<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>Pointwise Gradient Estimate of the Ritz Projection</title></titleStmt>
			<publicationStmt>
				<publisher>SIAM</publisher>
				<date>06/30/2024</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10529787</idno>
					<idno type="doi">10.1137/23M1571800</idno>
					<title level='j'>SIAM Journal on Numerical Analysis</title>
<idno>0036-1429</idno>
<biblScope unit="volume">62</biblScope>
<biblScope unit="issue">3</biblScope>					

					<author>Lars Diening</author><author>Julian Rolfes</author><author>Abner J Salgado</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[Let Ω ⊂ R n be a convex polytope (n ≤ 3). The Ritz projection is the best approximation, in the W 1,2 0 -norm, to a given function in a finite element space. When such finite element spaces are constructed on the basis of quasiuniform triangulations, we show a pointwise estimate on the Ritz projection. Namely, the gradient at any point in Ω is controlled by the Hardy-Littlewood maximal function of the gradient of the original function at the same point. From this estimate, the stability of the Ritz projection on a wide range of spaces that are of interest in the analysis of PDEs immediately follows. Among those are weighted spaces, Orlicz spaces, and Lorentz spaces.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><p>1. Introduction. To approximate solutions of partial differential equations (PDEs), in particular, those that are second order and elliptic, the finite element method has emerged as the method of choice. A finite element scheme is nothing but a Galerkin approximation with a particular choice of finite dimensional subspace (piecewise polynomials subject to a triangulation of the domain) and a particular basis. It is fair to say that the study of the properties of finite element schemes for second order linear elliptic second order equations in an energy setting has reached a state of maturity. In short, the Ritz projection, which is the best approximation in the W 1,2 0 -norm (see section 2 for notation), possesses optimal approximation properties when these are measured in the energy norm, which usually is a norm equivalent to the W 1,2 -norm. This reduces the numerical analysis of a finite element scheme to a question of approximation theory, and this is usually resolved by constructing a suitable interpolant.</p><p>On the other hand, the study of the properties of the Ritz projection in nonenergy norms has been the subject of intensive study with many classical results, recent progress, and still some open questions. We refer the reader to the introductions of <ref type="bibr">[15]</ref> and <ref type="bibr">[7]</ref> for some historical accounts. It is fair to say that the development of this subject is obscured by technicalities, and it is far from settled. Nevertheless, apart from the intrinsic interest such estimates may present, these become important when dealing, for instance, with nonlinear or coupled problems, or even when in a linear problem the data is sufficiently rough that the functional setting that provides well-posedness is no longer the energy one (see, for instance, <ref type="bibr">[10]</ref>), or when the energy norm is not equivalent to the usual W 1,2 -norm (see <ref type="bibr">[22]</ref>).</p><p>The purpose of this work is to make a contribution in this direction. In this work we concentrate on the Ritz projection subject to homogeneous Dirichlet boundary values. Homogeneity is merely a matter of convenience. It might be possible to consider Neumann boundary conditions, see <ref type="bibr">[31]</ref> for results in this direction. In fact one would only need suitable Green's function estimates. Treating mixed boundary values, however, is difficult in this context since the regularity corresponds to the one of a slit domain. For such domains the needed estimates of the Green's functions are not valid. We show that, over quasiuniform meshes, the gradient of the Ritz projection at any point in the domain is controlled by the Hardy-Littlewood maximal operator of the gradient of the original function at the same point. This pointwise estimate not only immediately implies stability of the Ritz projection in any function space where the maximal operator is bounded butalso elucidates the action of the Ritz projection, i.e., finite element approximation. It is a sort of averaging procedure.