Abstract
We introduce a new quantification of nonuniform ellipticity in variational problems via convex duality, and prove higher differentiability and 2d-smoothness results for vector valued minimizers of possibly degenerate functionals. Our framework covers convex, anisotropic polynomials as prototypical model examples—in particular, we improve in an essentially optimal fashion Marcellini’s original results (Marcellini in Arch Rat Mech Anal 105:267–284, 1989).
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Motivated by the regularity theory for elliptic systems driven by differential operators that are homogeneous polynomials in the derivative symbols obtained by Douglis and Nirenberg [28], Ladyzhenskaya and Ural’tseva [56], and Morrey [68], we introduce a new class of autonomous variational integrals of the type
governed by a nonuniformly elliptic integrand \(F\in {\text {C}}^{2}(\mathbb {R}^{N\times n})\) in the sense that the related ellipticity ratio
might blow up on large values of the derivative variable. We obtain higher differentiability and low-dimensional smoothness results for local minimizers, whose standard definition reads as follows:
Definition 1.1
A map \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^N)\) is a local minimizer of functional \(\mathscr {F}\) if for any open subset \(\Omega '\Subset \Omega \) it satisfies \(\mathscr {F}(u;\Omega ')<\infty \) and \(\mathscr {F}(u;\Omega ')\le \mathscr {F}(w;\Omega ')\) for all \(w\in u+{\text {W}}^{1,p}_{0}(\Omega ',\mathbb {R}^N)\).
More precisely, we focus on a large class of convex integrands satisfying so-called (p, q)-growth conditions, according to Marcellini’s foundational works [59, 61, 63],
for all \(z\in \mathbb {R}^{N\times n}\), and some exponents \(2\le p\le q<\infty \), including anisotropic convex polynomials as model example. This family of integrands is characterized by a potentially wild behavior of the ellipticity ratio at infinity, and the rate of blow up of \(\mathcal {R}_{F}\) as \(|z|\rightarrow \infty \) is measured in terms of the stress tensor, i.e.:
We shall refer to functionals as (1.1) with integrands satisfying (1.2), (1.3) as Legendre (p, q)-nonuniformly elliptic integrals, terminology justified since (1.3) comes as the quantification of the interaction between F and its Fenchel conjugate \(F^{*}\), the strong convexity of F, and the (p, q)-growth conditions in (1.2), see Sect. 3 for more details, and [29, 71] for the abstract convex analytic setting. We also stress that by convexity, in our local setting, the pointwise definition of functional \(\mathscr {F}\) in (1.1) coincides with its Lebesgue-Serrin-Marcellini extension [55, 59, 62]. The class of nonuniformly elliptic problems we propose falls into the (slightly larger) realm of functionals with controlled (p, q)-growth, first studied by Marcellini [59, 61, 63] in connection to some delicate issues of compressible elasticity including the phenomenon of cavitation [2, 59, 62]. Roughly speaking, variational integrals with (p, q)-growth are driven by possibly anisotropic integrands satisfying the unbalanced growth condition (1.2), whose ellipticity ratio can at most be controlled via
By convexity, (1.2) and (1.3), it follows that Legendre (p, q)-nonuniform ellipticity (1.3) implies the usual (p, q)-nonuniform ellipticity (1.4). The crucial aspect of these problems is the subtle, quantitative relation existing between the rate of blow up of the ellipticity ratio and the regularity of minima: in fact, Marcellini [60, 63] and Giaquinta [38] exhibited examples of (p, q)-nonuniformly elliptic functionals with unbounded minimizers provided that the exponents (p, q) violate a closeness condition of the type
where \(\texttt {o}(n)\searrow 0\) as \(n\rightarrow \infty \). It is then natural to relate the regularity of minima to the possibility of slowing down the rate of blow up of the ellipticity ratio by choosing p and q not too far apart, cf. (1.5): in a nutshell, this was Marcellini’s approach [61] to local Lipschitz continuity for scalar minimizers of certain anisotropic energies such as
where
for exponents
After Marcellini’s initial success, (p, q)-nonuniformly elliptic functionals have been the object of intensive investigation: considerable attention has been focused on full regularity [4,5,6, 8,9,10,11,12, 15, 17, 20, 23,24,25, 30,31,32, 46, 74], boundary regularity [1, 13, 19, 26, 49, 51, 52], nonlinear potential theoretic results [3, 5, 8, 19, 22, 24,25,26,27] and partial regularity [22, 27, 34, 42, 44, 57, 73, 75,76,77], see also [58] for a reasonable survey - recently, (p, q)-nonuniformly elliptic regularity theory found interesting applications to nonlinear homogenization [21, 72]. Note in particular that Marcellini’s variational approach, originally designed to handle polynomial rates of nonuniformity, turns out to fit also more general, possibly nonautonomous problems whose ellipticity ratio blows up faster than powers; see [65] for the very first results on functionals at fast exponential growth, [5, 24] for more recent progress, and [33] for applications to weak KAM theory. Back to (p, q)-nonuniform ellipticity, it is worth mentioning that the optimal value of the threshold to impose on the ratio q/p guaranteeing gradient regularity is known in the autonomous setting only in the scalar case [46, 63]. It covers boundedness of minima, leaving open the delicate matter of gradient regularity. In this respect, at first Marcellini [61] determined two bounds on the size of q/p for Lipchitz continuity of minima: first, (1.7), formulated for the model (1.6), that in its general form reads as
which was found in [63, 64] for minima belonging to higher energy classes (e.g. \({\text {W}}^{1,q}\)-minimizers), and second, the more restrictive (1.9), that allows deriving gradient boundedness for \({\text {W}}^{1,p}\)-minimizers. Actually, this second setting is the natural framework to investigate, as given the growth/coercivity prescribed in (1.2) it is only possible to prove the existence of minima in \({\text {W}}^{1,p}\). The constraint given by (1.9) is far from being optimal: under natural growth conditions and in the vectorial case, Carozza et al. [17] obtained gradient higher differentiability and as a consequence that minima belong to \({\text {W}}^{1,q}_\textrm{loc}\), provided that
by means of convex duality methods. The earlier results did not use duality theory and required stronger natural growth conditions, see in particular [31]. When considering Lipschitz continuity of \({\text {W}}^{1,p}\)-minima in the scalar case, a first, substantial improvement is due to Bella and Schäffner [6, 8] in the genuine controlled growth (p, q)-setting, updating (1.9) to
by using a refinement of De Giorgi’s iteration technique via optimization on radial cut-off functions, leading to the application of Sobolev inequality on spheres rather than on balls. Under Legendre (p, q)-growth conditions, we are able to further relax these bounds, both with regard to obtaining higher differentiability results and to proving Lipschitz regularity in the scalar case. More precisely, for \({\text {W}}^{1,p}\)-minimizers of Legendre (p, q)-nonuniformly elliptic integrals we allow a faster blow-up rate of the ellipticity ratio, quantifiable in terms of exponents (p, q) as
We remark that this constraint had previously been obtained by Bella and Schäffner [6] provided the minimizer is a priori assumed to be of class \({\text {W}}^{1,q}\). Our approach is based on the duality between the gradient of minima and the related stress tensors, a technique that in this context finds its roots in [17]. This is at the heart of the main result of this paper, which is gradient higher differentiability for minimizers of variational integrals satisfying Legendre (p, q)-growth.
Theorem 1.2
Under assumptions (1.12), and (2.1), (2.2), let \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) be a local minimizer of functional \(\mathscr {F}\). Then,
the second inclusion in (1.13) holding in the nondegenerate case \(\mu >0\) in (2.1)\(_{3}\). In particular, if \(B\Subset \Omega \) is any ball with radius less than one,
holds true with \(c\equiv c(n,N,L,L_{\mu },p,q)\), and \({\texttt {b}}\equiv {\texttt {b}}(n,p,q)\).
We immediately refer to Sect. 2.2 for the precise description of our assumptions and of the various quantities mentioned above. The main building block in the proof of Theorem 1.2 and the possibility of enlarging the ellipticity of F by means of convex duality arguments in such a way that it quantitatively controls both gradient of minima and stress tensor, is granted by our Legendre (p, q)-growth, thus yielding essentially optimal results, [60]. As a consequence of the boost of integrability earned via Sobolev embedding from Theorem 1.2, we also derive a lower order regularity result in the spirit of Campanato [14, Theorem 1.IV].
Corollary 1.3
Under assumptions (1.12), and (2.1), (2.2), let \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) be a local minimizer of functional \(\mathscr {F}\). The following holds:
-
(i.)
if \(p>n-2\) and \(n\ge 3\), then \(u\in {\text {C}}^{0,1-\frac{n-2}{p}}_{{\text {loc}}}(\Omega , \mathbb {R}^{N})\);
-
(ii.)
if in addition
$$\begin{aligned} 2\le p\le q\le \frac{np}{n-2}\quad \text{ if } \ \ n\ge 3\qquad \text{ and }\qquad 2\le p\le q<\infty \quad \text{ if } \ \ n=2,\nonumber \\ \end{aligned}$$(1.15)then \(\nabla u\in \textrm{L}^{q}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N\times n})\);
-
(iii.)
if \(n=2\) then \(u\in {\text {C}}^{0,\beta _{0}}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\cap {\text {W}}^{1,m}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) and \(F'(\nabla u)\in \textrm{L}^{m}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N \times n})\) for all \(\beta _{0}\in (0,1)\), \(m\in [1,\infty )\).
The higher differentiability granted by Theorem 1.2 allows deriving full regularity in two ambient space dimensions.
Theorem 1.4
In ambient space dimension \(n=2\), suppose that (1.12), and (2.1), (2.2) are in force. Then, any local minimizer \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) of functional \(\mathscr {F}\) has locally Hölder continuous gradient. In particular, on all balls \(B\Subset \Omega \), the \(\textrm{L}^{\infty }\)-\(\textrm{L}^{p}\) type estimate
holds with \(c\equiv c(N,L,L_{\mu },p,q)\) and \({\texttt {b}}\equiv {\texttt {b}}(p,q)\). Moreover, given any two open subsets \(\Omega _{2}\Subset \Omega _{1}\Subset \Omega \) with \(\,\textrm{dist}(\Omega _{2},\partial \Omega _{1})\approx \,\textrm{dist}(\Omega _{1},\partial \Omega ) \), there exists \({\texttt {t}}\equiv {\texttt {t}}(N,L,L_{\mu },p,q,\mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2},\partial \Omega _{1}))>1\) such that \(V_{p}(\nabla u), V_{1,q'} (F'(\nabla u))\in {\text {W}}^{1,2{\texttt {t}}}(\Omega _{2},\mathbb {R}^{N\times 2})\) and the reverse Hölder type inequality
is verified for any ball \(B\Subset \Omega _{2}\), with \(c\equiv c(N,L,L_{\mu },p,q,\mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2}, \partial \Omega _{1}))\), \({\texttt {b}}\equiv {\texttt {b}}(p,q)\). Finally,
-
if \(\mu >0\) in (2.1), then the gradient of minima is locally Hölder continuous up to any exponent less than one;
-
if \(\mu >0\) in (2.1), and integrand F is real analytic, then u is real analytic.
Theorems 1.2–1.4 and Corollary 1.3 cover a large number of models, including (1.6)–(1.8), and more generally, convex polynomials. Actually, restating Corollary 1.3 and Theorem 1.4 for vector-valued minimizers of functionals driven by strongly convex, even polynomials, we extend to the nonuniformly elliptic framework early results of Morrey [67, 68].
Theorem 1.5
Let \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) be a local minimizer of functional \(\mathscr {F}\), where the integrand F is a convex, even polynomial with nonnegative homogeneous components, lowest homogeneity degree larger or equal than p, and satisfying the lower bound in (2.1\(_{3}\)). The following holds:
-
(i.) in three space dimensions \(n=3\), minima are locally \((p-1)/p\)-Hölder continuous;
-
(ii.) in two space dimensions \(n=2\), if \(\mu >0\) in (2.1)\(_{3}\) then minima are real analytic.
It is worth highlighting that convex, even polynomials with nonnegative homogeneous components automatically satisfy our Legendre (p, q)-nonuniform ellipticity condition,Footnote 1 see Sect. 3.1—in particular, Theorem 1.5 holds with no restriction on the degree of the polynomial. In the scalar setting, combining Theorem 1.2 with a homogenised Moser iteration argument, we obtain Lipschitz regularity of minima.
Theorem 1.6
Under Assumptions (1.12), and (2.1), (2.2), let \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega )\) be a local minimizers of functional \(\mathscr {F}\). Then, \(u\in {\text {W}}^{1,\infty }_{{\text {loc}}}(\Omega )\). In particular, whenever \(B\Subset \Omega \) is a ball with radius \({\texttt {r}}\in (0,1]\), the Lipschitz bound
holds with \(c\equiv c(n,L,L_{\mu },p,q)\) and \({\texttt {b}}\equiv {\texttt {b}}(n,p,q)\).
We remark that if \(p=2\), the bound in (1.12) corresponds precisely to the one violated by the counterexample in [63, Theorem 6.1]. This highlights the sharpness of our results for polynomial-type integrals with quadratic growth from below. Notice further that in the vectorial case, and when the ambient space dimension n is larger than two, minimizers need not be locally Lipschitz. In fact, this is already the case when the integrand is smooth and uniformly elliptic, as shown by Šverák and Yan [79] for dimensions \(n \ge 3\), \(N \ge 5\). More recently, Mooney and Savin [66] gave an example of a smooth and uniformly elliptic integrand on \(2 \times 3\) matrices, so dimensions \(n=3\), \(N=2\), such that the corresponding variational integral admits a minimizer that is Lipschitz but not \({\text {C}}^1\). It seems likely that their approach can be adapted to give examples of non-Lipschitz minimizers also in those dimensions. The remainder of the section is devoted to a description of some key technical points appearing throughout the paper that we believe could be of wider use. We end with a quick outline of the structure of the paper.
