Special Topic: Elliptic Operators

Introduction

We discuss how to use Fourier series to prove the regularity of Elliptic operators, following this excellent note https://math.mit.edu/~vwg/classnotes-spring05.pdf

We use a different normalization convention than other notes for Fourier transform. Recall that by derivative theorem we have

F(αf)(ξ)=(2πi)αξαFf(ξ).\mathcal F(\partial_\alpha f )(\xi) = (2\pi i)^{|\alpha|} \xi^\alpha \mathcal Ff(\xi).

We adopt two normalization conventions to simplify the notations:

  • We define the Fourier transform to be Ff(ξ)=Rnf(x)eixξdm(x)\mathcal Ff(\xi) = \int_{\mathbb{R}^n} f(x)e^{-ix\xi} dm(x), here we write xξx\xi instead of x,ξ\langle x,\xi\rangle for simplicity. Consequently, let TT be a 2π2\pi-

High-dimensional Fourier series

In our study of Fourier series we can summerize the theory development as follows:

  • For every smooth periodic function ff on R\mathbb{R} with period 2π2\pi, the Fourier transform Ff=kZf^(k)δk\mathcal Ff = \sum_{k\in \mathbb{Z}} \hat{f}(k)\delta_k, where f^(k)=12πππf(x)eikx\hat{f}(k) = \frac{1}{2\pi}\int_{-\pi}^{\pi}f(x)e^{-ikx}.
  • The sequence f^(k)\hat f(k) satisfies the rapidly decreasing condition: For every rr, there exists constant CrC_r such that f^(k)Crkr|\hat{f}(k)|\leq C_r |k|^{-r}. We also call such a sequence smooth.
  • The inversion theorem: f(x)=F1Ff(x)=F1kZf^(k)δk=kZf^(k)F1δk=kZf^(k)eikxf(x) = \mathcal F^{-1}\mathcal Ff(x) = \mathcal F^{-1}\sum_{k\in \mathbb{Z}}\hat f(k)\delta_k = \sum_{k\in \mathbb{Z}}\hat f(k)\mathcal F^{-1}\delta_k = \sum_{k\in \mathbb{Z}}\hat f(k)e^{ikx}.

The theory generalizes to nn-dimensional case as well. We restrict our discussion to smooth case.

Differential operators

A differential operator of order rr is an opeartor P:C(T)C(T)P:\mathscr{C}^{\infty}(T) \to \mathscr{C}^{\infty}(T) of the form Pu=αraα(x)DαuPu = \sum_{|\alpha|\leq r} a_\alpha (x) D^{\alpha}u.

Main Theorem

Let PP be an elliptic operator on TnT^n. Then PP is right-invertible module smoothing operator, PQ=I+smoothingPQ=I+\text{smoothing}. Precisely, there exists an operator Q:C(X)C(X)Q:\mathscr{C}^{\infty}(X)\to \mathscr{C}^{\infty}(X) and a smoothing operator TKT_K such that PQ=ITKPQ = I-T_K.

We know that after Fourier transform, a differential operator becomes an algebraic object. To construct such an operator, we take Fourier transform, do some algebra on the “frequency domain” and construct a corresponding algebraic object of QQ. Then we are done by inversion. The algebraic object corresponding to differential operators in the frequency domain are symbols.

Some Calculus of Symbols

Motivation

Lemma
Dαeikx=kαeikx.D^{\alpha}e^{ikx} = k^\alpha e^{ikx}.

\square

Let P=αmaα(x)DαP = \sum_{|\alpha|\leq m} a_\alpha(x)D^{\alpha} be a differential operator. Let ff be a smooth function on TnT^n, then we identify ff with F1Ff=kZnf^(k)eikx\mathcal F^{-1}\mathcal F f = \sum_{k\in \mathbb{Z}^n}\hat f(k) e^{-ikx}. Then

Pf=αmkZnf^(k)aα(x)Dαeikx=kZ(αmaα(x)kα)f^(k)eikx.Pf = \sum_{|\alpha|\leq m}\sum_{k\in \mathbb{Z}^n}\hat{f}(k)a_\alpha(x)D^{\alpha}e^{ikx} = \sum_{k\in \mathbb{Z}}(\sum_{|\alpha|\leq m}a_\alpha(x)k^{\alpha})\hat{f}(k)e^{ikx}.

