betaFeedback

WorksMathematics

A singular minimizer of a uniformly convex Hessian-dependent functional

Abstract

We construct a smooth, uniformly convex Hessian-dependent energy functional, which admits a singular C1,1C^{1,1} minimizer on R5\mathbb{R}^5 that is a two-homogeneous normalization of Cartan's five-dimensional cubic. Our construction relies on constructing a quadratic null Lagrangian for the Hessian that is suitably non-degenerate on the image of the Hessian for the candidate minimizer. This complements a partial regularity result previously obtained by the authors in this setting, and demonstrates that one cannot hope for a fourth-order version of classical second-order elliptic regularity theory à la De Giorgi-Nash-Moser.

Introduction

In this article, we consider smooth uniformly convex functionals of the form

EF(v;Ω)=∫ΩF(D2v) dx,v:Ω⊂Rn⟶R(1)\mathcal{E}_F(v;\Omega)=\int_{\Omega}F(D^2v)\,dx,\qquad v:\Omega\subset\mathbb{R}^{n}\longrightarrow\mathbb{R} \tag*{(1)}

on a bounded domain Ω\Omega, with F∈C∞(Sym⁡(n))F\in C^{\infty}(\operatorname{Sym}(n)) satisfying

λ∣N∣2≤D2F(M)[N,N]≤Λ∣N∣2(2)\lambda\lvert N\rvert^2\le D^2F(M)[N,N]\le\Lambda\lvert N\rvert^2 \tag*{(2)}

for all M,N∈Sym⁡(n)M,N\in\operatorname{Sym}(n), for some 0<λ≤Λ<∞0<\lambda\le\Lambda<\infty. The associated Euler–Lagrange equation satisfied by critical points is div⁡div⁡DF(D2v)=0\operatorname{div}\operatorname{div}DF(D^2v)=0, understood weakly as

∫ΩDF(D2v):D2φ dx=0(3)\int_{\Omega}DF(D^2v):D^2\varphi\,dx=0 \tag*{(3)}

for every φ∈Cc∞(Ω)\varphi\in C^\infty_c(\Omega).

The upper bound in (2) gives quadratic growth of FF and linear growth of DFDF, so both the functional and its first variation are well-defined on W2,2(Ω)W^{2,2}(\Omega). Moreover, uniform convexity of EF\mathcal{E}_F implies that a weak solution of (3) is the unique minimizer in v+W02,2(Ω)v+W^{2,2}_0(\Omega).

Hessian-dependent energies and their associated fourth-order Euler–Lagrange equations occur naturally in elasticity and models of folding [48, 33, 31, 21], in liquid crystals [1], in the affine Plateau problem [49], and in the study of canonical metrics that are minimizers or critical points of various curvature-dependent energies [Wil65, Don02, Pan08, GV15]. Our particular motivation comes from Hamiltonian stationary Lagrangian submanifolds, see for instance [42, 43] for the variational formulation, and [45] for the constrained area minimization problem. Recall that the volume of a Lagrangian gradient graph {(x,Dv(x)):x∈Ω}⊆Rn×Rn\{(x,Dv(x)):x\in\Omega\}\subseteq\mathbb{R}^{n}\times\mathbb{R}^{n} is

∫Ωdet⁡(I+(D2v)2) dx,(4)\int_{\Omega}\sqrt{\det\left(I+(D^{2}v)^{2}\right)}\,dx, \tag*{(4)}

and the vanishing of its first variation under compactly supported Hamiltonian variations gives the Hamiltonian stationary equation for the potential vv (see [14 eqn (1.1)]). Note that the energy (4) is smooth and uniformly convex when ∣D2v∣|D^{2}v| is sufficiently small, but is not globally uniformly convex on Sym⁡(n)\operatorname{Sym}(n), see e.g. [8] for a more detailed discussion.

For gradient-dependent energy functionals of the form ∫ΩΦ(Dv) dx\int_{\Omega}\Phi(Dv)\,dx, smoothness of scalar minimizers under the corresponding assumptions follows from De Giorgi–Nash theory [16, 37], applied to the second order Euler–Lagrange equation. Equation (3) is instead fourth order, and this argument is no longer available. This raises the natural question of whether one expects scalar minimizers of Hessian-dependent energies (1) to also be smooth. Although a Schauder theory was developed by Bhattacharya & Warren in [11], demonstrating that C2,αC^{2,\alpha} critical points of (1) are smooth, this left open the question of whether W2,2W^{2,2} critical points are C2,αC^{2,\alpha}. When n=2n=2, this is indeed true, as demonstrated by Nečas [38 Théorème 3].

In general dimensions, merely ε\varepsilon-regularity results for W2,∞W^{2,\infty} critical points were known; see [5] for the case of small L∞L^{\infty} Hessian excess (or oscillation), and our earlier work [9] for the case of small LpL^{p} Hessian excess, i.e., small BMO seminorm. The latter in particular yields a partial regularity result for bounded-Hessian critical points: there exists p0∈(2,3)p_{0}\in(2,3) such that the critical point vv is smooth outside the relatively closed set

\Sigma_{v}=\left\{x\in\Omega:\liminf_{r\downarrow0}\fint_{B_{r}(x)}\left|D^{2}v-(D^{2}v)_{B_{r}(x)}\right|^{p_{0}}\,dy>0\right\}, \tag*{(5)}

which satisfies the dimension bound

dim⁡HΣv≤n−p0.(6)\dim_{\mathcal{H}}\Sigma_{v}\leq n-p_{0}. \tag*{(6)}

Here (h)_{B}=\fint_{B}h denotes the average over BB.

In this article, we show that singular minimizers for energies of the form (1) do exist in general. More precisely, we construct a uniformly convex energy functional of the form (1) on the space Sym⁡(5)\operatorname{Sym}(5) of symmetric 5×55\times5 matrices, for which the associated (unique) scalar minimizer on R5\mathbb{R}^{5} is a 2-homogeneous function that is not a quadratic polynomial, and thus even lies in W2,∞W^{2,\infty}, but not in C2C^{2} at the origin. This provides a first example of a singular minimizer in this setting.

Before providing the precise statement of our main theorem, we recall the Cartan isoparametric cubic PP on R5\mathbb{R}^{5} (see e.g. [12]), which is the three-homogeneous polynomial

P(x)=x13+3x12(x32+x42−2x52−2x22)+332(x2x32−x2x42+2x3x4x5).(7)P(x)=x_{1}^{3}+\frac{3x_{1}}{2}\left(x_{3}^{2}+x_{4}^{2}-2x_{5}^{2}-2x_{2}^{2}\right)+\frac{3\sqrt{3}}{2}\left(x_{2}x_{3}^{2}-x_{2}x_{4}^{2}+2x_{3}x_{4}x_{5}\right). \tag*{(7)}

where x=(x1,…,x5)x=(x_{1},\ldots,x_{5}). We then consider the two-homogeneous extension of its link in S4\mathbb{S}^{4}, i.e.

u(x)={P(x)∣x∣,x≠0,0,x=0.(8)u(x)= \begin{cases} \frac{P(x)}{|x|}, & x\ne0,\\ 0, & x=0. \end{cases} \tag*{(8)}

Note that uu is smooth away from the origin and has uniformly bounded Hessian on R5∖{0}\mathbb{R}^{5}\setminus\{0\} thanks to the zero-homogeneity of the latter, but uu is not twice differentiable at the origin. Indeed, if it were, its two-homogeneity would force it to agree with its quadratic Taylor polynomial, but uu is odd. In particular, it is easy to check that 0∈Σu0\in\Sigma_{u} as defined in (5).

Our main result is the following, in which we demonstrate the existence of an energy functional of the form (1.1) for which the above function uu is a global minimizer.

Theorem 1.1. There exists F∈C∞(Sym⁡(5))F\in C^{\infty}(\operatorname{Sym}(5)) satisfying (1.2) for a suitable choice of ellipticity constants λ,Λ\lambda,\Lambda, such that the function uu in (1.8) is a weak solution of (1.3) on B1⊂R5B_{1}\subset\mathbb{R}^{5}. Moreover, for every φ∈W02,2(B1)\varphi\in W_{0}^{2,2}(B_{1}),

EF(u+φ;B1)−EF(u;B1)≥λ2∫B1∣D2φ∣2 dx.(9)\mathcal{E}_{F}(u+\varphi;B_{1})-\mathcal{E}_{F}(u;B_{1})\geq\frac{\lambda}{2}\int_{B_{1}}|D^{2}\varphi|^{2}\,dx. \tag*{(9)}

In particular, uu is the unique minimizer with the smooth data u=Pu=P and ∂νu=2P\partial_{\nu}u=2P on ∂B1\partial B_{1}.

For u∈W2,∞(B1)u\in W^{2,\infty}(B_{1}), we will henceforth let

Reg⁡(u):={x: there exists a neighborhood U∋x with u∈C2(U)},\operatorname{Reg}(u):=\{x:\text{ there exists a neighborhood }U\ni x\text{ with }u\in C^{2}(U)\},
Sing⁡(u):=B1∖Reg⁡(u).\operatorname{Sing}(u):=B_{1}\setminus\operatorname{Reg}(u).

Note that Sing⁡(u)\operatorname{Sing}(u) is closed and is contained in Σu\Sigma_{u} as defined in (5). As a consequence of Theorem 1.1, we easily produce examples with codimension 55 singular sets in any dimension.

Corollary 1.2. For each n≥5n\geq5, there exists a smooth integrand F:Sym⁡(n)→RF:\operatorname{Sym}(n)\to\mathbb{R} satisfying (1.2) for some λ(n),Λ(n)\lambda(n),\Lambda(n), and a minimizer u∈W2,∞(B1)u\in W^{2,\infty}(B_{1}) such that

Sing⁡(u)=Σu=({0}×Rn−5)∩B1.(10)\operatorname{Sing}(u)=\Sigma_{u}=(\{0\}\times\mathbb{R}^{n-5})\cap B_{1}. \tag*{(10)}

In particular, dim⁡HSing⁡(u)=n−5\dim_{\mathcal{H}}\operatorname{Sing}(u)=n-5.

Remark 1.3. Note that strict convexity alone is insufficient for an ε\varepsilon-regularity theorem. This stems from the fact that, unlike in the uniformly convex case, there is no longer a Schauder theory for C2,αC^{2,\alpha}-regular critical points in general. Indeed, for each even integer p≥4p \ge4, an elementary one-dimensional example is

Fp(m)=mp,v(t)=sgn⁡(t)∣t∣2+1p−1,F_p(m)=m^p,\qquad v(t)=\operatorname{sgn}(t)|t|^{2+\frac{1}{p-1}},

which corresponds to the pp-biLaplacian; cf. [2 Remark 1.1] (see also [34]) for a corresponding class of examples for the pp-Laplacian. It is easy to check that FF is smooth and strictly convex but not uniformly convex, and vv lies in the class C2,1p−1C^{2,\frac{1}{p-1}} but not in C3C^3.

