Professional Documents
Culture Documents
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 1 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 2 / 52
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 3 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 4 / 52
Time-Frequency Time-Frequency
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 5 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 6 / 52
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 7 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 8 / 52
Can We Find an Ideal TFD? WVD
Wigner-Ville Distribution Marginal
Z Z Z
R( ) = s (t)s(t + )dt = s (t )s(t)dt P(t, )dt = |S()|2
Z
= s t s t+ dt
2 2 Z Z Z
1
P(t, )d = Rt ( ) exp(j )d d
Now define: 2
Z Z
1
1
Z = Rt ( ) exp(j )dd
P(t, ) = Rt ( ) exp(j )d. 2
2 1
Z
= Rt ( ) 2( )d
where 2
Rt ( ) = s t s t+ = Rt (0) = |s(t)|2 .
2 2
Note that Rt ( ) = Rt ( ) in this definition. Why? (answer real) This is Wigner-Ville Distribution (WVD).
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 9 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 10 / 52
Ex 1.1: Sinusoid
Generalized time-dependent autocorrelation function:
s(t) = ej0 t , W (t, ) = ( 0 )
Z Z
1
Rt ( ) = exp(j(u t))(, )s u s u+ dud Ex 1.2: Impulse:
2 2 2
i.e. s(t) = 2(t t0 ), W (t, ) = (t t0 )
Z
1
Rt ( ) = r (u t, )s u s u+ du Ex 1.3: Pure chirp (linear FM):
2 2 2
where 2 /2+j
Z s(t) = ejt 0t
.
r (u, ) = exp(ju)(, )d.
what is the instantaneous frequency?
For symmetry, the kernel function must take:
W (t, ) = ( t 0 ).
(, ) = (, )
Great!
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 11 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 12 / 52
Why TFD? Examples WVD the tale/sum of two signals
Ex 1.4: Sinusoid with Gaussian amplitude (AM):
s(t) = s1 (t) + s2 (t)
1/4 t 2 /2+j0 t
s(t) = (/) e Rt ( ) =?
2 /2+j t
+ A2 (2 /)1/4 e2 t 2
.
More messy!!!
Also not positive.
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 17 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 18 / 52
Definitions: Z
2
If total energy is preserved, i.e., T = t2 = (t < t >)2 |s(t)|2 dt
Z Z
2
P(t, )ddt = 1, B = 2 = ( < >)2 |S()|2 d
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 23 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 24 / 52
Uncertainty Principle Uncertainty Principle
Proof Proof cont
Without loss of generality, assume the signal is zero mean time and Now use Schwarz inequality
2
zero mean frequency, i.e., we can normalize a signal as following
Z Z Z
2 2
|f (x)| dx |g(x)| dx f (x)g(x)dx ,
Snew (t) = ej<>(t+<t>) sold (t+ < t >).
we have Z 2
The bandwidth is thus t2 2
0
ts (t)s (t)dt .
Z Z
2 = 2
|S()| dt = 2
|s0 (t)|2 dt. Now note s(t) = A(t)ej(t) ,
1d 1
The duration is ts (t)s0 (t) = tA0 A + jt0 A2 = (tA2 ) A2 + jt0 (t)A2 .
2 dt 2
Z
t2 = t 2 |s(t)|2 dt.
Take integration, the first term is zero (why? perfect differential).
Therefore Assume signal is normalized, then
Z Z
2 2
t2 2 = t 2 |s(t)|2 dt |s0 (t)|2 dt.
Z
2 2
0
1 1 2
t ts (t)s (t)dt = + jCovt = + Covt .
2 4
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 25 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 26 / 52
cts(t) = s0 (t).
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 27 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 28 / 52
Uncertainty Principle Uncertainty Principle
Examples: Chirp Examples: Chirp cont - Calculation tricks
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 29 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 30 / 52
1d
Bandwidth in general: s(t) = [jt + t + 0 ]s(t).
j dt
Z
2 = ( < >)2 |S()|2 d Therefore,
Z
2
s (t)(jt + t + 0 )s(t)dt
Z
1d <> =
= s (t) < > s(t)dt
j dt r Z
2
(jt + t + 0 )et dt = 0 .
2
=
Z
1d
= < > s(t) dt
j dt
Also,
2 + 2
r Z
2
< > =2
|jt + t + 0 |2 et dt = + 02 ,
2
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 31 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 32 / 52
Uncertainty Principle The Short-Time Fourier Transform
Examples: Chirp cont Spectrogram
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 33 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 34 / 52
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 35 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 36 / 52
The Spectrogram The Spectrogram
Marginals Total Energy
Time:
Z
P(t) = PSP (t, )d
Z
1 0
Z Z Z Z
= s( )h( t)s ( 0 )h ( 0 t)ej( ) d d 0 d ESP = PSP (t, )dtd = 2
|s(t)| dt |h(t)|2 dt
2
Z
= s( )h( t)s ( 0 )h ( 0 t)( 0 )d d 0
i.e., STFT preserves the energy (Parsevals identity).
Not difficult to prove STFT preserves the inner product, i.e.,
Z
= |s( )|2 |h( t)|2 d Z Z Z
s(t)f (t)dt = St ()Ft ()dtd
Frequency:
Z
P() = |S( 0 )|2 |H( 0 )|2 d 0
Not good!
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 37 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 38 / 52
R
STFT is invertible. Let h(t)dt = 1, then
2
Z
s(t) = s(t)h(t )d Choose Gaussian window: h(t) = (/)1/4 et /2 .
Z Z Recall uncertainty: most compact joint time-frequency.
1 Gabor Transforms
= S ()ej dd
2
Admissible low pass window
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 39 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 40 / 52
The Short-Time Fourier Transform The Short-Time Fourier Transform
Examples The tale/sum of two signals
Sinusoid: s(t) = ej0 t . Using Gabor transform, STFT is s(t) = s1 (t) + s2 (t)
Z
1
( 0 )2
1
St () = e j(0 )t
exp St () = ej st ( )d
()1/4 2 2 Z
1
= ej [s1 ( ) + s2 ( )]h( t)d
PSP (t, ) = |St ()|2 . 2
= S1,t () + S2,t ().
Impulse: s(t) = 2(t t0 ).
2 /2 Note the spectrogram is then:
St () = h(t0 t)ejt0 = (/)1/4 ejt0 e(tt0 )
PSP (t, ) = |St ()|2 = |S1,t () + S2,t ()|2
PSP (t, ) = |St ()|2 .
= |S1,t ()|2 + |S2,t ()|2 + S1,t ()S2,t () + S1,t ()S2,t ()
Recall uncertainty principle.
= |S1,t ()|2 + |S2,t ()|2 + 2<{S1,t ()S2,t ()}
See Figs. 7.1 and 8.5 in Cohens book. Note that the scale a is related to frequency. The larger is a, the
How about a different window? smaller is frequency.
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 43 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 44 / 52
Continuous Wavelet Transform Continuous Wavelet Transform
Parsevals Theorem Parsevals Theorem:Proof
Note
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 47 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 48 / 52
Continuous Wavelet Transform Continuous Wavelet Transform
Examples Relation to STFT
a
a =h ej a (tb) .
a a
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 49 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 50 / 52
Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 51 / 52 Dr. X.-P. Zhang (CASPAL, Ryerson) DSP II September 14, 2017 52 / 52