Haar wavelets and multiresolution analysis

Jordan Bell
April 3, 2014

1 Introduction

Let

ψ(x)={0x<0,10x<12,-112x<1,0x1.

For n,k, we define

ψn,k(x)=2n/2ψ(2nx-k),x.

L2() is a complex Hilbert space with the inner product

f,g=f(x)g(x)¯𝑑x.

We will prove that ψ satisfies the following definition of an orthonormal wavelet.11 1 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 303, Definition 6.4.1.

Definition 1 (Orthonormal wavelet).

If ΨL2(), Ψn,k(x)=2n/2Ψ(2nx-k), and the set {Ψn,k:n,k} is an orthonormal basis for L2(), then Ψ is called an orthonormal wavelet.

Lemma 2.

{ψn,k:n,k} is an orthonormal set in L2().

Proof.

If n,n,k,k, then

ψn,k(x)ψn,k(x)¯𝑑x = 2n/2ψ(2nx-k)2n/2ψ(2nx-k)𝑑x
= 2(n-n)/2ψ(x-k)ψ(2n-nx-k)𝑑x
= 2(n-n)/2δk,k01ψ(x)ψ(2(n-n)/2x)𝑑x
= δk,kδn,n,

hence {ψn,k:n,k} is an orthonormal set. ∎

Bessel’s inequality states that if is an orthonormal set in a Hilbert space H, then for any fH we have e|f,e|2f22, from which it follows that ef,eeH. To say that a subset of a Hilbert space H is an orthonormal basis is equivalent to saying that is an orthonormal set and that

idH=eee

in the strong operator topology. In other words, for to be an orthonormal basis of H means that is an orthonormal set and that for every fH we have

f=ef,ee.

From Lemma 2 and Bessel’s inequality, we know that for each fL2(),

n,k|f,ψn,k|2f22,n,kf,ψn,kψn,kL2().

We have not yet proved that f is equal to the series n,kf,ψn,kψn,k, and this will not be accomplished until later in this note.

2 Coarser sigma-algebras

For n,k, let

In,k=[k2n,k+12n),

and let n be the σ-algebra generated by {Ik,n:k}. =kIn,k, and if kk then In,kIn,k=. If n<n then

nn,

where is the σ-algebra of Lebesgue measurable subsets of . An element of L2(,n) is an element of L2(,) that is constant on each set In,k, k. In other words, an element of L2(,n) is a function f: such that if k then the image f(In,k) is a single element of and such that

f22=|f(x)|2𝑑x=kIn,k|f(x)|2𝑑x=k12n|f(In,k)|2<.

If n<n, then

L2(,n)L2(,n)L2(,).

3 Integral kernels

We define

ϕ(x)={0x<0,10x<1,0x1.

For n we define

Kn(x,y)=2nkϕ(2nx-k)ϕ(2ny-k),x,y.

We have

Kn(x,y){0,2n}.

Kn(x,y)=2n if and only if there is some k such that 2nx-k,2ny-k[0,1), equivalently there is some k with 2nx,2ny[k,k+1), which is equivalent to there being some k such that

x,y[k2n,k+12n)=In,k.

We define

Pnf(x)=Kn(x,y)f(y)𝑑y.

If x then there is a unique kx with xIn,kx, and

Pnf(x)=2nIn,kxf(y)𝑑y. (1)

It is straightforward to check that L2(,n) is a closed subspace of L2(,), and in the following theorem we prove that Pn is the orthogonal projection onto L2(,n).

Lemma 3.

If n, then Pn is the orthogonal projection of L2(,) onto L2(,n).

Proof.

For each k, the function Pnf is constant on the interval In,k, and using (1) and the Cauchy-Schwarz inequality,

Pnf22 =kIn,k|Pnf(x)|2𝑑x
=kIn,k|2nIn,kf(y)𝑑y|2𝑑x
=2nk|In,kf(y)𝑑y|2
2nk(In,k|f(y)|2𝑑y)(In,k𝑑y)
=kIn,k|f(y)|2𝑑y
=|f(y)|2𝑑y.

