BITS Pilani presentation

BITS Pilani Dr. Mukesh Kumar Rohil

Department of Computer Science & Information Systems
Pilani Campus
 BITS Pilani
Pilani Campus

Course: Design & Analysis of Algorithms

Lecture 04-05: Polynomial Representations, Evaluating a
Polynomial using Horners’ Rule, Interpolation using Gaussian
Elimination, Discrete Fourier Transform, Fast Fourier Transform
Algorithm
Learning Objectives

• Polynomial Representations
• Evaluating a Polynomial using Horners’ Rule
• Interpolation using Gaussian Elimination
• Discrete Fourier Transform
• Fast Fourier Transform Algorithm

BITS Pilani, Pilani Campus

Polynomial Representation
• We write that
degree(A) = k.
• Any integer strictly
greater than the degree
of a polynomial is a
degree-bound of that
polynomial.
• Therefore, the degree
of a polynomial of
degree-bound n may be
any integer between 0
and (n -1), inclusive.
BITS Pilani, Pilani Campus
Operations on Polynomials

• Division:
C(x) = A(x)/B(x)
• Reciprocal
• C(x) = 1/A(x)

(J-k)
BITS Pilani, Pilani Campus
Representing Polynomials

1. By coefficients form

2. By point values form

BITS Pilani, Pilani Campus

 Representing Polygon:
Coefficient Form
• A polynomial f(x) = a0 + a1x + … + an-1xn-1 is represented as
(a0, a1, …, an-1) .

For coefficient form the Time complexity of:

Adding two polynomial = Θ(n)

Multiplying a polynomial by a constant C0 = Θ(n) (∵ C0(a0, a1,

…, an-1) = (C0a0, C0a1, …, C0an-1)

Multiplying two polynomials = Θ(n2) (∵ (a0, a1, …, an-1) x (b0, b1,

…, bn-1) = (a0b0 + a0b1 + a1b0, …., an-1bn-1) using
convolutional algorithms i.e. Iterative Multiply)
BITS Pilani, Pilani Campus
 Representing Polynomial:
Point Value Form
• A polynomial f(x) = a0 + a1x + … + an-1xn-1 is represented as
{(x0, f(x0)), (x1, f(x1)), …, (xn-1, f(xn-1))} where all xi‘s are
different.
For Point Value form the Time complexity of:
Adding two polynomial = Θ(n) (Second polynomial also
specified at xi’s)

Multiplying a polynomial by a constant C0 = Θ(n)

Multiplying two polynomials = Θ(n) (∵ multiplication is

represented as {(x0, f1(x0)f2(x0)), (x1, f1(x1)f2(x1)), …, (xn-1,
f1(xn-1)f2(xn-1))})
BITS Pilani, Pilani Campus
 Evaluating Polynomials by
Horner’s Rule
a0 + a1x + … + an-1xn-1 • Horner’s rule takes Θ(n)
= a0 + x(a1x + x(a2 + … x(an-2 time.
+ x(an-1)) …)) • Applying Horner’s rule n
times will take Θ(n2)
• Using the Horner’s Rule a time.
polynomial can be
converted from • Example:
coefficient to point-value 6 + 7x + 2x2 = 6 + x(7+x(2))
form in Θ(n2) time.

BITS Pilani, Pilani Campus

 Interpolating using Gaussian
Elimination
• Interpolating polynomial a0 - a1 + a2 = 1 ……….(1)
means finding/deriving a0 = 6 ……….(2)
coefficients-form from a0 + a1 + a2 = 15 ……….(3)
the Point-value form of  a0 = 6, a1 = 7 and a2 = 2.
polynomial. Divide Steps:
• Using Gaussian First eliminate a0,
Elimination we can Then eliminate a1,
perform this conversion then
in Θ(n3) time. Conquer Steps:
• Example: find a2
f(x) = {(-1, 1), (0, 6), (1, 15)} = (a0, And then substitute to solve
a1, a2) = a0 + a1x + a2 x2 backwards.)
BITS Pilani, Pilani Campus
 Gaussian Elimination: Time
Complexity
Divide Step: • There is only one sub-
• Putting values of xn in problem of size (n-1)
previous equations to  T(n) = T(n-1) + Cn(n-1) +
get new (n-1) equations. bn for n > 1
• Hence, total complexity T(1) = d for n = 1
of divide = (n-1)Cn =
Cn(n-1)
Conquer Step:
Computing xn given the
values of x1, x2, … xn-1
Complexity = bn
BITS Pilani, Pilani Campus
 Gaussian Elimination: Time
Complexity
Root Level = 0 T(n) = (Cn2 + (b-c)n) + (C(n-1)2 + (b-
At Level Complexity is c)(n-1)) + (C(n-2)2 + (b-c)(n-2)) + …
0 is Cn2 + (b-c)n + (C(n-i)2 + (b-c)(n-i)) + … + (c22 +
1 is C(n-1)2 + (b-c)(n-1) (b-c).2) + d
2 is C(n-2)2 + (b-c)(n-2) = [Cn(n+1)(2n+1)]/6 – C + (b-
: c)n(n+1)/2 – (b-c) + d
i is C(n-i)2 + (b-c)(n-i)
:
= Θ(n3)
(n-1) is d
Leaf Level = (n-1)

