The filterbanks developed in the module on the filterbanks interpretation of
the DWT start with the signal representation
c
0
n
n∈Z
c
0
n
n
and break the representation down into wavelet
coefficients and scaling coefficients at lower resolutions
(i.e., higher levels
kk). The question remains: how do
we get the initial coefficients
c
0
n
c
0
n
?
From their definition, we see that the scaling coefficients
can be written using a convolution:
c
0
n=〈φt−n,xt〉=∫−∞∞φt−nxtdt=φ−t*xt|
t=
n
′
c
0
n
φ
t
n
x
t
t
φ
t
n
x
t
t
n
′
φ
t
x
t
(1)
which suggests that the proper initialization of wavelet
transform is accomplished by passing the continuous-time input
xt
x
t
through an analog filter with impulse response
φ−t
φ
t
and sampling its output at integer times (
Figure 1).
Practically speaking, however, it is very difficult to build
an analog filter with impulse response
φ−t
φ
t
for typical choices of scaling function.
The most often-used approximation is to set
c
0
n=xn
c
0
n
x
n
. The sampling theorem implies that this would be
exact if
φt=sinπtπt
φ
t
t
t
, though clearly this is not
correct for general
φt
φ
t
. Still, this technique is somewhat justified if we
adopt the view that the principle advantage of the wavelet
transform comes from the multi-resolution capabilities implied
by an iterated perfect-reconstruction filterbank (with
good filters).