Progressive Data Access for Regular Grids 161
filters of even length, which introduce considerably more complexity, or han-
dling N =2
n
(see [1, 12]). For clarity of exposition, assume that the filter
is whole-sample symmetric about h[0]. That is, h[n]=h[−n], and h[0] is not
repeated. For example, if L = 3, the center point is h[0], and h[−1] = h[1].
Moreover, the input signal must be made a whole-sample symmetric about the
first (n =0)andlast(n = N −1) samples. The motivation for this symmetry
is straightforward: the output samples for n<0andn ≥ N are redundant
and need not be explicitly stored. Consider the calculation in a general lin-
ear transform of the expansion coefficient a[−1], by the convolution of x with
h[−n] for whole-sample symmetric x,andh,withL =3:
a[−1] = h[0]x[−2] + h[1]x[−1] + h[2]x[0].
Due to symmetry, h[n]=h[−n], and x[n]=x[−n]. Therefore:
a[1] = a[−1] = h[0]x[2] + h[1]x[1] + h[2]x[0],
and a[−1] is redundant and need not be explicitly stored!
Things become a little more complicated with the DWT, which operates as
a dual channel convolution filter, followed by downsampling. Here, centering
the filter must be done carefully for each channel, such that symmetry is
preserved after downsampling, and the total number of coefficients output by
the two channels, equals the number of input coefficients. For even N , each
channel outputs exactly N/2 samples.
Symmetric filters combined with a symmetric signal extension provide a
straightforward mechanism for dealing with finite length signals and the DWT.
Unfortunately, with the exception of the Haar wavelet, there are no orthogonal
wavelets with compact support possessing both the property of symmetry and
perfect reconstruction. The solution to this dilemma is the relaxation of the
orthogonality requirement and the introduction of biorthogonal wavelets. For
biorthogonal wavelets, the properties of Equation 8.11 no longer hold. Differ-
ent analysis scaling and wavelet basis functions,
˜
φ(t)and
˜
ψ(t), and synthesis
scaling and wavelet basis, φ(t)andψ(t), must be introduced. Similarly, new
analysis,
˜
h
0
and
˜
h
1
, and synthesis filters, g
0
and g
1
, will appear.
From a filter bank perspective, h
1
no longer relates to h
0
by a simple
expression. However, the following cross relationship between synthesis and
analysis filters hold:
˜
h
0
[n]=(−1)
n
g
1
(N − 1 − n),g
0
[n]=(−1)
n
˜
h
1
(N − 1 − n), (8.23)
where N is the support size of the filter.
With the analysis filters no longer related to each other by Equation 8.15,
the support of these respective filters need not be the same. The support sizes
of
˜
h
0
and
˜
h
1
are then denoted as L
0
and L
1
, respectively, leading to new
analysis equations: