Operadores Elípticos de Quarta Ordem Com Condições de Contorno de Wentzell em Domínios de Lipschitz

J. Evol. Equ.
(2024)24:86
© 2024 The Author(s)
Journal of Evolution
https://doi.org/10.1007/s00028-024-01015-z Equations
Elliptic fourth-order operators with Wentzell boundary conditions

on Lipschitz domains
David Ploß
Abstract. For bounded domains with Lipschitz boundary , we investigate boundary value problems
for elliptic operators with variable coefficients of fourth-order subject to Wentzell (or dynamic) boundary
conditions. Using form methods, we begin by showing general results for an even wider class of operators
of type

B∗ B 0
A= b ,
−N B γ
where B is associated to a quadratic form b and N b an abstractly defined co-normal Neumann trace.
Even in this general setting, we prove generation of an analytic semigroup on the product space H :=
L 2 () × L 2 (). Using recent results concerning weak co-normal traces, we apply our abstract theory to
the elliptic fourth-order case and are able to fully characterize the domain in terms of Sobolev regularity for
operators in divergence form B = − div Q∇ with Q ∈ C 1,1 (, Rd×d ), also obtaining Hölder-regularity
of solutions. Finally, we also discuss asymptotic behavior and (eventual) positivity.
1. Introduction
Wentzell, or dynamic boundary conditions, appear in a multitude of physical ap-

plications and pose a mathematically challenging problem. Given a bounded domain
with boundary , they model the interchange of free energy of a physical system
between and . The main issue with modeling this interchange is that the energy
flux is represented by an integral over the domain, which cannot “see” the boundary as
it is a set of Lebesgue measure zero. This is usually resolved by considering functions
in a product space, e.g., H := L 2 () × L 2 (), and a related operator A for which the
action in the interior of the domain and on the boundary is decoupled (cf. [1,3,15]).
The connection between interior and boundary is then encoded by a coupling condition
in the definition of the domain of A.
Mathematics Subject Classification: 35K35 (primary); 47A07, 47D06, 35B65 (secondary)

Keywords: Boundary value problem, Lipschitz boundary, Wentzell boundary condition, Generalized
trace, Analytic semigroup, Eventual positivity.
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID
258734477—SFB 1173.
The author thanks Robert Denk and Markus Kunze for helpful comments and discussions.
0123456789().: V,-vol
86 Page 2 of 43 D. Ploß J. Evol. Equ.
A detailed discussion of a physical interpretation of these boundary conditions in

comparison to classical ones can be found in [19] for the heat and the wave equation.
Other instances where these boundary conditions occur are the Stefan problem with
surface tension (see [16, Section 1]), and climate models including coupling between
the deep ocean and the surface (see [13, Section 2]) where they incorporate the external
energy transported into the ocean by the sun. Furthermore, they are used in the Cahn–
Hilliard equation describing spinodal decomposition of binary polymer mixtures (see
[27, Section 1]) in order to model effects close to the boundary, e.g., that one of
the agents is more attracted to the boundary than the other, which might lead to
further separation effects. Contrary to the first examples whose leading part is given by
(variations of) the Laplacian, the Cahn–Hilliard equation is based on the Bi-Laplacian,
an operator of order 4 which fits into the setting of the present work.
We are going to consider a general class of operators. As prototype and main ap-
plication, we study the following system of fourth order:
∂t u + B(α B)u = 0 in (0, ∞) × , (1.1)

tr B(α B)u − β∂νQ (α B)u − βδ tr(α B)u − γ tr u = 0 on (0, ∞) × , (1.2)
∂νQ u + δ tr u = 0 on (0, ∞) × , (1.3)
u|t=0 = u 0 in . (1.4)
Here B is given by B = − div Q∇, where Q ∈ C 1,1 (, Rd×d ) is uniformly positive
Q
definite and is a Lipschitz domain. Let ∂ν denote its corresponding co-normal
derivative given by ν, Q∇u (cf. (3.2) below), and assume α, β, γ , δ to be bounded,
real-valued functions. The precise smoothness assumption of the coefficients will be
specified later on (cf. Hypotheses 2.9 and 3.1). In (1.1)–(1.4), it is implicitly assumed
that the initial value u 0 is sufficiently smooth to have a trace on the boundary and that
this trace is used as an initial condition for u on the boundary.
Note that, as Eq. (1.1) is of fourth order with respect to x ∈ , we have to impose
two boundary conditions. Here, we have chosen the Robin boundary condition (1.3)
in addition to the Wentzell boundary condition (1.2).
The main mathematical challenge in tackling Wentzell boundary conditions lies
in the fact that the elliptic operator that governs the equation in the interior itself
appears in the boundary condition, and the standard condition B(α B)u ∈ L 2 () is
not sufficient to guarantee existence of the trace. In order to decouple this system and
circumvent this issue, we rewrite the Wentzell boundary condition (1.2) as a dynamic
boundary condition using B(α B)u = −∂t u from (1.1). Then we rename u to u 1 and
replace the time derivative ∂t u 1 in the boundary condition by the time derivative ∂t u 2
of an independent function u 2 that lives on the boundary. Even though u 2 is formally
independent of u 1 , we think of u 2 as the trace of u 1 ; this condition will be incorporated
into the domain of our operator D(A), later. We thus obtain the following decoupled
J. Evol. Equ. Elliptic fourth-order operators Page 3 of 43 86
version of (1.1)–(1.4):
∂t u 1 + B(α B)u 1 = 0 in (0, ∞) × , (1.5)

∂t u 2 + β∂νQ (α B)u 1 + βδ tr(α B)u 1 + γ u 2 = 0 on (0, ∞) × , (1.6)
∂νQ u 1 + δu 2 = 0 on (0, ∞) × , (1.7)
u 1 |t=0 = u 1,0 in , (1.8)
u 2 |t=0 = u 2,0 on . (1.9)
Note that, as u 2 is independent of u 1 , we have to impose an additional initial condition

for u 2 . If, however, the initial value u 0 in (1.4) is smooth enough, we can put u 1,0 =
u 0 and u 2,0 = u 0 | . However, this coupling condition for the initial value is not
mandatory. The Hilbert space theory established below will allow to prescribe non-
continuous initial data, even the extreme case of u 1,0 = 0 and u 2,0 ∈ L 2 () arbitrary
is allowed. This is especially useful to model situations where in the beginning the
entire energy only lives on the boundary and slowly dissipates into the interior of the
domain over time. Rewritten as a Cauchy problem, for u = (u 1 , u 2 ) ∈ H we obtain
∂t u + Au = 0 for u(t, ·) in H, (1.10)

u|t=0 = u0 for u0 ∈ H, (1.11)
where A is given by

B(α B) 0
Q (1.12)
β(∂ν (α B) + δ tr(α B)) γ
on a suitable domain D(A) that incorporates (1.7) and the coupling condition u 2 =
tr u 1 . In order to construct a solution for (1.5)–(1.9), the main idea is to obtain an
analytic semigroup (generated by −A), whose smoothing effects will allow us to
recover the original system with Wentzell boundary conditions. Using this decoupling
idea and form methods to tackle Wentzell boundary conditions, has proven to be
a suitable approach for the second-order case, e.g., the Laplace operator subject to
Wentzell boundary conditions. A series of papers starting in 2003 has shown generation
results concerning an analytic semigroup for the decoupled system on L 2 ()× L 2 (),
using the classical Beurling–Deny criteria [3]. These results were then extended to
the L p -scale, and later also to general second-order elliptic operators on Lipschitz
domains, where also Hölder continuity of the solution was deduced, see [23] and [29].
Under additional smoothness assumptions also spaces of continuous functions were
considered in [3]; see also [14] and [4] where generation of an analytic semigroup
was shown in an abstract perturbation framework. For higher-order elliptic operators
the extension procedure to the L p -scale does not work, because the Beurling–Deny
criteria are in general not fulfilled (see also Proposition 4.15). Less results are available
and they typically rely on being in a smooth setting. For fourth-order equations with
sufficiently smooth coefficients in C 4 -domains, it was shown in [17, Theorem 2.1]
that the related operator in the product space is essentially self-adjoint. For the Cahn–
Hilliard equation, classical well-posedness was shown in [27, Theorem 5.1] in the
L 2 -setting, and in [26, Theorem 2.1] in the L p -setting. Again the domain and the
coefficients were assumed to be (sufficiently) smooth, and the methods do not carry
over to Lipschitz domains.
In [11], the Lipschitz-case was solved for the Bi-Laplacian using weak Green’s
formulae and the theory of quasi-boundary triples [6, Chapter 8]. In the present paper,
however, we choose a more abstract approach which deals with a larger class of
systems and does not depend on the theory of boundary triples. It contains the results
of [11] as a special case, also giving a simpler proof employing recent developments
on co-normal derivatives in Lipschitz domains for operators with variable coefficients
[5].
To that end, in Sect. 2.1, we begin by addressing very general forms b whose asso-
ciated operators will take the role of B in (1.12). To proceed, we start by considering
the Neumann case δ = 0 and define an abstract Neumann trace N b that fits into the
form approach and is connected to Green’s second formula. Afterward, in Sect. 2.2, we
investigate the quadratic form a on the product space H = L 2 () × L 2 () to which
the operator A is associated. Based on the analysis of that form, we can show that
the operator A is self-adjoint and −A is the generator of a strongly continuous and
analytic semigroup (T(t))t≥0 (Theorem 2.12). This will also show that the operator A
indeed governs a generalized version of (1.5)–(1.9) with δ = 0, given by (2.4)–(2.8).
We will explain that we can also obtain a solution of the Wentzell system in the original
formulation generalizing (1.1)–(1.3) with initial condition (1.4). If u 2,0 is not the trace
of u 1,0 , there are some subtleties concerning the initial values, see Remark 2.14. In
Sect. 3, we return to the main application where B is a second-order elliptic operator in
divergence form, identifying the associated operator B and its minimal and maximal
realization, as well as the normal trace N b (Sect. 3.2). In Sect. 3.3, we finally collect
our results for the fourth-order system, which culminate in Corollary 3.20 where we
precisely identify the operator A and its domain in terms of Sobolev regularity. After
extending our results to the Robin case δ > 0, we obtain that the operator A indeed
governs the system precisely as formulated in (1.5)–(1.9).
In Sect. 4, we briefly discuss higher regularity for smoother domains and coefficients
(Sect. 4.1) before undertaking further investigations of the operator A in the original
setting. One of the main results of this section is Theorem 4.8, which states that for
every element (u 1 , u 2 ) of D(A ∞ ) the function u 1 is Hölder continuous and u 2 is the
trace of u 1 . As the semigroup T is analytic, it follows that for positive time the solution
of (1.5)–(1.9) is Hölder continuous and satisfies the Wentzell boundary condition in
a pointwise sense. Moreover, this result implies regularity of the eigenfunctions of
the operator A and is used later on. In Sect. 4.3, we show that the operator A has
compact resolvent and thus a decomposition into a basis consisting of eigenfunctions
of A. This allows us to describe the semigroup in terms of the eigenfunctions and
to characterize the asymptotic behavior of the semigroup. In particular, we study its
positivity properties: It turns out that the generated semigroup is neither positive nor
L ∞ -contractive (Proposition 4.15) as the operator does not satisfy the Beurling–Deny
criteria. However, as shown for the for the semigroup generated by the Bi-Laplacian
in [11], in the case γ = δ = 0 our semigroup is again eventually (strongly) positive in
the sense of [9] and [8] (Theorem 4.18). We close the article by showing that the same
abstract approach can be used to obtain abstract results for higher-order operators,
e.g., (− )4k .
2. The abstract setting
Aim of this section is to establish a solution theory for Wentzell boundary conditions
for higher-order operators which can be represented by nested forms, i.e. two quadratic
forms where the operator associated to the first one is used to construct the second.
We fix the following setting:
Let ⊂ Rd be a domain with Lipschitz boundary . We denote the inner products
in L 2 () and L 2 () by

f, g := f g dx and f, g := f g dS,

respectively, and write · and · for the induced norms. We denote the standard
Sobolev spaces by H s () for s ≥ 0. By slight abuse of notation, we will also write

d
∇u, ∇v := ∂ j u∂ j v dx
j=1
whenever u, v ∈ H 1 (). For fractional orders we may either use complex interpola-
tion, or, equivalently, restriction. For negative orders we employ duality. Additionally
for an elliptic operator of second-order B, we introduce the space H Bs () for the
space of functions u ∈ H s () such that Bu belongs to L 2 () (cf. Definition 3.12).
We endow H Bs () with the canonical norm
u 2
H Bs () := u 2
H s () + Bu 2
(u ∈ H Bs ()).
The Dirichlet-trace on C ∞ (), defined by u → u | , and its extension to any Sobolev

space H s () for s > 21 is denoted by tr.
Taking a brief look back to the Bi-Laplacian case (cf. [11]), where the form a(u, v) =
N u, N v is considered on the domain
{u = (u 1 , u 2 ) ∈ H | u 1 ∈ D( N ), u 2 = tr u 1 },
we recall that its associated operator is given by

2
0
A= .
−∂ν 0
In order to tackle the general system, in the form a, we are going to replace the Neumann
Laplacian N by a more general operator B N . To that end, we introduce a second form
that somehow operates on a “lower level”. More precisely, N is naturally associated
to the form b(u, v) = ∇u, ∇v on L 2 () with form domain D(b) = H 1 (), so we
may generalize this form. In order to distinguish between a and b terminologically,
we will call a the primary form and b the subsidiary form.
At first, we will establish our theory for quite general subsidiary quadratic forms
b whose associated operators are not necessarily differential operators in divergence
form or even of second order.
2.1. Abstract realizations of the lower-order operator
Recall, that if a form b : D(b) × D(b) → C (D(b) ⊂ H ) is densely-defined,

semi-bounded by λ ∈ R, closed, and continuous in the sense of [24, Chapter 1], its
associated operator A satisfies that λ − A generates an analytic contraction semigroup
on H. We call such forms generating. Note that, in this terminology, b is semi-bounded
by λ if the shifted form bλ (u, v) = b(u, v) + λ u, v H is accretive. Furthermore, A
will be self-adjoint if a is also symmetric.
Definition 2.1. Consider the Hilbert space H = L 2 (). We call a form b : D(b) ×
D(b) → C admissible, if it is a generating, symmetric form on H such that for some
ρ ∈ (0, 1)
1
Cc∞ () ⊆ (D(b), · b) ⊆ H 2 +ρ () (2.1)
holds, where the latter embedding is continuous and dense.

Remark 2.2. The continuous embedding into the space H 1/2+ρ () is assumed to
ensure existence of the Dirichlet trace. The space H 1/2+ρ () can be replaced by any
space on which the Dirichlet trace exists and is bounded, and its range embeds densely
1/2
into L 2 (), as for example H or some variant of it, if such an embedding of the
form domain is known. However, in this abstract setting, we want to avoid spaces
depending on specific operators.
Next, we introduce two operators, connected to the subsidiary form b. We think of
them as realizations of a certain “genera” operator B subject to Neumann or Dirichlet
boundary conditions. While this is indeed true in the setting of elliptic differential
operators on domains (cf. Bmax in Definition 3.12), we point out that in the abstract
setting considered here, it is unclear which manner to define such an operator would be
the most sensible. Therefore, in this section the operator B0∗ will be used as a suitable
substitute for the formally undefined operator B, which only appears terminologically
in the following definition.
Definition 2.3. Let b : D(b) × D(b) → L 2 () be an admissible form.
(i) Denote by λb the maximal semi-bound λ ∈ R such that Re b(u, u) ≥ λ u 2,

i.e. we have u 2b = Re b(u, u) + (1 − λb ) u 2 for u ∈ D(b).
(ii) The operator B N associated to b on L 2 () is called the Neumann realization of

B.
(iii) The Dirichlet realization of B, denoted by B D , is the associated operator to the
from b D , given by the restriction of b to {u ∈ D(b) | tr u = 0}.
Proposition 2.4. If b is an admissible form, then b D is generating and symmetric, so
B D is well defined and self-adjoint. Furthermore, we have
1/2+ρ
D(B N ) ∩ D(B D ) = {u ∈ D(B N ) | tr u = 0} = D(B N ) ∩ H0 () (2.2)
and
B N u = B D u for u ∈ D(B D ) ∩ D(B N ).
Proof. The restricted form b D is clearly symmetric, and densely defined as the test
functions still lie in D(b D ). By definition, we have · b = · b D on D(b D ), from
which continuity and semi-boundedness follow. In order to show that (D(b D ), . b D )
is complete as well, take a Cauchy-sequence u n with respect to · b D = · b . As
b is a closed form, u n converges to some u ∈ D(b), and by continuous embedding
also in H 1/2+ρ (). By continuity of the Dirichlet trace, the traces converge as well,
whence tr u = 0 and u ∈ D(b D ) as desired. Hence it is also generating.
We verify the second part in (2.2) first: In Lipschitz domains we have the identity
{u ∈ H s () | tr u = 0} = H0s () for all s ∈ (1/2, 3/2) (cf. [5, Equation (3.7)]).
1/2+ρ
So we directly obtain D(B N ) ∩ H0 () = {u ∈ D(B N ) | tr u = 0} due to
D(B N ) ⊂ D(b) ⊂ H 1/2+ρ ().
For the remaining identity assume u ∈ D(B N ) with tr u = 0. Hence there is
an f N ∈ L 2 () such that, for all v ∈ D(b), f N , v = b(u, v) holds. But now
u ∈ D(b D ) and in particular for all v ∈ D(b D ) ⊂ D(b), we have
f N , v = b(u, v) = b D (u, v),
which shows u ∈ D(B D ) and B D u = f N = B N u. The converse is trivial.

