68

7.2 Discrete Fourier Transform(DFT)

(See section 9.3 of textbook)
• In real world,
– We must often perform numerical computation since the real signals can be rarely
described in mathematical expression.
– For numerical computations to be possible, both signal and spectrum are discrete.
⇒ We require the discrete spectrum calculation for discrete signals.
⇒ In numerical computation, both signals spectrums are periodic and discrete.

• Review
– Fourier Transform (FT): Continuous and Non-periodic in time domain ⇔Non-periodic
and Discrete in freq. domain

– Fourier Series (FS) : Continuous and Periodic in time domain ⇔Non-periodic and
Discrete in freq. domain

– Discrete Time Fourier Transform (DTFT) : Discrete and Non-periodic in time domain
⇔Periodic and Continuous in freq. domain
 69

• Development of numerical computation of discrete signal spectrums

– Starting from a periodic continuous-time signal x(t) = x(t + T0 ) (that is, whose
spectrum is discrete.)

– Discretizing x(t) = x(t + T0 ) with sampling time interval Ts where N = T0 /Ts

⇒ Approximation of Fourier series for a discrete & periodic signa becomes

∫ To
1 ∑
N −1
1
Xk = x(t)e jkω0 t
dt = lim x(nTs )e−jkω0 n·Ts · Ts
To 0 Ts →0 N Ts n=0
∑
N −1
x(nTs ) −jkω0 nTs
∼
= ·e
n=0
N
– Normalize time and frequency :
∑
N −1 ∑
N −1
−jnkω̂0
X[k] = x[n] · e = x[n] · e−j2πnk/N
n=0 n=0

where x[n] = x(nTs )/N, ω̂0 = ω0 · Ts = 2π/T0 · Ts = 2π/N, and n = t/Ts

– Recovery of the discrete signal x[n] from discrete spectrum X[k]:


∑
N −1  N :n=l 1 ∑
N −1
j(n−l) 2π 2π
Orthogonality : e N
k
= ⇒ x[n] = X[k]ejnk N
 0 : n ̸= l N
k=0 k=0

∗ The periodicity of signal and spectrum :

∑
N −1 ∑
n−1
−jn(k+N ) 2π
x[n]e−jnk N e−jn2π = X[k] ⇒ X[k + N ] = X[k]
2π
⋄ X[k + N ] = x[n]e N =
n=0 n=0

⋄ In similiar way, x[n] = x[n + N ]

 70

• Discrete Fourier Transform (DFT)

Discrete Fourier Transform (DFT)

∑
N −1
- DFT : X[k] = x[n]e−jk2πn/N
n=0

1 ∑
N −1
- IDFT :x[n] = X[k]ejk2πn/N
N k=0
where
- x[n] = {x(nTs )/N }, X[k] = { T1o X(kω0 )}
- ω0 = 2π/T0 , ωs = 2π/Ts = 2πfs
- N = T0 /Ts , ω̂0 = ω0 · Ts = 2π/N
- x[n] = x[n + mN ], X[k] = X[k + mN ]

∗ FFT(Fast Fourier Transform) is a fast computation of DFT when N = 2n .

– EX: DFT of x[n] = ej(2πk0 /N )n for n = 0, · · · , N − 1


∑
N −1 ∑
N −1  N :k=k
−j(2π/N )kn j(2π/N )(k0 −k)n 0
X[k] = x1 [n]e = e = = N δ[k − k0 ]
 0 : k ̸= k0
n=0 n=0

( 2πk n ) 1 { }
ej(2πk0 n/N ) + e−j(2πk0 n/N ) , where k0 < N .
0
– EX: DFT of x[n] = cos =
N 2
N
X[k] = {δ[k − k0 ] + δ[k + k0 ]}
2

• Frequency sampling ⇒In frequency domain, DFT is the discrete version of DTFT.
∑
N −1
2π
X[k] = x[n]ej N kn = X(ej ω̂ ) 2π
ω̂=k N
n=0
71
 72

• Using the DFT in real world

1) Properly truncate or window the target signal to have a proper signal duration.
⇒Signal duration is To .
2) Determine the signal bandwidth fB as observing the aliasing effect.