When combined with (1.6), this leaves open the question of the optimal p∈[p0,5]p \in[p_0,5] such that dim⁡HΣu≤n−p\dim_{\mathcal{H}}\Sigma_u \le n-p for any minimizer uu of an energy functional of the form (1.1).

One may compare our partial regularity results to those for vectorial minimizers of gradient-dependent energies. Indeed, one may view (1.1) as a gradient-dependent energy on gradient vector fields. The regularity theory for vectorial critical points uu of uniformly elliptic gradient-dependent integrands has been the subject of extensive study; the same dimension estimate (1.6) on the singular set is known for some p0∈(2,3)p_0 \in(2,3) in that setting [26, 25, 35, 24]; while singular minimizers of uniformly convex first-gradient integrals do indeed exist. The first result in the latter direction is due to Nečas [39], where a smooth, uniformly convex gradient-dependent energy functional admitting a singular minimizer is constructed in sufficiently high dimension. In subsequent work [29], Hao, Leonardi & Nečas produced an example with an isolated singularity in dimension five. Šverák & Yan [46] then proceeded to demonstrate the existence of a Lipschitz (see also [47] for non-Lipschitz examples) singular minimizer of a suitable gradient-dependent energy from R3\mathbb{R}^3 to R5\mathbb{R}^5 via first constructing a suitable null Lagrangian, and then finding a uniformly convex function whose derivative matches that of the null Lagrangian on gradients. More recently, Mooney & Savin [36] obtained a singular minimizer u:R3→R2u:\mathbb{R}^3 \to\mathbb{R}^2 with an isolated singularity.

In the vector-valued setting, energy functionals satisfying weaker notions of convexity such as quasiconvexity and polyconvexity are also frequently considered, since they appear more naturally in many physical contexts; all these weaker notions coincide with classical convexity on scalar functions. For minimizers of strongly quasiconvex functionals, a significantly weaker bound on the singular set is currently known, see [20, 32, 15]. Meanwhile, in the recent work [27], Guerra & Tione constructed a Lipschitz singular minimizer in two dimensions for a smooth strongly polyconvex integral.

When compared to the aforementioned vector-valued setting, the additional gradient structure of vectorial minimizers that we have herein both allows us to potentially hope for better regularity and structure than that for gradient-dependent energies on vector-valued functions, and prevents us from using the singular examples (which are not gradients) constructed in the latter setting. One can also see the difference at the level of the Euler-Lagrange equation, for which our test vector fields also have a gradient structure, therefore yielding the scalar weak fourth-order equation div⁡div⁡DF(D2u)=0\operatorname{div}\operatorname{div} DF(D^2u)=0, rather than the second-order system associated with arbitrary vector-valued variations.

The construction of the energy and associated singular minimizer herein is instead motivated by partial regularity results for Hamiltonian stationary graphs. Building on earlier regularity results [14, 10, 4 Bha26, 7] in that setting, together with Orriols, the present authors proved a sharp codimension-five partial regularity result for Lipschitz Hamiltonian stationary graphs with locally W1,2W^{1,2} Lagrangian angle in [8]. This in particular yields sharpness of the previously known partial regularity [19, 17] for Lipschitz special Lagrangian graphs. The sharpness is realized by a construction of a singular Lipschitz graphical special Lagrangian cone in dimension five, using an ansatz derived from the two-homogeneous normalization of Cartan’s isoparametric cubic (1.8). Constructions modeled on the Cartan cubic have previously appeared in results yielding singular solutions to uniformly elliptic second order equations by Nadirashvili & Vlăduţ [NV07] and Nadirashvili, Tkachev & Vlăduţ [40] (see also [NV13]), as well as in results producing new examples of singular minimal hypercones [22] and solutions to the one-phase Bernoulli free boundary problem [30, 50 FTZ26, 13].

The novelty herein is to realize (1.8) itself as a minimizer of a suitable energy functional (1.1). The freedom of tailoring the energy to this particular singular function gives us an analogue to that in [40], though the variational problem considered here is different. This is somewhat in contrast to what we do in [8], where instead the PDE is fixed to be the zero-phase special Lagrangian equation, which is compensated by the flexibility of seeking a solution that is a real analytic reparameterization of the level sets of the Cartan cubic foliation.

Overview of proof & structure of paper

In order to find an energy functional FF satisfying (1.2) for which (1.8) is a minimizer, we follow the null-Lagrangian approach of Šverák & Yan [46]. Namely, we seek a quadratic null Lagrangian LL (see § 4) satisfying the uniform non-degeneracy condition

L(M−N)≥σ∣M−N∣2for M,N∈K:={D2u(ξ):ξ∈S1(3)},(11)L(M-N) \ge\sigma|M-N|^2 \qquad\text{for } M,N \in K := \{D^2u(\xi): \xi\in\mathcal{S}_1(3)\}, \tag*{(11)}

for a suitable choice of σ>0\sigma>0, where uu is our candidate singular minimizer (1.8). To this end, we explore some key identities satisfied by the Cartan cubic to first find two canonical quadratic null-Lagrangians L0L_0, L1L_1 that have the same SO(3)SO(3) conjugation action invariance as our candidate singular minimizer.

These computations are performed in Section 3 and the beginning of Section 4. Much like in [36], we proceed to check that the degeneracies of L0L_0 and L1L_1 compensate one another. The remainder of Section 4 is then dedicated to identifying the degenerate directions of L1L_1 in KK, and showing that L0L_0 is uniformly nondegenerate there, thus yielding the desired non-degeneracy (1.11) for a suitable linear combination L=L0+bL1L=L_0+bL_1.

This is a sufficient condition in order to guarantee the existence of an extension to a uniformly convex integrand FF satisfying

F=LandDF=DLon K,F=L \quad\text{and} \quad DF=DL \quad\text{on }K,

as already observed in [46, 36]. In Section 5, we recall this extension result and verify that it applies herein and yields an extension satisfying our desired properties, therefore allowing us to conclude our main results.

The fact that LL is a null-Lagrangian whose gradient agrees with that of FF on KK then immediately yields the weak Euler-Lagrange equation

div⁡div⁡DF(D2u)=div⁡div⁡DL(D2u)=0\operatorname{div}\operatorname{div}DF(D^2u)=\operatorname{div}\operatorname{div}DL(D^2u)=0

for our choice of uu by construction.

Acknowledgements

A.B. acknowledges the support of NSF grant DMS-2350290, the Simons Foundation grant MPS-TSM-00002933, and a Bill Guthridge Fellowship from UNC-Chapel Hill. A.S. is grateful for the generous support of Dr. Max Rössler, the Walter Haefner Foundation, and the ETH Zürich Foundation. This research was conducted during the period A.S. served as a Clay Research Fellow.

Note on LLM Usage

The authors came up with the proof strategy and ideas underlying this paper in 2024, namely, to find an appropriate null Lagrangian that respects the same symmetries as some candidate singular minimizer, and then extend to a uniformly convex energy. The only missing ingredient at that stage was a suitable singular candidate for a minimizer. Following the completion of the recent work [8], the authors realized that the Cartan potential P(X)/∣X∣P(X)/|X| provides precisely such a candidate, thereby completing the strategy. At this stage, AI was used as a computational aid for some of the algebraic calculations involved in identifying suitable quadratic null Lagrangians. The authors reworked these computations extensively, substantially simplifying, modifying, and entirely rewriting the arguments. The idea behind the examples in Remark 1.3 was suggested by AI. All the text in this paper was written by the authors.

Notation

Unless otherwise specified, positive constants cc and CC depend only on the normalization of PP in (1.8); note that there is no dependence on nn since we fix n=5n=5. We write Sym⁡(n)\operatorname{Sym}(n) for the space of symmetric n×nn \times n matrices, while Sym⁡0(n)\operatorname{Sym}_{0}(n) denotes the trace-free subspace of these. We let tr⁡M\operatorname{tr} M denote the trace of an n×nn \times n matrix MM. We write S(n)\mathcal{S}(n) to denote the space of self-adjoint linear operators from Sym⁡0(n)\operatorname{Sym}_{0}(n) to Sym⁡0(n)\operatorname{Sym}_{0}(n), which may be identified with Sym⁡(n(n+1)2−1)\operatorname{Sym}\left(\frac{n(n+1)}{2}-1\right). We write either ⟨X,Y⟩=tr⁡(XY)\langle X,Y\rangle=\operatorname{tr}(XY) or X:YX:Y for the Frobenius inner product on Sym⁡(n)\operatorname{Sym}(n). We let ∣⋅∣|\cdot| denote the norm induced by the Frobenius inner product, and ∥⋅∥op\|\cdot\|_{\mathrm{op}} to denote the operator norm; we may equip both Sym⁡0(n)\operatorname{Sym}_{0}(n) and S(n)\mathcal{S}(n) with both. Note that nn−1∥⋅∥op≤∣⋅∣≤C(n)∥⋅∥op\sqrt{\frac{n}{n-1}}\|\cdot\|_{\mathrm{op}}\leq|\cdot|\leq C(n)\|\cdot\|_{\mathrm{op}} on Sym⁡0(n)\operatorname{Sym}_{0}(n). The identity operators on Sym⁡0(n)\operatorname{Sym}_{0}(n) and Rn\mathbb{R}^{n} are denoted by Id⁡Sym⁡0(n)\operatorname{Id}_{\operatorname{Sym}_{0}(n)} and Id⁡n\operatorname{Id}_{n}, respectively. We let S1(n):={X∈Sym⁡0(n):∣X∣=1}\mathcal{S}_{1}(n):=\{X\in\operatorname{Sym}_{0}(n):|X|=1\} denote the unit sphere in Sym⁡0(n)\operatorname{Sym}_{0}(n). For X,Y∈Sym⁡0(n)X,Y\in\operatorname{Sym}_{0}(n), we further use X⊗YX\otimes Y to denote the tensor product, which can be seen as a linear operator on Sym⁡0(n)\operatorname{Sym}_{0}(n), given by (X⊗Y)Z=(X:Z)Y(X\otimes Y)Z=(X:Z)Y.