Remark 2.5. (i) Note that b D itself is not admissible, as D(b D ) cannot be embedded
densely into H 1/2+ρ () due to the continuity of the trace, which is the reason
why pure Dirichlet-Wentzell boundary conditions (without Neumann term) can
not be handled via this method.
(ii) In the following, we are going to assume that D(B D ) ∩ D(B N ) is dense in
L 2 (), which is useful to define realizations of B that are in a sense minimal
(or maximal) and still densely defined. We will see below, in Proposition 2.11,
that this is a very natural assumption for considering the primary form a, as it
will ensure that it will be densely defined as well. The simplest way to ensure
this density will be to demand that D(B D ) ∩ D(B N ) contains the test functions.
However, this excludes the case of operators of the form − div Q∇ for Q only
in L ∞ (, Rd×d ). As we assume more regularity on Q in this article, anyway,
this will be no restriction for us, though.
Next, we introduce two notions of generalized weak Neumann traces, the first of
which is connected to a generalization of Green’s first formula, while the second is
closer related to the abstract notion of the associated operator and Green’s second
formula.
Definition 2.6. Let b be an admissible form. Assume that D(B D ) ∩ D(B N ) is dense
in L 2 (). Define B0 = B N | D(B D )∩D(B N ) . Let N b : D(N b ) ⊆ L 2 () → L 2 () be
the linear operator defined by
D(N b ) := {u ∈ D(b) ∩ D(B0∗ ) |

∃g ∈ L 2 () ∀v ∈ D(b) : B0∗ u, v
− b(u, v) = g, tr v }
and N b u := g. Let furthermore N b : D(N ) ⊆ L 2 () → L 2 () be the linear

operator defined by
D(N b ) := {u ∈ D(B0∗ ) |

∃g ∈ L 2 ()∀v ∈ D(B N ) : B0∗ u, v
− u, B N v = g, tr v }
and N b u := g.
We want to point out a subtlety concerning the signs: In comparison to the usual weak
Neumann trace (cf. (3.2) below) N b and N b generalize −∂ν , as B0∗ is a generalized
version of − .
We begin with a very simple observation that will prove to be quite useful to show
equality of different traces.
Lemma 2.7. Consider two linear operators S1 : D(S1 ) ⊆ V → W , S2 : D(S2 ) ⊆
V → W on vector spaces V, W . If S1 ⊆ S2 , S1 is surjective, and ker(S2 ) ⊆ D(S1 ),
then S1 = S2 .
Proof. Let u ∈ D(S2 ). As S1 is surjective there is some v ∈ D(S1 ) with S2 v = S1 v =
S2 u. So u − v ∈ ker(S2 ) ⊆ D(S1 ) whence also u = v + (u − v) ∈ D(S1 ) and
S2 u = S1 u. This shows S2 ⊆ S1 and thus equality.
We come to our first main result, which shows that the traces N b and N b are well
defined, i.e. the assigned element g ∈ L 2 () is indeed unique in both cases.
Theorem 2.8. Let b be an admissible form. Assume that D(B D ) ∩ D(B N ) is dense in
L 2 (). Then B0 , B N , N b , and N b (cf. Definitions 2.3 and 2.6) satisfy the following
properties.
(i) B0 is a densely defined, symmetric, and closed operator. Therefore, B0∗ and B0∗∗
are well defined and we have B0 ⊆ B0∗ as well as B0∗∗ = B0 .
(ii) tr(D(B N )) is dense in L 2 ().
(iii) N b and N b are well defined, linear operators. We have N b | D(b)∩D(N b ) = N b
and ker N b = ker N b = D(B N ), which also shows that N b and N b are
densely defined.
(iv) If N b is surjective, we have N b = N b .

Proof. (i) The density follows by assumption, the closedness follows as both B D
and B N are closed by default (as b and b D are generating) and coincide on the
intersection. Furthermore, as a restriction of a self-adjoint operator, B0 has to be
symmetric.
(ii) Let f ∈ L 2 () and ε > 0. As H ρ () is dense in L 2 (), we find a function
u ∈ H ρ () with u − f 2 ≤ ε. Because tr : H 1/2+ρ () → H ρ () is
bounded (denote its operator norm by M) and surjective (cf. [18, Equation 2.7]),
we find a function u ∈ H 1/2+ρ () with tr u = u . As b is admissible, the
domain of the subsidiary form D(b) is densely and continuously embedded in
H 1/2+ρ (). Hence, there is a function û ∈ D(b) satisfying
û − u 2
H 1/2+ρ ()
≤ M −2 ε.
As for any generating form it is known that the domain of the associated operator
is a form core (cf. [24, Lemma 1.25]), one may further approximate and even
find a function ū ∈ D(B N ) such that
ū − û 2
H 1/2+ρ ()
≤ C ū − û 2
b ≤ M −2 ε.
Altogether, we have
tr ū − f 2
≤ 2 tr u − f 2
+ 2 tr u − tr ū 2

≤ 2ε + 2M 2 ū − û + û − u 2
H 1/2+ρ ()
≤ 10ε
as desired.
(iii) The linearity of the operators is obvious. Concerning the well-definedness, as-
sume there were two elements g1 , g2 ∈ L 2 () satisfying the defining condi-
tions, respectively. Then we have g1 , tr v = g2 , tr v in particular for all
v ∈ D(B N ), and hence g1 − g2 , tr v = 0. But as tr(D(B N )) is dense in
L 2 () due to (ii), this implies g1 = g2 . For u ∈ D(b), v ∈ D(B N ) we have
b(u, v) = b(v, u) = B N v, u = u, B N v . This shows N b ⊆ N b .
Concerning the restriction, we assume u ∈ D(N b ) ∩ D(b). Then, as before,
there is a g ∈ L 2 () such that, for all v ∈ D(B N ) ⊆ D(b), B0∗ u, v −
u, B N v = g, tr v holds. As u ∈ D(b) and v ∈ D(B N ), this implies the
L 2 ()-function g also satisfies
∗
B0 u, v − b(u, v) = g, tr v (2.3)
for all v ∈ D(B N ). However, D(B N ) is a form core for b, so for any v ∈ D(b)
there is a sequence (vn )n ∈ D(B N ) with vn → v with respect to · b and
due to the admissibility of b also in H 1/2+ρ (), whence the trace converges as
well. Using this approximation, Formula (2.3) can be extended to all v ∈ D(b),
which indeed proves N b | D(b)∩D(N b ) = N b . Next we show N b (and thus
N b ) is densely defined. As B0 ⊆ B N and B N is self-adjoint, we have B N =

B N∗ ⊆ B0∗ which exists due to (i). Hence for u ∈ D(B N ) (which is a dense
subset of L 2 ()) we have B0∗ u, v − b(u, v) = B0∗ u, v − u, B N v =
B N u, v − u, B N v = 0 for all v ∈ D(B N ). So N b u = 0 for any u ∈
D(B N ). Furthermore, if u ∈ D(N b ) and N b u = 0, then for all v ∈ D(B N ) we
have B0∗ u, v −u, B N v = 0. This, however, is the definition of u ∈ D(B N∗ )
and shows B N∗ u = B0∗ u. As B N is self-adjoint, this means u ∈ D(B N ). Hence
D(B N ) ⊆ ker N b ⊆ ker N b ⊆ D(B N ), which shows equality, and in particular
that both operators are densely defined.
(iv) This is an immediate consequence of (iii) and Lemma 2.7.

2.2. The system on the product space
Next we introduce a primary form, which will be connected to the generalized

system of (1.5)–(1.9), i.e
∂t u 1 + B0∗ (α B N )u 1 = 0 in (0, ∞) × , (2.4)

b
∂t u 2 − βN (α B N )u 1 + γ u 2 = 0 on (0, ∞) × , (2.5)
N bu1 = 0 on (0, ∞) × , (2.6)
u 1 |t=0 = u 1,0 in , (2.7)
u 2 |t=0 = u 2,0 on . (2.8)
We define the following regularity assumptions for the coefficients.

Hypothesis 2.9. Let ⊆ Rd be a bounded domain with Lipschitz boundary . Con-
sider α ∈ L ∞ (, R) and β, γ , δ ∈ L ∞ (, R) such that there exists a constant η > 0
with α ≥ η almost everywhere on and β ≥ η almost everywhere on . Furthermore,
let δ ≥ 0.
Definition 2.10. Let b be an admissible form. Let B D and B N defined as in Defini-
tion 2.3 and assume Hypothesis 2.9.
(i) Let H := L 2 (, λd )× L 2 (, β −1 dS) be the Hilbert space, where λd denotes the
d-dimensional Lebesgue measure and dS the surface measure on , endowed
with the canonical inner product
u, vH = u 1 , v1 + u 2 , v2 ,β , (2.9)
where u = (u 1 , u 2 ), v = (v1 , v2 ) ∈ H and

u 2 , v2 ,β = β −1 u 2 , v2 = β −1 (x)u 2 (x) · v2 (x)dS.

(ii) Let D(B D ) ∩ D(B N ) be dense in L 2 (). Then, we define the primary form
a : D(a) × D(a) → C as
a(u, v) := α B N u 1 , B N v1 + γ u 2 , v2 ,β
for all u, v ∈ D(a) where
D(a) := {u = (u 1 , u 2 ) ∈ H | u 1 ∈ D(B N ), u 2 = tr u 1 }.
Proposition 2.11. Let b be an admissible form. We assume Hypothesis 2.9 and the
density of D(B D ) ∩ D(B N ) in L 2 (). Then, the primary form a is densely defined.
Proof. We may assume without loss of generality that β ≡ 1, otherwise switch to

an equivalent norm. Next we exploit the density of D(B N ) ∩ D(B D ) in L 2 (). As
D(B N ) ∩ D(B D ) × {0} ⊆ D(a), we have L 2 () × {0} ⊆ D(a). In order to show
{0} × L 2 () ⊆ D(a), we use Theorem 2.8 (ii) which yields that tr D(B N ) is dense
in L 2 (). Hence, given a function f ∈ L 2 () and some number ε > 0, there is
an element ū 1 of D(B N ) such that tr ū 1 − f 2 < ε. Finally, we pick a function
w ∈ D(B N ) ∩ D(B D ) such that ū 1 − w 2 ≤ ε and put u = (ū 1 − w, tr(ū 1 − w)) =
(ū 1 − w, tr ū 1 ). Then, by construction, we have
u − (0, f ) 2
H = ū 1 − w 2
+ tr ū 1 − f 2
≤ 2ε.
As f was arbitrary, {0} × L 2 () ⊆ D(a). Since D(a) is a vector space, we may
combine our two results and obtain D(a) = H.
Theorem 2.12. Let b be an admissible form. Assume that Hypothesis 2.9 holds and
that D(B D ) ∩ D(B N ) is dense in L 2 (). Then, the corresponding operators B0 , B N ,
N b , N b and the form a (cf. Definitions 2.3, 2.6, and 2.10) satisfy the following
properties.
(i) a is a generating, symmetric form. Hence the operator A associated to a on H
is self-adjoint and −A generates an analytic semigroup T on H.
(ii) A is given by

B0∗ (α B N ) 0
A= (2.10)
−βN b (α B N ) γ
on
D(A) = {u ∈ H | u 1 ∈ D(B N ), α B N u 1 ∈ D(N b ), u 2 = tr u 1 }.
(iii) In particular, for u0 = (u 1,0 , u 2,0 ) ∈ H the Cauchy problem (2.4)–(2.8) pos-
sesses a unique solution, which is given by u(t) = T(t)(u 1,0 , u 2,0 ) for t > 0. If
N b is additionally surjective, we may replace N b by N b in (ii) and (2.4)–(2.8).
Proof. (i) We begin by showing that a is a generating, symmetric form. As b is

admissible, we have D(b) ⊆ H 1/2+ρ (), hence also D(B N ) ⊆ H 1/2+ρ ()
and the condition tr u 1 = u 2 makes sense. The density of D(a) has been shown
in Proposition 2.11. Because of γ ∈ L ∞ () the form is semi-bounded due to
√ √
a(u, u) = α BN u1, α BN u1
+ γ u 2 , u 2 ,β ≥ − γ ∞ u H.
2
The symmetry is trivial as α, β, γ are real-valued. Next we consider the induced

norm u 2a = a(u, u) + (1 + γ ∞ ) u 2H . With respect to this norm the form
is continuous as
√ √
|a(u, v)| ≤ α B N u 1 α B N v1 + γ ∞ u 2 ,β v2 ,β ≤ 2 u a v a .
√
Note that by definition we have α B N u 1 2 ≤ u 2a , as well as γ ∞ u 2 2,β ≤
u 2a . Finally, we show the closedness of the form. Let (un )n ⊆ D(a) be a · a -
Cauchy sequence, where un = (u n1 , u n2 ). We have to prove that this sequence
converges with respect to · a . Let us first note that because α is bounded from
below by η > 0, for a certain constant C, we have
1 √
u1 2
B ≤ α BN u1 2
+ u1 2
≤C u 2
a
η
whenever u = (u 1 , u 2 ) ∈ D(a). It follows that (u n1 )n is a Cauchy sequence with
respect to · B . As B N is closed, we find some u ∈ D(B N ) such that u n1 → u in
L 2 () and B N u n1 → B N u in L 2 (). Next observe that for u ∈ D(B N ) ⊆ D(b),
by definition of the associated operator and Young’s inequality, we have
u 2 ≤ C u 2b = C(b(u, u) + (1 − λb ) u 2 ) = C((1 − λb ) u 2 + B N u, u )

H 1/2+ρ ()
BN u 2 + u 2 ) = C
≤ C( u 2
B
for some constant C ≥ 1. Combining this with the above, we observe that
u 1 is also convergent in H 1/2+ρ () whence, by the continuity of the trace,
n
u n2 = tr u n1 → tr u in L 2 (). Setting u = (u, tr u), we see that u ∈ D(a) and

un → u with respect to · a . This proves closedness of the form. Hence a
is a generating, symmetric form with a corresponding associated self-adjoint
operator A such that −A generates an analytic semigroup on H.
(ii) At first we define

B0∗ (α B N ) 0
C=
−βN b (α B N ) γ
on
D(C) = {u ∈ H | u 1 ∈ D(B N ), α B N u 1 ∈ D(N b ), u 2 = tr u 1 }.
We want to show C = A. We begin by showing C ⊆ A. So let u ∈ D(C) ⊆

D(a), i.e. u 1 ∈ D(B N ), α B N u 1 ∈ D(N b ) and tr u 1 = u 2 . Then we have for
all v ∈ D(a)
a(u, v) = α B N u 1 , B N v1 + γ u 2 , v2 ,β

= B0∗ (α B N u 1 ), v1 − N b (α B N u 1 ), tr v1 + γ u 2 , v2 ,β

∗ b
= B0 (α B N u 1 ), v1 + −βN (α B N u 1 ) + γ u 2 , v2 = Cu, vH .
,β
For the reverse direction let u ∈ D(A) and Au = f. Then u ∈ D(a) and for
any v ∈ D(a) we have a(u, v) = f, vH . In particular, for all v ∈ D(B0 ) × {0}
(and thus for all v1 ∈ D(B0 )) we have
f 1 , v1 = f, vH = a(u, v) = α B N u 1 , B N v1 = α B N u 1 , B0 v1 .
This shows that α B N u 1 is in D(B0∗ ) and f 1 = B0∗ (α B N u 1 ) by definition of the

adjoint. So for all v1 ∈ D(B N ) we have
f 2 , tr v1 ,β = a(u, v) − f 1 , v1

= α B N u 1 , B N v1 − B0∗ (α B N u 1 ), v1
+ γ u 2 , tr v1 ,β
or

β −1 ( f 2 − γ u 2 ), tr v1 = α B N u 1 , B N v1 − B0∗ (α B N u 1 ), v1

for all v1 ∈ D(B N ), which shows α B N u 1 ∈ D(N b ) (and u ∈ D(C )) as well

as
−N b (α B N u 1 ) = β −1 ( f 2 − γ u 2 )
or, equivalently, f 2 = −βN b (α B N u 1 ) + γ u 2 , which shows Cu = f = Au.