Let p(x,ξ)=αmaα(x)ξαp(x,\xi) = \sum_{|\alpha|\leq m}a_\alpha(x)\xi^{\alpha} be the polynomial in RHS, we can write Pf=kZna(x,k)f^(k)eikxPf = \sum_{k\in \mathbb{Z}^n}a(x,k)\hat{f}(k)e^{ikx}. Therefore, via Fourier transform, we can identify PP with its polynomial p(x,ξ)p(x,\xi).

Conversely, with a polynomial αmaα(x)ξα\sum_{|\alpha|\leq m}a_\alpha(x)\xi^{\alpha} we can define a differential operator by replacing ξα\xi^\alpha by DαD^\alpha.

Observe that for a general function a(x,ξ)a(x,\xi), not necessarily a polynomial, we can still define an operation Taf=kZna(x,k)f^(k)eikxT_a f = \sum_{k\in \mathbb{Z}^n}a(x,k)\hat{f}(k)e^{ikx} for fC(Tn)f\in \mathscr{C}^{\infty}(T^n), and TafT_a f will be C(T)\mathscr{C}^{\infty}(T) when the ξ part satisfy sufficient decreasing condition. Such a function a(x,ξ)a(x,\xi) is called a symbol, and the corresponding TaT_a is called a pseudo differential operator.

Symbols and pseudodifferential operators

Definition

A function a:Tn×RnCa:T^n \times \mathbb R^n\to \mathbb{C} is a symbol of degree mm, which we denote by SmS^m, provided that for all multiindices α,β\alpha,\beta, DxαDξβa(x,ξ)Cα,βξmβ|D_x^\alpha D_\xi^\beta a(x,\xi)| \leq C_{\alpha,\beta}\langle\xi \rangle^{m - \beta|}.

  • We use ξ:=1+ξ2\langle \xi \rangle:=\sqrt{1+|\xi|^2} instead of ξ|\xi| to measure the growth condition to avoid the blowing up issue of ξnegative|\xi|^{\text{negative}}.
Example
  • p(x,ξ)=αmaα(x)ξαp(x,\xi) = \sum_{|\alpha|\leq m} a_\alpha(x)\xi^\alpha is in SmS^m.
  • ((1+ξ2))m(\sqrt(1+|\xi|^2))^m is in SmS^m, which is not a polynomial.

Note that mm can take nevative value. Acutally this is the main interesting case. If aSma\in S^{m} for positive mm, then DξαaS0D^\alpha_\xi a \in S^{0} for α|\alpha| big enough. Also, (Sm)mZ(S^m)_{m\in \mathbb{Z}} is a decreasing family. SmSm1....S^m\supset S^{m-1}\supset ..... Let SS^{-\infty} be the decreasing limit mZSm\cap_{m\in \mathbb{Z}}S^{m}.

\square

Lemma
  • aSl,bSm    abSm+la\in S^l, b\in S^m \implies ab\in S^{m+l}.
  • aSm    DxαDξβaSmβa\in S^m \implies D_x^{\alpha}D_{\xi}^\beta a \in S^{m - |\beta|}. Taking derivatives decrease growth condition. Note that as stated, taking derivative on xx is irrevalent.

\square

Definition

A pseudodifferential operator is an operator T:C(Tn)C(Tn)T:\mathscr{C}^{\infty}(T^n)\to \mathscr{C}^{\infty}(T^n) given by some aSma\in S^m for some mm, so that Tf=kZf^(k)a(x,k)eikxTf = \sum_{k\in \mathbb{Z}}\hat{f}(k)a(x,k)e^{ikx}. We call aa the symbol of TT. We usually let TaT_a to denote a pseudodifferential operator with symbol aa specified.

\square

It follows that a differential operator is a pseudodifferential operator with positive degree. But how about negative degree? Consider the example a(x,k)=k2a(x,k) = k^{-2}, so Taf^(k)=k2f^(k)\widehat{T_af}(k) = k^{-2}\hat{f}(k), so f^(k)=D2Taf^(k)\hat f(k) = \widehat{D^2T_af}(k) by derivative theorem. This indicates that if aa is negative degree, then TaT_a will make the function more regular. In particular, if aSa\in S^{-\infty}, then TaT_a will be a “smoothing operator”. We’ll make these intuitions precise later. Now let’s study a first algebraic property of pseudodifferential operators.

Composition of pseudodifferential operators

The idea to construct the right inverse QQ is to find out the symbol of QQ. We already know the symbol of PP is p(x,ξ)p(x,\xi), so the first question is to understand what is PQPQ in the symbol world.

