TOPOLOGIES OF TEST/DISTRIBUTION SPACES Mark asked me for some references on the topologies of these spaces, I am still thinking about a good one but in the meantime here are some notes. Let me add a little more background. The conventional wisdom is that you do not really need to know about these topologies, that it is enough to know the estimates – derived below – for continuity of linear functionals (i.e. the definition of distributions) and continuity of linear maps between the various spaces. On the other hand I realize that enquiring minds want to know and understand these things, so here goes. First, we started with the space S(Rn ) which has a countable collection of norms that fix the topology – not surprisingly these are called countably-normed spaces. Having got this far we know actually that we can take Hilbert norms instead of the weighted supremum type we started with since for u ∈ S(Rn ), (1) sup |Dα [(1 + |x|2 )k/2 u]| ≤ Ck k(1 + |ξ|2 )p (1 +\ |x|2 )p/2 ukL2 |α|≤k ≤ Cp sup |(1 + |x|2 )n/2 Dα [(1 + |x|2 )p/2 u]|, p > k + n/2. |α|≤p Here I have used the Sobolev embedding theorem to bound the C k norm in terms of a Sobolev norm and then Cauchy-Schwartz to bound that in terms of a supremum norm again. Increasing the power from k to p inside the integral uses boundedness properties we know as does the fact that the last norm is bounded again by a higher version of the first norm. So, either way S(Rn ) can be written as an intersection of Hilbert or Banach spaces. The completeness is not really important – you can replace the spaces by Hilber/Banach space by S(Rn ) again just keeping the norms. Still, it means that S(Rn ) can be written as a ‘directed limit’ – a projective limit – of Hilbert spaces H (k) = {u ∈ L2 (Rn ); (1 + |ξ|)k/2 (1 +\ |x|2 )k/2 u ∈ L2 }, (2) . . . −→ H (N ) −→ H (N −1) −→ . All the maps here are injections and S(Rn ) is the intersection of the spaces but that means an element of S(Rn ) can be identified with a sequence (u0 , u1 , . . . ) such that the image of uN is always uN −1 under these injections. The topology on S(Rn ) is a metric topology where a set O ⊂ S(Rn ) is open if for each point (sequence) in O there is an N and a ball around the corresponding uN contains all sequences (points in S(Rn ) with u0N in that ball are in O. Transferring the norms to S(Rn ) as is sensible this is what we know to be the topology from the metric X kukk (3) d(u, v) = 2−k 1 + kukk k n and S(R ) is complete with respect to this metric. Now, why have I made this more complicated? To try to make things more systematic as regards the other spaces. If this is not helping, maybe stop reading! 1 2 TOPOLOGIES OF TEST/DISTRIBUTION SPACES For C ∞ (Ω) we can write things in a similar way, starting with an exhaustion by compact subsets Kj b Ω which are increasing, Kj+1 ⊃ Kj and such that for any K b Ω, there is a j such that K ⊂ Kj . In fact it is better to assume that Kj is the closure of its interior. You can arrange this by replacing one exhaustion by the closures of their interiors and you will get another. In fact it is even better to arrange that the Kj have smooth boundaries, that can be done but let’s not bother. So we can define C ∞ (Kj ) as the space of smooth functions on the interior of Kj all the derivatives of which are bounded (this is a nicer space if Kj has smooth boundary, like a closed ball). These are complete countably normed spaces rather like S(Rn ), with C k norms and so on. In fact we can replace C ∞ (Kj ) by C j (Kj ) with the fairly obvious definition. Again C ∞ (Ω) is a directed limit since we have restriction maps . . . −→ C ∞ (KN ) −→ C ∞ (KN −1 ) −→ or (4) . . . −→ C N (KN ) −→ C N −1 (KN −1 ) −→ . Again an element of C ∞ (Ω) can be identified with a sequence which is consistent under these restriction maps. The major difference is that the maps are not injective (in general, you could insist that each Kj is contained in the interior of Kj+1 and then it is always true). This means that the norms on C N (KN ) correspond to seminorms on C ∞ (Ω) but apart from that things are pretty much the same. The topology has the same definition and is a complete metric topology. These are Fréchet spaces – complete metric spaces where the metric comes from a countable number of seminorms with no common null space. You can always assume the seminorms grow with j by replacing they jth by the sum of the first j. Now, for the third space Cc∞ (Ω) ⊂ Cc∞ (Rn ) consisting of the smooth functions on n R which vanish outside a compact subset of Ω. This is a union over any compact exhaustion [ (5) Cc∞ (Ω) = Cc∞ (Kj ), Cc∞ (K) = {u ∈ Cc∞ (Rn ); supp(u) ⊂ K}. j Cc∞ (Kj ) are closed subsets of S(Rn ) so they are already countably normed Here the Fréchet spaces. Again there are maps but this time the ‘go the other way’ (6) Cc∞ (K0 ) −→ Cc∞ (K1 ) −→ . . . −→ Cc∞ (Kj ) −→ . . . and are again continuous injections (since they are bigger closed susbsets of the same space). Now we are looking at a inductive limit since the spaces are getting bigger. An element of C ∞ (Ω) can be thought of as a consistent sequence going to the right and just starting somewhere, not at the beginning. This is just the settheoretic union in this case. A set in I(Ω) is open if it meets each of the Cc∞ (Kj ) in an open set. The result is not a metric space. Sometimes these are call ‘LF’ or ‘ILF’ spaces as ‘inductive limits of Fréchet spaces’. These definitions unravel when you start to look at the dual spaces – the spaces of continuous linear maps U : F −→ C. Of course this just means that the inverse image of the unit ball is open, but since the map is linear it just means (using the linearity again to translate around) that 0 ∈ F is an interior point of U −1 (|z| < 1). For a Fréchet space this means that there is some seminorm and constant such that (7) kU (f )kk ≤ Ckf kk TOPOLOGIES OF TEST/DISTRIBUTION SPACES 3 which we already knew. It is perfectly reasonable just to take this as the definition of a tempered distribution (for F = S(Rn )) or a distribution with compact support in Ω (for F = C ∞ (Ω)) and move on. For the LF space such as Cc∞ (Ω) continuity reduces, from the definition of an open set, to continuity of U : Cc∞ (Ω) −→ C reduces to continuity of every restriction U : Cc∞ (Kj ) −→ C and this means that for each K (it is enough to take Kj so there are really only countably many conditions) there is a k and C such that (8) |U (f )| ≤ Ckf kk ∀ f with supp(f ) ⊂ K. So in some sense we always get the same condition but you have to watch the quantifiers carefully! Now, we need to know about continuity of linear maps between these spaces L : F1 −→ F2 . The result is pretty much the same but I better write it out unless someone would like to do it for me – this is a matter of examining open sets and what happens to them and again it is only an issue of deciding when 0 is an interior point of the inverse image but now of a ‘basis’ of open sets containing 0 in the image space, so that you get every open set containing 0 in the end. What about topologies on the dual spaces, what are they like? We already talked about S 0 (Rn ). From the discussion above each U ∈ S 0 (Rn ) is continous with respect to some (admissible, i.e. continuous) norm on S(R). So we can let B−k be those which satisfy (9) |U (f )| ≤ CU kf kk , kU k−k = sup |U (f )| < ∞ kf kk ≤1;f ∈S(Rn ) defines a norm on B−k with respect to which it is a Banach space. In our case if we take the weighted Sobolev norms above, these are Hilbert spaces and you can see that they are (negatively weighted/order) Sobolev spaces. Then [ (10) S 0 (Rn ) = B−k k is written as an inductive limit of Banach (in this case Hilbert) spaces. So it is a simpler case than the LF spaces above – in fact not significantly simpler. So this gives S 0 (Rn ) a topology with respect to which it is complete. There is however more to the story than that, even for S 0 (Rn ), since there are other topologies. For instance there is the weak topology. Each of the elements f ∈ S(Rn ) fixes a seminorm on S 0 (Rn ) by pairing (11) S 0 (Rn ) 3 U 7−→ |U (f )|. So, now we have a space with uncountably many seminorms, and none is really stronger than the others so there is no hope of dropping down to a countable subcollection. What to do? Well, they define a topology. Declare the sets formed by a finite intersection of semi-norm balls with centres at Uj to be open (12) {V ∈ S 0 (Rn ); |(V − Uj )(fj )| < j , f1 , . . . , fN ∈ S(Rn ), j > 0} to be open. Any finite intersection of these is then also open since it is of the same form. It follows that arbitrary unions of these sets form a topology – since the finite intersection of such an arbitray union is a union of finite intersections (check basis of topology ...). 4 TOPOLOGIES OF TEST/DISTRIBUTION SPACES Exercise 1. Show that the distributional pairing (13) S(Rn ) × S 0 (Rn ) 3 (f, U ) −→ U (f ) ∈ C is continuous with respect to either the product of the Fréchet and weak topologies or the products of the Fréchet and inductive limit topologies and that it in both cases it is a ‘perfect pairing’ in that the restricted to either factor it gives all continous linear maps on the other. Meaning of course that the continous linear functionals on S 0 (Rn ) with respect to the weak topology or the inductive limit topology are precisely the elements of S(Rn ).