(iii) This follows from (i) and (ii) by standard semigroup theory. For the last part we
use Theorem 2.8 (iv).

Remark 2.13. Theorem 2.12 (iii) states that the semigroup T governs the system (2.4)–
(2.8). We observe that u 1 also solves the corresponding non-decoupled problem with
Wentzell boundary conditions
∂t u + B0∗ (α B N )u = 0 in (0, ∞) × , (2.11)

tr B0∗ (α B N )u b
+ βN (α B N )u − γ tr u = 0 on (0, ∞) × , (2.12)
b
N u=0 on (0, ∞) × , (2.13)
u|t=0 = u 0 in . (2.14)
As the semigroup is analytic, the solution is C ∞ in time so that

(u(t))t>0 = (T(t)(u 1,0 , u 2,0 ))t>0 satisfies Eqs. (2.4) and (2.5) in a classical (in time)
sense. Concerning the initial system (2.11)–(2.14), we immediately see that u = u 1
solves Eqs. (2.11), (2.13) and (2.14).
The question remains in which way the Wentzell boundary condition (2.12) is
satisfied. But as u ∈ C((0, ∞), D(A)2 ) due to the analyticity of the semigroup,
naturally for all t > 0 the functions u(t, ·) are in D(A 2 ) and thus we have
tr B0∗ (α B N )u = tr(Au)1 = (Au)2 = −βN b (α B N )u + γ tr u,
which shows (2.12). In fact, the analyticity even yields u(t, ·) ∈ D(A ∞ ) for t > 0.
Remark 2.14. We point out that the system (2.11)–(2.14) has to be interpreted in such a
way that u 0 is sufficiently smooth to have a trace, say u 0 ∈ H 1/2+ρ (); in this setting,
the solutions of (2.11)–(2.14) are in a one-to-one correspondence with the solutions
of (2.4)–(2.8) with u 1,0 = u 0 | and u 2,0 = u 0 | . In our semigroup approach, however,
u 2,0 can be chosen independently of u 1,0 and, by the above, all of these solutions are
(distinct!) solutions of (2.11)–(2.14). In a way, choosing u 2,0 different from tr u 1,0
corresponds precisely to having some free energy on the boundary, which was a main
motivation to consider Wentzell boundary conditions in the first place.
3. Application to strongly elliptic operators in divergence form
In this section, we will specify the operator B to be a strongly elliptic second-order

operator in divergence form and return to the investigation of the system (1.1)–(1.4).
We consider δ = 0 at first and deal with the Robin case at the end of Sect. 3.3. We
begin by settling the precise regularity assumptions on the matrix Q and recalling
some facts concerning different realizations of co-normal traces.
3.1. Co-normal traces
Hypothesis 3.1. Assume Q ∈ C 1,1 (, Rd×d ) to be symmetric and uniformly positive
definite, which means there is some open superset ⊆ Rd containing such that
Q ∈ C (, R
1,1 d×d ) is symmetric and satisfies for some κ Q > 0
, ξ ∈ Cd ).
Q(x)ξ, ξ Cd ≥ κ Q |ξ |2 (x ∈ (3.1)
Remark 3.2. The regularity Q ∈ C 1,1 (, ¯ Rd×d ) is not necessary for all the subsequent
steps, part of the theory can be done using only W 1,∞ -regularity. However C 1,1 is the
regularity from [5, Chapter 11], and thus used when we establish higher regularity and
a precise identification of the occurring traces and the domain of our operator. For a
finer distinction in regularity, we refer to [25, Hypothesis 2.5 and Section 3.1].
Definition 3.3. Let ⊆ Rd be a Lipschitz domain with outward normal ν. We

consider the following notions of strong traces:
¯ Rd×d ), we
(i) For a real-valued, uniformly positive definite matrix Q ∈ W 1,∞ (,
denote the co-normal Neumann trace of a function u ∈ C ∞ () by τ N u :=
Q
ν · tr Q∇u, where we read the operator tr component-wise.

(ii) For any function δ ∈ L ∞ (), we will call τδQ = τ NQ + δ tr the (co-normal) Robin
trace.
It is known, that the Dirichlet trace extends by continuity to a bounded linear sur-
jective operator

1 3
tr : H () → H
s s−1/2
() for all s ∈ ,
2 2
(cf. [18, Equation (2.7)]). In fact, this operator is even a retraction, i.e. there ex-
ists a continuous right-inverse. Even for smooth domains, however, the continuity
of tr : H s () → H s−1/2 () does not hold for the endpoint case s = 21 , see [21,
Theorem 1.9.5]. However, one can include the cases s = 21 and s = 23 by replacing
H s () by H s (). In particular, it was shown in [18, Lemma 2.3] that the smooth
3/2
trace extends to a retraction tr : H () → H 1 ().
Next we consider the weak definition of the (co-normal) Neumann trace.

D(∂νQ ) := u ∈ Hdiv1
Q∇ () | there exists a g ∈ L () such that
2

div Q∇u, v + Q∇u, ∇v = g, tr v for all v ∈ H 1 () ,
(3.2)
Q Q
where we set ∂ν u = g. Naturally, one wants to know whether ∂ν coincides with
Q 3/2
an extension of τ N . For Q = Id one has ∂ν = τ N : H () → L 2 (), see [18,
Lemma 2.4]. For Q = Id the properties of such a possible extension were much less
clear for some time. In the recent preprint [5], those issues were resolved. We recall
their central result for our case ( [5, Corollary 11.28]) adapted to the notation we are
going to use.
Lemma 3.4. Let ⊆ Rd be a Lipschitz domain. Let B be a formal second-order
differential operator acting on elements in L 2 () in a distributional sense via

d
Bu = ∂i qi j (x)∂ j u,
i, j=1
where the matrix Q = (qi j ) is given as in Hypothesis 3.1. Let B denote its L 2 ()-
realization (cf. Definition 3.12 below). Then, the co-normal Neumann trace defined
by u → ν · tr(Q∇u) for smooth functions extends uniquely to
γ Qs : H Bs () → H s−3/2 () (3.3)
for all s ∈ [ 21 , 23 ], forming a compatible family in s. Furthermore, for all s ∈ [ 21 , 23 ],

we have the following:
(i) The generalized Neumann traces in (3.3) are surjective. In fact, there are bounded
linear operators
ϒQ
s
: H s−3/2 () → H Bs (), (3.4)
which are also compatible with each other for different s, and right inverses to
the Neumann trace, meaning for all ψ ∈ H s−3/2 () we have γ Qs (ϒ Qs ψ) = ψ.
(ii) For any f ∈ H Bs () and h ∈ H B2−s () the following Green’s formula holds:
tr h, γ Qs f − γ Qs h, tr f
H 3/2−s ()×(H 3/2−s ()) H s−1/2 ()×(H 1/2−s ())
= h, B f − Bh, f .
1/2
(iii) ker(γ Qs ) ⊆ H 3/2 (), ker(tr) ⊆ H 3/2 (), and for any u ∈ H B () with either
γ Qs u = 0 or tr u = 0, there is some C > 0 such that
u 2
H 3/2 ()
≤C u 2
+ Bu .
2
Proof. This is [5, Corollary 11.25 and 11.28].

Q
We are going to verify the compatibility to the weak formulation ∂ν in Theo-
rem 3.14, later.
3.2. On the second-order operator
With the trace results from the last section, we are going to be able to identify the
operator A for the case B = − div Q∇, and to fully describe its domain in precise
terms of Sobolev regularity. The underlying subsidiary form is given as follows.
Definition 3.5. Assume Hypothesis 3.1. Set D(b) := H 1 (), and let b : D(b) ×
D(b) → C be given by
b(u, v) = Q∇u, ∇v (3.5)
for u, v ∈ D(b). Also, set b D (u, v) := b(u, v) for u, v ∈ D(b D ) = H01 ().
Lemma 3.6. Assume Hypothesis 3.1. Then the subsidiary form b (cf. Definition 3.5)
is admissible in the sense of Definition 2.1, whence B D and B N are well defined
in the sense of Definition 2.3. Furthermore, we have Cc∞ () ⊆ D(B N ) ∩ D(B D ).
Hence, the intersection D(B N ) ∩ D(B D ) lies dense in L 2 () and the assumptions of
Definition 2.6 and Theorem 2.8 are satisfied, as well.
Proof. We choose ρ = 1/2 in Eq. (2.1) and have Cc∞ () ⊆ D(b) = H 1 (), whence
the form b is also densely defined. It is accretive, as Q is uniformly positive definite.
As Q is also bounded, we have
u 2
H1
≤ b(u, u) + u 2
≤C u H1.
So b is closed and continuous, and therefore b is generating. It is symmetric, as Q is

symmetric and real-valued. Hence the Neumann realization B N , as the associated oper-
ator to b, and the Dirichlet realization, as the associated operator to b D = b| H 1 () , are
0
well defined and we have D(B N )∩H01 () = D(B D )∩D(B N ) (cf. Proposition 2.4). As
Cc∞ () ⊆ H01 () we only need to show that Cc∞ () ⊆ D(B N ). So let ϕ ∈ Cc∞ (),
which implies that Q∇ϕ ∈ (H 1 ()d ) as C 1,1 (, Rd×d ) ⊂ W 1,∞ (, Rd×d ). Now,
we may use the following version of Green’s formula taken from [5, Corollary 4.5],
which holds for their case of ε = 1/2, as maps from H 1 () to H −1 (). For all
v ∈ D(b) = H 1 () we have that
Q∇ϕ, ∇v + div Q∇ϕ, v = ν · tr Q∇ϕ, tr v = 0

as ∇ϕ = 0 close to the boundary, which shows
b(ϕ, v) = − div Q∇ϕ, v
for all v ∈ D(b) and thus ϕ ∈ D(B N ) and B N ϕ = − div Q∇ϕ.
So we may define B0 , B0∗ , N b , and N b as stated in Theorem 2.8. Furthermore, the

next Lemma shows that we are in the situation where N b = N b holds.
Lemma 3.7. Assume Hypothesis 3.1. Let b be the admissible form from Definition 3.5.
Then the corresponding operator N b (cf. Definition 2.6) is surjective.
Proof. The surjectivity of N b is a special case of [23, Lemma 3.8]. There it is shown
that for any g ∈ L 2 () there is a u ∈ H 1 () and a λ ∈ R such that for all v ∈ C 1 ()
we have b(u, v)+λu, v = g, tr v . Approximation in H 1 () yields this result for
all v ∈ H 1 (). Hence u ∈ D(B0∗ ) as for v ∈ D(B0 ) we have − λu, v = b(u, v) =
u, B N v = u, B0 v and B0∗ u = −λu, so u ∈ D(N b ) and −N b u = g.
Next we show that the operator B0∗ is in some sense the maximal L 2 -realization of
− div Q∇, and −N b coincides with the usual co-normal derivative ∂ν , which extends
ν · tr Q∇. So we begin by verifying that both operators act as desired on smooth
functions.
Proposition 3.8. Assume Hypothesis 3.1. Let b be the admissible form from Defini-
tion 3.5 and B0 , N b be the corresponding induced operators in the sense of Defini-
tion 2.6. Then, if ϕ ∈ C ∞ (), we have B0∗ ϕ = − div Q∇ϕ and N b ϕ = −ν · tr Q∇ϕ.
Proof. For any function ϕ ∈ C ∞ (), so as seen above Q∇ϕ is an element of

(H 1 ())d . So again by [5, Corollary 4.5], for all v ∈ H 1 (), we have
Q∇ϕ, ∇v + div Q∇ϕ, v = ν · tr Q∇ϕ, tr v .
This means, for all functions v ∈ D(B0 ) ⊆ D(B N ), that
− div Qϕ, v = Q∇ϕ, ∇v = b(ϕ, v) = ϕ, B0 v ,
which shows ϕ ∈ D(B0∗ ) and B0∗ ϕ = − div Q∇ϕ. Furthermore, we even have ϕ ∈
D(N b ) and N b ϕ = −ν · tr Q∇ϕ.
We point out that div Q∇ does not map test functions onto test functions. Hence,
there is no distributional realization of that operator and the largest space we can
work on is H −2 (), the dual of H02 (), whence we use the L 2 ()-realization of that
version. To that end, we establish an elliptic regularity result on Rd on the level of test
functions.
In the following we denote by BUCs () the space of uniformly bounded continuous
functions whose derivatives are continuous and uniformly bounded up to order s, as
well.
Lemma 3.9. Let Q satisfy Hypothesis 3.1. Then, there is a symmetric, uniformly
∈ BUC1 (Rd , Rd×d ) of Q. Furthermore, for all s ∈ [0, 2],
positive definite extension Q
there is a λ0 > 1 such that for any λ ≥ λ0 there exists Cλ > 0 for which
ϕ H s (Rd )

≤ Cλ (λ − div Q∇)ϕ H s−2 (Rd )
holds for all ϕ ∈ Cc∞ ().

Proof. We first construct the extension. As Q is uniformly positive definite in some
which contains , there is a C ∞ -domain with ⊆ ⊆
open superset , which
can be constructed by approximation with mollified functions in the supremum norm.
As is smooth with smooth boundary , there is a small tubular neighborhood
ε = {x ∈ Rd | dist(x, ) < ε}
of , which can be parameterized by the normal vector, i.e. there is a smooth bijective
map
γ : (−ε, ε) × → ε ; (h, x ) → γ (h, x ) := x + h · ∂ν (x ).
Choosing ε > 0 small enough, it is possible to guarantee ε ∈ \, so Q is defined

on ∪ ε . For the extension, let ψ ∈ C ∞ (R, [0, 1]) be a strictly decreasing function
satisfying ψ = 1 on (−∞, −ε ε ∞
2 ) and ψ = 0 on ( 2 , ∞). Then, we obtain ϕ ∈ C (R )
d

by setting ϕ = 1 on \ε , ϕ = 0 on R \ (ε ∪ ), and ϕ(x) := ψ(h) for
d

x = γ (h, x ) ∈ ε . Hence, we can define Q(x) = (1 − ϕ(x))Q(x ∗ ) + ϕ(x)Q(x)
for an arbitrary but fixed x ∗ ∈ , and the new matrix Q ∈ BUC1 (Rd , Rd×d ) still
satisfies (3.1) as the set of uniform positive definite matrices is convex.
For the ellipticity estimate, we use parabolic theory from [12]. The constructed
extension Q satisfies their assumption (S2) (its entries are BUC1 (Rd )-functions which
are constant for large |x|), and a simple calculation shows that due to the uniform
positive definiteness, the resulting operator λ − div Q∇ is also parameter-elliptic.
Hence the assumptions of [12, Lemma 3.14] are satisfied for σ = 0 and r = |s−1| =
1 and, we obtain the estimate in the λ-dependent spaces by [12, Lemma 3.14]. As the
constant is allowed to depend on λ, this finishes the proof as we switch to λ-independent
spaces.
Remark 3.10. A thorough comparison of regularities will show that a strict application
of [12, Lemma 3.14] would need qi j ∈ BUC2 (). However, as our operator is in
divergence form, the coefficients do not need to be multipliers in H s−2 (Rd ) but only
in H s−1 (Rd ), whence one can deduce that BUC1 () is actually sufficient as |s−1| ≤ 1
in our case. For details (cf. [25, Section 7.2.1]).
Now we can show that the “minimal” realization of − div ∇ Q is well defined and its
domain is given by H02 (). In this context we understand minimality in the sense that
it is the smallest closed operator that acts like − div ∇ Q on the space of test functions.
Proposition 3.11. Assume Hypothesis 3.1. Let b be the admissible form from Defini-
tion 3.5 and let B0 and B N be the corresponding induced operators in the sense of
Definitions 2.3 and 2.6. Let Bmin be the closure of
B N |Cc∞ () = B0 |Cc∞ () = (− div ∇ Q)|Cc∞ () .
Then the following holds.