Lemma

Let PP be a differential operator, TaT_a be a pseudodifferential operator for aSra\in S^r. Then PTaPT_a is a pseudodifferential operator with symbol

p(x,D+ξ)a(x,ξ)=βm1β!ξβp(x,ξ)Dxβa(x,ξ).(cmp)p(x,D+\xi)a(x,\xi) = \sum_{|\beta|\leq m}\frac{1}{\beta !}\partial_\xi^\beta p(x,\xi)D_x^{\beta}a(x,\xi). \tag{cmp}

To find the symbol of PTaPT_a we need to find its action on the Fourier coefficients of a function. For fC(Tn)f\in \mathscr{C}^{\infty}(T^n), we have

PTaf=P(kZna(x,k)f^(k)eikx)=kZnf^(k)P(a(x,k)eikx)=kZnf^(k)eikx[eikxP(a(x,k)eikx)]\begin{split} PT_af &= P(\sum_{k\in \mathbb{Z}^n}a(x,k)\hat{f}(k)e^{ikx}) \\ & = \sum_{k\in \mathbb{Z}^n}\hat{f}(k)P(a(x,k)e^{ikx}) \\ &= \sum_{k\in \mathbb{Z}^n}\hat{f}(k)e^{ikx}[e^{-ikx}P(a(x,k)e^{ikx})] \end{split}

It follows that by definition, the symbol of PTaPT_a is eiξxP[a(x,ξ)eiξx]e^{-i\xi x}\cdot P[a(x,\xi)e^{-i\xi x}]. It sufficies to show that the latter is given by equation (cmp)(cmp) above.

We make a more general computation which will be used later. Consider P(eitfu)P(e^{itf}u) for smooth ff and uu. First we look at DjD_j, we define δj\delta_j such that Dj(eitfu)=eitfδjuD_j(e^{itf}u) = e^{itf}\delta_j u. This allows us to simply switch the eitfe^{itf} out to avoid extra term. It’s straightforward to find that δju=Dju+tfxju\delta_j u = D_j u + t \frac{\partial f}{\partial x_j} \cdot u, and we just write δj=Dj+tfxj\delta_j = D_j + t \frac{\partial f}{\partial x_j}. Similarly, DjDr(eitfu)=Dj(eitfδru)=eitfδjδruD_jD_r(e^{itf}u) = D_j(e^{itf}\delta_r u) = e^{itf}\delta_j\delta_r u. It follows that

Dα(eitfu)=D1α1Dnαn(eitfu)=eitfδ1α1δnαnu=eitf(D1+tfx1)α1(Dn+tfxn)αnu. \begin{split} D^\alpha(e^{itf}u) &= D_1^{\alpha_1}\cdots D_n^{\alpha_n}(e^{itf} u) \\ & = e^{itf}\delta_1^{\alpha_1}\cdots \delta_n^{\alpha_n}u \\ & = e^{itf} (D_1+t \frac{\partial f}{\partial x_1})^{\alpha_1} \cdots (D_n+t\frac{\partial f}{\partial x_n})^{\alpha_n}u. \end{split}

Now we apply the formula to the case f(x)=ξx=ξ1x1++ξnxnf(x) = \xi x = \xi_1x_1+\cdots+\xi_n x_n and t=1t = 1 we get

eiξxP(a(x,ξ)eiξx)=αmaα(x)(D1+ξ1)α1(Dn+ξn)αn(a(x,ξ))=αmaα(x)(D+ξ)αa(x,ξ)=p(x,D+ξ)a(x,ξ). \begin{split} e^{-i\xi x}P(a(x,\xi)e^{i\xi x}) &= \sum_{|\alpha|\leq m }a_\alpha(x)(D_1+\xi_1)^{\alpha_1}\cdots(D_n+\xi_n)^{\alpha_n}(a(x,\xi)) \\ &= \sum_{|\alpha|\leq m}a_\alpha(x)(D+\xi)^{\alpha}a(x,\xi)\\ &=p(x,D+\xi)a(x,\xi). \end{split}

Recall that pp is a polynomial function of ξ, so (totally algebraic) global Taylor expansion givecs p(x,η+ξ)=βm1β1!βn!ξβp(x,ξ)ηβp(x,\eta+\xi) = \sum_{|\beta|\leq m} \frac{1}{\beta_1!\cdots \beta_n!} \partial_\xi^{\beta}p(x,\xi) \eta^{\beta} (as polynomial), so note that the DD is taking on xx variable,

p(x,D+ξ)a(x,ξ)=βm1β!ξβp(x,ξ)Dxβa(x,ξ).p(x,D+\xi)a(x,\xi) = \sum_{|\beta|\leq m} \frac{1}{\beta!}\partial_\xi^{\beta} p(x,\xi)D^{\beta}_x a(x,\xi).

