The Calderón problem for the Fractional Dirac Operator

The Calderón problem asks whether one can determine the electrical conductivity of a medium by making voltage and current measurements at the boundary. In the anisotropic case, that is when the conductivity depends on direction. It is modelled by a positive-definite symmetric matrix. It was shown in [24] that this problem is equivalent to determining a Riemannian metric from the associated Dirichlet to Neumann (DN) map associated with harmonic functions. Therefore this problem can be considered on a compact Riemannian manifold with boundary and arises also in the AdS/CFT correspondence [5]. See [33], [34] for more details and other results.

The study of the fractional Calderón problem was initiated in [15] where the unknown potential in the fractional Schrödinger equation on a bounded domain in the Euclidean space was determined from exterior measurements. An important generalization was in the work [13] where the Euclidean Laplacian is replaced by the Laplace-Beltrami operator of a Riemannian metric. Following these works, inverse problems of recovering lower order terms for fractional elliptic equations have been studied extensively, see for example [14], [13], [30], [29], [2], [4], [7], [6], [8], [9], [10], [25], [26], [28] for some of the contributions. In all of those papers, it is assumed that the leading order coefficients are known.

In the article [12], it was solved the fractional anisotropic Calderón problem on closed Riemannian manifolds of dimensions two and higher that the knowledge of the local source-to-solution map for the fractional Laplacian, given an arbitrary small open nonempty a priori known subsef a smooth closed connected Riemannian manifold, determines the Riemannian manifold up to an isometry. This can be viewed as a nonlocal analog of the anisotropic Calderón problem in the setting of closed Riemannian manifolds, which is open in dimensions three and higher.

We consider in this paper the anisotropic Calderon problem for the fractional Dirac operator acting on sections of a Hermitian vector bundle over a smooth closed connected Riemannian manifold $M$ of dimension $m\geq 2$. Just from local measurements confined to an arbitrary non-empty open set, we show that the fractional Dirac operator determines the smooth structure, Riemannian metric, Hermitian bundle and connection, and its Clifford module, up to a isometry which fixes the set in question. We will first briefly motivate the concept of a generalized Dirac operator, and its fractional powers, before explaining the types of measurements we allow, and stating our results.

In his attempts to quantize electromagnetism, Dirac defined the Dirac operator while trying to find a 1st order differential operator such that its square was the “Laplacian” on Minkowski space (in fact this was a factorization of the d’Alembertian). The data of the Clifford algebra (see §2) entered into his definition after he realized this was not possible for a scalar operator. On a vector bundle $E$ over a Riemannian manifold $M$, the definition of a generalized Dirac operator $\not{D}$ is motivated by the desire to capture this property that $\not{D}\circ\not{D}=\Delta_{E}$ is a generalized Laplace-type operator on $E$, i.e. the Bochner Laplacian $\nabla^{*}\nabla$ of some connection up to lower order terms. This factorization of $\Delta_{E}$ implies $\not{D}$ is elliptic and hence admits a spectral resolution $\{\varphi_{k},\lambda_{k}\}$ of $L^{2}(M;E)$-orthonormal eigensections, with discrete eigenvalues accumulating only at $\pm\infty$. From the spectral theorem we can define for $\alpha\in(0,1)$, the fractional Dirac operator by $\not{D}^{\alpha}\varphi_{k}=\lambda_{k}^{\alpha}\varphi_{k}$. From unique solvability of the “Poisson equation”,

for data $f$ orthogonal to $\operatorname{Ker}(\not{D})$, we can consider the operator which maps a source function $f\in{\mathcal{C}}^{\infty}_{0}(\mathcal{O};E|_{\mathcal{O}})\cap\operatorname{Ker}(\not{D})^{\perp}$ to the solution,

for $f\perp\operatorname{Ker}(\not{D})$. We refer to this operator as the source-to-solution map. We prove,

Let $(M^{m},E^{n},h_{E},g)$ be a Hermitian vector bundle of rank $n$ over smooth compact Riemannian manifold without boundary. We further call this Hermitian bundle a Clifford module if $E\to M$ is bundle over $M$ equipped with bundle morphism, known as Clifford multiplication,

from the unique Clifford module on $T^{*}M$ induced by $g$ to $\operatorname{End}(E)$. As a vector bundle, $\mathbb{C}l(T^{*}M,g)$ is isomorphic to $\Lambda^{\bullet}T^{*}M$, and as a module has an algebra operation determined by the relation

We say the Hermitian metric is compatible with Clifford multiplication if

and a choice of Hermitian connection $\nabla^{E}$ on $E$ is is compatible with Clifford multiplication if

where $\nabla^{g}$ is the Levi-Civita connection of $g$.

