Indian Institute of Science

E9−252: Mathematical Methods and Techniques in Signal Processing

Instructor: Shayan G. Srinivasa
Homework #7 Solutions, Fall 2017
Solutions prepared by Priya J Nadkarni
Late submission policy: Points scored = Correct points scored × e−d , d = # days late
Assigned date: Oct. 16th 2017 Due date: Oct. 23rd 2017 by end of the day

PROBLEM 1:
Derive wavelet decomposition of a signal using m-adic Haar wavelets. (10 points)
Solution:
First let us obtain the m-adic Haar wavelets. Let the basis of Vj be {φ(mj t − k)}∞
k=−∞ , where,

1 k ≤ t < k + 1
φ(mj t − k) = mj mj
0 else


m−1
X
Note that φ(mj t − k) = φ(mj+1 t − mk + a). As it needs m signals in Vj+1 to obtain Vj ,
a=0
thus the dimensionality of Vj+1 with respect to Vj is m. Thus, we need m − 1 wavelets to obtain
Vj+1 using Vj . Let us define the wavelets in W0 as follows:
i


 x 0≤t<
m



ψi (t) = y i i+1 (1)
i ≤t<
m m




0 else


As we need the wavelets to be orthogonal to each other, consider ψi (t) and ψj (t), where i < j,
then,
Z ∞
ψi (t)ψj (t)dt = 0
−∞
x2 i xyi
=⇒ + = 0
m m
xi + yi
=⇒ = 0
m
As we need the wavelets to be orthogonal to φ(t),
Z ∞
φ(t)ψi (t)dt = 0
−∞
xi yi
=⇒ + = 0
m m
xi + yi
=⇒ = 0
m
Thus, considering x = 1 =⇒ yi = −i. Thus,
i


1 0≤t<
m



ψi (t) = −i i ≤ t < i + 1
m m




0 else

 Z ∞
j(j + 1)
These form the basis of the wavelet space W0 . Note that |ψj (t)|2 = ψj2 (t)dt =
.
−∞ m
Let us denote this by αj2 . As φ(t) and {ψj (t)}m−1
j=1 are orthogonal and lie in V1 , they form a
basis for V1 as it has dimension of m. Thus, we can represent φ(mt) as follows:
m−1
X
φ(mt) = a0 φ(t) + ai ψi (t).
i=1

where the coefficients are obtained as follows:

a0 = hφ(mt), φ(t)i
1
ai = hφ(mt), ψi (t)i
αi
In general,
1


 j<i
m


hφ(mt − j), ψi (t)i = −i
j=i



 m
0 j>i


1
Thus, ai = As we have obtained V1 from V0 and W0 , we can similarly obtain the higher
mαi
resolution subspaces and obtain a wavelet decomposition.
Generalization: The wavelets in Wj are given by,
i


 1 0 ≤ t < j+1
m



ψi (mj t) = −i i
≤ t <
i+1

 m j+1 mj+1


0 else

Z ∞
j(j + 1)
2
Note that |ψj (t)| = ψj2 (t)dt = . Let us denote this by αj2 . As φ(mj t) and
−∞ mj
m−1
{ψk (mj t)}k=1 are orthogonal and lie in Vj+1 , they form a basis for it as it has dimension of m.
Thus, we can represent φ(mj+1 t) as follows:
m−1
X
j+1 j
φ(m t) = a0 φ(m t) + ai ψi (mj t).
i=1

where the coefficients are obtained as follows:

1
a0 =
mj
1
ai =
mj αi

PROBLEM 2:
Let Wj be the space of all functions with basis ψ(2j t − k) where k ∈ Z. Prove Vj+1 = Vj ⊕ Wj .
(5 points)
Solution:
Any element in Vj+1 is given by:
∞
X
f (t) = al φ(2j+1 t − l)
l=−∞
The basis function is given by:

k k+1
1 ≤t<
φ(2j+1 t − k) = 2j+1 2j+1
0 else


Let us consider the basis functions φ(2j t − k) and ψ(2j t − k) of Vj and Wj respectively, then,