Therefore, Pn:L2(,)L2(,n). Moreover, the left-hand side of the above inequality is equal to Pnf22 and the right-hand side is equal to f22, hence we have Pnf2f2, giving Pn1.

If fL2(,n), then

Pnf(x) =Kn(x,y)f(y)𝑑y
=2nIn,kxf(y)𝑑y
=f(In,kx)
=f(x),

hence if fL2(,n) then Pnf=f. ∎

For n, we define

Ln=Kn+1-Kn,

and the following lemma gives a different expression for Ln.22 2 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 293, §6.3.2.

Lemma 4.

If n, then

Ln(x,y)=kψn,k(x)ψn,k(y),x,y.
Proof.

ψ(2nx-k)=1 means that 02nx-k<12, which is equivalent to k2nx<k+122n, which is equivalent to 2k2n+1x<2k+12n+1, which is equivalent to xIn+1,2k. ψ(2nx-k)=-1 means that 122nx-k<1, which is equivalent to k+122nx<k+12n, and this is equivalent to xIn+1,2k+1. ψ(2nx-k)=0 if and only if xIn+1,2kIn+1,2k+1. Therefore,

ψn,k(x)ψn,k(y)={2n(x,y)In+1,2k×In+1,2kIn+1,2k+1×In+1,2k+1,-2n(x,y)In+1,2k×In+1,2k+1In+1,2k+1×In+1,2k,0otherwise.

If there is no k such that (x,y)In,k×In,k, then Ln(x,y)=0. Otherwise, suppose that k and that (x,y)In,k×In,k. We have

In,k=In+1,2kIn+1,2k+1.

If (x,y)In+1,2k×In+1,2k, then

Ln(x,y)=Kn+1(x,y)-Kn(x,y)=2n+1-2n=2n;

if (x,y)In+1,2k+1×In+1,2k+1, then

Ln(x,y)=Kn+1(x,y)-Kn(x,y)=2n+1-2n=2n;

if (x,y)In+1,2k×In+1,2k+1, then

Ln(x,y)=Kn+1(x,y)-Kn(x,y)=0-2n=-2n;

and if (x,y)In+1,2k+1×In+1,2k, then

Ln(x,y)=Kn+1(x,y)-Kn(x,y)=0-2n=-2n.

It follows that

Ln(x,y)=kψn,k(x)ψn,k(y).

4 Continuous functions

Let C0() denote those continuous functions f: such that if ϵ>0 then there is some compact subset K of such that xK implies that |f(x)|<ϵ. We say that an element of C0() is a continuous function that vanishes at infinity. Let Cc() denote the set of continuous functions f: such that

supp(f)={x:f(x)0}¯

is a compact set.

In the following lemma, we prove that the larger the intervals over which we average a continuous function vanishing at infinity, the smaller the supremum of the averaged function.33 3 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 295, Lemma 6.3.2.

Lemma 5.

If fC0(), then Pnf0 as n-.

Proof.

If gCc() and x, then

|Png(x)| =|Kn(x,y)g(y)𝑑y|
=|supp(g)Kn(x,y)g(y)𝑑y|
supp(g)Kn(x,y)|g(y)|𝑑y
supp(g)2n|g(y)|𝑑y
2nμ(supp(g))g,

hence

Png2nμ(supp(g))g. (2)

If fC0() and ϵ>0 then there is some gCc() with f-g<ϵ. Hence,

PnfPn(f-g)+Png.

If x, then

|Pn(f-g)(x)|=2n|In,kx(f-g)(y)𝑑y|2nIn,kx|(f-g)(y)|𝑑yf-g,

hence Pn(f-g)f-g. Using this and (2) we obtain

Pnff-g+2nμ(supp(g))g<ϵ+2nμ(supp(g))g.

Hence,

lim supn-Pnflim supn-(ϵ+2nμ(supp(g))g)=ϵ.

This is true for every ϵ>0, so

limn-Pnf=0.

Lemma 6.

If fL2(), then Pnf20 as n-.

Proof.

