für Mathematik

in den Naturwissenschaften
Leipzig

Studying nonlinear pde by geometry in

matrix space

by

Bernd Kirchheim, Stefan Müller, and Vladimı́r Šverák

Preprint no.: 49 2002

 Studying nonlinear pde by geometry in matrix
space
Bernd Kirchheim, Stefan Müller
Max Planck Institute for Mathematics in the Sciences
Inselstr. 22–26
D–04103 Leipzig, Germany
{bk,sm}@mis.mpg.de

Vladimı́r Šverák
Department of Mathematics
University of Minnesota
Minneapolis, MN 55455, USA
[email protected]

Abstract
We outline an approach to study the properties of nonlinear partial
differential equations through the geometric properties of a set in the
space of m × n matrices which is naturally associated to the equation.
In particular, different notions of convex hulls play a crucial role. This
work draws heavily on Tartar’s work on oscillations in nonlinear pde
and compensated compactness and on Gromov’s work on partial dif-
ferential relations and convex integration. We point out some recent
successes of this approach and outline a number of open problems,
most of which seem to require a better geometric understanding of the
different convexity notions.

Contents
1 Introduction 3

2 Convex integration and rank-one convex hulls 7

2.1 Convex integration for open sets . . . . . . . . . . . . . . . . 7
2.2 Closed sets and in-approximations . . . . . . . . . . . . . . . 8

1
3 Elliptic systems with nowhere smooth solutions 9
3.1 Reduction to first order systems . . . . . . . . . . . . . . . . . 9
3.2 Tk -configurations . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Embedding a T4 -configuration . . . . . . . . . . . . . . . . . . 11
3.4 Families of T4 -configurations and dimension counting . . . . . 14
3.5 Polyconvex examples and obstructions to T4 . . . . . . . . . . 17
3.6 Beyond Tk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Tools to study rank-one convex hulls 21

4.1 Rank-one convex, quasiconvex and polyconvex hulls . . . . . 21
4.2 The dual objects: laminates and gradient Young measures . . 22
4.3 Localization and extension . . . . . . . . . . . . . . . . . . . . 23

5 The simplest polyconvex integrand 25

5.1 A rank-one foliation of Σ . . . . . . . . . . . . . . . . . . . . 26
5.2 Separating functions on Σ . . . . . . . . . . . . . . . . . . . . 27
5.3 Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.4 More general polyconvex integrands . . . . . . . . . . . . . . 30

6 Separate convexity 31
6.1 Separate convexity in R2 . . . . . . . . . . . . . . . . . . . . . 31
6.2 Separate convexity in R2 ⊕ R . . . . . . . . . . . . . . . . . . 32
6.3 Separate convexity in R2 ⊕ R2 . . . . . . . . . . . . . . . . . . 43
6.4 A laminate without discrete approximations . . . . . . . . . . 46

7 Local hulls, degenerate sets and hyperbolic conservation

laws 49

8 Outlook 51

2
1 Introduction
The purpose of this paper is to outline a connection between nonlinear par-
tial differential equations and simply stated but largely unexplored ques-
tions about certain convex hulls, such as the rank-one convex hull or the
separately convex hull, in the space of m × n matrices. While many of
the underlying ideas are old and go back to the pioneering work of Tartar
[Ta 79, Ta 83], Gromov [Gr 73, Gr 86, Sp 98] and DiPerna [DP 85] there
have been a number of recent new successes of this approach including the
construction of elliptic and parabolic 2 × 2 systems with nowhere C 1 solu-
tions [MS 98, MS 99, MRS 02], an analysis of Lipschitz maps with finitely
many gradients [Ki 01a], the existence of solutions in mathematical models
of martensitic phase transitions [BJ 87, CK 88, MS 96] and a large number of
other applications of Gromov’s method of convex integration and its variants
and extensions, see e.g. [DM 97, DM 99, Ki 01b, Sy 01, MSy 01] for further
discussion and references. At the same time a theory of the relevant convex-
ity notions in matrix space is beginning to emerge [MP 98, Ki 01b, Ko 01]
even though many basic questions remain open.
In a nutshell, the situation can be described as follows. Many nonlinear
systems of pdes for a map u : Ω ⊂ Rn → Rd can be naturally expressed as
a combination of a linear systems of pdes


n
A(Dv) := Ai ∂i v = 0 (1)
i=1

and a pointwise nonlinear constraint

v(x) ∈ K ⊂ Rd a.e. (2)
Then one considers the cone Λ related to one dimensional solutions
v(x) = h(x, ξ) which is defined by
 


n
Λ := ξ ∈ R : ∃a ∈ R
n d
ξi Ai a = 0 . (3)
i=1

Equivalently Λ characterizes the directions of one dimensional high fre-

quency oscillations compatible with (1). Given a cone Λ we say that K
is lamination convex (with respect to Λ) if for any two points A, B ∈ K
with B − A ∈ Λ the whole segment [A, B] belongs to K. The lamination
convex hull K lc,Λ is the smallest lamination convex set containing K (Gro-
mov [Gr 86], who works in the more general setting of jet bundles, calls this
the P -convex hull).

3
 The key point in Gromov’s method of convex integration (which is a
far reaching generalization of the work of Nash [Na 54] and Kuiper [Ku 55]
on isometric immersions) is that (1) and (2) admit many interesting so-
lutions provided that K lc,Λ is sufficiently large. In applications to elliptic
and parabolic systems we always have K lc,Λ = K so that Gromov’s ap-
proach does not directly apply. It turns out, however, that for the con-
struction of Lipschitz (rather than C 1 ) solutions one can work with the
Λ-convex hull K Λ , defined by duality. More precisely for a compact set K
a point does not belong to K Λ if and only if there exists a Λ-convex func-
tion which separates it from K. A crucial fact is that K Λ can be much
larger than the hull K lc,Λ . This difference already arises for a set consist-
ing of four matrices which form a so-called T4 -configuration (see Section 3.2
below). This surprising fact was observed independently in different con-
texts [Sch 74, AH 86, NM 91, CT 93, Ta 93], we learned it from Tartar. In
connection with suitable approximations and general position arguments it
leads to surprising consequences. We illustrate this by three examples.
Theorem 1 ([MS 99]) (elliptic systems with nowhere C 1 solutions) Let
Ω be the unit ball in R2 . There exists a smooth function φ : R2×2 → R
which is strongly quasiconvex and satisfies |D2 φ| ≤ C, and a Lipschitz map
w : Ω → R2 , which is a weak solution of the elliptic system
− div Dφ(∇w) = 0, (4)
such that w is not C 1 in any open subset of Ω. Moreover the system (4)
admits solutions with compact support.
We remark in passing that this counterexample is quite different from
the classical counterexamples [BDG 69, DG 68, GM 68] and their more re-
cent extensions [HLN 96, SY 00] which are all based on singularities at a
point or more generally a set of lower dimension. Scheffer [Sch 74] used
T4 -configurations as a basis of counterexamples to regularity. He proved
a weaker version of Theorem 1 with w in the Sobolev space W 1,1 and φ
satisfying the Legendre-Hadamard condition (7). Unfortunately, the work
[Sch 74] has not appeared in a journal and the original ideas there remained
largely unknown and had to be re-discovered by various authors.
Note that (4) is the Euler-Lagrange equation of the functional

I(w) = φ(∇w) dx. (5)
Ω

By a classical result of Evans [Ev 86] minimizers of I are smooth outside a

closed set of measure zero if φ satisfies the assumptions stated in Theorem 1

4
(the same is true for local minimizers, see [KT 01]). Thus general stationary
points of I can behave much worse than minimizers.
We recall that a (continuous) integrand φ : Rm×n → R is called strongly
quasiconvex if
 
φ(X + ∇η) − φ(X) dx ≥ c |∇η|2 dx (6)
Tn Tn

for some c > 0, all X ∈ Rm×n and all periodic Lipschitz maps η : T n → Rm
(equivalently one can consider test functions η on bounded domains with
zero boundary conditions). If (6) holds with c = 0 we say that φ is qua-
siconvex. Using small amplitude test functions η(x), Taylor expansion and
Fourier transform one easily sees that (6) implies that φ is uniformly rank-
one convex and one obtains the Legendre-Hadamard (or strong ellipticity)
condition
D 2 φ(X)(a ⊗ b, a ⊗ b) ≥ c|a|2 |b|2 , (7)
so that (4) is indeed an elliptic system. Recently L. Székelyhidi has shown
that the conclusion of Theorem 1 also holds for a suitable strictly polyconvex
integrand φ, i.e. a strictly convex function of F and det F .
The failure of regularity can be extended to parabolic systems with
smooth initial data and a small and Hölder continuous right hand side.

Theorem 2 ([MRS 02]) (parabolic systems with nowhere C 1 solutions)

Let Ω be the unit ball in R2 . Let η > 0, T > 0, α ∈ (0, 1). Then there
exists a function φ : R2×2 → R such that φ is strongly quasiconvex, smooth
and |D 2 φ| ≤ C, a function f ∈ C α (Ω × [0, T ]; R2 ) with ||f ||C α < η and a
Lipschitz solution w : Ω × [0, T ] → R of the parabolic system

∂t w − div Dφ(∇w) = f in Ω × (0, T ) (8)

and

w(·, 0) ≡ 0, w(t, x) = 0 for x ∈ ∂Ω (9)

such that w is nowhere C 1 in Ω × (0, T ).

This system exhibits some other unusual features, such as failure of unique-
ness and of the energy inequality.
Our last example concerns Lipschitz maps whose gradient takes only
finitely many values (except on a set of measure zero).

5
Theorem 3 ([Ki 01a]) (maps with finitely many gradients). Let Ω be the
unit ball in R2 . There exist five matrices A1 , . . . , A5 ∈ R2×2 with

rk(Ai − Aj ) = 1 (10)

and a Lipschitz map u : Ω → R2 which satisfies

∇u ∈ {A1 , . . . , A5 } a.e. (11)

and ∇u ≡ Ai .
Interestingly, the corresponding statement for four matrices turns out to
be false [CK 00]. The condition (10) rules out trivial maps which depend
only on one direction and whose gradient takes two values. It also implies
that the sets Ωi = {x : ∇(x) = Ai } must be very complicated. Indeed if Ωi
and Ωj meet at a smooth (or rectifiable) boundary then a straightforward
blow-up argument shows that Ai − Aj must have rank one and the common
boundary of Ωi and Ωj is flat with normal b, where Ai − Aj = a ⊗ b. In this
context (10) can be seen as an ellipticity condition for the partial differential
relation (11). Nonetheless, as in Theorem 1, ellipticity is not strong enough
to rule out large scale oscillations of ∇u. These are, in are certain sense,
encoded in the T4 -configurations alluded to above (see Section 3.2 below).
The rest of this paper is organized as follows. We specialize to the
situation that v ∈ Rm×n and that the differential constraint (1) is simply
curl v = 0 (where the curl is taken along rows). Then the combination
of (1) and (2) leads to the first order partial differential relation ∇u ∈ K
and Λ is the cone of rank-one matrices. In Section 2 we review the general
results on convex integration and reduce the existence of (highly oscillatory)
solutions to the computation of the rank-one convex hull of K. To illustrate
this idea we outline in Section 3 the main ideas of the proof of Theorem 1.
One first finds one T4 -configuration in the relevant set K and then uses a
dimension counting argument to show that the abstract conditions reviewed
in Section 2 are satisfied.
The constructions related to Theorems 1–3 all use the simplest set K
which has the property that K lc,Λ = K, but K Λ is much bigger, the T4 -
configuration. We hope that a better understanding of the geometry of
rank-one convexity will lead to new applications and to insights how to
formulate structure conditions on φ which guarantee regularity results for
elliptic systems (so far nothing is known beyond monotonicity or quasimono-
tonicity [Fu 87, Zh 88, Ha 95] of Dφ).
To this end we first recall in Section 4 some general tools to study rank-
one convex hulls and then consider in Sections 5–7 a number of case studies.

6
In Section 5 we study the set K related to the simplest polyconvex inte-
grand φ(F ) = (det F )2 and we show that its rank-one convex hull is trivial.
Interestingly, it is not known whether the same holds for the set related to
the integrand ε|F |2 + (det F )2 for small positive ε. Already the restriction
to diagonal matrices leads to interesting questions about separately convex
functions which we discuss in Section 6. In Section 7 we discuss an exam-
ple related to compactness for hyperbolic conservation laws. It leads to a
set K which is degenerate in the sense that while K contains no rank-one
connections its tangent spaces do. The study of such sets was initiated in
a pioneering paper by DiPerna [DP 85], following the program outlined by
Tartar [Ta 79, Ta 83]. Finally in Section 8 we give a brief outlook.

2 Convex integration and rank-one convex hulls

For simplicity we will mostly consider the situation v ∈ Rm×n and the
simplest linear differential constraint curl v = 0 (where the curl is taken by
rows). If we restrict attention to simply connected domains Ω in Rn then
the combination of (1) and (2) reduces to the first order partial differential
relation
∇u ∈ K a.e. in Ω, (12)
where K ⊂ Rm×n is given and where we seek a map u : Ω ⊂ Rn → Rm .
In this case the cone Λ defined by (3) is simply the cone of rank-one
matrices and for brevity we use the notation K lc := K lc,Λ and K rc := K Λ
for the lamination convex hull and the rank-one convex hull. A function
f : Rm×n → R is rank-one convex if it is convex along each rank-one line
A + ta ⊗ b. For a compact set K the rank-one convex hull K rc consists of
all points which cannot be separated by rank-one convex functions, i.e.
K rc := {A ∈ Rm×n : f (A) ≤ sup f, for all rank-one convex f }. (13)
K

For a general set E we set


E rc := K rc . (14)
K⊂E compact

If m = 1 or n = 1 then K rc and K lc agree with the convex hull.

2.1 Convex integration for open sets

The key result in the theory of convex integration is that the partial dif-
ferential relation (12) admits many solutions if K rc is large. Here we just

7
recall the relevant results and refer to [MS 99] for the proofs and further
discussion. We first consider the case that K is open. In the following we
say that a map u : Ω → Rm is piecewise affine if it is Lipschitz and there
exists finite or countably many open sets Ωi such that u is affine on Ωi and
the union of the Ωi has full measure.

Theorem 4 ([MS 99], Thm. 3.1) Let K ⊂ Rm×n be open and let L ⊂
K rc be compact. Let u0 : Ω → Rm be a piecewise affine map with ∇u0 ∈ L
a.e. Then there exists a piecewise affine map u : Ω → Rm such that

∇u ∈ K a.e. in Ω, (15)

u = u0 on ∂Ω. (16)

Remark. In fact there exist many such solutions u. One can show that
u0 admits a fine approximation by solutions of (15), i.e. for each continuous
function η with η > 0 in Ω there exists a solution u with |u − u0 |(x) < η(x).

