8/12/2019 Crit2 Updated
1/39
Wavelets
Ingrid Daubechies
8/12/2019 Crit2 Updated
2/39
Construction of scaling and wavelet
functions
8/12/2019 Crit2 Updated
3/39
8/12/2019 Crit2 Updated
4/39
8/12/2019 Crit2 Updated
5/39
Time and frequency resolution
8/12/2019 Crit2 Updated
6/39
Time and frequency perspective for
scaling function
Frequencies of x(t) inside main lobe are
emphasized with respect to frequenciesoutside the main lobe
As we go from V1 to V+1, more and
more frequencies are emphasized as
peak always remains on zero
frequency
8/12/2019 Crit2 Updated
7/39
Time and frequency relation for the wavelet function
8/12/2019 Crit2 Updated
8/39
8/12/2019 Crit2 Updated
9/39
Time and frequency perspective for
wavelet function
Localization in time gets better and better as we go
fromW1
to W+1
Localization in frequency gets poorer and poorer as we
go from W1
to W+1
Along with the bandwidth, the centre frequency also
shifts
Different bands with increasing bandwidth are
emphasized as we go fromW1
to W+1
Haar scaling function aspires to become low pass
filter and wavelet function aspires to be band pass
filter
8/12/2019 Crit2 Updated
10/39
Therefore two important observations can be made as
The ratio of bandwidth to center frequency of remains
constant. As we go up the ladder of MRA, we deal with having higher
center frequency and larger bandwidth.
Similarly, as we go down the ladder, possesses lower center
frequency and smaller bandwidth.
(.)
(.)
(.)
8/12/2019 Crit2 Updated
11/39
() and () are bounded in both domains in aweaker sense as we focus on main lobe.
Main lobe has certain amount of energy. Then() and () are localized in time andfrequency both.
Variance is important statistical property that
is very useful in calculating spread of a givenfunction, which is indicative of concentration ofenergy of a function within certain band (intime as well as frequency domain).
Is it possible to have finite variances in bothfrequency as well as time domainsimultaneously?
8/12/2019 Crit2 Updated
12/39
Haar wavelet, it is somewhat
concentrated in frequency, but well
concentrated in time.
Daubechies function, as we go at higher
order, we get a somewhat better filtering
operation that is better frequency
localization.
8/12/2019 Crit2 Updated
13/39
Uncertainty Principle
8/12/2019 Crit2 Updated
14/39
Uncertainty Principle
8/12/2019 Crit2 Updated
15/39
Uncertainty Principle
8/12/2019 Crit2 Updated
16/39
Th d d it f ti S d it d
8/12/2019 Crit2 Updated
17/39
The second density function: Squared magnitude
response
8/12/2019 Crit2 Updated
18/39
Why go ahead/ away from Haar?
If we want to get some
meaningful
uncertainty, some
meaningful bound, we
must at least considercontinuous functions.
8/12/2019 Crit2 Updated
19/39
Time Bandwidth Product should be minimum for the function to be
compact in both the domains. But for MRA we translate as well as
scale the basis function. Therefore it is necessary to look at the
properties of this product.
8/12/2019 Crit2 Updated
20/39
The time bandwidth product is thus a
robust measure of combined time and
frequency spread of a signal. It isessentially a property of the shape of the
waveform. time-bandwidth product of the
Haar scaling function was
What is the minimumvalue of this product?
8/12/2019 Crit2 Updated
21/39
8/12/2019 Crit2 Updated
22/39
Cauchy Schwarz inequality theorem states that
8/12/2019 Crit2 Updated
23/39
For any complex number :
8/12/2019 Crit2 Updated
24/39
8/12/2019 Crit2 Updated
25/39
8/12/2019 Crit2 Updated
26/39
Ideal function to achieve the TBP
8/12/2019 Crit2 Updated
27/39
Can the TBP of the Haar function
be improved ?
( )* ( )t t
2
20.13
0.3
t
TBP
Just by cascading the scaling
function once with itself the TBP
has reduced from infinity to 0.3 .
TBP can be further reduced by
cascading multiple times.
8/12/2019 Crit2 Updated
28/39
Time frequency tilingOccupancy of x(t) in time-frequency plane can be thought as being
around t0, the center in time, from t0 +tto t0 - ton the
horizontal axis. On the vertical axis we would like to center it at 0,the
frequency center, and we would spread it between0
Thus min area of the signal
spread in both the domains
should be 2 units
Area of the rectangle =
8/12/2019 Crit2 Updated
29/39
Time frequency tiling for wavelet transform
8/12/2019 Crit2 Updated
30/39
Resolution of Time & Frequency
Time
Frequenc
y
Better time
resolution;
Poor
frequency
resolution
Betterfrequency
resolution;
Poor time
resolutionEach box represents a equal portion
Resolution in STFT is selected once for entireanalysis
8/12/2019 Crit2 Updated
31/39
Continuous Wavelet Transform (CWT)
Wavelet Transform :
Find projection on the basis function and its translates
Sum up all these projections to give the piecewise
constant representation of the continuous function x(t).
In the Frequency Domain:
8/12/2019 Crit2 Updated
32/39
Continuous Wavelet Transform (CWT)
Continuous Wavelet Transform (CWT):
8/12/2019 Crit2 Updated
33/39
Continuous Wavelet Transform (CWT):Reconstruction
We will try to reconstruct the original signal x(t) bysumming the components of CWT along the unit
vector in its direction i.e. its basis function.
According to Parsevals Theorem
Continuous Wavelet Transform (CWT):
8/12/2019 Crit2 Updated
34/39
Continuous Wavelet Transform (CWT):
Reconstruction
Substituting back in the earlier integral
Continuous Wavelet Transform (CWT):
8/12/2019 Crit2 Updated
35/39
Continuous Wavelet Transform (CWT):
Reconstruction
Let
Let
8/12/2019 Crit2 Updated
36/39
8/12/2019 Crit2 Updated
37/39
8/12/2019 Crit2 Updated
38/39
8/12/2019 Crit2 Updated
39/39