(i) Bmin is well defined (i.e. the closure of the underlying graph yields a single-
valued operator).
(ii) On the space Cc∞ () the graph norm u + B0∗ u is equivalent to the full
H 2 ()-norm, whence D(Bmin ) = H02 ().
(iii) We have (− div Q∇)| H 2 () = Bmin = B0 | H 2 () .
0 0
Proof. (i) The operators B N |Cc∞ () and B0 |Cc∞ () are closable due to the self-
adjointness of B N . Thus Bmin is well defined.
(ii) Let ϕ ∈ Cc∞ () be an arbitrary test function. Then B0∗ ϕ = − div Q∇ϕ by
Proposition 3.8 and thus ϕ 2 + B0∗ ϕ 2 ≤ ϕ H 2 () . For the reverse inequality
we use the matrix Q from Lemma 3.9. To that end, we extend ϕ by zero to the
whole space and write e0 ϕ for this extension. Choose any fixed λ > λ0 where λ0
is taken from Lemma 3.9. Then we have (λ − div Q∇)e 0 ϕ = −e0 (div Q∇ϕ) +
e (λϕ), as supp(div Q∇ϕ) ⊆ supp ϕ ⊆ . Now Lemma 3.9 with s = 2 yields

0
there is some Cλ > 0 such that
ϕ H 2 () = e0 ϕ H 2 (Rd ) ≤ Cλ e0 (div Q∇ϕ) Rd + λ e0 ϕ Rd

≤ λCλ ( B0∗ ϕ + ϕ ),
with Cλ independent of the choice of ϕ. Thus we have

· · D(B0∗ )
H02 () = Cc∞ () H 2 ()
= Cc∞ () = D(Bmin ).
(iii) As Bmin is closed, the first equality follows straight-forward. Now let u ∈
H02 () = D(Bmin ) ⊆ D(b D ). Then there is a sequence of test functions such
that ϕn → u with respect to the H 2 ()-norm. For ϕn we have for all v ∈ H 1 ()
Q∇ϕn , ∇v = − div Q∇ϕn , v .
Due to H 2 -convergence this also holds for u, and div Q∇ϕn converges to div Q∇u ∈
L 2 (). So by definition u ∈ D(B N ) ∩ D(B D ) = D(B0 ) and B0 u = B N u =
− div Q∇u = Bmin u, which shows Bmin ⊂ B0 .

Now we use duality to define Bmax . Recall that for a function u ∈ L 2 (), the
induced regular distribution [u] acts on a function ϕ via [u](ϕ) = u, ϕ .
Definition 3.12. Assume Hypothesis 3.1. Let B : L 2 () → (H02 ()) be defined
by
Bu(ϕ) := u, − div Q∇ϕ

for all ϕ ∈ H02 (), and define Bmax as its L 2 ()-realization, i.e. we let
D(Bmax ) := {u ∈ L 2 () | Bu ∈ (L 2 ()) },
and identify Bmax u with g ∈ L 2 (), where g is the unique element for which
(Bu)(ϕ) = [g](ϕ) holds.
It can also be verified that this definition is compatible with the representations

Bu = − i, j ∂i qi j ∂ j u = − div Q∇u, where each derivative is considered as weak
derivative, and the multiplication with the coefficients in H −1 () (cf. [25, Lemma 3.17]).
Now we can characterize Bmax by duality as follows:
Proposition 3.13. Assume Hypothesis 3.1. Let b be the admissible subsidiary form
defined as in Definition 3.5 and let B0 , Bmin , Bmax be the corresponding induced
operators (cf. Definitions 2.6, 3.12, and Proposition 3.11). Then, we have Bmax =
∗ , as well as B ∗
(B0 | H 2 () )∗ = Bmin
0 max = Bmin . In particular, this shows that Bmax is
closed.
Proof. Let u ∈ D(Bmax ). This, equivalently, means u ∈ L 2 () and there is an f ∈
L 2 () such that f, ϕ = Bu(ϕ) = u, − div Q∇ϕ = u, B0 ϕ for all ϕ ∈
H02 (). By definition this means u ∈ D((B0 | H 2 () )∗ ) and f = (B0 | H 2 () )∗ u. The
0 0
second assertion follows directly as Bmin is closed and the restriction of the self-adjoint
operator B N , and therefore symmetric.
In a final step we remove the restriction to H02 by showing that the smooth functions
are actually a core of Bmax .
To that end, we restate the definition of the spaces H Bs () from (cf. Sect. 3.1) in
a more precise manner by setting H Bs () := {u ∈ H s () | Bu ∈ L 2 ()} for
s ≥ 0 equipped with the norm u H s () + Bmax u , and in particular we have
D(Bmax ) = H B0 (). We will also write · B instead of · D(Bmax ) or · H 0 () . It
B
might be more accurate to call those spaces H Bs max (), but we refrain from doing so
for sake of readability, also emphasizing the fact that Bmax actually takes the role of
the abstract operator B that remained undefined in Sect. 2.
Furthermore, we even may explicitly characterize D(N b ) and N b as we are in the
setting of [5, Chapter 11.4], whence we have Lemma 3.4 at our disposal and existence,
continuity, and surjectivity of γ Qs : H Bs () → H s−3/2 () is assured for all s ∈ [ 21 , 23 ].
All those Neumann traces are continuous extensions of ν · tr Q∇ on C ∞ () to H Bs ()
by the density which we also ascertain below.
Theorem 3.14. Assume Hypothesis 3.1. Let b be the admissible subsidiary form
defined as in Definition 3.5 and let B0 , B D , B N , Bmin , Bmax , N b be the corresponding
induced operators (cf. Definitions 2.3, 2.6, 3.12, and Proposition 3.11). Let ∂νQ and
3/2
γ Q be the weak and strong Neumann traces from (3.2) and (3.3). Then we have:
(i) For all s ≥ 0, the space C ∞ () is dense in H Bs ().
(ii) We have B0 = Bmin (i.e. D(B0 ) = H02 () and Cc∞ () is a core of B0 ) and
B0∗ = Bmax .
(iii) We have N b = −∂ν = −γ Q , so in particular D(N b ) = D(∂ν ) = H B ().

Q 3/2 Q 3/2
3/2 3/2
(iv) D(B D ) ⊆ H B (), D(B N ) ⊆ H B ().
Proof. (i) In the case s ≥ 2 the space H Bs () coincides with H s (). The cases
s ∈ [0, 2) will follow by adapting the proof of [5, Lemma 2.13], where this
density was shown for B = . Consider Ḣ s () := {u ∈ H s (Rd ) | supp u ∈ }
(cf. [5, Section 2.3]) and the map
ι : H Bs () → H s () × L 2 (), u → ι(u) = (u, Bmax u),
which is an isometric isomorphism from H Bs () to the closed subspace ι(H Bs ())
as Bmax is a closed operator. Let be any functional in (H Bs ()) , then ◦ι−1 is
a linear, bounded functional on ι(H Bs ()), which can be extended to a functional
∈ (H s () × L 2 ()) = Ḣ −s () × L 2 () using Hahn-Banach’s theorem.

Hence, by [5, p.27–29], there are representatives h 1 ∈ Ḣ −s (), h 2 ∈ L 2 ()
such that, given any u ∈ H Bs (), for any F ∈ H s (Rd ), G ∈ L 2 (Rd ) satisfying
F| = u and G| = Bmax u we have
(u) = h 1 , F H −s (Rd )×H s (Rd ) + e0 h 2 , G .

Rd
Note that for s = 0, we may replace Ḣ 0 () with the zero extension of L 2 ()-
functions and the dual pairing with the standard inner product on L 2 (). In
particular, if we take u = ϕ| , for any ϕ ∈ Cc∞ (Rd ), we obtain

(ϕ| ) = h 1 , ϕ H −s (Rd )×H s (Rd ) + e0 h 2 , − div Q∇ϕ
Rd

due to (− div Q∇ϕ)| ∗
= (− div Q∇ϕ)| = B0 ϕ| = Bmax ϕ| by Proposi-

tion 3.8, where Q once more denotes the extended matrix from Lemma 3.9.
In order to show the desired density, we assume that for any ϕ ∈ C ∞ () we
had (ϕ) = 0, and deduce that this implies = 0. By definition, however,
(ϕ) = 0 means that we have

h 1 , ϕ H −s (Rd )×H s (Rd ) = e0 h 2 , div Q∇ϕ 0 h 2 )(ϕ)
= (div Q∇e
Rd
for all ϕ ∈ Cc∞ (Rd ) and by density for all ϕ ∈ H 2 (Rd ). At first we consider
0 h 2 = (− div ◦m Q ◦ ∇)(e0 h 2 ) as an element of H −2 (Rd ) consisting
− div Q∇e
of the separate mappings ∇ : L 2 (Rd ) → (H −1 (Rd ))d , m Q : (H −1 (Rd ))d →
(H −1 (Rd ))d , and div : (H −1 (Rd ))d → H −2 (Rd ) where m Q denotes the
multiplication with Q in H −1 (Rd ). With the usual identification of dual spaces,
we obtain
0h2, ϕ
div Q∇e 0 h 2 )(ϕ)
= (div Q∇e
H −2 (Rd )×H 2 (Rd )
= h 1 , ϕ H −s (Rd )×H s (Rd ) .
0 h 2 ∈ H −2 (Rd ), or for some large λ ≥ λ0 also −h 1 +λe0 h 2 =

So h 1 = div Q∇e

(λ − div Q∇)e 0 h . Now by Lemma 3.9 applied with 2 − s ∈ (0, 2], we have
2
e0 h 2 H 2−s (Rd ) ≤ Cλ λe0 h 2 − h 1 H −s (Rd ) ,
which shows that e0 h 2 ∈ H 2−s (Rd ) and as supp e0 h 2 ∈ , also e0 h 2 ∈

Ḣ 2−s ().
However the space of zero extensions of Cc∞ () lies dense in Ḣ 2−s () (cf. [5,
(2.82)]), and there is a sequence of functions (ψn )n ⊆ Cc∞ () such that e0 ψn
0 ψn converges to
converges to e0 h 2 in H 2−s (Rd ), which shows that div Q∇e
0 −s
div Q∇e h 2 = h 1 in H (R ). But then, for any u ∈ H B () and F ∈ H s (Rd )
d s
such that F| = u, we obtain
(u) = h 1 , F H −s (Rd )×H s (Rd ) + e0 h 2 , e0 Bmax u

Rd
= lim div Q∇e ψn , F 0

+ ψn , Bmax u
n→∞ H −s (Rd )×H s (Rd )
∗
= lim − Bmin ψn , u + ψn , Bmin u = 0.
n→∞
Hence, already vanishes on H Bs (); and we have shown that |C ∞ () = 0
implies = 0, which yields the desired density by a standard corollary to
Hahn–Banach.
(ii) Let v ∈ D(B0 ) and u ∈ D(Bmax ) be arbitrary. Because of (i) there is a sequence
of functions ϕn in C ∞ () such that B0∗ ϕn → Bmax u and ϕn → u in L 2 ().
Hence for all v ∈ D(B0 ) we have
0 = −N b ϕn , tr v = −N b ϕn , tr v

= v, B0∗ ϕn − B0 v, ϕn → v, Bmax u − B0 v, u ,
so
v, Bmax u − B0 v, u = 0
for all u ∈ D(Bmax ). This shows v ∈ D(Bmax ∗ ) and B v = B ∗ v = B

0 max min v,
which together with Propositions 3.11 (iii) and 3.13 shows that B0 = Bmin and
B0∗ = Bmax as claimed.
(iii) N b = −∂ν follows from B0∗ = Bmax and the definition of the weak co-
Q
normal (3.2). We show γ Q ⊆ −N b ⊆ γ Q1 first. So let u ∈ H B (). By (i) there

3/2 3/2
is a sequence (ϕn )n ⊆ C ∞ () that converges to u in H B (). As in the proof

3/2
of Proposition 3.8 we have Bmax ϕn , v − b(ϕn , v) = −ν · tr Q∇ϕn , tr v =

3/2
−γ Q ϕn , tr v for all v ∈ H 1 (). Hence, as the sequence (ϕn )n in particular
converges in H B1 (), we may take the limit and obtain for all v ∈ H 1 ()
Bmax u, v − b(u, v) = lim Bmax ϕn , v − b(ϕn , v)
n→∞
3/2 3/2
= lim −γ Q ϕn , tr v = −γ Q u, tr v .
n→∞
As γ Q u ∈ L 2 (), by definition we have u ∈ D(N b ) and N b u = −γ Q u. For

3/2 3/2
the second inclusion let u ∈ D(N b ), then for all v ∈ H 1 () we have
Bmax u, v − b(u, v) = N b u, tr v = N b u, tr v

H −1/2 ()×H 1/2 ()
as v ∈ H 1 () and the Dirichlet trace maps continuously from H 1 () to H 1/2 ()
(see Definition 3.3).
Next, recall that D(N b ) ⊆ H B1 () by definition. Using the density of C ∞ ()
in H B1 (), we find a sequence (ϕn )n ⊆ C ∞ () that converges in H B1 () toward
u. So continuity of γ Q1 from H 1 () to H −1/2 () yields
Bmax u, v − b(u, v) = lim Bmax ϕn , v − b(ϕn , v)

n→∞
= lim −γ Q1 ϕn , tr v
n→∞
= lim −γ Q1 ϕn , tr v −1/2
n→∞ H ()×H 1/2 ()
= −γ Q1 u, tr v .
H −1/2 ()×H 1/2 ()
Hence we have
N b u, tr v = −γ Q1 u, tr v
H −1/2 ()×H 1/2 () H −1/2 ()×H 1/2 ()
for all v ∈ H 1 (). As the Dirichlet trace is surjective onto H 1/2 (), we obtain
N b u, ψ = −γ Q1 u, ψ
H −1/2 ()×H 1/2 () H −1/2 ()×H 1/2 ()
for all ψ ∈ H 1/2 () by taking any solution of tr v = ψ. Thus −γ Q1 u = N b u

on H −1/2 (), which in particular yields ker γ Q ⊆ ker N b ⊆ ker γ Q1 . By
3/2
3/2 3/2 3/2

Lemma 3.4 we have ker γ Qs ⊆ H B = D(γ Q ) for all s ∈ [ 21 , 23 ]. As γ Q is
Nb
3/2
also surjective onto L 2 (),
we have = γQ
by Proposition 2.7.
(iv) This follows from Lemma 3.4, once more. We have that u ∈ H B1 () and either
3/2
tr u = 0 or γ Qs u = 0 for any s ∈ [ 21 , 23 ] implies u ∈ H B ().