1.1 Techniques
From the technical point of view, the main contribution of this paper is threefold. First, we introduce a new way to control the rate of blow up of the ellipticity ratio for superlinear, possibly degenerate functionals, that we refer to as the Legendre (p, q)-nonuniform ellipticity condition (1.3). As already mentioned, (1.3) is slightly more restrictive than the genuine (p, q)-nonuniform ellipticity (1.4), nonetheless it covers the same models, i.e. convex polynomials or anisotropic energies, under essentially optimal bounds on exponents (p, q), cf. (1.12)—in particular, in two and three space dimensions the ellipticity ratio is allowed to blow up arbitrarily fast, meaning for polynomials that no upper bound on the degree is required. Our results crucially rely on a subtle interplay between the gradient of minima and the stress tensor, that can be handled via convex duality. This is a natural strategy, and convex duality tools have already been employed by Zhikov et al. [50] in connection to homogenized elasticity theory, by Seregin [78] in the theory of plasticity, and by Carozza et al. [16], and Koch and Kristensen [53] on the validity of the Euler-Lagrange system. The key aspect in our approach is the duality between the gradient of minima and the related stress tensor that, via (1.3) and rather elementary convex duality arguments, can be made quantitative, so that the ellipticity ratio of the integrand F is enlarged as a natural consequence of the combination between convexity and extremality conditions—in particular, primal and dual problems can be handled simultaneously. The second main novelty consists in a streamlining of Bella and Schäffner’s optimization trick [6], that allows us to work directly on spheres rather than on solid balls and eventually leads to the (surprisingly simple) proof of gradient higher differentiability of \({\text {W}}^{1,p}\)-minimizers under the constraint (1.12). This brings us to the account of the third main novelty of this paper, that is a new proof of full regularity for vector-valued minimizers of possibly degenerate, nonuniformly elliptic functionals in two space dimensions, obtained by means of a monotonicity argument in the spirit of Frehse [36], or via a renormalized version of classical Gehring–Giaquinta and Modica Lemma [37, 39], following the useful point of view introduced by Beck and Mingione [5]. This is a well-known issue in the multidimensional Calculus of Variations, handled for p-Laplacian type problems via Gehring–Giaquinta and Modica Lemma applied right after differentiating the Euler–Lagrange system of \(\mathscr {F}\), see previous contributions by Nečas [69], Giaquinta and Modica [39], and Campanato [14]. The corresponding 2d-smoothness result for nondegenerate, genuinely (p, q)-nonuniformly elliptic integrals, is a more recent achievement of Bildhauer and Fuchs [9], whose strategy strongly relies on the existence in \(\textrm{L}^{2}\) of second derivatives—a distinctive feature of minima of nondegenerate integrals that dramatically fails already for the degenerate p-Laplacian [81]. For this reason, Theorem 1.4 comes by no means as an adaptation of the arguments in [9], and covers degenerate problems. In this respect, we offer two independent proofs of 2d-smoothness. The first one follows the arguments outlined in the proof of Theorem 1.2 to derive a monotonicity formula resulting in a logarithmic Morrey-type decay for the \(\textrm{L}^{2}\)-norm of certain nonlinear functions of the gradient, eventually implying \({\text {C}}^{1}\)-regularity of minima. The second one is based on a quantitative version of Gehring–Giaquinta and Modica lemma [37, 39], applied to the differentiated Euler-Lagrange system after bounding the ellipticity ratio via a power of the \(\textrm{L}^{\infty }\)-norm of the gradient, so that it becomes uniformly elliptic—constants are then carefully tracked at each stage of the proof. Both our arguments are inspired by the principle that a suitable control on the rate of blow up of the ellipticity ratio associated to nonuniformly elliptic functionals gives access to a technical toolbox that has classically been employed in the Lipschitz regularity theory for uniformly elliptic structures, such as De Giorgi’s level sets technique [5, 8, 24,25,26], homogenized Moser’s iteration [23] and now monotonicity formula and Gehring–Giaquinta and Modica lemma. Notice that after minor modifications, Theorem 1.4 embraces also genuine (p, q)-nonuniformly elliptic integrals, cf. Remark 5.3 below.
1.1.1 Outline of the paper
In Sect. 2 we describe our notation and collect a number of auxiliary results that will be helpful at various stages of the paper. In Sect. 3 we derive several important consequences of our Legendre (p, q)-growth via convex duality arguments, and provide examples of integrands satisfying it, notably in the form of even, convex polynomials. In Sect. 4 we establish higher differentiability and higher integrability results for a suitable nonlinear function of the gradient of minima, and of the related stress tensor. Section 5 contains the proof of full regularity in the two-dimensional case, while finally Sect. 6 contains the proof of Lipschitz regularity in the scalar setting.
2 Preliminaries
In this section we display our notation, describe the main assumptions governing the functional \(\mathscr {F}\), and collect some basic results that will be helpful throughout the paper.
2.1 Notation
In this paper, \(\Omega \subset \mathbb {R}^n\), \(n\ge 2\), will always be an open domain. We denote by c a general constant larger than one, possibly depending on various parameters related to the problem under investigation. We will still denote by c distinct occurrences of the constant c from line to line. Specific instances will be marked with symbols \(c_*, {\tilde{c}}\) or the like. Significant dependencies on certain parameters will be outlined by putting them in parentheses, i.e. \(c\equiv c(n,p)\) means that c depends on n and p. Sometimes we shall employ symbols "\(\lesssim \)", "\(\approx \)" or "\(\gtrsim \)" to indicate that an inequality holds up to multiplicative constants depending on basic parameters governing our problem. By \( B_r(x_0):= \{x \in \mathbb {R}^n : |x-x_0|< r\}\) we indicate the open ball with center in \(x_0\) and radius \(r>0\), while \(Q_{r}(x_{0}):=\left\{ x\in \mathbb {R}^{n}:\max _{i\in \{1,\cdots ,n\}}|x_{i}-x_{0;i}|<r\right\} \) denotes the open cube with half-side length equal to r - when clear from the context, we shall avoid specifying radius or center, i.e., \(B \equiv B_r \equiv B_r(x_0)\), \(Q\equiv Q_{r}\equiv Q_{r}(x_{0})\); this happens in particular with concentric balls or concentric cubes. In particular, with B being a given ball with radius \({\texttt {r}}\) and \(\gamma \) being a positive number, we denote by \(\gamma B\) the concentric ball with radius \(\gamma {\texttt {r}}\) and by \(B/\gamma := (1/\gamma )B\). For a number \(t\in [1,\infty ]\), its conjugate exponent \(t'\in [1,\infty ]\) is defined as \(t':=(1-1/t)^{-1}\) if \(t>1\), \(t':=\infty \) if \(t=1\), and \(t'=1\) if \(t=\infty \), while, given \({\texttt {n}}\in \mathbb {N}\), its \({\texttt {n}}\)-dimensional Sobolev exponent, and its \({\texttt {n}}\)-dimensional "lower" Sobolev exponent are given by \(t^{*}_{{\texttt {n}}}:={\texttt {n}}t/({\texttt {n}}-t)\) if \(1\le t<{\texttt {n}}\) and \(t^{*}_{{\texttt {n}}}:=\text{ any } \text{ number } \text{ larger } \text{ than } t\) if \(t\ge {\texttt {n}}\), and \(t_{*;{\texttt {n}}}:=\max \{1,{\texttt {n}}t/({\texttt {n}}+t)\}\) respectively - here, \({\texttt {n}}\) will be either chosen as \({\texttt {n}}=n\) or \({\texttt {n}}=n-1\). Moreover, if \(A \subset \mathbb {R}^{n}\) is a measurable set with bounded positive Lebesgue measure \(|A|\in (0,\infty )\), and \(g :A \rightarrow \mathbb {R}^{d}\), \(d\ge 1\), is a measurable map, we set to indicate its integral average, while if \(g\in \textrm{L}^{\gamma }(B,\mathbb {R}^{d})\) for some \(\gamma >1\), we shorten its averaged norm as . Furthermore, with \(z\in \mathbb {R}^{k}\), and \(\mu \ge 0\), we set \(\ell _{\mu }(z):=(\mu ^{2}+|z|^{2})^{1/2}\). Finally, if t is any parameter, we indicate by \(\texttt {o}(t)\) a quantity that is infinitesimal as \(t\rightarrow 0\) or \(t\rightarrow \infty \).
2.2 Structural Assumptions
Throughout the paper, \(F:\mathbb {R}^{N\times n}\rightarrow \mathbb {R}\), \(n\ge 2\), \(N\ge 1\), is an autonomous integrand satisfying
for all \(z, \xi \in \mathbb {R}^{N \times n}\), some absolute constants \(L> 1\), \(\mu \in [0,1]\), and exponents (p, q) satisfying (1.12). If \(\mu =0\) in (2.1), we need to prescribe a (rather natural, see Section 3 below) limitation on the rate of blow up of the ellipticity ratio:
where \(L_{0}>1\) is an absolute constant. Moreover, (2.1)\(_{3}\) assures the strict convexity of F, and, combined with (2.1)\(_{2}\) guarantees also that
for \(c\equiv c(L,p,q)\), see [61, Lemma 2.1]. Let us point out that in the nondegenerate case \(\mu >0\) in (2.1), the constraint in (2.2) comes as a consequence of (2.1)\(_{3}\) and (2.3) up to constants depending on \((L,p,q,\mu )\), so overall,
with \(L_{\mu }\equiv L_{0}\) if \(\mu =0\) and \(L_{\mu }=\mu ^{2-p}c(L,p,q)\) if \(\mu >0\) - in other words, thanks to (2.2) the dependency on \(\mu \) occurs only if \(\mu >0\).
Remark 2.1
Assumption (2.2) deserves some comment in relation to (2.1)\(_{3}\) and to (2.4), where a dependency on \(\mu \) appears in the constants. The restriction in (2.2) is needed for general degenerate integrands to control the rate of blow up of the ellipticity ratio, i.e., \(\mu =0\) in (2.1)\(_{3}\), while if \(\mu >0\) it is not needed at all as the amount of information contained in (2.1)\(_{3}\) suffices to derive the bound in (2.4) (with explicit dependency on \(\mu \)). This seems to be unavoidable, given the strong inhomogeneity displayed in the growth of \(F''\), cf. (2.1)\(_{3}\). On the other hand, if more homogeneous growth conditions are imposed on \(F''\), like those usually appearing in the literature on (p, q)-nonuniformly elliptic problems [5, 8, 23, 24, 26], prescribing that \( |F''(z)|\le L \ell _{\mu }(z)^{p-2}+L\ell _{\mu }(F'(z))^{\frac{q-2}{q-1}}\) replaces the right-hand side of (2.1)\(_{3}\), then (2.2) can be discarded and no dependency on \(\mu \) appears in (2.4).
Remark 2.2
If \(p=q\) our results are classical, see [14, 41] and references therein, so given that our approach is of interpolative nature, to avoid trivialities throughout the paper we shall permanently work assuming the strict inequality \(p<q\).
2.3 Functional Analytical Tools for Degenerate Problems
The vector field \(V_{\mu ,\gamma }:\mathbb {R}^{N \times n}\rightarrow \mathbb {R}^{N \times n}\), defined as
for all \(z\in \mathbb {R}^{N\times n}\), which encodes the scaling features of the p-Laplace operator, is a useful tool to handle singular or degenerate problems. A couple of helpful related equivalences are the following:
holding for any \(1<p<\infty \), \(z_{1},z_{2}\in \mathbb {R}^{N\times n}\), all measurable subsets \(A\subset \mathbb {R}^{n}\) (resp. \(A\subset \mathbb {R}^{n-1}\)) with positive, finite n-dimensional Lebesgue measure (resp. \((n-1)\)-dimensional Hausdorff measure), functions \(w\in \textrm{L}^{p\gamma /2}(A,\mathbb {R}^{N\times n})\), up to constants depending only on \((n,N,\gamma ,p)\), cf. [45, Section 2], and [40, (2.6)]. Let us also recall Sobolev embedding theorem and Sobolev–Poincaré inequalities on spheres, cf. [6, Section 3].
Lemma 2.3
Let \(\partial B_{\varrho }(x_{0})\) be an \((n-1)\)-dimensional sphere with \(n\ge 2\), \(k\in \mathbb {N}\) be a number, and \(w\in {\text {W}}^{1,1}(\partial B_{\varrho }(x_{0}),\mathbb {R}^{k})\) be a function. Then
-
if \(w\in {\text {W}}^{1,2}(\partial B_{\varrho }(x_{0}),\mathbb {R}^{k})\), then
(2.7)with \(c\equiv c(n,k)\);
-
if \(w\in {\text {W}}^{1,p}(\partial B_{\varrho }(x_{0}),\mathbb {R}^{k})\) for some \(1<p<\infty \), then
(2.8)for \(c\equiv c(n,p,k)\).
We will also need an "unbalanced" version of Poincaré inequality, [6, Section 3].
Lemma 2.4
Let \(B_{\varrho }(x_{0})\subset \mathbb {R}^{n}\), \(n\ge 2\), be a ball, \(\sigma >0\) a number, and \(w\in {\text {W}}^{1,2}(B_{\varrho }(x_{0}))\) be any function. Then
with \(c\equiv c(n,\sigma )\).
We further record classical Sobolev–Morrey embedding theorem with sharp bounding constant, obtained by applying [18, Theorem 1.1] componentwise.
Proposition 2.5
Let \(w\in W^{1,p}(\mathbb {R}^{n},\mathbb {R}^{k})\) with \(p>n\ge 2\), \(k\ge 1\) be a function such that \(|\,\textrm{supp }(w)|<\infty \). Then
We conclude this section with the "simple but fundamental" iteration lemma, [41, Chapter 6].
