• Aucun résultat trouvé

4.2.1 Multiresolution analysis

We consider a pair{φ, ψ}offather andmother wavelets defining a multiresolution analysis (MRA), see Cohen [2003]. This is the case for the celebrated Daubechies wavelets. MRA wavelets are widely used in practice, since they allow a fast computation of the discrete wavelet transform (DWT) using the pyramidal algorithm; see for instance Mallat [1998].

We adopt the engineering convention under which large values of the scale indexj corre-spond to coarse scales, that is, the analyzing wavelets are defined as{ψj,k, j ∈Z, k∈Z} with ψj,k(t) = 2j/2ψ(2jt−k), j ∈ Z, k ∈ Z. The DWT of X can be interpreted as follows. Using the father wavelet (orscaling function) φand the interpolated functionX defined in (2.30), the wavelet coefficients are defined as

Wj,k def= Z

X(t)ψj,k(t) dt, j≥1, k∈Z. (4.1) Based on the quadrature mirror filters h and g defined respectively in (2.45) and (2.48), one can compute the wavelet coefficients as follows

A0,k =Xk, k∈Z, (4.2)

Aj,k = X n=−∞

h[2k−n]Aj1,n=

2(h ⋆ Aj1,·)

k, k∈Z, j≥1, (4.3) Wj,k =

X n=−∞

g[2k−n]Aj1,n =

2(g ⋆ Aj1,·)

k, k∈Z, j≥1, (4.4) where {h[k]}k∈Z and {g[k]}k∈Z denote the impulse response of h and g, ⋆ denotes the convolution operator and↓2 the down-sampling operator by factor 2. We assume that ψ is compactly supported and denote by M the number of vanishing moments (M ≥ 1).

This corresponds to the following assumption on the filtersh and g.

Assumption 1. The filtershandg have support{−T+ 1, . . . ,0}andg has M vanishing moments, that is,

X0 t=T+1

g[t]t = 0 for all ℓ= 0,1, . . . , M−1. (4.5)

This assumption is supposed to hold throughout the chapter. It follows from the finite support assumption that a finite set of wavelet coefficients can be computed exactly from a finite sampleX0, . . . , Xn1. In this case the pyramidal algorithm is applied with k= 0, . . . , n0−1 in (4.2) and k= 0, . . . , nj−1 in (4.3) and (4.4), where

n0 =n and nj ={⌊(nj1−T)/2⌋+ 1}+, j≥1, (4.6) and⌊x⌋denotes the largest integer at most equal toxand x+ the non-negative part ofx, max(x,0).

4.2.2 pyramidal algorithm based on approximation coefficients incre-ments

The DWT (4.2–4.4) with input {Xk,0 ≤k < n}, provides{Wj,k, j ≥1,0≤k < nj} as output. We now wish to provide the same output with

[∆MX] k, M ≤k < n.

To this end we define the approximation coefficients increments at successive orders ℓ= 0,1, . . . by A(ℓ)j,k = [∆Aj,·]k, that isA(0)j,k =Aj,k and for ℓ≥0,A(ℓ+1)j,k =A(ℓ)j,k−A(ℓ)j,k1. Because X may not be stationary but becomes stationary after differencing sufficiently many times, it is a key step to derive a pyramidal algorithm in whichAj,kin (4.2) and (4.3) is replaced byA(ℓ)j,k withℓ the large enough. By assumption 1 we may take ℓ=M. This new scheme is stated in Lemma 2.2.1.

4.3 Second order properties in the wavelet domain

4.3.1 Assumptions and notation

LetXbe anM(d) process with memory parameterdand with generalized spectral density f. By Definition 2.1.10 there exists K such that ∆KX is second order stationary. By possibly increasingKto remove possibly remaining polynomial trends, we further assume that∆KX is centered. In fact, we will not use in this section that X is anM(d) process but only the two following assumptions.

(H-i) TheKth order increment process ∆KX is centered and covariance stationary.

(H-ii) Assumption 1 holds withM ≥K.

Let us now introduce some notation valid all along the chapter. For any scales i and j, we denote by Mi,j def= (Mi,j[k, k],0≤k≤ni−1,0≤k ≤nj−1) the covariance matrix of the wavelet coefficients at scalesiand j, respectively,

Mi,j[k, k]def= Cov(Wi,k, Wj,k). (4.7)

Following Moulines et al. [2007b], we pool blocks of 2u wavelet coefficients at scale j−u within column vectors denoted by

Wj,k(u)def= [Wju,2uk, Wju,2uk+1, . . . , Wju,2uk+2u1]T , (4.8) and define thebetween-scale process at scalej, timek and scale differenceu, as

{[WTj,k(u), Wj,k]T}kZ.

Observe thatWj,k(u) is a 2u-dimensional vector of wavelet coefficients at scalei=j−u and involves all possible translations of the position index 2uk by v = 0,1, . . . ,2u −1.

The indexu in (4.8) denotes the scale differencej−i≥0 between the finest scale iand the coarsest scale j. By convention Wj,k(0) (u = 0) coincides with the scalar Wj,k. By [Moulines et al., 2007b, Corollary 1],{[Wj,kT (u), Wj,k]T}k∈Z is then covariance stationary.

