LECTURE 7, 18.155, 29 SEPTEMBER 2011 (1) Last time I defined the Sobolev spaces of integral order in terms of weak derivatives. To understand the L2 based spaces, H k (Rn ) = {u ∈ L2 (Rn ); ∂ α u ∈ L2 (Rn ) ∀ |α| ≤ k} we will use the Fourier transform. Just as with derivatives last time, this will be defined ‘weakly’ (so the definition above says – u in L2 ,→ S 0 (Rn ) and its derivatives, ∂ α u ∈ S 0 (Rn ) happen to be in (the image of) L2 ⊂ S 0 (Rn ). Let me first remind you of the process. (2) We showed earlier (I think you did all the work) that multiplication by functions of slow growth gives a map on S(Rn ) – if f ∈ C ∞ (Rn ) and for each k there exists N such that |∂ α | ≤ CN (1 + |x|)N for all |α| ≤ k then ×f : S(Rn ) −→ S(Rn ). We can see that this has a ‘weak definition’ as well. Namely, if we regard ψ ∈ S(Rn ) as a distribution, denoted for the moment Z Uψ (φ) = ψφ then we get the identity Z Uf ψ (φ) = f ψφ = Uψ (f φ). We want to extend multiplication by f to S 0 (Rn ), consistent with its action on S(Rn ), so we want f Uψ = Uf ψ and then we see how to do it in general, just define f u(φ) = u(f φ) ∀ u ∈ S 0 (Rn ), φ ∈ S(Rn ). This makes sense since f u ∈ S 0 (Rn ), because of the continuity on S of multiplication by f. Clearly this gives a linear map from S 0 (Rn ) to itself. All trivial, but we do the same thing for the Fourier transform (3) So we need to get a ‘weak’ formulation of the Fourier transform on S(Rn ). What is ψ̂ as a distribution? For one last time lets denote it Uψ̂ and compute away: Z Z Uψ̂ (φ) = ψ̂φ = ψ̂(ξ)φ(ξ). 1 2 LECTURE 7, 18.155, 29 SEPTEMBER 2011 Now, we can substitute the definition of ψ̂ and not have to worry too much about convergence, since everything is in S : Z Z Z Z −ix·ξ e−ix·ξ ψ(x)φ(ξ)dxdξ. ψ̂(ξ) = e ψ(x)dx =⇒ ψ̂φ = We can exchange the order of integration without much trouble (do a limiting argument if you are not sure) and conclude that Z Z Z Z −ix·ξ ψ̂φ = ψ(x) e φ(ξ)dξdx = ψ φ̂. This is the weak formulation of the Fourier transform we want, it says that Uψ̂ (φ) = Uψ (φ̂). Thus, we can extend the Fourier transform to all of S 0 (Rn ) by defining (1) û(φ) = u(φ̂) ∀ φ ∈ S(Rn ). Proposition 1. The Fourier transform of tempered distributions, defined by (1) is a bijection on S 0 (Rn ). Proof. First observe that if u ∈ S 0 (Rn ) then (1) does indeed define an element û ∈ S 0 (Rn ). In fact û = u◦F : S(Rn ) −→ C is the composite of continuous linear maps, so is itself continuous an linear hence an element of S 0 (Rn ). The fact that F is a bijection on S(Rn ) means that Fu = û defines a bijection – since û = 0 means u(φ̂) = 0 for all φ and hence u(ψ) = 0 for all ψ ∈ S(Rn ). So it is an injective linear map and its inverse is given by the inverse of the Fourier transform on S(Rn ), v = û =⇒ u = Gv, Gv(φ) = v(Gφ) ∀ v ∈ S 0 (Rn ), φ ∈ S(Rn ). This follows from the identity c = u(φ). G û(φ) = û(Gφ) = u(Gφ) (4) For the moment I will not discuss continuity of the Fourier transform, but it is in fact continuous in the natural topologies on S 0 (Rn ). We do want to observe that the relationship of derivation and Fourier transform, which we found on S(Rn ) carries over to S 0 (Rn ) : d α u = i|α| ξ α û, x β u = (−i)|β| ∂ β û. ∂d LECTURE 7, 18.155, 29 SEPTEMBER 2011 3 These follow from the weak definition; for instance α u(φ) = (∂ α u)(φ̂) = u((−1)|α| ∂ α φ̂) ∂d α u) = i|α| û(ξ α φ) = i|α| ξ α û(φ) = (−1)|α| u((−i)α ξd using the identities on S in the middle. (5) Next we consider the restriction of F from S 0 (Rn ) to L2 (Rn ); in fact we will discuss this by extension from S(Rn ). Going back to the identity (2), or (3), we can set φ = µ̂ where µ ∈ S(Rn ) so Z Z ψ̂φ = ψ̂ µ̂. Now, µ is the inverse Fourier transform of its Fourier transform:Z −n µ(x) = (2π) eix·ξ µ̂(ξ)dξ. Taking complex conjutates we see that Z −n b = (2π)−n φ̂. µ(x) = (2π) e−ix·ξ µ̂(ξ)dξ =⇒ µ(x) = (2π)−n µ (2) Substituting this in (3) gives the Plancherel/Parseval identity (I am hopelessly confused about which is which) Z Z n φ̂µ̂ = (2π) ψµ ∀ φ, µ ∈ S(Rn ). Now, from this it follows that Proposition 2. The Fourier transform extends by continuity (using the density of S(Rn ) in L2 (Rn ) to an essentially isometric isomorphism F : L2 (Rn ) −→ L2 (Rn ); (3) kFukL2 = (2π)−n/2 kukL2 . Proof. The norm identity follows from the discussion above when u ∈ S(Rn ). Then if S(Rn ) 3 uk → u in L2 (Rn ) it follows, applying the identity to the norm of differences, that the sequence Fuk is Cauchy, and hence convergent, in L2 (Rn ) and (3) holds for the limit – from which it follows that F : L2 −→ L2 is well-defined and satisfies this identity. It is necessarily injective (since Fuk = 0 implies u = 0) and has inverse given by G to which the same extension discussion applies. (6) So, now we have a simple ‘characterization’ of the image under the Fourier transform of L2 (Rn ) ⊂ S 0 (Rn ), namely it is L2 (Rn ) again. Thus if you believe that F ‘interchanges growth and regularity’ then L2 has the same growth and regularity! We can 4 LECTURE 7, 18.155, 29 SEPTEMBER 2011 use this basic result to get a rather useful characterization of the Fourier transforms of the L2 -based Sobolev spaces H k (Rn ) for k ∈ N. Before doing so, let’s recall some simple estimates. The usual notation |x| for x ∈ Rn is the Euclidean norm, |x|2 = x21 + · · · + x2n . We could also consider the ‘l1 ’ norm |x1 | + · · · + |xn |. As norms on finite dimensional vector spaces these are equivalent, namely (4) 1 (|x1 | + · · · + |xn |) ≤ |x| ≤ |x1 | + · · · + |xn |. n I also introduced the notatio 1 hxi = (1 + |x|2 ) 2 which has the virtue of being a smooth function. In fact this, and its powers, satisfy symbol estimates (5) ∂xα hxhm ≤ Cm,α hxim−|α| . The point to note is that the three functions (6) (1 + |x|), (1 + |x1 | + |x2 | + · · · + |xn |), hxh are all the ‘same size’ on Rn – each is bounded by a positive multiple of the other two, and hence above an below and the same is true for any fixed power m (the same for all three of course). You can check this from (4) and 1 (1 + |x|)2 ≤ (1 + |x|2 ) ≤ (1 + |x|)2 . 2 Since the quotients are bounded continuous functions for u ∈ 2 L and m > 0 the three conditions (7) (1 + |x1 | + · · · + |xn |)m u ∈ L2 , (1 + |x|)m u ∈ L2 , hxim u ∈ L2 are all equivalent. As already noted the advantage of the last one is that the ‘multiplier’ is smooth and of slow growth so hxim u ∈ S 0 (Rn ) for any u ∈ S 0 (Rn ) makes sense, an observation which will be used below. (7) Now for the characterization of L2 -based Sobolev spaces Proposition 3. If k ∈ N and u ∈ S 0 (Rn ) then u ∈ H k (Rn ) if and only if u ∈ L2 (Rn ) and (8) hξik û ∈ L2 (Rn ). LECTURE 7, 18.155, 29 SEPTEMBER 2011 5 Proof. Suppose first that u ∈ H k (Rn ), for some fixed k ∈ N. Then by definition u ∈ L2 (Rn ) and the (weak) derivatives ∂ α u ∈ L2 (Rn ) for all |α| ≤ k. From the discussion of the Fourier tranform we conclude that α u = iξ α û ∈ L2 (Rn ). ∂d From this it follows that |ξ α |û ∈ L2 (Rn ) and hence that X ( |ξ α |)û ∈ L2 (Rn ). |α|≤k From the inequalities above this in turn implies (8). Conversely, (8) implies that ξ α û ∈ L2 (Rn ) for |α| ≤ k since then xiα hξi−k is a bounded continuous function, so a multiplier on L2 (Rn ). Reversing the argument above, ∂xα u(x) ∈ L2 (Rn ) and hence u ∈ H k (Rn ). (8) In fact the condition (8) alone on u ∈ L2 (Rn ) implies that u ∈ H k (Rn ) as follows from the argument above. However, this makes sense with k replaced by any m ∈ R, positive or negative and not necessarily integral. So we may simply define, consistently with the earlier definition, H m (Rn ) = {u ∈ S 0 (Rn ); hξim û ∈ L2 (Rn ). For m < 0 this does not imply that u ∈ L2 (Rn ) and in general the elements of negative Sobolev spaces are not functions.