Lemma 2.6
Let \(h:[\varrho _{0},\varrho _{1}]\rightarrow \mathbb {R}\) be a non-negative and bounded function, and let \(\theta \in (0,1)\), \(A,B,\gamma _{1},\gamma _{2}\ge 0\) be numbers. Assume that \(h(t)\le \theta h(s)+A(s-t)^{-\gamma _{1}}+B(s-t)^{-\gamma _{2}}\) holds for all \(\varrho _{0}\le t<s\le \varrho _{1}\). Then the following inequality holds \(h(\varrho _{0})\le c(\theta ,\gamma _{1},\gamma _{2})[A(\varrho _{1}-\varrho _{0})^{-\gamma _{1}}+B(\varrho _{1}-\varrho _{0})^{-\gamma _{2}}].\)
3 Growth Conditions and Convex Duality
In this section we recall certain elementary notions from convex analysis [29, 47, 48, 71], and derive important consequences of the structural assumptions (2.1). We start by recording some basic facts on convex functions. The Fenchel conjugate of an integrand \(F :\mathbb {R}^{N \times n}\rightarrow \mathbb {R}\) is defined as the extended real-valued integrand \( \mathbb {R}^{N \times n}\ni \xi \mapsto F^{*}( \xi ) := \sup _{z \in \mathbb {R}^{N \times n}} \left( \langle z , \xi \rangle - F(z) \right) .\) Recall that the Fenchel conjugate \(F^*\) is real-valued precisely when F is super-linear at infinity, so precisely when \(F(z)/|z| \rightarrow \infty \) as \(|z| \rightarrow \infty \). In fact there is perfect symmetry here. Recall that the convex and lower semicontinuous envelope of F coincides with the double Fenchel conjugate \(F^{**} \equiv ( F^{*})^{*}\), and so, when F is convex, it is real-valued and super-linear precisely when its Fenchel conjugate \(F^*\) is so. For later reference, we also recall the Fenchel-Young inequality stating that
holds for all z, \(\xi \in \mathbb {R}^{N \times n}\). It is a direct consequence of the definition of Fenchel conjugation and it holds for any proper integrand F. Equality in (3.1) holds precisely when \(\xi \in \partial F(z)\), the subdifferential of F at z. In particular, we emphasize that if F is \({\text {C}}^1\)-regular, equality holds precisely when F is convex at z and \(\xi = F^{\prime }(z)\). It is worth highlighting that if F satisfies the (p, q)-growth conditions (2.1)\(_{2}\), these are transformed under Fenchel conjugation into
for some \(c\equiv c(L,p,q)\ge 1\), , cf. [17, Section 2] for more details. Let us point out that the integrand F is super-linear, strictly convex and \({\text {C}}^1\)-regular precisely when its Fenchel conjugate \(F^{*}\) is so. In this case we also have that both derivatives are homeomorphisms of \(\mathbb {R}^{N \times n}\) and that \((F^{*})^{\prime } = ( F^{\prime })^{-1}\). Let us record one of the implications, namely that when F is real-valued, super-linear, strictly convex and \({\text {C}}^1\)-regular, then so is \(F^*\) and
We further record the duality relations that exist between strict convexity and smoothness for an integrand and its Fenchel conjugate.
Lemma 3.1
Let \(F :\mathbb {R}^{N \times n}\rightarrow \mathbb {R}\cup \{ \infty \}\) be a convex and proper integrand. Then F is real-valued, super-linear and strictly convex if and only if its Fenchel conjugate \(F^*\) is real-valued, super-linear and \({\text {C}}^1\)-regular.
The following well-known result [48, Corollary 4.2.10] and its consequences play an important role in this paper.
Lemma 3.2
Let \(F\in {\text {C}}^{1}(\mathbb {R}^{N \times n})\) be super-linear and strictly convex. If \(F\in {\text {C}}^{2}(B_{r}(z_{0}))\) for some ball \(B_{r}(z_{0})\subset \mathbb {R}^{N \times n}\), and \(\textrm{det} (F^{\prime \prime }(z)) \ne 0\) for all \(z \in B_{r}(z_{0})\), then the Fenchel conjugate \(F^*\) is super-linear, strictly convex and \({\text {C}}^1\)-regular. In addition, \(F^*\in {\text {C}}^{2}(F^{\prime }(B_{r}(z_{0})))\) with \(( F^*)^{\prime \prime }(F^{\prime }(z) ) = F^{\prime \prime }(z)^{-1}\) for all \(z\in B_{r}(z_{0})\).
We now turn our attention to the growth/ellipticity condition (2.1)\(_{3}\), that is a quantitative form of Legendreness for real-valued \({\text {C}}^2\)-regular integrands. Recall indeed that Rockafellar in [71, Chapt. V] highlighted a special class of convex functions, called there functions of Legendre type, for their useful features in optimization theory. In the context considered in this paper, these are the integrands that, together with their Fenchel conjugates, are real-valued and strictly convex. Here it is clear that the lower bound in (2.1)\(_3\) quantifies the strict convexity of F, whereas the upper bound quantifies the strict convexity of \(F^*\) as we specify below.
Lemma 3.3
Let \(F\in {\text {C}}^{2}(\mathbb {R}^{N \times n})\) be an integrand satisfying the Legendre (p, q)-growth condition (2.1)\(_3\) for exponents \(2 \le p \le q < \infty \). Then \(F^{*}\in {\text {C}}^{1}(\mathbb {R}^{N \times n})\cap {\text {C}}^{2}(\mathbb {R}^{N \times n}\setminus \{F^{\prime }(0)\})\) and
for all \(z\in \mathbb {R}^{N \times n}\setminus \{0\}\), \(\xi \in \mathbb {R}^{N \times n}\). If \(\mu >0\) in (2.1)\(_3\), then \(F^*\) is \({\text {C}}^2\)-regular and (3.4) holds for all \(z\in \mathbb {R}^{N \times n}\), \(\xi \in \mathbb {R}^{N \times n}\).
Proof
The lower bound in (2.1)\(_3\) implies in a routine manner that F is strictly convex and super-linear, and it also guarantees that the Hessian matrix \(F^{\prime \prime }(z)\) is invertible for all \(z \in \mathbb {R}^{N \times n}\setminus \{0\}\). We can now infer from Lemma 3.2 that \(F^*\) is \({\text {C}}^2\)-regular away from \(F'(0)\) and that \(( F^{*})^{\prime \prime } (F^{\prime }(z )) = F^{\prime \prime }(z)^{-1}\) for all \(z \in \mathbb {R}^{N \times n}\setminus \{0\}\). In order to derive the double bound (3.4) we note that (2.1)\(_3\) yields that all the eigenvalues of \(F^{\prime \prime }(z)\) belong to the interval
and given that the eigenvalues of \(( F^{*} )^{\prime \prime }(F^{\prime }(z)) = F^{\prime \prime }(z)^{-1}\) are their reciprocals, they must all belong to the interval
Since \((2-q)/(q-1)=q^{\prime }-2\) this concludes the proof of (3.4). The final claim follows easily from the same considerations, and Lemma 3.2. \(\square \)
We conclude this part with a monotonicity property that follows from the Legendre (p, q)-growth condition (2.1)\(_3\). While this type of result is probably not surprising to the experts, we are not aware of any record of it in the literature. It is instrumental for the approach of this paper and we remark that it allows us to use the full power of convex duality without having to go through the dual variational problem [50].
Corollary 3.4
Let \(F\in {\text {C}}^{2}(\mathbb {R}^{N\times n})\) be an integrand that satisfies the Legendre (p, q)-growth condition (2.1)\(_{3}\). Then the monotonicity inequalities
hold for all \(z,z_{1},z_{2}\in \mathbb {R}^{N\times n}\), with \(c\equiv c(L,p,q)\). Moreover, given any \(\mathbb {R}^{N\times n}\)-valued, differentiable function w defined on an open set \(\Omega \subset \mathbb {R}^{n}\), it holds that
for all \(s\in \{1,\cdots ,n\}\), and some \(c\equiv c(L,p,q)\).
Proof
If \(\mu \in (0,1]\) in (2.1)\(_{3}\), inequality (3.5)\(_{1}\) comes as a direct consequence of Lemma 3.3 and (2.6)\(_{1,2}\), with the stated dependency of the constant c. In fact, letting \(\xi _{1}:=F^{\prime }(z_{1})\), \(\xi _{2}:=F^{\prime }(z_{2})\), we rewrite
for \(c\equiv c(L,p,q)\) which, together with the lower bound in (2.1)\(_{3}\), yields (3.5)\(_{1}\). On the other hand, if \(\mu =0\) we apply Lemma 3.3 and (2.6)\(_{1,2}\) to derive (3.5)\(_{1}\) for all \(z_{1},z_{2}\in \mathbb {R}^{N\times n}\) such that zero does not belong to the closed segment with endpoints \(z_{1},z_{2}\). Because the terms on the two sides of (3.5)\(_{1}\) are continuous in \(z_{1},z_{2}\) we obtain the general case by approximation. The bound in (3.5)\(_{2}\) comes as a straightforward consequence of (3.1) and Lemma 3.3. Indeed, observe that up to replacingFootnote 2F(z) by \(F(z)-F(0)-\langle F^{\prime }(0),z\rangle \) for all \(z\in \mathbb {R}^{N\times n}\), we can assume that \(F(0)=0\), and \(F^{\prime }(0)=0\), observe that this and the definition of \(F^{*}\) imply that \(F^{*}(0)=0\), set \(\xi :=F^{\prime }(z)\) and bound
with \(c\equiv c(L,q)\), and (3.5)\(_{2}\) follows including also the p-ellipticity information from (2.1)\(_{3}\). Finally, (3.6) is a direct consequence of (3.5\(_{1}\)) and standard difference quotients arguments. \(\square \)
As a direct consequence of Corollary 3.4, we deduce \(\textrm{L}^{q'}\)-integrability of the stress tensor for any function having \(\mathscr {F}\) locally finite.
Corollary 3.5
Let \(B\Subset \Omega \) be any ball with radius smaller than one, and \(v\in {\text {W}}^{1,p}(B,\mathbb {R}^{N})\) be a function such that \(\mathscr {F}(v;B)<\infty \). Then \(F'(\nabla v)\in \textrm{L}^{q'}(B;\mathbb {R}^{N})\) with
with \(c\equiv c(n,L,p,q)\).
Proof
A direct computation gives
where we also used Young inequality with conjugate exponents \((p,p')\), and it is \(c\equiv c(L,p,q)\). Integrating the content of the previous display on B we obtain (3.7), and the proof is complete. \(\square \)
3.1 Convex Polynomials
In this section we prove that for a reasonably large class of integrands, the usual natural growth conditions and the quantified Legendre one actually coincide. In this respect, let us first point out that a straightforward consequence of (2.3), and (2.1)\(_{3}\) yield
for all \(z\in \mathbb {R}^{N \times n}\), and some positive \(c\equiv c(L,q)\), that is the \((q-2)\)-growth of \(F''\) available in the controlled (p, q)-growth case. It is easy to see that the converse is false, that is, the condition on the right-hand side of (2.1)\(_{3}\) is strictly stronger than (3.8) when \(q>2\). However, if we restrict our attention to a special family of polynomials we can show that the two growth conditions actually coincide. To this aim, let us recall that, given a number \({\texttt {d}}\in \mathbb {N}\), for our purposes a polynomial of degree \({\texttt {d}}\) is a smooth real-valued function \(P:\mathbb {R}^{N\times n}\rightarrow \mathbb {R}\) that coincides with its Taylor expansion of order \({\texttt {d}}\). For convenience, we shall group by homogeneity the various terms appearing in the polynomial, thus ultimately expressing it as the sum of homogeneous components:
for all \(z\in \mathbb {R}^{N\times n}\), where we used the definition of polynomial to explicitly identify \(P_{s}\), that from now on, will be referred to as homogeneous components of the polynomial. We are mostly interested in convex, even polynomials, resulting as the sum of nonnegative homogeneous components. In particular, the evenness condition implies that the polynomial has even degree \(2{\texttt {d}}\), and in (3.9) only 2s-homogeneous terms appear. Before proving that convex, even polynomials of degree \(2{\texttt {d}}\) verify the quantified Legendre growth in (2.1)\(_{3}\) with \(q=2{\texttt {d}}\), let us recall the Euler relation for homogeneous functions: whenever \({\texttt {g}}\in {\text {C}}^{1}(\mathbb {R}^{N\times n})\) is positively \(\kappa \)-homogeneous for some \(\kappa > 1\), then
Next, a technical lemma.
Lemma 3.6
Let \(Q,H\in {\text {C}}^{1}(\mathbb {R}^{N\times n})\) be functions such that H is convex, even, and s-homogeneous for some \(s\ge 2\), and \(\langle Q^{\prime }(z), z\rangle \ge 0\) for all \(z\in \mathbb {R}^{N\times n}\). Then there exists \(\delta \equiv \delta (H,s)\in [0,1)\), depending only on the structure of H such that
Proof
Without loss of generality we can assume that H is not identically 0. Convexity and s-homogeneity assure that \(H(z)\ge 0\) for all \(z\in \mathbb {R}^{N\times n}\). Moreover, as \(H^{1/s}\) is convex, 1-homogeneous and nonnegative cf. [71, Corollary 15.3.2], it is the support function of some compact, convex, symmetric set \(K\subset \mathbb {R}^{N\times n}\), i.e., \(H(z)=\left( \sup _{\xi \in K}\xi \cdot z\right) ^s\) for all \(z\in \mathbb {R}^{N\times n}\). We then set \(\mathbb {V}:=\text {span}(K)\), decompose \(\mathbb {R}^{N\times n}=\mathbb {V}\oplus \mathbb {V}^{\perp }\), and accordingly split \(\mathbb {R}^{N\times n}\ni z=z_{1}+z_{2}\in \mathbb {V}\oplus \mathbb {V}^{\perp }\), so that \(H(z)\equiv H(z_{1})\), \(H^{\prime }(z)\equiv H^{\prime }(z_{1})\) for all \(z\in \mathbb {R}^{N\times n}\). Being K symmetric, zero belongs to the relative interior of K, and as H is not identically 0 there exists a positive radius \(\sigma _{H}>0\) such that \({\bar{B}}_{\sigma _{H}}(0)\cap \mathbb {V}\subset K\) and
Back to (3.11), if \(z\in \mathbb {V}^{\perp }\) there is nothing to prove as \(H^{\prime }(z)=0\), so for the rest of the proof we will take \(z\in \mathbb {V}\setminus \{0\}\). We then decompose \(H^{\prime }\) and \(Q^{\prime }\) in an orthonormal fashion along z - for \(G\in \{H,Q\}\) we have:
A direct computation then shows that
Now, if \(|E_{H}(z)|=0\) there is nothing to prove, otherwise by homogeneity we bound
where we used that by continuity (\(H\in {\text {C}}^{1}(\mathbb {R}^{N\times n})\)) and thanks to (3.12), the infimum of \(H/|H^{\prime }|\) taken over \(\partial B_{\sigma _{H}}(0)\cap \mathbb {V}\) is strictly positive, thus \(\delta \in [0,1)\). \(\square \)
Now we are ready to state the main result of this section.
Proposition 3.7
Let \({\texttt {d}}\in \mathbb {N}\), and P be an even, convex polynomial of degree \(2{\texttt {d}}\) with nonnegative homogeneous components. Then for each \(i\in \{0,\ldots ,2{\texttt {d}}-2\}\), there exists \(c\equiv c(n,N,P,{\texttt {d}})>0\) such that
Proof
The proof relies on an induction argument on the degree of the polynomial.
3.2 Base step: \({\texttt {d}}=1\)
Recalling that P is even, by definition we have \(P(z)=P(0)+ 2^{-1}\langle P^{\prime \prime }(0)z,z\rangle \), with \(P^{\prime \prime }(0)\) being positive semi-definite, and in particular not identically zero. Therefore we trivially have \(|P^{\prime \prime }(z)|=|P^{\prime \prime }(0)|\le |P^{\prime \prime }(0)|(1+|P^{\prime }(z)|^{0})\), that is (3.13) with \({\texttt {d}}=1\), and \(c_{0}:=|P^{\prime \prime }(0)|>0\).
3.3 Inductive Step
Assume now that (3.13) holds for all polynomials of degree 2j whenever \(1\le j\le {\texttt {d}}\), and let \(P:\mathbb {R}^{N\times n}\rightarrow \mathbb {R}\) be a convex, even polynomial of degree \(2{\texttt {d}}+2\), meaning in particular that \(P^{(2{\texttt {d}}+2)}(0)\not \equiv 0\), and the \((2{\texttt {d}}+2)\)-homogeneous part of P is given by
so that \(P=Q+H\), where \(Q:\mathbb {R}^{N\times n}\rightarrow \mathbb {R}\) is an even polynomial of degree \(2\tilde{{\texttt {d}}}\le 2{\texttt {d}}\). By homogeneity we have
pointwise in \(z\in \mathbb {R}^{N\times n}\), thus H is convex. But then \(z\mapsto H^{1/(2{\texttt {d}}+2)}(z)\) is convex, 1-homogeneous, and nonnegative, see [71, Corollary 15.3.2], so it must be the support function for a compact and convex subset \(K\subset \mathbb {R}^{N\times n}\), i.e.:
Since any nonempty, convex, closed set is uniquely determined by its support function, and \(H\not \equiv 0\), then K is symmetric, thus 0 belongs to the relative interior of K. Set \(\mathbb {V}:=\text {span}(K)\), and let \(\sigma _{P}>0\) be the largest number such that \({\bar{B}}_{\sigma _{P}}(0)\cap \mathbb {V}\subset K\) - \(\sigma _{P}\) is positive as 0 is in the relative interior of K. After splitting \(z=z_{1}+ z_{2}\in \mathbb {V}\otimes \mathbb {V}^{\perp }\), the formulation of H as the support function of K in (3.14) yields that \(H^{(i+2)}(z)\equiv H^{(i+2)}(z_{1})\) for all \(i\in \{0,\cdots ,2{\texttt {d}}\}\), so we bound
so for some \(c\equiv c(n,N,P,{\texttt {d}})\ge 1\) it is
where we used the \((2{\texttt {d}}+1)\)-homogeneity of \(H^{\prime }\). The very definition of H now yields, for \(i\in \{0,\ldots ,2{\texttt {d}}\}\),
with \(c\equiv c(n,N,P,{\texttt {d}})\). We then jump back to the decomposition \(P=Q+H\). Recalling that \(H\equiv H(z_{1})\), on \(\mathbb {V}^{\perp }\) it is \(P=Q\), so \(\left. Q\right| _{\mathbb {V}^{\perp }}\) is an even, convex polynomial of degree \(2\tilde{{\texttt {d}}}\le 2{\texttt {d}}\). In particular, by the induction hypothesis we have
Next, Taylor-expanding about \(z_{2}\), we obtain
where \(c\equiv c(n,N,P,{\texttt {d}})\) and we used Young inequality with conjugate exponents \(\left( \frac{2{\texttt {d}}-i}{2{\texttt {d}}-i-s},\frac{2{\texttt {d}}-i}{s}\right) \), and observed that for \(s\ge 0\) and \(i\ge -1\), it is
Via Taylor expansion, we further estimate
for \(c\equiv c(n,N,P,{\texttt {d}})\) - here we used Young inequality with conjugate exponents \(\left( \frac{2{\texttt {d}}+1}{2{\texttt {d}}+1-s},\frac{2{\texttt {d}}+1}{s}\right) \). The content of the previous display then gives
so choosing \(\varepsilon \equiv \varepsilon (n,N,P,{\texttt {d}})\in (0,1)\) sufficiently small we end up with
for \(c\equiv c(n,N,P,{\texttt {d}})\). Merging estimates (3.18) and (3.20) we eventually get
with \(c\equiv c(n,N,P,{\texttt {d}})\). Finally, recalling that Q is (at most) the sum of 2s-homogeneous terms of P with \(s\in \{0,\cdots ,{\texttt {d}}\}\), i.e.:
the nonnegativity of the \(P_{s}\)’s imply
Keeping in mind that H is convex, even and \((2{\texttt {d}}+2)\)-homogeneous, by Lemma 3.6 we find \(\delta \equiv \delta (n,N,P,{\texttt {d}})\in [0,1)\) such that
Combining the last four displays we end up with (3.13)\(_{{\texttt {d}}+1}\) and the proof is complete. \(\square \)
As a direct consequence of Lemma 3.6 and Proposition 3.7 we can show that whenever a convex, even, positively s-homogeneous function is added to an integrand satisfying Legendre (p, q)-growth conditions, the sum verifies Legendre \((p,\max (s,q))\)-growth conditions.
Corollary 3.8
Let \(Q,H\in {\text {C}}^{2}(\mathbb {R}^{N\times n})\) be functions such that Q satisfies (2.1)\(_{3}\), for some exponents \(2\le p\le q\), and H is convex, even, and s-homogeneous for some \(s\ge 2\). Then the sum \(Q+H\) satisfies \((p,\max \{q,s\})\)-Legendre growth conditions.
Proof
The lower bound in (2.1)\(_{3}\) is preserved by the sum \(Q+H\) as Q satisfies (2.1)\(_{3}\), and H is convex. Concerning the upper bound, if \(z=0\), there is nothing to prove, otherwise, being H convex, s-homogeneous (\(s\ge 2\)), and even, we can repeat verbatim the construction detailed in the "Inductive step" of the proof of Proposition 3.7 up to (3.15) to decompose \(\mathbb {R}^{N\times n}=\mathbb {V}\oplus \mathbb {V}^{\perp }\), so that on \(\mathbb {V}\) the double bound in (3.15) holds true for the map H. Splitting \(\mathbb {R}^{N\times n}\ni z=z_{1}+z_{2}\in \mathbb {V}\oplus \mathbb {V}^{\perp }\), we control
with \(c\equiv c(n,N,L,p,q,s,H)\). \(\square \)
Next, we focus on those convex, even polynomials that are strongly p-elliptic in the sense of (2.1)\(_{3}\) (left-hand side), and show that in the degenerate case \(p>2\), \(\mu =0\), the bound in (2.2) holds true.
Corollary 3.9
Let \({\texttt {d}}_{0},{\texttt {d}}\in \mathbb {N}\), be such that \(2< 2{\texttt {d}}_{0}<2{\texttt {d}}\), and let P be an even, convex polynomial of degree \(2{\texttt {d}}\) with nonnegative homogeneous components \(P_{s}\) such that \(P_{s}\equiv 0\) for all \(s\in \{1,\cdots ,2({\texttt {d}}_{0}-1)\}\). Then (2.2) is verified with \(q=2{\texttt {d}}\), for all \(2< p\le 2{\texttt {d}}_{0}\), that is,
with \(c\equiv c(n,N,p,{\texttt {d}},P)\).
Proof
By definitionFootnote 3 of P,
with \(c\equiv c(P,{\texttt {d}})\). Now, keeping in mind that \(2< p\le 2{\texttt {d}}_{0}\), if \(|z|\in (0,1)\) we bound
for \(c\equiv c(P,p,{\texttt {d}})\). We next observe that, whenever \(|z|\ge 1\), the \((2s-1)\)-homogeneity of \(P_{2s}^\prime \) gives
for \(c\equiv c(P,{\texttt {d}})\), so we estimate
with \(c\equiv c(P,p,{\texttt {d}})\), and the proof is complete. \(\square \)
Remark 3.10
Thanks to Corollary 3.8, we immediately see that the class of integrands featuring Legendre (p, q)-growth conditions embraces the anisotropic polynomial examples introduced by Marcellini in [61]. In fact, repeated applications of Corollary 3.8 to the integrand in (1.6)–(1.8) yield that P satisfies \((p,q_{n})\)-Legendre growth conditions, and the prototypical examples of integrands with (p, q)-growth are covered.
3.4 An Abstraction of the Quantified Legendre Condition
Let F, \(G :\mathbb {R}^{N \times n}\rightarrow \mathbb {R}\) be two strictly convex, \({\text {C}}^1\) integrands that are super-linear, i.e.:
Consequently also the Fenchel conjugates \(F^*\) and \(G^*\) are strictly convex, \({\text {C}}^1\) and super-linear, that is, F and G real-valued, are super-linear Legendre integrands. The goal of this subsection is the following result and its corollary, that can be seen as an abstract counterpart of (2.1)\(_{3}\), covering more general convexity conditions than the usual power type one.
Proposition 3.11
The difference of the Fenchel conjugates \(F^{*}-G^{*}\), is convex if and only if
holds for all z, \(z_{0} \in \mathbb {R}^{N \times n}\).
The condition can be recast as a second order condition if we interpret it in the viscosity sense:
Corollary 3.12
The difference of the Fenchel conjugates \(F^{*}-G^{*}\), is convex if and only if
holds as quadratic forms in the viscosity sense for all \(z_{0} \in \mathbb {R}^{N \times n}\).
Proof of Proposition 3.11
Note that the derivatives \(F^\prime \) and \(G^\prime \) are homeomorphisms of \(\mathbb {R}^{N \times n}\) with inverses \((F^{*})^\prime \) and \((G^{*})^\prime \), respectively. Fix \(z_0 \in \mathbb {R}^{N \times n}\) and put \(\xi _{0} = F^{\prime }(z_{0})\). Now \(F^{*}-G^{*}\) is convex at \(\xi _0\) if and only if
holds for all \(\xi \in \mathbb {R}^{N \times n}\), that is,
for all \(\xi \in \mathbb {R}^{N \times n}\). Because \((F^{*})^{\prime }( \xi _{0})=z_0\) we get by Fenchel conjugation that this inequality is equivalent to
for all \(z \in \mathbb {R}^{N \times n}\). Recalling that \(F^{**} = F\), \(G^{**} = G\) we can rewrite this as
for all \(z \in \mathbb {R}^{N \times n}\). Because \(F(z_{0})+F^{*}(\xi _{0})=\langle z_{0}, \xi _{0}\rangle \) and \(G^{*}(\xi _{0})+G( (G^{*})^{\prime }( \xi _{0})) = \langle \xi _{0}, (G^{*})^{\prime }( \xi _{0})\rangle \) the last inequality can be rewritten as \(F(z+z_{0})-F(z_{0}) \le G(z+(G^{*})^{\prime }( \xi _{0})) - G( (G^{*})^{\prime }( \xi _{0}))\) for all \(z \in \mathbb {R}^{N \times n}\). Consequently, using that \(\xi _{0}=F^{\prime }(z_{0})\) and \(G^{\prime }( (G^{*})^{\prime }(F^{\prime }(z_{0})))=\xi _{0}\) we arrive at the inequality
for all \(z \in \mathbb {R}^{N \times n}\), as required.
4 Higher Differentiability Under Legendre (p, q)-Growth
In this section we derive some regularity results for local minimizers of variational integrals under Legendre (p, q)-growth. Our focus is on the higher differentiability of minima, that will subsequently be used to prove finer regularity in low dimension and in the scalar setting.
4.1 Approximation Scheme
Our approximation scheme is rather basic and aims at correcting two relevant structural issues of the integrand F: unbalanced growth and possible degeneracy. Let \(B\equiv B_{{\texttt {r}}}(x_{B})\Subset \Omega \) be a ball with radius \(0<\texttt {r}\le 1\). We regularize \(u\in {\text {W}}^{1,p}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N})\) via convolution against a sequence of mollifiers \(\{\phi _{\varepsilon }\}_{\varepsilon >0}\subset {\text {C}}^{\infty }_{c}(\mathbb {R}^{n})\) thus determining a sequence \(\{\tilde{u}_{\varepsilon }\}_{\varepsilon>0}:=\{u*\phi _{\varepsilon }\}_{\varepsilon >0}\subset {\text {C}}^{\infty }_{{\text {loc}}}(\Omega ,\mathbb {R}^N)\), set
define integrand
and introduce the family of approximating functionals
By (2.1)\(_{3}\) and basic direct methods, the Dirichlet problem
admits a unique solution \(u_{\varepsilon }\in \left( \tilde{u}_{\varepsilon }+{\text {W}}^{1,q}_{0}(B,\mathbb {R}^N)\right) \), verifying by minimality the integral identity
The convergence features of the sequence of minima obtained solving the approximating problems in (4.3) are well known, see e.g. [15, Section 3]:
Let us quickly show that \(F_{\varepsilon }\in {\text {C}}^{2}(\mathbb {R}^{N\times n})\) enjoys both growth/ellipticity features of the q-Laplacian type, and uniform Legendre (p, q)-growth.
Lemma 4.1
With (2.1) in force, let \(F_{\varepsilon }\) be the integrand defined in (4.2). Then
for all \(z,\xi \in \mathbb {R}^{N\times n}\) and some \(\Lambda \equiv \Lambda (n,N,L,p,q)\).