Key identities for PP and uu

We begin by recording some useful elementary identities satisfied by the Cartan cubic PP in (1.7) and its associated two-homogeneous normalization uu in (1.8). First of all, note that the three-homogeneity of PP in particular yields

DP(X)⋅X=3P(X),D2P(X)X=2DP(X).(12)DP(X)\cdot X=3P(X),\qquad D^{2}P(X)X=2DP(X). \tag*{(12)}

Moreover, recall that via the map

Q(x):=(2x13x33x43x3−x1+3x23x53x43x5−x1−3x2),Q(x):= \begin{pmatrix} 2x_{1} & \sqrt{3}x_{3} & \sqrt{3}x_{4}\\ \sqrt{3}x_{3} & -x_{1}+\sqrt{3}x_{2} & \sqrt{3}x_{5}\\ \sqrt{3}x_{4} & \sqrt{3}x_{5} & -x_{1}-\sqrt{3}x_{2} \end{pmatrix},

we may identify R5\mathbb{R}^{5} with the space Sym⁡0(3)\operatorname{Sym}_{0}(3) of symmetric trace-free matrices. Moreover, since

P(x)=12det⁡Q(x),∣x∣2=16tr⁡(Q(x)2),P(x)=\frac{1}{2}\det Q(x),\qquad|x|^{2}=\frac{1}{6}\operatorname{tr}\left(Q(x)^{2}\right),

the conjugation action ρR(Q):=RQRT\rho_{R}(Q):=RQR^{T}, R∈SO⁡(3)R\in\operatorname{SO}(3), induces an orthogonal action on R5\mathbb{R}^{5} under which PP and uu are invariant. In particular, we may henceforth work with isometric coordinates X=Q(x)/6X=Q(x)/\sqrt{6}, thus identifying PP with a function on Sym⁡0(3)\operatorname{Sym}_{0}(3) (not relabeled), so that ∣X∣2=∣x∣2|X|^{2}=|x|^{2} and P(X)=36det⁡X=6tr⁡(X3)P(X)=3\sqrt{6}\det X=\sqrt{6}\operatorname{tr}(X^{3}).

We begin with the following lemma, which provides us with some useful basic algebraic properties for PP that arise from its symmetries.

Lemma 3.1. For every X,Y∈Sym⁡0(3)X,Y \in\operatorname{Sym}_0(3),

DP(X)=36(X2−∣X∣23Id⁡3),(13)DP(X)=3\sqrt{6}\left(X^2-\frac{|X|^2}{3}\operatorname{Id}_3\right), \tag*{(13)}
D2P(X)Y=36(XY+YX−23tr⁡(XY)Id⁡3),(14)D^2P(X)Y=3\sqrt{6}\left(XY+YX-\frac{2}{3}\operatorname{tr}(XY)\operatorname{Id}_3\right), \tag*{(14)}
ΔP(X)=0,∣DP(X)∣2=9∣X∣4,∣P(X)∣≤∣X∣3,(15)\Delta P(X)=0,\qquad|DP(X)|^2=9|X|^4,\qquad|P(X)|\leq|X|^3, \tag*{(15)}
D2P(X)DP(X)=18∣X∣2X,(16)D^2P(X)DP(X)=18|X|^2X, \tag*{(16)}
⟨D2P(X),D2P(Y)⟩=126⟨X,Y⟩,∣D2P(X)∣=126∣X∣.(17)\langle D^2P(X),D^2P(Y)\rangle=126\langle X,Y\rangle,\qquad|D^2P(X)|=\sqrt{126}|X|. \tag*{(17)}

Proof. The identities (3.2), (3.3), and the first two identities in (3.4) all follow directly from differentiating P(X)=36det⁡XP(X)=3\sqrt{6}\det X, while the final estimate in (3.4) follows from (3.2), the three-homogeneity relation in (3.1), and Cauchy-Schwarz. As for the identity (3.5), this is an immediate consequence of differentiating the second identity in (3.4).

Finally, to obtain the relations in (3.6), we combine the elementary Hilbert space identity

⟨A,B⟩=14(∥A+B∥2−∥A−B∥2),\langle A,B\rangle=\frac{1}{4}\left(\|A+B\|^2-\|A-B\|^2\right),

with Bochner’s identity and (3.4).

Let us now use the identities and estimates in Lemma 3.1 in order to characterize the behavior of the Hessian of uu. Firstly, note that by the two-homogeneity of uu, it suffices to understand D2uD^2u on the unit sphere in Sym⁡0(3)\operatorname{Sym}_0(3) (which corresponds to S4S^4 by our isometric choice of coordinates). Differentiating uu, we find

D2u(X)=D2P(X)∣X∣−1−DP(X)⊗X∣X∣3−X∣X∣3⊗DP(X)−P(X)∣X∣−3Id⁡Sym⁡0(3)+3P(X)∣X∣−5X⊗X,D^2u(X)=D^2P(X)|X|^{-1}-DP(X)\otimes\frac{X}{|X|^3}-\frac{X}{|X|^3}\otimes DP(X)-P(X)|X|^{-3}\operatorname{Id}_{\operatorname{Sym}_0(3)}+3P(X)|X|^{-5}X\otimes X,

and thus

D2u(X)=D2P(X)−DP(X)⊗X−X⊗DP(X)−P(X)Id⁡Sym⁡0(3)+3P(X)X⊗X(18)D^2u(X)=D^2P(X)-DP(X)\otimes X-X\otimes DP(X)-P(X)\operatorname{Id}_{\operatorname{Sym}_0(3)}+3P(X)X\otimes X \tag*{(18)}

for any X∈S1(3):={X∈Sym⁡0(3):∣X∣=1}X\in\mathcal{S}_1(3):=\{X\in\operatorname{Sym}_0(3):|X|=1\}.

Next, we record some important properties of D2uD^2u that will be particularly important later, when extending the non-degenerate quadratic null Lagrangian that we end up with to a uniformly convex integrand.

Lemma 3.2. For X∈S1(3)X\in\mathcal{S}_1(3),

Δu(X)=−8P(X),∣D2u(X)∣2=72+38P(X)2.(19)\Delta u(X)=-8P(X),\qquad|D^2u(X)|^2=72+38P(X)^2. \tag*{(19)}

In particular,

72≤∣D2u(X)∣≤CHess,(20)\sqrt{72} \leq\lvert D^{2}u(X)\rvert\leq C_{\mathrm{Hess}}, \tag*{(20)}

where CHess=max⁡S1(3)∣D2u∣<∞C_{\mathrm{Hess}}=\max_{\mathcal{S}_{1}(3)}\lvert D^{2}u\rvert<\infty only depends on PP.

Proof. Let X∈S1(3)X\in\mathcal{S}_{1}(3). To see that Δu=−8P\Delta u=-8P, we simply take the trace in (18), and combine with the properties ΔP=0\Delta P=0 and (12), as well as the fact that the dimension of Sym⁡0(3)\operatorname{Sym}_{0}(3) is 5. As for the identity for ∣D2u(X)∣2\lvert D^{2}u(X)\rvert^{2}, we may rewrite (18) as D2u(X)=D2P(X)−A0+B0D^{2}u(X)=D^{2}P(X)-A_{0}+B_{0}, with

A0=DP(X)⊗X+X⊗DP(X),B0=−P(X)Id⁡Sym⁡0(3)+3P(X)X⊗X.(21)A_{0}=DP(X)\otimes X+X\otimes DP(X), \qquad B_{0}=-P(X)\operatorname{Id}_{\operatorname{Sym}_{0}(3)}+3P(X)X\otimes X. \tag*{(21)}

Expanding ⟨D2u(X),D2u(X)⟩\langle D^{2}u(X),D^{2}u(X)\rangle via this decomposition, and invoking Lemma 3.1, (12) and the identity ⟨a⊗b,c⊗d⟩=(a⋅c)(b⋅d)\langle a\otimes b,c\otimes d\rangle=(a\cdot c)(b\cdot d), we obtain

∣D2u(X)∣2=∣D2P(X)∣2+∣A0∣2+∣B0∣2−2⟨D2P(X),A0⟩+2⟨D2P(X),B0⟩−2⟨A0,B0⟩\lvert D^{2}u(X)\rvert^{2}=\lvert D^{2}P(X)\rvert^{2}+\lvert A_{0}\rvert^{2}+\lvert B_{0}\rvert^{2}-2\langle D^{2}P(X),A_{0}\rangle+2\langle D^{2}P(X),B_{0}\rangle-2\langle A_{0},B_{0}\rangle

where

∣D2P(X)∣2=126,∣A0∣2=18+18P(X)2,∣B0∣2=8P(X)2,\lvert D^{2}P(X)\rvert^{2}=126,\qquad\lvert A_{0}\rvert^{2}=18+18P(X)^{2},\qquad\lvert B_{0}\rvert^{2}=8P(X)^{2},
⟨D2P(X),A0⟩=36,⟨D2P(X),B0⟩=18P(X)2,⟨A0,B0⟩=12P(X)2.\langle D^{2}P(X),A_{0}\rangle=36,\qquad\langle D^{2}P(X),B_{0}\rangle=18P(X)^{2},\qquad\langle A_{0},B_{0}\rangle=12P(X)^{2}.

This yields the second identity in (19) as desired. The bounds in (20) follow immediately.

We conclude this section with the following lemma, which guarantees smoothness and embeddedness of the set KK given by

K:={D2u(X):X∈S1(3)},(22)K:=\{D^{2}u(X):X\in\mathcal{S}_{1}(3)\}, \tag*{(22)}

cf. (11). This is one of the properties that we will require to extend our final quadratic null Lagrangian to a uniformly convex integrand.

Lemma 3.3. The map D2u:S1(3)→S(3)D^{2}u:\mathcal{S}_{1}(3)\to\mathcal{S}(3) satisfies

∣D2u(X)−D2u(Y)∣2≥452∣X−Y∣2X,Y∈S1(3).(23)\lvert D^{2}u(X)-D^{2}u(Y)\rvert^{2}\geq\frac{45}{2}\lvert X-Y\rvert^{2}\qquad X,Y\in\mathcal{S}_{1}(3). \tag*{(23)}

In particular, D2u∣S1(3)D^{2}u|_{\mathcal{S}_{1}(3)} is a smooth embedding, i.e. KK is a smooth embedded submanifold.

Proof. Fix X,Y∈S1(3)X,Y\in\mathcal{S}_{1}(3). Let us directly compute the expansion of ∣D2u(X)−D2u(Y)∣2\lvert D^{2}u(X)-D^{2}u(Y)\rvert^{2}, analogously to the computation in the proof of Lemma 3.2. Let A0A_{0} and B0B_{0} be as in (21), and note that D2PD^{2}P is linear. This time, it will be more convenient to keep a piece of the D2PD^{2}P term with A0−B0A_{0}-B_{0}, in order to ensure that we end up with a positive lower bound while crudely estimating the square term involving A0−B0A_0-B_0 below by zero. More precisely, we have

∣D2u(X)−D2u(Y)∣2=34∣D2P(X−Y)∣2+∣12D2P(X−Y)−(A0(X)−A0(Y))+B0(X)−B0(Y)∣2−⟨D2P(X−Y),A0(X)−A0(Y)⟩+⟨D2P(X−Y),B0(X)−B0(Y)⟩.\lvert D^2u(X)-D^2u(Y)\rvert^2=\frac{3}{4}\lvert D^2P(X-Y)\rvert^2 +\left\lvert\frac{1}{2}D^2P(X-Y)-(A_0(X)-A_0(Y))+B_0(X)-B_0(Y)\right\rvert^2 -\langle D^2P(X-Y),A_0(X)-A_0(Y)\rangle +\langle D^2P(X-Y),B_0(X)-B_0(Y)\rangle.

By Lemma 3.1, the symmetry of D3PD^3P (cf. (27)), and (3.1), we have

∣D2P(X−Y)∣2=126∣X−Y∣2,\lvert D^2P(X-Y)\rvert^2=126\lvert X-Y\rvert^2,
⟨D2P(X−Y),A0(X)⟩=2⟨D2P(X)DP(X),X−Y⟩=36⟨X,X−Y⟩,\langle D^2P(X-Y),A_0(X)\rangle=2\langle D^2P(X)DP(X),X-Y\rangle=36\langle X,X-Y\rangle,
⟨D2P(X−Y),B0(X)⟩=3P(X)⟨D2P(X)X,X−Y⟩=6P(X)⟨DP(X),X−Y⟩,\langle D^2P(X-Y),B_0(X)\rangle=3P(X)\langle D^2P(X)X,X-Y\rangle=6P(X)\langle DP(X),X-Y\rangle,

Inserting these into the above identity, and throwing away the square term involving A0−B0A_0-B_0, yields

∣D2u(X)−D2u(Y)∣2≥1172∣X−Y∣2+6⟨P(X)DP(X)−P(Y)DP(Y),X−Y⟩.(24)\lvert D^2u(X)-D^2u(Y)\rvert^2\geq\frac{117}{2}\lvert X-Y\rvert^2+6\langle P(X)DP(X)-P(Y)DP(Y),X-Y\rangle. \tag*{(24)}

In order to bound the final term on the right-hand side above, we proceed via the Fundamental Theorem of Calculus, using a lower bound on D(P(X)DP(X))D(P(X)DP(X)). More precisely, notice that

⟨P(X)DP(X)−P(Y)DP(Y),X−Y⟩=∫01D(P(Y+t(X−Y))DP(Y+t(X−Y)))[X−Y,X−Y] dt.\langle P(X)DP(X)-P(Y)DP(Y),X-Y\rangle =\int_0^1D\bigl(P(Y+t(X-Y))DP(Y+t(X-Y))\bigr)[X-Y,X-Y]\,dt.

Moreover, note that D(P(Z)DP(Z))=DP(Z)⊗DP(Z)+P(Z)D2P(Z)D(P(Z)DP(Z))=DP(Z)\otimes DP(Z)+P(Z)D^2P(Z). Again invoking the symmetry of D3PD^3P, we have

∣⟨D2P(Z)W,W⟩∣=2∣⟨DP(W),Z⟩∣≤6∣Z∣∣W∣2.\lvert\langle D^2P(Z)W,W\rangle\rvert=2\lvert\langle DP(W),Z\rangle\rvert\leq6\lvert Z\rvert\lvert W\rvert^2.

Combining this with the fact that DP(Z)⊗DP(Z)DP(Z)\otimes DP(Z) is positive semidefinite and that Z∈[X,Y]⊂{W∈Sym⁡0(3):∣W∣≤1}Z\in[X,Y]\subset\{W\in\operatorname{Sym}_0(3):\lvert W\rvert\leq1\}, and exploiting the final inequality (3.4), we therefore obtain

D(P(Z)DP(Z))≥−6Id⁡Sym⁡0(3).D(P(Z)DP(Z))\geq-6\operatorname{Id}_{\operatorname{Sym}_0(3)}.

Inserting this lower bound into the integrated identity above, and combining with (3.13), we arrive at (23) as desired, which in turn tells us that D2uD^2u is an injective immersion on S1(3)\mathcal{S}_1(3). The fact that it is a smooth embedding follows immediately by compactness of S1(3)\mathcal{S}_1(3).

A non-degenerate quadratic null Lagrangian

We are now in a position to construct a quadratic null Lagrangian for the Hessian, i.e. a quadratic form LL for which the weak double-divergence form equation

∫DL(D2v):D2φ=0for every φ∈Cc∞(B1)\int DL(D^2v):D^2\varphi=0 \qquad\text{for every }\varphi\in C^\infty_c(B_1)

holds for every v∈Wloc2,2v\in W^{2,2}_{\mathrm{loc}}. See [3] for more details about the theory of null Lagrangians. Taking the Fourier transform (as a tempered distribution) of div⁡div⁡DL(D2(vχ))\operatorname{div}\operatorname{div}DL(D^2(v\chi)) for χ\chi a smooth compactly supported function that is identically 1 on supp⁡φ\operatorname{supp}\varphi, we immediately see that a sufficient condition for LL to be a null Lagrangian is thus the algebraic relation

L(ξ⊗ξ)=0for every ξ∈Sym⁡0(3)≅R5.(25)L(\xi\otimes\xi)=0 \qquad\text{for every }\xi\in\operatorname{Sym}_0(3)\cong\mathbb{R}^5. \tag*{(25)}

Inspired by the computations in [46 Proof of Lemma 2.1, Section 3.1], our construction of a suitable null Lagrangian will rely on understanding the irreducible SO(3)SO(3)-invariant subspaces. The expression (18) suggests a canonical decomposition of S(3)\mathcal{S}(3) into

S(3)=D2P(Sym⁡0(3))⊕V0⊕⟨Id⁡Sym⁡0(3)⟩,\mathcal{S}(3)=D^2P(\operatorname{Sym}_0(3))\oplus V_0\oplus\langle\operatorname{Id}_{\operatorname{Sym}_0(3)}\rangle,

where

V0:=(D2P(Sym⁡0(3))⊕⟨Id⁡Sym⁡0(3)⟩)⊥=(D2P(Sym⁡0(3)))⊥∩{M∈S(3):tr⁡M=0},\begin{aligned} V_0:=\left(D^2P(\operatorname{Sym}_0(3))\oplus\langle\operatorname{Id}_{\operatorname{Sym}_0(3)}\rangle\right)^\perp\\ &=\left(D^2P(\operatorname{Sym}_0(3))\right)^\perp\cap\{M\in\mathcal{S}(3):\operatorname{tr}M=0\}, \end{aligned}

where the orthogonal complement is within S(3)\mathcal{S}(3). It is easy to verify that these SO(3)SO(3)-invariant subspaces are irreducible. It is a classical fact (a particular case of Schur’s Lemma) from invariant operator theory (see e.g. [51]) that any SO(3)SO(3)-invariant quadratic form LL on S(3)\mathcal{S}(3) must therefore be of the form

L=a1∣M1∣2+a2∣M2∣2+a3∣M3∣2,L=a_1|M_1|^2+a_2|M_2|^2+a_3|M_3|^2,

for M1∈D2P(Sym⁡0(3))M_1\in D^2P(\operatorname{Sym}_0(3)), M2∈V0M_2\in V_0, M3∈⟨Id⁡Sym⁡0(3)⟩M_3\in\langle\operatorname{Id}_{\operatorname{Sym}_0(3)}\rangle. This decomposition is precisely what we will make use of in order to make appropriate choices of quadratic null Lagrangians.

Now, let us characterize (D2P(Sym⁡0(3)))⊥\left(D^2P(\operatorname{Sym}_0(3))\right)^\perp in a more convenient way. Let T:S(3)→Sym⁡0(3)\mathcal{T}:\mathcal{S}(3)\to\operatorname{Sym}_0(3) be the adjoint operator of D2PD^2P, namely

⟨T(M),Y⟩=⟨M,D2P(Y)⟩for M∈S(3), Y∈Sym⁡0(3),\langle\mathcal{T}(M),Y\rangle=\langle M,D^2P(Y)\rangle\qquad\text{for }M\in\mathcal{S}(3),\ Y\in\operatorname{Sym}_0(3),

so that

(D2P(Sym⁡0(3)))⊥={M∈S(3):T(M)=0}.\left(D^2P(\operatorname{Sym}_0(3))\right)^\perp=\{M\in\mathcal{S}(3):\mathcal{T}(M)=0\}.

In conclusion,

V0={M∈S(3):tr⁡M=0, T(M)=0}.V_0=\{M\in\mathcal{S}(3):\operatorname{tr}M=0,\ \mathcal{T}(M)=0\}.

Let us further understand how T\mathcal{T} acts on KK in terms of X∈S1(3)X \in\mathcal{S}_1(3); this will be crucial in order to verify the non-degeneracy conditions required from our choices of quadratic null Lagrangians.

Recalling (3.7), this reduces to characterizing its individual action on each of the components of the Hessian therein. First of all, in light of (3.4) and (3.6) from Lemma 3.1, we have

T(D2P(X))=126X,∥T∥op=126,T(Id⁡Sym0(3))=0.(26)\mathcal{T}(D^2P(X)) = 126X,\qquad\|\mathcal{T}\|_{\mathrm{op}} = \sqrt{126},\qquad\mathcal{T}(\operatorname{Id}_{\mathrm{Sym}_0(3)}) = 0. \tag*{(26)}

As for the action of T\mathcal{T} on rank-one tensors, the symmetry of D3PD^3P as a constant 3-tensor yields

⟨T(X⊗X),Y⟩=⟨D2P(Y)X,X⟩=D3P(0)[Y,X,X]=⟨D2P(X)X,Y⟩,(27)\langle\mathcal{T}(X \otimes X),Y\rangle= \langle D^2P(Y)X,X\rangle= D^3P(0)[Y,X,X] = \langle D^2P(X)X,Y\rangle, \tag*{(27)}

and thus, combining with the homogeneity (3.1) of D2PD^2P we have

T(X⊗X)=2DP(X).(28)\mathcal{T}(X \otimes X) = 2DP(X). \tag*{(28)}

Finally, for X∈S1(3)X \in\mathcal{S}_1(3), (3.5) additionally yields

T(DP(X)⊗X+X⊗DP(X))=36X.(29)\mathcal{T}\bigl(DP(X) \otimes X + X \otimes DP(X)\bigr) = 36X. \tag*{(29)}

Combining (4.2), (4.4) and (4.5), we therefore obtain

T(D2u(X))=90X+6P(X)DP(X).(30)\mathcal{T}(D^2u(X)) = 90X + 6P(X)DP(X). \tag*{(30)}

Let us now use the above decomposition into SO(3)SO(3)-irreducible invariant subspaces in order to find a pair of null Lagrangians whose degeneracies compensate one another. We begin by identifying two quadratic null Lagrangians arising naturally from the computations in the preceding section. Consider two quadratic forms, each of which depends only on two components MiM_i in a decomposition (M1,M2,M3)∈D2P(Sym⁡0(3))⊕V0⊕⟨Id⁡Sym⁡0(3)⟩(M_1,M_2,M_3) \in D^2P(\operatorname{Sym}_0(3)) \oplus V_0 \oplus\langle\operatorname{Id}_{\operatorname{Sym}_0(3)}\rangle, for instance

L1(M)=a1∣M1∣2+a3∣M3∣2,L_1(M) = a_1|M_1|^2 + a_3|M_3|^2,
L0(M)=b2∣M2∣2+b3∣M3∣2.L_0(M) = b_2|M_2|^2 + b_3|M_3|^2.

