next up previous contents index
Next: Unit-Economy via the Two Up: Multiresolution Analysis as Hierarchical Previous: Resolution Spaces as Hierarchical   Contents   Index


Resolution Analysis as a Decomposition into Subspaces

Fourth, decompose each resolution subspace $ \mathbf V_{k}$ into its subsequent resolution subspace $ \mathbf V_{k+1}$ and its corresponding orthogonal complement, the subspace of details $ \mathbf O_{k+1}$ :

$\displaystyle \mathbf V_{k}=\mathbf V_{k+1}\oplus \mathbf O_{k+1}$ (284)

Given the fact that $ \mathbf V_{k}$ is spanned by the $ k$ th resolution basis, Eq.(2.83), the meaning of such a decomposition consists of exhibiting an alternative o.n. basis part of whose elements span $ \mathbf V_{k+1}$ , while the remainder spans its orthogonal complement. This decomposition is achieved as follows:

Recall that every square integrable function $ f(t)$ can be approximated as an optimal element in $ \mathbf V_{k}$ . This optimal approximation, which in Section 1.5.4 was identified as the least squares approximation of $ f(t)$ in the subspace $ \mathbf V_{k}$ , is uniquely expressed in terms of any orthonormal basis. Following Eq.(1.8), and using the o.n. basis, Eq.(2.82), one has has the projection of $ f(t)$ onto $ \mathbf V_{k}$ :

$\displaystyle P_{\mathbf V_{k}}f(t)=\sum_{\ell=-\infty}^\infty 2^{-k}\phi(2^{-k}t-\ell) \, \langle \phi(2^{-k}u-\ell),f(u)\rangle$ (285)

This is the least squares approximation of $ f(t)$ based on the subspace $ \mathbf V_{k}$ , or more briefly the $ \mathbf V_{k}$ -least squares approximation. The next (less refined) approximation is the projection of $ f(t)$ onto the subspace $ \mathbf
V_{k+1}\subset\mathbf V_{k}$ :

$\displaystyle P_{\mathbf V_{k+1}}f(t)=\sum_{\ell=-\infty}^\infty 2^{-(k+1)}\phi(2^{-(k+1)}t-\ell) \, \langle \phi(2^{-(k+1)}u-\ell),f(u)\rangle~.$ (286)

Here $ P_{\mathbf V_{k}}$ and $ P_{\mathbf V_{k+1}}$ are the projection operators onto $ \mathbf V_{k}$ and $ \mathbf V_{k+1}$ respectively.

Let us compare the two o.n. bases for the two resolution spaces $ \mathbf V_{k}$ and $ \mathbf V_{k+1}$ . We shall presently see that they are two families of o.n. wave packets identified already on page [*] by Eq.(2.64):

$\displaystyle Q_{0\ell}^{\varepsilon}(t)=
\left\{ \begin{array}{ccl}
Q_{0\ell}^...
...Q_{0\ell}^{\varepsilon'}(t)&~~~&\varepsilon=\varepsilon'
\end{array}\right.~~.
$

Here nad throughout the ensuing development we always let

$\displaystyle \varepsilon'=2^{-k},\quad k=0,\pm 1,\pm 2,\cdots~~.
$

Using Eqs.(2.82) and (2.80) one finds that the $ \mathbf V_{k}$ -family members are

$\displaystyle \mathbf V_{k}:\, \sqrt{2^{-k}}\,\phi(2^{-k}t-\ell)$ $\displaystyle =\sqrt{2^{-k}} \frac{1}{2\pi}\int_{-\pi}^\pi e^{i\omega(2^{-k}t-\ell)}d\omega$    
  $\displaystyle =\frac{1}{\sqrt{2\pi}}\frac{1}{\sqrt{2\varepsilon'}}\int^{\varepsilon'}_{-\varepsilon'} e^{-2\pi i\ell\omega/2\varepsilon'} e^{i\omega t} d\omega$    
  $\displaystyle =Q_{0\ell}^{2\varepsilon'}(t)~,$ (287)

and
$\displaystyle \Delta t$ $\displaystyle =$ $\displaystyle \frac{2\pi}{2\varepsilon'}=2^k$  
$\displaystyle \Delta\omega$ $\displaystyle =$ $\displaystyle 2\varepsilon'=\frac{2\pi}{2^k}~.$  

By contrast, the $ \mathbf V_{k+1}$ -family members, which are twice as wide in the temporal domain and twice as narrow in the frequency domain, are

$\displaystyle \mathbf V_{k+1}:\, \sqrt{2^{-(k+1)}}\,\phi(2^{-(k+1)}t-\ell)$ $\displaystyle =\sqrt{2^{-(k+1)}} \frac{1}{2\pi}\int_{-\pi}^\pi e^{i\omega(t2^{-(k+1)}-\ell)}d\omega$    
  $\displaystyle =\frac{1}{\sqrt{2\pi}}\frac{1}{\sqrt{\varepsilon'}} \int^{\vareps...
...'/2}_{-\varepsilon'/2} e^{-2\pi i\ell\omega/\varepsilon'} e^{i\omega t} d\omega$    
  $\displaystyle =Q_{0\ell}^{\varepsilon'}(t)$ (288)

and
$\displaystyle \Delta t$ $\displaystyle =$ $\displaystyle \frac{2\pi}{\varepsilon'}=2\times2^k$  
$\displaystyle \Delta\omega$ $\displaystyle =$ $\displaystyle \varepsilon'=\frac{1}{2}\times\frac{2\pi}{2^k}~.$  