If ϵ>0 then there is some gCc() such that f-g2<ϵ. Say supp(g)[-K,K]. If 2m>K, then we have by (1) and because supp(g)I-m,-1I-m,0,

P-mg22 =|2-mI-m,kxg(y)𝑑y|2𝑑x
=2m|2-mI-m,-1g(y)𝑑y|2+2m|2-mI-m,0g(y)𝑑y|2
=2-m|-K0g(y)𝑑y|2+2-m|0Kg(y)𝑑y|2
2-mμ([-K,0])g22+2-mμ([0,K])g22
=2K2-mg22.

Therefore, when 2m>K we have P-mg22-m22Kg2, and so, as the operator norm of P-m on L2() is 1,

P-mf2 P-m(f-g)2+P-mg2
f-g2+P-mg2
<ϵ+2-m22Kg2.

Thus, if ϵ>0 then

lim supmP-mf2ϵ.

This is true for all ϵ>0, so we obtain

limmP-mf2=0.

The following lemma shows that if fCc(), then Pnf converges to f in the L2 norm and in the L norm as n.44 4 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 296, Lemma 6.3.3.

Lemma 7.

If fCc(), then Pnff in the L2 norm and in the L norm as n.

Proof.

Suppose that supp(f)[-2M,2M] for M0. f is uniformly continuous on the compact set [-2M,2M], thus, if ϵ>0 then there is some δ>0 such that x,y[-2M,2M] and |x-y|<δ imply that |f(x)-f(y)|<ϵ2M. Let 2-nδ. For each x, there is some kx such that xIn,kx and we have

|Pnf(x)-f(x)| =|2nIn,kxf(y)𝑑y-f(x)|
=2n|In,kxf(y)-f(x)dy|
2nIn,kx|f(y)-f(x)|𝑑y
<2nIn,kxϵ2M𝑑y
=ϵ2M.

This tells us that if 2-nδ then Pnf-fϵ2M. Therefore, if ϵ>0 then for sufficiently large n we have Pnf-fϵ2M, showing that

limnPnf-f=0.

Furthermore, if n0 then

Pnf-f22=|Pnf(x)-f(x)|2𝑑x=-2M2M|Pnf(x)-f(x)|2𝑑x22MPnf-f2,

and because Pnf-f0 as n we get Pnf-f20 as n. ∎

From Lemma 4, we get

(Pn+1-Pn)f(x) =Kn+1(x,y)f(y)𝑑y-Kn(x,y)f(y)𝑑y
=Ln(x,y)f(y)𝑑y
=kψn,k(x)ψn,k(y)f(y)dy
=kf,ψn,kψn,k(x),

thus

Pn+1-Pn=kψn,kψn,k (3)

in the strong operator topology. Using (3), we obtain for n0 that

Pn+1 =P0+j=0nPj+1-Pj
=P0+j=0nkψj,kψj,k

in the strong operator topology. For n<0,

Pn =P0-j=-n-1Pj+1-Pj
=P0-j=-n-1kψj,kψj,k

in the strong operator topology.

We have already shown in Lemma 2 that {ψn,k:n,k} is an orthonormal set in L2(), and we now prove that it is an orthonormal basis for L2().

Theorem 8.

In the strong operator topology,

idL2()=n,kψn,kψn,k.
Proof.

Let fL2() and suppose ϵ>0. By Lemma 6, there is some M such that mM implies that P-mf2<ϵ2. There is some gCc() satisfying f-g2<ϵ6, and by Lemma 7 there is some N such that nN implies that Png-g2<ϵ6. Hence, if nN then

Pnf-f2 Pnf-Png2+Png-g2+g-f2
2f-g2+Png-g2
<2ϵ6+ϵ6
=ϵ2.

Therefore, if mM and nN, then

(Pn-P-m-idL2()f2Pnf-f2+P-mf2<ϵ2+ϵ2=ϵ.