Proof
The first three bounds in (4.6) are a straightforward consequence of (2.1), (2.3) and the very definition of \(F_{\varepsilon }\), so we focus on (4.6)\(_{4}\). We first prove that there exists a constant \(\tilde{c}\equiv \tilde{c}(L,p,q)\) such that if \(|z|\ge \tilde{c}\), then \(\langle F'(z),z\rangle >0\). In fact, by Young inequality with conjugate expoenents \((p,p')\) we have
with \(c',c''\equiv c',c''(L,p,q)\), thus letting \(\tilde{c}:=\max \{(2c''/c')^{1/p},1\}\equiv \tilde{c}(L,p,q)\ge 1\) we get
Next, by (4.6\(_{3}\)) we have
while if \(|z|\ge \tilde{c}\) we compute
therefore
for \(c\equiv c(n,N,L,p,q)\). Combining (4.8) and (4.10) we obtain (4.6)\(_{4}\) and the proof is complete. \(\square \)
Thanks to (4.6)\(_{1,2,3}\), by-now classical regularity theory [41, Chapter 8] yields that
in particular, by (4.11) via difference quotients arguments, (4.4) can be differentiated: system
holds for any \(w\in {\text {W}}^{1,2}(B,\mathbb {R}^N)\) such that \(\,\textrm{supp }(w)\Subset B\), and all \(s\in \{1,\cdots ,n\}\). Finally, we record for later use the elementary energy estimates
4.2 Proof of Theorem 1.2
We start by recording the uniform \(L^{q'}\)-bound on the stress tensors \(\{F'(\nabla u_{\varepsilon })\}_{\varepsilon >0}\):
for \(c\equiv c(n,L,p,q)\). We fix a ball \(B_{r}(x_{0})\subseteq B\), let \(\varrho \in (0,3r/4)\), introduce the following symbols:
and notice that by (4.11) all the above quantities are finite. Before proceeding further, a brief remark about notation: since all balls considered from now on will be centered at \(x_{0}\), we shall omit denoting it and simply write \(B_{\varrho }(x_{0})\equiv B_{\varrho }\). With \(0<\delta <\min \{1,\,\textrm{dist}(B_{\varrho },\partial B)/100\}\), we regularize the characteristic function of \(B_{\varrho }\) by convolution against a sequence \(\{\phi _{\delta }\}_{\delta >0}\subset {\text {C}}^{\infty }_{c}(\mathbb {R}^{n})\) of standard mollifiers , thus defining \(\{\eta _{\delta }\}_{\delta>0}:=\{\mathbbm {1}_{B_{\varrho }}*\phi _{\delta }\}_{\delta >0}\subset {\text {C}}^{\infty }_{c}(B_{\varrho +2\delta })\), test (4.12) against \(w_{\delta }:=\eta _{\delta }(\partial _{s}u_{\varepsilon }-(\partial _{s}u_{\varepsilon })_{\partial B_{\varrho }})\), which is admissible by (4.11), and sum on \(s\in \{1,\cdots ,n\}\). We obtain:
Recalling that \(\eta _{\delta }\rightharpoonup ^{*} \mathbbm {1}_{B_{\varrho }}\) in \(\textrm{BV}\), we let \(\delta \rightarrow 0\) in the above display to derive via Cauchy Schwarz inequality, and Hölder inequality with conjugate exponents \(\left( q/(q-p),q/p\right) \),
with \(c\equiv c(n,N,L,L_{\mu },p,q)\). We then notice that, up to choosing \(2^{*}_{n-1}\) so large that \(2^{*}_{n-1}>2q/p\) when \(n=2\) or \(n=3\), by (1.12) it is \(2q/p<2^{*}_{n-1}\), so we can write
as to control via Hölder inequality with conjugate exponents \(\left( 1/\lambda ,1/(1-\lambda )\right) \),
for \(c\equiv c(n,N,L,p,q).\) Merging (4.15) and (4.17) we get
where we set
and it is \(c\equiv c(n,N,L,L_{\mu },p,q)\). Next, we fix parameters \(r/2\le \tau _{2}<\tau _{1}\le 3r/4<\,\textrm{dist}(x_{0},\partial B)\), and, for \(\texttt {K}_{\varepsilon }\in \{\texttt {H}_{\varepsilon },\texttt {P}_{\varepsilon },\texttt {Q}_{\varepsilon },\texttt {S}_{\varepsilon }\}\), set
From the definition we get, since \(\texttt {K}_{\varepsilon }(x_{0}; \cdot )\) is absolutely continuous and increasing, that \(\mathcal {L}^{1}\left( \!(\tau _{2},\tau _{1})\setminus \mathcal {I}_{\texttt {K}_{\varepsilon }}\!\right) {<}(\tau _{1}-\tau _{2})/4\), and consequently \(\mathcal {L}^{1}\left( \mathcal {I}_{\texttt {H}_{\varepsilon }}\cap \mathcal {I}_{\texttt {P}_{\varepsilon }}\cap \mathcal {I}_{\texttt {Q}_{\varepsilon }}\cap \mathcal {I}_{\texttt {S}_{\varepsilon }}\!\right) >0\). Therefore we can pick \(\varrho \in \mathcal {I}_{\texttt {H}_{\varepsilon }}\cap \mathcal {I}_{\texttt {P}_{\varepsilon }}\cap \mathcal {I}_{ \texttt {Q}_{\varepsilon }}\cap \mathcal {I}_{\texttt {S}_{\varepsilon }}\) for which also (4.18) holds. Hereby
with \(c\equiv c(n,N,L_{\mu },p,q)\). Now we observe that
so we can apply Young inequality with conjugate exponents \(\left( \frac{2q}{2q-\lambda p},\frac{2q}{\lambda p}\right) \) and (2, 2) to have
where we set
and it is \(c\equiv c(n,N,L_{\mu },p,q)\). Lemma 2.6 then yields
for \(c\equiv c(n,N,L,L_{\mu },p,q)\). Finally, we use (4.5\(_{2,3}\)) and (2.1\(_{1}\)) to send \(\varepsilon \rightarrow 0\) in (4.21) , and fix \(B_{r}(x_{0})\equiv B\) to conclude with (1.14). A standard covering argument then yields that \(V_{\mu ,p}(\nabla u), V_{1,q'}(F'(\nabla u))\in {\text {W}}^{1,2}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N\times n})\), and (1.13)\(_{1}\) is proven. Now we only need to prove the validity of (1.13\(_{2}\)). To this end, we assume \(\mu >0\) in (2.1)\(_{3}\), observe that thanks to (4.21) and (2.5), sequence \(\{u_{\varepsilon }\}_{\varepsilon >0}\) of solutions to problem (4.3) is now uniformly bounded also in \(W^{2,2}(B_{r/2},\mathbb {R}^{N})\), and observe that by (4.6\(_{3}\)) and (4.11) it is
where we set \(\bigodot (\mathbb {R}^{N\times n})\times \mathbb {R}^{N\times n}\ni (A,z)\mapsto \mathcal {D}(A,z):=\langle Az,z\rangle \), \(\mathbb {R}^{N\times n}\ni z\mapsto G''(z):=F''(z)\ell _{\mu }(z)^{2-p}\), and \(V_{\varepsilon }^{s}:=\ell _{\mu }(\nabla u_{\varepsilon })^{\frac{p-2}{2}}\partial _{s}\nabla u_{\varepsilon }\). Function \(\mathcal {D}\) is continuous by definition, and, if the first argument is restricted to positive semi-definite forms, it is also nonnegative. Notice that by (2.1)\(_{1}\) and (4.5)\(_{3}\), it is \(G''(\nabla u_{\varepsilon })\rightarrow F''(\nabla u)\ell _{\mu }(\nabla u)^{2-p}\) in \(\mathcal {L}^{n}\)-measure. Moreover, observing that
being \(p\ge 2\), for each \(s\in \{1,\cdots ,n\}\) by (4.21) we have \(V_{\varepsilon }^{s}\rightharpoonup \ell _{\mu }(\nabla u)^{\frac{p-2}{2}}\partial _{s}\nabla u\) weakly in \(\textrm{L}^{2}(B_{r/2},\mathbb {R}^{N\times n})\). By [70, Theorem 4.1 and Remark 4.2], up to nonrelabelled subsequences, each of the sequences \(\{G''(\nabla u_{\varepsilon }),V_{\varepsilon }^{s}\}_{\varepsilon >0}\) generates a Young measure and, via [70, Lemma 5.19], it is \(\{G''(\nabla u_{\varepsilon }),V_{\varepsilon }^{s}\}_{\varepsilon >0}{\mathop {\rightarrow }\limits ^{\textbf{Y}}}\{\nu _{x}^{s}\}\) where \(\nu _{x}^{s}=\delta _{G''(\nabla u)}\otimes \kappa ^{s}_{x}\), and \(\{\kappa ^{s}_{x}\}\) is a Young measure generated by \(\{V_{\varepsilon }^{s}\}_{\varepsilon >0}\). Therefore
By weak lower semicontinuity for Young measures [70, Proposition 4.6], we have
so we can apply Jensen inequality [70, Lemma 5.11] to derive
A standard covering argument then gives (1.13\(_{2}\)) and completes the proof.
Remark 4.2
Theorem 1.2 is valid a fortiori if the integrand governing \(\mathscr {F}\) is of the form \(\mathbb {R}^{N\times n}\ni z\mapsto \tilde{F}(z):=F(z)+a_{0}(1+|z|^{2})^{q/2}\) with F as in (2.1) and some constant \(a_{0}\ge 0\). In this case estimate (1.14) includes also the informations coming from the q-elliptic term, i.e.:
with \(c\equiv c(n,N,L,L_{\mu },p,q)\).
4.3 Proof of Corollary 1.3
The proof of statements (ii.)–(iii.) comes as a direct consequence of Sobolev–Morrey embedding theorem and Theorem 1.2. Concerning the result in (i.), Theorem 1.2 and Sobolev embedding theorem yields that \(\nabla u\in \textrm{L}^{\frac{np}{n-2}}_{{\text {loc}}}(\Omega ,\mathbb {R}^{N\times n})\) and, if \(p>n-2\), it is \(np/(n-2)>n\), thus Sobolev-Morrey embedding theorem allows to complete the proof.
4.3.1 Proof of Theorem 1.5, (i.)
Proposition 3.7 and Corollary 3.9 yield that Corollary 1.3 applies to the class of convex polynomials in the statement of Theorem 1.5, and in three space dimensions \(n=3\) the local Hölder continuity result in (i.) immediately follows with no upper restriction on the polynomial degree. Part (ii.) of Theorem 1.5 will be proven in Section 5.3.1 below.
5 Smoothness in Two Space Dimensions
In this section we offer two independent proofs of gradient boundedness in 2d for vector-valued minimizers of functional \(\mathscr {F}\), based on an inhomogeneous monotonicity formula, and on a renormalized Gehring–Giaquinta and Modica type lemma. From this, we deduce full regularity in 2d. As already mentioned in Section 1.1, our approach allows handling simultaneously degenerate and nondegenerate problems and entails new results already in the genuine (p, q)-setting, see Remark 5.3 below. For the sake of clarity, we shall divide the remainder of the section in three parts, beginning by detailing the two arguments leading to gradient boundedness and concluding by proving the full regularity statement, leading to the proof of Theorem 1.4.
5.1 Proof of Theorem 1.4: 2d-Gradient Boundedness via Monotonicity Formula
We look back at Sect. 4.2, and exploit the low dimension—if \(n=2\) we are allowed to choose \(2^{*}_{1}\) arbitrarily large—to study the asymptotics of the exponents in (4.20). We have
so we can increase \(2^{*}_{1}\) in such a way that \(\alpha _{0}-\kappa _{1}\ge -4q/p\) and \(\kappa _{2}+1\le 2q/p\), and be more precise on estimate (4.21), that now becomes
for \(c\equiv c(N,L,L_{\mu },p,q)\). By (4.14), (5.1), and Trudinger inequality [80, Theorem 2], there exists a parameter , with constants implicit in "\(\approx \)" depending on \((N,L,L_{\mu },p,q)\) such that
again up to constants depending on \((N,L,L_{\mu },p,q)\)—in fact, an inspection of the proof of [80, Theorem 2] shows that thanks to (5.1) and (4.14) the constant \({\texttt {d}}\), that is proportional to the quantity can be suitably enlarged to replace any dependency on \(\varepsilon \) with a dependency on \((N,L,L_{\mu },p,q)\). In particular, whenever \(B_{\varrho }\subseteq B/2\) we have by Jensen’s inequality,
for \(c\equiv c(N,L,L_{\mu },p,q)\). Next, taking \(x_{0}\in B/4\), and \(\varrho \in (0,{\texttt {r}}/6)\), we fix a ball \(B_{\varrho }(x_{0})\Subset B/2\) and refine estimate (4.15) as
where we also used Hölder inequality with conjugate exponents \(\left( q/(q-p),q/p\right) \), and it is \(c\equiv c(N,L,L_{\mu },p,q)\). By the trace theorem [43, Section 2.3], we control
for \(c\equiv c(N,L,L_{\mu },p,q)\), therefore we can update (5.4) to
with \(c\equiv c(N,L,L_{\mu },p,q)\). We then let , pick any \(\sigma \in (0,{\texttt {r}}/6^{4})\), and integrate the inequality in (5.5) on \(\varrho \in (\sigma ,{\texttt {r}}/6)\) to derive
Next, we set \(\beta :=\left( q{\texttt {B}}^{-1}/(q+p)\right) ^{2q/(q+p)}\), \(\gamma :=(q+p)/(2q)\), and pass to the exponentials in the previous display to recover
for \(c\equiv c(N,L,L_{\mu },p,q)\). We then let \(\varepsilon \rightarrow 0\) above and use (4.5)\(_{3}\) to secure
where we used that \((\beta \log ({\texttt {r}}/\sigma ))^{\gamma +2}\exp \left\{ -(\beta \log ({\texttt {r}}/\sigma ))^{\gamma }\right\} \rightarrow 0\) as \(\sigma \rightarrow 0\), and it is \({\texttt {b}}:=(\gamma +2)\left( \frac{q-p}{p}\right) +\frac{2q}{p}\), \(c\equiv c(N,L,L_{\mu },p,q)\). Since (5.6) holds true for all balls \(B_{\sigma }(x_{0})\Subset B/2\), \(x_{0}\in B/4\), from a variant of Morrey lemma [35, Lemma 1.1], see also [36, page 287], we deduce that \(V_{\mu ,p}(\nabla u)\), \(V_{1,q'}(F'(\nabla u))\) are continuous on B/8 and therefore bounded, and a standard covering argument eventually yields that \(F'(\nabla u),\ \nabla u\in \textrm{L}^{\infty }_{{\text {loc}}}(\Omega ,\mathbb {R}^{N\times 2})\).
