A Walking Tour of Microlocal Analysis

Je Schonert August 10, 2006

Abstract We summarize some of the basic principles of microlocal analysis and their applications. After reviewing distributions, we then dene pseudodierential operators, their symbols, and the pseudolocal property. This then leads to the fundamental notion of microlocal analysis: the wave front set of a distribution. The wave front set will then be used to analyze the problem of the propagation of singularities.

Introduction

The techniques of microlocal analysis were developed in the 1960s and 70s as part of the study of linear partial dierential equations. Many of the ideas are due to the work of Hormander, Kohn and Nirenberg, and Maslov, in which they generalized existing notions from analysis to investigate distributions and their singularities. Indeed, much of microlocal analysis is concerned with shifting the study of a distributions singularities from the base space to the cotangent bundle. Such a study will allow us to answer basic questions such as when the product of two distributions is well-dened, as well as extending and strengthening some standard theorems from dierential equations, such as elliptic regularity.

Review of Distributions

In elementary calculus, one is immediately confronted with functions that are not dierentiable. The introduction of distributions seeks to remedy this shortcoming by 1

providing the smallest set of objects such that every member is innitely dierentiable in a sense to be made precise later (hence the term generalized function). In this section, we will give a precise denition and topology of this space of functions. Fix an open set Rn , and consider the space of smooth functions with compact support Cc (). We would like to give this space a suitable topology under which it will be complete. This can be accomplished by introducing a family of seminorms that range over all compact K : PK,m = where sup
||m,xK

|D f (x)|,

(1)

|| ; l1 = ||, and supp(f ) K. 1 2 k Because this family of seminorms {P j } separates points in that P j (x) for all j implies x = 0, they turn Cc () into a locally convex space, such that the natural metrizable topology is Hausdor. It is well-known that Cc (Rn ) is not complete under this topology, but it can be made so by equipping it with the strict inductive limit topology1 . To get the largest space of such functions contained in CC (), we take the union of all sets of smooth functions whose support is contained in a compact subset of . We dene this space to be D() = K Cc (K), where K is compact. We have now made the space of test functions D() into a Frchet space, e which gives us an acceptable notion of what it means for a sequence to converge in this space, i.e. a notion of continuity. More precisely, a sequence {j } D() converges to D() if supp(j ) is contained in some compact K for all j, and Pm,k (j ) 0. A distribution is then dened to be a continuous linear functional on this space of test functions. The space of all distributions is denoted D (). In other words, a linear functional T on D() is a distribution if and only for every compact K , there is a constant C > 0 and n N such that D = T () C sup
||n,xK

|D (x)|, for D().

(2)

Even though we have dened distributions on open subsets of Rn , they can easily be extended to a smooth manifold M . The construction proceeds in the obvious way, by dening distributions on the homeomorphic images of coordinate patches. More specically, if (Uj , j ) is an arbitrary chart, then consider the distribution uj
1

See p. 146 of [4] for a discussion of strict inductive limits

D (j (Uj )). A distribution on M is given by all such local representatives that satisfy uj (j 1 ) = ui on i (Ui Uj ). It is clear that this denition coincides with our i previous denition in the case that M is an open subset of Rn . Furthermore, it can be shown that this denition is independent of the charts we choose. Finally, we dene the notion of a weak derivative, a construct that allows every distribution to be innitely dierentiable. The basic idea in this denition is that since the test functions are smooth and vanish outside a suitably large space, then we can integrate by parts and discard the boundary terms. More formally, the weak derivative of T D () is dened as the distribution D T that satises (D T )(x)(x)dx = (1)|| T (x)(D )(x)dx D(). (3)

Pseudodierential Operators and Symbols

In this section we dene some of the basic properties of pseudodierential operators and the symbols associated with them. To motivate the denition of symbols, con sider the following result from ordinary real analysis. Let p(x, D) = ||k a (x)Dx be a dierential operator, where the coecients a (x) are smooth functions, Dx = 1 n Dx1 . . . Dxn , and Dxj = ixj . If f is a function of rapid decrease, then the Fourier inversion formula is valid, and we can write f as an integral of its f . Now apply p(x, D) to f to get: p(x, D)f (x) = 1 (2)n p(x, )eix f ()d. (4)