2.2 Closed sets and in-approximations

One crucial step in convex integration is the passage from open sets K to
general sets (which may have high codimension). It is now understood that
at least for Lipschitz solutions this can be done in different ways, e.g. by the
Baire category theorem [DM 97, DM 99] (for earlier applications to odes see
e.g. [Ce 80, DP 82, DP 91]), a refinement of it using Baire-1 functions or
the Banach-Mazur game [Ki 01b] or by direct construction [Sy 01, MSy 01].
As in [MS 96] we follow here Gromov’s original approach based on an in-
approximation. Whatever approach one uses the basic theme is the same
as in the construction of continuous, nowhere differentiable functions: at
each step of the construction one adds a highly oscillatory correction whose
frequency is much larger and whose amplitude is much smaller than those
of the previous corrections. This leads to strong convergence of the gradient
in L1 but typically to a very irregular limiting Lipschitz map.
For simplicity we consider only compact sets K (for the application to
elliptic systems discussed in Section 3 below it suffices to intersect the set
in (24) with a large ball).

Definition 5 We say that a sequence of open sets {Ui } is an in-approxi-

mation of a compact set K if Ui ⊂ Ui+1
rc and sup
X∈Ui dist(X, K) → 0 as
i → ∞.

8
Theorem 6 ([MS 99], Thm. 3.2.) Suppose that the compact set K ad-
mits an in-approximation by open sets {Ui }. Let u0 : Ω → Rm be a C 1 map
which satisfies
∇u0 ∈ U1 in Ω. (17)
Then there exists a Lipschitz map u : Ω → Rm such that

∇u ∈ K a.e. in Ω, (18)

u = u0 on ∂Ω. (19)

Remark. Again each such u0 admits in fact a fine approximation by

solutions u of (18).
An illustrative example is given by equidimensional isometric (Lipschitz)
immersions. In this case K = O(n) = {F ∈ Rn×n : F T F = Id}. We have
K rc = K lc = {F : F T F ≤ Id} and an in-approximation of K is given by
Ui = {F : λi Id < F T F < Id}, where 0 < λ1 < λ2 < . . . 1, λi → 1. Let
||F || = sup{|F x| : |x| ≤ 1} denote the operator norm of F with respect to
the Euclidean norm in Rn . Then any C 1 map u0 with sup ||∇u0 || < 1 can
be approximated in C 0 by Lipschitz maps u with ∇u ∈ O(n) which satisfy
the same boundary conditions. For other examples, including applications
to models of solid-solid phase transitions in crystals, see [MS 98, DM 99].

3 Elliptic systems with nowhere smooth solutions

In this section we sketch the proof of Theorem 1, following partly the expo-
sition in [MRS 02].

3.1 Reduction to first order systems

For definiteness let Ω be the unit ball in R2 . We seek solutions w : Ω → R2
of the system
− div Dφ(∇w) = 0, (20)
where
φ strongly quasiconvex, smooth, |D2 φ| ≤ C. (21)
In particular φ satisfies the Legendre-Hadamard condition (7). Now the
condition (20) is equivalent to the existence of a potential W such that
Dφ(∇w)J = ∇W where J is the 90◦ rotation. If we introduce
 
w
u= , u : Ω ⊂ R2 → R4 (22)
W

9
then (20) is equivalent to

∇u ∈ K ⊂ R4×2 , (23)

where
   
X 0 −1
K= : Y = Dφ(X)J, X ∈ R2×2 , J= . (24)
Y 1 0

3.2 Tk -configurations
To construct ‘wild’ solutions of (23) and hence (20) we have to show that
K rc is sufficiently large so that in-approximation of K can be constructed.
A first attempt might be to show that K lc is large. This, however, is doomed
since the Legendre-Hadamard condition (7) implies that K lc = K. Indeed
if A = X a⊗n
Y and A + b⊗n belong to K, then Dφ(X + ta ⊗ n) − Dφ(X) =
(b ⊗ n)J T = b ⊗ Jn. Hence Dφ(X + ta ⊗ n) − Dφ(X), a ⊗ n = 0 and
this contradicts the strict convexity of the map t → φ(X + ta ⊗ n), see also
[Ba 80].
A crucial observation is that there are simple sets which have a nontrivial
rank-one convex hull K rc even though K lc is trivial.

Definition 7 Let k ≥ 4 and consider k-tuples M = (M1 , . . . , Mk ) of matri-

ces Mj ∈ Rm×n . We say that M is a Tk -configuration if there exist rank-one
matrices C1 , . . . , Ck with kj=1 Cj = 0, scalars κ1 , . . . , κk with κj > 0 and
matrices Pj ∈ Rm×n such that the relations

Pj+1 − Pj = Cj , Mj − Pj+1 = κj Cj

hold, where the index j is counted modulo k (see Fig. 1).

Let M be a Tk -configuration and let K = {M1 , . . . , Mk }. One easily

sees that every rank-one convex function which vanishes on K must also
be nonpositive at all the Pj (see Fig. 1). Hence the Pj belong to K rc . On
the other hand there may be no rank-one connection in K. The simplest
example arises already for k = 4 in diagonal 2 × 2 matrices. One may take
   
3 0 1 0
M1 = −M3 = , M2 = −M4 = .
0 −1 0 3

We emphasize that in general a T4 -configuration need not lie in a plane. To

construct ‘wild’ solutions of (23) we will show in the next subsections that
K contains sufficiently many T4 -configurations.

10
 M2
M3 P4
P3

P1 C1 P2 M1

M4
Figure 1: T4 -configuration with P1 = P , P2 = P + C1 , P3 = P + C1 + C2 ,
P4 = P + C1 + C2 + C3 . The lines indicate rank-1 connections. Note that
the figure need not be planar

3.3 Embedding a T4 -configuration

The first observation is that there exists a strongly quasiconvex and smooth
function φ : R2×2 → R (with |D2 φ| ≤ C) such that the set
   
X 0 −1
K= :X ∈R 2×2
⊂ R ,J =
4×2
Dφ(X)J 1 0

admits a T4 -configuration M0 with Mi0 ∈ K, see [MS 99], Lemma 4.3. One
may take
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
3 0 1 0 −3 0 −1 0
⎜ −1 ⎟ ⎜ ⎟ ⎜ 1⎟ ⎜ ⎟
M10 = ⎜
0 ⎟ 0 ⎜ 0 3 ⎟ 0 ⎜ 0 ⎟ , M 0 = ⎜ 0 −3⎟ . (25)
⎝0 −1⎠ , M2 = ⎝0 3⎠ , M3 = ⎝ 0 1⎠ 4 ⎝ 0 −3⎠
3 0 1 0 −3 0 −1 0

To illustrate the geometric ideas behind the construction of such an

integrand we first sketch the construction of a φ which satisfies at least
the Legendre-Hadamard condition (7). In this case it essentially suffices to
construct φ on diagonal 2 × 2 matrices. Then one can use general extension

11
arguments in the spirit of Proposition 13 below to conclude. Let
     
s s 0 σ
f =φ , = ∇f.
t 0 t τ

We look for T4 -configurations in the set

⎛ ⎞
s 0
⎜ 0 t ⎟
⎜ ⎟.
⎝ 0 −σ(s, t) ⎠
τ (s, t) 0

The Legendre-Hadamard condition reduces to

∂σ ∂τ
≥ c, ≥ c, c > 0. (26)
∂s ∂t
If we drop the constraint that (σ, τ ) is a gradient we can easily embed a
T4 -configuration in the plane {σ = τ = 0}. It suffices to make the ansatz

σ(s, t) = s − g(t), τ (s, t) = t − h(s).

Then the Legendre-Hadamard condition imposes no constraint on g and h

and hence we can embed in the plane {σ = τ = 0} any set which is a
graph both over s and over t. For definiteness we chose the T4 -configuration
(−3, 1), (1, 3), (3, −1), (−1, −3).
One can check, however, that the plane {σ = τ = 0} cannot contain a T4 -
configuration if (σ, τ ) is a gradient (it suffices to integrate around the inner
square of the T4 -configuration and to use (26) to obtain a contradiction). To
overcome this difficulty we tilt the T4 -configuration, i.e. we look for a linear
map R2 → R4×2 which preserves rank-one connections. A natural choice is
⎛ ⎞
s 0
⎜ 0 t ⎟
(s, t) → ⎜ ⎟
⎝ 0 µt ⎠ ,
µs 0

where µ is a constant. This leads to the conditions σ = −µt, τ = µs or

equivalently
       
s s s 3
∇f = µJ , for =J k
, k ∈ {0, 1, 2, 3}, (27)
t t t −1

12
 t

−3 a 3 s

Q
−a

Figure 2: Construction of an integrand for which the set K contains a T4 -

configuration. The function f0 is bilinear in the quadrant Q = {(s, t) : s >
a, t > −a} and vanishes outside Q. It vanishes in particular on the rotated
quadrants J k Q, where k = 1, 2, 3 and where J is the 90◦ rotation.

where J is the 90◦ rotation. We make an ansatz for f which reflects the 90◦
rotation symmetry of the planar T4 -configuration:
 
3   
s s
f = f0 J k + λ(s2 + t2 ), where λ > 0.
t t
k=0

Then it suffices to verify (27) for k = 0 and this is equivalent to

 
3
∇f (x), x = 0, with x = . (28)
−1
We now make the choice (see Fig. 2)

 
s
f0 = (s − a)+ (t − a)+ , where 1 < a < 3
t

13
and where s+ = max(s, 0). In the quadrant {s > a, t > −a} we have
f (s, t) = f0 (s, t) + λ(s2 + t2 ) and one easily checks that (28) is satisfied if,
for example, a = 5/4, λ = 1/20. Multiplying of f by a suitable factor we
thus see that the T4 -configuration (25) lies in K.
To construct a quasiconvex φ for which K contains (25) we first use
that the rank-one convex function s+ t+ on diagonal 2 × 2 matrices can be
extended to a quasiconvex function on symmetric 2×2 matrices, see [Sv 92b].
Then φ can be extended to all 2 × 2 matrices by adding a high quadratic
penalty for the skew-symmetric part, see [MS 99] for the details.

3.4 Families of T4 -configurations and dimension counting

In order to construct an in-approximation of K, which is needed for the
application of Theorem 6, we first show that

K := K × K × K × K ⊂ (R4×2 )4 (29)

contains not only the special T4 -configuration M0 but an eight-dimensional

family of T4 -configurations. Then we will show that the corresponding corner
points Pj cover an open set in the eight dimensional space R4×2 and use this
fact to construct the in-approximation.
The dimension of the set of T4 -configurations can be guessed by a sim-
ple parameter count. First note that the rank-one cone in R4×2 is a five
dimensional manifold (away from its vertex). In view of Definition 7 the set

M = {M ∈ (R4×2 )4 : M is a T4 − configuration} ⊂ R32 . (30)

involves 4 × 5 + 4 × 1 + 8 = 32 parameters (namely the Cj , the κj and P1 )

which are subject to 8 constraints (namely Cj = 0). Hence we expect M
to be a 24 dimensional manifold in the neighbourhood of M0 . Since K is 16
dimensional, we expect the intersection to be have dimension 8, as desired.
To verify this and to actually construct the in-approximation we consider
the maps

πj : M ∩ K −→ R4×2
(M1 , M2 , M3 , M4 ) −→ Pj
µj : M ∩ K −→ R4×2
(M1 , M2 , M3 , M4 ) −→ Mj

14
Let TM 0 K be the tangent space of K at Mj0 , let Q⊥
j denote the projection
j
onto its orthogonal complement and define the map

ψj : M ∩ K −→ R4×2
(M1 , M2 , M3 , M4 ) −→ (Mj , Q⊥
j (Pj − Pj ))
0

Proposition 8 ([MS 99]) There exists a choice of φ such that M0 ∈ K and

(i) in a neighbourhood of M0 the sets M and K intersect transversely in an

eight dimensional manifold,

(ii) πj and ψj are local diffeomorphisms from a neighbourhood of M0 in

M ∩ K to open sets in R4×2 .

With this result at hand one can construct an in-approximation as follows

(see Fig. 3). Let O1 ⊂ O2 ⊂ . . . be (small) open neighbourhoods of M0 in
M ∩ K which are diffeomorphic to eight dimensional balls. For 1/2 < λ2 <
λ3 < . . . < 1 consider the maps λi µj + (1 − λi )πj and define for i ≥ 2 the
sets Ui by

Ui = ∪4j=1 Uij , Uij = (λi µj + (1 − λi )πj )(Oi ).

Using the non-degeneracy of ψj one can show that the Ui are open (if λ2
is chosen sufficiently close to 1), see [MS 99]. Let U1 = U2rc . Then {Ui }i≥1
is an in-approximation of K. In fact it is an in-approximation of the set
K intersected with a small neighbourhood of the set {M10 , M20 , M03 , M04 }.
Moreover U1 contains the points P10 , P20 , P30 , P40 . Hence it also contains 0 .
This shows that (23) admits a non-trivial solution with zero boundary
conditions whose gradient is always close to the set {M10 , M20 , M03 , M04 }. It
is easy to see that one can achieve in addition 0 ∈ K. Hence extension of
the solution by zero yields a solution of (4) with compact support.
To see that this solution is nowhere C 1 one has to trace back the general
construction used in the proof of Theorem 6 a bit more carefully, see [MS 99]
for the details. The main point is that at each step of the construction a
(locally) affine map is replaced by a piecewise affine map whose gradient
takes values near all of the four points Mi0 . This leads to a limit map whose
gradient has an oscillation of order 1 in every open set.

15
 M20

U22

M30 U23
1111111
0000000
0000000
1111111 M
0000000
1111111
P1 Q
0000000
1111111
0000000
1111111
0000000
1111111 M10
0000000
1111111
P10 U21 U31 U41
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
U24

M40

Figure 3: Schematic illustration of the sets Uij ⊂ R4×2 . The solid (resp.
dashed, or dotted) lines through the point M10 are the sets µ1 (O2 ), µ1 (O3 ),
µ1 (O4 ), respectively, i.e. the projections of the sets Oi ⊂ (R4×2 )4 to the first
component. The sets µ1 (Oi ) are not open in R4×2 since they are contained
in K. The shaded set is π1 (O4 ) and is open in R4×2 and the sets Ui1 =
[(1 − λi )π1 − λi µ1 ](Oi ) are also open. A typical point Q in U41 is given by
(1 − λ4 )π1 (M) + λ4 M1 , where M = (M1 , M2 , M3 , M4 ) ∈ O4 . In particular
Q lies on the rank-one segment [P1 , M1 ] and hence in the rank-one convex
hull of {M1 , M2 , M3 , M4 }. It also lies in the rank-one convex hull of four
points in U5 which are close the Mi .

