Z t s
⟨DEr2(u(r)), u′(r)⟩V∗×Vdr
= lim
h→0−
Z t s
⟨DEr2(u0+ (KShu′)(r)),(Shu′)(r)⟩V∗×V dr
= lim
h→0
Es2(u0+ (KShu′)(s))− Et2(u0+ (KShu′)(t))
=Es2(u(s))− Et2(u(t)) (6.4.5)
for all s, t∈[0, T]. Second, it has been shown in Emmrich & Thalhammer [77, Lemma 6] that
−
Z t 0
⟨u′′(r) + E(u(r)), u′(r)⟩V∗×V dr
≤ 1
2|v0|2− 1
2|u′(t)|2 +E1(u0)− E1(u(t))
for almost everyt ∈(0, T). The latter inequality together with (6.4.5) implies (6.4.3), which completes the proof.
6.5 Proof of Theorem 6.1.4
Letu0 ∈U, v0 ∈H and (τn)n∈N be a vanishing sequence of positive step sizes. Let (uk0)k∈N ⊂ U ∩V and (v0k)k∈N ⊂ V be such that uk0 → u0 in U and vk0 → v0 in H as k → ∞. We let k ∈ N be fixed and we denote the interpolations associated with the initial data uk0 andvk0 again by (5.3.1)-(5.3.3) and (6.3.1). Henceforth, we suppress the dependence of the interpolations on k for simplicity. By the previous lemma, there exists a subsequence (relabeled as before) of the interpolations and limit
functionsu∈Cw([0, T];U)∩W1,∞(0, T;H)∩W1,q(0, T;V∗)∩W2,r∗(0, T;U∗+V∗) (notice that uk0 ∈ U ∩V) and u(0) = uk0 in U and u′(0) = v0k in H such that the convergences (6.4.1) hold, where we again suppress the dependence of the limit functions onk. First, we prove that the inclusion (6.1.3) holds. To do so, we note that the Euler–Lagrange equation (6.3.7) reads
Vbτ′n(t) +ητn(t) + DEtτn(t)(Uτn(t)) +Sτn(t) = 0 inU∗+V∗,
ηn(t)∈∂VΨUτn(t)(Vτn(t)) (6.5.1)
for allt∈(0, T), where Sτn(t) = B(tτn(t), Vτn(t), Uτn(t))−fτn(t), t∈[0, T]. Testing equation (6.5.1) with w∈Lmax{2,q}(0, T;U ∩V), we obtain
Z T 0
⟨Vbτ′n(r) +ητn(r) + DEtτn(s)(Uτn(s)) +Sτn(r), w(r)⟩(U∗+V∗)×(U∩V)dr= 0.
Then, with the aid of the convergences (6.4.1), we are allowed to pass to the limit in the weak formulation obtaining
Z T 0
⟨u′′(r) +η(r) + DEs(u(s)) +B(t, u(r), u′(r))−f(r), w(r)⟩(U∗+V∗)×(U∩V)dr= 0 for allw∈Lmax{2,q}(0, T;U∩V). Then, by a density argument and the fundamental lemma of calculus of variations, we deduce
u′′(t) +η(t) + DEt(u(t)) +B(t, u(t), u′(t)) =f(t) in U∗+V∗
for a.e. t ∈ (0, T). We shall identify the weak limit η as subgradient of the dissipation potential almost everywhere, i.e, η(t) ∈∂VΨu(t)(u′(t)) for almost every t∈(0, T). For that purpose, we will employ Lemma 2.4.4 with fn(t, v) =ΨU
τn(t)(v) andf(t, v) = Ψu(t)(v) for all v ∈X =V andn ∈N. Assumption (2.4.1) is already fulfilled by Condition (6.Ψc). Hence, it remains to show that
lim sup
n→∞
Z T 0
⟨ηn(t), Vτn(t)⟩V∗×V dt ≤
Z T 0
⟨η(t), u′(t)⟩V∗×V dt. (6.5.2) In order to show the latter limes superior estimate, we use the fact thatητn can be expressed through the remaining terms of the Euler–Lagrangeequation (6.5.1).
Therefore, we will split the integral on the left-hand side of (6.5.2) and note first that
−
Z t 0
⟨Vbτ′n(r), Vτn(r)⟩V∗×V dr
=−
Z t 0
⟨Vbτ′n(r),Vbτn(r)⟩V∗×V dr+
Z t 0
⟨Vbτ′n(r),Vbτn(r)−Vτn(r)⟩V∗×V dr
= 1
2|v0|2− 1
2|Vbτn(t)|2+
Z t 0
⟨Vbτ′n(r),Vbτn(r)−Vτn(r)⟩V∗×V dr
≤ 1
2|v0|2−1
2|Vbτn(t)|2,
where we used the fundamental theorem of calculus for the absolutely continuous
We continue with the term involving the derivative of the energy functional and start with the linear part:
−
which can be shown in the same way as above by using the strong positivity of E.