We will see that (4) eventually will dene a pseudodierential operator as a map from Cc (X) to C . To make this denition more precise, however, we rst need to specify what p(x, ) is. Since we have applied a dierential operator that is a polynomial in derivatives, (4) suggests that p(x, ) (called the characteristic polynomial of the dierential operator) is not just an arbitrary function, but a polynomial in some vector. Since p(x, D) takes derivatives with respect to x, every time we apply p(x, D) to f (x), we pick up more powers of ; note that there are no factors of i present due to our denition of Dx . In particular, p(x, ) can be replaced by ||k a (x) , which is now a type of function instead of a dierential operator. We would like this p(x, ) to satisfy a certain bound such that every time we dierentiate it, we lose a degree of smoothness. This leads to the denition of symbols, a class of functions that plays an important role in microlocal analysis. 3

Denition 3.1 Let Rq be open, , [0, 1], m R, and n a positive integer. m Then the space of symbols of order m and type (, ), denoted S, ( Rn ), is the n n set of all u C ( R ) such that for all compact K R and all Nq and Nn , there is a constant C = C(K, , ) such that
|x u(x, )| C(1 + ||)m||+|| .

(5)

m We denote the intersection of S, for all real m by S, . m Denition 3.2 Let p(x, ) S, . Then a pseudodierential operator A is a function from Cc (X) to D (X) dened by (4). The space of all such operators is denoted Lm . , m e S, ( Rn ) can be made into a Frchet space by introducing the seminorms: |x y u(x, y)| . n (1 + |y|)m||+|| (x,y)KR

PK,, (u) =

sup

According to our denition, pseudodierential operators are only functions on Cc (X); we do not know if they can be dened on more general spaces, e.g. the space of distributions. In order to determine when such an extension is allowed, we must consider the kernel of the operator. In the above expression for pseudodierential operators, expand the function f () in terms of f () using the denition of the transform. This suggests that the kernel of the pseudodierential operator A, denoted KA , is given by an oscillatory integral of the form 1 (2)n p(x, )ei(xy) ,

where p(x, ) is a symbol of order m. This is indeed the case, and it can be made more rigorous using the Schwartz kernel theorem2 , which establishes a one-one corre spondence between distributions on X Y and linear maps Cc (Y ) D (X). The correspondence that denes the kernel can be written as < Au, v >=< KA , u(y)v(x) >= 1 (2)n u(y)p(x, )v(x)ei(yx) dydxd. (6)

This leads to the following important lemma:

2

For a statement and proof of the Schwartz kernel theorem, see 5.2 of [3].

Lemma 3.3 If > 0, then KA is a smooth function o the diagonal in Rn Rn . Proof : For an arbitrary > 0, we can integrate by parts -times the integral for KA and discard the boundary terms to get (y x) KA =
ei(yx) D p(x, )d.

(7)

This integral converges when is chosen large enough so that m + n || < 0. If we dierentiate the above equation q-times, where m + q + n || < 0, then the integral also converges, implying that (y x) KA C q (Rn Rn ).2 This lemma implies that if KA C (X Y ), then A can be extended to a continuous map from Dc (Y ) to C (X) (Dc denotes the space of compactly supported distributions). Before stating an important corollary of this lemma, we need some denitions that will allow us to characterize singularities of a distribution: Denition 3.4 If u D (X), then the singular support of u, denoted sing supp u, is the smallest closed subset of X on which u is not C . Denition 3.5 A Lm () is properly supported if supp KA = C X Y is , proper, i.e. if the projections x : C X and y : C Y are proper maps. It is clear that if A Lm () is properly supported, then it is a continuous map , from D () to D (). Corollary 3.6 Let A Lm and u be a compactly supported distribution on . If , > 0, then sing supp Au sing supp u. Proof : Let A Lm () and u Dc (). Consider some open set U in sing supp u , and a test function =1 on sing supp u. We can localize u by multiplying it by the test function, which gives two new distributions u1 = u and u2 = (1 )u so that u = u1 + u2 . Let KA be the kernel associated with A. By the lemma above, KA (x, y) is C o the diagonal, i.e. when x V and y V . Since Au is just KA smeared / with u, we have Au1 = KA (x, y)u1 (y)dy. (8)
supp u1

