arXiv reader

On the Stability of Inverse Conductivity Problem for Polyhedral Inclusions under a Single Measurement

Chun-Hsiang Tsou

arXiv
2605.17484v2 · PDF · 1 July 2026
Categories
math.AP
Licence
CC BY 4.0 The site shows the text.

Abstract

In this paper, we study the stability of the inverse conductivity problem of determining a convex polyhedral inclusion embedded in a homogeneous isotropic medium from a single boundary measurement. The main tools in our analysis are singularity decomposition for elliptic equations in non-smooth domains, propagation of smallness, and microlocal analysis. Combining these tools, we establish a logarithmic stability estimate for the Hausdorff distance between inclusions in terms of the measurement error.

Items 25

Assumption

Assumption 1.1 (Geometric)

1XDK

Throughout this paper, we impose the following geometric conditions, and the corresponding parameters are regarded as a priori data.

  • The Lipschitz domain \Omega is arcwise connected, with Lipschitz constant L_\Omega>0.

  • The conductivity contrast k>1, i.e., the inclusion is more conductive than the surrounding medium.

  • The inclusions are at distance at least \delta_0>0 from the boundary \partial\Omega, i.e., D\subset \Omega_{\delta_0} where \Omega_{\delta_0}:=\left\lbrace {x\in\Omega} \; \middle| \; {{\rm dist}(x,\partial\Omega)>\delta_0} \right\rbrace.

  • The interior domain \Omega_{\delta_0} is also arcwise connected.

  • Each inclusion D is a convex polyhedron with edges of length at least l_0>0.

  • The openings of all vertices are assumed to lie in a fixed range [\vartheta_m,\vartheta_M]\subset (0,2\pi).

  • The openings of all edges are assumed to lie in a fixed range [a_m,a_M]\subset (0,\pi).

  • The Dirichlet boundary condition f\in H^{1/2}(\partial\Omega) is nontrivial, with \|f\|_{H^{1/2}(\partial\Omega)}>0.

Notations

Notations

75TQ
  • We will use the symbols “\lesssim” and “\gtrsim” to denote the inequality up to a multiplicative constant depending only on a priori data.

  • We denote by d_H(D,D') the Hausdorff distance between D and D', which is defined as follows,

    \begin{aligned} d_H(D,D')=\max\left\lbrace \sup_{x\in D}{\rm dist}(x,D'),\; \sup_{x'\in D'}{\rm dist}(x',D) \right\rbrace. \end{aligned}
Theorem

Theorem 1.2

YWC8

Assume the geometrical and analytical conditions assumption:geometric,assumption:singularity hold. Let u and u' be solutions to (?) with respect to D and D'. We suppose that \|\partial_\nu(u-u')\|_{H^{-1/2}(\Gamma_0)}\leq \varepsilon for \varepsilon>0 being sufficiently small. Then, there exists positive constants \kappa and C depending only on a priori data such that

\begin{aligned} d_H(D,D')\leq C \left(\ln|\ln \varepsilon|\right)^{-\kappa}. \end{aligned}
Uses Assumption 1.1, Notations · Used by Proof
Theorem

Theorem 2.1

3NZH

Let u be the solution to (?). Then, u admits the following decomposition into vertex-singular part, edge-singular part and a regular part.

\begin{aligned} u=\sum_{{\bf v}\;{\rm vertex}} u_{{\bf v}}+\sum_{E \;{\rm edge}} u_E+u_{\rm reg}. \end{aligned}
Assumption

Assumption 2.2 (Singularity)

M6PJ

Throughout the paper, we will assume that the following conditions hold.

  1. We follow case b) of Theorem 7.1 in [Dauge1999]. That is, the parameter s is chosen such that s_0\leq s <s_0+1 in order to have sufficient regularity for u_{\rm reg} and to avoid the “shadow terms” in condition c) for singular functions. Furthermore, from the estimates of the first singularity exponents \frac{1}{2}<\mu_1<1<\lambda_1, we have s_0=\mu_1>\frac{1}{2} and s_0+1>\frac{3}{2}. With this condition, we will choose s close enough to \mu_1+1 such that,

    \begin{aligned} \frac{3}{2}< s <\mu_1+1. \end{aligned}

    As an immediate consequence, the regular part u_{\rm reg} has at least H^{5/2+\epsilon} regularity in each subdomain, where \epsilon:=s-\frac{3}{2}>0.

  2. There is no \log terms in the vertex or edge singular expansions. That is, K_{\bf v,\lambda}=K_{E,\mu}=0 for any vertex \bf v and any edge E.

  3. For any edge E, the first edge singularity is not degenerate, i.e., \gamma_{E,\mu_1} is not identically null on E.

  4. For any adjacent vertex-edge pair ({\bf v}, E) the first interaction coefficient a_{E,\mu_1}^{{\bf v},\lambda_1} in the asymptotic expansion (?) is not zero.

Lemma

Lemma 3.1 (Three sphere inequality)

9VUH

Let 0<r_1<r_2<r_3. There exists a constant \widetilde{\alpha}\in (0,1) such that for any harmonic function w\in H^1_{loc}(2.4^n) with n\geq 2, it holds that

\begin{aligned} \sup_{|x|\leq r_2}|w(x)|\leq \left(\sup_{|x|\leq r_1}|w(x)|\right)^{\widetilde{\alpha}}\left(\sup_{|x|\leq r_3}|w(x)|\right)^{1-\widetilde{\alpha}}. \end{aligned}

The exponent \widetilde{\alpha} depends only on n and the ratios r_1/r_2, r_2/r_3.

Proposition

Proposition 3.2

0VM6

Let r>0 and let U\subset 2.4^3 be a bounded and connected domain. Consider x,y\in U and assume that there exists a rectifiable curve \gamma in U with endpoints x and y such that {\rm dist}(\gamma,\partial U)>4r. Consider also a bounded harmonic function w in U. We assume furthermore that w is bounded with \|w\|_{L^\infty(B(x,r))}\leq 1 and \|w\|_{L^\infty(U)}= T>0. Then, it holds that

\begin{aligned} \|w\|_{L^\infty(B(y,r))}\leq T \|w\|_{L^\infty(B(x,r))}^{\widetilde{\alpha}^{|\gamma|/r+1}}, \end{aligned}

where |\gamma| denotes the length of \gamma.

Proof of Proposition 3.2

See Lemma 3.2 in [moi5].

Proposition

Proposition 3.3

MFW9

Let U be a Lipschitz connected domain. Let \Sigma\subset \partial U be an open portion of the boundary. Consider a harmonic function w in U. We make the following assumptions

  • for any y\in U, there exists a rectifiable curve \gamma in U connecting y and a point x\in \Sigma,

  • the length of \gamma is bounded by a uniform constant L>0,

  • \gamma does not touch the boundary of U outside \Sigma, i.e., {\rm dist}(\gamma,\partial U\setminus \Sigma)>4r for some r>0,

  • w is bounded with \|w\|_{L^\infty(U)}=T>0,

  • \|w\|_{H^{1/2}(\Sigma)}+\|\partial_{\nu} w\|_{H^{-1/2}(\Sigma)}\leq \varepsilon with 0< \varepsilon<1.