5.2 Proof of Theorem 1.4: 2d-Gradient Boundedness via Renormalized Gehring–Giaquinta and Modica Lemma
Our first move is a quantitative version of classical Gehring lemma [37], after Giaquinta and Modica [39].
Lemma 5.1
Assume that \(\varphi \) is a nonnegative, decreasing function on \([t_{0},\infty )\), infinitesimal as \(t\rightarrow \infty \), and satisfying
for some absolute constants \(c_{0}\ge 1\), \({\texttt {M}}\ge 1\), \({\texttt {m}}\in (0,1)\) and all \(t\ge t_{0}\). There exists a positive number
such that
Proof
We assume for the moment that there is \(\kappa \ge 1\) such that \(\varphi (t)=0\) whenever \(t\ge \kappa \). We will remove this restriction in the end. With \({\texttt {d}}>0\) we set
For some \({\texttt {t}}>1>{\texttt {m}}\) to be determined, integrating by parts we have
We now fix \({\texttt {t}}>1\) so close to one that
and (5.8) is satisfied. The above choice of \({\texttt {t}}\) allows us to conclude that
and we are done in the case \(\varphi (t)\equiv 0\) for \(t\ge \kappa \). Let us take care of the general case. Observe that being \(\varphi \) nonincreasing on \([t_{0},\infty )\), we have for any given \(T\ge \kappa \):
so letting \(T\rightarrow \infty \) above we get
Next, we set \(\varphi _{\kappa }(t):=\varphi (t)\) if \(t\le \kappa \), and \(\varphi _{\kappa }(t):=0\) if \(t>\kappa \). For \(t\le \kappa \) we have
while if \(t>\kappa \) the above relation is trivially true, given that \(\varphi _{\kappa }(t)\equiv 0\) if \(t>\kappa \). Thanks to what we have just proved, we have
The conclusion now follows sending \(\kappa \rightarrow \infty \). \(\square \)
The previous lemma is instrumental to establish a self-improving property for functions satisfying suitable reverse Hölder inequalities.
Lemma 5.2
Let \(Q_{1}(0)\subset \mathbb {R}^{n}\) be the unitary cube centered at the origin and \(v\in \textrm{L}^{1}(Q_{1}(0))\) be a nonnegative function verifying
for all cubes \(Q_{\varrho }(x_{0})\Subset Q_{1}(0)\), some exponent \({\texttt {m}}\in (0,1)\), and absolute constants \({\hat{c}},{\texttt {M}}\ge 1\). There exists a constant \(c_{*}\equiv c_{*}(n,{\hat{c}})\ge 1\), and a positive number \({\texttt {t}}\in (1,2)\) as in (5.8), i.e.:
such that
In particular, for any fixed threshold \({\texttt {s}}_{0}\ge 1\), the constant \(c_{*}\) in (5.15) can be taken so large that
up to update dependency \(c_{*}\equiv c_{*}(n,{\hat{c}},{\texttt {s}}_{0})\).
Proof
We ask the reader to have [41, Section 6.4] at hand as we shall carefully track the dependency of the constants appearing in the various arguments leading to the proof of [41, Theorem 6.6]. To be as close as possible to the setting of [41, Section 6.4], we let \({\texttt {d}}(x):=\,\textrm{dist}(x,\partial Q_{1})\), introduce function \({\texttt {F}}(x):={\texttt {d}}(x)^{n}v(x)\), and denoting by \(\tilde{P}\Subset Q_{1}\) a cube and P the cube concentric to \(\tilde{P}\) with half side, and rearrange (5.14) in terms of \({\texttt {F}}\) as
holding for all cubes \(\tilde{P}\Subset Q_{1}\), with \(c\equiv c(n,{\hat{c}})\), that is [41, inequality (6.47)]. We next look into [41, Lemma 6.2], that relies on a Calderón-Zygmund type argument combined with integral estimates on level sets. Introducing the superlevel set \(\Phi _{t}:=\left\{ x\in Q_{1}:{\texttt {F}}(x)>t\right\} \) for all
and carefully tracking the occurrences of constant \({\texttt {M}}\) along the proof of [41, Lemma 6.2] (fix \(\lambda \approx {\texttt {M}}\) up to dimensional constants there) we end up with
for \(c_{*}\equiv c_{*}(n,{\hat{c}})\). Notice that there is no loss of generality in arbitrarily enlarging the value of \(c_{*}\), and in particular to take it larger than the assigned \({\texttt {s}}_{0}\), thus fixing dependencies \(c_{*}\equiv c_{*}(n,{\hat{c}},{\texttt {s}}_{0})\). In the light of [41, Lemma 6.3], inequality (5.19) can be rewritten as
where we set \(\varphi (t):=\int _{\Phi _{t}}{\texttt {F}}^{{\texttt {m}}}\,\textrm{d}x\). Notice that by definition, \(\varphi \) is nonincreasing on \([t_{0},\infty )\) and, since \({\texttt {F}}\in \textrm{L}^{1}(Q_{1})\), we see that \(\varphi (t)\rightarrow 0\) as \(t\rightarrow \infty \), so the assumptions of Lemma 5.1 are verified—in particular, (5.20) is exactly (5.7) with \(c_{0}\equiv c_{*}\), and \(t_{0}\) as defined in (5.18), thus (5.9) holds true with exponent \({\texttt {t}}\) resulting from (5.8) that is the one in (5.15), and, via [41, Lemma 6.3] we can write
On the other hand, on \(Q_{1}\setminus \Phi _{t_{0}}\) we have
therefore we can conclude with
and, coming back to the function v we eventually get
where we used (5.18) and (5.8)\(_{1}\) to remove any dependency of the bounding constants on \({\texttt {t}}\). The proof is complete. \(\square \)
We return to Dirichlet problem (4.3), and observe that conditions (4.6)\(_{1,2,3}\) guarantee that [14, Theorem V] applies and
We then let \(Q\equiv Q_{r_{0}}(x_{Q})\Subset B/16\) be any cube with half-side length \(r_{0}\in (0,1]\), scale the whole problem on \(Q_{1}\) by letting \(v_{\varepsilon }(x):=u_{\varepsilon }(x_{Q}+r_{0}x)/r_{0}\), \(v(x):=u(x_{Q}+r_{0}x)/r_{0}\), observe that the amount of regularity in (5.21) is obviously preserved by the blown up map \(v_{\varepsilon }\) so that the definition
makes sense, and recall that system (4.12) holds verbatim for the scaled maps \(v_{\varepsilon }\) on \(Q_{1}\). We then let \(Q_{\varrho }(x_{0})\Subset Q_{1}\) be any cube with half side length \(\varrho \in (0,1]\), \(\eta \in {\text {C}}^{1}_{c}(Q_{\varrho }(x_{0}))\) be such that \(\mathbbm {1}_{Q_{\varrho /2}(x_{0})}\le \eta \le \mathbbm {1}_{Q_{\varrho }(x_{0})}\) and \(|\nabla \eta |\lesssim \varrho ^{-1}\), and test (4.12) against \(w_{\varepsilon }:=\eta ^{2}(\partial _{s}v_{\varepsilon }-(\partial _{s} v_{\varepsilon })_{Q_{\varrho }(x_{0})})\) to get, after using Cauchy–Schwarz inequality, (3.6), (2.6)\(_{1,3}\), (2.4), and Sobolev-Poincaré inequality,
for \(c\equiv c(N,L,L_{\mu },p,q)\). From (5.21)-(5.23) we see that the assumptions of Lemma 5.2 are satisfied with
therefore we obtain an exponent \({\texttt {t}}\in (1,2)\) as in (5.15) such that , which in particular implies
Next, we let \(\eta _{0}\in {\text {C}}^{1}_{c}(Q_{1/2})\) such that \(\mathbbm {1}_{Q_{1/4}}\le \eta _{0}\le \mathbbm {1}_{Q_{1/2}}\) and \(|\nabla \eta _{0}|\le 2\), observe that we have \(\eta _{0}V_{1,q^{\prime }}(F^{\prime }(\nabla v_{\varepsilon }))\in {\text {W}}^{1,2{\texttt {t}}}(\mathbb {R}^{2},\mathbb {R}^{N\times 2})\) and via (5.25), bound
We then apply Proposition 2.5, and use (5.26) and (5.15) to deduce that
Before proceeding further, let us make explicit the values of the constants/exponents involving \({\texttt {t}}\) above by means of (5.15), (5.16) and (5.24). We have
where we used that \(1/(2{\texttt {t}})^{\prime }\), \(({\texttt {t}}-1)/ {\texttt {t}}\) are decreasing with respect to both \({\texttt {M}}_{\varepsilon }\) and \(c_{*}\). Incorporating this information in (5.27), scaling back on Q, and letting \({\texttt {U}}_{\varepsilon }:=|\nabla V_{\mu ,p}(\nabla u_{\varepsilon })|^{2}+|\nabla V_{1,q^{\prime }}(F^{\prime }(\nabla u_{\varepsilon }))|^{2}+ \gamma _{\varepsilon }|\nabla V_{1,q}(\nabla u_{\varepsilon })|^{2}\), by (2.5) we obtain after standard manipulations
with \(c\equiv c(N,L,L_{\mu },p,q)\). Next, we fix parameters \(2^{-3}{} \texttt {r}\le \tau _{2}<\tau _{1}\le 2^{-5/2}{} \texttt {r}\), correspondingly, cubes concentric with B, i.e.: \(B/8\subset Q_{2^{-3}{} \texttt {r}}(x_{B})\subset Q_{\tau _{2}}(x_{B})\subset Q_{\tau _{1}}(x_{B})\subset Q_{2^{-5/2}{} \texttt {r}}(x_{B})\Subset B/4\), and observe that for any \(x_{Q}\in Q_{\tau _{2}}(x_{B})\), cube \(Q_{(\tau _{1}-\tau _{2})/8}(x_{Q})\Subset Q_{\tau _{1}}(x_{B})\), so, keeping in mind that there is no loss of generality in assuming that \(\Vert F^{\prime }(\nabla u_{\varepsilon }) \Vert _{\textrm{L}^{\infty }(Q_{2^{-3}{} \texttt {r}}(x_{B}))}\ge 1\) (otherwise the proof would be finished already), we apply (5.28) with \(Q\equiv Q_{(\tau _{1}-\tau _{2})/8}(x_{Q})\) and \({\texttt {M}}_{\varepsilon }\equiv \Vert F'(\nabla u_{\varepsilon }) \Vert _{\textrm{L}^{\infty }(Q_{\tau _{1}}(x_{B}))}^{(q-p)/(q-1)}\), which verifies (5.22), to get
for \(c\equiv c(N,L,L_{\mu },p,q)\). Taking the supremum for all \(x_{Q}\in Q_{\tau _{2}}(x_{B})\), and setting \(\theta _{1}:=(4q-3p)/(4q-p),\ \theta _{2}:=4(q-p)/(4q-p)\in (0,1)\), by Young inequality we obtain
with \(c\equiv c(N,L,L_{\mu },p,q)\). We can then apply Lemma 2.6, (4.14), (4.5)\(_{1,2}\), and (4.21) (keep Remark 4.2 in mind), to conclude with
with \(c\equiv c(N,L,L_{\mu },p,q)\), \({\texttt {b}}\equiv {\texttt {b}}(p,q)\). Now, by (4.5\(_{3}\)) we know that \(\nabla u_{\varepsilon }\rightarrow \nabla u\) strongly in \(\textrm{L}^{p}(B,\mathbb {R}^{N})\), so by (2.1)\(_{1}\) (up to subsequences) \(F^{\prime }(\nabla u_{\varepsilon })\rightarrow F^{\prime }(\nabla u)\) almost everywhere. This means that we can send \(\varepsilon \rightarrow 0\) in (5.30) and pass to the weak\(^{*}\)-limit to deduce
for \(c\equiv c(N,L,L_{\mu },p,q)\), \({\texttt {b}}\equiv {\texttt {b}}(p,q)\), which, together with (2.3) gives (1.16) (with possibly different exponent \({\texttt {b}}\equiv {\texttt {b}}(p,q)\)), and a standard covering argument leads to \(\nabla u\in \textrm{L}^{\infty }_{{\text {loc}}}(\Omega ,\mathbb {R}^{N\times 2})\), and, whenever \(\Omega _{2}\Subset \Omega _{1}\Subset \Omega \) are open sets as in the statement of Theorem 1.4, it is
5.3 Proof of Theorem 1.4: Full Regularity
Once it is known, by either of the arguments in the previous two sections, that minima of \(\mathscr {F}\) are locally Lipschitz continuous, the nonuniform ellipticity of the integrand F prescribed by (2.1)\(_{3}\) becomes immaterial. The gradient boundedness, Theorem 1.2, and a simple difference quotients argument guarantee that we can recover (4.12) with F, u, instead of \(F_{\varepsilon }\), \(u_{\varepsilon }\), so, fixing open sets \(\Omega _{2}\Subset \Omega _{1}\Subset \Omega \) as in (5.31), and cubes \(Q_{\varrho }(x_{0})\Subset Q\Subset \Omega _{2}\), with \(B/16\Subset Q\equiv Q_{2^{-15/4}{} \texttt {r}}(x_{Q})\Subset B/8\), computations analogous to those in (5.23) eventually yield
where we set \({\texttt {M}}:=1+\Vert F^{\prime }(\nabla u) \Vert _{\textrm{L}^{\infty }(Q)}^{(q-p)/(q-1)}\), and it is \(c\equiv c(N,L,L_{\mu },p,q, \mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2},\partial \Omega _{1}))\), \({\texttt {m}}=1/2\). We can now apply classical Gehring lemma [41, Theorem 6.6] to prove the existence of an exponent \({\texttt {t}}\equiv {\texttt {t}}(N,L,L_{\mu },p,q, \mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2},\partial \Omega _{1}))\in (1,2)\), cf. (5.15), such that
where we also used (1.14), and it is \(c\equiv c(N,L,L_{\mu },p,q,\mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2},\partial \Omega _{1})\), \({\texttt {b}}\equiv {\texttt {b}}(p,q)\), and (1.17) is proven. By Sobolev Morrey embedding theorem we further obtain that \(V_{\mu ,p}(\nabla u)\in {\text {C}}^{0,1-1/{\texttt {t}}}(Q/2,\mathbb {R}^{N\times 2})\), which in turn implies that \(\nabla u\in {\text {C}}^{0,\beta _{0}}(Q/2,\mathbb {R}^{N\times 2})\), \(\beta _{0}\equiv \beta _{0}(N,L,L_{\mu },p,q,,\mathscr {F}(u;\Omega _{1}),\,\textrm{dist}(\Omega _{2},\partial \Omega _{1}))\in (0,1)\). With this last information at hand, if \(\mu >0\) in (2.1)\(_{3}\) we can follow a well-known strategy [41, Section 8.8] to conclude with gradient Hölder continuity up to any exponent less than one. If in addition to the nondegeneracy condition \(\mu >0\), the integrand governing \(\mathscr {F}\) is also real analytic, by-now standard theory [67, Chapter 6], [68] yields that u is real analytic. A standard covering argument finishes the proof.