</p><p>Our presentation is organized as follows. In section 2 we introduce notation. The statement of our main result, Theorem 3.1, is presented in section 3. Here we also collect a list of corollaries. Some of these recover known results, whereas others are truly new and may find application in the finite element approximation of, for instance, nonlinear elliptic problems with nonstandard growth conditions <ref type="bibr">[8]</ref>. The proof of our main result is the content of section 4. For clarity, this proof is split into several steps that comprise the bulk of this section.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.">Notation and preliminaries.</head><p>We begin by introducing some notation and specifying the framework under which we shall operate. The relation A B means that there is a constant c for which A &#8804; cB. The value of this constant may change at each occurrence. More importantly, this constant does not depend on A, B, or discretization parameters. A &#8776; B means that A B and B A.</p><p>Throughout our work, &#8486; &#8834; R n , n &#8804; 3, is a bounded convex polytope. While convexity is essential for our arguments, the dimensional restriction is merely an artifact of our methods. Given x &#8712; R n , we denote its Euclidean norm by |x|. By B(x, r) we denote the open ball with center x &#8712; R n and radius r &gt; 0. For a measurable set E &#8834; R n we denote by |E| its Lebesgue measure. L 0 (&#8486;) denotes the collection of functions &#8486; &#8594; R that are measurable. For p &#8712; [1, &#8734;] and k &#8712; N we denote by L p (&#8486;) and W k,p (&#8486;), respectively, the usual Lebesgue and Sobolev spaces. The subspace of W k,p (&#8486;) that consists of functions vanishing on the boundary is denoted by W k,p 0 (&#8486;). We immediately notice that, whenever w &#8712; W k,p 0 (&#8486;), its extension to R n \ &#8486; by zero, denoted by w, is such that w &#8712; W k,p (R n ). For this reason, whenever necessary, we shall make this extension by zero without explicit mention or change of notation. By L 1 loc (R n ) we denote the space of locally integrable functions. For</p><p>for all x &#8712; R n . With this notation M [f ] readily extends to vector valued functions. If X is a normed space, we shall denote by &#8226; X its norm. If this norm comes from an inner product, this will be denoted by &#8226;, &#8226; X . We shall make no distinction between scalar and vector valued functions or their spaces, as this will be clear from the context. For &#945; &#8712; (0, 1] we let C 0,&#945; (&#8486;) denote the space of H&#246;lder continuous functions with seminorm</p><p>Let T = {T h } h&gt;0 be a quasiuniform family of conforming triangulations of &#8486; in the sense of <ref type="bibr">Ciarlet [3,</ref><ref type="bibr">p. 124]</ref> where, for h &gt; 0, the triangulation T h has mesh size h. For k &#8712; N we denote by</p><p>the Lagrange space of degree k, where P k is the space of polynomials of degree at most k. We set</p><p>We comment that this mapping is the orthogonal projections onto V h with respect to the W 1,2 0 (&#8486;)-seminorm. The following local error estimate for R h can be found in [7, <ref type="bibr">Theorem 1]</ref>. In fact, it holds for more general families of triangulations than quasiuniform ones. (&#8486;) and T be a quasiuniform family of triangulations of a polytype &#8486;. Let z &#8712; &#8486; and h &gt; 0. Define D = &#8486; &#8745; B d (z) with d &#8805; k 0 h, where k 0 is sufficiently large. We have, for every</p><p>where the implicit constant is independent of w, h, and z.</p><p>Proof. As mentioned before, this is essentially [7, <ref type="bibr">Theorem 1]</ref>. However, in that result, as stated, the point z is where &#8711;(w -R h w) L &#8734; (&#8486;) is attained. One merely needs to examine the proof to see that this point may be arbitrary.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.">Statement of the main result and corollaries.</head><p>We are now in position to state the main result of our work.