16
3.5 Polyconvex examples and obstructions to T4
Can one carry out the construction outlined above also for an integrand φ
which is uniformly polyconvex (i.e. φ(X) = g(X, det X) where g is uniformly
convex) ? It turns out that T4 -configurations are no longer sufficient.
Proposition 9 Suppose that φ : R2×2 is strictly polyconvex. Then the set
K given by (24) does not contain any T4 -configuration.
Nonetheless Székelyhidi [Sz 02] has shown that there exists a uniformly poly-
convex φ such that the elliptic system (4) admits a Lipschitz, nowhere C 1
solution. His proof uses the fact that one can embed sufficiently many T5 -
configurations in K.
The statement of the proposition above is in fact the corollary of a
slightly more general result. It gives some insight into the consequences
of the monotonicity condition on the gradient of the convex function g rep-
resenting our polyconvex integrand. It states that Kφ can not support a
discrete laminate whose second R2×2 -coordinate is just a linear image of
the first R2×2 -coordinate, see Proposition 10 below. As we will see this in
particular rules out the existence of a T4 configuration in Kφ .
We begin with a little algebra. Let
φ(X) = g(X, det X) for X ∈ R2×2 ,
where g : R5 → R is strictly convex. Using that
∇φ(X) = ∇X G(X, det X) + G,5 (X, det X) cof X,
and writing ρ(X) = G,5 (X, det X) we obtain from the strict monotonicity
of ∇G that for any two different X, X̃ ∈ R2×2
0 < ∇φ(X̃) − ∇φ(X) − ρ(X̃) cof X̃ + ρ(X) cof X, X̃ − X
+ (ρ(X̃) − ρ(X))(det X̃ − det X).
We abbreviate ρ = ρ(X), Y = ∇φ(X), ρ̃ = ρ(X̃) and Ỹ = ∇φ(X̃). Expand-
ing the difference of the determinants we conclude
0 < Ỹ − Y, X̃ − X + ρ cof X − ρ̃ cof X̃, X̃ − X
+ (ρ̃ − ρ)(cof X, X̃ − X + det(X̃ − X))
= Ỹ − Y, X̃ − X − ρ̃(cof X̃ − cof X), X̃ − X + (ρ̃ − ρ) det(X̃ − X)
= Ỹ − Y, X̃ − X − 2ρ̃ det(X̃ − X) + (ρ̃ − ρ) det(X̃ − X).
Therefore, we have for any X = X̃
0 < ∇φ(X̃) − ∇φ(X), X̃ − X − (ρ(X) + ρ(X̃)) det(X − X̃). (31)

17
Proposition 10 Suppose that φ : R2×2 → R is strictly polyconvex, and that
for  
Xi
K= ; i = 1, . . . , n ⊂ Kφ
Yi
there is a matrix A ∈ R2×2 such that

(Yi − Yj ) = A(Xi − Xj ) for all i, j ≤ n.

Then K rc = K.

Proof. From the definition of K we obtain ∇φ(Xi ) = −Yi J. Since for

all X ∈ R2×2 the equation XJ = J cof X holds, we infer from (31) that for
any i = j

0 < −A(Xi − Xj )J, Xi − Xj  − (ρ(Xi ) + ρ(Xj )) det(Xi − Xj )

= −AJ cof(Xi − Xj ), Xi − Xj  − (ρ(Xi ) + ρ(Xj )) det(Xi − Xj )
= tr((−AJ) cof (Xi − Xj )(Xi − Xj )T ) − (ρ(Xi ) + ρ(Xj )) det(Xi − Xj )
= det(Xi − Xj )(tr(−AJ) − ρ(Xi ) − ρ(Xj )).

We denote
1
σ(X) = −( tr(AJ) + ρ(X))
2
and have therefore

det(Xi − Xj )(σ(Xi ) + σ(Xj )) > 0 for i = j.

Lemma 11 below ensures now that in case K rc = K we find a closed cycle

i0 , i1 , . . . , il , il+1 ∈ {1, . . . , n} with i0 = il , i1 = il+1 such that

det(Xik − Xik−1 ) det(Xik+1 − Xik ) < 0 for all k = 1, . . . , l.

Up to a shift of indices we can therefore suppose that

(−1)k (σ(Xik ) + σ(Xik+1 )) > 0 if k = 1, . . . , l

and that l is even. Summing over this cycle we get

−σ(Xi1 ) + σ(Xi2 ) − σ(Xi2 ) · · · + σ(Xil+1 ) > 0.

In other words, 0 > 0 - this contradiction finishes our proof.

18
Lemma 11 Assume a laminate µ ∈ Mrc (R2×2 ) is supported in a finite set
{X1 , . . . , Xn } satisfying det(Xi − Xj ) = 0 if i = j. (See Section 4.2 for the
definition of Mrc .) Then there is a closed cycle

i0 , i1 , . . . , il , il+1 ∈ {1, . . . , n} with i0 = il , i1 = il+1

satisfying

det(Xik − Xik−1 ) det(Xik+1 − Xik ) < 0 for k = 1, . . . , l.

Proof. We claim that

for each i ≤ n there are j, k : det(Xj − Xi ) det(Xk − Xi ) < 0. (32)

Indeed, if det Xj − Xi , has a fixed sign for all j different from i then
one can show that the laminate µ is either a Dirac mass at Xi or does not
charge Xi (see Proposition 15 below). This contradicts the hypothesis that
the support of µ is exactly {X1 , . . . , Xn }.
A little combinatorial argument will now show that (32) implies the
existence of the required cycle. Indeed, (32) certainly enforces the existence
of

i1 , . . . , il , il+1 with i1 = il+1 , (−1)k det(Xik+1 − Xik ) < 0 if 1 ≤ k ≤ l.

Consequently, the only problem that might occur is that l is odd and so
det(Xi2 − Xi1 ), det(Xil − Xi1 ) > 0. But, we observe that then necessarily

det(Xik − Xi1 ) > 0 also for all k ∈ {3, . . . , l − 1}.

Indeed, if this inequality fails then we extend the starting negative connec-
tion i1 , ik into the adjacent positive connection and keep then running in
this direction inside the already built cycle back to i1 . In other words, for
even k we can take ik , i1 , i2 , . . . , ik , i1 as the desired cycle and if k is odd
then il , i1 , ik , ik+1 , . . . il , i1 does the job.
However, by (32) also Xi1 has to have a negative connection, so we find

Xi0 with det(Xi0 − Xi1 ) < 0 and Xi0 ∈

/ {Xi1 , . . . , Xil }.

The same reasoning as before now shows that

det(Xi0 − Xik ) < 0 for all k = 2, . . . , l.

Next, we find Xi−1 with det(Xi−1 − Xik ) > 0 if k ≥ 0, and so on. Obviously,
repeating this argument finally leads to a contradiction, because our set
{X1 , . . . , Xn } is finite.

19
 Proof of Proposition 9. In light of Proposition 10 we only need to show
that for any T4 -configuration Mi = (Xi , Yi ) ∈ R2×2 × R2×2 there exists an
A ∈ R2×2 such that Yi − Yj = A(Xi − Xj ). According to Definition 7 we
find ni , xi , yi ∈ R2 and κi > 0 for i = 1, . . . , 4 such that
4  

   
xi xi+1 xi
⊗ ni = 0 and Mi+1 − Mi = (κi+1 + 1) ⊗ ni+1 − κi ⊗ ni .
yi yi+1 yi
i=1

So we are done, if we find A ∈ R2×2 with yi = Axi for all i. For this
purpose we notice that any consecutive ni , ni+1 are linearly independent and
that x1 , x2 , x3 , x4 span the whole R2 . Indeed, the first statement is obvious
from rk(Mi − Mi+1 ) > 1. So suppose the second fails, then all matrices
X1 , . . . , X4 one contained in a single rank-one plane S ⊂ R2×2 . Of course,
φ must be strictly convex on S which gives the monotonicity condition

0 < ∇φ(Xi ) − ∇φ(Xj ), Xi − Xj  = −(Yi − Yj )J, Xi − Xj 

for all i = j. Because Y J, X = (X11 Y12 − X12 Y11 ) + (X21 Y22 − X22 Y21 )
we see that a certain sum of quadratic minors is negative on all differences
in the set {M1 , . . . , M4 }. By [Sv 93] this even implies that all polyconvex
measures on {M1 , . . . , M4 } have to be Dirac masses. So it is clear that after
reshuffling the indices if necessary, we can assume that {n1 , n2 } and {x3 , x4 }
are both bases for R2 .
Next, we observe that
 
v1
u1 ⊗ v1 + u2 ⊗ v2 = (u1 u2 ) ,
v2

if the ui ’s are column and the vi ’s are row vectors from R2 . So we have
   
n1 n3
−(x1 x2 ) = (x3 x4 ) .
n2 n4
In other words our zero sum assumption implies
  −1
n3 n1
(x1 x2 ) = (x3 x4 )B and (y1 y2 ) = (y3 y4 )B, for B = − .
n4 n2
Thus, if we define A by yj = Axj for j = 3, 4 then

(y1 y2 ) = (y3 y4 )B = [A(x3 x4 )]B = A[(x3 x4 )B] = A(x1 x2 )

as required.

20
3.6 Beyond Tk
In the above examples it is enough to embed a T4 -configuration (or a T5 -
configuration) to show that K rc is sufficiently rich. The same strategy essen-
tially works for Theorems 2 and 3. To understand more general examples
and to find structure conditions which would exclude ‘wild’ solutions one
would like to compute (or at least estimate) K rc rather than just trying to
embed Tk -configurations (we will see below examples that K rc can be non-
trivial even if K contains no Tk -configuration). Ultimately this will require
a deeper understanding of the geometry of rank-one convexity. Due to the
high dimensions of the rank-one cone and the surrounding space this seems
rather difficult at the moment. To build up some intuition we discuss below
some examples where by different means one can reduce the dimensionality
of the problem and thus gain some geometric insight.
One interesting example is set K related to the simplest polyconvex
integrand

ε|F |2 + (det F )2 , ε ≥ 0.

For ε = 0 we show in Section 5 that K rc = K. Interestingly, it is not known

whether the same holds for the seemingly better case ε > 0, see Section 6
for some partial results. In both cases no partial regularity results for weak
solutions are known, except for the case of local minimizers.
In Section 7 we raise the question what one can say about the local
rank-one convex hull. One issue is whether one can formulate higher order
conditions which give triviality of the hull when the tangent space of K
contains rank-one lines. We first review some general tools to study rank-
one convex hulls.

4 Tools to study rank-one convex hulls

In this section we collect some definitions and tools which are useful for the
study of rank-one convex hulls and related hulls. For more detailed accounts
see e.g. [Da 89, Pe 97, Mu 99].

4.1 Rank-one convex, quasiconvex and polyconvex hulls

In the following we always assume that m, n ≥ 2 since otherwise all the con-
vexity notions introduced below agree with ordinary convexity. A function
f : Rm×n → R is rank-one convex if is convex on each rank-one line, it is

21
polyconvex if it can be expressed as a convex functions of minors (subdeter-
minants) and it is quasiconvex if

f (A + ∇η) − f (A) dx ≥ 0 (33)
Tn

for all A ∈ Rm×n and for all periodic Lipschitz maps η : T n → Rm . We have
the following implications:

f polyconvex ⇒ f quasiconvex ⇒ f rank-one convex. (34)

For a compact set we define the rank-one convex, quasiconvex and poly-
convex hull as the set of those points which can not be separated by the
corresponding class of functions.

K ∗ := {A ∈ Rm×n : f (A) ≤ sup f : f is ∗}, ∗ = {rc,qc,pc}. (35)

K

For a general set E we set


E ∗ := K ∗, ∗ = {rc,qc,pc}. (36)
K⊂E compact

In view of (34)
K pc ⊃ K qc ⊃ K rc . (37)
As we have seen in Sections 2 and 3 above a large rank-one convex hull
allows one to construct many solutions of ∇u ∈ K. The quasiconvex convex
hull is related to the stability of the partial differential relation under weak
convergence. More specifically if
∗
∇u(j)  ∇u and dist(∇u(j) , K) → 0 in L1 (38)

then
∇u ∈ K̄ qc , (39)
and this property characterizes K qc for compact K. The polyconvex hull
provides an upper bound for both K qc and K rc .

4.2 The dual objects: laminates and gradient Young mea-

sures
Let P(K) denote the set of probability
 measures supported on K. For
µ ∈ P(K) we denote by µ̄ = A dµ(A) its barycentre. We consider the

22
following subsets of P(K) which satisfy a Jensen’s inequality with respect
to the above convexity notions.

M (K) = {µ ∈ P(K) : f (A) dµ(A) ≥ f (µ̄), ∀f rc}
rc


M (K) = {µ ∈ P(K) : f (A) dµ(A) ≥ f (µ̄), ∀f qc}
qc


M (K) = {µ ∈ P(K) : f (A) dµ(A) ≥ f (µ̄), ∀f pc}
pc


= {µ ∈ P(K) : M (A) dµ(A) = M (µ̄), for all minors M }.

For a compact set K we have

K ∗ = {µ̄ : µ ∈ M∗ (K)}, ∗ ∈ {rc,qc,pc}.

The elements of Mqc (Rm×n ) are called homogeneous gradient Young

measures since they arise as distribution functions of the gradients of pe-
riodic functions. More precisely let µ ∈ Mqc (Rm×n ) be a measure with
compact support and let Ā = µ̄ be its barycentre. Then there exists peri-
odic maps u(j) : T n → R 
m with uniform Lipschitz constant such that the

measures µj defined by Rm×n η(A)dµj (A) := T n η(Ā + ∇u(j) ) dx converge
weak∗ to µ, see [KP 91, Sy 99]. For a compact set K the set Mqc (K) is
trivial (i.e. contains only Dirac masses) if and only if all sequences ∇u(j)
satisfying (38) converge strongly in L1 .
The elements of Mrc (K) are called laminates and they can be obtained
as weak∗ limits of so called laminates of finite order (with uniformly bounded
support), see [Pe 93, MS 99]. The class L of laminates of finite order is
defined by inductive splitting along rank-one segments as follows. First,
each Dirac mass δA belongs to L. Second, suppose that λi ≥ 0, m i=1 λi = 1,
that ν = λi δAi ∈ L and that Am = (1 − s)B1 + sB2 is a rank-one convex
m−1
combination of B1 and B2 . Then µ = i=1 λi δAi + (1 − s)λm δB1 + sλm δB2
also belongs to L. Thus laminates are the distribution functions of gradients
which arise from essentially one-dimensional constructions.

4.3 Localization and extension

Rank-one convexity is a local property (while qc and pc are not, at least for
m ≥ 3, see [Kr 99]) and this greatly simplifies the construction of separating
functions. We first recall that it suffices to separate locally.

23
Proposition 12 ([MS 99], Lemma 2.3) Let K be compact and let U be
an open neighbourhood of K rc . Suppose that f : U → R is rank-one convex
in U , i.e. convex on each rank-one segment entirely contained in U . Then
there exists a rank-one convex function F : Rm×n → R which agrees with f
in a neighbourhood of K rc .

One can also extend from lower dimensional sets, see [Sv 92a] for the
details.
Proposition 13 Let L be a subspace Rm×n and suppose that f : L → R
is C 2 and is rank-one convex on L. Let δ > 0 and let E ⊂ L be compact.
Then there exists a rank-one convex function F : Rm×n → R which satisfies
supE |F − f | ≤ δ.
If K rc has several components then the rank-one convex hull can be
computed for each piece separately.
Proposition 14 ([Ki 01b], Thm. 4.7) Let K be a compact set.
(i) Let B be bounded. Then

K rc ∩ B = [(K ∩ B) ∪ (K rc ∩ ∂B)]rc ∩ B. (40)

(ii) Let C1 , . . . Ck be disjoint compact sets and suppose K rc ⊂ ∪i Ci . Then

K rc = ∪i (K ∩ Ci )rc .
Part (ii) appears already in [Pe 93, MP 98] (see [Ma 01] for a detailed
proof) and it can easily be deduced from part (i) by taking B as a sufficiently
small neighbourhood Ui of Ci which does not intersect the other Cj . The
following application of the proposition will be useful later.