This proves the symbol of PTaP\circ T_a is βm1β!ξβp(x,ξ)Dxβa(x,ξ).\sum_{|\beta|\leq m} \frac{1}{\beta!}\partial_\xi^{\beta} p(x,\xi)D^{\beta}_x a(x,\xi).

Regularity of pseudodifferential operator

Let’s think about pseudodifferential operators again. The action of TaT_a is given by “modifying Fourier coefficients” by multiplication, we know that such operation is essentially convolution, by convolution theorem. In particular, if aa doesn’t depend on xx, then TaT_a is precisely convolution with the inverse Fourier transform of (a(k))kZ(a(k))_{k\in \mathbb{Z}}, in other words, Taf=Tn[kZa(k)eik(xy)]f(y)dμ(y)T_a f = \int_{T^n}[\sum_{k\in \mathbb{Z}}a(k)e^{-ik(x-y)}]f(y)d\mu(y).

Lemma

The series kZnkm\sum_{k\in \mathbb{Z}^n}\langle k \rangle^{-m} converges if m>nm>n (note that >> is required not ).

Proof. By comparing with integrals it sufficies to show that Rnξmdm(ξ)\int_{\mathbb R^n} \langle \xi \rangle^{-m}dm(\xi) converges. Use polar coordinate and change of variable formula we get dm(ξ)=rn1sin(φ1)sin(φn2)dm(\xi) = r^{n-1}\sin(\varphi_1)\cdots \sin(\varphi_{n-2}) (see Lecture 20 for the notations). Then

Rnξmdm(ξ)(2π)n10+rn1rmdr\int_{\mathbb R^n}\langle \xi \rangle^{-m}dm(\xi) \leq (2\pi)^{n-1}\int_{0}^{+\infty}r^{n-1}\langle r\rangle^{-m}dr

And the RHS is integrable on R\mathbb{R} when n1m<1n-1-m<-1 so m>nm>n.

\square

Lemma

For aSla\in S^l, let Ka(x,y)=kZna(x,k)eik(xy)K_a(x,y) = \sum_{k\in \mathbb{Z}^n}a(x,k)e^{ik(x-y)}.

  • If l<nl<-n then Taf=TnKa(x,y)f(y)dμ(y)T_a f = \int_{T^n}K_a(x,y)f(y)d\mu(y).
  • If l<(m+n)l<-(m+n) then Ka(x,y)Cm(Tn×Tn)K_a(x,y)\in \mathscr{C}^{m}(T^n\times T^n).

Proof.

DxαDyβ[a(x,k)eik(xy)]=Dxα[a(x,k)eikx]Dyβeiky=eikx(D+k)αa(x,k)kβeikyCkl+α+β.\begin{split} |D_x^{\alpha}D_y^{\beta}[a(x,k)e^{ik(x-y)}]| &= |D_x^{\alpha}[a(x,k)e^{ikx}]D_y^{\beta}e^{-iky}|\\ &= |e^{ikx}(D+k)^{\alpha}a(x,k) \cdot k^{\beta} e^{-iky}|\\ &\leq C \langle k\rangle ^{l+|\alpha|+|\beta|}. \end{split}

It follows that when α+βm    l+α+β<n|\alpha|+|\beta|\leq m \implies l+|\alpha|+|\beta|<-n, so DxαDyβKa(x,y)D_x^{\alpha}D_y^{\beta}K_a(x,y) converges absolutely. This implies DxαDyβKa(x,y)D_x^{\alpha}D_y^{\beta}K_a(x,y) is continuous on Tn×TnT^n\times T^n for every α+βm|\alpha|+|\beta|\leq m, i.e. Ka(x,y)K_a(x,y) is in Cm(Tn×Tn)\mathscr{C}^{m}(T^n\times T^n).

Smoothing operator

If aSa\in S^{-\infty} then TaT_a is a smoothing operator.