</p><p>Theorem 3.1 (pointwise estimate). Let &#8486; &#8834; R n , for n &#8712; {2, 3}, be a convex polytope and T = {T h } h&gt;0 be a family of conforming and quasiuniform triangulations of &#8486;. For every u &#8712; W 1,1 0 (&#8486;) and almost every z &#8712; &#8486; we have</p><p>where the implicit constant is independent of z, u, and h and depends on T only through its shape regularity constants.</p><p>Before we embark on the proof of this result, we immediately mention that it implies the stability of the Ritz projection in any space where the Hardy-Littlewood maximal operator is bounded. For the sake of completeness we present a far from exhaustive list of examples: (weighted) L p spaces (see section 3.3), Lorentz spaces (see sections 3.1 and 3.4); Orlicz spaces (see section 3.2) and (weighted) variable exponent spaces (see section 3.5).</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1.">Lorentz spaces.</head><p>Let &#181; be a measure on &#8486;, p &#8712; [1, &#8734;), and q &#8712; [1, &#8734;]. The Lorentz spaces are defined as</p><p>and</p><p>is the distribution function of f . We recall that,</p><p>Finally, if &#181; is the Lebesgue measure, we simply denote these spaces by L p,q (&#8486;).</p><p>Corollary 3.2 (Lorentz stability). In the setting of Theorem 3.1 assume, in addition, that p &#8712; (1, &#8734;) and q &#8712; (1, &#8734;], or that p = 1 and q = &#8734;. Then we have</p><p>where the implicit constant is independent of u and h. In particular, for p &#8712; (1, &#8734;], we have</p><p>Proof. Consider first the case p = 1 and q = &#8734;. Owing to, for instance, [12, Theorem 2.1.6] we have</p><p>For p &gt; 1 it suffices to invoke [17, Theorems A, section 5.2].</p><p>We comment that the boundedness of the Ritz projection in W 1,p spaces has already been presented in <ref type="bibr">[23,</ref><ref type="bibr">15,</ref><ref type="bibr">7]</ref>. Thus, the case p &#8712; (1, &#8734;) of Corollary 3.2 can also be obtained by the Marcinkiewicz interpolation theorem as presented in [2, Theorem 5.3.2].</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.2.">Orlicz spaces.</head><p>Another new result is stability in Orlicz spaces. We say that &#981; : (0, &#8734;) &#8594; (0, &#8734;) is an Orlicz function if it is nonnegative and increasing and</p><p>If &#981; is an Orlicz function and, in addition, it is convex and satisfies</p><p>then we say that it is an N-function.</p><p>For an N-function &#981;, we define its corresponding Orlicz space as</p><p>We refer the reader to <ref type="bibr">[18]</ref> for further properties of such spaces. Given an N-function &#981;, we say that &#981; &#8712; &#8711; 2 if there exists a &gt; 1 such that</p><p>Corollary 3.3 (Orlicz stability). In the setting of Theorem 3.1 let &#981; &#8712; &#8711; 2 . Then</p><p>where the implicit constant is independent of u and h.</p><p>Proof. According to [17, Theorem 1.2.1(v)], if &#981; &#8712; &#8711; 2 , then the maximal function is bounded on L &#981; (&#8486;). Apply Theorem 3.1 to conclude.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>Remark 3.4 (Simonenko indices). Given an N-function &#981; define</head><p>The upper and lower Simonenko indices of &#981; are, respectively,</p><p>We comment that &#981; &#8712; &#8711; 2 implies p - &#981; &gt; 1 so that the condition in Corollary 3.3 is consistent with the results of Corollary 3.2.</p><p>On the other hand, we say that an N-function is power-like if p + &#981; &lt; &#8734;. According to <ref type="bibr">[26]</ref> (see also <ref type="bibr">[16]</ref>), the space L &#981; (&#8486;) is an intermediate space between L p (&#8486;) and L q (&#8486;) provided the Simonenko indices satisfy</p><p>Thus, in the case of &#981; &#8712; &#8711; 2 and power-like, the results of Corollary 3.3 could be obtained by interpolation. Since, however, we are not assuming p + &#981; &lt; &#8734;, this is truly a new result.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.3.">Muckenhoupt weighted spaces.</head><p>Next we extend the results of <ref type="bibr">[10]</ref> to the optimal range of indices. We recall that a function 0</p><p>where the supremum is over all cubes Q &#8838; &#8486; with sides parallel to the coordinate axes. Weighted Lebesgue spaces are defined, for p &#8712; (1, &#8734;) and &#969; &#8712; A p , as </p><p>where the implicit constant is independent of u and h.