Then, for all y\in U and all sufficiently small r>0, one has

\begin{aligned} |w(y)|\leq C T \varepsilon^{\widetilde{\alpha}^{L/r+1}}, \end{aligned}

with a constant C depending only on a priori data.

Uses Notations · Used by Proof
Proof of Proposition 3.3

This follows from the arguments in Remark 6.3 of [Alessandrini2009]. Using Lemma 6.1 and Theorem 6.2 in that paper, we can extend w to an exterior neighborhood V of \Sigma such that the extension \widetilde{w} satisfies \|\widetilde{w}\|_{H^1(V)}\leq C\|w\|_{H^{1/2}(\Sigma)}. Moreover, \widetilde{w} solves \Delta \widetilde{w}=\widetilde{f} in U\cup V with a sufficiently small right-hand side \widetilde{f}. Then, interior elliptic regularity yields \|\widetilde{w}\|_{L^\infty(B(x_0,r_0))}\leq \varepsilon for some x_0\in V and r_0>0, where r_0 depends only on the Lipschitz character of \Sigma. Finally, we apply prop:p1 to this modified equation to deduce the desired estimate.

Proposition

Proposition 3.4

K8U3

Let u and u' be solutions to (?) corresponding to D and D', respectively. Let {\bf v} \in \partial D be a vertex of D, and let h>0 satisfy B({\bf v},h)\subset\Omega_{\delta_0} and B({\bf v},h)\cap D'=\varnothing. Then there exist \alpha \in (0,1) and T>0, depending only on a priori data, such that for 0<\varepsilon<\varepsilon_m,

\begin{aligned} \forall x\in B({\bf v},h)\setminus \overline{D},\quad & |(u-u')(x)|\leq T\left(\ln |\ln \varepsilon|\right)^{-\alpha};\\ &|\nabla (u-u')(x)|\leq \frac{T}{{\rm dist}(x,\partial D)}\left(\ln |\ln \varepsilon|\right)^{-\alpha}. \end{aligned}
Proof of Proposition 3.4

Using the convexity of D and D', there exist affine hyperplanes P' and P such that P' separates B({\bf v},h) from D', and P\cap \overline{(D\cup D')}={\bf v}. Denote by P'_\pm the two half-spaces determined by P', with D'\subset P'_-. Similarly, let P_- be the half-space determined by P such that D\cup D'\subset P_-, and let P_+ be the opposite half-space. Then the region V=(P_+\cap P'_+ \cap \Omega_{\delta_0})\cup (\Omega\setminus \overline{\Omega_{\delta_0}}) is connected and satisfies \Gamma_0\subset \partial V. Consequently, we can find a rectifiable curve \gamma_{0,{\bf v}} connecting \bf v to a point x_0\in \Gamma_0 such that \gamma_{0,{\bf v}} stays at distance at least r_m:=\frac{1}{3}\min\{h,\delta_0\} from \partial\Omega\cup \partial (D\cup D')\setminus (\Gamma_0 \cup B({\bf v},h)). We also define the domain U by

\begin{aligned} U=B({\bf v},h)\cup \left\lbrace {y\in\Omega} \; \middle| \; {{\rm dist}(y,\gamma_{0, {\bf v}})<r_m} \right\rbrace. \end{aligned}

Then we choose r<\min\{r_0,r_m\} small enough to apply prop:aa2 to the harmonic function u-u' in the domain U, along a curve obtained by slightly modifying \gamma_{0,{\bf v}}. Here the quantity r_0 depends only on the Lipschitz character of \Sigma, which comes from the proof of prop:aa2. It remains to derive Hölder estimates for u-u' and its gradient on U in terms of the distance to \partial D.

The well-known result of De Giorgi-Nash-Moser (Theorem 8.24 in [Gilbarg-Trudinger]) states that solutions to strongly elliptic equations with bounded coefficients are Hölder continuous at interior points. Consequently, u-u'\in\mathcal{C}^{\alpha_0}(\Omega_{\delta_0}) for some \alpha_0>0. We also introduce T_0=\|u-u'\|_{\mathcal{C}^{\alpha_0}(\Omega_{\delta_0})}.

For the gradient \nabla (u-u'), we expect blow-up near the boundary \Gamma_{D\cup D'}. To control the blow-up rate, we use the decomposition of u and u' into singular and regular parts as stated in th:SingDecom.

We recall that the regular part u_{\rm reg} has H^{5/2+\epsilon} regularity in D\cup (\Omega\setminus \overline{D}). By applying Sobolev embedding H^{3/2+\epsilon}\hookrightarrow \mathcal{C}^{\epsilon} in 2.4^3, we can deduce that \nabla u_{\rm reg} is Hölder continuous in U.

For the singular parts, it is sufficient to work in local spherical or cylindrical coordinates near vertices or edges. Using the estimates of the singularity exponents \lambda and \mu, we deduce that the singularity of \nabla u_{\bf v} is at most of type \rho^{\lambda_1-1} and the singularity of \nabla u_E is at most of type r^{\mu_1-1}, as shown in lemma:esti-grad-edge. Consequently, the function x\mapsto {\rm dist}(x,\partial D)\nabla u(x) is \mathcal{C}^{\alpha_1} Hölder continuous in U with 0<\alpha_1\leq \min\{\epsilon,\lambda_1, \mu_1\}.

The same arguments hold for u'. Hence, we can deduce that {\rm dist}(x,\partial D)\nabla (u-u')(x) is Hölder continuous in U, and thus we can set T_1=\|{\rm dist}(\cdot,\partial D)\nabla (u-u')\|_{\mathcal{C}^{\alpha_1}(U)}. Next, we set \alpha=\min\{\alpha_0,\alpha_1\} and T=\max\{T_0,T_1\} and remark that those parameters depend only on a priori data.

The remainder of the proof consists in estimating (3.3) as r\rightarrow0^+ by using the Hölder continuity of these functions. The argument follows Proposition 3.4, Corollary 3.5, and Proposition 3.7 in [moi5]. The value of \varepsilon_m is given by (3.4) in [moi5].

Proposition

Proposition 4.1

HSS3

Let u_0\in H^1_{loc}(2.4^3) be harmonic. Then

\begin{aligned} (k-1)\int_{F_{\pm}}u_0\partial_\nu u d\sigma= &\int_{S^e \cup S^i\cup A_{\pm}} \left(u_0\partial_\nu (u-u')-(u-u')\partial_\nu u_0\right)d\sigma. \end{aligned}
Used by Proof
Proof of Proposition 4.1

The proof follows by applying Green’s formula in the regions \widetilde{\Omega}\setminus \widetilde{D} and \widetilde{D}, together with the transmission conditions in (?).