We know that the kernel is smooth, so apply the operator D to both sides to get D Au1 =
supp u1

D KA (x, y)u1 (y)dy. 5

(9)

This implies that D A is smooth o the diagonal as well, hence it is smooth outside of V . Since u2 is Cc on , it follows that Au2 C (). Because pseudodierential operators are linear maps, Au1 +Au2 = A(u1 +u2 ) = Au, implying that Au is smooth outside of V . But since we chose V to be an arbitary open subset of sing supp u, this implies that Au is smooth outside of sing supp u.2 This corollary is often called the pseudolocal property of operators, and it will play an important role in the discussion of wave front sets below. Now we give a generalized notion of an elliptic dierential operator in the context of pseudodierential operators. Recall that if p(x, D) is a dierential operator with smooth coecients p(x, D) = a (x)Dx , (10)
||m

then the principal part of p(D) is dened to be Pm (x, D) =

||=m m Similarly, if p(x, ) S, () as above, then we dene the principal symbol to be the symbol a (x) . (12) pm (x, ) = ||=m a (x)Dx .

(11)

The operator p(x, D) is said to be elliptic if pm (x, ) = 0 for all 0 = Rn . Pseudodierential operators generalize this idea by extending the same notion to the associated symbol. An operator p(x, D) Lm , is said to be elliptic at the point , (x0 , 0 ) if there is some conical neighborhood V of (x0 , 0 ) and a positive constant C such that |p(x, )| C(1 + ||)m (13) for all (x, ) V and || C. If we restrict our attention to elliptic operators, we can make the estimate of the pseudolocal property more precise by showing that the singular supports are equal. Before proving this theorem, we rst need a denition: Denition 3.7 If P is a pseudodierential operator of order m, then a parametrix for P is a properly supported pseudodierential operator Q L such that P QI , L and QP I L . It is true that if P is elliptic, then P has a parametrix , , Q Lm (See [1] p. 298 for details). , 6

Theorem 3.8 If P Lm () is elliptic and u is a distribution on , then sing supp , (P u)=sing supp (u). Proof : We already know that sing supp (P u) sing supp (u), which amounts to saying that if u D () is smooth on some open V , then P u is smooth on V . To show the converse inclusion, we need to show that if P u C (V ), then u|V is smooth. To show this, let Q be a properly supported parametrix for P . Then the restriction of its kernel KQ to V V is also properly supported. Because P u|V C (V ), it follows that QP U |V C (V ) and KQ is smoothing (i.e. KQ u C ). As a result, u|V C (V ). So we have shown that sing supp (u) sing supp (P u) for the case of an elliptic operator. Combining this with Corollary 3.6 shows that equality must hold. 2 Note that a key result used in the proof was that every elliptic operator has a parametrix and has a kernel that is smooth o the diagonal.

Microlocal Analysis

It is well known that the decay properties of the Fourier transform of a distribution are related to its smoothness. More specically, if u D (), then we can localize u by multiplying it by a test function and considering u. By a basic theorem of Fourier analysis and the Paley-Wiener theorem, u is an entire holomorphic function, and it is smooth if it is of rapid decrease. The idea of the wave front set is to characterize both the points and the directions in which this condition fails to hold. Even though this construct is most easily motivated by considering the Fourier transforms of localized distributions, we nd it more satisfying to use our above ideas of pseudodierential operators. Instead, our denition of the smoothness of u will be the following: (In this section, all pseudodierential operators are assumed to be properly supported) Denition 4.1 If u D () and (x, ) T , then u is smooth on a neighborhood of (x, ) if there is some A Lm () that is elliptic at (x, ), and such that Au C (). , This suggests that the set of points where u is not smooth is related to the set of points where some pseudodierential operator A is not elliptic. Indeed, this is the case. We now dene the characteristic variety of A Lm () to be , Char A = {(x, ) T |A is not elliptic at (x, )}. 7