</p><p>Copyright &#169; by SIAM. Unauthorized reproduction of this article is prohibited. Downloaded 08/02/24 to 160.36.108.101 . Redistribution subject to SIAM license or copyright; see <ref type="url">https://epubs.siam.org/terms-privacy</ref> Proof. It suffices to recall that, provided &#969; &#8712; A p , the Hardy-Littlewood maximal operator is bounded on weighted spaces; see <ref type="bibr">[12,</ref><ref type="bibr">Theorem 7.1.9(b)</ref>].</p><p>As we mentioned above, this result generalizes <ref type="bibr">[10,</ref><ref type="bibr">Corollary 3.3]</ref>, where such an estimate is obtained, but with &#969; &#8712; A p/2 , a strictly smaller class.</p><p>3.4. Weighted Lorentz spaces. Let &#969; be a weight. Here we are concerned with weighted Lorentz spaces L p,q (&#969;, &#8486;); i.e., the measure in (3.2) is &#181; = &#969; dx.</p><p>Corollary 3.6 (weighted stability). In the setting of Theorem 3.1 let p &#8712; (1, &#8734;), q &#8712; (1, &#8734;], and &#969; &#8712; A p . Then</p><p>where the implicit constant is independent of u and h.</p><p>Proof. According to [17, Theorem 5.2.1], given the range of exponents, we have that M : L p,q (&#969;, &#8486;) &#8594; L p,q (&#969;, &#8486;) boundedly if &#969; &#8712; A p .</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.5.">Weighted variable exponent spaces.</head><p>As a final application we mention weighted variable exponent spaces. A variable exponent is p &#8712; L 0 (&#8486;) such that p(&#8486;) &#8834; [1, &#8734;]. Given a variable exponent and a weight 0 &#8804; &#969; &#8712; L 1 loc (&#8486;) we define weighted variable exponent Lebesgue spaces as</p><p>We refer the reader to <ref type="bibr">[6,</ref><ref type="bibr">9]</ref> for an extensive treatise on these spaces. Given a variable exponent p, we say that p &#8712; P log (&#8486;) if</p><p>and, there is p &#8734; &#8805; 1 such that</p><p>If p is a variable exponent, then p &#8242; is its H&#246;lder conjugate, that is, the variable exponent that satisfies 1 p(x)</p><p>for almost every x &#8712; &#8486;. We say that the weight &#969; satisfies the generalized Muckenhoupt condition, denoted by &#969; &#8712; A, if</p><p>&#8776; |Q| for every cube Q with sides parallel to the coordinate axes. Here &#967; Q is the characteristic function of Q.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>Remark 3.7 (A versus</head><p>Thus, we see that &#969; &#8712; A is equivalent to &#181; = &#969; p &#8712; A p .</p><p>Copyright &#169; by SIAM. Unauthorized reproduction of this article is prohibited. Downloaded 08/02/24 to 160.36.108.101 . Redistribution subject to SIAM license or copyright; see <ref type="url">https://epubs.siam.org/terms-privacy</ref> Corollary 3.8 (variable exponent stability). Under the assumptions of Theorem 3.1 let p &#8712; P log (&#8486;) with ess inf x&#8712;&#8486; p(x) &gt; 1, and &#969; &#8712; A. Then</p><p>, where the implicit constant is independent of u and h.</p><p>Proof. Under the given assumptions the Hardy-Littlewood maximal operator is bounded on L p(&#8226;) &#969; (&#8486;); see <ref type="bibr">[4]</ref>, <ref type="bibr">[9,</ref><ref type="bibr">Theorem 4.3.8]</ref>, and [9, Theorem 5.8.6].</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.6.">Other extensions and variations.</head><p>As we mentioned after Theorem 3.1, the list we have provided is not exhaustive. For instance, under certain conditions, one can also assert the boundedness in Orlicz-Musielak spaces <ref type="bibr">[5]</ref>.</p><p>On the other hand, there are some spaces where the stability remains open. Notable examples are H 1 (&#8486;), the atomic Hardy space, and L 1 (&#8486;).</p><p>4. Proof of the main result. We now focus on the proof of Theorem 3.1. The technique that we shall follow will be a combination of weighted norm inequalities, as in <ref type="bibr">[23]</ref>, and local estimates, as presented in <ref type="bibr">[7]</ref>. We shall also rely on some estimates on the Green's function that hold, for n &#8712; {2, 3}, in convex polytopes. Proposition 4.1 (Green's function estimates). Let &#8486; &#8834; R n , with n &#8712; {2, 3}, be a convex polytope and G : &#8486; &#215; &#8486; &#8594; R be the Green's function associated to this domain. Then, for every i &#8712; {1, . . . , n},</p><p>In addition, there is &#945; &#8712; (0, 1], depending only on the inner angles of &#8486;, such that, for every i, j &#8712; {1, . . . , n} and all x, y, &#958; &#8712; &#8486;, we have</p><p>Proof. The first bound can be found in <ref type="bibr">[14,</ref><ref type="bibr">Theorem 3.3(iv)</ref>] for n &#8805; 3 and [11, Proposition 1 (9)] for n = 2.