Construction of the inverse: bootstrap

We want to prove that for elleptic differential operator PP of order mm, there exists a pseudodifferential operator TaT_a which is almost right inverse of PP (i.e. PTa=Ismoothing PT_a = I-\text{smoothing }). Since PP is positive order, then to be inverse, TaT_a should be SmS^{-m}.

Let pm(x,ξ)p_m(x,\xi) be the top homogeneous term of pp, pm(x,ξ)=α=maα(x)ξαp_m(x,\xi) = \sum_{|\alpha| = m}a_\alpha(x)\xi^{\alpha}. PP is elliptic implies pm(x,ξ)0p_m(x,\xi)\neq 0 when ξ0\xi\neq 0. Remember we want to find some symbol with order m-m, the raw ingadient we have is 1pm(x,ξ)\frac{1}{p_m(x,\xi)}, the growth condition is ξm\langle \xi\rangle^{-m}. To make it a symbol we need to deal with the issue at zero, we can modify it with a smooth “high-pass” filter. Let ρ be a smooth cut off function on R\mathbb{R} with ρ(t)0\rho(t)\equiv 0 for t<1t<1 and ρ(t)1\rho(t)\equiv 1 for t>2t>2. Then a0:=ρ(ξ)1pm(x,ξ)a_0:=\rho(|\xi|)\frac{1}{p_m(x,\xi)} is a symbol in SmS^{-m}.

Let Tr1=IPTa0T_{r_1} = I - PT_{a_0}, or r1=1pa0r_1=1-p\circ a_0. Then r1r_1 is order -1, because the degree mm part of Ta0T_{a_0} and PP are reverse to each other. One can also use the composition formula to see it. By (cmp), the symbol of PTa0PT_{a_0} is

pa0:=p(x,D+ξ)a0(x,ξ)=αmaα(x)(D+ξ)α(ρ(ξ)1α=maα(x)ξα)=α=maα(x)ξαρ(ξ)1α=maα(x)ξα+order 1=ρ(ξ)+ order 1 \begin{split} p\circ a_0 &:=p(x,D+\xi)a_0(x,\xi) \\ &= \sum_{|\alpha|\leq m}a_\alpha(x)(D+\xi)^{\alpha}\left(\rho(|\xi|)\frac{1}{\sum_{|\alpha| = m}a_\alpha(x)\xi^{\alpha}}\right) \\ & = \sum_{|\alpha|=m} a_\alpha(x)\xi^{\alpha} \cdot \rho(|\xi|) \cdot \frac{1}{\sum_{|\alpha| = m}a_\alpha(x)\xi^{\alpha}}+ \text{order }\leq -1 \\ & = \rho(|\xi|)+\text{ order }\leq -1 \end{split}

Observe that, if we let a1=r1a0a_1 = r_1a_0, r2=r1p(r1a0)=1pa0pa1r_2 = r_1 - p\circ(r_1a_0) = 1-p\circ a_0-p\circ a_1, then r2r_2 is order -2, as stated in the lemma below.

Lemma

For bSib\in S^i, bp(a0b)b - p\circ(a_0b) is in Si1S^{i-1}.

The prove of this lemma is the same as the case pa0p\circ a_0. Here we need PP to be elliptic to make sure the mm-order term of PP cancel with a0a_0.

Continue this process we get a sequence riS(i+1)r_i\in S^{-(i+1)} and aiSm(i+1)a_i\in S^{m-(i+1)} so that ri=1[pa0++pai1]r_i =1-[ p\circ a_0 + \cdots + p\circ a_{i-1}]. Eventually we expect to arrive at some a=i=0aia = \sum_{i=0}^{\infty}a_i and rSr\in S^{-\infty} such that r=1par = 1-p\circ a. Then we are done if we can show such aa is in SmS^{m}. The lemma below is slightly different from the ideal case but satisfies our need.

Lemma (The Asymptotic summation theorem)

Let (bi)i=0+(b_i)_{i=0}^{+\infty} be a sequence of symbols such that biSmib_i\in S^{m-i}. Then there exists bSmb\in S^{m} such that for every NN, bSNSm(N+1)b-S_N\in S^{m-(N+1)}, where SN=i=0NbiS_N = \sum_{i=0}^{N}b_i is the NN-th partial sum.

Proof. Fix 0<ϵ<10<\epsilon < 1. Since biSmib_i\in S^{m-i}, then by definition bib_i has the growth condition

bi(x,ξ)<Ci,0ξmi=Ci,0ξm+ϵiξϵ.|b_i(x,\xi)|<C_{i,0}\langle \xi\rangle^{m-i} = C_{i,0}\frac{\langle\xi \rangle^{m+\epsilon -i}}{\langle \xi\rangle^{\epsilon}}.

