Stability of L p L^{p} Dirichlet solvability under small bi-Lipschitz transformations of domains Thanks: J. Zhuge is partially supported by NNSF of China (No. 12494541, 12288201, 12471115). Joseph Feneuil, Linhan Li, Jinping Zhuge Abstract . We show that small bi-Lipschitz deformations of a Lipschitz domain (with possibly large Lipschitz constant) preserve the solvability of the Dirichlet problem for the Laplacian with boundary data in L p L^{p} , for the same value of p > 1 p>1 . As a consequence, for all p ∈ ( 1 , ∞ ) p\in(1,\infty) , we obtain the solvability of the L p L^{p} Dirichlet problem for small Lipschitz perturbations of convex domains, thereby unifying two fundamentally different settings in which such results were previously known: convex and C 1 C^{1} domains. The key ingredient and novelty of our approach is a construction of a change of variables based on a non-constant basis derived from the Green function, which encodes the geometry of the base domain. Keywords: L^{p} Dirichlet problem, Laplacian, bi-Lipschitz change of variables, perturbations, Green function, Lipschitz domains, convex domains, quasiconvex domains. 1. Introduction 1.1. Motivation and statement of main results We consider the Dirichlet problem for the Laplacian in a domain \Omega\subset\mathbb{R}^{n} . The solvability of the Dirichlet problem with boundary data in L^{p} —or (D)_{p} for simplicity, see Definition 2.8 —is closely tied to the geometry of the underlying domain and its boundary. A substantial body of recent work (see e.g. [ HM14 , HMUT14 , AHM+20 ] ) has culminated in identifying the necessary and sufficient geometric conditions for solving the L^{p} Dirichlet problem for some (large) p>1 . Roughly speaking, the right geometric condition is some weak connectivity of the domain in addition to uniform rectifiability of the boundary of the domain, the latter being a strictly more general notion than a union of Lipschitz graphs. In contrast, the right geometric notion for the solvability of the L^{p} Dirichlet problem for all p\in(1,\infty) is far from clear. It remains an open question: Question 1.1 . What are the necessary and sufficient conditions on the domain for solving the L^{p} Dirichlet problem (D)_{p} for all p>1 ? Noticeably, there are two fundamentally different types of domains on which the L^{p} Dirichlet problem for the Laplacian is solvable for all p\in(1,\infty) : convex domains, and domains with “very flat” boundaries. Let us be more precise. For convex domains, as one has a pointwise bound on the gradient of the Green function up to the boundary, one can obtain (D)_{p} for the full range of p using the Hardy-Littlewood maximal function (this is observed in [ She06 ] ). Combining this idea with Schauder theory, we can extend the result slightly beyond convex domains, namely, to domains whose boundary is (locally) the image of a convex graph under a C^{1,\alpha} diffeomorphism (see Appendix B ). However, since the argument relies heavily on boundary C^{1} estimates of solutions, reducing the regularity of the C^{1,\alpha} diffeomorphism to only C^{1} is beyond reach of this method. On the other hand, it is known that for any p>1 , the L^{p} Dirichlet problem is solvable in Lipschitz domains with sufficiently small Lipschitz constant, and in particular, in bounded C^{1} domains. This follows from the result of [ JK82b ] , which shows that on C^{1} domains, the logarithm of the Poisson kernel ( \log k ) has vanishing mean oscillation (VMO) on the boundary. Alternatively, one can deduce this result from [ DPP07 ] , which treats a more general class of elliptic operators. Also in this direction, [ KT97 ] generalizes the result to a broader class of domains whose boundaries are not necessarily given (locally) by graphs of functions. Specifically, Kenig and Toro showed that on vanishing chord arc domains —which can be viewed as a generalization of Lipschitz graph domains with vanishing Lipschitz constant—the logarithm of the Poisson kernel is in \rm{VMO}(\partial\Omega) , which implies that the L^{p} Dirichlet problem for the Laplacian is solvable for all p\in(1,\infty) . Remarkably, there is also a converse to this result ( [ KT03 ] ): assuming that the domain is a chord-arc domain with sufficiently small constant, if \log k\in\rm{VMO}(\partial\Omega) , then the domain is a vanishing chord arc domain. We can see that “flatness” of the boundary—such as the vanishing chord-arc condition— cannot be the right geometric criterion characterizing the solvability of (D)_{p} for all p>1 . Indeed, convex domains can be Lipschitz domains with very large Lipschitz constants, and since the Lipschitz constant is invariant under dilation, the boundaries of such domains can remain far from flat at every scale—yet the L^{p} Dirichlet problem is solvable for all p>1 . In particular, this means that neither convex nor vanishing chord-arc is a necessary condition for solving the L^{p} Dirichlet problem for all p>1 . The first step to understand the right geometric condition would thus require unifying these two conditions. One of our contributions in this paper is identifying a class of domains which includes both convex domains and Lipschitz domains with vanishing Lipschitz constants, and on which the L^{p} Dirichlet problem is solvable for all p>1 (see Corollary 1.10 ). Notably, this class of domains is invariant under any C^{1} diffeomorphism. We postpone the precise definition and statement until the end of this section. Our approach is to study the following question, which can be viewed as the first step toward understanding Question 1.1 . Question 1.2 . Let p\in(1,\infty) . What kinds of domain perturbations preserve (D)_{p} for the same value of p ? While there are many perturbation results for operators that preserve the solvability of the L^{p} Dirichlet problem—see e.g. [ Dah86 , Esc96 , DPP07 , MPT13 , MPT14 , CHM19 , AHMT23 , AAA+11 , BHLG+24 ] , see also perturbation results for operators that do not preserve the same value of p [ FKP91 , CHMT20 , DSU24 , FP22 , Hof25 ] —to our knowledge there has been no perturbation theory concerning domains , or at least, it has not usually been viewed in this way. On the other hand, one can apply certain perturbation results for operators to obtain corresponding perturbation results for domains, which is likely one of the original motivations for studying perturbation theory for operators. However, even when restricted to small Lipschitz perturbations in the transversal direction of a Lipschitz graph domain, existing theory suffers from one of the following 3 limitations: • it fails to preserve the same value of p , • it does not yield results in the range 10 and a M -Lipschitz function g:\mathbb{R}^{n-1} such that \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R},\,t>g(x)\}. Let p\in(1,\infty) be such that the L^{p} Dirichlet problem for the Laplacian is solvable in \Omega_{0} . There exists \epsilon_{0}>0 such that if \Phi:\,\mathbb{R}^{n}\to\mathbb{R}^{n} is a bi-Lipschitz map whose Jacobian matrix \nabla\Phi satisfies \|\nabla\Phi-I\|_{\infty}\leq\epsilon_{0} , then the L^{p} Dirichlet problem for the Laplacian is solvable in \Omega=\Phi(\Omega_{0}) . Moreover, \epsilon_{0}=\epsilon_{0}(n,p,C_{p},M) depends only on the dimension n , the parameter p , the reverse Hölder constant C_{p} in ( 2.13 ), and the Lipschitz constant M . Some remarks are in order. Remark 1.4 . Under the assumption of Theorem 1.3 , the perturbed domain \Omega=\Phi(\Omega_{0}) takes the form of \Omega=\left\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R}:t>g(x+\psi(x))+h(x)\right\}, where \psi:\mathbb{R}^{n-1}\to\mathbb{R}^{n-1} and h:\mathbb{R}^{n-1}\to\mathbb{R} are mappings that satisfy |\nabla\psi|+|\nabla h|\lesssim\epsilon_{0} . We provide a proof of this observation in Appendix A . Remark 1.5 . In view of Remark 1.4 , the class of allowable domain perturbations includes, in particular, small Lipschitz perturbations in the transversal direction. More precisely, our result applies to any perturbation of \Omega_{0} of the form (1.6) \Omega=\left\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R}:t>g(x)+h(x)\right\}, where h:\mathbb{R}^{n-1}\to\mathbb{R} is any \epsilon_{0} -Lipschitz function. On the other hand, the class of perturbations covered by our theorem is strictly more general than the class described by ( 1.6 ), as the function g(x+\psi(x))-g(x) may not have small Lipschitz constant even if \psi does. Remark 1.7 . When restricted to p\geq 2 , Theorem 1.3 can be derived from known arguments and results. In fact, since both \Omega_{0} and the perturbed domain \Omega are Lipschitz domains, the result can be obtained by using the solvability of the L^{p} Dirichlet problem for the Laplacian in Lipschitz domains for all p\in[2,\infty) , which is established in [ Dah77 ] and [ Ver84 ] . However, none of the available methods apply to the range 11 and does not distinguish between the cases p\geq 2 and p<2 . Equipped with Theorem 1.3 , we can now introduce the class of domains that unifies convex and C^{1} domains. Roughly speaking, these domains are locally the image of a convex domain under a bi-Lipschitz map close to the identity. Specifically, we define the following. Definition 1.8 . We say that a bounded domain \Omega\subset\mathbb{R}^{n} is strongly quasiconvex if there exists M>0 such that for any \epsilon_{0}>0 , there exists r_{0}>0 such that for any x_{0}\in\partial\Omega , there exists an coordinate basis \{e_{1},\dots,e_{n}\} , an M -Lipschitz convex function g_{x_{0}} and a bi-Lipschitz map \Phi_{x_{0}}:\,\mathbb{R}^{n}\to\mathbb{R}^{n} whose Jacobian matrix \nabla\Phi_{x_{0}} satisfies \|\nabla\Phi_{x_{0}}-I\|_{\infty}\leq\epsilon_{0} such that \Omega\cap B(x_{0},r_{0})=\Phi_{x_{0}}\Big(\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R},\,t>g_{x_{0}}(x)\}\Big)\cap B(x_{0},r_{0}). It is obvious from definition that any convex domain and any Lipschitz domain with vanishing Lipschitz constant (or any bounded C^{1} domain) are strongly quasiconvex. We refer to this class of domains as strongly quasiconvex because the condition resembles that of quasiconvex domains in the literature (see, for instance, [ JLW10 , Zhu20 , ZZ23 ] ). Remark 1.9 . The class of strongly quasiconvex domains is stable under any C^{1} diffeomorphism (not necessarily with small constants). We give a sketch of the proof of this claim in Appendix A . In particular, a C^{1} diffeomorphism transforms a convex domain into a strongly quasiconvex domain. This suggests that the class of strongly quasiconvex domains is a natural setting to consider. By an appropriate localization argument (see Section 4.2 ), one can deduce the our main corollary from Theorem 1.3 . Corollary 1.10 . Let \Omega\subset\mathbb{R}^{n} be a bounded strongly quasiconvex domain. Then the L^{p} Dirichlet problem for -\Delta in \Omega is solvable for all p\in(1,\infty) . By the duality between the Dirichlet problem and the Regularity problem (see Definition 2.10 ) in Lipschitz domains [ She07 ] , the L^{p} Regularity problem is solvable whenever the L^{p^{\prime}} Dirichlet problem is solvable and the L^{q} Regularity problem is solvable for some q>1 . Since the latter is known to hold for the Laplacian in Lipschitz domains ( [ Ver84 , DK87 ] ), we obtain the same range of p for which the L^{p} Regularity problem is solvable as for the Dirichlet problem. Corollary 1.11 . Let \Omega\subset\mathbb{R}^{n} be a bounded strongly quasiconvex domain. Then the L^{p} Regularity problem for -\Delta in \Omega is solvable for all p\in(1,\infty) . 1.2. Existing methods and their limitations If one views Lipschitz domains with sufficiently small constant as a small Lipschitz perturbation of flat sets, and vanishing chord arc domains as a small perturbation of Lipschitz domains with vanishing constant, then the solvability results mentioned at the beginning of the previous section on (D)_{p} for all p>1 on these two types of domains can be interpreted as perturbation results for domains that preserve the solvability of the L^{p} Dirichlet problem, for the same value of p>1 . One feature common to these two results is that both involve small perturbations from “flat sets”. Since our base domain may have very large Lipschitz constant—and thus its boundary far from being flat—existing methods either fail to preserve the same value of p , or they do not produce the result in the range 10 depending on the Lipschitz domain. Moreover, this range is optimal, in the sense that there are counterexamples for every p<2 (see [ Ken86 ] ). Thus the method of layer potentials reveals a sharp dichotomy between the cases p\geq 2 and 1\varphi(x)\right\} is to transform it into a boundary value problem on the upper half-space \mathbb{R}^{n}_{+} for a different operator via a bi-Lipschitz change of variables (see Proposition 2.17 for a precise formulation of such an example). There are two types of elliptic operators in the form L=-\diver A\nabla arsing from two different bi-Lipschitz change of variables: (1) “ t -independent” operators, which are operators whose coefficients are independent of the transversal variable, that is, A(x,t)=A(x) ; (2) “DKP” operators, which are operators whose coefficients satisfy a Carleson condition, namely, \dist(\cdot,\partial\Omega)|\nabla A|\in\CMs (see Definition 2.1 ). The “ t -independent” operators are associated with the simple flattening (or lifting) map: \rho_{0}:\mathbb{R}^{n}_{+}\to\Omega,\quad(x,t)\mapsto(x,t+\varphi(x)), and the “DKP” operators are associated with the Nečas-Kenig-Stein map (see [ KP01 ] ): \rho_{1}:\mathbb{R}^{n}_{+}\to\Omega,\quad(x,t)\mapsto(x,ct+\eta_{t}*\varphi(x)), where \eta_{\epsilon} is a standard mollifier, and c is a constant chosen to make sure that the map is 1-1. For the map \rho_{0} , the Lipschitz constant of the boundary is reflected as the L^{\infty} norm of the pull-back operator, and so Theorem 1.3 in the special case when the perturbed domain \Omega has the form ( 1.6 ) can be reduced to the stability of the solvability of the L^{p} Dirichlet problem for symmetric, real, t -independent operators L_{0}=-\diver A_{0}\nabla and L_{1}=-\diver A_{1}\nabla in \mathbb{R}^{n}_{+} , under the perturbation \left\|A_{0}-A_{1}\right\|_{L^{\infty}}\leq\epsilon. This question is studied for p=2 in [ AAA+11 ] and for general p in [ HMM15 ] via the method of layer potentials. One of the remarkable features of both works is that their theory applies to operators with complex coefficients. When restricted to the real-coefficient setting, however, the method faces the same limitations discussed above for layer potentials. In particular, it cannot be applied, at least directly, to obtain the range 11 for DKP operators (with possibly large Carleson constant) on \mathbb{R}^{n}_{+} , and hence also on Lipschitz domains, since the DKP condition is stable under bi-Lipschitz changes of variables. Later, [ DPP07 ] showed that for any p>1 , if the Carleson constant of the coefficients of a DKP operator is sufficiently small , then (D)_{p} is solvable on \mathbb{R}^{n}_{+} , and therefore on Lipschitz domains with sufficiently small Lipschitz constant. Both works employ clever and delicate integration-by-parts arguments that depend crucially on a distinguished variable t ; as a result, one is inevitably led to the flat case \mathbb{R}^{n}_{+} . If we apply \rho_{1} to the two Lipschitz domains \Omega_{0} and \Omega , we get two DKP operators with large Carleson constants, for which we only know (D)_{p} for some large p . Although their Carleson constants are close to each other in the special case when \Omega is a small Lipschitz perturbation of \Omega_{0} in the transversal direction, they are not the appropriate Carleson perturbations to each other. In particular, the existing perturbation theory on operators do not apply in this setting. 