</p><p>In the case n = 3, the H&#246;lder estimates on the first and mixed derivatives can be found in <ref type="bibr">[15, formula (1.4)</ref>]. When n = 2 [10, Lemma 2.1] presents a proof of the estimate for the mixed derivative. The estimate on the first derivative follows the same proof presented in <ref type="bibr">[15]</ref>.</p><p>Notice that the H&#246;lder estimates on derivatives of G are the only instances where our dimensional restriction plays a role. As soon as the estimates in Proposition 4.1 are valid for more dimensions, the proof of Theorem 3.1 follows verbatim. <ref type="bibr">[20,</ref><ref type="bibr">21]</ref> relies on the construction of a regularized distance function and its properties. Here we rephrase some of the properties of such a function that may help elucidate the reason for its use. For K, &#947; &gt; 0 to be chosen we define &#981; 1 : R n &#8594; R by</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4.1.">Approximation of identity. The technique of weighted norms</head><p>Copyright &#169; by SIAM. Unauthorized reproduction of this article is prohibited. Downloaded 08/02/24 to 160.36.108.101 . Redistribution subject to SIAM license or copyright; see <ref type="url">https://epubs.siam.org/terms-privacy</ref> where c 1 is such that R n &#981; 1 (x) dx = 1. Now, for &#491; &gt; 0 and z &#8712; &#8486;, we define</p><p>Notice that the family {&#981; &#491; } &#491;&gt;0 is an approximation of the identity.</p><p>Lemma 4.2 (convolution estimate). For every &#491; &gt; 0, z &#8712; &#8486;, and f &#8712; L 0 (&#8486;) we have</p><p>where the constant is independent of &#491;, z, and f . Here, f is extended by zero outside of &#8486;.</p><p>Proof. Since the function &#981; 1 is radial and decreasing, it suffices to invoke Theorem 2.2 of section 2.2 in <ref type="bibr">[27]</ref>; see also <ref type="bibr">[12,</ref><ref type="bibr">Theorem 2.1.10]</ref>.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4.2.">Regularized Green's function.</head><p>To establish our main estimate we shall rely on a pointwise representation. We fix h &gt; 0 and let z &#8712; &#8486; be such that z &#8712; T for some T &#8712; T h . Owing to shape regularity, see <ref type="bibr">[19, formula (12)</ref>], [28, Lemma 2.2] and <ref type="bibr">[25]</ref>; there is a function</p><p>Fix l &#8712; {1, . . . , n}. The regularized Green's function is</p><p>Owing to the fact that the right-hand side in (4.2) is compactly supported in &#8486;, we can, using Proposition 4.1, obtain some H&#246;lder regularity for g z . This is the content of the following result. Proposition 4.3 (estimates on g z ). Let z &#8712; T &#8712; T h and g z solve (4.2). Then, for every i &#8712; {1, . . . , n} and all x, y / &#8712; T , x = y, we have</p><p>where the exponent &#945; &#8712; (0, 1] is the same as in Proposition 4.1. Moreover,</p><p>Proof. We begin by using the pointwise representation of g z in terms of the Green's function G, and the fact that &#948; z is supported on T to obtain</p><p>We now invoke Proposition 4.1 to obtain</p><p>Copyright &#169; by SIAM. Unauthorized reproduction of this article is prohibited. Downloaded 08/02/24 to 160.36.108.101 . Redistribution subject to SIAM license or copyright; see <ref type="url">https://epubs.siam.org/terms-privacy</ref> as claimed.</p><p>To obtain the second estimate we observe that &#948; z is supported on T and use its scaling properties to assert that, for any i &#8712; {1, . . . , n}, we have</p><p>All estimates have been proved.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>4.3.</head><p>Step 1: Pointwise representation. We now begin with the proof of Theorem 3.1 per se. Owing to the properties of &#948; z we have that</p><p>where we used (4.2) and the definition of the Ritz projection (2.3). From the definition of &#948; z it follows immediately that</p><p>On the other hand, we estimate the second term as</p><p>.</p><p>Owing to Lemma 4.2,</p><p>Thus, if we define</p><p>we see that the heart of the matter is to provide a uniform, in h, estimate for this quantity.</p><p>In summary, the rest of the proof consists in showing the following result.