152 High Performance Visualization
2
4
1
−80
−60
−40
−20
0
20
40
60
(a)
2
4
1
12
−80
−60
−40
−20
0
20
40
60
(b)
2
4
1
12
−80
−60
−40
−20
0
20
40
60
(c)
FIGURE 8.2: PDA based on multiresolution. The original signal containing
1024 samples (a), reduced to 1/8
th
resolution using subsampling, analogous
to the Z-order curve (b); and the Haar wavelet approximation coefficients (c).
a reconstruction of an approximation with 1, 2, 4, or 8 samples, but with-
out any storage overhead. In the context of data read from disk, the coarsest
approximation, c
0
, can be retrieved by reading a single sample, the next ap-
proximation level, c
1
, by reading one additional coefficient, d
0
, and applying
Equation 8.6, and the next, by reading two more coefficients, and so on. Multi-
ple dimensions can be easily generalized by averaging neighboring points along
each dimension to generate approximation coefficients, and generate detail co-
efficients by taking the difference between the approximation coefficients and
each of the points used to find its average.
Mathematically, this is the 1D, unnormalized, Haar wavelet transform.
The Haar wavelet transform, and other wavelet transforms that are discussed
later, constructs a multiresolution representation of a signal. In one dimension,
each multiresolution level j contains, on average, half the samples of the j +1
level approximation, with coarser level samples created by averaging finer
level samples. Note, while the Haar transform exhibits what is known in the
digital signal processing world as the property of perfect reconstruction, due
to limited floating point precision the signal reconstructed from the c
j
and
d
j
coefficients will not, as was the case with the Z-order curve, be an exact
copy of the original signal. Moreover, even if the input signal consists entirely
of the integer data the division by two in Equation 8.4 will result in floating
point coefficients. However, integer-to-integer wavelet transforms do exist [3],
but they are not discussed here.
Figure 8.2 shows a qualitative comparison of a 1D sampling of the x-
component of vorticity obtained from a 1024
3
simulation by Taylor-Green
turbulence [7]. The original signal is seen in Figure 8.2a. The signal generated
by nearest neighbor sampling (analogous to Z-order curve approximation) at
1/8
th
resolution is shown in Figure 8.2b, and at the same coarsened resolution,
approximated with the Haar wavelet, as in Figure 8.2c.
Equations 8.4–8.6 provide everything that is needed to support a PDA
scheme based purely on a multiresolution hierarchy, not suffering from the