Polynomials and the

Fast Fourier Transform (FFT)

Algorithm Design and Analysis

(Week 7)

Battle Plan
• Polynomials
– Algorithms to add, multiply and evaluate polynomials
– Coefficient and point-value representation
• Fourier Transform
– Discrete Fourier Transform (DFT) and inverse DFT to
translate between polynomial representations
– “A Short Digression on Complex Roots of Unity”
– Fast Fourier Transform (FFT) is a divide-and-conquer
algorithm based on properties of complex roots of unity

2
 Polynomials
• A polynomial in the variable 𝑥 is a representation of
a function
𝐴 𝑥 = 𝑎𝑛−1 𝑥 𝑛−1 + ⋯ + 𝑎2 𝑥 2 + 𝑎1 𝑥 + 𝑎0
𝑛−1 𝑗
as a formal sum 𝐴 𝑥 = 𝑗=0 𝑎𝑗 𝑥 .
• We call the values 𝑎0 , 𝑎1, … , 𝑎𝑛−1 the coefficients of
the polynomial
• 𝐴 𝑥 is said to have degree 𝑘 if its highest nonzero
coefficient is 𝑎𝑘 .
• Any integer strictly greater than the degree of a
polynomial is a degree-bound of that polynomial

Examples
• 𝐴 𝑥 = 𝑥 3 − 2𝑥 − 1
– 𝐴(𝑥) has degree 3
– 𝐴(𝑥) has degree-bounds 4, 5, 6, … or all values > degree
– 𝐴(𝑥) has coefficients (−1, −2, 0, 1)
• 𝐵 𝑥 = 𝑥3 + 𝑥2 + 1
– 𝐵(𝑥) has degree 3
– 𝐵(𝑥) has degree bounds 4, 5, 6, … or all values > degree
– 𝐵(𝑥) has coefficients (1, 0, 1, 1)

4
 Coefficient Representation
• A coefficient representation of a polynomial
𝐴 𝑥 = 𝑛−1 𝑗
𝑗=0 𝑎𝑗 𝑥 of degree-bound 𝑛 is a vector of
coefficients 𝑎 = 𝑎0 , 𝑎1 , … , 𝑎𝑛−1 .
• More examples
– 𝐴 𝑥 = 6𝑥 3 + 7𝑥 2 − 10𝑥 + 9 (9, −10, 7, 6)
– 𝐵 𝑥 = −2𝑥 3 + 4𝑥 − 5 (−5, 4, 0, −2)
• The operation of evaluating the polynomial 𝐴(𝑥) at
point 𝑥0 consists of computing the value of 𝐴 𝑥0 .
• Evaluation takes time Θ(𝑛) using Horner’s rule
𝐴(𝑥0 ) = 𝑎0 + 𝑥0 (𝑎1 + 𝑥0 (𝑎2 + ⋯ + 𝑥0 𝑎𝑛−2 + 𝑥0 𝑎𝑛−1 ⋯ ))

Adding Polynomials
• Adding two polynomials represented by the
coefficient vectors 𝑎 = (𝑎0 , 𝑎1 , … , 𝑎𝑛−1 ) and
𝑏 = (𝑏0 , 𝑏1 , … , 𝑏𝑛−1 ) takes time Θ(𝑛).
• Sum is the coefficient vector 𝑐 = (𝑐0 , 𝑐1 , … , 𝑐𝑛−1 ),
where 𝑐𝑗 = 𝑎𝑗 + 𝑏𝑗 for 𝑗 = 0,1, … , 𝑛 − 1.
• Example
𝐴 𝑥 = 6𝑥 3 + 7𝑥 2 − 10𝑥 + 9 (9, −10, 7, 6)
𝐵 𝑥 = − 2𝑥 3 + 4𝑥 − 5 (−5, 4, 0, −2)
3 2
𝐶(𝑥) = 4𝑥 + 7𝑥 − 6𝑥 + 4 (4, −6,7,4)