Given a Hermitian bundle with connection both compatible with Clifford multiplication we can construct a generalized Dirac operator $\not{D}$, defined by

and call the collected data $(M,E,g,h_{E},\nabla^{E})$ a Dirac bundle. The generalized Dirac operator is a Dirac operator in the usual sense by arising as the ‘square root’ of a generalized Laplace operator on $E$; compatibility of the Hermitian metric and connection implies

thus $\not{D}$ has principal symbol

hence $\not{D}^{2}:=\Delta_{E}$ is a principally scalar multiple of the metric $g$. Further, we see from this computation that both $\not{D}$ and $\not{D}^{2}$ are principally scalar elliptic differential operators (of orders 1 and 2 respectively).

In this section we give two equivalent definitions of the fractional Dirac operator and use one to define the source-to-solution operator associated to an a priori known open set $\mathcal{O}\subset M$. Then, following [12], we show that knowledge of this source-to-solution operator determines the Heat kernel on $\mathcal{O}$.

From the symbol calculation above we see that $\not{D}$ is a symmetric operator. Because $M$ is a closed manifold, $\not{D}:{\mathcal{C}}^{\infty}(M;E)\to{\mathcal{C}}^{\infty}(M;E)$ is essentially self-adjoint on its core domain of smooth sections with a self-adjoint extension to $H^{1}(M;E)$. Unlike $\not{D}^{2}$, $\not{D}$ fails to be a non-negative operator. On the other hand, its discrete spectrum (excluding the zero eigenvalue) is in correspondence with the spectrum of $\not{D}^{2}$; the discrete eigenvalues of $\not{D}$ come in positive and negative pairs $\{\pm\lambda_{k}\}$ (corresponding to an eigenvalue $\lambda_{k}^{2}$ of $\not{D}^{2}$) which we index by their absolute values

for the distinct eigenvalues of $\not{D}$, and denote $d_{k}$ the multiplicity of $\lambda_{k}$. Let $\{\varphi_{k_{j}}\}_{j=1}^{d_{k}}$ be an $L^{2}(M;E)$-orthonormal basis for the eigenspace $\operatorname{Ker}(\not{D}-\lambda_{k})$ corresponding to $\lambda_{k}$, and denote $\pi_{k}:L^{2}(M;E)\to\operatorname{Ker}(\not{D}-\lambda_{k})$ for the orthogonal projection onto the corresponding eigenspace, written as

for all $k=0,1,\ldots$. Here $\langle\cdot,\cdot\rangle_{L^{2}(M;E)}$ is the $L^{2}$-inner product on sections of $E$ induced by our choice of $h_{E}$.

Fix $\alpha\in(0,1)$. Given this spectral resolution of $\not{D}$ we can define the Fractional Dirac operator $\not{D}^{\alpha}:{\mathcal{C}}^{\infty}(M;E)\to{\mathcal{C}}^{\infty}(M;E)$

which extends to an unbounded self-adjoint operator on $L^{2}(M;E)$ with domain $\mathcal{D}(\not{D}^{\alpha})=H^{\alpha}(M;E)$.

Unlike the scalar case, the nullspace of $\not{D}$ may include more than just the constant functions. Say that dim $\operatorname{Ker}(\not{D})=d_{0}$ has orthonormal basis $\{\varphi_{j}^{0}\}_{j=1}^{d_{0}}$, then we can solve the equation

for $f\in{\mathcal{C}}^{\infty}_{0}(\mathcal{O};E)$, with $\mathcal{O}\subset M$ open, whenever we impose

for a unique solution $u=u^{f}\in{\mathcal{C}}^{\infty}(M;E)$ defined by the condition that $u^{f}\perp_{h_{E}}\operatorname{Ker}(\not{D})$. Associated to the equation (3.1) we can define the local source-to-source solution map $L_{M,g,\mathcal{O},\ldots}$ by

We can give an equivalent (spectral theoretic) definition for the Fractional Dirac operator via the heat kernel of the square of the Dirac operator. First note that we can define from the integral formula for the Gamma function,

thus we write

Using this formula, we have $\not{D}^{\alpha}=\not{D}\circ\Delta_{E}^{\frac{\alpha-1}{2}}=\Delta_{E}^{\frac{\alpha-1}{2}}\not{D}$, i.e.

using the fact that $\not{D}e^{-t\Delta_{E}}=e^{-t\Delta_{E}}\not{D}$ (from uniqueness of the solution to the heat equation).