3.3. On the fourth-order system
Applying the above results to the primary form a and its associated operator we
obtain the following solution theorems for fourth-order system with Wentzell bound-
ary conditions. We are able to solve the following system, as N b is surjective by
Lemma 3.7.
Theorem 3.15. Assume Hypotheses 2.9 and 3.1. Let H = L 2 (, λd )× L 2 (, β −1 dS)
and B = − div ∇ Q. Then for u0 = (u 1,0 , u 2,0 ) ∈ H the Cauchy problem
∂t u 1 + B(α B)u 1 = 0 in (0, ∞) × , (3.6)

∂t u 2 + β∂νQ (α B)u 1 + γ u2 = 0 on (0, ∞) × , (3.7)
∂νQ u 1 =0 on (0, ∞) × , (3.8)
u 1 |t=0 = u 1,0 in , (3.9)
u 2 |t=0 = u 2,0 on (3.10)
possesses a unique solution, which is given by u(t) = T(t)(u 1,0 , u 2,0 ) for t > 0 where
T(t) is the analytic semigroup generated by −A given as in (2.10). Furthermore, we
have
3/2 3/2 3/2
D(A) = {u ∈ H | u 1 ∈ H B (), α div Q∇u 1 ∈ H B (), tr u 1 = u 2 , γ Q u 1 = 0}.
Proof. This is a direct consequence of Theorems 2.8 and 2.12, whose assumptions are
validated by Lemmata 3.6 and 3.7. The identification of the operators and characteri-
zation of the domain follow from Theorem 3.14.
Remark 3.16. As in Remark 2.13, we observe that u 1 also solves the corresponding
non-decoupled problem with Wentzell boundary conditions
∂t u + B(α B)u = 0 in (0, ∞) × , (3.11)

tr B(α B∇)u − β∂νQ (α B)u − γ tr u = 0 on (0, ∞) × , (3.12)
∂νQ u = 0 on (0, ∞) × , (3.13)
u|t=0 = u 0 in . (3.14)
Finally, we also want to add the case δ > 0. The main idea is to compare the
machinery of the form b with that of bδ defined by
bδ (u, v) = Q∇u, ∇v + δu, v (3.15)
for 0 ≤ δ ∈ L ∞ () on D(bδ ) = H 1 ().

Proposition 3.17. Under Hypotheses 2.9 and 3.1, the subsidiary form bδ (defined via
(3.15)) is admissible. We denote its associated operator by B N ,δ .
Proof. As 0 ≤ δ ∈ L ∞ (), the calculations are similar to the proof of Lemma 3.6.
Note that the norm · bδ is also equivalent to the full H 1 ()-norm as the trace is
continuous from H 1 () to L 2 ().
Under the given smoothness conditions we have Cc∞ () ⊆ D(B N ,δ ) similar as in
Lemma 3.6 due to the fact that b = bδ for test functions. Naturally, one can consider
the restriction of the form to H01 () once more, but there the form also coincides with
the previous form b D , so the Dirichlet realization is independent of δ. This also shows
that also D(B N ,δ ) ∩ D(B D,δ ) is dense in L 2 () and Theorem 2.8 is applicable to bδ ,
as well.
Theorem 3.18. Assume Hypotheses 2.9 and 3.1. Let H = L 2 (, λd )×L 2 (, β −1 dS),
and b and bδ be the admissible subsidiary forms defined via Definition 3.5 and Formula
(3.15), respectively. Let B0 , Bmin , Bmax , N b be the corresponding induced operators
with respect to b; and B0,δ , Bmin,δ , Bmax,δ , N bδ be the corresponding induced oper-
ators with respect to bδ (cf. Definitions 2.6, 3.12, and Proposition 3.11). Also let aδ
denote the primary form corresponding to bδ (cf. Definition 2.10 (ii)). Let ∂νQ and
3/2
γ Q be the weak and strong Neumann traces from (3.2) and (3.3). Then, we have the
following results.
(i) N bδ is surjective.
(ii) B0,δ = B0 , and (B0,δ )∗ = Bmax,δ = Bmax = B0∗ = − div Q∇ considered as
L 2 -realization of a map from L 2 () to H −2 ().
(iii) N bδ = −γ Q − δ · tr, and the operator associated to aδ on H is given by
3/2

div Q∇(α div Q∇) 0
Aδ = 3/2 (3.16)
−β(γ Q + δ tr)(α div Q∇) γ
on
3/2 3/2
D(Aδ ) = {u ∈ H | u 1 ∈ H B (),α div Q∇u 1 ∈ H B (),
3/2
tr u 1 = u 2 , γ Q u 1 + δ tr u 1 = 0}.
Proof. (i) The surjectivity of N bδ follows from [23, Lemma 3.7/3.8], just as in the
Neumann case, because the theory there also contains the Robin case (cf. [23,
Equation (2.3)]). So N bδ = N bδ .
(ii) For u ∈ H01 and f ∈ L 2 () we have f, v = Q∇u, ∇v for all v ∈ H 1 ()
if and only if we have f, v = Q∇u, ∇v + δ tr u, tr v for all v ∈ H 1 ()
due to tr u = 0. Thus D(B N ) ∩ H01 = D(B N ,δ ) ∩ H01 and the operators B N and
B N ,δ coincide there, which shows B0 = B0,δ . Taking adjoints, this carries over
to B0∗ . As B0,δ = B0 = (− div Q∇)| H 2 we also have Bu = Bδ u, so also their
0
L 2 -realizations Bmax and Bmax,δ must coincide.
(iii) Due to the surjectivity shown in (i) we have N bδ = N bδ by Theorem 2.8 (iv).
Next we show N bδ = N b − δ tr u. Assume u ∈ D(N
b ). Hence u ∈ H 1 ()
and for all v ∈ H 1 () we have N b u, tr v = B0∗ u, v − b(u, v). So,

equivalently, N b − δ tr u, tr v = (B0,δ )∗ u, v − bδ (u, v), which shows
D(N b ) = D(N bδ ) and N bδ = N b − δ tr = −γ Q − δ tr by Theorem 3.14 (iii).
3/2
Furthermore, by Theorem 2.12, the associated operator Aδ is given by

(B0,δ )∗ (α B N ,δ ) 0
Aδ =
−βN bδ (α B N ,δ ) γ
on
bδ
D(Aδ ) = {u ∈ H | u 1 ∈ D(B N ,δ ), α B N ,δ u 1 ∈ D(N ), u 2 = tr u 1 }.
Using (i), (ii) and Theorem 3.14 (iii) and (iv) yields the result. Note that ker N bδ =
D(B N ,δ ) because of Theorem 2.8 (iii) applied to bδ .

Remark 3.19. After having verified H B ()-regularity for the trace N b , it can be
3/2
deduced that (L 2 (), tr, N bδ ) actually is a quasi-boundary triple for the operator
Bmax | D(N b ) in the sense of [7], which generalizes the results from their Section 4.2 to
Lipschitz domains. A detailed proof can be found in [25, Section 3.4].
We may finally collect the main result for our original system (1.5)–(1.9). The
notations N b and γ Q were useful in context with the general theory. In the following,
3/2
Q
however, we will write ∂ν , again, which is closer to classical notation. Note that due
to our results we have ∂ν = γ Q = −N b = −N b , anyway.
Q 3/2
Theorem 3.20. Assume Hypotheses 2.9 and 3.1. Let H = L 2 (, λd )×L 2 (, β −1 dS).
Q
Write B = − div Q∇ and ∂ν for the unique extension of ν · tr Q∇ to H 3/2 (). Then
for u0 = (u 1,0 , u 2,0 ) ∈ H the Cauchy problem
∂t u 1 + B(α B)u 1 = 0 in (0, ∞) × , (3.17)

∂t u 2 + β∂νQ (α B)u 1 + βδ tr(α B)u 1 + γ u 2 = 0 on (0, ∞) × , (3.18)
∂νQ u 1 + δu 2 = 0 on (0, ∞) × , (3.19)
u 1 |t=0 = u 1,0 in , (3.20)
u 2 |t=0 = u 2,0 on (3.21)
possesses a unique solution, which is given by u(t) = T(t)(u 1,0 , u 2,0 ) where T(t) is
the analytic semigroup generated by −Aδ given as in (3.16). For t > 0, we have u(t) ∈
D(Aδ∞ ), whence u 1 solves the original system with Wentzell boundary conditions
given by
∂t u + B(α B)u = 0 in (0, ∞) × , (3.22)

tr B(α B)u − β∂νQ (α B)u − βδ tr(α B)u − γ tr u = 0 on (0, ∞) × , (3.23)
∂νQ u + δ tr u = 0 on (0, ∞) × , (3.24)
u|t=0 = u 0 in . (3.25)
4. Further properties of the solution in the fourth-order case
In this section, we present results concerning regularity and long-time behavior of

our solution. We begin with a regularity result in a smoother situation. Then we return
to our situation with Lipschitz domains and rougher coefficients.
4.1. Higher regularity for smoother cases
Even with smooth coefficients and boundary, we cannot expect that for u ∈ D(A)
the first component u 1 belongs to H 4 (). However, using the theory of [12], we can
deduce u 1 ∈ H 7/2 ().
To that end let us recall the notion of parameter-ellipticity. Therein, we use the
standard form λ+A B for parameter-elliptic boundary value problems for a moment,
which is not to be confused with our operators A and B.
Definition 4.1. Let ⊆ C be a closed sector in the complex plane with vertex at
the origin. Using the standard convention D := −i∇ for parabolic boundary value
problems, let A and B = (B1 , ..., Bm ) be formally given by

A(x, D) := aα (x)D α and B j (x, D) := bβ (x) tr D β ( j = 1, . . . , m),
|α|≤2m |β|≤m j
where m j < 2m
(i) We define the principal symbols of A and B j as

a0 (x, ξ ) := aα (x)ξ α and b0, j (x, ξ ) := b jβ (x)ξ β ( j = 1, . . . , m),
|α|=2m |β|=m j
respectively.
(ii) We call the family λ − A(x, D) parameter-elliptic in if the principal symbol
a0 (x, ξ ) satisfies
|λ − a0 (x, ξ )| ≥ C |λ| + |ξ |2m (x ∈ , λ ∈ , ξ ∈ Rd , (ξ, λ) = 0) (4.1)
for some constant C > 0.

(iii) The boundary value problem λ+A B is called parameter-elliptic in if λ −
A(x, D) is parameter-elliptic in , and the Shapiro–Lopatinskii condition holds,
i.e.:
Let x0 ∈ ∂ be an arbitrary point of the boundary; rewrite the boundary value
problem (λ−a0 (x0 , D), b0,1 (x0 , D), . . . , b0,m (x0 , D)) in the coordinate system
associated with x0 obtained from the original one by a rotation after which the
positive xd -axis has the direction of the interior normal to ∂ at x0 . Then, for
all ξ ∈ Rd−1 and λ ∈ with (ξ , λ) = 0, the trivial solution w = 0 is the only
stable solution of the ordinary differential equation on the half-line
(λ − a0 (x0 , ξ , Dd ))w(xd ) = 0 (xd ∈ (0, ∞)),

b0, j (x0 , ξ , Dd )w(0) = 0 ( j = 1, . . . , m).
So we rewrite our system as a parameter-elliptic boundary value problem. Recall

that any solution of (1.5)–(1.9) satisfies u ∈ D(A), which shows B(α B)u 1 ∈ L 2 (),
β∂νQ (α B)u 1 ∈ L 2 , and ∂νQ u 1 + δ tr u 1 = 0. Thus the first component of any solution
of (1.5)–(1.9) in particular satisfies
λu 1 + B(α B)u 1 = f := λu 1 + B(α B)u 1 in (0, ∞) × , (4.2)
−β∂νQ (α B)u 1 = g := −β∂νQ (α B)u 1 on (0, ∞) × , (4.3)
∂νQ u 1 + δ tr u 1 = 0 on (0, ∞) × (4.4)
with ( f, g) ∈ H = L 2 () × L 2 (). It remains to check that all assumptions of

[12, Corollary 4.10] are satisfied for this system, where we have τ = m j + 1/ p and
thus |r | = |k1 | + 1 = 1, |k2 | + 1 = 3. In order to ensure aα ∈ BUC1 (),
b1β ∈ BUC1 (), and b2β ∈ BUC3 (), and the sufficient boundary regularity, we
make the following assumptions, which also imply Hypotheses 2.9 and 3.1.
Hypothesis 4.2. Let Q ∈ BUC4 (, ¯ Rd×d ) be symmetric and uniformly positive def-
inite, which means there is some open superset ⊆ Rd containing such that
Q ∈ BUC (, R4 d×d ) is symmetric and satisfies for some κ Q > 0
, ξ ∈ Cd ).
Q(x)ξ, ξ Cd ≥ κ Q |ξ |2 (x ∈
Let α ∈ BUC3 (), β ∈ BUC1 (), and δ ∈ BUC3 (), such that there exists a
constant η > 0 with α ≥ η almost everywhere on and β ≥ η almost everywhere on
. Furthermore, let δ ≥ 0.
The final assumption to check is parameter-ellipticity:
⎛ ⎞
λ + Bα B
⎜ ⎟
Lemma 4.3. Assuming Hypothesis 4.2 the system ⎝−β∂νQ (α B)⎠ is parameter-elliptic
Q
∂ν + δ tr
in θ for θ ∈ (0, π ).
Proof. The parameter-ellipticity for the family λ + Bα B is simple as we have

a0 (x, ξ ) := symb0 [Bα B](x, ξ ) = (ξ j q jk (x)ξk )α(x) (ξ j q j k (x)ξk ) > 0
j,k j ,k
for all |ξ | = 0 as the matrix Q = (q jk ) jk is symmetric and uniformly positive defi-

nite. Note for the first line that all the terms where the derivative hits the coefficient
are of lower order. Hence, we can exploit homogeneity and boundedness of the do-
main, which shows parameter-ellipticity in any closed sector that does not contain the
negative real line.
A similar calculation shows

a (x, ξ ) := symb0 [B](x, ξ ) = ξ j q j k (x)ξk ,
j ,k

b0,1 (x, ξ ) := symb0 [−β∂νQ (α B)](x, ξ ) = iβ(x) qdk ξk α(x)a (x, ξ ),
k

b0,2 (x, ξ ) := symb0 [∂νQ ](x, ξ ) = −i qdk (x)ξk .
k
In order to verify the Shapiro–Lopatinskii condition, we need to show that the ODE
below only has the trivial solution. For this, the operators are locally transformed
into the half-space at every fixed point x0 ∈ . It is a known fact that this coor-
dinate transformation leaves the coefficients of the main symbol invariant (cf. [30,
Satz 10.3]). Furthermore, we would like to note that, as we only consider a fixed x0 ,
the coefficients commute with all derivatives and we can simply pass from divergence
to non-divergence form, which shows we may investigate the system (4.5) below. We
ξ
write = −i∂ d
, where ξ = (ξ , ξd ) ∈ Rd as usual, and interpret the first compo-
nents as multiplication, so w = (ξ1 w, . . . , ξd−1 w, −i∂d w). Then, for xd ∈ (0, ∞),
and λ ∈ θ , ξ ∈ Rd−1 satisfying (λ, ξ ) = 0, we assume
(λ + a0 (x0 , ))w(xd ) = 0,
b0,1 (x0 , )w(0) = 0, (4.5)
b0,2 (x0 , )w(0) = 0.
Note that integration by parts yields

j u, v = u, jv − iu d (0) · vd (0).
j L 2 (0,∞) j L 2 (0,∞)
Multiplying the first line with w in L 2 ((0, ∞)) and using integration by parts and
(4.5), we obtain

0 = |λ| w 2
L 2 (0,∞)
+ j q jk (x0 )k α(x0 )a (x0 , )w, w L 2 (0,∞)
j,k

= |λ| w 2
L 2 (0,∞)
+ q jk (x0 )k α(x0 )a (x0 , )w, j w L 2 (0,∞)
j.k
− β −1 (x0 )b0,1 (x0 , )w(0) · w(0)

= |λ| w 2L 2 (0,∞) + k α(x0 )a (x0 , )w, q jk (x0 ) j w L 2 (0,∞)
j,k

= |λ| w 2
L 2 (0,∞)
+ α(x0 )a (x0 , )w, j qk j (x0 )k w
k, j L 2 (0,∞)

− α(x0 )a (x0 , )w(0) · b0,2 (x0 , )w(0)

= |λ| w 2L 2 (0,∞) + α(x0 )a (x0 , )w 2L 2 (0,∞) .
Note that we used that the matrix Q = (q jk ) is symmetric and real-valued and that
α > 0, β > 0. If λ = 0, this implies that w = 0 as desired, since θ ⊆ C\(−∞, 0).
If λ = 0, and thus by assumption ξ = 0, we have a (x0 , )w = 0.
Multiplying with w in L 2 (0, ∞) once more, we obtain

