University of Leeds, School of Mathematics Analysis (wavelets and operators): Some useful results and formulae (1) J.R. Partington March 2003 Please let me know of any errors, omissions or obscurities! R∞ Lp spaces: functions in Lp (R) are measurable and satisfy kf kp = ( −∞ |f (t)|p dt)1/p < ∞, regarding two functions as the same if they are equal almost everywhere. R ∞ These are 2 Banach spaces, and L (R) is a Hilbert space with inner product hf, gi = −∞ f (t)g(t) dt. For L∞ (R), take the norm to be the (essential) supremum of |f (t)|. FourierP series: L2 (0, T ) has an orthonormal basis, en (t) = T −1/2 exp(2πint/T ), n ∈ Z, ∞ so f = n=−∞ hf, en ien , converging in L2 norm. R∞ Fourier transforms: for f ∈ L1 (R) define fˆ(w) = −∞ f (t) exp(−2πiwt) dt, for w ∈ R, and write F f = fˆ. This extends to a bounded linear operator F from L2 (R) onto L2 (R), and kfˆk2 = kf k2 (Plancherel). Moreover, (fˆ)ˆ(t) = f (−t), that is, R∞ f (t) = −∞ fˆ(w) exp(2πiwt) dw, a.e., at least if fˆ ∈ L1 (R) ∩ L2 (R). We get as a corollary that hf, gi = hfˆ, ĝi, for f , g ∈ L2 (R). Write F̌ for the inverse Fourier transform of F . Haar wavelets: let V0 ⊂ L2 (R) be the closed subspace of all functions f which are constant on all intervals (k, k + 1), k ∈ Z. Let φ(t) = χ(0,1) (t) and φk (t) = φ(t − k) for k ∈ Z. Then (φk )k∈Z is an orthonormal basis of V0 . Now for j ∈ Z let Vj be the space of functions constant on all intervals (k/2j , (k + 1)/2j ), k ∈ Z. S T Then . . . ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ . . .. Also Vj = L2 (R) and Vj = {0}. Now f (t) ∈ Vj ⇐⇒ f (2−j t) ∈ V0 , and Vj has o.n.b. consisting of 2j/2 φ(2j t − k), k ∈ Z. Any chain of subspaces with these properties is called a multiresolution approximation or multiresolution analysis (MRA) of L2 (R). Define the Haar wavelet ψ(t) = φ(2t) − φ(2t − 1). The functions ψk (t) = ψ(t − k), k ∈ Z form an o.n.b. for a space W0 such that V0 ⊕ W0 = V1 (orthogonal direct sum). Then Vj ⊕ Wj = Vj+1 , where Wj has o.n.b. ψj,k (t) = 2j/2 ψ(2j t − k), k ∈ Z. Finally L2 (R) = . .P . ⊕ W−2P ⊕ W−1 ⊕ W0 ⊕ W1 ⊕ . . . and has orthonormal basis (ψj,k )j,k∈Z . ∞ ∞ Hence f = j=−∞ k=−∞ hf, ψj,k iψj,k , converging in L2 norm. Band-limited functions: f ∈ L2 (R) is time-limited, if ∃T > 0 s.t. f (t) = 0 a.e. for |t| > T ; and f is band-limited if fˆ is time-limited. Paley–Wiener space: P W (b) = {f ∈ L2 (R) : fˆ(w) = 0 a.e. for |w| > b}. Such functions are restrictions to R of entire functions satisfying |f (z)| ≤ kfˆk1 e2πby , for z = x + iy ∈ C. Hence the only function both time- and band-limited is 0 a.e. Whittaker–Kotel’nikov–Shannon sampling theorem: for f ∈ P W (b), we have f (t) = P∞ (1/2b) n=−∞ f (n/2b)kn/2b (t), where ks (t) = sin 2πb(t − s)/(π(t − s)), the reproducing √ kernel function. Moreover, the functions (1/ 2b)kn/2b (t) form an o.n.b. of P W (b) and hf, ks i = f (s) for all s ∈ R: this is the reproducing kernel Hilbert space (r.k.H.s.) property. Other r.k.H.s. include the Hardy space H 2 on the open unit disc D, where kw (z) = 1/(1 − wz); also the space Cn on the set {1, 2, . . . , n}. In general ks (t) = kt (s) and ks (s) > 0 unless all functions vanish at s. Now define V0 = P W (1/2) and Vj = P W (2j−1 ) for j ∈ Z. These give a multiresolution approximation, as defined above, and an orthonormal basis of Vj is given by 2j/2 φ(2j t − k), k ∈ Z, where φ(t) = sin πt/(πt) = k0 (t). Then the Littlewood–Paley wavelet is ψ(t) where ψ̂(w) = χ(1/2,1) (w), i.e., 1 ψ(t) = (sin 