HIDDEN CONVEXITY AND DAFERMOS’ PRINCIPLE FOR SOME

DISPERSIVE EQUATIONS

DMITRY VOROTNIKOV
arXiv:2501.05389v1 [math.AP] 9 Jan 2025

Abstract. We discover an abstract structure behind several nonlinear dispersive equations (in-
cluding the NLS, NLKG and GKdV equations with generic defocusing power-law nonlinearities)
that is reminiscent of hyperbolic conservation laws. The underlying abstract problem admits an
“entropy” that is formally conserved. The entropy is determined by a strictly convex function
that naturally generates an anisotropic Orlicz space. For such problems, we introduce the dual
matrix-valued variational formulation in the spirit of Y. Brenier. Comm. Math. Phys. (2018)
364(2) 579-605 and prove several results related to existence, uniqueness and consistency on
large time intervals. As an application, we show that no subsolution of the problems that fit into
our framework is able to dissipate the total entropy earlier or faster than the strong solution on
the interval of existence of the latter. This result (we call it Dafermos’ principle) is new even for
“isotropic” problems such as the incompressible Euler system.

Keywords: nonlinear PDE, convex duality, entropy, anisotropic Orlicz space, generalized optimal
transport, Dafermos’ criterion
MSC [2020] 35D99, 35L90, 37K58, 47A56, 49Q99

1. Introduction
Cauchy problems for nonlinear evolutionary PDEs can have infinitely many weak solutions, cf.
[24, 23, 22, 14]. Many of such problems possess a physically relevant quantity (a Hamiltonian,
energy or entropy depending on the context) that should be formally conserved along the flow,
but can fail to do it for the weak solutions. There is some consensus that for physically relevant
solutions this quantity (for definiteness, hereafter we refer to it as the “total entropy”) cannot exceed
its initial value, cf. [21]. No universal selection criterion for weak solutions has been found, but
in physics, information theory, chemistry and biology there exist numerous variational principles
that can potentially assist. The most famous ones are Prigogine’s principle also known as the
minimum entropy production principle [50] that is applicable to open systems and Ziegler’s principle
also known as the maximum entropy production principle [58] that is suitable for closed systems.
Other notable principles are due to Onsager, Gyarmati, Berthelot, Swenson, Lotka, Enskog, Kohler,
Haken, Paltridge, Malkus, Veronis and Jaynes, see [46] for a survey. Such principles can be employed
both for derivation of physically relevant systems of PDEs and for selection of physically relevant
solutions [31]. In this vein, Dafermos [19, 20] suggested to assume that the physically relevant
weak solutions dissipate the total entropy earlier and faster than the irrelevant ones. Typically,
this kind of entropy1 is conserved for smooth solutions but can dissipate for weak solutions due to
shocks etc. It is important that the criterion acts locally near the “bifurcation” moment of time,
i.e., before a certain moment the total entropies of a “good” and a “bad” solution are equal (this
1
The wording comes from the theory of conservation laws and might be confusing for some readers. For example,
smooth solutions of the heat equation dissipate the Boltzmann entropy, so this entropy is never conserved. In this
paper, as we have already agreed, the “total entropy” is a formally conserved quantity.
1
2 D. VOROTNIKOV

stage is optional and the discrepancy can already occur at the initial moment of time), but shortly
after that moment the total entropy of a “bad” solution becomes larger than the total entropy of a
“good” solution; however, it is permitted that, as time elapses, the total entropy of a “bad” solution
returns to being smaller than or equal to the one of a “good” solution.
Appropriateness of Dafermos’ principle was examined for various PDEs, see, e.g., [33, 20, 26,
16, 18]. Dafermos’ criterion has recently been applied [29, 15] at the (wider) level of subsolutions
to some problems of fluid dynamics. Numerical applications of the criterion have recently been
produced in [35, 36, 37].
For systems of hyperbolic conservation laws
∂t  = −∂ (F()), (1.1)
where F : Rn → Rn is a given flux function, there exist classical selection criteria (Lax’ shock
condition, the Kruzhkov-Lax entropy criterion, the vanishing viscosity criterion and some others)
that under some assumptions are compatible with Dafermos’ principle, see [19, 41]. For equations
of Lagrangian and continuum mechanics, a certain least action admissibility criterion has recently
been advocated in [30]. A related least action principle was discussed in [28].
For the incompressible Euler and some related equations, Brenier [11, 12] suggested to search for
the solution that minimizes the time average of the kinetic energy. This problem might not always
admit a solution, but it generates a dual variational problem that has better convexity features.
He found an explicit formula that relates the smooth solutions of the incompressible Euler existing
on a small time interval to the solutions of the corresponding dual problem. In [13], this technique
was applied to the multi-stream Euler-Poisson system. In [57], we developed Brenier’s approach by
finding structures in nonlinear quadratic PDEs that permit to define similar dual problems with
decent properties. It was also discussed in [11, 12, 57] that dual problems of this kind have strong
analogy with the optimal transportation problems.
A similar duality scheme has been (rather independently) proposed by Acharya and several
collaborators, see [1, 2, 52] and the references therein. The recent preprint [3] builds upon both
approaches. Among other results, the authors of [3] prove the -convergence of the dual problem for
the incompressible Navier-Stokes system to the corresponding dual problem for the incompressible
Euler. From the perspective of the analogy with the optimal transport, this result is reminiscent of
the -convergence results from [42, 6, 48]. A notable feature of [3] is that their construction of the
dual problem involves a “base state” that can be regarded as an “initial guess” for the solution of
the primal problem. (From this point of view, the “base states” in [11, 12, 13, 57] are identically
zero).
It was clear from the very beginning [11, 1] that the nonlinearity in a PDE does not need to
be quadratic in order to implement the construction. However, the majority of rigorous results for
this kind of problems has until now been obtained for quadratic nonlinearities. Of course, for many
relevant systems the nonlinearity fails to be quadratic, for instance, for multidimensional systems
of conservation laws
∂t  = − div(F()), (1.2)
where F is a prescribed matrix function, and for various nonlinear dispersive equations. Remarkably,
rigorous study of Brenier’s duality scheme for multidimensional systems of conservation laws (1.2)
that admit a convex entropy, initiated in [11], is still pending.
In this paper, we extend Brenier’s approach to systems of PDEs that can be rewritten in the
abstract form
∂t  = L(F()), (1.3)
 DAFERMOS’ PRINCIPLE 3

where the matrix function F has some convexity and positivity properties to be specified below in
Assumptions 3.2 and 3.3, and L is a vector-valued differential operator of any order. To fix the
ideas, we will restrict ourselves to unknown vector functions (t, ) with  varying in the periodic
box Ω = Td . Note that (1.3) is strongly reminiscent of (1.1) and (1.2). We will also need to assume
that (1.3) admits a strictly convex anisotropic entropy function K() that is formally conserved
along the flow as described in Assumptions 3.1 and 3.7. As we will see, the NLS, NLKG and GKdV
equations with generic defocusing power-law nonlinearities can be recast in our abstract form.
As an application of our duality scheme, we rigorously prove the following variant of Dafermos’
principle: no subsolution of (1.3) can dissipate the total entropy earlier or faster than the strong
solution on the interval of existence of the latter (Theorem 4.3). Our Dafermos’ principle is new
even for quadratic problems that fit into our framework, including the incompressible Euler equation
[11], the ideal incompressible MHD, the multidimensional Camassa-Holm, the Zakharov-Kuznetsov
system and many others (various examples can be found in [57]).
Our entropy K generates an anisotropic Orlicz space, hence our procedure will heavily rely on the
theory of anisotropic Orlicz spaces, see, e.g., [17]. In particular, we will prove existence of solutions
to the dual problem that belong to an anisotropic Orlicz space. Notably, some authors employed
isotropic Orlicz spaces for the study of nonlinear Schrödinger and generalized KdV equations, cf.
[32], but we have never heard of any application of anisotropic Orlicz spaces — that appear so
naturally in our generic approach — to this sort of PDEs.
Until recently, an unsettling detail of the duality scheme was that the majority of consistency
results — that advocate the possibility to view solutions to the dual problems as certain “dual
variational solutions‘” to PDEs that admit the considered duality approach — were only obtained
on small time intervals, cf. [11, Theorems 2.3 and 3.1], [57, Theorem 3.8], [13, Theorem 5.1].
In this paper, we are able to guarantee the consistency on large time intervals by introducing
appropriate time-dependent weights to the problem. Global in time consistency theorems for the
dual formulations of incompressible Euler and Navier-Stokes equations have recently been obtained
in [3]. The latter results heavily rely on a suitable choice of the “base state”. Nevertheless, our
consistency results seem to be of a different nature, and, furthermore, the two attitudes complement
each other to a certain degree, see Remark 4.4 for a thorough comparison.
We believe that it is potentially possible to prove the weak-strong uniqueness property in the
sense that existence of a (suitably defined) strong solution to (1.3) implies that the solution to
the dual problem is unique in a certain anisotropic Orlicz space. The corresponding result in the
quadratic case F() =  ⊗  was established in [57, Section 5]. As the first step towards this
conjecture, we will show that the strong solution is always unique. The proof is unusual because it
relies on the analysis of a specific “Jeffreys divergence”.
To the best of our knowledge, global in time solvability for the considered dispersive PDEs is
merely known under significant restrictions on the exponents in the power laws or on the size of the
initial data, cf. [55, 44, 27]. Our Theorem 5.4 and Remark 5.5 provide existence of certain “dual
variational solutions” for these problems without such restrictions.
The paper consists of seven sections (including the Introduction) and two appendices. In Section
2, we list some necessary facts about N-functions and anisotropic Orlicz spaces. In Section 3, we
describe our abstract formalism. In Section 4, we establish the consistency of the dual problem,
and explain how it can be secured on large time intervals. As an application, we prove the already
mentioned version of Dafermos’ principle. In Section 5, we show that the dual problem always has
a solution that belongs to a specific anisotropic Orlicz space. In Section 6, we obtain uniqueness
of strong solutions. In Section 7, we demonstrate how the NLS, NLKG and GKdV equations
4 D. VOROTNIKOV

with generic defocusing power-law nonlinearities (as well as scalar conservation laws) fit into our
framework. In Appendix A, we prove a technical anisotropic variant of Hardy’s inequality (we do
not claim much novelty here, but we failed to find it in the existing literature). In Appendix B,
we employ an unsophisticated geometric point of view in order to illustrate the analogy of the dual
problems in question with the optimal transport theory.

2. Anisotropic Orlicz spaces

In this section we collect some basic necessary facts about N-functions and anisotropic Orlicz
spaces.
Definition 2.1 (Young function). A convex function  : R → R+ is called a Young function
provided  is even, (s) = 0 if and only if s = 0, and
(s) (s)
lim = 0, lim = +∞.
s→0 s s→+∞ s
Note that the Legendre transform ∗ of a Young function is well defined and is also a Young
function.
Definition 2.2 (N-function). A convex function K : Rn → R+ is called an N-function provided K
is even, and there exist two Young functions − and + such that
− (|y|) ≤ K(y) ≤ + (|y|), y ∈ Rn .
It follows from the definition that if K is an N-function, then
(i) K(y) = 0 if and only if y = 0;
(ii) The Legendre transform K∗ is well-defined and is also an N-function.
Definition 2.3 (Δ2 -condition). An N-function K : Rn → R+ satisfies the Δ2 -condition provided2
K(2y) ® K(y) + 1, y ∈ Rn .
Assumption 2.4. Henceforth in this section, we fix an N-function K. We always assume that both
K and K∗ satisfy the Δ2 -condition, and all the facts are claimed under this assumption.
Let Q be either an interval, or a flat torus (periodic box), or the Cartesian product of an interval
and a torus, equipped with the Lebesgue measure d.
Definition 2.5 (Modular and anisotropic Orlicz class). The integral
Z
K(ƒ ()) d
Q

is called the modular of a measurable function ƒ : Q → Rn . Denote by LK (Q; Rn ) the set of

functions ƒ whose modular is finite.
Note that LK (Q; Rn ) is a linear space, cf. [17, Lemma 3.3.1], and it is called an anisotropic Orlicz
|| p
space. For K() = p , p > 1, the spaces LK (Q; Rn ) are classical Lebesgue spaces Lp (Q; Rn ) .
There are two different ways to define a relevant norm on LK (Q; Rn ).