1 k ≤ t < k + 1
j
φ(2 t − k) = 2j 2j
0 else

k k 1


 1 j
≤ t < j
+ j+1
2 2 2



j k 1 k+1
ψ(2 t − k) = −1 + j+1 ≤ t < j+1


 2 j 2 2

0 else


Thus, if l is odd,
l−1 l−1
φ(2j t − ) − ψ(2j t − )
φ(2j+1 − l) = 2 2
2
Thus, if l is even,
l l
φ(2j t − ) + ψ(2j t − )
φ(2j+1 − l) = 2 2
2
Thus,
∞
X ∞
X ∞
X
j+1 j j
f (t) = al φ(2 t − l) = a2l (φ(2 t − l) + (ψ(2 t − l)) + a2l+1 (φ(2j t − l) − (ψ(2j t − l))
l=−∞ l=−∞ l=−∞
X∞
= (a2l + a2l+1 )φ(2j t − l) + (a2l − a2l+1 ψ(2j t − l))
l=−∞

Thus, every signal can be expressed in basis of Vj and ψj which are orthogonal to each other.
Thus, the direct sum of the two spaces add upto Vj+1 .

PROBLEM 3:
Obtain the Haar wavelet decomposition for the signal f (t) using the Haar basis. Indicate the
signal dimension at each subspace. Sketch the waveforms explicitly at each subspace. Obtain
the reconstructed signal in functional form after nulling out any spike of (1/8)th unit of time.
Analyze using Fourier Transform. How much of energy is lost in the recovered signal?(10 points)
1

3
 0≤t<


 4
1 3


−1 ≤t<


f (t) = 4 8
3 5
2 ≤t<





 8 8

 5
0 ≤t<1

8

Solution:  th
1
The maximum resolution of the function is unit of time. The function can be written
8
 Level Signal Dimension
1 2
2 4
3 5
4 5

Table 1: Signal Dimension

as:
f (t) = 3φ(8t) + 3φ(8t − 1) − φ(8t − 2) + 2φ(8t − 3) + 2φ(8t − 4)
3 3 1
= (φ(4t) + ψ(4t)) + (φ(4t) − ψ(4t)) − (φ(4t − 1) + ψ(4t − 1)) + (φ(4t − 1) − ψ(4t − 1)) + (φ(4t −
2 2 2
1 3
= 3φ(4t) + φ(4t − 1) − ψ(4t − 1) + φ(4t − 2) + ψ(4t − 2)
2 2
3 1 1 3
= (φ(2t) + ψ(2t)) + (φ(2t) − ψ(2t)) + (φ(2t − 1) + ψ(2t − 1)) − ψ(4t − 1) + ψ(4t − 2)
2 4 2 2
7 5 1 3
= φ(2t) + ψ(2t) + (φ(2t − 1) + ψ(2t − 1)) − ψ(4t − 1) + ψ(4t − 2)
4 4 2 2
7 5 1 1 3
= (φ(t) + ψ(t)) + ψ(2t) + ( (φ(t) − ψ(t)) + ψ(2t − 1)) − ψ(4t − 1) + ψ(4t − 2)
8 4 2 2 2
9 5 5 3
= φ(t) + ψ(t) + ψ(2t) + ψ(2t − 1)) − ψ(4t − 1) + ψ(4t − 2)
8 8 4 2

th
The signal dimension at each level is: After nulling out the 18 spike but suppressing ψ(4t−1)
and ψ(4t − 2) , we obtain,
9 5 5 1
g(t) = φ(t) + ψ(t) + ψ(2t) + ψ(2t − 1))
8 8 4 2
9 9 9 9 5 5 5
= φ(4t) + φ(4t − 1) + φ(4t − 2) + φ(4t − 3) + φ(4t) + φ(4t − 1) − φ(4t − 2)
8 8 8 8 8 8 8
5 5 5 1 1
− ψ(4t − 3) + φ(4t) − φ(4t − 1) + φ(4t − 2) − φ(4t − 3)
8 4 4 2 2
1
= 3φ(4t) + φ(4t − 1) + φ(4t − 2)
2
1


 3 0≤t<


 4
1 1 1


≤t<



= 2 4 2
1 3
1 ≤t<





 2 4

 3
0 ≤t<1

4
Z ∞
Energy lost = f 2 (t) − g 2 (t)dt
−∞
 2
1 1 1 1 1 1 1
= ( × 32 + × (−1)2 + × 22 − ( × 32 + × + × 12 )
4 8 4 4 4 2 4
27 41 13
= − =
8 16 16
Fourier Analysis:
  −j2πf   −j6πf   −j10πf
3 f 1 f 1 f
G(f ) = sinc e 8 + sinc e 8 + sinc e 8
4 4 8 4 4 4