For m,n>0, we have

Pn+1-P-m =j=0nkψj,kψj,k+j=-m-1kψj,kψj,k
=j=-mnkψj,kψj,k

in the strong operator topology. ∎

5 Other function spaces

Let Cb() denote those continuous functions that are bounded. We have

Cc()C0()Cb()C().
Lemma 9.

If n and fCb(), then Pnff.

Proof.

If x, then there is a unique kx with xIn,kx, and

|Pnf(x)|=|2nIn,kxf(y)𝑑y|2nIn,kx|f(y)|𝑑yf.

Theorem 10.

If fC0(), then the series n,kf,ψn,kψn,k converges to f uniformly on .

Proof.

If ϵ>0 then there is some gCc() with f-g<ϵ6. By Lemma 5, there is some M such that mM implies that P-mg<ϵ3, hence

P-mf P-mf-P-mg+P-mg
f-g+P-mg
<ϵ6+ϵ3
=ϵ2.

By Lemma 7, there is some N such that nN implies that Png-g<ϵ6, hence

Pnf-f Pnf-Png+Png-g+g-f
2f-g+Png-g
<ϵ2.

Therefore, if nN and mM, then

Pnf-P-mf-fPnf-f+P-mf<ϵ2+ϵ2=ϵ.

The following theorem states that Pn is an operator on Lp() with operator norm 1.55 5 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 297, Lemma 6.3.9. In particular, it asserts that if fLp() then the averaged function Pnf is also an element of Lp().

Theorem 11.

If 1p<, n, and fLp(), then Pnfpfp.

Proof.

Let 1p+1q=1, so q=pp-1. (If p=1 then q=.) If x, then there is a unique kx with xIn,kx, and using Hölder’s inequality we get

|Pnf(x)| =|2nIn,kxf(y)𝑑y|
2n(In,kx|f(y)|p𝑑y)1/p(μ(In,kx))1/q
=2n(In,kx|f(y)|p𝑑y)1/p2-n/q.

Therefore, if k then

In,k|Pnf(x)|p𝑑x In,k2np2-np/qIn,kx|f(y)|p𝑑y𝑑x
=In,k2np2-np/qIn,k|f(y)|p𝑑y𝑑x
=2-n2np2-np/qIn,k|f(y)|p𝑑y
=In,k|f(y)|p𝑑y.

We obtain

Pnfpp =kIn,k|Pnf(x)|p𝑑x
kIn,k|f(y)|p𝑑y
=|f(y)|p𝑑y
=fpp,

giving Pnfpfp. ∎

6 Multiresolution analysis

For a, we define ma: by ma(x)=ax, and we define τa: by τa(x)=x-a.

Definition 12 (Multiresolution analysis).

A multiresolution analysis of L2(R) is a set {Vn:n} of closed subspaces of the Hilbert space L2() and a function ΦL2() satisfying

  1. 1.

    If n, then fVn if and only if fm2Vn+1.

  2. 2.

    VnVn+1.

  3. 3.

    nVn¯=L2().

  4. 4.

    nVn={0}.

  5. 5.

    {Φτk:k} is an orthonormal basis for V0.

It is straightforward to prove the following theorem using what we have established so far.

Theorem 13.

The closed subspaces {L2(,n):n} of L2() and the function ϕ=χ[0,1) is a multiresolution analysis of L2().

The following lemma shows that if Pn is the projection onto Vn, where Vn is a closed subspace of a multiresolution analysis of L2(), then Pn0 in the strong operator topology as n-.66 6 Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets, p. 313, Lemma 6.4.28.

Lemma 14.

If {Vn:n} and ΦL2() is a multiresolution analysis of L2(), Pn:L2()Vn is the orthogonal projection onto Vn, and fL2(), then

limn-Pnf=0.
Proof.

Define Φn,k(x)=2n/2Φ(2nx-k). The set {Φ0,k:k} is an orthonormal basis for V0, and one checks that the set {Φn,k:k} is an orthonormal basis for Vn. Therefore

Pn=kΦn,kΦn,k

in the strong operator topology.