Then we can choose λi\lambda_i big enough such that when ξ>λi|\xi|>\lambda_i, ξϵ>2i\langle\xi \rangle^{\epsilon} > 2^{i}, so bi(x,ξ)<12iξm+ϵi|b_i(x,\xi)|<\cdot\frac{1}{2^{i}}\cdot\langle \xi\rangle^{m+\epsilon-i}.

The idea is to take the small part for bib_i for every ii, and use smooth high-pass filter to cut off the remaining part, then sum them up. Let ρC(R)\rho\in \mathscr{C}^{\infty}(\mathbb R) be a smooth function such that

ρ(t)={0,t<11,t>2 \rho(t) = \begin{cases} 0, t<1\\ 1, t>2\end{cases}

Then ρ(ξλi)\rho(\frac{|\xi|}{\lambda_i}) is 0 on ξ<λi|\xi|<\lambda_i and 1 on ξ>2λi|\xi|>2\lambda_i, so ρ(ξλi)bi(x,ξ)\rho(\frac{|\xi|}{\lambda_i})b_i(x,\xi) is a filtered symbol which takes only small parts of bib_i.

Let b=i=0ρ(ξλi)bi(x,ξ)b = \sum_{i=0}^{\infty}\rho(\frac{|\xi|}{\lambda_i})b_i(x,\xi). Then we have by direct computation

b=SN1+rN1b = S_{N-1} + r_{N-1}

where rN1=i=0N1(ρ(ξλi)1)bi+ρ(ξλN)bN+i=N+1+ρ(ξλi)bir_{N-1} = \sum_{i=0}^{N-1}(\rho(\frac{|\xi|}{\lambda_i})-1)b_i+\rho(\frac{|\xi|}{\lambda_N})b_N+\sum_{i=N+1}^{+\infty}\rho(\frac{|\xi|}{\lambda_i})b_i. Next we show that the error term rN1r_{N-1} is in SmNS^{m-N}.

The first term is compact supported so doesn’t matter, the second term is in SmNS^{m-N}; The third term is dominated by i=N+1+12iξm+ϵiξm+ϵ12N+1ξ(N+1)\sum_{i=N+1}^{+\infty}\frac{1}{2^{i}}\langle \xi \rangle^{m+\epsilon -i}\leq \langle \xi \rangle^{m+\epsilon}\frac{1}{2^{N+1}}\langle \xi\rangle^{-(N+1)} so it is O(ξmN)O(\langle \xi\rangle^{m-N} ) because ϵ<1\epsilon < 1.

To show that the third term is in SmNS^{m-N} we need to look at derivatives as well.

Fix M>0M>0, we can choose λi,M>0\lambda_{i,M} > 0 such that DxαDξβbi(x,ξ)12iξm+ϵβi|D_x^{\alpha}D_\xi^{\beta}b_i(x,\xi)|\leq \frac{1}{2^{i}}\langle \xi\rangle^{m+\epsilon-|\beta|-i} for every α+βM|\alpha|+|\beta|\leq M. The previous case is the M=0M=0 case. So by replacing λi\lambda_i with λi,M\lambda_{i,M} we can get new bb and rNr_N with

DxαDξβrNCNξmβND_x^{\alpha}D_\xi^{\beta}r_N \leq C_N\langle\xi \rangle^{m-|\beta|-N}

for every α+βM|\alpha|+|\beta|\leq M.

Next we need to eleminate dependence of bb on MM. Run the argument for every MM we get a infinite matrix of positive numbers λi,M\lambda_{i,M}. Now we choose its diagnal, let λi=λi,i\lambda_i = \lambda_{i,i}. Then create bb with such λi\lambda_i will give us desired decreasing property for every α,β\alpha, \beta, hence the rNr_N of this bb satisfiess rNSmNr_N\in S^{m-N}.

\square

There exists aSma\in S^{-m} such that a=i=0+ai+Sa= \sum_{i=0}^{+\infty}a_i + S^{-\infty}. This means a=i=0+ai+ra = \sum_{i=0}^{+\infty} a_i + r for some rSr\in S^{-\infty}. It follows that PTa=1TrPT_a = 1-T_r for some rSr\in S^{-\infty}.