Using this definition of $\not{D}^{-\alpha}$ we can easily extend the proof of [12, Thm 1.5] to our setting:

Remark: The requirement for well-posed of the source-to-source solution maps that $f$ be orthogonal to the space Harmonic sections is subtle: the local bundle isomorphism $\phi$ is already sufficient to ensure that $\text{dim}\;\operatorname{Ker}(\not{D}_{1})=\text{dim}\operatorname{Ker}(\not{D}_{2})$. If $f\in\operatorname{Ker}(\not{D}_{2})$ then its restriction to $\mathcal{O}_{2}$ pullbacks to a section of $E_{1}|_{\mathcal{O}_{1}}$, and in particular is also an element of $\operatorname{Ker}(\not{D}_{1})$ as the Dirac operator commutes with pullback. Similarly pullback by the inverse proves the reverse inclusion. The possibility that a non-trivial Harmonic section vanishes on the open sets in consideration is ruled out by the unique continuation principle for the Dirac operator: if $\not{D}f=0$ and $f|_{\mathcal{O}}=0$ then $f=0$ (see e.g. [3, Ch. 8]).

Have shown that the source-to-source solution operators determine the Heat kernel on the open neighborhood $\mathcal{O}$. Using the Kannai transmutation formula

the local determination of $e^{-t\Delta_{E_{j}}}(x,y)$ and this equality imply

for all $t>0$ and $x\in\mathcal{O}$.

The benefit of this last equality is that the operator in (4.1) lets us represent the solution of the initial value problem,

for data $f\in{\mathcal{C}}^{\infty}_{0}(\mathcal{O}\times M_{j};E_{j})$ via the formula

Associated to (4.2), we define the source-to-source solution map for the wave operator on $\mathcal{O}$ by

where $u_{j}(t,x)$ is the unique solution to (4.2). And by the equality in (4.1) we have equality of the source-to-solution maps for the wave operators for $j=1,2$,

Now we can move to the statement of our main theorem.

We follow the proof of [17], (which itself was an application to closed manifolds of the boundary control method of Belishev [1] combined with the crucial unique continuation method of Tataru [32]) extended to the bundle case. The general structure of the proof is as follows

1. (1)

   The source-to-source solution operator $(\mathcal{O},L_{\mathscr{D},\mathcal{O}}^{\text{wave}})$ determines the distance function $d_{g}$ on $\mathcal{O}\times\mathcal{O}$

2. (2)

   The source-to-source solution data $(\mathcal{O},g|_{\mathcal{O}},h_{E}|_{\mathcal{O}},L_{\mathscr{D},\mathcal{O}}^{\text{wave}})$ determines the distance data $(\mathcal{O},g|_{\mathcal{O}},R(M))$

3. (3)

   The distance data $(\mathcal{O},g|_{\mathcal{O}},R(M))$ determines the topology, smooth structure and Riemannian structure of $(M,g)$

4. (4)

   New: The distance data determines the isomorphism class of $E\to M$ and Hermitian metric $h_{E}$

5. (5)

   New: The distance data determines the Hermitian connection $\nabla^{E}$, and thus the homomorphism of clifford multiplication from the identity $[\nabla^{E},\mathpzc{cl}(\theta)]=\mathpzc{cl}(\nabla^{g}\theta)$

This uses several facts about the solution to the linear wave equation on sections of $E$: the first is its finite speed of propagation.

To give the statement of the relevant unique continuation principle we also define

for the domain of dependence of the wave equation. As a result of the Carleman estimates established in [11], the proof of the unique continuation theorem is as in the scalar case since the wave equation being considered is principally scalar (i.e. $\partial_{t}^{2}-\Delta_{E}$) see section 2.5 of [20],

Given the proper form of unique continuation we should be able to prove density of solutions with sources from the set,

Having proven this fact, the proofs of 1)  2) and 3) from [17] generalize immediately to the bundle-valued case. One lemma they depend on is the Blagovestchenskii identity:

Another such lemma is that the source-to-source solution operator determines the distance function:

Let $\gamma_{y,\xi}(t)$ be the time-$t$ flow of the unique unit speed geodesic starting at $y$ with initial velocity vector $\xi$. Several facts about geometric determination in our setting will rely on the notion of cut time,

Note that we have

for $\mathcal{O}\subset M$ open and non-empty. In other words any point in $M$ can be reached by geodesics originating in $\mathcal{O}$ which do not meet the cut locus.

Further, the cut distance and distance between points can be determined using only the source-to-source solution data

namely as shown by [17] in the scalar case,