Algorithm 2: Divide-and-conquer
Divide-and-conquer
split the problem into two roughly equal subproblems, which
are then solved recursively
patch together the two solutions of the subproblems to arrive
at a solution for the whole problem
The maximum subsequence sum can be
Entirely in the left half of the input
Entirely in the right half of the input
It crosses the middle and is in both halves
Algorithm 2 (cont’d)
The first two cases can be solved recursively
For the last case:
find the largest sum in the first half that includes the last
element in the first half
the largest sum in the second half that includes the first
element in the second half
add these two sums together
Algorithm 2 …
O(1)
T(m/2)
T(m/2)
O(m)
O(1)
Algorithm 2 (cont’d)
Recurrence equation
T (1) 1
N
T (N ) 2T ( ) N
2
2 T(N/2): two subproblems, each of size N/2
N: for “patching” two solutions to find solution to
whole problem
Algorithm 2 (cont’d)
N
Solving the recurrence: T (N ) 2T ( ) N
2
N
4T ( ) 2N
4
N
8 T ( ) 3N
8
N
2 k T ( k ) kN
2
With k=log N (i.e. 2k = N), we have
T ( N ) N T (1) N log N
N log N N
Thus, the running time is O(N log N)
faster than Algorithm 1 for large data sets