2Hereafter, we write b ®  for two scalar expressions  and b to indicate that b ≤ C with a uniform constant C.
 DAFERMOS’ PRINCIPLE 5

Definition 2.6 (Luxemburg and Orlicz norms). The Luxemburg norm of ƒ ∈ LK (Q; Rn ) is defined
as
ƒ ()
 Z   
kƒ kLK (Q) := inf λ > 0 : K d ≤ 1 .
Q λ
The Orlicz norm of ƒ ∈ LK (Q; Rn ) is defined as
Z Z 
∗
|||ƒ |||LK (Q) := sp ƒ () · g() d : K (g()) d ≤ 1 .
Q Q

Both norms are indeed norms. Moreover, LK (Q; Rn ) equipped with any of these norms is a
separable reflexive Banach space, see [17]. Both norms are equivalent with uniform constants:
kƒ kLK (Q) ≤ |||ƒ ||| LK (Q) ≤ 2kƒ kLK (Q) . (2.1)
Furthermore, the strong topology associated with both norms coincides with the modular topology
[17], i.e.,
Z
kƒk kLK (Q) → 0 ⇔ K(ƒk ()) d → 0 (2.2)
Q
as k → +∞.
Proposition 2.7 (Hölder inequality [17]). For every ƒ ∈ LK (Q; Rn ) and g ∈ LK∗ (Q; Rn ) one has
Z
ƒ () · g() d ≤ 2kƒ kLK (Q) kgkLK∗ (Q) .
Q

Another useful estimate, [17, Lemma 3.1.14], is

 Z 
kƒ kLK (Q) ≤ mx 1, K(ƒ ()) d . (2.3)
Q

Proposition 2.8 (Duality [17]). Bounded linear functionals on LK (Q; Rn ) can be represented in
the form
Z
ƒ 7→ ƒ () · g() d
Q
for some g ∈ LK∗ (Q; Rn ). Consequently, LK∗ (Q; Rn ) is the dual space of LK (Q; Rn ) .
We will need the following (refined) embedding of anisotropic Orlicz spaces.

Proposition 2.9. Let K, K̃ be two N-functions. Assume that there exists r > 0 such that K() ≤
K̃() for || ≥ r. Then for any ƒ ∈ LK̃ (Q; Rn ) we have ƒ ∈ LK (Q; Rn ) and
kƒ kLK (Q) ≤ Cr kƒ kLK̃ (Q)
where Cr merely depends on r and Q.
The proof of Proposition 2.9 is omitted because the classical proof for the isotropic case, cf. [39,
Theorem 13.3], is applicable here as well, mutatis mutandis.
Proposition 2.10 (Density of smooth functions). Smooth functions are dense in LK (Q; Rn ).
6 D. VOROTNIKOV

Proof. This fact is well known, so we just give a sketch of the proof. It follows from [17, Lemma
|| p
2.3.16] that there exist r > 0 and p > 1 such that K() ≤ p for || ≥ r. By Proposition 2.9,
Lp (Q) ⊂ LK (Q) and k · kLK (Q) ≤ Cr k · kLp (Q) . Simple functions are dense in LK (Q; Rn ) by [17,
Corollary 3.4.12]. On the other hand, in any Lp -neighborhood (and hence in any LK -neighborhood)
of a simple function there is a smooth function. 
We will also need the following variant of Krasnoselskii’s theorem about continuity of Nemytskii
operators.
Proposition 2.11. Let η : Rn → Rm be a continuous function such that the Nemytskii operator
N : LK (Q; Rn ) → L1 (Q; Rm ), N ()() = η(())
is well-defined. Then N is continuous.
We omit the proof because in the original proof of Krasnoselskii, see [38, Chapter 1, Theorem
2.1], it is possible to substitute the anisotropic Orlicz source LK for the Lebesgue source Lp without
destroying the validity of the argument3.
We finish this section with a statement about mollification. We restrict ourselves to the case
when Q is a d-dimensional periodic box. Let ρ be a smooth compactly supported probability
density on Rd . Set ρk () := kρ(k), k ∈ N. For any fixed ƒ ∈ LK (Q; Rn ), the convolution ρk ∗ ƒ
is well-defined, smooth, and belongs to LK (Q; Rn ).
Proposition 2.12. We have ρk ∗ ƒ → ƒ strongly in LK (Q; Rn ) as k → +∞.
Proof. We give just a brief sketch of the proof. It follows from the density of smooth functions that
the “translation” map
 7→ ƒ (· − )
is well-defined and continuous from R to LK (Q; Rn ). We conclude by mimicking the “isotropic”
d

Donaldson-Trudinger argument from [25, Lemma 2.1]. 

3. The abstract setting

Let us start with fixing some basic notation. In the what follows, Ω is the periodic box Td ,
d ∈ N, equipped with the Lebesgue measure d. For two measurable functions ,  : Ω → Rm ,
m ∈ N, we will use the shortcut
Z
(, ) := () · () d.
Ω
We will also use the notations RN×N , RN×N
s
and RN×N
+
for the spaces of N × N matrices, symmetric
matrices and nonnegative-definite matrices, resp. For A, B ∈ RsN×N , we write A  B when A − B ∈
RN×N
+
(the Loewner order [45]; we stress that the notation A ≥ B is reserved for a less restrictive
order that will introduced in Assumption 3.3 below), and A : B for the scalar product of A and B
generated by the Frobenius norm | · |. The symbol  will stand for the identity matrix of a relevant
size.
As anticipated in the Introduction, we are interested in the problem
∂t  = L(F()), (0, ·) = 0 . (3.1)
3The case of anisotropic targets is more tricky and lies beyond this discussion; for instance, the proof for isotropic
Orlicz sources and targets that can be found in [39] does not work in the anisotropic case.
 DAFERMOS’ PRINCIPLE 7

Here 0 : Ω → Rn , n ∈ N, is the initial datum,  : [ 0, T] × Ω → Rn is an unknown vector function,

F : Rn → RN×N
s
is a prescribed C2 -smooth matrix function with some convexity and positivity properties to be
specified below, and L is a vector-valued differential operator4 with constant coefficients
ν X
X N X
L() = bmα ∂α m ,  = 1, . . . , n,
j=0 ,m=1 |α|=j

where α is a generic multiindex and () ∈ RN×N

s
.
Assumption 3.1 (Entropy). Define the “entropy” function for our problem by
1
K : Rn → R, K() := Tr(F() − F(0)).
2
Assume that K is strictly convex (but not necessarily uniformly convex). Assume also that K is an
N-function, and both K and its Legendre transform K∗ satisfy the Δ2 -condition.
To any given  ∈ Rn we associate the “sharp” vector
 # := ∇K(),
where the gradient ∇ is naturally taken w.r.t. . Note that 0# = 0 because ∇K(0) = 0. Since K
is strictly convex, the map
∇K :  7→  #
is C1 -smooth and injective. Moreover,
 = ∇K∗ ( # )
for any  # ∈ ∇K(Rn ) = Rn .
It follows from [17, Lemma 2.3.16] that there exist p, q > 1 and r > 0 such that
 · # # · 
≤ p, ≥q
K() K∗ ( # )
for || ≥ r. Consequently,
K∗ ( # ) ® K() + 1,  ∈ Rn . (3.2)
Similarly,
K() ® K∗ ( # ) + 1,  ∈ Rn . (3.3)
Assumption 3.2 (Λ-convexity). We will require some convexity of the matrix function F. The
classical notion of convexity of matrix-valued functions is the Loewner convexity [40, 9], namely,
F : Rn → RN×Ns
is Loewner convex if F : P is a convex function for any matrix P ∈ RN×N+
. This is
too restrictive for our purposes, cf. Remark 7.1 and Appendix B, so we will merely assume that F
is Λ-convex in the following more relaxed sense. Let  : Ω → Rn be an arbitrary smooth function,
and let L∗ be the differential operator adjoint to L. Let Λ ⊂ RN×N
s
be the smallest linear subspace
of RsN×N ∗
independent of  and containing  such that L (()) ∈ Λ,  ∈ Ω. We say that F is

4As a matter of fact, L can be a more general (and possibly nonlocal) linear operator; it is easy to see from the
considerations below that it is enough to assume that L commutes with the mollification L(ρk ∗ M) = ρk ∗ L(M) and
is a closed linear operator with a dense domain D(L) in a certain space.
8 D. VOROTNIKOV

Λ-convex if F : P is a convex function for any P ∈ Λ ∩ RN×N +

. Of course, any Loewner convex
matrix function is Λ-convex, but not vice versa. Notably, it follows from the Λ-convexity of F that
K is convex.
Assumption 3.3 (Λ-order). We will also require some positivity of the matrix function F. For
this purpose, we will use the less restrictive Λ-order instead of the Loewner order. Namely, denote
RΛN×N = { ∈ RN×N
s
|  : P ≥ 0, ∀P ∈ Λ ∩ RN×N
+
}.
We will assume that
F(Rn ) ⊂ RN×N
Λ
. (3.4)
For A, B ∈ RsN×N , we write B ≤ A when A − B ∈ RN×N
ΛObviously, B  A implies B ≤ A, and the
.
converse is true only if Λ = RN×N
s
. Employing some simple algebra, it is easy to derive from F ≥ 0
that
|πΛ F()| ® K() + 1, (3.5)
where πΛ : RN×N
s
→ Λ denotes the orthogonal projection w.r.t. the Frobenius product. We will also
need to assume that pointwise control similar to (3.5) holds at the level of partial derivatives
F () := F (),  = 1, . . . , n.
More exactly, we assume that
|πΛ ∂ F()| ® |∇K()| + 1. (3.6)
Remark 3.4. In some examples, including the incompressible Euler equation, cf. [11, 57], Λ coin-
cides with RN×N
s
, hence in this case the Λ-order and the Λ-convexity coincide with their Loewner
counterparts. For the dispersive equations considered in Section 7, the subspace Λ is significantly
smaller than RN×N
s
.

Remark 3.5. The “flux” F() can be viewed as an “entropy matrix” that very vaguely resembles the
entropy matrix from [51]. Notably, the definition of convexity of matrix-valued functions employed
in [51] is very restrictive, it is much stronger than the Loewner convexity.

Remark 3.6. The quadratic flux F() =  ⊗  and the corresponding entropy K() = 21 ||2
obviously satisfy Assumptions 3.1, 3.2 and 3.3 for any subspace Λ ⊂ Rsn×n . Note that in this
case N = n and  # = . This applies to the incompressible Euler and other quadratic examples
mentioned in the Introduction.
Assumption 3.7 (Conservativity). We will focus on the situation when L satisfies the “formal
conservativity condition”
(F(), L∗ ( # )) = 0 (3.7)
provided  # : Ω → Rn is a smooth function (remember that  = ∇K∗ ( # )).
Problem (3.1) admits the following natural weak formulation:
ZT
(, ∂t ) + (F(), L∗ ) dt + (0 , (0, ·)) = 0
 
(3.8)
0
for all smooth vector fields  : [ 0, T] × Ω → Rn , (T, ·) ≡ 0.
 DAFERMOS’ PRINCIPLE 9

Remark 3.8 (Formulation in terms of the sharp variable). In the conservative case (3.1) admits a
formally equivalent formulation in terms of the “sharp” unknown variable  # :
∂t ( # ) + L∗ ( # ) : ∂ F(∇K∗ ( # )) = 0,  # (0, ·) = 0# := (0 )# ,  = 1, . . . , n. (3.9)
Indeed, let  : [ 0, T] × Ω → Rn be a smooth test function, s ∈ R be a real number, and  be a
solution to (3.1). Then due to the conservativity
(F( + s), L∗ (∇K( + s))) = 0. (3.10)
Taking the derivative w.r.t. s at s = 0 yields
n
X
(∂ F() , L∗ ( # )) + (L(F()), ∂ (∇K()) ) = 0.
 
(3.11)
=1

Using (3.1), we recast (3.11) in the form

n
X n
X
(∂ F() : L∗ ( # ),  ) + (∂t m , ∂m K() ) = 0. (3.12)
=1 ,m=1
Pn
By the chain rule, ∂t ( # ) = ∂t (∂ K()) = m=1
∂m K()∂t m . Thus (3.12) becomes
n
X
(∂ F() : L∗ ( # ) + ∂t ( # ) ,  ) = 0, (3.13)
=1

which immediately implies (3.9) due to arbitrariness of  and the relation  = ∇K∗ ( # ).
Let us now rewrite problem (3.8) in terms of the test functions B := L∗  and E := ∂t  (note
that the conservativity (3.7) is not needed at this stage). We first observe that
ZT ZT
(0 , (0)) = − (0 , ∂t ) dt = − (0 , E) dt. (3.14)
0 0
The link between B and E can alternatively be described by the conditions
∂t B = L∗ E, B(T, ·) ≡ 0. (3.15)
Hence, (3.8) becomes
Z T
[ ( − 0 , E) + (F(), B)] dt = 0 (3.16)
0
for all smooth vector fields B : [ 0, T] × Ω → RN×N
s
, E : [ 0, T] × Ω → Rn satisfying the constraints
(3.15).
Observe that (3.15) can be rewritten in the following weak form
ZT
[ (B, ∂t Ψ) + (E, LΨ)] dt = 0 (3.17)
0

for all smooth matrix fields Ψ : [ 0, T] × Ω → RN×N

s
, Ψ(0) = 0.
Motivated by the discussion above, we adopt the following definitions, where we tacitly assume
0 ∈ LK (Ω; Rn ).
10 D. VOROTNIKOV

Definition 3.9 (Weak solutions). A function

 ∈ LK ((0, T) × Ω; Rn ) (3.18)
is a weak solution to (3.1) if it satisfies (3.16) for all pairs
(E, B) ∈ LK∗ ((0, T) × Ω; Rn ) × L∞ ((0, T) × Ω; RN×N
s
) (3.19)
meeting the constraint (3.17).
Note that if  is a weak solution, then F() ∈ L1 ((0, T) × Ω; RN×N
s
) due to (3.5), so (3.16)
makes sense.
Definition 3.10 (Subsolutions). A pair of functions
(, M) ∈ LK ((0, T) × Ω; Rn ) × L1 ((0, T) × Ω; RN×N
s
), F() ≤ M, (3.20)
is a subsolution to (3.1) if it satisfies
ZT
[ ( − 0 , E) + (M, B)] dt = 0 (3.21)
0
for all pairs
(E, B) ∈ LK∗ ((0, T) × Ω; Rn ) × L∞ ((0, T) × Ω; RN×N
s
)
meeting the constraint (3.17).
The second inequality in (3.20) is understood in the sense of the Λ-order a.e. in (0, T) × Ω.
Obviously, if  is a weak solution, then (, F()) is a subsolution. Accordingly, the subsolution
entropy function that complies with Assumption 3.1 is
1
M 7→ Tr(M − F(0)), M ∈ RN×N
Λ
.
2
Observe that for a weak solution  the modular (that we will call the total entropy)
Z
K(t) := K((t, )) d. (3.22)
Ω
belongs to L1 (0, T) and thus is finite for almost all times. Similarly, the total entropy
1 1
Z
K̃(t) := Tr [ M(t, ) − F(0)] d = (M(t, ·) − F(0), ). (3.23)
2 Ω 2
of a subsolution also belongs to L1 (0, T).
Remark 3.11. For the incompressible Euler equation, in view of Remark 3.4 and some observations
made in [21, 57], Definition 3.10 is equivalent to the conventional definition of a subsolution that,
among other applications, is used in the theory of convex integration [23, 22].
Our primal problem is to search for the weak solutions to (3.1) that minimize a suitable weighted
time integral of the total entropy5 (3.22).
More exactly, fix a positive scalar function ϟ(t) that is bounded away from 0 and ∞ on [ 0, T].
RT RT
Let Ϟ(t) := t ϟ(s) ds. The idea is to search for the weak solutions that minimize 0 ϟ(t)K(t) dt.

5See also Remark 3.18.

 DAFERMOS’ PRINCIPLE 11

Remark 3.12 (“Rough” Dafermos’ criterion). A typical weight is ϟ(t) := exp(−γt) with large
γ > 0, cf. Remark 4.2. In this case our selection criterion can be viewed as a “rough” Dafermos’
principle, because we prioritize the solutions that dissipate the total entropy as fast and as early as
possible and that at later times do not admit dramatic exponential growth (of order exp(γt)) of
the total entropy.
We will now employ the ”sharp” formulation (3.9) in order to define strong solutions to (3.1).
Definition 3.13 (Strong solutions). Assume that 0 ∈ LK (Ω; Rn ). A function  satisfying (3.18)
and
∂t (Ϟ # ) ∈ LK∗ ((0, T) × Ω; Rn ), ϞL∗ ( # ) ∈ L∞ ((0, T) × Ω; RsN×N ) (3.24)
is a strong solution to (3.1) if it is a weak solution and
∂t ( # ) + L∗ ( # ) : ∂ F(∇K∗ ( # )) = 0,  # (0, ·) = 0# (3.25)
a.e. in (0, T) × Ω and in Ω, resp.
Remark 3.14. Let us explain why (3.25) makes sense for  from the regularity class (3.18), (3.24).
Firstly, (3.2) and (3.18) imply  # ∈ LK∗ ((0, T) × Ω; Rn ). The first equality in (3.25) is equivalent
to
ϟ( # ) + ∂t (Ϟ # ) + ϞL∗ ( # ) : ∂ F(∇K∗ ( # )) = 0 (3.26)
a.e. in (0, T) × Ω. By (3.6), πΛ ∂ F() is pointwise controlled by ∇K() =  # , therefore all
members of (3.26) belong to LK∗ ((0, T) × Ω; Rn ). We moreover claim that
 # ∈ C([ 0, T); LK∗ (Ω; Rn )) (3.27)
and that all members in the second equality in (3.25) belong to LK∗ (Ω; Rn ). Indeed, since
K∗ (0# ) ® K(0 ) + 1, we have 0# ∈ LK∗ (Ω; Rn ). But
1 ϟ
∂t  # = ∂t (Ϟ # ) +  # ∈ LK∗ ,oc ([ 0, T) × Ω; Rn ) ⊂ L1oc ([ 0, T); L1 (Ω)),
Ϟ Ϟ
whence
 # ⊂ ACoc ([ 0, T); L1 (Ω)).
Leveraging convexity of K∗ , for every t ∈ [ 0, T) and small h > 0 we obtain
‚ Z t+h Œ
1
Z Z
∗ # # ∗ #
K ( (t + h, ) −  (t, )) d = K h∂t  (s, ) ds d
Ω Ω h t
1 t+h ∗
Z Z
K h∂t  # (s, ) ds d


≤
Ω h t
Z t+h
1
Z
K∗ h∂t  # (s, ) + (1 − h)0 ds d


=
Ω h t
1 t+h
Z Z
hK∗ ∂t  # (s, ) + (1 − h)K∗ (0) ds d



≤
Ω h t
Z Z t+h
K∗ ∂t  # (s, ) ds d → 0


= (3.28)
Ω t
12 D. VOROTNIKOV

as h → 0 due to absolute continuity of the Lebesgue integral. This secures the right-continuity of
the map  # : [ 0, T) → LK∗ (Ω); the left-continuity is proved in a similar fashion.
Remark 3.15. Definition 3.13 is actually independent of the weight function Ϟ. Indeed, let 
be a strong solution, ϟ1 be another positive scalar function bounded away from 0 and ∞, and
RT
Ϟ1 = t ϟ1 (s) ds. Obviously, Ϟ(t) ® Ϟ1 (t) ® Ϟ(t). Hence,

Ϟ1 L∗ ( # ) ∈ L∞ ((0, T) × Ω; RN×N
s
),
and
Ϟ1 Ϟ1
∂t (Ϟ1  # ) = −ϟ1  # + Ϟ1 ∂t ( # ) = −ϟ1  # + ϟ # +
∂t (Ϟ # ).
Ϟ Ϟ
Since LK∗ ((0, T) × Ω; Rn ) is a linear space invariant w.r.t. multiplication by bounded functions,
∂t (Ϟ1  # ) ∈ LK∗ ((0, T) × Ω; Rn ).
The following lemma shows that the strong solutions soar above the rest of the weak solutions.
Lemma 3.16. If (3.7) is assumed, then for any strong solutionR  the total entropy is continuous
in time for t ∈ [ 0, T) and is conserved, i.e., K(t) = K(0) = Ω K(0 ) d.
φ
Proof. Let φ(t) be an arbitrary smooth compactly supported function on (0, T). Observe that Ϟ
is bounded, Lipschitz and compactly supported on (0, T). Let us prove that
ZT
(K(), ∂t φ) dt = 0. (3.29)
0
We first claim that
Z T
(F(), L∗ (φ # )) dt = 0 (3.30)
0
and
Z T
( − 0 , ∂t (φ # )) − (K(∇K∗ ( # )), ∂t φ) dt = 0.
 
(3.31)
0
Taking (3.30) and (3.31) for granted, and employing (3.16) with
φ
B = L∗ (φ # ) = ϞL∗ ( # ) ∈ L∞ ((0, T) × Ω; Rn ),
Ϟ
φ φ
E = ∂t (φ # ) = ∂t ( )Ϟ # + ∂t (Ϟ # ) ∈ LK∗ ((0, T) × Ω; Rn ),
Ϟ Ϟ
we infer
ZT
( − 0 , ∂t (φ # )) dt = 0
0
and, consequently,
Z T
(K(∇K∗ ( # )), ∂t φ) dt = 0.
0
We have proved (3.29), i.e., that the total entropy is conserved in the sense of distributions. In
order to prove that it is conserved in the classical sense, we just need to establish the continuity
 DAFERMOS’ PRINCIPLE 13

of the total entropy w.r.t. time. By Remark 3.14,  # ∈ C([ 0, T); LK∗ (Ω; Rn )). Proposition 2.11
and (3.3) imply that the Nemytskii operator
N1 :  7→ K(∇K∗ ())
is continuous from LK∗ (Ω; Rn ) to L1 (Ω). Consequently, K() ∈ C([ 0, T); L1 (Ω)), whence K(t) ∈
C([ 0, T)).
It remains to prove (3.30) and (3.31). Without loss of generality (continuing by zero outside
[ 0, T]) we may assume that  # and φ are defined on the torus Q := 2T T1 × Ω. Consider the
mollifications k# := ρk ∗  # , where ρk are as in Section 2.
We now prove (3.31). Due to the presence of the cut-off function φ, the behaviour of  # for
small T − t does not affect the validity of (3.31), and thus without loss of generality we may assume
that
1 ϟ
∂t  # = ∂t (Ϟ # ) +  # ∈ LK∗ (Q; Rn ).
Ϟ Ϟ
Then
k# →  # , ∂t k# → ∂t  # , ∂t (φk# ) → ∂t (φ # ) (3.32)
strongly in LK∗ (Q; Rn ) as k → +∞. Set k := ∇K∗ (k# ). Observe now that
ZT
” —
(k , ∂t (φk# )) − (K(∇K∗ (k# )), ∂t φ) dt = 0. (3.33)
0
Indeed, integration by parts gives
ZT
” —
(k , ∂t (φk# )) − (K(k ), ∂t φ) dt
0
Z T ” —
=− (∂t k , φk# ) − (∇K(k ) · ∂t k , φ) dt
0
Z T ” —
=− (∂t k , φk# ) − (k# · ∂t k , φ) dt = 0. (3.34)
0
By (3.3),
Z Z
∗
K(∇K ()) ® K∗ () + 1. (3.35)
Q Q
Proposition 2.11 implies that the Nemytskii operator
N1 :  7→ K(∇K∗ ())
is continuous from LK∗ (Q; Rn ) to L1 (Q). Hence, we can pass to the limit in the second term of
(3.33).
On the other hand, it follows from (3.32) and (3.35) that the sequence k = ∇K∗ (k# ) is bounded
in LK (Q; Rn ) and a.e. converges to . Passing to a subsequence if necessary, we infer k → 
weakly in LK (Q; Rn ). We can now legitimately pass to the limit in the first term of (3.33).
Moreover,
ZT
(0 , ∂t (φk# )) dt = 0 (3.36)
0
14 D. VOROTNIKOV

since φ is compactly supported in (0, T) and 0 does not depend on t. Since 0 ∈ LK (Ω; Rn ) ⊂
LK (Q; Rn ), we can pass to the limit in (3.36) as well.
As a result, (3.33) and (3.36) imply (3.31).
Let us prove (3.30). Due to the presence of φ, the behaviour of  # for small T − t is again not
relevant, and we may assume that
L∗ ( # ) ∈ L∞ (Q; RN×N
s
).
Passing to a subsequence if necessary, we get
L∗ (k# ) → L∗ ( # ) (3.37)
weakly-∗ in L∞ (Q; RN×N
s
) as k → +∞. On the other hand, (3.7) yields
ZT ZT
(F(∇K∗ (k# )), L∗ (φk# )) dt = (F(∇K∗ (k# )), L∗ (k# ))φ(t) dt = 0. (3.38)
0 0
Estimates (3.5), (3.35) and Proposition 2.11 imply that the Nemytskii operator
N2 :  7→ F(∇K∗ ())
is continuous from LK∗ (Q; Rn ) to L1 (Q; RN×N
s
). Hence,

F(k ) = F(∇K∗ (k# )) → F(∇K∗ ( # )) = F()

strongly in L1 (Q; RN×N
s
), and passing to the limit in (3.38) we obtain (3.30). 
RT
The suggested approach of selecting the weak solutions that minimize 0 ϟ(t)K(t) dt — or,
1R
equivalently, 2 (0,T)×Ω ϟ(t) Tr F(t) d dt — can be implemented via the saddle-point problem
Z T
1

I(0 , T) = inf sp ( − 0 , E) + (F(), ϟ + 2B) dt. (3.39)
 E,B: (3.17) 2
0

The infimum in (3.39) is taken over all  ∈ LK ((0, T) × Ω; Rn ), and the supremum is taken over
all pairs (E, B) satisfying (3.19) and the linear constraint (3.17).
The dual problem is
Z T
1

J (0 , T) = sp inf ( − 0 , E) + (F(), ϟ + 2B) dt, (3.40)
E,B: (3.17)  0 2
where , E, B are varying in the same function spaces as above.
It is also relevant to consider the corresponding problems for the subsolutions:
Z T
1

Ĩ(0 , T) = inf sp ( − 0 , E) + (M, ϟ + 2B) dt. (3.41)
,M:F()≤M E,B: (3.17) 0 2

The infimum in (3.39) is taken over all  ∈ LK ((0, T) × Ω; Rn ) and M ∈ L1 ((0, T) × Ω; RN×NΛ ),
and the supremum is taken over all pairs (E, B) satisfying (3.19) and the linear constraint (3.17).
The problem dual to (3.41) is
Z T
1

J˜ (0 , T) = sp inf ( − 0 , E) + (M, ϟ + 2B) dt, (3.42)
E,B: (3.17) ,M:F()≤M 0 2
 DAFERMOS’ PRINCIPLE 15

where , M, E, B are varying in the same function spaces as above.