Since TrT_r is smoothing operator we have finished the proof.

The *Kernel-Range relation

Let WW be an inner product space. T:WWT:W\to W be a linear operator, and TT^* be its adjoint operator given by requiring (Tf,g)=(f,Tg)(Tf,g) = (f,T^*g) on WW. Let Ker(T):={wW:Tw=0}\mathrm{Ker}(T):=\{w\in W:Tw=0\} be the kernel of TT and Ran(T):={wW:w=Tw for some wW}\mathrm{Ran}(T):=\{w\in W:w=Tw' \text{ for some }w'\in W\} be the range of TT. Then we automatically have

Ker(T)=Ran(T),\mathrm{Ker}(T^*) = \mathrm{Ran}(T)^{\perp},

this is pure linear algebra and no Hilbert space theory involved.

Taking \perp both side we get Ker(T)=(Ran(T))\mathrm{Ker}(T^*)^{\perp} = (\mathrm{Ran}(T^*)^{\perp})^{\perp}, where the RHS is equal to Ran(T)\overline{\mathrm{Ran}(T)} when WW is a Hilbert space. But we are looking for a condition to imply Ker(T)=Ran(T)\mathrm{Ker}(T^*)^{\perp} = \mathrm{Ran}(T), with no closure allowed, since WW will be space of smooth functions and taking closure in L2L^2 will violate regularity. Here is a handy lemma which satisfies our needs.

Lemma (Key Linear algebra)

If there exists a finite dimensional subspace VWV\subset W such that VRan(T)V^{\perp}\supset \mathrm{Ran}(T), then

(1) Ker(T)=(Ran(T)V)V\mathrm{Ker}(T^*) = (\mathrm{Ran}(T)\cap V)^{\perp _V}, here V\perp_V means taking orthogonal complement in VV. It is equal to (Ran(T)V)V(\mathrm{Ran}(T)\cap V)^{\perp}\cap V. By direct sum decomposition of the finite dimensional VV we get V=Ker(T)(Ran(T)V).V = \mathrm{Ker}(T^*)\oplus (\mathrm{Ran(T)\cap V}).

(2) Consequently, Ker(T)=Ran(T)\mathrm{Ker}(T^*)^{\perp} = \mathrm{Ran}(T).

Proof.

(1)

  • First we show that Ker(T)=Ran(T)V\mathrm{Ker}(T^*) = \mathrm{Ran}(T)^{\perp}\cap V.

Let h1,,hrh_1,\dots,h_r be an orthonormal basis of VV. For any gKer(T)=Ran(T)g\in \mathrm{Ker}(T^*)=\mathrm{Ran}(T)^{\perp}, let g=PVg+g2g = P_V g+g_2, where PVg=i=1r(g,hi)hiP_V g = \sum_{i=1}^r (g,h_i)h_i be the projection of gg to VV. Then g2V    g2Ran(T)g_2\perp V \implies g_2\in \mathrm{Ran}(T). So g2=0g_2 = 0. It follows that g=PVg    gVg = P_V g \implies g\in V. So gRan(T)Vg\in \mathrm{Ran}(T)^{\perp}\cap V, this proves \subset direction.

By the ^*kernel-range relation we have Ker(T)=Ran(T)Ran(T)V\mathrm{Ker(T^*)} = \mathrm{Ran}(T)^{\perp}\supset \mathrm{Ran}(T)^{\perp}\cap V. So Ker(T)=Ran(T)V\mathrm{Ker}(T^*) = \mathrm{Ran}(T)^{\perp}\cap V proved.

  • Next we show that Ran(T)V=(Ran(T)V)V\mathrm{Ran}(T)^{\perp}\cap V = (\mathrm{Ran}(T)\cap V)^{\perp V}.

For fRan(T)f\in \mathrm{Ran}(T), write f=PVf+f2f=P_V f+f_2. Then f2V    f2Ran(T)f_2\perp V\implies f_2\in \mathrm{Ran}(T), so PVfRan(T)VP_V f \in \mathrm{Ran}(T)\cap V. If gVg\in V and gRan(T)Vg\perp \mathrm{Ran}(T)\cap V, then gPVfg\perp P_V f. But f2Vf_2\perp V, so gf2g\perp f_2. So gfg\perp f. We have showed that g(Ran(T)V)V    gRan(T)Vg\in (\mathrm{Ran}(T)\cap V)^{\perp V}\implies g\in \mathrm{Ran}(T)^{\perp}\cap V, the \supset direction.

