LECTURE 8, 18.155, 4 OCTOBER, 2011 This week: Sobolev Embedding and Schwartz Structure Theorems. We now know that the Fourier transform ‘extends’ to an isomorphism of S 0 (Rn ). This was done by using the formula, for a pair of functions in S(Rn ) Z Z φ̂ψ = φψ̂ to define the Fourier transform by (1) û(φ) = u(φ̂), u ∈ S 0 (Rn ), φ ∈ S(Rn ) since we know that S(Rn ) 3 φ 7−→ φ̂ ∈ S(Rn ) is continuous (in fact it is a topological isomorphism). Then we showed that the Fourier transform restricts from S 0 (Rn ) (on which it is a bijection, but we have not given any continuity statement) to a bijection on L2 (Rn ) which is essentially an isometric isomorphism, i.e. (2π)−n/2 F is unitary on L2 (Rn ) – this follows from by the version of (1) Z Z −n φψ = (2π) φ̂ψ̂ ∀ φ, ψ ∈ S(Rn ). Next we showed that the L2 based Sobolev spaces are readily characterized in terms of the Fourier transforms of their elements: u ∈ H k (Rn ) ⇐⇒ hξik û ∈ L2 . Here k ∈ N but we can use this as a definition of H k (Rn ) for all k ∈ R where I usually replace the integer-looking k by s or m. Then it follows that these spaces decrease with increasing order 0 H s (Rn ) ⊂ H s (Rn ) ⇐⇒ s ≥ s0 which may be unfortunate but that is the way things are. From this we easily deduce a form of ‘weak=strong’ for derivatives. Namely, u ∈ L2 (Rn ) is said to have a strong derivative, v = ∂j u if there is a sequence un ∈ S(Rn ) (or Cc∞ (Rn ) it is the same in the end) such that uk → u in L2 and ∂j uk → v in L2 (Rn ). Such a strong derivative (of any order) is automatically a weak derivative since we can integrate by parts before taking the limit so u(−∂j φ) = lim uk (−∂j φ) = lim (∂j uk )(φ) = v(φ). k→∞ k→∞ 1 2 LECTURE 8, 18.155, 4 OCTOBER, 2011 In particular if a strong derivative exists then it is unique since weak derivatives are already unique. Proposition 1 (weak=strong). If k ∈ N then S(Rn ) ⊂ H k (Rn ) is dense (in the norm on H k ) and each u ∈ H k (Rn ) has strong derivatives or all orders |α| ≤ k. Proof. Take u ∈ H k (Rn ), so û ∈ L2 (Rn ) and û(ξ) = hξi−k w, w ∈ L2 (Rn ) by the result above. Since S(Rn ) is dense in L2 we can choose a sequence wl → w in L2 . Then choose ul ∈ S(Rn ) by ûl = hξi−k wk where we use the fact that h·i−k is a multiplier on S and that the Fourier transform is an isomorphism. The norm on u ∈ H k is equivalent to kwkL2 when u and w are related as above, so wl → w in L2 implies ul → u in H k (Rn ) and we have the density. Since ξ α ûl = ξ α hξh−k wl , the same argument shows that ∂ α ul converges in L2 , so strong derivatives exist for |α| ≤ k. Of course if s ≥ k then u ∈ H s (Rn ) has strong derivatives of any order up to k – but what about the ‘extra’ s − k ‘derivatives’ ? We will get to this later. From this discussion the basic result relating ‘L2 -derivatives’ to ordinary derivatives also follows. Theorem 1 (Sobolev embedding). If k ∈ N then n H s (Rn ) ⊂ C0k (Rn ) if s > + k. 2 This result is sharp. The spaces on the right are the spaces of functions with k continuous, bounded derivatives each vanishing at infinity. Proof. The main point here is that hξi−s L2 (Rn ) ⊂ L1 (Rn ) if (and only if) s > n2 . By this I mean the space of those v ∈ S 0 (Rn ) such that v = h·i−s w, w ∈ L2 . Indeed, this is just the fact that h·i−p ∈ L1 (Rn ) iff p > n. So, if we apply the Cauchy-Schwarz inequality, Z Z |v| = hξi−s |w| ≤ ki·i−2s kL2 kwkL2 the result follows. Well, we really need to check that the integral makes sense first! This however is the usual argument by continuity – take an approximating sequence wn → w with wn ∈ Cc0 (Rn ) for which the computation does make sense, and hence deduce that vn = h·i−s wn → v a.e. is an approximating sequence in L1 , so v ∈ L1 and the estimate holds. Now, we apply this to v = û, where by assumption u ∈ H s (Rn ) means that w(ξ) = hξis û ∈ L2 (Rn ). Thus, we see that v ∈ L1 and so, LECTURE 8, 18.155, 4 OCTOBER, 2011 3 as you showed in the last problem set, n . 