0= j q j k (x0 )k w, w L 2 (0,∞)
j ,k

= q j k (x0 )k w, j w L 2 (0,∞)
+ b0,2 (x0 , )w(0) · w(0)
j ,k

= q j k (x0 )k w, j w L 2 (0,∞)
j ,k
∞ ∞
= Q(w)(xd ), (w)(xd )Cd dxd ≥ κ Q |(w)(xd )|2 dxd
0 0
= κ Q (|ξ |2 w 2
L 2 (0,∞)
+ ∂d w 2
L 2 (0,∞)
).
In the last step we used that Q is uniformly positive definite (cf. (3.1)). As ξ = 0,
we also have w = 0 in this case. Hence altogether the system is parameter-elliptic in
every sector smaller than π .
This shows H 7/2 -regularity of any solution:
Corollary 4.4. Assume Hypothesis 4.2. Let H be as in Definition 2.10 (i) and set
B := − div Q∇. Then the domain of Aδ (cf. Formula (3.16)) is given by
Q
D(Aδ ) = {u ∈ H | u 1 ∈ H 7/2 (), B(α B)u 1 ∈ L 2 (), ∂ν + δ tr u 1 = 0, u 2 = tr u 1 }.
Proof. Due to Hypothesis 4.2 and Lemma 4.3 all assumptions of [12, Corollary 4.10]
are satisfied, hence we obtain
u1 7/2
Hλ ()
≤ f L 2 () + g 0 ()
B22,λ <∞
0 () = L 2 () = L 2 ().

due to B22,λ
λ
4.2. Hölder regularity
We show next that on Lipschitz domains and with coefficients as in Hypotheses 2.9
and 3.1 the solution u(t, ·) = (u 1 (t, ·), u 2 (t, ·)) of (1.1)–(1.4) satisfies that u 1 (t, ·) is
Hölder continuous for every t > 0. This implies that tr u 1 (t, ·) = u 2 (t, ·) also holds
in a classical sense.
Recall that T(t) maps H into D(Aδ∞ ) for any t > 0 as it is analytic. It is not
to be expected, however, that u(t, ·) ∈ D(Aδ∞ ) implies u 1 (t, ·) ∈ H 3/2+ε () for
any ε > 0, as such a gain in differentiability does not even necessarily hold for the
much simpler Neumann Laplacian due to possible non-convex corners (cf. [20]). So
Hölder-continuity cannot be derived by Sobolev embedding directly in high dimen-
sions. However, we can use a bootstrapping idea on the integrability.
To that end we use the regular spaces L p () and L p () where the coefficient β
is not included, and write · , p and · , p for the occurring norms, respectively.
In the case p = 2, the index is dropped. Note that the L p -spaces are nested as our
domain is bounded. Furthermore, recall that C 0,ϑ () refers to the space of ϑ-Hölder
continuous functions on , and note that every function u ∈ C 0,ϑ () can be extended
uniquely to a (Hölder) continuous function on .
As a preparation, we establish some further results concerning weak solutions of
the inhomogeneous Neumann problem
(λ − div Q∇)u = f˜ in ,
(4.6)
∂νQ u = g̃ on .
Proposition 4.5. Assume Hypotheses 2.9 and 3.1 and let B = − div ∇ Q. For f˜ ∈
L 2 (), g̃ ∈ L 2 () the following holds.
(i) For λ > 0, (4.6) has a unique weak solution, by which we mean a function
u ∈ H 1 () such that
bλ (u, v) = Q∇u, ∇v + λu, v = f˜, v + g̃, tr v (4.7)
3/2
for all v ∈ H 1 (). Furthermore, u ∈ H B (), and we have the estimate
u 2
3/2 ≤ C( f˜ 2
+ g̃ ).
2
H B ()
(ii) Let f˜ ∈ L d/2+ε (), g̃ ∈ L d−1+ε () for some ε > 0. Then, for any λ ∈ R,
any weak solution of (4.6) satisfies u ∈ C 0,ϑ () for some ϑ ∈ (0, 1) and the
estimate
u C 0,ϑ () ≤C u + f˜ , d2 +ε + g̃ ,d−1+ε . (4.8)
If λ > 0, we can drop the u -term on the right-hand side.

(iii) Let f˜ ∈ L p (), g̃ ∈ L p () for some p > 2. Then, for λ > 0, the unique
solution u of (4.6) satisfies (u, tr u) ∈ L ϕ( p) () × L ϕ( p) () and

u ,ϕ( p) + tr u ,ϕ( p) ≤ C0 f˜ , p + g̃ , p (4.9)
where

d−2
p if p ∈ (2, d),
ϕ( p) := d− p
∞ if p ∈ [d, ∞).
Proof. (i) We construct a solution candidate by collecting properties of a weak

solution. At first we observe that for any such weak solution u ∈ H 1 () we
have, given any v ∈ D(B0 ) ⊆ D(B N ),
u, B0 v = b(u, v) = bλ (u, v) − λ u, v = f˜ − λu, v .

Hence u ∈ D(B0∗ ) and f˜ = (λ + B0∗ )u. By Theorem 3.14 (ii), we also have
(λ+ Bmax )u = (λ+ B0∗ u) = f˜ by, so the first line of (4.6) holds in L 2 (), where
− div Q∇ is seen as L 2 ()-realization of an object in H −2 (). Furthermore,

for all v ∈ H 1 () we obtain

−B0∗ u, v + b(u, v) = − f˜, v + bλ (u, v) = g̃, tr v ,

which shows u ∈ and D(N b ) −N b u

= g̃. However, Theorem 3.14 (iii) yields
D(N ) = H B () and γ Q u = ∂ν u = −N b u = g̃. So when we subtract v =
b 3/2 3/2 Q
3/2 3/2 3/2

ϒ N g̃ where ϒ N is the continuous right-inverse of γ Q from Lemma 3.4 (i),
the difference u − v solves the Neumann problem
(λ − div Q∇)(u − v) = f˜ + div Q∇v − λv in ,
3/2
(4.10)
γ Q (u − v) = 0 on .
Thus u − v ∈ D(B N ), as well as (λ − div Q∇)(u − v) = (λ + B N )(u − v).
In conclusion, we have shown that any weak solution of (4.6) satisfies (λ +
B N )(u − v) = f˜ + div Q∇v − λv. Hence, a suitable solution candidate is given
by ũ := (λ + B N )−1 ( f˜ + div Q∇ϒ N g̃ − λϒ N g̃) + ϒ N g̃, which is well
3/2 3/2 3/2
defined due to (−∞, 0) ∈ ρ(B N ).

Finally, we verify ũ indeed is a solution and satisfies the regularity estimate. By
Lemma 3.4 (iii), we have
ũ − v 2
3/2 ≤ C( ũ − v 2
+ B N (ũ − v) )
H B ()
= C ũ − v 2
BN
= C (λ + B N )−1 ( f˜ + div Q∇v − λv) 2
BN
≤C f˜ 2
+ v 2
3/2
H B ()
(where the constant C is generic) and thus by the above and the continuity of
3/2
ϒN
ũ 2
3/2 ≤C f˜ 2
+ v 2
3/2 ≤ C( f˜ 2
+ g̃ ).
2
H B () H B ()
Now, naturally, ũ solves (4.6) in a strong sense by construction. By definition of

−N b = ∂νQ it also satisfies (4.7) and is in particular a weak solution with all the
desired properties.
The uniqueness is straightforward: If we had two weak solutions u, w ∈ H 1 ()
satisfying (4.7) for all v ∈ H 1 (), we would have bλ (u − w, v) = 0 for all
v ∈ H 1 (). Now λ + B N is the associated operator to bλ , whence u − w in
D(B N ) and (λ + B N )(u − w) = 0. Again, by (−∞, 0) ⊆ ρ(B N ), we have
u − w = 0.
(ii) This is [22, Theorem 3.1.6] applied to A(x, u, p) = Q · p, a(x, u, p) = λu.
Note that their Assumption 2.9.1 is satisfied, and we are in the situation of [22,
Remark 3.1.7]. If λ > 0, we can estimate
u ≤ u 3/2
H B ()
≤ C( f˜ 2
+ g̃ )
2
≤ C( f˜ 2
d + g̃ d−1+ε, ).
2
2 +ε,
(iii) Let λ > 0. By (i) and (ii) the unique solution satisfies the two estimates

u ,2 + tr u ,2 ≤ C u H 3/2 () ≤ C f˜ ,2 + g̃ ,2 , (4.11)
B
as well as
u ,∞ + tr u ,∞ ≤ u C 0,ϑ () ≤C f˜ ,d + g̃ ,d . (4.12)
More precisely, the solution operator Rλ that maps ( f˜, g̃) to (u, tr u) is well
defined and continuous from X 0 := L 2 () × L 2 () to Y0 := L 2 () × L 2 ()
as well as from X 1 := L d () × L d () to Y1 := L ∞ () × L ∞ (). By complex
interpolation, we obtain that Rλ is also continuous from [X 0 , X 1 ]θ to [Y0 , Y1 ]θ
for all θ ∈ (0, 1). To identify the interpolation spaces, recall from [28, Theo-
rem 1.18.1] that complex interpolation of tuples of L p -spaces yields the tuple
of interpolated spaces in the sense of
[L p0 () × L q0 (), L p1 () × L q1 ()]θ

= [L p0 (), L p1 ()]θ × [L q0 () × L q1 ()]θ
for all p0 , p1 , q0 , q1 ∈ [1, ∞]. Moreover, we have the equality [L p0 (),

θ
L p1 ()]θ = L p () (and a similar equality for ) for 1p = 1−θ p0 + p1 in the
sense of equivalent norms, see [28, Theorem 1.18.6/2]. From this, we obtain for
all θ ∈ (0, 1) the continuity of Rλ : X θ → Yθ where X θ := L p ()× L p () and
θ
Yθ := L ϕ( p) () × L ϕ( p) () with p and ϕ( p) being defined by 1p = 1−θ
2 + d
d( p−2)
and ϕ(1p) = 1−θ
2 . For p ∈ (2, d), the first equation yields θ = (d−2) p , and the
second equation gives
2 d −2
ϕ( p) = = p.
1−θ d−p
This proves the assertion for p ∈ (2, d). For p ≥ d the statement follows directly
from (4.12).

However, the estimate we actually would like to make use of would be of type (4.9)
for solutions of the inhomogeneous Robin problem with λ = 0, i.e.
− div Q∇u = f in ,
(4.13)
∂νQ u + δ tr u = g on ,
because in order to obtain higher regularity for the Wentzell problem we decouple it
into two underlying Robin problems of precisely that form. Though Hölder continuity
for the Robin case is also established in [22, Example 4.2.7], this is not helpful in
our situation, since an explicit estimate of type (4.8) is not given there due to the
complexity of the bootstrapping argument, and we cannot deduce (4.9), as before.
To avoid this obstacle, we rewrite the Robin problem into a Neumann problem, to
which we apply Proposition 4.5. The price we pay is that the solution u appears on
the right-hand side and we have to assume a priori that its integrability is as high as
the data’s.
Lemma 4.6. Assume Hypotheses 2.9 and 3.1. Let d ≥ 2, p ∈ (2, ∞). Let B =
3/2
− div ∇ Q. Then, there is a constant C0 > 0 such that whenever u ∈ H B () is
a weak solution of (4.13) with f, u ∈ L p (), as well as g, tr u ∈ L p (), we have
(u, tr u) ∈ L ϕ( p) () × L ϕ( p) () and

u ,ϕ( p) + tr u ,ϕ( p) ≤ C0 u , p + f , p + g , p + tr u , p
where

d−2
p if p ∈ (2, d),
ϕ( p) := d− p
(4.14)
∞ if p ∈ [d, ∞).
Proof. Let u ∈ H 1 () be a weak solution of (4.13). Then, given any v ∈ D(B0 ) ⊆
D(B N ),
u, B0 v = b(u, v) = bδ (u, v) = f, v .
Hence, u ∈ D(B0∗ ), f = B0∗ u = (B0,δ )∗ u by Theorem 3.18 (ii). Once more, we have
Bmax u = B0∗ u = f by Theorem 3.14 (ii), so the first line of (4.13) holds in L 2 (),
where − div Q∇ is seen as L 2 ()-realization of an object in H −2 (). Furthermore,
for all v ∈ H 1 () we obtain

−(B0,δ )∗ u, v + bδ (u, v) = g, tr v
for all v ∈ H 1 (), which shows u ∈ D(N bδ ) and (by Theorem 3.18) −N bδ u =
∂ν u + δ tr u = g. Therefore, u also solves (4.6) with λ = 1, f˜ = f + u ∈ L 2 (), and
Q
g̃ = g − δ tr u ∈ L 2 (), whence it must coincide with this problem’s unique solution.

Hence, Proposition 4.5 is applicable and as, due to the extra assumption, (u, tr u) is
also an element of L p () × L p (), so is ( f˜, g̃). Then, by Proposition 4.5 we have

u ,ϕ( p) + tr u ,ϕ( p) ≤ C f˜ , p + g̃ , p

≤ C0 u , p + f , p + g , p + tr u , p
as desired.
We obtain the following corollary about the integrability of elements of D(Aδ ),
where ϕ(r ) is defined as in (4.14).
Corollary 4.7. Assume Hypotheses 2.9 and 3.1 and let H = L 2 (, λd )×L 2 (, β −1 dS).
Take B = − div ∇ Q and let Aδ be given by (3.16). Let r > 2. If u ∈ D(Aδ ) ∩
(L r ()×L r ()), Aδ u ∈ L r ()×L r () and (α div Q∇u 1 , tr α div Q∇u 1 ) ∈ L r ()×
L r (), then u ∈ L ϕ(r ) ()×L ϕ(r ) () and (α div Q∇u 1 , tr α div Q∇u 1 ) ∈ L ϕ(r ) ()×
L ϕ(r ) ().
Proof. By Theorem 3.18, we have for u ∈ D(Aδ )
(Aδ u)1 = div Q∇α(div Q∇u 1 ),

(Aδ u)2 = −β(∂νQ + δ tr)(α div Q∇u 1 ) + γ u 2 .
Thus, if u satisfies the assumption of this corollary, then w = −α div Q∇u 1 solves
the inhomogeneous Robin problem
− div Q∇w = (Aδ u)1 ∈ L r ()

(∂νQ + tr δ)w = β −1 (Aδ u)2 − β −1 γ u 2 ∈ L r ().
As (α div Q∇u 1 , tr α div Q∇u 1 ) ∈ L r () × L r () by assumption, so is (w, tr w).