For R>0, let fR=fχ[-R,R]. If 2nR<12, then, using the Cauchy-Schwarz inequality,

PnfR22 =k|PnfR,Φn,k|2
=k|fR,Φn,k|2
=k|fR,χ[-R,R]Φn,k|2
k(-RR|fR(x)|2𝑑x)(-RR|Φn,k(x)|2𝑑x)
=fR22k-RR|Φn,k(x)|2𝑑x
=fR22k2n-RR|Φ(2nx-k)|2𝑑x
=fR22k-2nR-k2nR-k|Φ(x)|2𝑑x
=fR22Un|Φ(x)|2𝑑x,

where

Un=k(-k-2nR,-k+2nR);

the intervals are disjoint because 2nR<12. Define Fn(x)=|Φ(x)|2χUn(x). For all x we have |Fn(x)||Φ(x)|2, and if x then

limn-Fn(x)|Φ(x)|2χ(x),

where =nUn. Thus by the dominated convergence theorem we get

limn-Fn(x)𝑑x=|Φ(x)|2χ(x)𝑑x=0,

because μ()=0. Therefore,

limn-PnfR2=0.

If ϵ>0 then there is some R such that f-fR2<ϵ. We have, because Pn is an orthogonal projection,

lim supn-Pnf2 lim supn-Pnf-PnfR2+lim supn-PnfR2
=lim supn-Pnf-PnfR2
lim supn-f-fR2
<ϵ.

This is true for all ϵ>0, so we obtain

limn-Pnf2=0.

If Sα,αI, are subsets of a Hilbert space H, we denote by αISα the closure of the span of αISα. If S is a subset of H, let S be the set of all xH such that yS implies that x,y=0. If Sn,n, are mutually orthogonal closed subspaces of a Hilbert space, we write

nSn=nSn.

The following theorem shows a consequence of Definition 12.

Theorem 15.

If {Vn:n} are the closed subspaces of a multiresolution analysis of L2() and Wn=Vn+1Vn, then

L2()=nWn.
Proof.

Because Wn=Vn+1Vn is the intersection of two closed subspaces, it is itself a closed subspace. Suppose that n<n, fWn,gWn. n+1n, and hence Vn+1Vn. Therefore

Wn=Vn+1VnVnVn+1.

But fWnVn+1 and gWnVn+1, so f,g=0. Therefore WnWn.

If fVn and f0, then there is a minimal N such that fVN; this minimal N exists because VnVn+1 and nVn={0}. We have

VN=VN-1WN-1,

hence f=fN-1+gN-1, with fN-1VN-1 and gN-1WN-1. Likewise,

VN-1=VN-2WN-2,

hence fN-1=fN-2+gN-2, with fN-2VN-2 and gN-2WN-2. In this way, for any M0 we obtain

f=fN-M+m=1MgN-m,

where fN-MVN-M and gN-mWN-m. Check that fN-M is the orthogonal projection of f onto VN-M. It thus follows from Lemma 14 that fN-M0 as M. Thus, for any ϵ>0 there is some M with fN-M2<ϵ and ffN-M+m=1MWN-m. Therefore, if fnVn then there is some gnWn satisfying f-g2<. Thus

nVn¯nWn,

and so

L2()=nWn.

7 The unit interval

L2([0,1)) is a Hilbert space with the inner product

f,g=01f(x)g(x)¯𝑑x.

If n0, then In,0=[0,12n) and In,2n-1=[1-12n,1), and we have

[0,1)=k=02n-1In,k.

Let n0, let 𝒢n be the σ-algebra generated by {In,k:0k2n-1}, and let 𝒢 be the σ-algebra of Lebesgue measurable subsets of [0,1). If n<n, then

𝒢n𝒢n𝒢.

An element of L2([0,1),𝒢n) is an element of L2([0,1),𝒢) that is constant on each set In,k,0k2n-1. Equivalently, an element of L2([0,1),𝒢n) is a function f:[0,1) that is constant on each set In,k,0k2n-1; because [0,1) is a union of finitely many In,k, any such function will be an element of L2([0,1),𝒢). It is apparent that

L2([0,1),𝒢n)L2([0,1),𝒢n)L2([0,1),𝒢).