In contrast, the process of wavelet coefficient at the finer scale j−u and the approx-imation coefficient at scale j {[WTj,k(u), Aj,k]T}kZ may not be stationary. Nevertheless, it follows from (H-i) and (H-ii) above that {A(M)0,k , k ∈ Z} is centered and covariance stationary, hence we conclude from Lemma 2.2.1 that{[WTj,k(u), A(M)j,k ]T}kZ is centered and covariance stationary with second order properties described below using iterative formula. Before that we need some additional notation.

Let us define the covariance functions δi,ji,j andνi,j by

γi,j[k]def= Cov(Wi,k, A(M)j,0 ), δi,j[k]def= Cov(Wi,k, Wj,0), 1≤i≤j , νi[k]def= Cov(A(M)i,k , A(Mi,0)), 0≤i .

When i=j, we denoteγi,ji and δi,ji. Using these definitions and the stationary structure in the wavelet coefficients, we obtain that

Mi,j[k, k] =δi,j[k−2jik], 1≤i≤j .

Similarly we introduce some notation for spectral densities. We denote byDj,u the cross-spectral density function of the between-scale process{[WTj,k(u), Wj,k]T}k∈Z and byBj,u the cross-spectral density function of the processn

[Wj,kT (u), A(Mj,k)]To

k∈Z. Hence, we have, for all 1≤i=j−u≤j,

i,j[2uk], . . . , δi,j[2u(k+ 1)−1]]T = Cov (Wj,k(u), Wj,0) = Z π

π

Dj,u(λ) eiλkdλ (4.9) [γi,j[2uk], . . . , γi,j[2u(k+ 1)−1]]T = Cov

Wj,k(u), A(M)j,0

= Z π

π

Bj,u(λ) eiλkdλ . (4.10) By conventionDj,0 (u= 0) denotes the spectral density of{Wj,k, k∈Z}. Finally we let Aj denote the spectral density function of the process {A(M)j,k }k∈Z. Hence, for allj ≥0,

νj[k] = Cov

A(Mj,k), A(Mj,0)

= Z π

π

Aj(λ) eiλkdλ . (4.11)

4.3.2 Iterative formula in spectral domain

Observe that A0 and ν0 are the spectral density and autocovariance function of ∆MX, respectively. In particular, if f denotes the generalized spectral density ofX,

A0(λ) =|1−e|2Mf(λ) and ν0(τ) = Z π

π

A0(λ)e

are easily derived for second order properties of X. We now wish to obtain the second order properties of the wavelet coefficients. This can be done either in the spectral domain, that is, by computingDj,u for allj ≥1 and 0≤u≤j, or in the time domain, that is by computingδi,j for all 1≤i≤j. Other spectral densities Aj and Bj,u, or covariances γi,j andνi,j are key extra quantities to achieve this goal. The following results indeed provide simple iterative to compute all these quantities simply starting fromA0 orν0.

The first result provides iterative formula for the spectral densities Aj, Dj,u(·), and Bj,u(·),u∈Z, based on the filterseh and eg defined in (2.56) and (2.55). The proof of Proposition 4.3.1 is postponed in Section 4.5

Remark 4.3.1. All the quantities in Proposition 4.3.1 can be successively computed from A0 as follows. The spectral densities Aj, j ≥ 1, can be computed from A0 using the iterative formula (4.12). The cross-spectral densitiesDj,0andBj,0,j≥1, can be computed from Aj, j ≥ 0, using (4.13) and (4.14). Finally, Bj,u and Dj,u, 1 ≤ j, 0 ≤ u ≤ j−1 can be computed from Bj,0, j≥1 using the iterative formula (4.15) and (4.16).

4.3.3 Iterative formula in temporal domain

The second result provides iterative formula for the covariance functionsνii,j and γi,j, 1≤i≤j, based on the filterseh and egdefined in (2.56) and (2.55).

Proposition 4.3.2. The following formula hold for allk∈Z and 1≤i≤j.

νj[k] = X

l,l∈Z

e

h[l]eh[lj1

2k−l+l

, (4.18)

δj[k] = X

l,l∈Z

e

g[l]eg[lj1

2k−l+l

, (4.19)

γj[k] = X

l,l∈Z

e

g[l]eh[lj1

2k−l+l

, (4.20)

γi,j+1[k] =X

l∈Z

eh[l]γi,j[k+ 2jil], (4.21) δi,j+1[k] =X

l∈Z

e

g[l]γi,j[k+ 2jil]. (4.22) The proof of Proposition 4.3.2 in postponed in Section 4.5

Remark 4.3.2. All the quantities in Proposition 4.3.2 can be successively computed from ν0 as follows. The autocovariance functionsνj, j≥1, can be computed fromν0 using the iterative formula (4.18). The covariance functions δj and γj can be computed from νj, j≥0, using (4.19) and (4.20). Finally γi,j and δi,j, 1≤i < j can be computed from γj, j≥1 using the iterative formula (4.21) and (4.22). Note also that the double summation and simple summation signs appearing in (4.18-4.22) only involve (T −M)2 and T−M non-vanishing terms, since eh and eg have lengths both equal to T −M.