Hence, by Lemma 4.6, (w, tr w) ∈ L ϕ(r ) ()×L ϕ(r ) (), which also implies div Q∇u 1 ∈
L ϕ(r ) () as the functions α, α −1 are bounded. Naturally L ϕ(r ) () ⊆ L r (). Since
u ∈ D(Aδ ) ∩ (L r () × L r ()), we also know that (∂νQ + δ tr)u 1 = 0 and (u, tr u) =
(u 1 , u 2 ) ∈ L r () × L r (), whence u 1 solves the homogeneous Robin problem
− div Q∇u 1 = − div Q∇u 1 ∈ L r (),

(∂νQ + δ tr)u 1 = 0 ∈ L r ().
Applying Lemma 4.6 once more yields u 1 ∈ L ϕ(r ) () and u 2 = tr u 1 ∈ L ϕ(r ) (), as
claimed.
We can now prove the main result of this section.
Theorem 4.8. Assume Hypotheses 2.9 and 3.1 and let H = L 2 (, λd )×L 2 (, β −1 dS).
Let Aδ be defined as in (3.16). Then, there is a ϑ ∈ (0, 1), such that for any
u ∈ D(Aδ∞ ), its first component u 1 is an element of C 0,ϑ (). In particular, this
shows Hölder continuity of u 1 (t, ·) for t > 0 due to the analyticity of the semigroup
T generated by −Aδ .
3/2
Proof. Let u ∈ D(Aδ ). Then u 1 , α div Q∇u 1 ∈ H () ⊆ H 1 (). Furthermore,
tr u 1 , tr(α div Q∇u 1 ) ∈ H 1 (). By Sobolev embedding (see [2, Theorem 4.12]), we
2d
obtain H 1 ⊆ L d−2 . Hence, for d ≤ 4, we have H 1 ⊆ L d , which shows
D(Aδ ) ⊆ {u ∈ L d () × L d () | α div Q∇u 1 , tr(α div Q∇u 1 ) ∈ L d () × L d ()}.
2d 2d
If d ≥ 5, we have u 1 , α div Q∇u 1 ∈ L d−2 () and tr u 1 , tr(α div Q∇u 1 ) ∈ L d−2 ().
This shows
D(Aδ ) ⊆ {u ∈ L r1 () × L r1 () | α div Q∇u 1 , tr(α div Q∇u 1 ) ∈ L r1 () × L r1 ()}
for r1 = d−2 .
2d
Inductively, we obtain
D(Aδk ) ⊆ {u ∈ L rk () × L rk () | α div Q∇u 1 , tr(α div Q∇u 1 ) ∈ L rk () × L rk ()},
where rk = ϕ(rk−1 ) = ϕ k−1 ( d−2

2d
). Indeed, assume this statement is true for some k
k+1
and consider u ∈ D(Aδ ). Then u ∈ D(Aδk ) ⊆ D(Aδ ) and Aδ u ∈ D(Aδk ). By in-
duction hypothesis, u, Aδ u ∈ L rk ()× L rk (), and α div Q∇u 1 , tr(α div Q∇u 1 ) ∈
L rk ()×L rk (). Hence, Corollary 4.7 yields u ∈ L ϕ(rk ) ()×L ϕ(rk ) () = L rk+1 ()×
L rk+1 () as well as
α div Q∇u 1 , tr(α div Q∇u 1 ) ∈ L ϕ(rk ) () × L ϕ(rk ) () = L rk+1 () × L rk+1 ().
From the structure of the map ϕ it is clear that (rk )k∈N is an increasing sequence that
tends to ∞. Hence for all d ∈ N we have found a k0 ∈ N such that
D(Aδk0 ) ⊆ {u ∈ L d () × L d () | α div Q∇u 1 , tr(α div Q∇u 1 ) ∈ L d () × L d ()}.
For any such u ∈ D(Aδk0 ), we have − div Q∇u 1 = − div Q∇u 1 =: f˜ ∈ L d () as
well as ∂ν u 1 = −δ tr u 1 =: g̃ ∈ L d () due to α ∈ L ∞ (), δ ∈ L ∞ (). Thus u 1 is
Q
a weak solution of (4.6) for λ = 0, and Proposition 4.5 (ii) implies u 1 ∈ C 0,ϑ () as
claimed.
Remark 4.9. The proof of Theorem 4.8 actually yields a number k0 ∈ N, depending
only on the dimension d, such that u ∈ D(Aδk0 ) implies u 1 ∈ C 0,ϑ (). The number
k0 we calculated there is not sharp, but the embedding certainly does not hold for
dimensions that are too large.
For example, for the case of Neumann boundary conditions, i.e. δ = 0, we can
2d
verify D(A) ⊆ C 0,ϑ () for d ≤ 6, simply as α div Q∇u 1 ∈ H 3/2 () ⊆ L d−3 () ⊆
L d/2+ε () from which the assertions follows from Proposition 4.5 (ii) for λ = 0 (cf.
[2, Theorem 7.34]).
But, at least for constant Q and α it is quite simple to construct functions in D(A)
which are not Hölder continuous for d ≥ 8. As Sobolev embeddings are sharp, we
know that for d ≥ 8 the Sobolev space H 4 () is not contained in L ∞ (). Now,
let be a smooth domain contained in . Let v be a function that lives in H 4 ( )
but not in L ∞ ( ). As is smooth there exists an extension of v (denoted by v
again) to H 4 (Rd ), which still cannot be in L ∞ (Rd ). Now let ϕ ∈ Cc∞ () such that
ϕ = 1 on . Then the function u = ϕ · v ∈ H 4 ()\L ∞ (). Moreover, it satisfies
the Neumann boundary condition ∂ν u = 0 as it is compactly supported on . Hence
(u, tr u) ∈ D(A) (as all the other regularity conditions are implied by H 4 -regularity
because Q and α are constant). However, (u, tr u) ∈ L ∞ () × L ∞ (), so u cannot
be Hölder continuous.
4.3. Asymptotic behavior and eventual (strong) positivity
In this section, we derive asymptotic properties of our solution. We are going to skip
the proofs whenever neither variable coefficients nor extra Robin-term are relevant and
the ideas can be carried over directly from [11, Chapter 6], as is the case for the next
two results.
Lemma 4.10. Assume Hypotheses 2.9 and 3.1. Then, the operator Aδ (cf. (3.16)) has
compact resolvent.
Corollary 4.11. Assume Hypotheses 2.9 and 3.1. There exists an orthonormal basis
(en )n of H = L 2 (, λd ) × L 2 (, β −1 dS) consisting of eigenfunctions of Aδ , say
Aδ en = λn en , where the sequence λn is increasing to ∞, allowing the representation
∞

Aδ f = λk f, ek H ek
k=1
for all f ∈ D(Aδ ). Moreover, as en ∈ D(Aδ∞ ), it has a Hölder continuous representa-

tive in the sense that there exists a function en ∈ C 0,ϑ () such that en = (en | , en | ).
Finally, for all f ∈ H, the semigroup T can be represented as
∞

u(t) = (u 1 (t), u 2 (t)) = T(t)f = e−λk t f, ek H ek . (4.15)
k=1
Lemma 4.12. Assume Hypotheses 2.9 and 3.1. Let Aδ be given by (3.16). Then, the
following holds.

(i) If γ = 0 almost everywhere, then ker(Aδ ) ⊆ span(1 , 1 ) and δ| tr u 1 |2 dS =
0.
(ii) If γ = 0, δ = 0 almost everywhere, then λ1 = 0 and ker(A) = span(1 , 1 ).
(iii) If γ , δ ≥ 0 and either γ > 0 or δ > 0 on a set of positive surface measure, then
λ1> 0 and we have ker(Aδ ) = {0}.
(iv) If γ dS < 0, then λ1 < 0.
Proof. In cases (i)–(iii), we have γ , δ ≥ 0. Hence, aδ is accretive, so we have λ1 ≥ 0.

Thus, whether λ1 = 0 or λ1 > 0 depends only on ker(Aδ ).
(i) Suppose γ = 0 almost everywhere, then u ∈ ker(Aδ ) implies u 1 ∈ ker(B N ,δ ).
More precisely, let u ∈ ker(Aδ ) ⊆ D(Aδ ) ⊆ D(aδ ). Then

0 = Aδ u, uH = aδ (u, u) = α|B N ,δ u 1 |2 dx.

It follows that α|B N ,δ u 1 |2 = 0 and hence, since α(x) ≥ η, u 1 ∈ ker(B N ,δ ). This

means
√
0 = Q∇u 1 , ∇u 1 + δu 1 , u 1 = Q∇u 1 2 + δ tr u 1 2
shows 2∇u 1 = 0, whence u 1 is constant (and, therefore, also u 2 = tr u 1 ). Moreover,

and
δ| tr u 1 | dS = 0.
(ii) If δ = 0, naturally also the converse holds as B N 1 = − div Q∇1 = 0 and
the constant functions satisfy the Neumann boundary condition, which shows λ1 = 0.
(iii) If γ = 0, δ > 0 on a set of positive surface measure 0 , due to (i), u 1 is
still constant. But now, we also find a set of positive measure ε ⊆ where δ > ε.
However, if we had u = c = 0 by (i), this would yield

0= δ| tr u 1 |2 dS ≥ δ| tr u 1 |2 dS ≥ εc2 S(ε ) > 0.
ε
Analogously, if γ > 0 on a set of positive measure and u ∈ ker(Aδ ), we calculate

0 = Aδ u, uH = a(u, u) = α| u 1 |2 dx + β −1 γ |u 2 |2 dS

≥ β −1 γ c2 dS ≥ β −1
∞ εc S(ε ) > 0,
2
ε
another contradiction.
, 1 ) ∈ D(aδ ) into the usual Rayleigh quotient, we obtain a
(iv) Plugging (1
negative value as β −1 γ dS < 0, and thus λ1 < 0.
This yields the following asymptotic behavior of the semigroup T.
Theorem 4.13. Assume Hypotheses 2.9 and 3.1. Let H = L 2 (, λd )×L 2 (, β −1 dS),
Aδ be given by (3.16), and T be the semigroup generated by −Aδ . Then the following
holds.
(i) If γ = 0, δ = 0 almost everywhere, then T(t)f − f̄ H ≤ e−λ2 t f H for all
f ∈ H, where

1 −1
f̄ := f 1 dx + β f 2 dS (1 , 1 ),
λd () + β −1 dS
and λ2 > 0 is the second eigenvalue of A.

(ii) If γ , δ ≥ 0 and γ > 0 or δ > 0 on a set of positive measure, then T(t)f H ≤
e−λ1 t f H holds for all f ∈ H. Thus, in this case, the semigroup T is exponen-
tially
stable.
(iii) If γ dS < 0, then T(t) L(H) = e−λ1 t → ∞ as t → ∞.
Proof. For (i) observe that in this case λ1 = 0 and f̄ = e−λ1 t f, e1 H e1 in view of
Lemma 4.12. Thus (4.15) and Parseval’s identity yield
∞
∞

T(t)f − f̄ 2
H = e−λk t |f, ek H |2 ≤ e−λ2 t |f, ek H |2 ≤ e−λ2 t f H.
2
k=2 k=2
This proves (i). In case (ii), we have λ1 > 0 (see again Lemma 4.12), and (ii) follows
by a similar computation. (iii) follows by considering an eigenvalue corresponding to
the eigenvalue λ1 .
we consider the space H =

Remark 4.14. For our following positivity investigations,

L 2 ( ∪˙ , μ), where μ is given by μ(A) = A∩ 1 dλ + A∩ β −1 dS, with positivity
cone
H + := {u ∈ H | u(x) ≥ 0 μ-a.e}. (4.16)
Once can easily show that H can be identified with our Hilbert space H = L 2 () ×
L 2 (, β −1 dS) from Definition 2.10 (i), and u with u = (u 1 , u 2 ).
Proposition 4.15. Assume Hypotheses 2.9 and 3.1, as well as γ ≥ 0. Let Aδ be given
by (3.16). Then the semigroup T generated by −Aδ is real, but neither positive nor
L ∞ -contractive.
Proof. The same arguments as in [11, Lemma 3.5] work as Robin and Neumann
boundary conditions are equal for test functions. Note that γ ≥ 0 is assumed, as the
invariance criteria from [24] used there are stated for accretive forms only.
Again however, we may show that the semigroup T is eventually strongly positive
in the sense of [9], [10] and [8], as the critical ingredient is the Hölder continuity
established in the previous section also for the variable coefficient case. We recall a
simplified version of the definition from [8, Section 1] and the criterion used there.
Definition 4.16. Let (, μ) be a σ -finite measure space and T a real C0 -semigroup
on the space H = L 2 (, μ) with positivity cone H + given in the sense of (4.16).
Then, T is called eventually strongly positive if there is some time t0 > 0 such that for
all f ∈ H+ \{0} and t ≥ t0 there is some ε > 0 for which T (t) f ≥ ε holds μ-almost
everywhere.
Theorem 4.17. Let (, μ) be a σ -finite measure space and T a real C0 -semigroup
generated by a self-adjoint operator A on H = L 2 (, μ). If D(A∞ ) ⊆ L ∞ (, μ),
then the following assertions are equivalent:
(i) T is eventually strongly positive.
(ii) ker(s(A) − A) is one-dimensional and contains a vector v such that v ≥ ε holds
μ-almost everywhere for some ε > 0.
Theorem 4.18. Assume Hypotheses 2.9 and 3.1, and additionally that γ = δ = 0.
Let A be given by (3.16) (for δ = 0). Then T, the semigroup generated by −A, is
eventually strongly positive in the sense of Definition 4.16.
Proof. We apply Theorem 4.17 for H defined as in Remark 4.14 and A = −A. As
β, β −1 are bounded, L ∞ ( ∪˙ , μ) can be identified with L ∞ () × L ∞ (). T is
real as a consequence of Proposition 4.15, and the operator A is self-adjoint due to
the symmetry of the form (see Theorem 2.12). Finally, we have that D(A ∞ ) embeds
into L ∞ () × L ∞ () by Theorem 4.8. Now, to deduce eventual strong positivity,
we only have to verify assertion (ii) of Theorem 4.17. But this a direct consequence
of Lemma 4.12 (ii) that yields s(−A) = λ1 = 0 as γ = 0, δ = 0, and (1 , 1 ) ∈
ker(A).
For a counterexample for eventual positivity in the case γ > 0, δ = 0, Q = Id we
refer the reader to [11, Section 7].
5. Application to systems of higher order
In this short excursion, we point out that the abstract theory established in Chapter 2
is not necessarily restricted to problems of order 4, though identification of the abstract
operators and regularity characterizations can be more difficult.
Proposition 5.1. Let ⊂ Rd be a bounded Lipschitz domain, H = L 2 (), and let

N denote the Neumann Laplacian. For k ∈ N consider the form bk =
k u, k v

on D(bk ) = D( N ). Then bk is admissible and B N = N , whence also Cc∞ () ⊂
k 2k
D(B N ) ∩ D(B D ) is satisfied, and we are in the situation of Theorem 2.12 and Re-
mark 2.13.
Proof. The form bk is symmetric and accretive by default. We have u 2bk = u 2 +
k u 2 . As k is a closed operator. D( k ) is a Hilbert space with respect to
N N
the graph norm and hence the form is closed. The continuity follows from Cauchy–
Schwarz’s inequality. Thus bk is generating and its associated operator B N is self-
adjoint. We have Cc∞ () ⊆ D( kN ) ⊆ D( N ) ⊆ H 1 () and may choose ρ = 1/2
in (2.1). For the last embedding we calculate u 2H 1 () = u, u ≤ C( u 2 +
u 2 ). That D( kN ) is continuously embedded in D( N ) follows by the closed
graph theorem applied to the identity id: D( kN ) → D( N ). D( kN ) is also dense
in H 1 () due to [24, Proof of Lemma 1.25] and the analyticity of the semigroup. We
begin by showing that B N coincides with 2k distributionally. Let u ∈ D(B ) and
k N
v ∈ Cc () ⊂ D( N ), then f, ϕ = N u, kN ϕ = ( kN )∗ u, kN ϕ = u, 2k ϕ
∞ 2k
as k ϕ ∈ D( kN ). Hence f = 2k u as N and thus kN is self-adjoint. So D(B N ) ⊂

{u ∈ D( kN ) | 2k N u ∈ L ()}. On the other
2 hand D( 2k ) ⊂ D(B N ) as for the
2k Nk
same reason for u ∈ D( N ) the relation N u, v =
2k
N u, N v holds for all
k
v ∈ D( kN ). So 2k ∗ 2k ∗
N ⊂ BN = BN ⊂ ( N ) = N .
2k
Now we obtain D(B0 ) = D(B D ) ∩ D(B N ) = D(B N ) ∩ H01 () by Proposition 2.4.
Thus D(B0 ) = {u ∈ D( 2k ∗
N ) | tr u = 0}. Now for u ∈ D(B0 ), there is a f ∈ L ()
2
such that for all v ∈ D(B0 ) we have u, B0 v = u, 2k ∗

v = f, v , so B0 is a
restriction of the maximal L -realization of the distribution 2k .
2
Hence (taking α = β = 1, γ = δ = 0 in Remark 2.13) we can find a solution of

the system
∂t u + 4k
u=0 in (0, ∞) × , (5.1)
bk
tr 4k
u+N ( 2k
)u = 0 on (0, ∞) × , (5.2)
bk
N u=0 on (0, ∞) × , (5.3)
u|t=0 = u 0 in , (5.4)
where N bk is given by the abstract definition

bk
D(N ) := {u ∈ D(B0∗ ) |
∃g ∈ L 2 ()∀v ∈ D( N)
2k
: 2k
u, v − u, 2k
v = g, tr v }

and N bk = g. The identification of N bk is more difficult here, but some progress

can be made using the theory of quasi-boundary triples (cf. [7], [6, Chapter 8.6]).
For all u ∈ H 0 () and v ∈ D( N ) we have
u, v − u, v = τ̃ N u, tr vG ×G0 , (5.5)
0
where τ̃ N is the extension of the Neumann trace from H 0 () to the space G0 introduce
there, which is the dual space of tr D( N ) equipped with a Hilbert space structure.
For this version of Green’s formula also see [11, Proposition 2.4]. This allows us to
characterize N bk at least on a subset of D(B0∗ ). Let
V = {u ∈ L 2 () | 2k
u, 2k−1
u ∈ L 2 (), j
u ∈ ker τ̃ N , for j = 0, ..., 2k − 2}.
Lemma 5.2. Let ⊂ Rd be a bounded Lipschitz domain, H = L 2 (), let bk be

the admissible form from Proposition 5.1, and N bk the corresponding trace operator
in the sense of Definition 2.6. Let u ∈ D(N bk ) ∩ V , then j u ∈ H () for
3/2
j = 0, ..., 2k − 1 and N bk u = ∂ν 2k−1 u.