Proposition

Proposition 4.2

B27I

The left-hand side of the integral identity (4.1) admits the following decomposition:

\begin{aligned} &\int_{F_\pm} u_0\partial_\nu u d\sigma = I_{F,1} + I_{F,2,\infty}-I_{F,2,c} + I_{F,3} + I_{F,4} + I_{F,5}. \end{aligned}

Furthermore, for \rho,h,h_z satisfying (?) and (?) and being sufficiently small, the following estimates hold for each term,

\begin{aligned} &|I_{F,1}|\lesssim \rho^{\lambda_1-1}h_z\tau^{-1}, & &|I_{F,2,\infty}| \gtrsim \rho^{\lambda_1+\varepsilon_*-\mu_1} h_z \tau^{-\mu_1},\\ &|I_{F,2,c}| \lesssim (\rho+h_z)^{1-\mu_1}\tau^{-\mu_1} e^{-\frac{\tau \widetilde{a} h}{2}}, & &|I_{F,3}| \lesssim \rho^{-(s-\mu_1-1/2)}h_z \tau^{-(s-1/2)},\\ &|I_{F,4}| \lesssim \rho^{-\widetilde{\mu_2}}h_z\tau^{-\widetilde{\mu_2}},& &|I_{F,5}| \lesssim h_z\tau^{-1}. \end{aligned}

Here, \varepsilon_*>0 is a small constant, \widetilde{\mu_2}=\mu_2 if s-\mu_2\geq 1/2 and \widetilde{\mu_2}=s-1/2 if s-\mu_2< 1/2.

Proof of Proposition 4.2

We split u into singular and regular parts as stated in th:SingDecom. We mainly focus on the first edge singular function. More precisely, we decompose

\begin{aligned} u=u_{\bf v}+v_1+\widetilde{v_1}+u_{E,2}+u_{\rm reg}, \end{aligned}

where

  • u_{\bf v} is the vertex singular function associated with \bf v,

  • \displaystyle v_1(\widetilde{r},\theta,z):=\gamma_{E,\mu_1}(z)\widetilde{r}^{\mu_1}\varphi_{E,\mu_1}(\theta),

  • \displaystyle \widetilde{v_1}=(\mathcal{K}[\gamma_{E,\mu_1}](\widetilde{r},z)-\gamma_{E,\mu_1}(z))\widetilde{r}^{\mu_1}\varphi_{E,\mu_1}(\theta),

  • u_{E,2} is the remaining edge singular functions with the leading singular exponent \mu_2,

  • u_{\rm reg} is the regular part.

We also extend the rays \Gamma_\pm to infinity and denote the corresponding faces by F_\pm^\infty. Straightforward computations give

\begin{aligned} \int_{F_\pm} u_0\partial_\nu u d\sigma = I_{F,1} + I_{F,2,\infty}-I_{F,2,c} + I_{F,3} + I_{F,4} + I_{F,5}. \end{aligned}

where

\begin{aligned} &I_{F,1} := \int_{F_{\pm}}u_0\partial_\nu u_{\bf v} d\sigma, & & I_{F,2,\infty} := \int_{F_{\pm}^\infty}u_0\partial_\nu v_1 d\sigma,\\ &I_{F,2,c} := \int_{F_{\pm}^\infty\setminus F_{\pm}}u_0\partial_\nu v_1 d\sigma,& & I_{F,3} := \int_{F_{\pm}}u_0\partial_\nu \widetilde{v_1} d\sigma,\\ &I_{F,4} := \int_{F_{\pm}}u_0\partial_\nu u_{E,2} d\sigma, & & I_{F,5} := \int_{F_{\pm}}u_0\partial_\nu u_{\rm reg} d\sigma. \end{aligned}

We first focus on the integral I_{F,2,\infty}. Since the variables are separated, we have

\begin{aligned} &\int_{F^\infty_\pm}u_0\partial_\nu v_1 d\sigma = \int_0^{h_z} \int_0^{+\infty} u_0(r,\pm \frac{a}{2},z)\partial_\nu v_1(r,\pm \frac{a}{2},z) dr dz,\\ &= \int_0^{h_z} \gamma_{E,\mu_1}(z) \delta(z)^{-\mu_1}dz \left( \varphi'_{E,\mu_1}(\frac{a}{2}) \int_0^{+\infty} r^{\mu_1-1} e^{z^+_a r} dr- \varphi'_{E,\mu_1}(-\frac{a}{2}) \int_0^{+\infty} r^{\mu_1-1} e^{z^-_a r} dr \right). \end{aligned}

By construction of the CGO solution, u_0(r,\pm \frac{a}{2},z) = e^{z^\pm_a r} with z^\pm_a=-\tau e^{\mp i\frac{a}{2}}. To evaluate the integral in r, we use a complex-analysis argument from Proposition 4.4 in [moi5]. Thus,

\begin{aligned} \int_0^{+\infty} r^{\mu_1-1} e^{z^{\pm}_a r} dr =\Gamma(\mu_1)(\frac{-1}{z^\pm_a})^{\mu_1}; \end{aligned}

Here \Gamma denotes Euler’s Gamma function. Hence,

\begin{aligned} \left| \varphi'_{E,\mu_1}(\frac{a}{2}) \int_0^{+\infty} r^{\mu_1-1} e^{z^+_a r} dr- \varphi'_{E,\mu_1}(-\frac{a}{2}) \int_0^{+\infty} r^{\mu_1-1} e^{z^-_a r} dr \right| \geq \sin(a\mu_1)\Gamma(\mu_1)\tau^{-\mu_1}. \end{aligned}

We now estimate the integral with respect to z, \displaystyle \int_0^{h_z} \gamma_{E,\mu_1}(z) \delta(z)^{-\mu_1}dz. We use the assumptions 3.) and 4.) in assumption:singularity and the asymptotic behavior (?) to derive that \displaystyle \gamma_{E,\mu_1}(\rho) \sim a_{E,\mu_1}^{{\bf v},\lambda_k} \rho^{\lambda_k} as \rho \to 0. This asymptotic behavior implies that \gamma_{E,\mu_1}(\rho) keeps the same sign as \rho \rightarrow 0, and we will assume that it is positive without loss of generality. By choosing \varepsilon_*>0, we have,

\begin{aligned} \frac{\gamma_{E,\mu_1}(\rho)}{\rho^{\lambda_1+\varepsilon_*}} \underset{{\rho}\rightarrow {0}}{\longrightarrow} +\infty. \end{aligned}

Consequently, there exists a constant C_{\varepsilon_*}>0 such that

\begin{aligned} \forall 0<\rho<C_{\varepsilon_*},\quad \gamma_{E,\mu_1}(\rho) \geq \rho^{\lambda_1+\varepsilon_*}. \end{aligned}

We remark that C_{\varepsilon_*} is a local constant, depending on the inclusion D, the edge E, the vertex \bf v, and the parameter \varepsilon_*.

