Introduction
1.1. Background. Let Ω ⊂ R n + 1 \Omega\subset\mathbb{R}^{n+1} Ω ⊂ R n + 1 , n ≥ 1 n \ge1 n ≥ 1 , be a bounded domain, and set σ : = H n ∣ ∂ Ω \sigma:= \mathcal{H}^n|_{\partial\Omega} σ := H n ∣ ∂ Ω . Given a function g g g on ∂ Ω \partial\Omega ∂ Ω with ∫ ∂ Ω g d σ = 0 \int_{\partial\Omega} g\,d\sigma= 0 ∫ ∂ Ω g d σ = 0 , the Neumann problem consists in finding a harmonic function u u u in Ω \Omega Ω whose outer normal derivative equals g g g on ∂ Ω \partial\Omega ∂ Ω . In Lipschitz domains one usually asks that ∇ u \nabla u ∇ u has non-tangential limits σ \sigma σ -almost everywhere on ∂ Ω \partial\Omega ∂ Ω and that the normal component of these limits equals g g g ,
2020 Mathematics Subject Classification. 35J25, 31B20, 42B37, 28A75, 35B27.
Key words and phrases. Neumann problem, non-tangential maximal function, bi-Lipschitz domains, chord-arc domains, Kenig’s problem, homogenization, potential flow.
M.M. was supported by IKERBASQUE and partially supported by the project PID2024-157724NB-I00, funded by MICIU/AEI/10.13039/501100011033/FEDER, EU, by IT-1615-22, and IT1875-26 (Basque Government). C.P. was supported by the Ramón y Cajal fellowship grant RYC2023-045598-I (MCIU, Spain), and partially supported by the grants PID2023-150984NB-I00 and PID2024-155320NB-I00 funded by MICIU/AEI/10.13039/501100011033/FEDER, EU X.T. was partially supported by MICIU (Spain) under the grant PID2024160507NB-I00. see [24 ] . In rougher domains it is more convenient to understand the boundary condition in the weak sense, that is, to ask that ∇ u ∈ L 2 ( Ω ) \nabla u \in L^{2}(\Omega) ∇ u ∈ L 2 ( Ω ) and
∫ Ω ∇ u ⋅ ∇ φ d X = ∫ ∂ Ω g φ d σ , for every φ ∈ C c ∞ ( R n + 1 ) . \int_{\Omega} \nabla u \cdot\nabla\varphi\,dX = \int_{\partial\Omega} g \varphi\,d\sigma, \qquad\text{for every } \varphi\in C_{c}^{\infty}(\mathbb{R}^{n+1}). ∫ Ω ∇ u ⋅ ∇ φ d X = ∫ ∂ Ω g φ d σ , for every φ ∈ C c ∞ ( R n + 1 ) .
In all the domains considered in this paper, for every g ∈ L 2 ( σ ) g \in L^{2}(\sigma) g ∈ L 2 ( σ ) with zero mean such a solution exists and is unique up to additive constants, see Subsection 1.2, and we call it the variational solution with datum g g g .
Definition 1.1. Given 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , we say that the L p L^{p} L p Neumann problem is solvable in Ω \Omega Ω , and we write that ( N p ) (N_{p}) ( N p ) holds in Ω \Omega Ω , if there exists C p < ∞ C_{p} < \infty C p < ∞ such that, for every g ∈ L p ( σ ) ∩ L 2 ( σ ) g \in L^{p}(\sigma) \cap L^{2}(\sigma) g ∈ L p ( σ ) ∩ L 2 ( σ ) with ∫ ∂ Ω g d σ = 0 \int_{\partial\Omega} g \,d\sigma= 0 ∫ ∂ Ω g d σ = 0 , the variational solution u u u with datum g g g satisfies
∥ N ( ∇ u ) ∥ L p ( σ ) ≤ C p ∥ g ∥ L p ( σ ) , (1) \|N(\nabla u)\|_{L^{p}(\sigma)} \leq C_{p} \|g\|_{L^{p}(\sigma)}, \tag*{(1)} ∥ N ( ∇ u ) ∥ L p ( σ ) ≤ C p ∥ g ∥ L p ( σ ) , (1) where N N N is the non-tangential maximal function, see (2) below.
In Lipschitz domains the L 2 L^{2} L 2 Neumann problem was solved by D. Jerison and C. Kenig [20 ] , by means of a Rellich identity which shows that the normal and the tangential derivatives of a harmonic function have comparable L 2 ( σ ) L^{2}(\sigma) L 2 ( σ ) norms, and by B. Dahlberg and C. Kenig [10 ] for 1 < p < 2 + ε 1 < p < 2 + \varepsilon 1 < p < 2 + ε , a range which is sharp, see [23 pp. 153-154] . Thus, in a Lipschitz domain the L p L^{p} L p Neumann problem can only fail for large values of p p p . We also refer to G. Verchota’s work [Ve84] for the method of layer potentials and the monograph [24 ] , as well as C. Kenig and J. Pipher [25 ] for related results for elliptic operators with non-smooth coefficients, and R. Brown [6 ] , who extended the H 1 H^{1} H 1 estimate of [10 ] to data in the Hardy spaces H p H^{p} H p with p < 1 p < 1 p < 1 close to 1.
Beyond Lipschitz domains much less is known. In planar chord-arc domains the solvability of the Neumann problem in L p L^{p} L p is equivalent to that of the regularity problem in L p L^{p} L p and of the Dirichlet problem in L p ′ L^{p'} L p ′ , see [21 , 33 ] . In higher dimensions, S. Hofmann, M. Mitrea and M. Taylor [18 ] solved the Neumann problem for every 1 < p < ∞ 1 < p < \infty 1 < p < ∞ in regular Semmes-Kenig-Toro domains whose outer unit normal has small BMO norm. We also refer to the work of J. J. Marín, J. M. Martell, D. Mitrea, I. Mitrea and M. Mitrea [MMMMM22] for related results in sufficiently flat Ahlfors regular domains. In every bounded chord-arc domain the Dirichlet problem is solvable in L q L^{q} L q for some 1 < q < ∞ 1 < q < \infty 1 < q < ∞ [9 Se90] , and the regularity problem is solvable in L p L^{p} L p for some 1 < p < ∞ 1 < p < \infty 1 < p < ∞ [32 ] . It has also been recently shown in [31 ] that this holds for elliptic operators satisfying the Dahlberg-Kenig-Pipher condition.
Whether the same holds for the Neumann problem when n ≥ 2 n \geq2 n ≥ 2 is a question posed by Kenig [24 Problem 3.2.2] , see also the ICM proceedings [40 Question 2.5] and [39 Question 7.5] . The first and third named authors proved in [33 ] the solvability of the Neumann problem in L p L^{p} L p , 1 < p < 2 1 < p < 2 1 < p < 2 , in chord-arc domains with very big pieces of superdomains on which the problem is uniformly solvable, under additional assumptions on the regularity problem and on the boundary. J. Feneuil and L. Li [11 ] and S. Hofmann and D. Sparrow [HS25] studied the Neumann function and the Poisson-Neumann problem in chord-arc domains. In the opposite direction, it has been shown in [29 ] that the solvability of the L p L^{p} L p Neumann problem in a corkscrew domain with Ahlfors regular boundary implies the uniform rectifiability of the boundary.
In this paper we answer Kenig’s question in the negative by constructing an explicit counterexample. Our domain is the image of a ball under a bi-Lipschitz homeomorphism of R n + 1 \mathbb{R}^{n+1} R n + 1 , and so the Neumann problem may indeed fail in a class of domains which is much smaller than that of chord-arc domains. It is obtained from a cube by removing, near one of its faces, several generations of thin “columns” at smaller and smaller scales, each generation standing on the “walls” of the previous one, see Figure 1 .
Figure 1. Schematic picture of the domain near the bottom face of the cube after a few generations of columns: (a) in perspective, (b) in a vertical section. The solids are not part of the domain. The first generation of columns (grey) stands on the bottom face; the second generation (blue) stands on selected squares of the walls of the first one, perpendicularly to them; a few columns of the third generation (green) are shown.
The boundary datum is very simple: it equals 1 and − 1 -1 − 1 on two opposite faces of the cube, and it vanishes elsewhere.
Main results
For a domain Ω ⊂ R n + 1 \Omega\subset\mathbb{R}^{n+1} Ω ⊂ R n + 1 , we write δ Ω ( X ) : = dist ( X , ∂ Ω ) \delta_{\Omega}(X) := \operatorname{dist}(X,\partial\Omega) δ Ω ( X ) := dist ( X , ∂ Ω ) and σ Ω : = H n ∣ ∂ Ω \sigma_{\Omega} := \mathcal{H}^{n}|_{\partial\Omega} σ Ω := H n ∣ ∂ Ω . For a > 0 a>0 a > 0 and x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω we consider the cone Γ Ω , a ( x ) : = { X ∈ Ω : ∣ X − x ∣ < ( 1 + a ) δ Ω ( X ) } \Gamma_{\Omega,a}(x):=\{X\in\Omega:|X-x|<(1+a)\delta_{\Omega}(X)\} Γ Ω , a ( x ) := { X ∈ Ω : ∣ X − x ∣ < ( 1 + a ) δ Ω ( X )} and, for a continuous function f f f in Ω \Omega Ω , the non-tangential maximal function
N a f ( x ) : = sup X ∈ Γ Ω , a ( x ) ∣ f ( X ) ∣ , N : = N 1 . (2) \mathcal{N}_{a}f(x):=\sup_{X\in\Gamma_{\Omega,a}(x)}|f(X)|,\qquad\mathcal{N}:=\mathcal{N}_{1}. \tag*{(2)} N a f ( x ) := X ∈ Γ Ω , a ( x ) sup ∣ f ( X ) ∣ , N := N 1 . (2)
If Ω \Omega Ω is the image of a ball (or, equivalently, of a cube) under a bi-Lipschitz homeomorphism of R n + 1 \mathbb{R}^{n+1} R n + 1 , then W 1 , 2 ( Ω ) W^{1,2}(\Omega) W 1 , 2 ( Ω ) has a bounded trace operator Tr : W 1 , 2 ( Ω ) → L 2 ( σ Ω ) \operatorname{Tr}:W^{1,2}(\Omega)\to L^{2}(\sigma_{\Omega}) Tr : W 1 , 2 ( Ω ) → L 2 ( σ Ω ) and a Poincaré inequality with boundary means, see Lemma 2.1 . Hence, for every g ∈ L 2 ( σ Ω ) g\in L^{2}(\sigma_{\Omega}) g ∈ L 2 ( σ Ω ) with zero mean there is a function u ∈ W 1 , 2 ( Ω ) u\in W^{1,2}(\Omega) u ∈ W 1 , 2 ( Ω ) , unique up to additive constants, such that
∫ Ω ∇ u ⋅ ∇ φ d X = ∫ ∂ Ω g Tr φ d σ Ω , for every φ ∈ W 1 , 2 ( Ω ) . (3) \int_{\Omega}\nabla u\cdot\nabla\varphi\,dX=\int_{\partial\Omega}g\operatorname{Tr}\varphi\,d\sigma_{\Omega},\qquad\text{for every }\varphi\in W^{1,2}(\Omega). \tag*{(3)} ∫ Ω ∇ u ⋅ ∇ φ d X = ∫ ∂ Ω g Tr φ d σ Ω , for every φ ∈ W 1 , 2 ( Ω ) . (3)
Since the restrictions to Ω \Omega Ω of the functions in C ∞ ( R n + 1 ) C^{\infty}(\mathbb{R}^{n+1}) C ∞ ( R n + 1 ) are dense in W 1 , 2 ( Ω ) W^{1,2}(\Omega) W 1 , 2 ( Ω ) , see Lemma 2.1 , u u u is the variational solution with datum g g g introduced in Subsection 1.1, and ( N p ) (N_{p}) ( N p ) is understood with this notion of solution. Notice that the class of data in the definition of ( N p ) (N_{p}) ( N p ) is dense in the subspace of L p ( σ Ω ) L^{p}(\sigma_{\Omega}) L p ( σ Ω ) of functions with zero mean; in any case, all the data constructed in this paper are bounded.
The main result of this paper is the following.
Theorem 1.2. For every n ≥ 2 n\geq2 n ≥ 2 there exist a bounded domain Ω ⊂ R n + 1 \Omega\subset\mathbb{R}^{n+1} Ω ⊂ R n + 1 , a bi-Lipschitz homeomorphism T : R n + 1 → R n + 1 T:\mathbb{R}^{n+1}\to\mathbb{R}^{n+1} T : R n + 1 → R n + 1 such that Ω = T ( B ( 0 , 1 ) ) \Omega=T(B(0,1)) Ω = T ( B ( 0 , 1 )) , and a function g ∈ L ∞ ( σ ) g\in L^{\infty}(\sigma) g ∈ L ∞ ( σ ) with ∣ g ∣ ≤ 1 |g|\leq1 ∣ g ∣ ≤ 1 and ∫ ∂ Ω g d σ = 0 \int_{\partial\Omega}g\,d\sigma=0 ∫ ∂ Ω g d σ = 0 , such that the variational solution u u u with datum g g g satisfies
∫ ∂ Ω N a ( ∇ u ) d σ = ∞ , for every a > 0. (4) \int_{\partial\Omega}\mathcal{N}_{a}(\nabla u)\,d\sigma=\infty,\qquad\text{for every }a>0. \tag*{(4)} ∫ ∂ Ω N a ( ∇ u ) d σ = ∞ , for every a > 0. (4) In particular, ( N p ) (N_{p}) ( N p ) fails in Ω \Omega Ω for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ .
Since Ω \Omega Ω is a bi-Lipschitz image of a ball, it is a two-sided chord-arc domain and ∂ Ω \partial\Omega ∂ Ω supports a weak ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality, see Lemma 9.3. Thus, the Dirichlet problem is solvable in Ω \Omega Ω in L q L^{q} L q for some 1 < q < ∞ 1<q<\infty 1 < q < ∞ , and the regularity problem in L q ′ L^{q'} L q ′ . For the Dirichlet problem, this goes back to B. Dahlberg [8 ] in Lipschitz domains, and to G. David and D. Jerison [9 ] and S. Semmes [Se90] in chord-arc domains. For the Neumann problem there is no such comparison between a domain and its subdomains, see Subsection 1.3. In corkscrew domains with Ahlfors regular boundaries, the solvability of the Dirichlet problem in L q L^{q} L q for some q < ∞ q<\infty q < ∞ was characterized geometrically in [2 ] , and the first and third named authors showed in [32 ] that it implies the solvability of the regularity problem in L q ′ L^{q'} L q ′ . All these results are quantitative, and bi-Lipschitz homeomorphisms of R n + 1 \mathbb{R}^{n+1} R n + 1 map chord-arc domains to chord-arc domains, see [13 , 27 ] and the proof of Lemma 9.3. So, if we deform the unit ball by a bi-Lipschitz homeomorphism of R n + 1 \mathbb{R}^{n+1} R n + 1 , the Dirichlet and the regularity problems remain solvable for some exponents, with constants depending only on n n n and on the bi-Lipschitz constant. Theorem 1.2 shows that the Neumann problem behaves in a substantially different way: ( N p ) (N_{p}) ( N p ) holds in the unit ball for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ , while it fails in its bi-Lipschitz image Ω \Omega Ω for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ . In other words, unlike the solvability of the Dirichlet and of the regularity problems, the solvability of the L p L^{p} L p Neumann problem for some p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) is not invariant under bi-Lipschitz homeomorphisms of R n + 1 \mathbb{R}^{n+1} R n + 1 .
This sharp difference is somewhat surprising. As mentioned above, in planar chord-arc domains the Neumann, the regularity and the Dirichlet problems are equivalent [21 , 33 ] . We also recall that, by the Lipschitz Schönflies theorem of P. Tukia [Tu80], every bounded planar domain bounded by a chord-arc curve is the image of the unit disc under a bi-Lipschitz homeomorphism of R 2 \mathbb{R}^{2} R 2 . Since in such domains the L p L^{p} L p Neumann problem is solvable for some p > 1 p>1 p > 1 , by the results just quoted, the analogue of Theorem 1.2 fails in the plane, and the assumption n ≥ 2 n\ge2 n ≥ 2 in that theorem cannot be removed. Moreover, in Lipschitz domains of any dimension the Neumann and the regularity problems go hand in hand: they are both solvable for 1 < p < 2 + ε 1<p<2+\varepsilon 1 < p < 2 + ε [20 Ve84, 10 ] , and in both cases the starting point is the Rellich identity. Theorem 1.2 shows that in higher dimensions the two problems decouple, even in domains with very good geometric properties, and that the Neumann problem may fail in a very strong way: the datum is bounded, while N ( ∇ u ) \mathcal{N}(\nabla u) N ( ∇ u ) is not even integrable on ∂ Ω \partial\Omega ∂ Ω .
As mentioned above, bi-Lipschitz images of a ball form a class of domains which is much smaller than that of chord-arc domains. Indeed, in higher dimensions the boundary of a chord-arc domain need not be a topological sphere and, even when it is, the existence of bi-Lipschitz parametrizations of surfaces with good geometric properties is a delicate matter. We refer, for instance, to [41 ] , [37 ] , and also [17 ] for the behaviour of several classes of domains under bi-Lipschitz maps. We also note that ∂ Ω \partial\Omega ∂ Ω is uniformly rectifiable, so that the result of [29 ] mentioned above is not contradicted: uniform rectifiability is necessary, but very far from sufficient, for the solvability of the Neumann problem. Something similar happens with the Rellich inequalities. The one-sided inequality, which bounds the normal derivative of a harmonic function by its tangential derivative, holds in corkscrew domains with uniformly rectifiable boundaries (see [12 ] and [32 ] ), while the reverse one fails in Ω \Omega Ω , see Remark 1.5-(c).
Moreover, we mention that Ω \Omega Ω is locally flat outside a very small set. Namely, there is a closed set K ⊂ ∂ Ω K\subset\partial\Omega K ⊂ ∂ Ω with H n ( K ) = 0 \mathcal{H}^{n}(K)=0 H n ( K ) = 0 such that every point of ∂ Ω ∖ K \partial\Omega\setminus K ∂ Ω ∖ K has a neighbourhood in which Ω \Omega Ω coincides with a half-space bounded by a hyperplane parallel to a coordinate hyperplane, see Remarks 8.1 and 9.2.
The domain Ω \Omega Ω is obtained as the limit of a decreasing sequence of domains, each of them being a cube modified finitely many times. From these finite stages we also get the following quantitative result.
Theorem 1.3. Let n ≥ 2 n\ge2 n ≥ 2 . There is L 0 ≥ 1 L_{0}\ge1 L 0 ≥ 1 such that, for every M > 0 M>0 M > 0 , there exist a bounded Lipschitz domain Ω M ⊂ R n + 1 \Omega_{M}\subset\mathbb{R}^{n+1} Ω M ⊂ R n + 1 , whose boundary consists of finitely many rectangles parallel to coordinate hyperplanes, an L 0 L_{0} L 0 -bi-Lipschitz homeomorphism G M G_{M} G M of R n + 1 \mathbb{R}^{n+1} R n + 1 with Ω M = G M ( ( − 1 , 1 ) n + 1 ) \Omega_{M}=G_{M}((-1,1)^{n+1}) Ω M = G M (( − 1 , 1 ) n + 1 ) , and a datum g M g_{M} g M which equals 1 1 1 and − 1 -1 − 1 on the two faces of Ω M \Omega_{M} Ω M contained in { x 1 = 1 } \{x_{1}=1\} { x 1 = 1 } and { x 1 = − 1 } \{x_{1}=-1\} { x 1 = − 1 } and vanishes elsewhere, such that the variational solution u M u_{M} u M satisfies
∥ N ( ∇ u M ) ∥ L 1 ( σ Ω M ) ≥ M . \|\mathcal{N}(\nabla u_{M})\|_{L^{1}(\sigma_{\Omega_{M}})}\ge M. ∥ N ( ∇ u M ) ∥ L 1 ( σ Ω M ) ≥ M . In particular, ∥ N ( ∇ u M ) ∥ L p ( σ Ω M ) ≥ c n , p M ∥ g M ∥ L p ( σ Ω M ) \|\mathcal{N}(\nabla u_{M})\|_{L^{p}(\sigma_{\Omega_{M}})}\ge c_{n,p}M\|g_{M}\|_{L^{p}(\sigma_{\Omega_{M}})} ∥ N ( ∇ u M ) ∥ L p ( σ Ω M ) ≥ c n , p M ∥ g M ∥ L p ( σ Ω M ) for every 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ , with c n , p > 0 c_{n,p}>0 c n , p > 0 independent of M M M .
Corollary 1.4. The domains Ω M \Omega_{M} Ω M of Theorem 1.3 are uniformly bi-Lipschitz equivalent to a ball, but their Lipschitz characters are not uniformly bounded.
Here, the Lipschitz character of a bounded Lipschitz domain consists of a bound for the slopes of its local Lipschitz graphs, the number of coordinate cylinders needed to cover its boundary, and the ratio between its diameter and the size of these cylinders. Indeed, the diameters and the surface measures of the boundaries of the domains Ω M \Omega_{M} Ω M are bounded above and below in terms of n n n and L 0 L_{0} L 0 . If the Lipschitz characters were uniformly bounded, the theorem of Jerison and Kenig [20 ] (recall that the boundaries of the domains Ω M \Omega_{M} Ω M are connected, being bi-Lipschitz images of ∂ Q 0 \partial Q_{0} ∂ Q 0 ) would give (1.1) for p = 2 p=2 p = 2 with a constant independent of M M M , which contradicts Theorem 1.3 .
We recall that the class of Lipschitz domains is not invariant under bi-Lipschitz homeomorphisms of the ambient space, see [17 ] . The situation in our construction is quite different: the domains Ω M \Omega_{M} Ω M are not Lipschitz graph domains at the scale of the oscillations, because the columns are tall, the oscillations are iterated at many scales, and the boundary condition is of Neumann type.
Remark 1.5. (a) The lower bound for N ( ∇ u ) \mathcal{N}(\nabla u) N ( ∇ u ) that we use is pointwise: on certain flat pieces of ∂ Ω \partial\Omega ∂ Ω we show that N a ( ∇ u ) ( x ) ≥ ∣ ∇ u ( x ) ∣ \mathcal{N}_a(\nabla u)(x) \ge|\nabla u(x)| N a ( ∇ u ) ( x ) ≥ ∣∇ u ( x ) ∣ , where ∇ u \nabla u ∇ u is continuous up to the boundary, and the right-hand side is obtained as a limit along the normal direction. Hence Theorems 1.2 and 1.3 also hold for the non-tangential maximal functions defined by means of L r L^r L r averages over Whitney balls, 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ , such as the one used in [33 ] , which goes back to [25 ] .
(b) Since ∂ Ω \partial\Omega ∂ Ω has finite measure and g g g is bounded with zero mean, g g g is a multiple of an atom of the Hardy space H 1 ( σ ) H^1(\sigma) H 1 ( σ ) . Thus (4) also shows that the endpoint estimate ∥ N ( ∇ u ) ∥ L 1 ( σ ) ≲ ∥ g ∥ H 1 ( σ ) \|\mathcal{N}(\nabla u)\|_{L^1(\sigma)} \lesssim\|g\|_{H^1(\sigma)} ∥ N ( ∇ u ) ∥ L 1 ( σ ) ≲ ∥ g ∥ H 1 ( σ ) fails.
(c) In Lipschitz domains, the Rellich identity shows that the normal and the tangential derivatives of a harmonic function u u u with N ( ∇ u ) ∈ L 2 ( σ ) \mathcal{N}(\nabla u) \in L^2(\sigma) N ( ∇ u ) ∈ L 2 ( σ ) have comparable L 2 ( σ ) L^2(\sigma) L 2 ( σ ) norms. In corkscrew domains with uniformly rectifiable boundaries, J. M. Gallegos [12 ] proved the one-sided estimate ∥ ∂ ν u f ∥ ≲ ∥ ∇ H f ∥ L 1 ( σ ) \|\partial_\nu u_f\| \lesssim\|\nabla_H f\|_{L^1(\sigma)} ∥ ∂ ν u f ∥ ≲ ∥ ∇ H f ∥ L 1 ( σ ) for the solution u f u_f u f of the Dirichlet problem with Lipschitz datum f f f , where ∇ H f \nabla_H f ∇ H f is the Hajłasz gradient of f f f and the left-hand side is the total variation of ∂ ν u f \partial_\nu u_f ∂ ν u f , while an L p L^p L p version of this estimate, for p > 1 p > 1 p > 1 , was obtained in [32 ] . By Theorem 1.2 , the reverse inequality fails in bi-Lipschitz images of a ball, in every L p L^p L p space with 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ . To see this, note that ∂ ν u = g \partial_\nu u = g ∂ ν u = g is bounded. On the other hand, u u u is smooth up to the boundary on a union of flat pieces of ∂ Ω \partial\Omega ∂ Ω , and the integral of ∣ ∇ τ u ∣ |\nabla_\tau u| ∣ ∇ τ u ∣ over this union is infinite, see § 9 .
Discussion of the proof, main ideas, and outline of the paper
For simplicity, in this discussion we restrict ourselves to R 3 \mathbb{R}^3 R 3 . In higher dimensions it is enough to take products with cubes, see § 10 . Our starting domain is the cube Q 0 = ( − 1 , 1 ) 3 Q_0 = (-1,1)^3 Q 0 = ( − 1 , 1 ) 3 , with the datum g g g equal to 1 1 1 on the face { x 1 = 1 } \{x_1 = 1\} { x 1 = 1 } , to − 1 -1 − 1 on the face { x 1 = − 1 } \{x_1 = -1\} { x 1 = − 1 } , and to 0 0 0 on the other four faces. The corresponding solution is just u 0 ( x ) = x 1 u_0(x) = x_1 u 0 ( x ) = x 1 , and on the bottom face { x 3 = − 1 } \{x_3 = -1\} { x 3 = − 1 } its gradient is tangent, with ∣ ∇ τ u 0 ∣ = 1 |\nabla_\tau u_0| = 1 ∣ ∇ τ u 0 ∣ = 1 .
We will only modify the cube near its bottom face, where g g g vanishes, and we will never touch the faces { x 1 = ± 1 } \{x_1 = \pm1\} { x 1 = ± 1 } . The normal derivative of the solution is then the same at all the stages of the construction, and if (1) is going to fail, it has to be because the tangential gradient gets large somewhere on the boundary. In a Lipschitz domain this cannot happen, by the Rellich identity. On the other hand, the Neumann problem has no maximum principle which would allow us to compare the solutions in a domain and in a subdomain, and in fact, if we remove a piece of the domain, the gradient of the solution near the boundary may increase. This is the effect that we will exploit: we look for a sequence of uniformly bi-Lipschitz modifications of the cube such that, after each of them, the integral of ∣ ∇ τ u ∣ |\nabla_\tau u| ∣ ∇ τ u ∣ over certain flat squares of the boundary is multiplied by a fixed factor γ > 1 \gamma> 1 γ > 1 . A schematic representation of the resulting domain is shown in Figure 1 .
The source of this gain is an elementary property of the planar flow of an ideal fluid around a square obstacle , and the main novelty of this paper is the way in which we bring it to the boundary of a domain in R 3 \mathbb{R}^3 R 3 . To do so, we remove from the domain a fine periodic collection of tall square columns, and then we repeat the operation on the walls of the new columns. The gain at each step is only a fixed factor slightly larger than 1 1 1 , so we need an unbounded number of generations, and we have to make sure that the gain is not lost along the way. Most of the difficulties lie in the following three points:
(i) The lower bound for the tangential gradient has to be proved on the rough part of the boundary, at the scale of the columns, whereas homogenization theory usually describes the behaviour of the solutions away from the roughness.
(ii) The domains have to stay uniformly bi-Lipschitz equivalent to a ball. To this end, the construction is rigid near the squares where the next generation will be built, and then, in each period of the array, these squares cannot have more area than the floor. Thus the area does not help, and the whole gain has to come from the flow.
(iii) Every insertion of columns modifies the solution everywhere, and how small the scale of an insertion must be depends, in a non-quantitative way, on the solution obtained before it. So the
scales are chosen one after the other, and the lower bounds obtained at some stage must survive all the subsequent insertions, which for Theorem 1.2 are infinitely many.
We now explain the construction a bit more in detail.
The basic computation relies on a planar toy model, and it is carried out in § 3 . Let E : = [ − 1 , 1 ] 2 E := [-1,1]^2 E := [ − 1 , 1 ] 2 . Given U ∈ R 2 U \in\mathbb{R}^2 U ∈ R 2 , we denote by V E U V_E^U V E U the velocity potential of an ideal fluid which flows around E E E and has velocity U U U at infinity, namely the harmonic function in R 2 ∖ E \mathbb{R}^2 \setminus E R 2 ∖ E with vanishing normal derivative on ∂ E \partial E ∂ E such that ∇ V E U ( y ) → U \nabla V_E^U(y) \to U ∇ V E U ( y ) → U as ∣ y ∣ → ∞ |y| \to\infty ∣ y ∣ → ∞ . Writing ∂ s \partial_s ∂ s for the derivative with respect to arc length on ∂ E \partial E ∂ E , the speed of the fluid along ∂ E \partial E ∂ E is ∣ ∂ s V E U ∣ |\partial_s V_E^U| ∣ ∂ s V E U ∣ , and we show that
1 H 1 ( ∂ E ) ∫ ∂ E ∣ ∂ s V E U ∣ d s = c ∗ ∣ U ∣ , where c ∗ = Γ ( 1 / 4 ) 2 2 π 3 / 2 ≈ 1.18 , (5) \frac{1}{\mathcal{H}^1(\partial E)}\int_{\partial E} |\partial_s V_E^U|\,\mathrm{d}s = c_*|U|,\qquad\text{where}\qquad c_*=\frac{\Gamma(1/4)^2}{2\pi^{3/2}}\approx1.18, \tag*{(5)} H 1 ( ∂ E ) 1 ∫ ∂ E ∣ ∂ s V E U ∣ d s = c ∗ ∣ U ∣ , where c ∗ = 2 π 3/2 Γ ( 1/4 ) 2 ≈ 1.18 , (5)
for every U ∈ R 2 U \in\mathbb{R}^2 U ∈ R 2 . That is, on average, the flow along the obstacle is faster than at infinity. Here c ∗ c_* c ∗ is the logarithmic capacity of E E E . It comes from the conformal map of the exterior of the unit disc onto R 2 ∖ E \mathbb{R}^2 \setminus E R 2 ∖ E , which transforms the boundary values of V E U V_E^U V E U into the function 2 c ∗ ∣ U ∣ cos ( θ − α ) 2c_*|U|\cos(\theta-\alpha) 2 c ∗ ∣ U ∣ cos ( θ − α ) on the unit circle (with U = ∣ U ∣ e i α U=|U|e^{i\alpha} U = ∣ U ∣ e i α ). The total variation of the latter is 8 c ∗ ∣ U ∣ 8c_*|U| 8 c ∗ ∣ U ∣ , and H 1 ( ∂ E ) = 8 \mathcal{H}^1(\partial E)=8 H 1 ( ∂ E ) = 8 . In the rest of the paper we only need the inequality c ∗ > 1 c_*>1 c ∗ > 1 , which can also be proved by potential theory, using that E E E contains the unit disc. We remark that for the unit disc the same quantity equals 4 / π 4/\pi 4/ π , which is even larger, but it is important for us that the obstacle has flat sides.
Next we go to three dimensions, and let Π : = E × [ 0 , λ ] \Pi:= E \times[0,\lambda] Π := E × [ 0 , λ ] , with λ \lambda λ large, and let us remove from the upper half-space the column Π \Pi Π and all its translates ( z , 0 ) + Π (z,0)+\Pi ( z , 0 ) + Π , z ∈ ς Z 2 z \in\varsigma\mathbb{Z}^2 z ∈ ς Z 2 . In the domain that remains, we consider the periodic solution of the Neumann problem whose gradient tends to ( U , 0 ) (U,0) ( U , 0 ) at infinity. Away from the floor and from the roofs of the columns, this function is almost independent of the vertical variable, and in each horizontal section it looks like the planar flow V E U V_E^U V E U . By (5) , the mean value of the tangential gradient on the walls of the columns is then close to c ∗ ∣ U ∣ c_*|U| c ∗ ∣ U ∣ .
Since the walls of one column have area 8 λ 8\lambda 8 λ and the floor of a period cell has area ς 2 \varsigma^2 ς 2 , we choose ς 2 ≈ λ \varsigma^2\approx\lambda ς 2 ≈ λ , so that both areas are comparable. In this way, roughly speaking, the columns turn the floor into walls, and the walls carry a larger tangential gradient . The details are given in § 4 , where we also fix the selected squares, that is, a family of small squares on the walls where the next generation of columns will be placed. We remark that this reduction to a planar problem only makes sense if the horizontal sections have dimension at least two, so that the fluid can flow around the obstacle. This agrees with the fact that the analogue of Theorem 1.2 fails in the plane, as mentioned above.
Now let u u u be the solution in some domain D D D , and let Q Q Q be a flat square of ∂ D \partial D ∂ D on which the datum vanishes, so that ∇ u = ( U ( y ) , 0 ) \nabla u=(U(y),0) ∇ u = ( U ( y ) , 0 ) on Q Q Q , with U ( y ) U(y) U ( y ) tangent to Q Q Q . We remove from D D D a copy of the periodic array above, scaled by a small factor ρ \rho ρ and placed over Q Q Q , and we denote by u ρ u_\rho u ρ the new solution. In § 7 we discuss that u ρ u_\rho u ρ is close to u + ρ ∑ j U j ( y ) P e j ( x / ρ ) u+\rho\sum_j U_j(y)P^{e_j}(x/\rho) u + ρ ∑ j U j ( y ) P e j ( x / ρ ) , where U ⋅ y + P U U\cdot y+P^U U ⋅ y + P U is the solution of the periodic problem in the complement of the columns. So, if M ρ \mathcal{M}_\rho M ρ is the union of the selected squares on the walls of the new columns, for ρ \rho ρ small we expect that
∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≈ ∣ M ∣ ∣ Y ∣ c ∗ ∫ Q ∣ ∇ τ u ∣ d σ , (6) \int_{\mathcal{M}_\rho} |\nabla_\tau u_\rho|\,d\sigma\approx\frac{|\mathcal{M}|}{|Y|}c_*\int_Q |\nabla_\tau u|\,d\sigma, \tag*{(6)} ∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≈ ∣ Y ∣ ∣ M ∣ c ∗ ∫ Q ∣ ∇ τ u ∣ d σ , (6)
where ∣ M ∣ |\mathcal{M}| ∣ M ∣ is the area of the selected squares in one period and ∣ Y ∣ = ℓ 2 |Y|=\ell^2 ∣ Y ∣ = ℓ 2 is the area of its floor. The corresponding lower bound, with c ∗ c_* c ∗ replaced by any smaller constant, is proved in § 7 . On the other hand, the energy of the solution only changes by O ( ρ ) O(\sqrt{\rho}) O ( ρ ) , since the gradient of the correction is of order one in a layer of thickness comparable to ρ \rho ρ .
This is the point where (i) comes into play. Approximations of this type are classical in homogenization, see for instance [5 , 22 , 1 ] , but there one studies the limit of the solutions, which here is just u u u . We need instead a lower bound for the tangential gradient on the selected squares of the new columns, where it is amplified by the factor γ \gamma γ . For this, the error of the approximation has to be much smaller than ρ \sqrt{\rho} ρ in the energy norm, that is, than the correction itself. In Lemma 6.4 we show that, as ρ → 0 \rho\to0 ρ → 0 , it is at most of order d ρ \sqrt{d\rho} d ρ , where d d d is the width of a strip along ∂ Q \partial Q ∂ Q in which we cut off the approximation. Since u ρ u_\rho u ρ does not depend on d d d , we can then let d → 0 d\to0 d → 0 , and this is enough.
Next we iterate, and new columns are put on the selected squares, then on the selected squares of these columns, and so on. Iterating modifications on flat squares of the faces of a cube is an idea that has been exploited in the literature before: T. Wolff used it to construct his snowflakes [Wo95] (see also [26 ] ), in connection with the dimension of harmonic measure. In Wolff’s construction, though, each modification is comparable in size to the square where it is placed, and its effect on harmonic measure is estimated by a perturbation argument. In our case the columns are much smaller than the squares, and the gain at each step is due to the planar flow past a square, through the homogenization argument described above.
Here we meet the difficulty (ii): for the domains to stay uniformly bi-Lipschitz equivalent to a ball, the distortions of the maps must not accumulate along the iteration. Indeed, if we compose k k k maps with bi-Lipschitz constant L L L , we only obtain the bound L k L^k L k , and choosing smaller and smaller scales does not help. For this reason, each selected square on a wall will be a rigid copy of a square of the floor. That is, the bi-Lipschitz map which creates one column is a rigid motion near the preimage of each selected square, and so the next generation is built where the accumulated map is an isometry, see § 6 and Appendix A, where we also use an elementary lemma on moving squares by affine maps. The drawback is that the preimages of the selected squares are disjoint squares in the floor of a period cell, and thus
∣ M ∣ ≤ A ∣ Y ∣ , for some A < 1. |\mathcal{M}| \le A|Y|,\qquad\text{for some } A<1. ∣ M ∣ ≤ A ∣ Y ∣ , for some A < 1.
Fortunately, since c ∗ > 1 c_*>1 c ∗ > 1 , we can still take A < 1 A<1 A < 1 so close to 1 that c ∗ A > 1 c_*A>1 c ∗ A > 1 . Then, by (6) , every generation multiplies the integral of the tangential gradient over the selected squares by a factor at least β \beta β , where β > 1 \beta>1 β > 1 is fixed (any β < c ∗ A \beta<c_*A β < c ∗ A is possible). Recall that the scale of each generation has to be small enough, depending on the solution obtained at the previous one, see Remark 6.6. In this way, after k k k generations we get a domain Ω k \Omega_k Ω k , uniformly bi-Lipschitz equivalent to the cube, whose datum is still bounded by 1, and such that
∫ ∂ Ω k N ( ∇ u k ) d σ ≥ ∫ C k ∣ ∇ τ u k ∣ d σ ≥ β k \int_{\partial\Omega_k} \mathcal{N}(\nabla u_k)\,d\sigma\ge\int_{C_k} |\nabla_\tau u_k|\,d\sigma\ge\beta^k ∫ ∂ Ω k N ( ∇ u k ) d σ ≥ ∫ C k ∣ ∇ τ u k ∣ d σ ≥ β k
where C k C_k C k denotes the union of the selected squares of the k k k -th generation. Notice that the area of C k C_k C k is at most A k A^k A k . Theorem 1.3 follows from this, as we show in § 8 . We also check in Appendix B that the domains Ω k \Omega_k Ω k are Lipschitz, although their Lipschitz characters cannot be uniformly bounded, by Corollary 1.4.
The proof of Theorem 1.2 requires one more step, since we need a single domain, and it is here that (iii) really matters. We consider a sequence of disjoint balls B j B_j B j , centred on the bottom face of the cube and accumulating at one point of it (see Figure 4 ), and in each B j B_j B j we perform a finite construction with enough generations to make the integral of the tangential gradient over the last selected squares in B j B_j B j at least 2. Every such construction perturbs the solution in the whole domain, including the balls which have been treated before. However, these perturbations are small in the energy norm, and near the selected squares the domain is flat and the datum vanishes. By Lemma 2.2 , the gradient of the solution on these squares is then perturbed only slightly. Thus, choosing the scales of the later constructions small enough, the changes of energy are summable and on every ball B j B_j B j we keep a lower bound 1, and summing over j j j gives (1.4). All this is done in § 9 .
Figure 4. The domain Ω \Omega Ω of Theorem 1.2 near the bottom face of the cube. In the balls B j B_j B j , which accumulate at x ∗ x_* x ∗ , we perform finite constructions with more and more generations. The construction in B j B_j B j produces a perturbation of the solution in the energy norm so small that the lower bounds for the tangential gradient on S 1 , … , S j − 1 S_1,\ldots,S_{j-1} S 1 , … , S j − 1 survive.
Concerning the structure of the paper, after some preliminaries in § 2 on variational solutions in bi-Lipschitz images of a cube and on flat pieces of the boundary, we follow the scheme described above, except that the geometric construction comes before the homogenization argument. Namely, the planar identity (1.5) is proved in § 3 , the periodic array of tall columns and the selected squares are studied in § 4 , the column map and the rigidity of the construction are discussed in § 6 , and the lower bound in (1.6) is proved in § 7 . In § 8 and § 9 we prove Theorems 1.3 and 1.2 in R 3 \mathbb{R}^3 R 3 , and in § 10 we treat higher dimensions and the geometric properties of the domains mentioned after Theorem 1.2 . Finally, Appendix A contains the proofs of the geometric lemmas of § 6 , including a lemma on moving squares by affine maps, and Appendix B the proof that the finite stages are Lipschitz domains.
Related works. The construction of our counterexample and the results of this paper are tightly related to the companion paper [30 ] by the first named author, which appears simultaneously with the present one. There, these ideas are combined with the techniques developed in [29 ] , and it is shown that, for n ≥ 2 n \ge2 n ≥ 2 , the L p L^p L p Neumann problem in the upper half-space R + n + 1 \mathbb{R}^{n+1}_+ R + n + 1 may fail for every 1 < p < ∞ 1 < p < \infty 1 < p < ∞ for elliptic operators — div ( A ∇ ) \operatorname{div}(A\nabla) div ( A ∇ ) with real coefficients which are independent of the transversal variable, or which satisfy the Dahlberg-Kenig-Pipher condition. In that setting the boundary is flat, and the role of the geometry is played by skew-symmetric perturbations of the identity matrix.
Acknowledgements. The paper was finalized during the first-named author’s visit to the Centre de Recerca Matemàtica (Bellaterra, Barcelona) for the program “Analysis of Free Boundary Problems” in the fall of 2026. He thanks the institute for its hospitality.
AI use. The authors used LLMs in the development of this work to explore mathematical ideas and develop proofs. More specifically, the identification of the obstructions to the L p L^p L p solvability of the Neumann problem and the main geometric ideas for the construction of the counterexample in Theorem 1.2 are due to the authors. Next, AI tools were used interactively to assist with the calculations, with the editing of the manuscript, and the proofreading. All the outputs were checked by the authors, who revised them with extensive edits and take full responsibility for the content of the paper.
Preliminaries
Notation
We collect here the notation used throughout the paper.
We write a ≲ b a \lesssim b a ≲ b if a ≤ C b a \le Cb a ≤ C b for some constant C > 0 C > 0 C > 0 , and a ≈ b a \approx b a ≈ b if a ≲ b ≲ a a \lesssim b \lesssim a a ≲ b ≲ a . The dependence of C C C on the parameters will be specified when it is not clear from the context.
If S ⊂ R n + 1 S \subset\mathbb{R}^{n+1} S ⊂ R n + 1 and μ ( S ) ∈ ( 0 , ∞ ) \mu(S) \in(0,\infty) μ ( S ) ∈ ( 0 , ∞ ) , where μ \mu μ is either the Lebesgue measure or the surface measure H n \mathcal{H}^{n} H n , we write
m S f : = μ ( S ) − 1 ∫ S f d μ . m_{S}f := \mu(S)^{-1} \int_{S} f\,d\mu. m S f := μ ( S ) − 1 ∫ S f d μ .
We write σ \sigma σ instead of σ V = H n ∣ ∂ V \sigma_{V} = \mathcal{H}^{n}|_{\partial V} σ V = H n ∣ ∂ V when the domain V V V is clear from the context.
On a flat piece of the boundary of a domain, ∇ τ f \nabla_{\tau} f ∇ τ f is the gradient of the restriction of f f f to the affine hyperplane containing that piece. For a Lipschitz function on a rectifiable boundary, ∇ τ f \nabla_{\tau} f ∇ τ f is the tangential gradient, defined σ \sigma σ -almost everywhere.
By the energy of a function f f f we mean ∥ ∇ f ∥ L 2 \lVert\nabla f\rVert_{L^{2}} ∥ ∇ f ∥ L 2 .
A map G : R n + 1 → R n + 1 G \colon\mathbb{R}^{n+1} \to\mathbb{R}^{n+1} G : R n + 1 → R n + 1 is L L L -bi-Lipschitz if it is a homeomorphism with Lip ( G ) ≤ L \operatorname{Lip}(G) \le L Lip ( G ) ≤ L and Lip ( G − 1 ) ≤ L \operatorname{Lip}(G^{-1}) \le L Lip ( G − 1 ) ≤ L . A rigid motion is an orientation-preserving isometry of the ambient space.
Points of R n + 1 \mathbb{R}^{n+1} R n + 1 are written x = ( y , t ) x = (y,t) x = ( y , t ) , with y ∈ R n y \in\mathbb{R}^{n} y ∈ R n and t ∈ R t \in\mathbb{R} t ∈ R , and R + n + 1 : = { t > 0 } \mathbb{R}^{n+1}_{+} := \{t > 0\} R + n + 1 := { t > 0 } . From § 3 to § 9 we have n = 2 n = 2 n = 2 .