These are convenient ansätze since we avoid having to work with linear combinations of all 3 of the MiM_i, while retaining enough flexibility to solve for the coefficients in order to satisfy the null Lagrangian condition (4.1). However, many other choices would also work, subject to satisfying the latter constraint. Moreover, note that we only expect to find two linearly independent quadratic null Lagrangians with SO(3)SO(3)-invariance, since there are three mutually orthogonal subspaces D2P(Sym⁡0(3))D^2P(\operatorname{Sym}_0(3)), V0V_0, ⟨Id⁡Sym⁡0(3)⟩\langle\operatorname{Id}_{\operatorname{Sym}_0(3)}\rangle from which to take linear combinations, but the null Lagrangian constraint removes one degree of freedom.

First of all, letting X∈Sym⁡0(3)X \in\operatorname{Sym}_0(3) be such that M1=D2P(X)M_1 = D^2P(X) and noting that T≡0\mathcal{T} \equiv0 on V0⊕⟨Id⁡Sym⁡0(3)⟩V_0 \oplus\langle\operatorname{Id}_{\operatorname{Sym}_0(3)} \rangle, (26) yields 126X=T(M)126X = \mathcal{T}(M), and thus thanks to (17) we obtain

∣M1∣2=∣D2P(X)∣2=126∣X∣2=1126∣T(M)∣2.|M_1|^2 = |D^2P(X)|^2 = 126|X|^2 = \frac{1}{126}|\mathcal{T}(M)|^2.

Secondly, note that (15) guarantees tr⁡M1=0\operatorname{tr} M_1 = 0, and we have tr⁡M2=0\operatorname{tr} M_2 = 0 by construction. Thus, we must have tr⁡M3=tr⁡M\operatorname{tr} M_3 = \operatorname{tr} M, and since dim⁡Sym⁡0(3)=5\dim\operatorname{Sym}_0(3) = 5, we obtain

∣M3∣2=15(tr⁡M)2.|M_3|^2 = \frac{1}{5}(\operatorname{tr} M)^2.

Orthogonality then allows us to recover ∣M2∣2|M_2|^2 via

∣M2∣2=∣M∣2−∣M1∣2−∣M3∣2.|M_2|^2 = |M|^2 - |M_1|^2 - |M_3|^2.

Taking M=ξ⊗ξM = \xi\otimes\xi and combining the above properties with (28), (15) and the null Lagrangian constraint (25) which requires L0L_0 and L1L_1 to vanish on rank-one tensors, we obtain the relations

27a1+15a3=0,(1−27−15)b2+15b3=0.\frac{2}{7}a_1 + \frac{1}{5}a_3 = 0, \qquad\left(1-\frac{2}{7}-\frac{1}{5}\right)b_2 + \frac{1}{5}b_3 = 0.

Taking a1=1a_1 = 1 and b2=−1b_2 = -1 yields

L1(M)=1126∣T(M)∣2−27(tr⁡M)2,L_1(M) = \frac{1}{126}|\mathcal{T}(M)|^2 - \frac{2}{7}(\operatorname{tr} M)^2,
L0(M)=1126∣T(M)∣2+57(tr⁡M)2−∣M∣2.L_0(M) = \frac{1}{126}|\mathcal{T}(M)|^2 + \frac{5}{7}(\operatorname{tr} M)^2 - |M|^2.

Again, this is simply a choice that we fix, for concreteness; any other one would work equally well.

By construction, we have therefore obtained the following.

Proposition 4.1. Both L0L_0 and L1L_1 are quadratic null Lagrangians for the Hessian. They are both invariant under the action M↦ρRMρR∗M \mapsto\rho_R M \rho_R^* induced by R∈SO(3)R \in SO(3).

We will see that although L1L_1 is strictly positive on every direction M−NM-N between distinct M,NM,N in the image KK of S1(3)\mathcal{S}_1(3) under D2uD^2u, this positivity is not uniform and indeed degenerates along certain infinitesimal directions as MM and NN approach one another. However, we will verify that L0L_0 remains uniformly non-degenerate on these directions. This complementary behavior is what allows us to find a linear combination of L0L_0 and L1L_1 that is uniformly non-degenerate on KK, as stated in the following key proposition; this is what we will need in order to extend this linearization to our nonlinear uniformly convex energy.

Proposition 4.2. There are constants b≥0b \ge0 and σ>0\sigma> 0 such that the quadratic null Lagrangian L=L0+bL1L=L_0+bL_1 satisfies

L(M−N)≥σ∣M−N∣2for M,N∈K,(31)L(M-N) \ge\sigma|M-N|^2 \qquad\text{for } M,N\in K, \tag*{(31)}

where KK is as in (22).

The remainder of this section will be dedicated to the proof of Proposition 4.2, which we subdivide into a number of intermediate lemmas. In the first of these, we show strict positivity of L1L_1 on KK.

Lemma 4.3. For distinct X,Y∈S1(3)X,Y\in\mathcal{S}_1(3),

L1(D2u(X)−D2u(Y))=1126∣T(D2u(X))−T(D2u(Y))∣2−82⋅27∣P(X)−P(Y)∣2>0.L_1(D^2u(X)-D^2u(Y))=\frac{1}{126}\left|\mathcal{T}(D^2u(X))-\mathcal{T}(D^2u(Y))\right|^2-8^2\cdot\frac{2}{7}|P(X)-P(Y)|^2>0.

Proof. Recalling (30) and (13), for X∈S1(3)X\in\mathcal{S}_1(3) we have

T(D2u(X))=90X+186P(X)(X2−13Id⁡3).(32)\mathcal{T}(D^2u(X))=90X+18\sqrt{6}P(X)\left(X^2-\frac{1}{3}\operatorname{Id}_3\right). \tag*{(32)}

Thus, if we choose an orthonormal basis {ei}i=13\{e_i\}_{i=1}^3 of eigenvectors for XX, i.e. Xei=λX,ieiXe_i=\lambda_{X,i}e_i, we in turn have X2ei=λX,i2eiX^2e_i=\lambda_{X,i}^2e_i and so {ei}\{e_i\} is also an eigenbasis for T(D2u(X))\mathcal{T}(D^2u(X)). Letting {μX,i}\{\mu_{X,i}\} denote the associated eigenvalues for T(D2u(X))\mathcal{T}(D^2u(X)) observe that the Rayleigh quotient characterization of the eigenvalues yields

∣μX,i−μY,i∣≤∥T(D2u(X))−T(D2u(Y))∥op.(33)|\mu_{X,i}-\mu_{Y,i}|\leq\left\|\mathcal{T}(D^2u(X))-\mathcal{T}(D^2u(Y))\right\|_{\mathrm{op}}. \tag*{(33)}

We would like to find a lower bound on ∣μX,i−μY,i∣|\mu_{X,i}-\mu_{Y,i}| in terms of XX directly, rather than via the adjoint T\mathcal{T}. List the eigenvalues {λX,i}i\{\lambda_{X,i}\}_i monotonically increasing in size, i.e. λX,1≤λX,2≤λX,3\lambda_{X,1}\leq\lambda_{X,2}\leq\lambda_{X,3}. We will proceed to express everything in terms of the middle one λX,2\lambda_{X,2}, though it could be done with any fixed one of them; the reason for this choice will become apparent later. Arrange X,Y∈S1(3)X,Y\in\mathcal{S}_1(3) so that λX,2≥λY,2\lambda_{X,2}\geq\lambda_{Y,2} without loss of generality. Now, recall that

P(X)=36det⁡X=36∏i=13λX,i.P(X)=3\sqrt{6}\det X=3\sqrt{6}\prod_{i=1}^{3}\lambda_{X,i}.

Since tr⁡X=∑i=13λX,i=0\operatorname{tr}X=\sum_{i=1}^{3}\lambda_{X,i}=0 and ∣X∣2=∑i=13λX,i2=1|X|^2=\sum_{i=1}^{3}\lambda_{X,i}^2=1, we obtain λX,1λX,3=λX,22−12\lambda_{X,1}\lambda_{X,3}=\lambda_{X,2}^2-\frac{1}{2}, and therefore

P(X)=36λX,2(λX,22−12)=:p(λX,2).P(X)=3\sqrt{6}\lambda_{X,2}\left(\lambda_{X,2}^2-\frac{1}{2}\right)=:p(\lambda_{X,2}).

Inserting this into the identity (32) and taking the e2e_2 component of the latter yields

∣μX,2−μY,2∣=∣90(λX,2−λY,2)+186[p(λX,2)(λX,22−13)−p(λY,2)(λY,22−13)]∣=∣90(λX,2−λY,2)+186∫λY,2λX,2dds[p(s)(s2−13)] ds∣.\begin{aligned} |\mu_{X,2}-\mu_{Y,2}|&=\left|90(\lambda_{X,2}-\lambda_{Y,2})+18\sqrt{6}\left[p(\lambda_{X,2})\left(\lambda_{X,2}^2-\frac{1}{3}\right)-p(\lambda_{Y,2})\left(\lambda_{Y,2}^2-\frac{1}{3}\right)\right]\right|\\ &=\left|90(\lambda_{X,2}-\lambda_{Y,2})+18\sqrt{6}\int_{\lambda_{Y,2}}^{\lambda_{X,2}}\frac{d}{ds}\left[p(s)\left(s^2-\frac{1}{3}\right)\right]\,ds\right|. \end{aligned}

Let us now compute the derivative inside the integral. Since p(s)=36s(s2−12)p(s)=3\sqrt{6}s\left(s^2-\frac{1}{2}\right), we have

186dds[p(s)(s2−13)]=182(5s4−52s2+16)=dds[−186⋅89p(s)−90s]+54s2(1+30s2).18\sqrt{6}\frac{d}{ds}\left[p(s)\left(s^2-\frac{1}{3}\right)\right]=18^2\left(5s^4-\frac{5}{2}s^2+\frac{1}{6}\right) =\frac{d}{ds}\left[-18\sqrt{6}\cdot\frac{8}{9}p(s)-90s\right]+54s^2\left(1+30s^2\right).

and thus

∣μX,2−μY,2∣=∣−166[p(λX,2)−p(λY,2)]+54∫λY,2λX,2s2(1+30s2) ds∣.(34)\left|\mu_{X,2}-\mu_{Y,2}\right|=\left|-16\sqrt{6}\left[p(\lambda_{X,2})-p(\lambda_{Y,2})\right]+54\int_{\lambda_{Y,2}}^{\lambda_{X,2}}s^2\left(1+30s^2\right)\,ds\right|. \tag*{(34)}

Now our choice of the middle eigenvalue λX,2\lambda_{X,2} comes into play. We claim that for this one, p′(s)≤0p'(s)\leq0 on the segment [λY,2,λX,2][\lambda_{Y,2},\lambda_{X,2}] and so in particular −[p(λX,2)−p(λY,2)]≥0-[p(\lambda_{X,2})-p(\lambda_{Y,2})]\geq0. Indeed, notice that p′(s)=36(3s2−12)p'(s)=3\sqrt{6}\left(3s^2-\frac{1}{2}\right), so

p′(λX,2)=36(3λX,22−12)=36(λX,2−λX,1)(λX,2−λX,3)≤0,p'(\lambda_{X,2})=3\sqrt{6}\left(3\lambda_{X,2}^2-\frac{1}{2}\right)=3\sqrt{6}(\lambda_{X,2}-\lambda_{X,1})(\lambda_{X,2}-\lambda_{X,3})\leq0,

where the second identity is again exploiting the relations coming from tr⁡X=0\operatorname{tr}X=0 and ∣X∣2=1|X|^2=1. This in turn implies that λX,2∈[−16,16]\lambda_{X,2}\in\left[-\frac{1}{\sqrt{6}},\frac{1}{\sqrt{6}}\right], and so the claim follows.