We also remark that the function \delta is equivalent to the distance between the point under consideration and the vertex \bf{v}. Then there exist constants C_\pm>0, which depends only on the geometry and thus considered as an a priori parameter, such that for all z\in (0,h_z),

\begin{aligned} C_- (\rho+z) \leq &\delta(z)\leq C_+ (\rho+z). \end{aligned}

We assume hereafter that

\begin{aligned} \rho+h_z<C_{\varepsilon_*}, \end{aligned}

Hence,

\begin{aligned} \left|\int_0^{h_z} \gamma_{E,\mu_1}(z) \delta(z)^{-\mu_1} dz \right| \geq \frac{C_+^{-\mu_1}}{\lambda_1+\varepsilon_*-\mu_1+1} \left((\rho+h_z)^{\lambda_1+\varepsilon_*-\mu_1+1}-\rho^{\lambda_1+\varepsilon_*-\mu_1+1}\right). \end{aligned}

We may choose \varepsilon_* smaller than a certain a priori constant, so that

\begin{aligned} |I_{F,2,\infty}|=\left| \int_{F_\pm^\infty} u_0 \partial_\nu v_1 d\sigma \right| \gtrsim h_z \rho^{\lambda_1+\varepsilon_*-\mu_1} \tau^{-\mu_1}. \end{aligned}

Next, we consider the integral I_{F,2,c} on the complementary faces. For s,x >0, we introduce the incomplete Gamma function

\begin{aligned} \Gamma(s,x)=\int_x^{+\infty} t^{s-1} e^{-t} dt. \end{aligned}

Using the inequality \displaystyle e^{-t}\leq e^{-t/2}e^{-x/2} for t\geq x, we deduce that

\begin{aligned} \Gamma(s,x)\leq 2^s \Gamma(x) e^{-x/2}. \end{aligned}

Then, by a simple change of variables, we obtain

\begin{aligned} \left|\int_h^{+\infty} r^{\mu_1-1} e^{z_a^\pm r} dr \right|& \leq \int_h^{+\infty} r^{\mu_1-1} e^{-\tau \widetilde{a} r} dr\leq (\tau \widetilde{a})^{-\mu_1} \Gamma(\mu_1, \tau \widetilde{a} h)\\ &\leq (\tau \widetilde{a})^{-\mu_1} 2^{\mu_1} \Gamma(\mu_1) e^{-\frac{\tau \widetilde{a} h}{2}}. \end{aligned}

Applying the estimates of singular coefficients (?), we deduce that

\begin{aligned} \left|\int_0^{h_z} \gamma_{E,\mu_1}\delta(z)^{-\mu_1}dz \right|&\leq C_+^{-\mu_1}\|\gamma_{E,\mu_1}\|_{L^2(E)} \frac{(\rho+h_z)^{1-\mu_1}-\rho^{1-\mu_1}}{1-\mu_1}\\ &\leq C_+^{-\mu_1}C\|u\|_{H^1(\Omega)}\frac{(\rho+h_z)^{1-\mu_1}}{1-\mu_1}. \end{aligned}

Consequently,

\begin{aligned} |I_{F,2,c}|&\leq 2\mu_1 (\tau \widetilde{a})^{-\mu_1} 2^{\mu_1} \Gamma(\mu_1) e^{-\frac{\tau \widetilde{a} h}{2}} C_+^{-\mu_1}C\|u\|_{H^1(\Omega)}\frac{(\rho+h_z)^{1-\mu_1}}{1-\mu_1}\\ &\lesssim \tau^{-\mu_1}(\rho+h_z)^{1-\mu_1}e^{-\frac{\tau \widetilde{a} h}{2}}. \end{aligned}

Applying lemma:K-gamma and the estimates of singular coefficients (?), we obtain

\begin{aligned} |I_{F,3}|&\leq \int_0^{h_z}\int_0^h |\mathcal{K}[\gamma_{E,\mu_1}](r,z)-\gamma_{E,\mu_1}(z)| r^{\mu_1-1} e^{-\tau \widetilde{a} r} dr dz \\ & \leq C'_{\phi,\mu_1} \|\gamma_{E,\mu_1}\|_{V_{-s}^{s-\mu_1}(E)}\rho^{-(s-\mu_1-1/2)} \int_0^{h_z}\int_0^h r^{s-\mu_1-1/2+\mu_1-1} e^{-\tau \widetilde{a} r} dr dz \\ &\lesssim h_z \rho^{-(s-\mu_1-1/2)} \int_0^{+\infty} r^{s-3/2} e^{-\tau\widetilde{a} r} dr \\ & \lesssim h_z \rho^{-(s-\mu_1-1/2)} (\tau\widetilde{a})^{-(s-1/2)}\Gamma(s-1/2)\lesssim h_z \rho^{-(s-\mu_1-1/2)} \tau^{-(s-1/2)}. \end{aligned}

The estimates of I_{F,1}, I_{F,4}, and I_{F,5} are relatively straightforward. Consider a point x=(r,\pm a/2, z)\in F_\pm.

  • It readily follows from the vertex singular functions and the estimates of singular coefficients that

    \begin{aligned} |\nabla u_{\bf v}(x)| & \lesssim \rho^{\lambda_1-1}. \end{aligned}

    We remark that \lambda_1>1.

  • For the remaining edge singular functions, we apply lemma:esti-grad-edge and (?) to obtain

    \begin{aligned} |\nabla u_{E,2}(x)| & \lesssim \frac{r^{\widetilde{\mu_2}-1}}{\rho^{\widetilde{\mu_2}}}, \end{aligned}

    where \widetilde{\mu_2}=\mu_2 if s-\mu_2\geq 1/2 and \widetilde{\mu_2}=s-1/2 if s-\mu_2< 1/2.

  • Combining the Sobolev embedding H^{3/2}\hookrightarrow L^\infty in 2.4^3 with (?), we deduce that

    \begin{aligned} \|\nabla u_{\rm reg}\|_{L^\infty(\Omega)} \leq C \|u_{\rm reg}\|_{H^{5/2}(\Omega)} \lesssim 1. \end{aligned}

Substituting these estimates into the integrals, we obtain

\begin{aligned} |I_{F,1}| &\lesssim \rho^{\lambda_1-1} \int_0^{h_z}\int_0^h e^{-\tau\widetilde{a} r} dr dz \lesssim \rho^{\lambda_1-1}h_z \frac{1-e^{-\tau\widetilde{a} h}}{\tau \widetilde{a}} \lesssim \rho^{\lambda_1-1}h_z \tau^{-1} ,\\ |I_{F,4}| &\lesssim \rho^{-\widetilde{\mu_2}} \int_0^{h_z}\int_0^h r^{\widetilde{\mu_2}-1} e^{-\tau\widetilde{a} r} dr dz\lesssim \rho^{-\widetilde{\mu_2}} h_z \int_0^{+\infty} r^{\widetilde{\mu_2}-1} e^{-\tau\widetilde{a} r} dr \\ &\lesssim \rho^{-\widetilde{\mu_2}}h_z (\tau\widetilde{a})^{-\widetilde{\mu_2}}\Gamma(\widetilde{\mu_2})\lesssim \rho^{-\widetilde{\mu_2}}h_z \tau^{-\widetilde{\mu_2}},\\ |I_{F,5}| &\lesssim h_z \frac{1-e^{-\tau\widetilde{a} h}}{\tau \widetilde{a}} \lesssim h_z \tau^{-1}. \end{aligned}

The claim follows by collecting the estimates (?), (?), (?), (?), (?) and (?).