6
 Multiplying Polynomials
• For polynomial multiplication, if 𝐴(𝑥) and 𝐵(𝑥) are
polynomials of degree-bound n, we say their product
𝐶(𝑥) is a polynomial of degree-bound 2𝑛 − 1.
• Example
6𝑥 3 + 7𝑥 2 − 10𝑥 + 9
− 2𝑥 3 + 4𝑥 − 5
− 30𝑥 3 2
− 35𝑥 + 50𝑥 − 45
24𝑥 4 + 28𝑥 3 − 40𝑥 2 + 36𝑥
− 12𝑥 6 − 14𝑥 5 + 20𝑥 4 − 18𝑥 3
− 12𝑥 6 − 14𝑥 5 + 44𝑥 4 − 20𝑥 3 − 75𝑥 2 + 86𝑥 − 45

Multiplying Polynomials
• Multiplication of two degree-bound n polynomials
𝐴(𝑥) and 𝐵(𝑥) takes time Θ 𝑛2 , since each
coefficient in vector 𝑎 must be multiplied by each
coefficient in vector 𝑏.
• Another way to express the product C(x) is
2𝑛−1 𝑗 𝑗
𝑗=0 𝑐𝑗 𝑥 , where 𝑐𝑗 = 𝑘=0 𝑎𝑘 𝑏𝑗−𝑘 .
• The resulting coefficient vector 𝑐 = (𝑐0 , 𝑐1 , … 𝑐2𝑛−1 )
is also called the convolution of the input vectors 𝑎
and 𝑏, denoted as 𝑐 = 𝑎⨂𝑏.

8
 Point-Value Representation
• A point-value representation of a polynomial 𝐴(𝑥)
of degree-bound 𝑛 is a set of 𝑛 point-value pairs
𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 , … , 𝑥𝑛−1 , 𝑦𝑛−1
such that all of the 𝑥𝑘 are distinct and 𝑦𝑘 = 𝐴(𝑥𝑘 )
for 𝑘 = 0, 1, … , 𝑛 − 1.
• Example 𝐴 𝑥 = 𝑥 3 − 2𝑥 + 1
– 𝑥𝑘 0, 1, 2, 3
* 0, 1 , 1, 0 , 2, 5 , 3, 22 +
– 𝐴(𝑥𝑘 ) 1, 0, 5, 22
• Using Horner’s method, 𝒏-point evaluation takes
time Θ(𝑛2 ).

Point-Value Representation
• The inverse of evaluation is called interpolation
– determines coefficient form of polynomial from point-
value representation
– For any set * 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 , … , 𝑥𝑛−1 , 𝑦𝑛−1 + of 𝑛 point-
value pairs such that all the 𝑥𝑘 values are distinct, there is
a unique polynomial 𝐴(𝑥) of degree-bound 𝑛 such that
𝑦𝑘 = 𝐴(𝑥𝑘 ) for 𝑘 = 0, 1, … , 𝑛 − 1.
• Lagrange’s formula
𝑛−1
𝑗≠𝑘(𝑥 − 𝑥𝑗 )
𝐴 𝑥 = 𝑦𝑘
𝑘=0 𝑗≠k(𝑥𝑘 − 𝑥𝑗 )
• Using Lagrange’s formula, interpolation takes time
Θ(𝑛2 ).

10
 Example
• Using Lagrange’s formula, we interpolate the point-
value representation 0, 1 , 1, 0 , 2, 5 , 3, 22 .
𝑥−1 (𝑥−2)(𝑥−3) 𝑥 3 −6𝑥 2 +11𝑥−6 −𝑥 3 +6𝑥 2 −11𝑥+6
– 1 (0−1)(0−2)(0−3) = =
−6 6
(𝑥−0)(𝑥−2)(𝑥−3)
– 0 (1−0)(1−2)(1−3) = 0
(𝑥−0)(𝑥−1)(𝑥−3) 𝑥 3 −4𝑥 2 +3𝑥 −15𝑥 3 +60𝑥 2 −45𝑥
– 5 (2−0)(2−1)(2−3) = 5 =
−2 6
(𝑥−0)(𝑥−1)(𝑥−2) 𝑥 3 −3𝑥 2 +2𝑥 22𝑥 3 −66𝑥 2 +44𝑥
– 22 (3−0)(3−1)(3−2) = 22 =
6 6
1
– 6𝑥 3 + 0𝑥 2 − 12𝑥 + 6
6
– 𝑥 3 − 2𝑥 + 1