Inserting this into (4.10) and combining with (4.9), we deduce that

∣T(D2u(X))−T(D2u(Y))∣2≥32∥T(D2u(X))−T(D2u(Y))∥op2|\mathcal{T}(D^2u(X))-\mathcal{T}(D^2u(Y))|^2\geq\frac{3}{2}\|\mathcal{T}(D^2u(X))-\mathcal{T}(D^2u(Y))\|_{\mathrm{op}}^2
≥9⋅162(P(X)−P(Y))2.\geq9\cdot16^2(P(X)-P(Y))^2.

Combining this with the Laplacian identity in Lemma 3.2, the conclusion follows.

Note that the positivity in Lemma 4.3 is not uniform, and might degenerate as X,YX,Y approach one another. Let us therefore identify these infinitesimal directions on which L1L_1 degenerates and show that on these, L0L_0 is bounded uniformly below.

Lemma 4.4. For X∈S1(3)X\in\mathcal{S}_1(3) and v∈TXS1(3)v\in T_X\mathcal{S}_1(3),

L1(D(D2u)(X)[v])=L1((D3u)(X)v)≥0,L_1(D(D^2u)(X)[v])=L_1((D^3u)(X)v)\geq0,

If v≠0v\ne0 and equality holds, then

L0(D3u(X)v)∣D3u(X)v∣2=c0>0,\frac{L_0(D^3u(X)v)}{|D^3u(X)v|^2}=c_0>0,

where c0c_0 is a universal constant depending on PP. Here, (D3u)(X)v=(D3u)(X)[v,⋅,⋅](D^3u)(X)v=(D^3u)(X)[v,\cdot,\cdot] is an element of S(3)\mathcal{S}(3).

Proof. Fix X∈S1(3)X \in\mathcal{S}_1(3) and v∈TXS1(3)v \in T_X\mathcal{S}_1(3). We will proceed to directly compute the behavior of L1L_1 on D3u(X)vD^3u(X)v. First of all, recall from definition that we have L1(D3u(X)v)=1126∣T(D3u(X)v)∣2−27(tr⁡[D3u(X)v])2L_1(D^3u(X)v)=\frac{1}{126}\lvert\mathcal{T}(D^3u(X)v)\rvert^2-\frac{2}{7}\left(\operatorname{tr}[D^3u(X)v]\right)^2. Using linearity of the trace operator and differentiating the Laplacian identity in Lemma 3.2 yields

tr⁡[D3u(X)v]=−8DP(X)v,(35)\operatorname{tr}[D^3u(X)v]=-8DP(X)v, \tag*{(35)}

and thus

L1(D3u(X)v)=1126∣T(D3u(X)v)∣2−1287(DP(X)v)2.(36)L_1(D^3u(X)v)=\frac{1}{126}\lvert\mathcal{T}(D^3u(X)v)\rvert^2-\frac{128}{7}\left(DP(X)v\right)^2. \tag*{(36)}

Noting that T\mathcal{T} is a linear operator, we may simply differentiate the identity (4.6) in order to analyze the behavior of T(D3u(X)v)\mathcal{T}(D^3u(X)v). This yields

T(D3u(X)v)=90v+6(DP(X)v)DP(X)+6P(X)D2P(X)v.(37)\mathcal{T}(D^3u(X)v)=90v+6(DP(X)v)DP(X)+6P(X)D^2P(X)v. \tag*{(37)}

Keeping in mind that we wish to realize (4.12) as a non-negative expression, let us try to complete the square. Combining (4.13) with Lemma 3.1 and the observation that ⟨X,v⟩=0\langle X,v\rangle=0, we obtain

⟨T(D3u(X)v),DP(X)⟩=90DP(X)v+6∣DP(X)∣2DP(X)v+6P(X)D2P(X)DP(X)v=90DP(X)v+54DP(X)v+108P(X)⟨X,v⟩=144DP(X)v.\begin{aligned} \langle\mathcal{T}(D^3u(X)v),DP(X)\rangle &=90DP(X)v+6\lvert DP(X)\rvert^2DP(X)v+6P(X)D^2P(X)DP(X)v\\ &=90DP(X)v+54DP(X)v+108P(X)\langle X,v\rangle\\ &=144DP(X)v. \end{aligned}

Thus,

L1((D3u)(X)v)=1126∣T(D3u(X)v)−16(DP(X)v)DP(X)∣2=1126∣90v+6P(X)D2P(X)v−10(DP(X)v)DP(X)∣2≥0\begin{aligned} L_1((D^3u)(X)v) &=\frac{1}{126}\left\lvert\mathcal{T}(D^3u(X)v)-16(DP(X)v)DP(X)\right\rvert^2\\ &=\frac{1}{126}\left\lvert90v+6P(X)D^2P(X)v-10(DP(X)v)DP(X)\right\rvert^2\\ &\ge0 \end{aligned}

as desired.

Let us now investigate when equality holds. Taking an inner product of the expression inside the square with XX and again invoking Lemma 3.1 and ⟨X,v⟩=0\langle X,v\rangle=0, a necessary condition for equality to hold is P(X)DP(X)v=0P(X)DP(X)v=0. Therefore, either P(X)=0P(X)=0 or DP(X)v=0DP(X)v=0. In the latter case, inserting this condition back into the squared expression above yields ∣90v+6P(X)D2P(X)v∣2=0\lvert90v+6P(X)D^2P(X)v\rvert^2=0, which in turn forces v=0v=0 thanks to the bounds on D2P(X)D^2P(X) and P(X)P(X) from Lemma 3.1.

It therefore remains to investigate the behavior of L0(D3u(X)v)∣D3u(X)v∣2\frac{L_0(D^3u(X)v)}{\lvert D^3u(X)v\rvert^2} when P(X)=0P(X)=0 and v≠0v\ne0. In this case, the expression in the square tells us that v=aDP(X)v=aDP(X) for some a≠0a\ne0. Without loss of generality, we may take a=1/3a=1/3 so that ∣v∣=1\lvert v\rvert=1. Once again using Lemma 3.1, this choice further yields DP(X)v=3DP(X)v=3 and D2P(X)v=6XD^2P(X)v=6X. Now, since

L0(D3u(X)v)=L1(D3u(X)v)+(tr⁡[D3u(X)v])2−∣D3u(X)v∣2(38)L_0(D^3u(X)v)=L_1(D^3u(X)v)+\left(\operatorname{tr}[D^3u(X)v]\right)^2-\lvert D^3u(X)v\rvert^2 \tag*{(38)}

and we know L1(D3u(X)v)=0L_1(D^3u(X)v)=0 and have already computed (tr⁡[D3u(X)v])2(\operatorname{tr}[D^3u(X)v])^2 in (35), the only thing left is to compute ∣D3u(X)v∣2|D^3u(X)v|^2.

Differentiating (18) in the direction vv and substituting in these identities together with P(X)=0P(X)=0 yields

D3u(X)v=D3P(X)v−3X⊗X−23DP(X)⊗DP(X)−3Id⁡Sym⁡0(3),D^3u(X)v=D^3P(X)v-3X\otimes X-\frac{2}{3}DP(X)\otimes DP(X)-3\operatorname{Id}_{\operatorname{Sym}_0(3)},

and thus, combining with the fact that D3P(X)v=D2P(v)D^3P(X)v=D^2P(v) since PP is a cubic polynomial, we obtain

∣D3u(X)v∣2=∣D2P(v)∣2+9∣X∣4+49∣DP(X)∣4+9∣Id⁡Sym⁡0(3)∣2−6⟨D3P(X)v,X⊗X⟩−43⟨D3P(X)v,DP(X)⊗DP(X)⟩−6ΔP(v)+4⟨X⊗X,DP(X)⊗DP(X)⟩+18∣X∣2+4∣DP(X)∣2.\begin{aligned} |D^3u(X)v|^2={}&|D^2P(v)|^2+9|X|^4+\frac{4}{9}|DP(X)|^4+9|\operatorname{Id}_{\operatorname{Sym}_0(3)}|^2\\ &-6\langle D^3P(X)v,X\otimes X\rangle-\frac{4}{3}\langle D^3P(X)v,DP(X)\otimes DP(X)\rangle-6\Delta P(v)\\ &+4\langle X\otimes X,DP(X)\otimes DP(X)\rangle+18|X|^2+4|DP(X)|^2. \end{aligned}

Now, recalling the tensor identity ⟨a⊗b,c⊗d⟩=(a⋅c)(b⋅d)\langle a\otimes b,c\otimes d\rangle=(a\cdot c)(b\cdot d) and differentiating (16) in the direction vv, we obtain

⟨D3P(X)v,DP(X)⊗DP(X)⟩+⟨D2P(X)(D2P(X)v),DP(X)⟩=36(X⋅v)(DP(X)⋅X)+18∣X∣2DP(X)v.\begin{aligned} \langle D^3P(X)v,DP(X)\otimes DP(X)\rangle+\langle D^2P(X)(D^2P(X)v),DP(X)\rangle\\ =36(X\cdot v)(DP(X)\cdot X)+18|X|^2DP(X)v. \end{aligned}

Combining this with the symmetry of D3P(X)D^3P(X), the properties in Lemma 3.1, the homogeneity relations (12), the fact that dim⁡Sym⁡0(3)=5\dim\operatorname{Sym}_0(3)=5, and ∣X∣=∣v∣=1|X|=|v|=1, we arrive at

∣D3u(X)v∣2=306.|D^3u(X)v|^2=306.

