Biorthogonal Wavelets
Ref: Rao & Bopardikar, Ch.4
Jyun-Ming Chen
Spring 2001
Why is orthogonality useful
x 1a1 2a 2
x 1 1 a 1 2 1
T T
a 2 1 2
T
1 x, a1 / a1 , a1 3 / 5
2 x, a 2 / a 2 , a 2 1/ 5
• Orthonormal bases further simplify the
computation
Ortho v. Non-Ortho Basis
Sum of projection vectors !?
Dual Basis
a 1 2 1 x 1 1 x 1a1 2a 2
T T
a1-a2 and b1-b2
a 2 1 2
T
1 x, b1 / a1 , b1 1 / 3 are biorthogonal
a1 , b1 1 2 x, b2 / a 2 , b2 1 / 3
a 2 , b2 1
a1 , b 2 0
a 2 , b1 0
Dual Bases
b1 2 / 3 1 / 3
T
b 2 1 / 3 2 / 3
T
Dual Basis (cont) Verify
duality !
a1 1 1 2 b1 1 0 0
T T
a 2 0 1 0 b 2 1 1 0
T T
• Dual basis may generate different spaces
– Here: a1-a2 and b1-b2 generate two different 2D
subspaces in Euclidean 3space.
• Semiorthogonal:
– For dual basis that generates the same subspace
• Orthogonal:
– Primal and dual are the same bases
Extend to Function Space
• MRA types:
– orthogonal, semiorthogonal, biorthognal
• Extend the concept to using biorthogonal
MRA
– More flexible design
– Lifting scheme: a general design method for
biorthogonal wavelets
Alternative Wavelets:
Biorthogonal Wavelets
Proposed by Cohen (1992)
Characteristics of Orthogonal Basis
• Decomposition and • Higher-order filters (with
reconstruction filters are more coefficients) have
FIR and have the same poor time-frequency
length localization
• Generally do not have
closed-form expressions • Desired property: perfect
• Usually not symmetric reconstruction FIR
(linear phase) symmetric (linear-phase)
• Haar wavelet is the only filters
real-valued wavelet that is – Not available in orthogonal
compactly supported, bases
symmetric and orthogonal
The Need for Biorthogonal Basis
• delegate the • compactly supported
responsibilities of symmetric analyzing
analysis and synthesis and synthesis wavelets
to two different and scaling functions
functions (in the
biorthogonal case) as
opposed to a single
function in the
orthonormal case
– more design freedom
Biorthogonal Scaling Functions
~ ~ ~
(t ) 2 h(n) (2t n) dual (t ) 2 h (n) (2t n)
n n
~ ~
(t ), (t k ) (k ) (2 t ), (2 k t n) 2k (n)
k
~
• Two sequences h(n) and h (n)serve as impulse
response of FIR filters
~
• Two sets of scaling functions (t ) and (t ) generate
subspaces Vk and V~k respectively
• The basis are orthogonal; the two MRAs are said
to be biorthogonal to each other
Dual MRA (cont)
~
• Basis of Vk and Vk
– Translated copy of appropriate dilation of
~
(t ) and (t )
V0 V1 V2 V3
~ ~ ~ ~
V0 V1 V2 V3
Function approximation in
subspaces
Coarser approx f 0 (t ) a(0, n) (t n)
n
~ coefficien ts obtained by
a(0, n) f (t ), (t n) ~
projecting f (t ) on V0
Finer approx f1 (t ) a(1, n) (2t n)
n
~