11

Adding Polynomials
• In point-value form, addition 𝐶 𝑥 = 𝐴 𝑥 + 𝐵(𝑥) is
given by 𝐶 𝑥𝑘 = 𝐴 𝑥𝑘 + 𝐵(𝑥𝑘 ) for any point 𝑥𝑘 .
– 𝐴: 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 , … , 𝑥𝑛−1 , 𝑦𝑛−1
– 𝐵: 𝑥0 , 𝑦0′ , 𝑥1 , 𝑦1′ , … , 𝑥𝑛−1 , 𝑦𝑛−1
′

– 𝐶: * 𝑥0 , 𝑦0 + 𝑦0′ , 𝑥1 , 𝑦1 + 𝑦1′ , … , 𝑥𝑛−1 , 𝑦𝑛−1 + 𝑦𝑛−1

′
+
• 𝐴 and 𝐵 are evaluated for the same 𝑛 points.
• The time to add two polynomials of degree-bound 𝑛
in point-value form is Θ(𝑛).

12
 Example
• We add 𝐶 𝑥 = 𝐴 𝑥 + 𝐵(𝑥) in point-value form
– 𝐴 𝑥 = 𝑥 3 − 2𝑥 + 1
– 𝐵 𝑥 = 𝑥3 + 𝑥2 + 1
– 𝑥𝑘 = (0, 1, 2, 3)
– 𝐴: 0, 1 , 1, 0 , 2, 5 , 3, 22
– 𝐵: * 0, 1 , 1, 3 , 2, 13 , 3, 37 +
– 𝐶: * 0, 2 , 1, 3 , 2, 18 , 3, 59 +

13

Multiplying Polynomials
• In point-value form, multiplication 𝐶 𝑥 = 𝐴 𝑥 𝐵(𝑥)
is given by 𝐶 𝑥𝑘 = 𝐴 𝑥𝑘 𝐵(𝑥𝑘 ) for any point 𝑥𝑘 .
• Problem: if 𝐴 and 𝐵 are of degree-bound 𝑛, then 𝐶 is
of degree-bound 2𝑛.
• Need to start with “extended” point-value forms for
𝐴 and 𝐵 consisting of 2𝑛 point-value pairs each.
– 𝐴: 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 , … , 𝑥2𝑛−1 , 𝑦2𝑛−1
– 𝐵: 𝑥0 , 𝑦0′ , 𝑥1 , 𝑦1′ , … , 𝑥2𝑛−1 , 𝑦2𝑛−1
′

– 𝐶: * 𝑥0 , 𝑦0 𝑦0′ , 𝑥1 , 𝑦1 𝑦1′ , … , 𝑥2𝑛−1 , 𝑦2𝑛−1 𝑦2𝑛−1

′
+
• The time to multiply two polynomials of degree-
bound 𝑛 in point-value form is Θ(𝑛).

14
 Example
• We multiply 𝐶 𝑥 = 𝐴 𝑥 𝐵(𝑥) in point-value form
– 𝐴 𝑥 = 𝑥 3 − 2𝑥 + 1
– 𝐵 𝑥 = 𝑥3 + 𝑥2 + 1
– 𝑥𝑘 = (−3, −2, −1, 0, 1, 2, 3) We need 7 coefficients!
– 𝐴: −3, −17 , −2, −3 , −1,1 , 0, 1 , 1, 0 , 2, 5 , 3, 22
– 𝐵: * −3, −20 , −2, −3 , −1, 2 , 0, 1 , 1, 3 , 2, 13 , 3, 37 +
– 𝐶: * −3,340 , −2,9 , −1,2 , 0, 1 , 1, 0 , 2, 65 , 3, 814 +

15

The Road So Far

𝑎0 , 𝑎1 , … , 𝑎𝑛−1 Ordinary multiplication Coefficient
Time Θ(𝑛2 )
𝑐0 , 𝑐1 , … , 𝑐𝑛−1
𝑏0 , 𝑏1 , … , 𝑎𝑛−1 representations

Evaluation Interpolation
Time Θ(𝑛2 ) Time Θ(𝑛2 )

𝐴 𝑥0 , 𝐵 𝑥0 𝐶 𝑥0
𝐴 𝑥1 , 𝐵 𝑥1 Pointwise multiplication 𝐶 𝑥1 Point-value
⋮ Time Θ(𝑛) ⋮ representations
𝐴 𝑥2𝑛−1 , 𝐵(𝑥2𝑛−1 ) 𝐶(𝑥2𝑛−1 )

• Can we do better?
– Using Fast Fourier Transform (FFT) and its inverse, we can do
evaluation and interpolation in time Θ(𝑛 log 𝑛).
• The product of two polynomials of degree-bound 𝑛 can
be computed in time Θ(𝑛 log 𝑛), with both the input and
output in coefficient form.
16
 Fourier Transform
• Fourier Transforms originate from signal processing
– Transform signal from time domain to frequency domain
Amplitude

Weight
Time Frequency
– Input signal is a function mapping time to amplitude
– Output is a weighted sum of phase-shifted sinusoids of
varying frequencies

17

Fast Multiplication of Polynomials

• Using complex roots of unity
– Evaluation by taking the Discrete Fourier Transform (DFT)
of a coefficient vector
– Interpolation by taking the “inverse DFT” of point-value
pairs, yielding a coefficient vector
– Fast Fourier Transform (FFT) can perform DFT and inverse
DFT in time Θ(𝑛 log 𝑛)
• Algorithm
1. Add 𝑛 higher-order zero coefficients to 𝐴(𝑥) and 𝐵(𝑥)
2. Evaluate 𝐴(𝑥) and 𝐵(𝑥) using FFT for 2𝑛 points
3. Pointwise multiplication of point-value forms
4. Interpolate 𝐶(𝑥) using FFT to compute inverse DFT

18
 Complex Roots of Unity
• A complex 𝒏th root of unity (1) is a complex number
𝜔 such that 𝜔𝑛 = 1.
• There are exactly 𝑛 complex 𝑛th root of unity
2𝜋𝑖𝑘
𝑒 𝑛 for 𝑘 = 0, 1, … , 𝑛 − 1

where 𝑒 𝑖𝑢 = cos 𝑢 + 𝑖 sin(𝑢). Here 𝑢 represents

an angle in radians.
2𝜋𝑖𝑘
• Using 𝑒 𝑛 = cos 2𝜋𝑘 + 𝑖 sin(2𝜋𝑘 𝑛), we can
𝑛
check that it is a root
2𝜋𝑖𝑘 𝑛
𝑒 𝑛 = 𝑒 2𝜋𝑖𝑘 = cos(2𝜋𝑘) +𝑖 sin(2𝜋𝑘) = 1
1 0

19

Examples
• The complex 4th roots of unity are
1, −1, 𝑖, −𝑖
where −1 = 𝑖.
• The complex 8th roots of unity are all of the above,
plus four more
1 𝑖 1 𝑖 1 𝑖 1 𝑖
+ , − ,− + , and − −
2 2 2 2 2 2 2 2
• For example
2
1 𝑖 1 2𝑖 𝑖 2
+ = + + =𝑖
2 2 2 2 2

20
 Principal 𝑛th Root of Unity
• The value
2𝜋𝑖
𝜔𝑛 = 𝑒 𝑛

is called the principal 𝒏th root of unity.

• All of the other complex 𝑛th roots of unity are powers
of 𝜔𝑛 .
• The 𝑛 complex 𝑛th roots of unity, 𝜔𝑛0 , 𝜔𝑛1 , … , 𝜔𝑛𝑛−1 ,
form a group under multiplication that has the same
structure as (ℤ𝑛 , +) modulo 𝑛.
• 𝜔𝑛𝑛 = 𝜔𝑛0 = 1 implies
𝑗 𝑗+𝑘 𝑗+𝑘 mod 𝑛
– 𝜔𝑛 𝜔𝑛𝑘 = 𝜔𝑛 = 𝜔𝑛
– 𝜔𝑛−1 = 𝜔𝑛𝑛−1
21