Proposition 15 Let K be a compact subset of {X ∈ R2×2 : det X > 0}.

Then (K ∪ {0})rc = K rc ∪ {0} and every laminate supported on K ∪ {0} is
either supported on K or is a Dirac mass at zero. In particular, if

E = {X ∈ R2×2 : det X = 1} ∪ {0}, then E rc = E.

Remark. Astala and Faraco [AF 02] have shown that the same assertion
holds for the quasiconvex hull and measures in Mqc (K ∪ {0}), i.e., gradient
Young measures. Their proof uses ideas from the theory of quasiregular
maps, in particular a careful analysis of the Beltrami equation.
Proof. By compactness there exist ε > 0, R > 0 such that det X ≥ ε and
|X| ≤ R for all X ∈ K. The polyconvex function f (X) = ε|X| − R det X

24
is ≤ 0 in K ∪ {0}, but is positive for 0 < |X| < 2ε/R. Thus we can apply
Proposition 14 with C1 = {X ∈ R2×2 : |X| ≤ ε/R} and C2 = {X ∈ R2×2 :
|X| ≥ 2ε/R} and we obtain (K ∪ {0})rc = K rc ∪ {0}.
Let µ be a laminate supported on K, with barycentre µ̄. Suppose first
µ̄ ∈ K rc . Let Ui be small neighbourhoods of Ci and define g : U = U1 ∪U2 →
R by g = −1 on U1 and g = 1 on U2 . Since g is constant on each component
of U it is trivially rank-one convex. By Proposition 12 there exists a rank-
one convex function g̃ : R2×2 → R which agrees with g on K. Hence
1 = g̃(µ̄) ≤ g̃ dµ = µ(K) − µ({0}). Thus µ must be supported on K. If
µ̄ = 0 one concludes similarly that µ = δ0 by starting from the function −g.

5 The simplest polyconvex integrand

Here we consider the simplest polyconvex integrand
1
φ(X) = (det X)2 .
2
The Euler-Lagrange equation div Dφ(∇w) = 0 is equivalent to the first order
partial differential relation
 
X
∇u ∈ K := ∈ R4×2 : Y = det X(cof X)J, X ∈ R2×2 , (41)
Y
where J is the 90◦ rotation.

Theorem 16 K rc = K.

To prove this result it is convenient to make the change of variables

X X
Y → −JY . This bijection maps rank-one lines onto rank-one lines and
therefore does not affect the computation of rank-one convex hulls. Since
−J cof XJ = X it thus suffices to consider the set
 
X
K̃ = ∈ R4×2 : Y = (det X)X, X ∈ R2×2 . (42)
Y

The set K̃ is contained in the cone

⎧⎛ ⎞ ⎫
⎪
⎪ α1 α2 ⎪
⎪
⎨⎜ ⎟ ⎬
β1 β2
Σ= ⎜ ⎝ γ1 γ2
⎟ : α ∧ γ = β ∧ δ = 0, α ∧ δ = β ∧ γ ,
⎠ (43)
⎪
⎪ ⎪
⎪
⎩ ⎭
δ1 δ2

25
where α ∧ γ = α1 γ2 − α2 γ1 etc. Since Σ is defined by minors it is polyconvex
and thus K̃ rc ⊂ Σ. To separate points in Σ \ K̃ we first use separating rank-
one convex functions defined only on Σ. Then we extend these function to
R4×2 . Since Σ is not smooth this requires some care.

5.1 A rank-one foliation of Σ

The construction of rank-one convex functions in Σ is largely simplified by
the fact that most of Σ can be decomposed into simpler sets which contain
all the rank-one connections.

Proposition 17 We have

Σ= Lλ ∪ L∞ ∪ N, (44)
λ∈R

where
   
X 0
Lλ = :X∈R 2×2
, L∞ = : Y ∈ R2×2 ,
λX Y

 
N = a ⊗ b : a ∈ R4 , b ∈ R2 .

Moreover if A, B ∈ Σ satisfy rk(A − B) = 1 then the rank-one line trough

A and B lies entirely in N or entirely in one of the spaces Lλ , where λ ∈
R ∪ {∞}.

Proof. Clearly N , Lλ and L∞ are contained in Σ. To see the converse

suppose suppose first that A = (α, β, γ, δ)T satisfies α ∧ β = 0 or α ∧ δ =
β ∧ γ = 0. Then α = 0 and β = 0 and the first two conditions in the
definition of Σ give γ = λα, δ = µβ while the last condition gives µ = λ.
Hence A ∈ Lλ . Now assume α ∧ β = α ∧ δ = β ∧ γ = 0. If γ ∧ δ = 0 then
α = λγ, β = µδ and this yields α = β = 0, so that A ∈ L∞ . Finally, if also
γ ∧ δ = 0 then all 2 × 2 minors of A are zero. Hence A ∈ N .
Since Σ is given by minors any rank-one line
   
X(t) X + tx ⊗ n
A(t) = =
Y (t) Y + ty ⊗ n

lies entirely in Σ if two points lie in Σ. If X(t) ≡ 0 then A(t) lies in L∞ .

Next assume that X(t) ≡ 0 but det X(t) ≡ 0. We claim that A(t) ∈ N .
Indeed X(t) = 0 for all but one value of t. If A(t) was not in N then

26
Y (t) = λ(t)X(t). Since det X(t) = 0 this yields again A(t) ∈ N for all
but one t and hence for all t. Finally assume that det X(t) ≡ 0. Then
det X(t) = 0 for all but one t. In particular Xn⊥ = 0 where n⊥ = Jn is
perpendicular to n. Moreover Y (t) = λ(t)X(t). Applying this identity to
n⊥ we deduce that λ(t) must be constant, so that the rank-one line lies in
one Lλ .

5.2 Separating functions on Σ

The foliation of Σ almost allows one to compute the rank-one convex hull
separately on all the spaces Lλ and on N . More precisely we make the
following ansatz for the separating function

Y ∈ Lλ ,
ϕ(λ) det Y if X
f (X, Y ) = h(X, Y ) − X (45)
0 if Y ∈ N ∪ L∞ .

Here h : Σ → R is rank-one convex and ϕ ∈ C0∞ (R). Since each rank-one

line stays either in Lλ or in N the function f is automatically rank-one
convex. On K̃ the expression for f simplifies since K̃ ∩ (N ∪ L∞ ) = {0} and
in K̃ ∩ Lλ we have λ = det X. We thus have

f (X, Y ) = h(X, X det X) − ϕ(det X)(det X)3 (46)

Now we specify h and φ in dependence of the point Ā = X̄Ȳ ∈ Σ \ K̃ we

seek to separate.
Case 1: Ā ∈ Lλ̄ , λ̄ ∈ {0, ∞}. The point A = X
Y belongs to K̃ ∩ Lλ̄ if
and only if

Y ∈ E := {Z ∈ R2×2 : det Z = λ̄3 } ∪ {0}.

By Proposition 15 the rank-one convex hull of E is trivial. Hence there

exists a rank-one convex function h̄ : R2×2 → R such that h̄(Ȳ ) = 1 and
h̄ ≤ 0 on E. Set h(X, Y ) = h̄(Y ). In view of (46) it suffices to show that
(for each R) there exists ϕ such that

ϕ(λ̄) = 0,

1
h̄(X det X) − ϕ(det X)(det X)3 ≤ , ∀X with |X| ≤ R
2
To see that this is possible first note that if det X = λ̄ then h̄(X det X) ≤ 0
by construction. Hence h̄(X det X) ≤ 1/2 for | det X − λ̄| ≤ δ0 (R). Simi-
larly for det X = 0 we have h̄(X det X) = 0. Hence h̄(X det X) ≤ 1/2 for

27
| det X| ≤ δ0 (R). For the remaining values of det X the desired inequality
can be achieved by a suitable choice of ϕ (note that det X ≤ R2 so that
there is no obstruction to choosing ϕ with compact support).
Case 2: Ā = X̄0 ∈ L0 \ N . We have det X̄ = 0 and we take

h(X, Y ) = det X sgn(det X̄).

Then we can choose ϕ such that

| det X̄|
|λ| − ϕ(λ)λ3 ≤ , for |λ| ≤ R2 . (47)
2
Case 3: Ā ∈ N ∪ L∞ . Let

h(X, Y ) = |Y |2 , ϕ = R2 ϕ̂,

where λ2 − ϕ̂(λ)λ3 ≤ R−2 |Ȳ |2 /2. On the set K̃ the function f is bounded
by

|X det X|2 − ϕ(det X)(det X)3 (48)

|Ā|2
≤ R2 [(det X)2 − ϕ̂(det X)(det X)3 ] ≤ . (49)
2

5.3 Extension
We show that the function f : Σ → R constructed above can be approxi-
mated (uniformly of compact subsets of Σ) by functions which are rank-1
convex in a neighbourhood of Σ. In view of Proposition 12 this will finish
the proof of Theorem 16.
The main point is to define an analogue of the parameter λ in (45) for a
general matrix A ∈ R4×2 . Consider the set
 
X cos α
Σ̃ = : X ∈ R2×2 , α ∈ (−π/2, π/2] .
X sin α
X
A point A = Y ∈ R4×2 has a unique best approximation πΣ̃ (A) in Σ̃ if
and only if

P (A) := (|X|2 − |Y |2 )2 + 4X, Y 2 = 0.

In this case the optimal angle ᾱ(A) is determined by the condition

   2 
cos 2ᾱ 1 |X| − |Y |2
= .
sin 2ᾱ P (A)1/2 2X, Y 

28
On the set P = 0 the closest point projection πΣ̃ is smooth. We define
λ̂(A) = tan ᾱ(A) and replace the ansatz (45) by

P (A)
fδ (A) = h(A) − ϕ(λ̂(A))η( ) det Y, (50)
δ4
where ϕ ∈ C0∞ (R), where 1 − η ∈ C0∞ (−1, 1) with η|(−1/2,1/2) = 1 and where
δ > 0 is a small parameter. Then fδ is well-defined on R4×2 and smooth.
We first claim that on Σ the function fδ is close to f and nearly rank-one
convex.
Clearly fδ = f on N since det Y = 0 on N . On each Lλ (including
λ = ∞) we have P (A) = |A|4 . Hence fδ = f on Σ \ Bδ and thus supΣ |f −
fδ | ≤ Cδ2 . Using again that P (A) = |A|4 on Lλ and the homogeneity of λ̂
and det it is easy to verify that for each rank-one line A + ta ⊗ b in Σ we
have D 2 fδ (A)(a ⊗ b, a ⊗ b) ≥ −C|a|2 |b|2 . Thus the proof is concluded by the
following extension result.
For A = (α, β, γ, δ)T ∈ R4×2 consider the minors

U1 (A) = α ∧ γ, U2 (A) = β ∧ δ, U3 (A) = α ∧ δ − β ∧ γ

and recall that Σ was defined as the set where all three minors vanish.

Proposition 18 Suppose that f ∈ C 2 (R4×2 ). Assume that for all rank-one

lines A + ta ⊗ b contained in Σ one has

D 2 fδ (A)(a ⊗ b, a ⊗ b) ≥ 0 if |A| ≥ δ,
D 2 fδ (A)(a ⊗ b, a ⊗ b) ≥ −C0 |a|2 |b|2 if |A| < δ.

Then there exists a smooth convex function g : R4×2 → R with

0 ≤ g(A) ≤ CC0 δ|A|

and for each ε > 0 there exists µ ≥ 0 such that the function

3
F := fδ + g + µ Ui2 + ε|A|2
i=1

is rank-one convex in a neighbourhood of Σ.

Proof. First note that there exists a smooth convex function g with
the properties stated in the proposition which satisfies D2 g(A) ≥ C0 Id
for |A| < δ. Fix ε > 0 and suppose the assertion of the proposition was

29
false. Then there exist points Ak → A with A ∈ Σ, rank-one directions
Bk = ak ⊗ bk with |Bk | = 1 and Bk → B, and µk → ∞ such that

3
2 2
D fδ (Ak )(Bk , Bk ) + D g(Ak )(Bk , Bk ) + 2µk DUi (Ak ), Bk 2 ≤ −2ε.
i=1

Taking the limit k → ∞ we see that

D 2 fδ (A)(B, B) + D2 g(A)(B, B) ≤ −2ε (51)

and DUi (A), B = 0. Hence A + tB is a rank-one line on Σ and thus (51)

leads to a contradiction with the hypotheses on D2 f and the choice of g.

Remark. The main idea behind formula (50) can be understood as
follows. Let Σreg = λ∈R Lλ \ {0}. We note that λ can be thought of as a
smooth function on Σreg defined by A ∈ Lλ(A) . For ϕ ∈ C0∞ (R) the function
A → ϕ(λ) can be smoothly extended from Σreg ∩ S(4×2)−1 to all of this unit
sphere. Such a function can now be extended to a smooth zero homogeneous
function ϕ̃ on R4×2 \ {0} which still agrees with the original function on all
of Σreg . Moreover, the function
 
X
A= → ϕ̃(A) det(Y )
Y

is 2-homogeneous on R4×2 , smooth away from 0 and rank-one convex on Σ.

The role of η and Proposition 18 is to handle the singularity at 0. Since
the second derivatives of ϕ̃(A) det(Y ) are bounded, this does not present a
problem.

5.4 More general polyconvex integrands

It seems natural to expect that Theorem 16 can be extended to all the
strictly polyconvex integrands
1 1
φ(F ) = ε|F |2 + (det F )2 , ε > 0, (52)
2 2
but whether this is true is not known. If we restrict X and Y to diagonal
matrices then the corresponding set K̃ becomes
⎧⎛ ⎞ ⎫
⎪
⎪ s 0 ⎪
⎪
⎨⎜ ⎟ ⎬
0 t
K̃diag = ⎜ ⎝ (st)s + εt
⎟ : s, t ∈ R .
⎠
⎪
⎪ 0 ⎪
⎪
⎩ ⎭
0 (st)t + εs

30
This set is contained in a four-dimensional subspace and on this subspace
rank-one convexity in R4×2 reduces to separate convexity in R2 ⊕ R2 . Even
rc = K̃
in this simplified setting it is not known whether K̃diag diag . In the next
section we will establish this result at least for finite sets (this in particular
implies that no Tk -configuration can be embedded in K̃diag ). To do so we
study separate convexity in more detail.

6 Separate convexity
6.1 Separate convexity in R2
This corresponds to the cone Λ = R × {0} ∪ {0} × R ⊂ R2 which arises if
we restrict the rank-one convex cone in 2 × 2 matrices to diagonal matrices.
This situation is relatively well understood. In particular every nontrivial
configuration must contain a T4 -configuration.

Proposition 19 ([Ta 93], Remark 10; [MP 98], Proposition 5.3) Let
K be a compact set in diagonal 2 × 2 matrices.
(i) Every point A ∈ K rc is contained in the rank-one convex hull of a
subset of K consisting of at most five points.
(ii) If K contains no rank-one connections but K rc = K then K must
contain a T4 -configuration.

There also exists an efficient algorithm for the computation of the rank-
one convex hull [MP 98]. Moreover on diagonal 2 × 2 matrices rank-one
convexity and quasiconvexity agree, in the sense that the spaces Mrc of
laminates and Mqc of gradient Young measures agree [Mu 99b].
Separate convexity in Rn = R ⊕ . . . ⊕ R which arises by restricting
rank-one convexity in Rn×n to diagonal matrices is already more subtle,
see [MP 98]. Here we are more interested in separate convexity in R2 ⊕ R2
and as an intermediate step we consider R2 ⊕ R.
One key tool is the following separation argument for sets which are
supported in two opposite quadrants [Ta 93]. In order to use the same
notions as in Section 4 we formally view separate convexity in Rm ⊕ Rn as
a special case of rank-one convexity (see (54)).

Lemma 20 For x0 , a ∈ Rm and y0 , b in Rn consider the generalized quad-

rants

Q±,± = {(x, y) ∈ Rm ⊕ Rn : ±x − x0 , a > 0, ±y − y0 , b > 0}

31
and the set

Q00 = {(x, y) ∈ Rm ⊕ Rn : x − x0 , a = y − y0 , b = 0}.

Let the set K ⊂ Q++ ∪Q−− ∪Q00 be compact and let µ be laminate supported
on K, with barycentre µ̄. Then one of the following three assertions holds
(i) µ̄ ∈ Q00 and supp µ ⊂ Q00 ,
(ii) µ̄ ∈ Q++ and supp µ ⊂ Q++ ∪ Q00 ,
(iii) µ̄ ∈ Q−− and supp µ ⊂ Q−− ∪ Q00 .
If, in addition, K ∩ Q++ and K ∩ Q−− are compact (e.g., if K is finite) then
in (ii) and (iii) one has supp µ ⊂ Q++ and supp µ ⊂ Q−− , respectively.

Proof. We may suppose x0 = y0 = 0. Consider the separately affine

function

f (x, y) = x, ay, b.


Then f (µ̄) = K f dµ ≥ 0. If f (µ̄) = 0 then supp µ ⊂ Q00 and hence
µ̄ ∈ Q00 . If f (µ̄) > 0 then µ̄ ∈ Q++ ∪ Q−− . Suppose µ̄ ∈ Q++ and consider
the separately convex function

g(x, y) = x, a+ y, b+ − x, ay, b.

where a+ = max(a, 0). Then g ≤ 0 on K and g < 0 on K ∩ Q−− . Since

0 = g(µ̄) ≤ g dµ it follows that supp µ ⊂ Q++ ∪Q00 . If K ∩Q++ is compact
we can find ε > 0 such that min(x, a, y, b) ≥ ε for all (x, y) ∈ K ∩Q++ . In
addition we may assume that µ̄1 , a ≥ ε and µ̄2 , b ≥ ε, where µ̄ = (µ̄1 , µ̄2 ).
Then we can replace x, a with x, a − ε and y, b with y, b − ε in the
definition of g and we conclude easily. The case µ̄ ∈ Q−− is analogous.

6.2 Separate convexity in R2 ⊕ R

Here we consider the cone

Λ = R2 × {0} ∪ {0} × R ⊂ R3 . (53)

We can view Λ as a subset of the rank-one cone in 3 × 2 matrices if we

identify R3 with the space
⎧⎛ ⎞ ⎫
⎨ x 0 ⎬
L = ⎝ y 0 ⎠ : x, y, z ∈ R . (54)
⎩ ⎭
0 z

32
 3

Figure 4: The figure on the left shows the T4 configuration embedded on a

non selfintersecting curve. The highly nonconvex spiral on the right, how-
ever, leads to a trivial rank-one convex hull

With this identification in mind we continue to write K rc instead of K Λ

for the rank-one convex hull. We wish to understand how the more com-
plicated geometry of separate convexity in R2 ⊕ R can be used to construct
examples which do not exist in R ⊕ R. We first summarize the results and
then turn to the proofs.

Proposition 21 Let L be given by (54). Suppose that K ⊂ L consists of

at most five points and contains no rank-one connection. If K rc = K then
K contains a T4 -configuration.

Proposition 22 Let L be given by (54). There exists a set K ⊂ L which

consists of six points, contains no T4 -configuration and has a non-trivial
rank-one convex hull K rc = K.

For the construction see Fig. 6. The short proof that the example con-
tains no Tk configuration is given below. An interesting class of sets with-
out rank-one connections are (monotone) graphs over curves without self-
intersections. For curves which are ‘spiral-like’ we can show that K rc is
trivial (see Fig. 4).

33
Proposition 23 Identify L given by (54) with R3 . Let γ : [0, T ] → R2
have a regular C 1 -image (i.e. if the curve is parametrized by arclength,
the derivative exists everywhere and varies continuously) which satisfies for
all t ∈ [0, T ) that γ(t) is not in the convex hull of {γ(s) : s > t}. Set
K = {(γ(t), t) : t ∈ [0, T ]}. Then K rc = K.

Without the hypotheses on the convex hull the result may fail, since
one can embed a T4 configuration (see Fig. 4). Interestingly, even with that
hypotheses the result can fail if we allow Lipschitz curves rather than C 1
curves, see Example 24 and Fig. 7 below.
Proof of Proposition 21. The proof shows that the simple geometric
separation argument, Lemma 20, becomes quite powerful when combined
with the localization formula (40) in Proposition 14 (i). We first eliminate
the simpler cases when K contains four or less points.
If K contains three or less points, then it is well-known that the absence
of rank-one connections implies K rc = K. For two points one can use a
suitable 2 × 2 minor to show this. We give a proof for three points for
our example since the same argument can be used for four points. Let
K = {P1 , P2 , P3 } with Pi = (pi , zi ) = (xi , yi , zi ). Since there are no rank-
one connections we can assume (after a permutation of indices if necessary)
that z1 < z2 < z3 . We claim that p1 ∈ {p2 , p3 }co . If this fails then there
exists a ∈ R2 such that pi − p1 , a > 0, for i = 2, 3. Thus the points P2 and
P3 lie in the generalized quadrant

Q++ = {(p, z) : p − p1 , a > 0, z − z1 > 0},

and K ⊂ Q++ ∪ Q00 . It follows from Lemma 20 that a nontrivial laminate

supported on K must be supported on {P2 , P3 }, but this possibility has
already been ruled out. Therefore p1 ∈ {p2 , p3 }co . Similarly we conclude
p3 ∈ {p1 , p2 }co . Thus p1 = p2 = p3 and K lies in a rank-one line, in
contradiction with our assumption.
If K = {P1 , P2 , P3 , P4 } we can again assume z1 < z2 < z3 < z4 and the
separation argument given above implies that p1 ∈ {p2 , p3 , p4 }co and p4 ∈
{p1 , p2 , p3 }co . Thus the convex set {p1 , p2 , p3 , p4 }co has only the two extreme
points p2 , p3 and hence {p1 , p2 , p3 , p4 } ⊂ {p2 , p3 }co . If {p1 , p2 }co ∩{p3 , p4 }co =
∅ we can again use the separation argument to obtain a contradiction. Hence
{p1 , p2 }co ∩ {p3 , p4 }co = ∅ and drawing the picture of K in the planar strip
[p2 , p3 ] × R ⊂ R2 × R we see the usual T4 -configuration (see Fig. 5).
Finally, if K contains five points, we suppose again z1 < z2 < ... < z5 .
As before, we infer p1 , p5 ∈ {p2 , p3 , p4 }co . If p2 , p3 and p4 lie on a line l then

34
 P4

P3

P2

P1

p2 p4 p1 p3

Figure 5: Finding the T4 -configuration in the plane spanned by [p2 , p3 ] and

the z-axis

K is contained in l × R and hence K must contain a T4 -configuration by

Proposition 19 (ii) for separate convexity in R2 (alternatively this can be
checked directly arguing as above and distinguishing a few cases).
Using separately convex functions of the type (p, z) → (p, a−b)± (z−c)±
that are chosen to vanish on K, we see that (p, z) ∈ K rc if

⎫
• z < z1 or z > z5 , or ⎬
• (z ∈ [z1 , z2 ) but p = p1 ) or (z ∈ (z4 , z5 ] but p = p5 ) or , (55)
⎭
• (z ∈ [z2 , z3 ) but p ∈ [p1 , p2 ]) or (z ∈ (z3 , z4 ] but p ∈ [p4 , p5 ]).

Now we will use the localization formula for the rank-one convex hull
(see Proposition 14(i))

K rc ∩ U = [(K ∩ U ) ∪ (K rc ∩ ∂U )]rc ∩ U, (56)

which is valid for any compact K and any bounded U . Let BR be an open
ball which contains K co and let

U = {(p, z) : z < z3 , dist(p, [p1 , p2 ]) < z3 − z} ∩ BR .

35
Then

∂U ⊂ ([p1 , p2 ] × {z3 }) ∪ {(p, z) : z < z3 , dist(p, [p1 , p2 ]) = z − z3 } ∪ ∂BR .

Thus (56) and (55) yield

K rc ∩ {(p, z) : z < z3 } = K rc ∩ U
⊂ U ∩ ({P1 , P2 } ∪ ([p1 , p2 ] × {z3 } ∩ K rc ))rc . (57)

Analogously, above the z3 -level we see

K rc ∩ {(p, z) : z > z3 } ⊂ ({P4 , P5 } ∪ ([p4 , p5 ] × {z3 } ∩ K rc ))rc . (58)

We claim that

C := K rc ∩ {(p, z) : z = z3 } = {P3 }.

Once this is shown were are done. Indeed (57) and (58) then imply that
K rc \ {(p, z) : z = z3 } = {P1 , P2 , P4 , P5 } since the rank-one convex hull of
three points without rank-one connections is trivial.
The compact convex set C is the convex hull of its extreme points.
Suppose that C = {P3 } and let P0 = (p0 , z3 ) ∈ C \ {P3 } be an extreme
point. We claim that p0 ∈ [p1 , p2 ] ∩ [p4 , p5 ]. To see this suppose first
p0 ∈ / [p1 , p2 ]. Choosing a ball Bε (P0 ) around P0 which is so small that
B̄ε (P0 ) ∩ (K ∪ ([p1 , p2 ] × R)) = ∅ we see from the localization formula that

P0 ∈ (K rc ∩ ∂Bε (P0 ))rc ⊂ [(C ∩ ∂Bε (P0 )) ∪ (K rc ∩ {(p, z) : z > z3 })]co .

Since in the computation of the convex hull points above {z = z3 } can not be
compensated for by points below {z = z3 } we get P0 ∈ (C ∩∂Bε (p0 ))co . This
contradicts the extremality of P0 in C. Similarly one shows p0 ∈ [p4 , p5 ].
For future reference we note that p0 = p2 . Otherwise p2 ∈ [p4 , p5 ] and thus
p5 = p2 , since p5 is in the convex hull of p2 , p3 and p4 and these three points
do not lie on a line. This contradicts the assumption that P2 and P5 are not
rank-one connected. Similarly p0 = p4 .
We next claim that the segments [p1 , p2 ] and [p4 , p5 ] intersect trans-
versely in a single point, which we denote again by p0 . Indeed, otherwise
p1 , p5 ∈ [p4 , p2 ] (since p1 , p5 ∈ {p2 , p3 , p4 }co ) and thus p4 , p1 , p5 , p2 form a
T4 -configuration in the plane generated by the line through p4 and p2 and
the z-axis. Thus the intersection must be transversal, p0 and P0 are uniquely
determined and C ⊂ [P0 , P3 ].
Moreover, the segment [p0 , p3 ] intersects both [p1 , p2 ] and [p4 , p5 ] in {p0 }
only, since else we would find a T4 -configuration over one of the edges [p2 , p3 ]

36
or [p4 , p3 ] (note that p0 ∈ [p2 , p3 ] implies p5 = p0 since p4 ∈
/ [p2 , p3 ]). Because
C ⊂ [P0 , P3 ] the inclusion (57) yields

K rc ∩ {(p, z) : z < z3 } ⊂ {P1 , P2 , P0 }rc = {P1 , P2 , P0 },

except when p0 = p1 . Suppose first p0 = p1 . Then K rc ∩ {(p, z) : z ∈

(z2 , z3 )} = ∅. As before this contradicts the fact P0 is not in K (i.e. different
from P3 ) but extreme in C. If p0 = p1 , we have p0 = p5 and we can use (58)
to find

K rc ∩ {(p, z) : z > z3 } ⊂ {P4 , P5 , P0 }rc = {P4 , P5 , P0 }

and to get the same contradiction. In conclusion we must have C := K rc ∩

{(p, z) : z = z3 } = {P3 } and the proof is finished.

Proof of Proposition 22. For the construction of the set K see Fig. 6. Now
suppose that the set K contained a Tk configuration (M1 , . . . , Mk ). Then
Mj+1 − Mj = (1 + κj+1 )Cj+1 − κj Cj , where the Cj are rank-one matrices.
Since there are no rank-one connections in K, the Cj must alternate between
vertical and horizontal vectors. Thus if Cj is horizontal, both Cj−1 and Cj+1
are vertical and thus the projections of Mj−1 , Mj and Mj+1 to the base plane
lie on a line. Fig. 6 shows that no three points in K have this property.

Proof of Proposition 23. If the conclusion fails then T > 0 and, perhaps
after cutting away unused parts of the curve, we can certainly find a laminate
µ which is not concentrated on a proper closed subinterval of [0, T ]. In other
words
(γ(0), 0), (γ(T ), T ) ∈ supp(µ) ⊂ K. (59)
The separation argument which is a central theme of this section now
tells that for any t ∈ (0, T ) the sets γ([0, t)) and γ((t, T ]) can not strictly lie
on two different sides of some line. Indeed, else there is an a ∈ R2 such that

γ(r), a < γ(t), a < γ(s), a for r < t < s. (60)

Then {(γ(s), s) : s ∈ [0, T ]} ⊂ Q++ ∪ Q−− ∪ Q00 , where

Q±± = {±γ(s) − γ(t), a > 0, ±(s − t) > 0}.

Thus Lemma 20 implies that supp µ is contained in R2 ×[t, T ], or in R2 ×[0, t]

or in R2 × {t}. This contradicts (59).
This shows that (60) can not occur, and we will finish our proof by
verifying that the assumption on γ anyhow enforces (60) to hold for some

37
This figure presents the six point con-
figuration with a nontrivial separately
convex hull in R2 ⊕ R. The set K con- 6
sists of {1, 2, 3, 4, 5, 6} - each point at
the corresponding height.
The picture below shows the projection 5 P 6
into the base plane - at each point it is
indicated for which height about this
4
given point we are in the separately P 1 6
convex hull of the set.
The six smaller pictures on the right
P 1
present the intersections of the hull
6
with a plane of given height 1, . . . , 6.
Here • denotes a point from the origi-
nal set, and × and ◦ denote points in 1
the hull, but not in K - the ×-points 2 P
are extreme on their vertical line, while
the ◦-points are not.
To prove that a separately convex func-
1
tion which vanishes on K can not be
positive in any of the × or ◦-points it
is enough to check that each ×-point is
not convex extreme in its correspond-
ing plane. A special feature is the oc-
curence of the auxiliar point P - this 4[4]
makes it difficult to find simple grid-
based algorithms to compute the hull.