Combining this with (38) and (35), we conclude that

L0(D3u(X)v)∣D3u(X)v∣2=270306=1517=:c0>0.\frac{L_0(D^3u(X)v)}{|D^3u(X)v|^2}=\frac{270}{306}=\frac{15}{17}=:c_0>0.

With the competing non-degeneracies at hand, we are now in a position to conclude the proof of Proposition 4.2.

Proof of Proposition 4.2. Our goal is to find a suitable linear combination L=L0+bL1L=L_0+bL_1 satisfying (31). Recall from Lemma 4.3 that L1(M−N)≥0L_1(M-N)\ge0 for every M,N∈KM,N\in K; it remains to compensate this with a piece of L0L_0 to obtain uniform nondegeneracy. Fix σ∈(0,c0)\sigma\in(0,c_0), where c0c_0 is as in Lemma 4.4, and set

U={(M,N)∈K×K:L0(M−N)≥σ∣M−N∣2}.U=\{(M,N)\in K\times K:L_0(M-N)\ge\sigma|M-N|^2\}.

By definition, on UU the desired lower bound holds for any choice of b≥0b \ge0, and so it suffices to restrict our attention to (K×K)∖U(K \times K) \setminus U. We claim that

η:=inf⁡(M,N)∈(K×K)∖UM≠NL1(M−N)∣M−N∣2>0.(39)\eta:= \inf_{\substack{(M,N)\in(K\times K)\setminus U\\M\ne N}} \frac{L_1(M-N)}{\lvert M-N\rvert^2} > 0. \tag*{(39)}

Suppose for a contradiction that η=0\eta= 0. Then there exists a sequence (Mj,Nj)∈K×K(M_j,N_j) \in K \times K for which the ratios in (39) converge to zero. Up to extracting a subsequence, we must therefore have Mj,Nj→M∈KM_j, N_j \to M \in K. By Lemma 3.3, the corresponding points Xj,Yj∈S1(3)X_j,Y_j \in\mathcal{S}_1(3) that respectively satisfy D2u(Xj)=MjD^2u(X_j)=M_j, D2u(Yj)=NjD^2u(Y_j)=N_j also converge (subsequentially) to the same element X∈S1(3)X \in\mathcal{S}_1(3). This induces a tangent direction v∈TXS1(3)v \in T_X\mathcal{S}_1(3) as follows:

Xj−Yj∣Xj−Yj∣⟶v,∣v∣=1,v∈TXS1(3),\frac{X_j-Y_j}{\lvert X_j-Y_j\rvert} \longrightarrow v,\qquad\lvert v\rvert=1,\qquad v\in T_X\mathcal{S}_1(3),

up to yet another subsequence. Taylor expanding between YjY_j and XjX_j yields

Mj−Nj∣Xj−Yj∣=D2u(Xj)−D2u(Yj)∣Xj−Yj∣⟶D3u(X)v.\frac{M_j-N_j}{\lvert X_j-Y_j\rvert} = \frac{D^2u(X_j)-D^2u(Y_j)}{\lvert X_j-Y_j\rvert} \longrightarrow D^3u(X)v.

Again invoking Lemma 3.3, we observe that the right-hand side is non-zero. However, by construction of our contradiction sequences, we have

L1(D3u(X)v)=0,L0(D3u(X)v)≤σ∣D3u(X)v∣2,L_1(D^3u(X)v)=0,\qquad L_0(D^3u(X)v)\le\sigma\lvert D^3u(X)v\rvert^2,

which contradicts Lemma 4.4, since σ<c0\sigma<c_0. Thus we indeed must have η>0\eta>0, as claimed.

Next, noticing that L0(M−N)≥−∣M−N∣2L_0(M-N)\ge-\lvert M-N\rvert^2 by definition, we obtain

(L0+bL1)(M−N)≥(−1+bη)∣M−N∣2(L_0+bL_1)(M-N)\ge(-1+b\eta)\lvert M-N\rvert^2

for (M,N)∈(K×K)∖U(M,N)\in(K\times K)\setminus U with M≠NM\ne N. Thus, letting b≥(1+σ)/ηb\ge(1+\sigma)/\eta, the desired nondegeneracy (31) follows for this choice of bb.

Proof of main results

The final step required to conclude our main results is an extension of LL from KK to a uniformly convex integrand FF whose linearization is LL on KK. This will be a consequence of the following lemma, which is a special case of [36 Lemma 2.1] where the function GG therein is taken to be quadratic. We refer the reader to the proof therein, and merely clarify why the extension constructed there also yields a uniformly bounded Hessian for FF, which is what we are assuming herein.

Lemma 5.1. Let KK be a compact smooth embedded submanifold of Rq\mathbb{R}^q. Suppose that L:Rq→RL:\mathbb{R}^q\to\mathbb{R} is a quadratic form satisfying

L(M−N)≥σ∣M−N∣2M,N∈K,L(M-N)\geq\sigma\lvert M-N\rvert^2 \qquad M,N\in K,

for some σ>0\sigma>0. Then there exists F∈C∞(Rq)F\in C^\infty(\mathbb{R}^q) satisfying the bounds (1.2) with λ=σ\lambda=\sigma and Λ=Λ(q,K,L,σ)<∞\Lambda=\Lambda(q,K,L,\sigma)<\infty, with the property that on KK both F=LF=L and DF=DLDF=DL hold.

Proof. If KK is nonempty, we may apply [36 Lemma 2.1] with G=LG=L, v=DLv=DL, and γ=σ\gamma=\sigma in the notation therein. This in addition yields the Hessian lower bound. Meanwhile, to obtain the uniform upper bound on the Hessian, we consider the function H0H_0 from Step 2 of [36 Section 4.1], given by

H0(M)=sup⁡N∈Kδ{F(N)+⟨DF(N),M−N⟩+34σ∣M−N∣2},H_0(M)=\sup_{N\in K_\delta}\left\{F(N)+\langle DF(N),M-N\rangle+\frac{3}{4}\sigma\lvert M-N\rvert^2\right\},

where KδK_\delta is a δ\delta-neighborhood of KK, for δ>0\delta>0 sufficiently small, and FF is the local extension obtained in [36 Proof of Lemma 2.1, Step 1]. Expanding the square in the right-hand side above, this may be rewritten as

H0(M)−34σ∣M∣2=sup⁡N∈Kδ{F(N)−⟨DF(N),N⟩+⟨DF(N)−32σN,M⟩+34σ∣N∣2}.H_0(M)-\frac{3}{4}\sigma\lvert M\rvert^2=\sup_{N\in K_\delta}\left\{F(N)-\langle DF(N),N\rangle+\left\langle DF(N)-\frac{3}{2}\sigma N,M\right\rangle+\frac{3}{4}\sigma\lvert N\rvert^2\right\}.

Noting that DF(N)−32σNDF(N)-\frac{3}{2}\sigma N is uniformly bounded over N∈KδN\in K_\delta, we deduce that H0−34σ∣⋅∣2H_0-\frac{3}{4}\sigma\lvert\cdot\rvert^2 is globally Lipschitz, with Lipschitz constant at most sup⁡N∈Kδ∣DF(N)−32σN∣\sup_{N\in K_\delta}\lvert DF(N)-\frac{3}{2}\sigma N\rvert.

Now, for fixed ε>0\varepsilon>0 suitably small, let Hε:=H0∗ρεH_\varepsilon:=H_0*\rho_\varepsilon for a standard mollifier ρε\rho_\varepsilon be as in [36 Proof of Lemma 2.1, Step 2]. Note that we may ensure ∥Dρε∥L1≤Cε−1\lVert D\rho_\varepsilon\rVert_{L^1}\leq C\varepsilon^{-1}. Taking the Hessian of both sides above, observe that this mollification satisfies the uniform Hessian bounds

sup⁡M∈Rq∥D2Hε(M)∥op≤32σ+Cε−1Lip⁡(H0−34σ∣M∣2)<∞.\sup_{M\in\mathbb{R}^q}\lVert D^2H_\varepsilon(M)\rVert_{\mathrm{op}}\leq\frac{3}{2}\sigma+C\varepsilon^{-1}\operatorname{Lip}\left(H_0-\frac{3}{4}\sigma\lvert M\rvert^2\right)<\infty.

The final extension is then an interpolation between H0H_0 on KδK_\delta and HεH_\varepsilon on Rq∖K2δ\mathbb{R}^q\setminus K_{2\delta}, so the above estimates indeed guarantee a uniform upper Hessian bound. □

Let us now collect everything together to conclude our main results.

Proof of Theorem 1.1. Let LL be the null Lagrangian from Proposition 4.2. Since Lemma 3.3 in addition tells us that the set KK is a smooth, compact embedded submanifold of S(3)≅Sym⁡(5)\mathcal{S}(3)\cong\operatorname{Sym}(5), we may apply Lemma 5.1 to obtain a smooth extension FF that satisfies (1.2) and F=LF=L, DF=DLDF=DL on KK. In particular, by construction we have

∫B1DF(D2u):D2φ dx=∫B1DL(D2u):D2φ dx=0for every φ∈Cc∞(B1).(40)\int_{B_1} DF(D^2u) : D^2\varphi\,dx = \int_{B_1} DL(D^2u) : D^2\varphi\,dx = 0 \qquad\text{for every } \varphi\in C^\infty_c(B_1). \tag*{(40)}

Since DF(D2u)DF(D^2u) is uniformly bounded, (5.1) remains valid for φ∈W02,2(B1)\varphi\in W^{2,2}_0(B_1) by density. Using Taylor’s Theorem with integral remainder together with the uniform convexity of FF (i.e. the lower bound in (1.2)), we obtain

F(D2u+D2φ)−F(D2u)=DF(D2u):D2φ+∫01(1−t)D2F(D2u+tD2φ)[D2φ,D2φ] dt≥DF(D2u):D2φ+λ2∣D2φ∣2,\begin{aligned} F(D^2u+D^2\varphi)-F(D^2u) &=DF(D^2u):D^2\varphi+\int_0^1(1-t)D^2F(D^2u+tD^2\varphi)[D^2\varphi,D^2\varphi]\,dt\\ &\ge DF(D^2u):D^2\varphi+\frac{\lambda}{2}|D^2\varphi|^2, \end{aligned}

from which (1.9) follows immediately after integration over B1B_1. The uniqueness of uu as a minimizer follows immediately from the uniform convexity (1.9). Note that the stated boundary data for uu is due to its two-homogeneity. □

Proof of Corollary 1.2. This follows immediately by splitting the variables in Rn\mathbb{R}^n as x=(x′,y)∈R5×Rn−5x=(x',y)\in\mathbb{R}^5\times\mathbb{R}^{n-5} and considering the candidate minimizer x↦u(x′)x\mapsto u(x') as a function on Rn\mathbb{R}^n. This function, which we still call uu at the risk of abusing notation, clearly satisfies (1.10) for its singular set, by construction. The associated uniformly convex integrand can be taken to be