Visualizing 8 Complex 8th Roots of Unity

imaginary
𝜔82 = 𝑖
𝑖
1 𝑖 1 𝑖
𝜔83 =− + 𝜔18 = +
2 2 2 2

𝜔84 = −1 𝜔80 = 𝜔88 = 1

−1 1 real

1 𝑖 1 𝑖
𝜔85 = − − 𝜔87 = −
2 2 2 2
−𝑖

𝜔86 = −𝑖

22
 Cancellation Lemma
• For any integers 𝑛 ≥ 0, 𝑘 ≥ 0, and 𝑏 > 0,
𝑑𝑘
𝜔𝑑𝑛 = 𝜔𝑛𝑘 .
• Proof
2𝜋𝑖 𝑑𝑘 2𝜋𝑖 𝑘
𝑑𝑘
𝜔𝑑𝑛 = 𝑒 𝑑𝑛 = 𝑒 𝑛 = 𝜔𝑛𝑘
𝑛
• For any even integer 𝑛 > 0, 𝜔𝑛 2 = 𝜔2 = −1.
6
• Example 𝜔24 = 𝜔4
6 6
= 𝑒 2𝜋𝑖 24 = 𝑒
6 2𝜋𝑖 2𝜋𝑖
– 𝜔24 = 𝑒 24 4 = 𝜔4

23

Halving Lemma
• If 𝑛 > 0 is even, then the squares of the 𝑛 complex
𝑛th roots of unity are the 𝑛 2 complex 𝑛 2th roots of
unity.
• Proof
2
– By the cancellation lemma, we have 𝜔𝑛𝑘 = 𝜔𝑛𝑘 2 for any
nonnegative integer 𝑘.
• If we square all of the complex 𝑛th roots of unity,
then each 𝑛 2th root of unity is obtained exactly twice
𝑘+𝑛 2 2
– 𝜔𝑛 2
= 𝜔𝑛2𝑘+𝑛 = 𝜔𝑛2𝑘 𝜔𝑛𝑛 = 𝜔𝑛2𝑘 = 𝜔𝑛𝑘
𝑘+𝑛
– Thus, 𝜔𝑛𝑘 and 𝜔𝑛 2
have the same square

24
 Summation Lemma
• For any integer 𝑛 ≥ 1 and nonzero integer 𝑘 not
𝑛−1 𝑗
divisible by 𝑛, 𝑗=0 𝜔𝑛𝑘 = 0.
• Proof
𝑛−1 𝑗 𝑥 𝑛 −1
– Geometric series 𝑗=0 𝑥 = 𝑥−1
𝑘 𝑛 𝑛 𝑘 −1
𝑛−1 𝑗 𝜔𝑛 −1 𝜔𝑛 1 𝑘 −1
– 𝑗=0 𝜔𝑛𝑘 = 𝑘 −1 = 𝑘 −1 = 𝑘 −1 =0
𝜔𝑛 𝜔𝑛 𝜔𝑛
– Requiring that 𝑘 not be divisible by 𝑛 ensures that the
denominator is not 0, since 𝜔𝑛𝑘 = 1 only when k is divisible
by 𝑛

25

Discrete Fourier Transform (DFT)

• Evaluate a polynomial 𝐴(𝑥) of degree-bound 𝑛 at the
𝑛 complex 𝑛th roots of unity, 𝜔𝑛0 , 𝜔𝑛1 , 𝜔𝑛2 , … , 𝜔𝑛𝑛−1 .
– assume that 𝑛 is a power of 2
– assume 𝐴 is given in coefficient form 𝑎 = (𝑎0 , 𝑎1 , … , 𝑎𝑛−1 )
• We define the results 𝑦𝑘 , for 𝑘 = 0, 1, … , 𝑛 − 1, by
𝑛−1 𝑘𝑗
𝑦𝑘 = 𝐴 𝜔𝑛𝑘 = 𝑗=0 𝑎𝑗 𝜔𝑛 .
• The vector 𝑦 = (𝑦0 , 𝑦1 , … , 𝑦𝑛−1 ) is the Discrete
Fourier Transform (DFT) of the coefficient vector
𝑎 = 𝑎0 , 𝑎1 , … , 𝑎𝑛−1 , denoted as 𝑦 = DFT𝑛 (𝑎).

