ME591

Random Data
Analysis and Processing

Lecture 4

Discrete FT,
Impulse Response,
FRF(Frequency Response Function)
ME591 Random Data  Prof. Yong-Hwa Park, KAIST
 Review: Fourier Transform, Finite Fourier transform, Fourier Series
x(t )

t X ( f ) =  x(t )e− j 2 ft dt
−

xT (t )
T
X T ( f ) = X ( f , T ) =  x(t )e − j 2 ft dt
0
T

xP (t )
1 T 1

− j 2 f k t
Ak = x (t ) e dt = X ( fk ,T )
T 0 T
T 2T 1
( f k = k f = k , k = 0, 1, 2,...)
2
T
1 X ( fk ,T )
(auto) power spectral density: S xx ( f k , T ) = Ak = T Ak = k = 0, 1, 2,.......
2 2
,
f T
2
 X ( f ,T ) 1
S xx ( f ) =  Rxx ( )e − j 2 f  d = lim = lim E[ X s ( f , T ) ], s = 1, 2,3,...
2
− T → T T → T

ME591 Random Data 2  Prof. Yong-Hwa Park, KAIST

 Discrete Fourier Transform
Discretized version of Finite Fourier Transform applied to a pseudo-random signal

T
X ( f , T ) =  x(t )e − j 2 ft dt
0


N
 kn 
X ( f k ) = X (k f ) =  xn exp  − j 2 t
n =1  N
Where, discretization parameters are:

T 1
t = f = n, k = 1, 2,..., N
N T

ME591 Random Data 3  Prof. Yong-Hwa Park, KAIST

 Continuous vs Discrete Fourier Transform
Continuous F.T. Discrete F.T.
non-periodic, random suitable function pseudo-random (record length T)
dt sampling time interval t = h
t (continuous) time tn = nt = nh; n = 1, 2,....., N (discrete)
infinite period T=Nh (finite)
1 1
f → df loop frequency, Hz f = =
T Nh
1 1
f (continuous) Frequency, Hz f k = k f = k =k (discrete)
T Nh
N N
 1
 (
−
)df frequency domain integration  ( )f =  ( ) Nh
k =1 k =1
N N

 ( )t =  ( )h

 ( )dt
−
time domain integration
n =1 n =1
  kn 
X ( f ) =  x(t )e
N
− j 2 ft X ( f k ) = X (k f ) = h xn exp  − j 2  1
dt Fourier Transform Xk  X ( fk )
− n =1  N h

 N
 kn 
x(t ) =  X ( f )e j 2 ft
df Inverse Fourier Transform xn = x(nt ) = f  X ( f k ) exp  j 2 
− k =1  N

continuous spectrum discrete

ME591 Random Data 4  Prof. Yong-Hwa Park, KAIST

 ‘Conjugate even’ Properties

xn + N = xn (periodic, pseudo-random assumption) for all n

X ( fk )
For X k = : X − k = X k* , X k + N = X k (periodic) for all k
h
N
↑Note 1 X N − k = X k* , X ( N /2) + k = X (*N /2) − k for k = 1,2,3,......,
2
xN = x0
• •
k 1 2 3 N/2-1 N/2 N/2+1 N-3 N-2 N-1 N
• •
Re{Xk} ar br cr gr Nyquist gr cr br ar mean
• •
Im{Xk} ai bi ci gi 0 -gi -ci -bi -ai 0

Real (time) FT pair Conjugate Even (freq.)

1 Note that X ( f ,T ) is the finite Fourier transform of x(t) defined over the time period (0, T),
X ( f k ,T )
X k is the computational version of the original discrete Fourier transform X ( f , T ) . Thus it holds that
k
Xk = .
h

ME591 Random Data 5  Prof. Yong-Hwa Park, KAIST

 Test for Invariance of PSD : Discrete Random Signal
PSD of pseudo-random signal with record length T:
 2 X ( f k , T ) 2  2h N
 kn 
2
N
Gxx ( f k )  2S xx ( f k ) = E  = E[  x(nh) exp  − j 2  ], k = 1, 2,3,....,
 T  N n =1  N 2

Random Input (normal distributed random signal)

Input: Normal distributed normalized random;
Sampling interval h = 1 ms;

Bandpass(BP)-filtered Output time (sec)

System: Butterworth BP filter; 4th, low 20 Hz, high 50 Hz;
xn : (discrete) pseudo random signal
No. of averaging = 100;
Hanning window with 50 % overlapping;

N, number of data for FFT, varied;

→ T, record length varied correspondingly.
time (sec)
ME591 Random Data 6  Prof. Yong-Hwa Park, KAIST
 h = 1ms
Min freq. of signal = 20 Hz → T1=0.05 sec
N = 128; MS = 0.0582 N = 512; MS = 0.0630 Rule of thumb T > 10 T1 = 0.5 sec
T=Nh, N> 500

PSD: Gxx ( f k )

Recall temporal mean square

frequency (Hz) frequency (Hz) Parseval’s theorem
(see section 5.2):
1 T
N = 2048; MS = 0.0613 N = 8192; MS = 0.0622  2x =
T 0
x 2 (t )dt



2
= Ak
k =−

PSD: Gxx ( f k ) =  S xx ( f )df
−

=  Gxx ( f )df
0

All integration (summation) tell

About “power of signal”, which
frequency (Hz) frequency (Hz) Is invariant to record length)

ME591 Random Data 7  Prof. Yong-Hwa Park, KAIST

 Impulse response and FRF
Delta Function:  (t )
 t = 0  
Properties:  (t ) = 
0 t  0
 −
 (t )dt = 1 
−
x(t ) (t − t0 )dt = x(t0 )


Fourier transform:  −
 (t ) e− j 2ft dt = 1 for all f Impulse (force)

  (t ) at =0
Inverse Fourier transform : −
e j 2ft
df =  (t ) 1
impulse response function, h(t)
0

Impulse Response Function of a linear system : h(t ) -1

0 10 20 30 40
  time (sec)
Convolution integral : y (t ) = − h( ) x(t −  )d = − h(t −  ) x( )d

Causal (Physically realizable) function : h ( ) = 0 for   0 → y (t ) =  h( ) x(t −  )d
0


Stability condition : 
−
h( ) d  

ME591 Random Data 8  Prof. Yong-Hwa Park, KAIST

 Causal function and linear superposition
 (t )  (t − 10)
1

force =0 =10

0
0 10 20 30 40
1
=0 impulse response function h(t) f1 =  (t )
h1

0
y1 = h(t )
-1
0 10 20 30 40
1
=10 f 2 =  (t − 10)
h2

0
y2 = h(t − 10)
-1
0 10 20 30 40
1 f =  (t ) +  (t − 10)
h1+h2

0 y = y1 + y2
-1 = h(t ) + h(t − 10)
0 10 20 30 40
Time
ME591 Random Data 9  Prof. Yong-Hwa Park, KAIST
 Properties of impulse function:
Dirac delta δ(t)
F0
Dirac Delta
F(t) function
(theory) Equal
impulses
F (t −  ) = F0 (t −  ) = 0, t   in practice
 
−
F (t −  )dt =  F0 (t −  )dt =F0
−

 t

ME591 Random Data 10  Prof. Yong-Hwa Park, KAIST

 Response to Arbitrary Input
Viewed as a series of impulses of magnitude F (ti ) t

xi Response at time t due to the ith impulse

xi (t ) = F (ti )th(t − ti ) for t  ti
ti t
F(t) Response from linear superposition
N
Impulses x(t ) =  F (ti )th(t − ti )
“the i-th impulse” i =1
F(ti) F (ti )t t → dt , ti →  
t 
x(t ) =  F ( )h(t −  )d =  F ( )h(t −  )d
0 −
t convolution integral

t1,t2 ,t3 ti t=tN t h(t ) = 0, F (t ) = 0 for t  0

ME591 Random Data 11  Prof. Yong-Hwa Park, KAIST

 Impulse response and FRF
Frequency Response Function :

H ( f ) =  h( )e− j 2f d
−
cf ) H (− f ) = H ( f )*: " Conjugate Even "
H ( f ) = H R ( f ) + jH I ( f ) = H ( f ) e j ( f ) cf ) e j ( f ) = cos  ( f ) + j sin  ( f )

Gain : H ( f ) = HR2 ( f ) + HI2 ( f )

 HI ( f ) 
Phase :  ( f ) = tan −1  
 R
H ( f ) 

Convolution vs. Fourier Transform : 

Y( f ) = H ( f )X ( f )  y (t ) =  x( )h(t −  )d

−

Transfer function : H ( s ) =  h( )e − s d , s =  + j , ( = 2 f )
0

Transfer function vs. Frequency response function: H ( f ) =  H ( s ) 

s = j = j 2 f

ME591 Random Data 12  Prof. Yong-Hwa Park, KAIST

 Bandwidth
fr 1
Q − factor = 
Looking into a single mode: f 2 − f1 2 r
Half-power-point Bandwidth : 3 7 2
H ( fr )
Br = f 2 − f1  2f r 2.5 H ( fr ) 6

2 2 1 2
5 Br = f 2 − f1  2f r
H ( f1 ) = H ( f 2 ) = H( fr ) 2 H ( fr )
2
2 4

power
gain

2
f r : resonance frequency; 1.5 H ( fr )
3 2
 : damping ratio
1
2

0.5 1
ar fr fr
H( f ) = 2
r −  2 + 2 j rr
0 0
( = 2 f ) 0 f1 1 f2 2 3 4 0 f1 1 f 2 2 3 4
frequency frequency

ME591 Random Data 13  Prof. Yong-Hwa Park, KAIST

 Highlight 4
⚫ Discrete Fourier Transform as a discretized version of Finite Fourier Transform (continuous)

⚫ Conjugate even complex spectrum corresponding to a real-valued time signal

⚫ Invariance of PSD and MS of stationary random signal to a record length T

⚫ Dirac delta function, causal function and impulse response function

⚫ Response to an arbitrary input in terms of convolution integral of force and impulse response function

⚫ Frequency response function (FRF) as a Fourier Transform of impulse response function (IRF)

ME591 Random Data 14  Prof. Yong-Hwa Park, KAIST