2πt − sin πt)/(πt). As before we obtain an o.n.b. of L2 (R), consisting of wavelets 2j/2 ψ(2j t − k), j, k ∈ Z, where for a fixed j they form an o.n.b. of a space Wj such that Vj ⊕ Wj = Vj+1 . A Riesz basis (φk ) in a Hilbert spacePH satisfies (i) P H is the closedPlinear span of 2 1/2 the (φk ), and (ii) ∃A, B > 0 s.t. A( k |ak | ) ≤ k ak φk k ≤ B( k |ak |2 )1/2 for all (ak ) ∈ ℓ2 (finite Riesz bases in finite-dimensional spaces are also allowed, indeed any algebraic basis is a Riesz basis in this case). Then (φk ) is a Riesz basis ⇐⇒ ∃U : H → H bounded, linear, invertible, such that (U φk ) is an orthonormal basis. P A frame (φk ) in H satisfies ∃A, B > 0 s.t. Akφk2 ≤ k |hφ, φk i|2 ≤ Bkφk2 for all φ ∈ H. Then o.n.b. =⇒ R.b. =⇒ frame. The numbers hφ, φk i are called the frame coefficients of φ. In a finite-dimensional space, a frame is just any finite spanning set. Now (φk ) is an infinite linear map T : ℓ2 → H (ONTO) P frame iff there is a bounded such that T ((ak )) = ak φk for all (ak ) ∈ ℓ2 . Define the mapping S = T T ∗ : H → H P by Sψ = hψ, φk iφk . Then S is a positive, bounded, invertible operator, and (φek ) = (S −1 φk ) is also a frame, the dual frame to (φk ). The dual frame to (φek ) is (φk ) once P more. Also ψ = hψ, φk iφek . We can invert S using P∞ −1 S = (2/(B + A)) n=0 (I − 2S/(B + A))n , which converges rapidly. Windowed Fourier transforms: take g ∈ L2 (R) of norm 1, often real and positive; write ft (u) = f (u)g(u − t), for f ∈ L2 (R) and t ∈ R. Define fe(w, t) = fˆt (w). This also ê equals e−2πiwt f (−t, w), where the RHS is formed using the window ĝ (time-frequency R∞ R∞ localization). Inversion formula: f (u) = −∞ −∞ fe(w, t)g(u − t)e2πiwu dw dt, when all integrals converge absolutely. Also kfek2 = kf k2 , i.e., we get a norm-preserving linear operator W : L2 (R) → L2 (R2 ) (not onto). This can be discretized: fe(mw0 , nt0 ) = hf, gmni, where gmn (u) = e2πimw0 u g(u − nt0 ). For suitable w0 , t0 > 0, these form a frame, and we can invert the discrete transform. Heisenberg’s inequality: for f ∈ L2 (R), ktf (t)k2 kwfˆ(w)k2 ≥ kf k22 /(4π), with equality for the Gaussian function. R∞ Wavelet transform: let ψ ∈ L2 (R) satisfy Cψ = −∞ (|ψ̂(w)|2 /|w|) dw < ∞, and write ψ x,y (t) = |x|−1/2 ψ((t−y)/x), for x, y ∈ R withRxR 6= 0. Define f ◦ (x, y) = hf, ψ x,y i. Better ‘zooming’ properties than the W.F.T. Also f ◦ (x, y)g ◦(x, y) dx dy/x2 = Cψ hf, gi, giving a weak inversion formula. The transform maps L2 (R) isometrically into a subspace H of L2 (R2 , µ), where dµ = dx dy/(Cψ x2 ), which is a r.k.H.s. on R2 with kernel k(s,t) (x, y) = hψ x,y , ψ s,ti. Discretising: write ψj,k (t) = aj/2 ψ(aj t − kb) for some fixed a > 1 and b > 0 and consider inner products hf, ψj,k i. Again we would like a frame (e.g. for the Mexican 2 hat function C(1 − x2 )e−x /2 , the choice a = 2 and b = 1 will do). Other wavelets can be constructed using a MRA (Vj )j∈Z , a father wavelet φ and a mother wavelet ψ, as in the two examples above. Wavelet Properties of ψ(t) Properties of ψ̂(w) Haar Compact support, discontinuous O(1/w), C ∞ ∞ Littlewood–Paley O(1/t), C Compact support, discontinuous ∞ Meyer Rapidly-decreasing, C Compact support, can be C ∞ k Battle–Lemarié Rapidly-decreasing, C O(1/wk ), C ∞ Daubechies Compact support, C k O(1/wk ), C ∞ Compact support of both ψ and ψ̂ is always impossible; exponential decay and smoothness for both isn’t possible for o.n.b. of wavelets, but is for frames (e.g. Mexican hat). 2