1[1...4]

5[5]

P[2...5]
6[3...6]

2[2]

Figure 6: The new C6 -configuration

38
t. The assumptions of Proposition 23 and (60) involve γ rather than K and
remain unchanged if we (monotonously) reparametrize γ and apply some
affine map of the plane to it. Therefore, due to the hypothesis γ(0) ∈ /
co
(γ((0, T ]) we can suppose in the sequel that

• γ : [0, T ] → S1 is continuous,

• γ(0) = (0, 0)

• γ2 (t) ≥ 0 for all t and hence γ2 (0) ≥ 0.

First, we assume γ2 (t) > 0 for all t ∈ (0, T ]. Then a simple compactness
argument gives that for δ > 0 but sufficiently small

γ(t), (δγ1 (0), 1) > 0 if t ∈ (0, T ].

Because also

γ (t), (δγ1 (0), 1) > 0 for all t positive but small enough,

the mean value theorem says that for k sufficiently large

1 1
γ(t), (δγ1 (0), 1) < γ( ), (δγ1 (0), 1) < γ(s), (δγ1 (0), 1) if t < < s.
k k
So we arrived at the impossible relation (60) and are done with the first
case.
Thus, we can in addition, perhaps after a reparametrisation and a re-
flection at the x2 axis, suppose that

γ(T0 ) = (−1, 0) for some T0 ∈ (0, T ].

Hence
γ(t) ∈
/ (0, ∞) × {0} for t > 0
since otherwise γ(0) ∈ [γ(t), γ(T0 )]. If γ (0) = (1, 0) then as before we find
δ > 0 such that

γ (0), (−δ, 1) > 0 and γ(t), (−δ, 1) > 0 for all t > 0.

Again we conclude that (60) holds for a = (−δ, 1) and t > 0 but suffi-
ciently small. So we are left with the most difficult case when

• γ (0) = (1, 0).

39
Now we can moreover require that T0 was chosen such that γ(T0 ) is the
point on γ((0, T ]) ∩ (R × {0}) which is closest to 0. Then

• γ(t) ∈
/ (−1, ∞) × {0} if t ∈ (0, T ].

We use again compactness arguments to choose a few more constants. First

fix

• δ0 > 0 with γ1 > 1/2 on [0, 2δ0 ], then

• η > 0 such that for any t ∈ [δ0 , T ] we have γ2 (t) > 2η or γ1 (t) ≤ −9/10.
Finally, pick

• δ1 ∈ (0, δ0 ) with 10(diam(K) + 10)γ2 (t) < η for all t ∈ [0, δ1 ].

d
Because dt log(γ2 ) has to be unbounded from above in any neighbourhood
of the root of γ2 at 0, we find t0 ∈ (0, δ1 ) with

γ2 (t0 ) γ2 (t0 )
> . (61)
γ1 (t0 ) γ1 (t0 ) + 12

Finally, we select t1 to maximize

γ2 (t) − γ2 (t0 ) 9
ct = under the constraint γ1 (t) ≤ − .
γ1 (t) − γ1 (t0 ) 10

The key observation is now that there is an ε > 0 such that

(i) ct > ct1 + 2ε if t ∈ [0, t0 ) or γ1 (t) > γ(t0 ), and

(ii) γ2 (t) ≥ γ2 (t0 ) + ct1 (γ1 (t) − γ1 (t0 )) if t ∈ (t0 , T ] and γ1 (t) ≤ γ1 (t0 ).

It is easy to check that these two conditions give for t = t0 and a = (−ct1 −
ε, 1) the forbidden separation (60).
Now for γ1 (t) ≤ −9/10 assertion (ii) is just a reformulation of the max-
imality property of ct1 . If γ1 (t) ∈ (−9/10, γ1 (t0 )] then t > t0 implies t > δ0 .
Thus γ2 (t) > η > γ2 (t0 ). Since ct1 ≥ cT0 > 0 this proves assertion (ii).
It remains to verify (i). First we note that ct − ct1 is larger then some
positive constant for t near t0 – this is just a consequence of (61), which
gives
γ2 (t0 ) γ2 (t0 ) γ2 (t0 ) − γ2 (t1 )
> > = ct1 .
γ1 (t0 ) 1
γ1 (t0 ) + 2 γ1 (t0 ) − γ1 (t1 )
If ct = ct1 and t ∈ [0, t0 ) then γ(t) ∈ {γ(t1 ), γ(t0 )}co . Combining this
with the above estimate for t near t1 we conclude that inf t∈[0,t0 ) ct − ct1 > 0.

40
 P2+ P+
P1+ 0

g+
X0
P2 P1 P0
0
g−

X1

P2− −
P1

Figure 7: A continuous curve with small future but large hull

Similarly inf t∈(t0 ,δ0 ] ct −ct1 > 0 follows from γ(t0 ) ∈

/ {γ(t1 ), γ(t)}co for t > t0 .
The last situation to deal with is γ1 (t) > γ1 (t0 ) but t > δ0 . Then the bound
is a consequence of

γ2 (t0 )
γ2 (t0 ) + ct1 (γ1 (t) − γ1 (t0 )) ≤ γ2 (t0 ) + diam(K) < η < γ2 (t) − η.
( 12 )

Example 24

We construct an example which shows that the conclusion of Proposition 23

need not to hold for Lipschitz curves. We fix in the plane the two rays
1 1
g+ = {(x, x) : x ≥ 0} and g− = {(x, − x) : x ≥ 0},
2 2
and the point P0 = (1, 0). For λ ∈ (0, 1) we define P1 = λP0 and the
two segments s+i = [Pi Pi ], where Pi ∈ g+ satisfies (Pi − Pi )(1, 2) for
+ + +

i = 0, 1. Together with the connecting segment c0 = [P1+ P0+ ] they form

41
M0 = s+ 0 ∪ c0 ∪ s1 . We use the linear map Aλ = diag(λ, −λ) to build the
+

selfsimilar set
∞

M= Mi ∪ {0} where Mi+1 = Aλ (Mi ),
i=0

consisting of the segments s− + + −

i+1 = Aλ (si ), si+1 = Aλ (si ) and ci+1 = Aλ (ci )
−
and containing the points Pi+1 = Aλ (Pi ), Pi+1 = Aλ (Pi+ ) and Pi+1 +
=
Aλ (Pi− ), see Figure 7.
Let X1 = (P1 + P1− )/2 and denote by X0 = X0 (λ) the intersection of
the line through X1 and P2+ (λ) with the line containing s+ 0 . It is clear
that for λ → 1− the point X0 (λ) tends to P0+ , but if λ gets smaller, it will
leave s+0 trough its endpoint P0 . Hence, we can find λ0 ∈ (0, 1) such that
X0 (λ0 ) = (P0 + P0+ )/2 = A−1
λ (X1 ).
Now we can consider any injective continuous curve γ : [0, T ] :→ R2 with
−
γ([0, T ]) = M = M (λ0 ) and γ(0) = 0. Because s+ i+1 ∪ ci+1 ∪ si+1 intersects
−
the convex hull of (s+i ∪ ci ∪ si ) only in an extreme point and since
− −
i+1 ∪ ci+1 ∪ si+1 and then through si ∪ ci ∪ si , (62)
γ first runs through s+ +

it is easy to verify that γ(t) never belongs to the convex hull of γ((t, T ]). On
the other hand, K = {(γ(t), t) : t ∈ [0, T ]} has a nontrivial rank-one convex
hull. To verify this note that X0 ∈ [P0+ , X1 ] and thus

(X0 , γ −1 (P2+ )) ∈ [(P2+ , γ −1 (P2+ )), (X1 , γ −1 (P2+ ))].

Now

(X1 , γ −1 (P2+ )) ∈ [(X1 , γ −1 (P3− )), (X1 , γ −1 (X1 ))]

which implies that

(X0 , γ −1 (P2+ )) ∈ (K ∪ {(X1 , γ −1 (P3− ))})rc .

Using the (Aλ )-selfsimilarity of M to iterate this we have

(−1)i
(X0 , γ −1 (P2+ )) ∈ (K ∪ {(Xi , γ −1 (Pi+2 ))})rc for all i.
(−1)i
Since limi→∞ (Xi , γ −1 (Pi+2 )) = (0, 0) ∈ K and any separately convex func-
tion is continuous, we get as required

(X0 , γ −1 (P2+ )) ∈ K rc \ K.

42
6.3 Separate convexity in R2 ⊕ R2
We consider the integrand
1
φ(X) = |X|2 + h(det X), X ∈ R2×2 , (63)
2
the corresponding Euler-Lagrange equation div Dφ(∇w) = 0 and its refor-
mulation as a first order partial differential relation
 
X
∇u ∈ Kφ = : X ∈ R2×2 . (64)
Dφ(X)J

Theorem 25 Suppose h : R → R is differentiable and strictly convex. Sup-

pose further that E is a finite subset of K and that the elements of E are of
the form X Y , where X is a diagonal 2 × 2 matrix. Then K
rc = K and the

laminates M (K) supported on K are Dirac masses.

rc

Proof. Geometrically the heart of the matter is again a good foliation

as in Section 5. After some transformation we will see that E lies in a
certain two-dimensional set Γ in some R2 ⊕ R2 , where the R2 ’s correspond
to the rank-one directions. The point is that Γ can be foliated by a one-
parameter family of curves whose projection to either R2 is a foliation (of
one quadrant) by strictly convex curves (see Step 4 below; the different
curves are distinguished by d = st). Then the separation result, Lemma 20,
in connection with the finiteness of E can be used once more to finish the
argument. First we need to make some normalizations and to show that E
lies in one quadrant.
Step 1: Reduction to h (0) = 0. Adding a linear term to h does not
affect the Euler-Lagrange equation since det is a null Lagrangian, i.e. for
all function w we have div(D det)∇w = div cof ∇w = 0. On the level of the
partial differential relation we have

 
X
Kφ+λ det = : X ∈ R2×2
Dφ(X)J + λ cof X J
   
X X
= : ∈ Kφ .
Y + λJX Y

Since X X
Y → Y +λJX is a linear isomorphism which preserves rank-one lines
we may suppose h (0) = 0.

43
 Step 2: Special separately convex functions. We now write out the
X
elements Y of Kφ for which X is a diagonal matrix more explicitly
⎛ ⎞
    s 0
X X ⎜ 0 t ⎟
= =⎜⎝
⎟.
⎠
Dφ(X)J X + h (det X) cof X J 0 −s − h (st)t
t + h (st)s 0

After exchanging rows and multiplying one row by −1 (both of which cor-
respond to making a linear change of the dependent variable u in (64)) we
may suppose that
⎧⎛ ⎞ ⎫
⎪
⎪ s 0 ⎪
⎪
⎨⎜ ⎟ ⎬
⎜ σ 0 ⎟
E⊂ ⎝ : s, t ∈ R, σ = t + h (st)s, τ = s + h (st)t . (65)
⎪
⎪ 0 t ⎠ ⎪
⎪
⎩ ⎭
0 τ

On the linear subspace

L = {A ∈ R4×2 : A12 = A22 = A31 = A41 = 0},

which contains E, rank-one convexity agrees with separate convexity in the

variables (A11 , A21 ) and (A32 , A42 ).
In particular, for each a, b ∈ R2 we have the separately convex functions

(A11 , A21 , A32 , A42 ) → (A11 , A21 ), a± (A32 , A42 ), b±

at our disposal.
Step 3: Restriction to generalized quadrants. We argue by contradic-
tion. If the claim fails then there is a minimal set E0 ⊂ E which supports a
nontrivial laminate µ0 . Our goal is to contradict minimality by a separation
argument similar to [Sv 92b, Ta 93]. We denote by P0 = (s0 , σ0 , t0 , τ0 ) the
centre of mass of µ0 .
Recall that strict convexity and the normalization condition h (0) = 0
imply that h (x)x > 0 for x = 0. Thus all (s, σ, t, τ ) ∈ E satisfy

|τ | ≥ |s|, τ s ≥ 0 and s = 0 ⇔ τ = 0. (66)

Thus we can apply Lemma 20 to the generalized quadrants

Q±± = {(s, σ, t, τ ) : ±s > 0, ±τ > 0}.

Since E0 is finite this shows that either s0 > 0, τ0 > 0 and s > 0, τ > 0
in E0 , or s0 < 0, τ0 < 0 and s < 0, τ < 0 on E0 , or s0 = τ0 = 0 and

44
s = τ = 0 on E0 . In the last case we have σ = t on E0 and therefore
the quadratic minor σt=Aˆ 21 A32 − A22 A31 has a (strictly) positive value at
all non vanishing differences of elements in E0 . This implies that even all
polyconvex measures supported on E0 must be Dirac masses [Sv 93] and thus
yields a contradiction. Hence we have shown that there exists a cs ∈ {−1, 1}
such that
cs s0 , cs τ0 > 0, cs s, cs τ > 0 on E0 .
Since the problem is invariant under the exchange of variables (s, τ ) ↔
(t, σ) we also find ct ∈ {−1, 1} such that
ct t0 , ct σ0 > 0 ct t, ct σ > 0 on E0 .
Multiplication of all variables by −1 leaves the right hand side of (65) in-
variant. Hence we may assume cs = 1. If we multiply only t and σ by −1
and replace h(x) by its reflection h(−x) then (65) remains invariant. Hence
we may suppose cs = ct = 1, i.e.
s0 , σ0 , t0 , τ0 > 0, s, σ, t, τ > 0 on E0 . (67)
Step 4: Separation. Let d = min{st : (s, σ, t, τ ) ∈ E0 }. Since E0 is
finite we have d > 0. Pick any point P1 ∈ E0 with s1 t1 = d. Projecting E0
into the (s, σ) and (t, τ ) planes we see that the set
E0s := {(s, σ) : (s, σ, t, τ ) ∈ E0 }
is contained in the (closed) epigraph of the strictly convex function
d
s →
+ h (d)s,
s
because for fixed s > 0 the expression on the right hand side is increasing
in d > 0. Since (s1 , σ1 ) lies on the graph of this strictly convex function it
is an extreme point of E0s . Thus there exists as ∈ R2 with
(s1 , σ1 ), as  < (s, σ), as  for (s, σ, t, τ ) ∈ E0 \ {P1 }. (68)
Here we used that fact that the projection from E0 to E0s is injective since
Kφ contains no rank-one connections (this in turn follows from the strict
convexity of h). In view of the invariance under the exchange of variables
(s, σ) ↔ (t, τ ) the same reasoning yields at ∈ R2 with
(t1 , τ1 ), at  < (t, τ ), at  for (s, σ, t, τ ) ∈ E0 \ {P1 }. (69)
Invoking once more Lemma 20 we deduce that µ0 is either supported on
{P1 } or on E0 \ {P1 }. This contradicts the minimality of µ0 and the proof
is finished.

45
6.4 A laminate without discrete approximations
One might hope to extend Theorem 25 to more general sets through an
approximation by finite sets. The following example shows, however, that
the rank-one convex hull can shrink drastically, when we pass from continua
to discrete sets. It was motivated by J.M. Ball’s construction of sets of
gradients with no rank-one connections ([Ba 90]) and is based on separate
convexity in R ⊕ R ⊕ R, which arises by restricting rank-one convexity in
R3×3 to diagonal matrices.

Proposition 26 We consider separate convexity in R ⊕ R ⊕ R and define

for t ∈ [0, 1]
⎛ ⎞
t
p(t) = ⎝ t ⎠ , pi (t) = p(t) + tei for i = 1, 2, 3.
t
Let Si = pi ([0, 1]). Then p(t) ∈ (S1 ∪ S2 ∪ S3 )rc for t ∈ [0, 1] and there
is only one laminate supported in S1 ∪ S2 ∪ S3 which generates this point.
Moreover, this laminate does not charge points. More precisely,
{µ ∈ Mrc (S1 ∪ S2 ∪ S3 ) : µ̄ = p(t)} = {µt },
where µ0 = δ0 and where for t > 0
 3
1 t

µt (A) = 3 χA (pi (τ ))τ 2 dτ if A ⊂ R ⊕ R ⊕ R.
t 0
i=1

In particular, if K is a compact subset of (S1 ∪ S2 ∪ S3 ) but pi (t0 ) ∈

/ K, then
/ K for t ∈ (t0 , 1].
p(t) ∈ rc

Proof of Proposition 26. In the sequel we identify R ⊕ R ⊕ R with R3

and suppose the measures µt are defined as above. First we show that for all
t0 ∈ [0, 1] and all f : R3 → R separately convex the Jensen-type inequality

f (p(t0 )) ≤ f dµt0 (70)

holds. Choosing for f constant and linear functions we can than conclude
that µt0 is a probability measure, µ̄t0 = p(t0 ) and µt0 ∈ Mrc (S1 ∪ S2 ∪ S3 ).
To establish (70) we notice that for any f : R3 → R separately convex,
for x ∈ R3 and t1 , t2 , t3 > 0 we have

f (x + p(t)) − f (x)

3
f (x + ti ei ) − f (x)
lim sup ≤ si , where si = . (71)
t→0+ t ti
i=1

46
Indeed, after addition of an affine function and translation we may sup-
pose that x = 0 = (s1 , s2 , s3 ). Suppose now that the upper limit of these
difference quotients was positive. Since separately convex functions are lo-
cally Lipschitz, a suitable subsequence of the rescaled functions fε (y) =
ε−1 (f (εy) − f (0)) converges (uniformly on compact subsets) to a limit f0
with f0 (p(1)) > 0. Moreover f0 is globally Lipschitz, separately convex and
satisfies f0 (tei ) ≤ 0 = f0 (0) for t > 0 and i = 1, 2, 3. This implies that for
any y ∈ R3 and i the function t → f0 (y + tei ) is nonincreasing. Otherwise
t → f0 (y + tei ) is bounded from below by a function with positive slope and
f0 (y + tei ) − f0 (tei ) tends to infinity as t → +∞, in contradiction with the
Lipschitz property of f0 . Using that t → f0 (y + tei ) is nonincreasing we
obtain
0 < f0 (p(1)) ≤ f0 (e1 + e2 ) ≤ f0 (e1 ) ≤ f0 (0) = 0,
a contradiction proving (71).
Since t → f (p(t)) is Lipschitz it is differentiable for almost every t ∈ [0, 1]
and (71) implies that

d 1
f (p(t)) ≤ (f (p1 (t)) + f (p2 (t)) + f (p3 (t)) − 3f (p(t))).
dt t
This is equivalent to
d
3f (p(t)) + t f (p(t)) ≤ f (p1 (t)) + f (p2 (t)) + f (p3 (t)),
dt
and therefore
 t0

3 
t30 f (p(t0 )) ≤ 2
f (pi (t))t dt = f d(t30 µt0 )
0 i=1

follows and ensures (70).

The uniqueness of µt0 requires a bit more effort, so assume there is an
other µ ∈ Mrc (S1 ∪ S2 ∪ S3 ) with µ̄ = p(t0 ). Since these two measures
differ, we find some
 
f : S1 ∪ S2 ∪ S3 → R such that f dµt0 > f dµ

and, after adding a constant and making a suitable approximation if neces-

sary, we also may assume that for i = 1, 2, 3

t → f (pi (t)) is C ∞ (R) with support in (η0 , 2] for some η0 > 0.

47
 Now, for ε > 0 let gε : R → R be the solution of
 1
gε (t) = t (f (p1 (t)) + f (p2 (t)) + f (p3 (t)) − 3gε (t)) − ε if t > η0 ,
− 4ε t if t ≤ η0 ,
∞
with gε (0) = 0. It is easily checked that gε is also C . Since limε→0+ gε (t) =
f dµt for t ∈ [0, 1] we can choose ε = ε0 > 0 such that

g = gε0 satisfies f dµ < g(t0 ).

For convenience we also introduce the smooth functions

 1
si (t) = t (f (pi (t)) − g(t)) if t > η0 ,
, i = 1, 2, 3,
ε
4 if t ≤ η0
and fix a finite constant c0 such that


3
10 |si (t)| + |si (t)| < c0 if t ∈ [−4, 4].
i=1

Finally, we define the auxiliary function


f1 (x) = (xj − xi ) if x ∈ R3 and xi = min xj ,
j
j∈{1,2,3}\{i}

and the function

⎛ ⎞
s1 (t)
F (x) = g(t) + x − p(t), ⎝ s2 (t) ⎠ + c0 f1 (x) for x ∈ R3 and t = min xj .
j
s3 (t)
Suppose we had already shown that F is separately convex in the open
cube (−4, 4)3 of R3 . Then Proposition 12 combined with Proposition 13
imply that F satisfies the Jensen inequality for all laminates supported in
(−2, 2)3 . Because F (pi (t)) = g(t) + tsi (t) = f (pi (t)) and F (p(t)) = g(t) for
t ∈ [0, 1], we obtain the contradiction
 
F dµ = f dµ < g(t0 ) = F (p(t0 )) = F (µ̄ )

finishing our proof.

Therefore, it remains to study separate convexity of F near an arbitrary
point x0 ∈ (−4, 4)3 . Essentially only x0 ∈ (0, 4)3 is interesting but the
general case is not more complicated. So, let t0 = minj x0j and we consider

h(r) = f (x0 + rei ) for i ∈ {1, 2, 3} fixed and r near 0.

48
 First, suppose t0 < x0i . Hence minj (x0 + rej ) ≡ t0 , and
h(r) = (x0i + r − t0 )si (t0 ) + c0 f1 (x0 + rei ) + const
is obviously convex in r.
Next, let t0 = x0i < minj =i x0j . Then minj (x0 + rei ) = t0 + r and


h(r) = g(t0 + r) + (x0j − t − r)sj (t0 + r) + c0 (x0j − t0 − r).
j =i j =i

Since g (t) = s1 (t) + s2 (t) + s3 (t) − ε0 , we have

h (r) = si (t0 + r) − ε0 + (x0j − t0 − r)sj (t0 + r) + c0 (2r + 2t0 − x0j )
j =i j =i

and


h (0) = si (t0 ) + −sj (t) + (x0j − t0 )sj (t0 ) + 2c0 ≥ 2c0 − c0 > 0.
j =i

So we are left with the last case when t0 = x0i = x0j ≤ x0k where {i, j, k} =
{1, 2, 3}. For r < 0 we compute h and h as before and obtain
lim h (r) = si (t0 ) − ε0 + (x0k − t0 )(si (t0 + r) − c0 ).
r→0−

For r > 0 we have

h(r) = g(t0 ) + (x0 + rei ) − p(t0 ), s(t0 ) + r(x0k − t0 )
and thus
lim h (r) = si (t0 ) + (x0k − t0 ) ≥ lim h (r) + ε0
r→0+ r→0−

and again h has a (local) subdifferential at zero. This finishes the proof.

7 Local hulls, degenerate sets and hyperbolic con-

servation laws
In this section we formulate some ‘local’ problems for the various hulls K ∗
and the classes of measures M∗ associated with a given set K ⊂ Rm×n .
As we sketch below, such problems are relevant for example in connection
with compactness properties of L∞ entropy solutions of l × 2 systems of
hyperbolic conservation laws.
Let us consider a smooth submanifold K ⊂ Rm×n . We introduce the
following properties of K.

49
(P1) Each point A ∈ K has a neighbourhood U ⊂ Rm×n such that (K ∩
Ū )rc = K ∩ Ū .

(P2) Each point A ∈ K has a neighbourhood U ⊂ Rm×n such that Mrc (K∩
Ū ) is trivial.

(P3) Each point A ∈ K has a neighbourhood U ⊂ Rm×n such that (K ∩

Ū )pc = K ∩ Ū .

(P4) Each point A ∈ K has a neighbourhood U ⊂ Rm×n such that Mpc (K∩
Ū ) is trivial.

A sufficient condition for (P1) and (P2) to be satisfied is that the tangent
spaces TA K do not contain any rank-one connections, see e.g. [Ta 83] (this
condition in fact implies the stronger assertion obtained by replacing ‘rc’
with ‘qc’). The same condition is also sufficient for (P3) and (P4) when
n = 2. This follows from the fact that a rank-one convex quadratic form on
Rm×n is polyconvex, see [Se 83].
If TA K does contain rank-one connections, the situation is more com-
plicated. For simplicity, assume that for some A0 ∈ K the rank-one cone
Λ and TA0 K intersect transversely (away from the origin). Then TA K ∩ Λ
behaves ‘well’ for A close to A0 , and one might speculate that there is some
natural ‘higher order’ condition which would imply (P1)–(P4), or at least
some of these properties, in a neighbourhood of A0 . This situation arises
in connection with compactness properties of L∞ entropy solutions of hy-
perbolic conservation laws, as was pointed out already in DiPerna’s work
[DP 85], where some very interesting examples are considered and shown to
have property (P4).
It seems that the problem of determining whether (P1)–(P4) are satisfied
remains open even in some simple and very natural situations. Consider for
example the following problem taken from the above paper by DiPerna. We
look at the one dimensional wave equation

ϕtt − a(ϕx )x = 0, (72)

where a is strictly convex, increasing and satisfies a(0) = 0. There is a
natural energy associated with (72):
 ξ
1
η(ϕt , ϕx ) = ϕ2t + F (ϕx ), where F (ξ) = a(s) ds.
2 0

50
One has ηt − qx = 0 for each regular solution of (72), where q = q(ϕt , ϕx ) =
ϕt a(ϕx ). Letting u = ϕt , v = ϕx we write (72) as a first order system

ut − a(v)x = 0,
(73)
vt − ux = 0.

L∞ entropy solutions of (73) can be defined as L∞ functions (u, v) satisfying

ut − a(v)x = 0,
vt − ux = 0, (74)
ηt − qx ≤ 0

in the sense of distributions. One can also add further entropies to (74),
to get a more restrictive class of solutions, see [DP 85]. In view of Murat’s
lemma [Mu 81] it is reasonable - at least in a first approximation - to replace
the inequality in (74) by an equality when studying compactness properties.
This enables us to introduce stream functions as follows.
(u, −a(v)) = (ψx1 , −ψt1 ),
(v, −u) = (ψx2 , −ψt2 ),
(η, −q) = (ψx3 , −ψt3 ).

This can be rewritten as

∇ψ ∈ K ⊂ R3×2 ,

with
⎧⎛ ⎞ ⎫
⎨ u a(v) ⎬
K= ⎝ v u ⎠ : u, v ∈ R .
⎩ ⎭
η(u, v) q(u, v)

As far as we know it is an open problem to determine for which functions a

the set K ⊂ R3×2 given above satisfies any of the conditions (P1)–(P4). We
refer the reader to the paper of DiPerna [DP 85] for a study of a situation
with additional entropies, in which one can determine that (P4) holds.

8 Outlook
Here we formulate some further questions and briefly mention a few related
results.

51
Question 1 Is it possible to characterize (or at least give some non-trivial
examples of ) smooth uniformly rank-one convex functions φ : R2×2 → R for
which Kφpc = Kφ ?

Perhaps we can expect that weak solutions of the Euler-Lagrange equa-

tion div Dφ(∇u) = 0 for φ with Kφpc = Kφ have some extra regularity
properties.
Characterizing φ with Kφqc = Kφ would be even more interesting but
that looks out of reach.

Question 2 Is it possible to identify some geometric properties of smooth

“elliptic” sets K ⊂ Rm×2 which imply a-priori estimates (for example C α -
estimates) for smooth solutions of ∇u ∈ K?

By a smooth elliptic set we mean a set set K which is a smooth sub-

manifold of Rm×2 such that the tangent spaces TA K contain no rank-one
connections.
While rank-one connections, T4 -configurations or, in general, non-trivial
measures in Mrc are obstructions to regularity of Lipschitz solutions of
∇u ∈ K, their rôle in connection with a-priori estimates for smooth solutions
is less clear. Of course, they do become relevant even in this setting if they
occur “infinitesimally”, but that is not possible for “elliptic” sets.

Question 3 Is it possible to develop some general methods which would

work for the local problems in Section 7?

The construction of examples by convex integration rests on the subtle

combination of one-dimensional constructions (and hence is closely related
to rank-one convexity).

Question 4 Can one extend convex integration using genuinely multi-di-

mensional constructions as building blocks ?

This would allow one to obtain interesting examples under the weaker (and
more natural) condition that the quasiconvex hull is sufficiently trivial. The
main difficulty is that it is much harder to ‘glue’ multidimensional construc-
tions. One interesting test case is an example by Šverák (see, e.g., [Mu 99],
Section 4.7, equation (4.25)) of a set K ⊂ R6×2 for which ∇u ∈ K admits
periodic solutions. It is not known whether solutions with compact support
exist. For possible extensions of convex integration see also [Gr 86], Section
2.4.12.

52
Question 5 Is there an effective algorithm to decide whether a given prob-
ability measure supported on a finite subset of R2×2 is a laminate?

Such an algorithm would enable us to effectively test (at least in a prob-

ability sense) whether rank-one convexity implies quasi-convexity on 2 × 2
matrices. For example, one would generate (by some random or pseudo-
random process) on a computer piecewise affine maps u : T 2 → R2 and
check whether the corresponding distribution µ of their gradients given by

1
f, µ = 2 f (∇u(x)) dx
|T | T 2

are laminates. At present no effective algorithm for this is known. For

the related question to determine numerically the rank-one convex hull of a
function, see [Do 02] and the references therein.
Many simple geometric questions about rank-one convexity are open,
even in low dimensions. In view of applications the case of 2 × 2 matrices is
particular interesting.

Question 6 Does each Tk -configuration in R2×2 contain some T4 -configu-

ration? More generally, does every compact set K ⊂ R2×2 with nontrivial
rank-one convex hull contain a T4 -configuration ?

This is even open for symmetric 2×2 matrices. The answer to both questions
is yes for diagonal 2 × 2 matrices (see Proposition 19). The answer to the
second question is no for 3 × 2 matrices (see Proposition 22). Székelyhidi
[Sz 02] constructs a T5 -configuration in R4×2 which does not contain a T4 -
configuration.
Many examples arise from nonlinear elasticity and thus have at least an
SO(n) symmetry.

Question 7 How can exploit symmetries efficiently ?

A nice example is the computation of the rank-one convex and polycon-

vex hull for an energy function which describes nematic elastomers [DD 02,
Si 02]. The role of discrete symmetries which arise in crystalline materials
is also important [Fr], but largely unexplored.

Question 8 Is there an efficient algorithm to decide if K rc = K ? Can one

at least efficiently check whether K contains a Tk configuration ?

53
A typical ingredient in efficient algorithms is the use of extreme points in
connection with the Krein-Milman theorem. With a sufficiently abstract
(dual) definition of extreme points the Krein-Milman theory holds in our
setting [Al 71, Kr 00, Zh 98]. The question is whether one can obtain a
more geometric characterization of extreme points. For a nearly optimal
result for 2 × 2 matrices see [Ki 01b], Thm. 4.20.

Question 9 Can one combine Theorems 1 and 3 ? In other words, is there

an elliptic system (with a quasiconvex or polyconvex energy function) which
admits a solution whose gradient takes only finitely many values ?

This might be easier if one goes beyond 2 × 2 systems. In this case v = X

Y ,
with X, Y ∈ Rn×n and A(Dv) = curl X
div Y .

Question 10 Is the quasiconvex hull of the set K related to φ(X) = (det X)2
trivial ?

This is related to the compactness and regularity properties of solutions to

the corresponding (degenerate) Euler-Lagrange equation

div[(det ∇w) cof ∇w] = 0.

Since div cof ∇w = 0 one can argue formally that all solutions must satisfy
det ∇w = const and this argument is correct for C 1 solutions. If one applies
the formal argument to the ‘linearized’ problem

div[f (x) cof ∇w] = 0

one would similarly conclude that f = const, at least if det ∇w = 0. There

exist, however, a Lipschitz map w : R2 → R2 with det ∇w ∈ {−1, 1} a.e. and
a non constant f with values in {−1, 1} which satisfy the above equation.
In this paper we have mainly studied sets, but the study of rank-one
convex functions is equally interesting. Many seemingly simple questions
are open.

Question 11 Can one characterize the traces of separately convex functions

f : R ⊕ R → R on the diagonal ?

Tartar showed that any smooth function can arise as a trace, see [Ta 93],
Remark 11 there. Motivated by results due to Šverák and Preiss mentioned
in [Ta 93], it was shown that C 1 is a necessary and C 1,α a sufficient condition
for being a trace. The lack of a precise characterization seems related to our

54
partial understanding of rank-one convexification procedures. The situation
is even less clear if we go into 3 or more dimensions.
Finally, throughout this work we have focused on oscillations effects only
and restricted attention to bounded sets K. If one drops this assumptions
one also needs to study the possible interaction of oscillation and concen-
tration effects and new tools are required, see e.g. [Ta 90]. More on the
technical side one can ask to which differential operators A one can extend
the general theory (for questions of compactness, A gradient Young mea-
sures, and relaxation see [FM 99, BFL 00]).

Acknowledgements
We thank Laszlo Székelyhidi for many interesting discussions, in particu-
lar on T5 configurations and on the laminate without discrete approxima-
tions. S.M. and V.S. wish to thank the members SFB 256 and the DFG for
their longstanding support and encouragement and for providing an atmo-
sphere of open and inspiring exchange. Special thanks go to W. Ballmann,
U. Hamenstädt and J. Lohkamp who organised a seminar on convex inte-
gration during V.S.’s stay in Bonn in 1992/93. S.M and V.S. were also
supported by a Max Planck Research award, and V.S. was supported by the
NSF grant DMS 9877055. B.K and S.M. would like to thank the IMA and
the Department of Mathematics at the University of Minnesota for their
hospitality during numerous visits.

References
[Al 71] E.M. Alfsen, Compact convex sets and boundary integrals,
Springer, 1971.
[AF 02] K. Astala and D. Faraco, Quasiregular mappings and Gradient
Young measures, to appear in Proc. Roy. Soc. Edinburgh.
[AH 86] R. Aumann and S. Hart, Bi-convexity and bi-martingales, Israel
J. Math. 54 (1986), 159–180.
[Ba 80] J.M. Ball, Strict convexity, strong ellipticity and regularity in
the calculus of variations, Math. Proc. Cambridge Phil. Soc. 87
(1980), 501–513.
[Ba 90] J.M. Ball, Sets of gradients with no rank-one connections, J.
Math. Pures Appl. 69 (1990), 241–259.

55
[BJ 87] J.M. Ball and R.D. James, Fine phase mixtures as minimizers of
energy, Arch. Rat. Mech. Anal. 100 (1987), 13–52.

[BDG 69] E. Bombieri, E. De Giorgi and E. Giusti, Minimal cones and the
Bernstein theorem, Invent. Math. 7 (1969), 243–269.

[BFL 00] A. Braides, I. Fonseca and G. Leoni, A-quasiconvexity: relaxation

and homogenization, ESAIM Control Optim. Calc. Var. 5 (2000),
539–577.

[CT 93] E. Casadio-Tarabusi, An algebraic characterization of quasi-

convex functions, Ricerche Mat. 42 (1993), 11–24.

[Ce 80] A. Cellina, On the differential inclusion x ∈ [−1, 1], Atti Accad.
Naz. Lincei Rend. Sci. Fis. Mat. Nat. 69, 1–6.

[CK 88] M. Chipot and D. Kinderlehrer, Equilibrium configurations of

crystals, Arch. Rat. Mech. Anal. 103 (1988), 237–277.

[CK 00] M. Chlebik and B. Kirchheim, Rigidity for the four gradient prob-
lem, Preprint MPI-MIS 35/2000.

[Da 89] B. Dacorogna, Direct methods of the calculus of variations,

Springer, 1989.

[DM 97] B. Dacorogna and P. Marcellini, General existence theorems for

Hamilton-Jacobi equations in the scalar and vectorial cases, Acta
Math. 178 (1997), 1–37.

[DM 99] B. Dacorogna and P. Marcellini, Implicit partial differential equa-

tions, Birkhäuser, 1999.

[Do 02] G. Dolzmann, Variational Methods for Crystalline Microstruc-

ture - Analysis and Computation, to appear as Lecture Notes in
Mathematics, Springer.

[DP 82] F.S. De Blasi and G. Pianigiani, Baire category approach to the
existence of solutions of multivalued differential equations in Ba-
nach spaces, Funkcial. Ekvac. 25 (1982), 153–162.

[DP 91] F.S. De Blasi and G. Pianigiani, Non convex valued differential
inclusions in Banach spaces, J. Math. Anal. Appl. 157 (1991),
469–494.

56
[DG 68] E. De Giorgi, Un esempio di estremali discontinue per un prob-
lema variazionale di tipo ellittico, Boll. U.M.I. 4 (1968), 135–137.

[DD 02] A. DeSimone and G. Dolzmann, Macroscopic response of nematic

elastomers via relaxation of a class of SO(3) invariant energies,
Arch. Rat. Mech. Anal. 161 (2002), 181–204.

[DP 85] R.J. DiPerna, Compensated compactness and general systems of

conservation laws, Trans. Amer. Math. Soc. 292 (1985), 383–420.

[Ev 86] L.C. Evans, Quasiconvexity and partial regularity in the calculus
of variations, Arch. Rat. Mech. Anal. 95 (1986), 227–252.

[FM 99] I. Fonseca and S. Müller, A-quasiconvexity, lower semicontinuity,

and Young measures, SIAM J. Math. Anal. 30 (1999), 1355–1390.

[Fr] G. Friesecke, personal communication.

[Fu 87] M. Fuchs, Regularity theorems for nonlinear systems of partial

differential equations under natural ellipticity conditions, Analy-
sis 7 (1987), 83–93.

[GM 68] E. Giusti and M. Miranda, Un esempio di soluzioni discontinue

per un problema di minimo relativo ad un integrale regolare del
calcolo delle variazioni, Boll. U.M.I 2 (1968), 1–8.

[Gr 73] M. Gromov, Convex integration of differential relations, Izw.

Akad. Nauk S.S.S.R 37 (1973), 329–343.

[Gr 86] M. Gromov, Partial differential relations, Springer, 1986.

[Ha 95] C. Hamburger, Quasimonotonicity, regularity and duality for non-

linear systems of partial differential equations, Ann. Mat. Pure
Appl. Ser. IV 169 (1995), 321–354.

[HLN 96] W. Hao, S. Leonardi, J. Nečas, An example of irregular solution

to a nonlinear Euler-Lagrange elliptic system with real analytic
coefficients, Ann. SNS Pisa Ser. IV 23 (1996), 57–67.

[KP 91] D. Kinderlehrer and P. Pedregal, Characterizations of Young mea-

sures generated by gradients, Arch. Rat. Mech. Anal. 115 (1991),
329–365.

57
[Ki 01a] B. Kirchheim, Deformations with finitely many gradients and sta-
bility of quasiconvex hulls, C. R. Acad. Sci. Paris Sr. I Math. 332
(2001), 289–294.

[Ki 01b] B. Kirchheim, Habilitation thesis, University Leipzig, 2001.

[Ko 01] J. Kolář, Non-compact lamination convex hulls, Preprint MPI-

MIS 17/2001.

[Kr 99] J. Kristensen, On the non-locality of quasiconvexity, Ann. Inst.

H. Poincaré Anal. Non Linéaire 16 (1999), 1–13.

[KT 01] J. Kristensen and A. Taheri, Partial regularity of strong local min-
imizers in the multi-dimensional calculus of variations, Preprint
MPI-MIS 59/2001.

[Kr 00] M. Kružı́k, Bauer’s maximum principle and hulls of sets, Calc.
Var. 11 (2000), 321–332.

[Ku 55] N.H. Kuiper, On C 1 isometric embeddings, I., Nederl. Akad.

Wetensch. Proc. A 58 (1955), 545–556.

[Ma 01] J. Matoušek, On directional convexity, Discrete Comput. Geom.

25 (2001), 389–403.

[MP 98] J. Matoušek and P. Plecháč, On functional separately convex

hulls, Discrete Comput. Geom. 19 (1998), 105–130.

[MS 96] S. Müller and V. Šverák, Attainment results for the two-well prob-
lem by convex integration, Geometric analysis and the calculus of
variations, (J. Jost, ed.), International Press, 1996, 239–251.

[Mu 99] S. Müller, Variational models for microstructure and phase tran-
sitions, in: Calculus of Variations and Geometric Evolution Prob-
lems, Cetaro 1996 (F. Bethuel, G. Huisken, S. Müller, K. Steffen,
S. Hildebrandt, and M. Struwe, eds.), Springer, Berlin, 1999.

[Mu 99b] S. Müller, Rank-one convexity implies quasiconvexity on diagonal

matrices, Int. Math. Research Not. 1999, 1087–1095.

[MS 98] S. Müller and V. Šverák, Unexpected solutions of first and sec-
ond order partial differential equations, Documenta Math., Spe-
cial volume Proc. ICM 1998, Vol. II, 691–702.

58
[MS 99] S. Müller and V. Šverák, Convex integration for Lipschitz map-
pings and counterexamples to regularity, Preprint MPI-MIS
26/1999.

[MRS 02] S. Müller, M.O. Rieger and V. Šverák, Parabolic systems with
nowhere smooth solutions, Preprint.

[MSy 01] S. Müller and M. Sychev, Optimal existence theorems for non-
homogeneous differential inclusions, J. Funct. Anal. 181 (2001),
447–475.

[Mu 81] F. Murat, L’injection du cone positif de H −1 dans W −1,q est

compacte pour tout q < 2, J. Math. Pures Appl. 60 (1981), 309–
322.

[Na 54] J. Nash, C 1 isometric embeddings, Ann. Math. 60 (1954), 383–

396.

[NM 91] V. Nesi and G.W. Milton, Polycrystalline configurations that

maximize electrical resistivity, J. Mech. Phys. Solids 39 (1991),
525–542.

[Pe 93] P. Pedregal, Laminates and microstructure, Europ. J. Appl. Math.

4 (1993), 121–149.

[Pe 97] P. Pedregal, Parametrized measures and variational principles,

Birkhäuser, 1997.

[Sch 74] V. Scheffer, Regularity and irregularity of solutions to nonlinear

second order elliptic systems of partial differential equations and
inequalities, Dissertation, Princeton University, 1974.

[Se 83] D. Serre, Formes quadratiques et calcul des variations, J. Math.

Pures Appl. 62 (1983), 177–196.

[Si 02] M. Šilhavý, Relaxation in a class of SO(n)-invariant energies re-

lated to nematic elastomers, Preprint.

[Sp 98] D. Spring, Convex integration theory, Birkhäuser, 1998.

[Sv 90] V. Šverák, Examples of rank-one convex functions, Proc. Roy.

Soc. Edinburgh A 114 (1990), 237–242.

[Sv 92a] V. Šverák, Rank-one convexity does not imply quasiconvexity,

Proc. Roy. Soc. Edinburgh 120 (1992), 185–189.

59
[Sv 92b] V. Šverák, New examples of quasiconvex functions, Arch. Rat.
Mech. Anal. 119 (1992), 293–300.

[Sv 93] V. Šverák, On Tartar’s conjecture, Ann. Inst. H. Poincaré Anal.

Non Linéaire 10 (1993), 405–412.

[Sv 95] V. Šverák, Lower semicontinuity of variational integrals and com-

pensated compactness, in: Proc. ICM 1994 (S.D. Chatterji, ed.),
vol. 2, Birkhäuser, 1995, 1153–1158.

[SY 00] V. Šverák and Xiadong Yan, A singular minimizer of a smooth

strongly convex functional in three dimensions, Calc. Var. 10
(2000), 213–221.

[Sy 99] M.A. Sychev, A new approach to Young measure theory, relax-
ation and convergence in energy, Ann. Inst. H. Poincaré Anal.
Non Linéaire 16 (1999), 773–812.

[Sy 01] M.A. Sychev, Comparing two methods of resolving homogeneous

differential inclusions, Calc. Var. 13 (2001), 213–229.

[Sz 02] L. Székelyhidi, PhD thesis, in preparation.

[Ta 79] L. Tartar, Compensated compactness and partial differential

equations, in: Nonlinear Analysis and Mechanics: Heriot-Watt
Symposium Vol. IV, (R. Knops, ed.), Pitman, 1979, 136–212.

[Ta 83] L. Tartar, The compensated compactness method applied to sys-

tems of conservations laws, in: Systems of Nonlinear Partial Dif-
ferential Equations, (J.M. Ball, ed.), NATO ASI Series, Vol. C111,
Reidel, 1983, 263–285.

[Ta 90] L. Tartar, H-measures, a new approach for studying homogeni-

sation, oscillations and concentration effects in partial differential
equations, Proc. Roy. Soc. Edinburgh Sect. A 115 (1990), 193–
230.

[Ta 93] L. Tartar, Some remarks on separately convex functions, in: Mi-
crostructure and phase transitions, IMA Vol. Math. Appl. 54,
(D. Kinderlehrer, R.D. James, M. Luskin and J.L. Ericksen, eds.),
Springer, 1993, 191–204.

[Zh 88] Kewei Zhang, On the Dirichlet problem for a class of quasilinear
elliptic systems of partial differential equations in divergence form,

60
 in: Partial differential equations (Tianjun, 1986), Lecture Notes
in Mathematics 1306, Springer, 1988.

[Zh 98] Kewei Zhang, On the structure of quasiconvex hulls, Ann. Inst.
H. Poincaré Analyse Non Linéaire 15 (1998), 663–686.

61