Sym⁡(n)∋M↦F(A)+12(∣M∣2−∣A∣2),\operatorname{Sym}(n)\ni M\mapsto F(A)+\frac{1}{2}(|M|^2-|A|^2),

where FF is the integrand on Sym⁡(5)\operatorname{Sym}(5) from Theorem 1.1 and AA is the upper left 5×55\times5 block of MM. It is immediate to see that this satisfies (1.2) with constants min⁡{λ,1}\min\{\lambda,1\} and max⁡{Λ,1}\max\{\Lambda,1\}, where λ,Λ\lambda,\Lambda are the constants associated to FF. It is easy to verify that this integrand satisfies the desired properties; we leave the details to the reader. □

References

Department of Mathematics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, USA

Email address: [email protected]

Institute for Theoretical Studies, ETH Zürich, Zürich, Switzerland

Email address: [email protected]

References

  1. [1]Patricio Aviles and Yoshikazu Giga, A mathematical problem related to the physical theory of liquid crystal configurations, Miniconference on Geometry/Partial Differential Equations, 2 (Canberra) (John E. Hutchinson and Leon M. Simon, eds.), Proceedings of the Centre for Mathematical Analysis, vol. 12, Australian National University, 1987, pp. 1–16.
  2. [2]Pierre Bousquet and Lorenzo Brasco, C^1 regularity of orthotropic p-harmonic functions in the plane, Anal. PDE 11 (2018), no. 4, 813–854.
  3. [3]John M. Ball, J. C. Currie, and Peter J. Olver, Null Lagrangians, weak continuity, and variational problems of arbitrary order, J. Funct. Anal. 41 (1981), no. 2, 135–174.DOI
  4. [4]Arunima Bhattacharya, Jingyi Chen, and Micah Warren, Regularity of Hamiltonian stationary equations in symplectic manifolds, Adv. Math. 424 (2023), Paper No. 109059.arxiv.org/abs/2108.00325
  5. [5]Arunima Bhattacharya, Regularity for critical points of convex functionals on Hessian spaces, Proc. Amer. Math. Soc. 150 (2022), no. 12, 5217–5230.arxiv.org/abs/2108.01319
  6. [7]Arunima Bhattacharya and W. Jacob Ogden, Optimal regularity for Hölder continuous Hamiltonian stationary Lagrangian graphs, 2025, Preprint, arXiv:2510.22756.arxiv.org/abs/2510.22756
  7. [8]Arunima Bhattacharya, Gerard Orriols, and Anna Skorobogatova, Sharp partial regularity of Hamiltonian stationary and special Lagrangian graphs, 2026, Preprint, arXiv:2608.27324.arxiv.org/abs/2608.27324
  8. [9]Arunima Bhattacharya and Anna Skorobogatova, Variational integrals on Hessian spaces: partial regularity for critical points, Nonlinear Anal. 255 (2025), Paper No. 113760.
  9. [10]Arunima Bhattacharya and Micah Warren, Interior Schauder estimates for the fourth order Hamiltonian stationary equation in two dimensions, Proceedings of the American Mathematical Society 147 (2019), no. 8, 3471–3477.DOI
  10. [11]——— , Regularity bootstrapping for 4th-order nonlinear elliptic equations, Int. Math. Res. Not. IMRN 2021 (2021), no. 6, 4324–4348.
  11. [12]Élie Cartan, Sur des familles remarquables d’hypersurfaces isoparamétriques dans les espaces sphériques, Math. Z. 45 (1939), 335–367.DOI
  12. [13]Matteo Carducci and Bozhidar Velichkov, A minimizing cone from Cartan polynomials in the Alt–Caffarelli problem, Preprint, arXiv:2609.35179.arxiv.org/abs/2609.35179
  13. [14]Jingyi Chen and Micah Warren, On the regularity of Hamiltonian stationary Lagrangian submanifolds, Adv. Math. 343 (2019), 316–352.DOI
  14. [15]Cristiana De Filippis, Jan Kristensen, Giuseppe Mingione, and Aidan Strong, The singular set of minima of multiple integrals, Preprint, arXiv:2609.11546.arxiv.org/abs/2609.11546
  15. [16]Ennio De Giorgi, Sulla differenziabilità e l’analiticità delle estremali degli integrali multipli regolari, Mem. Accad. Sci. Torino Cl. Sci. Fis. Mat. Nat. (3) 3 (1957), 25–43.
  16. [17]Bryan Dimler, Partial regularity for lipschitz solutions to the minimal surface system, Calc. Var. Partial Differential Equations 62 (2023), no. 9, 260.arxiv.org/abs/2306.12586
  17. [19]J. P. Duggan, Regularity theorems for varifolds with mean curvature, Indiana University Mathematics Journal 35 (1986), no. 1, 117–144.
  18. [20]Lawrence C. Evans, Quasiconvexity and partial regularity in the calculus of variations, Arch. Rational Mech. Anal. 95 (1986), no. 3, 227–252.
  19. [21]Gero Friesecke, Richard D. James, and Stefan Müller, A hierarchy of plate models derived from nonlinear elasticity by Gamma-convergence, Archive for Rational Mechanics and Analysis 180 (2006), no. 2, 183–236.DOI
  20. [22]Benjy Firester and Raphael Tsiamis, Topology of minimal surfaces in the sphere from capillarity, 2026, Preprint, arXiv:2604.04928.arxiv.org/abs/2604.04928
  21. [24]Mariano Giaquinta and Enrico Giusti, Nonlinear elliptic systems with quadratic growth, Manuscripta Mathematica 24 (1978), 323–349.DOI
  22. [25]Enrico Giusti, Precisazione delle funzioni di H^{1,p} e singolarità delle soluzioni deboli di sistemi ellittici non lineari, Boll. Un. Mat. Ital. (4) 2 (1969), 71–76. MR 0243190
  23. [26]Enrico Giusti and Mario Miranda, Sulla regolarità delle soluzioni deboli di una classe di sistemi ellittici quasi-lineari, Arch. Rational Mech. Anal. 31 (1968), no. 3, 173–184.DOI
  24. [27]André Guerra and Riccardo Tione, A singular minimizer of a smooth strongly polyconvex functional in two dimensions, 2026, Preprint, arXiv:2609.32404.arxiv.org/abs/2609.32404
  25. [29]Wenge Hao, Salvatore Leonardi, and Jindřich Nečas, An example of irregular solution to a nonlinear Euler–Lagrange elliptic system with real analytic coefficients, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 23 (1996), no. 1, 57–67.
  26. [30]Guanghao Hong and Lizhou Wang, A singular critical point for some variational free boundary problem in five dimensions, Journal of Mathematical Analysis and Applications 426 (2015), no. 2, 855–863.DOI
  27. [31]Weimin Jin and Robert V. Kohn, Singular perturbation and the energy of folds, J. Nonlinear Sci. 10 (2000), no. 3, 355–390.DOI
  28. [32]Jan Kristensen and Giuseppe Mingione, The singular set of Lipschitzian minima of multiple integrals, Arch. Rational Mech. Anal. 184 (2007), no. 2, 341–369.DOI
  29. [33]R. D. Mindlin, Micro-structure in linear elasticity, Arch. Rational Mech. Anal. 16 (1964), 51–78.DOI
  30. [34]Connor Mooney, Minimizers of convex functionals with small degeneracy set, Calc. Var. Partial Differential Equations 59 (2020), no. 2, Paper No. 74, 19 pp.arxiv.org/abs/1903.03103
  31. [35]Charles B. Morrey, Jr., Multiple integrals in the calculus of variations, Die Grundlehren der mathematischen Wissenschaften, vol. 130, Springer-Verlag, Berlin–Heidelberg–New York, 1966.DOI
  32. [36]Connor Mooney and Ovidiu Savin, Some singular minimizers in low dimensions in the calculus of variations, Arch. Rational Mech. Anal. 221 (2016), no. 1, 1–22.DOI
  33. [37]John Nash, Continuity of solutions of parabolic and elliptic equations, Amer. J. Math. 80 (1958), no. 4, 931–954.DOI
  34. [38]Jindřich Nečas, Sur la régularité des solutions variationnelles des équations elliptiques non-linéaires d’ordre 2k en deux dimensions, Annali della Scuola Normale Superiore di Pisa, Classe di Scienze (3) 21 (1967), no. 3, 427–457.
  35. [39]———, Example of an irregular solution to a nonlinear elliptic system with analytic coefficients and conditions for regularity, Theory of Nonlinear Operators (Proc. Fourth Internat. Summer School, Berlin, 1975), Abh. Akad. Wiss. DDR, Abt. Math.-Natur.-Tech., no. 1977, 1, Akademie-Verlag, Berlin, 1977, pp. 197–206.DOI
  36. [40]Nikolai Nadirashvili, Vladimir Tkachev, and Serge Vlăduţ, A non-classical solution to a Hessian equation from Cartan isoparametric cubic, Adv. Math. 231 (2012), no. 3–4, 1589–1597.arxiv.org/abs/1111.0329
  37. [42]Yong-Geun Oh, Second variation and stabilities of minimal Lagrangian submanifolds in Kähler manifolds, Invent. Math. 101 (1990), 501–519.DOI
  38. [43]——, Volume minimization of Lagrangian submanifolds under Hamiltonian deformations, Math. Z. 212 (1993), 175–192.DOI
  39. [45]Richard Schoen and Jon G. Wolfson, Minimizing area among Lagrangian surfaces: the mapping problem, J. Differential Geom. 58 (2001), no. 1, 1–86.DOI
  40. [46]Vladimír Šverák and Xiaodong Yan, A singular minimizer of a smooth strongly convex functional in three dimensions, Calc. Var. Partial Differential Equations 10 (2000), no. 3, 213–221.DOI
  41. [47]——, Non-Lipschitz minimizers of smooth uniformly convex functionals, Proc. Natl. Acad. Sci. USA 99 (2002), no. 24, 15269–15276.pubmed.ncbi.nlm.nih.gov/12403822
  42. [48]R. A. Toupin, Elastic materials with couple-stresses, Archive for Rational Mechanics and Analysis 11 (1962), no. 1, 385–414.DOI
  43. [49]Neil S. Trudinger and Xu-Jia Wang, The affine Plateau problem, Journal of the American Mathematical Society 18 (2005), no. 2, 253–289.DOI
  44. [50]Lizhou Wang, Singular critical points for variational free boundary problem from isoparametric hypersurfaces, Communications in Contemporary Mathematics 18 (2016), no. 3, 1650010.DOI
  45. [51]Hermann Weyl, The classical groups: Their invariants and representations, Princeton University Press, Princeton, NJ, 1939.

Paper details

vlnc:W_bujy5d2d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents