Professional Documents
Culture Documents
1 Fourier Series
· · · = f (t − 2T ) = f (t − T ) = f (t) = f (t + T ) = f (f + 2T ) = · · · (1)
Any periodic function may be represented as a series expansion of sines and cosines, in a
Fourier series,
∞ ∞
1 X 2πqt X 2πqt
f (t) = a0 + aq cos + bq sin , (2)
2 q=1 T q=1 T
where the Fourier coefficients, aq and bq are given by the Fourier integrals,
2 Z to +T 2πqt
aq = f (t) cos dt , q = 0, 1, 2, . . . (3)
T to T
2 Z to +T 2πqt
bq = f (t) sin dt , q = 0, 1, 2, . . . (4)
T to T
and the time to is arbitrary. This section emphasizes the connection between real and
complex-valued notations for an oscillation. The Fourier series (2) may be represented using
complex exponential notation.
∞
1 2πqt 2πqt
X
f (t) = a0 + aq cos + bq sin (5)
2 q=1 T T
q=∞
2πq
X
= Fq exp i t (6)
q=−∞ T
q=∞
Fq eiωq t
X
= (7)
q=−∞
∗
as long as Fq = F−q .
2 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
Proof:
q=∞
X
f (t) = Fq eiωq t (8)
q=−∞
q=0
X q=∞
X
= Fq eiωq t + Fq eiωq t (9)
q=−∞ q=1
q=∞ n=∞
F−q e−iωq t +
X X
= Fq eiωq t (10)
q=0 q=1
∞ h i
Fq eiωq t + F−q e−iωq t
X
= F0 + (11)
q=1
∞ h i
(Fq0 + iFq00 ) (cos ωq t + i sin ωq t) + (Fq0 − iFq00 ) (cos ωq t − i sin ωq t)
X
= f0 + (12)
q=1
∞ h i
2Fq0 cos ωq t − 2Fq00 sin ωq t
X
= F0 + (13)
q=1
So, the real part of Fq , Fq0 , is half of aq , the imaginary part of Fq , Fq00 , is half of −bq , and
∗
Fq = F−q . The complex coefficients may be found directly from
1 Z to +T 2πiqt
Fq = f (t) exp − dt . (14)
T to T
The function H(ω) is called the frequency response function for the dynamic system relating
the input f (t) to the output x(t).
Because the oscillator is linear, if the response to f1 (t) is x1 (t), and the response to f2 (t)
is x2 (t), then the response to c1 f1 (t) + c2 f2 (t) is c1 x1 (t) + c2 x2 (t). More generally, then,
q=∞
1
Fq eiωq t
X
x(t) = (17)
q=−∞ (k − mωq2 ) + i(cωq )
where ωq = 2πq/T .
q
Note that ωq is not the same symbol as ωn ; ωn = k/m, ωq = 2πq/T .
The series expansion for the response x(t) converges with fewer terms than the Fourier
series for the external forcing, because |H(ωq )| decreases as 1/ωq2 .
Dynamic Periodic Response to Periodic Forcing 3
2 Fourier Transforms
where the Fourier coefficients, Fq , have the same units as f (t), and are given by the Fourier
integral,
1 Z T /2
Fq = f (t) e−iωq t dt , (19)
T −T /2
in which the interval of integration arbitrarily starts at −T /2.
Now, consider a change of variables, by defining a frequency increment, ∆ω, and a scaled
amplitude, F (ωq ).
∆ 2π
∆ω = ω1 = (ωq = q ∆ω) (20)
T
∆ 2π
F (ωq ) = T Fq = Fq (21)
∆ω
Where the scaled amplitude, F (ωq ), has units of f (t) · t or f (t)/ω.
1 q=∞
F (ωq ) eiωq t ∆ω ,
X
f (t) = (22)
2π q=−∞
Z T /2
F (ωq ) = f (t) e−iωq t dt . (23)
−T /2
1 Z∞
f (t) = F (ω) eiωt dω , (24)
2π Z −∞
∞
F (ω) = f (t) e−iωt dt . (25)
−∞
These expressions are the famous Fourier transform pair. Equation (24) is commonly called
the inverse Fourier transform and equation (25) is commonly called the forward Fourier
transform. They differ only by the sign of the exponent and the factor of 2π.
By convention, the forward fast Fourier transform (FFT) of an N -point time series of
duration T (xk = x((k − 1)∆t), k = 1, · · · , N ) scales the N , complex-valued, Fourier am-
plitudes/coefficients as follows: FFT(x) = X(ωq )/∆t = N Xq , where the Fourier-transform
frequencies, ωq , are given by ωq = 2πq/T , and are sorted as follows:
q = 0, · · · , N/2, −N/2 + 1, · · · , −1.
4 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
3 Fourier Approximation
Any periodic function may be approximated as a truncated series expansion of sines and
cosines, as a Fourier series,
Q Q
˜ 1 X 2πqt X 2πqt
f (t) = a0 + aq cos + bq sin , (26)
2 q=1 T q=1 T
where the Fourier coefficients, aq and bq may be found by solving the Fourier integrals,
2 Z T /2 2πqt
aq = f (t) cos dt , q = 0, 1, 2, . . . , Q (27)
T −T /2 T
2 Z T /2 2πqt
bq = f (t) sin dt , q = 0, 1, 2, . . . , Q (28)
T −T /2 T
The Fourier approximation (26) may also be represented using complex exponential notation.
q=Q
1 q=Q
f˜(t) = Fq eiωq t = F (ωq ) eiωq t
X X
(29)
q=−Q T q=−Q
where eiωq t = cos ωq t + i sin ωq t, ωq = 2πq/T , Fq = (aq − ibq )/2, F−q = Fq∗ , and
1 Z T /2
Fq = f (t) e−iωq t dt . (30)
T −T /2
Xq = H(ωq ) Fq , (31)
4 Examples
In the following examples, the external forcing is periodic with a period, T , of 2π (sec-
onds). For functions periodic in 2π seconds, the frequency increment in the Fourier series is
1 radian/second. In this case, the frequency index number, q, is also the frequency, ωq , in
radians per second, otherwise, ωq = 2πq/T .
1/ωn2
H(ω) = , (34)
1 − Ω2 + i 2ζΩ
In the (a) plots, the solid line represents the exact form of f (t), the dashed lines represent
the real-valued form of the Fourier approximation and the complex-valued form of the Fourier
approximation, and the circles represent 2Q sample points of the function f (t) for use in fast
Fourier transform (FFT) computations. The two dashed lines are exactly equal.
In the (b) plots, the impulse-lines show the values of the Fourier coefficients, Fq , found by
evaluating the Fourier integral, equation (30), and the circles represent the Fourier coefficients
returned by the FFT.
In the (c) plots, the red solid lines show the cosine terms of the Fourier series and the
blue dashed lines show the sine terms of the Fourier series.
The (d) and (e) plots show the magnitude and phase of the transfer function, H(ω) as a
solid line, the Fourier coefficients, Fq , as the green circles, and H(ωq )Fq as the blue circles.
In the (f) plots, the solid line represents x(t) as computed using equation (32), the circles
represent the real part of the result of the inverse FFT calculation and the dots represent
the imaginary part of the result of the inverse FFT calculation.
The numerical details involved in the correct representation of periodic functions and
frequency indexing for FFT computations are provided in the attached Matlab code.
6 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
1.5 0
-0.1
1
-0.3
0
-0.4
-0.5
-0.5
-1
-0.6
-1.5 -0.7
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(a) time, t, (s) (b) frequency, ω, (rad/s)
1.5 1
1
0.1
cosine and sine terms
frequency response
0.5
0 0.01
-0.5
0.001
-1
-1.5 1e-04
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(c) time, t, (s) (d) frequency, ω, (rad/s)
200 0.025
0.02
150
0.015
100
phase lead, (degrees)
0.01
response, x, (m)
50
0.005
0 0
-0.005
-50
-0.01
-100
-0.015
-150
-0.02
-200 -0.025
0 2 4 6 8 10 12 14 16 -3 -2 -1 0 1 2 3
(e) frequency, ω, (rad/s) (f) time, t, (s)
Figure 1. (a): — = f ext (t); - - = f˜ext (t). (b): red solid = <(Fq ); blue dashed = =(Fq );
*: Fq from FFT. (c): red solid = component cosine terms, blue dashed = component sine
terms. (d): — = |H(ω)|; *: |Fq | and |H(ωq )Fq |. (e): — = phase lead of H(ω); *: phase
leads of Fq and H(ωq )Fq . (f): — = x(t); *: x(t) from IFFT; · · · : =(x(t)) from IFFT.
Dynamic Periodic Response to Periodic Forcing 7
4 0.6
3 0.4
1 0
0 -0.2
-1 -0.4
-2 -0.6
-3 -0.8
-4 -1
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(a) time, t, (s) (b) frequency, ω, (rad/s)
2 1
1.5
1 0.1
cosine and sine terms
frequency response
0.5
0 0.01
-0.5
-1 0.001
-1.5
-2 1e-04
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(c) time, t, (s) (d) frequency, ω, (rad/s)
150 0.04
100
0.02
50
phase lead, (degrees)
0
response, x, (m)
0
-0.02
-50
-0.04
-100
-0.06
-150
-200 -0.08
0 2 4 6 8 10 12 14 16 -3 -2 -1 0 1 2 3
(e) frequency, ω, (rad/s) (f) time, t, (s)
Figure 2. (a): — = f ext (t); - - = f˜ext (t). (b): red solid = <(Fq ); blue dashed = =(Fq );
*: Fq from FFT. (c): red solid = component cosine terms, blue dashed = component sine
terms. (d): — = |H(ω)|; *: |Fq | and |H(ωq )Fq |. (e): — = phase lead of H(ω); *: phase
leads of Fq and H(ωq )Fq . (f): — = x(t); *: x(t) from IFFT; · · · : =(x(t)) from IFFT.
8 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
2 0.7
1.5 0.6
0.5 0.4
0 0.3
-0.5 0.2
-1 0.1
-1.5 0
-2 -0.1
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(a) time, t, (s) (b) frequency, ω, (rad/s)
1.5 1
1
0.1
cosine and sine terms
frequency response
0.5
0.01
0.001
-0.5
1e-04
-1
-1.5 1e-05
-3 -2 -1 0 1 2 3 0 2 4 6 8 10 12 14 16
(c) time, t, (s) (d) frequency, ω, (rad/s)
200 0.02
150 0.015
100 0.01
phase lead, (degrees)
response, x, (m)
50 0.005
0 0
-50 -0.005
-100 -0.01
-150 -0.015
-200 -0.02
0 2 4 6 8 10 12 14 16 -3 -2 -1 0 1 2 3
(e) frequency, ω, (rad/s) (f) time, t, (s)
Figure 3. (a): — = f ext (t); - - = f˜ext (t). (b): red solid = <(Fq ); blue dashed = =(Fq );
*: Fq from FFT. (c): red solid = component cosine terms, blue dashed = component sine
terms. (d): — = |H(ω)|; *: |Fq | and |H(ωq )Fq |. (e): — = phase lead of H(ω); *: phase
leads of Fq and H(ωq )Fq . (f): — = x(t); *: x(t) from IFFT; · · · : =(x(t)) from IFFT.
Dynamic Periodic Response to Periodic Forcing 9
71 end
72
73 i f strcmp( type , ’ sawtooth ’ ) % s a w t o o t h wave
74 b = -(2./ q ) .* ( -1).ˆ q ;
75 f_true = t ; f_true (1) = 0; f_true (2* P +1) = 0;
76 f_samp = ts ;
77 f_samp (1) = 0; % g e t p e r i o d i c i t y r i g h t , f (− p i ) = 0
78 end
79
80 i f strcmp( type , ’ triangle ’ ) % t r i a n g l e wave
81 a = -(2./( pi * q .ˆ2)) .* ( ( -1).ˆ q - 1 );
82 f_true = pi /2 - t .* sign ( t );
83 f_samp = pi /2 - ts .* sign ( ts );
84 end
85
86
87 f_approxR = a * cos (q ’* t ) + b * s i n (q ’* t ); % r e a l Fourier s e r i e s
88
89 Fq = ( a - i * b )/2; % complex F o u r i e r c o e f f i c i e n t s
90
91 % Complex F o u r i e r s e r i e s ( p o s i t i v e and n e g a t i v e e x p o n e n t s )
92 f_approxC = Fq * exp( i * wq ’* t ) + conj ( Fq ) * exp( - i * wq ’* t );
93
94
95 % The i m a g n i a r y p a r t o f t h e complex F o u r i e r s e r i e s i s e x a c t l y z e r o !
96 i m a g _ o v e r _ re al _ 1 = max(abs(imag( f_approxC ))) / max(abs( r e a l ( f_approxC )))
97
98
99 % The f a s t F o u r i e r t r a n s f o r m (FFT) method
100 % In Matlab , t h e f o r w a r d F o u r i e r t r a n s f o r m has a n e g a t i v e e x p o n e n t .
101 % According t o a c o n v e n t i o n o f t h e FFT method , i n d e x number 1 i s f o r time = d t
102 % t s (Q+1)=0 ; t s (2∗Q) = (Q−1)∗ d t = T/2− d t ; t s ( 1 ) = −Q∗ d t = −T/2 ; t s (Q) = −d t
103 % . . . s e e t h e t a b l e a t t h e end o f t h i s f i l e .
104
105 F_FFT = f f t ( [ f_samp ( Q +1:2* Q ) f_samp (1: Q ) ] ) / (2* Q ); % Fourier coeff ’ s
106
107
108 % c o e f = [ Fq ’ F FFT ( 2 :Q+1) ’ ] % d i s p l a y the Fourier c o e f f i c i e n t s
109
110 % −−−−−−−− P l o t t i n g −−−−−−−−−−−
111
112 f i g u r e (1);
113 plot (t , f_true , ’ -r ’ , ts , f_samp , ’* r ’ , t , f_approxR , ’ -c ’ , t , f_approxC , ’ -b ’ )
114 xlabel ( ’ time , t , ( s ) ’)
115 ylabel ( ’ function , f ( t ) , ( N ) ’)
116 axis (1.05*[ - pi , pi ])
117 legend ( ’ true ’ , ’ sampled ’ , ’ real Fourier approx ’ , ’ complex Fourier approx ’)
118 plot ( wq , r e a l ( Fq ) , ’ˆ r ’ , wq ,imag( Fq ) , ’ˆ b ’ , ...
119 wq , r e a l ( F_FFT (2: Q +1)) , ’* r ’ , wq ,imag( F_FFT (2: Q +1)) , ’* b ’ )
120 ylabel ( ’ Fourier coefficients , F ( w ) , ( N ) ’)
121 xlabel ( ’ frequency , {/ Symbol w } , ( rad / s ) ’)
122 legend ( ’ real ( F ) ’ , ’ imag ( F ) ’ , ’ real ( FFT ( f )) ’ , ’ imag ( FFT ( f )) ’)
123
124
125 f i g u r e (2);
126 plot (t ,a ’* ones (1 ,2* P +1).* cos (q ’* t ) , ’ -r ’ , t ,b ’* ones (1 ,2* P +1).* s i n (q ’* t ) , ’ -b ’ );
127 xlabel ( ’ time , t , ( s ) ’)
128 ylabel ( ’ cosine and sine terms ’)
129 axis (1.05*[ - pi , pi ])
130
131
132 % =======================================================================
133 % Steady−s t a t e r e s p o n s e o f a SDOF o s c i l l a t o r t o g e n e r a l p e r i o d i c f o r c i n g .
134
135 % Complex−v a l u e d f r e q u e n c y r e s p o n s e f u n c t i o n , H(w) , has u n i t s o f [m/N]
136 H = (1/ wn ˆ2) ./ ( 1 - ( wq / wn ).ˆ2 + 2* i * z *( wq / wn ) );
137
138 % Complex F o u r i e r s e r i e s ( p o s i t i v e and n e g a t i v e e x p o n e n t s )
139 x_approxC = ( H .* Fq ) * exp( i * wq ’* t ) + ( conj ( H ) .* conj ( Fq )) * exp( - i * wq ’* t );
140
141 % In Matlab , t h e i n v e r s e F o u r i e r t r a n s f o r m has a p o s i t i v e e x p o n e n t .
Dynamic Periodic Response to Periodic Forcing 11
142 % . . . s e e t h e t a b l e a t t h e end o f t h i s f i l e f o r t h e s o r t i n g c o n v e n t i o n .
143 x_FFT = i f f t ( [ (1/ wn ˆ2) , H , conj ( H (Q -1: -1:1)) ] .* F_FFT ) * (2* Q );
144
145 f i g u r e (3); % p l o t frequency response functions
146 semilogy ( wq ,abs( Fq ) , ’* g ’ , wq ,abs( H ) , ’ -r ’ , wq ,abs( H .* Fq ) , ’* b ’ )
147 xlabel ( ’ frequency , {/ Symbol w } , ( rad / s ) ’)
148 ylabel ( ’ frequency response ’)
149 legend ( ’| F | ’ , ’| H | ’ , ’| H * F | ’)
150
151 plot ( wq , angle ( Fq )*180/ pi , ’* g ’ , wq , angle ( H )*180/ pi , ’ -r ’ , wq , angle ( H .* Fq )*180/ pi , ’* b ’)
152 xlabel ( ’ frequency , {/ Symbol w } , ( rad / s ) ’)
153 ylabel ( ’ phase lead , ( degrees ) ’)
154
155 f i g u r e (4); % p l o t time r e s p o n s e
156 ts = [ ts ( Q +1:2* Q ) ts (1: Q ) ]; % re−s o r t t h e time a x i s
157 plot ( t , x_approxC , ’ -r ’ , ts , r e a l ( x_FFT ) , ’* b ’ , ts , imag( x_FFT ) , ’. b ’ )
158 axis (1.05*[ - pi , pi ])
159 xlabel ( ’ time , t , ( s ) ’)
160 ylabel ( ’ response , x , ( m ) ’)
161 legend ( ’x ( t ) - complex Fourier series ’ , ’ real ( x ( t )) - FFT ’ , ’ imag ( x ( t )) - FFT ’)
162
163
164 % The i m a g i n a r y p a r t i s e x a c t l y z e r o f o r t h e complex e x p o n e n t s e r i e s method !
165 i m a g _ o v e r _ rea l _ 2 = max(abs(imag( x_approxC ))) / max(abs( r e a l ( x_approxC )))
166
167 % The i m a g i n a r y p a r t i s p r a c t i c a l l y z e r o f o r t h e FFT method !
168 i m a g _ o v e r _ rea l _ 3 = max(abs(imag( x_FFT ))) / max(abs( r e a l ( x_FFT )))
169
170 % −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
171 % SORTING o f t h e FFT c o e f f i c i e n t s number o f d a t a p o i n t s = N = 2∗Q
172 % note : f max = 1/(2∗ d t ) ; d f = f max /(N/2) = 1/(N∗ d t ) = 1/T;
173 %
174 %
175 % index time frequency
176 % ===== ==== =========
177 % 1 0 0
178 % 2 dt df
179 % 3 2∗ d t 2∗ d f
180 % 4 3∗ d t 3∗ d f
181 % : : :
182 % N/2−1 (N/2−2)∗ d t f max −2∗ d f = ( N/2−2)∗ d f
183 % N/2 (N/2−1)∗ d t f max−d f = ( N/2−1)∗ d f
184 % N/2+1 (N/2)∗ d t +/− f max = ( N/2 ) ∗ d f
185 % N/2+2 (N/2+1)∗ d t −f max+d f = (−N/2+1)∗ d f
186 % N/2+3 (N/2+2)∗ d t −f max+2∗ d f = (−N/2+2)∗ d f
187 % : : : :
188 % N−2 (N−3)∗ d t −3∗ d f
189 % N−1 (N−2)∗ d t −2∗ d f
190 % N (N−1)∗ d t −d f
191 % −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
12 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
5 Random Loading
The frequency content of a random process is characterized by its power spectral density.
The power spectral density of a random process of finite duration, T , is the expectation of
the process’s squared Fourier amplitudes divided by its duration
M M Z T
1 1 1 X 1 1 X
Pxx (ωq ) = E[|X(ωq )|2 ] = lim |Xm (ωq )|2 = lim xm (t)e−iωq t dt .
T T M →∞ M m=1 T M →∞ M m=1 0
(38)
where xm (t) is a record in a sample of M records representative of the process. Unit-intensity
white noise, u(t), is a random process with a power spectral density of unity over all frequen-
cies, so E[|U (ωq )|2 ] = T . If u(t) is periodic in T then E[|U (ωq )|2 ] = |Um (ωq )|2 ∀ n, and
there is no uncertainty in the Fourier magnitudes, |U (ωq )|. In such a situation, the uncer-
tainty lies entirely in the ratio of the real to the imaginary parts of U (ωq ), i.e., the phase,
φq = tan−1 (U 00 (ωq )/U 0 (ωq )). The complex Fourier amplitudes of unit-intensity noise with
period T (and no spectral magnitude variability) may therefore be computed from
√ √
U (ωq ) = U 0 (ωq ) + iU 00 (ωq ) = T cos φq + i T sin φq ; (39)
If u(t) is periodic (or if u(t) has finite duration and is assumed to be periodic) the
response y(t) of a system with frequency response H(ω) to u(t) is also periodic and may be
approximated from the finite series
1 q=Q
H(ωq ) U (ωq ) eiωq t .
X
y(t) = (40)
T q=−Q
where ωq = 2πq/T . If U (ω) is the Fourier transform of a unit-intensity random process for
which E[|U (ωq )|2 ] = |U (ωq )|2 , the periodic response may be computed from
q q=Q
X
y(t) = 2 ∆f |H(ωq )| cos(ωq t + θq ) , (41)
q=1
where θq is uniformly distributed between −π and π. This response has a period T and
∆f = 1/T .
Since the frequency response, H(ω) represents the input-output relationship of real-valued
functions, H(−ω) = H ∗ (ω), U (−ω) = U ∗ (ω), and Y (−ω) = Y ∗ (ω). The system frequency
response function may be represented in terms of real and imaginary parts, H(ω) = H 0 (ω) +
iH 00 (ω) where H 0 (ω) = |H(ω)| cos ψ(ω) and H 00 (ω) = |H(ω)| sin ψ(ω), and where |H(ω)| is
the magnitude of the frequency response function and ψ(ω) is the phase of the frequency
response function. The following derivation makes use of the notation U (ωq )/T = Uq and
H(ωq )/T = Hq .
Dynamic Periodic Response to Periodic Forcing 13
Breaking the sum of equation (40) into negative and non-negative values of the index q,
1 q=−1 1 q=Q
H(ωq ) U (ωq ) eiωq t + H(ωq ) U (ωq ) eiωq t
X X
y(t) =
T q=−Q T q=0
−1
X Q
X
= T Hq Uq (cos ωq t + i sin ωq t) + T Hq Uq (cos ωq t + i sin ωq t)
q=−Q q=0
Q Q
Hq∗ Uq∗ (cos ωq t
X X
= T − i sin ωq t) + T Hq Uq (cos ωq t + i sin ωq t)
q=1 q=0
= T H0 U0 (cos ω0 t + i sin ω0 t) +
Q
(Hq0 Uq0 − Hq00 Uq00 − i(Hq0 Uq00 + Hq00 Uq0 ))(cos ωq t − i sin ωq t) +
X
T
q=1
Q
(Hq0 Uq0 − Hq00 Uq00 + i(Hq0 Uq00 + Hq00 Uq0 ))(cos ωq t + i sin ωq t)
X
T
q=1
Q
Hq0 Uq0 cos ωq t − Hq00 Uq00 cos ωq t − i(Hq0 Uq00 + Hq00 Uq0 ) cos ωq t +
X
= T H0 U0 + T
q=1
− iHq0 Uq0 sin ωq t + iHq00 Uq00 sin ωq t − (Hq0 Uq00 + Hq00 Uq0 ) sin ωq t +
Q
Hq0 Uq0 cos ωq t − Hq00 Uq00 cos ωq t + i(Hq0 Uq00 + Hq00 Uq0 ) cos ωq t +
X
T
q=1
iHq0 Uq0 sin ωq t − iHq00 Uq00 sin ωq t − (Hq0 Uq00 + Hq00 Uq0 ) sin ωq t
Q
(Hq0 Uq0 − Hq00 Uq00 ) cos ωq t − (Hq0 Uq00 + Hq00 Uq0 ) sin ωq t
X
= T H0 U0 + 2T
q=1
√ XQ
= T H0 U0 + 2 T (Hq0 cos φq − Hq00 sin φq ) cos ωq t − (Hq0 sin φq + Hq00 cos φq ) sin ωq t
q=1
√ XQ
= T H0 U0 + 2 T Hq0 (cos φq cos ωq t − sin φq sin ωq t) − Hq00 (cos φq sin ωq t + sin φq cos ωq t)
q=1
√ XQ
= T H0 U0 + 2 T Hq0 cos(ωq t + φq ) − Hq00 sin(ωq t + φq )
q=1
√ XQ
= T H0 U0 + 2 T |Hq | cos ψq cos(ωq t + φq ) − |Hq | sin ψq sin(ωq t + φq )
q=1
√ XQ
= T H0 U0 + 2 T |Hq | cos(ωq t + φq + ψq ) (42)
q=1
Since cosine is periodic in [−π : π] and since φ is uniformly distributed within [−π : π], and
assuming either H0 = 0 or U0 = 0, equation (42) simplifies to equation (41)
q q=Q
X
y(t) = 2 ∆f |H(ωq )| cos(ωq t + θq )
q=1
where θq is uniformly distributed in [−π : π]. Note, again, that over a period T = 2π/ω1 =
1/(∆f ), this response is periodic. With frequencies ωq = 2πq/T , and 0 ≤ t < T , the
14 CEE 699.04, ME 555.05. System Identification – Duke University – Fall 2013 – H.P. Gavin
magnitude of the Fourier transform of the response to unit-intensity random input, y(t), as
computed from equation (41) recovers |H(ω)| exactly: E[|Y (ωq )|2 ] = |Y (ωq )|2 = |H(ωq )|2 .
The response of a linear system to non-periodic or persistent white noise is not periodic.
A strictly-proper linear system,
has the frequency response H(ω) = C(iωI − A)−1 B, which, in turn, has a realization
" #
A B
H(ω) ∼ . (45)
C 0
Its response may be simulated in discrete-time via numerical integration methods. If the
time series u(tk ) and y(tk ) have N points, simulating white noise response of the system (45)
requires the generation of N uncorrelated standard Gaussian random values, u(tk ). Sampled
unit Gaussian white noise with√ time step ∆t is normally distributed with a mean of zero and
a standard deviation of 1/ ∆t, has a mean square value of 1/(∆t), and a power spectral
density of 1 over a frequency range −1/(2∆t) < f < 1/(2∆t).