The converse inclusion is automatic because (Ran(T)V)(\mathrm{Ran}(T)\cap V)^{\perp} is bigger than Ran(T)\mathrm{Ran}(T)^{\perp}.

(2) The *kernel-range relation implies Ran(T)Ker(T)\mathrm{Ran}(T)\subset \mathrm{Ker}(T^*)^{\perp}.

Just need to prove Ker(T)Ran(T)\mathrm{Ker}(T^*)^{\perp}\subset \mathrm{Ran}(T). For fWf\in W such that fKer(T)f\perp \mathrm{Ker}(T^*), write f=PVf+f2f = P_V f + f_2 so f2Ran(T)f_2\in \mathrm{Ran}(T). Then f2Ker(T)f_2\perp \mathrm{Ker}(T^*), so PVfKer(T)P_V f \perp \mathrm{Ker}(T^*). By (1) we have PVfRan(T)P_V f\in \mathrm{Ran}(T). This implies fRan(T).f\in \mathrm{Ran}(T).

The Fredholm theorem for Elliptic operators

Theorem

Let TKT_K be a smoothing operator on XX. Then ITKI-T_K satisfies

(1) Ker(ITK)\mathrm{Ker}(I-T_K) is finite dimensional.

(2) For fC(X)f\in \mathscr{C}^{\infty}(X), fRan(ITK)f\in \mathrm{Ran}(I-T_K) if and only if fKer(ITK)f\perp \mathrm{Ker}(I-T_{K^*}).

Finite rank case

A smoothing operator TFT_F is called finite rank provided F(x,y)=i=1Nfi(x)gi(y)F(x,y) = \sum_{i=1}^N f_i(x)g_i(y).

Lemma

Let FF be a finite rank smoothing operator. Then

(1) Ker(ITF)\mathrm{Ker}(I-T_F) is finite dimensional.

(2) If (f,gi)=0(f,g_i) = 0 for every i=1,,Ni=1,\dots,N, then fRan(ITF)f\in \mathrm{Ran}(I-T_F).

Approximation by finite rank

Lemma

There exists smoothing LL such that ITLI-T_{L} is invertible and

(1) (ITL)(ITK)=ITF1(I-T_L)(I-T_K) = I-T_{F_1}

(2) (ITK)(ITL)=ITF2(I-T_K)(I-T_L) = I-T_{F_2}

for some finite rank smothing F1,F2F_1,F_2.

Proof of theorem

Let L,F1,F2L,F_1,F_2 be the one in the approximation by finite rank lemma above.

Proof of (1):

If fKer(ITK)f\in \mathrm{Ker}(I-T_K), then (ITL)(ITK)f=(ITL)0=0(I-T_L)(I-T_K)f = (I-T_L) 0 = 0, so fKer(ITF1)f\in \mathrm{Ker}(I-T_{F_1}). By the finite rank case Ker(1TF)\mathrm{Ker}(1-T_F) is fintie dimensional. Since ITLI-T_L is invertible this implies Ker(ITK)\mathrm{Ker}(I-T_K) is also finite dimensional.

Proof of (2):

It’s sufficient to prove fKer(ITK)    fRan(ITK)f\perp\mathrm{Ker}(I-T_{K^*})\implies f\in \mathrm{Ran}(I-T_K). The other direction is automatic by the *kernel-range relation.

Let F2=i=1Nfi(x)gi(y)F_2 = \sum_{i=1}^N f_i(x)g_i(y). It’s sufficient to prove that fKer(ITK)f\perp \mathrm{Ker}(I-T_{K^*}) implies fRan(ITF)f\in \mathrm{Ran}(I-T_F). Because fRan(1TF2)    f=(ITK)(ITL)gf\in \mathrm{Ran}(1-T_{F_2}) \implies f = (I-T_K)(I-T_{L})g for some gg, so fRan(ITK)f\in \mathrm{Ran}(I-T_K).

Take V=span{g1,,gN}V = \mathrm{span}\{g_1,\dots,g_N\}. By the lemma in finite rank case, fVf\perp V implies fRan(ITF2)    fRan(ITK)f\in \mathrm{Ran}(I-T_{F_2})\implies f\in \mathrm{Ran}(I-T_K). Then apply the Key linear algebra lemma above to T=ITKT = I-T_{K} and VV we conclude that fRan(ITF2)f\in \mathrm{Ran}(I-T_{F_2}).