We write Q 0 : = ( − 1 , 1 ) n + 1 Q_{0} := (-1,1)^{n+1} Q 0 := ( − 1 , 1 ) n + 1 . We remark that all the domains that we construct are images of Q 0 Q_{0} Q 0 under bi-Lipschitz homeomorphisms of R n + 1 \mathbb{R}^{n+1} R n + 1 . To pass from the cube to the ball it suffices to use the radial map
B ( X ) : = ∣ X ∣ 2 ∣ X ∣ ∞ X ( X ≠ 0 ) , B ( 0 ) = 0. (7) B(X) := \frac{|X|_{2}}{|X|_{\infty}}X \quad(X \ne0), \qquad B(0)=0. \tag*{(7)} B ( X ) := ∣ X ∣ ∞ ∣ X ∣ 2 X ( X = 0 ) , B ( 0 ) = 0. (7)
where ∣ ⋅ ∣ 2 |\cdot|_{2} ∣ ⋅ ∣ 2 and ∣ ⋅ ∣ ∞ |\cdot|_{\infty} ∣ ⋅ ∣ ∞ are the Euclidean and the maximum norms. It maps B ( 0 , 1 ) B(0,1) B ( 0 , 1 ) onto Q 0 Q_{0} Q 0 . Both B B B and its inverse X ↦ ∣ X ∣ ∞ ∣ X ∣ 2 X X \mapsto\frac{|X|_{\infty}}{|X|_{2}}X X ↦ ∣ X ∣ 2 ∣ X ∣ ∞ X are homogeneous of degree one, and their restrictions to the unit sphere are Lipschitz, because ( n + 1 ) − 1 / 2 ≤ ∣ X ∣ ∞ ≤ 1 (n+1)^{-1/2} \le|X|_{\infty} \le1 ( n + 1 ) − 1/2 ≤ ∣ X ∣ ∞ ≤ 1 there. Hence both maps are globally Lipschitz. Thus, if Ω = G ( Q 0 ) \Omega= G(Q_{0}) Ω = G ( Q 0 ) for a bi-Lipschitz homeomorphism G G G , then Ω = T ( B ( 0 , 1 ) ) \Omega= T(B(0,1)) Ω = T ( B ( 0 , 1 )) with T = G ∘ B T = G \circ B T = G ∘ B .
Variational solutions and flat pieces of the boundary
Since the domains that we construct are bi-Lipschitz images of Q 0 Q_{0} Q 0 , the basic properties of Sobolev functions in the cube transfer to them, with constants depending only on n n n and the bi-Lipschitz constant. In particular, the variational solutions in (3) are well defined, as the following lemma shows.
Lemma 2.1. Let V = G ( Q 0 ) V = G(Q_{0}) V = G ( Q 0 ) , where G : R n + 1 → R n + 1 G \colon\mathbb{R}^{n+1} \to\mathbb{R}^{n+1} G : R n + 1 → R n + 1 is L L L -bi-Lipschitz. Then, there exists a constant C > 0 C > 0 C > 0 depending only on n n n and L L L such that the following properties hold:
(a) The restrictions to V V V of functions in C c ∞ ( R n + 1 ) C^{\infty}_{c}(\mathbb{R}^{n+1}) C c ∞ ( R n + 1 ) are dense in W 1 , 2 ( V ) W^{1,2}(V) W 1 , 2 ( V ) , and the restriction to ∂ V \partial V ∂ V of these functions extends to a bounded trace operator Tr : W 1 , 2 ( V ) → L 2 ( σ V ) \operatorname{Tr} \colon W^{1,2}(V) \to L^{2}(\sigma_{V}) Tr : W 1 , 2 ( V ) → L 2 ( σ V ) with norm at most C C C .
(b) For every φ ∈ W 1 , 2 ( V ) \varphi\in W^{1,2}(V) φ ∈ W 1 , 2 ( V ) , with c : = m ∂ V ( Tr φ ) c := m_{\partial V}(\operatorname{Tr}\varphi) c := m ∂ V ( Tr φ ) ,
∥ φ − c ∥ L 2 ( V ) + ∥ Tr φ − c ∥ L 2 ( σ V ) ≤ C ∥ ∇ φ ∥ L 2 ( V ) . (8) \lVert\varphi-c\rVert_{L^{2}(V)}+\lVert\operatorname{Tr}\varphi-c\rVert_{L^{2}(\sigma_{V})}\le C\lVert\nabla\varphi\rVert_{L^{2}(V)}. \tag*{(8)} ∥ φ − c ∥ L 2 ( V ) + ∥ Tr φ − c ∥ L 2 ( σ V ) ≤ C ∥ ∇ φ ∥ L 2 ( V ) . (8) (c) For every g ∈ L 2 ( σ V ) g \in L^{2}(\sigma_{V}) g ∈ L 2 ( σ V ) with ∫ g d σ V = 0 \int g\,d\sigma_{V}=0 ∫ g d σ V = 0 there is a solution u ∈ W 1 , 2 ( V ) u \in W^{1,2}(V) u ∈ W 1 , 2 ( V ) of (3) , unique up to additive constants, and ∥ ∇ u ∥ L 2 ( V ) ≤ C ∥ g ∥ L 2 ( σ V ) \lVert\nabla u\rVert_{L^{2}(V)} \le C\lVert g\rVert_{L^{2}(\sigma_{V})} ∥ ∇ u ∥ L 2 ( V ) ≤ C ∥ g ∥ L 2 ( σ V ) .
Proof. All the assertions are well known in Q 0 Q_{0} Q 0 , and we transfer them to V V V . Given φ ∈ W 1 , 2 ( V ) \varphi\in W^{1,2}(V) φ ∈ W 1 , 2 ( V ) , let v : = φ ∘ G v := \varphi\circ G v := φ ∘ G . Since G G G is L L L -bi-Lipschitz, the chain rule and the change of variables formula show that ∥ φ ∥ W 1 , 2 ( V ) ≈ ∥ v ∥ W 1 , 2 ( Q 0 ) \|\varphi\|_{W^{1,2}(V)} \approx\|v\|_{W^{1,2}(Q_0)} ∥ φ ∥ W 1 , 2 ( V ) ≈ ∥ v ∥ W 1 , 2 ( Q 0 ) and ∥ ∇ φ ∥ L 2 ( V ) ≈ ∥ ∇ v ∥ L 2 ( Q 0 ) \|\nabla\varphi\|_{L^2(V)} \approx\|\nabla v\|_{L^2(Q_0)} ∥∇ φ ∥ L 2 ( V ) ≈ ∥∇ v ∥ L 2 ( Q 0 ) , with constants depending on n n n and L L L . The same holds for the L 2 L^2 L 2 norms on ∂ V \partial V ∂ V and on ∂ Q 0 \partial Q_0 ∂ Q 0 , because the Jacobian of G ∣ ∂ Q 0 G|_{\partial Q_0} G ∣ ∂ Q 0 lies between L − n L^{-n} L − n and L n L^n L n almost everywhere, by the area formula. Hence it makes sense to define Tr φ : = ( Tr v ) ∘ G − 1 \operatorname{Tr}\varphi:= (\operatorname{Tr}v)\circ G^{-1} Tr φ := ( Tr v ) ∘ G − 1 , and the properties in (a) follow from the analogous ones in the cube by standard arguments.
For (b), let b : = m Q 0 v b := m_{Q_0}v b := m Q 0 v . In the cube, the trace and Poincaré inequalities give
∥ v − b ∥ L 2 ( Q 0 ) + ∥ Tr v − b ∥ L 2 ( ∂ Q 0 ) ≲ ∥ ∇ v ∥ L 2 ( Q 0 ) . \|v-b\|_{L^2(Q_0)}+\|\operatorname{Tr}v-b\|_{L^2(\partial Q_0)}\lesssim\|\nabla v\|_{L^2(Q_0)}. ∥ v − b ∥ L 2 ( Q 0 ) + ∥ Tr v − b ∥ L 2 ( ∂ Q 0 ) ≲ ∥∇ v ∥ L 2 ( Q 0 ) .
We transport this inequality to V V V , and we replace b b b by c c c using
∣ c − b ∣ = ∣ m ∂ V ( Tr φ − b ) ∣ ≤ σ V ( ∂ V ) − 1 / 2 ∥ Tr φ − b ∥ L 2 ( σ V ) . |c-b|=\left|m_{\partial V}(\operatorname{Tr}\varphi-b)\right|\le\sigma_V(\partial V)^{-1/2}\|\operatorname{Tr}\varphi-b\|_{L^2(\sigma_V)}. ∣ c − b ∣ = ∣ m ∂ V ( Tr φ − b ) ∣ ≤ σ V ( ∂ V ) − 1/2 ∥ Tr φ − b ∥ L 2 ( σ V ) .
together with the fact that ∣ V ∣ |V| ∣ V ∣ and σ V ( ∂ V ) \sigma_V(\partial V) σ V ( ∂ V ) are bounded above and below in terms of n n n and L L L .
To prove (c), note that by (8) the norm ∥ ∇ φ ∥ L 2 ( V ) \|\nabla\varphi\|_{L^2(V)} ∥∇ φ ∥ L 2 ( V ) is equivalent to the W 1 , 2 W^{1,2} W 1 , 2 norm on the closed subspace { φ : m ∂ V ( Tr φ ) = 0 } \{\varphi:m_{\partial V}(\operatorname{Tr}\varphi)=0\} { φ : m ∂ V ( Tr φ ) = 0 } , and the functional φ ↦ ∫ g Tr φ d σ V \varphi\mapsto\int g\operatorname{Tr}\varphi\,d\sigma_V φ ↦ ∫ g Tr φ d σ V has norm at most C C C times ∥ g ∥ L 2 ( σ V ) \|g\|_{L^2(\sigma_V)} ∥ g ∥ L 2 ( σ V ) there, again by (8) . The Riesz representation theorem yields the existence of a function u u u in this subspace satisfying (3) for test functions in the subspace, and hence for all test functions, because g g g has zero mean. The energy bound is obtained by taking u u u itself as a test function, and the uniqueness by testing the equation for the difference of two solutions with this difference. □
For instance, when ∣ g ∣ ≤ 1 |g|\le1 ∣ g ∣ ≤ 1 , Lemma 2.1 yields
∥ ∇ u ∥ L 2 ( V ) ≤ C σ V ( ∂ V ) 1 / 2 ≤ C ′ , \|\nabla u\|_{L^2(V)}\le C\sigma_V(\partial V)^{1/2}\le C', ∥∇ u ∥ L 2 ( V ) ≤ C σ V ( ∂ V ) 1/2 ≤ C ′ ,
where C ′ C' C ′ only depends on n n n and L L L .
The next lemma deals with flat pieces of the boundary where the datum vanishes. On such pieces the solution can be extended by reflection, and so it is smooth up to the boundary. All our lower bounds will be proved there.
Lemma 2.2. Let V V V be as in Lemma 2.1 , let u u u be the variational solution with datum g g g , let x ∈ ∂ V x\in\partial V x ∈ ∂ V , r > 0 r>0 r > 0 , and suppose that there is an open half-space H H H such that
V ∩ B ( x , r ) = H ∩ B ( x , r ) V\cap B(x,r)=H\cap B(x,r) V ∩ B ( x , r ) = H ∩ B ( x , r ) and g = 0 g=0 g = 0 on ∂ V ∩ B ( x , r ) \partial V\cap B(x,r) ∂ V ∩ B ( x , r ) . Then the even reflection of u u u across ∂ H \partial H ∂ H is harmonic in B ( x , r ) B(x,r) B ( x , r ) . In particular, ∇ u \nabla u ∇ u is continuous up to ∂ V ∩ B ( x , r ) \partial V\cap B(x,r) ∂ V ∩ B ( x , r ) , it is tangent to ∂ V \partial V ∂ V there, and
N a ( ∇ u ) ( z ) ≥ ∣ ∇ u ( z ) ∣ = ∣ ∇ τ u ( z ) ∣ for every z ∈ ∂ V ∩ B ( x , r ) and a > 0. (9) N_a(\nabla u)(z)\ge|\nabla u(z)|=|\nabla_\tau u(z)|\qquad\text{for every }z\in\partial V\cap B(x,r)\text{ and }a>0. \tag*{(9)} N a ( ∇ u ) ( z ) ≥ ∣∇ u ( z ) ∣ = ∣ ∇ τ u ( z ) ∣ for every z ∈ ∂ V ∩ B ( x , r ) and a > 0. (9) Moreover, if v v v is the variational solution in another domain V ′ V' V ′ satisfying the same assumptions in B ( x , r ) B(x,r) B ( x , r ) , with the same half-space H H H , then
sup z ∈ ∂ V ∩ B ( x , r / 2 ) ∣ ∇ u ( z ) − ∇ v ( z ) ∣ ≤ C n r − n + 1 2 ∥ ∇ ( u − v ) ∥ L 2 ( V ∩ B ( x , r ) ) . (10) \sup_{z\in\partial V\cap B(x,r/2)}|\nabla u(z)-\nabla v(z)|\le C_n r^{-\frac{n+1}{2}}\|\nabla(u-v)\|_{L^2(V\cap B(x,r))}. \tag*{(10)} z ∈ ∂ V ∩ B ( x , r /2 ) sup ∣∇ u ( z ) − ∇ v ( z ) ∣ ≤ C n r − 2 n + 1 ∥∇ ( u − v ) ∥ L 2 ( V ∩ B ( x , r )) . (10)
Proof. After a rigid motion, we may and will assume that H = { t > 0 } H=\{t>0\} H = { t > 0 } . Let u ^ ( y , t ) : = u ( y , ∣ t ∣ ) \widehat{u}(y,t):=u(y,|t|) u ( y , t ) := u ( y , ∣ t ∣ ) for ( y , t ) ∈ B ( x , r ) (y,t)\in B(x,r) ( y , t ) ∈ B ( x , r ) . If φ ∈ C 0 ∞ ( B ( x , r ) ) \varphi\in C^\infty_0(B(x,r)) φ ∈ C 0 ∞ ( B ( x , r )) , we apply (3) to the test function φ ( y , t ) + φ ( y , − t ) \varphi(y,t)+\varphi(y,-t) φ ( y , t ) + φ ( y , − t ) , restricted to V V V , whose trace vanishes on ∂ V ∩ B ( x , r ) \partial V\cap B(x,r) ∂ V ∩ B ( x , r ) . Since g = 0 g=0 g = 0 on ∂ V ∩ B ( x , r ) \partial V\cap B(x,r) ∂ V ∩ B ( x , r ) , after the change of variables ( y , t ) ↦ ( y , − t ) (y,t)\mapsto(y,-t) ( y , t ) ↦ ( y , − t ) in B ( x , r ) ∩ { t < 0 } B(x,r)\cap\{t<0\} B ( x , r ) ∩ { t < 0 } , we obtain
∫ B ( x , r ) ∇ u ^ ⋅ ∇ φ = 0. \int_{B(x,r)}\nabla\widehat{u}\cdot\nabla\varphi=0. ∫ B ( x , r ) ∇ u ⋅ ∇ φ = 0.
Thus, since u ^ ∈ W 1 , 2 ( B ( x , r ) ) \widehat{u}\in W^{1,2}(B(x,r)) u ∈ W 1 , 2 ( B ( x , r )) by construction, it is a harmonic function in B ( x , r ) B(x,r) B ( x , r ) by Weyl’s lemma. It follows that ∇ u \nabla u ∇ u is continuous up to ∂ V ∩ B ( x , r ) \partial V\cap B(x,r) ∂ V ∩ B ( x , r ) , and since u ^ \widehat{u} u is even in t t t , ∂ t u = 0 \partial_tu=0 ∂ t u = 0 there. Let e t = ( 0 , … , 0 , 1 ) e_t=(0,\ldots,0,1) e t = ( 0 , … , 0 , 1 ) . If z ∈ ∂ V ∩ B ( x , r ) z\in\partial V\cap B(x,r) z ∈ ∂ V ∩ B ( x , r ) and h > 0 h>0 h > 0 is small, then δ V ( z + h e t ) = h \delta_V(z+he_t)=h δ V ( z + h e t ) = h and ∣ ( z + h e t ) − z ∣ = h < ( 1 + a ) h |(z+he_t)-z|=h<(1+a)h ∣ ( z + h e t ) − z ∣ = h < ( 1 + a ) h , so that z + h e t ∈ Γ V , a ( z ) z+he_t\in\Gamma_{V,a}(z) z + h e t ∈ Γ V , a ( z ) . Hence N a ( ∇ u ) ( z ) ≥ ∣ ∇ u ( z + h e t ) ∣ \mathcal{N}_a(\nabla u)(z)\geq|\nabla u(z+he_t)| N a ( ∇ u ) ( z ) ≥ ∣∇ u ( z + h e t ) ∣ , and letting h → 0 h\to0 h → 0 we obtain (9) .
We finally prove (10) . We note that V ∩ B ( x , r ) = V ′ ∩ B ( x , r ) V\cap B(x,r)=V'\cap B(x,r) V ∩ B ( x , r ) = V ′ ∩ B ( x , r ) by assumption. Then, we define v ^ ( y , t ) : = v ( y , ∣ t ∣ ) \widehat{v}(y,t):=v(y,|t|) v ( y , t ) := v ( y , ∣ t ∣ ) , and observe that the components of ∇ ( u ^ − v ^ ) \nabla(\widehat{u}-\widehat{v}) ∇ ( u − v ) are harmonic in B ( x , r ) B(x,r) B ( x , r ) . Hence, if z ∈ ∂ V ∩ B ( x , r / 2 ) z\in\partial V\cap B(x,r/2) z ∈ ∂ V ∩ B ( x , r /2 ) , then B ( z , r / 2 ) ⊂ B ( x , r ) B(z,r/2)\subset B(x,r) B ( z , r /2 ) ⊂ B ( x , r ) , and the mean value property in B ( z , r / 2 ) B(z,r/2) B ( z , r /2 ) , the Cauchy-Schwarz inequality and the fact that the reflection doubles the squared L 2 L^2 L 2 norm give
∣ ∇ ( u − v ) ( z ) ∣ ≲ r − n + 1 2 ∥ ∇ ( u ^ − v ^ ) ∥ L 2 ( B ( z , r / 2 ) ) ≤ 2 r − n + 1 2 ∥ ∇ ( u − v ) ∥ L 2 ( V ∩ B ( x , r ) ) . |\nabla(u-v)(z)|\lesssim r^{-\frac{n+1}{2}}\|\nabla(\widehat{u}-\widehat{v})\|_{L^2(B(z,r/2))}\leq\sqrt{2}r^{-\frac{n+1}{2}}\|\nabla(u-v)\|_{L^2(V\cap B(x,r))}. ∣∇ ( u − v ) ( z ) ∣ ≲ r − 2 n + 1 ∥∇ ( u − v ) ∥ L 2 ( B ( z , r /2 )) ≤ 2 r − 2 n + 1 ∥∇ ( u − v ) ∥ L 2 ( V ∩ B ( x , r )) .
Lemma 2.3. Under the assumptions of Lemma 2.2, if u u u is not constant and P ⊂ ∂ V ∩ B ( x , r ) P\subset\partial V\cap B(x,r) P ⊂ ∂ V ∩ B ( x , r ) is a closed n n n -dimensional cube, then ∫ P ∣ ∇ τ u ∣ d σ > 0 \int_P|\nabla_\tau u|\,d\sigma>0 ∫ P ∣ ∇ τ u ∣ d σ > 0 .
Proof. We argue by contradiction, and assume that the integral vanishes. Then, by continuity, ∇ τ u = 0 \nabla_\tau u=0 ∇ τ u = 0 in the relative interior of P P P , and so u u u equals a constant c c c there. As in the proof of Lemma 2.2, assume H = { t > 0 } H=\{t>0\} H = { t > 0 } , and let w : = u ^ − c w:=\widehat{u}-c w := u − c , which is harmonic in B ( x , r ) B(x,r) B ( x , r ) , even in t t t , and vanishes on the relative interior P ′ P' P ′ of P P P , which is an open subset of the plane { t = 0 } \{t=0\} { t = 0 } .1 Take a ball B ′ ⊂ B ( x , r ) B'\subset B(x,r) B ′ ⊂ B ( x , r ) centred on P ′ P' P ′ with B ′ ∩ { t = 0 } ⊂ P ′ B'\cap\{t=0\}\subset P' B ′ ∩ { t = 0 } ⊂ P ′ . By the Schwarz reflection principle for the Dirichlet problem, the odd reflection of w ∣ B ′ ∩ { t > 0 } w|_{B'\cap\{t>0\}} w ∣ B ′ ∩ { t > 0 } is harmonic in B ′ B' B ′ ; it coincides with w w w in B ′ ∩ { t > 0 } B'\cap\{t>0\} B ′ ∩ { t > 0 } , and therefore in B ′ B' B ′ , by analyticity. Hence w w w is both even and odd in t t t in B ′ B' B ′ , that is, w ≡ 0 w\equiv0 w ≡ 0 in B ′ B' B ′ . In other words, we have used the uniqueness of the Cauchy problem for the Laplace equation, since u − c u-c u − c has vanishing Dirichlet and Neumann data on P ′ P' P ′ . By unique continuation, u ≡ c u\equiv c u ≡ c in the connected open set V V V , which is a contradiction.
Potential flow past a square
In this section we prove the planar identity (5) , which is the source of the gain in the whole construction. We identify R 2 \mathbb{R}^2 R 2 with C \mathbb{C} C and we set
E : = [ − 1 , 1 ] 2 . E:=[-1,1]^2. E := [ − 1 , 1 ] 2 .
Given U ∈ R 2 U\in\mathbb{R}^2 U ∈ R 2 , we look for the velocity potential of the flow of an ideal fluid with velocity U U U at infinity around the obstacle E E E , that is, a harmonic function V E U V_E^U V E U in R 2 ∖ E \mathbb{R}^2\setminus E R 2 ∖ E with ∂ ν V E U = 0 \partial_\nu V_E^U=0 ∂ ν V E U = 0 on ∂ E \partial E ∂ E and ∇ V E U ( y ) → U \nabla V_E^U(y)\to U ∇ V E U ( y ) → U as ∣ y ∣ → ∞ |y|\to\infty ∣ y ∣ → ∞ . For the unit disc this function is explicit, ( r + r − 1 ) ∣ U ∣ cos ( θ − α ) (r+r^{-1})|U|\cos(\theta-\alpha) ( r + r − 1 ) ∣ U ∣ cos ( θ − α ) in polar coordinates, where U = ∣ U ∣ e i α U=|U|e^{i\alpha} U = ∣ U ∣ e i α , and we transport it to the exterior of E E E by a conformal map.
Let D : = { ∣ ζ ∣ < 1 } \mathbb{D}:=\{|\zeta|<1\} D := { ∣ ζ ∣ < 1 } be the unit disc, and let
f : C ∖ D ‾ → C ∖ E f:\mathbb{C}\setminus\overline{\mathbb{D}}\to\mathbb{C}\setminus E f : C ∖ D → C ∖ E
be the conformal map with f ( ∞ ) = ∞ f(\infty)=\infty f ( ∞ ) = ∞ such that
f ( ζ ) = c ∗ ζ + O ( 1 ) as ζ → ∞ , f(\zeta)=c_*\zeta+O(1)\qquad\text{as }\zeta\to\infty, f ( ζ ) = c ∗ ζ + O ( 1 ) as ζ → ∞ ,
where c ∗ > 0 c_*>0 c ∗ > 0 is the logarithmic capacity of E E E , see Remark 3.1.
Remark 3.1. We note that, by standard complex analysis, a map f f f with the properties above exists and is unique. Indeed, if g g g is the conformal map of D \mathbb{D} D onto the simply connected domain { 0 } ∪ { 1 / y : y ∈ C ∖ E } \{0\}\cup\{1/y:y\in\mathbb{C}\setminus E\} { 0 } ∪ { 1/ y : y ∈ C ∖ E } with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 and g ′ ( 0 ) > 0 g'(0)>0 g ′ ( 0 ) > 0 , given by the Riemann mapping theorem, then f ( ζ ) : = 1 / g ( 1 / ζ ) f(\zeta):=1/g(1/\zeta) f ( ζ ) := 1/ g ( 1/ ζ ) has the required properties, with c ∗ = 1 / g ′ ( 0 ) c_*=1/g'(0) c ∗ = 1/ g ′ ( 0 ) . Moreover, if f 1 f_1 f 1 and f 2 f_2 f 2 are two such maps, then ζ ↦ \zeta\mapsto ζ ↦ 1 / ( f 2 − 1 ∘ f 1 ) ( 1 / ζ ) 1/(f_2^{-1} \circ f_1)(1/\zeta) 1/ ( f 2 − 1 ∘ f 1 ) ( 1/ ζ ) is a conformal automorphism of D \mathbb{D} D fixing 0 0 0 , that is, a rotation, and the normalization at ∞ \infty ∞ forces it to be the identity. Recall also that log ∣ f − 1 ∣ \log|f^{-1}| log ∣ f − 1 ∣ is the Green function of C ∖ E \mathbb{C} \setminus E C ∖ E with pole at ∞ \infty ∞ , and that
log ∣ f − 1 ( y ) ∣ = log ∣ y ∣ − log c ∗ + o ( 1 ) as y → ∞ , \log|f^{-1}(y)| = \log|y| - \log c_* + o(1) \qquad\text{as } y \to\infty, log ∣ f − 1 ( y ) ∣ = log ∣ y ∣ − log c ∗ + o ( 1 ) as y → ∞ ,
which is why c ∗ c_* c ∗ is the logarithmic capacity of E E E , see [35 Chapter 5] . Since ∂ E \partial E ∂ E is a Jordan curve, f f f extends to a homeomorphism of C ∖ D \mathbb{C} \setminus\mathbb{D} C ∖ D onto C ∖ int E \mathbb{C} \setminus\operatorname{int} E C ∖ int E , see [34 ] .
We now transport the potential of the disc to the exterior of E E E by means of f f f .
Lemma 3.2. For U ∈ R 2 ≡ C U \in\mathbb{R}^2 \equiv\mathbb{C} U ∈ R 2 ≡ C , let V E U : C ∖ int E → R V_E^U:\mathbb{C} \setminus\operatorname{int} E \to\mathbb{R} V E U : C ∖ int E → R be defined by
V E U ( f ( ζ ) ) : = c ∗ Re ( U ‾ ζ + U ζ − 1 ) = c ∗ ∣ U ∣ ( r + r − 1 ) cos ( θ − α ) , ζ = r e i θ , U = ∣ U ∣ e i α . (11) V_E^U(f(\zeta)) := c_* \operatorname{Re}(\overline{U}\zeta+ U\zeta^{-1}) = c_* |U|(r+r^{-1})\cos(\theta-\alpha), \qquad\zeta= r e^{i\theta},\ U=|U|e^{i\alpha}. \tag*{(11)} V E U ( f ( ζ )) := c ∗ Re ( U ζ + U ζ − 1 ) = c ∗ ∣ U ∣ ( r + r − 1 ) cos ( θ − α ) , ζ = r e i θ , U = ∣ U ∣ e i α . (11) Then V E U V_E^U V E U depends linearly on U U U , it is harmonic in R 2 ∖ E \mathbb{R}^2 \setminus E R 2 ∖ E , continuous up to ∂ E \partial E ∂ E , smooth up to the open sides of E E E , with ∇ V E U ∈ L 2 ( B ( 0 , R ) ∖ E ) \nabla V_E^U \in L^2(B(0,R)\setminus E) ∇ V E U ∈ L 2 ( B ( 0 , R ) ∖ E ) for every R > 1 R>1 R > 1 , and it satisfies
∫ R 2 ∖ E ∇ V E U ⋅ ∇ φ = 0 for every φ ∈ W 1 , 2 ( R 2 ∖ E ) with bounded support . \int_{\mathbb{R}^2\setminus E} \nabla V_E^U \cdot\nabla\varphi= 0 \qquad\text{for every } \varphi\in W^{1,2}(\mathbb{R}^2\setminus E)\text{ with bounded support}. ∫ R 2 ∖ E ∇ V E U ⋅ ∇ φ = 0 for every φ ∈ W 1 , 2 ( R 2 ∖ E ) with bounded support . The function d U ( y ) : = V E U ( y ) − U ⋅ y d^U(y):=V_E^U(y)-U\cdot y d U ( y ) := V E U ( y ) − U ⋅ y satisfies
∣ d U ( y ) ∣ ≲ ∣ U ∣ ∣ y ∣ − 1 for y ∈ R 2 ∖ E , ∣ ∇ d U ( y ) ∣ ≲ ∣ U ∣ ∣ y ∣ − 2 for ∣ y ∣ ≥ 3. (12) |d^U(y)|\lesssim|U|\,|y|^{-1}\qquad\text{for } y\in\mathbb{R}^2\setminus E,\qquad|\nabla d^U(y)|\lesssim|U|\,|y|^{-2}\qquad\text{for } |y|\ge3. \tag*{(12)} ∣ d U ( y ) ∣ ≲ ∣ U ∣ ∣ y ∣ − 1 for y ∈ R 2 ∖ E , ∣∇ d U ( y ) ∣ ≲ ∣ U ∣ ∣ y ∣ − 2 for ∣ y ∣ ≥ 3. (12) Moreover, if ∂ s \partial_s ∂ s denotes the derivative with respect to arc length along ∂ E \partial E ∂ E , we have
∫ ∂ E ∣ ∂ s V E U ∣ d H 1 = 8 c ∗ ∣ U ∣ , and c ∗ > 1. (13) \int_{\partial E}|\partial_s V_E^U|\,d\mathcal{H}^1=8c_*|U|,\qquad\text{and}\qquad c_*>1. \tag*{(13)} ∫ ∂ E ∣ ∂ s V E U ∣ d H 1 = 8 c ∗ ∣ U ∣ , and c ∗ > 1. (13)
Proof. We define W ( ζ ) : = c ∗ Re ( U ‾ ζ + U ζ − 1 ) W(\zeta):=c_*\operatorname{Re}(\overline{U}\zeta+U\zeta^{-1}) W ( ζ ) := c ∗ Re ( U ζ + U ζ − 1 ) . This function is harmonic in { ∣ ζ ∣ > 1 } \{|\zeta|>1\} { ∣ ζ ∣ > 1 } , smooth up to the unit circle, and its radial derivative vanishes on the circle. Thus V E U = W ∘ f − 1 V_E^U=W\circ f^{-1} V E U = W ∘ f − 1 is harmonic in R 2 ∖ E \mathbb{R}^2\setminus E R 2 ∖ E and continuous up to ∂ E \partial E ∂ E . By the conformal invariance of the Dirichlet integral, if we write A R : = { 1 < ∣ ζ ∣ < R } A_R:=\{1<|\zeta|<R\} A R := { 1 < ∣ ζ ∣ < R } , then
∫ f ( A R ) ∣ ∇ V E U ∣ 2 = ∫ A R ∣ ∇ W ∣ 2 < ∞ . \int_{f(A_R)}|\nabla V_E^U|^2=\int_{A_R}|\nabla W|^2<\infty. ∫ f ( A R ) ∣∇ V E U ∣ 2 = ∫ A R ∣∇ W ∣ 2 < ∞.
Since every set B ( 0 , R ′ ) ∖ E B(0,R')\setminus E B ( 0 , R ′ ) ∖ E is contained in f ( A R ) f(A_R) f ( A R ) for R R R large enough, we deduce that ∇ V E U ∈ L 2 ( B ( 0 , R ′ ) ∖ E ) \nabla V_E^U\in L^2(B(0,R')\setminus E) ∇ V E U ∈ L 2 ( B ( 0 , R ′ ) ∖ E ) for every R ′ > 1 R'>1 R ′ > 1 . Let us now check the weak equation. We take φ ∈ C c ∞ ( R 2 ) \varphi\in C^\infty_c(\mathbb{R}^2) φ ∈ C c ∞ ( R 2 ) and R R R so large that φ ∘ f \varphi\circ f φ ∘ f vanishes in { ∣ ζ ∣ ≥ R / 2 } \{|\zeta|\ge R/2\} { ∣ ζ ∣ ≥ R /2 } . Then φ ∘ f \varphi\circ f φ ∘ f is bounded and
∫ A R ∣ ∇ ( φ ∘ f ) ∣ 2 = ∫ f ( A R ) ∣ ∇ φ ∣ 2 < ∞ , \int_{A_R}|\nabla(\varphi\circ f)|^2=\int_{f(A_R)}|\nabla\varphi|^2<\infty, ∫ A R ∣∇ ( φ ∘ f ) ∣ 2 = ∫ f ( A R ) ∣∇ φ ∣ 2 < ∞ ,
so that φ ∘ f ∈ W 1 , 2 ( A R ) \varphi\circ f\in W^{1,2}(A_R) φ ∘ f ∈ W 1 , 2 ( A R ) . We do not want to use any regularity of f f f up to the unit circle, and so we approximate φ ∘ f \varphi\circ f φ ∘ f in W 1 , 2 ( A R ) W^{1,2}(A_R) W 1 , 2 ( A R ) by functions ξ k ∈ C ∞ ( A R ‾ ) \xi_k\in C^\infty(\overline{A_R}) ξ k ∈ C ∞ ( A R ) vanishing near { ∣ ζ ∣ = R } \{|\zeta|=R\} { ∣ ζ ∣ = R } . As W W W is smooth up to the unit circle and ∂ r W = 0 \partial_rW=0 ∂ r W = 0 there, an integration by parts gives ∫ A R ∇ W ⋅ ∇ ξ k = 0 \int_{A_R}\nabla W\cdot\nabla\xi_k=0 ∫ A R ∇ W ⋅ ∇ ξ k = 0 , and letting k → ∞ k\to\infty k → ∞ we obtain
∫ R 2 ∖ E ∇ V E U ⋅ ∇ φ = ∫ A R ∇ W ⋅ ∇ ( φ ∘ f ) = 0. \int_{\mathbb{R}^2\setminus E}\nabla V_E^U\cdot\nabla\varphi=\int_{A_R}\nabla W\cdot\nabla(\varphi\circ f)=0. ∫ R 2 ∖ E ∇ V E U ⋅ ∇ φ = ∫ A R ∇ W ⋅ ∇ ( φ ∘ f ) = 0.
Now let φ ∈ W 1 , 2 ( R 2 ∖ E ) \varphi\in W^{1,2}(\mathbb{R}^2\setminus E) φ ∈ W 1 , 2 ( R 2 ∖ E ) have bounded support, say in B ( 0 , R ′ − 1 ) B(0,R'-1) B ( 0 , R ′ − 1 ) for some R ′ > 2 R'>2 R ′ > 2 . Since B ( 0 , R ′ ) ∖ E B(0,R')\setminus E B ( 0 , R ′ ) ∖ E is a bounded Lipschitz domain, φ \varphi φ is the limit in W 1 , 2 ( B ( 0 , R ′ ) ∖ E ) W^{1,2}(B(0,R')\setminus E) W 1 , 2 ( B ( 0 , R ′ ) ∖ E ) of restrictions of functions in C c ∞ ( B ( 0 , R ′ ) ) C^\infty_c(B(0,R')) C c ∞ ( B ( 0 , R ′ )) . As ∇ V E U ∈ L 2 ( B ( 0 , R ′ ) ∖ E ) \nabla V_E^U\in L^2(B(0,R')\setminus E) ∇ V E U ∈ L 2 ( B ( 0 , R ′ ) ∖ E ) , we can pass to the limit in the previous equation, and we get the weak equation for φ \varphi φ . Reflecting evenly across the open sides of E E E , as in the proof of Lemma 2.2, we see that V E U V_E^U V E U is smooth up to these sides. Concerning (12) , since − E = E -E=E − E = E , the map ζ ↦ − f ( − ζ ) \zeta\mapsto-f(-\zeta) ζ ↦ − f ( − ζ ) has the same properties as f f f , and so it coincides with f f f , by the uniqueness in Remark 3.1. Thus f f f is odd, and its Laurent expansion at ∞ \infty ∞ has no constant term, that is, f ( ζ ) = c ∗ ζ + O ( ∣ ζ ∣ − 1 ) f(\zeta)=c_*\zeta+O(|\zeta|^{-1}) f ( ζ ) = c ∗ ζ + O ( ∣ ζ ∣ − 1 ) as ζ → ∞ \zeta\to\infty ζ → ∞ . Since U ⋅ f ( ζ ) = Re ( U ‾ f ( ζ ) ) U\cdot f(\zeta)=\operatorname{Re}(\overline{U}f(\zeta)) U ⋅ f ( ζ ) = Re ( U f ( ζ )) , by (11) we get
d U ( f ( ζ ) ) = Re ( U ‾ ( c ∗ ζ − f ( ζ ) ) + c ∗ U ζ − 1 ) = O ( ∣ U ∣ ∣ ζ ∣ − 1 ) . d^U(f(\zeta))=\operatorname{Re}\left(\overline{U}\left(c_*\zeta-f(\zeta)\right)+c_*U\zeta^{-1}\right)=O(|U||\zeta|^{-1}). d U ( f ( ζ )) = Re ( U ( c ∗ ζ − f ( ζ ) ) + c ∗ U ζ − 1 ) = O ( ∣ U ∣∣ ζ ∣ − 1 ) .
Together with the facts that ∣ f ( ζ ) ∣ ≈ ∣ ζ ∣ |f(\zeta)|\approx|\zeta| ∣ f ( ζ ) ∣ ≈ ∣ ζ ∣ and that ∣ d U ∣ ≤ C ∣ U ∣ |d^U|\le C|U| ∣ d U ∣ ≤ C ∣ U ∣ on compact sets, this yields the first estimate in (12) . For the second one, we apply interior estimates to the harmonic function d U d^U d U in the balls B ( y , ∣ y ∣ / 2 ) B(y,|y|/2) B ( y , ∣ y ∣/2 ) with ∣ y ∣ ≥ 3 |y|\ge3 ∣ y ∣ ≥ 3 , which do not intersect E E E , and we use the first estimate.
Let us prove (13) . On the unit circle we have W ( e i θ ) = 2 c ∗ ∣ U ∣ cos ( θ − α ) W(e^{i\theta})=2c_*|U|\cos(\theta-\alpha) W ( e i θ ) = 2 c ∗ ∣ U ∣ cos ( θ − α ) , and the total variation of this function for θ ∈ [ 0 , 2 π ] \theta\in[0,2\pi] θ ∈ [ 0 , 2 π ] is 8 c ∗ ∣ U ∣ 8c_*|U| 8 c ∗ ∣ U ∣ . A homeomorphic change of parameter does not affect the total variation of a continuous function on a closed curve, and so V E U V_E^U V E U also has total variation 8 c ∗ ∣ U ∣ 8c_*|U| 8 c ∗ ∣ U ∣ along ∂ E \partial E ∂ E . Next, let Σ \Sigma Σ be a closed side of E E E . On Σ \Sigma Σ the function V E U V_E^U V E U is continuous, and in the interior of Σ \Sigma Σ it is C 1 C^1 C 1 . Therefore, on each closed subinterval of the interior of Σ \Sigma Σ , the total variation of V E U V_E^U V E U coincides with the integral of ∣ ∂ s V E U ∣ |\partial_sV_E^U| ∣ ∂ s V E U ∣ . On the other hand, by continuity at the endpoints, any partition of Σ \Sigma Σ can be approximated by partitions of closed subintervals of the interior. So, if we take closed subintervals exhausting the interior of Σ \Sigma Σ , their total variations increase to the total variation on Σ \Sigma Σ , which by monotone convergence equals ∫ Σ ∣ ∂ s V E U ∣ d H 1 \int_\Sigma|\partial_sV_E^U|\,d\mathcal{H}^1 ∫ Σ ∣ ∂ s V E U ∣ d H 1 . Summing over the four sides, we get the identity in (13) , and also the integrability of ∂ s V E U \partial_sV_E^U ∂ s V E U on ∂ E \partial E ∂ E .
Finally, we check that c ∗ > 1 c_*>1 c ∗ > 1 . Recall from Remark 3.1 that c ∗ c_* c ∗ is the logarithmic capacity of E E E , and let μ E \mu_E μ E be the equilibrium measure of E E E . It is a probability measure supported on ∂ E \partial E ∂ E with finite logarithmic energy, and so it has no atoms. Moreover, by Frostman’s theorem, its logarithmic potential is equal to log c ∗ \log c_* log c ∗ in the interior of E E E , that is,
∫ ∂ E log ∣ x − z ∣ d μ E ( z ) = log c ∗ for every x ∈ int E . \int_{\partial E}\log|x-z|\,d\mu_E(z)=\log c_* \qquad\text{for every }x\in\operatorname{int}E. ∫ ∂ E log ∣ x − z ∣ d μ E ( z ) = log c ∗ for every x ∈ int E .
Taking x = 0 x=0 x = 0 , we get log c ∗ = ∫ ∂ E log ∣ z ∣ d μ E ( z ) \log c_*=\int_{\partial E}\log|z|\,d\mu_E(z) log c ∗ = ∫ ∂ E log ∣ z ∣ d μ E ( z ) . Now ∣ z ∣ ≥ 1 |z|\ge1 ∣ z ∣ ≥ 1 for every z ∈ ∂ E z\in\partial E z ∈ ∂ E , with equality only at the midpoints of the four sides. Since μ E \mu_E μ E has no atoms, the integrand is positive μ E \mu_E μ E -almost everywhere, and therefore log c ∗ > 0 \log c_*>0 log c ∗ > 0 . □
Remark 3.3. The Schwarz-Christoffel formula gives f ′ ( ζ ) = c ∗ ( 1 + ζ − 4 ) 1 / 2 f'(\zeta)=c_*(1+\zeta^{-4})^{1/2} f ′ ( ζ ) = c ∗ ( 1 + ζ − 4 ) 1/2 , so that
f ( ζ ) = c ∗ ( ζ + ∑ k ≥ 1 ( 1 / 2 k ) ζ 1 − 4 k 1 − 4 k ) . f(\zeta)=c_*\left(\zeta+\sum_{k\ge1}\binom{1/2}{k}\frac{\zeta^{1-4k}}{1-4k}\right). f ( ζ ) = c ∗ ( ζ + k ≥ 1 ∑ ( k 1/2 ) 1 − 4 k ζ 1 − 4 k ) .
Imposing that the side length of E E E be 2 gives c ∗ = Γ ( 1 / 4 ) 2 / ( 2 π 3 / 2 ) ≈ 1.18034 c_*=\Gamma(1/4)^2/(2\pi^{3/2})\approx1.18034 c ∗ = Γ ( 1/4 ) 2 / ( 2 π 3/2 ) ≈ 1.18034 , which is the classical value of the logarithmic capacity of a square of side 2, see [35 ] , Table 5.1, p. 135]. The rest of the paper uses only c ∗ > 1 c_*>1 c ∗ > 1 , which just expresses the fact that the logarithmic capacity of E E E is larger than that of the unit disc D ⊆ E \mathbb{D}\subseteq E D ⊆ E . By (13) , the mean value of the speed ∣ ∂ s V E U ∣ |\partial_sV_E^U| ∣ ∂ s V E U ∣ on ∂ E \partial E ∂ E is c ∗ ∣ U ∣ c_*|U| c ∗ ∣ U ∣ , and so on average the fluid moves faster along ∂ E \partial E ∂ E than at infinity. For the unit disc, the mean value of the speed on the unit circle is even larger, namely δ ∣ U ∣ / ( 2 π ) = ( 4 / π ) ∣ U ∣ ≈ 1.27 ∣ U ∣ \delta|U|/(2\pi)=(4/\pi)|U|\approx1.27|U| δ ∣ U ∣/ ( 2 π ) = ( 4/ π ) ∣ U ∣ ≈ 1.27∣ U ∣ , although its logarithmic capacity is 1. However, in our construction it is convenient to use a square because its sides are flat: new “columns” will be inserted on squares contained in the walls of the columns of the previous generation, see Figure 1 .
It is convenient to express the L 1 L^1 L 1 norm in (13) by duality, testing against a fixed finite family of smooth functions supported away from the vertices of E E E . We parametrize each side of E E E by the arc length s ∈ [ − 1 , 1 ] s\in[-1,1] s ∈ [ − 1 , 1 ] , measured from its midpoint, counterclockwise.
Lemma 3.4. For every η > 0 \eta> 0 η > 0 , there exist w ∈ ( 0 , 1 / 32 ) w \in(0,1/32) w ∈ ( 0 , 1/32 ) and finitely many smooth functions ψ 1 , … , ψ m \psi_1,\ldots,\psi_m ψ 1 , … , ψ m on ∂ E \partial E ∂ E , with ∣ ψ j ∣ ≤ 1 |\psi_j| \le1 ∣ ψ j ∣ ≤ 1 and with support in the set where ∣ s ∣ < 1 − 2 w |s| < 1 - 2w ∣ s ∣ < 1 − 2 w on each side, such that
max 1 ≤ j ≤ m ∫ ∂ E ∂ s V E U ψ j d H 1 ≥ 8 ( c ∗ − η ) ∣ U ∣ , for every U ∈ R 2 . (14) \max_{1\le j\le m}\int_{\partial E} \partial_s V_E^U \psi_j\,d\mathcal{H}^1 \ge8(c_*-\eta)|U|,\qquad\text{for every } U\in\mathbb{R}^2. \tag*{(14)} 1 ≤ j ≤ m max ∫ ∂ E ∂ s V E U ψ j d H 1 ≥ 8 ( c ∗ − η ) ∣ U ∣ , for every U ∈ R 2 . (14)
Proof. By homogeneity we assume, without loss of generality, that ∣ U ∣ = 1 |U|=1 ∣ U ∣ = 1 . Since ∂ s V E U \partial_s V_E^U ∂ s V E U is integrable and continuous on the open sides, there is a smooth function ψ U \psi_U ψ U with ∣ ψ U ∣ ≤ 1 |\psi_U|\le1 ∣ ψ U ∣ ≤ 1 , compactly supported in the open sides, for instance, a smooth approximation of the sign of ∂ s V E U \partial_s V_E^U ∂ s V E U , such that
∫ ∂ E ∂ s V E U ψ U d H 1 > 8 c ∗ − 4 η . \int_{\partial E}\partial_s V_E^U\psi_U\,d\mathcal{H}^1>8c_*-4\eta. ∫ ∂ E ∂ s V E U ψ U d H 1 > 8 c ∗ − 4 η .