5.3.1 Proof of Theorem 1.5, (ii.)
The proof of the second claim of Theorem 1.5 is a direct consequence of Proposition 3.7 and Theorem 1.4 formulated for minima of variational integrals governed by nondegenerate, analytic integrands. Recalling the content of Section 4.3.1, the proof of Theorem 1.5 is complete.
Remark 5.3
Our techniques apply also to the genuine (p, q)-nonuniform ellipticity conditions, i.e.
in the sense of bilinear forms, with \(\mu \in [0,1]\), and \(2\le p< q<2p\), the usual bound in two space dimensions. To adapt the arguments in Sects. 5.1, 5.2 to the (p, q)-framework, we recall that (5.32) implies
accordingly, the following (very minor) modifications need to be (respectively) implemented—higher differentiability results in the spirit of Theorem 1.2 that allow controlling the resulting right-hand side term being available in [30, 32].
-
Since (5.33) is in force, in (4.15) Hölder inequality needs to be applied with conjugate exponents \((p/(q-p),p/(2p-q))\), so in the last line of (4.15), term \({\texttt {P}}_{\varepsilon }'(\varrho )^{\frac{q-p}{2p}}\) replaces \({\texttt {S}}_{\varepsilon }'(\varrho )^{\frac{q-p}{2q}}\), and appears instead of . As a consequence, \(V_{\mu ,p}(\nabla u_{\varepsilon })\) replaces \(V_{1,q^{\prime }}(F^{\prime }(\nabla u_{\varepsilon }))\) everywhere in (5.4), and the term in parenthesis that needs to be controlled via trace theorem, now is , so in (5.2) the argument of the exponential must be replaced by \(|V_{\mu ,p}(\nabla u_{\varepsilon })|^{2}\) (any control on the stress tensor \(F^{\prime }(\nabla u_{\varepsilon })\) being lost in the genuine (p, q)-case).
-
In (5.22) the lower bound on \({\texttt {M}}_{\varepsilon }\) needs to be replaced by \(\max \{1,\Vert \nabla u_{\varepsilon } \Vert _{\textrm{L}^{\infty }(Q)}^{q-p}\}\); in (5.25) there must be \(|\nabla V_{\mu ,p}(\nabla u_{\varepsilon })|\) instead of \(|\nabla V_{1,q^{\prime }}(F^{\prime }(\nabla u_{\varepsilon }))|\) that is no longer available; and in (5.24)\(_{1}\) we need that \({\texttt {s}}_{0}= p/(2p-q)\) to be able to apply Young inequality and reabsorbe terms in (5.29).
Very recently, Schäffner [74] improved the assumptions for the validity of Lipschitz regularity for scalar local minimizers of (p, q)-nonuniformly elliptic functionals in two space dimensions by updating the bound on exponents (p, q) to \(q<3p\). The approach in [74] relies on a fine argument based on 1-d interpolation and maximum principle, that are in general not available in the vectorial setting.
6 The Scalar Case
This section is devoted to the proof, based on a renormalized [23] version of the Moser iteration designed in [6, 23, 24, 61], of local Lipschitz continuity for scalar minimizers of functional \(\mathscr {F}\) under Legendre (p, q)-growth conditions. Given the quite technical content of this last part of the paper, for simplicity we split it in two steps, ultimately yielding the proof of Theorem 1.6.
6.1 An Abstract \({\text {W}}^{1,2}\) to \(\textrm{L}^{\infty }\) Iteration
Here we derive an abstract \(\textrm{L}^{\infty }\)-\({\text {W}}^{1,2}\) estimate valid for any sufficiently regular function satisfying a suitable reverse Hölder type inequality for arbitrarily large powers.
Lemma 6.1
Let \(-1\le \alpha _{0}<\infty \), \(0<\gamma _{0}<1\) be numbers, \(v\in \textrm{L}^{\infty }_{{\text {loc}}}(B_{1}(0))\) be a nonnegative function such that \(v^{(\alpha _{0}+2)/2}\in {\text {W}}^{1,2}_{{\text {loc}}}(B_{1}(0))\) and, fixed balls \(B_{1/8}(0)\subset B_{\sigma }(0)\Subset B_{1}(0)\), \(\sigma \in (1/8,1]\) arbitrary, for all nonnegative \(\eta \in {\text {C}}^1_{c}(B_{\sigma }(0))\), any \(\alpha \ge \alpha _{0}\), and some absolute constants \(c_{0}\ge 1\), \({\texttt {M}}\ge 1\), inequality
holds true with
Then, for every ball \(B_{1/8}(0)\subseteq B_{\tau _{2}}(0)\subset B_{\tau _{1}}(0)\subseteq B_{\sigma }(0)\), the \(\textrm{L}^{\infty }\)-\({\text {W}}^{1,2}\) estimate
is verified, where
and it is \(c,{\texttt {a}}_{0}\equiv c,{\texttt {a}}_{0}(n,\alpha _{0},\gamma _{0})\).
Proof
We fix parameters \(1/8\le \tau _{2}<\tau _{1}\le \sigma < 1\), and, for \(i\in \mathbb {N}\cup \{0\}\) introduce radii \(\varrho _{i}:=\tau _{2}+2^{-i}(\tau _{1}-\tau _{2})\), balls \(B_{i}:=B_{\varrho _{i}}(0)\), cut-off functions \(\eta _{i}\in {\text {C}}^{1}_{c}(B_{i})\) such that \(\mathbbm {1}_{B_{i+1}}\le \eta _{i}\le \mathbbm {1}_{B_{i}}\), and apply [6, Lemma 3], see also [54, Lemma 3], to the right-hand side of (6.1) with \(\eta = \eta _{i}\). We obtain
for all \(\delta \in (0,1]\). With \(\gamma \in (0,1]\) as in (6.4), we record that
define \(\tilde{v}:=v^{\frac{\gamma (\alpha +2)}{2}}\), and estimate
for \(c\equiv c(n,\gamma _{0})\). We plug (6.7) in (6.5), choose \(\delta =\gamma \), and restore the original notation to get
where we also used that \(\varrho _{i}\ge 2^{-3}\) for all \(i\in \mathbb {N}\cup \{0\}\), and it is \(c\equiv c(n,\gamma _{0})\). Merging (6.8) and (6.1) we obtain,
with \(c\equiv c(n,c_{0},\gamma _{0})\). We have only one degree of freedom left in (6.9), that is \(\alpha \ge \alpha _{0}\ge -1\), so we introduce sequence \(\{\alpha _{i}\}_{i\in \mathbb {N}\cup \{0\}}\), recursively defined as
The above position immediately implies that
Set , and for \(i\in \mathbb {N}\) rearrange (6.9) as
with \(c\equiv c(n,c_{0},\gamma _{0})\). Iterating (6.10) over \(j\in \{0,\cdots ,i-1\}\) we end up with
Let us examine the behavior as \(i\rightarrow \infty \) of the various quantities appearing in (6.11). We have
and
for \(c,{\texttt {a}}_{0}\equiv c,{\texttt {a}}_{0}(n,\alpha _{0},\gamma _{0})\). We then send \(i\rightarrow \infty \) in (6.11) to conclude with (6.3) and the proof is complete. \(\square \)
6.2 A Caccioppoli Type Inequality for Powers
For \(z\in \mathbb {R}^{n}\) and \(\alpha \ge -1\), we introduce functions
and recall that whenever w is a twice differentiable function, the inequalities
holds up to constants depending only on p. Next, we look back at the family of approximating integrals defined in Sect. 4.1, and record that by (4.6)\(_{1,2,3}\), and classical scalar regularity results [41, Chapter 8] it is \(u_{\varepsilon }\in {\text {W}}^{1,\infty }_{{\text {loc}}}(B)\cap {\text {W}}^{2,2}_{{\text {loc}}}(B)\). Then we scale \(u_{\varepsilon }\) in such a way that function \(v_{\varepsilon }(x):=(u_{\varepsilon }(x_{0}+{\texttt {r}}x)-(u_{\varepsilon })_{B})/{\texttt {r}}\) minimizes \(\mathscr {F}\) on \(B_{1}(0)\), name \(v(x):=(u(x_{0}+{\texttt {r}}x)-(u)_{B})/{\texttt {r}}\), and observe that by construction it is
Now we are ready to prove a homogenized Caccioppoli type inequality involving arbitrary powers of \({\texttt {L}}(\nabla v_{\varepsilon })\).
Lemma 6.2
For all nonnegative \(\eta \in {\text {C}}^{1}_{c}(B_{1/4}(0))\), \(\alpha \ge -1\) and all constants \({\texttt {M}}_{\varepsilon }\ge 1\) such that
the Caccioppoli type inequality for powers
holds with \(c\equiv c(n,N,L,p,q)\), and \({\texttt {A}}_{\alpha }\) as in (6.2).
Proof
With \(\eta \in {\text {C}}^{1}_{c}(B_{1/4})\) being a nonnegative cut-off function, \(\alpha > -1\), and \(s\in \{1,\cdots ,n\}\), we test (4.12) against \(w_{s,\alpha }:=\eta ^{2}{\texttt {L}}(\nabla v_{\varepsilon })^{\alpha +1}\partial _{s}v_{\varepsilon }\), admissible by (6.14), to get
We immediately bound below
for \(c\equiv c(n,L,p,q)\). Now notice that
so we control
Finally, via Cauchy–Schwarz and Young inequalities we obtain
with \(c\equiv c(n,L,p,q)\). Merging the content of the two previous displays, we derive
with \(c\equiv c(n,L,p,q)\). Let us briefly discuss the case \(\alpha =-1\): in (6.17) term \(\text{(II) }\) vanishes, so we only need to estimate term \(\text{(III) }\) in a slightly different way than what we did before. Via Cauchy Schwarz and Young inequalities we get
for \(c\equiv c(n,L,p,q)\), thus
Finally, with (6.20), (6.17) at hand and keeping in mind the validity of (6.14), we can pick any constant \({\texttt {M}}_{\varepsilon }\ge 1\) satisfying (6.15) and turn (6.20), (6.18) into (6.16). The proof is complete. \(\square \)
6.3 Proof of Theorem 1.6
We start by observing that there is no loss of generality in assuming
otherwise the proof would be already finished. We then fix parameters \(1/8\le \tau _{2}<\tau _{1}\le 1/4\), corresponding to concentric balls \(B_{1/8}\subseteq B_{\tau _{2}}\subset B_{\tau _{1}}\subseteq B_{1/4}\), set \({\texttt {M}}_{\varepsilon }:=\Vert F_{\varepsilon }^{\prime }(\nabla v_{\varepsilon }) \Vert _{\textrm{L}^{\infty }(B_{\tau _{1}})}^{\frac{q-p}{q-1}}\) and notice that (6.16) holds in particular for all nonnegative \(\eta \in {\text {C}}^{1}_{c}(B_{\tau _{1}})\) and our choice of \({\texttt {M}}_{\varepsilon }\) matches (6.15) (keep in mind (6.22)). Moreover, by (6.14) we see that \({\texttt {L}}(\nabla v_{\varepsilon })\in \textrm{L}^{\infty }_{{\text {loc}}}(B_{1})\), and \({\texttt {l}}_{-1}(\nabla v_{\varepsilon })\in {\text {W}}^{1,2}_{{\text {loc}}}(B_{1})\). This, together with (6.16) allows applying Lemma 6.1 to \({\texttt {L}}(\nabla v_{\varepsilon })\), with \(\sigma =\tau _{1}\), \(\alpha _{0}=-1\), \(\gamma _{0}\) being any positive number less than p/q (say \(\gamma _{0}=p/(2q)<1\)), \({\texttt {A}}_{\alpha }\) as in (6.2), and \({\texttt {M}}={\texttt {M}}_{\varepsilon }\) to get
for \(c\equiv c(n,L,p,q)\), \({\texttt {a}}_{0}\equiv {\texttt {a}}_{0}(n,p,q)\). Now, from (1.12) and our choice of \(\gamma _{0}\) we deduce that
so in (6.23) we can apply Young inequality with conjugate exponents \(\bigg (\!\frac{p(1-\gamma )}{\gamma (q-p)}, \frac{p(1-\gamma )}{p-\gamma q}\! \bigg )\) to have
Lemma 2.6 then yields
thus via (3.5)\(_{1}\),
with \(c\equiv c(n,L,p,q)\). We then scale back on B to deduce via (6.13) and (4.21),
for \(c\equiv c(n,L,L_{\mu },p,q)\) and \({\texttt {b}}\equiv {\texttt {b}}(n,p,q)\) (possibly different from the one appearing in (4.21)). Finally, we send \(\varepsilon \rightarrow 0\) in the above display and use (4.5\(_{3}\)) to conclude with (1.18). A standard covering argument completes the proof.