a(1, n) 2 f (t ), (2t n)
~ ~
f (t ), (2t n) a(1, n) (2t n), (2t n)
1
a(1, n)
2
Relation between Finer and
Coarser Coefficients
~ ~ ~ ~
a(0, n) f (t ), (t n) (t n) 2 h (m) (2t 2n m)
m
~ ~
a(0, n) 2 h (m) f (t ), (2t 2n m)
m
~
a(1, n) 2 f (t ), (2t n)
~
a(1,2n m) 2 f (t ), (2t 2n m)
n( 2 n m )
~
a(0, n) h (m)a(1,2n m)
m
~
a(0, n) a(1, m)h (m 2n)
m
Biorthogonal Wavelets
(t )dt 0
~(t )dt 0
~ (t )
(t k ) : k Z spans W0 • Two sets of wavelets (t ) and
~
generate subspaces kW and W
(t k ) : k Z spans W0 respectively
~ k
~
• The basis are orthogonal; the two
MRAs are said to be
Require: biorthogonal to each other
(t ),~(t k ) (k ) Dual
~
(t ), (t n) 0 wavelet dual scaling fns
~(t ), (t n) 0 dual wavelet scaling fns
Two-scale relations of wavelet:
primal and dual
(t ) 2 g (n) (2t n)
n
~
~(t ) 2 g (n) (2t n)
~
n
Function Projection
detail fn : g 0 (t ) f1 (t ) f 0 (t ) b(0, n) (t n)
n
b(0, n) g 0 (t ),~ (t n) f1 (t ) f 0 (t ),~(t n)
f1 (t ),~ (t n) f 0 (t ),~ (t n)
f (t ),~(t n)
1
f1 (t ) a(1, m) (2t m) f 0 (t ) a(0, n) (t n)
m n
~
(t n) 2 g (l ) (2t 2n l )
~ ~
and ~(t ), (t n) 0
l
m=2n+l b(0, n) a(1, m) g~(m 2n)
m
Function Reconstruction
f1 (t ) f 0 (t ) g 0 (t )
a (0, l ) (t l ) b(0, l ) (t l )
l l
2 a(0, l )h(m) (2t 2l m) 2 b(0, l ) g (m) (2t 2l m)
m l m l
Substituti ng n 2l m
f1 (t ) 2 a(0, l )h(n 2l ) (2t n) 2 b(0, l ) g (n 2l ) (2t n)
n l n l
f1 (t ) a (1, n) (2t n)
n
Hence
a (1, n) 2 a (0, l )h(n 2l ) 2 b(0, l ) g (n 2l )
l l
Filter Bank
Primal and Dual MRA
(biorthogonal)
VN ~
VN
VN-1 WN-1
~ ~
VN-1 WN-1
VN-2 WN-2 ~ ~
VN-2 WN-2
VN-3 WN-3 ~ ~
VN-3 WN-3
~ ~
Vk Wk Vk Wk
Filter Relations
(between primal and dual)
~
(t ), (t n) (n)
~ ~
4 h( p)h (q) (2t p), (2t 2n q)
p q
~
2 h( p)h (q) ( p 2n q)
p q
~
2 h ( m 2 n ) h ( m) only term left :
m
p 2n q 0
~ ( n)
m
h ( m 2 n ) h ( m)
2
p 2n q
Similarly,
( n)
(t ),~(t n) 0
m
~
g (m 2n) g (m)
2
Filter Relations (cont)
(t ),~(t n) 0
~
4 h( p) g~(q) (2t p), (2t 2n q)
p q
2 h( p) g~(q) ( p 2n q) 2 h(m 2n) g~(m)
p q m
m
h ( m 2 n ) ~(m) 0
g
Similarly,
~
~
(t ), (t n) 0 h ( m 2 n ) g ( m) 0
m
Design of Biorthogonal Wavelets
• because there is quite a bit of freedom in
designing the biorthogonal wavelets, there
are no set steps in the design procedure. …
• Lifting (Sweldens 94): a scheme for custom-
design biorthogonal wavelets
Special Cases:
orthogonal and semiorthogonal
• Common property: VN
Vk Wk Vk 1 Vk Wk VN-1 WN-1
• Differences:
VN-2 WN-2
– if orthogonal: scaling
functions (and wavelets) of
the same level are VN-3 WN-3
orthogonal to each other
– If semiorthogonal, wavelets
~ ~
of different levels are Vk Wk and Wk Wk
orthogonal (from nested
space) Dual and
primal are the
same