From here, it is easy to dene the wave front set of a distribution, a concept that plays a central role in microlocal analysis. Denition 4.2 If u D (), then the wave front set of u, denoted WF(u), is given by {Char A A Lm , and Au C }. 1,0 It is immediate that the wave front set is a renement of the notion of singular support considered above, as it also describes singular directions. Moreover, the wave front set also has the advantage of being a closed conic subset of the cotangent bundle. Because of this, it is clear that we can recover the singular support of a distribution by merely projecting the wave front set onto its rst factor: Lemma 4.3 Let : T X X be the projection map dened by (x, ) = x for all (x, ) T X. Then (WF(u)) =sing supp u. Using the following denition, we also have a microlocal analogue of the pseudolocal property from above; Denition 4.4 For P Lm , dene the microsupport of P , denoted supp(P ), to , be the complement of the set on which P is smoothing. Theorem 4.5 If P Lm (), u D (), then WF(P u) WF(u)supp(P ). , Proof :We will only sketch the steps here; see e.g. [1] for a complete treatment. If (x0 , 0 ) supp(P ), then there is a Q Lm such that Q is elliptic at (x0 , 0 ) and / 1,0 the microsupports of Q and P are disjoint. Then QP L and (x0 , 0 ) WF(P u). / , m Now assume that (x0 , 0 ) WF(u), which implies there is a Q L1,0 such that / Qu C and Q is elliptic on a cone around 0 . The theorem will follow once we / prove the following claim: there are R, S Lm such that (x0 , 0 ) Char(R) and 1,0 RP SQ L, . To see that it is sucient to prove this claim, note that since Q, SQ, and RP SQ are smoothing, this implies RP is smoothing. It then follows that (x0 , 0 ) WF(P u).2 / We can extend the result on elliptic operators to wave front sets by simply replacing the singular support with the wave front set. We then have the following theorem: Theorem 4.6 If P Lm () is elliptic, then WF(P u)=WF(u) u D (). ,

Proof : If Q is a parametrix for P , then by denition QP u u C . Since WF(QP u u) = , it is clear that WF(QP u)=WF(u). But by theorem 3.4, we know that WF(P u) WF(u) and WF(u)=WF(QP u) WF(P u). Hence, equality must hold. 2 In order to impose a suitable topology on the set of distributions whose wave front set is bounded above, it is necessary to introduce a new collection of seminorms. If T X is a closed cone, then denote the space of all distributions whose wave front set lies in by D (X). We say that a sequence of distributions {uj } D (X) converges to u D (X) if uj u weakly and PN,,U (uj u) 0, where P is a seminorm dened by P (u) = sup |u|(1 + ||)N (14)
U

for some and closed cone U Rn . The nal property of wave front sets that we will give is perhaps the most useful one: dening the product of distributions. The tensor product of two distributions u1 and u2 is dened whenever they depend on dierent sets of variables, but the pointwise product is often ill-dened. We will see that whenever the two wave front sets have a particular form, the product is unambiguously dened. Let f : X Y be a smooth map, where X and Y are open subsets of Rn . The pull-back of f , denoted f , is dened in the usual way by f ()(x) = (f (x)).
Cc (X)

Denition 4.7 If u, v D (), then the product uv is dened to be the pull-back of the tensor product by the diagonal map : that sends x to (x, x). We can now formulate a precise condition for when the product of two distributions is well-dened: Theorem 4.8 Let u1 , u2 D (X), with respective wave front sets WF(u1 ), WF(u2 ). Then u1 u2 is dened whenever the composite wave front set WF(u1 ) WF(u2 ) = {(x, 1 + 2 ) |(x, 1 ) WF(u1 ); (x, 2 ) WF(u2 )} does not contain an element of the form (x, 0). We conclude this section with examples of this theorem([5]): i) If f : U V is C , then WF(f ) = , and so, as we expect, the product of smooth functions is well-dened.

ii) Let x be the usual delta distribution, i.e. x f = f (x) for f D(X). It is easy to see that WF()={(0, ) | = 0}. Now we want to consider two delta distributions (x1 ) and (x2 ). We know that WF((x1 )) = {(0, x2 ; , 0) |x2 R, = 0} WF((x2 )) = {(x1 , 0; 0, ) |x1 R, = 0}. Since we know that = 0, it is clear that (x1 )(x2 ) exists and is given by (x1 , x2 ) that acts on f (x1 , x2 ) by f (x1 , x2 )(x1 , x2 )dx1 dx2 = f (0, 0). (15)

1 iii) Let P( x ) be the Cauchy principle part integral given by

P(1/x) : f lim
0 |x|

f (x) dx. x

(16)

1 Then for the operator P( x ) i(x), we can rewrite it as

P(1/x) i(x) = lim

0

1 . x+i

(17)

Its wave front set is given by {(0, ) | > 0}.

1 By the above theorem, the product of P( x ) with itself clearly exists.

Propagation of Singularities

Let (M, ) be a symplectic manifold, and H : M R a C r map. We dene the Hamiltonian vector eld on M generated by H to be the vector eld XH determined by (XH , ) = dH. (18)

10

Note that the nondegeneracy of guarantees the existence of XH . Due to a theorem of Darboux ([2] p.98), we can always choose local canonical coordinates (q1 , . . . , qn , p1 , . . . , pn ) so that n =
i=1

dqi dpi .

(19)

It turns out that the principal symbol of a pseudodierential operator A can also dene Hamiltonian vector eld in a similar way (A must actually be a special type of operator; see p. 35 of [GS] for details). In the above canonical coordinates, the Hamiltonian eld generated by the principal symbol P is
n

XP =
i

P P . pi qi qi pi

(20)

An integral curve3 : [a, b] M is called a bicharacteristic strip of XH if H vanishes along , i.e. if is an integral curve of XH through H 1 (0). We now have assembled the necessary geometrical tools to state the propagation of singularities theorem. The motivation for this result is that since the wave front set is the set of points where P Lm is not elliptic, we have that WF(u) P 1 (0) = {(x, ) 1,0 T X |P (x, ) = 0} . Our goal is to formulate conditions on certain subsets of P 1 (0) so that they are of the form WF(u) for some u. This idea may be stated precisely as follows: Theorem 5.1 Let A Lm and let Pm be its associated real, positive principal symbol , of degree m. Furthermore, let : [a, b] T X be a bicharacteristic strip of XPm . If u D (X) satises ([a, b]) WF(Au) = , then either ([a, b]) WF(u) or ([a, b]) WF(u) = . The proof of this theorem is rather involved, and the reader is referred to 8 of [2] for details. The importance of this theorem is that if A(x, D)u = f and (x0 , 0 ) is a point in p1 (0), then the entire bicharacteristic strip beginning at (x0 , 0 ) is A contained in WF(u). So the singularities of u are literally owed along the vector eld generated by pA .
Recall that an integral curve of a vector eld X is a C k curve such that (t) = X((t)) for all t in the domain of .
3

11

References
[1] G. Folland, Introduction to Partial Dierential Equations, 2nd edition (Princeton University Press, Princeton, 1995). [2] A. Grigis and J. Sjostrand, Microlocal Analysis for Dierential Operators (Cambridge University Press, Cambridge, 1994). [3] L. Hormander, The Analysis of Linear Partial Dierential Operators I (SpringerVerlag, Berlin, 1983). [4] M. Reed and B. Simon, Methods of Modern Mathematical Physics, Volume I: Functional Analysis (Academic Press, San Diego, 1980). [5] M. Reed and B. Simon, Methods of Modern Mathematical Physics, Volume II: Fourier Analysis and Self-Adjointness(Academic Press, San Diego, 1974). [6] M. Taylor, Pseudodierential Operators (Princeton University Press, Princeton, 1981).

12