If ∣ U ′ ∣ = 1 |U'|=1 ∣ U ′ ∣ = 1 , by (13) and the linearity of U ↦ V E U U\mapsto V_E^U U ↦ V E U , we have
∣ ∫ ∂ E ( ∂ s V E U − ∂ s V E U ′ ) ψ U d H 1 ∣ ≤ ∫ ∂ E ∣ ∂ s V E U − U ′ ∣ d H 1 = 8 c ∗ ∣ U − U ′ ∣ . \left|\int_{\partial E}\left(\partial_sV_E^U-\partial_sV_E^{U'}\right)\psi_U\,d\mathcal{H}^1\right|\le\int_{\partial E}\left|\partial_sV_E^{U-U'}\right|\,d\mathcal{H}^1=8c_*|U-U'|. ∫ ∂ E ( ∂ s V E U − ∂ s V E U ′ ) ψ U d H 1 ≤ ∫ ∂ E ∂ s V E U − U ′ d H 1 = 8 c ∗ ∣ U − U ′ ∣.
Thus, if moreover ∣ U − U ′ ∣ ≤ η / ( 2 c ∗ ) |U-U'|\le\eta/(2c_*) ∣ U − U ′ ∣ ≤ η / ( 2 c ∗ ) , we obtain
∫ ∂ E ∂ s V E U ′ ψ U d H 1 ≥ 8 ( c ∗ − η ) . \int_{\partial E}\partial_sV_E^{U'}\psi_U\,d\mathcal{H}^1\ge8(c_*-\eta). ∫ ∂ E ∂ s V E U ′ ψ U d H 1 ≥ 8 ( c ∗ − η ) .
By the compactness of the unit circle, finitely many of the functions ψ U \psi_U ψ U suffice; these are the functions ψ j \psi_j ψ j , and w w w is chosen so small that all of them are supported where ∣ s ∣ < 1 − 2 w |s|<1-2w ∣ s ∣ < 1 − 2 w .
:::
A periodic array of tall square columns
In this section we study the Neumann problem in the upper half-space from which a periodic array of tall square columns has been removed, assuming that the gradient of the solution at infinity is a constant horizontal vector U U U .
The main idea is that, far from its top and its bottom, a tall column looks like an infinite one, and so, in each horizontal section, the solution should be close to the planar solution of Section 3. Then, by (13) , the modulus of the tangential gradient on the walls of the column should have mean value close to c ∗ ∣ U ∣ c_*|U| c ∗ ∣ U ∣ , which is larger than ∣ U ∣ |U| ∣ U ∣ because c ∗ > 1 c_*>1 c ∗ > 1 . We choose the period so that the area of the floor of a period cell is comparable to the area of the walls of one column. In this way, the columns transform a piece of the floor, where the tangential gradient has modulus ∣ U ∣ |U| ∣ U ∣ , into walls, where its mean value is about c ∗ ∣ U ∣ c_*|U| c ∗ ∣ U ∣ . At the end of the section, we select some small squares on the walls, on which the next generation of columns will be built.
The parameters and the period cells
We first fix, once and for all, numbers η > 0 \eta>0 η > 0 and A ∈ ( 1 / 2 , 1 ) A\in(1/2,1) A ∈ ( 1/2 , 1 ) such that
γ ∗ : = A ( c ∗ − 2 η ) > 1. (15) \gamma_*:=A(c_*-2\eta)>1. \tag*{(15)} γ ∗ := A ( c ∗ − 2 η ) > 1. (15)
This is possible because c ∗ > 1 c_*>1 c ∗ > 1 ; for instance, one can take η = 10 − 2 \eta=10^{-2} η = 1 0 − 2 and A = 9 / 10 A=9/10 A = 9/10 .
Next, we take the functions ψ 1 , … , ψ m \psi_1,\ldots,\psi_m ψ 1 , … , ψ m and the number w w w given by Lemma 3.4 for this η \eta η . Then, let λ ≥ 32 \lambda\ge32 λ ≥ 32 be a large integer, to be fixed in Lemma 4.2, and set
Π : = E × [ 0 , λ ] , R : = { x ∈ ∂ E × [ 0 , λ ] : ∣ s ( x ) ∣ ≤ 1 − 2 w , 8 ≤ t ≤ λ − 8 } , \Pi:=E\times[0,\lambda],\qquad\mathcal{R}:=\{x\in\partial E\times[0,\lambda]: |s(x)|\le1-2w,\ 8\le t\le\lambda-8\}, Π := E × [ 0 , λ ] , R := { x ∈ ∂ E × [ 0 , λ ] : ∣ s ( x ) ∣ ≤ 1 − 2 w , 8 ≤ t ≤ λ − 8 } ,
where, as in Section 3, s ( x ) s(x) s ( x ) is the arc length coordinate on the side of E E E containing the horizontal projection of x x x , and t t t is the vertical component of x x x . Thus R \mathcal{R} R is the union of four closed rectangles, one in the middle of each lateral face of Π \Pi Π , and ∣ R ∣ : = H 2 ( R ) = 8 ( 1 − 2 w ) ( λ − 16 ) |\mathcal{R}|:=\mathcal{H}^2(\mathcal{R})=8(1-2w)(\lambda-16) ∣ R ∣ := H 2 ( R ) = 8 ( 1 − 2 w ) ( λ − 16 ) .
We choose the period ς \varsigma ς so that
ς 2 = ( 1 − w ) 2 ∣ R ∣ A and Y : = ( − ς 2 , ς 2 ) 2 . (16) \varsigma^2=\frac{(1-w)^2|\mathcal{R}|}{A}
\qquad\text{and}\qquad
Y:=\left(-\frac{\varsigma}{2},\frac{\varsigma}{2}\right)^2. \tag*{(16)} ς 2 = A ( 1 − w ) 2 ∣ R ∣ and Y := ( − 2 ς , 2 ς ) 2 . (16)
In particular, the bounds λ / 2 ≤ λ − 16 ≤ λ \lambda/2\leq\lambda-16\leq\lambda λ /2 ≤ λ − 16 ≤ λ yield
4 ( 1 − w ) 2 ( 1 − 2 w ) A λ ≤ ς 2 ≤ 8 A λ , (17) \frac{4(1-w)^2(1-2w)}{A}\lambda\leq\varsigma^2\leq\frac{8}{A}\lambda, \tag*{(17)} A 4 ( 1 − w ) 2 ( 1 − 2 w ) λ ≤ ς 2 ≤ A 8 λ , (17)
so that ς ≈ λ \varsigma\approx\sqrt{\lambda} ς ≈ λ , with constants depending only on A A A , and we assume λ \lambda λ so large that ς > 32 \varsigma>32 ς > 32 . The reason for the factor ( 1 − w ) 2 / A (1-w)^2/A ( 1 − w ) 2 / A will be clear in Subsection 4.4: the selected squares will have total area at most ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣ (1-w)^2|\mathcal{R}|=A|Y| ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣ in each period cell.
Finally, let
P : = R + 3 ∖ ⋃ z ∈ ς Z 2 ( ( z , 0 ) + Π ) , P Y : = P ∩ ( Y × R ) . P:=\mathbb{R}_+^3\setminus\bigcup_{z\in\varsigma\mathbb{Z}^2}\bigl((z,0)+\Pi\bigr),\qquad P_Y:=P\cap(Y\times\mathbb{R}). P := R + 3 ∖ z ∈ ς Z 2 ⋃ ( ( z , 0 ) + Π ) , P Y := P ∩ ( Y × R ) .
We say that a function on P P P is periodic if it is invariant under the translations by ( z , 0 ) (z,0) ( z , 0 ) , z ∈ ς Z 2 z\in\varsigma\mathbb{Z}^2 z ∈ ς Z 2 . See Figure 2 .
Figure 2. (a) A period cell P Y P_Y P Y : a column Π \Pi Π of height λ \lambda λ on a floor of side ς ≈ λ \varsigma\approx\sqrt{\lambda} ς ≈ λ , with the selected squares P i P_i P i (blue) on its walls. (b) Since their total area is less than ∣ Y ∣ |Y| ∣ Y ∣ , the squares P i P_i P i can be rearranged into disjoint squares D i D_i D i of the floor Y Y Y .
The cell problem
We first introduce the space in which we will solve the cell problem. Namely, we denote by H \mathcal{H} H the completion of the space of periodic functions φ ∈ C ∞ ( P ‾ ) \varphi\in C^\infty(\overline{P}) φ ∈ C ∞ ( P ) which vanish for t t t large, with respect to the norm ∥ ∇ φ ∥ L 2 ( P Y ) \|\nabla\varphi\|_{L^2(P_Y)} ∥∇ φ ∥ L 2 ( P Y ) (here, as usual, t t t stands for the vertical coordinate in R 3 \mathbb{R}^3 R 3 ).
Since the sets P Y ∩ { t < T } P_Y\cap\{t<T\} P Y ∩ { t < T } are bounded and connected, the Poincaré inequality in each of them shows that every element of H \mathcal{H} H can be identified with a function in W l o c 1 , 2 ( P ‾ ) W_{\mathrm{loc}}^{1,2}(\overline{P}) W loc 1 , 2 ( P ) , which is unique up to an additive constant.
We look for the solution of the Neumann problem in P P P with gradient ( U , 0 ) (U,0) ( U , 0 ) at infinity of the form V U = U ⋅ y + P U V^U=U\cdot y+P^U V U = U ⋅ y + P U . Each of the two terms plays a different role: the linear function U ⋅ y U\cdot y U ⋅ y has the right behaviour at infinity, while the periodic function P U P^U P U corrects the boundary condition on the walls of the columns.
Lemma 4.1. For every U ∈ R 2 U \in\mathbb{R}^{2} U ∈ R 2 there exists a unique P U ∈ H P^{U} \in\mathcal{H} P U ∈ H , depending linearly on U U U , such that
∫ P Y ∇ P U ⋅ ∇ φ = − ∫ P Y ∩ { t < λ } U ⋅ ∇ y φ for every φ ∈ H . (18) \int_{P_Y} \nabla P^{U} \cdot\nabla\varphi= -\int_{P_Y\cap\{t<\lambda\}} U \cdot\nabla_y\varphi\qquad\text{for every }\varphi\in\mathcal{H}. \tag*{(18)} ∫ P Y ∇ P U ⋅ ∇ φ = − ∫ P Y ∩ { t < λ } U ⋅ ∇ y φ for every φ ∈ H . (18) Equivalently, the function V U ( y , t ) : = U ⋅ y + P U ( y , t ) V^{U}(y,t):=U\cdot y+P^{U}(y,t) V U ( y , t ) := U ⋅ y + P U ( y , t ) satisfies
∫ P Y ∇ V U ⋅ ∇ φ = 0 for every periodic φ ∈ C ∞ ( P ‾ ) vanishing for large t . (19) \int_{P_Y} \nabla V^{U}\cdot\nabla\varphi=0 \qquad\text{for every periodic }\varphi\in C^{\infty}(\overline{P})\text{ vanishing for large }t. \tag*{(19)} ∫ P Y ∇ V U ⋅ ∇ φ = 0 for every periodic φ ∈ C ∞ ( P ) vanishing for large t . (19) Moreover, ∫ P ∇ V U ⋅ ∇ φ = 0 \int_{P}\nabla V^{U}\cdot\nabla\varphi=0 ∫ P ∇ V U ⋅ ∇ φ = 0 for every φ ∈ W 1 , 2 ( P ) \varphi\in W^{1,2}(P) φ ∈ W 1 , 2 ( P ) with bounded support. The additive constant in P U P^{U} P U can be chosen so that P U ∈ L 2 ( P Y ) P^{U}\in L^{2}(P_Y) P U ∈ L 2 ( P Y ) , and then there exists C λ C_{\lambda} C λ , depending on λ \lambda λ (and on A A A and w w w , through ς \varsigma ς ), such that
∫ P Y ( ∣ P U ∣ 2 + ∣ ∇ P U ∣ 2 ) ≤ C λ ∣ U ∣ 2 . (20) \int_{P_Y}\left(|P^{U}|^{2}+|\nabla P^{U}|^{2}\right)\leq C_{\lambda}|U|^{2}. \tag*{(20)} ∫ P Y ( ∣ P U ∣ 2 + ∣∇ P U ∣ 2 ) ≤ C λ ∣ U ∣ 2 . (20) Finally, V U V^{U} V U is smooth up to the open faces of ∂ P \partial P ∂ P , and all its derivatives are bounded by ∣ U ∣ |U| ∣ U ∣ times a constant on every compact subset of an open face.
Thus V U V^{U} V U is the solution of the periodic Neumann problem in P P P with gradient ( U , 0 ) (U,0) ( U , 0 ) at infinity. Indeed, interior estimates, periodicity, and the finiteness of the gradient energy imply that ∇ P U ( y , t ) → 0 \nabla P^{U}(y,t)\to0 ∇ P U ( y , t ) → 0 uniformly in y y y as t → + ∞ t\to+\infty t → + ∞ . Moreover, P U = U 1 P e 1 + U 2 P e 2 P^{U}=U_{1}P^{e_{1}}+U_{2}P^{e_{2}} P U = U 1 P e 1 + U 2 P e 2 by linearity. Notice also that P U P^{U} P U and ∇ V U \nabla V^{U} ∇ V U are periodic and
V U ( y + ς z , t ) = V U ( y , t ) + ς U ⋅ z , for z ∈ Z 2 , V^{U}(y+\varsigma z,t)=V^{U}(y,t)+\varsigma U\cdot z,\qquad\text{for }z\in\mathbb{Z}^{2}, V U ( y + ς z , t ) = V U ( y , t ) + ς U ⋅ z , for z ∈ Z 2 ,
so V U V^{U} V U is, in general, not periodic. Moreover, if φ ∈ W 1 , 2 ( P ) \varphi\in W^{1,2}(P) φ ∈ W 1 , 2 ( P ) has bounded support, then Lemma 4.1 and the divergence theorem give
∫ P ∇ P U ⋅ ∇ φ = − ∫ P ( U , 0 ) ⋅ ∇ φ = − ∫ ∂ P φ ( U , 0 ) ⋅ ν P d H 2 , (21) \int_{P}\nabla P^{U}\cdot\nabla\varphi=-\int_{P}(U,0)\cdot\nabla\varphi=-\int_{\partial P}\varphi(U,0)\cdot\nu_{P}\,d\mathcal{H}^{2}, \tag*{(21)} ∫ P ∇ P U ⋅ ∇ φ = − ∫ P ( U , 0 ) ⋅ ∇ φ = − ∫ ∂ P φ ( U , 0 ) ⋅ ν P d H 2 , (21)
where ν P \nu_{P} ν P is the outer unit normal of P P P . Thus, ∂ ν P P U = − ( U , 0 ) ⋅ ν P \partial_{\nu_{P}}P^{U}=-(U,0)\cdot\nu_{P} ∂ ν P P U = − ( U , 0 ) ⋅ ν P on ∂ P \partial P ∂ P in the weak sense, that is, P U P^{U} P U solves the Neumann problem in P P P with datum − ( U , 0 ) ⋅ ν P -(U,0)\cdot\nu_{P} − ( U , 0 ) ⋅ ν P , which vanishes on the floor and on the roofs of the columns. In other words, P U P^{U} P U compensates the normal derivative of U ⋅ y U\cdot y U ⋅ y on the walls: the linear function U ⋅ y U\cdot y U ⋅ y has the right gradient at infinity, but it does not satisfy the homogeneous Neumann condition on the walls, and adding P U P^{U} P U corrects this. For this reason, as is customary in homogenization theory, we refer to P U P^{U} P U as the cell corrector .
Proof of Lemma 4.1 . The right-hand side of (18) is a bounded functional on H \mathcal{H} H , whose norm is controlled by ∣ U ∣ ∣ P Y ∩ { t < λ } ∣ 1 / 2 |U||P_Y\cap\{t<\lambda\}|^{1/2} ∣ U ∣∣ P Y ∩ { t < λ } ∣ 1/2 . So the existence and the uniqueness of P U P^{U} P U follow from the Riesz representation theorem, and the linearity follows from the uniqueness. If φ \varphi φ is smooth, periodic and vanishes for large t t t , then
∫ Y U ⋅ ∇ y φ ( y , t ) d y = 0 for t > λ , \int_{Y}U\cdot\nabla_y\varphi(y,t)\,dy=0\qquad\text{for }t>\lambda, ∫ Y U ⋅ ∇ y φ ( y , t ) d y = 0 for t > λ ,
by periodicity, because every horizontal section of P Y P_Y P Y above the roof of the column is the whole square Y Y Y . Hence the right-hand side of (18) equals − ∫ P Y U ⋅ ∇ y φ -\int_{P_Y}U\cdot\nabla_y\varphi − ∫ P Y U ⋅ ∇ y φ , which gives (19) . If φ ∈ W 1 , 2 ( P ) \varphi\in W^{1,2}(P) φ ∈ W 1 , 2 ( P ) has bounded support, its periodization φ p e r ( x ) : = ∑ z ∈ ς Z 2 φ ( x + ( z , 0 ) ) \varphi^{\mathrm{per}}(x):=\sum_{z\in\varsigma\mathbb{Z}^{2}}\varphi(x+(z,0)) φ per ( x ) := ∑ z ∈ ς Z 2 φ ( x + ( z , 0 )) is periodic and vanishes for large t t t , and so the periodicity of ∇ V U \nabla V^{U} ∇ V U , (19) and a density argument yield
∫ P ∇ V U ⋅ ∇ φ = ∫ P Y ∇ V U ⋅ ∇ φ p e r = 0. \int_{P}\nabla V^{U}\cdot\nabla\varphi=\int_{P_Y}\nabla V^{U}\cdot\nabla\varphi^{\mathrm{per}}=0. ∫ P ∇ V U ⋅ ∇ φ = ∫ P Y ∇ V U ⋅ ∇ φ per = 0.
In Y × ( λ , ∞ ) Y \times(\lambda,\infty) Y × ( λ , ∞ ) , the function P U P^{U} P U is periodic and harmonic. Hence its horizontal mean m ( t ) : = m Y ( P U ( ⋅ , t ) ) m(t):=m_Y(P^{U}(\cdot,t)) m ( t ) := m Y ( P U ( ⋅ , t )) , t > λ t>\lambda t > λ , satisfies
m ′ ′ ( t ) = m Y ( ∂ t 2 P U ( ⋅ , t ) ) = − m Y ( Δ y P U ( ⋅ , t ) ) = 0 , m''(t)=m_Y(\partial_t^2P^{U}(\cdot,t))=-m_Y(\Delta_yP^{U}(\cdot,t))=0, m ′′ ( t ) = m Y ( ∂ t 2 P U ( ⋅ , t )) = − m Y ( Δ y P U ( ⋅ , t )) = 0 ,
by periodicity, and so m m m is affine. Since
∣ m ′ ( t ) ∣ 2 = ∣ m Y ( ∂ t P U ( ⋅ , t ) ) ∣ 2 ≤ m Y ( ∣ ∂ t P U ( ⋅ , t ) ∣ 2 ) |m'(t)|^2=|m_Y(\partial_tP^{U}(\cdot,t))|^2\leq m_Y(|\partial_tP^{U}(\cdot,t)|^2) ∣ m ′ ( t ) ∣ 2 = ∣ m Y ( ∂ t P U ( ⋅ , t )) ∣ 2 ≤ m Y ( ∣ ∂ t P U ( ⋅ , t ) ∣ 2 )
and the energy of P U P^{U} P U is finite, m ′ m' m ′ is square integrable in ( λ , ∞ ) (\lambda,\infty) ( λ , ∞ ) . Thus m m m is constant, and we choose the additive constant in P U P^{U} P U so that m ≡ 0 m\equiv0 m ≡ 0 . The Riesz representation theorem also gives ∥ ∇ P U ∥ L 2 ( P Y ) ≤ ∣ U ∣ ∣ P Y ∩ { t < λ } ∣ 1 / 2 \|\nabla P^{U}\|_{L^2(P_Y)}\leq|U||P_Y\cap\{t<\lambda\}|^{1/2} ∥∇ P U ∥ L 2 ( P Y ) ≤ ∣ U ∣∣ P Y ∩ { t < λ } ∣ 1/2 , which is the bound for ∇ P U \nabla P^{U} ∇ P U in (20) .
For the L 2 L^2 L 2 norm of P U P^{U} P U , note that, since m ≡ 0 m\equiv0 m ≡ 0 , the Poincaré inequality on the torus Y Y Y in each horizontal section gives
∫ P Y ∩ { t > λ } ∣ P U ∣ 2 ≤ ( ς / 2 π ) 2 ∫ P Y ∩ { t > λ } ∣ ∇ y P U ∣ 2 . \int_{P_Y\cap\{t>\lambda\}}|P^{U}|^2\leq(\varsigma/2\pi)^2\int_{P_Y\cap\{t>\lambda\}}|\nabla_yP^{U}|^2. ∫ P Y ∩ { t > λ } ∣ P U ∣ 2 ≤ ( ς /2 π ) 2 ∫ P Y ∩ { t > λ } ∣ ∇ y P U ∣ 2 .
Moreover, P U P^{U} P U has zero mean on Y × ( λ , λ + 1 ) Y\times(\lambda,\lambda+1) Y × ( λ , λ + 1 ) , and the Poincaré inequality on the bounded connected set P Y ∩ { t < λ + 1 } P_Y\cap\{t<\lambda+1\} P Y ∩ { t < λ + 1 } gives the L 2 L^2 L 2 bound in the rest of P Y P_Y P Y . Finally, near an open flat face of ∂ P \partial P ∂ P , the function V U V^{U} V U satisfies a homogeneous Neumann condition, and so its even reflection across that face is harmonic. The last assertion readily follows from interior estimates and (20) . □
The walls see the planar flow
The next lemma makes precise the claim that, in the middle part R \mathcal R R of the walls, V U V^{U} V U behaves like the planar function V E U V_E^{U} V E U . For j = 1 , … , m j=1,\ldots,m j = 1 , … , m we denote by Υ j \Upsilon_j Υ j the horizontal tangent vector field on R \mathcal R R given by Υ j ( x ) = ψ j ( s ( x ) ) e s \Upsilon_j(x)=\psi_j(s(x))e_s Υ j ( x ) = ψ j ( s ( x )) e s , where ψ j \psi_j ψ j is one of the functions introduced in Lemma 3.4 and e s e_s e s is the unit horizontal tangent vector to ∂ E \partial E ∂ E in the counterclockwise direction. Note that ∇ τ V ⋅ Υ j = ψ j ∂ s V \nabla_\tau V\cdot\Upsilon_j=\psi_j\partial_sV ∇ τ V ⋅ Υ j = ψ j ∂ s V on R \mathcal R R for every function V V V which is C 1 C^1 C 1 up to the walls of Π \Pi Π .
Lemma 4.2. If λ \lambda λ is large enough, depending on η \eta η , A A A , w w w and the functions ψ j \psi_j ψ j , then for every U ∈ R 2 U\in\mathbb{R}^2 U ∈ R 2 there is j ∈ { 1 , … , m } j\in\{1,\ldots,m\} j ∈ { 1 , … , m } such that
∫ R ∇ τ V U ⋅ Υ j d H 2 ≥ ( c ∗ − 2 η ) ∣ U ∣ ∣ R ∣ . (22) \int_{\mathcal R}\nabla_\tau V^{U}\cdot\Upsilon_j\,d\mathcal H^2\geq(c_\ast-2\eta)|U||\mathcal R|. \tag*{(22)} ∫ R ∇ τ V U ⋅ Υ j d H 2 ≥ ( c ∗ − 2 η ) ∣ U ∣∣ R ∣. (22) In particular, ∫ R ∣ ∇ τ V U ∣ d H 2 ≥ ( c ∗ − 2 η ) ∣ U ∣ ∣ R ∣ \int_{\mathcal R}|\nabla_\tau V^{U}|\,d\mathcal H^2\geq(c_\ast-2\eta)|U||\mathcal R| ∫ R ∣ ∇ τ V U ∣ d H 2 ≥ ( c ∗ − 2 η ) ∣ U ∣∣ R ∣ .
Proof. We first note that it suffices to prove (22) , which implies the last assertion because ∣ Υ j ∣ ≤ 1 |\Upsilon_j|\leq1 ∣ Υ j ∣ ≤ 1 . We fix U ∈ R 2 U\in\mathbb{R}^2 U ∈ R 2 , and let j j j be given by Lemma 3.4, so that
∫ ∂ E ∂ s V E U ψ j d H 1 ≥ 8 ( c ∗ − η ) ∣ U ∣ . (23) \int_{\partial E}\partial_sV_E^{U}\,\psi_j\,d\mathcal H^1\geq8(c_\ast-\eta)|U|. \tag*{(23)} ∫ ∂ E ∂ s V E U ψ j d H 1 ≥ 8 ( c ∗ − η ) ∣ U ∣. (23)
We compare V U V^{U} V U with the planar solution V E U V_E^{U} V E U , extended to be constant in t t t and cut off far from the column in the horizontal directions and near its roof. For the horizontal cutoff we take a radial function χ ς ∈ C ∞ ( R 2 ) \chi_{\varsigma}\in C^\infty(\mathbb{R}^2) χ ς ∈ C ∞ ( R 2 ) such that χ ς = 1 \chi_{\varsigma}=1 χ ς = 1 on B ( 0 , ς / 8 ) B(0,\varsigma/8) B ( 0 , ς /8 ) , supp χ ς ⊂ B ( 0 , ς / 4 ) \operatorname{supp}\chi_{\varsigma}\subset B(0,\varsigma/4) supp χ ς ⊂ B ( 0 , ς /4 ) and ∣ ∇ χ ς ∣ ≲ ς − 1 |\nabla\chi_{\varsigma}|\lesssim\varsigma^{-1} ∣∇ χ ς ∣ ≲ ς − 1 , and for the vertical one a function χ λ ∈ C ∞ ( R ) \chi_\lambda\in C^\infty(\mathbb{R}) χ λ ∈ C ∞ ( R ) such that χ λ = 1 \chi_\lambda=1 χ λ = 1 on ( − ∞ , λ − 8 ] (-\infty,\lambda-8] ( − ∞ , λ − 8 ] , χ λ = 0 \chi_\lambda=0 χ λ = 0 on [ λ − 2 , ∞ ) [\lambda-2,\infty) [ λ − 2 , ∞ ) and ∣ χ λ ′ ∣ ≤ 1 |\chi_\lambda'|\leq1 ∣ χ λ ′ ∣ ≤ 1 . With d U : = V E U − U ⋅ y d^{U}:=V_E^{U}-U\cdot y d U := V E U − U ⋅ y as in Lemma 3.2, we let q q q be the periodic extension of χ ς d U \chi_{\varsigma}d^{U} χ ς d U , and we define
V ~ ( y , t ) : = U ⋅ y + χ λ ( t ) q ( y ) , e : = V U − V ~ . \widetilde V(y,t):=U\cdot y+\chi_\lambda(t)q(y),\qquad e:=V^{U}-\widetilde V. V ( y , t ) := U ⋅ y + χ λ ( t ) q ( y ) , e := V U − V .
Observe that V ~ = V E U \widetilde V=V_E^{U} V = V E U on R \mathcal R R , where both cutoffs are equal to 1 1 1 . Moreover, e = P U − χ λ q ∈ H e=P^{U}-\chi_\lambda q\in\mathcal H e = P U − χ λ q ∈ H , because χ λ q \chi_\lambda q χ λ q is periodic, has finite energy in P Y P_Y P Y , and vanishes for t ≥ λ − 2 t\geq\lambda-2 t ≥ λ − 2 . Since V U = V ~ + e V^U=\widetilde V+e V U = V + e , we have
∫ R ∇ τ V U ⋅ Υ j d H 2 = ∫ R ∇ τ V ~ ⋅ Υ j d H 2 + ∫ R ∇ τ e ⋅ Υ j d H 2 . \int_{\mathcal R}\nabla_\tau V^U\cdot\Upsilon_j\,d\mathcal H^2=\int_{\mathcal R}\nabla_\tau\widetilde V\cdot\Upsilon_j\,d\mathcal H^2+\int_{\mathcal R}\nabla_\tau e\cdot\Upsilon_j\,d\mathcal H^2. ∫ R ∇ τ V U ⋅ Υ j d H 2 = ∫ R ∇ τ V ⋅ Υ j d H 2 + ∫ R ∇ τ e ⋅ Υ j d H 2 .
We will compute the first integral on the right-hand side explicitly. The second one is an error term, and we will bound it in terms of the energy of e e e . More precisely, we claim that
∫ R ∇ τ V ~ ⋅ Υ j d H 2 ≥ ( c ∗ − η ) ∣ U ∣ ∣ R ∣ , (24) \int_{\mathcal R}\nabla_\tau\widetilde V\cdot\Upsilon_j\,d\mathcal H^2\ge(c_*-\eta)|U||\mathcal R|, \tag*{(24)} ∫ R ∇ τ V ⋅ Υ j d H 2 ≥ ( c ∗ − η ) ∣ U ∣∣ R ∣ , (24)
∣ ∫ R ∇ τ e ⋅ Υ j d H 2 ∣ ≲ λ ∥ ∇ e ∥ L 2 ( P Y ) , (25) \left|\int_{\mathcal R}\nabla_\tau e\cdot\Upsilon_j\,d\mathcal H^2\right|\lesssim\sqrt{\lambda}\|\nabla e\|_{L^2(P_Y)}, \tag*{(25)} ∫ R ∇ τ e ⋅ Υ j d H 2 ≲ λ ∥∇ e ∥ L 2 ( P Y ) , (25)
and
∥ ∇ e ∥ L 2 ( P Y ) ≲ ∣ U ∣ log ς . (26) \|\nabla e\|_{L^2(P_Y)}\lesssim|U|\sqrt{\log\varsigma}. \tag*{(26)} ∥∇ e ∥ L 2 ( P Y ) ≲ ∣ U ∣ log ς . (26)
Let us check first that these three estimates imply the lemma. Indeed, from them we get
∫ R ∇ τ V U ⋅ Υ j d H 2 ≥ ( c ∗ − η ) ∣ U ∣ ∣ R ∣ − C ∣ U ∣ λ log ς , \int_{\mathcal R}\nabla_\tau V^U\cdot\Upsilon_j\,d\mathcal H^2\ge(c_*-\eta)|U||\mathcal R|-C|U|\sqrt{\lambda\log\varsigma}, ∫ R ∇ τ V U ⋅ Υ j d H 2 ≥ ( c ∗ − η ) ∣ U ∣∣ R ∣ − C ∣ U ∣ λ log ς ,
and C λ log ς ≤ η ∣ R ∣ C\sqrt{\lambda\log\varsigma}\le\eta|\mathcal R| C λ log ς ≤ η ∣ R ∣ if λ \lambda λ is large enough, because ∣ R ∣ ≈ λ |\mathcal R|\approx\lambda ∣ R ∣ ≈ λ and ς ≈ λ \varsigma\approx\sqrt{\lambda} ς ≈ λ , by (16) and (17) .
The main term. Since V ~ = V U \widetilde V=V^U V = V U on R \mathcal R R and ψ j \psi_j ψ j is supported where ∣ s ∣ < 1 − 2 w |s|<1-2w ∣ s ∣ < 1 − 2 w , we have
∫ R ∇ τ V ~ ⋅ Υ j d H 2 = ( λ − 16 ) ∫ ∂ E ∂ s V U ψ j d H 1 . \int_{\mathcal R}\nabla_\tau\widetilde V\cdot\Upsilon_j\,d\mathcal H^2=(\lambda-16)\int_{\partial E}\partial_sV^U\psi_j\,d\mathcal H^1. ∫ R ∇ τ V ⋅ Υ j d H 2 = ( λ − 16 ) ∫ ∂ E ∂ s V U ψ j d H 1 .
By (23) , the right-hand side is at least 8 ( c ∗ − η ) ∣ U ∣ ( λ − 16 ) 8(c_*-\eta)|U|(\lambda-16) 8 ( c ∗ − η ) ∣ U ∣ ( λ − 16 ) , which is not smaller than ( c ∗ − η ) ∣ U ∣ ∣ R ∣ (c_*-\eta)|U||\mathcal R| ( c ∗ − η ) ∣ U ∣∣ R ∣ , because ∣ R ∣ = 8 ( 1 − 2 w ) ( λ − 16 ) |\mathcal R|=8(1-2w)(\lambda-16) ∣ R ∣ = 8 ( 1 − 2 w ) ( λ − 16 ) . Thus (24) is proved.
The error term. Given an integer k k k with 0 ≤ k ≤ λ − 1 0\le k\le\lambda-1 0 ≤ k ≤ λ − 1 , we consider the sets
S k : = ∂ E × ( k , k + 1 ) , O k : = ( ( − 2 , 2 ) 2 ∖ E ) × ( k , k + 1 ) ⊂ P Y . S_k:=\partial E\times(k,k+1),\qquad O_k:=\bigl((-2,2)^2\setminus E\bigr)\times(k,k+1)\subset P_Y. S k := ∂ E × ( k , k + 1 ) , O k := ( ( − 2 , 2 ) 2 ∖ E ) × ( k , k + 1 ) ⊂ P Y .
The set O k − ( 0 , k ) O_k-(0,k) O k − ( 0 , k ) is the same connected Lipschitz domain for every k k k , and S k S_k S k is contained in ∂ O k \partial O_k ∂ O k . So the trace and the Poincaré inequalities in this domain give, with constants independent of k k k ,
∥ φ − m O k φ ∥ L 2 ( S k ) ≲ ∥ ∇ φ ∥ L 2 ( O k ) (27) \|\varphi-m_{O_k}\varphi\|_{L^2(S_k)}\lesssim\|\nabla\varphi\|_{L^2(O_k)} \tag*{(27)} ∥ φ − m O k φ ∥ L 2 ( S k ) ≲ ∥∇ φ ∥ L 2 ( O k ) (27)
for every φ ∈ W 1 , 2 ( O k ) \varphi\in W^{1,2}(O_k) φ ∈ W 1 , 2 ( O k ) .
Suppose now that 8 ≤ k < λ − 8 8\le k<\lambda-8 8 ≤ k < λ − 8 . By Lemmas 4.1 and 3.2 , the functions V U V^U V U and V ~ \widetilde V V , and so also e e e , are smooth up to the open faces of Π \Pi Π . We integrate by parts in s s s on each horizontal section of R ∩ S k \mathcal R\cap S_k R ∩ S k . There are no boundary terms, because ψ j \psi_j ψ j vanishes near the endpoints of the horizontal segments of R \mathcal R R . Moreover, ∫ ∂ E ∂ s ψ j d H 1 = 0 \int_{\partial E}\partial_s\psi_j\,d\mathcal H^1=0 ∫ ∂ E ∂ s ψ j d H 1 = 0 , because ψ j \psi_j ψ j is compactly supported in the open sides of E E E , and so we may subtract the constant m O k e m_{O_k}e m O k e from e e e . Together with (27) , this gives
∣ ∫ R ∩ S k ∇ τ e ⋅ Υ j d H 2 ∣ = ∣ ∫ R ∩ S k ( e − m O k e ) ∂ s ψ j d H 2 ∣ ≲ ∥ ∇ e ∥ L 2 ( O k ) , \left|\int_{\mathcal R\cap S_k}\nabla_\tau e\cdot\Upsilon_j\,d\mathcal H^2\right|=\left|\int_{\mathcal R\cap S_k}(e-m_{O_k}e)\partial_s\psi_j\,d\mathcal H^2\right|\lesssim\|\nabla e\|_{L^2(O_k)}, ∫ R ∩ S k ∇ τ e ⋅ Υ j d H 2 = ∫ R ∩ S k ( e − m O k e ) ∂ s ψ j d H 2 ≲ ∥∇ e ∥ L 2 ( O k ) ,
where the implicit constant depends on the fixed test functions through ∥ ∂ s ψ j ∥ L 2 ( R ∩ S k ) \|\partial_s\psi_j\|_{L^2(\mathcal R\cap S_k)} ∥ ∂ s ψ j ∥ L 2 ( R ∩ S k ) . Up to a set of measure zero, R \mathcal R R is the union of the strips R ∩ S k \mathcal R\cap S_k R ∩ S k with 8 ≤ k < λ − 8 8\le k<\lambda-8 8 ≤ k < λ − 8 , and the sets O k O_k O k are pairwise disjoint. Hence, summing the last estimate over these values of k k k (there are fewer than λ \lambda λ of them) and applying the Cauchy–Schwarz inequality, we get (25) .
The energy of e e e . Let φ ∈ C ∞ ( P ‾ ) \varphi\in C^\infty(\overline P) φ ∈ C ∞ ( P ) be a periodic function which vanishes for t t t large. Then (19) gives
∫ P Y ∇ e ⋅ ∇ φ = − ∫ P Y ∇ V ~ ⋅ ∇ φ , \int_{P_Y}\nabla e\cdot\nabla\varphi=-\int_{P_Y}\nabla\widetilde V\cdot\nabla\varphi, ∫ P Y ∇ e ⋅ ∇ φ = − ∫ P Y ∇ V ⋅ ∇ φ ,
and to estimate the right-hand side we use the form of V ~ \widetilde V V . Note that ∇ V ~ = ( U + χ λ ∇ q , χ λ ′ q ) \nabla\widetilde V=(U+\chi_\lambda\nabla q,\ \chi_\lambda' q) ∇ V = ( U + χ λ ∇ q , χ λ ′ q ) , with χ λ = 0 \chi_\lambda=0 χ λ = 0 for t ≥ λ − 2 t\geq\lambda-2 t ≥ λ − 2 . For t > λ t>\lambda t > λ we have ∫ Y U ⋅ ∇ y φ ( y , t ) d y = 0 \int_Y U\cdot\nabla_y\varphi(y,t)\,dy=0 ∫ Y U ⋅ ∇ y φ ( y , t ) d y = 0 by periodicity, and for t < λ t<\lambda t < λ the divergence theorem in Y ∖ E Y\setminus E Y ∖ E gives
∫ Y ∖ E U ⋅ ∇ y φ ( y , t ) d y = − ∫ ∂ E φ ( ⋅ , t ) U ⋅ ν E d H 1 . \int_{Y\setminus E}U\cdot\nabla_y\varphi(y,t)\,dy=-\int_{\partial E}\varphi(\cdot,t)U\cdot\nu_E\,d\mathcal{H}^{1}. ∫ Y ∖ E U ⋅ ∇ y φ ( y , t ) d y = − ∫ ∂ E φ ( ⋅ , t ) U ⋅ ν E d H 1 .
where ν E \nu_E ν E stands for the outer unit normal of E E E . For a periodic function ξ \xi ξ in Y ∖ E Y\setminus E Y ∖ E , we denote
J ( ξ ) : = ∫ Y ∖ E ( U + ∇ q ) ⋅ ∇ ξ . J(\xi):=\int_{Y\setminus E}(U+\nabla q)\cdot\nabla\xi. J ( ξ ) := ∫ Y ∖ E ( U + ∇ q ) ⋅ ∇ ξ .
Putting together the previous identities, we obtain
∫ P Y ∇ e ⋅ ∇ φ = − ∫ 0 λ χ λ ( t ) J ( φ ( ⋅ , t ) ) d t + ∫ 0 λ ( 1 − χ λ ( t ) ) ∫ ∂ E φ U ⋅ ν E d H 1 d t − ∫ P Y χ λ ′ ( t ) q ∂ t φ . \int_{P_Y}\nabla e\cdot\nabla\varphi=-\int_0^\lambda\chi_\lambda(t)J(\varphi(\cdot,t))\,dt+\int_0^\lambda(1-\chi_\lambda(t))\int_{\partial E}\varphi U\cdot\nu_E\,d\mathcal{H}^{1}\,dt-\int_{P_Y}\chi_\lambda'(t)q\,\partial_t\varphi. ∫ P Y ∇ e ⋅ ∇ φ = − ∫ 0 λ χ λ ( t ) J ( φ ( ⋅ , t )) d t + ∫ 0 λ ( 1 − χ λ ( t )) ∫ ∂ E φ U ⋅ ν E d H 1 d t − ∫ P Y χ λ ′ ( t ) q ∂ t φ .
The term I is the error due to the horizontal cutoff χ ς \chi_\varsigma χ ς . The term II comes from the upper part of the walls, where χ λ < 1 \chi_\lambda<1 χ λ < 1 , and so V ~ \widetilde V V does not satisfy the Neumann condition. The term III comes from the vertical cutoff χ λ \chi_\lambda χ λ . We will show that
∣ I ∣ + ∣ I I ∣ + ∣ I I I ∣ ≲ ∣ U ∣ log ς ∥ ∇ φ ∥ L 2 ( P Y ) . (28) |I|+|II|+|III|\lesssim|U|\sqrt{\log\varsigma}\,\|\nabla\varphi\|_{L^2(P_Y)}. \tag*{(28)} ∣ I ∣ + ∣ I I ∣ + ∣ I I I ∣ ≲ ∣ U ∣ log ς ∥∇ φ ∥ L 2 ( P Y ) . (28)
By density, the resulting bound for ∫ P Y ∇ e ⋅ ∇ φ \int_{P_Y}\nabla e\cdot\nabla\varphi ∫ P Y ∇ e ⋅ ∇ φ extends to all φ ∈ H \varphi\in\mathcal{H} φ ∈ H , and choosing φ = e \varphi=e φ = e we get (26) .
Let us start with I. We claim that
∣ J ( ξ ) ∣ ≲ ∣ U ∣ ς − 1 ∥ ∇ ξ ∥ L 2 ( Y ∖ E ) (29) |J(\xi)|\lesssim|U|\varsigma^{-1}\|\nabla\xi\|_{L^2(Y\setminus E)} \tag*{(29)} ∣ J ( ξ ) ∣ ≲ ∣ U ∣ ς − 1 ∥∇ ξ ∥ L 2 ( Y ∖ E ) (29)
for every periodic ξ ∈ W 1 , 2 ( Y ∖ E ) \xi\in W^{1,2}(Y\setminus E) ξ ∈ W 1 , 2 ( Y ∖ E ) .
Taking this for granted for the moment, and using that λ ≈ ς \sqrt{\lambda}\approx\varsigma λ ≈ ς , by the Cauchy–Schwarz inequality in t t t we get
∣ I ∣ ≤ ∫ 0 λ ∣ J ( φ ( ⋅ , t ) ) ∣ d t ≲ ∣ U ∣ ς − 1 λ ∥ ∇ φ ∥ L 2 ( P Y ) ≲ ∣ U ∣ ∥ ∇ φ ∥ L 2 ( P Y ) . |I|\leq\int_0^\lambda|J(\varphi(\cdot,t))|\,dt\lesssim|U|\varsigma^{-1}\sqrt{\lambda}\,\|\nabla\varphi\|_{L^2(P_Y)}\lesssim|U|\|\nabla\varphi\|_{L^2(P_Y)}. ∣ I ∣ ≤ ∫ 0 λ ∣ J ( φ ( ⋅ , t )) ∣ d t ≲ ∣ U ∣ ς − 1 λ ∥∇ φ ∥ L 2 ( P Y ) ≲ ∣ U ∣∥∇ φ ∥ L 2 ( P Y ) .
To show (29) , we will use the equation satisfied by V E U V_E^U V E U and the fact that ∇ V E U = U + ∇ d U \nabla V_E^U=U+\nabla d^U ∇ V E U = U + ∇ d U . Indeed, the function χ ς ξ \chi_\varsigma\xi χ ς ξ , extended by zero, belongs to W 1 , 2 ( R 2 ∖ E ) W^{1,2}(\mathbb{R}^2\setminus E) W 1 , 2 ( R 2 ∖ E ) and has bounded support, and so it is a test function for the exterior problem in Lemma 3.2. Then we get
∫ Y ∖ E χ ς ∇ d U ⋅ ∇ ξ = ∫ Y ∖ E ∇ d U ⋅ ∇ ( χ ς ξ ) − ∫ Y ∖ E ξ ∇ d U ⋅ ∇ χ ς = − ∫ Y ∖ E U ⋅ ∇ ( χ ς ξ ) − ∫ Y ∖ E ξ ∇ d U ⋅ ∇ χ ς . \int_{Y\setminus E}\chi_\varsigma\nabla d^U\cdot\nabla\xi=\int_{Y\setminus E}\nabla d^U\cdot\nabla(\chi_\varsigma\xi)-\int_{Y\setminus E}\xi\nabla d^U\cdot\nabla\chi_\varsigma=-\int_{Y\setminus E}U\cdot\nabla(\chi_\varsigma\xi)-\int_{Y\setminus E}\xi\nabla d^U\cdot\nabla\chi_\varsigma. ∫ Y ∖ E χ ς ∇ d U ⋅ ∇ ξ = ∫ Y ∖ E ∇ d U ⋅ ∇ ( χ ς ξ ) − ∫ Y ∖ E ξ ∇ d U ⋅ ∇ χ ς = − ∫ Y ∖ E U ⋅ ∇ ( χ ς ξ ) − ∫ Y ∖ E ξ ∇ d U ⋅ ∇ χ ς .
Since ∇ q = χ ς ∇ d U + d U ∇ χ ς \nabla q=\chi_\varsigma\nabla d^U+d^U\nabla\chi_\varsigma ∇ q = χ ς ∇ d U + d U ∇ χ ς , we can write
J ( ξ ) = ∫ Y ∖ E ( U + χ ς ∇ d U + d U ∇ χ ς ) ⋅ ∇ ξ = ∫ Y ∖ E U ⋅ ∇ ( ( 1 − χ ς ) ξ ) − ∫ A r ξ + ∫ A d U ∇ χ ς ⋅ ∇ ξ , J(\xi)=\int_{Y\setminus E}(U+\chi_\varsigma\nabla d^U+d^U\nabla\chi_\varsigma)\cdot\nabla\xi=\int_{Y\setminus E}U\cdot\nabla((1-\chi_\varsigma)\xi)-\int_A r\xi+\int_A d^U\nabla\chi_\varsigma\cdot\nabla\xi, J ( ξ ) = ∫ Y ∖ E ( U + χ ς ∇ d U + d U ∇ χ ς ) ⋅ ∇ ξ = ∫ Y ∖ E U ⋅ ∇ (( 1 − χ ς ) ξ ) − ∫ A r ξ + ∫ A d U ∇ χ ς ⋅ ∇ ξ ,
where r : = ∇ χ ς ⋅ ∇ d U r:=\nabla\chi_\varsigma\cdot\nabla d^U r := ∇ χ ς ⋅ ∇ d U and A : = B ( 0 , ς / 4 ) ∖ B ( 0 , ς / 8 ) A:=B(0,\varsigma/4)\setminus B(0,\varsigma/8) A := B ( 0 , ς /4 ) ∖ B ( 0 , ς /8 ) . Since ( 1 − χ ς ) ξ (1-\chi_\varsigma)\xi ( 1 − χ ς ) ξ is periodic and vanishes near E E E , the first integral on the right hand side is zero. We also claim that r r r has zero mean. To see this, we test the exterior problem with χ ς \chi_\varsigma χ ς , and we get
∫ A r = − ∫ R 2 ∖ E U ⋅ ∇ χ ς = ∫ ∂ E U ⋅ ν E d H 1 = 0. \int_A r=-\int_{\mathbb{R}^2\setminus E}U\cdot\nabla\chi_\varsigma=\int_{\partial E}U\cdot\nu_E\,d\mathcal{H}^{1}=0. ∫ A r = − ∫ R 2 ∖ E U ⋅ ∇ χ ς = ∫ ∂ E U ⋅ ν E d H 1 = 0.
On the other hand, (12) gives ∥ r ∥ L 2 ( A ) ≲ ∣ U ∣ ς − 2 \|r\|_{L^2(A)}\lesssim|U|\varsigma^{-2} ∥ r ∥ L 2 ( A ) ≲ ∣ U ∣ ς − 2 and ∥ d U ∇ χ ς ∥ L 2 ( A ) ≲ ∣ U ∣ ς − 1 \|d^U\nabla\chi_\varsigma\|_{L^2(A)}\lesssim|U|\varsigma^{-1} ∥ d U ∇ χ ς ∥ L 2 ( A ) ≲ ∣ U ∣ ς − 1 . In the second integral we may replace ξ \xi ξ by ξ \xi ξ minus its mean on A A A , and then (29) follows from the Poincaré inequality in A A A , a set of diameter comparable to ς \varsigma ς .
We turn now to II. Its integrand is zero unless λ − 8 < t < λ \lambda-8<t<\lambda λ − 8 < t < λ , so that only the eight strips S k S_k S k with λ − 8 ≤ k ≤ λ − 1 \lambda-8\le k\le\lambda-1 λ − 8 ≤ k ≤ λ − 1 contribute. Since ∫ ∂ E U ⋅ ν E d H 1 = 0 \int_{\partial E}U\cdot\nu_E\,d\mathcal{H}^1=0 ∫ ∂ E U ⋅ ν E d H 1 = 0 , in each of these strips we may replace φ \varphi φ by φ − m O k φ \varphi-m_{O_k}\varphi φ − m O k φ , and then (27) gives ∣ I I ∣ ≲ ∣ U ∣ ∥ ∇ φ ∥ L 2 ( P Y ) |\mathrm{II}|\lesssim|U|\|\nabla\varphi\|_{L^2(P_Y)} ∣ II ∣ ≲ ∣ U ∣∥∇ φ ∥ L 2 ( P Y ) .
Finally, we deal with III. By (12) , ∣ q ( y ) ∣ ≲ ∣ U ∣ ∣ y ∣ − 1 |q(y)|\lesssim|U||y|^{-1} ∣ q ( y ) ∣ ≲ ∣ U ∣∣ y ∣ − 1 in Y ∖ E Y\setminus E Y ∖ E , and q q q vanishes outside B ( 0 , ς / 4 ) B(0,\varsigma/4) B ( 0 , ς /4 ) . Since ∣ y ∣ ≥ 1 |y|\ge1 ∣ y ∣ ≥ 1 in Y ∖ E Y\setminus E Y ∖ E , this gives
∥ q ∥ L 2 ( Y ∖ E ) 2 ≲ ∣ U ∣ 2 ∫ 1 ς d r r = ∣ U ∣ 2 log ς . \|q\|_{L^2(Y\setminus E)}^2\lesssim|U|^2\int_1^\varsigma\frac{dr}{r}=|U|^2\log\varsigma. ∥ q ∥ L 2 ( Y ∖ E ) 2 ≲ ∣ U ∣ 2 ∫ 1 ς r d r = ∣ U ∣ 2 log ς .
As ∣ χ λ ′ ∣ ≤ 1 |\chi'_\lambda|\le1 ∣ χ λ ′ ∣ ≤ 1 and χ λ ′ \chi'_\lambda χ λ ′ is supported in [ λ − 8 , λ − 2 ] [\lambda-8,\lambda-2] [ λ − 8 , λ − 2 ] , the Cauchy–Schwarz inequality yields ∣ I I I ∣ ≲ ∣ U ∣ log ς ∥ ∇ φ ∥ L 2 ( P Y ) |\mathrm{III}|\lesssim|U|\sqrt{\log\varsigma}\|\nabla\varphi\|_{L^2(P_Y)} ∣ III ∣ ≲ ∣ U ∣ log ς ∥∇ φ ∥ L 2 ( P Y ) . This proves (28) , and completes the proof of the lemma. □
From now on we fix λ \lambda λ so that Lemma 4.2 holds. This also fixes ς \varsigma ς .
Remark 4.3. The proof explains why we need tall columns. When we replace V U V^U V U by the planar solution, we make an error which is concentrated near the ends of the column and in the far field, and whose energy is only O ( ∣ U ∣ log ς ) O(|U|\sqrt{\log\varsigma}) O ( ∣ U ∣ log ς ) , which grows very slowly with the height of the column. By a trace inequality on unit strips, its effect on the walls is at most O ( λ log ς ) O(\sqrt{\lambda}\log\varsigma) O ( λ log ς ) , while the main term grows like λ \lambda λ . Note also that the period must grow like λ \sqrt{\lambda} λ , so that the selected area matches A A A times the area of the floor.
The selected squares
Let ℓ > 0 \ell>0 ℓ > 0 be small. On each of the four rectangles of R \mathcal{R} R we take the closed squares of side ℓ \ell ℓ of the standard grid of mesh ℓ \ell ℓ (in the coordinates ( s , t ) (s,t) ( s , t ) ) which are contained in the rectangle, and we shrink each of them to the concentric closed square of side
a : = ( 1 − w ) ℓ . a:=(1-w)\ell. a := ( 1 − w ) ℓ .
We denote these smaller squares by P 1 , … , P N P_1,\ldots,P_N P 1 , … , P N , their centres by x 1 , … , x N x_1,\ldots,x_N x 1 , … , x N , and we set M : = ⋃ i P i \mathcal{M}:=\bigcup_i P_i M := ⋃ i P i . They are pairwise disjoint, separated from each other and from the edges of Π \Pi Π , and, by (16) ,
∣ M ∣ : = H 2 ( M ) ≤ ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣ . (30) |\mathcal{M}|:=\mathcal{H}^2(\mathcal{M})\le(1-w)^2|\mathcal{R}|=A|Y|. \tag*{(30)} ∣ M ∣ := H 2 ( M ) ≤ ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣. (30)
Lemma 4.4. Let 1 < γ < γ ∗ 1<\gamma<\gamma_\ast 1 < γ < γ ∗ . If ℓ \ell ℓ is small enough, then there are finitely many smooth tangent vector fields Υ ~ 1 , … , Υ ~ m \widetilde{\Upsilon}_1,\ldots,\widetilde{\Upsilon}_m Υ 1 , … , Υ m on ∂ Π \partial\Pi ∂ Π , compactly supported in the relative interior of M \mathcal{M} M , with ∣ Υ ~ j ∣ ≤ 1 |\widetilde{\Upsilon}_j|\le1 ∣ Υ j ∣ ≤ 1 , such that for every U ∈ R 2 U\in\mathbb{R}^2 U ∈ R 2
max 1 ≤ j ≤ m ∫ M ∇ τ V U ⋅ Υ ~ j d H 2 ≥ γ ∣ U ∣ ∣ Y ∣ . (31) \max_{1\le j\le m}\int_{\mathcal{M}}\nabla_\tau V^U\cdot\widetilde{\Upsilon}_j\,d\mathcal{H}^2\ge\gamma|U||Y|. \tag*{(31)} 1 ≤ j ≤ m max ∫ M ∇ τ V U ⋅ Υ j d H 2 ≥ γ ∣ U ∣∣ Y ∣. (31) In particular, ∫ M ∣ ∇ τ V U ∣ d H 2 ≥ γ ∣ U ∣ ∣ Y ∣ \int_{\mathcal{M}}|\nabla_\tau V^U|\,d\mathcal{H}^2\ge\gamma|U||Y| ∫ M ∣ ∇ τ V U ∣ d H 2 ≥ γ ∣ U ∣∣ Y ∣ .
Proof. We fix a function ζ ∈ C c ∞ ( ( − 1 / 2 , 1 / 2 ) 2 ) \zeta\in C^\infty_c((-1/2,1/2)^2) ζ ∈ C c ∞ (( − 1/2 , 1/2 ) 2 ) such that 0 ≤ ζ ≤ 1 0\le\zeta\le1 0 ≤ ζ ≤ 1 and θ : = ∫ ζ > γ / γ ∗ \theta:=\int\zeta>\gamma/\gamma_\ast θ := ∫ ζ > γ / γ ∗ . In the coordinates ( s , t ) (s,t) ( s , t ) , we let ζ ℓ ( x ) = ζ ( ( x − x i ) / a ) \zeta_\ell(x)=\zeta((x-x_i)/a) ζ ℓ ( x ) = ζ (( x − x i ) / a ) on P i P_i P i and ζ ℓ = 0 \zeta_\ell=0 ζ ℓ = 0 elsewhere, and we set Υ ~ j : = ζ ℓ Υ j \widetilde{\Upsilon}_j:=\zeta_\ell\Upsilon_j Υ j := ζ ℓ Υ j . For every continuous function h h h on R \mathcal{R} R , the integral ∫ M ζ ℓ h \int_{\mathcal{M}}\zeta_\ell h ∫ M ζ ℓ h is a Riemann sum, and by the uniform continuity of h h h ,
∫ M ζ ℓ h d H 2 = ∑ i a 2 ∫ ζ ( z ) h ( x i + a z ) d z → ℓ → 0 θ ( 1 − w ) 2 ∫ R h d H 2 , \int_{\mathcal{M}}\zeta_\ell h\,d\mathcal{H}^2
=\sum_i a^2\int\zeta(z)h(x_i+az)\,dz
\xrightarrow{\ell\to0}\theta(1-w)^2\int_{\mathcal{R}}h\,d\mathcal{H}^2, ∫ M ζ ℓ h d H 2 = i ∑ a 2 ∫ ζ ( z ) h ( x i + a z ) d z ℓ → 0 θ ( 1 − w ) 2 ∫ R h d H 2 ,
since the grid squares of side ℓ \ell ℓ cover R \mathcal{R} R up to a set of area O ( ℓ ) O(\ell) O ( ℓ ) . We apply this to the finitely many functions h = ∇ τ V e k ⋅ Υ j h=\nabla_\tau V^{e_k}\cdot\Upsilon_j h = ∇ τ V e k ⋅ Υ j , k = 1 , 2 k=1,2 k = 1 , 2 , j = 1 , … , m j=1,\ldots,m j = 1 , … , m . By linearity in U U U , for ℓ \ell ℓ small enough we obtain, for all U U U and j j j ,
∫ M ∇ τ V U ⋅ Υ ~ j ≥ θ ( 1 − w ) 2 ∫ R ∇ τ V U ⋅ Υ j − ( θ γ ∗ − γ ) ∣ U ∣ ∣ Y ∣ . (32) \int_{\mathcal{M}}\nabla_\tau V^U\cdot\widetilde{\Upsilon}_j
\ge\theta(1-w)^2\int_{\mathcal{R}}\nabla_\tau V^U\cdot\Upsilon_j-(\theta\gamma_\ast-\gamma)|U||Y|. \tag*{(32)} ∫ M ∇ τ V U ⋅ Υ j ≥ θ ( 1 − w ) 2 ∫ R ∇ τ V U ⋅ Υ j − ( θ γ ∗ − γ ) ∣ U ∣∣ Y ∣. (32)
Choosing j j j as in Lemma 4.2 and using ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣ (1-w)^2|\mathcal{R}|=A|Y| ( 1 − w ) 2 ∣ R ∣ = A ∣ Y ∣ , (32) concludes the proof. □
We now fix γ : = ( 1 + γ ∗ ) / 2 \gamma:= (1+\gamma_*)/2 γ := ( 1 + γ ∗ ) /2 , so that in particular γ ∈ ( 1 , γ ∗ ) \gamma\in(1,\gamma_*) γ ∈ ( 1 , γ ∗ ) . The side ℓ \ell ℓ will be fixed in Lemma 5.1 , where we may need to decrease it; this is allowed, because Lemma 4.4 holds for every ℓ \ell ℓ small enough. We also set
C Υ : = max j ∥ div τ T ~ j ∥ L ∞ < ∞ . (33) C_{\Upsilon}:=\max_j\left\|\operatorname{div}_{\tau}\widetilde{T}_j\right\|_{L^\infty}<\infty. \tag*{(33)} C Υ := j max div τ T j L ∞ < ∞. (33)
This constant depends on ℓ \ell ℓ (it is of order ℓ − 1 \ell^{-1} ℓ − 1 ), and it is fixed together with ℓ \ell ℓ in Lemma 5.1 .
Remark 4.5. In each period cell, a square of the floor of area ∣ Y ∣ |Y| ∣ Y ∣ , on which the linear function U ⋅ y U\cdot y U ⋅ y has tangential gradient U U U , is replaced by the selected squares, of total area at most A ∣ Y ∣ A|Y| A ∣ Y ∣ , on which the integral of ∣ ∇ τ V U ∣ |\nabla_{\tau}V^{U}| ∣ ∇ τ V U ∣ is at least γ ∣ U ∣ ∣ Y ∣ \gamma|U||Y| γ ∣ U ∣∣ Y ∣ . Thus the integral of the modulus of the tangential gradient grows at least by the factor γ > 1 \gamma>1 γ > 1 , while the area is multiplied by a factor at most A < 1 A<1 A < 1 .
The column map and the rigidity of the construction
The domains of the construction are obtained by successive insertions of columns, and we need them to be bi-Lipschitz images of the cube, with a constant independent of the number of insertions. This is not automatic, for two reasons:
If we compose k k k maps with bi-Lipschitz constant L L L , we only obtain the bound L k L^k L k .
Choosing smaller and smaller insertion scales does not help, because all the conjugates of a fixed map by similarities have the same bi-Lipschitz constant.
The way out is the following simple observation: let Ψ \Psi Ψ and G G G be bi-Lipschitz maps with constants at most L L L , and suppose that G G G coincides with an isometry on the preimage of the region where Ψ \Psi Ψ differs from the identity. Then, Ψ ∘ G \Psi\circ G Ψ ∘ G is also bi-Lipschitz with constant at most L L L , because around every point it coincides either with G G G or with Ψ \Psi Ψ composed with an isometry. Hence, our strategy is the following:
We build a single map F F F which creates one column, and which is an isometry near the preimage of every selected square.
The next generation of columns is inserted on the selected squares, and so it is always inserted where the accumulated map is an isometry.
However, this rigidity has a cost. Indeed, the map F F F sends the boundary of the half-space onto the boundary of the domain with one column, and it maps some disjoint squares D i D_i D i of the floor rigidly onto the selected squares P i P_i P i on the walls. Hence, in each period cell, the total area of the selected squares is at most the area of the floor. So we cannot hope to gain anything from the area, which is why we need A < 1 A<1 A < 1 in § 4 : the gain has to come from the flow, that is, it crucially relies on the fact that c ∗ > 1 c_*>1 c ∗ > 1 .
The column map
The next lemma provides the map F F F which creates one column. Its proof, which contains an explicit construction of F F F , is given in Appendix A, see also Figure 5 .
Figure 5. The column map F = F 2 ∘ F 1 ∘ Θ F=F_2\circ F_1\circ\Theta F = F 2 ∘ F 1 ∘ Θ (not to scale). (a) The squares D i D_i D i in the floor Y ′ Y' Y ′ . (b) The map Θ \Theta Θ moves them onto thin rectangles T i T_i T i in the square annulus { 1 < ∣ y ∣ ∞ < 1 + ε 0 } \{1<|y|_\infty<1+\varepsilon_0\} { 1 < ∣ y ∣ ∞ < 1 + ε 0 } . (c) The map F 1 F_1 F 1 creates a truncated pyramid of height λ \lambda λ , whose lateral faces contain the images of the T i T_i T i . (d) The map F 2 F_2 F 2 makes the lateral faces vertical. Near each T i T_i T i , the map F 2 ∘ F 1 F_2\circ F_1 F 2 ∘ F 1 is affine and undoes the distortion of Θ \Theta Θ , so that F F F maps D i D_i D i rigidly onto P i P_i P i .
Lemma 5.1. After decreasing ℓ \ell ℓ if necessary, there exist an open square Y ′ ⊆ Y Y'\subseteq Y Y ′ ⊆ Y centred at the origin, pairwise disjoint closed squares D 1 , … , D N ⊂ Y ′ × { 0 } D_1,\ldots,D_N\subset Y'\times\{0\} D 1 , … , D N ⊂ Y ′ × { 0 } of side a a a and with sides parallel to the axes, pairwise disjoint connected open sets W i ⊃ D i W_i\supset D_i W i ⊃ D i , and a bi-Lipschitz homeomorphism F F F of R 3 \mathbb{R}^3 R 3 such that
F = id outside Z : = Y ′ × ( − 5 λ , 6 λ ) , (34) F=\operatorname{id}\quad\text{outside }Z:=Y'\times(-5\lambda,6\lambda), \tag*{(34)} F = id outside Z := Y ′ × ( − 5 λ , 6 λ ) , (34) F ( R + 3 ) = R + 3 ∖ Π , (35) F(\mathbb{R}^3_+)=\mathbb{R}^3_+\setminus\Pi, \tag*{(35)} F ( R + 3 ) = R + 3 ∖ Π , (35) F = R i in W i , where R i is a rigid motion with R i ( D i ) = P i . (36) F=R_i\text{ in }W_i,\quad\text{where }R_i\text{ is a rigid motion with }R_i(D_i)=P_i. \tag*{(36)} F = R i in W i , where R i is a rigid motion with R i ( D i ) = P i . (36) The linear part of every R i R_i R i is a signed permutation matrix. We fix ℓ \ell ℓ and F F F from now on, and we denote by L F ≥ 1 L_F\geq1 L F ≥ 1 a common bound for Lip ( F ) \operatorname{Lip}(F) Lip ( F ) and Lip ( F − 1 ) \operatorname{Lip}(F^{-1}) Lip ( F − 1 ) .
Inserting columns on a flat square
Given a domain D ⊂ R 3 D \subset\mathbb{R}^3 D ⊂ R 3 , we say that a closed square P ⊂ ∂ D P \subset\partial D P ⊂ ∂ D is flat if there are an open set V P ⊃ P V_P \supset P V P ⊃ P and a rigid motion R P R_P R P whose linear part is a signed permutation matrix, such that
D ∩ V P = R P ( R + 3 ) ∩ V P , R P − 1 ( P ) ⊂ R 2 × { 0 } is a square with sides parallel to the axes . (37) D \cap V_P = R_P(\mathbb{R}^3_+) \cap V_P,\qquad R_P^{-1}(P) \subset\mathbb{R}^2 \times\{0\}\text{ is a square with sides parallel to the axes}. \tag*{(37)} D ∩ V P = R P ( R + 3 ) ∩ V P , R P − 1 ( P ) ⊂ R 2 × { 0 } is a square with sides parallel to the axes . (37)
In this case we call V P V_P V P a flat neighbourhood of P P P .
Next we fix a small ρ > 0 \rho> 0 ρ > 0 . Identifying R P − 1 ( P ) R_P^{-1}(P) R P − 1 ( P ) with a subset of R 2 \mathbb{R}^2 R 2 , we denote by I ρ ( P ) I_\rho(P) I ρ ( P ) the set of points b ∈ ρ Z 2 b \in\rho\mathbb{Z}^2 b ∈ ρ Z 2 such that b + ρ Y ′ ⊂ R P − 1 ( P ) b+\rho Y' \subset R_P^{-1}(P) b + ρ Y ′ ⊂ R P − 1 ( P ) . For every b ∈ I ρ ( P ) b \in I_\rho(P) b ∈ I ρ ( P ) we consider the similarity τ b ( y , t ) : = ( b + ρ y , ρ t ) \tau_b(y,t):=(b+\rho y,\rho t) τ b ( y , t ) := ( b + ρ y , ρt ) and the map
Ψ b : = R P ∘ τ b ∘ F ∘ τ b − 1 ∘ R P − 1 , \Psi_b := R_P \circ\tau_b \circ F \circ\tau_b^{-1} \circ R_P^{-1}, Ψ b := R P ∘ τ b ∘ F ∘ τ b − 1 ∘ R P − 1 ,
which is the identity outside the box R P τ b ( Z ) R_P\tau_b(Z) R P τ b ( Z ) . For ρ \rho ρ small enough, all these boxes are contained in V P V_P V P , and since Y ′ ⋐ Y Y' \Subset Y Y ′ ⋐ Y their closures are pairwise disjoint.
Definition 5.3. We define the insertion map Ψ \Psi Ψ as Ψ b \Psi_b Ψ b in R P τ b ( Z ) R_P\tau_b(Z) R P τ b ( Z ) for every b ∈ I ρ ( P ) b \in I_\rho(P) b ∈ I ρ ( P ) , and as the identity elsewhere. We say that Ψ ( D ) \Psi(D) Ψ ( D ) is obtained from D D D by inserting columns of scale ρ \rho ρ on P P P .
We use the same definition for finitely many flat squares with pairwise disjoint flat neighbourhoods. The new selected squares are R P τ b ( P i ) R_P\tau_b(P_i) R P τ b ( P i ) , with b ∈ I ρ ( P ) b \in I_\rho(P) b ∈ I ρ ( P ) and i = 1 , … , N i=1,\ldots,N i = 1 , … , N .
Lemma 5.4. The insertion map Ψ \Psi Ψ is a bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 with Lip ( Ψ ) ≤ L F \operatorname{Lip}(\Psi) \le L_F Lip ( Ψ ) ≤ L F and Lip ( Ψ − 1 ) ≤ L F \operatorname{Lip}(\Psi^{-1}) \le L_F Lip ( Ψ − 1 ) ≤ L F , and
Ψ ( D ) = D ∖ ⋃ b ∈ I ρ ( P ) R P τ b ( Π ) . \Psi(D)=D\setminus\bigcup_{b\in I_\rho(P)}R_P\tau_b(\Pi). Ψ ( D ) = D ∖ b ∈ I ρ ( P ) ⋃ R P τ b ( Π ) . Moreover, every new selected square P ′ = R P τ b ( P i ) P'=R_P\tau_b(P_i) P ′ = R P τ b ( P i ) is flat in Ψ ( D ) \Psi(D) Ψ ( D ) , with flat neighbourhood V P ′ = Ψ ( R P τ b ( W i ) ) V_{P'}=\Psi(R_P\tau_b(W_i)) V P ′ = Ψ ( R P τ b ( W i )) , and in R P τ b ( W i ) R_P\tau_b(W_i) R P τ b ( W i ) the map Ψ \Psi Ψ coincides with a rigid motion whose linear part is a signed permutation matrix. The neighbourhoods V P ′ V_{P'} V P ′ of the new selected squares are connected and pairwise disjoint.
The proof of Lemma 5.4 is given in Appendix A.
The rigidity induction
We now iterate the insertions. We start from a domain Ω 0 = G 0 ( Q 0 ) \Omega_0=G_0(Q_0) Ω 0 = G 0 ( Q 0 ) and a finite family C 0 \mathcal{C}_0 C 0 of flat closed squares in ∂ Ω 0 \partial\Omega_0 ∂ Ω 0 , with the following properties:
Lip ( G 0 ) ≤ L F \operatorname{Lip}(G_0)\le L_F Lip ( G 0 ) ≤ L F and Lip ( G 0 − 1 ) ≤ L F \operatorname{Lip}(G_0^{-1})\le L_F Lip ( G 0 − 1 ) ≤ L F ;
the squares P ∈ C 0 P\in\mathcal{C}_0 P ∈ C 0 have pairwise disjoint connected flat neighbourhoods V P V_P V P ;
G 0 G_0 G 0 coincides with a rigid motion in G 0 − 1 ( V P ) G_0^{-1}(V_P) G 0 − 1 ( V P ) for every P ∈ C 0 P\in\mathcal{C}_0 P ∈ C 0 .
The basic example is G 0 = id G_0=\operatorname{id} G 0 = id , Ω 0 = Q 0 = ( − 1 , 1 ) 3 \Omega_0=Q_0=(-1,1)^3 Ω 0 = Q 0 = ( − 1 , 1 ) 3 , and C 0 = { C 0 } \mathcal{C}_0=\{C_0\} C 0 = { C 0 } , where C 0 = [ − 1 2 , 1 2 ] 2 × { − 1 } C_0=[-\frac12,\frac12]^2\times\{-1\} C 0 = [ − 2 1 , 2 1 ] 2 × { − 1 } is a square in the bottom face of Q 0 Q_0 Q 0 , with flat neighbourhood ( − 3 4 , 3 4 ) 2 × ( − 2 , 0 ) \left(-\frac34,\frac34\right)^2\times(-2,0) ( − 4 3 , 4 3 ) 2 × ( − 2 , 0 ) .
Suppose now that Ω k = G k ( Q 0 ) \Omega_k=G_k(Q_0) Ω k = G k ( Q 0 ) and a finite family C k \mathcal{C}_k C k of flat selected squares, with pairwise disjoint flat neighbourhoods, have been constructed. We insert columns of a small scale ρ k \rho_k ρ k on all the squares of C k \mathcal{C}_k C k , by an insertion map Ψ k \Psi_k Ψ k , and we set
Ω k + 1 : = Ψ k ( Ω k ) , G k + 1 : = Ψ k ∘ G k . (38) \Omega_{k+1}:=\Psi_k(\Omega_k),\qquad G_{k+1}:=\Psi_k\circ G_k. \tag*{(38)} Ω k + 1 := Ψ k ( Ω k ) , G k + 1 := Ψ k ∘ G k . (38)
We let C k + 1 \mathcal{C}_{k+1} C k + 1 be the family of the new selected squares, and C k : = ⋃ P ∈ C k P \mathcal{C}_k:=\bigcup_{P\in\mathcal{C}_k}P C k := ⋃ P ∈ C k P . The scales ρ k \rho_k ρ k will be chosen in § 8 , and the next lemma holds for any choice of sufficiently small scales. Notice that the selected squares of each column lie on its four walls, and that the next generation of columns is inserted on all of them, perpendicularly to the walls. Thus the direction of the columns changes from one generation to the next, see Figure 1 . The induction is illustrated in Figure 3 .
Figure 3. The rigidity induction, in a vertical section. (a) A selected square P ∈ C k P \in C_k P ∈ C k (blue) and its flat neighbourhood V P V_P V P (dashed), where G k G_k G k is rigid; the map Ψ k \Psi_k Ψ k differs from the identity only in the dotted boxes. (b) Each new selected square P ′ P' P ′ has a flat neighbourhood V P ′ V_{P'} V P ′ inside one of the boxes. (c) Near P ′ P' P ′ we have, up to a rotation, the same picture as in (a), with G k + 1 G_{k+1} G k + 1 rigid in V P ′ V_{P'} V P ′ .
Lemma 5.5. For every k ≥ 0 k \ge0 k ≥ 0 , we have:
(a) Lip ( G k ) ≤ L F \operatorname{Lip}(G_k) \le L_F Lip ( G k ) ≤ L F and Lip ( G k − 1 ) ≤ L F \operatorname{Lip}(G_k^{-1}) \le L_F Lip ( G k − 1 ) ≤ L F .
(b) For every P ∈ C k P \in C_k P ∈ C k , the flat neighbourhood V P V_P V P is connected, and G k G_k G k coincides with a rigid motion in G k − 1 ( V P ) G_k^{-1}(V_P) G k − 1 ( V P ) .
(c) Ω k + 1 ⊂ Ω k \Omega_{k+1} \subset\Omega_k Ω k + 1 ⊂ Ω k , and H 2 ( C k ) ≤ A k H 2 ( C 0 ) \mathcal{H}^2(C_k) \le A^k \mathcal{H}^2(C_0) H 2 ( C k ) ≤ A k H 2 ( C 0 ) .
The proof of Lemma 5.5 is given in Appendix A.
By Lemmas 5.5 and 2.1 , the trace and Poincaré inequality (2.2) and the energy bound for the solutions hold in all the domains Ω k \Omega_k Ω k with constants independent of k k k and of the scales ρ k \rho_k ρ k . Moreover, in Appendix B we prove that every Ω k \Omega_k Ω k is a Lipschitz domain whose boundary consists of finitely many rectangles parallel to the coordinate planes.
Inserting columns at a small scale
In this section we prove the lower bound in (1.6). Roughly speaking, we show that if we insert columns of scale ρ \rho ρ on a flat square where the datum vanishes, then, in the limit as ρ → 0 \rho\to0 ρ → 0 , the integral of the tangential gradient of the new solution over the new selected squares is at least γ \gamma γ times the integral of the tangential gradient of the old solution over the square. The precise statement is Lemma 6.1 below.
To describe the new solution, we use a “two-scale expansion”, as in the theory of homogenization [5 ] . That is, up to a small error, we write the new solution as the sum of two terms:
The old solution.
A correction term of size ρ \rho ρ which, near each column, is a rescaled copy of the cell corrector P U P^U P U of Lemma 4.1 , with U U U equal to the local value of the old tangential gradient.
We note that the main difficulty is to show that the error is small enough on the boundary, at the scale of the columns. This discussion is expanded after the statement of Lemma 6.1.
The setting and the insertion lemma
Throughout this section we fix the following objects:
A domain D = G ( Q 0 ) ⊂ R 3 D = G(Q_0) \subset\mathbb{R}^3 D = G ( Q 0 ) ⊂ R 3 , where G G G is an L F L_F L F -bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 .
A datum g g g on ∂ D \partial D ∂ D as in Lemma 2.1 , and the corresponding variational solution u u u .
A flat closed square Q ⊂ ∂ D Q \subset\partial D Q ⊂ ∂ D , in the sense of (5.4), with a connected flat neighbourhood V Q V_Q V Q , such that G G G is a rigid motion in G − 1 ( V Q ) G^{-1}(V_Q) G − 1 ( V Q ) .
After a rigid motion whose linear part is a signed permutation matrix, we may assume that R Q = id R_Q = \operatorname{id} R Q = id , so that Q ⊂ R 2 × { 0 } Q \subset\mathbb{R}^2 \times\{0\} Q ⊂ R 2 × { 0 } . Then there is h 0 ∈ ( 0 , 1 ) h_0 \in(0,1) h 0 ∈ ( 0 , 1 ) such that, with Q ~ : = { y : dist ( y , Q ) < h 0 } \widetilde{Q} := \{y : \operatorname{dist}(y,Q) < h_0\} Q := { y : dist ( y , Q ) < h 0 } , we have
D ∩ ( Q ~ × ( − h 0 , h 0 ) ) = Q ~ × ( 0 , h 0 ) . D \cap\left(\widetilde{Q} \times(-h_0,h_0)\right) = \widetilde{Q} \times(0,h_0). D ∩ ( Q × ( − h 0 , h 0 ) ) = Q × ( 0 , h 0 ) .
We also assume that the datum vanishes near Q Q Q , that is, g = 0 g = 0 g = 0 on Q ~ × { 0 } \widetilde{Q} \times\{0\} Q × { 0 } . Then, by Lemma 2.2 , u u u is smooth up to Q ~ × { 0 } \widetilde{Q} \times\{0\} Q × { 0 } , with ∂ t u = 0 \partial_t u = 0 ∂ t u = 0 there. We set U ( y ) : = ∇ y u ( y , 0 ) U(y) := \nabla_y u(y,0) U ( y ) := ∇ y u ( y , 0 ) , so that ∣ ∇ τ u ∣ = ∣ U ∣ |\nabla_\tau u| = |U| ∣ ∇ τ u ∣ = ∣ U ∣ on Q Q Q . Taking h 0 h_0 h 0 smaller if necessary, we may also assume that
∣ ∇ u ( y , t ) − ( U ( y ) , 0 ) ∣ ≤ C u t , ∣ U ( y ) ∣ + ∣ ∇ U ( y ) ∣ ≤ C u , for y ∈ Q , 0 < t < h 0 , (39) |\nabla u(y,t) - (U(y),0)| \le C_u t,\qquad|U(y)| + |\nabla U(y)| \le C_u,\qquad\text{for } y \in Q,\ 0 < t < h_0, \tag*{(39)} ∣∇ u ( y , t ) − ( U ( y ) , 0 ) ∣ ≤ C u t , ∣ U ( y ) ∣ + ∣∇ U ( y ) ∣ ≤ C u , for y ∈ Q , 0 < t < h 0 , (39)
for some constant C u C_u C u depending on u u u and Q Q Q .
Next we define the new domain and the new solution. For ρ > 0 \rho> 0 ρ > 0 small, we insert columns of scale ρ \rho ρ on Q Q Q as in Subsection 5.2, and we denote by D ρ = Ψ ρ ( D ) D_\rho= \Psi_\rho(D) D ρ = Ψ ρ ( D ) the resulting domain. Writing I ρ = I ρ ( Q ) I_\rho= I_\rho(Q) I ρ = I ρ ( Q ) and K b : = ( b , 0 ) + ρ Π K_b := (b,0) + \rho\Pi K b := ( b , 0 ) + ρ Π , we have D ρ = D ∖ ⋃ b ∈ I ρ K b D_\rho= D \setminus\bigcup_{b\in I_\rho} K_b D ρ = D ∖ ⋃ b ∈ I ρ K b . The datum is not changed on the old part of ∂ D ρ \partial D_\rho ∂ D ρ , and it is zero on the new faces. More precisely, we let g ρ g_\rho g ρ be equal to g g g on ∂ D ∩ ∂ D ρ \partial D \cap\partial D_\rho ∂ D ∩ ∂ D ρ and to 0 0 0 on the walls and the roofs of the columns K b K_b K b , and we let u ρ u_\rho u ρ be the corresponding variational solution in D ρ D_\rho D ρ ; notice that ∫ ∂ D ρ g ρ d σ D ρ = ∫ ∂ D g d σ D = 0 \int_{\partial D_\rho} g_\rho\,d\sigma_{D_\rho} = \int_{\partial D} g\,d\sigma_D = 0 ∫ ∂ D ρ g ρ d σ D ρ = ∫ ∂ D g d σ D = 0 .
We stress that the constants in the trace and Poincaré inequalities in D ρ D_\rho D ρ do not depend on ρ \rho ρ . Indeed, by Lemma 5.5, applied with G 0 = G G_0 = G G 0 = G and C 0 = { Q } C_0 = \{Q\} C 0 = { Q } , we have D ρ = ( Ψ ρ ∘ G ) ( Q 0 ) D_\rho= (\Psi_\rho\circ G)(Q_0) D ρ = ( Ψ ρ ∘ G ) ( Q 0 ) , with Lip ( Ψ ρ ∘ G ) \operatorname{Lip}(\Psi_\rho\circ G) Lip ( Ψ ρ ∘ G ) and Lip ( ( Ψ ρ ∘ G ) − 1 ) \operatorname{Lip}((\Psi_\rho\circ G)^{-1}) Lip (( Ψ ρ ∘ G ) − 1 ) bounded by L F L_F L F . Thus, by Lemma 2.1 ,
∥ Tr φ − c ∥ L 2 ( σ D ρ ) + ∥ φ − c ∥ L 2 ( D ρ ) ≤ C ∥ ∇ φ ∥ L 2 ( D ρ ) , c = m ∂ D ρ Tr φ , (40) \|\operatorname{Tr}\varphi-c\|_{L^2(\sigma_{D_\rho})}+\|\varphi-c\|_{L^2(D_\rho)}
\le C\|\nabla\varphi\|_{L^2(D_\rho)},\qquad c=m_{\partial D_\rho}\operatorname{Tr}\varphi, \tag*{(40)} ∥ Tr φ − c ∥ L 2 ( σ D ρ ) + ∥ φ − c ∥ L 2 ( D ρ ) ≤ C ∥∇ φ ∥ L 2 ( D ρ ) , c = m ∂ D ρ Tr φ , (40)
with C C C independent of ρ \rho ρ .
Finally, for b ∈ I ρ b \in I_\rho b ∈ I ρ we consider the following sets, which describe D ρ D_\rho D ρ near the column K b K_b K b :
S b S_b S b is the union of the walls and the roof of K b K_b K b .
M b : = ( b , 0 ) + ρ M M_b := (b,0) + \rho M M b := ( b , 0 ) + ρM is the union of the selected squares on the walls of K b K_b K b .
N b : = ( ( b + ρ Y ) × ( 0 , ( λ + 1 ) ρ ) ) ∖ K b N_b := \left((b+\rho Y)\times(0,(\lambda+1)\rho)\right)\setminus K_b N b := ( ( b + ρ Y ) × ( 0 , ( λ + 1 ) ρ ) ) ∖ K b is the box of height ( λ + 1 ) ρ (\lambda+1)\rho ( λ + 1 ) ρ over the cell b + ρ Y b+\rho Y b + ρ Y , minus the column K b K_b K b .
Thus M b ⊂ S b ⊂ ∂ N b M_b \subset S_b \subset\partial N_b M b ⊂ S b ⊂ ∂ N b , and the sets N b N_b N b are pairwise disjoint. We denote by M ρ : = ⋃ b ∈ I ρ M b M_\rho:= \bigcup_{b\in I_\rho} M_b M ρ := ⋃ b ∈ I ρ M b the union of the new selected squares.
Lemma 6.1 (Insertion lemma). With the notation above, for all sufficiently small ρ > 0 \rho> 0 ρ > 0 we have
∥ ∇ ( u ρ − u ) ∥ L 2 ( D ρ ) ≲ ρ . (41) \|\nabla(u_\rho-u)\|_{L^2(D_\rho)} \lesssim\sqrt{\rho}. \tag*{(41)} ∥∇ ( u ρ − u ) ∥ L 2 ( D ρ ) ≲ ρ . (41)
Moreover,
lim inf ρ → 0 ∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≥ γ ∫ Q ∣ ∇ τ u ∣ d σ . (42) \liminf_{\rho\to0}\int_{M_\rho}|\nabla_\tau u_\rho|\,d\sigma
\ge\gamma\int_Q|\nabla_\tau u|\,d\sigma. \tag*{(42)} ρ → 0 lim inf ∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≥ γ ∫ Q ∣ ∇ τ u ∣ d σ . (42)
*The same holds if columns are inserted on finitely many flat squares Q 1 , … , Q r Q_1,\ldots,Q_r Q 1 , … , Q r as above, with pairwise disjoint flat neighbourhoods, with ∫ Q ∣ ∇ τ u ∣ d σ \int_Q|\nabla_\tau u|\,d\sigma ∫ Q ∣ ∇ τ u ∣ d σ replaced by ∑ i ∫ Q i ∣ ∇ τ u ∣ d σ \sum_i\int_{Q_i}|\nabla_\tau u|\,d\sigma ∑ i ∫ Q i ∣ ∇ τ u ∣ d σ in (6.4). In (6.3), the implicit constant depends on u u u , on the squares and on the parameters fixed in § 4 and § 6 . It does not depend on ρ \rho ρ , and thus neither on the number of columns, which is of order ρ − 2 \rho^{-2} ρ − 2 . +#+#+#+#+#+
We first explain the idea of the proof. It is based on the two-scale ansatz
u ρ ≈ u + ρ ∑ j = 1 2 U j ( y ) P e j ( x ρ ) , (43) u_\rho\approx u+\rho\sum_{j=1}^{2}U_j(y)P^{ej}\left(\frac{x}{\rho}\right), \tag*{(43)} u ρ ≈ u + ρ j = 1 ∑ 2 U j ( y ) P e j ( ρ x ) , (43)
where P e j P^{ej} P e j is the cell corrector of Lemma 4.1 . On the walls of the columns, the normal derivatives of the correctors cancel those of the linear functions y j y_j y j (see (21) ), and their gradients are square integrable in a period cell, by Lemma 4.1 . For this reason, most of the energy of the correction term in (43) is concentrated in a layer of thickness comparable to ρ \rho ρ over Q Q Q . The proof of Lemma 6.1 has two parts:
The main term. Close to the column K b K_b K b , the gradient of the right-hand side of (43) is approximately ( ∇ V U ( b ) ) ( x − ( b , 0 ) ρ ) (\nabla V^{U(b)})\left(\frac{x-(b,0)}{\rho}\right) ( ∇ V U ( b ) ) ( ρ x − ( b , 0 ) ) . So, on the selected squares M b \mathcal{M}_b M b , its tangential gradient is given by Lemma 4.4, and this is where the factor γ \gamma γ comes from.
The error. For a rigorous argument we need to bound the error e e e in (43) on the boundary, at the scale of the columns. By scaling, a trace inequality at scale ρ \rho ρ , summed over the ≈ ρ − 2 \approx\rho^{-2} ≈ ρ − 2 columns, loses a factor ρ − 1 / 2 \rho^{-1/2} ρ − 1/2 . Therefore we need ∥ ∇ e ∥ L 2 \|\nabla e\|_{L^2} ∥∇ e ∥ L 2 to be much smaller than ρ \sqrt{\rho} ρ .
The second part is the delicate one. Indeed, the correction term itself has energy ≈ ρ \approx\sqrt{\rho} ≈ ρ , because its gradient is of order one in a layer of thickness ≈ ρ \approx\rho ≈ ρ , and so we need to show that the ansatz (43) is correct to leading order in the energy norm, up to an error which is a small multiple of ρ \sqrt{\rho} ρ . To this end, we cut off the ansatz near ∂ Q \partial Q ∂ Q and above the layer where the columns live, at distances d d d and h = d 2 h=d^2 h = d 2 respectively, which are fixed while ρ → 0 \rho\to0 ρ → 0 . In Lemma 6.4 we show that the resulting error e e e satisfies
∥ ∇ e ∥ L 2 ≤ C d ρ + o ( ρ ) as ρ → 0 , \|\nabla e\|_{L^2}\le C\sqrt{d}\rho+o(\sqrt{\rho})\quad\text{as }\rho\to0, ∥∇ e ∥ L 2 ≤ C d ρ + o ( ρ ) as ρ → 0 ,
with C C C independent of d d d . Since u ρ u_\rho u ρ does not depend on d d d , at the end we can let d → 0 d\to0 d → 0 .
Remark 6.2. The ansatz (43) is the first-order boundary layer expansion used in the homogenization of problems in domains with rough boundaries, see for instance [4 , 22 ] for flows over rough walls and [1 ] for the Laplace equation with Neumann conditions on an oscillating boundary. We need instead precise information on the trace of ∇ u ρ \nabla u_\rho ∇ u ρ on the rough part of the boundary itself, at the scale of the columns, and this is the reason why the error has to be much smaller than ρ \sqrt{\rho} ρ in the energy norm.
Two Poincaré inequalities
Lemma 6.3. For every φ ∈ W 1 , 2 ( D ρ ) \varphi\in W^{1,2}(D_\rho) φ ∈ W 1 , 2 ( D ρ ) and b ∈ I ρ b\in\mathcal{I}_\rho b ∈ I ρ ,
∥ φ − m N b φ ∥ L 2 ( S b ) ≲ ρ ∥ ∇ φ ∥ L 2 ( N b ) . (44) \|\varphi-m_{N_b}\varphi\|_{L^2(S_b)}\lesssim\sqrt{\rho}\,\|\nabla\varphi\|_{L^2(N_b)}. \tag*{(44)} ∥ φ − m N b φ ∥ L 2 ( S b ) ≲ ρ ∥∇ φ ∥ L 2 ( N b ) . (44) Moreover, if λ ρ < h < h 0 / 2 \lambda\rho<h<h_0/2 λ ρ < h < h 0 /2 , the layer L h : = D ρ ∩ ( Q × ( 0 , 2 h ) ) L_h:=D_\rho\cap(Q\times(0,2h)) L h := D ρ ∩ ( Q × ( 0 , 2 h )) satisfies
∥ φ − c ∥ L 2 ( L h ) ≲ h ∥ ∇ φ ∥ L 2 ( D ρ ) , c = m ∂ D ρ Tr φ . (45) \|\varphi-c\|_{L^2(L_h)}\lesssim\sqrt{h}\,\|\nabla\varphi\|_{L^2(D_\rho)},\qquad c=m_{\partial D_\rho}\operatorname{Tr}\varphi. \tag*{(45)} ∥ φ − c ∥ L 2 ( L h ) ≲ h ∥∇ φ ∥ L 2 ( D ρ ) , c = m ∂ D ρ Tr φ . (45)
Proof. Consider the connected Lipschitz domain N : = ( Y × ( 0 , λ + 1 ) ) ∖ Π N:=(Y\times(0,\lambda+1))\setminus\Pi N := ( Y × ( 0 , λ + 1 )) ∖ Π , which does not depend on ρ \rho ρ , and the part S : = ∂ Π ∩ { t > 0 } S:=\partial\Pi\cap\{t>0\} S := ∂ Π ∩ { t > 0 } of its boundary. In N N N , the trace and Poincaré inequalities give
∥ φ − m N φ ∥ L 2 ( S ) ≲ ∥ ∇ φ ∥ L 2 ( N ) . \|\varphi-m_N\varphi\|_{L^2(S)}\lesssim\|\nabla\varphi\|_{L^2(N)}. ∥ φ − m N φ ∥ L 2 ( S ) ≲ ∥∇ φ ∥ L 2 ( N ) .
To get (44) , we just rescale, taking into account that surface measure scales like ρ 2 \rho^2 ρ 2 , volume like ρ 3 \rho^3 ρ 3 , and gradients like ρ − 1 \rho^{-1} ρ − 1 .
Concerning (45) , by Lemma 2.1 we can assume that φ \varphi φ is the restriction to D ρ D_\rho D ρ of a function in C ∞ ( R 3 ) C^\infty(\mathbb{R}^3) C ∞ ( R 3 ) . Given y ∈ Q y\in Q y ∈ Q , the vertical line through y y y enters L h L_h L h at a point ( y , r ( y ) ) ∈ ∂ D ρ (y,r(y))\in\partial D_\rho ( y , r ( y )) ∈ ∂ D ρ , with r ( y ) ∈ { 0 , λ ρ } r(y)\in\{0,\lambda\rho\} r ( y ) ∈ { 0 , λ ρ } , which belongs either to the floor or to the roof of a column. Then the fundamental theorem of calculus gives
∫ r ( y ) 2 h ∣ φ ( y , t ) − c ∣ 2 d t ≤ 4 h ∣ φ ( y , r ( y ) ) − c ∣ 2 + 8 h 2 ∫ r ( y ) 2 h ∣ ∂ t φ ( y , t ) ∣ 2 d t . \int_{r(y)}^{2h}|\varphi(y,t)-c|^2\,dt\le4h|\varphi(y,r(y))-c|^2+8h^2\int_{r(y)}^{2h}|\partial_t\varphi(y,t)|^2\,dt. ∫ r ( y ) 2 h ∣ φ ( y , t ) − c ∣ 2 d t ≤ 4 h ∣ φ ( y , r ( y )) − c ∣ 2 + 8 h 2 ∫ r ( y ) 2 h ∣ ∂ t φ ( y , t ) ∣ 2 d t .
Since the points ( y , r ( y ) ) (y,r(y)) ( y , r ( y )) belong to horizontal pieces of ∂ D ρ \partial D_\rho ∂ D ρ , after integrating in y ∈ Q y \in Q y ∈ Q we obtain
∥ φ − c ∥ L 2 ( L h ) 2 ≤ 4 h ∥ Tr φ − c ∥ L 2 ( σ D ρ ) 2 + 8 h 2 ∥ ∇ φ ∥ L 2 ( D ρ ) 2 , \|\varphi-c\|_{L^2(\mathcal{L}_h)}^2 \le4h\|\operatorname{Tr}\varphi-c\|_{L^2(\sigma D_\rho)}^2+8h^2\|\nabla\varphi\|_{L^2(D_\rho)}^2, ∥ φ − c ∥ L 2 ( L h ) 2 ≤ 4 h ∥ Tr φ − c ∥ L 2 ( σ D ρ ) 2 + 8 h 2 ∥∇ φ ∥ L 2 ( D ρ ) 2 ,
and then (45) is a consequence of (40) . □
The approximate solution
Here we construct the approximate solution, namely the ansatz (43) cut off near ∂ Q \partial Q ∂ Q and above the columns, and we show that, in the energy norm, the error is small compared with the correction term if d d d is small. The two cutoffs act at the scales d d d and h : = d 2 h:=d^2 h := d 2 , where d ∈ ( 0 , 1 ) d\in(0,1) d ∈ ( 0 , 1 ) is so small that h < h 0 / 2 h<h_0/2 h < h 0 /2 . We stress that d d d does not depend on ρ \rho ρ : we first let ρ → 0 \rho\to0 ρ → 0 with d d d fixed, and only at the end of the proof of Lemma 6.1 we let d → 0 d\to0 d → 0 . Accordingly, we assume that ρ \rho ρ is so small, depending on d d d , that ( λ + 1 ) ρ < h (\lambda+1)\rho<h ( λ + 1 ) ρ < h and 10 ς ρ < d 10\varsigma\rho<d 10 ς ρ < d . In particular, the columns lie below the height h h h , and the cells, of side ς ρ \varsigma\rho ς ρ , are much smaller than d d d . We consider the following two cutoff functions:
the horizontal cutoff ζ ∈ C ∞ ( int Q ) \zeta\in C^\infty(\operatorname{int}Q) ζ ∈ C ∞ ( int Q ) , with 0 ≤ ζ ≤ 1 0\le\zeta\le1 0 ≤ ζ ≤ 1 and ∣ ∇ ζ ∣ ≲ d − 1 |\nabla\zeta|\lesssim d^{-1} ∣∇ ζ ∣ ≲ d − 1 , such that, for y ∈ Q y\in Q y ∈ Q , ζ ( y ) = 1 \zeta(y)=1 ζ ( y ) = 1 if dist ( y , ∂ Q ) ≥ d \operatorname{dist}(y,\partial Q)\ge d dist ( y , ∂ Q ) ≥ d and ζ ( y ) = 0 \zeta(y)=0 ζ ( y ) = 0 if dist ( y , ∂ Q ) ≤ d / 2 \operatorname{dist}(y,\partial Q)\le d/2 dist ( y , ∂ Q ) ≤ d /2 ;
the vertical cutoff ϑ ∈ C ∞ ( [ 0 , ∞ ) ) \vartheta\in C^\infty([0,\infty)) ϑ ∈ C ∞ ([ 0 , ∞ )) , with 0 ≤ ϑ ≤ 1 0\le\vartheta\le1 0 ≤ ϑ ≤ 1 and ∣ ϑ ′ ∣ ≲ h − 1 |\vartheta'|\lesssim h^{-1} ∣ ϑ ′ ∣ ≲ h − 1 , such that ϑ = 1 \vartheta=1 ϑ = 1 on [ 0 , h ] [0,h] [ 0 , h ] and ϑ = 0 \vartheta=0 ϑ = 0 on [ 2 h , ∞ ) [2h,\infty) [ 2 h , ∞ ) .
For j = 1 , 2 j=1,2 j = 1 , 2 , we denote by p j p_j p j the rescaled cell correctors, and by a j a_j a j the slowly varying coefficients of the ansatz, cut off by ζ \zeta ζ and ϑ \vartheta ϑ , that is,
p j ( x ) : = ρ P e j ( x ρ ) and a j ( y , t ) : = ϑ ( t ) ζ ( y ) U j ( y ) . p_j(x):=\rho P^{e_j}\left(\frac{x}{\rho}\right)
\qquad\text{and}\qquad
a_j(y,t):=\vartheta(t)\zeta(y)U_j(y). p j ( x ) := ρ P e j ( ρ x ) and a j ( y , t ) := ϑ ( t ) ζ ( y ) U j ( y ) .
Then, we set
X ρ : = ∑ j = 1 2 a j p j and e : = u ρ − u − X ρ . X_\rho:=\sum_{j=1}^{2}a_jp_j
\qquad\text{and}\qquad
e:=u_\rho-u-X_\rho. X ρ := j = 1 ∑ 2 a j p j and e := u ρ − u − X ρ .
Thus X ρ X_\rho X ρ is the correction term in (43) , cut off by ζ \zeta ζ near ∂ Q \partial Q ∂ Q and by ϑ \vartheta ϑ above the columns, and we call it the corrector . The function e e e is the error of the approximation. Notice that X ρ X_\rho X ρ and e e e depend on d d d , but u ρ u_\rho u ρ does not.
We now check that the corrector is well defined. Since 10 ς ρ < d 10\varsigma\rho<d 10 ς ρ < d , every cell b + ρ Y b+\rho Y b + ρ Y meeting the support of ζ \zeta ζ is a complete cell, that is, b ∈ I ρ b\in\mathcal{I}_\rho b ∈ I ρ . Hence, in the support of the functions a j a_j a j , the domain D ρ D_\rho D ρ coincides with the scaled periodic domain ρ P \rho P ρP . So the functions p j p_j p j are well defined there, and X ρ X_\rho X ρ , extended by zero, belongs to W 1 , 2 ( D ρ ) W^{1,2}(D_\rho) W 1 , 2 ( D ρ ) .
Next we collect the scaling properties of the functions p j p_j p j . To this end, we consider the following two regions:
S ρ S_\rho S ρ is the union of the half-prisms ( b + ρ Y ) × ( 0 , ∞ ) (b+\rho Y)\times(0,\infty) ( b + ρ Y ) × ( 0 , ∞ ) , b ∈ I ρ b\in\mathcal{I}_\rho b ∈ I ρ , minus the columns.
S ρ ′ S'_\rho S ρ ′ is the union of those half-prisms with dist ( b + ρ Y , ∂ Q ) < 2 d \operatorname{dist}(b+\rho Y,\partial Q)<2d dist ( b + ρ Y , ∂ Q ) < 2 d , minus the columns. We call the corresponding cells the strip cells .
There are ≲ ρ − 2 \lesssim\rho^{-2} ≲ ρ − 2 cells, and ≲ d ρ − 2 \lesssim d\rho^{-2} ≲ d ρ − 2 strip cells. By a change of variables, on each half-prism the squared L 2 L^2 L 2 norms of ∇ p j \nabla p_j ∇ p j and of p j p_j p j equal ρ 3 \rho^3 ρ 3 and ρ 5 \rho^5 ρ 5 times those of ∇ P e j \nabla P^{e_j} ∇ P e j and P e j P^{e_j} P e j on P Y P_Y P Y . Let also
T ( R ) : = ( ∑ j = 1 2 ∫ P Y ∩ { t > R } ∣ ∇ P e j ∣ 2 ) 1 / 2 for R > 0 , T(R):=\left(\sum_{j=1}^{2}\int_{P_Y\cap\{t>R\}}|\nabla P^{e_j}|^2\right)^{1/2}
\qquad\text{for }R>0, T ( R ) := ( j = 1 ∑ 2 ∫ P Y ∩ { t > R } ∣∇ P e j ∣ 2 ) 1/2 for R > 0 ,
which does not depend on ρ \rho ρ and d d d , and tends to 0 0 0 as R → ∞ R\to\infty R → ∞ , because ∇ P e j ∈ L 2 ( P Y ) \nabla P^{e_j}\in L^2(P_Y) ∇ P e j ∈ L 2 ( P Y ) . Hence, by Lemma 4.1,
∥ ∇ p j ∥ L 2 ( S ρ ) ≲ ρ , ∥ ∇ p j ∥ L 2 ( S ρ ′ ) ≲ ρ d , ∥ ∇ p j ∥ L 2 ( S ρ ∩ { t > h } ) ≲ ρ T ( h / ρ ) , \|\nabla p_j\|_{L^2(S_\rho)}\lesssim\sqrt{\rho},
\qquad
\|\nabla p_j\|_{L^2(S'_\rho)}\lesssim\sqrt{\rho d},
\qquad
\|\nabla p_j\|_{L^2(S_\rho\cap\{t>h\})}\lesssim\sqrt{\rho}\,T(h/\rho), ∥∇ p j ∥ L 2 ( S ρ ) ≲ ρ , ∥∇ p j ∥ L 2 ( S ρ ′ ) ≲ ρ d , ∥∇ p j ∥ L 2 ( S ρ ∩ { t > h }) ≲ ρ T ( h / ρ ) ,
∥ p j ∥ L 2 ( S ρ ) ≲ ρ 3 / 2 . (46) \|p_j\|_{L^2(S_\rho)}\lesssim\rho^{3/2}. \tag*{(46)} ∥ p j ∥ L 2 ( S ρ ) ≲ ρ 3/2 . (46)
In words, the first column of (46) gives the total size of p j p_j p j and ∇ p j \nabla p_j ∇ p j , the second one shows that the strip cells only carry a fraction d \sqrt{d} d of the energy, and the third one shows that, for fixed h h h , the energy above the height h h h is o ( ρ ) o(\sqrt{\rho}) o ( ρ ) as ρ → 0 \rho\to0 ρ → 0 .
The main estimate of this subsection is the following. It says that, up to an error of order d ρ \sqrt{d\rho} d ρ , the ansatz (43) is correct to leading order in the energy norm, which is what we needed, see Subsection 6.1.
Lemma 6.4. There is a constant C C C , independent of ρ \rho ρ and d d d , such that, if ρ > 0 \rho> 0 ρ > 0 is small enough, depending on d d d ,
∥ ∇ e ∥ L 2 ( D ρ ) ≤ C ρ ( d + T ( d 2 / ρ ) d + ρ d 2 ) . (47) \|\nabla e\|_{L^2(D_\rho)} \le C\sqrt{\rho}\left(\sqrt{d}+\frac{T(d^2/\rho)}{d}+\frac{\sqrt{\rho}}{d^2}\right). \tag*{(47)} ∥∇ e ∥ L 2 ( D ρ ) ≤ C ρ ( d + d T ( d 2 / ρ ) + d 2 ρ ) . (47) In particular, lim sup ρ → 0 ρ − 1 / 2 ∥ ∇ e ∥ L 2 ( D ρ ) ≤ C d \limsup_{\rho\to0}\rho^{-1/2}\|\nabla e\|_{L^2(D_\rho)} \le C\sqrt{d} lim sup ρ → 0 ρ − 1/2 ∥∇ e ∥ L 2 ( D ρ ) ≤ C d .
Proof. The residual identity. Fix φ ∈ C ∞ ( R 3 ) \varphi\in C^\infty(\mathbb{R}^3) φ ∈ C ∞ ( R 3 ) . Since e = u ρ − u − X ρ e=u_\rho-u-X_\rho e = u ρ − u − X ρ , we compute one by one the integrals over D ρ D_\rho D ρ of ∇ u ρ ⋅ ∇ φ \nabla u_\rho\cdot\nabla\varphi ∇ u ρ ⋅ ∇ φ , ∇ u ⋅ ∇ φ \nabla u\cdot\nabla\varphi ∇ u ⋅ ∇ φ and ∇ X ρ ⋅ ∇ φ \nabla X_\rho\cdot\nabla\varphi ∇ X ρ ⋅ ∇ φ . The first one is easy:
∫ D ρ ∇ u ρ ⋅ ∇ φ = ∫ ∂ D ρ g ρ φ d σ = ∫ ∂ D g φ d σ , \int_{D_\rho}\nabla u_\rho\cdot\nabla\varphi=\int_{\partial D_\rho}g_\rho\varphi\,d\sigma=\int_{\partial D}g\varphi\,d\sigma, ∫ D ρ ∇ u ρ ⋅ ∇ φ = ∫ ∂ D ρ g ρ φ d σ = ∫ ∂ D g φ d σ ,
since g = 0 g=0 g = 0 on the part of ∂ D \partial D ∂ D that has been removed.
For the second integral, let ν \nu ν be the outer unit normal of D ρ D_\rho D ρ , and notice that on S b S_b S b it points into the column K b K_b K b . By Lemma 2.2, the even reflection of u u u across { t = 0 } \{t=0\} { t = 0 } is harmonic near K b K_b K b , and ∂ t u = 0 \partial_tu=0 ∂ t u = 0 on the bottom of K b K_b K b . Hence, integrating by parts in each K b K_b K b , we get
∫ D ρ ∇ u ⋅ ∇ φ = ∫ D ∇ u ⋅ ∇ φ − ∑ b ∫ K b ∇ u ⋅ ∇ φ = ∫ ∂ D g φ d σ + ∑ b ∫ S b φ ∇ u ⋅ ν d σ . \int_{D_\rho}\nabla u\cdot\nabla\varphi=\int_D\nabla u\cdot\nabla\varphi-\sum_b\int_{K_b}\nabla u\cdot\nabla\varphi=\int_{\partial D}g\varphi\,d\sigma+\sum_b\int_{S_b}\varphi\nabla u\cdot\nu\,d\sigma. ∫ D ρ ∇ u ⋅ ∇ φ = ∫ D ∇ u ⋅ ∇ φ − b ∑ ∫ K b ∇ u ⋅ ∇ φ = ∫ ∂ D g φ d σ + b ∑ ∫ S b φ ∇ u ⋅ ν d σ .
In the last identity we have used that ∫ K b ∇ u ⋅ ∇ φ = − ∫ S b φ ∇ u ⋅ ν d σ \int_{K_b}\nabla u\cdot\nabla\varphi=-\int_{S_b}\varphi\nabla u\cdot\nu\,d\sigma ∫ K b ∇ u ⋅ ∇ φ = − ∫ S b φ ∇ u ⋅ ν d σ , since − ν -\nu − ν is the outer unit normal of K b K_b K b on S b S_b S b and there is no contribution from the bottom of K b K_b K b . From the first two identities, it follows that
∫ D ρ ∇ ( u ρ − u ) ⋅ ∇ φ = − ∑ b ∫ S b φ ∇ u ⋅ ν d σ . \int_{D_\rho}\nabla(u_\rho-u)\cdot\nabla\varphi=-\sum_b\int_{S_b}\varphi\nabla u\cdot\nu\,d\sigma. ∫ D ρ ∇ ( u ρ − u ) ⋅ ∇ φ = − b ∑ ∫ S b φ ∇ u ⋅ ν d σ .
We now deal with the third integral. The function y j + p j ( x ) = ρ V e j ( x / ρ ) y_j+p_j(x)=\rho V^{e_j}(x/\rho) y j + p j ( x ) = ρ V e j ( x / ρ ) solves the Neumann problem in ρ P \rho P ρP , and by Lemma 4.1 we can use a j φ a_j\varphi a j φ , which has bounded support, as a test function. So
∫ ∇ p j ⋅ ∇ ( a j φ ) = − ∫ ∂ y j ( a j φ ) , \int\nabla p_j\cdot\nabla(a_j\varphi)=-\int\partial_{y_j}(a_j\varphi), ∫ ∇ p j ⋅ ∇ ( a j φ ) = − ∫ ∂ y j ( a j φ ) ,
and by the divergence theorem the right-hand side equals − ∫ S b a j φ ν j d σ -\int_{S_b}a_j\varphi\nu_j\,d\sigma − ∫ S b a j φ ν j d σ , because ν \nu ν is vertical on the floor and on the roofs. (This is just (21) , rescaled.) Now we write ∇ X ρ = ∑ j ( a j ∇ p j + p j ∇ a j ) \nabla X_\rho=\sum_j(a_j\nabla p_j+p_j\nabla a_j) ∇ X ρ = ∑ j ( a j ∇ p j + p j ∇ a j ) , we use the identity
a j ∇ p j ⋅ ∇ φ = ∇ p j ⋅ ∇ ( a j φ ) − φ ∇ p j ⋅ ∇ a j , a_j\nabla p_j\cdot\nabla\varphi=\nabla p_j\cdot\nabla(a_j\varphi)-\varphi\nabla p_j\cdot\nabla a_j, a j ∇ p j ⋅ ∇ φ = ∇ p j ⋅ ∇ ( a j φ ) − φ ∇ p j ⋅ ∇ a j ,
and we sum over j j j , taking into account that ∑ j a j ν j = ζ ( U , 0 ) ⋅ ν \sum_j a_j\nu_j=\zeta\,(U,0)\cdot\nu ∑ j a j ν j = ζ ( U , 0 ) ⋅ ν on S b S_b S b (because ϑ = 1 \vartheta=1 ϑ = 1 there). This gives
∫ D ρ ∇ X ρ ⋅ ∇ φ = − ∑ b ∫ S b φ ζ ( U , 0 ) ⋅ ν d σ − ∫ D ρ f φ + ∫ D ρ G ⋅ ∇ φ , \int_{D_\rho}\nabla X_\rho\cdot\nabla\varphi=-\sum_b\int_{S_b}\varphi\zeta\,(U,0)\cdot\nu\,d\sigma-\int_{D_\rho}f\varphi+\int_{D_\rho}G\cdot\nabla\varphi, ∫ D ρ ∇ X ρ ⋅ ∇ φ = − b ∑ ∫ S b φζ ( U , 0 ) ⋅ ν d σ − ∫ D ρ f φ + ∫ D ρ G ⋅ ∇ φ ,
where
f : = ∑ j ∇ p j ⋅ ∇ a j and G : = ∑ j p j ∇ a j . f:=\sum_j\nabla p_j\cdot\nabla a_j\qquad\text{and}\qquad G:=\sum_jp_j\nabla a_j. f := j ∑ ∇ p j ⋅ ∇ a j and G := j ∑ p j ∇ a j .
Let q b : = ( ∇ u − ζ ( U , 0 ) ) ⋅ ν q_b:=(\nabla u-\zeta(U,0))\cdot\nu q b := ( ∇ u − ζ ( U , 0 )) ⋅ ν on S b S_b S b . Subtracting the identity for X ρ X_\rho X ρ from the one for u ρ − u u_\rho-u u ρ − u , we obtain
∫ D ρ ∇ e ⋅ ∇ φ = − ∑ b ∫ S b q b φ d σ + ∫ D ρ f φ − ∫ D ρ G ⋅ ∇ φ . \int_{D_\rho}\nabla e\cdot\nabla\varphi=-\sum_b\int_{S_b}q_b\varphi\,d\sigma+\int_{D_\rho}f\varphi-\int_{D_\rho}G\cdot\nabla\varphi. ∫ D ρ ∇ e ⋅ ∇ φ = − b ∑ ∫ S b q b φ d σ + ∫ D ρ f φ − ∫ D ρ G ⋅ ∇ φ .
If we take φ = 1 \varphi= 1 φ = 1 in a neighbourhood of D ρ ‾ \overline{D_\rho} D ρ , we see that ∑ b ∫ S b q b = ∫ f \sum_b \int_{S_b} q_b = \int f ∑ b ∫ S b q b = ∫ f . Therefore, for any constant c c c ,
∫ D ρ ∇ e ⋅ ∇ φ = − ∑ b ∫ S b q b ( φ − c ) d σ + ∫ D ρ f ( φ − c ) − ∫ D ρ G ⋅ ∇ φ . (48) \int_{D_\rho} \nabla e \cdot\nabla\varphi= -\sum_b \int_{S_b} q_b(\varphi-c)\,d\sigma+ \int_{D_\rho} f(\varphi-c) - \int_{D_\rho} G\cdot\nabla\varphi. \tag*{(48)} ∫ D ρ ∇ e ⋅ ∇ φ = − b ∑ ∫ S b q b ( φ − c ) d σ + ∫ D ρ f ( φ − c ) − ∫ D ρ G ⋅ ∇ φ . (48)
Roughly speaking, I measures how far u u u plus the correction is from satisfying the new Neumann condition on the walls of the columns, while II and III are due to the derivatives of the slowly varying coefficients a j a_j a j . The estimates below and a density argument show that (48) is valid for all φ ∈ W 1 , 2 ( D ρ ) \varphi\in W^{1,2}(D_\rho) φ ∈ W 1 , 2 ( D ρ ) . From now on we take c = m ∂ D ρ Tr φ c=m_{\partial D_\rho}\operatorname{Tr}\varphi c = m ∂ D ρ Tr φ , we assume that ∥ ∇ φ ∥ L 2 ( D ρ ) = 1 \|\nabla\varphi\|_{L^2(D_\rho)}=1 ∥∇ φ ∥ L 2 ( D ρ ) = 1 , and we estimate I, II and III separately.
Estimate of I. For x ∈ S b x\in S_b x ∈ S b we split
q b = v b ⋅ ν + r b , where v b : = ( ( 1 − ζ ( b ) ) U ( b ) , 0 ) . q_b=v_b\cdot\nu+r_b,\qquad\text{where }v_b:=((1-\zeta(b))U(b),0). q b = v b ⋅ ν + r b , where v b := (( 1 − ζ ( b )) U ( b ) , 0 ) .
The constant vector v b v_b v b is horizontal, and so its flux through the boundary of K b K_b K b vanishes, as well as its flux through the bottom of K b K_b K b . Hence ∫ S b v b ⋅ ν = 0 \int_{S_b}v_b\cdot\nu=0 ∫ S b v b ⋅ ν = 0 , and by (44) ,
∣ ∫ S b ( v b ⋅ ν ) ( φ − c ) ∣ = ∣ ∫ S b ( v b ⋅ ν ) ( φ − m N b φ ) ∣ ≲ ∣ S b ∣ 1 / 2 ρ ∥ ∇ φ ∥ L 2 ( N b ) ≲ ρ 3 / 2 ∥ ∇ φ ∥ L 2 ( N b ) . \left|\int_{S_b}(v_b\cdot\nu)(\varphi-c)\right|
=\left|\int_{S_b}(v_b\cdot\nu)(\varphi-m_{N_b}\varphi)\right|
\lesssim|S_b|^{1/2}\sqrt{\rho}\,\|\nabla\varphi\|_{L^2(N_b)}
\lesssim\rho^{3/2}\|\nabla\varphi\|_{L^2(N_b)}. ∫ S b ( v b ⋅ ν ) ( φ − c ) = ∫ S b ( v b ⋅ ν ) ( φ − m N b φ ) ≲ ∣ S b ∣ 1/2 ρ ∥∇ φ ∥ L 2 ( N b ) ≲ ρ 3/2 ∥∇ φ ∥ L 2 ( N b ) .
As v b v_b v b vanishes outside the strip cells, by the Cauchy-Schwarz inequality the total contribution of these terms is ≲ ρ 3 / 2 ( d ρ − 2 ) 1 / 2 = ρ d \lesssim\rho^{3/2}(d\rho^{-2})^{1/2}=\sqrt{\rho d} ≲ ρ 3/2 ( d ρ − 2 ) 1/2 = ρ d . On the other hand, for x = ( y , t ) ∈ S b x=(y,t)\in S_b x = ( y , t ) ∈ S b we have
r b ( x ) = ( ∇ u ( y , t ) − ( U ( y ) , 0 ) ) ⋅ ν + ( ( 1 − ζ ( y ) ) ( U ( y ) − U ( b ) ) , 0 ) ⋅ ν + ( ( ζ ( b ) − ζ ( y ) ) U ( b ) , 0 ) ⋅ ν , r_b(x)=\bigl(\nabla u(y,t)-(U(y),0)\bigr)\cdot\nu+\bigl((1-\zeta(y))(U(y)-U(b)),0\bigr)\cdot\nu+\bigl((\zeta(b)-\zeta(y))U(b),0\bigr)\cdot\nu, r b ( x ) = ( ∇ u ( y , t ) − ( U ( y ) , 0 ) ) ⋅ ν + ( ( 1 − ζ ( y )) ( U ( y ) − U ( b )) , 0 ) ⋅ ν + ( ( ζ ( b ) − ζ ( y )) U ( b ) , 0 ) ⋅ ν ,
and ∣ x − ( b , 0 ) ∣ ≲ ρ |x-(b,0)|\lesssim\rho ∣ x − ( b , 0 ) ∣ ≲ ρ . So, by (39) , the first two terms are O ( ρ ) O(\rho) O ( ρ ) on every cell, while the last one is O ( ρ / d ) O(\rho/d) O ( ρ / d ) , because of ∇ ζ \nabla\zeta ∇ ζ . Outside the strip cells, ζ = 1 \zeta=1 ζ = 1 on the cell and only the first term survives. Thus ∣ r b ∣ ≲ ρ |r_b|\lesssim\rho ∣ r b ∣ ≲ ρ on the cells which are not strip cells, and ∣ r b ∣ ≲ ρ / d |r_b|\lesssim\rho/d ∣ r b ∣ ≲ ρ / d on the strip cells. Each set S b S_b S b has area ( 8 λ + 4 ) ρ 2 (8\lambda+4)\rho^2 ( 8 λ + 4 ) ρ 2 , and so the total area of the sets S b S_b S b is ≲ 1 \lesssim1 ≲ 1 , and that of the sets S b S_b S b in the strip cells is ≲ d \lesssim d ≲ d , so that
( ∑ b ∥ r b ∥ L 2 ( S b ) 2 ) 1 / 2 ≲ ρ + ρ d − 1 / 2 . \left(\sum_b\|r_b\|_{L^2(S_b)}^2\right)^{1/2}\lesssim\rho+\rho d^{-1/2}. ( b ∑ ∥ r b ∥ L 2 ( S b ) 2 ) 1/2 ≲ ρ + ρ d − 1/2 .
By the Cauchy-Schwarz inequality and (40) , we conclude that
∣ I ∣ ≲ ρ d + ρ d − 1 / 2 . |I|\lesssim\sqrt{\rho d}+\rho d^{-1/2}. ∣ I ∣ ≲ ρ d + ρ d − 1/2 .
Estimate of II. The derivatives of a j a_j a j split into three parts,
∇ a j = ϑ ′ ζ ( ∇ U j , 0 ) + ϑ U j ( ∇ ζ , 0 ) + ϑ ′ ζ U j e t , \nabla a_j=\vartheta'\zeta(\nabla U_j,0)+\vartheta U_j(\nabla\zeta,0)+\vartheta'\zeta U_j e_t, ∇ a j = ϑ ′ ζ ( ∇ U j , 0 ) + ϑ U j ( ∇ ζ , 0 ) + ϑ ′ ζ U j e t ,
coming from ∇ U j \nabla U_j ∇ U j , from ∇ ζ \nabla\zeta ∇ ζ and from ϑ ′ \vartheta' ϑ ′ , of sizes ≲ 1 \lesssim1 ≲ 1 , ≲ d − 1 \lesssim d^{-1} ≲ d − 1 (on the strip cells) and ≲ h − 1 \lesssim h^{-1} ≲ h − 1 (where t > h t>h t > h ), respectively. By (46) ,
∥ f ∥ L 2 ( D ρ ) ≤ ∑ j ∥ ∣ ∇ p j ∣ ∣ ∇ a j ∣ ∥ L 2 ( D ρ ) ≲ ρ + d − 1 ρ d + h − 1 ρ T ( h / ρ ) ≲ ρ / d + d − 2 ρ T ( d 2 / ρ ) . \|f\|_{L^2(D_\rho)}\leq\sum_j\|\lvert\nabla p_j\rvert\lvert\nabla a_j\rvert\|_{L^2(D_\rho)}
\lesssim\sqrt{\rho}+d^{-1}\sqrt{\rho d}+h^{-1}\sqrt{\rho}\,T(h/\rho)
\lesssim\sqrt{\rho/d}+d^{-2}\sqrt{\rho}\,T(d^2/\rho). ∥ f ∥ L 2 ( D ρ ) ≤ j ∑ ∥ ∣ ∇ p j ∣ ∣ ∇ a j ∣ ∥ L 2 ( D ρ ) ≲ ρ + d − 1 ρ d + h − 1 ρ T ( h / ρ ) ≲ ρ / d + d − 2 ρ T ( d 2 / ρ ) .
Since f f f is supported in the layer L h L_h L h , Lemma 6.3 gives
∣ I I ∣ ≤ ∥ f ∥ L 2 ( D ρ ) ∥ φ − c ∥ L 2 ( L h ) ≲ h ( ρ / d + ρ T ( d 2 / ρ ) d 2 ) = ρ d + ρ T ( d 2 / ρ ) d . |II|\leq\|f\|_{L^2(D_\rho)}\|\varphi-c\|_{L^2(L_h)}
\lesssim\sqrt{h}\left(\sqrt{\rho/d}+\frac{\sqrt{\rho}\,T(d^2/\rho)}{d^2}\right)
=\sqrt{\rho d}+\frac{\sqrt{\rho}\,T(d^2/\rho)}{d}. ∣ I I ∣ ≤ ∥ f ∥ L 2 ( D ρ ) ∥ φ − c ∥ L 2 ( L h ) ≲ h ( ρ / d + d 2 ρ T ( d 2 / ρ ) ) = ρ d + d ρ T ( d 2 / ρ ) .
Estimate of III. Since ∣ ∇ a j ∣ ≲ 1 + d − 1 + h − 1 ≲ d − 2 |\nabla a_j|\lesssim1+d^{-1}+h^{-1}\lesssim d^{-2} ∣∇ a j ∣ ≲ 1 + d − 1 + h − 1 ≲ d − 2 , by (46) we get
∣ I I I ∣ ≤ ∥ G ∥ L 2 ( D ρ ) ≤ ∑ j ∥ p j ∥ L 2 ( S ρ ) ∥ ∇ a j ∥ L ∞ ≲ ρ 3 / 2 d − 2 . |III|\leq\|G\|_{L^2(D_\rho)}\leq\sum_j\|p_j\|_{L^2(S_\rho)}\|\nabla a_j\|_{L^\infty}\lesssim\rho^{3/2}d^{-2}. ∣ I I I ∣ ≤ ∥ G ∥ L 2 ( D ρ ) ≤ j ∑ ∥ p j ∥ L 2 ( S ρ ) ∥∇ a j ∥ L ∞ ≲ ρ 3/2 d − 2 .
Finally, we gather the three estimates. Since ρ , d < 1 \rho,d<1 ρ , d < 1 , we have ρ d − 1 / 2 + ρ 3 / 2 d − 2 ≤ 2 ρ d − 2 \rho d^{-1/2}+\rho^{3/2}d^{-2}\le2\rho d^{-2} ρ d − 1/2 + ρ 3/2 d − 2 ≤ 2 ρ d − 2 , and so
∣ ∫ D ρ ∇ e ⋅ ∇ φ ∣ ≲ ρ ( d + T ( d 2 / ρ ) d + ρ d 2 ) ∥ ∇ φ ∥ L 2 ( D ρ ) \left|\int_{D_\rho}\nabla e\cdot\nabla\varphi\right|\lesssim\sqrt{\rho}\left(\sqrt{d}+\frac{T(d^2/\rho)}{d}+\frac{\sqrt{\rho}}{d^2}\right)\|\nabla\varphi\|_{L^2(D_\rho)} ∫ D ρ ∇ e ⋅ ∇ φ ≲ ρ ( d + d T ( d 2 / ρ ) + d 2 ρ ) ∥∇ φ ∥ L 2 ( D ρ )
for every φ ∈ W 1 , 2 ( D ρ ) \varphi\in W^{1,2}(D_\rho) φ ∈ W 1 , 2 ( D ρ ) . The choice φ = e \varphi=e φ = e gives (47) , and the last assertion of the lemma follows because T ( R ) → 0 T(R)\to0 T ( R ) → 0 as R → ∞ R\to\infty R → ∞ .
Remark 6.5. The order of the limits is important. For fixed d d d , the last two terms in (47) are o ( ρ ) o(\sqrt{\rho}) o ( ρ ) as ρ → 0 \rho\to0 ρ → 0 . The second one comes from the part of the correctors above the height h h h , and for it we only need that T ( R ) → 0 T(R)\to0 T ( R ) → 0 as R → ∞ R\to\infty R → ∞ , that is, that the correctors have finite energy; no rate of decay is needed. The first term, ρ d \sqrt{\rho d} ρ d , is due to the strip along ∂ Q \partial Q ∂ Q where the ansatz is cut off, and to the gradient of ζ \zeta ζ paired with functions in the layer of thickness h h h , which gives ρ h / d = ρ d \sqrt{\rho h/d}=\sqrt{\rho d} ρ h / d = ρ d . It is not o ( ρ ) o(\sqrt{\rho}) o ( ρ ) , but it is a small multiple of ρ \sqrt{\rho} ρ if d d d is small, and this is enough for us, because u ρ u_\rho u ρ does not depend on d d d .
Proof of the insertion lemma
Since ∇ X ρ = ∑ j a j ∇ p j + G \nabla X_\rho=\sum_j a_j\nabla p_j+G ∇ X ρ = ∑ j a j ∇ p j + G and ∣ a j ∣ ≲ 1 |a_j|\lesssim1 ∣ a j ∣ ≲ 1 , by (46) and the bound for ∥ G ∥ L 2 ( D ρ ) \|G\|_{L^2(D_\rho)} ∥ G ∥ L 2 ( D ρ ) in the proof of Lemma 6.4 ,
∥ ∇ X ρ ∥ L 2 ( D ρ ) ≤ ∑ j ∥ a j ∇ p j ∥ L 2 + ∥ G ∥ L 2 ≲ ρ + ρ 3 / 2 d − 2 . \|\nabla X_\rho\|_{L^2(D_\rho)}\le\sum_j\|a_j\nabla p_j\|_{L^2}+\|G\|_{L^2}\lesssim\sqrt{\rho}+\rho^{3/2}d^{-2}. ∥∇ X ρ ∥ L 2 ( D ρ ) ≤ j ∑ ∥ a j ∇ p j ∥ L 2 + ∥ G ∥ L 2 ≲ ρ + ρ 3/2 d − 2 .
Together with Lemma 6.4 , since u ρ − u = X ρ + e u_\rho-u=X_\rho+e u ρ − u = X ρ + e , this gives ∥ ∇ ( u ρ − u ) ∥ L 2 ( D ρ ) ≤ C d ρ \|\nabla(u_\rho-u)\|_{L^2(D_\rho)}\le C_d\sqrt{\rho} ∥∇ ( u ρ − u ) ∥ L 2 ( D ρ ) ≤ C d ρ for ρ \rho ρ small enough, where C d C_d C d depends on d d d . As u ρ u_\rho u ρ does not depend on d d d , we can apply this with any fixed value of d d d , and we get (41) .
For the lower bound (42) , we use only the cells which are far from ∂ Q \partial Q ∂ Q ,
J ρ : = { b ∈ I ρ : dist ( b + ρ Y , ∂ Q ) ≥ d } . \mathcal{J}_\rho:=\{b\in\mathcal{I}_\rho:\operatorname{dist}(b+\rho Y,\partial Q)\ge d\}. J ρ := { b ∈ I ρ : dist ( b + ρ Y , ∂ Q ) ≥ d } .
For b ∈ J ρ b\in\mathcal{J}_\rho b ∈ J ρ we have ζ = ϑ = 1 \zeta=\vartheta=1 ζ = ϑ = 1 in N b N_b N b , and we compare u + X ρ u+X_\rho u + X ρ with the frozen ansatz
w b ( x ) : = u ( b , 0 ) + ρ V U ( b ) ( x − ( b , 0 ) ρ ) , w_b(x):=u(b,0)+\rho V^{U(b)}\left(\frac{x-(b,0)}{\rho}\right), w b ( x ) := u ( b , 0 ) + ρ V U ( b ) ( ρ x − ( b , 0 ) ) ,
which is the cell solution with the value of the old tangential gradient at the point b b b , rescaled to the cell of b b b . Since ∇ w b = ( U ( b ) , 0 ) + ∑ j U j ( b ) ∇ p j \nabla w_b=(U(b),0)+\sum_j U_j(b)\nabla p_j ∇ w b = ( U ( b ) , 0 ) + ∑ j U j ( b ) ∇ p j by periodicity, in N b N_b N b we have
∇ ( u + X ρ − w b ) = ( ∇ u − ( U ( b ) , 0 ) ) + ∑ j ( U j ( y ) − U j ( b ) ) ∇ p j + ∑ j p j ( ∇ U j , 0 ) . \nabla(u+X_\rho-w_b)=(\nabla u-(U(b),0))+\sum_j(U_j(y)-U_j(b))\nabla p_j+\sum_jp_j(\nabla U_j,0). ∇ ( u + X ρ − w b ) = ( ∇ u − ( U ( b ) , 0 )) + j ∑ ( U j ( y ) − U j ( b )) ∇ p j + j ∑ p j ( ∇ U j , 0 ) .
By Lemma 4.1 , the gradients ∇ p j ( x ) = ( ∇ P e j ) ( x / ρ ) \nabla p_j(x)=(\nabla P^{e_j})(x/\rho) ∇ p j ( x ) = ( ∇ P e j ) ( x / ρ ) are bounded on M b \mathcal{M}_b M b , because M \mathcal{M} M is a compact subset of the open walls of Π \Pi Π . Also, ∣ p j ∣ ≲ ρ |p_j|\lesssim\rho ∣ p j ∣ ≲ ρ on M b \mathcal{M}_b M b . Then (39) gives
∣ ∇ τ ( u + X ρ − w b ) ∣ ≲ ρ on M b . (49) |\nabla_\tau(u+X_\rho-w_b)|\lesssim\rho\qquad\text{on }\mathcal{M}_b. \tag*{(49)} ∣ ∇ τ ( u + X ρ − w b ) ∣ ≲ ρ on M b . (49)
Next we bring in the test fields of Lemma 4.4. Let j ( b ) j(b) j ( b ) be the index provided by that lemma for U = U ( b ) U=U(b) U = U ( b ) . We transport the corresponding test field to M b \mathcal{M}_b M b , that is, we set Υ ~ b ( x ) : = Υ ~ j ( b ) ( ( x − ( b , 0 ) ) / ρ ) \widetilde{\Upsilon}_b(x):=\widetilde{\Upsilon}_{j(b)}((x-(b,0))/\rho) Υ b ( x ) := Υ j ( b ) (( x − ( b , 0 )) / ρ ) . Then ∣ Υ ~ b ∣ ≤ 1 |\widetilde{\Upsilon}_b|\le1 ∣ Υ b ∣ ≤ 1 and ∣ div τ Υ ~ b ∣ ≤ C Υ / ρ |\operatorname{div}_\tau\widetilde{\Upsilon}_b|\le C_\Upsilon/\rho ∣ div τ Υ b ∣ ≤ C Υ / ρ , and by scaling, Lemma 4.4 gives
∫ M b ∇ τ w b ⋅ Υ ~ b d σ = ρ 2 ∫ M ∇ τ V U ( b ) ⋅ Υ ~ j ( b ) d H 2 ≥ γ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ . \int_{\mathcal{M}_b}\nabla_\tau w_b\cdot\widetilde{\Upsilon}_b\,d\sigma
=\rho^2\int_{\mathcal{M}}\nabla_\tau V^{U(b)}\cdot\widetilde{\Upsilon}_{j(b)}\,d\mathcal{H}^2
\ge\gamma|U(b)|\rho^2|Y|. ∫ M b ∇ τ w b ⋅ Υ b d σ = ρ 2 ∫ M ∇ τ V U ( b ) ⋅ Υ j ( b ) d H 2 ≥ γ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣.
On M b \mathcal{M}_b M b we write ∇ τ u ρ = ∇ τ w b + ∇ τ ( u + X ρ − w b ) + ∇ τ e \nabla_\tau u_\rho=\nabla_\tau w_b+\nabla_\tau(u+X_\rho-w_b)+\nabla_\tau e ∇ τ u ρ = ∇ τ w b + ∇ τ ( u + X ρ − w b ) + ∇ τ e , and we get
∫ M b ∣ ∇ τ u ρ ∣ ≥ ∫ M b ∇ τ u ρ ⋅ Υ ~ b ≥ γ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ ∣ M b ∣ − ∣ ∫ M b ∇ τ e ⋅ Υ ~ b ∣ , \int_{\mathcal{M}_b}|\nabla_\tau u_\rho|
\ge\int_{\mathcal{M}_b}\nabla_\tau u_\rho\cdot\widetilde{\Upsilon}_b
\ge\gamma|U(b)|\rho^2|Y|-C\rho|\mathcal{M}_b|-\left|\int_{\mathcal{M}_b}\nabla_\tau e\cdot\widetilde{\Upsilon}_b\right|, ∫ M b ∣ ∇ τ u ρ ∣ ≥ ∫ M b ∇ τ u ρ ⋅ Υ b ≥ γ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ ∣ M b ∣ − ∫ M b ∇ τ e ⋅ Υ b ,
by (49) . To estimate the last term, we integrate by parts on each square of M b \mathcal{M}_b M b ; there are no boundary terms, because Υ ~ b \widetilde{\Upsilon}_b Υ b is compactly supported in the interior of the squares, and ∫ M b div τ Υ ~ b = 0 \int_{\mathcal{M}_b}\operatorname{div}_\tau\widetilde{\Upsilon}_b=0 ∫ M b div τ Υ b = 0 . By (44) ,
∣ ∫ M b ∇ τ e ⋅ Υ ~ b ∣ = ∣ ∫ M b ( e − m N b e ) div τ Υ ~ b ∣ ≲ ρ − 1 ∣ M b ∣ 1 / 2 ρ ∥ ∇ e ∥ L 2 ( N b ) ≲ ρ ∥ ∇ e ∥ L 2 ( N b ) . \left|\int_{\mathcal{M}_b}\nabla_\tau e\cdot\widetilde{\Upsilon}_b\right|
=\left|\int_{\mathcal{M}_b}(e-m_{N_b}e)\operatorname{div}_\tau\widetilde{\Upsilon}_b\right|
\lesssim\rho^{-1}|\mathcal{M}_b|^{1/2}\sqrt{\rho}\|\nabla e\|_{L^2(N_b)}
\lesssim\sqrt{\rho}\|\nabla e\|_{L^2(N_b)}. ∫ M b ∇ τ e ⋅ Υ b = ∫ M b ( e − m N b e ) div τ Υ b ≲ ρ − 1 ∣ M b ∣ 1/2 ρ ∥∇ e ∥ L 2 ( N b ) ≲ ρ ∥∇ e ∥ L 2 ( N b ) .
Here we used that the functions u ρ u_\rho u ρ , u u u and X ρ X_\rho X ρ are smooth up to the walls of the columns near M b \mathcal{M}_b M b : u ρ u_\rho u ρ by Lemma 2.2 , u u u because these walls lie in the interior of D D D , and X ρ X_\rho X ρ by Lemma 4.1 . Notice that this estimate only uses the trace and Poincaré inequalities; no homogeneous Neumann condition for e e e is needed. We sum over b ∈ J ρ b \in\mathcal{J}_\rho b ∈ J ρ , recalling that M b ⊂ M ρ \mathcal{M}_b \subset\mathcal{M}_\rho M b ⊂ M ρ , that the sets N b N_b N b are disjoint, and that # J ρ ≲ ρ − 2 \#\mathcal{J}_\rho\lesssim\rho^{-2} # J ρ ≲ ρ − 2 . Then, the Cauchy-Schwarz inequality yields
∫ M ρ ∣ ∇ τ u ρ ∣ ≥ ∑ b ∈ J ρ ∫ M b ∣ ∇ τ u ρ ∣ ≥ γ ∑ b ∈ J ρ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ ∑ b ∈ J ρ ∣ M b ∣ − C ρ ∑ b ∈ J ρ ∥ ∇ e ∥ L 2 ( N b ) ≥ γ ∑ b ∈ J ρ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ − C ρ ( # J ρ ) 1 / 2 ∥ ∇ e ∥ L 2 ( D ρ ) ≥ γ ∑ b ∈ J ρ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ − C ρ − 1 / 2 ∥ ∇ e ∥ L 2 ( D ρ ) . \begin{aligned}
\int_{\mathcal{M}_\rho} |\nabla_\tau u_\rho| &\ge\sum_{b\in\mathcal{J}_\rho} \int_{\mathcal{M}_b} |\nabla_\tau u_\rho| \\
&\ge\gamma\sum_{b\in\mathcal{J}_\rho} |U(b)|\rho^2|Y|-C\rho\sum_{b\in\mathcal{J}_\rho}|\mathcal{M}_b|-C\sqrt{\rho}\sum_{b\in\mathcal{J}_\rho}\|\nabla e\|_{L^2(N_b)} \\
&\ge\gamma\sum_{b\in\mathcal{J}_\rho}|U(b)|\rho^2|Y|-C\rho-C\sqrt{\rho}(\#\mathcal{J}_\rho)^{1/2}\|\nabla e\|_{L^2(D_\rho)} \\
&\ge\gamma\sum_{b\in\mathcal{J}_\rho}|U(b)|\rho^2|Y|-C\rho-C\rho^{-1/2}\|\nabla e\|_{L^2(D_\rho)}.
\end{aligned} ∫ M ρ ∣ ∇ τ u ρ ∣ ≥ b ∈ J ρ ∑ ∫ M b ∣ ∇ τ u ρ ∣ ≥ γ b ∈ J ρ ∑ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ b ∈ J ρ ∑ ∣ M b ∣ − C ρ b ∈ J ρ ∑ ∥∇ e ∥ L 2 ( N b ) ≥ γ b ∈ J ρ ∑ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ − C ρ ( # J ρ ) 1/2 ∥∇ e ∥ L 2 ( D ρ ) ≥ γ b ∈ J ρ ∑ ∣ U ( b ) ∣ ρ 2 ∣ Y ∣ − C ρ − C ρ − 1/2 ∥∇ e ∥ L 2 ( D ρ ) .
where in the third line we used that ∑ b ∣ M b ∣ ≤ ∣ M ρ ∣ ≲ 1 \sum_b|\mathcal{M}_b|\le|\mathcal{M}_\rho|\lesssim1 ∑ b ∣ M b ∣ ≤ ∣ M ρ ∣ ≲ 1 . The cells b + ρ Y b+\rho Y b + ρ Y , b ∈ J ρ b\in\mathcal{J}_\rho b ∈ J ρ , are disjoint, they have area ρ 2 ∣ Y ∣ \rho^2|Y| ρ 2 ∣ Y ∣ , and they cover Q Q Q up to a set of area ≲ d + ρ \lesssim d+\rho ≲ d + ρ . Since ∣ U ∣ |U| ∣ U ∣ is bounded and Lipschitz on Q Q Q , by (39) , the last sum is a Riemann sum which differs from ∫ Q ∣ U ∣ = ∫ Q ∣ ∇ τ u ∣ \int_Q|U|=\int_Q|\nabla_\tau u| ∫ Q ∣ U ∣ = ∫ Q ∣ ∇ τ u ∣ by at most C ( ρ + d ) C(\rho+d) C ( ρ + d ) . Therefore, by Lemma 6.4 ,
lim inf ρ → 0 ∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≥ γ ∫ Q ∣ ∇ τ u ∣ d σ − C d , \liminf_{\rho\to0}\int_{\mathcal{M}_\rho}|\nabla_\tau u_\rho|\,d\sigma\ge\gamma\int_Q|\nabla_\tau u|\,d\sigma-C\sqrt{d}, ρ → 0 lim inf ∫ M ρ ∣ ∇ τ u ρ ∣ d σ ≥ γ ∫ Q ∣ ∇ τ u ∣ d σ − C d ,
where C C C does not depend on d d d (recall that d < 1 d<1 d < 1 ). The left-hand side does not depend on d d d , because neither u ρ u_\rho u ρ nor M ρ \mathcal{M}_\rho M ρ do. So, letting d → 0 d\to0 d → 0 , we obtain (42) .
When the columns are inserted on finitely many squares Q i Q_i Q i , the argument is the same, with X ρ X_\rho X ρ replaced by the sum of the corresponding correctors, which have disjoint supports if ρ \rho ρ is small. Indeed, the residual identities of the different squares add up to a single identity for e e e , the sets N b N_b N b associated with different squares are disjoint, and all the estimates hold with constants depending on the finite family. So it is enough to use the same d d d for all the squares, and to take ρ \rho ρ smaller than the thresholds required by each of them. □
Remark 6.6. In short, Lemma 6.1 tells us that, after one insertion, the integral of ∣ ∇ τ u ρ ∣ |\nabla_\tau u_\rho| ∣ ∇ τ u ρ ∣ over the new selected squares is, in the limit ρ → 0 \rho\to0 ρ → 0 , at least γ \gamma γ times the integral of ∣ ∇ τ u ∣ |\nabla_\tau u| ∣ ∇ τ u ∣ over the old square, while the energy only changes by O ( ρ ) O(\sqrt{\rho}) O ( ρ ) . We stress that the lemma is purely qualitative in ρ \rho ρ : the speed of convergence in (42) depends on the old solution u u u , and so on all the previous steps of the construction. For this reason, the scales of the successive insertions have to be chosen one after the other.
The finite construction
In this section we iterate the insertion lemma and we prove Theorem 1.3 in R 3 \mathbb{R}^3 R 3 . All the domains below are of the form D = G ( Q 0 ) D=G(Q_0) D = G ( Q 0 ) , where G G G is obtained from the identity by finitely many insertions as in § 6 , so that Lip ( G ) , Lip ( G − 1 ) ≤ L F \operatorname{Lip}(G),\operatorname{Lip}(G^{-1})\le L_F Lip ( G ) , Lip ( G − 1 ) ≤ L F by Lemma 5.5. When we insert columns in such a domain, the datum is kept on the unchanged part of the boundary and it is set equal to zero on the new faces. Since the datum always vanishes near the squares where columns are inserted, it keeps zero mean and the same L p L^p L p norms.
Proposition 7.1 (Amplification). Let D = G ( Q 0 ) D=G(Q_0) D = G ( Q 0 ) , where G G G is obtained from the identity by finitely many insertions as in § 6 , so that Lip ( G ) , Lip ( G − 1 ) ≤ L F \operatorname{Lip}(G),\operatorname{Lip}(G^{-1})\le L_F Lip ( G ) , Lip ( G − 1 ) ≤ L F , let u u u be a nonconstant variational solution
in D D D with datum g g g , and let P ⊂ ∂ D P \subset\partial D P ⊂ ∂ D be a flat closed square with a connected flat neighbourhood V P V_P V P , such that g = 0 g=0 g = 0 on ∂ D ∩ V P \partial D \cap V_P ∂ D ∩ V P and G G G is a rigid motion in G − 1 ( V P ) G^{-1}(V_P) G − 1 ( V P ) . Then, for every M > 0 M>0 M > 0 and ε > 0 \varepsilon>0 ε > 0 , one can perform finitely many successive insertions of columns, all supported in a compact subset of V P V_P V P , with the following properties: the resulting domain D ′ = G ′ ( Q 0 ) ⊂ D D'=G'(Q_0)\subset D D ′ = G ′ ( Q 0 ) ⊂ D satisfies Lip ( G ′ ) \operatorname{Lip}(G') Lip ( G ′ ) , Lip ( G ′ − 1 ) ≤ L F \operatorname{Lip}(G'^{-1})\le L_F Lip ( G ′ − 1 ) ≤ L F and G ′ = G G'=G G ′ = G outside G − 1 ( V P ) G^{-1}(V_P) G − 1 ( V P ) , and there is a finite union S S S of flat selected squares contained in V P V_P V P such that the solution u ′ u' u ′ in D ′ D' D ′ satisfies
∫ S ∣ ∇ τ u ′ ∣ d σ ≥ M and ∥ ∇ ( u ′ − u ) ∥ L 2 ( D ′ ) ≤ ε . (50) \int_S \lvert\nabla_\tau u' \rvert\,d\sigma\ge M
\qquad\text{and}\qquad
\lVert\nabla(u'-u)\rVert_{L^2(D')} \le\varepsilon. \tag*{(50)} ∫ S ∣ ∇ τ u ′ ∣ d σ ≥ M and ∥ ∇ ( u ′ − u ) ∥ L 2 ( D ′ ) ≤ ε . (50)
The datum of u ′ u' u ′ is equal to g g g on ∂ D ∩ ∂ D ′ \partial D\cap\partial D' ∂ D ∩ ∂ D ′ , and to 0 0 0 on ∂ D ′ ∖ ∂ D \partial D'\setminus\partial D ∂ D ′ ∖ ∂ D .
Proof. Let us fix 1 < β < γ 1<\beta<\gamma 1 < β < γ . By Lemma 2.3, applied to a small square centred at an interior point of P P P , we have I 0 : = ∫ P ∣ ∇ τ u ∣ > 0 I_0:=\int_P\lvert\nabla_\tau u\rvert>0 I 0 := ∫ P ∣ ∇ τ u ∣ > 0 . Let k ≥ 1 k\ge1 k ≥ 1 be such that β k I 0 ≥ M \beta^k I_0\ge M β k I 0 ≥ M . We construct domains D 0 = D ⊃ D 1 ⊃ ⋯ ⊃ D k = D ′ D_0=D\supset D_1\supset\cdots\supset D_k=D' D 0 = D ⊃ D 1 ⊃ ⋯ ⊃ D k = D ′ , with solutions u 0 = u , u 1 , … , u k = u ′ u_0=u,u_1,\ldots,u_k=u' u 0 = u , u 1 , … , u k = u ′ and families of flat selected squares C 0 = { P } , C 1 , … , C k \mathcal{C}_0=\{P\},\mathcal{C}_1,\ldots,\mathcal{C}_k C 0 = { P } , C 1 , … , C k , such that
I j + 1 : = ∫ C j + 1 ∣ ∇ τ u j + 1 ∣ ≥ β I j and ∥ ∇ ( u j + 1 − u j ) ∥ L 2 ( D j + 1 ) ≤ ε k , I_{j+1}:=\int_{C_{j+1}}\lvert\nabla_\tau u_{j+1}\rvert\ge\beta I_j
\qquad\text{and}\qquad
\lVert\nabla(u_{j+1}-u_j)\rVert_{L^2(D_{j+1})}\le\frac{\varepsilon}{k}, I j + 1 := ∫ C j + 1 ∣ ∇ τ u j + 1 ∣ ≥ β I j and ∥ ∇ ( u j + 1 − u j ) ∥ L 2 ( D j + 1 ) ≤ k ε ,
where C j = ⋃ Q ∈ C j Q C_j=\bigcup_{Q\in\mathcal{C}_j}Q C j = ⋃ Q ∈ C j Q . Suppose that D j D_j D j has already been constructed. Then we insert columns of scale ρ j \rho_j ρ j on all the squares of C j \mathcal{C}_j C j . By Lemma 5.4 (or by assumption, if j = 0 j=0 j = 0 ), these squares are flat and have pairwise disjoint flat neighbourhoods, and they lie on faces where the datum vanishes. By Lemma 6.1, the integral of the tangential gradient of the new solution over the new selected squares has lower limit at least γ I j > β I j \gamma I_j>\beta I_j γ I j > β I j as ρ j → 0 \rho_j\to0 ρ j → 0 , and the energy change tends to zero. So both requirements are satisfied if we choose ρ j \rho_j ρ j small enough. By Lemma 5.5, applied with G 0 = G G_0=G G 0 = G and C 0 = { P } \mathcal{C}_0=\{P\} C 0 = { P } , the hypotheses of Lemma 6.1 are satisfied at every step, all the maps have bi-Lipschitz constants at most L F L_F L F , and all the insertions are supported in a compact subset of V P V_P V P if the scales are small enough.
After k k k steps we take S : = C k S:=C_k S := C k , so that ∫ S ∣ ∇ τ u ′ ∣ = I k ≥ β k I 0 ≥ M \int_S\lvert\nabla_\tau u'\rvert=I_k\ge\beta^k I_0\ge M ∫ S ∣ ∇ τ u ′ ∣ = I k ≥ β k I 0 ≥ M . Moreover, since the domains decrease,
∥ ∇ ( u ′ − u ) ∥ L 2 ( D ′ ) ≤ ∑ j = 0 k − 1 ∥ ∇ ( u j + 1 − u j ) ∥ L 2 ( D j + 1 ) ≤ ε . \lVert\nabla(u'-u)\rVert_{L^2(D')}
\le\sum_{j=0}^{k-1}\lVert\nabla(u_{j+1}-u_j)\rVert_{L^2(D_{j+1})}
\le\varepsilon. ∥ ∇ ( u ′ − u ) ∥ L 2 ( D ′ ) ≤ j = 0 ∑ k − 1 ∥ ∇ ( u j + 1 − u j ) ∥ L 2 ( D j + 1 ) ≤ ε .
The iteration in the proof of Proposition 7.1 can be seen as a reiterated homogenization procedure. We remark that each scale is chosen after the previous ones, and it depends on the solution obtained at the previous step, see Remark 6.6.
Proof of Theorem 1.3 for n = 2 n=2 n = 2 . Let Q 0 = ( − 1 , 1 ) 3 Q_0=(-1,1)^3 Q 0 = ( − 1 , 1 ) 3 , and let g 0 g_0 g 0 be equal to 1 1 1 on the face { x 1 = 1 } \{x_1=1\} { x 1 = 1 } , to − 1 -1 − 1 on the face { x 1 = − 1 } \{x_1=-1\} { x 1 = − 1 } , and to 0 0 0 on the other faces. Then u 0 ( x ) = x 1 u_0(x)=x_1 u 0 ( x ) = x 1 is the variational solution, and ∣ ∇ τ u 0 ∣ = 1 \lvert\nabla_\tau u_0\rvert=1 ∣ ∇ τ u 0 ∣ = 1 on the bottom face { x 3 = − 1 } \{x_3=-1\} { x 3 = − 1 } . The square C 0 = [ − 1 2 , 1 2 ] 2 × { − 1 } C_0=[-\frac12,\frac12]^2\times\{-1\} C 0 = [ − 2 1 , 2 1 ] 2 × { − 1 } is flat and ∫ C 0 ∣ ∇ τ u 0 ∣ = 1 \int_{C_0}\lvert\nabla_\tau u_0\rvert=1 ∫ C 0 ∣ ∇ τ u 0 ∣ = 1 . Its flat neighbourhood ( − 3 4 , 3 4 ) 2 × ( − 2 , 0 ) (-\frac34,\frac34)^2\times(-2,0) ( − 4 3 , 4 3 ) 2 × ( − 2 , 0 ) does not meet the faces where g 0 ≠ 0 g_0\ne0 g 0 = 0 .
Now let M > 0 M>0 M > 0 . We apply Proposition 7.1 with P = C 0 P=C_0 P = C 0 and any ε > 0 \varepsilon>0 ε > 0 , and we get a domain Ω M = G M ( Q 0 ) \Omega_M=G_M(Q_0) Ω M = G M ( Q 0 ) with Lip ( G M ) , Lip ( G M − 1 ) ≤ L F \operatorname{Lip}(G_M),\operatorname{Lip}(G_M^{-1})\le L_F Lip ( G M ) , Lip ( G M − 1 ) ≤ L F (this will give Theorem 1.3 with L 0 : = L F L_0:=L_F L 0 := L F ). Its datum g M g_M g M is ± 1 \pm1 ± 1 on the two unchanged faces in { x 1 = ± 1 } \{x_1=\pm1\} { x 1 = ± 1 } and 0 0 0 elsewhere, and we also get a finite union S S S of flat selected squares such that ∫ S ∣ ∇ τ u M ∣ ≥ M \int_S\lvert\nabla_\tau u_M\rvert\ge M ∫ S ∣ ∇ τ u M ∣ ≥ M . Every point of S S S has a neighbourhood in which Ω M \Omega_M Ω M is a half-space and the datum vanishes. By Lemma 2.2, N ( ∇ u M ) ≥ ∣ ∇ τ u M ∣ \mathcal{N}(\nabla u_M)\ge\lvert\nabla_\tau u_M\rvert N ( ∇ u M ) ≥ ∣ ∇ τ u M ∣ on S S S , and therefore
∥ N ( ∇ u M ) ∥ L 1 ( σ ) ≥ ∫ S ∣ ∇ τ u M ∣ d σ ≥ M . \lVert\mathcal{N}(\nabla u_M)\rVert_{L^1(\sigma)}
\ge\int_S\lvert\nabla_\tau u_M\rvert\,d\sigma
\ge M. ∥ N ( ∇ u M ) ∥ L 1 ( σ ) ≥ ∫ S ∣ ∇ τ u M ∣ d σ ≥ M .
Since σ ( ∂ Ω M ) ≤ L F 2 σ ( ∂ Q 0 ) \sigma(\partial\Omega_M) \le L_F^2 \sigma(\partial Q_0) σ ( ∂ Ω M ) ≤ L F 2 σ ( ∂ Q 0 ) and ∥ g M ∥ L p ( σ ) = 8 1 / p \lVert g_M\rVert_{L^p(\sigma)} = 8^{1/p} ∥ g M ∥ L p ( σ ) = 8 1/ p , Hölder’s inequality gives the L p L^p L p estimate in Theorem 1.3 . The facts that Ω M \Omega_M Ω M is a Lipschitz domain and that its boundary consists of finitely many rectangles parallel to the coordinate planes are proved in Appendix B. □
Remark 7.2. The proof gives more precise information. After k k k generations, the selected squares satisfy H 2 ( C k ) ≤ A k \mathcal{H}^2(C_k) \le A^k H 2 ( C k ) ≤ A k , by Lemma 5.5, and ∫ C k ∣ ∇ τ u k ∣ ≥ β k \int_{C_k} |\nabla_\tau u_k| \ge\beta^k ∫ C k ∣ ∇ τ u k ∣ ≥ β k . Thus the mean value of ∣ ∇ τ u k ∣ |\nabla_\tau u_k| ∣ ∇ τ u k ∣ on C k C_k C k is at least ( β / A ) k (\beta/A)^k ( β / A ) k , and Hölder’s inequality gives
∥ N ( ∇ u k ) ∥ L p ( σ ) ∥ g k ∥ L p ( σ ) ≥ 8 − 1 / p β k A − k / p ′ \frac{\lVert\mathcal{N}(\nabla u_k)\rVert_{L^p(\sigma)}}{\lVert g_k\rVert_{L^p(\sigma)}} \ge8^{-1/p}\beta^k A^{-k/p'} ∥ g k ∥ L p ( σ ) ∥ N ( ∇ u k ) ∥ L p ( σ ) ≥ 8 − 1/ p β k A − k / p ′
( 1 ≤ p < ∞ ) (1 \le p < \infty) ( 1 ≤ p < ∞ ) .
For large p p p the growth comes mostly from the concentration of the selected squares, and for p = 1 p = 1 p = 1 it comes only from the flow.
The proof of Theorem 1.2 in R 3 \mathbb{R}^3 R 3
The idea of the proof of Theorem 1.2 for n = 2 n = 2 n = 2 is to perform infinitely many finite constructions, in disjoint balls accumulating at one point of the bottom face of the cube, see Figure 4 . Each construction changes the solution everywhere, but it only produces a small perturbation in the energy norm. Near the selected squares of the previous constructions, such a perturbation changes the gradient on the boundary only slightly, by Lemma 2.2 . Thus, if the size of each perturbation is chosen after the previous constructions, the lower bounds obtained in each ball survive in the limit.
Proof of Theorem 1.2 for $n = 2. Let Q 0 Q_0 Q 0 , g 0 g_0 g 0 and u 0 = x 1 u_0 = x_1 u 0 = x 1 be as in the proof of Theorem 1.3 ; thus g 0 = ± 1 g_0 = \pm1 g 0 = ± 1 on the faces { x 1 = ± 1 } \{x_1 = \pm1\} { x 1 = ± 1 } and g 0 = 0 g_0 = 0 g 0 = 0 on the other faces of Q 0 Q_0 Q 0 . For j ≥ 1 j \ge1 j ≥ 1 , let B j : = B ( c j , r j ) B_j := B(c_j,r_j) B j := B ( c j , r j ) , with
c j : = ( 2 − j , 0 , − 1 ) and r j : = 2 − 3 j − 10 . c_j := (2^{-j},0,-1) \qquad\text{and}\qquad r_j := 2^{-3j-10}. c j := ( 2 − j , 0 , − 1 ) and r j := 2 − 3 j − 10 .
The closed balls B j ‾ \overline{B_j} B j are centred on the bottom face of Q 0 Q_0 Q 0 and do not meet the other faces. They are pairwise disjoint, because ∣ c j − c k ∣ ≥ 2 − j − 1 > r j + r k |c_j-c_k| \ge2^{-j-1} > r_j+r_k ∣ c j − c k ∣ ≥ 2 − j − 1 > r j + r k for k > j k>j k > j , and they only accumulate at x ∗ : = ( 0 , 0 , − 1 ) x_* := (0,0,-1) x ∗ := ( 0 , 0 , − 1 ) . We also let B ∗ : = B ( 0 , 1 / 2 ) B_* := B(0,1/2) B ∗ := B ( 0 , 1/2 ) , which is contained in Q 0 Q_0 Q 0 and disjoint from all the B j B_j B j .
The inductive construction. We are going to construct, by induction, domains Ω 0 = Q 0 ⊃ Ω 1 ⊃ Ω 2 ⊃ ⋯ \Omega_0 = Q_0 \supset\Omega_1 \supset\Omega_2 \supset\cdots Ω 0 = Q 0 ⊃ Ω 1 ⊃ Ω 2 ⊃ ⋯ of the form Ω j = G j ( Q 0 ) \Omega_j = G_j(Q_0) Ω j = G j ( Q 0 ) , the variational solutions u j u_j u j in Ω j \Omega_j Ω j with datum g 0 g_0 g 0 on the faces { x 1 = ± 1 } \{x_1 = \pm1\} { x 1 = ± 1 } and zero elsewhere, and finite unions S j ⊂ ∂ Ω j ∩ B j S_j \subset\partial\Omega_j \cap B_j S j ⊂ ∂ Ω j ∩ B j of flat squares. Assume that Ω j − 1 \Omega_{j-1} Ω j − 1 is already constructed and that it coincides with Q 0 Q_0 Q 0 in B j B_j B j . Note that u j − 1 u_{j-1} u j − 1 is not constant, since testing (1.3) with the function x 1 x_1 x 1 we get ∫ ∂ 1 u j − 1 = 8 \int\partial_1u_{j-1} = 8 ∫ ∂ 1 u j − 1 = 8 . We let Σ j \Sigma_j Σ j be the closed square of side r j ′ / 2 r'_j/2 r j ′ /2 in the bottom face centred at c j c_j c j . This square is flat, with flat neighbourhood B j B_j B j , and G j − 1 = id G_{j-1} = \operatorname{id} G j − 1 = id in B j B_j B j . Then we apply Proposition 7.1 with P = Σ j P = \Sigma_j P = Σ j , M = 2 M = 2 M = 2 and ε = ε j \varepsilon= \varepsilon_j ε = ε j , where
0 < ε j ≤ 2 − j 1 + max 1 ≤ i < j κ i σ ( S i ) . (51) 0 < \varepsilon_j \le\frac{2^{-j}}{1+\max_{1\le i<j}\kappa_i \sigma(S_i)}. \tag*{(51)} 0 < ε j ≤ 1 + max 1 ≤ i < j κ i σ ( S i ) 2 − j . (51)
and each constant κ i \kappa_i κ i is defined in (8.3) below, as soon as S i S_i S i has been constructed (for j = 1 j = 1 j = 1 , the maximum over the empty set is 0). We obtain a map H j H_j H j , which is the identity outside a compact subset of B j B_j B j and satisfies Lip ( H j ) , Lip ( H j − 1 ) ≤ L F \operatorname{Lip}(H_j), \operatorname{Lip}(H_j^{-1}) \le L_F Lip ( H j ) , Lip ( H j − 1 ) ≤ L F (indeed, H j H_j H j is the composition of the insertion maps performed in B j B_j B j ; since Ω j − 1 ∩ B j = Q 0 ∩ B j \Omega_{j-1} \cap B_j = Q_0 \cap B_j Ω j − 1 ∩ B j = Q 0 ∩ B j , these are also the insertion maps of a construction starting from Q 0 Q_0 Q 0 with C 0 = { Σ j } C_0 = \{\Sigma_j\} C 0 = { Σ j } , and Lemma 5.5(a), applied with G 0 = id G_0 = \operatorname{id} G 0 = id , bounds their composition), the domain Ω j = H j ( Ω j − 1 ) \Omega_j = H_j(\Omega_{j-1}) Ω j = H j ( Ω j − 1 ) , with G j : = H j ∘ G j − 1 G_j := H_j \circ G_{j-1} G j := H j ∘ G j − 1 , and a finite union S j S_j S j of flat squares in B j B_j B j , such that
∫ S j ∣ ∇ τ u j ∣ d σ ≥ 2 and ∥ ∇ ( u j − u j − 1 ) ∥ L 2 ( Ω j ) ≤ ε j . (52) \int_{S_j} \lvert\nabla_\tau u_j\rvert\,d\sigma\ge2
\qquad\text{and}\qquad
\lVert\nabla(u_j-u_{j-1})\rVert_{L^2(\Omega_j)} \le\varepsilon_j. \tag*{(52)} ∫ S j ∣ ∇ τ u j ∣ d σ ≥ 2 and ∥ ∇ ( u j − u j − 1 ) ∥ L 2 ( Ω j ) ≤ ε j . (52)
Finally, we choose r j ′ > 0 r'_j > 0 r j ′ > 0 such that, for every x ∈ S j x \in S_j x ∈ S j , the domain Ω j \Omega_j Ω j coincides with a half-space in B ( x , 2 r j ′ ) ⊂ B j B(x,2r'_j) \subset B_j B ( x , 2 r j ′ ) ⊂ B j and the datum vanishes there. Since the later constructions do not touch B j B_j B j , the same holds for all the domains Ω k \Omega_k Ω k , k ≥ j k \ge j k ≥ j . We set κ j : = C 2 ( 2 r j ′ ) − 3 / 2 \kappa_j := C_2(2r'_j)^{-3/2} κ j := C 2 ( 2 r j ′ ) − 3/2 , where C 2 C_2 C 2 is the constant in (2.4) for n = 2 n = 2 n = 2 , so that (2.4), applied with r = 2 r j ′ r = 2r'_j r = 2 r j ′ in the balls B ( x , 2 r j ′ ) B(x,2r'_j) B ( x , 2 r j ′ ) , x ∈ S j x \in S_j x ∈ S j , gives
sup S j ∣ ∇ v − ∇ w ∣ ≤ κ j ∥ ∇ ( v − w ) ∥ L 2 ( Ω j ∩ B j ) (53) \sup_{S_j}\lvert\nabla v-\nabla w\rvert\le\kappa_j \lVert\nabla(v-w)\rVert_{L^2(\Omega_j\cap B_j)} \tag*{(53)} S j sup ∣ ∇ v − ∇ w ∣ ≤ κ j ∥ ∇ ( v − w ) ∥ L 2 ( Ω j ∩ B j ) (53)
whenever v v v and w w w are variational solutions in domains which coincide with Ω j \Omega_j Ω j in B j B_j B j , with data vanishing in the balls B ( x , 2 r j ′ ) B(x,2r'_j) B ( x , 2 r j ′ ) , x ∈ S j x \in S_j x ∈ S j . This completes the inductive step. Observe that ε j \varepsilon_j ε j depends only on the steps 1 , … , j − 1 1,\ldots,j-1 1 , … , j − 1 .
The limit domain. Let G : = H j G := H_j G := H j in B j B_j B j for every j j j , and G : = id G := \operatorname{id} G := id outside ⋃ j B j \bigcup_j B_j ⋃ j B j . The map G G G is a bijection, because each H j H_j H j maps B j B_j B j onto itself. It is also continuous at x ∗ x_* x ∗ , since ∣ H j ( x ) − x ∣ ≤ 2 r j → 0 \lvert H_j(x)-x\rvert\le2r_j \to0 ∣ H j ( x ) − x ∣ ≤ 2 r j → 0 . The same argument as in the proof of Lemma 5.4 , with balls in place of boxes, shows that Lip ( G ) , Lip ( G − 1 ) ≤ L F \operatorname{Lip}(G), \operatorname{Lip}(G^{-1}) \le L_F Lip ( G ) , Lip ( G − 1 ) ≤ L F . Here the inverse of G G G is obtained by patching together the maps H j − 1 H_j^{-1} H j − 1 . Notice that the segment [ x , x ′ ] [x,x'] [ x , x ′ ] may meet infinitely many balls, but this is not a problem, since only the balls which contain x x x or x ′ x' x ′ appear in the estimate. We stress that this uniform bound does not come from the smallness of the insertion scales, but from the rigidity of the construction: in each ball B j B_j B j it is given by Lemma 5.5, and the balls are pairwise disjoint. Then we set Ω : = G ( Q 0 ) \Omega:= G(Q_0) Ω := G ( Q 0 ) . In a ball B i B_i B i , the domains Ω j \Omega_j Ω j with j < i j < i j < i coincide with Q 0 Q_0 Q 0 , whereas Ω \Omega Ω and the domains Ω j \Omega_j Ω j with j ≥ i j \ge i j ≥ i coincide with H i ( Q 0 ) H_i(Q_0) H i ( Q 0 ) . Outside ⋃ i B i \bigcup_i B_i ⋃ i B i , all these domains coincide with Q 0 Q_0 Q 0 . Checking separately the points in each B j B_j B j and the points outside all of them (including the point x ∗ x_* x ∗ , which lies on ∂ Q 0 \partial Q_0 ∂ Q 0 and is fixed by G G G ), we see that
Ω = ⋂ j Ω j and ∣ Ω j ∖ Ω ∣ ≤ ∑ i > j ∣ B i ∣ → j → ∞ 0. \Omega= \bigcap_j \Omega_j
\qquad\text{and}\qquad
\lvert\Omega_j \setminus\Omega\rvert\le\sum_{i>j}\lvert B_i\rvert\xrightarrow[j\to\infty]{} 0. Ω = j ⋂ Ω j and ∣ Ω j ∖ Ω ∣ ≤ i > j ∑ ∣ B i ∣ j → ∞ 0.
The limit solution. We normalize the functions u j u_j u j so that m B ∗ u j = 0 m_{B_*}u_j=0 m B ∗ u j = 0 . By Lemma 2.1 , ∥ ∇ u j ∥ L 2 ( Ω j ) ≲ 1 \|\nabla u_j\|_{L^2(\Omega_j)}\lesssim1 ∥∇ u j ∥ L 2 ( Ω j ) ≲ 1 , uniformly in j j j . By (52) , for k > j k>j k > j ,
∥ ∇ ( u k − u j ) ∥ L 2 ( Ω ) ≤ ∑ l = j + 1 k ε l ≤ 2 − j . \|\nabla(u_k-u_j)\|_{L^2(\Omega)}\leq\sum_{l=j+1}^{k}\varepsilon_l\leq2^{-j}. ∥∇ ( u k − u j ) ∥ L 2 ( Ω ) ≤ l = j + 1 ∑ k ε l ≤ 2 − j .
By the Poincaré inequality in Ω \Omega Ω (Lemma 2.1 ) and the normalization on B ∗ ⊂ Ω B_*\subset\Omega B ∗ ⊂ Ω , the restrictions u j ∣ Ω u_j|_\Omega u j ∣ Ω form a Cauchy sequence in W 1 , 2 ( Ω ) W^{1,2}(\Omega) W 1 , 2 ( Ω ) , and we call u u u its limit. Indeed, let h ∈ W 1 , 2 ( Ω ) h\in W^{1,2}(\Omega) h ∈ W 1 , 2 ( Ω ) with m B ∗ h = 0 m_{B_*}h=0 m B ∗ h = 0 , and let c = m ∂ Ω Tr h c=m_{\partial\Omega}\operatorname{Tr}h c = m ∂ Ω Tr h . Then ∣ c ∣ = ∣ m B ∗ ( h − c ) ∣ ≤ ∣ B ∗ ∣ − 1 / 2 ∥ h − c ∥ L 2 ( Ω ) |c|=|m_{B_*}(h-c)|\leq|B_*|^{-1/2}\|h-c\|_{L^2(\Omega)} ∣ c ∣ = ∣ m B ∗ ( h − c ) ∣ ≤ ∣ B ∗ ∣ − 1/2 ∥ h − c ∥ L 2 ( Ω ) , and so (8) gives ∥ h ∥ W 1 , 2 ( Ω ) ≲ ∥ ∇ h ∥ L 2 ( Ω ) \|h\|_{W^{1,2}(\Omega)}\lesssim\|\nabla h\|_{L^2(\Omega)} ∥ h ∥ W 1 , 2 ( Ω ) ≲ ∥∇ h ∥ L 2 ( Ω ) . It suffices to apply this to h = u k − u j h=u_k-u_j h = u k − u j . Now, for φ ∈ C c ∞ ( R 3 ) \varphi\in C^\infty_c(\mathbb{R}^3) φ ∈ C c ∞ ( R 3 ) , the equation for u j u_j u j reads
∫ Ω j ∇ u j ⋅ ∇ φ = ∫ { x 1 = 1 } ∩ ∂ Q 0 φ d σ − ∫ { x 1 = − 1 } ∩ ∂ Q 0 φ d σ . \int_{\Omega_j}\nabla u_j\cdot\nabla\varphi=\int_{\{x_1=1\}\cap\partial Q_0}\varphi\,d\sigma-\int_{\{x_1=-1\}\cap\partial Q_0}\varphi\,d\sigma. ∫ Ω j ∇ u j ⋅ ∇ φ = ∫ { x 1 = 1 } ∩ ∂ Q 0 φ d σ − ∫ { x 1 = − 1 } ∩ ∂ Q 0 φ d σ .
and its right-hand side is independent of j j j . Since ∣ ∫ Ω j ∖ Ω ∇ u j ⋅ ∇ φ ∣ ≲ ∥ ∇ φ ∥ ∞ ∣ Ω j ∖ Ω ∣ 1 / 2 → 0 \left|\int_{\Omega_j\setminus\Omega}\nabla u_j\cdot\nabla\varphi\right|\lesssim\|\nabla\varphi\|_\infty|\Omega_j\setminus\Omega|^{1/2}\to0 ∫ Ω j ∖ Ω ∇ u j ⋅ ∇ φ ≲ ∥∇ φ ∥ ∞ ∣ Ω j ∖ Ω ∣ 1/2 → 0 , by the uniform energy bound, and ∫ Ω ∇ u j ⋅ ∇ φ → ∫ Ω ∇ u ⋅ ∇ φ \int_\Omega\nabla u_j\cdot\nabla\varphi\to\int_\Omega\nabla u\cdot\nabla\varphi ∫ Ω ∇ u j ⋅ ∇ φ → ∫ Ω ∇ u ⋅ ∇ φ , we can pass to the limit. Then, by the density of smooth functions in W 1 , 2 ( Ω ) W^{1,2}(\Omega) W 1 , 2 ( Ω ) and the continuity of the trace (Lemma 2.1 ), u u u is the variational solution in Ω \Omega Ω with datum g g g , where g = ± 1 g=\pm1 g = ± 1 on the faces { x 1 = ± 1 } \{x_1=\pm1\} { x 1 = ± 1 } of Q 0 Q_0 Q 0 and g = 0 g=0 g = 0 on the rest of ∂ Ω \partial\Omega ∂ Ω . Obviously, ∣ g ∣ ≤ 1 |g|\leq1 ∣ g ∣ ≤ 1 and ∫ g d σ = 0 \int g\,d\sigma=0 ∫ g d σ = 0 .
The lower bound. Let i ≥ 1 i\geq1 i ≥ 1 . If we let k → ∞ k\to\infty k → ∞ in the bound for ∥ ∇ ( u k − u j ) ∥ L 2 ( Ω ) \|\nabla(u_k-u_j)\|_{L^2(\Omega)} ∥∇ ( u k − u j ) ∥ L 2 ( Ω ) obtained above, with j = i j=i j = i , we obtain ∥ ∇ ( u − u i ) ∥ L 2 ( Ω ) ≤ ∑ l > i ε l \|\nabla(u-u_i)\|_{L^2(\Omega)}\leq\sum_{l>i}\varepsilon_l ∥∇ ( u − u i ) ∥ L 2 ( Ω ) ≤ ∑ l > i ε l . In B i B_i B i the domain Ω \Omega Ω coincides with Ω i \Omega_i Ω i , and both data vanish in the balls B ( x , 2 r i ′ ) B(x,2r_i') B ( x , 2 r i ′ ) , x ∈ S i x\in S_i x ∈ S i . Hence we can apply (53) with j = i j=i j = i , v = u v=u v = u and w = u i w=u_i w = u i , and, by (51) , we get
σ ( S i ) sup S i ∣ ∇ u − ∇ u i ∣ ≤ κ i σ ( S i ) ∑ l > i ε l ≤ ∑ l > i 2 − l ≤ 1. \sigma(S_i)\sup_{S_i}|\nabla u-\nabla u_i|\leq\kappa_i\sigma(S_i)\sum_{l>i}\varepsilon_l\leq\sum_{l>i}2^{-l}\leq1. σ ( S i ) S i sup ∣∇ u − ∇ u i ∣ ≤ κ i σ ( S i ) l > i ∑ ε l ≤ l > i ∑ 2 − l ≤ 1.
So (52) yields ∫ S i ∣ ∇ τ u ∣ ≥ 2 − 1 = 1 \int_{S_i}|\nabla_\tau u|\geq2-1=1 ∫ S i ∣ ∇ τ u ∣ ≥ 2 − 1 = 1 . Since the sets S i S_i S i are pairwise disjoint and, by Lemma 2.2 , N a ( ∇ u ) ≥ ∣ ∇ τ u ∣ N_a(\nabla u)\geq|\nabla_\tau u| N a ( ∇ u ) ≥ ∣ ∇ τ u ∣ on each of them, we conclude that
∫ ∂ Ω N a ( ∇ u ) d σ ≥ ∑ i ≥ 1 ∫ S i ∣ ∇ τ u ∣ d σ = ∞ \int_{\partial\Omega}N_a(\nabla u)\,d\sigma\geq\sum_{i\geq1}\int_{S_i}|\nabla_\tau u|\,d\sigma=\infty ∫ ∂ Ω N a ( ∇ u ) d σ ≥ i ≥ 1 ∑ ∫ S i ∣ ∇ τ u ∣ d σ = ∞
for every a > 0 a>0 a > 0 . As σ ( ∂ Ω ) < ∞ \sigma(\partial\Omega)<\infty σ ( ∂ Ω ) < ∞ , it follows that ∥ N ( ∇ u ) ∥ L p ( σ ) = ∞ \|N(\nabla u)\|_{L^p(\sigma)}=\infty ∥ N ( ∇ u ) ∥ L p ( σ ) = ∞ for all p ≥ 1 p\geq1 p ≥ 1 , whereas g ∈ L ∞ ( σ ) g\in L^\infty(\sigma) g ∈ L ∞ ( σ ) . Thus ( N p ) (N_p) ( N p ) fails in Ω \Omega Ω for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ .
To conclude, note that Ω = G ( Q 0 ) = ( G ∘ B ) ( B ( 0 , 1 ) ) \Omega=G(Q_0)=(G\circ B)(B(0,1)) Ω = G ( Q 0 ) = ( G ∘ B ) ( B ( 0 , 1 )) , where B B B is the map in (7) , and so Theorem 1.2 is proved for n = 2 n=2 n = 2 .
There is a closed set K ⊂ ∂ Ω K\subset\partial\Omega K ⊂ ∂ Ω with H 2 ( K ) = 0 \mathcal{H}^2(K)=0 H 2 ( K ) = 0 such that every x ∈ ∂ Ω ∖ K x\in\partial\Omega\setminus K x ∈ ∂ Ω ∖ K has a neighbourhood in which Ω \Omega Ω coincides with a half-space bounded by a plane parallel to a coordinate plane. Indeed, let K 0 K_0 K 0 be the union of the edges of Q 0 Q_0 Q 0 , and, for j ≥ 1 j\geq1 j ≥ 1 , let K j K_j K j be the union of the edges of ∂ Ω j \partial\Omega_j ∂ Ω j contained in B j B_j B j , which is a finite union of segments. The set K : = K 0 ∪ { x ∗ } ∪ ⋃ j K j K:=K_0\cup\{x_*\}\cup\bigcup_jK_j K := K 0 ∪ { x ∗ } ∪ ⋃ j K j is closed, because a convergent sequence of points of ⋃ j K j \bigcup_jK_j ⋃ j K j with unbounded indices converges to x ∗ x_* x ∗ . It is a countable union of segments and a point, so H 2 ( K ) = 0 \mathcal{H}^2(K)=0 H 2 ( K ) = 0 . Every point of ∂ Ω ∖ K \partial\Omega\setminus K ∂ Ω ∖ K lies in the relative interior of a face of Q 0 Q_0 Q 0 or of a column, which gives the claim. The analogous property for n ≥ 3 n\geq3 n ≥ 3 is stated in Remark 9.2.
In the proof of Proposition 7.1 we may take k k k as the least integer with β k I 0 ≥ M \beta^k I_0\geq M β k I 0 ≥ M . With this choice, the number of generations used in the ball B j B_j B j tends to infinity as j → ∞ j\to\infty j → ∞ , although we do not need this fact. Indeed, if x ∈ Σ j x \in\Sigma_j x ∈ Σ j , then B ( x , r j / 4 ) ⊂ B j B(x,r_j/4) \subset B_j B ( x , r j /4 ) ⊂ B j , and (10) , applied to u j − 1 u_{j-1} u j − 1 and to v = 0 v=0 v = 0 with r = r j / 4 r=r_j/4 r = r j /4 , gives
∣ ∇ u j − 1 ( x ) ∣ ≲ r j − 3 / 2 ∥ ∇ u j − 1 ∥ L 2 ( Ω j − 1 ) ≲ r j − 3 / 2 . |\nabla u_{j-1}(x)| \lesssim r_j^{-3/2}\|\nabla u_{j-1}\|_{L^2(\Omega_{j-1})} \lesssim r_j^{-3/2}. ∣∇ u j − 1 ( x ) ∣ ≲ r j − 3/2 ∥∇ u j − 1 ∥ L 2 ( Ω j − 1 ) ≲ r j − 3/2 .
Hence ∫ Σ j ∣ ∇ τ u j − 1 ∣ d σ ≲ r j 1 / 2 → 0 \int_{\Sigma_j}|\nabla_\tau u_{j-1}|\,d\sigma\lesssim r_j^{1/2}\to0 ∫ Σ j ∣ ∇ τ u j − 1 ∣ d σ ≲ r j 1/2 → 0 .
Higher dimensions and geometric properties
Products with cubes
Let n ≥ 3 n\ge3 n ≥ 3 and d : = n − 2 d:=n-2 d := n − 2 , and write the points of R n + 1 \mathbb{R}^{n+1} R n + 1 as ( x , z ) (x,z) ( x , z ) , with x ∈ R 3 x\in\mathbb{R}^3 x ∈ R 3 and z ∈ R d z\in\mathbb{R}^d z ∈ R d . Let I d : = ( − 1 , 1 ) d I_d:=(-1,1)^d I d := ( − 1 , 1 ) d .
Proposition 9.1. Let Ω = G ( Q 0 ) ⊂ R 3 \Omega=G(Q_0)\subset\mathbb{R}^3 Ω = G ( Q 0 ) ⊂ R 3 and u u u be the domain and the solution constructed in § 9 (or in § 8 ), with datum g g g . Then Ω ♯ : = Ω × I d = G ♯ ( ( − 1 , 1 ) n + 1 ) \Omega^\sharp:=\Omega\times I_d=G^\sharp((-1,1)^{n+1}) Ω ♯ := Ω × I d = G ♯ (( − 1 , 1 ) n + 1 ) , where G ♯ ( x , z ) : = ( G ( x ) , z ) G^\sharp(x,z):=(G(x),z) G ♯ ( x , z ) := ( G ( x ) , z ) , and the function u ♯ ( x , z ) : = u ( x ) u^\sharp(x,z):=u(x) u ♯ ( x , z ) := u ( x ) is the variational solution in Ω ♯ \Omega^\sharp Ω ♯ with datum g ♯ ( x , z ) : = g ( x ) g^\sharp(x,z):=g(x) g ♯ ( x , z ) := g ( x ) on ∂ Ω × I d \partial\Omega\times I_d ∂ Ω × I d and g ♯ : = 0 g^\sharp:=0 g ♯ := 0 on the rest of ∂ Ω ♯ \partial\Omega^\sharp ∂ Ω ♯ . Moreover, ∣ g ♯ ∣ ≤ 1 |g^\sharp|\le1 ∣ g ♯ ∣ ≤ 1 , ∫ g ♯ d H n = 0 \int g^\sharp\,d\mathcal{H}^n=0 ∫ g ♯ d H n = 0 , and, for every a > 0 a>0 a > 0 and every flat selected square S ⊂ ∂ Ω S\subset\partial\Omega S ⊂ ∂ Ω ,
N a ( ∇ u ♯ ) ( x , z ) ≥ ∣ ∇ τ u ( x ) ∣ for all ( x , z ) ∈ S × I d . (54) N_a(\nabla u^\sharp)(x,z)\ge|\nabla_\tau u(x)|\qquad\text{for all }(x,z)\in S\times I_d. \tag*{(54)} N a ( ∇ u ♯ ) ( x , z ) ≥ ∣ ∇ τ u ( x ) ∣ for all ( x , z ) ∈ S × I d . (54)
Proof. The bounds Lip ( G ♯ ) ≤ max { Lip ( G ) , 1 } \operatorname{Lip}(G^\sharp)\le\max\{\operatorname{Lip}(G),1\} Lip ( G ♯ ) ≤ max { Lip ( G ) , 1 } and Lip ( ( G ♯ ) − 1 ) ≤ max { Lip ( G − 1 ) , 1 } \operatorname{Lip}((G^\sharp)^{-1})\le\max\{\operatorname{Lip}(G^{-1}),1\} Lip (( G ♯ ) − 1 ) ≤ max { Lip ( G − 1 ) , 1 } are elementary, and so G ♯ G^\sharp G ♯ is L F L_F L F -bi-Lipschitz, because L F ≥ 1 L_F\ge1 L F ≥ 1 . Next, up to a set of zero H n \mathcal{H}^n H n measure, ∂ Ω ♯ \partial\Omega^\sharp ∂ Ω ♯ is the disjoint union of ∂ Ω × I d \partial\Omega\times I_d ∂ Ω × I d and Ω × ∂ I d \Omega\times\partial I_d Ω × ∂ I d , and on these two pieces H n \mathcal{H}^n H n coincides with the product measures σ × L d \sigma\times\mathcal{L}^d σ × L d and L 3 × H d − 1 ∣ ∂ I d \mathcal{L}^3\times\mathcal{H}^{d-1}|_{\partial I_d} L 3 × H d − 1 ∣ ∂ I d , respectively. For the first piece, this is a consequence of the area formula. Indeed, if γ \gamma γ is a Lipschitz parametrization of a piece of ∂ Ω \partial\Omega ∂ Ω (for instance, the restriction of G G G to a face of Q 0 Q_0 Q 0 ), then the Gram matrix of ( y , z ) ↦ ( γ ( y ) , z ) (y,z)\mapsto(\gamma(y),z) ( y , z ) ↦ ( γ ( y ) , z ) is block diagonal, with blocks the Gram matrix of γ \gamma γ and the identity. The second piece is just a finite union of flat pieces. Thus ∫ g ♯ d H n = 2 d ∫ g d σ = 0 \int g^\sharp\,d\mathcal{H}^n=2^d\int g\,d\sigma=0 ∫ g ♯ d H n = 2 d ∫ g d σ = 0 and ∥ g ♯ ∥ L p ( H n ) = 2 d / p ∥ g ∥ L p ( σ ) \|g^\sharp\|_{L^p(\mathcal{H}^n)}=2^{d/p}\|g\|_{L^p(\sigma)} ∥ g ♯ ∥ L p ( H n ) = 2 d / p ∥ g ∥ L p ( σ ) . For φ ∈ C ∞ ( R n + 1 ) \varphi\in C^\infty(\mathbb{R}^{n+1}) φ ∈ C ∞ ( R n + 1 ) , Fubini’s theorem and the equation for u u u give
∫ Ω ♯ ∇ u ♯ ⋅ ∇ φ = ∫ I d ∫ Ω ∇ u ( x ) ⋅ ∇ x φ ( x , z ) d x d z = ∫ I d ∫ ∂ Ω g ( x ) φ ( x , z ) d σ ( x ) d z = ∫ ∂ Ω ♯ g ♯ φ d H n . \begin{aligned}
\int_{\Omega^\sharp}\nabla u^\sharp\cdot\nabla\varphi
=\int_{I_d}\int_\Omega\nabla u(x)\cdot\nabla_x\varphi(x,z)\,dx\,dz \\
&=\int_{I_d}\int_{\partial\Omega}g(x)\varphi(x,z)\,d\sigma(x)\,dz
=\int_{\partial\Omega^\sharp}g^\sharp\varphi\,d\mathcal{H}^n.
\end{aligned} ∫ Ω ♯ ∇ u ♯ ⋅ ∇ φ = ∫ I d ∫ Ω ∇ u ( x ) ⋅ ∇ x φ ( x , z ) d x d z = ∫ I d ∫ ∂ Ω g ( x ) φ ( x , z ) d σ ( x ) d z = ∫ ∂ Ω ♯ g ♯ φ d H n .
and by density this holds for every φ ∈ W 1 , 2 ( Ω ♯ ) \varphi\in W^{1,2}(\Omega^\sharp) φ ∈ W 1 , 2 ( Ω ♯ ) . Finally, if x ∈ S x\in S x ∈ S and z ∈ I d z\in I_d z ∈ I d , then Ω ♯ \Omega^\sharp Ω ♯ coincides with a half-space near ( x , z ) (x,z) ( x , z ) , where g ♯ = 0 g^\sharp=0 g ♯ = 0 , and (54) follows from Lemma 2.2 . □
Proof of Theorems 1.2 and 1.3 for n ≥ 3 n\ge3 n ≥ 3 . For Theorem 1.2 , by (54) and § 9 ,
∫ ∂ Ω ♯ N a ( ∇ u ♯ ) d H n ≥ 2 d ∑ i ∫ S i ∣ ∇ τ u ∣ d σ = ∞ , \int_{\partial\Omega^\sharp}N_a(\nabla u^\sharp)\,d\mathcal{H}^n
\ge2^d\sum_i\int_{S_i}|\nabla_\tau u|\,d\sigma=\infty, ∫ ∂ Ω ♯ N a ( ∇ u ♯ ) d H n ≥ 2 d i ∑ ∫ S i ∣ ∇ τ u ∣ d σ = ∞ ,
and Ω ♯ = ( G ♯ ∘ B ) ( B ( 0 , 1 ) ) \Omega^\sharp=(G^\sharp\circ\mathcal{B})(B(0,1)) Ω ♯ = ( G ♯ ∘ B ) ( B ( 0 , 1 )) .
For Theorem 1.3 , we take the products of the domains of § 8 with I d I_d I d ; they are Lipschitz domains by Lemma B.2, and the lower bound follows from (54) in the same way: the L 1 L^1 L 1 norm of the non-tangential maximal function of the gradient is at least 2 d M 2^dM 2 d M , the boundary has H n \mathcal{H}^n H n measure at most L F n H n ( ∂ ( − 1 , 1 ) n + 1 ) L_F^n\mathcal{H}^n(\partial(-1,1)^{n+1}) L F n H n ( ∂ ( − 1 , 1 ) n + 1 ) , the L p L^p L p norm of the datum is ( 2 d ⋅ 8 ) 1 / p (2^d\cdot8)^{1/p} ( 2 d ⋅ 8 ) 1/ p , and we conclude by Hölder’s inequality. □
Remark 9.2. The local flatness of Remark 8.1 also holds for the domain Ω ♯ \Omega^\sharp Ω ♯ constructed above for Theorem 1.2 , with the exceptional set
K ♯ : = ( K × [ − 1 , 1 ] d ) ∪ ( ∂ Ω × ∂ [ − 1 , 1 ] d ) ∪ ( Ω ‾ × E d ) , K^\sharp:=(K\times[-1,1]^d)\cup(\partial\Omega\times\partial[-1,1]^d)\cup(\overline{\Omega}\times E_d), K ♯ := ( K × [ − 1 , 1 ] d ) ∪ ( ∂ Ω × ∂ [ − 1 , 1 ] d ) ∪ ( Ω × E d ) ,
where K K K is the set of Remark 8.1, and E d E_d E d is the union of the faces of [ − 1 , 1 ] d [-1,1]^d [ − 1 , 1 ] d of dimension at most d − 2 d-2 d − 2 (empty if d = 1 d=1 d = 1 ). It is closed, and H n ( K ♯ ) = 0 \mathcal{H}^n(K^\sharp)=0 H n ( K ♯ ) = 0 . Indeed, since K K K is a countable union of segments and a point, the first set is a countable union of ( n − 1 ) (n-1) ( n − 1 ) -dimensional rectangles and one d d d -dimensional cube, and the other two are bi-Lipschitz images of finite unions of ( n − 1 ) (n-1) ( n − 1 ) -dimensional rectangles. Outside K ♯ K^\sharp K ♯ , the domain Ω ♯ \Omega^\sharp Ω ♯ is locally a half-space bounded by a hyperplane parallel to a coordinate hyperplane.
Geometric properties of bi-Lipschitz images of a cube
We now show that bi-Lipschitz images of a cube have the geometric properties mentioned in the introduction. Let us first recall the definitions. Let V ⊂ R n + 1 V\subset\mathbb{R}^{n+1} V ⊂ R n + 1 be a domain, and write Δ ( x , r ) : = B ( x , r ) ∩ ∂ V \Delta(x,r):=B(x,r)\cap\partial V Δ ( x , r ) := B ( x , r ) ∩ ∂ V for x ∈ ∂ V x\in\partial V x ∈ ∂ V and r > 0 r>0 r > 0 .
V V V is uniform if there is C ≥ 1 C\geq1 C ≥ 1 such that any two points X , X ′ ∈ V X,X'\in V X , X ′ ∈ V can be joined by a rectifiable curve γ ⊂ V \gamma\subset V γ ⊂ V with H 1 ( γ ) ≤ C ∣ X − X ′ ∣ \mathcal{H}^1(\gamma)\leq C|X-X'| H 1 ( γ ) ≤ C ∣ X − X ′ ∣ and
δ V ( Z ) ≥ C − 1 min { H 1 ( γ X , Z ) , H 1 ( γ Z , X ′ ) } \delta_V(Z)\geq C^{-1}\min\{\mathcal{H}^1(\gamma_{X,Z}),\mathcal{H}^1(\gamma_{Z,X'})\} δ V ( Z ) ≥ C − 1 min { H 1 ( γ X , Z ) , H 1 ( γ Z , X ′ )} for every Z ∈ γ Z\in\gamma Z ∈ γ , where γ X , Z \gamma_{X,Z} γ X , Z and γ Z , X ′ \gamma_{Z,X'} γ Z , X ′ are the two subarcs of γ \gamma γ between Z Z Z and its endpoints. As is well known, this is equivalent to the interior corkscrew and Harnack chain conditions, see for instance the appendix of [3 ] .
∂ V \partial V ∂ V is n n n -Ahlfors regular if there is C ≥ 1 C\geq1 C ≥ 1 such that C − 1 r n ≤ σ V ( Δ ( x , r ) ) ≤ C r n C^{-1}r^n\leq\sigma_V(\Delta(x,r))\leq Cr^n C − 1 r n ≤ σ V ( Δ ( x , r )) ≤ C r n for every x ∈ ∂ V x\in\partial V x ∈ ∂ V and 0 < r < diam ∂ V 0<r<\operatorname{diam}\partial V 0 < r < diam ∂ V .
V V V satisfies the exterior corkscrew condition if there is C ≥ 1 C\geq1 C ≥ 1 such that, for every x ∈ ∂ V x\in\partial V x ∈ ∂ V and 0 < r < diam ∂ V 0<r<\operatorname{diam}\partial V 0 < r < diam ∂ V , the ball B ( x , r ) B(x,r) B ( x , r ) contains a ball of radius C − 1 r C^{-1}r C − 1 r contained in R n + 1 ∖ V ‾ \mathbb{R}^{n+1}\setminus\overline{V} R n + 1 ∖ V .
V V V is a chord-arc domain if it is uniform, ∂ V \partial V ∂ V is n n n -Ahlfors regular, and V V V satisfies the exterior corkscrew condition. It is two-sided chord-arc if, moreover, R n + 1 ∖ V ‾ \mathbb{R}^{n+1}\setminus\overline{V} R n + 1 ∖ V is also a uniform domain.
Finally, ∂ V \partial V ∂ V supports a weak ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality if there are constants C , Λ ≥ 1 C,\Lambda\geq1 C , Λ ≥ 1 such that, for all Lipschitz functions f f f on ∂ V \partial V ∂ V , all x ∈ ∂ V x\in\partial V x ∈ ∂ V and all 0 < r < diam ∂ V 0<r<\operatorname{diam}\partial V 0 < r < diam ∂ V , we have
1 σ ( Δ ( x , r ) ) ∫ Δ ( x , r ) ∣ f − m Δ ( x , r ) f ∣ d σ ≤ C r σ ( Δ ( x , Λ r ) ) ∫ Δ ( x , Λ r ) ∣ ∇ τ f ∣ d σ . (55) \frac{1}{\sigma(\Delta(x,r))}\int_{\Delta(x,r)}|f-m_{\Delta(x,r)}f|\,d\sigma\leq\frac{Cr}{\sigma(\Delta(x,\Lambda r))}\int_{\Delta(x,\Lambda r)}|\nabla_\tau f|\,d\sigma. \tag*{(55)} σ ( Δ ( x , r )) 1 ∫ Δ ( x , r ) ∣ f − m Δ ( x , r ) f ∣ d σ ≤ σ ( Δ ( x , Λ r )) C r ∫ Δ ( x , Λ r ) ∣ ∇ τ f ∣ d σ . (55)
Lemma 9.3. Let G G G be an L L L -bi-Lipschitz homeomorphism of R n + 1 \mathbb{R}^{n+1} R n + 1 , and let V = G ( Q 0 ) V=G(Q_0) V = G ( Q 0 ) . Then V V V is a two-sided chord-arc domain with connected boundary and connected exterior, and ∂ V \partial V ∂ V supports a weak ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality. All the constants depend only on n n n and L L L .
Proof. The cube Q 0 Q_0 Q 0 has all these properties, with dimensional constants, and bi-Lipschitz homeomorphisms of the ambient space preserve them, with constants depending on L L L . Let us sketch the details. Since G ( ∂ Q 0 ) = ∂ V G(\partial Q_0)=\partial V G ( ∂ Q 0 ) = ∂ V , we have L − 1 δ Q 0 ( X ) ≤ δ V ( G ( X ) ) ≤ L δ Q 0 ( X ) L^{-1}\delta_{Q_0}(X)\leq\delta_V(G(X))\leq L\delta_{Q_0}(X) L − 1 δ Q 0 ( X ) ≤ δ V ( G ( X )) ≤ L δ Q 0 ( X ) , and similarly in the complements. As G G G is a homeomorphism of R n + 1 \mathbb{R}^{n+1} R n + 1 with G ( Q 0 ) = V G(Q_0)=V G ( Q 0 ) = V , we also have G ( R n + 1 ∖ Q 0 ‾ ) = R n + 1 ∖ V ‾ G(\mathbb{R}^{n+1}\setminus\overline{Q_0})=\mathbb{R}^{n+1}\setminus\overline{V} G ( R n + 1 ∖ Q 0 ) = R n + 1 ∖ V , which implies that ∂ V \partial V ∂ V and R n + 1 ∖ V ‾ \mathbb{R}^{n+1}\setminus\overline{V} R n + 1 ∖ V are connected. Moreover, G G G distorts the lengths of curves at most by the factor L L L . Thus G G G maps corkscrew balls to sets containing corkscrew balls with comparable radii, and it maps the curves of the definition of uniform domains (“cigars”) to curves with the same properties, both in Q 0 Q_0 Q 0 and in its complement. More precisely, if γ \gamma γ is a curve as in the definition of uniform domains joining X , X ′ ∈ Q 0 X,X'\in Q_0 X , X ′ ∈ Q 0 , then G ( γ ) G(\gamma) G ( γ ) joins G ( X ) G(X) G ( X ) and G ( X ′ ) G(X') G ( X ′ ) , it satisfies
H 1 ( G ( γ ) ) ≤ L H 1 ( γ ) ≤ C L 2 ∣ G ( X ) − G ( X ′ ) ∣ , \mathcal{H}^1(G(\gamma))\leq L\mathcal{H}^1(\gamma)\leq CL^2|G(X)-G(X')|, H 1 ( G ( γ )) ≤ L H 1 ( γ ) ≤ C L 2 ∣ G ( X ) − G ( X ′ ) ∣ ,
and, for Z ∈ γ Z\in\gamma Z ∈ γ ,
δ V ( G ( Z ) ) ≥ L − 1 δ Q 0 ( Z ) ≥ ( C L ) − 1 min { H 1 ( γ X , Z ) , H 1 ( γ Z , X ′ ) } ≥ ( C L 2 ) − 1 min { H 1 ( G ( γ X , Z ) ) , H 1 ( G ( γ Z , X ′ ) ) } . \begin{aligned}
\delta_V(G(Z))\geq L^{-1}\delta_{Q_0}(Z)\geq(CL)^{-1}\min\{\mathcal{H}^1(\gamma_{X,Z}),\mathcal{H}^1(\gamma_{Z,X'})\} \\
&\geq(CL^2)^{-1}\min\{\mathcal{H}^1(G(\gamma_{X,Z})),\mathcal{H}^1(G(\gamma_{Z,X'}))\}.
\end{aligned} δ V ( G ( Z )) ≥ L − 1 δ Q 0 ( Z ) ≥ ( C L ) − 1 min { H 1 ( γ X , Z ) , H 1 ( γ Z , X ′ )} ≥ ( C L 2 ) − 1 min { H 1 ( G ( γ X , Z )) , H 1 ( G ( γ Z , X ′ ))} .
The same argument applies in the complement of Q 0 Q_0 Q 0 , which is also a uniform domain. Concerning the corkscrew balls, if B ( X , s ) ⊂ Q 0 B(X,s)\subset Q_0 B ( X , s ) ⊂ Q 0 , then B ( G ( X ) , s / L ) ⊂ G ( B ( X , s ) ) B(G(X),s/L)\subset G(B(X,s)) B ( G ( X ) , s / L ) ⊂ G ( B ( X , s )) , and similarly in the complement. Moreover, by the area formula, the Jacobian of G ∣ ∂ Q 0 G|_{\partial Q_0} G ∣ ∂ Q 0 is between L − n L^{-n} L − n and L n L^n L n . Since, for x 0 ∈ ∂ Q 0 x_0 \in\partial Q_0 x 0 ∈ ∂ Q 0 and r > 0 r > 0 r > 0 , we also have
G ( B ( x 0 , r / L ) ∩ ∂ Q 0 ) ⊂ Δ ( G ( x 0 ) , r ) ⊂ G ( B ( x 0 , L r ) ∩ ∂ Q 0 ) , G(B(x_0,r/L)\cap\partial Q_0)\subset\Delta(G(x_0),r)\subset G(B(x_0,Lr)\cap\partial Q_0), G ( B ( x 0 , r / L ) ∩ ∂ Q 0 ) ⊂ Δ ( G ( x 0 ) , r ) ⊂ G ( B ( x 0 , L r ) ∩ ∂ Q 0 ) ,
we deduce that Ahlfors regularity is preserved. It remains to deal with the Poincaré inequality.
To begin with, ∂ Q 0 \partial Q_0 ∂ Q 0 satisfies (55) with some dimensional constants C 0 C_0 C 0 and Λ 0 \Lambda_0 Λ 0 . When the radius is small, this is a consequence of the Euclidean ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality, since Δ ( x 0 , r ) \Delta(x_0,r) Δ ( x 0 , r ) is the image of a ball of R n \mathbb{R}^n R n by a bi-Lipschitz map with dimensional constants. For large radii, we cover ∂ Q 0 \partial Q_0 ∂ Q 0 by finitely many pieces of this type, overlapping in sets of positive measure, we compare the means on overlapping pieces, and we add up the local inequalities. Now we consider a Lipschitz function f f f on ∂ V \partial V ∂ V , a point x = G ( x 0 ) ∈ ∂ V x=G(x_0)\in\partial V x = G ( x 0 ) ∈ ∂ V and 0 < r < diam ∂ V 0<r<\operatorname{diam}\partial V 0 < r < diam ∂ V . Let h : = f ∘ G h:=f\circ G h := f ∘ G and D : = B ( x 0 , L r ) ∩ ∂ Q 0 D:=B(x_0,Lr)\cap\partial Q_0 D := B ( x 0 , L r ) ∩ ∂ Q 0 , so that G − 1 ( Δ ( x , r ) ) ⊂ D G^{-1}(\Delta(x,r))\subset D G − 1 ( Δ ( x , r )) ⊂ D . For any constant b b b , the area formula gives
∫ Δ ( x , r ) ∣ f − m Δ ( x , r ) f ∣ d σ ≤ 2 ∫ Δ ( x , r ) ∣ f − b ∣ d σ ≤ 2 L n ∫ D ∣ h − b ∣ d H n . \int_{\Delta(x,r)}|f-m_{\Delta(x,r)}f|\,d\sigma\le2\int_{\Delta(x,r)}|f-b|\,d\sigma\le2L^n\int_D|h-b|\,d\mathcal{H}^n. ∫ Δ ( x , r ) ∣ f − m Δ ( x , r ) f ∣ d σ ≤ 2 ∫ Δ ( x , r ) ∣ f − b ∣ d σ ≤ 2 L n ∫ D ∣ h − b ∣ d H n .
We choose b = m D h b=m_Dh b = m D h , and we apply (55) to h h h on ∂ Q 0 \partial Q_0 ∂ Q 0 , with the larger set D ∗ : = B ( x 0 , Λ 0 L r ) ∩ ∂ Q 0 D^*:=B(x_0,\Lambda_0Lr)\cap\partial Q_0 D ∗ := B ( x 0 , Λ 0 L r ) ∩ ∂ Q 0 . (If L r ≥ diam ∂ Q 0 Lr\ge\operatorname{diam}\partial Q_0 L r ≥ diam ∂ Q 0 , then D = D ∗ = ∂ Q 0 D=D^*=\partial Q_0 D = D ∗ = ∂ Q 0 , and we use the Poincaré inequality on the whole of ∂ Q 0 \partial Q_0 ∂ Q 0 .) Almost everywhere on ∂ Q 0 \partial Q_0 ∂ Q 0 we have ∣ ∇ τ h ∣ ≤ L ∣ ∇ τ f ∣ ∘ G |\nabla_\tau h|\le L|\nabla_\tau f|\circ G ∣ ∇ τ h ∣ ≤ L ∣ ∇ τ f ∣ ∘ G , and G ( D ∗ ) ⊂ Δ ( x , Λ 0 L 2 r ) G(D^*)\subset\Delta(x,\Lambda_0L^2r) G ( D ∗ ) ⊂ Δ ( x , Λ 0 L 2 r ) . Changing variables back to ∂ V \partial V ∂ V , and comparing the measures of D D D , D ∗ D^* D ∗ , Δ ( x , r ) \Delta(x,r) Δ ( x , r ) and Δ ( x , Λ 0 L 2 r ) \Delta(x,\Lambda_0L^2r) Δ ( x , Λ 0 L 2 r ) by Ahlfors regularity, we obtain (55) for f f f , with Λ = Λ 0 L 2 \Lambda=\Lambda_0L^2 Λ = Λ 0 L 2 and C C C depending only on n n n and L L L . Alternatively, the two-sided chord-arc property also follows from the bi-Lipschitz invariance of uniform domains [27 ] , which is a particular case of their invariance under quasiconformal homeomorphisms of the ambient space, due to F. Gehring and B. Osgood [13 ] , and the Poincaré inequality from [38 ] , where it is shown that the boundary of every two-sided chord-arc domain supports a weak ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality. The invariance of weak Poincaré inequalities on Ahlfors regular spaces under bi-Lipschitz maps is also standard, see [16 He01] . □
By Lemma 9.3, the domains of Theorems 1.2 and 1.3 are two-sided chord-arc domains whose boundaries support a weak ( 1 , 1 ) (1,1) ( 1 , 1 ) -Poincaré inequality, uniformly. In particular, their boundaries are uniformly rectifiable, the harmonic measure belongs to A ∞ ( σ ) A_\infty(\sigma) A ∞ ( σ ) [9 Se90] , and the regularity problem is solvable in L p L^p L p for some p > 1 p>1 p > 1 [32 ] . Concerning Lipschitz graphs, the domains Ω M ⊂ R n + 1 \Omega_M\subset\mathbb{R}^{n+1} Ω M ⊂ R n + 1 of Theorem 1.3 have local Lipschitz graphs with slopes at most n \sqrt{n} n , by Lemma B.2. However, as explained after Corollary 1.4, their Lipschitz character cannot be uniformly bounded. What degenerates is the size of the coordinate cylinders in which the boundary is a Lipschitz graph.
Appendix A. The bi-Lipschitz parametrizations
In this appendix we prove the lemmas of § 6 on the column map, on the insertion maps and on the rigidity induction. The arguments only involve the fixed column pattern and the placement of its copies, and they do not use any estimate for the Neumann solutions.
An explicit map
Recall that Π = E × [ 0 , λ ] \Pi=E\times[0,\lambda] Π = E × [ 0 , λ ] , with E = [ − 1 , 1 ] 2 E=[-1,1]^2 E = [ − 1 , 1 ] 2 . We first construct an explicit bi-Lipschitz map F 0 ( R + 3 ) = R + 3 ∖ Π F_0(\mathbb{R}^3_+)=\mathbb{R}^3_+\setminus\Pi F 0 ( R + 3 ) = R + 3 ∖ Π , which is affine near the preimages of the selected squares, and then we precompose it with a diffeomorphism of the half-space, so that it becomes rigid there. The map F 0 F_0 F 0 , represented in Figure 5 , is the composition of two maps:
The first one, F 1 F_1 F 1 , pushes up a square of the floor slightly larger than E E E , and creates a tall truncated pyramid with steep lateral faces.
Then F 2 F_2 F 2 contracts each horizontal section of the truncated pyramid onto the square E E E , and in this way the lateral faces of the pyramid become the vertical walls of Π \Pi Π .
The precise definitions are given below, and the properties of F 1 F_1 F 1 and F 2 F_2 F 2 that we need are collected in Lemma A.1. Let w w w be as in § 4 . We fix 0 < ε 0 < w / 8 0<\varepsilon_0<w/8 0 < ε 0 < w /8 , and we consider the function
ϕ ( y ) : = min { 1 , ( 1 + ε 0 − ∣ y ∣ ∞ ) + ε 0 } . \phi(y):=\min\left\{1,\frac{(1+\varepsilon_0-\lvert y\rvert_\infty)_+}{\varepsilon_0}\right\}. ϕ ( y ) := min { 1 , ε 0 ( 1 + ε 0 − ∣ y ∣ ∞ ) + } .
Thus ϕ = 1 \phi=1 ϕ = 1 on E E E and ϕ = 0 \phi=0 ϕ = 0 outside ( 1 + ε 0 ) E (1+\varepsilon_0)E ( 1 + ε 0 ) E , while between these two squares ϕ \phi ϕ is an affine function of ∣ y ∣ ∞ \lvert y\rvert_\infty ∣ y ∣ ∞ . We also take ψ ∈ C c ∞ ( − 1 − 4 λ , 5 λ + 1 ) \psi\in C^\infty_c(-1-4\lambda,5\lambda+1) ψ ∈ C c ∞ ( − 1 − 4 λ , 5 λ + 1 ) with 0 ≤ ψ ≤ 1 0\leq\psi\leq1 0 ≤ ψ ≤ 1 , ψ = 1 \psi=1 ψ = 1 on [ − 1 , λ + 1 ] [-1,\lambda+1] [ − 1 , λ + 1 ] and ∣ ψ ′ ∣ ≤ ( 2 λ ) − 1 \lvert\psi'\rvert\leq(2\lambda)^{-1} ∣ ψ ′ ∣ ≤ ( 2 λ ) − 1 , and we define
F 1 ( y , t ) : = ( y , t + λ ϕ ( y ) ψ ( t ) ) . F_1(y,t):=(y,t+\lambda\phi(y)\psi(t)). F 1 ( y , t ) := ( y , t + λ ϕ ( y ) ψ ( t )) .
To define F 2 F_2 F 2 , we denote
ω ( t ) : = 1 + ε 0 ( 1 − t λ ) ∈ ( 1 , 1 + ε 0 ) for t ∈ ( 0 , λ ) , \omega(t):=1+\varepsilon_0\left(1-\frac{t}{\lambda}\right)\in(1,1+\varepsilon_0)\qquad\text{for }t\in(0,\lambda), ω ( t ) := 1 + ε 0 ( 1 − λ t ) ∈ ( 1 , 1 + ε 0 ) for t ∈ ( 0 , λ ) ,
and we extend ω \omega ω continuously to R \mathbb{R} R by ω = 1 \omega=1 ω = 1 on [ λ , ∞ ) [\lambda,\infty) [ λ , ∞ ) and on ( − ∞ , − 2 ] (-\infty,-2] ( − ∞ , − 2 ] , ω = 1 + ε 0 \omega=1+\varepsilon_0 ω = 1 + ε 0 on [ − 1 , 0 ] [-1,0] [ − 1 , 0 ] , and affinely on [ − 2 , − 1 ] [-2,-1] [ − 2 , − 1 ] . For every t t t , let g t : R → R g_t:\mathbb{R}\to\mathbb{R} g t : R → R be the odd increasing piecewise affine function which is the identity on [ 0 , 1 − w ] ∪ [ 3 , ∞ ) [0,1-w]\cup[3,\infty) [ 0 , 1 − w ] ∪ [ 3 , ∞ ) and maps affinely
[ 1 − w , ω ( t ) − w 2 ] → [ 1 − w , 1 − w 2 ] , [ ω ( t ) − w 2 , 2 ] → [ 1 − w 2 , 3 − ω ( t ) ] , \left[1-w,\omega(t)-\frac{w}{2}\right]\to\left[1-w,1-\frac{w}{2}\right],\qquad
\left[\omega(t)-\frac{w}{2},2\right]\to\left[1-\frac{w}{2},3-\omega(t)\right], [ 1 − w , ω ( t ) − 2 w ] → [ 1 − w , 1 − 2 w ] , [ ω ( t ) − 2 w , 2 ] → [ 1 − 2 w , 3 − ω ( t ) ] ,
[ 2 , 3 ] → [ 3 − ω ( t ) , 3 ] , [2,3]\to[3-\omega(t),3], [ 2 , 3 ] → [ 3 − ω ( t ) , 3 ] ,
and we write h t : = g t − id h_t:=g_t-\operatorname{id} h t := g t − id . To glue the maps g t g_t g t to the identity, we also take an even function χ ∈ C c ∞ ( − 3 , 3 ) \chi\in C^\infty_c(-3,3) χ ∈ C c ∞ ( − 3 , 3 ) with χ = 1 \chi=1 χ = 1 on [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] and ∣ χ ′ ∣ ≤ 2 \lvert\chi'\rvert\leq2 ∣ χ ′ ∣ ≤ 2 . We then set
F 2 ( y 1 , y 2 , t ) : = ( y 1 + χ ( y 2 ) h t ( y 1 ) , y 2 + χ ( y 1 ) h t ( y 2 ) , t ) , F 0 : = F 2 ∘ F 1 . F_2(y_1,y_2,t):=(y_1+\chi(y_2)h_t(y_1),y_2+\chi(y_1)h_t(y_2),t),\qquad F_0:=F_2\circ F_1. F 2 ( y 1 , y 2 , t ) := ( y 1 + χ ( y 2 ) h t ( y 1 ) , y 2 + χ ( y 1 ) h t ( y 2 ) , t ) , F 0 := F 2 ∘ F 1 .
Lemma A.1. The maps F 1 F_1 F 1 and F 2 F_2 F 2 satisfy the following properties.
(a) The map F 1 F_1 F 1 is a bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 , it coincides with the identity outside [ − 2 , 2 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) [-2,2]^2\times(-1-4\lambda,5\lambda+1) [ − 2 , 2 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) , and
F 1 ( R + 3 ) = { ( y , t ) : t > λ ϕ ( y ) } . F_1(\mathbb{R}^3_+)=\{(y,t):t>\lambda\phi(y)\}. F 1 ( R + 3 ) = {( y , t ) : t > λ ϕ ( y )} . Moreover, the set { 0 < t ≤ λ ϕ ( y ) } \{0<t\leq\lambda\phi(y)\} { 0 < t ≤ λ ϕ ( y )} is a truncated square pyramid of height λ \lambda λ , and its horizontal section at height t ∈ ( 0 , λ ) t\in(0,\lambda) t ∈ ( 0 , λ ) is the square { ∣ y ∣ ∞ ≤ ω ( t ) } \{\lvert y\rvert_\infty\leq\omega(t)\} {∣ y ∣ ∞ ≤ ω ( t )} .
(b) For every t t t , g t g_t g t is an odd increasing bijection of R \mathbb{R} R with g t ( ω ( t ) ) = 1 g_t(\omega(t))=1 g t ( ω ( t )) = 1 , and g t = id g_t=\operatorname{id} g t = id if t ≤ − 2 t\leq-2 t ≤ − 2 or t ≥ λ t\geq\lambda t ≥ λ . Moreover, for all r , t , t ′ ∈ R r,t,t'\in\mathbb{R} r , t , t ′ ∈ R ,
∣ h t ( r ) ∣ ≤ ε 0 , ∣ h t ′ ( r ) ∣ ≤ 2 ε 0 w , ∣ h t ( r ) − h t ′ ( r ) ∣ ≤ ε 0 ∣ t − t ′ ∣ . \lvert h_t(r)\rvert\leq\varepsilon_0,\qquad\lvert h_t'(r)\rvert\leq\frac{2\varepsilon_0}{w},\qquad\lvert h_t(r)-h_{t'}(r)\rvert\leq\varepsilon_0\lvert t-t'\rvert. ∣ h t ( r )∣ ≤ ε 0 , ∣ h t ′ ( r )∣ ≤ w 2 ε 0 , ∣ h t ( r ) − h t ′ ( r )∣ ≤ ε 0 ∣ t − t ′ ∣ . (c) The map F 2 F_2 F 2 is a bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 which preserves every horizontal plane and coincides with the identity outside [ − 3 , 3 ] 2 × ( − 2 , λ ) [-3,3]^2\times(-2,\lambda) [ − 3 , 3 ] 2 × ( − 2 , λ ) . Moreover,
F 2 ( y 1 , y 2 , t ) = ( g t ( y 1 ) , g t ( y 2 ) , t ) for ( y 1 , y 2 ) ∈ [ − 2 , 2 ] 2 . F_2(y_1,y_2,t)=(g_t(y_1),g_t(y_2),t)\qquad\text{for }(y_1,y_2)\in[-2,2]^2. F 2 ( y 1 , y 2 , t ) = ( g t ( y 1 ) , g t ( y 2 ) , t ) for ( y 1 , y 2 ) ∈ [ − 2 , 2 ] 2 .
Proof. We first prove (a). The support property is clear, because ϕ \phi ϕ vanishes outside ( 1 + ε 0 ) E (1+\varepsilon_0)E ( 1 + ε 0 ) E and ψ \psi ψ vanishes outside ( − 1 − 4 λ , 5 λ + 1 ) (-1-4\lambda,5\lambda+1) ( − 1 − 4 λ , 5 λ + 1 ) . Since 1 2 ≤ ∂ t ( t + λ ϕ ψ ) ≤ 3 2 \frac{1}{2}\leq\partial_t(t+\lambda\phi\psi)\leq\frac{3}{2} 2 1 ≤ ∂ t ( t + λ ϕ ψ ) ≤ 2 3 and ∣ ∇ y ( λ ϕ ψ ) ∣ ≤ λ / ε 0 \lvert\nabla_y(\lambda\phi\psi)\rvert\leq\lambda/\varepsilon_0 ∣ ∇ y ( λ ϕ ψ )∣ ≤ λ / ε 0 , the map F 1 F_1 F 1 is Lipschitz. Moreover, for fixed y y y , the function t ↦ t + λ ϕ ( y ) ψ ( t ) t\mapsto t+\lambda\phi(y)\psi(t) t ↦ t + λ ϕ ( y ) ψ ( t ) has derivative between 1 2 \frac{1}{2} 2 1 and 3 2 \frac{3}{2} 2 3 , and so it is an increasing bijection of R \mathbb{R} R . As F 1 F_1 F 1 does not change the first two coordinates, it follows that F 1 F_1 F 1 is a bijection of R 3 \mathbb{R}^3 R 3 . Concerning the inverse, suppose that F 1 ( y , t ) = ( y , t ′ ) F_1(y,t)=(y,t') F 1 ( y , t ) = ( y , t ′ ) and F 1 ( z , s ) = ( z , s ′ ) F_1(z,s)=(z,s') F 1 ( z , s ) = ( z , s ′ ) . Since ϕ \phi ϕ is ε 0 − 1 \varepsilon_0^{-1} ε 0 − 1 -Lipschitz, we have
∣ t ′ − s ′ ∣ ≥ ∣ t + λ ϕ ( y ) ψ ( t ) − s − λ ϕ ( z ) ψ ( s ) ∣ − λ ∣ ϕ ( y ) − ϕ ( z ) ∣ ∣ ψ ( s ) ∣ ≥ 1 2 ∣ t − s ∣ − λ ε 0 − 1 ∣ y − z ∣ . \begin{aligned}
\lvert t'-s'\rvert\geq\left\lvert t+\lambda\phi(y)\psi(t)-s-\lambda\phi(z)\psi(s)\right\rvert-\lambda\lvert\phi(y)-\phi(z)\rvert\lvert\psi(s)\rvert\\
&\geq\frac{1}{2}\lvert t-s\rvert-\lambda\varepsilon_0^{-1}\lvert y-z\rvert.
\end{aligned} ∣ t ′ − s ′ ∣ ≥ ∣ t + λ ϕ ( y ) ψ ( t ) − s − λ ϕ ( z ) ψ ( s ) ∣ − λ ∣ ϕ ( y ) − ϕ ( z )∣ ∣ ψ ( s )∣ ≥ 2 1 ∣ t − s ∣ − λ ε 0 − 1 ∣ y − z ∣ .
that is, ∣ t − s ∣ ≤ 2 ∣ t ′ − s ′ ∣ + 2 λ ε 0 − 1 ∣ y − z ∣ |t-s|\le2|t'-s'|+2\lambda\varepsilon_0^{-1}|y-z| ∣ t − s ∣ ≤ 2∣ t ′ − s ′ ∣ + 2 λ ε 0 − 1 ∣ y − z ∣ . Hence F 1 − 1 F_1^{-1} F 1 − 1 is also Lipschitz. Since F 1 F_1 F 1 preserves the vertical lines, it is increasing in t t t , and F 1 ( y , 0 ) = ( y , λ ϕ ( y ) ) F_1(y,0)=(y,\lambda\phi(y)) F 1 ( y , 0 ) = ( y , λ ϕ ( y )) , we deduce that F 1 ( R + 3 ) = { t > λ ϕ ( y ) } F_1(\mathbb{R}^3_+)=\{t>\lambda\phi(y)\} F 1 ( R + 3 ) = { t > λ ϕ ( y )} . Finally, by the definitions of ϕ \phi ϕ and ω \omega ω , for t ∈ ( 0 , λ ) t\in(0,\lambda) t ∈ ( 0 , λ ) we have λ ϕ ( y ) ≥ t \lambda\phi(y)\ge t λ ϕ ( y ) ≥ t if and only if ∣ y ∣ ∞ ≤ ω ( t ) |y|_\infty\le\omega(t) ∣ y ∣ ∞ ≤ ω ( t ) .
We now turn to (b). Being odd, increasing, piecewise affine, and equal to the identity on [ 3 , ∞ ) [3,\infty) [ 3 , ∞ ) , the function g t g_t g t is a bijection of R \mathbb{R} R . The middle map in its definition is the translation by 1 − ω ( t ) 1-\omega(t) 1 − ω ( t ) , which gives g t ( ω ( t ) ) = 1 g_t(\omega(t))=1 g t ( ω ( t )) = 1 . Also, if ω ( t ) = 1 \omega(t)=1 ω ( t ) = 1 , that is, if t ≤ − 2 t\le-2 t ≤ − 2 or t ≥ λ t\ge\lambda t ≥ λ , all three maps are the identity. The slopes of the other two pieces are w / 2 ω ( t ) − 1 + w / 2 \frac{w/2}{\omega(t)-1+w/2} ω ( t ) − 1 + w /2 w /2 and ω ( t ) \omega(t) ω ( t ) , and this easily implies that ∣ h t ∣ ≤ ε 0 |h_t|\le\varepsilon_0 ∣ h t ∣ ≤ ε 0 and ∣ h t ′ ∣ ≤ 2 ε 0 / w |h'_t|\le2\varepsilon_0/w ∣ h t ′ ∣ ≤ 2 ε 0 / w . For the Lipschitz bound in t t t , we set δ : = ω ( t ) − 1 ∈ [ 0 , ε 0 ] \delta:=\omega(t)-1\in[0,\varepsilon_0] δ := ω ( t ) − 1 ∈ [ 0 , ε 0 ] . If r ≥ 0 r\ge0 r ≥ 0 , then h t ( r ) = 0 h_t(r)=0 h t ( r ) = 0 for r ≤ 1 − w r\le1-w r ≤ 1 − w and for r ≥ 3 r\ge3 r ≥ 3 , and on the intervals [ 1 − w , ω ( t ) − w / 2 ] [1-w,\omega(t)-w/2] [ 1 − w , ω ( t ) − w /2 ] , [ ω ( t ) − w / 2 , 2 ] [\omega(t)-w/2,2] [ ω ( t ) − w /2 , 2 ] and [ 2 , 3 ] [2,3] [ 2 , 3 ] we have, respectively,
h t ( r ) = − δ ( r − 1 + w ) δ + w / 2 , h t ( r ) = − δ , h t ( r ) = δ ( r − 3 ) . h_t(r)=-\frac{\delta(r-1+w)}{\delta+w/2},\qquad h_t(r)=-\delta,\qquad h_t(r)=\delta(r-3). h t ( r ) = − δ + w /2 δ ( r − 1 + w ) , h t ( r ) = − δ , h t ( r ) = δ ( r − 3 ) .
The three expressions coincide at the breakpoints, and each of them, as a function of δ \delta δ , has derivative bounded by 1 1 1 in absolute value (for the first one, use that 0 ≤ r − 1 + w ≤ δ + w / 2 0\le r-1+w\le\delta+w/2 0 ≤ r − 1 + w ≤ δ + w /2 ). As h t h_t h t is odd and ω \omega ω is ε 0 \varepsilon_0 ε 0 -Lipschitz, we conclude that ∣ h t ( r ) − h t ′ ( r ) ∣ ≤ ε 0 ∣ t − t ′ ∣ |h_t(r)-h_{t'}(r)|\le\varepsilon_0|t-t'| ∣ h t ( r ) − h t ′ ( r ) ∣ ≤ ε 0 ∣ t − t ′ ∣ .
Let us prove (c). Since χ \chi χ vanishes outside ( − 3 , 3 ) (-3,3) ( − 3 , 3 ) , the support property is a consequence of (b), and the formula in [ − 2 , 2 ] 2 [-2,2]^2 [ − 2 , 2 ] 2 is immediate, because χ = 1 \chi=1 χ = 1 on [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] . To show that F 2 F_2 F 2 is bi-Lipschitz, we fix t t t and we write the horizontal part of F 2 F_2 F 2 as id + H t \operatorname{id}+H_t id + H t . From (b) we deduce that the ℓ 1 \ell^1 ℓ 1 norm of every row and every column of D y H t D_yH_t D y H t is at most κ : = 2 ε 0 / w + 2 ε 0 < 1 / 2 \kappa:=2\varepsilon_0/w+2\varepsilon_0<1/2 κ := 2 ε 0 / w + 2 ε 0 < 1/2 . Then the Schur test, ∥ D y H t ∥ 2 → 2 ≤ ∥ D y H t ∥ 1 → 1 1 / 2 ∥ D y H t ∥ ∞ → ∞ 1 / 2 \|D_yH_t\|_{2\to2}\le\|D_yH_t\|_{1\to1}^{1/2}\|D_yH_t\|_{\infty\to\infty}^{1/2} ∥ D y H t ∥ 2 → 2 ≤ ∥ D y H t ∥ 1 → 1 1/2 ∥ D y H t ∥ ∞ → ∞ 1/2 , gives ∥ D y H t ∥ ≤ κ \|D_yH_t\|\le\kappa ∥ D y H t ∥ ≤ κ . In particular, H t H_t H t is κ \kappa κ -Lipschitz. By the contraction mapping principle, for every x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 the equation y = x − H t ( y ) y=x-H_t(y) y = x − H t ( y ) has a unique solution, and so id + H t \operatorname{id}+H_t id + H t is a bijection of R 2 \mathbb{R}^2 R 2 . Moreover, ∣ H t ( y ) − H t ′ ( y ) ∣ ≤ 2 ε 0 ∣ t − t ′ ∣ |H_t(y)-H_{t'}(y)|\le\sqrt{2}\varepsilon_0|t-t'| ∣ H t ( y ) − H t ′ ( y ) ∣ ≤ 2 ε 0 ∣ t − t ′ ∣ , by (b). Thus, if y + H t ( y ) = x y+H_t(y)=x y + H t ( y ) = x and y ′ + H t ′ ( y ′ ) = x ′ y'+H_{t'}(y')=x' y ′ + H t ′ ( y ′ ) = x ′ , then
∣ y − y ′ ∣ ≤ ∣ x − x ′ ∣ + ∣ H t ( y ) − H t ′ ( y ) ∣ + ∣ H t ′ ( y ) − H t ′ ( y ′ ) ∣ ≤ ∣ x − x ′ ∣ + 2 ε 0 ∣ t − t ′ ∣ + κ ∣ y − y ′ ∣ , \begin{aligned}
|y-y'|\le|x-x'|+|H_t(y)-H_{t'}(y)|+|H_{t'}(y)-H_{t'}(y')| \\
&\le|x-x'|+\sqrt{2}\varepsilon_0|t-t'|+\kappa|y-y'|,
\end{aligned} ∣ y − y ′ ∣ ≤ ∣ x − x ′ ∣ + ∣ H t ( y ) − H t ′ ( y ) ∣ + ∣ H t ′ ( y ) − H t ′ ( y ′ ) ∣ ≤ ∣ x − x ′ ∣ + 2 ε 0 ∣ t − t ′ ∣ + κ ∣ y − y ′ ∣ ,
that is, ∣ y − y ′ ∣ ≤ ( 1 − κ ) − 1 ( ∣ x − x ′ ∣ + 2 ε 0 ∣ t − t ′ ∣ ) |y-y'|\le(1-\kappa)^{-1}(|x-x'|+\sqrt{2}\varepsilon_0|t-t'|) ∣ y − y ′ ∣ ≤ ( 1 − κ ) − 1 ( ∣ x − x ′ ∣ + 2 ε 0 ∣ t − t ′ ∣ ) . This bound for F 2 − 1 F_2^{-1} F 2 − 1 , together with the analogous (easier) one for F 2 F_2 F 2 , shows that F 2 F_2 F 2 is bi-Lipschitz. □
Lemma A.2. The map F 0 F_0 F 0 is a bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 with F 0 = id F_0=\operatorname{id} F 0 = id outside [ − 3 , 3 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) [-3,3]^2\times(-1-4\lambda,5\lambda+1) [ − 3 , 3 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) , and
F 0 ( R + 3 ) = R + 3 ∖ Π . (56) F_0(\mathbb{R}^3_+)=\mathbb{R}^3_+\setminus\Pi. \tag*{(56)} F 0 ( R + 3 ) = R + 3 ∖ Π. (56) Moreover, on the open set U r : = { 1 + 4 ε 0 / λ < y 1 < 1 + ε 0 ( 1 − 4 / λ ) , ∣ y 2 ∣ < 1 − w , ∣ t ∣ < 1 } U_r:=\{1+4\varepsilon_0/\lambda<y_1<1+\varepsilon_0(1-4/\lambda),\ |y_2|<1-w,\ |t|<1\} U r := { 1 + 4 ε 0 / λ < y 1 < 1 + ε 0 ( 1 − 4/ λ ) , ∣ y 2 ∣ < 1 − w , ∣ t ∣ < 1 } ,
F 0 ( y 1 , y 2 , t ) = A r ( y 1 , y 2 , t ) : = ( 1 + ε 0 t λ , y 2 , t + λ ε 0 ( 1 + ε 0 − y 1 ) ) . (57) F_0(y_1,y_2,t)=A_r(y_1,y_2,t):=\left(1+\frac{\varepsilon_0t}{\lambda},y_2,t+\frac{\lambda}{\varepsilon_0}(1+\varepsilon_0-y_1)\right). \tag*{(57)} F 0 ( y 1 , y 2 , t ) = A r ( y 1 , y 2 , t ) := ( 1 + λ ε 0 t , y 2 , t + ε 0 λ ( 1 + ε 0 − y 1 ) ) . (57) The affine map A r A_r A r has determinant 1 1 1 , and it maps { t = 0 } \{t=0\} { t = 0 } onto the plane { y 1 = 1 } \{y_1=1\} { y 1 = 1 } of the right wall of Π \Pi Π and R + 3 \mathbb{R}^3_+ R + 3 onto { y 1 > 1 } \{y_1>1\} { y 1 > 1 } . Moreover, A r ( U r ∩ { t = 0 } ) A_r(U_r\cap\{t=0\}) A r ( U r ∩ { t = 0 }) contains the right rectangle of R \mathcal{R} R . By symmetry, analogous statements hold for the other three walls.
Proof. The fact that F 0 F_0 F 0 is a bi-Lipschitz homeomorphism of R 3 \mathbb{R}^3 R 3 , equal to the identity outside [ − 3 , 3 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) [-3,3]^2\times(-1-4\lambda,5\lambda+1) [ − 3 , 3 ] 2 × ( − 1 − 4 λ , 5 λ + 1 ) , follows from Lemma A.1(a) and (c). Let us check (A.1). Since F 2 F_2 F 2 preserves the horizontal planes, it is enough to show that, for each t t t , the map id + H t \operatorname{id}+H_t id + H t sends the horizontal section of F 1 ( R + 3 ) F_1(\mathbb{R}^3_+) F 1 ( R + 3 ) at height t t t onto that of R + 3 ∖ Π \mathbb{R}^3_+\setminus\Pi R + 3 ∖ Π . For t ≤ 0 t\le0 t ≤ 0 both sections are empty. For t ≥ λ t\ge\lambda t ≥ λ we have g t = id g_t=\operatorname{id} g t = id , and both sections are R 2 \mathbb{R}^2 R 2 if t > λ t>\lambda t > λ , and R 2 ∖ Π \mathbb{R}^2\setminus\Pi R 2 ∖ Π if t = λ t=\lambda t = λ . For 0 < t < λ 0<t<\lambda 0 < t < λ , the section of F 1 ( R + 3 ) F_1(\mathbb{R}^3_+) F 1 ( R + 3 ) is { ∣ y ∣ ∞ > ω ( t ) } \{|y|_\infty>\omega(t)\} { ∣ y ∣ ∞ > ω ( t )} , by Lemma A.1(a). In [ − 2 , 2 ] 2 [-2,2]^2 [ − 2 , 2 ] 2 the map id + H t \operatorname{id}+H_t id + H t coincides with ( g t ( y 1 ) , g t ( y 2 ) ) (g_t(y_1),g_t(y_2)) ( g t ( y 1 ) , g t ( y 2 )) , and so, by Lemma A.1 (b), it maps { ∣ y ∣ ∞ ≤ ω ( t ) } \{|y|_\infty\le\omega(t)\} { ∣ y ∣ ∞ ≤ ω ( t )} onto E E E . Since id + H t \operatorname{id}+H_t id + H t is a bijection, it maps { ∣ y ∣ ∞ > ω ( t ) } \{|y|_\infty> \omega(t)\} { ∣ y ∣ ∞ > ω ( t )} onto R 2 ∖ E \mathbb{R}^2 \setminus E R 2 ∖ E , as wanted.
Now let ( y , t ) ∈ U r (y,t) \in\mathcal{U}_r ( y , t ) ∈ U r . Then ∣ y ∣ ∞ = y 1 ∈ ( 1 , 1 + ε 0 ) |y|_\infty=y_1 \in(1,1+\varepsilon_0) ∣ y ∣ ∞ = y 1 ∈ ( 1 , 1 + ε 0 ) , so that ϕ ( y ) = ( 1 + ε 0 − y 1 ) / ε 0 \phi(y)=(1+\varepsilon_0-y_1)/\varepsilon_0 ϕ ( y ) = ( 1 + ε 0 − y 1 ) / ε 0 , and ψ ( t ) = 1 \psi(t)=1 ψ ( t ) = 1 . Thus F 1 ( y , t ) = ( y , t ′ ) F_1(y,t)=(y,t') F 1 ( y , t ) = ( y , t ′ ) with t ′ = t + λ ε 0 ( 1 + ε 0 − y 1 ) ∈ ( 3 , λ − 3 ) t'=t+\frac{\lambda}{\varepsilon_0}(1+\varepsilon_0-y_1)\in(3,\lambda-3) t ′ = t + ε 0 λ ( 1 + ε 0 − y 1 ) ∈ ( 3 , λ − 3 ) , and a direct computation gives ω ( t ′ ) = y 1 − ε 0 t / λ \omega(t')=y_1-\varepsilon_0t/\lambda ω ( t ′ ) = y 1 − ε 0 t / λ . Hence y 1 y_1 y 1 lies in the interval [ ω ( t ′ ) − w / 2 , 2 ] [\omega(t')-w/2,2] [ ω ( t ′ ) − w /2 , 2 ] , where g t ′ g_{t'} g t ′ is the translation by 1 − ω ( t ′ ) 1-\omega(t') 1 − ω ( t ′ ) , so that g t ′ ( y 1 ) = 1 + ε 0 t / λ g_{t'}(y_1)=1+\varepsilon_0t/\lambda g t ′ ( y 1 ) = 1 + ε 0 t / λ , while g t ′ ( y 2 ) = y 2 g_{t'}(y_2)=y_2 g t ′ ( y 2 ) = y 2 because ∣ y 2 ∣ < 1 − w |y_2|<1-w ∣ y 2 ∣ < 1 − w .
This proves (A.2). The remaining assertions are immediate because the preimage of the point ( 1 , y 2 , t ′ ) (1,y_2,t') ( 1 , y 2 , t ′ ) of the right wall under A r A_r A r is ( 1 + ε 0 ( 1 − t ′ / λ ) , y 2 , 0 ) (1+\varepsilon_0(1-t'/\lambda),y_2,0) ( 1 + ε 0 ( 1 − t ′ / λ ) , y 2 , 0 ) , and 8 ≤ t ′ ≤ λ − 8 8\le t'\le\lambda-8 8 ≤ t ′ ≤ λ − 8 and ∣ y 2 ∣ ≤ 1 − 2 w |y_2|\le1-2w ∣ y 2 ∣ ≤ 1 − 2 w on R \mathcal{R} R .
Making the map rigid
The map F 0 F_0 F 0 is affine near the preimage T i T_i T i of each selected square P i P_i P i , but it is far from rigid there: it stretches the thin rectangle T i T_i T i onto the square P i P_i P i . We will precompose F 0 F_0 F 0 with a diffeomorphism Θ \Theta Θ of the half-space which maps a square D i D_i D i of side a a a onto T i T_i T i by the inverse of this stretching. Then F 0 ∘ Θ F_0\circ\Theta F 0 ∘ Θ maps D i D_i D i rigidly onto P i P_i P i . The squares D i D_i D i have the same size as the P i P_i P i , and they have to be disjoint, so we need enough room in the floor, which is where we use A < 1 A<1 A < 1 . The existence of Θ \Theta Θ is an elementary but somewhat tedious fact, see Lemma A.3 below.
Lemma A.3 (Moving squares). Let Q ⊂ R 2 Q\subset\mathbb{R}^2 Q ⊂ R 2 be an open square, and let D 1 , … , D N ⊂ Q × { 0 } D_1,\ldots,D_N\subset Q\times\{0\} D 1 , … , D N ⊂ Q × { 0 } be pairwise disjoint closed squares. For i = 1 , … , N i=1,\ldots,N i = 1 , … , N , let L i L_i L i be an orientation-preserving affine map of R 3 \mathbb{R}^3 R 3 with L i ( R + 3 ) = R + 3 L_i(\mathbb{R}^3_+)=\mathbb{R}^3_+ L i ( R + 3 ) = R + 3 , such that the sets L i ( D i ) L_i(D_i) L i ( D i ) are pairwise disjoint and contained in Q × { 0 } Q\times\{0\} Q × { 0 } . Then there is a C ∞ C^\infty C ∞ diffeomorphism Θ \Theta Θ of R 3 \mathbb{R}^3 R 3 , equal to the identity outside a compact subset of Q × ( − 1 , 1 ) Q\times(-1,1) Q × ( − 1 , 1 ) , such that Θ ( R + 3 ) = R + 3 \Theta(\mathbb{R}^3_+)=\mathbb{R}^3_+ Θ ( R + 3 ) = R + 3 and Θ = L i \Theta=L_i Θ = L i in a neighbourhood of D i D_i D i for every i i i . In particular, Θ \Theta Θ is bi-Lipschitz.
We remark that the lemma above is a version of the isotopy extension theorem of differential topology, see for instance [15 Chapter 8] , applied to affine motions of disjoint neighbourhoods of the squares D i D_i D i , with the additional requirement that the half-space is preserved.
We realize the affine maps L i L_i L i as the final positions of affine motions L i , s L_{i,s} L i , s , s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , which move small neighbourhoods of the squares D i D_i D i without collisions. Then Θ \Theta Θ is the time-one map of the flow of a vector field which follows these motions near the squares and vanishes far from them, see Figure 6 .
Figure 6. The construction of the map Θ \Theta Θ of Lemma A.3 , seen from above. The squares K i ⊃ D i K_i \supset D_i K i ⊃ D i are shrunk, moved one at a time without collisions, deformed by linear maps, and expanded onto the parallelograms ℓ i ( K i ) \ell_i(K_i) ℓ i ( K i ) . The map Θ \Theta Θ is the time-one map of the flow of a vector field which follows these motions near the squares (see the proof of Lemma A.3 ).
Since we also want Θ \Theta Θ to preserve the upper half-space and to coincide with the given affine maps near the squares, we give the details.
Proof of Lemma A.3 . The map L i L_i L i is affine, orientation-preserving, and preserves R + 3 \mathbb{R}^{3}_{+} R + 3 , and so it is of the form
L i ( y , t ) = ( ℓ i ( y ) + t v i , c i t ) , ℓ i ( y ) = A i y + b i , A i ∈ G L + ( 2 , R ) , v i ∈ R 2 , c i > 0. L_i(y,t) = (\ell_i(y) + t v_i, c_i t), \qquad\ell_i(y) = A_i y + b_i, \qquad A_i \in GL^{+}(2,\mathbb{R}),\ v_i \in\mathbb{R}^{2},\ c_i > 0. L i ( y , t ) = ( ℓ i ( y ) + t v i , c i t ) , ℓ i ( y ) = A i y + b i , A i ∈ G L + ( 2 , R ) , v i ∈ R 2 , c i > 0.
Let K i K_i K i be closed squares containing D i D_i D i in their interiors. We take them so close to D i D_i D i that they are pairwise disjoint and compactly contained in Q Q Q , and so that the same holds for the parallelograms ℓ i ( K i ) \ell_i(K_i) ℓ i ( K i ) .
Step 1: planar affine motions. First we construct smooth paths of affine maps a i , s a_{i,s} a i , s , s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] , with a i , 0 = id a_{i,0} = \operatorname{id} a i , 0 = id and a i , 1 = ℓ i a_{i,1} = \ell_i a i , 1 = ℓ i , such that for every s s s the sets a i , s ( K i ) a_{i,s}(K_i) a i , s ( K i ) are pairwise disjoint and compactly contained in Q Q Q . Denote by q i q_i q i the centre of K i K_i K i , let q i ′ = ℓ i ( q i ) q'_i = \ell_i(q_i) q i ′ = ℓ i ( q i ) , and choose auxiliary points p i ∈ Q p_i \in Q p i ∈ Q , all distinct and different from all the q j q_j q j and q j ′ q'_j q j ′ . The motion consists of five stages, each of them reparametrized so that all its time derivatives vanish at its endpoints:
shrink every K i K_i K i around q i q_i q i by a factor μ ∈ ( 0 , 1 ) \mu\in(0,1) μ ∈ ( 0 , 1 ) ;
move the shrunk copies, one at a time, from q i q_i q i to p i p_i p i , along smooth paths in Q Q Q which avoid the current positions of the other centres;
move them, one at a time, from p i p_i p i to q i ′ q'_i q i ′ in the same way;
deform the linear part of each copy from μ id \mu\operatorname{id} μ id to μ A i \mu A_i μ A i , along a path in G L + ( 2 , R ) GL^{+}(2,\mathbb{R}) G L + ( 2 , R ) (which is connected: if A i = Rot ( α i ) S i A_i = \operatorname{Rot}(\alpha_i)S_i A i = Rot ( α i ) S i is the polar decomposition of A i A_i A i , where Rot ( α ) \operatorname{Rot}(\alpha) Rot ( α ) denotes the rotation by the angle α \alpha α and S i S_i S i is positive definite, one can take the path s ↦ Rot ( s α i ) ( ( 1 − s ) id + s S i ) s \mapsto\operatorname{Rot}(s\alpha_i)((1-s)\operatorname{id}+sS_i) s ↦ Rot ( s α i ) (( 1 − s ) id + s S i ) , s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] );
expand every copy around q i ′ q'_i q i ′ by the factor μ − 1 \mu^{-1} μ − 1 .
The paths in (ii) and (iii) exist because Q Q Q minus finitely many points is path connected. Note that all the paths of centres and of linear parts are chosen before μ \mu μ . As there are finitely many of them, they stay at positive distance from the other current centres and from ∂ Q \partial Q ∂ Q , and the linear parts are bounded. Therefore, for μ \mu μ small enough, the shrunk copies remain inside Q Q Q and do not collide in stages (ii)–(iv). In stages (i) and (v) there are no collisions either, because by convexity each copy stays inside K i K_i K i in stage (i), and inside ℓ i ( K i ) \ell_i(K_i) ℓ i ( K i ) in stage (v).
Step 2: spatial affine motions and collars. We extend the planar motions to R 3 \mathbb{R}^{3} R 3 by
L i , s ( y , t ) : = ( a i , s ( y ) + s t v i , c i s t ) , L_{i,s}(y,t) := (a_{i,s}(y) + s t v_i, c_i^{s}t), L i , s ( y , t ) := ( a i , s ( y ) + s t v i , c i s t ) ,
so that L i , 0 = id L_{i,0}=\operatorname{id} L i , 0 = id , L i , 1 = L i L_{i,1}=L_i L i , 1 = L i , and every L i , s L_{i,s} L i , s preserves R + 3 \mathbb{R}^{3}_{+} R + 3 . By compactness in s s s , the sets a i , s ( K i ) a_{i,s}(K_i) a i , s ( K i ) stay at distance bounded below from each other and from ∂ Q \partial Q ∂ Q , and the coefficients of the maps L i , s L_{i,s} L i , s are uniformly bounded. Hence, if τ > 0 \tau>0 τ > 0 is small enough, the moving collars
K i ( s ) : = L i , s ( K i × [ − τ , τ ] ) \mathcal{K}_i(s):=L_{i,s}\left(K_i\times[-\tau,\tau]\right) K i ( s ) := L i , s ( K i × [ − τ , τ ] )
are pairwise disjoint for each s s s , and they are all contained in a fixed compact subset of Q × ( − 1 , 1 ) Q\times(-1,1) Q × ( − 1 , 1 ) .
Step 3: the flow. The sets { ( s , X ) : s ∈ [ 0 , 1 ] , X ∈ K i ( s ) } \{(s,X):s\in[0,1],\,X\in\mathcal{K}_i(s)\} {( s , X ) : s ∈ [ 0 , 1 ] , X ∈ K i ( s )} are compact and pairwise disjoint, and so we can find smooth functions η i ( s , X ) \eta_i(s,X) η i ( s , X ) with pairwise disjoint supports, such that η i = 1 \eta_i=1 η i = 1 in a neighbourhood of the i i i -th set and the supports in X X X are contained in a fixed compact subset of Q × ( − 1 , 1 ) Q\times(-1,1) Q × ( − 1 , 1 ) . We consider the time-dependent vector field
X s ( X ) : = ∑ i η i ( s , X ) ( ∂ s L i , s ) ( L i , s − 1 ( X ) ) . \mathcal{X}_s(X):=\sum_i\eta_i(s,X)\left(\partial_sL_{i,s}\right)\left(L_{i,s}^{-1}(X)\right). X s ( X ) := i ∑ η i ( s , X ) ( ∂ s L i , s ) ( L i , s − 1 ( X ) ) .
This vector field is smooth and has compact support. Hence its flow Φ s \Phi_s Φ s is a family of diffeomorphisms of R 3 \mathbb{R}^3 R 3 , which are the identity outside a compact subset of Q × ( − 1 , 1 ) Q\times(-1,1) Q × ( − 1 , 1 ) . Given X ∈ K i × [ − τ , τ ] X\in K_i\times[-\tau,\tau] X ∈ K i × [ − τ , τ ] , the curve s ↦ L i , s ( X ) s\mapsto L_{i,s}(X) s ↦ L i , s ( X ) stays in K i ( s ) \mathcal{K}_i(s) K i ( s ) , where η i = 1 \eta_i=1 η i = 1 and the other cutoffs vanish, and thus it solves the equation of the flow. By uniqueness, Φ 1 = L i \Phi_1=L_i Φ 1 = L i in K i × [ − τ , τ ] K_i\times[-\tau,\tau] K i × [ − τ , τ ] , which is a neighbourhood of D i D_i D i . On the other hand, the vertical component of X s ( y , t ) \mathcal{X}_s(y,t) X s ( y , t ) is t ∑ i η i ( s , y , t ) log c i t\sum_i\eta_i(s,y,t)\log c_i t ∑ i η i ( s , y , t ) log c i , that is, t t t times a bounded function. So the vertical coordinate gets multiplied by a positive factor along every trajectory, which shows that Φ s ( R + 3 ) = R + 3 \Phi_s(\mathbb{R}^3_+)=\mathbb{R}^3_+ Φ s ( R + 3 ) = R + 3 . Finally, we let Θ : = Φ 1 \Theta:=\Phi_1 Θ := Φ 1 . □
Proof of Lemma 5.1. Since A < 1 A<1 A < 1 and ς > 32 \varsigma>32 ς > 32 , we can choose an open square Y ′ ⋐ Y Y'\Subset Y Y ′ ⋐ Y centred at the origin with [ − 3 , 3 ] 2 ⊂ Y ′ [-3,3]^2\subset Y' [ − 3 , 3 ] 2 ⊂ Y ′ and ∣ Y ′ ∣ > A ∣ Y ∣ |Y'|>A|Y| ∣ Y ′ ∣ > A ∣ Y ∣ . Let δ > 0 \delta>0 δ > 0 be such that ( 1 + δ ) 2 A ∣ Y ∣ < ∣ Y ′ ∣ (1+\delta)^2A|Y|<|Y'| ( 1 + δ ) 2 A ∣ Y ∣ < ∣ Y ′ ∣ . The closed squares of side a a a centred at the points of the grid ( 1 + δ ) a Z 2 (1+\delta)a\mathbb{Z}^2 ( 1 + δ ) a Z 2 are pairwise disjoint, and at least ∣ Y ′ ∣ ( 1 + δ ) − 2 a − 2 − C a − 1 |Y'|(1+\delta)^{-2}a^{-2}-Ca^{-1} ∣ Y ′ ∣ ( 1 + δ ) − 2 a − 2 − C a − 1 of them are contained in Y ′ Y' Y ′ . On the other hand, by (30) , the number of selected squares is N = ∣ M ∣ a − 2 ≤ A ∣ Y ∣ a − 2 N=|\mathcal{M}|a^{-2}\le A|Y|a^{-2} N = ∣ M ∣ a − 2 ≤ A ∣ Y ∣ a − 2 . Thus, if ℓ \ell ℓ (and so a a a ) is small enough, we can choose N N N of these grid squares, and we call them D 1 , … , D N D_1,\ldots,D_N D 1 , … , D N . This is the only point where we need ∣ M ∣ < ∣ Y ∣ |\mathcal{M}|<|Y| ∣ M ∣ < ∣ Y ∣ .
Suppose that P i P_i P i lies on the right wall of Π \Pi Π , and let R i R_i R i be the rigid motion with linear part ( y 1 , y 2 , t ) ↦ ( t , y 2 , − y 1 ) (y_1,y_2,t)\mapsto(t,y_2,-y_1) ( y 1 , y 2 , t ) ↦ ( t , y 2 , − y 1 ) which maps D i D_i D i onto P i P_i P i ; it maps R + 3 \mathbb{R}^3_+ R + 3 onto { y 1 > 1 } \{y_1>1\} { y 1 > 1 } . By Lemma A.2, the map L i : = A r − 1 ∘ R i L_i:=A_r^{-1}\circ R_i L i := A r − 1 ∘ R i is affine, orientation-preserving, with L i ( R + 3 ) = R + 3 L_i(\mathbb{R}^3_+)=\mathbb{R}^3_+ L i ( R + 3 ) = R + 3 , and L i ( D i ) = A r − 1 ( P i ) = : T i L_i(D_i)=A_r^{-1}(P_i)=:T_i L i ( D i ) = A r − 1 ( P i ) =: T i is a thin closed rectangle in U r ∩ { t = 0 } U_r\cap\{t=0\} U r ∩ { t = 0 } . We proceed in the same way for the squares on the other walls. The rectangles T i T_i T i are pairwise disjoint and contained in [ − 3 , 3 ] 2 × { 0 } ⊂ Y ′ × { 0 } [-3,3]^2\times\{0\}\subset Y'\times\{0\} [ − 3 , 3 ] 2 × { 0 } ⊂ Y ′ × { 0 } . Let Θ \Theta Θ be the map given by Lemma A.3 with Q = Y ′ Q=Y' Q = Y ′ , and set
F : = F 0 ∘ Θ . F:=F_0\circ\Theta. F := F 0 ∘ Θ.
Then (34) holds because F 0 F_0 F 0 and Θ \Theta Θ are the identity outside Z Z Z , (35) follows from Θ ( R + 3 ) = R + 3 \Theta(\mathbb{R}^3_+)=\mathbb{R}^3_+ Θ ( R + 3 ) = R + 3 and (A.1), and, if each W i W_i W i is a small enough open cube containing D i D_i D i , then the sets W i W_i W i are pairwise disjoint and contained in Z Z Z , Θ = L i \Theta=L_i Θ = L i in W i W_i W i , and L i ( W i ) ⊂ U r L_i(W_i)\subset U_r L i ( W i ) ⊂ U r for the squares on the right wall (and similarly for the other walls). Thus F = A r ∘ L i = R i F=A_r\circ L_i=R_i F = A r ∘ L i = R i in W i W_i W i , by (A.2). □
Insertion maps.
Proof of Lemma 5.4. Since Ψ b \Psi_b Ψ b is the conjugate of F F F by a similarity, its bi-Lipschitz constants are at most L F L_F L F . Also, Ψ b \Psi_b Ψ b is a homeomorphism which is the identity outside the box R P τ b ( Z ) R_P\tau_b(Z) R P τ b ( Z ) , and so it maps this box onto itself. Hence Ψ \Psi Ψ is a homeomorphism, whose inverse is obtained by patching together the maps Ψ b − 1 \Psi_b^{-1} Ψ b − 1 . To estimate ∣ Ψ ( x ) − Ψ ( x ′ ) ∣ |\Psi(x)-\Psi(x')| ∣Ψ ( x ) − Ψ ( x ′ ) ∣ when x x x and x ′ x' x ′ lie in different boxes, we denote by x ˉ \bar{x} x ˉ the point where the segment [ x , x ′ ] [x,x'] [ x , x ′ ] leaves the box of x x x , and by x ˉ ′ \bar{x}' x ˉ ′ the point where it enters the box of x ′ x' x ′ . These points are fixed by Ψ \Psi Ψ , and so
∣ Ψ ( x ) − Ψ ( x ′ ) ∣ ≤ L F ∣ x − x ˉ ∣ + ∣ x ˉ − x ˉ ′ ∣ + L F ∣ x ˉ ′ − x ′ ∣ ≤ L F ∣ x − x ′ ∣ . |\Psi(x)-\Psi(x')|\le L_F|x-\bar{x}|+|\bar{x}-\bar{x}'|+L_F|\bar{x}'-x'|\le L_F|x-x'|. ∣Ψ ( x ) − Ψ ( x ′ ) ∣ ≤ L F ∣ x − x ˉ ∣ + ∣ x ˉ − x ˉ ′ ∣ + L F ∣ x ˉ ′ − x ′ ∣ ≤ L F ∣ x − x ′ ∣.
The other cases are similar (or easier), and the same argument applies to Ψ − 1 \Psi^{-1} Ψ − 1 . By (35) and since τ b ( R + 3 ) = R + 3 \tau_b(\mathbb{R}^{3}_{+})=\mathbb{R}^{3}_{+} τ b ( R + 3 ) = R + 3 , we have Ψ ( D ) ∩ R P τ b ( Z ) = R P τ b ( ( R + 3 ∖ Π ) ∩ Z ) \Psi(D)\cap R_{P\tau_b}(Z)=R_{P\tau_b}((\mathbb{R}^{3}_{+}\setminus\Pi)\cap Z) Ψ ( D ) ∩ R P τ b ( Z ) = R P τ b (( R + 3 ∖ Π ) ∩ Z ) , which gives the formula for Ψ ( D ) \Psi(D) Ψ ( D ) . Next, in R P τ b ( W i ) R_{P\tau_b}(W_i) R P τ b ( W i ) the map Ψ \Psi Ψ equals Ψ r i g : = R P ( τ b R i τ b − 1 ) R P − 1 \Psi^{\mathrm{rig}}:=R_P(\tau_bR_i\tau_b^{-1})R_P^{-1} Ψ rig := R P ( τ b R i τ b − 1 ) R P − 1 , by (36) , and τ b R i τ b − 1 \tau_bR_i\tau_b^{-1} τ b R i τ b − 1 is a rigid motion with the same linear part as R i R_i R i , since τ b \tau_b τ b is a similarity. Since R P τ b ( W i ) ⊂ V P R_{P\tau_b}(W_i)\subset V_P R P τ b ( W i ) ⊂ V P , we get
Ψ ( D ) ∩ V P ′ = Ψ r i g ( R P ( R + 3 ) ∩ R P τ b ( W i ) ) = Ψ r i g R P ( R + 3 ) ∩ V P ′ , \Psi(D)\cap V_{P'}=\Psi^{\mathrm{rig}}(R_P(\mathbb{R}^{3}_{+})\cap R_{P\tau_b}(W_i))=\Psi^{\mathrm{rig}}R_P(\mathbb{R}^{3}_{+})\cap V_{P'}, Ψ ( D ) ∩ V P ′ = Ψ rig ( R P ( R + 3 ) ∩ R P τ b ( W i )) = Ψ rig R P ( R + 3 ) ∩ V P ′ ,
and ( Ψ r i g R P ) − 1 ( P ′ ) = τ b ( D i ) (\Psi^{\mathrm{rig}}R_P)^{-1}(P')=\tau_b(D_i) ( Ψ rig R P ) − 1 ( P ′ ) = τ b ( D i ) is a square in R 2 × { 0 } \mathbb{R}^{2}\times\{0\} R 2 × { 0 } with sides parallel to the axes. Hence P ′ P' P ′ is flat in Ψ ( D ) \Psi(D) Ψ ( D ) , with the rigid motion Ψ r i g ∘ R P \Psi^{\mathrm{rig}}\circ R_P Ψ rig ∘ R P in (37) . Finally, the sets V P ′ V_{P'} V P ′ are connected and pairwise disjoint because the sets W i W_i W i are, the boxes are disjoint, and Ψ \Psi Ψ is a homeomorphism. □
The rigidity induction
Proof of Lemma 5.5. We prove (a) and (b) by induction, noting that for k = 0 k=0 k = 0 they hold by assumption. Suppose that they hold for some k k k . Let Z k Z_k Z k be the union of the open boxes R P τ b ( Z ) R_{P\tau_b}(Z) R P τ b ( Z ) used in the definition of Ψ k \Psi_k Ψ k , so that Ψ k = id \Psi_k=\operatorname{id} Ψ k = id outside Z k ‾ \overline{Z_k} Z k . Each of these boxes is contained in V P V_P V P for some P ∈ C k P\in\mathcal{C}_k P ∈ C k , and so G k G_k G k is a rigid motion in each connected component of G k − 1 ( Z k ) G_k^{-1}(Z_k) G k − 1 ( Z k ) , by (b). Since G k − 1 G_k^{-1} G k − 1 is Lipschitz, the preimage under G k G_k G k of a null set is a null set, and so, for almost every x x x , the chain rule gives
D G k + 1 ( x ) = { D G k ( x ) if G k ( x ) ∉ Z k ‾ , D Ψ k ( G k ( x ) ) O x if G k ( x ) ∈ Z k ‾ , DG_{k+1}(x)=
\begin{cases}
DG_k(x) & \text{if }G_k(x)\notin\overline{Z_k},\\
D\Psi_k(G_k(x))\,O_x & \text{if }G_k(x)\in\overline{Z_k},
\end{cases} D G k + 1 ( x ) = { D G k ( x ) D Ψ k ( G k ( x )) O x if G k ( x ) ∈ / Z k , if G k ( x ) ∈ Z k ,
where O x O_x O x is an orthogonal matrix. In both cases we have ∥ D G k + 1 ( x ) ∥ ≤ L F \lVert DG_{k+1}(x)\rVert\le L_F ∥ D G k + 1 ( x )∥ ≤ L F , and this implies that Lip ( G k + 1 ) ≤ L F \operatorname{Lip}(G_{k+1})\le L_F Lip ( G k + 1 ) ≤ L F . To see this, notice that the mollifications of G k + 1 G_{k+1} G k + 1 have derivatives bounded by L F L_F L F , and so they are L F L_F L F -Lipschitz, by integration along segments in the convex set R 3 \mathbb{R}^{3} R 3 . Since they converge locally uniformly to G k + 1 G_{k+1} G k + 1 , the claim follows. The same argument works for G k + 1 − 1 = G k − 1 ∘ Ψ k − 1 G_{k+1}^{-1}=G_k^{-1}\circ\Psi_k^{-1} G k + 1 − 1 = G k − 1 ∘ Ψ k − 1 , because Ψ k − 1 \Psi_k^{-1} Ψ k − 1 maps Z k ‾ \overline{Z_k} Z k onto itself, and so (a) holds for k + 1 k+1 k + 1 .
Let us turn to (b). Let P ′ = R P τ b ( P i ) ∈ C k + 1 P'=R_{P\tau_b}(P_i)\in\mathcal{C}_{k+1} P ′ = R P τ b ( P i ) ∈ C k + 1 be a new selected square, where P ∈ C k P\in\mathcal{C}_k P ∈ C k and R P τ b ( Z ) ⊂ V P R_{P\tau_b}(Z)\subset V_P R P τ b ( Z ) ⊂ V P . By Lemma 5.4 , we have G k + 1 − 1 ( V P ′ ) = G k − 1 ( R P τ b ( W i ) ) G_{k+1}^{-1}(V_{P'})=G_k^{-1}(R_{P\tau_b}(W_i)) G k + 1 − 1 ( V P ′ ) = G k − 1 ( R P τ b ( W i )) . In this set G k G_k G k is rigid, by (b), and Ψ k \Psi_k Ψ k is rigid in R P τ b ( W i ) R_{P\tau_b}(W_i) R P τ b ( W i ) . Thus G k + 1 G_{k+1} G k + 1 is rigid in G k + 1 − 1 ( V P ′ ) G_{k+1}^{-1}(V_{P'}) G k + 1 − 1 ( V P ′ ) , and V P ′ V_{P'} V P ′ is connected by Lemma 5.4 , which proves (b) for k + 1 k+1 k + 1 .
The inclusion Ω k + 1 ⊂ Ω k \Omega_{k+1}\subset\Omega_k Ω k + 1 ⊂ Ω k is clear from Lemma 5.4 . By (30) , the new selected squares built on each cell R P τ b ( Y ~ ) R_{P\tau_b}(\widetilde{Y}) R P τ b ( Y ) have total area at most A A A times the area of the cell, and the cells have pairwise disjoint interiors and are contained in the squares of C k \mathcal{C}_k C k . Thus H 2 ( C k + 1 ) ≤ A H 2 ( C k ) \mathcal{H}^{2}(\mathcal{C}_{k+1})\le A\mathcal{H}^{2}(\mathcal{C}_k) H 2 ( C k + 1 ) ≤ A H 2 ( C k ) . □
Appendix B. The finite stages are Lipschitz domains
In this appendix we show that the domains of Theorem 1.3 are Lipschitz domains, with local graphs whose slopes are bounded only in terms of the dimension. The proof relies on an elementary lemma, which applies to all the local models of the boundaries of these domains.
Lemma B.1. Let V ~ ⊂ R m \widetilde{V}\subset\mathbb{R}^{m} V ⊂ R m be a nonempty open cone with vertex at the origin, such that V ~ ≠ R m \widetilde{V}\ne\mathbb{R}^{m} V = R m and
V ~ + [ 0 , ∞ ) m ⊂ V ~ . (58) \widetilde{V}+[0,\infty)^{m}\subset\widetilde{V}. \tag*{(58)} V + [ 0 , ∞ ) m ⊂ V . (58) Let v = m − 1 / 2 ( 1 , … , 1 ) v=m^{-1/2}(1,\ldots,1) v = m − 1/2 ( 1 , … , 1 ) , and let H = v ⊥ H=v^{\perp} H = v ⊥ . Then there is a function f : H → R f:H\to\mathbb{R} f : H → R with Lip ( f ) ≤ m − 1 \operatorname{Lip}(f)\le\sqrt{m-1} Lip ( f ) ≤ m − 1 such that V ~ = { h + s v : h ∈ H , s > f ( h ) } \widetilde{V}=\{h+sv:h\in H,\ s>f(h)\} V = { h + s v : h ∈ H , s > f ( h )} .
The assumptions of Lemma B.1 are satisfied, for instance, by every set obtained from the half-spaces { x i > 0 } \{x_i>0\} { x i > 0 } by finitely many unions and intersections. Indeed, such a set is an open cone with vertex at the origin, it contains the open positive orthant and it does not contain the point ( − 1 , … , − 1 ) (-1,\ldots,-1) ( − 1 , … , − 1 ) , and (B.1) holds because it holds for each of the half-spaces { x i > 0 } \{x_i > 0\} { x i > 0 } and it is preserved by unions and intersections.
Proof. If h ∈ H h \in H h ∈ H , then h i 2 ≤ ( m − 1 ) ∑ j ≠ i h j 2 h_i^2 \le(m-1)\sum_{j\ne i} h_j^2 h i 2 ≤ ( m − 1 ) ∑ j = i h j 2 because h i = − ∑ j ≠ i h j h_i=-\sum_{j\ne i}h_j h i = − ∑ j = i h j , and so ∣ h i ∣ ≤ ( m − 1 ) / m ∣ h ∣ |h_i|\le\sqrt{(m-1)/m}\,|h| ∣ h i ∣ ≤ ( m − 1 ) / m ∣ h ∣ for every i i i . Therefore, the coordinates of h + m − 1 ∣ h ∣ v h+\sqrt{m-1}\,|h|\,v h + m − 1 ∣ h ∣ v are all nonnegative. In other words,
h + m − 1 ∣ h ∣ v ∈ [ 0 , ∞ ) m for every h ∈ H . (59) h+\sqrt{m-1}\,|h|\,v\in[0,\infty)^m \qquad\text{for every }h\in H. \tag*{(59)} h + m − 1 ∣ h ∣ v ∈ [ 0 , ∞ ) m for every h ∈ H . (59)
Now, V ~ \widetilde V V is a nonempty open cone invariant under the addition of vectors with nonnegative coordinates, and so it contains the open positive orthant: if p ∈ V ~ p\in\widetilde V p ∈ V and q ∈ ( 0 , ∞ ) m q\in(0,\infty)^m q ∈ ( 0 , ∞ ) m , then q = t p + ( q − t p ) ∈ V ~ q=tp+(q-tp)\in\widetilde V q = tp + ( q − tp ) ∈ V for t > 0 t>0 t > 0 small. Also, as V ~ ≠ R m \widetilde V\ne\mathbb{R}^m V = R m , no point q q q with all coordinates negative belongs to V ~ \widetilde V V . Otherwise, every x ∈ R m x\in\mathbb{R}^m x ∈ R m could be written as x = t q + ( x − t q ) x=tq+(x-tq) x = tq + ( x − tq ) , with t > 0 t>0 t > 0 large and x − t q ∈ [ 0 , ∞ ) m x-tq\in[0,\infty)^m x − tq ∈ [ 0 , ∞ ) m , and so x ∈ V ~ x\in\widetilde V x ∈ V . Next, fix h ∈ H h\in H h ∈ H . On the line { h + s v } \{h+sv\} { h + s v } , the set of s s s with h + s v ∈ V ~ h+sv\in\widetilde V h + s v ∈ V is an open half-line ( f ( h ) , ∞ ) (f(h),\infty) ( f ( h ) , ∞ ) : it is invariant under translations to the right, it contains all large s s s , it contains no very negative s s s , and it is open. If h , h ′ ∈ H h,h'\in H h , h ′ ∈ H and s > f ( h ) s>f(h) s > f ( h ) , then by (B.2)
h ′ + ( s + m − 1 ∣ h − h ′ ∣ ) v = ( h + s v ) + ( ( h ′ − h ) + m − 1 ∣ h ′ − h ∣ v ) ∈ V ~ , h'+\left(s+\sqrt{m-1}\,|h-h'|\right)v=(h+sv)+\left((h'-h)+\sqrt{m-1}\,|h'-h|\,v\right)\in\widetilde V, h ′ + ( s + m − 1 ∣ h − h ′ ∣ ) v = ( h + s v ) + ( ( h ′ − h ) + m − 1 ∣ h ′ − h ∣ v ) ∈ V ,
and so f ( h ′ ) ≤ f ( h ) + m − 1 ∣ h − h ′ ∣ f(h')\le f(h)+\sqrt{m-1}\,|h-h'| f ( h ′ ) ≤ f ( h ) + m − 1 ∣ h − h ′ ∣ . By symmetry, $\operatorname{Lip}(f)\le\sqrt{m-1}.
Lemma B.2. Let Ω k = G k ( Q 0 ) ⊂ R 3 \Omega_k=G_k(Q_0)\subset\mathbb{R}^3 Ω k = G k ( Q 0 ) ⊂ R 3 be a domain obtained by finitely many insertions as in § 6 , and let d ≥ 0 d\ge0 d ≥ 0 . Then Ω k × ( − 1 , 1 ) d \Omega_k\times(-1,1)^d Ω k × ( − 1 , 1 ) d is a bounded Lipschitz domain, whose boundary is a finite union of rectangles parallel to the coordinate hyperplanes, and whose local Lipschitz graphs have slopes at most d + 2 \sqrt{d+2} d + 2 .
Proof. At every finite stage of the construction there are only finitely many columns. Every insertion takes place in a flat neighbourhood, where the domain is a half-space, and the new columns are separated from each other and from the edges of the old ones. As a consequence, each point x ∈ ∂ Ω k x\in\partial\Omega_k x ∈ ∂ Ω k has a neighbourhood in which Ω k \Omega_k Ω k coincides, up to a translation and a signed permutation of the coordinates, with one of the following cones, all of which satisfy the assumptions of Lemma B.1 :
{ x 3 > 0 } , { x 1 > 0 } ∩ { x 3 > 0 } , ⋂ i = 1 3 { x i > 0 } , \{x_{3}>0\}, \qquad\{x_{1}>0\}\cap\{x_{3}>0\}, \qquad\bigcap_{i=1}^{3}\{x_{i}>0\}, { x 3 > 0 } , { x 1 > 0 } ∩ { x 3 > 0 } , i = 1 ⋂ 3 { x i > 0 } ,
{ x 1 > 0 } ∪ { x 2 > 0 } , { x 3 > 0 } ∩ ( { x 1 > 0 } ∪ { x 2 > 0 } ) , ⋃ i = 1 3 { x i > 0 } . \{x_{1}>0\}\cup\{x_{2}>0\}, \qquad\{x_{3}>0\}\cap\left(\{x_{1}>0\}\cup\{x_{2}>0\}\right), \qquad\bigcup_{i=1}^{3}\{x_{i}>0\}. { x 1 > 0 } ∪ { x 2 > 0 } , { x 3 > 0 } ∩ ( { x 1 > 0 } ∪ { x 2 > 0 } ) , i = 1 ⋃ 3 { x i > 0 } .
The first three models occur at the faces, edges and vertices of the cube, and also on the faces of the columns and on the edges where the columns meet the floor. The other three occur at the vertical edges of the columns and at the edges of their roofs, at the vertices where the columns meet the floor, and at the vertices of the roofs (see Figure 7 ). Similarly, at every point of ∂ ( Ω k × ( − 1 , 1 ) d ) \partial(\Omega_k\times(-1,1)^d) ∂ ( Ω k × ( − 1 , 1 ) d ) the product domain locally coincides, after a translation and a signed permutation, with the product of one of these models (or of R 3 \mathbb{R}^{3} R 3 ) with a set of the form ⋂ i ∈ I { z i > 0 } \bigcap_{i\in I}\{z_i>0\} ⋂ i ∈ I { z i > 0 } , I ⊂ { 1 , … , d } I\subset\{1,\ldots,d\} I ⊂ { 1 , … , d } , which is again obtained from coordinate half-spaces of R d + 3 \mathbb{R}^{d+3} R d + 3 by finitely many unions and intersections. By Lemma B.1 , the product domain is a Lipschitz graph domain near each boundary point, with slope at most d + 2 \sqrt{d+2} d + 2 , and by compactness it is a Lipschitz domain. □
Figure 7. The local models at the non-flat boundary points created by the columns: (i) an edge where a wall meets the floor, (ii) a vertical edge, or an edge of the roof, (iii) a vertex where a column meets the floor, (iv) a vertex of the roof. The grey solid is the complement of the domain, and the arrow is the vector v v v of Lemma B.1 .
We emphasize that there is no uniform lower bound for the size of the neighbourhoods in the proof: they are comparable to the scale of the smallest columns, which tends to zero along the construction. This is consistent with Corollary 1.4.
References
Mihalis Mourgoglou, Departamento de Matemáticas, Universidad del País Vasco (UPV/EHU), Barrio Sarriena S/N 48940 Leioa, Spain and, Ikerbasque, Basque Foundation for Science, Bilbao, Spain. Email address: michail.mourgoglou@ehu.eus
Carmelo Puliatti, Departament de Matemàtiques, Universitat Autònoma de Barcelona, 08193 Bellaterra (Barcelona), Catalonia. Email address: carmelo.puliatti@uab.cat
Xavier Tolsa, ICREA, 08010 Barcelona, Catalonia; Departament de Matemàtiques, Universitat Autònoma de Barcelona, 08193 Bellaterra, Catalonia; and Centre de Recerca Matemàtica, 08193 Bellaterra, Catalonia. Email address: xavier.tolsa@uab.cat