BITS Pilani, Pilani Campus

 Discrete Fourier Transform (DFT): Roots
of unity for algebraic structure Ring

• For any positive integer • 3rd roots of unity =

n, the roots of unity −1+𝑖√3 −1−𝑖√3
are the complex U3 = {1, , }
2 2
number solutions (n in • All nth roots of unity =
number) to the 2πk𝑖/n
Un = { 𝑒 | k ∈ {1, 2, …, n}}
equation xn=1.
For roots: Use Euler equation
• If n is even, there are
e2π𝑖 = cos(2π) + 𝑖sin(2π) = 1
two real solutions 1
and -1. (e2π𝑖)k = 1k = 1 (k, a +ve integer)
• If n is odd, there is one xn =1 = e2π𝑖 = e4π𝑖 = … = e2kπ𝑖
real solution i.e. 1. Raise each to power of (1/n).
BITS Pilani, Pilani Campus
 Discrete Fourier Transform
(DFT)
• The DFT of the vector (a0, Example:
a1, a2, …, an-1) DFT((6,7,2))=DFT((a0,a1,a2))
[corresponding to the • f(x) = 6 + 7x + 2x2
vector form of f(x)] is the
• Three terms so use U3
vector (𝑓(𝜔𝑛0 ), 𝑓(𝜔𝑛1 ), …,
−1+𝑖√3 −1−𝑖√3
𝑓(𝜔𝑛𝑛−1 )) [corresponding => U3 = {1, , }
2 2
to the point-value form
• Substitute each element
of f(x)] where ωn = e2πi/n =
of U3 in the f(x).
nth root of unity.
• DFT((6,7,2)) =
• Complexity of computing
DFT using Horner’s rule is (15, (3+5√3i)/2, (3-5√3i)/2)
Θ(n2). BITS Pilani, Pilani Campus
Inverse DFT

• The DFT-1 of the vector (y0, y1, y2, …, yn-1) is the vector (a0,
a1, a2, …, an-1) if (y0, y1, y2, …, yn-1) = (𝑓(𝜔𝑛0 ), 𝑓(𝜔𝑛1 ), …,
𝑓(𝜔𝑛𝑛−1 )), where ωn = e2πi/n = nth root of unity.
Example: DFT-1(15, (3+5√3i)/2, (3-5√3i)/2) = (6, 7, 2)

We write the DFT equation as:

[y0 y1 y2 ... yn-1]T = A [a0 a1 a2 ... an-1]T ……. (1)
𝑦0 1 1 1 .. 1 𝑎0
𝑦1 1 𝑤𝑛1 𝑤𝑛2 .. 𝑤𝑛𝑛−1 𝑎1
𝑦2 = 1 2(𝑛−1) 𝑎2
𝑤𝑛2 𝑤𝑛4 .. 𝑤𝑛
: : : : .. : :
𝑦𝑛−1 1 𝑤𝑛𝑛−1
2(𝑛−1)
𝑤𝑛
(𝑛−1)(𝑛−1)
. . 𝑤𝑛 𝑎𝑛−1
BITS Pilani, Pilani Campus
 Inverse DFT

We pre-multiply both sides of the equation (1) in the preceding

slide by a matrix B, we get:
B[y0 y1 y2 ... yn-1]T = BA [a0 a1 a2 ... an-1]T= C[a0 a1 a2 ... an-1]T

1 1 1 .. 1
−(𝑛−1)
1 𝑤𝑛−1 𝑤𝑛−2 .. 𝑤𝑛
−2(𝑛−1)
Where, 𝐵 = 1 𝑤𝑛−2 𝑤𝑛−4 .. 𝑤𝑛
: : : .. :
−(𝑛−1) −2(𝑛−1) −(𝑛−1)(𝑛−1)
1 𝑤𝑛 𝑤𝑛 . . 𝑤𝑛
(i, j)th entry of matrix BA (i.e. C) is given by:
(𝑗−𝑖) 2(𝑗−𝑖) (𝑛−1)(𝑗−𝑖)
Cij = 1 + 𝑤𝑛 + 𝑤𝑛 + … + 𝑤𝑛