1.3. Main ideas of our approach The takeaway from the previous discussions is that we do not want to reduce a Lipschitz domain to the flat case \mathbb{R}^{n}_{+} because the small perturbation of the domain is lost in the flattening process. The central idea—and the key novelty—of our approach is to construct a map \rho directly from the Lipschitz graph domain \Omega_{0}=\left\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R}:t>g(x)\right\} to the perturbed domain \Omega=\Phi(\Omega_{0}) , and to design \rho in such a way that the small perturbation of the domains is appropriately captured in the pull-back operator L_{\rho} so that it enjoys the same solvablity of the L^{p} Dirichlet problem as the Laplacian on \Omega_{0} . The first challenge that we face here is that we lose the special variable t (or equivalently, the special direction e_{n} ) when constructing a map directly from \Omega_{0} . Note that this issue is not due to the level of generality of the perturbation we consider: even when the perturbation occurs only in directions transversal to the boundary of \Omega_{0} , i.e., when the perturbed domain takes the form ( 1.6 ), one can no longer consider maps such as \rho_{0} or \rho_{1} mentioned in Section 1.2 by viewing points in \Omega as the image of points in \mathbb{R}^{n}_{+} . Our idea is to use the Green function with pole at infinity (see Lemma 2.5 for the definition) for the Laplacian in the base domain—denoted by G —to capture the geometry of \Omega_{0} . This is not too surprising, as the zero set of the Green function is \partial\Omega_{0} , and the level sets of G gives a family of surfaces that covers \Omega_{0} . We therefore consider the vector field \nabla G on \Omega_{0} , and build an orthonormal basis \left\{v_{1},\dots,v_{n}\right\} for any point X\in\Omega_{0} , with v_{n}=\frac{\nabla G}{|\nabla G|} . To make sure that it is possible, we show in Lemma 3.2 that \nabla G is non-degenerate on \Omega_{0} , and moreover, that \frac{G(X)}{|\nabla G(X)|} behaves like the distance from X to \partial\Omega_{0} . What is even better is that the gradient of \frac{G}{|\nabla G|} is ‘almost parallel’ to v_{n} at every X\in\Omega_{0} , in the sense that their difference satisfies a Carleson condition— see Theorem 3.10 . Although the result in Theorem 3.10 looks delicate, we remark that little additional work is required to obtain it as the difficult estimates for the second derivatives of the Green function have already been established in [ FL23 ] and [ DLM22 ] , and can also be found implicitly in [ Azz19 ] . This implies that we can think of the function t(X):=\frac{G(X)}{|\nabla G(X)|}\quad\text{for }X\in\Omega_{0} as the transversal variable t in \mathbb{R}^{n}_{+} , and hence we have an appropriate special direction for \Omega_{0} —although the direction depends on the location in \Omega_{0} . Next, we carefully design \rho so that the pull-back operator L_{\rho}=-\diver A_{\rho}\nabla from the Laplacian on \Omega under the map \rho is a small Carleson perturbation of an intermediate elliptic operator L_{0}:=-\diver A_{0}\nabla on \Omega_{0} , and that L_{0} enjoys the same solvability of the L^{p} Dirichlet problem as the Laplacian on \Omega_{0} . Since small Carleson perturbations of operators preserve solvability of the L^{p} Dirichlet problem (see Theorem 2.16 for a precise statement), this process ensures that L_{\rho} has the same Dirichlet solvability as the Laplacian on \Omega_{0} , and thus the Laplacian has the same Dirichlet solvability on \Omega_{0} and \Omega . The major work and challenge is to construct the intermediate operator L_{0} so that it is close to both the Laplacian on \Omega_{0} and the pull-back operator L_{\rho} . This construction—together with the construction of the map \rho —is carried out in Sections 3.2 and 3.3 , and the properties of L_{0} is summarized in Lemma 4.2 . One of the key properties that we want for the coefficient matrix A_{0} of the intermediate operator L_{0} is the following: we construct another orthonormal basis \left\{w_{1},\dots,w_{n}\right\} from \left\{v_{1},\dots,v_{n}\right\} , so that when written in these two bases , the Jacobian matrix of \rho is close—up to a Carleson perturbation—to a matrix in block form. It will imply that A_{\rho} is close to A_{0} , which, when written in the basis \left\{v_{1},\dots,v_{n}\right\} , is of the form \begin{bmatrix}*&\mathbf{0}^{T}\\ \mathbf{0}&1\end{bmatrix},\quad\text{where }\mathbf{0}=0_{\mathbb{R}^{n-1}}=(0,\dots,0). This implies that A_{0}v_{n}=v_{n} , which, thanks to the direct relation between v_{n} and the Green function, will be enough to show that L_{0} and -\Delta have the same L^{p} Dirichlet solvability. We note that reducing matters to a block-form matrix or to a small perturbation of such a matrix, is not new in this area (see e.g. [ HKMP15 , Fen22 , DFM19 , DHP23 ] ) and is often the core of the matter, though the idea is sometimes obscured by intricate integration-by-parts arguments. But again, in all previous work, such reductions rely crucially on the structure of \mathbb{R}^{n}_{+} and are based on the Euclidean basis \left\{e_{1},\dots,e_{n}\right\} that is constant. Let us now discuss the main ideas behind the construction of the basis \left\{w_{1},\dots,w_{n}\right\} from \left\{v_{1},\dots,v_{n}\right\} , as well as the map \rho . We want \left\{w_{1},\dots,w_{n}\right\} to be a small perturbation of \left\{v_{1},\dots,v_{n}\right\} , capturing the perturbation caused by the bi-Lipschitz map \Phi . In particular, the vector w_{n} should be an analog of v_{n} for the perturbed domain \Omega=\Phi(\Omega_{0}) , in the sense that it should play the role as a normal vector to the new boundary \partial\Omega that propagates into \Omega . This leads us to look at the perturbation of the boundary, which can be described by the vector-valued function h:\mathbb{R}^{n-1}\to\mathbb{R}^{n} defined by h(z):=\Phi(z,g(z))-(z,g(z)) (see ( 3.17 )). However, this function is not smooth enough (it is Lipschitz) to allow us control the resulting Jacobian, and so we introduce a smoothing of h for any point X\in\Omega_{0} , which is defined as \lambda(X) (see ( 3.16 )). We then define the map \rho:\Omega_{0}\to\Omega to be \rho(X)=X-t(X)v_{n}(X)+\lambda(X)+a(X)t(X)w_{n}(X), or equivalently, as in ( 3.40 ). Here, one should think of X-t(X)v_{n}(X) as the projection of X\in\Omega_{0} onto \partial\Omega_{0} , and so X-t(X)v_{n}(X)+\lambda(X) is near the boundary of \Omega smoothed at scale t(X) . The map \rho preserves the special direction of \partial\Omega due to our construction of w_{n} . The role of a(X) (defined in ( 3.36 )) is essentially to make the nn -th entry of the matrix A_{0} equal to 1, which simplifies the step comparing L_{0} to -\Delta . In the simpler case when the perturbed domain takes the form ( 1.6 ), experts familiar with [ DFM19 ] may wish to compare our construction of \rho with the change of variables introduced there. In [ DFM19 ] , a modification of the Nečas–Kenig–Stein map \rho_{1} is introduced to show that the L^{p} Dirichlet problem is solvable for some p>1 for a class of degenerate elliptic operators in domains with low-dimensional boundaries. One of the key advantages of their modification, even in the co-dimension 1 case, is that it is better adapted to the geometry of the target domain, in that it tends to preserve the orthogonal direction. In this respect, our construction is similar in spirit, but many new ideas are required as the argument must now be carried out relative to a non-constant basis \{v_{1},\dots,v_{n}\} , must also accommodate a perturbation of the domain, and we consider perturbations much more general than those occur only in directions transversal to the boundary of the base domain. We conclude the introduction by noting that our approach appears robust enough to apply to more general elliptic operators. For instance, we are hopeful that our results could be extended to operators satisfying the DKP condition, although substantial modifications would be required. Our results may also extend to more general domains whose boundaries are not necessarily given by graphs, following the ideas of [ KT03 ] . In a different direction, our method might be adapted to address Question 1.2 for the Neumann problem; however, due to the inherent difficulty of the Neumann problem, new ideas would be needed. Finally, a precise characterization of the necessary condition in Question 1.1 remains open and is likely to be highly challenging. 2. Preliminaries 2.1. Notations and definitions • For a scalar function f:\mathbb{R}^{n}\to\mathbb{R} , we adopt the convention that \nabla f is represented as a column vector, while a point X\in\mathbb{R}^{n} is represented as a row vector. • The dot product of two vectors of the same dimension yields a scalar. We denote it either by the usual dot “ \cdot ” or by the inner product notation \langle\cdot,\cdot\rangle . • We use |\cdot| to denote the norm of vectors or matrices. We use \|\cdot\| to denote norms on function spaces. When dealing with vector- or matrix-valued functions, \|\cdot\| should be understood as \||\cdot|\| , i.e., we first take the pointwise vector/matrix norm and then the function norm. • We denote by I_{ij} the entries of the identity matrix, i.e., I_{ij}=\begin{cases}1,&\text{if }i=j,\\ 0,&\text{if }i\neq j.\end{cases} • B_{\mathbb{R}^{n-1}}(x,r) denotes the ball in \mathbb{R}^{n-1} centered at x\in\mathbb{R}^{n-1} with radius r . We sometimes omit the subscript when the ambient space is clear from the context. • \delta(X) denotes \dist(X,\partial\Omega) when the underlying domain \Omega is clear from the context. • Given a domain \Omega and a point X\in\Omega , we denote B_{X}:=B(X,\delta(X)/2) , unless specified otherwise. • For x\in\partial\Omega , r>0 , we denote by \Delta(x,r) the boundary ball B(x,r)\cap\partial\Omega . • We denote by \pi the orthogonal projection on the first (n-1) -coordinates, i.e. \pi:\mathbb{R}^{n}\to\mathbb{R}^{n-1},\,(x_{1},\dots,x_{n-1},x_{n})\mapsto(x_{1},\dots,x_{n-1}). We use the concept of Carleson measures throughout the article. Definition 2.1 ( \CMs and \CM ) . Let \Omega be a Lipschitz domain in \mathbb{R}^{n} . We say a quantity f defined on \Omega satisfies the Carleson measure condition with sup (denoted by f\in\CMs ) if \sup_{\frac{1}{2}B_{X}}|f|^{2}\frac{1}{\delta(X)}dX is the density of a Carleson measure on \Omega , that is, there is some constant M>0 such that \iint_{B(x,r)\cap\Omega}\left(\sup_{Z\in\frac{1}{2}B_{Y}}|f(Z)|\right)^{2}\frac{dY}{\delta(Y)}\leq Mr^{n-1}\quad\text{for all }x\in\partial\Omega,r>0. If we want to emphasize the constant M , we write f\in\CMs(M) . We say a quantity f defined on \Omega satisfies the Carleson measure condition (denoted by f\in\CM ) if |f(X)|^{2}\frac{1}{\delta(X)}dX is the density of a Carleson measure on \Omega . Remark 2.2 . Observe that if f\in\CMs(M) , then there exists C>0 such that |f(X)|\leq CM^{1/2} for X\in\Omega . Next, we introduce the Green function and elliptic measures. For a uniformly elliptic operator L=-\diver A\nabla , one can construct the Green function G_{L} on a Lipschitz domain (cf. [ GW82 ] , and [ HK07 ] for unbounded domains) as well as the elliptic measure \omega_{L} . Although these can be done in more general settings, we only give the definitions and state the related results in Lipschitz domains and for symmetric uniformly elliptic operators in order to lighten the notations and avoid introducing extra geometry. We start with the L -elliptic measure with a pole in \Omega . For X\in\Omega , there exists a unique Borel probability measure \omega^{X}_{L} on \partial\Omega , such that for f\in C^{\infty}(\partial\Omega) if \partial\Omega is bounded and f\in C_{c}^{\infty}(\partial\Omega) if \partial\Omega is unbounded, the function (2.3) u(X)=\int_{\partial\Omega}f(y)\,d\omega_{L}^{X}(y) is the unique weak solution to the Dirichlet problem Lu=0 in \Omega , u|_{\partial\Omega}=f satisfying u\in C(\overline{\Omega}) , and if \Omega is unbounded, additionally u(X)\to 0 as |X|\to\infty in \Omega . We call \omega^{X}_{L} the L -elliptic measure (or the elliptic measure corresponding to L ) with pole at X . We can associate to L=-\diver A\nabla a unique Green function in \Omega , G_{L}(X,Y):\Omega\times\Omega\setminus\text{diag}(\Omega\times\Omega)\to\mathbb{R} with the following properties: G_{L}(\cdot,Y)\in W^{1,2}_{\rm loc}(\Omega\setminus\{Y\})\cap C(\overline{\Omega}\setminus\{Y\}) , G_{L}(\cdot,Y)\big|_{\partial\Omega}\equiv 0 for any Y\in\Omega , and LG_{L}(\cdot,Y)=\delta_{Y} in the weak sense in \Omega , that is, (2.4) \iint_{\Omega}A(X)\,\nabla_{X}G_{L}(X,Y)\cdot\nabla\varphi(X)\,dX=\varphi(Y),\qquad\text{for any }\varphi\in C_{c}^{\infty}(\Omega). In particular, G_{L}(\cdot,Y) is a weak solution to LG_{L}(\cdot,Y)=0 in \Omega\setminus\{Y\} . Given Y\in\Omega , we write G_{L}^{Y}=G_{L}(\cdot,Y) and call it the Green function with pole at Y . For unbounded domains, it is convenient to define the Green function and elliptic measure with pole at infinity . One can prove the following lemma as in [ KT99 , Lemma 3.7, Corollary 3.2] . Lemma 2.5 (Elliptic measure and Green function with pole at infinity) . Let \Omega be a Lipschitz graph domain and let L=-\diver{A\nabla} be a symmetric elliptic operator on \Omega . Then for any fixed Z_{0}\in\Omega , there exists a unique function G_{L}\in C(\overline{\Omega}) such that \begin{cases}LG_{L}=0\quad\text{in }\Omega\\ G_{L}>0\quad\text{in }\Omega\\ G_{L}=0\quad\text{on }\partial\Omega,\\ \end{cases} and G_{L}(Z_{0})=1 . In addition, for any sequence \left\{X_{k}\right\}_{k} of points in \Omega such that \left|X_{k}\right|\to\infty , there exists a subsequence (which we still denote by \left\{X_{k}\right\}_{k} ) such that if we set u_{k}(Y):=\frac{G_{L}(Y,X_{k})}{G_{L}(Z_{0},X_{k})}, then (2.6) u_{k}\text{ converges uniformly to }G_{L}\qquad\text{in any compact sets in }\overline{\Omega}\text{ as }k\to\infty. Moreover, there exists a unique locally finite positive Borel measure \omega_{L} on \partial\Omega such that the Riesz formula (2.7) \int_{\partial\Omega}f(y)\,d\omega_{L}(y)=-\iint A(Y)\nabla G_{L}(Y)\cdot\nabla F(Y)\,dY holds whenever f\in C_{c}^{\infty}(\partial\Omega) and F\in C_{c}^{\infty}(\Omega) are such that F|_{\partial\Omega}=f . We call \omega_{L} the elliptic measure with pole at infinity (normalized at Z_{0} ), and G_{L} the Green function with pole at infinity (normalized at Z_{0} ). Definition 2.8 ( (D)_{p}^{L} ) . Let p\in(1,\infty) . We say that the L^{p} Dirichlet problem (D)_{p} - or (D)_{p}^{L} when we mention the operator - is solvable in \Omega if there is C>0 such that for any f\in C^{\infty}_{0}(\mathbb{R}^{n}) , the solution u_{f} to Lu_{f}=0 in \Omega defined with the help of the elliptic measure by (2.9) u_{f}(X):=\int_{\partial\Omega}f(y)\,d\omega_{L}^{X}(y) satisfies \|N(u_{f})\|_{L^{p}(\partial\Omega)}\leq C\|f\|_{L^{p}(\partial\Omega)}, where N(u)(x)=N_{a}(u)(x):=\sup_{Y\in\gamma(x)}|u(Y)|,\text{ and }\gamma(x)=\gamma_{a}(x)=\left\{Y\in\Omega:\,|Y-x|0 such that for any f\in C^{\infty}_{0}(\mathbb{R}^{n}) , the solution u_{f} to Lu_{f}=0 in \Omega defined as in ( 2.9 ) satisfies \|\widetilde{N}(\nabla u_{f})\|_{L^{p}(\partial\Omega)}\leq C\|\nabla f\|_{L^{p}(\partial\Omega)}, where \widetilde{N}(u)(x)=\sup_{Y\in\gamma(x)}\mathchoice{{\vbox{\hbox{$\textstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.675pt}}{{\vbox{\hbox{$\scriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.02336pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-2.53589pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-1.7709pt}}\!\int\!\!\!\!\!\int_{B_{Y}}|u(Z)|dZ. 2.2. Some properties on (D)_{p} It is well-known that the elliptic measure and Green function with pole at infinity can be compared via the comparison principle (see, for example, [ Ken94 ] ). This is also true when the pole is at infinity. As before, we state the results only in the specific settings relevant to our work, rather than in their full generality. Lemma 2.11 (Lemmas 3.3 and 3.4 in [ DFM23 ] ) . Let \Omega_{0} be a Lipschitz graph domain, and let L=-\diver A\nabla be a symmetric uniformly elliptic operator on \Omega_{0} . Then there exists C>0 such that the Green function with pole at infinity G_{L} and the elliptic measure \omega_{L} with pole at infinity verifies C^{-1}\delta(X)^{n-2}G_{L}(X)\leq\omega_{L}(4B_{X})\leq\omega_{L}(8B_{X})\leq C\delta(X)^{n-2}G_{L}(X)\qquad\text{ for all }X\in\Omega_{0}, where B_{X}=B(X,\delta(X)/2) . In particular, \omega_{L} is a doubling measure. Moreover, the constant C=C(n,M,C_{ellip}) depends only the dimension n , the Lipschitz constant M of \Omega_{0} , and the ellipticity constant C_{ellip} of L . There is also a finite-pole version of the lemma, which we omit here. The elliptic measure and Green function with pole at infinity can be used to characterize the solvability of the Dirichlet problem. Theorem 2.12 . Let \Omega_{0} be a Lipschitz graph domain, and write \sigma for the surface measure on \partial\Omega_{0} . Let L=-\diver A\nabla be a symmetric uniformly elliptic operator, and p\in(1,\infty) . Then the following are equivalent: (i) The L^{p} Dirichlet problem (D)_{p}^{L} is solvable in \Omega_{0} . (ii) The elliptic measure with pole at infinity \omega_{L} is absolutely continuous with respect to \sigma , and there exists C_{p}>0 such that the function \kappa defined as \kappa(x):=\limsup_{X\in\gamma(x)\atop X\to x}\frac{G_{L}(X)}{\delta(X)} satisfies (2.13) \left(\fint_{B(x,r)\cap\partial\Omega_{0}}\kappa^{p^{\prime}}(y)\,d\sigma(y)\right)^{\frac{1}{p^{\prime}}}\leq C_{p}\fint_{B(x,r)\cap\partial\Omega_{0}}\kappa(y)\,d\sigma(y)\qquad\text{ for all }x\in\partial\Omega_{0},\,r>0. Here p^{\prime} stands for the Hölder conjugate of p . When \Omega_{0} is a bounded Lipschitz domain, we take a point X_{0}\in\Omega_{0} such that \dist(X_{0},\partial\Omega_{0})\approx\diam(\Omega) . Then (i) is equivalent to (ii’) The elliptic measure \omega_{L}^{X_{0}} is absolutely continuous with respect to \sigma , and there exists C_{p}>0 such that the function \kappa^{X_{0}} defined as \kappa^{X_{0}}(x):=\limsup_{X\in\gamma(x)\atop X\to x}\frac{G_{L}^{X_{0}}(X)}{\delta(X)}, satisfies (2.14) \left(\fint_{B(x,r)\cap\partial\Omega_{0}}\left(\kappa^{X_{0}}\right)^{p^{\prime}}(y)\,d\sigma(y)\right)^{\frac{1}{p^{\prime}}}\leq C_{p}\fint_{B(x,r)\cap\partial\Omega_{0}}\kappa^{X_{0}}(y)\,d\sigma(y)\qquad\text{for }x\in\partial\Omega,\,10 such that the function k:=\frac{d\omega_{L}}{d\sigma} satisfies ( 2.13 ), or, in the bounded case, k^{X_{0}}:=\frac{d\omega_{L}^{X_{0}}}{d\sigma} satisfies ( 2.14 ). We claim that k and \kappa are comparable, and that k^{X_{0}} and \kappa^{X_{0}} are comparable, with the implicit constants depending only the dimension n , the Lipschitz constant M of \Omega_{0} , and the ellipticity constant C_{ellip} of L . Note that the equivalence follows once this claim is justified. We only prove the unbounded (i.e. Lipschitz graph domain) case as the bounded case is similar. For any X\in\gamma(x) , write s_{X} for \delta(X) . Then by Lemma 2.11 , (2.15) \frac{G_{L}(X)}{\delta(X)}\approx\frac{\omega_{L}(\Delta(x,s_{X}))}{\sigma(\Delta(x,s_{X}))} since \sigma(\Delta(x,s_{X}))\approx s_{X}^{n-1}=\delta(X)^{n-1} . This implies that k(x)=\liminf_{s\to 0}\frac{\omega_{L}(\Delta(x,s))}{\sigma(\Delta(x,s))}\lesssim\limsup_{X\in\gamma(x)\atop X\to x}\frac{G_{L}(X)}{\delta(X)}\lesssim\limsup_{s\to 0}\frac{\omega_{L}(\Delta(x,s))}{\sigma(\Delta(x,s))}=k(x)\quad\sigma-\text{a.e. }x\in\partial\Omega_{0} by the assumption that \omega_{L}\ll\sigma . Hence we have that \kappa\approx k a.e.. \square The next result that we need concerns the stability of the Dirichlet problem under small Carleson perturbations. The earliest result of this kind was established by Dahlberg in [ Dah86 ] , which proves stability under Carleson perturbations with vanishing trace. The result below follows directly from [ CHM19 , Theorem 1.1] together with Theorem 2.12 . We note that the conclusion remains valid without assuming symmetry of the coefficient matrices; however, we state it in this form since it suffices for our application, and we are not aware of an appropriate reference in the general case. Theorem 2.16 . Let \Omega_{0} be a Lipschitz graph domain, and let L_{0}:=-\diver A_{0}\nabla and L_{1}:=-\diver A_{1}\nabla be symmetric uniformly elliptic operators. Let p\in(1,\infty) such that the Dirichlet problem for L_{0} in \Omega_{0} with boundary data in L^{p}(\partial\Omega_{0}) is solvable . There exists \epsilon_{1}>0 such that if A_{1}-A_{0}\in\CMs(\epsilon_{0}) , then the Dirichlet problem for L_{1} in \Omega with data in L^{p}(\partial\Omega_{0}) is solvable, meaning that the reverse Hölder bounds ( 2.13 ) holds with some constant C^{\prime}_{p} . Moreover, the constants \epsilon_{1}=\epsilon_{1}(n,p,C_{p},M,C_{ellip}) and C^{\prime}_{p}=C^{\prime}_{p}(n,p,C_{p},M,C_{ellip}) depends only on the dimension n , the parameter p , the reverse Hölder constant C_{p} in ( 2.13 ) for the L_{0} in \Omega_{0} , the Lipschitz constant M of \Omega_{0} , and the ellipticity constant C_{ellip} of L_{0} . We also need the stability of the Dirichlet problem under bi-Lispchitz change of variables. Proposition 2.17 . Let \Omega_{0} and \Omega be two Lipschitz domains in \mathbb{R}^{n} , and let \rho:\Omega_{0}\to\Omega be a bi-Lipschitz map that satisfies \rho(\partial\Omega_{0})=\partial\Omega . Then the solvability of the L^{p} Dirichlet problem is stable under \rho . That is, let L=-\diver A\nabla be an elliptic operator on \Omega , and let L_{\rho} be the pull-back of L under \rho (a.k.a. the conjugate of L by \rho ), then the L^{p} Dirichlet problem for L is solvable in \Omega if and only if the L^{p} Dirichlet problem for L_{\rho} is solvable in \Omega_{0} . Proof. This result is folklore; we sketch its proof for completeness. We only show that if the L^{p} Dirichlet problem is solvable for L in \Omega , then the L^{p} Dirichlet problem is solvable for L_{\rho} in \Omega_{0} , as the other direction is similar. Suppose the Lipschitz constant of \rho and \rho^{-1} is bounded by M . Let g\in L^{p}(\partial\Omega_{0}) be arbitrary, and let f=g\circ\rho^{-1} . Then f\in L^{p}(\partial\Omega) , and so by assumption, (2.18) \left\|N_{b}(u_{f})\right\|_{L^{p}(\partial\Omega)}\leq C\left\|f\right\|_{L^{p}(\partial\Omega)} for some C>0 , where u_{f} is defined as in ( 2.9 ), and b>0 is any aperture which will be determined later. One can verify that the solution v_{g} given by v_{g}(X):=\int_{\partial\Omega_{0}}g(y)d\omega_{L_{\rho}}^{X}(y) satisfies v_{g}=u_{f}\circ\rho . Our goal is to show that \left\|N(v_{g})\right\|_{L^{p}(\partial\Omega_{0})}\leq C^{\prime}\left\|g\right\|_{L^{p}(\partial\Omega_{0})} for some C^{\prime}>0 . Observe that for any x\in\partial\Omega_{0} , we have N(v_{\rho})(x)=\sup_{Y\in\gamma(x)}|u_{f}(\rho(Y))|\leq\sup_{Y\in\gamma_{M^{2}}(\rho(x))}|u_{f}(Y)|=N_{2M^{2}}(u_{f})(\rho(x)), where \gamma(x)=\left\{Y\in\Omega_{0}:\,|Y-x|<2\dist(Y,\partial\Omega_{0})\right\} is a cone in \Omega_{0} with vertex x and aperture 1, and \gamma_{2M^{2}}(\rho(x)) is a cone in \Omega with vertex at \rho(x) and aperture M^{2} . In fact, for Y\in\gamma(x) , since \rho is bi-Lipschitz with constant M , |\rho(Y)-\rho(x)|\leq M|Y-x|\leq 2M\dist(Y,\partial\Omega_{0})\leq 2M^{2}\dist(\rho(Y),\rho(\partial\Omega_{0})), which implies that \rho(Y)\in\gamma_{2M^{2}}(\rho(x)) . Therefore, taking b=2M^{2} in ( 2.18 ), we get that \left\|N(v_{\rho})\right\|_{L^{p}(\partial\Omega_{0})}\leq\left\|N_{M^{2}}(u_{f})\circ\rho\right\|_{L^{p}(\partial\Omega_{0})}\leq C_{M}\left\|N_{M^{2}}(u_{f})\right\|_{L^{p}(\partial\Omega)}\leq C_{M}\left\|f\right\|_{L^{p}(\partial\Omega)}\leq C_{M}\left\|g\right\|_{L^{p}(\partial\Omega_{0})}, which proves that the L^{p} Dirichlet problem is solvable for L_{\rho} in \Omega_{0} . ∎ 2.3. Estimates for the Green function. Lemma 2.19 . Let \Omega_{0} be a domain in \mathbb{R}^{n} and let \Delta u=0 in \Omega_{0} . Then there exists C>0 such that |\nabla u(X)|\leq\frac{C}{\delta(X)}\left(\fint_{B_{X}}|u(Y)|^{2}dY\right)^{1/2}\qquad\text{ for all }X\in\Omega_{0}, where B_{X}:=B(X,\delta(X)/2) . If in addition, u\geq 0 in \Omega_{0} , then |\nabla u(X)|\leq Cu(X)/\delta(X)\qquad\text{ for all }X\in\Omega_{0}. In particular, if \Omega_{0} is unbounded and if G is the Green function with pole at infinity for the Laplacian, as defined in Lemma 2.5 , then there exists C>0 such that |\nabla G(X)|\leq CG(X)/\delta(X)\qquad\text{ for all }X\in\Omega_{0}. Proof: This lemma is a classical consequence of the mean value property of harmonic functions, and also holds for solutions to more general elliptic PDEs. We include the short proof for completeness. Since \partial_{i}u is harmonic in \Omega_{0} for each i\in\{1,\dots,n\} , the De Giorgi-Nash-Moser estimate gives that |\nabla u(X)|^{2}:=\sum_{i=1}^{n}|\partial_{i}u(X)|^{2}\lesssim\sum_{i=1}^{n}\fint_{\frac{1}{2}B_{X}}|\partial_{i}u|^{2}dY=\fint_{\frac{1}{2}B_{X}}|\nabla u|^{2}dY. Using the Caccioppoli inequality, we further get |\nabla u(X)|^{2}\lesssim\frac{1}{\delta(X)^{2}}\fint_{B_{X}}u^{2}(Y)dY, which is the first claim. If u\geq 0 in \Omega_{0} , then the right-hand side above is further bounded by \frac{u(X)^{2}}{\delta(X)^{2}} by the Harnack inequality. The lemma follows. \square Theorem 2.20 . Let \Omega_{0} and G be as in Lemma 2.19 . Then there holds \dfrac{\delta^{2}|\nabla^{2}G|}{G}\in\CMs(M), where M depends on the Lipschitz character of \Omega_{0} . Proof: This result is essentially a special case of the main result of [ FL23 ] (see alternatively [ DLM22 , Corollary 1.17] and [ Azz19 ] ). Namely, [ FL23 , Theorem 1.13] gives that \dfrac{\delta^{2}|\nabla^{2}G|}{G}\in\CM(M) , that is, (2.21) \iint_{B(x,r)\cap\Omega_{0}}\left|\frac{\nabla^{2}G(Y)}{G(Y)}\right|^{2}\delta(Y)^{3}dY\leq Mr^{n-1}\quad\text{for all }x\in\partial\Omega_{0},\,r>0. So we only need to show that we can take supreme in \frac{1}{2}B_{Y} in the integrand, where B_{Y}=B(Y,\delta(Y)/2) . Since \partial_{i}\partial_{j}G is harmonic in \Omega_{0} for each i,j\in\left\{1,\dots,n\right\} , the De Giorgi-Nash-Moser estimate gives that \sup_{Z\in\frac{1}{2}B_{Y}}|\nabla^{2}G(Z)|\lesssim\mathchoice{{\vbox{\hbox{$\textstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.675pt}}{{\vbox{\hbox{$\scriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.02336pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-2.53589pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-1.7709pt}}\!\int\!\!\!\!\!\int_{\frac{1}{2}B_{Y}}|\nabla^{2}G(Z)|dZ. Therefore, applying Harnack inequality to G in \frac{1}{2}B_{Y} and Hölder’s inequality, we get that \iint_{B(x,r)\cap\Omega_{0}}\sup_{Z\in\frac{1}{2}B_{Y}}\left|\frac{\delta(Z)^{2}\nabla^{2}G(Z)}{G(Z)}\right|^{2}\frac{dY}{\delta(Y)}\lesssim\iint_{B(x,r)\cap\Omega_{0}}\mathchoice{{\vbox{\hbox{$\textstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.675pt}}{{\vbox{\hbox{$\scriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-4.02336pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-2.53589pt}}{{\vbox{\hbox{$\scriptscriptstyle\mkern-1.5mu{-}\mkern-7.5mu{-}$}}\kern-1.7709pt}}\!\int\!\!\!\!\!\int_{Z\in\frac{1}{2}B_{Y}}\frac{\left|\nabla^{2}G(Z)\right|^{2}}{G(Z)^{2}}\delta(Z)^{3}dZdY\\ =\iint_{Z\in B(x,2r)\cap\Omega_{0}}\frac{\left|\nabla^{2}G(Z)\right|^{2}}{G(Z)^{2}}\delta(Z)^{3}\left(\iint_{|Y-Z|<\delta(Y)/4}\delta(Y)^{-n}dY\right)dZ by Fubini’s theorem. Note that |Y-Z|<\delta(Y)/4 implies that \frac{4}{5}\delta(Z)\leq\delta(Y)\leq\frac{4}{3}\delta(Z) and that |Y-Z|\leq\frac{1}{3}\delta(Z) . Hence, \iint_{|Y-Z|<\delta(Y)/4}\delta(Y)^{-n}dY\lesssim\delta(Z)^{-n}\iint_{B(Z,\frac{\delta(Z)}{3})}dY\lesssim 1, from which it follows that for all x\in\partial\Omega_{0} and r>0 , \iint_{B(x,r)\cap\Omega_{0}}\sup_{Z\in\frac{1}{2}B_{Y}}\left|\frac{\delta(Z)^{2}\nabla^{2}G(Z)}{G(Z)}\right|^{2}\frac{dY}{\delta(Y)}\lesssim\iint_{B(x,2r)\cap\Omega_{0}}\frac{\left|\nabla^{2}G(Z)\right|}{G(Z)}^{2}\delta(Z)^{3}dZ\lesssim Mr^{n-1} by ( 2.21 ). The lemma follows. \square 3. The change of variables - the heart of the matter 3.1. Set-up estimates on Lipschitz graph domains Throughout this section, \Omega_{0} is an unbounded domain in \mathbb{R}^{n} , and G is the Green function for the Laplacian with pole at infinity in \Omega_{0} . Definition 3.1 . We say that the domain \Omega_{0} satisfies the \nabla G ’s non-degeneracy condition—called ( H1 )—if there exists C_{1}>0 such that (H1) G(X)\leq C_{H}\delta(X)|\nabla G(X)|\qquad\text{ for all }X\in\Omega_{0}. We say that the domain \Omega satisfies the preferred direction condition—called ( H2 )—if there exists an orthonormal basis \{e_{1},\dots,e_{n}\} and a constant C_{2}>0 such that (H2) \left<\nabla G(X),e_{n}\right>\geq(C_{2})^{-1}|\nabla G(X)|\quad\text{ for }X\in\Omega_{0}. Lemma 3.2 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} be a Lipschitz graph domain. Then \Omega_{0} satisfies ( H1 )–( H2 ). Proof: Recall that we write \partial_{n}G for \left<\nabla G(X),e_{n}\right> . We can assume without loss of generality that g(0)=0 , that is, 0\in\partial\Omega_{0} . Step 1: Show that \partial_{n}G\geq 0 in \Omega_{0} . This step carries all the key arguments, and ( H1 )–( H2 ) is actually a corollary of this statement. Step 1a: Convergence to G . For Y\in\Omega_{0} , let G^{Y} be a smooth version of Green function with pole at Y . Specifically, if G(X,Y) is the Green function in \Omega_{0} , and if \eta_{t} is a mollifier, meaning that \eta\in C^{\infty}(B_{\mathbb{R}^{n}}(0,1/2)) , \eta\geq 0 , \int_{\mathbb{R}^{n}}\eta=1 , and \eta_{t}(Z):=t^{-n}\eta(Z/t) , then G^{Y}(X):=\int_{\Omega_{0}}\eta_{\delta(Y)}(Y-Z)G(X,Z)\,dZ. Let Y_{i} be the point (0,2^{i}) . From the construction of the Green function with pole at infinity, and in particular, ( 2.6 ), we have that (3.3) \frac{G(.,Z_{i})}{G(Y_{0},Z_{i})}\text{ converges to }\frac{G}{G(Y_{0})}\text{ uniformly on compacts of $\overline{\Omega_{0}}$}. as i\to\infty , as long as Z_{i}\in B(Y_{i},\delta(Y_{i})/2) . Moreover, the rate of convergence does not depend on the choice of Z_{i}\in B(Y_{i},\delta(Y_{i})/2) . As a consequence (3.4) \frac{G^{Y_{i}}}{G^{Y_{i}}(Y_{0})}\text{ converges to }\frac{G}{G(Y_{0})}\text{ uniformly on compacts of $\overline{\Omega_{0}}$}. The De Giorgi-Nash-Moser estimate and the Caccioppoli inequality show then the convergence of the gradient, that is (3.5) \frac{\partial_{n}G^{Y_{i}}}{G^{Y_{i}}(Y_{0})}\text{ converges to }\frac{\partial_{n}G}{G(Y_{0})}\text{ uniformly on compacts of $\Omega_{0}$}. Indeed, take K\Subset\Omega_{0} be compact, and chose i\in\mathbb{N} large enough such that |Y_{i}|\geq 4(\diam K+\dist(K,\partial\Omega_{0}) . Then for any X\in K , since both G^{Y_{i}} and G are solutions to \Delta u=0 in B(X,\delta(X)) , so is \frac{\partial_{n}G^{Y_{i}}}{G^{Y_{i}}(Y_{0})}-\frac{\partial_{n}G}{G(Y_{0})} . Therefore, by the De Giorgi-Nash-Moser estimate and the Caccioppoli inequality, we have \left|\frac{\partial_{n}G^{Y_{i}}(X)}{G^{Y_{i}}(Y_{0})}-\frac{\partial_{n}G(X)}{G(Y_{0})}\right|\lesssim\left(\fint_{B_{X}/2}\left|\frac{\partial_{n}G^{Y_{i}}(Z)}{G^{Y_{i}}(Y_{0})}-\frac{\partial_{n}G(Z)}{G(Y_{0})}\right|^{2}\,dZ\right)^{\frac{1}{2}}\\ \lesssim\frac{1}{\dist(K,\partial\Omega_{0})}\left(\fint_{B_{X}}\left|\frac{G^{Y_{i}}(Z)}{G^{Y_{i}}(Y_{0})}-\frac{G^{Y_{i}}(Z)}{G^{Y_{i}}(Y_{0})}\right|^{2}\,dZ\right)^{\frac{1}{2}}\\ \lesssim\frac{1}{\dist(K,\partial\Omega_{0})}\sup_{Z\in K^{*}}\left|\frac{G^{Y_{i}}(Z)}{G^{Y_{i}}(Y_{0})}-\frac{G(Z)}{G(Y_{0})}\right|\longrightarrow 0\text{ as }i\to\infty. where K^{*}:=\bigcup_{Z\in K}\overline{B(Z,\delta(Z)/2)} , and the final convergence is by ( 3.4 ). Step 1b: Representation of \partial_{n}G^{Y} . We write H^{Y} for the function H^{Y}(X):=\int_{\Omega_{0}}\left(\partial_{n}\eta_{\delta(Y)}\right)(Y-Z)G(X,Z)dZ. We claim that (3.6) \partial_{n}G+H^{Y}\text{ is a weak solution to $-\Delta u=0$ in $\Omega_{0}$.} First, G^{Y} and H^{Y} are smooth functions, since there both weak solution to -\Delta u=f with some f\in C^{\infty}_{0}(\Omega) . More precisely, -\Delta G^{Y}=\eta_{\delta(Y)}(Y-\cdot) , and -\Delta H^{Y}=\left(\partial_{n}\eta_{\delta(Y)}\right)(Y-\cdot) . Hence, \partial_{n}G^{Y} and H^{Y} are in W^{1,2}_{loc}(\Omega_{0}) . Second, for \varphi\in C^{\infty}_{0}(\Omega) , we have by the PDE satisfied by G^{Y} and H^{Y} , \int_{\Omega_{0}}\nabla[\partial_{n}G^{Y}+H^{Y}]\cdot\nabla\varphi\,dZ=-\int_{\Omega_{0}}\nabla G^{Y}\cdot\nabla\partial_{n}\varphi\,dZ+\int_{\Omega_{0}}\left(\partial_{n}\eta_{\delta(Y)}\right)(Y-Z)\varphi(Z)\,dZ\\ =\int_{\Omega_{0}}\big(-\eta_{\delta(Y)}(Y-Z)\partial_{n}\varphi(Z)+\left(\partial_{n}\eta_{\delta(Y)}\right)(Y-Z)\varphi(Z)\big)\,dZ. Integration by parts in Z_{n} gives that -\int_{\Omega_{0}}\eta_{\delta(Y)}(Y-Z)\partial_{n}\varphi(Z)dZ=\int_{\Omega_{0}}\partial_{Z_{n}}\left(\eta_{\delta(Y)}(Y-Z)\right)\varphi(Z)dZ=-\int_{\Omega_{0}}(\partial_{n}\eta_{\delta(Y)})(Y-Z)\varphi(Z)dZ, which implies that \int_{\Omega_{0}}\nabla[\partial_{n}G^{Y}+H^{Y}]\cdot\nabla\varphi=0 as claimed. Since \Omega_{0} is a Lipschitz domain, the L^{2} Dirichlet problem is solvable for -\Delta in \Omega_{0} (see [ Dah77 ] or [ JK81 ] 2 2 2 An attentive reader might notice that both references are for bounded Lipschitz domain, but the same result holds for unbounded domain. Alternatively, since we actually only need L^{p} estimate for some p>1 , we can apply [ HKMP15 ] for unbounded domains. ). So, by Theorem 1.22 in [ MPT22 ] , the Poisson-Dirichlet and Poisson-regularity problem are also solvable in L^{2} . It implies that \|N(\nabla G^{Y})\|_{L^{2}(\partial\Omega_{0})}+\|N(H^{Y})\|_{L^{2}(\partial\Omega_{0})}<+\infty. In particular, \partial_{n}G^{Y} has a trace in L^{2}(\partial\Omega_{0}) . The L^{2} solvability of the Dirichlet problem in \Omega_{0} also implies that the harmonic extension u_{\partial_{n}G^{Y}} of this trace in \Omega_{0} verifies \|N(u_{\partial_{n}G^{Y}})\|_{L^{2}(\partial\Omega_{0})}\lesssim\left\|Tr(\partial_{n}G^{Y})\right\|_{L^{2}(\partial\Omega_{0})}<+\infty. Note that the non-tangential bounds of \partial_{n}G^{Y} , H^{Y} , and u_{\partial_{n}G^{Y}} guarantees the almost everywhere non-tangential limit of the functions to their traces, and that \Tr(\partial_{n}G^{Y}+H^{Y}-u_{\partial_{n}G^{Y}})=\Tr(H^{Y})=0. Consequently, \partial_{n}G^{Y}+H^{Y}-u_{\partial_{n}G^{Y}} is a function with non-tangential bounds and zero trace, and is a solution to -\Delta u=0 in \Omega_{0} by ( 3.6 ). Uniqueness of the Dirichlet problem (see [ Ken94 , Theorem 1.4.4] , or [ HLMP22 , Section 6] for unbounded domains) gives the formula: \partial_{n}G^{Y}=u_{\partial_{n}G^{Y}}-H^{Y}\quad\text{in }\Omega_{0}. Step 1c: \partial_{n}G\geq 0 . First, note that \partial_{n}G is non-negative on the boundary \partial\Omega_{0} . Indeed, for x\in\partial\Omega , since G^{Y} is positive in \Omega_{0} and since x+he_{n}\in\Omega_{0} for h>0 , we have \partial_{n}G^{Y}(x)=\lim_{h\to 0\atop h>0}\frac{G^{Y}(x+he_{n})}{h}\geq 0. As a consequence, the harmonic extension u_{\partial_{n}G^{Y}} of \Tr(\partial_{n}G^{Y}) - which is given by the integral of \Tr(\partial_{n}G^{Y}) with respect to a positive measure (the harmonic measure) - is non-negative in \Omega_{0} . Second, we compare H^{Y} and G^{Y} outside of B_{Y} . Observe that H^{Y}(X)=-\int_{B_{Y}}\partial_{Z_{n}}\left(\eta_{\delta(Y)}(Y-Z)\right)G(X,Z)dZ=\int_{B_{Y}}\eta_{\delta(Y)}(Y-Z)\partial_{Z_{n}}G(X,Z)dZ by integration by parts. Therefore, |H^{Y}(X)|\leq\int_{B_{Y}}\eta_{\delta(Y)}(Y-Z)|\nabla_{Z}G(X,Z)|dZ. If B_{X} and B_{Y} do not intersect, we apply Lemma 2.19 to G(X,\cdot) in \Omega_{0}\setminus B_{X} . For Z\in B_{Y} , \dist(Z,\partial(\Omega_{0}\setminus B_{X}))=\min\left\{\delta(Z),\dist(Z,B_{X})\right\} . Obviously, \delta(Z)\geq\delta(Y)/2 . We check that \dist(Z,B_{X})\geq\max\left\{\delta(X)/2,|Y-X|-\delta(X)/2-\delta(Y)/2\right\} . If \delta(X)<\delta(Y)/4 , then |Y-X|\geq\delta(Y)-\delta(X) , and so |Y-X|-\delta(X)/2-\delta(Y)/2\geq\delta(Y)/8 . This implies that \dist(Z,B_{X})\geq\delta(Y)/8 , and thus for Z\in B_{Y} , \dist(Z,\partial(\Omega_{0}\setminus B_{X}))\geq\delta(Y)/8 . So we have that |H^{Y}(X)|\lesssim\int_{B_{Y}}\eta_{\delta(Y)}(Y-Z)\frac{G(X,Z)}{\dist(Z,\partial(\Omega_{0}\setminus B_{X}))}dZ\lesssim\int_{B_{Y}}\eta_{\delta(Y)}(Y-Z)\frac{G(X,Z)}{\delta(Y)}dZ=\frac{G^{Y}(X)}{\delta(Y)}. It means that for any X\in\Omega_{0} and any c>0 , there exists i_{0}:=i_{0}(X,c) such that for i\geq i_{0} (3.7) |H^{Y_{i}}(X)|\lesssim 2^{-i}G^{Y_{i}}(X)\lesssim cG^{Y_{i}}. For c>0 , using successively Step 1a, Step 1b, the positivity of u_{\partial_{n}G^{Y_{i}}} and ( 3.7 ) we have then that \frac{\partial_{n}G(X)}{G(X)}=\lim_{i\to\infty}\frac{\partial_{n}G^{Y_{i}}(X)}{G^{Y_{i}}(X)}=\lim_{i\to\infty}\frac{u_{\partial_{n}G^{Y_{i}}}(X)-H^{Y_{i}}(X)}{G^{Y_{i}}(X)}\geq-c Since the bound is for all c>0 , we conclude that \frac{\partial_{n}G(X)}{G(X)}\geq 0 as desired. Step 2: Show that \partial_{n}G(X)\geq cG(X)/\delta(X) for X\in\Omega_{0} . From Step 1, we get that for any Lipschitz graph domain \Omega_{0} and any Green function with pole at infinity G on \Omega_{0} , we have (3.8) \partial_{n}G>0. Indeed, \partial_{n}G is a non-zero non-negative harmonic function in \Omega_{0} , so the Harnack inequality forces \partial_{n}G to be positive. We want to prove that there exists c>0 (depending only on n and the Lipschitz constant of g ) such that (3.9) \partial_{n}G(X)\geq cG(X)/\delta(X)\qquad\text{ for all }X\in\Omega_{0}. This claim easily yields ( H1 )–( H2 ). To show ( 3.9 ), it suffices to show that there is c>0 depending only on n and the Lipschitz constant of g , such that \partial_{n}G(0,1)>cG(0,1) (recall that 0\in\partial\Omega_{0} ). We can assume without loss of generality that G(0,1)=1 . By the Harnack inequality, 00 (depending only on n and the Lipschitz constant of g ) such that G(0,c_{0})\leq 1/2 . So we have that \frac{1}{2}\leq G(0,1)-G(0,c_{0})=\int_{c_{0}}^{1}\partial_{n}G(0,s)ds. By the mean value theorem, there exists s_{0}\in(c_{0},1) such that \partial_{n}G(0,s_{0})\geq 1/2 . Finally, since \partial_{n}G is harmonic and nonnegative, the Harnack inequality implies that \partial_{n}G(0,1)\geq c for some c>0 . \square Theorem 3.10 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} be a Lipschitz graph domain. Then the Green function G for -\Delta with pole at infinity defined in Lemma 2.5 verifies \delta\nabla\left(\frac{\nabla G}{|\nabla G|}\right)\in\CMs(M), and \nabla\left(\frac{G}{|\nabla G|}\right)-\frac{\nabla G}{|\nabla G|}\in\CMs(M), where M depends on the Lipschitz constant of g . Proof: This theorem is a simple consequence of Theorem 2.20 and ( H1 ), the latter being given by Lemma 3.2 . Indeed, ( H1 ) guarantees that |\nabla G| never becomes 0, meaning that we can differentiate \nabla G/|\nabla G| without subtilities and we have \left|\delta\nabla\left(\frac{\nabla G}{|\nabla G|}\right)\right|\leq 2\frac{\delta|\nabla^{2}G|}{|\nabla G|}\lesssim\frac{\delta^{2}|\nabla^{2}G|}{G}\in\CMs by ( H1 ) and then Theorem 2.20 . Similarly, \left|\nabla\left(\frac{G}{|\nabla G|}\right)-\frac{\nabla G}{|\nabla G|}\right|\leq\frac{G|\nabla^{2}G|}{|\nabla G|^{2}}\lesssim\frac{\delta^{2}|\nabla^{2}G|}{G}\in\CMs again by ( H1 ) and Theorem 2.20 . The theorem follows. \square We need now an orthonormal basis whose last element is \nabla G/|\nabla G| . Lemma 3.11 . Let \Omega_{0} satisfy ( H1 )–( H2 ). There exists n vectors functions v_{1},\dots,v_{n} on \Omega_{0} such that: (i) For each X\in\Omega_{0} , the collection \{v_{1}(X),\dots,v_{n}(X)\} is an orthonormal basis of \mathbb{R}^{n} . (ii) The last vector function is v_{n}:=\frac{(\nabla G)^{T}}{|\nabla G|} , so in particular v_{i}(X)\cdot\nabla G(X)=0\qquad\text{ for each }X\in\Omega\text{ and }i\in\{1,\dots,n-1\}. (iii) There exists C>0 such that, for each i\in\{1,\dots,n\} and X\in\Omega_{0} , |\nabla v_{i}(X)|\leq C|\nabla v_{n}(X)|. Proof: The result by itself is not surprising, since we can simply see it as the variant of the Gram-Schmidt process, and can probably already be found in the literature. For completeness, we sketch the argument. Let \{e_{1},\dots,e_{n}\} be the Cartesian orthonormal basis from condition ( H2 ). For any fixed X\in\Omega_{0} , we define the vectors v_{i}(X) by a decreasing induction in i as follows. For i\in\{n-1,\dots,1\} , (3.12) \tilde{v}_{i}(X):=e_{i}-\sum_{i\geq c for some c\in(0,1) independent of X . By the Parseval’s identity, (3.13) \sum_{i=1}^{n-1}|\left|^{2}\leq 1-c^{2}. We want to prove that (3.14) |\tilde{v}_{i}(X)|\geq c\quad\text{for }X\in\Omega_{0},\,i\in\left\{1,\dots,n-1\right\}. To this objective, note that by the Gram-Schmidt process, |\tilde{v}_{i}| is the distance from e_{i} to the linear space spanned by \{e_{i+1},\cdots,e_{n-1},v_{n}\} . Thus \displaystyle|\tilde{v}_{i}|^{2} \displaystyle=\inf_{\alpha_{k}\in\mathbb{R},i+1\leq k\leq n}\left \displaystyle=\inf_{\alpha_{k}\in\mathbb{R},i+1\leq k\leq n}\Big\{1+\sum_{k=i+1}^{n}\alpha_{k}^{2}-2\alpha_{n}\left-\sum_{k=i+1}^{n-1}2\alpha_{k}\alpha_{n}\left\Big\}. Using |2\alpha_{k}\alpha_{n}\left|\leq\alpha_{k}^{2}+\alpha_{n}^{2}|\left|^{2} and |2\alpha_{n}\left|\leq(1-c^{2})+(1-c^{2})^{-1}\alpha_{n}^{2}|\left|^{2} , we then have \displaystyle|\tilde{v}_{i}|^{2} \displaystyle\geq\inf_{\alpha_{n}\in\mathbb{R}}\Big\{1+\alpha_{n}^{2}-(1-c^{2})-(1-c^{2})^{-1}\alpha_{n}^{2}|\left|^{2}-\alpha_{n}^{2}\sum_{k=i+1}^{n-1}|\left|^{2}\Big\} \displaystyle\geq\inf_{\alpha_{n}\in\mathbb{R}}\Big\{c^{2}+\alpha_{n}^{2}-\alpha_{n}^{2}(1-c^{2})^{-1}\sum_{k=i}^{n-1}|\left|^{2}\Big\} \displaystyle\geq c^{2}, where we have used ( 3.13 ) in the last inequality. It remains to justify (iii) . In fact, the uniform non-degeneracy of \tilde{v}_{i} ( 3.14 ) guarantees that ( 3.12 ) is well-defined and thus, each v_{i} is a C^{\infty} function of v_{n} . Then the chain rule implies that |\nabla v_{i}|\lesssim|\nabla v_{n}| . The lemma follows. \square Corollary 3.15 . Let \Omega_{0} be a Lipschitz graph domain. Then the construction of \{v_{1},\dots,v_{n}\} given by Lemma 3.11 satisfies: \delta\nabla v_{i}\in\CMs(M), where M depends on the Lipschitz constant of \partial\Omega_{0} . Proof: Simple consequence of Lemma 3.11 and Theorem 3.10 . \square 3.2. Quantifying small Lipschitz perturbations of Lipschitz graph domains. In the rest of the section, \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R},\,t>g(x)\} is a Lipschitz graph domain, with a (possibly large) constant M . G is the Green function with pole at \infty for the Laplacian on \Omega_{0} . On top of that, we set \Omega:=\Phi(\Omega_{0}) , where \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} is a bi-Lipschitz map satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon , where \epsilon>0 will be small. We want to construct a “better” bi-Lipschitz change of variable \rho from \Omega_{0} to \Omega , which preserves the structure of \Omega_{0} . By preserving the structure of \Omega_{0} , we mean that at least the Jacobian \nabla\rho is close to identity, in the sense that the errors are Carleson measures on \Omega_{0} with small constant. We first need a mollifier. We let \eta\in C^{\infty}_{0}(\mathbb{R}^{n-1}) be a radial function such that \eta\geq 0 , \int_{\mathbb{R}^{n-1}}\eta\,dx=1 , and \eta\equiv 0 outside B_{\mathbb{R}^{n-1}}(0,1) . Then, for t>0 , we define \eta_{t}(x):=t^{1-n}\eta(x/t) . We view \Omega as a perturbation of \Omega_{0} , and the quantity that we want to control is \lambda:\Omega_{0}\to\mathbb{R}^{n} , which is a smooth version of the perturbation defined by (3.16) \lambda(X):=\int_{\mathbb{R}^{n-1}}h(z)\,\eta_{t(X)}\big(y(X)-z\big)\,dz\quad\text{for }X\in\Omega_{0}, where (3.17) h(z):=\Phi(z,g(z))-(z,g(z)), and (3.18) t(X):=\frac{G(X)}{|\nabla G(X)|},\qquad y(X):=\pi\Big(X-\frac{G(X)(\nabla G(X))^{T}}{|\nabla G(X)|^{2}}\Big)=\pi\Big(X-t(X)v_{n}(X)\Big), and v_{n}:=\frac{(\nabla G)^{T}}{|\nabla G|} as in Lemma 3.11 . Recall that \pi is the orthogonal projection on the first (n-1) -coordinates. We emphasize that in our definition, t:\Omega_{0}\to\mathbb{R} is a function and should not be confused with the variable t\in\mathbb{R} . We deliberately choose this notation, as we wish to view the function as playing the same role, in \Omega_{0} , of the special variable t in \mathbb{R}^{n}_{+} . In fact, by Lemma 2.19 , ( H1 ), Theorem 3.10 , and Corollary 3.15 , we have that (3.19) t\approx\delta:=\dist(.,\partial\Omega_{0}); (3.20) t\nabla v_{i}\in\CMs(CM)\quad\text{for }i\in\left\{1,\dots,n\right\}, and (3.21) \nabla t-v_{n}^{T}\in\CMs(CM). Note that h:\mathbb{R}^{n-1}\to\mathbb{R}^{n} is Lipschitz, since (3.22) |\nabla h(z)|\leq|\nabla\Phi((z,g(z))-I|\sqrt{1+|\nabla g(z)|^{2}}\leq C_{M}\epsilon. The main result of this subsection is: Theorem 3.23 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g:\mathbb{R}^{n-1}\to\mathbb{R} is a M -Lipschitz function, and let \Phi:\,\mathbb{R}^{n}\to\mathbb{R}^{n} be a bi-Lipschitz change of variable satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon . Then the quantity \lambda defined in ( 3.16 )–( 3.18 ) has the following properties: (i) \|\nabla\lambda\|_{\infty}\leq C_{M}\epsilon . (ii) \partial_{v_{n}}\lambda\in\CMs(C_{M}\epsilon) . (iii) There exists a decomposition \nabla\lambda=:\mathbf{B}+\mathbf{C} verifying \|\mathbf{B}\|_{\infty}\leq C_{M}\epsilon , v_{n}\mathbf{B}=\mathbf{0}_{\mathbb{R}^{n}} , and |\mathbf{C}|+t|\nabla\mathbf{B}|\in\CMs(C_{M}\epsilon) . The constant C_{M}>0 in (i), (ii) and (iii) is independent of \epsilon , meaning that it depends only on M , the dimension n , the choice of \eta , and the constant in ( H1 ). Proof: Step 0: Preliminaries. In the proof, we write t for t(X) , y for y(X) , v_{n} for v_{n}(X) , and G for G(X) ; this considerably lightens the notation. Now, ( 3.16 ) becomes \lambda(X):=\int_{\mathbb{R}^{n-1}}h(z)\eta_{t}\big(y-z\big)\,dz. Building on estimates ( 3.20 ), ( 3.21 ) and the observation that \nabla y=\pi(\nabla X)-\pi(\nabla(tv_{n}))=I_{n\times(n-1)}-\pi\left((\nabla t-v_{n}^{T})v_{n}\right)-\pi(v_{n}^{T}v_{n})-\pi(t\nabla v_{n}), where I_{n\times(n-1)} is the n\times(n-1) matrix that takes the value 1 on the diagonal and 0 elsewhere, we also have (3.24) \big[\nabla y-I_{n\times(n-1)}+(v_{n})^{T}\pi(v_{n})\big]\in\CMs(CM). In particular, the estimates ( 3.20 ), ( 3.21 ) and ( 3.24 ) imply that (see Remark 2.2 ) (3.25) \|\nabla t\|_{\infty}+\|t\nabla v_{n}\|_{\infty}+\|\nabla y\|_{\infty}\leq C(1+M^{1/2}). Step 1: proof of (i). Now, we can compute \nabla\lambda . Take a\in\mathbb{R}^{n} to be defined later. We have, since \int\eta_{t}=1 , (3.26) \nabla\lambda(X)=\nabla(\lambda-a)=\nabla_{X}\left(\int_{\mathbb{R}^{n-1}}[h(z)-a]\eta_{t}\big(y-z\big)\,dz\right)\\ =(\nabla t)\left(\int_{\mathbb{R}^{n-1}}[h(z)-a]\partial_{t}\eta_{t}\big(y-z\big)\,dz\right)+\left(\int_{\mathbb{R}^{n-1}}[h(z)-a](\nabla y)\nabla_{y}\eta_{t}\big(y-z\big)\,dz\right), where (\nabla y)\nabla_{y}\eta_{t} is seen as a matrix multiplication ( \nabla y has dimensions n\times(n-1) and \nabla_{y}\eta_{t} is a column vector of dimension n-1 ). Check that \nabla_{y}\eta_{t} and \partial_{t}\eta_{t} are both supported in B_{\mathbb{R}^{n-1}}(0,t) , and verify \|\nabla_{y}\eta_{t}\|_{\infty}+\|\partial_{t}\eta_{t}\|_{\infty}\lesssim t^{-n} . So we obtain that |\nabla\lambda(X)|\lesssim t^{-n}\left(\int_{B(y,t)}|h(z)-a|\,dz\right)(|\nabla t|+|\nabla y|)\lesssim\frac{M^{1/2}+1}{t}\fint_{B(y,t)}|h(z)-a|\,dz by ( 3.25 ). We choose a:=\fint_{B(y,t)}h\,dz , and since h is \epsilon -Lipschitz, we get that |h(z)-a|\leq\epsilon t for all z\in B(y,t) . So |\nabla\lambda(X)|\leq C_{M}\epsilon. with a constant independent of X . Conclusion (i) of the Theorem follows. Step 2: proof of (ii). We need to be finer. We compute \displaystyle v_{n}\nabla\lambda(X) \displaystyle=\left(\int_{\mathbb{R}^{n-1}}h(z)\partial_{t}\eta_{t}\big(y-z\big)\,dz\right)(\partial_{v_{n}}t)+\left(\int_{\mathbb{R}^{n-1}}h(z)v_{n}(\nabla y)\nabla_{y}\eta_{t}\big(y-z\big)\,dz\right) \displaystyle=:I_{1}+I_{2}. Observe that \nabla_{y}\eta_{t}(y-z)=-\nabla_{z}\eta_{t}(y-z) , and so \int_{\mathbb{R}^{n-1}}v_{n}(\nabla y)\nabla_{y}\eta_{t}\big(y-z\big)\,dz=0, which implies that I_{2}=\int_{\mathbb{R}^{n-1}}\left(h(z)-a^{\prime}\right)\big[v_{n}(\nabla y)\nabla_{y}\eta_{t}\big(y-z\big)\big]\,dz\quad\text{for any }a^{\prime}\in\mathbb{R}^{n}. Note that ( 3.24 ) gives that v_{n}(\nabla y)=v_{n}\big(\nabla y-I_{n\times(n-1)}+(v_{n})^{T}\pi(v_{n})\big)\in\CMs(CM). Hence |I_{2}|\leq\underbrace{\left(\int_{B(y,t)}|h(z)-a^{\prime}||\nabla_{y}\eta_{t}(y-z)|\,dz\right)}_{\leq C\epsilon\text{ as in Step 1}}|v_{n}(\nabla y)|\in\CMs(CM\epsilon). As for I_{1} , we use ( 3.25 ) to get that |I_{1}|\lesssim(1+M^{1/2})\Big|\underbrace{\int_{\mathbb{R}^{n-1}}h(z)\,\partial_{t}\eta_{t}\big(y-z\big)\,dz}_{I_{11}}\Big| To deal with I_{11} , we observe that \partial_{t}\eta_{t}(z)=\diver_{z}[\tilde{\eta}_{t}](z), with \tilde{\eta}(z)=-z\eta(z) and \tilde{\eta}_{t}(z)=t^{1-n}\tilde{\eta}(z/t) . Denote by (z_{1},\dots,z_{n-1}) the coordinates of z . By the divergence theorem, the identity implies that for any a\in\mathbb{R} , \int_{\mathbb{R}^{n-1}}(a+z_{i})\partial_{t}\eta_{t}\big(y-z\big)\,dz=-\int_{\mathbb{R}^{n-1}}\nabla_{z}z_{i}\cdot\tilde{\eta}_{t}(y-z)\,dz=0 since \tilde{\eta} is odd (because \eta is radial). We deduce that for any constant vector a\in\mathbb{R}^{n} and for any constant matrix B\in\mathbb{R}^{(n-1)\times n} , we have \int_{\mathbb{R}^{n-1}}(a+zB)\,\partial_{t}\eta_{t}\big(y-z\big)\,dz=0, and hence I_{11}=\int_{\mathbb{R}^{n-1}}[h(z)-a-zB]\,\partial_{t}\eta_{t}\big(y-z\big)\,dz. Consequently, (3.27) |I_{11}|=|I_{11}(X)|\lesssim\beta\left(y(X),t(X)\right), where \beta(y,t):=\frac{1}{t}\inf_{a\in\mathbb{R}^{n},B\in\mathbb{R}^{(n-1)\times n}}\fint_{B(y,t)}|h(z)-a-zB|\,dz. Observe that the function (y,t)\mapsto\beta(y,t) enjoys the following property: for (y,t) and (y^{\prime},t^{\prime})\in\mathbb{R}^{n-1}\times(0,\infty) satisfying the inclusion of balls B(y^{\prime},t^{\prime})\subset B(y,Kt) for some K\geq 1 and t^{\prime}\geq ct for some c>0 , we have \beta(y^{\prime},t^{\prime})\leq C_{n,K,c}\beta(y,Kt) . Moreover, since h is a C\epsilon -Lipschitz function, Dorronsoro’s Theorem ( [ Dor85 ] ) gives that \beta^{2}(y,t)dy\,dt/t is a Carleson measure on \mathbb{R}^{n-1}\times(0,\infty) , with Carleson norm smaller than C\epsilon , that is, (3.28) \sup_{(x,r)\in\mathbb{R}^{n-1}\times(0,\infty)}r^{1-n}\int_{B_{\mathbb{R}^{n-1}}(x,r)}\int_{0}^{r}\beta(y,t)^{2}\frac{dt}{t}dy\leq C\epsilon. We want to show that ( 3.28 ) implies that |I_{1}|\in\CMs(C_{M}\epsilon) . To do that, consider that map \Upsilon:(y,t)\mapsto(y,t+g(y)) . It is easy to see that \Upsilon is a bi-Lipschitz map from \mathbb{R}^{n-1}\times(0,\infty) to \Omega_{0} . For an arbitrary X\in\Omega_{0} , we reserve the notations y=y(X) , t=t(X) . We denote X^{\prime}=\Upsilon(y,t)=(y,t+g(y))\in\Omega_{0},\quad y^{\prime}=y(X^{\prime}),\text{ and }t^{\prime}=t(X^{\prime}). We claim that the point X^{\prime} is close to X in the sense that the coefficients \beta(y^{\prime},t^{\prime}) can be controlled by a multiple of \beta(y,Kt) for some K\geq 1 . In fact, since g is M -Lipschitz, \delta(X^{\prime})\approx t , which implies that t\approx t^{\prime} because \delta(X^{\prime})\approx t^{\prime} and \delta(X)\approx t by ( 3.19 ). Next, note that \pi(X^{\prime})=y , and so y^{\prime}=y-t^{\prime}\pi(v_{n}(X^{\prime})) . This implies that |y^{\prime}-y|=|t^{\prime}\pi(v_{n}(X^{\prime}))|\leq t^{\prime}\approx t , and hence B(y^{\prime},t^{\prime})\subset B(y,Kt) for some K\geq 1 . We deduce that \beta(y^{\prime},t^{\prime})\lesssim\beta(y,Kt) from the property of \beta . We now consider an arbitrary point Z\in W_{X^{\prime}} , where W_{X^{\prime}}=B(X^{\prime},\delta(X^{\prime})/4) is a Whitney ball in \Omega_{0} centered at X^{\prime}=\Upsilon(y,t) . Obviously, t(Z)\approx\delta(Z)\approx\delta(X^{\prime})\approx t^{\prime} , and |y(Z)-y^{\prime}|\leq|Z-X^{\prime}|+|t(Z)v_{n}(Z)-t^{\prime}v_{n}(X^{\prime})|\lesssim t^{\prime}. This again implies that B(y(Z),t(Z))\subset B(y^{\prime},Kt^{\prime}) for some K\geq 1 , and that \beta(y(Z),t(Z))\lesssim\beta(y^{\prime},Kt^{\prime}) . Therefore, by ( 3.27 ) and the above estimates on \beta , we have that for any x\in\mathbb{R}^{n-1} and r>0 , \int_{B(x,r)}\int_{0}^{r}\sup_{Z\in W_{\Phi(y,t)}}|I_{11}(Z)|^{2}\,\frac{dt}{t}\,dy\lesssim\int_{B(x,r)}\int_{0}^{r}\sup_{Z\in W_{\Phi(y,t)}}|\beta(y(Z),t(Z))|^{2}\,\frac{dt}{t}\,dy\\ \lesssim\int_{B(x,r)}\int_{0}^{r}|\beta(y^{\prime},Kt^{\prime})|^{2}\,\frac{dt}{t}\,dy\lesssim\int_{B(x,r)}\int_{0}^{r}|\beta(y,K^{2}t)|^{2}\,\frac{dt}{t}\,dy\lesssim\epsilon\,r^{n-1}, where we have used ( 3.28 ) in the last inequality. Using again the fact that \Upsilon is bi-Lipschitz, we obtain that |I_{1}|\in\CMs(C_{M}\epsilon) as desired. Step 2 follows. Step 3: proof of (iii). The computations are similar to Step 2, although more involved. Morally, we will proceed as in Step 2, but for all the coordinates instead just for v_{n} . The terms that we can control by Carleson measures at the ‘first derivative’ level will be put in \mathbf{C} . The others will be placed in \mathbf{B} and we shall differentiate them an additional time to get the control in terms of Carleson measure. We have \displaystyle\nabla\lambda(X)= \displaystyle(\nabla t)\left(\int_{\mathbb{R}^{n-1}}h(z)\partial_{t}\eta_{t}\big(y-z\big)\,dz\right)+\left(\int_{\mathbb{R}^{n-1}}(\nabla y)\nabla_{y}\eta_{t}\big(y-z\big)h(z)\,dz\right) \displaystyle= \displaystyle(\nabla t)\left(\int_{\mathbb{R}^{n-1}}h(z)\partial_{t}\eta_{t}\big(y-z\big)\,dz\right) \displaystyle+\left(\int_{\mathbb{R}^{n-1}}\big(\nabla y-I_{n\times(n-1)}+(v_{n})^{T}\pi(v_{n})\big)\nabla_{y}\eta_{t}\big(y-z\big)h(z)\,dz\right) \displaystyle+\left(\int_{\mathbb{R}^{n-1}}\big(I_{n\times(n-1)}-(v_{n})^{T}\pi(v_{n})\big)\nabla_{y}\eta_{t}\big(y-z\big)h(z)\,dz\right)=:J_{1}+J_{2}+J_{3}. We write \mathbf{C}:=J_{1}+J_{2} , and \mathbf{B}:=J_{3} . The control on \mathbf{C} is like in Step 2: specifically, J_{1} can be treated like I_{1} , J_{2} is treated like I_{2} (using, instead of v_{n}(\nabla y)\in\CMs(CM) like for I_{2} , the Carleson estimate ( 3.24 )). For the L^{\infty} -bound on \mathbf{B} , we observe that 0=\int_{\mathbb{R}^{n-1}}\big(I_{n\times(n-1)}-(v_{n})^{T}\pi(v_{n})\big)\nabla_{y}\eta_{t}\big(y-z\big)\,dz, and so \mathbf{B}=\int_{\mathbb{R}^{n-1}}\big(I_{n\times(n-1)}-(v_{n})^{T}\pi(v_{n})\big)\nabla_{y}\eta_{t}\big(y-z\big)[h(z)-a^{\prime}]\,dz\quad\text{for any }a^{\prime}\in\mathbb{R}^{n}. Then the estimate \|\mathbf{B}\|_{\infty}\leq C_{M}\epsilon is a simpler variant of the L^{\infty} -bound of \nabla\lambda treated in Step 1. Moreover, observe that v_{n}\big(I_{n\times(n-1)}-(v_{n})^{T}\pi(v_{n})\big)=\pi(v_{n})-v_{n}(v_{n})^{T}\pi(v_{n})=0, which entails that v_{n}\mathbf{B}=0 . We want now to derive a Carleson measure estimate for \nabla\mathbf{B} . Let \{\mathbf{b}_{j}\}_{1\leq j\leq n} be the n rows of \mathbf{B} , and we want thus to study the quantities \nabla\mathbf{b}_{j} . Write v_{n}=:\big((v_{n})_{1},\dots,(v_{n})_{n}\big) and let \{e_{1},\dots,e_{n}\} be the Cartesian orthonormal basis. Then \mathbf{b}_{j}(X)=\int_{\mathbb{R}^{n-1}}h(z)\Big<\pi(e_{j}-(v_{n})_{j}v_{n}),\nabla_{y}\eta_{t}(y-z)\Big>\,dz. Differentiating in X_{i} , we have by the chain rule that |\partial_{i}\mathbf{b}_{j}(X)|\leq\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}\partial_{t}\eta_{t}\big(y-z\big)h(z)\,dz\right||\pi(e_{j}-(v_{n})_{j}v_{n})||\partial_{i}t|\\ +\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}^{2}\eta_{t}\big(y-z\big)h(z)\,dz\right||\pi(e_{j}-(v_{n})_{j}v_{n})||\nabla y|\\ +\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}\eta_{t}(y-z)h(z)\,dz\right||\partial_{i}[(v_{n})_{j}v_{n}]|=:K_{1}+K_{2}+K_{3}. The term K_{3} is simple: we have that K_{3}=\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}\eta_{t}(y-z)[h(z)-a]\,dz\right||\partial_{i}[(v_{n})_{j}v_{n}]|\quad\text{for any }a\in\mathbb{R}^{n}, and we can prove that \left|\int_{\mathbb{R}^{n-1}}[h(z)-a]\nabla_{y}\eta_{t}(y-z)\,dz\right|\lesssim\epsilon, with the same strategy as in Step 1, and ( 3.20 ) gives that t\,\partial_{i}[(v_{n})_{j}v_{n}]\in\CMs(CM) . Hence t|K_{3}|\in\CMs(CM\epsilon) as desired. For the two other terms K_{1} and K_{2} , observe that \|\pi(e_{j}-(v_{n})_{j}v_{n})\|_{\infty}\leq 2 and \|\partial_{i}t\|_{\infty}+\|\nabla y\|_{\infty}\lesssim(1+M) by ( 3.25 ). So proving that t\left(K_{1}+K_{2}\right)\in\CMs(C_{M}\epsilon) reduces to proving that t|K_{11}|+t|K_{21}|\in\CMs(C_{M}\epsilon) , where K_{11}:=\int_{\mathbb{R}^{n-1}}\nabla_{y}\partial_{t}\eta_{t}\big(y-z\big)h(z)\,dz and K_{12}:=\int_{\mathbb{R}^{n-1}}\nabla_{y}^{2}\eta_{t}\big(y-z\big)h(z)\,dz. The strategy is the same as the one used for I_{11} . Observe that \nabla_{y}\eta_{t}(y-z)=-\nabla_{z}\eta_{t}(y-z)\text{ and }\int_{\mathbb{R}^{n-1}}\partial_{t}\eta_{t}\,dz=\int_{\mathbb{R}^{n-1}}\nabla_{y}\eta_{t}\,dz=0. Hence, by the divergence theorem, for any a\in\mathbb{R}^{n} and any constant matrix B\in\mathbb{R}^{(n-1)\times n} , \int_{\mathbb{R}^{n-1}}(a+zB)\nabla_{y}\partial_{t}\eta_{t}(y-z)\,dz=\int_{\mathbb{R}^{n-1}}(a+zB)\nabla^{2}_{y}\eta_{t}(y-z)\,dz=0, and thus |K_{11}|\leq\inf_{a\in\mathbb{R}^{n},B\in\mathbb{R}^{(n-1)\times n}}\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}\partial_{t}\eta_{t}\big(y-z\big)\big[h(z)-(a+zB)\big]\,dz\right|\\ \lesssim\inf_{a\in\mathbb{R}^{n},B\in\mathbb{R}^{(n-1)\times n}}\frac{1}{t^{2}}\fint_{B(y,t)}|h(z)-(a+zB)|\,dz=\frac{\beta(y,t)}{t} and similarly |K_{12}|\leq\inf_{a\in\mathbb{R}^{n},B\in\mathbb{R}^{(n-1)\times n}}\left|\int_{\mathbb{R}^{n-1}}\nabla_{y}^{2}\eta_{t}\big(y-z\big)\big[h(z)-(a+zB)\big]\,dz\right|\\ \lesssim\inf_{a\in\mathbb{R}^{n},B\in\mathbb{R}^{(n-1)\times n}}\frac{1}{t^{2}}\fint_{B(y,t)}|h(z)-(a+zB)|\,dz=\frac{\beta(y,t)}{t}. Then we can argue as in Step 2 starting from ( 3.28 ) to obtain that t|K_{11}|+t|K_{21}|\in\CMs(C_{M}\epsilon) . The theorem follows. \square 3.3. Construction of the change of variables Let \{e_{1},\dots,e_{n}\} be the Cartesian basis, \{v_{i},\dots,v_{n}\} be the orthonormal basis constructed in Lemma 3.11 , and \lambda be the perturbation defined in ( 3.16 )–( 3.18 ). We define then the vector \bar{w}_{i} as (3.29) \bar{w}_{i}(X):=v_{i}(X)[I+\mathbf{B}(X)]\quad\text{for }X\in\Omega_{0}, where \mathbf{B} is the matrix-valued function from Theorem 3.23 , (iii) . Lemma 3.30 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g:\mathbb{R}^{n-1}\to\mathbb{R} is a M -Lipschitz function, and let \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} be a bi-Lipschitz change of variables satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon . There exists \epsilon_{2}>0 such that if \epsilon\in(0,\epsilon_{2}] , then there exists an orthonormal basis \{w_{1},\dots,w_{n}\} satisfying (i) For each i\in\{1,\dots,n\} , \Span\{w_{1},\dots,w_{i}\}=\Span\{\bar{w}_{1},\dots,\bar{w}_{i}\} , (ii) There exists C>0 such that, for any i,j\in\{1,\dots,n\} , we have 3 3 3 Recall that I_{ij} are the coefficients of the n\times n identity matrix. |\left-I_{ij}|\leq C\epsilon , (iii) There exists C>0 such that, for any i,j\in\{1,\dots,n\} , we have t\left|\nabla\langle w_{i},v_{j}\rangle\right|\in\CMs(C\epsilon), where t is as defined in ( 3.18 ). In the lemma, \epsilon_{2} and C depend on M (on top of the the dimension n and the choice of \eta ). Proof: The lemma is obtained again with the Gram-Schmidt process. We define by induction (3.31) \tilde{w}_{1}=\bar{w}_{1},\quad\tilde{w}_{i}:=\bar{w}_{i}-\sum_{1\leq kw_{k}\quad\text{for }2\leq i\leq n,\quad\text{ and }\quad w_{i}:=\dfrac{\tilde{w}_{i}}{|\tilde{w}_{i}|}. We prove by induction on i\in\{1,\dots,n\} that (1) \left<\tilde{w}_{i},v_{i}\right>\geq 1-C\epsilon , (2) \big||\left|-1\big|\leq C\epsilon , (3) for all j\in\{1,\dots,n\}\setminus\{i\} , |\left|\leq C\epsilon , and (4) for all j\in\{1,\dots,n\} , we have t\nabla\left\in\CMs(C\epsilon) . Note that (1) implies that |\tilde{w}_{i}|\geq 1-C\epsilon , which entails that the construction ( 3.31 ) is well defined. It is obvious from the construction that (i) holds and that \left\{w_{1},\dots,w_{n}\right\} is an orthonormal basis. Initial Step: We have \tilde{w}_{1}=\bar{w}_{1} , and so \left<\tilde{w}_{1},v_{1}\right>=1+\left\geq 1-C\epsilon by the L^{\infty} bound of \mathbf{B} in (iii) of Theorem 3.23 , which yields conclusion (1). We also have |\tilde{w}_{1}|\leq|v_{1}|+|\mathbf{B}|\leq 1+C\epsilon Altogether, \big||\tilde{w}_{1}|-1\big|\leq C\epsilon , and then \big||\left|-1\big|\leq\left|\frac{|\left<\tilde{w}_{1},v_{1}\right>|-1}{|\tilde{w}_{1}|}\right|+\left|\frac{|\tilde{w}_{1}|-1}{|\tilde{w}_{1}|}\right|\lesssim\epsilon as long as \epsilon_{2} is chosen small enough such that |\tilde{w}_{1}|\geq 1-C\epsilon_{2}\geq\frac{1}{2} . Since w_{1} is a unit vector and \{v_{1},\dots,v_{n}\} is an orthonormal base, the fact that \big||\left|-1\big|\leq C\epsilon immediately implies that \left|\leq C\epsilon when j\neq 1 . So we have proved (2) and (3) when i=1 . Finally |\nabla\left<\tilde{w}_{1},v_{j}\right>|=|\nabla\big[\left\big]|\lesssim|\nabla\mathbf{B}|+|\mathbf{B}|(|\nabla v_{1}|+|\nabla v_{j}|), meaning that t\nabla\left<\tilde{w}_{1},v_{j}\right>\in\CMs(C\epsilon) by Theorem 3.23 and Corollary 3.15 . Therefore, t\,\nabla\left[|\tilde{w}_{1}|^{2}\right]=t\,\nabla\left[\sum_{i=1}^{n}|\left<\tilde{w}_{1},v_{j}\right>|^{2}\right]\in\CMs(C\epsilon) and, since |\tilde{w}_{1}|\geq\frac{1}{2} by our choice of \epsilon_{2} , t\,\nabla\big[\left\big]=t\,\nabla\left[\frac{\left<\tilde{w}_{1},v_{j}\right>}{\sqrt{|\tilde{w}_{1}|^{2}}}\right]\in\CMs(C\epsilon). The case i=1 follows. Induction step: Let i\in\{2,\dots,n\} and assume that (1)–(4) are true for all 1\leq k=1+\left-\sum_{k|^{2}-\sum_{k\left\\ \geq 1-i\,|\mathbf{B}|-\sum_{k|^{2}\geq 1-C\epsilon by (iii) of Theorem 3.23 and (3) of the induction hypothesis. A similar argument yields that |\tilde{w}_{i}|\leq 1+C\epsilon , which means, with the same reasoning as in the initiation step, that conclusions (2) and (3) hold too once we take \epsilon_{2} small enough. To prove (4), we observe that \left<\tilde{w}_{i},v_{j}\right>=I_{ij}+\left-\sum_{k\left-\sum_{k\left\\ =I_{ij}+\left-\sum_{k\left-\sum_{k\left\left, where we have written w_{k}=\sum_{\ell=1}^{n}\langle w_{k},v_{\ell}\rangle v_{\ell} in the last equality. Direct computation shows that |\nabla\left<\tilde{w}_{i},v_{j}\right>|\lesssim|\nabla\mathbf{B}|+|\mathbf{B}|\left(\sum_{\ell=1}^{n}|\nabla v_{\ell}|\right)+(1+|\mathbf{B}|)\sum_{k|, and thus t\nabla\left<\tilde{w}_{i},v_{j}\right>\in\CMs(C\epsilon) by Theorem 3.23 and (4) of the induction hypothesis. With the same argument as in the initiation step, we deduce that t\nabla\left<\tilde{w}_{i},v_{j}\right>\in\CMs(C\epsilon) The induction step and the lemma follows. \square In addition to properties (i)-(iii) in Lemma 3.30 , we shall also need the following estimate for \nabla w_{i} : for i\in\left\{1,\dots,n\right\} , (3.32) t\left|\nabla w_{i}\right|\in\CMs(CM), which can be deduced by expressing w_{i}=\sum_{j=1}^{n}\langle w_{i},v_{j}\rangle v_{j} and then using ( 3.20 ) and (iii) of Lemma 3.30 . Let \mathcal{O}=\mathcal{O}(X) be the rotation matrix that sends the orthonormal basis \{v_{1},\dots,v_{n}\} constructed in Lemma 3.11 to the orthonormal basis \{w_{1},\dots,w_{n}\} constructed in Lemma 3.30 . Written in matrix multiplication, it means that (3.33) w_{i}=v_{i}\mathcal{O}\text{ for all }i\in\left\{1,\dots,n\right\},\quad\text{or }W=V\mathcal{O}, where V is the n\times n (orthogonal) matrix with row vectors v_{1} , \dots , v_{n} , and W is the n\times n (orthogonal) matrix with row vectors w_{1} , \dots , w_{n} . Lemma 3.34 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g:\mathbb{R}^{n-1}\to\mathbb{R} is a M -Lipschitz function. Let \epsilon_{2} be the value given in Lemma 3.30 , and let \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} be a bi-Lipschitz change of variables satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon_{2} . Then the matrix \mathcal{O} defined as in ( 3.33 ) verifies the following conditions: (i) |\mathcal{O}-I_{n\times n}|\leq C\epsilon , (ii) t\nabla\mathcal{O}\in\CMs(C\epsilon) , where t is defined as in ( 3.18 ). In both cases, the constant C depends on M , but not on \epsilon . Proof: This lemma is actually a simple corollary of Lemma 3.30 . Let \mathcal{O}_{v} be the transition matrix that express \left\{w_{1},\dots,w_{n}\right\} in terms of \left\{v_{1},\dots,v_{n}\right\} , that is, \mathcal{O}_{v}V=W , and in particular, its ij -th entry is \langle w_{i},v_{j}\rangle . Lemma 3.30 then gives that |\mathcal{O}_{v}-I_{n\times n}|\leq C\epsilon and t\nabla\mathcal{O}_{v}\in\CMs(C\epsilon) . Since \mathcal{O}=V^{-1}\mathcal{O}_{v}V , we have that |\mathcal{O}-I_{n\times n}|=|\mathcal{O}_{v}-I_{n\times n}|\leq C\epsilon . By Corollary 3.15 , t\left|\nabla V\right|\in\CMs(C_{M}) . So we get that t|\nabla\mathcal{O}|=t|\nabla(\mathcal{O}-I_{n\times n})|=t\left|\nabla\left[V^{T}(\mathcal{O}_{v}-I_{n\times n})V\right]\right|\leq 2t|\nabla V||\mathcal{O}_{v}-I_{n\times n}|+t|\nabla O_{v}|\in\CMs(C\epsilon). The lemma follows. \square We set the matrix function A_{||}:\,\Omega_{0}\to\mathbb{R}^{(n-1)^{2}} and the function a:\Omega_{0}\to\mathbb{R} to be such that (3.35) A_{||}(X):=\begin{bmatrix}\left<\bar{w}_{1},w_{1}\right>&\dots&\left<\bar{w}_{1},w_{n-1}\right>\\ \vdots&\ddots&\vdots\\ \left<\bar{w}_{n-1},w_{1}\right>&\dots&\left<\bar{w}_{n-1},w_{n-1}\right>\end{bmatrix}=(\left<\bar{w}_{i},w_{j}\right>)_{1\leq i,j\leq n-1} and (3.36) a(X):=\det(A_{||}(X)). Lemma 3.37 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g:\mathbb{R}^{n-1}\to\mathbb{R} is a M -Lipschitz function. Let \epsilon_{2} be the value given in Lemma 3.30 , and let \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} be a bi-Lipschitz change of variables satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon_{2} . Then the matrix J_{0} defined as (3.38) J_{0}(X):=\begin{bmatrix}A_{||}(X)&\mathbf{0}^{T}\\ \mathbf{0}&a(X)\end{bmatrix},\quad\text{where }\mathbf{0}=0_{\mathbb{R}^{n-1}}=(0,\dots,0),\quad\text{for }X\in\Omega_{0} verifies \|J_{0}-I\|_{\infty}\leq C\epsilon and t|\nabla J_{0}|\in\CMs(C\epsilon) , where t is defined as in ( 3.18 ). The constant C depends on M but not on \epsilon . Proof: Recall that \bar{w}_{i}:=v_{i}(I+\mathbf{B}). So \|\left<\bar{w}_{i},w_{j}\right>-I_{ij}\|_{\infty}\leq\|\left-I_{ij}\|_{\infty}+\|\mathbf{B}\|_{\infty}\lesssim\epsilon by (ii) of Lemma 3.30 and (iii) of Theorem 3.23 . It proves that \|A_{||}-I_{(n-1)\times(n-1)}\|_{\infty}\lesssim\epsilon . Moreover, we have (3.39) t\Big|\nabla\big[\left<\bar{w}_{i},w_{j}\right>\big]\Big|\lesssim t\Big|\nabla\big[\left\big]\Big|+t\|\mathbf{B}\|_{\infty}\left(|\nabla v_{i}|+|\nabla w_{j}|\right)+t|\nabla\mathbf{B}|\in\CMs(C\epsilon) by Lemma 3.30 , Theorem 3.23 , Corollary 3.15 , and ( 3.32 ). It proves that t|\nabla A_{||}|\in\CMs(C\epsilon) . Since a is the determinant of A_{||}=(a_{ij})_{1\leq i,j\leq n-1} , it can be written as \displaystyle a-1 \displaystyle=\left(\prod_{i=1}^{n-1}a_{ii}\right)-1+\sum_{\sigma\in\mathfrak{S}_{n-1}\atop\sigma\neq id}\sgn(\sigma)\prod_{i=1}^{n-1}a_{i,\sigma(i)} \displaystyle=\sum_{k=1}^{n-1}(a_{kk}-1)\left(\prod_{i=k+1}^{n-1}a_{ii}\right)+\sum_{\sigma\in\mathfrak{S}_{n-1}\atop\sigma\neq id}\sgn(\sigma)\prod_{i=1}^{n-1}a_{i,\sigma(i)}, where we have used the convention that \prod_{i=n}^{n-1}a_{ii}=1 . Observe that each product above contains at least one element of A_{||}-I . Combined to the fact that the entries of A_{||} are uniformly bounded (by 1+C\epsilon ), we get |a-1|\lesssim\|A_{||}-I\|_{\infty}\lesssim\epsilon. Since a is a linear combination of product of elements of A_{||} , we also easily have t|\nabla a|\lesssim\sum_{i=1}^{n-1}\sum_{j=1}^{n-1}t\Big|\nabla\big[\left<\bar{w}_{i},w_{j}\right>\big]\Big|\in\CMs(C\epsilon) by ( 3.39 ). The lemma follows. \square Given \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g:\mathbb{R}^{n-1}\to\mathbb{R} is a M -Lipschitz function. Given a bi-Lipschitz map \Phi satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon , we now define the change of variables \rho from \Omega_{0} to \mathbb{R}^{n} as (3.40) \rho(X)=X-t(X)v_{n}(X)+\lambda(X)+a(X)t(X)v_{n}(X)\mathcal{O}(X), where \lambda(X) and t(X) are defined in ( 3.16 )–( 3.18 ), a defined in ( 3.36 ), and \mathcal{O} defined in ( 3.33 ). With a lighter notation, it becomes \displaystyle\rho(X) \displaystyle=X-tv_{n}+\lambda+atv_{n}\mathcal{O} (3.41) \displaystyle=X+\lambda+atv_{n}(\mathcal{O}-I)+(a-1)tv_{n}. Let \Omega=\Phi(\Omega_{0}) . Observe that for X\in\partial\Omega_{0} , t(X)=0 , y(X)=\pi(X) , and \lambda(X)=h(\pi(X)) , where h is defined as in ( 3.17 ). So \rho(X)=X+h(\pi(X))=\Phi(X) for X\in\partial\Omega_{0} , which implies that \rho(\partial\Omega_{0})=\partial\Omega . We want to prove that \rho is a bijection between \Omega_{0} and \Omega . Moreover, we want to derive estimates for the Jacobian matrix of the map \rho in the bases \{v_{1},\dots,v_{n}\} and \{w_{1},\dots,w_{n}\} , which we define as the matrix J . That is, the ij -th entry of the Jacobian matrix J is (3.42) J_{ij}=\left<\partial_{v_{i}}\rho,w_{j}\right>. Theorem 3.43 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g is a M -Lipschitz function. There exists \epsilon_{2}=\epsilon_{2}(M)\in(0,1] such that for any \epsilon\in[0,\epsilon_{2}] and for any bi-Lipschitz change of variables \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon , the map \rho defined in ( 3.40 ) is a diffeomorphism between \Omega_{0} and \Omega:=\Phi(\Omega_{0}) . Moreover, the Jacobian matrix J defined in ( 3.42 ) satisfies (i) \|J-I_{n\times n}\|_{\infty}\leq C\epsilon , (ii) J_{1}:=J-J_{0}\in\CMs(C\epsilon) , where the Carleson condition is with respect to domain \Omega_{0} . In both cases, the constant C>0 depends on M , but not on \epsilon . Proof: Step 1: We prove that \rho(\Omega_{0})\subset\Omega when \epsilon_{2} is sufficiently small. Fix any X\in\Omega_{0} . First, we claim that (3.44) |\lambda(X)-h(\pi(X))|\leq Ct(X)\epsilon, where we recall that h is the C_{M}\epsilon_{2} -Lipschitz function defined as h(z):=\Phi(z,g(z))-(z,g(z)) . Indeed, from the expression of \lambda in ( 3.16 ), and since \int\eta_{t}=1 , we have |\lambda(X)-h(\pi(X))|=\int_{\mathbb{R}^{n-1}}|h(z)-h(\pi(X))|\eta_{t}(y-z)\,dz\lesssim\fint_{B(y,t)}|h(z)-h(\pi(X))|dz\\ \lesssim\epsilon\max\{t,|y-\pi(X)|\} since h is \epsilon -Lipschitz. However, |y-\pi(X)|=|\pi(X-tv_{n})-\pi(X)|\leq t , so |\lambda(X)-h(\pi(X))|\lesssim\epsilon t as desired. We have then, by ( 3.3 ), that (3.45) |\rho(X)-X-h(\pi(X))|\leq|\lambda(X)-h(\pi(X)|+at|\mathcal{O}-I_{n\times n}|+t|a-1|\lesssim\epsilon t by ( 3.44 ), (i) of Lemma 3.34 , and Lemma 3.37 . We now look at the point X+h(\pi(X)) . Let z=\pi(X) and let s be the distance between X and its projection on \partial\Omega_{0} in the direction of e_{n} . Then we can write X=(z,g(z))+(0,s) , which shows that X+h(\pi(X))=\Phi(z,g(z))+(0,s). So if y\in\mathbb{R}^{n} , we have |X+h(\pi(X))-\Phi(y,g(y))|=|(z-y,g(z)-g(y)+s)+h(z)-h(y)|\\ \geq\sqrt{|z-y|^{2}+|s+g(z)-g(y)|^{2}}-C\epsilon_{2}|z-y|, since h is C\epsilon_{2} -Lipschitz. If moreover, |y-z|\geq s/2M , then of course |X+h(\pi(X))-\Phi(y,g(y))|\geq(1-C\epsilon_{2})|z-y|\gtrsim s, if \epsilon_{2} is small enough. But if |y-z|\leq s/2M , then |g(z)-g(y)|\leq s/2 , meaning that |X+h(\pi(X))-\Phi(y,g(y))|\geq|s|-|g(z)-g(y)|\gtrsim s. Altogether, we have \dist(X+h(\pi(X)),\partial\Omega)\gtrsim s . With the obvious or previously known estimates, we obtain that (3.46) Z:=X+h(\pi(X))\in\mathbb{R}^{n}\setminus\partial\Omega\quad\text{ and }\quad\dist(Z,\partial\Omega)\approx s\approx\dist(X,\partial\Omega_{0})\approx t(X), with the implicit constants independent of \epsilon_{2}\leq 1 . The combination of ( 3.45 ) and ( 3.46 ) gives that (3.47) \rho(X)\in B(Z,C\epsilon_{2}t)\subset B(\rho(X),\dist(X,\partial\Omega)/2)\subset\Omega, if \epsilon_{2} is small enough. Step 2: Decomposition of J . Direct computation shows that \partial_{v_{i}}\rho=v_{i}+(\partial_{v_{i}}\lambda)+atv_{n}\partial_{v_{i}}\mathcal{O}+(\partial_{v_{i}}t)(av_{n}\mathcal{O}-v_{n})+t\big[\partial_{v_{i}}v_{n}\big](a\mathcal{O}-I)+(\partial_{v_{i}}a)tv_{n}\mathcal{O}. We have v_{i}+(\partial_{v_{i}}\lambda)=\bar{w}_{i}+v_{i}\mathbf{C} , where we recall that \nabla\lambda=\mathbf{B}+\mathbf{C} is the decomposition given in Theorem 3.23 and \bar{w}_{i} is given in ( 3.29 ). Moreover, v_{n}\mathcal{O}=w_{n} by the definition on \mathcal{O} in ( 3.33 ). So we can write (3.48) \partial_{v_{i}}\rho=\bar{w}_{i}+v_{i}\mathbf{C}+atv_{n}\partial_{v_{i}}\mathcal{O}+\left<\nabla t-v_{n}^{T},v_{i}\right>(aw_{n}-v_{n})+\left(aw_{n}-v_{n})\\ +t\big[\partial_{v_{i}}v_{n}\big](a\mathcal{O}-I)+(\partial_{v_{i}}a)tw_{n}. We claim that the entries of the matrix J_{0} (defined in ( 3.38 )) are (3.49) (\widetilde{J_{0}})_{ij}:=\Big<\bar{w}_{i}+\left(aw_{n}-v_{n}),w_{j}\Big>. Indeed, • When i\in\{1,\dots,n-1\} and j\in\{1,\dots,n\} , we have \left=0 and thus (\widetilde{J_{0}})_{ij}=\left<\bar{w}_{i},w_{j}\right>. So (\widetilde{J_{0}})_{ij}=(A_{||})_{ij}=(J_{0})_{ij} when i=0 whenever \ell>k , which means particularly that \langle\bar{w}_{i},w_{n}\rangle=0 when i=1 , so (\widetilde{J_{0}})_{nj}:=\left=aI_{nj}. So the last line of \widetilde{J_{0}} equals that of J_{0} . By ( 3.48 ) and ( 3.49 ), one sees that the matrix J_{1} has entries (J_{1})_{ij}:=\Big(aw_{n}-v_{n})+t\big[\partial_{v_{i}}v_{n}\big](a\mathcal{O}-I)+(\partial_{v_{i}}a)tw_{n},\,w_{j}\Big>. To prove that J_{1}\in\CMs(C\epsilon) , observe that |(J_{1})_{ij}|\lesssim\underbrace{|\mathbf{C}|}_{\in\CMs(C\epsilon)}+\underbrace{t|\nabla\mathcal{O}|}_{\in\CMs(C\epsilon)}+\underbrace{|\nabla t-v_{n}^{T}|}_{\in\CMs(C)}\underbrace{\left(|a-1|+|w_{n}-v_{n}|\right)}_{\leq C\epsilon}\\ +\underbrace{\left(|a-1|+|\mathcal{O}-I|\right)}_{\leq C\epsilon}\underbrace{t|\nabla v_{n}|}_{\in\CMs(C)}+\underbrace{t|\nabla a|}_{\in\CMs(C\epsilon)}\in\CMs(C\epsilon), by Theorem 3.23 , ( 3.20 ), ( 3.21 ), Lemma 3.30 , Lemma 3.34 , and Lemma 3.37 . We have proved that J satisfies the conclusion (ii) of the Theorem. To prove (i) , observe that J_{1}\in\CMs(C\epsilon) implies in particular \|J_{1}\|_{\infty}\lesssim\epsilon . In addition, \|J_{0}-I\|_{\infty}\lesssim\epsilon by Lemma 3.37 . So we conclude that \|J-I\|_{\infty}\leq\|J_{0}-I\|_{\infty}+\|J_{1}\|_{\infty}\lesssim\epsilon, as desired. Step 3: \rho is a diffeomorphism. Note that \rho is smooth on \Omega_{0} , because it is essentially constructed from the Green function G which is itself smooth on \Omega_{0} . So we just need to prove that \rho is a bijection. The first key element is that the Jacobian matrix of \rho is invertible and close to identity. Indeed, by ( 3.42 ) and ( 3.33 ), the Jacobian matrix \nabla\rho of \rho is (3.50) \nabla\rho=PJP^{-1}\mathcal{O}, where P is the orthogonal matrix that changes the Cartesian base \{e_{1},\dots,e_{n}\} into \{v_{1},\dots,v_{n}\} . Step 2 and Lemma 3.34 give us that then \|J-I_{n\times n}\|_{\infty}+\|\mathcal{O}-I_{n\times n}\|_{\infty}\lesssim\epsilon, so if \epsilon_{2} is small enough, (3.51) \|\nabla\rho-I_{n\times n}\|_{\infty}\leq\frac{1}{2}. It means that the map \rho is locally invertible at any point. This implies the degree of \rho defined as N_{\rho}(\mathcal{Z}):=\#\{X\in\Omega_{0},\,\rho(X)=\mathcal{Z}\} is locally constant (positive, integer, not necessarily finite). Therefore N_{\rho} is constant on connected components of \mathbb{R}^{n}\setminus\partial\Omega . We want to show that N_{\rho}\equiv 1 on \Omega and N_{\rho}\equiv 0 on \mathbb{R}^{n}\setminus\overline{\Omega} . One strategy is to construct the family of bi-Lipschitz changes of variables \Phi_{\theta}:=\theta\Phi+(1-\theta)I. One can see from our explicit construction that \rho=\rho_{\Phi} is continuous in \Phi , so \rho_{\Phi_{\theta}} is continuous in \theta , and eventually that N_{\rho_{\theta}}(\Phi_{\theta}(X))\in\mathbb{Z} is continuous in \theta . Since \Phi_{0} and then \rho_{0} are the identity maps, we have that N_{\rho_{\theta}}(\Phi_{\theta}(X))=N_{\rho_{0}}(\Phi_{0}(X))=\left\{\begin{array}[]{ll}1&\text{ if }X\in\Omega_{0}\\ 0&\text{ if }X\in\mathbb{R}^{n}\setminus\overline{\Omega_{0}}.\end{array}\right. The theorem follows. \square 4. Proof of the main theorem and corollary 4.1. Proof of Theorem 1.3 When \rho is a bi-Lipschitz change of variables from \Omega_{0} to \Omega , we say that L_{\rho} is the conjugate of -\Delta by \rho if for any harmonic function u in \Omega , u_{\rho}:=u\circ\rho is a weak solution to L_{\rho}u_{\rho}=0 in \Omega_{0} . It is a simple exercise to see (see for instance (6.21) in [ DFM19 ] ) that L_{\rho}:=-\diver A_{\rho}\nabla with (4.1) A_{\rho}(X):=\det(\nabla\rho(X))(\nabla\rho)^{-T}(X)(\nabla\rho)^{-1}(X) where \nabla\rho is the Jacobian matrix of \rho . From there, we have the following result: Lemma 4.2 . Let \Omega_{0}:=\{(x,t)\in\mathbb{R}^{n},\,t>g(x)\} , where g is a M -Lipschitz function, and write G for the Green function with pole at infinity for -\Delta on \Omega_{0} . There exist constants \epsilon_{2}=\epsilon_{2}(n,M)\in(0,1] and C_{2}=C_{2}(n,M)>0 such that the following holds. For any \epsilon\in[0,\epsilon_{2}] , take any bi-Lipschitz map \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} satisfying \|\nabla\Phi-I\|_{\infty}\leq\epsilon , and set \Omega=\Phi(\Omega_{0}) . Let \rho=\rho_{\Phi} be the map defined in ( 3.40 ). Let L_{\rho} be the conjugate of -\Delta by \rho . There exists a symmetric operator L_{0}:=-\diver A_{0}\nabla on \Omega_{0} such that (i) for all \xi\in\mathbb{R}^{n} and X\in\Omega_{0} (4.3) \frac{1}{2}|\xi|^{2}\leq A_{0}(X)\xi\cdot\xi\leq 2|\xi|^{2}; (ii) \|A_{0}-I\|_{\infty}\leq C_{2}\epsilon ; (iii) |A_{\rho}-A_{0}|\in\CMs(C_{2}\epsilon) , where the Carleson condition is with respect to domain \Omega_{0} ; (iv) L_{0}G=0 in \Omega_{0} . Proof: Take \epsilon_{2} small enough such that Theorem 3.43 applies. Let P=P(X) be the orthogonal matrix that changes the base \{e_{1},\dots,e_{n}\} into \{v_{1},\dots,v_{n}\} . Recall that \nabla\rho is given by \nabla\rho=PJP^{-1}\mathcal{O} as in ( 3.50 ). So the matrix A_{\rho} is then given by A_{\rho}(X)=\det(J)PJ^{-T}J^{-1}P^{-1}, because the matrices P and \mathcal{O} are orthogonal. The choice of A_{0} —hence of L_{0} —is given by A_{0}:=\det(J_{0})P(J_{0})^{-T}(J_{0})^{-1}P^{-1}. From the construction and the fact that P is an orthogonal matrix, it follows that A_{0} is symmetric. Recall that \|J_{0}-I\|_{\infty}\leq C\epsilon by Lemma 3.37 . So—assuming that \epsilon_{2} is small enough—the matrix A_{0} is elliptic, and since the determinant is a multilinear form, we have \|\det(J_{0})-1\|_{\infty}\leq C\epsilon. By writing A_{0}-I=(\det(J_{0})-\det(I))PJ_{0}^{-T}J_{0}^{-1}P^{-1}+P(J_{0}^{-1}+I)J_{0}^{-1}(I-J_{0})P^{-1}, the above estimates give that \|A_{0}-I\|_{\infty}\leq\|\det(J_{0})-\det(I)\|_{\infty}\|P(J_{0})^{-T}(J_{0})^{-1}P^{-1}\|_{\infty}\\ +\|P\|_{\infty}\|P^{-1}\|_{\infty}\|(J_{0})^{-1}\|_{\infty}\big(\|I\|_{\infty}+\|(J_{0})^{-1}\|_{\infty}\big)\|J_{0}-I\|_{\infty}\leq C\epsilon because all the considered matrices are uniformly bounded in X\in\Omega_{0} . So by choosing \epsilon_{2} small enough, we can guarantee that A_{0} satisfies the ellipticity condition ( 4.3 ). Since we also have that J-J_{0}\in\CMs(C\epsilon) and \|J-J_{0}\|_{\infty}\leq C\epsilon by Lemma 3.43 , the multilinearity of the determinant also gives that |\det(J)-\det(J_{0})|\in\CMs(C\epsilon). We conclude that |A_{\rho}-A_{0}|\leq|\det(J)-\det(J_{0})|\|PJ^{-T}J^{-1}P^{-1}\|_{\infty}\\ +\|P\|_{\infty}\|P^{-1}\|_{\infty}\|\det(J_{0})\|_{\infty}\|J^{-1}\|_{\infty}\|(J_{0})^{-1}\|_{\infty}\big(\|J^{-1}\|_{\infty}+\|(J_{0})^{-1}\|_{\infty}\big)|J_{0}-J|\\ \in\CMs(C\epsilon). It remains to prove that L_{0}G=0 . This is true because of our careful construction of \rho . More precisely, we constructed \rho to preserve the structure of \Omega_{0} , so that G is close to be a solution to L_{\rho}u=0 . We claim that for j\in\left\{1,\dots,n\right\} , (4.4) \left=I_{jn}. Indeed, \left is the coefficients of the last column of A_{0} in the base \{v_{1},\dots,v_{n}\} . The matrix A_{0} in the base \{v_{1},\dots,v_{n}\} is given by A_{0,v}:=P^{-1}A_{0}P=\det(J_{0})(J_{0})^{-T}(J_{0})^{-1}. However, the matrix J_{0} - defined in ( 3.38 ) - is diagonal by block, so it is easy to compute the last column of A_{0,v} : it is given by (0,\dots,0,\det(J_{0})a^{-2})^{T}=(0,\dots,0,1)^{T} since \det(J_{0})=a\det(A_{||})=a^{2} . The claim ( 4.4 ) gives that A_{0}v_{n}^{T}=v_{n}^{T} , which implies that A_{0}\nabla G=\nabla G since v_{n}=\nabla G^{T}/|\nabla G| . It means that L_{0}G=-\diver A_{0}\nabla G=-\Delta G=0\quad\text{in }\Omega_{0}, since G is the Green function (for -\Delta ) with pole at infinity. The lemma follows. \square Proof of Theorem 1.3 . Let \Omega_{0}\subset\mathbb{R}^{n} and p\in(1,\infty) be as in the theorem. Let G_{-\Delta} and \omega_{-\Delta} be the Green function and harmonic measure with pole at infinity for \Omega_{0} defined in Lemma 2.5 . Since the L^{p} Dirichlet problem for -\Delta is solvable in \Omega_{0} , there is a constant C_{p}>0 such that the reverse Hölder inequality ( 2.13 ) holds. • Let \epsilon_{1}=\epsilon_{1}(n,p,C_{p},M,2) be the constant given by Theorem 2.16 , that is the constant under which Carleson perturbations preserves the solvability of the Dirichlet problem in L^{p} if as long as the operator in \Omega_{0} before perturbation verifies the reverse Hölder bound ( 2.13 ) and is uniformly elliptic with constant 2 . • Let \epsilon_{2}=\epsilon(n,M) and C_{2}=C_{2}(n,M) be the constants given by Lemma 4.2 . Define \epsilon_{0}:=\min\{\epsilon_{2},\epsilon_{1}/C_{2}\}. Take then a bi-Lipschitz change of variables \Phi verifying \|\nabla\Phi-I\|_{\infty}\leq\epsilon_{2} and set \Omega:=\Phi(\Omega_{0}) . We can construct the bi-Lipschitz change of variables \rho as in ( 3.40 ) that maps \Omega_{0} to \Omega , and then the conjugate L_{\rho} of -\Delta by \rho as in ( 4.1 ). Take then L_{0} as in Lemma 4.2 . With all the necessary tools being introduced, the argument is then straightforward. (1) Since L_{0}G_{-\Delta}=-\Delta G_{-\Delta}=0 in \Omega_{0} , one has G_{-\Delta}=G_{L_{0}} , that is, G_{-\Delta} is the Green function with pole at infinity for L_{0} in \Omega_{0} . It means in particular that L_{0} satisfies the reverse Hölder inequality ( 2.13 ) with the same constant C_{p} as the Laplacian. (2) By Lemma 4.2 , we have that L_{0} is uniformly elliptic with constant 2, and |A_{\rho}-A_{0}|\in\CMs(C_{2}\epsilon_{0}) . But since C_{2}\epsilon_{0}\leq\epsilon_{1} , Theorem 2.16 gives that the Dirichlet problem (for L_{\rho} in \Omega_{0} ) is also solvable for boundary data in L^{p} . (3) Since a bi-Lipschitz change of variables preserves the solvability of the L^{p} Dirichlet problem, see Proposition 2.17 , the L^{p} Dirichlet problem for -\Delta in \Omega is also solvable. Succinctly, the proof is {\lx@inpgf@ignorespaces(D_{p})_{-\Delta,\Omega_{0}}} {\lx@inpgf@ignorespaces(D_{p})_{L_{0},\Omega_{0}}} {\lx@inpgf@ignorespaces(D_{p})_{L_{\rho},\Omega_{0}}} {\lx@inpgf@ignorespaces(D_{p})_{-\Delta,\Omega}} \scriptstyle{\lx@inpgf@ignorespaces\tiny G_{-\Delta}=G_{L_{0}}} Lem 4.2 (iii) \scriptstyle+ Thm 2.12 \scriptstyle{\lx@inpgf@ignorespaces\tiny A_{0}-A_{\rho}\in\CMs(\epsilon_{1})} Thm 2.16 \scriptstyle{\lx@inpgf@ignorespaces\tiny\rho\text{ bi-Lipschitz}} Prop 2.17 Theorem 1.3 follows. \square 4.2. Proof of Corollary 1.10 Let \Omega\subset\mathbb{R}^{n} be a bounded strongly quasiconvex domain. Take X_{0} to be a center of \Omega , that is, a point such that \dist(X_{0},\partial\Omega)\approx\diam(\Omega) , and set M to be the Lipschitz constant that appears in Definition 1.8 . Let G^{X_{0}} and \omega^{X_{0}} be, respectively, the Green function and harmonic measure with pole at X_{0} for \Omega . Let p\in(1,\infty) . Let C_{p}=C_{p}(M) be the reverse Hölder constant which is uniform to M -Lipschitz convex domains. Let \epsilon_{0}:=\epsilon_{0}(n,p,C_{p},M)>0 be as in Theorem 1.3 , and then r_{0}>0 as in Definition 1.8 . Without loss of generality, we can assume that r_{0}<\frac{\dist(X_{0},\partial\Omega)}{2} . We will prove that there exists m\in\mathbb{N} and C>0 such that for any x_{0}\in\partial\Omega and any r\in(0,2^{-m}r_{0}] , we have the reverse Hölder inequality (4.5) \left(\fint_{B(x_{0},r)\cap\partial\Omega}\left(\kappa^{X_{0}}\right)^{p^{\prime}}\,d\sigma\right)^{\frac{1}{p^{\prime}}}\leq C\fint_{B(x_{0},r)\cap\partial\Omega}\kappa^{X_{0}}\,d\sigma, where \kappa^{X_{0}}(x):=\limsup_{X\in\gamma(x)\atop X\to x}\frac{G^{X_{0}}(X)}{\delta(X)} is defined as in ( 2.14 ). The claim ( 4.5 ) is sufficient to prove Corollary 1.10 . Indeed, by Theorem 2.12 , the Corollary would follow if we had ( 4.5 ) for all r\in(0,\diam\Omega) , instead of of only r\in(0,2^{-m}r_{0}] . However, the reverse Hölder bounds for r\in(2^{-m}r_{0},\diam\Omega) can easily deduced from the reverse Hölder bounds for r=2^{-m}r_{0} and a covering argument. Take x_{0}\in\partial\Omega . Let g=g_{x_{0}} be the M -Lipschitz function and \Phi=\Phi_{x_{0}} be the bi-Lipschitz change of variables given by Definition 1.8 (for a coordinate system that also depend on x_{0} ), that is such that \Omega\cap B(x_{0},r_{0})=\underbrace{\Phi_{x_{0}}(\{(x,t)\in\mathbb{R}^{n},\,t>g_{x_{0}}(x)\})}_{=:\Omega_{x_{0}}}\cap B(x_{0},r_{0}). We write \sigma_{x_{0}} for the surface measure on \partial\Omega_{x_{0}} , G_{x_{0}} for the Green function for -\Delta with pole at infinity in the domain \Omega_{x_{0}} normalized so that (4.6) G_{x_{0}}(X_{x_{0},m})=G^{X_{0}}(X_{x_{0},m}), where X_{x_{0},m}\in B(x_{0},r_{0}/2)\cap\Omega satisfies \dist(X_{x_{0},m},\partial\Omega)=\dist(X_{x_{0},m},\partial\Omega_{x_{0}})\approx|X_{x_{0},m}-x_{0}|\approx 2^{-m}r_{0}, m\geq 1 to be determined. We write \kappa_{x_{0}} for \kappa_{x_{0}}(x):=\limsup_{X\in\gamma_{x_{0}}(x)\atop X\to x}\frac{G_{x_{0}}(X)}{\dist(X,\partial\Omega_{x_{0}})},\qquad x\in\partial\Omega_{x_{0}}. Here, \gamma_{x_{0}}(x):=\{Z\in\Omega_{x_{0}},\,|Z-x|\leq 2\dist(Z,\partial\Omega_{x_{0}}) as expected. By definition, the domain \Omega_{0,x_{0}}:=\left\{(x,t)\in\mathbb{R}^{n}:\,t>g_{x_{0}}(x)\right\} is a convex Lipschitz graph domain, and so Theorem B.6 applies. It means that the L^{p} Dirichlet problem for -\Delta is solvable in \Omega_{0,x_{0}} , and that the constant C_{p} in ( 2.13 ) is independent of x_{0} even though \Omega_{0,x_{0}} itself depends on x_{0} . Therefore, by Theorem 1.3 , the L^{p} Dirichlet problem for -\Delta is solvable in \Omega_{x_{0}} . Then Theorem 2.12 entails that there exists a constant C^{\prime}_{p}:=C^{\prime}_{p}(n,p,C_{p},M) —so independent of x_{0} —such that we have the reverse Hölder bound (4.7) \left(\fint_{B(x,r)\cap\partial\Omega_{x_{0}}}|\kappa_{x_{0}}|^{p^{\prime}}\,d\sigma\right)^{\frac{1}{p^{\prime}}}\leq C^{\prime}_{p}\fint_{B(x,r)\cap\partial\Omega_{x_{0}}}|\kappa_{x_{0}}|\,d\sigma\qquad\text{ for all }x\in\partial\Omega_{x_{0}},\,r>0. This last bound is not too far from our objective ( 4.5 ), we just need to drop the dependence of \kappa in x_{0} . We do this using the comparison principle. Observe first that for x and r such that B(x,r)\subset B(x_{0},r_{0}) , one has B(x,r)\cap\partial\Omega_{x_{0}}=B(x,r)\cap\partial\Omega , and that (4.8) \dist(Z,\partial\Omega_{x_{0}})=\dist(Z,\partial\Omega)\quad\text{ whenever $Z\in B(x_{0},3r_{0}/2)$, $\dist(Z,\partial\Omega)0 small enough so that C\epsilon_{0}(M+1)<1 , we can apply the implicit function theorem to solve the above equation and get that t=f(x) for some Lipschitz function f:\mathbb{R}^{n-1}\to\mathbb{R} . Hence, \partial\Omega can be expressed as \left\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R}:t=g(x+\tilde{\psi}(x,f(x)))+\tilde{h}(x,f(x))\right\}. We let \psi(x)=\tilde{\psi}(x,f(x)) , h(x)=\tilde{h}(x,f(x)) , and check that they have small gradients as desired. ∎ Proof of Remark 1.9 . Let \Omega be a strongly quasiconvex domain and let x_{0}=0\in\partial\Omega . By definition, for any \epsilon_{0}>0 , there exists r>0 , a bi-Lipschitz map \Phi with \left\|\nabla\Phi-I\right\|\leq\epsilon_{0} , and an M -Lipschitz function g such that \Omega\cap B(0,r)=\Phi(\{t>g(x)\})\cap B(0,r) . Let \Psi:\mathbb{R}^{n}\to\mathbb{R}^{n} be a C^{1} diffeomorphism. We want to show that \Psi(\Omega) is strongly quasiconvex. We can assume without loss of generality that \Psi(0)=0=\Phi(0) . For some r_{0}\in(0,r) , we have \Psi(\Omega)\cap B(0,r_{0})=\Psi\circ\Phi(\{t>g(x)\})\cap B(0,r_{0}). Define a nondegenerate linear transform L:\mathbb{R}^{n}\to\mathbb{R}^{n} , L(X)=\nabla\Psi(0)X=X_{i}\,\partial_{i}\Psi_{j}(0) . Write \Psi\circ\Phi(X)=\Psi\circ\Phi\circ(L^{-1}(L(X))). Note that the L transforms convex domains into convex domains. Thus up to a change of coordinates, L(\{t>g(x)\})=\{t>\hat{g}(x)\} , where \hat{g} is also a convex function. Hence \Psi(\Omega)\cap B(0,r_{0})=\Psi\circ\Phi\circ L^{-1}(\{t>\hat{g}(x)\}). Finally note that \Phi(X)=X+O(\epsilon_{0})|X| is close to the identity transform and that L^{-1} is close to \Psi^{-1} in a neighborhood of 0 since \Psi(X)=LX+o(|X|) (using the fact that \Psi is C^{1} ). Thus by a messy but straightforward computation, we can show that |\nabla(\Psi\circ\Phi\circ L^{-1})-I|\leq C\epsilon_{0} if r_{0} is small enough. This means that \Psi(\Omega) is strongly quasiconvex. ∎ Appendix B Solvability of the L^{p} Dirichlet problem under C^{1,\alpha} diffeomorphism of convex domains In this appendix, we show by a simple method that if the boundary of a bounded domain \Omega (or an unbounded Lipschitz graph domain) is locally (or globally, in the unbounded case)“a C^{1,\alpha} diffeomorphism of a convex gragh”, then the L^{p} Dirichlet problem (D)_{p} is solvable for all 1\phi(x)\}) , where \phi:\mathbb{R}^{n-1}\to\mathbb{R} is a convex function with Lipschitz constant M , and \Phi:\mathbb{R}^{n}\to\mathbb{R}^{n} is a C^{1,\alpha} diffeomorphism. Assume \phi(0)=0 (i.e., 0\in\partial\Omega ). Let L=-\diver A\nabla be an elliptic operator with coefficients in C^{\beta}(\Omega) for some \beta>0 . Let u be a weak solution to Lu=0 in \Omega\cap B_{1}(0) with u=0 on \partial\Omega\cap B_{1}(0) . Then (B.2) \|\nabla u\|_{L^{\infty}(\Omega\cap B_{1/2})}\leq C\|u\|_{L^{\infty}(\Omega\cap B_{1})}, where C>0 depends on n , M , the ellipticity constant, \left\|\Phi\right\|_{C^{1,\alpha}(B_{1})} , and \left\|A\right\|_{C^{\beta}(\Omega\cap B_{1})} . Proof. Let v=u\circ\Phi . Then v satisfies L_{\Phi}v=-\diver(A_{\Phi}\nabla v)=0 in \Phi^{-1}(\Omega\cap B_{1}) , where A_{\Phi}=|\det\nabla\Phi|(\nabla\Phi)^{-T}(A\circ\Phi)(\nabla\Phi)^{-1}. Clearly, A_{\Phi} also satisfies the uniform ellipticity condition and C^{\alpha_{1}} continuous (since \Phi is a C^{1,\alpha} diffeomorphism), with \alpha_{1}=\min\{\alpha,\beta\}>0 . Hence, the desired estimate ( B.2 ) is equivalent to (B.3) \|\nabla v\|_{L^{\infty}(\Phi^{-1}(\Omega\cap B_{1/2}))}\leq C\|v\|_{L^{\infty}(\Phi^{-1}(\Omega\cap B_{1}))}. But now \Phi^{-1}(\Omega\cap B_{1})=\{(x,t):t>\phi(x)\}\cap\Phi^{-1}(B_{1}) . By assumption, the part \{(x,t):t=\phi(x)\}\cap\Phi^{-1}(B_{1}) is convex. The gradient estimate ( B.3 ) over convex boundaries is more or less well-known; see [ GT83 , Chapter 14.2] for elliptic operator in nondivergence form. We provide a proof for our elliptic operator in divergence form for the reader’s convenience. Up to a translation and rotation, we can assume that \Phi(0)=0 , \phi(0)=0 , and by the convexity of \phi , that \phi(x)\geq 0 for x\in\mathbb{R}^{n-1} . We claim: for t>0 and (0,t)\in\Phi^{-1}(B_{3/4})\subset\subset\Phi^{-1}(B_{1}) , we have (B.4) |v(0,t)|\leq Ct\|v\|_{L^{\infty}(\Phi^{-1}(\Omega\cap B_{1}))} This is proved by the comparison principle. We construct a barrier function in the following way. Let M=\|v\|_{L^{\infty}(\Phi^{-1}(\Omega\cap B_{1}))} . Let w be a barrier function satisfying L_{\Phi}(w)=0 in \{t>0\}\cap\Phi^{-1}(B_{1}) , w=0 on \{t=0\}\cap\Phi^{-1}(B_{1}) and w=M on \{t>0\}\cap\partial\Phi^{-1}(B_{1}) . Clearly, 0\leq w\leq M in \{t>0\}\cap\Phi^{-1}(\Phi) . The key fact we will use from the convexity of \phi(x) is that \{t>\phi(x)\}\cap\Phi^{-1}(B_{1})\subset\{t>0\}\cap\Phi^{-1}(B_{1}) (due to the fact \phi(x)\geq 0 ). Therefore, we have |v|=0\leq w on the convex part of the boundary \{t=\phi(x)\}\cap\Phi^{-1}(B_{1}) and |v|\leq M=w on the rest of the boundary \{t>\phi(x)\}\cap\partial\Phi^{-1}(B_{1}) . By the comparison principle, we have |v|\leq w in \{t>\phi(x)\}\cap\Phi^{-1}(B_{1}) . By the Schauder estimate of w over a flat boundary \{t=0\} , we have |w(x,t)|\leq CtM for any (x,t)\in\{t>0\}\cap\Phi^{-1}(B_{3/4}) . This particularly implies ( B.4 ). Finally, by applying the previous argument for ( B.4 ) to all the boundary points on \{t=\phi(x)\}\cap\Phi^{-1}(B_{3/4}) , we actually have |v(X)|\leq C\delta(X)M , where \delta(X) is the distance from X to the boundary \{t=\phi(x)\} . Hence, by the interior Schauder estimate, for any X\in\{t>\phi(x)\}\cap\Phi^{-1}(B_{1/2}) , we have |\nabla v(X)|\leq C\delta(X)^{-1}\|v\|_{L^{\infty}(B(X,c\delta(X)))}\leq CM. This ends the proof of ( B.3 ). ∎ Remark B.5 . The previous lemma (and thus the next theorem) is also valid if \Phi is a C^{1,{\rm Dini}} -type diffeomorphism, thanks to the well-known gradient estimate for elliptic operators with Dini continuous coefficients; see [ Lie86 ] . However, the argument does not work if \Phi is merely C^{1} , due to the classical failure of the boundedness of gradient for the elliptic operator L_{\Phi}=-\diver(A_{\Phi}\nabla) with continuous coefficients. Theorem B.6 . Let \Omega be either a bounded Lipschitz domain, or an unbounded Lipschitz graph domain in a form of \Omega=\left\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R}:t>\phi_{0}(x)+\phi_{1}(x)\right\} , where \phi_{0} is a convex function and \|\nabla\phi_{0}\|_{L^{\infty}(\mathbb{R}^{n-1})}+\|\phi_{1}\|_{L^{\infty}(\mathbb{R}^{n-1})}+\|\nabla\phi_{1}\|_{L^{\infty}(\mathbb{R}^{n-1})}\leq M . Suppose that there exists r_{0}>0 such that for any x_{0}\in\partial\Omega , there exists a coordinate basis \{e_{1},\dots,e_{n}\} , an M -Lipschitz convex function g_{x_{0}} and a C^{1,\alpha} diffeomporphism \Phi_{x_{0}}:\,\mathbb{R}^{n}\to\mathbb{R}^{n} with |\nabla\Phi_{x_{0}}|+|\nabla\Phi_{x_{0}}^{-1}|\leq M such that \Omega\cap B(x_{0},r_{0})=\Phi_{x_{0}}\Big(\{(x,t)\in\mathbb{R}^{n-1}\times\mathbb{R},\,t>g_{x_{0}}(x)\}\Big)\cap B(x_{0},r_{0}). Then (D)_{p} for -\Delta is solvable for all 11 , there exists a constant C>0 such that (B.7) \left(\frac{1}{\sigma(\Delta(X,r))}\int_{\Delta(X,r)}k(y)^{q}d\sigma(y)\right)^{1/q}\leq\frac{C}{\sigma(\Delta(X,r))}\int_{\Delta(X,r)}k(y)d\sigma(y)=C\frac{\omega(\Delta(X,r))}{\sigma(X,r)} for all X\in\partial\Omega , r>0 , where k(y)=\frac{d\omega}{d\sigma}(y) and \omega is the harmonic measure with pole at infinity (see Lemma 2.5 ). Since \phi is Lipschitz with constant at most M , for any y=(y^{\prime},\phi(y^{\prime}))\in\Delta(X,r) , we have k(y)=\lim_{s\to 0}\frac{\omega(\Delta(y,s))}{\sigma(\Delta(y,s))}\lesssim\limsup_{s\to 0}\frac{G(y^{\prime},s+\phi(y^{\prime}))}{s} by Lemma 2.11 , where G is the Green function for -\Delta with pole at infinity. Apply Lemma B.1 to u=G in B(X,r_{0})\cap\Omega , we get that (for s>0 sufficiently small) (B.8) \frac{G(y^{\prime},s+\phi(y^{\prime}))}{s}\leq\frac{C\left\|G\right\|_{L^{\infty}(B(X,r)\cap\Omega)}}{r}, for any 0r_{0} , we need a large-scale Lipschitz estimate, which has been established in [ Zhu21 , Theorem 4.2] if the boundary graph is a “convex + bounded” function (i.e., \partial\Omega=\{t=\phi_{0}(x)+\phi_{1}(x)\} ), and in our case can be formed as \left\|G\right\|_{L^{\infty}(B(X,r_{0})\cap\Omega)}\leq\frac{C\left\|G\right\|_{L^{\infty}(B(X,r)\cap\Omega)}}{r} for all r>r_{0} . Combing the local estimate with the large-scale Lipschitz estimate, we obtain ( B.8 ) for all r>s>0 . By the Bourgain estimate (see e.g. [ JK82a , Lemma 4.4] ), one has \left\|G\right\|_{L^{\infty}(B(X,2r)\cap\Omega)}\leq G(A_{X,r}) , where A_{X,r} is a corkscrew point relative to X and r , that is, A_{X,r}\in B(X,Mr)\cap\Omega satisfies \dist(A_{X,r},\partial\Omega)\approx|A_{X,r}-X|\approx r . By Lemma 2.11 again, we have that \frac{G(A_{X,r})}{r}\approx\frac{\omega(\Delta(X,r))}{\sigma(X,r)} . Putting everything together, one has \left(\frac{1}{\sigma(\Delta(X,r))}\int_{\Delta(X,r)}k(y)^{q}d\sigma(y)\right)^{1/q}\leq\frac{C\left\|G\right\|_{L^{\infty}(B(X,2r)\cap\Omega)}}{r}\leq C\frac{\omega(\Delta(X,r))}{\sigma(X,r)} as desired. Case 2: when \Omega is bounded. It suffices to show that for any q>1 , there exists a constant C>0 such that (B.9) \left(\frac{1}{\sigma(\Delta(X,r))}\int_{\Delta(X,r)}k^{X_{0}}(y)^{q}d\sigma(y)\right)^{1/q}\leq\frac{C}{\sigma(\Delta(X,r))}\int_{\Delta(X,r)}k^{X_{0}}(y)d\sigma(y)=C\frac{\omega^{X_{0}}(\Delta(X,r))}{\sigma(X,r)} for all X\in\partial\Omega , 0g_{X}(x)\right\}) where g_{X} is convex with Lipschitz constant at most M . By a covering argument, it suffices to show ( B.9 ) for 0