Remark 3.17 (Simple observations about the optimal values). Denote
1 1
Z Z
K0 := Tr F(0 ) d = K(0) + Tr F(0) d. (3.43)
2 Ω 2 Ω
Since inf sp ≥ sp inf, one has
I(0 , T) ≥ Ĩ(0 , T) ≥ J˜ (0 , T), I(0 , T) ≥ J (0 , T) ≥ J˜ (0 , T).
Note that the sp in (3.39) is always +∞ if  is not a weak solution, whence I(0 , T) = +∞ if
there are no weak solutions. On the other hand, if (3.7) holds, and there exists a strong solution
, then the corresponding sp in (3.39) is equal to
T T
1 1
Z Z  Z 
(F(), ϟ) dt = ϟ K(t) + Tr F(0) d dt
2 0 0 2 Ω
T T
1
Z  Z  Z
= ϟ K(0) + Tr F(0) d dt = ϟK0 dt = Ϟ(0)K0 ,
0 2 Ω 0

which yields I(0 , T) ≤ Ϟ(0)K0 . Finally, testing by (E, B) = (0, 0) we see that

1
Z T
J˜(0 , T) ≥ inf (M, ϟ) dt ≥
2 ,M:F()≤M 0
1 Ϟ(0)|Ω| Tr F(0)
Z  
inf ϟ K() + Tr F(0) d dt ≥ ≥ 0.
,M:F()≤M (0,T)×Ω 2 2

It is easy to see (taking into account that M + A ≥ M for any A  0) that if ϟ + 2B is not
non-negative-definite on a set of positive Lebesgue measure in (0, T) × Ω, then the inf in (3.42)
equals −∞. Hence, any solution to (3.42) necessarily satisfies
ϟ + 2B  0 .e. in (0, T) × Ω. (3.44)
Consider the nonlinear functional
K : LK∗ ((0, T) × Ω; Rn ) × L∞ ((0, T) × Ω; RN×N
s
)→R
defined by the formula
T
1
Z  
K(E, B) = inf (z, E) + (M, ϟ + 2B) dt, (3.45)
F(z)≤M 0 2
where the infimum is taken over all pairs (z, M) ∈ LK ((0, T) × Ω; Rn ) × L1 ((0, T) × Ω; RN×N
s
).
Then (3.42) is equivalent to
ZT
J˜(0 , T) = sp −(0 , E) dt + K(E, B), (3.46)
E,B: (3.17),(3.44) 0

the supremum is taken over all pairs (E, B) belonging to the class (3.19).
16 D. VOROTNIKOV

Remark 3.18 (A variant of our theory based on the relative entropy). Inspired by [3], is possible
to develop a variant of our theory for which the primal problem consists in minimizing a suitable
time integral of the relative entropy (also known as the Bregman divergence)
Z
K((t, )) − K(̃(t, )) − ̃ # (t, ) · ((t, ) − ̃(t, )) d.
 
D(t) := (3.47)
Ω

Here
̃ : [ 0, T] × Ω → Rn
is a fixed “base state”. (Since K(0) = 0 and 0# = 0, the setting that we have adopted in this
paper corresponds to ̃ ≡ 0.) The implementation of this idea lies beyond the scope of this paper,
but it is very plausible that an existence theorem (similar to Theorem 5.4) and a global in time
consistency result similar to [3, Theorem 3.6], cf. Remark 4.4, should hold for the corresponing
dual problem. An interesting task would be to obtain consistency results in which ̃ would neither
be identically zero (as in Theorem 4.1) nor coincide with the strong solution (as in [3, Theorem
3.6]).

4. Consistency and Dafermos’ principle

The following theorem shows that a strong solution determines a solution to the optimization
problem (3.46), and vice versa. This advocates the possibility to view the maximizers of (3.46) as
“dual variational solutions‘” to (3.1), cf. [11, 57, 3], see also Remark 5.5.

Theorem 4.1 (Consistency). Let 0 ∈ LK (Ω; Rn ). Let  be a strong solution to (3.1) satisfying

ϟ  −2Ϟ(t)L∗ ( # ) a.e. in (0, T) × Ω. (4.1)

Then I(0 , T) = J˜(0 , T) = Ĩ(0 , T) = J (0 , T) = Ϟ(0)K0 . The pair (E+ , B+ ) defined by

B+ = L∗ , E+ = ∂t , (4.2)

where
 = Ϟ # , (4.3)
belongs to the class (3.19) and maximizes (3.46). Moreover, one can invert these formulas and
express  in terms of E+ as follows
‚ ZT Œ
∗
1
(t, ) = ∇K (−E+ )(s, ) ds , t < T. (4.4)
Ϟ(t) t

Proof. By construction, the pair (E+ , B+ ) belongs to LK∗ ((0, T)× Ω; Rn )× L∞ ((0, T)× Ω; RN×N
s
).
It follows from Ϟ(T) = 0 that this pair verifies (3.17). Moreover, (4.1) implies (3.44) for B+ . We
now claim that
1
(ϟ + 2B+ ) : ∂ F() + (E+ ) = 0. (4.5)
2
 DAFERMOS’ PRINCIPLE 17

Indeed, using (3.9) we compute

1
(ϟ + 2B+ ) : ∂ F() + (E+ )
2
1
= (ϟ + 2ϞL∗ ( # )) : ∂ F() + (−ϟ( # ) + Ϟ∂t ( # ) )
2
= ϟ∂ K() + ϞL∗ ( # ) : ∂ F() + (−ϟ( # ) + Ϟ∂t ( # ) )
= Ϟ(L∗ ( # : ∂ F() + ∂t ( # ) ) = 0. (4.6)

On the other hand, since  is in particular a weak solution, it satisfies (3.16) with test functions
(E+ , B+ ). Thus we have
Z T
[ ( − 0 , E+ ) + (F(), B+ )] dt = 0. (4.7)
0

Hence, by (4.5),

T n T
1X
Z Z
[ −(0 , E+ ) + (F(), B+ )] dt = ( (ϟ + 2B+ ), ∂ F()) dt. (4.8)
0 2 =1 0

Employing (4.8) and Lemma 3.16, we obtain

T n T
1 1X
Z Z
(F(), ϟ + 2B+ ) dt − ( ∂ F(), ϟ + 2B+ ) dt
2 0 2 =1 0
T
1
Z   
= (0 , E+ ) + ϟ(t) K((t)) + Tr F(0), 1 dt
0 2
T
1
Z   
= (0 , E+ ) + ϟ(t) K(0 ) + Tr F(0), 1 dt
0 2
Z T
= (0 , E+ ) dt + Ϟ(0)K0 . (4.9)
0

By Remark 3.17, we have I(0 , T) ≤ Ϟ(0)K0 . Hence, it suffices to show that

Z T
−(0 , E+ ) dt + K(E+ , B+ ) ≥ Ϟ(0)K0 , (4.10)
0

so that there is no duality gap.

1
The Λ-convexity of F implies that the function y 7→ 2 F(y) : (ϟ + 2B+ (t, )) is convex w.r.t.
y ∈ Rn for a.e. fixed pair of parameters (t, ) ∈ (0, T) × Ω. Now (4.5), (4.9) and (3.44) for B+
18 D. VOROTNIKOV

yield
ZT
−(0 , E+ ) dt + K(E+ , B+ )
0
Z T
= −(0 , E+ ) dt
0
T n
– ™
1X 1
Z
+ inf − (z  ∂ F(), ϟ + 2B+ ) + (M, ϟ + 2B+ ) dt
F(z)≤M 0 2 =1 2
Z T
≥ −(0 , E+ ) dt
0
T n
– ™
1X 1
Z
+ inf − (z  ∂ F(), ϟ + 2B+ ) + (F(z), ϟ + 2B+ ) dt
z∈LK ((0,T)×Ω;Rn ) 0 2 =1
2

Z T
= −(0 , E+ ) dt
0
T n
– ™
1X 1
Z
+ − ( ∂ F(), ϟ + 2B+ ) + (F(), ϟ + 2B+ ) dt
0 2 =1
2
= Ϟ(0)K0 .
The penultimate equality follows from the well-known fact that, given a convex function  : Rn → R,
the function y 7→ (y) − y · ∇(ξ) attains its minimum at y = ξ.
Finally, since Ϟ(T) = 0, we deduce from (4.2) and (4.3) that
ZT ZT
E+ (s, ) ds = ∂s (Ϟ(s) # (s, )) ds = −Ϟ(t) # (t, ), (4.11)
t t
which yields (4.4). 
Remark 4.2 (Limitation (4.1) can be overcome by adapting the weight ϟ). At first glance, condition
(4.1) indicates that the interval for the which the consistency holds can be smaller than the interval
[ 0, T) on which the strong solution6 exists. However, the weight ϟ can be selected to guarantee
the consistency on any interval [ 0, T1 ], T1 < T. Indeed, by Remark 3.15 with ϟ1 ≡ 1, for any
strong solution on [ 0, T) we have that (T − t)L∗ ( # (t, )) is essentially bounded. Therefore, for
any T1 < T there exists γ > 0 such that
γ  −2L∗ ( # ) (4.12)
a.e. on (0, T1 ) × Ω. It is easy to see from Definition 3.13 that  (more accurately, the restriction
|[ 0,T1 ]×Ω ) is a strong solution on [ 0, T1 ]. Consider the weight ϟ(t) := exp(−γt) and let
RT
Ϟ(t) := t 1 ϟ(s) ds ≥ 0 (note that we are now working on the interval [ 0, T1 ]). Observe that
γϞ ≤ ϟ.
6We say “the strong solution” because strong solutions are unique, see Theorem 6.1.
 DAFERMOS’ PRINCIPLE 19

Because of (4.12), this implies that (4.1) holds on (0, T1 ) × Ω.