26
 Fast Fourier Transform (FFT)
• Fast Fourier Transform (FFT) takes advantage of the
special properties of the complex roots of unity to
compute DFT𝑛 (a) in time Θ(𝑛 log 𝑛).
• Divide-and-conquer strategy
– define two new polynomials of degree-bound 𝑛 2, using
even-index and odd-index coefficients of 𝐴(𝑥) separately
𝑛
– 𝐴 0 𝑥 = 𝑎0 + 𝑎2 𝑥 + 𝑎4 𝑥 2 + ⋯ + 𝑎𝑛−2 𝑥 2−1
𝑛
– 𝐴 1 𝑥 = 𝑎1 + 𝑎3 𝑥 + 𝑎5 𝑥 2 + ⋯ + 𝑎𝑛−1 𝑥 2−1

– 𝐴 𝑥 = 𝐴 0 𝑥 2 + 𝑥𝐴 1 (𝑥 2 )

27

Fast Fourier Transform (FFT)

• The problem of evaluating 𝐴(𝑥) at 𝜔𝑛0 , 𝜔𝑛1 , … , 𝜔𝑛𝑛−1
reduces to
1. evaluating the degree-bound 𝑛 2 polynomials 𝐴 0 (𝑥)
and 𝐴 1 (𝑥) at the points 𝜔𝑛0 2 , 𝜔𝑛1 2 , … , 𝜔𝑛𝑛−1 2
2. combining the results by 𝐴 𝑥 = 𝐴 0 𝑥 2 + 𝑥𝐴 1 (𝑥 2 )
• Why bother?
– The list 𝜔𝑛0 2 , 𝜔𝑛1 2 , … , 𝜔𝑛𝑛−1 2 does not contain 𝑛
distinct values, but 𝑛 2 complex 𝑛 2th roots of unity
– Polynomials 𝐴 0 and 𝐴 1 are recursively evaluated at the
𝑛 complex 𝑛 th roots of unity
2 2
– Subproblems have exactly the same form as the original
problem, but are half the size

28
 Example
• 𝐴 𝑥 = 𝑎0 + 𝑎1 𝑥 + 𝑎2 𝑥 2 + 𝑎3 𝑥 3 of degree-bound 4
– 𝐴 𝜔40 =𝐴 1 = 𝑎0 + 𝑎1 + 𝑎2 + 𝑎3
– 𝐴 𝜔41 =𝐴 𝑖 = 𝑎𝑜 + 𝑎1 𝑖 − 𝑎2 − 𝑎3 𝑖
– 𝐴 𝜔42 =𝐴 −1 = 𝑎0 − 𝑎1 + 𝑎2 − 𝑎3
– 𝐴 𝜔43 =𝐴 −𝑖 = 𝑎0 − 𝑎1 𝑖 + 𝑎2 + 𝑎3 𝑖
• Using 𝐴 𝑥 = 𝐴 0 𝑥 2 + 𝑥𝐴 1 𝑥 2
– 𝐴 𝑥 = 𝑎0 + 𝑎2 𝑥 2 + 𝑥 𝑎1 + 𝑎3 𝑥 2
– 𝐴 𝜔40 = 𝐴 1 = 𝑎0 + 𝑎2 + 1(𝑎1 + 𝑎3 )
– 𝐴 𝜔41 = 𝐴 𝑖 = 𝑎0 − 𝑎2 + 𝑖(𝑎1 − 𝑎3 )
– 𝐴 𝜔42 = 𝐴 −1 = 𝑎0 + 𝑎2 − 1 𝑎1 + 𝑎3
– 𝐴 𝜔43 = 𝐴 −𝑖 = 𝑎0 − 𝑎2 − 𝑖(𝑎1 − 𝑎3 )
29