2 For higher derivatives, we just need to use density an iteration as usual. Take a sequence as in the preceding result, ul → u in H s (Rn ) with ul ∈ S(Rn ). Then, ∂ α ul is a convergent sequence in H s−|α| (Rn ), so if s − |α| > n2 it follows that ∂ α ul → ∂ α u ∈ C00 (Rn ) where the notation is for the weak (or strong) derivative of u. That is, we conclude that the weak derivatives up to order k are continuous. In fact we have been through these arguments before, the fact that the sequence converges with respect to the C k norm (supremum of derivatives up to order k) means the limit is in C0k (Rn ) – one just needs to quote the completeness of C0k (Rn ). u = G û ∈ C00 (Rn ), if u ∈ H s (Rn ), s > Now, we ‘know’ that the Fourier transform exchanges regularity and growth. Let’s look at another (extreme) example of this. Definition 1. An element u ∈ S 0 (Rn ) is said to be of compact support (or have compact support, or be compactly supported) if there exists R > 0 such that u(φ) = 0 ∀ φ ∈ S(Rn ) s.t. φ(x) = 0 in |x| < R. We will discuss the notion of the support of a distribution soon. Another way of looking at this property is to say that there exists χ ∈ Cc∞ (Rn ) such that χu = u. Indeed, take χ ∈ Cc∞ (Rn ) to be identically equal to 1 in say |x| < R + 1 and to vanish identically in |x| > R + 2. Then if φ ∈ S(Rn ) it follows that (1−χ)φ = 0 in |x| < R and (1−χ)φ ∈ S(Rn ) so u((1 − χ)φ) = 0 for all φ ∈ S(Rn ) which is u(φ) = (χu)(φ) for all φ ∈ S(Rn ). Conversely, if χu = u with χ ∈ Cc∞ (Rn ) it follows χ(x) = 0 in |x| > R − 1 for some R and then if φ ∈ S(Rn ), and φ = 0 in |x| < R it follows that χφ ≡ 0 and hence χu(φ) = u(χφ) = 0 which is the the definition. One of the Assignment Questions is: Discuss the Paley-WienerSchwartz Theorem. Here is an extremely watered-down version (the true version characterizes the Fourier transforms of compactly supported distributions in terms of entire functions of several complex variables). Lemma 1. If u ∈ S 0 (Rn ) has compact support then û ∈ C ∞ (Rn ) is a function of slow growth. Proof. The important properties of a distribution of compact support are (duh!) it is a distribution and it has compact support. The first 4 LECTURE 8, 18.155, 4 OCTOBER, 2011 means that there are constants C and N such that X |u(φ)| ≤ CkφkN = C sup |xα ∂ β φ|. |α|+|β|≤N The second means u = χu so u(φ) = u(χφ) and we see what happens to the right side when φ is replaced by χφ : X sup |xα ∂ β (χφ)| ≤ C sup |∂ β φ| |β|≤N |x|≤T where T is such that χ = 0 in |x| ≥ T. So, for compactly supported distributions we get a better estimate in terms of the C N seminorm on a fixed set |x| ≤ T |u(φ)| ≤ CkφkC N ({|x|≤T }) . So, what is the Fourier transform of u, well, it is defined by û(φ) = u(φ̂) Let’s consider the function involved in the definition of the Fourier transform, exp(ix · ξ) as a function of x. It is smooth, but of course not in S(Rn ). However, if χ ∈ Cc∞ (Rn ) then χ(x) exp(ix · ξ) ∈ S(Rn ). So we can define a function of ξ by F (ξ) = u(χ(x) exp(ix · ξ) ∈ C ∞ (Rn ). First, it is continuous, since as ξ varies, χ(x) exp(−ix · ξ) varies continuously in S(Rn ) and u is continuous, being a distribution. Moreover, we can take the difference quotient and see that h−1 (χ(x) exp(−ix · (ξ + hej )) − χ(x) exp(−ix · ξ)) → −ixχ exp(ix · ξ) converges in S(Rn ). Iterating this argument gives the smoothness of F (ξ). Finally then, we claim that is that F (ξ) is the Fourier transform of u. It has slow growth, since the norm estimate above shows |F (ξ)| ≤ ChξiN so it is a tempered distribution. Moreover, if we integrate against χ(ξ)φ̂ where φ ∈ S(Rn ), Z F (ξ)φ(ξ)dξ = u(χφ̂) = u(φ̂) = û(φ) showing that F (ξ) = û in S 0 (Rn ). The first step is the crucial one. Think about computing u(χφ̂) in terms of φ. Let’s assume that φ itself has compact support. Then the integral for φ̂ is a Riemann integral and so is the limit of a sum Z X φ̂(x) = exp(−iξ · x)φ(ξ) = lim 2−nj exp(−iξk,j · x)φ(ξk,j ) j→∞ k LECTURE 8, 18.155, 4 OCTOBER, 2011 5 Here ξk,j is an ordering of the finite number of points in 2−j Zn in a big ball containing the support of φ and the constant is the volume of the corresponding cube(s). As a function of x in a compact set this approximation to the Riemann integral actually converges uniformly with all its derivatives – just differentiate to see this. So X χ(x)2−nj exp(−iξk,j · x)φ(ξk,j ) → χ(x)φ̂(x) in S(Rn ). k By the continuity of u this implies X exp(−iξk,j · x)φ(ξk,j ) u(χφ̂) = lim u(χ(x)2−nj j→∞ k → χ(x)φ̂(x) in S(Rn ) Z X =⇒ lim F (ξkj )φ = F (ξ)φ(ξ) j→∞ k again by the properties of Riemann integrals. So, for χ of compact support, û(φ) = F (φ) but both sides are tempered distributions so by density of Cc∞ (Rn ) in S(Rn ) û = F. I said we would check that the extension of the Fourier transform (etc) to S 0 (Rn ) was forced by continuity here is the first version. We can say that a sequence uj ∈ S 0 (Rn ) converges weakly to u ∈ S 0 (Rn ) if uj (φ) → u(φ) in C for all φ ∈ S(Rn ). The limit is certainly unique and we can use the Fourier transform on S(Rn ) to prove that S(Rn ) ⊂ S 0 (Rn ) is weakly dense. I did not finish the proof of this last result.