Proof. As we have j u ∈ L 2 () for j = 0, ..., 2k and D(B0 ) ⊂ D(B N ) = D( 2k )
N
for u ∈ D(N bk ) ∩ V and v ∈ D( 2kN ) we have

2k−1
g, tr v = 2k
u, v − u, 2k
v = τ̃ N j
u, tr 2k−1− j
v
G0 ×G0
j=0
= τ̃ N 2k−1
u, tr v .
G0 ×G0
N ) is dense in H (), tr D(
Now as D( 2k 1 2k ) is dense in L 2 (). Thus we have
N
g = τ̃ N 2k−1 u ∈ L () and
2
2k
u, v − 2k−1
u, v = τ̃ N 2k−1
u, tr v = g, tr v
G0 ×G0
for all v ∈ D( 2k N ), and by approximation also for all v ∈ D( N ) (as D( N ) is

2k
3/2
a core of N ). Finally, we obtain 2k−1 u ∈ H () and N bk u = ∂ν 2k−1 u for
3/2
u ∈ D(N bk ) ∩ V by [11, Proposition 2.4 (ii)] and similarly j u ∈ H () for all
j = 0, ..., 2k − 1.
Funding Open Access funding enabled and organized by Projekt DEAL.
Data availability Data sharing is not applicable to this article as no datasets were
generated or analyzed during the current study.
Declarations
Conflict of interest The author declares that he has no conflict of interest.
Open Access. This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative
Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use
is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission
directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/
by/4.0/.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims
in published maps and institutional affiliations.
REFERENCES
[1] Herbert Amann and Joachim Escher. Strongly continuous dual semigroups. Annali di Matematica
Pura ed Applicata, 171:41–62, 1996.
[2] Robert A. Adams and John J. F. Fournier. Sobolev spaces, volume 140 of Pure and Applied Math-
ematics (Amsterdam). Elsevier/Academic Press, Amsterdam, 2003.
[3] Wolfgang Arendt, Giorgio Metafune, Diego Pallara, and Silvia Romanelli. The Laplacian with
Wentzell-Robin boundary conditions on spaces of continuous functions. Semigroup Forum,
67(2):247–261, 2003.
[4] Tim Binz and Klaus-Jochen Engel. Operators with Wentzell boundary conditions and the Dirichlet-
to-Neumann operator. Math. Nachr., 292(4):733–746, 2019.
[5] Jussi Behrndt, Fritz Gesztesy, and Marius Mitrea. Sharp boundary trace theory and Schrödinger
Operators on Bounded Lipschitz Domains. 2022.
[6] Jussi Behrndt, Seppo Hassi, and Henk de Snoo. Boundary value problems, Weyl functions, and
differential operators, volume 108 of Monographs in Mathematics. Birkhäuser/Springer, Cham,
2020.
[7] Jussi Behrndt and Till Micheler. Elliptic differential operators on Lipschitz domains and abstract
boundary value problems. Journal of Functional Analysis, 267(10):3657–3709, 2014.
[8] Daniel Daners and Jochen Glück. A criterion for the uniform eventual positivity of operator semi-
groups. Integral Equations Operator Theory, 90(4):Paper No. 46, 19, 2018.
[9] Daniel Daners, Jochen Glück, and James B. Kennedy. Eventually positive semigroups of linear
operators. J. Math. Anal. Appl., 433(2):1561–1593, 2016.
[10] Daniel Daners, Jochen GlÃck, and James B. Kennedy. 2016. Eventually and asymptotically positive
semigroups on Banach lattices. J. Differential Equations, 261(5):2607–2649,
[11] Robert Denk, Markus Kunze, and David Ploß. The Bi-Laplacian with Wentzell boundary conditions
on Lipschitz domains. Integral Equations Operator Theory, 93: Paper No. 13, 26pp, 2021.
[12] Robert Denk, David Ploß, Sophia Rau, and Jörg Seiler. 2023, Boundary value problems with rough
boundary data. Journal of Differential Equations, 366:85–131, .
[13] Dıaz JI, Tello L. 2008, On a climate model with a dynamic nonlinear diffusive boundary condition.
Discrete Contin. Dyn. Syst. Ser. S, 1(2):253–262,
[14] Klaus-Jochen Engel and Genni Fragnelli. Analyticity of semigroups generated by operators with
generalized Wentzell boundary conditions. Adv. Differential Equations, 10(11):1301–1320, 2005.
[15] Klaus-Jochen Engel. Second order differential operators on C[0, 1] with Wentzell-Robin boundary
conditions. In Evolution equations, volume 234 of Lecture Notes in Pure and Appl. Math., pages
159–165. Dekker, New York, 2003.
[16] Joachim Escher, Jan Prüss, and Gieri Simonett. Analytic solutions for a Stefan problem with Gibbs-
Thomson correction. J. Reine Angew. Math., 563:1–52, 2003.
[17] Angelo Favini, Gisèle Ruiz Goldstein, Jerome A. Goldstein, and Silvia Romanelli. 2008. Fourth
order operators with general Wentzell boundary conditions. Rocky Mountain J. Math., 38(2):445–
460,
[18] Fritz Gesztesy and Marius Mitrea. Generalized Robin boundary conditions, Robin-to-Dirichlet
maps, and Krein-type resolvent formulas for Schrödinger operators on bounded Lipschitz domains.
In Perspectives in partial differential equations, harmonic analysis and applications, volume 79 of
Proc. Sympos. Pure Math., pages 105–173. Amer. Math. Soc., Providence, RI, 2008.
[19] Goldstein GR.. 2006, Derivation and physical interpretation of general boundary conditions. Adv.
Differential Equations, 11(4):457–480,
[20] VA Kondrat’eV. 1967, Boundary problems for elliptic equations in domains with conical or angular
points. Trans. Moscow Math. Soc, 16(227-313):129–133,
[21] Jacques-Louis Lions and Enrico Magenes. Non-homogeneous boundary value problems and appli-
cations. Vol. I. Springer-Verlag, New York-Heidelberg, 1972.
[22] Robin Nittka. Elliptic and parabolic problems with Robin boundary conditions on Lipschitz do-
mains. PhD thesis, Universität Ulm, 2010. https://doi.org/10.18725/OPARU-1790.
[23] Robin Nittka. Regularity of solutions of linear second order elliptic and parabolic boundary value
problems on Lipschitz domains. J. Differential Equations, 251(4-5):860–880, 2011.
[24] El Maati Ouhabaz. Analysis of heat equations on domains, volume 31 of London Mathematical
Society Monographs Series. Princeton University Press, Princeton, NJ, 2005.
[25] David Ploß. Wentzell boundary conditions for elliptic fourth-order operators. PhD thesis, Universität
Konstanz, Konstanz, 2024.
[26] Jan Prüss, Reinhard Racke, and Songmu Zheng. 2006, Maximal regularity and asymptotic behavior
of solutions for the Cahn-Hilliard equation with dynamic boundary conditions. Ann. Mat. Pura
Appl. 185(4):627–648,
[27] Reinhard Racke and Songmu Zheng. The Cahn-Hilliard equation with dynamic boundary condi-
tions. Adv. Differential Equations, 8(1):83–110, 2003.
[28] Hans Triebel. Interpolation theory, function spaces, differential operators. Johann Ambrosius Barth,
Heidelberg, 1995.
[29] Mahamadi Warma. Parabolic and elliptic problems with general Wentzell boundary condition on
Lipschitz domains. Commun. Pure Appl. Anal., 12(5):1881–1905, 2013.
[30] J. Wloka. Partial differential equations. Cambridge University Press, Cambridge, 1987. Translated
from the German by C. B. Thomas and M. J. Thomas.
David Ploß
Department of Mathematics
Karlsruhe Institute of Technology
Englerstraße 2
76131 Karlsruhe
Germany
E-mail: [email protected]
Accepted: 16 September 2024

Operadores Elípticos de Quarta Ordem Com Condições de Contorno de Wentzell em Domínios de Lipschitz

Uploaded by

Copyright:

Available Formats

Operadores Elípticos de Quarta Ordem Com Condições de Contorno de Wentzell em Domínios de Lipschitz

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Operadores Elípticos de Quarta Ordem Com Condições de Contorno de Wentzell em Domínios de Lipschitz

Uploaded by

Copyright:

Available Formats

J. Evol. Equ.

Elliptic fourth-order operators with Wentzell boundary conditions

Wentzell, or dynamic boundary conditions, appear in a multitude of physical ap-

Mathematics Subject Classification: 35K35 (primary); 47A07, 47D06, 35B65 (secondary)

A detailed discussion of a physical interpretation of these boundary conditions in

∂t u + B(α B)u = 0 in (0, ∞) × , (1.1)

∂t u 1 + B(α B)u 1 = 0 in (0, ∞) × , (1.5)

Note that, as u 2 is independent of u 1 , we have to impose an additional initial condition

∂t u + Au = 0 for u(t, ·) in H, (1.10)

2. The abstract setting

The Dirichlet-trace on C ∞ (), defined by u → u | , and its extension to any Sobolev

we recall that its associated operator is given by

Recall, that if a form b : D(b) × D(b) → C (D(b) ⊂ H ) is densely-defined,

holds, where the latter embedding is continuous and dense.

(ii) The operator B N associated to b on L 2 () is called the Neumann realization of

B N u = B D u for u ∈ D(B D ) ∩ D(B N ).

 f N , v = b(u, v) = b D (u, v),

which shows u ∈ D(B D ) and B D u = f N = B N u. The converse is trivial.

D(N b ) := {u ∈ D(b) ∩ D(B0∗ ) |

and N b u := g. Let furthermore N b : D(N ) ⊆ L 2 () → L 2 () be the linear

(iv) If N b is surjective, we have N b = N b .

N b ) is densely defined. As B0 ⊆ B N and B N is self-adjoint, we have B N =

Next we introduce a primary form, which will be connected to the generalized

∂t u 1 + B0∗ (α B N )u 1 = 0 in (0, ∞) × , (2.4)

We define the following regularity assumptions for the coefficients.

where u = (u 1 , u 2 ), v = (v1 , v2 ) ∈ H and

for all u, v ∈ D(a) where

Proof. We may assume without loss of generality that β ≡ 1, otherwise switch to

D(A) = {u ∈ H | u 1 ∈ D(B N ), α B N u 1 ∈ D(N b ), u 2 = tr u 1 }.

Proof. (i) We begin by showing that a is a generating, symmetric form. As b is

The symmetry is trivial as α, β, γ are real-valued. Next we consider the induced

u 2 ≤ C u 2b = C(b(u, u) + (1 − λb ) u 2 ) = C((1 − λb ) u 2 + B N u, u )

u n2 = tr u n1 → tr u in L 2 (). Setting u = (u, tr u), we see that u ∈ D(a) and

D(C) = {u ∈ H | u 1 ∈ D(B N ), α B N u 1 ∈ D(N b ), u 2 = tr u 1 }.

We want to show C = A. We begin by showing C ⊆ A. So let u ∈ D(C) ⊆

 f 1 , v1  = f, vH = a(u, v) = α B N u 1 , B N v1  = α B N u 1 , B0 v1  .

This shows that α B N u 1 is in D(B0∗ ) and f 1 = B0∗ (α B N u 1 ) by definition of the

for all v1 ∈ D(B N ), which shows α B N u 1 ∈ D(N b ) (and u ∈ D(C )) as well

or, equivalently, f 2 = −βN b (α B N u 1 ) + γ u 2 , which shows Cu = f = Au.

∂t u + B0∗ (α B N )u = 0 in (0, ∞) × , (2.11)

As the semigroup is analytic, the solution is C ∞ in time so that

tr B0∗ (α B N )u = tr(Au)1 = (Au)2 = −βN b (α B N )u + γ tr u,

3. Application to strongly elliptic operators in divergence form

In this section, we will specify the operator B to be a strongly elliptic second-order

3.1. Co-normal traces

Definition 3.3. Let ⊆ Rd be a Lipschitz domain with outward normal ν. We

ν · tr Q∇u, where we read the operator tr component-wise.

γ Qs : H Bs () → H s−3/2 () (3.3)

for all s ∈ [ 21 , 23 ], forming a compatible family in s. Furthermore, for all s ∈ [ 21 , 23 ],

Proof. This is [5, Corollary 11.25 and 11.28].

3.2. On the second-order operator

b(u, v) = Q∇u, ∇v (3.5)

So b is closed and continuous, and therefore b is generating. It is symmetric, as Q is

Q∇ϕ, ∇v + div Q∇ϕ, v = ν · tr Q∇ϕ, tr v = 0

as ∇ϕ = 0 close to the boundary, which shows

b(ϕ, v) = − div Q∇ϕ, v

for all v ∈ D(b) and thus ϕ ∈ D(B N ) and B N ϕ = − div Q∇ϕ.

So we may define B0 , B0∗ , N b , and N b as stated in Theorem 2.8. Furthermore, the

Proof. For any function ϕ ∈ C ∞ (), so as seen above Q∇ϕ is an element of

f N , v = b(u, v) = b D (u, v),

u 2 ≤ C u 2b = C(b(u, u) + (1 − λb ) u 2 ) = C((1 − λb ) u 2 + B N u, u )

f 1 , v1 = f, vH = a(u, v) = α B N u 1 , B N v1 = α B N u 1 , B0 v1 .

b(u, v) = Q∇u, ∇v (3.5)

Q∇ϕ, ∇v + div Q∇ϕ, v = ν · tr Q∇ϕ, tr v = 0

b(ϕ, v) = − div Q∇ϕ, v

Q∇ϕ, ∇v + div Q∇ϕ, v = ν · tr Q∇ϕ, tr v .

− div Qϕ, v = Q∇ϕ, ∇v = b(ϕ, v) = ϕ, B0 v ,

ε = {x ∈ Rd | dist(x, ) < ε}

γ : (−ε, ε) × → ε ; (h, x ) → γ (h, x ) := x + h · ∂ν (x ).

Choosing ε > 0 small enough, it is possible to guarantee ε ∈ \, so Q is defined

Q∇ϕn , ∇v = − div Q∇ϕn , v .

Bu(ϕ) := u, − div Q∇ϕ

D(Bmax ) := {u ∈ L 2 () | Bu ∈ (L 2 ()) },

(u) = h 1 , F H −s (Rd )×H s (Rd ) + e0 h 2 , G .

0 h 2 ∈ H −2 (Rd ), or for some large λ ≥ λ0 also −h 1 +λe0 h 2 =

(u) = h 1 , F H −s (Rd )×H s (Rd ) + e0 h 2 , e0 Bmax u

of Proposition 3.8 we have Bmax ϕn , v − b(ϕn , v) = −ν · tr Q∇ϕn , tr v =

Bmax u, v − b(u, v) = N b u, tr v = N b u, tr v

Bmax u, v − b(u, v) = lim Bmax ϕn , v − b(ϕn , v)

bδ (u, v) = Q∇u, ∇v + δu, v (3.15)

|λ − a0 (x, ξ )| ≥ C |λ| + |ξ |2m (x ∈ , λ ∈ , ξ ∈ Rd , (ξ, λ) = 0) (4.1)

(λ − a0 (x0 , ξ , Dd ))w(xd ) = 0 (xd ∈ (0, ∞)),

for all |ξ | = 0 as the matrix Q = (q jk ) jk is symmetric and uniformly positive defi-

− β −1 (x0 )b0,1 (x0 , )w(0) · w(0)

u, B0 v = b(u, v) = bλ (u, v) − λ u, v = f˜ − λu, v .