Recursive FFT Algorithm

RECURSIVE-FFT 𝑎
1 𝑛 ← 𝑙𝑒𝑛𝑔𝑡ℎ,𝑎- 𝑛 is a power of 2
2 if 𝑛 = 1
3 then return 𝑎 basis of recursion
2𝜋𝑖
4 𝜔𝑛 ← 𝑒 𝑛 𝜔𝑛 is principal 𝑛th root of unity
5 𝜔←1
6 𝑎 0 ← (𝑎0 , 𝑎2 , … , 𝑎𝑛−2 )
7 𝑎 1 ← (𝑎1 , 𝑎3 , … , 𝑎𝑛−1 )
0
8 𝑦 0 ← RECURSIVE-FFT 𝑎 0 𝑦𝑘 = 𝐴 0 𝜔𝑛𝑘 2 = 𝐴 0 (𝜔𝑛2𝑘 )
1
9 𝑦 1 ← RECURSIVE-FFT 𝑎 1 𝑦𝑘 = 𝐴 1 𝜔𝑛𝑘 2 = 𝐴 1 (𝜔𝑛2𝑘 )
10 for 𝑘 ← 0 to 𝑛 2 − 1
0 1
11 do 𝑦𝑘 ← 𝑦𝑘 + 𝜔𝑦𝑘
0 1 𝑘+ 𝑛 2
12 𝑦𝑘+(𝑛 2) ← 𝑦𝑘 − 𝜔𝑦𝑘 since −𝜔𝑛𝑘 = 𝜔𝑛
13 𝜔 ← 𝜔𝜔𝑛 compute 𝜔𝑛𝑘 iteratively
14 return 𝑦

30
 Why Does It Work?
• For 𝑦0 , 𝑦1 , … 𝑦𝑛 2−1 (line 11)
0 1
𝑦𝑘 = 𝑦𝑘 + 𝜔𝑛𝑘 𝑦𝑘
0
=𝐴 𝜔𝑛2𝑘 + 𝜔𝑛𝑘 𝐴 1 𝜔𝑛2𝑘
=𝐴 𝜔𝑛𝑘
• For 𝑦𝑛 2 , 𝑦𝑛 2+1 , … , 𝑦𝑛−1 (line 12)
𝑦𝑘+𝑛 2 = 𝑦𝑘0 − 𝜔𝑛𝑘 𝑦𝑘1
𝑘+ 𝑛 2 𝑘+ 𝑛 2
= 𝑦𝑘0 + 𝜔𝑛 𝑦𝑘1 since −𝜔𝑛𝑘 = 𝜔𝑛
𝑘+ 𝑛
0
=𝐴 𝜔𝑛2𝑘 + 𝜔𝑛 2
𝐴 1 𝜔𝑛2𝑘
𝑘+ 𝑛
0
=𝐴 𝜔𝑛2𝑘+𝑛 + 𝜔𝑛 2
𝐴 1 𝜔𝑛2𝑘+𝑛
since 𝜔𝑛2𝑘+𝑛 = 𝜔𝑛2𝑘
𝑘+ 𝑛 2
= 𝐴 𝜔𝑛

31

Input Vector Tree of RECURSIVEFFT(𝑎)

(𝑎0 , 𝑎1 , 𝑎2 , 𝑎3 , 𝑎4 , 𝑎5 , 𝑎6 , 𝑎7 )

(𝑎0 , 𝑎2 , 𝑎4 , 𝑎6 ) (𝑎1 , 𝑎3 , 𝑎5 , 𝑎7 )

(𝑎0 , 𝑎4 ) (𝑎2 , 𝑎6 ) (𝑎1 , 𝑎5 ) (𝑎3 , 𝑎7 )

(𝑎0 ) (𝑎4 ) (𝑎2 ) (𝑎6 ) (𝑎1 ) (𝑎5 ) (𝑎3 ) (𝑎7 )

32
 Interpolation
• Interpolation by computing the inverse DFT, denoted
by a = DFT𝑛−1 (𝑦).
• By modifying the FFT algorithm, we can compute
DFT𝑛−1 in time Θ(𝑛 log 𝑛).
– switch the roles of 𝑎 and 𝑦
– replace 𝜔𝑛 by 𝜔𝑛−1
– divide each element of the result by 𝑛

33