Proposition

Proposition 4.3

T2FF

For \tau>0 large enough, the following estimate holds,

\begin{aligned} &\left|\int_{S^e \cup S^i\cup A_{\pm}} \left(u_0\partial_\nu (u-u')-(u-u')\partial_\nu u_0\right)d\sigma\right|\\ &\lesssim h^{-1}\tau^{-2}+\rho^{-\mu_1}\tau^{-(\mu_1+1)}+h_z e^{-\tau \widetilde{b} h} (h^{\mu_1}\rho^{-\mu_1}+1+h\tau)\\ &\quad +h h_z \left(\|\nabla(u-u')\|_{L^\infty(S^e)}+\tau\|u-u'\|_{L^\infty(S^e)}\right). \end{aligned}
Proof of Proposition 4.3

First, we estimate the integrals on the upper and lower faces A_\pm. The normal vectors on these faces are parallel to the z-axis. Therefore, \partial_\nu u_0=\partial_z u_0=0 on these regions, and hence

\begin{aligned} \int_{A_\pm} (u-u')\partial_\nu u_0 d\sigma =0. \end{aligned}

We now estimate \partial_z u and \partial_z u' on A_\pm. Since the region \widetilde{\Omega} is at distance at least h from the polyhedron D', interior elliptic regularity implies that u' is analytic there. Applying the gradient estimate (e.g. Theorem 3.9 in [Gilbarg-Trudinger]), we have

\begin{aligned} \|\nabla u'\|_{L^\infty(\widetilde{\Omega})} \leq C h^{-1} \|u'\|_{L^\infty(\Omega)}\lesssim h^{-1}. \end{aligned}

For \nabla u, we must account for singularities. We decompose u into singular and regular parts by th:SingDecom, then estimate each term.

\begin{aligned} u=u_{\bf v}+u_{E} +u_{\rm reg}. \end{aligned}

Since singular exponents are ordered increasingly, it is sufficient to consider the strongest singularity. For vertex singular functions, it is straightforward that \displaystyle |\partial_z u_{\bf v}(x)| \leq C \rho^{\lambda_1-1}. For edge singular functions, combining lemma:esti-grad-edge with (?), we have

\begin{aligned} |\partial_z u_{E}(x)| \leq C \frac{\widetilde{r}^{\mu_1-1}}{\delta(z)}\lesssim \rho^{-\mu_1}r^{\mu_1-1}. \end{aligned}

For the regular part, the estimate comes directly from (?). Combining those estimates, we have,

\begin{aligned} \|\partial_z (u-u')\|_{L^\infty(\widetilde{\Omega})} \lesssim (h^{-1} +\rho^{-\mu_1}r^{\mu_1-1}). \end{aligned}

Calculating the integral over A_\pm^i,

\begin{aligned} \int_{A_\pm^i}|u_0(x)| r^{\mu_1-1}d\sigma &= \int_{-a/2}^{a/2}\int_0^h e^{-\tau r\cos \theta} r^{\mu_1} dr d\theta \leq \int_0^h\int_{-a/2}^{a/2} e^{-\tau \widetilde{a} r} r^{\mu_1} dr d\theta \\ &\leq a \int_0^{+\infty} e^{-\tau \widetilde{a} r} r^{\mu_1} dr \leq C \tau ^{-(\mu_1+1)}\Gamma(\mu_1+1). \end{aligned}

So,

\begin{aligned} \left|\int_{A_\pm^i} u_0 \partial_\nu(u-u')d\sigma \right|\lesssim (h^{-1}\tau^{-2}+\rho^{-\mu_1}\tau^{-(\mu_1+1)}). \end{aligned}

To estimate the integrals on A_\pm^e, we split A_\pm^e into two parts. Denote by B_{\tau} the two-dimensional disk on \Pi, centered at O with radius L\tau^{-1}. Here, L>1 is chosen such that \arcsin(1/L)<(\pi-a)/4. With this construction, we deduce that

\begin{aligned} (r,\theta)\in A_-^e\setminus B_{\tau} \Rightarrow L/\tau <r< h \;\; {\rm and}\;\; \theta\in (a/2,(\pi+a)/4)\cup (-(\pi+a)/4,-a/2). \end{aligned}

Introducing \widetilde{\beta}:=\cos(\frac{\pi+a}{4})>0, we have u_0(r,\theta,z)\leq e^{-\tau \widetilde{\beta} r} on A_-^e\setminus B_{\tau}. We now compute

\begin{aligned} \int_{A_-^e\cap B_{\tau}} |u_0(x)|r^{\mu_1-1} d\sigma &\leq e\int_{B_{\tau}} r^{\mu_1-1} d\sigma= 2\pi e \int_0^{L\tau^{-1}} r^{\mu_1} dr = \frac{2\pi e}{\mu_1+1} (L\tau^{-1})^{\mu_1+1}\\ &\leq C\tau^{-(\mu_1+1)},\\ \int_{A_-^e\setminus B_{\tau}} |u_0(x)|r^{\mu_1-1} d\sigma &\leq 2\int_{a/2}^{(\pi+a)/4} \int_{L\tau^{-1}}^h r^{\mu_1} e^{-\tau \widetilde{\beta} r} dr d\theta \leq C \int_{0}^{+\infty} r^{\mu_1} e^{-\tau \widetilde{\beta} r} dr \\ &\leq C \tau^{-(\mu_1+1)}\Gamma(\mu_1+1). \end{aligned}

Hence, it follows that

\begin{aligned} \left|\int_{A_\pm^e} u_0 \partial_\nu (u-u')d\sigma\right|\lesssim h^{-1}\tau^{-2}+\rho^{-\mu_1}\tau^{-(\mu_1+1)}. \end{aligned}

Next, we estimate the integrals on the inner lateral face S^i. One readily checks that |u_0(x)|\leq e^{-\tau\widetilde{b}h} on S^i, with \widetilde{b}=\cos((\pi+a)/4)>0, once the following condition is satisfied:

\begin{aligned} \tau\geq \tau_0=\frac{1}{h\sin((\pi-a)/4)}. \end{aligned}

We assume this condition throughout the remainder of the proof. Then,

\begin{aligned} &\left|\int_{S^i} u_0 \partial_\nu (u-u')-(u-u')\partial_\nu u_0 d\sigma\right| \leq \left(\|\partial_\nu(u-u')\|_{L^\infty(S^i)} + \tau\|u-u'\|_{L^\infty(S^i)}\right)|S^i| e^{-\tau \widetilde{b} h}. \end{aligned}

The quantity \|u-u'\|_{L^\infty(S^i)} can be estimated by Hölder norms that depend only on a priori data. For \|\partial_\nu(u-u')\|_{L^\infty(S^i)}, we apply lemma:esti-grad-edge to obtain \|\partial_\nu u\|_{L^\infty(S^i)}\lesssim h^{\mu_1-1}\rho^{-\mu_1}, and we use (Proof) for \nabla u'. Consequently, we obtain the following estimate for the integrals on S^i:

\begin{aligned} \left|\int_{S^i} u_0 \partial_\nu (u-u')-(u-u')\partial_\nu u_0 d\sigma\right|\lesssim hh_z e^{-\tau \widetilde{b} h} (h^{\mu_1-1}\rho^{-\mu_1}+h^{-1}+\tau). \end{aligned}

Finally, we estimate the integrals on the lateral face S^e. Since we use propagation of smallness to estimate u-u' and \nabla(u-u') on S^e, the following bounds are obtained by straightforward calculations:

\begin{aligned} &\left|\int_{S^e} u_0 \partial_\nu (u-u')- (u-u')\partial_\nu u_0 d\sigma\right| \\ &\leq |S^e|\left(\|u_0\|_{L^\infty(S^e)}\|\nabla (u-u')\|_{L^\infty(S^e)} +\|\nabla u_0\|_{L^\infty(S^e)}\|(u-u')\|_{L^\infty(S^e)}\right) \\ &\lesssim h h_z \left(\|\nabla(u-u')\|_{L^\infty(S^e)}+\tau\|u-u'\|_{L^\infty(S^e)}\right). \end{aligned}

The proof is complete by collecting the results (?), (?), (?), and (?).

Proof

Proof (Proof of th:main)

SSN1

We apply the integral identity (4.1) with the CGO solution u_0 constructed in (?). The left-hand side of (4.1) is estimated using prop:lowerbound, and the right-hand side is estimated using prop:upperbound. Collecting these estimates, we obtain

\begin{aligned} \rho^{\lambda_1+\varepsilon_*-\mu_1} h_z \tau^{-\mu_1} &\lesssim \rho^{\lambda_1-1}h_z\tau^{-1} + (\rho+h_z)^{1-\mu_1}\tau^{-\mu_1} e^{-\frac{\tau \widetilde{a} h}{2}}\\ & \quad + \rho^{-(s-\mu_1-1/2)}h_z \tau^{-(s-1/2)} + \rho^{-\widetilde{\mu_2}}h_z\tau^{-\widetilde{\mu_2}} + h_z\tau^{-1}\\ & \quad + h^{-1}\tau^{-2}+\rho^{-\mu_1}\tau^{-(\mu_1+1)}+h_z e^{-\tau \widetilde{b} h} (h^{\mu_1}\rho^{-\mu_1}+1+h\tau)\\ &\quad +h h_z \left(\|\nabla(u-u')\|_{L^\infty(S^e)}+\tau\|u-u'\|_{L^\infty(S^e)}\right). \end{aligned}

To estimate \|u-u'\|_{L^\infty(S^e)} and \|\nabla(u-u')\|_{L^\infty(S^e)}, we apply propagation of smallness from prop:propa-small. Introduce the notation

\begin{aligned} \zeta(\varepsilon):=(\ln |\ln \varepsilon|)^{-\alpha}. \end{aligned}

Noting that the common Hölder norm T depends only on a priori data and that the surface S^e has a distance 1/\tau to D, we have,

\begin{aligned} \|u-u'\|_{L^\infty(S^e)} &\lesssim \zeta(\varepsilon),\\ \|\nabla(u-u')\|_{L^\infty(S^e)} &\lesssim \tau \zeta(\varepsilon). \end{aligned}

Recall that \rho,h, and h_z must satisfy the constraints (?),(?) and (?), also that D' is at distance at least h from \widetilde{\Omega}. Thus, we choose

\begin{aligned} h_z=h,\quad \rho=\frac{h}{\ell},\quad h=\min\left\lbrace C_{\varepsilon_*}\left(1+\frac{1}{\ell}\right)^{-1} ,\frac{\ell^2}{2\ell+1},\;\frac{1}{2}d_H(D,D')\right\rbrace. \end{aligned}

Then the estimate becomes

\begin{aligned} h^{\lambda_1+\varepsilon_*-\mu_1+1} \tau^{-\mu_1} &\lesssim h^{\lambda_1}\tau^{-1} + h^{1-\mu_1}\tau^{-\mu_1} e^{-\frac{\tau \widetilde{a} h}{2}} + h^{3/2-s+\mu_1} \tau^{-(s-1/2)} + h^{1-\widetilde{\mu_2}}\tau^{-\widetilde{\mu_2}} + h\tau^{-1}\\ & \quad + h^{-1}\tau^{-2}+h^{-\mu_1}\tau^{-\mu_1-1}+h e^{-\tau \widetilde{b} h} (1+h\tau) +h^2 \tau \zeta(\varepsilon). \end{aligned}

Next, we use the usual inequalities e^{-x}\leq x^{-1} and e^{-x}\leq x^{-2} for all x>0 to deduce that

\begin{aligned} h^{\lambda_1+\varepsilon_*-\mu_1+1} \tau^{-\mu_1} &\lesssim h^{\lambda_1}\tau^{-1} + h^{-\mu_1}\tau^{-\mu_1-1} + h^{3/2-s+\mu_1} \tau^{-(s-1/2)} + h^{1-\widetilde{\mu_2}}\tau^{-\widetilde{\mu_2}} + h\tau^{-1}\\ & \quad + h^{-1}\tau^{-2}+h^{-\mu_1}\tau^{-\mu_1-1}+\tau^{-1} +h^2 \tau \zeta(\varepsilon). \end{aligned}

Recall the estimates on singular exponents: 0<\lambda_1<s-1/2, 0<\mu_1<1, and 3/2\leq s <\mu_1+1. Combining these with \tau\geq 1 and h\leq 1, we deduce that

\begin{aligned} h^{\lambda_1+\varepsilon_*-\mu_1+1} \tau^{-\mu_1} &\lesssim h^{-1}\tau^{-1}+ h^{1-\widetilde{\mu_2}}\tau^{-\widetilde{\mu_2}}+h^2 \tau \zeta(\varepsilon). \end{aligned}

Recalling that s=3/2+\epsilon with 0<\epsilon<\mu_1-1/2 and using the definition of \widetilde{\mu_2}, we distinguish the following two cases.

  • If s-\mu_2\geq 1/2, then -\widetilde{\mu_2}=-\mu_2 and 1-\widetilde{\mu_2}=1-\mu_2\geq 3/2-s=\epsilon>-1/2.

  • If s-\mu_2< 1/2, then -\widetilde{\mu_2}=1/2-s=-1-\epsilon <-1 and 1-\widetilde{\mu_2}=-\epsilon >-1/2.

Taking into account that \mu_2>1, the estimate becomes

\begin{aligned} h^{\lambda_1+\varepsilon_*-\mu_1+1} \tau^{-\mu_1} &\lesssim h^{-1}\tau^{-1}+h^2 \tau \zeta(\varepsilon). \end{aligned}

By minimizing the right-hand side with respect to \tau, we choose

\begin{aligned} \tau=\tau_e:=\left(h^3\zeta(\varepsilon)\right)^{-\frac{1}{2}}. \end{aligned}

For \varepsilon sufficiently small, one can verify that \tau_e\geq \tau_0, where \tau_0 is given in (?).

Substituting this choice into the estimate and solving for h, we obtain

\begin{aligned} h \lesssim \zeta(\varepsilon)^{\frac{1}{2(\lambda_1+\varepsilon_*)+\mu_1+1}}\leq \zeta(\varepsilon)^{\frac{1}{3\lambda_1+\mu_1+1}}, \end{aligned}

where the last inequality holds by choosing \varepsilon_*<\lambda_1/2.

Finally, as \displaystyle \zeta(\varepsilon)\underset{{\varepsilon}\rightarrow {0}}{\longrightarrow} 0, we can deduce that for \varepsilon small enough, the right-hand side above is less than both C_{\varepsilon_*}\left(1+\frac{1}{\ell}\right)^{-1} and \frac{\ell^2}{2\ell+1}. Hence, the stability estimate (1.2) follows with the exponent \kappa given by

\begin{aligned} \kappa:= \frac{-\alpha }{3\lambda_1+\mu_1+1}. \end{aligned}

The exponent \kappa is bounded below by a constant depending only on a priori data, using \alpha=\min\{\alpha_0,\epsilon,\lambda_1,\mu_1\} from prop:propa-small and the bounds \mu_1\in (\frac{1}{2},1), \lambda_1\in (1,\sqrt{2k}).

The proof is complete.

Remark

Remark 5.1

HGSP

The assumption 4.) in assumption:singularity can be relaxed to certain order k\in\mathbb{N} such that the interaction coefficient a_{E,\nu}^{{\bf v},\lambda_k} is non-zero from this rank. In that setting, it is sufficient to replace \lambda_1 by \lambda_k accordingly in the proofs in this paper. The stability remains valid with a weaker exponent \kappa in (1.2).

Remark

Remark 5.2

93IO

By letting \varepsilon\rightarrow 0, we can deduce the uniqueness result for the inverse conductivity problem for convex polyhedral inclusions. As a significant improvement over the previous result [Barcelo1994], the assumption on the boundary data can be relaxed to a generic condition as described here.

Lemma

Lemma A.1

SWMH

Let \frac{1}{2}<\mu<s and let u_{E,\mu} be given by (?) with coefficient \gamma\in V_{-s}^{s-\mu}(E). Then, in the region \widetilde{\Omega}, there exists a constant C_{\phi,\mu}>0 such that

\begin{aligned} |\nabla u_{E,\mu}(r,\theta,z)| \leq C_{\phi,\mu} \|\gamma\|_{V_{-s}^{s-\mu}(E)}\frac{r^{\widetilde{\mu}-1}}{\rho^{\widetilde{\mu}}}. \end{aligned}

Here, \rho denotes the distance from the point (r,\theta,z) to its closest endpoint of E. Also, \widetilde{\mu}=\mu if s-\mu\geq 1/2 and \widetilde{\mu}=s-1/2 if s-\mu<1/2.

Proof of Lemma A.1

By a change of variables stretching E into 2.4 and the embedding (?), we have \widetilde{\gamma}\in H^{s-\mu}(2.4). We compute \partial_z u_{E,\mu} via the chain rule:

\begin{aligned} \frac{\partial}{\partial z}u_{E,\mu}(r,\theta,z)&= \left(\frac{\partial \widetilde{z}}{\partial z}\frac{\partial}{\partial \widetilde{z}}+\frac{\partial \widetilde{r}}{\partial z}\frac{\partial}{\partial \widetilde{r}}\right) u_{E,\mu}(r,\theta,z)\\ &= \frac{1}{\delta(z)}\left(\frac{\partial}{\partial \widetilde{z}}\mathcal{K}[\gamma]\right)\widetilde{r}^{\mu}\varphi(\theta)-\frac{\delta'(z)}{\delta(z)} \left(\widetilde{r}^{\mu+1} \frac{\partial}{\partial \widetilde{r}}\mathcal{K}[\gamma]+\mu \widetilde{r}^{\mu}\mathcal{K}[\gamma]\right) \varphi(\theta)\\ \frac{\partial}{\partial r} u_{E,\mu}(r,\theta,z) &= \frac{\partial \widetilde{r}}{\partial r}\frac{\partial}{\partial \widetilde{r}} u_{E,\mu}(r,\theta,z)=\frac{1}{\delta(z)}\left(\frac{\partial}{\partial \widetilde{r}}\mathcal{K}[\gamma]\right)\widetilde{r}^{\mu}\varphi(\theta)+\frac{\mu}{\delta(z)} \mathcal{K}[\gamma]\widetilde{r}^{\mu-1}\varphi(\theta),\\ \frac{1}{r} \frac{\partial}{\partial \theta} u_{E,\mu}(r,\theta,z) &= \frac{1}{\delta(z)}\mathcal{K}[\gamma]\widetilde{r}^{\mu-1}\varphi'(\theta). \end{aligned}

By the convolution structure of \mathcal{K}, we have:

\begin{aligned} \frac{\partial}{\partial \widetilde{z}}\mathcal{K}[\gamma](\widetilde{r},z)& = \frac{\partial}{\partial \widetilde{z}}\int_2.4 \frac{1}{\widetilde{r}}\phi(\frac{\widetilde{z}-t}{\widetilde{r}})\widetilde{\gamma}(t)dt=\frac{1}{\widetilde{r}}\int_2.4 \frac{1}{\widetilde{r}}\phi'(\frac{t}{\widetilde{r}})\widetilde{\gamma}(\widetilde{z}-t) dt,\\ \frac{\partial}{\partial \widetilde{r}}\mathcal{K}[\gamma](\widetilde{r},z)&= -\int_2.4 \left(\frac{1}{\widetilde{r}^2}\phi(\frac{t}{\widetilde{r}})+\frac{t}{\widetilde{r}^3}\phi'(\frac{t}{\widetilde{r}})\right)\widetilde{\gamma}(\widetilde{z}-t)dt=-\frac{1}{\widetilde{r}}\int_2.4 \frac{1}{\widetilde{r}}(t\phi)'(\frac{t}{\widetilde{r}})\gamma(\widetilde{z}-t)dt. \end{aligned}

Note that \phi',\;(t\phi)'\in \mathcal{S}(2.4). We apply the Sobolev embedding H^{s-\mu}(2.4)\hookrightarrow L^q(2.4) with \frac{1}{q}=\frac{1}{2}-(s-\mu) if s-\mu<1/2 and q=\infty if s-\mu\geq 1/2.

  • Case s-\mu <1/2. Denoting by q' the conjugate exponent of q, it follows from Hölder’s inequality that

    \begin{aligned} \left|\mathcal{K}[\gamma](\widetilde{r},z)\right| &\leq \widetilde{r}^{-1/q} \|\phi\|_{L^{q'}(2.4)} \|\widetilde{\gamma}\|_{L^q(2.4)}=\widetilde{r}^{(s-\mu)-1/2} \|\phi\|_{L^{q'}(2.4)} \|\widetilde{\gamma}\|_{L^q(2.4)},\\ \left|\frac{\partial}{\partial \widetilde{z}}\mathcal{K}[\gamma](\widetilde{r},z)\right| &\leq \widetilde{r}^{(s-\mu)-3/2} \|\phi'\|_{L^{q'}(2.4)} \|\widetilde{\gamma}\|_{L^q(2.4)},\\ \left|\frac{\partial}{\partial \widetilde{r}}\mathcal{K}[\gamma](\widetilde{r},z)\right| &\leq \widetilde{r}^{(s-\mu)-3/2} \|(t\phi)'\|_{L^{q'}(2.4)} \|\widetilde{\gamma}\|_{L^q(2.4)}. \end{aligned}
  • Case s-\mu\geq 1/2. Using Young’s inequality, we have

    \begin{aligned} |\mathcal{K}[\gamma](\widetilde{r},z)| &\leq \|\phi\|_{L^1(2.4)} \|\widetilde{\gamma}\|_{L^\infty(2.4)},\\ \left|\frac{\partial}{\partial \widetilde{z}}\mathcal{K}[\gamma](\widetilde{r},z)\right| &\leq \widetilde{r}^{-1} \|\phi'\|_{L^1(2.4)} \|\widetilde{\gamma}\|_{L^\infty(2.4)},\\ \left|\frac{\partial}{\partial \widetilde{r}}\mathcal{K}[\gamma](\widetilde{r},z)\right| &\leq \widetilde{r}^{-1} \|(t\phi)'\|_{L^1(2.4)} \|\widetilde{\gamma}\|_{L^\infty(2.4)}. \end{aligned}

Define \widetilde{\mu}:=\mu if s-\mu\geq 1/2 and \widetilde{\mu}=s-1/2 if s-\mu<1/2. Substituting these estimates into the expressions for derivatives of u_{E,\mu}, we obtain

\begin{aligned} \left|\frac{\partial}{\partial z} u_{E,\mu}(r,\theta,z)\right| &\leq \|\widetilde{\gamma}\|_{L^q(2.4)} \left(\frac{\widetilde{r}^{\widetilde{\mu}-1}}{\delta(z)} \|\phi'\|_{L^{q'}(2.4)} + \frac{|\delta'(z)| \widetilde{r}^{\widetilde{\mu}}}{\delta(z)}(\|(t\phi)'\|_{L^{q'}(2.4)}+\mu \|\phi\|_{L^{q'}(2.4)})\right). \end{aligned}

The same estimate holds for \frac{\partial}{\partial r} u_{E,\mu}.

Note that the cut-off function \eta and its support \ell satisfy \ell \leq 1/2 and \eta(\widetilde{r})=0 if \widetilde{r}\geq 2\ell. Hence we may assume \widetilde{r}\leq 1. Also, \delta is equivalent to the distance to the endpoints of E, which implies \delta(z)\geq C_{\delta}\rho for some constant C_\delta>0. Then the estimates become

\begin{aligned} &\left|\frac{\partial}{\partial z} u_{E,\mu}(r,\theta,z)\right|,\; \left|\frac{\partial}{\partial r} u_{E,\mu}(r,\theta,z)\right|\\ &\leq \|\widetilde{\gamma}\|_{L^q(2.4)} C_\delta^{-\widetilde{\mu}}\left( \|\phi'\|_{L^{q'}(2.4)} + |\delta'(z)|(\|(t\phi)'\|_{L^{q'}(2.4)}+\mu \|\phi\|_{L^{q'}(2.4)})\right)\frac{r^{\widetilde{\mu}-1}}{\rho^{\widetilde{\mu}}},\\ & \frac{1}{r} \left|\frac{\partial}{\partial \theta} u_{E,\mu}(r,\theta,z)\right| \leq \mu\|\widetilde{\gamma}\|_{L^q(2.4)} C_\delta^{-\widetilde{\mu}} \|\phi\|_{L^{q'}(2.4)}\frac{r^{\widetilde{\mu}-1}}{\rho^{\widetilde{\mu}}}. \end{aligned}

Thus, the claim follows by combining these estimates with (?).

Lemma

Lemma A.2

SKB1

Let 1/2<\mu and let \mu+1/2<s<\mu+1. Then, there exists a constant C'_{\phi,\mu}>0 such that

\begin{aligned} \forall z\in E,\; \widetilde{r}>0,\quad |\mathcal{K}[\gamma](\widetilde{r},z)-\gamma(z)|\leq C'_{\phi,\mu} \widetilde{r}^{s-\mu-1/2}\|\gamma\|_{V_{-s}^{s-\mu}(E)}. \end{aligned}
Proof of Lemma A.2

Using the fact \displaystyle \int_2.4 \phi(t)dt=1, we have

\begin{aligned} \mathcal{K}[\gamma](\widetilde{r},z)-\widetilde{\gamma}(\widetilde{z})=\int_2.4 \frac{1}{\widetilde{r}} \phi(\frac{t}{\widetilde{r}})(\widetilde{\gamma}(\widetilde{z}-t)-\widetilde{\gamma}(\widetilde{z})) dt. \end{aligned}

We split this integral into two parts, one for |t|\leq \widetilde{r} and the other for |t|>\widetilde{r}.

Since s-\mu>1/2 and \widetilde{g}\in H^{s-\mu}(2.4), it follows from Sobolev embedding that \widetilde{\gamma}\in \mathcal{C}^{s-\mu-1/2}(2.4), and for some constant C>0 that

\begin{aligned} \forall x,y\in2.4,\quad |\widetilde{\gamma}(x)-\widetilde{\gamma}(y)| &\leq C \|\widetilde{\gamma}\|_{H^{s-\mu}(2.4)}|x-y|^{s-\mu-1/2}. \end{aligned}

Then, we can estimate the first part of the integral as follows,

\begin{aligned} \left|\int_{|t|\leq \widetilde{r}} \frac{1}{\widetilde{r}} \phi\left(\frac{t}{\widetilde{r}}\right) (\widetilde{\gamma}(\widetilde{z}-t)-\widetilde{\gamma}(\widetilde{z})) dt \right| &\leq C\|\widetilde{\gamma}\|_{H^{s-\mu}(2.4)}\widetilde{r}^{s-\mu-1/2}\|\phi\|_{L^1(2.4)}. \end{aligned}

For the second part, since \phi is rapidly decreasing, there exists C'>0 such that \displaystyle |x^2\phi(x)| \leq C'. Then, we can estimate the second part of the integral as follows,

\begin{aligned} \left|\int_{|t|>\widetilde{r}} \frac{1}{\widetilde{r}} \phi\left(\frac{t}{\widetilde{r}}\right) (\widetilde{\gamma}(\widetilde{z}-t)-\widetilde{\gamma}(\widetilde{z})) dt \right| &\leq 2C'\|\widetilde{\gamma}\|_{L^\infty(2.4)}\int_{\widetilde{r}}^{+\infty} \frac{\widetilde{r}^2}{t^2} dt \leq 2C'\|\widetilde{\gamma}\|_{L^\infty(2.4)}\widetilde{r}. \end{aligned}

The claim follows by combining these estimates with the embedding (?).

Source and licence

The text on this page is from arXiv:2605.17484v2 by Chun-Hsiang Tsou, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2605/2605.17484.jsonl.