</p><p>Proposition 4.4 (uniform estimate). In the setting of Theorem 3.1 there are K &gt; k 0 and &#947; &#8712; (0, &#945;) such that, if &#981; 1 is defined as in (4.1), we have</p><p>where the constant is independent of h &gt; 0, and G h was defined in (4.3). Here k 0 is as in Proposition 2.1, and &#945; is as in Proposition 4.1.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>4.4.</head><p>Step 2: Dyadic decomposition. Fix z &#8712; &#8486;. Define, for j &#8712; N 0 , d j = 2 j Kh. We decompose the domain &#8486; into the following annuli:</p><p>with the convention that, for j &lt; 0, B j = &#8709;. For S &#8834; &#8486; we also define</p><p>The use of this dyadic decomposition lies in the fact that on each annulus the regularized distance function &#981; h,z is almost constant. Lemma 4.5 (distance estimates). Assume that K &gt; 2. For all j &#8805; 0 we have</p><p>where the implicit constants are independent of h. As a consequence, for j &#8805; 3 we have</p><p>where &#945; &#8712; (0, 1] is as in Proposition 4.1.</p><p>Proof. The estimate on &#981; h,z follows by definition. The estimate on the distance between A + j and N h (&#8486; \ A ++ j ) does so as well. On the other hand, if j &#8805; 3, x, y &#8712; A ++ j , and &#958; &#8712; T , then |x -&#958;|, |y -&#958;| &#8776; d j . We can then refine the estimate of Proposition 4.3 to conclude</p><p>4.5.</p><p>Step 3: Reduction to interpolation and duality. Now fix some z &#8712; &#8486; and define</p><p>.</p><p>Let j &#8712; N 0 now be such that</p><p>.</p><p>Using the distance estimates of Lemma 4.5, we can also assert that</p><p>where k 0 was introduced in Proposition 2.1. Then we have diam A j &#8776; d j &#8805; k 0 h, so that with a simple covering argument we may obtain that</p><p>where &#928; h is, for instance, the so-called Scott-Zhang interpolant <ref type="bibr">[24]</ref>, or any other interpolant satisfying local stability and approximation properties. The first two terms will be handled using interpolation estimates, whereas the last one is controlled by duality.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>4.6.</head><p>Step 4: Bound of I + II via interpolation estimates. It is our goal now to bound I + II using the approximation properties of &#928; h and the regularity of g z . This regularity, however, depends on the distance between A + j and z. If j &#8805; 3, then we can invoke the estimate in Lemma 4.5 to see that</p><p>As a consequence, since 0 &lt; &#947; &lt; &#945;,</p><p>If, on the other hand, j &lt; 3, we use the second bound of Proposition 4.3 to obtain</p><p>In this case then we get</p><p>Gathering (4.6) and (4.7), we arrive at</p><p>Step 5: Bound of III by duality. We bound III by duality. Define</p><p>Then, by Galerkin orthogonality,</p><p>An application of H&#246;lder's inequality then allows us to conclude that</p><p>Notice that if, in this last estimate, the term that is multiplying G h is sufficiently small, then it could be absorbed on the left-hand side in <ref type="bibr">(4.5)</ref>. This possibility is explored in the following result.</p><p>Lemma 4.6 (duality bound). Let S j be defined as in (4.9) and &#947; &#8712; (0, &#945;). There is a constant, independent of j, z, and h, such that</p><p>where w v &#8712; W 1,2 0 (&#8486;) is the solution to (4.10) and &#945; is as in Proposition 4.1. Proof. Let v &#8712; S j be arbitrary. Using Lemma 4.5 and scaling, we have</p><p>where, since &#8486; is convex, we used a regularity estimate on w v .</p><p>To control the norm in &#8486; \ A ++ j observe that, owing to the estimates of Proposition 4.1, for every i &#8712; {1, . . . , n} we have for every x, y &#8712; &#8486; \ A ++ To shorten notation let e w = w v -&#928; h w v . We use that &#981; h,z L 1 (&#8486;) = 1 and the recently obtained regularity estimate to proceed as follows:</p><p>We combine both bounds to conclude. With Lemma 4.6 at hand we conclude that </p><p>where k 0 is defined as in Proposition 2.1; and &#947; &#8712; (0, &#945;), with &#945; as in Proposition 4.1. Notice the hidden constant here is independent of z. Since G h = sup z&#8712;&#8486; G h,z , we can now, if necessary, choose an even bigger K to conclude the proof of Proposition 4.4 and, as a consequence, that of Theorem 3.1.</p></div></body>
		</text>
</TEI>