We check that L2([0,1),𝒢n) is a complex vector space of dimension 2n.

In,k=In+1,2kIn+1,2k+1. If xIn+1,2k, then 2k2n+1x<2k+12n+1, so k2nx<k2n+12n+1, hence 02nx-k<12. If xIn+1,2k+1, then 2k+12n+1x<2k+22n+1, hence k2n+12n+1x<k+12n, and so 122nx-k<1. Thus, if xIn+1,2k then

ψn,k(x)=2n/2ψ(2nx-k)=2n/2

and if xIn+1,2k+1 then

ψn,k(x)=2n/2ψ(2nx-k)=-2n/2.

Otherwise xIn,k, for which ψn,k(x)=0. It follows that ψn,kL2([0,1),𝒢n+1).

Theorem 16.

If

0={χ[0,1)}

and, for n0,

n+1={ψn,k:0k2n-1},

then

n=0Nn

is an orthonormal basis of L2([0,1),𝒢N).

Proof.

It follows from Lemma 2 that n=1Nn is orthonormal in L2([0,1)), as it is a subset of an orthonormal set. If 0nN then nL2([0,1),𝒢N), hence n=1Nn is orthonormal in L2([0,1),𝒢N). If 0<nN and 0k2n-1-1, then ψn-1,kn and

ψn-1,k,χ[0,1) =01ψn-1,k(x)χ[0,1)(x)¯𝑑x
=01ψn-1,k(x)𝑑x
=In,2kψn-1,k(x)𝑑x+In,2k+1ψn-1,k(x)𝑑x
=In,2k2(n-1)/2𝑑x+In,2k+1-2(n-1)/2dx
=0.

Therefore, n=0Nn is orthonormal in L2([0,1),𝒢N).

|0|=1, and if n1 then |n|=2n-1. Therefore the number of elements of n=0Nn is

1+n=1N2n-1=1+n=0N-12n=2N.

As dimL2([0,1),𝒢N)=2N, the orthonormal set n=0Nn is an orthonormal basis for L2([0,1),𝒢N). ∎

By Theorem 16, if N0 then n=0Nn is an orthonormal set in L2([0,1)). Hence

=n=0n

is an orthonormal set in L2([0,1)): if f,g then there is some N with f,gn=0Nn, which is an orthonormal set. The following theorem shows that is an orthonormal basis for the Hilbert space L2([0,1)).77 7 John K. Hunter and Bruno Nachtergaele, Applied Analysis, p. 177, Lemma 7.13.

Theorem 17.

is an orthonormal basis for L2([0,1)).

Proof.

If fL2([0,1)) and ϵ>0 then there is some gC([0,1]) with f-g2<ϵ2. g is uniformly continuous on the compact set [0,1], so there is some δ>0 such that |x-y|<δ implies that |g(x)-g(y)|<ϵ2. Let 2-nδ, and define h:[0,1) by

h(x)=k=02n-1g(k2n)χIn,k(x).

If x[0,1) then there is a unique kx,0kx2n-1, with xIn,kx, and for this kx we have |x-kx2n|<2-nδ, and hence

|g(x)-h(x)|=|g(x)-g(kx2n)|<ϵ2.

Therefore g-hϵ2.

We have hL2([0,1),𝒢n), and

f-h2f-g2+g-h2<ϵ2+g-hϵ.

We have shown that if fL2([0,1)) and ϵ>0 then there is some n and some hL2([0,1),𝒢n) with f-h2<ϵ. This tells us that n=0L2([0,1),𝒢n) is a dense subset of L2([0,1)). Since is orthonormal and span=n=0L2([0,1),𝒢n), is an orthonormal basis for L2([0,1)). ∎

8 References

Useful references on wavelets and multiresolution analysis are Mark A. Pinsky, Introduction to Fourier Analysis and Wavelets; P. Wojtaszczyk, A Mathematical Introduction to Wavelets; Yves Meyer, Wavelets and Operators; Eugenio Hernández and Guido Weiss, A First Course on Wavelets.