Figure: Partitioning of phase space by a nested set of band limited orthonormal basis functions. The tall thin rectangles and the shaded rectangles are the phasespace cells of the basis functions which span $ \mathbf V_{k}$ and $ \mathbf V_{k+1}$ respectively. These two sets of phase space cells are reproduced respectively on the l.h.s. and r.h.s. of Figure 2.22.
\begin{figure}\centering
\epsfig{file=twoCellArrays,scale=.6}\end{figure}
These two bases are represented by two overlapping arrays of phase space cells, as in Figure 2.21. The phase space cells referring to the $ \mathbf V_{k}$ -basis are taller and skinnier than those referring to $ \mathbf V_{k+1}$ . Furthermore, the phase space domain of the $ \mathbf V_{k+1}$ -basis is a horizontal strip which is contained entirely within that of the $ \mathbf V_{k}$ -basis. Consequently, the phase space domain of the $ \mathbf V_{k}$ -basis gets partitioned into three mutually exclusive and jointly exhaustive domains:

The mutual exclusivity of these three strips, together with the fact that their union equals the strip generated by the $ \mathbf V_{k}$ -basis, implies that $ \mathbf V_{k}$ is spanned by two fundamental bases. Besides the one given by Eq.(2.87),

$\displaystyle \mathbf V_k=span\{Q_{0\ell}^{2\varepsilon'}(t):\, \ell=0,\pm1,\cdots\}~,
$

there also is
$\displaystyle \mathbf V_k=span\!\!\!\!\!\!\!\!$   $\displaystyle \{Q_{0\ell}^{\varepsilon'}(t):\, \ell=0,\pm1,\cdots\}\cup$  
       
    $\displaystyle \{P_{1\ell}^{\varepsilon'/2}(t):\, \ell=0,\pm1,\cdots\}\cup
\{P_{-2\,\ell}^{\varepsilon'/2}(t):\, \ell=0,\pm1,\cdots\}~.$  

Figure: Two alternative partitionings of the same phase space domain of $ \mathbf V_{k}$ . The three different horizontal strips in the right hand partitioning refer to the three orthogonal subspaces $ \mathbf
O^+_{k+1}$ , $ \mathbf V_{k+1}$ , and $ \mathbf O^-_{k+1}$ .
\begin{figure}\centering
\epsfig{file=threestrips,scale=.5}\end{figure}
As one can see from Figure 2.22, this corresponds to the union of the three strips mentioned above. Here the $ P$ 's are the familiar o.n. wave packets defined by Eq.(2.61):

$\displaystyle P_{1\ell}^{\varepsilon'/2}(t)=$ $\displaystyle \frac{1}{\sqrt{2\pi}}\frac{1}{\sqrt{\varepsilon'/2}} \int^{\varep...
...}_{\varepsilon'/2} e^{-2\pi i\ell\omega/(\varepsilon'/2)} e^{i\omega t} d\omega$ (289)
$\displaystyle P_{-2\,\ell}^{\varepsilon'/2}(t)=$ $\displaystyle \frac{1}{\sqrt{2\pi}}\frac{1}{\sqrt{\varepsilon'/2}} \int\limits^...
...{-\varepsilon'} e^{-2\pi i\ell\omega/(\varepsilon'/2)} e^{i\omega t} d\omega ~,$ (290)

and
$\displaystyle \Delta t$ $\displaystyle =$ $\displaystyle \frac{2\pi}{\varepsilon'/2}=2\times2^{(k+1)}$  
$\displaystyle \Delta\omega$ $\displaystyle =$ $\displaystyle \varepsilon'/2=\frac{1}{2}\times\frac{2\pi}{2^{(k+1)}}$  

for both the positive and negative frequency wave packets. Due to the mutual orthogonality of all the $ P$ 's and $ Q$ 's combined, every band limited function $ f\in\mathbf V_{k}$ is a unique linear combination of these elements. Thus we have identified two alternative bases of $ \mathbf V_{k}$ . The first one consists of the elements exhibited by Eq.(2.87). The second one consists of the elements exhibited by Eqs.(2.88)-(2.90).

This fact is reexpressed by the statement that $ \mathbf V_{k}$ is the direct sum of the subspaces

$\displaystyle \mathbf O_{k+1}^+$ $\displaystyle \equiv$ $\displaystyle span \{P_{1\ell}^{\varepsilon'/2}(t):\, \ell=0,\pm1,\cdots\}$ (291)
$\displaystyle \mathbf V_{k+1}$ $\displaystyle \equiv$ $\displaystyle span\{Q_{0\ell}^{\varepsilon'}(t):\, \ell=0,\pm1,\cdots\}$ (292)
$\displaystyle \mathbf O_{k+1}^-$ $\displaystyle \equiv$ $\displaystyle span\{P_{-2\,\ell}^{\varepsilon'/2}(t):\, \ell=0,\pm1,\cdots\}~,$ (293)

or, symbolically, that

$\displaystyle \mathbf V_{k}=\mathbf V_{k+1}\oplus\mathbf O_{k+1}^+\oplus\mathbf O_{k+1}^-~,
$

which is the same as Eq.(2.84), provided one sets

$\displaystyle \mathbf O_{k+1}=\mathbf O_{k+1}^+\oplus\mathbf O_{k+1}^- ~,
$

the direct sum of the positive and negative frequency subspaces othogonal to the $ (k+1)$ st resolution space $ \mathbf V_{k+1}$ .


next up previous contents index
Next: Unit-Economy via the Two Up: Multiresolution Analysis as Hierarchical Previous: Resolution Spaces as Hierarchical   Contents   Index
Ulrich Gerlach 2007-04-05