As an application of Theorem 4.1 we establish a variant of Dafermos’ principle for (3.1) affirm-
ing that a strong solution dissipates the total entropy not slower and not later than any other
subsolution.
Theorem 4.3 (Dafermos’ principle). Let 0 ∈ LK (Ω;R Rn ) . Let  be a strong solution to (3.1) on
the interval [ 0, T] with total entropy K(t) = K(0) = Ω K(0 ) d, and let (, M) be a subsolution
1
on [ 0, T] with total entropy K̃(t) = 2 (M(t, ·) − F(0), ). Then for any 0 ≤ t0 < t1 ≤ T it cannot
simultaneously be that K̃(t) ≤ K(t) for a.a. t ∈ (0, t1 ) and K̃(t) < K(t) for a.a. t ∈ (t0 , t1 ). In
particular, it is impossible that K̃(t) < K(t) for a.a. t ∈ (0, ε), ε > 0.
Proof. We just prove the first claim, to obtain the second one it suffices to put t0 = 0 and t1 = ε.
Assume that a subsolution (, M) satisfies K̃(t) ≤ K(0) for a.a. t ∈ (0, t1 ) and K̃(t) < K(0)
for a.a. t ∈ (t0 , t1 ) for some 0 ≤ t0 < t1 ≤ T. W.l.o.g. t1 < T. Fix any T1 ∈ (t1 , T). Let γ
be a sufficiently large number that satisfies (4.12) a.e. on (0, T1 ) × Ω and other lower bounds to
RT
be determined below. As in Remark 4.2, we set ϟ(t) := exp(−γt) and Ϟ(t) := t 1 ϟ(s) ds, and
infer that (4.1) holds on (0, T1 ) × Ω.
We now claim that Z T1
ϟ(t)(K̃(t) − K(t)) dt < 0, (4.13)
t0
which implies that
Z T1
ϟ(t)(K̃(t) − K(t)) dt < 0. (4.14)
0
Since  is a strong solution, its total entropy K(t) = K(0), so (4.14) is equivalent to
Z T1 Z T1
ϟ(t)K̃(t) dt < ϟ(t)K(0) dt = Ϟ(0)K(0).
0 0
Hence,
T1 T1
1 1
Z Z
ϟ(t)(M(t, ·), ) dt < ϟ(t)(F(0), ) dt + Ϟ(0)K(0) = Ϟ(0)K0 . (4.15)
2 0 2 0
It follows from Theorem 4.1 that
Ĩ(0 , T1 ) = Ϟ(0)K0 . (4.16)
It is easy to conclude from Definition 3.10 that the restriction (, M)|(0,T1 )×Ω is a subsolution on
(0, T1 ). Thus in order to get a contradiction it is enough to put together (3.41) with T = T1 ,
(4.15) and (4.16).
It remains to prove (4.13). Let
Z T1
ε := ϟ(t)(K̃(t) − K(t)) dt.
t1
Then our claim (4.13) becomes
Z t1
ϟ(t)(K̃(t) − K(t)) dt < −ε. (4.17)
t0
20 D. VOROTNIKOV

If ε ≤ 0, (4.17) is obvious, so let us assume ε > 0. Observe now that

Z T1
ε ≤ ϟ(t1 ) |K̃ (t) − K(t)| dt ≤ Cε ϟ(t1 ),
0

where Cε does not depend on γ. Fix any point t2 ∈ (t0 , t1 ). Let γ be so large that
Z t2
exp(γ(t1 − t2 )) (K(t) − K̃(t)) dt ≥ Cε .
t0

Then we conclude that

Z t1 Z t2
ϟ(t)(K̃(t) − K(t)) dt < ϟ(t)(K̃(t) − K(t)) dt
t0 t0
Z t2 Z t2
≤ ϟ(t2 ) (K̃(t) − K(t)) dt = exp(−γt2 ) (K̃(t) − K(t)) dt
t0 t0
≤ − exp(−γt1 )Cε ≤ −ε.


Remark 4.4 (Comparison with [3]). The results of this section are new even for the quadratic
case F() =  ⊗ , and are consequently applicable to the incompressible Euler equation, cf.
[11, 57]. Another global in time consistency result for the dual formulation of the incompressible
Euler has recently been obtained in [3, Theorem 3.6]. Nevertheless, our results (Theorem 4.1 and
Remark 4.2) seem to be of a different nature, and, furthermore, the two attitudes complement
each other to a certain degree. Indeed, [3, Theorem 3.6] states that if the strong solution  to the
incompressible Euler coincides with the “base state” ̃, then the solution of the dual problem is
identically zero (in the variables used in [3]). Thus the information about the strong solution is
contained not in the solution of the dual problem but in the “base state” only. Moreover, the proof
of [3, Theorem 3.6] ignores the formal conservativity of the problem (and therefore can be extended
to the Navier-Stokes). In contrast, our “base state” is identically zero (see Remark 3.18), and the
information about the strong solution is contained in the solution (B+ , E+ ) of the dual problem.
This information can be retrieved by formula (4.4) that strongly relies on the formal conservativity
(3.7).

5. Solvability of the dual problem

In this section we prove solvability of the dual problem (3.46) under an additional technical
assumption on the operator L.
Definition 5.1 (Strong trace condition). The operator L is said to satisfy the strong trace condition
if there exists a uniform constant c such that for any ζ ∈ D(L∗ ) such that the eigenvalues of the
matrix −L∗ ζ() are uniformly bounded from above by a constant k for a.e.  ∈ Ω, the eigenvalues
of the matrix L∗ ζ() are also uniformly bounded from above a.e. in Ω by ck.
Remark 5.2. The strong trace condition is particularly satisfied provided
L(q) = 0 (5.1)
 DAFERMOS’ PRINCIPLE 21

for any smooth scalar function q(). Indeed, it suffices to observe that in this case the trace of the
matrix function L∗ ζ vanishes almost everywhere because

(Tr(L∗ ζ), q) = (L∗ ζ, q) = (ζ, L(q)) = 0.

Remark 5.3. Although the strong trace condition is not very restrictive and holds in many situations,
cf. Section 7, see also [57], it fails for the system of conservation laws (1.2) with a symmetric flux
matrix F. Indeed, the adjoint of − div is the symmetric part of the Jacobian matrix, and there is
no way to control its eigenvalues from above if they are bounded from below. However, in many
cases it possible to recast the problem in a form that admits the strong trace condition. In Section
7 we will show how to implement this for the scalar conservation laws.

Theorem 5.4 (Existence in anisotropic Orlicz spaces). Assume that L satisfies the strong trace
condition. Then for any 0 ∈ LK (Ω; Rn ) there exists a maximizer (E, B) to (3.46) in the class
(3.19), and

Ϟ(0)|Ω| Tr F(0)
0≤ ≤ J˜(0 , T) < +∞.
2

Proof. It suffices to consider the pairs (E, B) that meet the restrictions (3.17), (3.44). Testing (3.46)
with E = 0, B = 0 as in Remark 3.17, we see that

Ϟ(0)|Ω| Tr F(0)
J˜(0 , T) ≥ ≥ 0.
2

Let (Em , Bm ) be a maximizing sequence. Since 0 ≤ J˜(0 , T), without loss of generality we may
assume that
Z T
0≤− (0 , Em ) dt + K(Em , Bm ). (5.2)
0

The eigenvalues of −Bm are uniformly bounded from above because ϟ + 2Bm  0. Since the
strong trace condition is assumed, a uniform L∞ bound on Bm follows directly from Definition 5.1.
In other words, ϟ + 2Bm  ϟc with some constant c > 0 a.e. in (0, T) × Ω. By the definition of
K in (3.45), we have

T
ϟc
Z • ˜
K(Em , Bm ) ≤ inf (z, Em ) + (M, ) dt
F(z)≤M 0 2
T
1
Z   
= inf (z, Em ) + ϟc, K(z) + Tr F(0) dt
z∈LK 0 2
Z T  
Em

cϞ(0)|Ω| Tr F(0)
∗
=− ϟc, K − dt + . (5.3)
0 ϟc 2
22 D. VOROTNIKOV

Employing the Fenchel–Young inequality we infer that

Z T ZT
cϞ(0)|Ω| Tr F(0)
 
∗
Em
ϟc, K − dt ≤ − (0 , Em ) dt +
0 ϟc 0 2
ZT
cϞ(0)|Ω| Tr F(0)
 
ϟc Em
= 20 , − dt +
0 2 ϟc 2
ZT ZT 
1 1 cϞ(0)|Ω| Tr F(0)
  
∗
Em
≤ (K(20 ), ϟc) dt + K − , ϟc dt + , (5.4)
0 2 0 2 ϟc 2
whence ZT 
1 Em
  
∗
K − , ϟc dt ≤ Ϟ(0)C, (5.5)
0 2 ϟc
where C depends only on the L1 -norm of K(20 ) that is finite. This together with (2.3) yields a
uniform LK∗ ((0, T) × Ω; Rn )-bound on Em . It follows from (5.3), (5.4) and (5.5) that the right-
hand side of (5.2) is uniformly bounded, whence J˜(0 , T) < +∞. The functional K is concave
and upper semicontinuous on LK∗ ((0, T) × Ω; Rn ) × L∞ ((0, T) × Ω; Rsn×n ) as an infimum of affine
RT
continuous functionals, cf. (3.45). The functional 0 (0 , ·) dt is a linear bounded functional on
LK∗ ((0, T) × Ω; Rn ). Consequently, every weak-∗ accumulation point of (Em , Bm ) is a maximizer
of (3.46). Note that the constraints (3.17), (3.44) are preserved by the limit. 
Remark 5.5 (Back to the original problem). Let (E, B) be any maximizer of (3.46) satisfying (3.19).
Mimicking (4.4), we can define a “generalized solution” to (3.1) by setting
ZT
#
1
 (t, ) := − E(s, ) ds, (t, ) := ∇K∗ ( # (t, )). (5.6)
Ϟ(t) t
This object  automatically belongs to the same regularity class as the strong solutions. Indeed,
∂t [Ϟ # ] = E ∈ LK∗ ((0, T) × Ω; Rn ).
On the other hand, since
1 1
, ®
Ϟ(t) T − t
our anisotropic Hardy inequality (see Appendix A) implies that  # ∈ LK∗ ((0, T) × Ω; Rn ), whence
 ∈ LK ((0, T) × Ω; Rn ) by (3.3). Finally,  # (t) ∈ D(L∗ ) for a.a. t ∈ (0, T), and
ϞL∗ ( # ) = B ∈ L∞ ((0, T) × Ω; Rsn×n ). (5.7)
Indeed, let ψ : (0, T) × Ω → RN×N
s
be an arbitrary smooth compactly supported matrix field, and
Rt
set Ψ(t, ·) := 0 ψ(τ, ·) dτ; integrating by parts and using (3.17), we deduce that
Z T Z T
(B(t, ·), ψ(t, ·)) dt = − (E(t, ·), LΨ(t, ·)) dt =
0 0
T T T
Z ‚Z Œ Z
=− E(τ, ·) dτ, Lψ(t, ·) dt = (Ϟ(t)L∗  # (t, ·), ψ(t, ·)) dt.
0 t 0
However, at this level of generality,  is not necessarily a strong solution.
 DAFERMOS’ PRINCIPLE 23

6. Uniqueness of strong solutions

We believe that it is possible to prove the weak-strong uniqueness property in the sense that
existence of a strong solution (as in Definition 3.13) implies that the corresponding solution (4.2)
to the dual problem (3.42) is unique in the class (3.19). The corresponding result in the quadratic
case F() =  ⊗  was established in [57, Section 5]. As the first step towards this conjecture, we
show here that a strong solution is always unique. The proof heavily relies on Λ-convexity of F.

Theorem 6.1 (Uniqueness of strong solutions). Let ,  be two strong solutions to (3.1) with the
same initial datum 0 ∈ LK (Ω; Rn ). Then, for every t ∈ [ 0, T), (t, ·) = (t, ·) a.e. in Ω.

Proof. To fix the ideas, we will assume that ,  are regular enough to perform the manipulations
below. (The proof in the general case follows exactly the same strategy with some tedious and
rather standard technicalities: in the (comparatively easy) quadratic case F() = ⊗  the rigorous
implementation can be found in [57, Lemma 5.2]). In order to avoid heavy notation, we will write
(t) and (t), and often even  and , instead of (t, ·) and (t, ·).
The key idea is to look at the evolution of the “Jeffreys divergence”7

ϛ(t) := ((t) − (t), # (t) −  # (t)) = ((t) − (t), ∇K()(t) − ∇K()(t)).

By strict convexity of K, ϛ(t) ≥ 0, and ϛ(t) = 0 if and only if (t) = (t). Obviously, ϛ(0) = 0,
and it is enough to show that
ϛ(t) = 0, t ∈ [ 0, T1 ] (6.1)

for any T1 < T.

Using (3.1) and (3.9), we compute

ϛ′ (t) = (∂t (t) − ∂t (t), # (t) −  # (t)) + ((t) − (t), ∂t # (t) − ∂t  # (t))
= (F() − F(), L∗ (# ) − L∗ ( # ))
n
X
+ ( −  , L∗ ( # ) : ∂ F() − L∗ (# ) : ∂ F())
=1
n
‚ Œ
X
∗ #
= −L ( ), F() − F() − ( −  ) : ∂ F()
=1
n
‚ Œ
X
∗ #
+ −L ( ), F() − F() − ( −  ) : ∂ F() . (6.2)
=1

It follows from Definition 3.13, cf. (3.24), that L∗ (# ) and L∗ ( # ) are essentially bounded on
[ 0, T1 ] × Ω. Thus there exists a constant γ > 0 such that γ − L∗ (# ) and γ − L∗ ( # )
are positive-definite matrix functions a.e. on [ 0, T1 ] × Ω, and for some uniform c > 0 we have

7The classical Jeffreys divergence corresponds to the case n = 1 and K() =  log , which is however ruled out
by our assumptions.
24 D. VOROTNIKOV

−L∗ (# ) − L∗ ( # )  c. Consequently,

n
‚ Œ
X
′ ∗ #
ϛ (t) = γ − L ( ), F() − F() − ( −  ) : ∂ F()
=1
n
‚ Œ
X
∗ #
+ γ − L ( ), F() − F() − ( −  ) : ∂ F()
=1
Xn
−γ ( −  , Tr[ ∂ F() − ∂ F()]) . (6.3)
=1

Due to the Λ-convexity of F, the functions y 7→ (γ − L∗ (# (t, ))) : F(y) and y 7→ (γ −
L∗ ( # (t, )))
: F(y) are convex w.r.t. y ∈ Rn for a.e. fixed pair of parameters (t, ) ∈ [ 0, T1 ] ×
Ω. On the other hand, it is easy to see that for any convex function  : Rn → R we have
0 ≤ (y) − (ξ) − (y − ξ) · ∇(ξ) ≤ (y − ξ) · (∇(y) − ∇(ξ)), y, ξ ∈ Rn . (6.4)
Leveraging (6.4), we derive from (6.3) that
n
‚ Œ
X
′ ∗ #
ϛ (t) ≤ γ − L ( ), ( −  ) : (∂ F() − ∂ F())
=1
n
‚ Œ
X
∗ #
+ γ − L ( ), ( −  ) : (∂ F() − ∂ F())
=1
n
X
− 2γ ( −  , ∂ K() − ∂ K())
=1
n
‚ Œ
X
( −  ) : (∂ F() − ∂ F()) − 2γ  − , # −  #


≤ (2γ + c),
=1
n
X
= (4γ + 2c) ( −  , ∂ K() − ∂ K()) − 2γϛ(t) = 2(c + γ)ϛ(t),
=1
and (6.1) follows by Grönwall’s inequality. 

7. Applications to PDEs
In this section present three relatively simple examples of dispersive equations to which our
abstract theory is germane (NLS, NLKG and GKdV equations with generic defocusing power-
law nonlinearities). We however believe that it is possible to go much beyond these examples by
considering PDEs of higher order (as, e.g., in [41, 10]), with nonlocal terms (e.g., of Benjamin-Ono
or Hartree type) or more general nonlinearities. In all these examples, we will be able to verify the
assumptions of Section 3 and of Theorem 5.4. Consequently, the theorems of Sections 4, 5 and 6
are fully applicable here.
To the best of our knowledge, global in time solvability for these PDEs is merely known under
significant restrictions on the exponents in the power laws or on the size of the initial data, cf.
[55, 44, 27]. Theorem 5.4 and Remark 5.5 provide existence of certain “dual variational solutions”
for these problems without such restrictions.
As a warm-up exercise, we show how our theory can be applied to scalar conservation laws.
 DAFERMOS’ PRINCIPLE 25

7.1. Scalar conservation laws. In this example we further restrict ourselves to Ω = T1 . Consider
the scalar conservation law
∂t  = −∂ (K()), (7.1)
where K : R → R is as in Assumption 3.1. Note that there is a tiny “anisotropy” because K does
not need to be even. As anticipated in Remark 5.3, the operator −∂ does not satisfy the strong
trace condition, so Theorem 5.4 is not directly applicable. In order to overcome this obstacle, we
let
n = 1, N = 2,
 ‹
K() K()
F() = .
K() K()
1
Obviously, K = 2 Tr(F − F(0)) and  # = K′ ().
We now define the operator L by the formula
 ‹
11 12
L = −∂ 12 .
12 22

Then we can rewrite (7.1) in the abstract form (3.1) with L and F that we have introduced.
Condition (5.1) is obviously true, so the strong trace condition holds. Moreover,

∗
1 0 ∂ κ
‹
L (κ) = .
2 ∂ κ 0

Let us check the conservativity condition (3.7). Remembering that we are now working on T1 ,
we easily see that

(F(), L∗ ( # )) = (K(), ∂ (K′ ())) = −(∂ , (K′ ())2 ) = 0.

It is clear that F  0 and F is Loewner convex. The subspace Λ consists of the elements of the
form
 ‹
11 12
.
12 11
The validity of Assumptions 3.2 and 3.3 is now obvious.

7.2. GKdV. We keep assuming Ω = T1 . Consider the “defocusing” generalized Korteweg-de Vries
equation [53, 27, 47] (GKdV)8

∂t  + ∂ξξξ  = ||α ∂ξ , (0) = 0 . (7.2)

The unknown is  : [ 0, T] × Ω → R, and α ≥ 1 is a prescribed constant (not necessarily integer).

Performing a simple change of variable  := ξ − t we rewrite this equation as follows:

∂t  + ∂  = (||α + 1)∂ , (0) = 0 . (7.3)

8To simplify the presentation we consider the real-valued equation, but the same approach works for the complex-
valued case.
26 D. VOROTNIKOV

Letting  := ∂  we can rewrite the problem in the following way:

1
∂t  = −∂  + ∂ (||α ) + ∂ , (7.4)
α+1
1
∂t  = −∂  + ∂ (||α ) + ∂ , (7.5)
α+1
(0) = 0 , (0) = 0 . (7.6)
Rigorously speaking, this is an extended system, but what is important is that it still possesses
an “entropy” that is formally conserved even if we do not assume the compatibility condition
0 = ∂ 0 .
We let
n = 2, N = 4,
 = (, ),
̄ = (1, , , ||α ),
1 ||α+2
K() = (2 + 2 ) + ,
2 (α + 2)(α + 1)

1 1
   
2
F() = ε̄ ⊗ ̄ + K() + 1 e1 ⊗ e1 + −ε + K() e2 ⊗ e2
2 2
1 1
   
2 2α
+ −ε + K() e3 ⊗ e3 + −ε|| + K() e4 ⊗ e4 ,
2 2
1
where small ε > 0 will be selected later. It is straightforward to check that K = 2
Tr(F − F(0))
and that K fits into the framework of Assumption 3.1. Furthermore,
 # = (z, ) = ( + (α + 1)−1 ||α , );
observe that the second component coincides with .
Define the operator L by the formula
11 12 13 14
 
1
‚ Œ
12 22 23 24  −1 −∂ 13 + α+1 ∂ 24 + ∂ 12
L =ε 1 .
13 23 33 34  −∂ 13 + α+1 ∂ 24 + ∂ 12
14 24 34 44
Then (7.4)–(7.5) can be written in the abstract form (3.1).
Condition (5.1) is obviously true, so the strong trace condition holds. Moreover,

 ‹ ε−1
κ
L∗ =
θ 2
0 −∂ κ + ∂ θ −∂ κ + ∂ θ 0
 
 −∂ κ + ∂ θ 1 1
0 0 − α+1 ∂ κ + ∂ θ
α+1   .
×
−∂ κ + ∂ θ 0 0 0 
1 1
0 − α+1 ∂ κ + α+1 ∂ θ 0 0
 DAFERMOS’ PRINCIPLE 27

Let us check the conservativity condition (3.7). We compute, integrating by parts where needed,

(F(), L∗ ( # )) =
1 1
 
α
(, −∂ z + ∂ ) + (, −∂ z + ∂ ) + || , − ∂ z + ∂ 
α+1 α+1
= (, −∂ z + ∂ ) + (∂ , −z + ∂ ) + (z − , −∂ z + ∂ )
= (∂ , ∂ ) + (z, −∂ z) = 0.
Observe that Λ consists of the elements of the form
11 12 13 0
 
12 11 0 24 

13 0 11 0 
0 24 0 11
Let us check that F ≥ 0 and that F is Λ-convex. Fix any matrix
p11 p12 p13 0
 
 12 p11
p 0 p24 
P=  0.
p13 0 p11 0 
0 p24 0 p11
Then
F() : P = εp11 + ε2 p11 + ε2 p11 + ε||2α p11
+ 2εp12 + 2εp13 + 2ε||α p24
1 1
 
2
+ K()p11 + p11 + −ε + K() p11
2 2
1 1
   
2 2α
+ −ε + K() p11 + −ε|| + K() p11
2 2
α+2
– ™
||
= ε + 1 + 2 + 2 + 2 p11 + 2εp12 + 2εp13 + 2ε||α p24 (7.7)
(α + 2)(α + 1)
Since P  0,
mx{|p12 |, |p13 |, |p24 |} ≤ p11 . (7.8)
Thus the function in (7.7) is non-negative provided ε is sufficiently small. The Hessian of the
function in (7.7) w.r.t.  = (, ) is
 
2p11 + 2||α p11 + 2εα(α + 1)||α−2 p24 0
.
0 2p11
This matrix is always non-negative-definite provided ε is less than or equal to a certain constant
that depends only on α.
We now we need to compute the partial derivatives of the non-diagonal components of πΛ F w.r.t.
the components of :
∂ F12 = ε,
∂ F13 = 0,
∂ F24 = ε(α + 1)||α ® ||α+1 + 1 ® | # | + 1,
28 D. VOROTNIKOV

∂ F12 = 0,
∂ F13 = ε,
∂ F24 = 0.
This immediately implies (3.6).
Remark 7.1 (Lack of Loewner convexity). It is interesting to observe the full Loewner convexity is
missing for the matrix function F above. Indeed, let
0 0 0 0
 
0 0 0 0
P=  0.
0 0 1 1
0 0 1 1
Then, for any α ≥ 1 and any fixed ε > 0, the function

α
1 2 2
||α+2
(, ) 7→ F() : P = K() + 2ε||  = ( +  ) + + 2ε||α 
2 (α + 2)(α + 1)
is obviously not convex on R2 . Similar counterexamples can be constructed for the systems discussed
in the sequel.
Remark 7.2 (Sharp formulation of GKdV). For any α ≥ 1, using the obtained expressions of the
partial derivatives of F, it is elementary to see that the “sharp” problem (3.9) for GKdV reads
∂t z + (1 + (|z|))(−∂ z + ∂ ) = 0,
∂t  − ∂ z + ∂  = 0,
where  is the inverse of the function s 7→ s1/ α (1 + (α + 1)−1 s).
7.3. Defocusing NLS. Let Ω = Td . Consider the nonlinear Schrödinger equation with a generic
defocusing power-law nonlinearity
i∂t Ψ = −ΔΨ + |Ψ|2q Ψ, Ψ(0) = Ψ0 , (7.9)
where q ≥ 12 is a given constant. The unknown is Ψ : [ 0, T] × Ω → C.
We first change the variable ψ := Ψe−it to rewrite this in the form
i∂t ψ = −Δψ + |ψ|2q ψ + ψ, ψ(0) = Ψ0 . (7.10)
We now let  = Reψ, b = Imψ, ϝ = ∇, ϐ = ∇b. Then the system becomes
∂t  = − div ϐ + (2 + b2 )q b + b, (7.11)
∂t b = div ϝ − (2 + b2 )q  − , (7.12)
2 2 q
∂t ϝ = −Δϐ + ∇(( + b ) b + b), (7.13)
∂t ϐ = Δϝ − ∇((2 + b2 )q  − ). (7.14)
As in the previous example, this is an extended system, but we will see that it still possesses an
“entropy” that is formally conserved.
We let
n = 2d + 2, N = 2d + 4,
 = (, b, ϝ, ϐ),
 DAFERMOS’ PRINCIPLE 29

̄ = (1, , b, ϝ, ϐ, (2 + b2 )q ),
1 1
 
2 2 2 2 2 2 q+1
K() = ϝ +ϐ + +b + ( + b ) ,
2 q+1

2 2
   
2
F() = ε̄ ⊗ ̄ + K() + 1 e1 ⊗ e1 + −ε + K() e2 ⊗ e2
N N

2
 d 
X 2

2 2
+ −εb + K() e3 ⊗ e3 + −ε|ϝm | + K() e3+m ⊗ e3+m
N m=1
N
d 
2 2
X   
2 2 2 2q
+ −ε|ϐm | + K() e3+d+m ⊗ e3+d+m + −ε( + b ) + K() eN ⊗ eN ,
m=1 N N
1
where small ε > 0 will be selected later. It is straightforward to check that K = 2
Tr(F − F(0))
and that
 # = (y, z, ϝ, ϐ) = ( + (2 + b2 )q , b + (2 + b2 )q b, ϝ, ϐ).
Define the operator L by the formula
 
11 12 13 A14 A15 16
 12 22 23 A24 A25 26  
− div A15 + 36 + 13

 13 23 33 A34 A35 36 
 
−1  div A 14 −  26 −  12 
LA ⊤ A⊤ A⊤ A46  = ε −ΔA15 + ∇(36 + 13 ) .

A44 A45
 14 24 34
A ⊤ A⊤ A⊤ A⊤

15 25 35 45
A55 A56  ΔA14 − ∇(26 + 12 )
16 26 36 A⊤
46
A⊤
56
66
Then (7.11)–(7.14) can be rewritten in the abstract form (3.1).
Condition (5.1) is obviously true, so the strong trace condition holds. Moreover,

κ
 
−1
 θ ε
L∗ =
η
 
2
ξ
 
0 −θ + div ξ κ − div η −∇θ + Δξ ∇κ − Δη 0
 −θ + div ξ 0 0 0 0 −θ + div ξ
 
 κ − div η 0 0 0 0 κ − div η 
×
(−∇θ + Δξ)⊤
.
 0 0 0 0 0 

 (∇κ − Δη)⊤ 0 0 0 0 0 
0 −θ + div ξ κ − div η 0 0 0
Let us check the conservativity condition (3.7). We compute, integrating by parts where needed,

(F(), L∗ ( # )) = (, −z + div ϐ) + (b, y − div ϝ) + (ϝ, −∇z + Δϐ) + (ϐ, ∇y − Δϝ)
+ ((2 + b2 )q , −z + div ϐ) + (b(2 + b2 )q , y − div ϝ)
= (y, −z + div ϐ) + (z, y − div ϝ) + (div ϝ, z) − (∇ϝ, ∇ϐ) − (div ϐ, y) + (∇ϐ, ∇ϝ) = 0.
30 D. VOROTNIKOV

Observe now that Λ consists of the elements of the form

 
11 12 13 A14 A15 0
 12 11 0 0 0 26 
 
 13 0 11 0 0 36 
 ⊤ ⊤ ⊤

A
 14
0 0  11  0 0 
A ⊤ 0 0 0 11  0 
15
0 26 36 0 0 11
Let us show that F ≥ 0 and that F is Λ-convex. Fix
 
p11 p12 p13 P14 P15 0
p12 p11 0 0 0 p26 
 
p13 0 p11 0 0 p36 
P=   0.
P⊤ 0⊤ 0⊤ p11  0 0 
 14 
P⊤ 0 0 0 p11  0 
15
0 p26 p36 0 0 p11
Then, in the same spirit as for the GKdV, we compute
1
 
2 2 2 2 2 2 q+1
F() : P = ε + 1 + ϝ + ϐ +  + b + ( + b ) p11
q+1
+ 2εp12 + 2εbp13 + 2εP14 ϝ + 2εP15 ϐ
+ 2ε(2 + b2 )q p26 + 2ε(2 + b2 )q bp36 . (7.15)
Since P  0, it is tedious but elementary to check that the function in (7.15) and its Hessian w.r.t.
 = (, b, ϝ, ϐ) are non-negative and non-negative-definite, resp., provided ε is controlled by a
constant that depends only on q ≥ 12 .
It remains to estimate the partial derivatives of the non-diagonal components of πΛ F w.r.t. the
components of . The only potentially problematic terms (because of the presence of a nonlinearity)
are
∂ F26 = ε(2 + b2 )q + 2εq2 (2 + b2 )q−1 ,
∂b F26 = 2εqb(2 + b2 )q−1 ,
∂ F36 = 2εqb(2 + b2 )q−1 ,
∂b F36 = ε(2 + b2 )q + 2εqb2 (2 + b2 )q−1 .
All of them are
2q+1
® (2 + b2 )q ® (2 + b2 ) 2 + 1 ® | # | + 1,
which yields validity of assumption (3.6).

8. Complex NLKG
Let Ω = Td . Consider the complex Klein-Gordon equation with a generic (defocusing) power-law
nonlinearity
∂tt ψ − Δψ + |ψ|2q ψ + ψ = 0, ψ(0) = ψ0 . (8.1)
1
Here q ≥ 2
is a given constant. The unknown is ψ : [ 0, T] × Ω → C.
 DAFERMOS’ PRINCIPLE 31

We let  = Reψ, b = Imψ, ϝ = ∇, ϐ = ∇b,  = ∂t ,  = ∂t b. Then the system can be

rewritten in the following form:

∂t  = , (8.2)
∂t b = , (8.3)
∂t ϝ = ∇, (8.4)
∂t ϐ = ∇, (8.5)
2 2 q
∂t  = div ϝ − ( + b )  − , (8.6)
∂t  = div ϐ − (2 + b2 )q b − b. (8.7)

As above, this is an extended system, but what matters is that it is formally conservative.
We let

n = 2d + 4, N = 2d + 6,  = (, b, ϝ, ϐ, , ),

̄ = (1, , b, ϝ, ϐ, , , (2 + b2 )q ),

1 1
 
2 2 2 2 2 2 2 2 q+1
K() =  + +ϝ +ϐ + +b + ( + b ) ,
2 q+ 1

2 2
   
2
F() = ε̄ ⊗ ̄ + K() + 1 e1 ⊗ e1 + −ε + K() e2 ⊗ e2
N N

2
 d 
X 2

+ −εb2 + K() e3 ⊗ e3 + −ε|ϝm |2 + K() e3+m ⊗ e3+m
N m=1
N
d 
2
X 
+ −ε|ϐm |2 + K() e3+d+m ⊗ e3+d+m
m=1 N
2
 
2
+ −ε + K() eN−2 ⊗ eN−2
N
2
 
2
+ −ε + K() eN−1 ⊗ eN−1
N
2
 
2 2 2q
+ −ε( + b ) + K() eN ⊗ eN ,
N

1
where small ε > 0 will be selected later. It is straightforward to check that K = 2
Tr(F − F(0)).
We have

 # = (y, z, ϝ, ϐ, , ) = ( + (2 + b2 )q , b + b(2 + b2 )q , ϝ, ϐ, , ).

32 D. VOROTNIKOV

We now set
 
11 12 13 A14 A15 16 17 18
 12 22 23 A24 A25 26 27 28  16
 
 
 13 23 33 A34 A35 36 37 38 
 ⊤ 17
A⊤ A⊤
  
A A44 A45 A46 A47 A48  ∇16
 
 14 24 34 −1 
L A ⊤ = ε .
 
 15 A⊤
25
A⊤
35
A⊤
45
A55 A56 A57 A58 
 
 ∇17 
 16 26 36 A⊤ A⊤ 66 67 68  div A14 − (28 + 12 )
  
46 56
A⊤ A⊤ div A15 − (38 + 13 )
 
 17 27 37 47 57
67 77 78 
18 28 38 A⊤
48
A⊤
58
68 78 88
Then (7.11)–(7.14) can be rewritten in the abstract form (3.1).
It is clear that
0 −χ −ω −∇χ −∇ω κ − div η θ − div ξ 0
 
κ  −χ 0 0 0 0 0 0 −χ 
 
θ  −ω 0 0 0 0 0 0 −ω


  ε−1  ⊤
η −∇χ 0 0 0 0 0 0 0 

L∗   = .
  
ξ

−∇ω ⊤ 0 0 0 0 0 0 0 
  2 
χ
κ − div η 0 0 0 0 0 0 0 
 
ω 
θ − div ξ 0 0 0 0 0 0 0

0 −χ −ω 0 0 0 0 0
The validity of Assumptions 3.2 and 3.3 and of condition (5.1) can be shown in a way similar to
the NLS case. Let us just check Assumption 3.7. We compute, integrating by parts where needed,

(F(), L∗ ( # )) = (, −) + (b, −) + (ϝ, −∇) + (ϐ, −∇) + (, y − div ϝ)
+ (, z − div ϐ) + ((2 + b2 )q , −) + (b(2 + b2 )q , −)
= (y, −) + (z, −) + (div ϝ, ) + (div ϐ, ) + (∇ϐ, ∇ϝ)
+ (, y − div ϝ) + (, z − div ϐ) = 0.
Acknowledgment. The author thanks Yann Brenier and Iván Moyano for inspiring discussions.

Appendix A. An anisotropic Hardy inequality

Proposition A.1. Let K be a C1 -smooth N-function such that K and K∗ satisfy the Δ2 -condition.
Given ƒ ∈ LK ((0, T) × Ω; Rn ), define the function
1 t
Z
F(t, ) := ƒ (s, ) ds.
t 0
Then
F ∈ LK ((0, T) × Ω; Rn )
and
kFkLK ((0,T)×Ω) ® kƒ kLK ((0,T)×Ω) . (A.1)
Proof. It follows from [17, Lemma 2.3.16] that there exist r > 0 and p > 1 such that
 · ∇K()
≥p (A.2)
K()
 DAFERMOS’ PRINCIPLE 33

for || ≥ r.
We first observe that
K(s) ≤ K()sp , s ∈ (0, 1),  ∈ Rn , |s| ≥ r. (A.3)
Indeed, fix s0 ∈ (0, 1) with |s0 | ≥ r, and consider the function
ϕ(s) = log K(s) − log(K()sp ), s ∈ [ s0 , 1].
Then
s · ∇K(s)
ϕ′ (s) = s−1 − s−1 p ≥ 0.
K(s)
p
Since ϕ(1) = 0, we infer K(s0 ) ≤ K()s0 .
It follows from (A.3) that
K() ≤ sK s−1/ p , s ∈ (0, 1), || ≥ r.


(A.4)


Denote Ks () := sK s−1/ p and ƒs (t, ) := ƒ (st, ), s ∈ (0, 1). Then
Z1
F(t, ) = ƒs (t, ) ds (A.5)
0
and
K() ≤ Ks () , || ≥ r. (A.6)
We now claim that
kƒs kLKs ((0,T)×Ω) ≤ s−1/ p kƒ kLK ((0,T)×Ω) . (A.7)
Indeed,
¨ Z   «
ƒ (ts, )
kƒs kLKs ((0,T)×Ω) = inf λ > 0 : Ks d dt ≤ 1
(0,T)×Ω λ
¨ «
ƒ (t, )
Z  
= inf λ > 0 : Ks d dt ≤ s
(0,sT)×Ω λ
¨ «
ƒ (t, )
Z  
≤ inf λ > 0 : Ks d dt ≤ s
(0,T)×Ω λ
¨ Z   «
ƒ (t, )
= inf λ > 0 : K d dt ≤ 1
(0,T)×Ω λs1/ p
= s−1/ p kƒ kLK ((0,T)×Ω) .
Employing convexity of the Luxemburg norm, Proposition 2.9, (A.5) and (A.6), and integrating
(A.7) w.r.t. s ∈ (0, 1), we conclude that
Z1
kFkLK ((0,T)×Ω) = ƒs ds
0 LK ((0,T)×Ω)
Z 1 Z 1
pCr
≤ kƒs kLK ((0,T)×Ω) ds ≤ Cr kƒs kLKs ((0,T)×Ω) ds ≤ kƒ kLK ((0,T)×Ω) .
0 0 p−1

34 D. VOROTNIKOV

Appendix B. Ballistic optimal transport

Let us briefly describe the link of our dual problem (3.46) with the optimal transportation
problems. We employ a geometric intuition that did not explictly appear in the previous works.
Let us start from the following heuristic setting. Let M be a complete, connected Riemannian
manifold. Let X : M → TM be a given vector field on M, and let T ∈ M be a prescribed point.
By a ballistic 9 geodesic problem we mean finding the curve10 γt ⊂ M that minimizes
1 T
Z
min 〈γ̇t , γ̇t 〉γt dt, (B.1)
γ̇ 0 =X(γ 0 ), γ T =T 2 0

i.e., we know the final position and the initial direction of an unknown geodesic.
Assume that the vector field X that determines the direction of the “howitzer” at every point
of M is a potential field, i.e., X = grd  for some  : M → R. Then (B.1) is equivalent to the
Hopf-Lax formula
1
min (0 ) + dst 2 (0 , T ), (B.2)
0 ∈M 2
where dst is the Riemannian distance on M (here we are merely interested in the minimizer 0
that determines the starting point of the unknown geodesic and not in the optimal value itself).
Formula (B.2) makes sense even if we have no manifold structure and M is merely a metric space,
not necessarily connected. This is particularly relevant for the metric spaces with “Riemannian
flavour”, see [49, p. 6] for a list of examples of such spaces. Problem (B.2) indeed appears in
infinite-dimensional and metric-geometric contexts, including the context of optimal transport.
The generalization to p > 1, i.e.,
1
min (0 ) + dst p (0 , T ) (B.3)
0 ∈M p
was studied in [5] in a metric context. Problems of this kind are called marginal entropy-transport
problems in [43]. They are crucial ingredients of de Giorgi’s minimizing movement scheme, also
known as the JKO scheme [34, 4].
Consider now the quadratic Hamilton-Jacobi equation
1
|∇ψ|2 = 0, ψ(0, ) = ψ0 (), (t, ) ∈ [ 0, T] × Ω.
∂t ψ +
2
Setting  = ∇ψ, we rewrite it in the form
1
∂t  + ∇ Tr( ⊗ ) = 0, (0) = ∇ψ0 . (B.4)
2
Note that (B.4) is a particular case of our abstract equation (3.1) with F() =  ⊗ , L = − 21 ∇ Tr.
Let ϟ ≡ 1. As explained in [57, Section 2.2], the dual problem (3.46) for the quadratic Hamilton-
Jacobi equation may be rewritten (this is formal but can be made rigorous) as
1 T
Z Z Z
− ψ0 ρ(0) d − ρ−1 |q|2 d dt → sp (B.5)
Ω 2 0 Ω

9The wording is borrowed from [7].

10The subscript t is not a time derivative but just the value at time t.
 DAFERMOS’ PRINCIPLE 35

subject to the constraints

∂t ρ + div q = 0, ρ(T) = 1, ρ ≥ 0. (B.6)
R
Rescaling if needed, we can assume T = 1 and |Ω| = 1. Defining the functional Ψ(ρ) := Ω ψ0 dρ
on the Wasserstein space P(Ω) of probability measures on Ω, multiplying by −1 and leveraging
the Benamou-Brenier formula [56, 8], we can recast (B.5), (B.6) in the Hopf-Lax form
1
min Ψ(ρ0 ) + W22 (ρ0 , ρ1 ). (B.7)
ρ 0 ∈P (Ω) 2
Here W2 is the quadratic Wasserstein distance [56], and dρ1 is the Lebesgue measure d on Ω.
Hence, in this very particular but instructive case, the dual problem (3.46) can be viewed as a
ballistic problem on the Wasserstein space.
The observations above can be generalized to the case of the p-Wasserstein space (and very likely
to the Orlicz-Wasserstein spaces [54, 5]). The generating Hamilton-Jacobi equation is
p−1
∂t  + ∇ Tr F() = 0
p
with
2−p
F() = || p−1  ⊗ . (B.8)
We omit the implementation but call attention to the details that the matrix function F in (B.8)
is Λ-convex but not Loewner convex, and that the resulting Hopf-Lax formula is similar to (B.3)
because the p-Wasserstein space is not ‘Riemannian-like” unless p = 2.

References
[1] A. Acharya. Variational principles for nonlinear PDE systems via duality. Quart. Appl. Math., 81(1):127–140,
2023.
[2] A. Acharya and A. N. Sengupta. Action principles for dissipative, non-holonomic Newtonian mechanics. Proc.
A., 480(2293):Paper No. 20240113, 21, 2024.
[3] A. Acharya, B. Stroffolini, and A. Zarnescu. Variational dual solutions for incompressible fluids. arXiv preprint
arXiv:2409.04911, 2024.
[4] L. Ambrosio, N. Gigli, and G. Savaré. Gradient flows in metric spaces and in the space of probability measures.
Lectures in Mathematics ETH Zürich. Birkhäuser Verlag, Basel, second edition, 2008.
[5] L. Ambrosio, N. Gigli, and G. Savaré. Density of lipschitz functions and equivalence of weak gradients in metric
measure spaces. Revista Matemática Iberoamericana, 29(3):969–996, 2013.
[6] A. Baradat and L. Monsaingeon. Small noise limit and convexity for generalized incompressible flows, Schrödinger
problems, and optimal transport. Arch. Ration. Mech. Anal., 235(2):1357–1403, 2020.
[7] A. Barton and N. Ghoussoub. Dynamic and stochastic propagation of the Brenier optimal mass transport.
European Journal of Applied Mathematics, 30(6):1264–1299, 2019.
[8] J.-D. Benamou and Y. Brenier. A computational fluid mechanics solution to the Monge-Kantorovich mass
transfer problem. Numer. Math., 84(3):375–393, 2000.
[9] J. M. Borwein and J. D. Vanderwerff. Convex functions: constructions, characterizations and counterexamples,
volume 109 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 2010.
[10] T. Boulenger and E. Lenzmann. Blowup for biharmonic NLS. Ann. Sci. Éc. Norm. Supér. (4), 50(3):503–544,
2017.
[11] Y. Brenier. The initial value problem for the Euler equations of incompressible fluids viewed as a concave
maximization problem. Comm. Math. Phys., 364(2):579–605, 2018.
[12] Y. Brenier. Examples of hidden convexity in nonlinear pdes. Preprint, 2020.
[13] Y. Brenier and I. Moyano. Relaxed solutions for incompressible inviscid flows: a variational and gravitational
approximation to the initial value problem. Philos. Trans. Roy. Soc. A, 380(2219):Paper No. 20210078, 12, 2022.
36 D. VOROTNIKOV

[14] A. C. Bronzi, M. C. Lopes Filho, and H. J. Nussenzveig Lopes. Wild solutions for 2d incompressible ideal flow
with passive tracer. Communications in Mathematical Sciences, 13(5):1333–1343, 2015.
[15] Á. Castro, D. Faraco, and B. Gebhard. Entropy solutions to macroscopic ipm. arXiv preprint arXiv:2309.03637,
2023.
[16] E. Chiodaroli and O. Kreml. On the energy dissipation rate of solutions to the compressible isentropic Euler
system. Arch. Ration. Mech. Anal., 214(3):1019–1049, 2014.
[17] I. Chlebicka, P. Gwiazda, A. Świerczewska-Gwiazda, and A. Wróblewska-Kamińska. Partial differential equa-
tions in anisotropic Musielak-Orlicz spaces. Springer Monographs in Mathematics. Springer, Cham, 2021.
[18] T. Cieślak and G. Jamróz. Maximal dissipation in Hunter-Saxton equation for bounded energy initial data. Adv.
Math., 290:590–613, 2016.
[19] C. M. Dafermos. The entropy rate admissibility criterion for solutions of hyperbolic conservation laws. J. Dif-
ferential Equations, 14:202–212, 1973.
[20] C. M. Dafermos. Maximal dissipation in equations of evolution. J. Differential Equations, 252(1):567–587, 2012.
[21] S. Daneri and L. Székelyhidi, Jr. Non-uniqueness and h-principle for Hölder-continuous weak solutions of the
Euler equations. Arch. Ration. Mech. Anal., 224(2):471–514, 2017.
[22] C. De Lellis and L. Székelyhidi, Jr. On admissibility criteria for weak solutions of the Euler equations. Arch.
Ration. Mech. Anal., 195(1):225–260, 2010.
[23] C. de Lellis and L. Székelyhidi Jr. The Euler equations as a differential inclusion. Annals of mathematics,
170(3):1417–1436, 2009.
[24] R. J. DiPerna. Uniqueness of solutions to hyperbolic conservation laws. Indiana Univ. Math. J., 28(1):137–188,
1979.
[25] T. K. Donaldson and N. S. Trudinger. Orlicz-Sobolev spaces and imbedding theorems. J. Functional Analysis,
8:52–75, 1971.
[26] E. Feireisl. Maximal dissipation and well-posedness for the compressible Euler system. J. Math. Fluid Mech.,
16(3):447–461, 2014.
[27] I. Friedman, O. Riaño, S. Roudenko, D. Son, and K. Yang. Well-posedness and dynamics of solutions to the
generalized KdV with low power nonlinearity. Nonlinearity, 36(1):584–635, 2023.
[28] B. Gebhard, J. Hirsch, and J. J. Kolumbán. On a degenerate elliptic problem arising in the least action principle
for Rayleigh-Taylor subsolutions. Ann. Inst. H. Poincaré C Anal. Non Linéaire, 41(6):1527–1594, 2024.
[29] B. Gebhard and J. J. Kolumbán. Relaxation of the Boussinesq system and applications to the Rayleigh-Taylor
instability. NoDEA Nonlinear Differential Equations Appl., 29(1):Paper No. 7, 38, 2022.
[30] H. Gimperlein, M. Grinfeld, R. J. Knops, and M. Slemrod. The least action admissibility principle. arXiv
preprint arXiv:2409.07191, 2024.
[31] J. Glimm, D. Lazarev, and G.-Q. G. Chen. Maximum entropy production as a necessary admissibility condition
for the fluid Navier–Stokes and Euler equations. SN Applied Sciences, 2:1–9, 2020.
[32] F. Höfer and N. A. Nikov. On growth of Sobolev norms for periodic nonlinear Schrödinger and generalised
Korteweg-de Vries equations under critical Gibbs dynamics. arXiv preprint arXiv:2412.08630, 2024.
[33] L. Hsiao. The entropy rate admissibility criterion in gas dynamics. J. Differential Equations, 38(2):226–238,
1980.
[34] R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck equation. SIAM J.
Math. Anal., 29(1):1–17, 1998.
[35] S.-C. Klein. Using the Dafermos entropy rate criterion in numerical schemes. BIT, 62(4):1673–1701, 2022.
[36] S.-C. Klein. Stabilizing discontinuous Galerkin methods using Dafermos’ entropy rate criterion: I—One-
dimensional conservation laws. J. Sci. Comput., 95(2):Paper No. 55, 37, 2023.
[37] S.-C. Klein. Stabilizing discontinuous Galerkin methods using Dafermos’ entropy rate criterion: II—Systems of
conservation laws and entropy inequality predictors. J. Sci. Comput., 100(2):Paper No. 42, 49, 2024.
[38] M. A. Krasnoselskii. Topological methods in the theory of nonlinear integral equations. The Macmillan Company,
New York, 1964.
[39] M. A. Krasnoselskii and J. B. Rutickii. Convex functions and Orlicz spaces. P. Noordhoff Ltd., Groningen, 1961.
[40] F. Kraus. Über konvexe Matrixfunktionen. Math. Z., 41(1):18–42, 1936.
[41] P. Krejčı́ and I. Straškraba. A uniqueness criterion for the Riemann problem. Hiroshima Math. J., 27(2):307–346,
1997.
[42] C. Léonard. From the Schrödinger problem to the Monge-Kantorovich problem. J. Funct. Anal., 262(4):1879–
1920, 2012.
 DAFERMOS’ PRINCIPLE 37

[43] M. Liero, A. Mielke, and G. Savaré. Optimal entropy-transport problems and a new Hellinger-Kantorovich
distance between positive measures. Invent. Math., 211(3):969–1117, 2018.
[44] F. Linares and G. Ponce. Introduction to nonlinear dispersive equations. Springer, 2014.
[45] K. Löwner. Über monotone Matrixfunktionen. Math. Z., 38(1):177–216, 1934.
[46] L. M. Martyushev and V. D. Seleznev. Maximum entropy production principle in physics, chemistry and biology.
Phys. Rep., 426(1):1–45, 2006.
[47] S. Masaki and J.-i. Segata. Existence of a minimal non-scattering solution to the mass-subcritical generalized
Korteweg–de Vries equation. Ann. Inst. H. Poincaré C Anal. Non Linéaire, 35(2):283–326, 2018.
[48] L. Monsaingeon, L. Tamanini, and D. Vorotnikov. The dynamical Schrödinger problem in abstract metric spaces.
Adv. Math., 426:Paper No. 109100, 2023.
[49] M. Muratori and G. Savaré. Gradient flows and Evolution Variational Inequalities in metric spaces. I: Structural
properties. Journal of Functional Analysis, 278(4):108347, 2020.
[50] I. Prigogine. Introduction to thermodynamics of irreversible processes. Interscience Publishers, New York-
London, revised edition, 1961.
[51] Y. Shenfeld. Matrix displacement convexity along density flows. Arch. Ration. Mech. Anal., 248(5):Paper No.
74, 41, 2024.
[52] S. Singh, J. Ginster, and A. Acharya. A hidden convexity of nonlinear elasticity. J. Elasticity, 156(3):975–1014,
2024.
[53] W. A. Strauss. Nonlinear scattering theory at low energy. J. Functional Analysis, 41(1):110–133, 1981.
[54] K.-T. Sturm. Generalized orlicz spaces and wasserstein distances for convex–concave scale functions. Bulletin
des sciences mathématiques, 135(6-7):795–802, 2011.
[55] C. Sulem and P.-L. Sulem. The nonlinear Schrödinger equation: self-focusing and wave collapse. Springer
Science & Business Media, 2007.
[56] C. Villani. Topics in optimal transportation. American Mathematical Soc., 2003.
[57] D. Vorotnikov. Partial differential equations with quadratic nonlinearities viewed as matrix-valued optimal
ballistic transport problems. Arch. Ration. Mech. Anal., 243(3):1653–1698, 2022.
[58] H. Ziegler. Some extremum principles in irreversible thermodynamics with application to continuum mechanics.
In Progress in solid mechanics, Vol. IV, pages 91–193. North-Holland, Amsterdam, 1963.

(D. Vorotnikov) University of Coimbra, CMUC, Department of Mathematics, 3001-501 Coimbra, Portu-
gal
Email address: mitvorot@mat.uc.pt