Data Availibility
There is no data attached to this paper.
Notes
Specifically, for polynomials p denotes the ellipticity exponent in (2.1\(_{3}\)), while q indicates the degree.
Here we are subtracting a null Lagrangian from the convex integrand F, so this does not affect variational problem (1.1).
By evenness all the homogeneous components of P of odd degree vanish.
References
Antonini, C. A., Cianchi, A., Ciraolo, G., Farina, A., Maz’ya, V. G.: Global second-order estimates in anisotropic elliptic problems. Preprint (2023). arXiv:2307.03052
Ball, J.M.: Discontinuous equilibrium solutions and cavitation in nonlinear elasticity. Phil. Trans. R. Soc. Lond. 306, 557–611, 1982
Baroni, P.: Riesz potential estimates for a general class of quasilinear equations. Calc. Var. PDE 53, 803–846, 2015
Baroni, P., Di Castro, A., Palatucci, G.: Intrinsic geometry and de Giorgi classes for certain anisotropic problems. Disc. Cont. Dyn. Systems 10(4), 647–659, 2017
Beck, L., Mingione, G.: Lipschitz bounds and nonuniform ellipticity. Comm. Pure Appl. Math. 73, 944–1034, 2020
Bella, P., Schäffner, M.: On the regularity of minimizers for scalar integral functionals with \((p, q)\)-growth. Anal. PDE 13(7), 2241–2257, 2020
Bella, P., Schäffner, M.: Local boundedness and Harnack inequality for solutions of linear nonuniformly elliptic equations. Comm. Pure Appl. Math. LXXIV, 0453–0477, 2021
Bella, P., Schäffner, M.: Lipschitz bounds for integral functionals with \((p, q)\)-growth conditions. Adv. Calc. Var. 17(2), 373–390, 2024
Bildhauer, M., Fuchs, M.: Two dimensional anisotropic variational problems. Calc. Var. PDE 16, 177–186, 2003
Bonfanti, G., Cellina, A., Mazolla, M.: The higher integrability and the validity of the Euler-Lagrange equation for solutions to variational problems. SIAM J. Control Optim. 50, 888–899, 2012
Bousquet, P., Brasco, L.: \(\operatorname{C}^{1}\) regularity of orthotropic \(p\)-harmonic functions in the plane. Anal. PDE 11(4), 813–854, 2018
Bousquet, P., Brasco, L.: Lipschitz regularity for orthotropic functionals with nonstandard growth conditions. Rev. Math. Iberoam. 36(7), 1989–2032, 2020
Bögelein, V., Duzaar, F., Marcellini, P., Scheven, C.: Boundary regularity for elliptic systems with \(p, q\)-growth. J. Math. Pures Appl. 159, 250–293, 2022
Campanato, S.: Hölder continuity of the solutions of some non-linear elliptic systems. Adv. Math. 48, 16–43, 1983
Carozza, M., Kristensen, J., Passarelli di Napoli, A.: Higher differentiability of minimizers of convex variational integrals. Ann. I. H. Poincaré - AN 28, 395–411, 2011
Carozza, M., Kristensen, J., Passarelli di Napoli, A.: On the validity of the Euler-Lagrange system. Comm. Pure Appl. Anal. 14, 1, 2015
Carozza, M., Kristensen, J., Passarelli di Napoli, A.: Regularity of minimizers of autonomous convex variational integrals. Ann. Sc. Norm. Sup. Pisa Cl. Sci. XIII, 1065–1089, 2014
Cianchi, A.: Sharp Morrey-Sobolev inequalities and the distance from the extremals. Trans. Am. Math. Soc. 360(8), 4335–4347, 2008
Cianchi, A., Maz’ya, V.G.: Global boundedness of the gradient for a class of nonlinear elliptic systems. Arch. Ration. Mech. Anal. 212(1), 129–177, 2014
Cianchi, A., Schäffner, M.: Local boundedness of minimizers under unbalanced Orlicz growth conditions. J. Diff. Equ. 401, 58–92, 2024
Clozeau, N., Gloria, A.: Quantitative nonlinear homogenization: control of oscillations. Arch. Ration. Mech. Anal. 247, 67, 2023
De Filippis, C.: Quasiconvexity and partial regularity via nonlinear potentials. J. Math. Pures Appl. 163, 11–82, 2022
De Filippis, C., Mingione, G.: On the regularity of minima of non-autonomous functionals. J. Geom. Anal. 30, 1584–1626, 2020
De Filippis, C., Mingione, G.: Lipschitz bounds and nonautonomous integrals. Arch. Ration. Mech. Anal. 242, 973–1057, 2021
De Filippis, C., Mingione, G.: Nonuniformly elliptic Schauder theory. Invent. math. 234, 1109–1196, 2023
De Filippis, C., Piccinini, M.: Borderline global regularity for nonuniformly elliptic systems. Int. Math. Res. Notices 20(2023), 17324–17376, 2023
De Filippis, C., Stroffolini, B.: Singular multiple integrals and nonlinear potentials. J. Funct. Anal. 285(2), 109952, 2023
Douglis, A., Nirenberg, L.: Interior estimates for elliptic systems of partial differential equations. Comm. Pure Appl. Math. 8(4), 503–538, 1955
Ekeland, I., Témam, R.: Convex analysis and variational problems. Classics in Applied Mathematics, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA 28, (1999).
Esposito, L., Leonetti, F., Mingione, G.: Higher integrability for minimizers of integral functionals with \((p, q)\) growth. J. Diff. Equ. 157, 414–438, 1999
Esposito, L., Leonetti, F., Mingione, G.: Regularity results for minimizers of irregular integrals with \((p, q)\) growth. Forum Math. 14, 245–272, 2002
Esposito, L., Leonetti, F., Mingione, G.: Sharp regularity for functionals with \((p, q)\) growth. J. Diff. Equ. 204(1), 5–55, 2004
Evans, L.C.: Some new PDE methods for weak KAM theory. Calc. Var. PDE 17, 159–177, 2003
Franceschini, F.: Partial regularity for \(BV^{\cal{B}}\) minimizers. Preprint (2023). arXiv:2310.20002
Frehse, J.: On Signorini’s problem and variational problems with thin obstacles. Ann. Sc. Norm. Sup. Pisa Cl. Sci. 2, 343–362, 1977
Frehse, J.: Two dimensional variational problems with thin obstacle. Math. Z. 143, 279–288, 1975
Gehring, F.W.: The \(\rm L ^{p}\)-integrability of the partial derivatives of a quasi conformal mapping. Acta Math. 130, 265–277, 1973
Giaquinta, M.: Growth conditions and regularity, a counterexample. Manuscr. Math. 59, 245–248, 1987
Giaquinta, M., Modica, G.: Regularity results for some classes of higher order nonlinear elliptic systems. J. Reine Ang. Math. 311(312), 145–168, 1979
Giaquinta, M., Modica, G.: Remarks on the regularity of the minimizers of certain degenerate functionals. Manuscr. Math. 57, 55–99, 1986
Giusti, E.: Direct Methods in the calculus of variations. World Scientific Publishing Co, River Edge (2003)
Gmeineder, F.: Partial regularity for symmetric quasiconvex functionals on \(BD\). J. Math. Pures Appl. 145, 83–129, 2021
Gmeineder, F., Kristensen, J.: Partial regularity for BV minimizers. Arch. Ration. Mech. Anal. 232, 1429–1473, 2019
Gmeineder, F., J. Kristensen, F.: Quasiconvex functionals of \((p,q)\)-growth and the partial regularity of relaxed minimizers. Preprint (2022). arXiv:2209.01613
Hamburger, C.: Regularity of differential forms minimizing degenerate elliptic functionals. J. Reine Angew. Math. 431, 7–64, 1992
Hirsch, J., Schäffner, M.: Growth conditions and regularity, an optimal local boundedness result. Commun. Contemp. Math. 23(3), 2050029, 2021
Hiriart-Urruty, J.-B., Lemaréchal, C.: Convex Analysis and Minimization Algorithms I—Fundamentals. Springer-Verlag, Berlin (1996)
Hiriart-Urruty, J.-B., Lemaréchal, C.: Convex Analysis and Minimization Algorithms II. Springer, Berlin (2010)
Irving, C., Koch, L.: Boundary regularity results for minimisers of convex functionals with \((p, q)\)-growth. Adv. Nonlinear Anal. 12, 20230110, 2023
Jikov, V.V., Kozlov, S.M., Oleinik, O.A.: Homogenization of Differential Operators and Integral Functionals. Springer-Verlag, Berlin (1994)
Koch, L.: Global higher integrability for minimisers of convex functionals with \((p, q)\)-growth. Calc. Var. PDE 60(2), 63, 2021
Koch, L.: Global higher integrability for minimisers of convex obstacle problems with \((p, q)\)-growth. Calc. Var. PDE 61(3), 88, 2022
Koch, L., Kristensen, J.: On the validity of the Euler-Lagrange system without growth assumptions. Preprint (2022). arXiv:2203.00333
Koch, L., Ruf, M., Schäffner, M.: On the Lavrentiev gap for convex, vectorial integral functionals. Preprint (2023). arXiv:2305.19934
Kristensen, J.: A necessary and sufficient condition for lower semicontinuity. Nonlinear Anal. 120, 43–56, 2015
Ladyzhenskaya, O.A., Ural’tseva, N.N.: Linear and Quasilinear Elliptic Equations. Academic Press, New York-London (1968)
Li, Z.: Partial regularity for \(BV\)\(\omega \)-minimizers of quasiconvex functionals. Calc. Var. PDE 61, 178, 2022
Marcellini, P.: Growth conditions and regularity for weak solutions to nonlinear elliptic PDEs. J. Math. Anal. Appl. 501(1), 124408, 2021
Marcellini, P.: On the definition and the lower semicontinuity of certain quasiconvex integrals. Ann. Inst. H. Poincaré Anal. Non Linéaire 3, 391–409, 1986
Marcellini, P.: Un example de solution discontinue d’un probléme variationel dans le cas scalaire. Preprint, Ist. U. Dini, Firenze (1987)-(1988).
Marcellini, P.: Regularity of minimizers of integrals of the calculus of variations with non-standard growth conditions. Arch. Rat. Mech. Anal. 105, 267–284, 1989
Marcellini, P.: The stored-energy for some discontinuous deformations in nonlinear elasticity. Partial Differential Equations and the Calculus of Variations vol. II, Birkhäuser Boston Inc., (1989).
Marcellini, P.: Regularity and existence of solutions of elliptic equations with p, q-growth conditions. J. Differ. Equ. 90, 1–30, 1991
Marcellini, P.: Regularity for elliptic equations with general growth conditions. J. Diff. Equ. 105, 296–333, 1993
Marcellini, P.: Regularity for some scalar variational problems under general growth conditions. J. Optim. Theory Appl. 90, 161–181, 1996
Mooney, C., Savin, O.: Some singular minimizers in low dimensions in the Calculus of Variations. Arch. Ration. Mech. Anal. 221, 1–22, 2016
Morrey, Jr., C. B.: Multiple integrals in the calculus of variations. Classics in Mathematics Springer-Verlag, Berlin (2008)
Morrey, C.B., Jr.: On the analyticity of the solutions of analytic non-linear elliptic systems of partial differential equations: part I. Analyticity in the interior. Amer. J. Math. 80, 198–218, 1958
Nečas, J.: Sur la régularité des solutions variationelles des équations elliptiques non-linéaires d’ordre \(2k\) en deux dimensions. Ann. Scu. Norm. Sup. Pisa III(21), 427–457, 1967
Rindler, F.: Calculus of Variations. Springer, Heidelberg (2018)
Rockafellar, R.T.: Convex Analysis. Princeton University Press, Princeton (1970)
Ruf, M., Schäffner, M.: New homogenization results for convex integral functionals and their Euler-Lagrange equations. Calc. Var. PDE 63, 32, 2024
Schäffner, M.: Higher integrability for variational integrals with non-standard growth conditions. Calc. Var. Partial Differ. Equ. 60, 1–16, 2021
Schäffner, M.: Lipschitz bounds for nonuniformly elliptic integral functionals in the plane. Proc. Amer. Math. Soc. (2024). proc/16878
Schmidt, T.: Regularity of minimizers of \(\operatorname{W}^{1, p}\)-quasiconvex variational integrals with \((p, q)\)-growth. Calc. Var. 32, 1–24, 2008
Schmidt, T.: Regularity theorems for degenerate quasiconvex energies with (p, q)-growth. Adv. Calc. Var. 1(3), 241–270, 2008
Schmidt, T.: Regularity of Relaxed Minimizers of Quasiconvex Variational Integrals with \((p, q)\)-growth. Arch. Ration. Mech. Anal. 193, 311–337, 2009
Seregin, G.: On differentiability properties of the stress tensor in the Coulomb-Mohr theory of plasticity. St. Petersburg Math. J. 4(6), 1257–1272, 1993
Šverák, V., Yan, X.: Non-Lipschitz minimizers of smooth uniformly convex functionals. PNAS 99(24), 15269–15276, 2002
Trudinger, N.S.: On Imbeddings into Orlicz spaces and some applications. J. Math. Mech. 17(5), 473–483, 1967
Ural’tseva, N.N.: Degenerate quasilinear elliptic systems. Zap. Na. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 7, 184–222, 1968
Acknowledgements
C. De Filippis is supported by the University of Parma via the project “Local vs Nonlocal: mixed type operators and nonuniform ellipticity”, CUP_D91B21005370003, and by the INdAM GNAMPA project “Problemi non locali: teoria cinetica e non uniforme ellitticità”, CUP_E53C22001930001. We thank the anonymous referee for her/his helpful comments that eventually lead to an improved presentation of our results.
Funding
Open access funding provided by Università degli Studi di Parma within the CRUI-CARE Agreement.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by A. Braides.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
De Filippis, C., Koch, L. & Kristensen, J. Quantified Legendreness and the Regularity of Minima. Arch Rational Mech Anal 248, 69 (2024). https://doi.org/10.1007/s00205-024-02008-5
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00205-024-02008-5