BITS Pilani, Pilani Campus

 Inverse DFT

For i=j, we get B[y0 y1 y2 ... yn-1]T = nI

Cij = 1 + 1 + .. + 1 = n [a0 a1 a2 ... an-1]T=
n[a0 a1 a2 ... an-1]T
Otherwise, we get
 [a0 a1 a2 ... an-1]T
𝑗−𝑖 𝑛
𝑤𝑛 −1
Cij = (𝑗−𝑖) = =(1/n)B[y0 y1 y2 ...yn-1]T
𝑤𝑛 −1
(𝑤𝑛𝑛 ) 𝑗−1 −1 1−1
= DFT-1(y0, y1, y2, ... yn-1)
(𝑗−𝑖) = (𝑗−𝑖) =0
𝑤𝑛 −1 𝑤𝑛 −1
Complexity of computing
1 0 0 0 DFT-1 (Inverse DFT) using
0 1 0 0 Horner’s rule for matrix
C = nI = n
: : ‫ ܆‬: multiplication = Θ(n2).
0 0 0 1 BITS Pilani, Pilani Campus
 Inverse DFT: Example

• DFT-1(15, (3+5√3i)/2, (3-5√3i)/2). For n = 3, we have

𝑤3−1 =e(-2*π*i/3) = (-1-√3i)/2, and 𝑤3−2 =(-1+√3i)/2

DFT-1(15, (3+5√3i)/2, (3-5√3i)/2) =

1 1 1 15 18 6
1
(1/3) 1 𝑤3−1 𝑤3−2 (3+5√3i)/2 = 21 = 7
3
1 𝑤3−2 𝑤3−4 (3−5√3i)/2) 6 2

BITS Pilani, Pilani Campus

 Fast Fourier Transform

• We separate f(x) into terms having f(𝑤𝑛𝑖 ) = fe((𝑤𝑛𝑖 )2 ) + 𝑤𝑛𝑖

odd and even powers of x as: fo((𝑤𝑛𝑖 )2)
f(x) = fe(x2) + xfo(x2)
= fe((𝑤𝑛2 )i ) + 𝑤𝑛𝑖
Where
fo((𝑤𝑛2 )i)
fe(x) = a0 + a2x + a4x2 + … + an-2x(n/2-1)
𝑖 𝑖
fo(x) = a1 + a3x + a5x2 + … + an-1x(n/2-1) = fe(𝑤𝑛/2 ) + 𝑤𝑛𝑖 fo(𝑤𝑛/2 )
We have,
𝑤𝑛2 =e2(2πi/n) = e2πi/(n/2) = 𝑤𝑛/2 • As i varies from 0 to
𝑛/2
𝑤𝑛 =e2πi/(n/2) = eπi = -1 (n-1), the values of
𝑖
𝑤𝑛/2 repeats at a
distance of n/2.
BITS Pilani, Pilani Campus
Fast Fourier Transform
𝑖+𝑛/2 𝑛/2 𝑖 𝑖
𝑤𝑛/2 = 𝑤𝑛/2 . 𝑤𝑛/2 = 𝑤𝑛/2 ……… 1
𝑖+𝑛/2 𝑛/2
𝑤𝑛 = 𝑤𝑛 . 𝑤𝑛𝑖 = −𝑤𝑛𝑖 ……… 2

Divide Step:
DFT(f(x)) (of size n) = DFT(fe(x)) + DFT(fo(x)) (each of size n/2)

Conquer Step:
Using equations (1) and (2), we write (for 0 ≤ i ≤ (n/2-1))
𝑖 𝑖
f(𝑤𝑛𝑖 ) = fe(𝑤𝑛/2 ) + 𝑤𝑛𝑖 f0(𝑤𝑛/2 )
𝑖+𝑛/2 𝑖 𝑖
f(𝑤𝑛 ) = fe(𝑤𝑛/2 ) - 𝑤𝑛𝑖 f0(𝑤𝑛/2 )

BITS Pilani, Pilani Campus

Fast Fourier Transform: Time
Complexity
• Both Divide and Conquer steps take time linear in input =
Cn
• There are two sub-problems of size n/2. We get the
following recurrence:

T(1) = d
T(n) = 2T(n/2) + Cn for n > 1

BITS Pilani, Pilani Campus

 Fast Fourier Transform: Time
Complexity
Root level = 0 • Divide + Conquer steps takes Cn (linear
At level: time).
• Recursive step has two sub-problems of
0 it is Cn. size n/2.
1 it is 2(Cn/2).
log 𝑛 −1
: T(n) = σ𝑖=0 𝐶𝑛 + dn = Cnlog2(n) +
i it is 2i(Cn/2i) dn
: = ϴ(nlog2(n))
Leaf level it is
d.2log2(n) = dn
Leaf level=(log2n – 1)
BITS Pilani, Pilani Campus
 Recursive FFT & Inverse FFT
S# Statement
Recursive_FFT(a) // a is an array • For Computing DFT using
1. N = a.length // n is a power of 2 Recursive_FFT:
2. If n == 1
3. return a
Line# Change to be made
4. wn = e2πi/n 4 wn = e(-2πi/n)
5. w=1
14 Return (y/n) but in sub-
6. A(0) = (a0, a2, …, an-2)
7. A(1) = (a1, a3, …, an-1)
problems (return y)
8. y(0) = recursive_FFT(A(0) )
9. y(1) = recursive_FFT(A(1) ) • For Inverse FFT (FFT-1):
10. For k=0 to (n/2 – 1)
11. yk = yk(0) + wyk(1)
Line#4: wn = e-2πin
12. yk+(n/2) = yk(0) - wyk(1) Line#14: return (1/n)
13. w=wwn
recursive_FFT-1(a)
14. Return y; BITS Pilani, Pilani Campus
Polynomial Multiplication
using FFT
Example: Given f1(x) = x + 2, and f2(x) = 2x + 3. Multiply.
1: Pad each polynomial with zeros so that number of terms
is the nearest power of 2. (Four terms in this example).
f1(x) = (2, 1, 0, 0), and f2(x) = (3,2,0,0)
2: Find DTF(f1(x)) and DFT(f2(x)). (Four terms in each.)
3: Find product of DFTs of two polynomials. The product is
obtained by multiplying corresponding entries of DFTs.
4: Find DFT-1(DTF(f1(x)).DTF(f2(x))). (i.e. DFT-1 of the
product computed in step 3 above).
5: The result of step 4 is the product of polynomials in the
coefficient form.

BITS Pilani, Pilani Campus

 Polynomial Multiplication
using FFT
Example: Given f1(x) = x + 2, and f2(x) = 2x + 3. Multiply.
1: f1(x) = (2, 1, 0, 0), and f2(x) = (3,2,0,0)
2: DTF(f1(x)) = (3, 2+i, 1, 2-i) and DFT(f2(x)) = (5, 3+2i, 1, 3-2i).
3: P_DFT = DTF(f1(x)). DTF(f2(x)) =
(3x5, (2+i)x(3+2i), 1x1, (2-i)(3-2i)) = (15, 4+7i, 1, 4 – 7i)
4: DFT-1(PDFT) = (6, 7, 2, 0)
5: f(x) = (6, 7, 2, 0) => f(x) = 6 + 7x + 2x2

Verify: f(x) = (x+2)(2x+3) = 2x2 + 4x + 3x +6 = 6 + 7x + 2x2

BITS Pilani, Pilani Campus

 Divide and Conquer Graph of
f1(x)
(2,1,0,0)

(2,0) (1,0)

(2) (0) (1) (0)

w2 = -1
(2+0, 2- 0) = (2, 2) (1+0, 1- 0) = (1, 1)
w4 = i
(2+1, 2+i, 2 – 1, 2 - i)
= (3, 2+i, 1, 2 – i)

BITS Pilani, Pilani Campus

 Divide and Conquer Graph of
f2(x)
(3,2,0,0)

(3,0) (2,0)

(3) (0) (2) (0)

w2 = -1
(3+0, 3- 0) = (3, 3) (2+0, 2- 0) = (2, 2)
w4 = i
(3+2, 3+2i, 3 – 2, 3 - 2i)
= (5, 3+2i, 1, 3 – 2i)

BITS Pilani, Pilani Campus

 Divide and Conquer Graph of
DFT-1(f1(x) f2(x))
(15, 4+7i, 1, 4-7i)

(15,1) (4+7i, 4-7i)

(15) (1) (4+7i) (4-7i)

w2-1= -1

((4+7i) + (4-7i),
(15+1, 15- 0) =
(4+7i) - (4-7i)) =
(16, 15)
(8, 14i)
W4-1 = -i
(1/4)(16+8, 14+14, 16– 8, 14 - 14)
= (6, 7, 2, 0)
BITS Pilani, Pilani Campus
 Thanks ….

Any questions please …..

Thanking you