3) Decide the number of data for the signal duration To .

From the Nyquist theorem, the initial sampling time Ts′ ≈ 1
.
⌈ ⌉ 2fB

So, the number of data N = TT0′ .

s

4) Tune Ts′ with N for obtaining the actual sampling time Ts . ⇒Ts = T0
N
.
5) Acquire the finte discrete signal x[n] with sampling time Ts .
6) Regard x[n] as the periodic signal x̃[n] = x[n + m · N ] for m = 0, ±1, ±2 . . . .
7) Perform N -point DFT
∑
N −1 ∑
N −1
−j( 2π
x[n]e−j( N )kn k = 0, 1, · · · , N − 1
2π
)kn
X[k] = x̃[n]e N =
n=0 n=0
 73


 |t| for |t| < 1
• Ex: A signal f (t) =
 0 for otherwise

Let F (ω) the Fourier spectrum of f (t). The bandwidth of F (ω) is |ω| < 7π/3. That is,
F (ω) w 0 for |ω| > 7π/3. We must compute f (t) and F (ω) numerically.

1) Decide sampling time interval Ts′ to avoid the aliasing on freq. domain.
1 1 6 3
From Nyquist theorem : Ts′ < = × =
2fB 2 7 7
2) Decide frequency-sampling interval ωs′ to avoid the aliasing on time-domain.
1 1
T0 = 2 ⇒ f0 < = ⇒ ω0 = π
T0 2
3) Determine the number of sampling points N .
⌈ ( 3 )⌉ ⌈ 14 ⌉
N = ⌈T0 /Ts′ ⌉ = 2/ = = ⌈4.666⌉ = 5
7 3
4) Decide the sampling time Ts for numerical computation.
2
Ts = T0 /N =
5
5) Plot discrete signal f[n]. (Remember we discretize both f (t) and F (ω).)

6) Compute the discrete Fourier Transform (DFT) F [k]

1 ∑ 1 ∑
2
−jk 2π
f [n]e−jk 5 n for k = 0, 1, 2, 3, 4
2π
n
F [k] = f [n]e N =
N n=<N >
5 n=−2

1 1{ −j 2π k j 2π
} 1{
−j 4π k j 4π
}
= f [0] + f [1]e 5 + f [−1]e 5 + f [2]e 5 + f [−2]e 5
5 5 5
2 4
Since f [0] = 0, f [1] = f [−1] = , f [2] = f [−2] =
5 5
4 [ 2π ] 8 [ 4π ]
= cos k + cos k
25 5 25 5
 74

 [ ( )]
 0.5 1 − cos 2πn 0≤n≤L−1
L
• Ex: DFT X[k] of x[n] =
 0 Otherwise
 75

• Properties of DFT
– DFT{ax[n] + bx[n]} = aX[k] + bX[k]
– For real x[n] X[k] = X ∗ [N − k]
– Circular time shift : DFT {xL(n − L)|N } = WNkL X[k]
– DFT{ej2πnM/N x[n]} = X[(k − M )N ]
∑
N −1
– Circular convolution : {x1 [n]~x
˜ 2 [n]} = x1 [k]x[(n − k)N ]
k=0
– DFT {x1 [n]~x
˜ 2 [n]} = X1 [k] · X2 [k]
– DFT{x[n] · y[n]} = 1
N
X[k]~Y
˜ [k]
– Parseval Theorem :
∑
N −1
1 ∑ ∗
N −1 ∑
N −1
1 ∑
n−1
x∗ [n]y[n] = X [k] · Y [k] ⇒ |x[n]|2 = |X[k]|2
n=0
N k=0 n=0
N k=0
76
77
78