MATH 414 Lecture 35

From Notes
Jump to navigation Jump to search

« previous | Wednesday, April 16, 2014 | next »


Wavelet Review

{2j2ϕ(2jx−k)}k∈ℤ is an orthonormal basis for Vj {2j2ψ(2jx−k)}k∈ℤ is an orthonormal basis for Wj

Vj=Vj−1⊕Wj−1


Decomposition and Reconstruction

For fj∈Vj, fj={∑k=−∞∞akjϕ(2jx−k)∑k=−∞∞akj−1ϕ(2j−1x−k)⏟fj−1+∑k=−∞∞bkj−1ψ(2j−1x−k)⏟wj−1


Projecting an arbitrary function f into function fj∈Vj involves sampling.

fj can then be decomposed into fj−1 and wj−1, and so on down to f0, w0, w1, …


Inversely, f0 and w0 can be used to reconstruct f1, f1 and w1 reconstruct f2, and so on up to fj


Questions:

  1. How do we get from akj's to akj−1's and bkj−1's?
  2. How do we get from akj−1's and bkj−1's back to akj's


Decomposition

pk=2∫−∞∞ϕ(2x−k)ϕ(x)dxqk=2∫−∞∞ϕ(2x−k)ψ(x)dx


wj−1=projWj−1(fj)=∑k∈ℤbkj−1ψ(2j−1x−k)=∑k∈ℤbkj−12j−122j−12ψ(2j−1x−k)⏟orthonormalbkj−12j−12=(∫−∞∞fj(x)ψ(2j−1x−k)dx)2j−12=⟨fj,ψj−1,k⟩bkj−1=2j−1∫−∞∞fj(x)ψ(2j−1x−k)dx

We already have fj=∑ℓ∈ℤaℓjϕ(2jx−ℓ), so

bkj−1=2j−1∫−∞∞ψ(2j−1x−k)fj(x)dx=2j−1∫−∞∞ψ(2j−1x−k)(∑ℓ∈ℤaℓjϕ(2jx−ℓ))=2j−1∑ℓ∈ℤaℓj∫−∞∞ψ(2j−1x−k)ϕ(2jx−ℓ)dx

Let u=2j−1x−k, then 2jx=2u+2k, and du=2j−1dx

Also recall qm=2∫−∞∞ψ(x)ϕ(2x−m)dx.

bkj−1=2j−1∑ℓ∈ℤaℓj∫−∞∞ψ(u)ϕ(2u+2k−ℓ)dx⏟2=12∑ℓ∈ℤaℓjqℓ−2k


Therefore, we are left with

akj−1=12∑ℓ∈ℤaℓjp2k−ℓbkj−1=12∑ℓ∈ℤaℓjq2k−ℓ

If we let λ be a filter sequence, we have

akj−1=(λ*aj)2k

Therefore, we can downsample:

akj−1=DL(aj)

Note that this does not depend on j!!

Likewise, bkj−1=DH(aj)

Projection

We are interested in coefficients akj that are projections of f(x) onto Vj. Therefore (let u=2jx−k)

akj=2j∫−∞∞f(x)ϕ(2jx−k)=∫−∞∞f(2−ju+2−jk)ϕ(u)du


"All" interesting wavelets have support contained in some finite interval [a,b]. Hence

akj=∫abf(2−ju+2−ju)ϕ(u)du

If j is large, then 2−ja≤2−ju≤2−jb, where 2−ja and 2−jb are small.

Since f∈L2 is continuous, f(2−ju+2−jk)≈f(2−jk) for small 2−ju. Now what happens next is crucial:

akj≈f(2−jk)∫abϕ(u)du⏟1

(the integral equality to 1 is the case in most MRAs)

This means that akj≈f(2−jk) is just a sampling of f at 2−jk.

The Wavelet Crime

Given samples at any level j, we can decompose aj into component parts aj−1, bj−1 by using the same downsampling and convolution filters