As for the nonlinear part, we obtain by employing the λ-convexity of Et2 that
−
+ tm−t
Now, we want to make use of the inequality (6.4.3). However, the aforementioned inequality only holds true for almost everyt ∈(0, T). Therefore, we take a sequence of increasing values (βl)l∈N,βi ∈(0, T) for all i∈N, converging toT for which (6.4.3) holds true. Then, choosingt =βl, we obtain with the convergences (6.4.1b), (6.4.1h), (6.4.1e), and (6.4.1g), the sequential weak lower semicontinuity of Et1 and | · | and
the continuity ofEt1, the limes superior condition and growth condition (6.Ed) on
∂tEt2 and Fatou’s Lemma that
Then, in view of the convergences (6.4.1m) and (6.4.1n), the Euler–Lagrange equation (6.5.1), we obtain
u satisfying the inclusion (6.1.3), and the initial values u(0) = uk0 ∈ U ∩V and u′(0) = vk0 ∈ V. Denote with (uk)k∈N the sequence of solutions to the associated sequence of initial values, and with (ηk)k∈N the subgradients of Ψuk(t)(u′k(t)). In the last step, we want to show that there exists a limit functionu which satisfies (6.1.3) and (6.1.4) as well as the intial valuesu(0) =u0 in U andu′(0) = v0 inH. We recall that uk0 →u0 in U and v0k → v0 in H as k → ∞. As in Chapter 5, the next steps are the following.
1. We derive a priori estimates based on the energy-dissipation inequality (6.1.4),
2. We show compactness of the sequences (uk)k∈N and (ηk)k∈N in appropriate Condition (6.Ed), (6.Bb), and (6.Bb), we obtain with the lemma of Gronwall (Lemma A.1.1)
Ad 2. With the same reasoning as for the interpolations, we obtain the convergences uk⇀ u∗ in L∞(0, T;U), (6.5.3a) uk−uk0 ⇀ u∗ −u0 in L∞(0, T;V), (6.5.3b) uk(t)⇀ u(t) in U for all t∈[0, T], (6.5.3c) uk(t)−uk0 ⇀ u(t)−u0 in V for all t∈[0, T], (6.5.3d) uk →u in Lr(0, T;Wf) for any r≥1, (6.5.3e) uk(t)→u(t) in Wf for all t∈[0, T], (6.5.3f) u′k(t)⇀ u∗ ′ in Lq(0, T;V)∩L∞(0, T;H), (6.5.3g) u′k(t)→u′ in Lp(0, T;H) for all p≥1, (6.5.3h) u′k(t)⇀ u′(t) in H for all t∈[0, T], (6.5.3i)
ητk
n ⇀ η in Lq∗(0, T;V∗), (6.5.3j) Euk⇀Eu in L2(0, T;U∗), (6.5.3k) DEt2(uk)→DEt2(u) in Lr(0, T;U∗) for any r≥1, (6.5.3l) u′′k ⇀ u′′ in Lmin{2,q∗}(0, T;U∗+V∗), (6.5.3m) B(·, uk, u′k)→B(·, u, u′) in Lr∗(0, T;V∗), (6.5.3n) and if Ψu satisfies (6.Ψd), then
u′k→u′ in Lmax{2,q}(0, T;U), uk →u in C([0, T];U).
Ad 3. Therefore, u ∈ Cw([0, T];U)∩W1,∞([0, T];H)∩W2,q∗(0, T;U∗ +V∗) with u−u0 ∈W1,q(0, T;V) and η∈Lq∗(0, T;V∗) satisfies the initial conditions u(0) = u0 inU and u′(0) =v0 in H. Along the same lines as for the interpolations, we obtain with Condition (6.Ψc) and Lemma 2.4.4 that u andη satisfy the inclusions (6.1.3).
It remains to show the energy-dissipation balance (6.1.4). The inequality 1
2|u′(t)|2+Et(u(t)) +
Z t 0
Ψu(r)(u′(r)) +Ψu(r)∗ (S(r)−DEr(u(r))−u′′(r))dr
≤ 1
2|v0|2+E0(u0) +
Z t 0
∂rEr(u(r)) dr+
Z t 0
⟨S(r), u′(r)⟩V∗×V dr,
for all t ∈ [0, T] with S(r) =f(r)−B(r, u(r), u′(r)) is obtained by passing to the limit ask → ∞while taking into account the convergences (6.4.1). Then, employing again (6.4.3) and theFenchel–Young inequality, we obtain
Z t 0
Ψu(r)(u′(r)) +Ψu(r)∗ (S(r)−DEr(u(r))−u′′(r))dr
≤ 1
2|v0|2− 1
2|u′(t)|2+E0(u0)− ET(u(t)) +
Z t 0
∂rEr(u(r)) dr +
Z t 0
⟨S(r), u′(r)⟩V∗×V dr
≤
Z t 0
⟨DEr(u(r))−u′′(r), u′(r)⟩V∗×V dr+
Z t 0
⟨S(r), u′(r)⟩V∗×V dr
=
Z t 0
⟨S(r)−DEr(u(r))−u′′(r), u′(r)⟩V∗×V dr
≤
Z t 0
Ψu(r)(u′(r)) +Ψu(r)∗ (S(r)−DEr(u(r))−u′′(r))dr
for almost everyt∈(0, T). Now, if V ,→U, then the inequality (6.4.3) indeed holds as equality for allt ∈[0, T] by the classical integration by parts formula. This shows (6.1.4), and hence the completion of the proof.
Remark 6.5.1 The proof of Theorem 6.1.4 reveals that one can consider dissipation potentials that depend on a parameter ε. In this case, the Condition (6.Ψa) is assumed to hold for every ε≥ 0 while Condition (6.Ψb) holds uniformly in ε≥ 0.
Condition (6.Ψc) can either be replaced with the Mosco-convergence Ψuεnn −→M Ψu0 for every sequenceun ⇀ uas ε↘0, or with a more general liminf estimate (2.4.1).
Applications
In this section, we want to apply the abstract results on linear and nonlinear inertial systems developed and proven in Chapter 5 and 6, respectively, to concrete examples.
We will give a sufficient number of examples to cover the range of applications from the abstract results. Since our main results are established in a nonsmooth setting, the examples with nonsmooth functionals, in particular, can not be cast into the framework of the existing results. Those nonsmooth functionals correspond to multi-valued equations or nonlinear constraints. We first start with examples for linearly damped inertial systems and continue with examples for nonlinearly damped inertial systems. We assume the same notation and function spaces as in Chapter 4.
7.1 Differential inclusion I A
In the first example, we consider a system, which can be treated in the Case (a) of the linearly damped intertial system, where the dissipation potential is given by the Dirichletenergy and the energy functional is a nonsmoothλ-convex function which to the best of the authors’ knowledge can not be treated with the abstract results known thus far. More precisely, we consider the initial-boundary value problem
(P1)
∂ttu−∆∂tu−∆pu+ (|u|2−1)u− ∇ ·p+b(x, t,u, ∂tu) =f inΩT, p(x, t)∈Sgn(∇u(x, t)) a.e. in ΩT,
u(x,0) =u0(x) onΩ, u′(x,0) = v0(x) onΩ,
u(x, t) = 0 on ∂Ω×[0, T],
where Sgn : Rd×m ⇒ Rd×m is the sign function defined in (4.0.1), f : Ω → Rm, b : Ω ×[0, T]×Rm ×Rm → Rm a Carathéodory function in the sense that b(x,·,·,·) is continuous for almost every x∈Ω and b(·, t,y,z) is measurable for all t∈[0, T] and y,z ∈Rm. Furthermore, b is assumed to satisfy the following growth condition: there exists a constantCb >0 and numbersq, r > 1 such that
|b(x, t,u,v)| ≤Cb(1 +|u|q−1+|v|r−1) for a.e. x∈Ω, t∈[0, T] and all u,v ∈Rm. Here, p, q, r≥1 are to be chosen in accordance with the assumptions.
Choosing the spaces U = W1,p0 (Ω)m∩L4(Ω)m, V = H10(Ω)m, W = Lmax{2,q}(Ω)m andH = L2(Ω)m equipped with the standard norms, we assume f ∈L2(0, T;V∗).
The energy functional E :V →(−∞,+∞] and the dissipation potential Ψ :V →R are given by
E(u) =
R
Ω
1
p|∇u(x)|p+|∇u(x)|+14(|u(x)|2−1)2dx if u∈dom(E), +∞ otherwise,
and
Ψ(v) = 1 2
Z
Ω
|∇v(x)|2dx,
respectively, whereas the perturbationB : [0, T]×W ×H →V∗ is defined by
⟨B(t,u,v),w⟩V∗×V =⟨B(t,u,v),w⟩W∗×W =
Z
Ω
b(x, t,u(x),v(x))·w(x) dx.
The Legendre–Fenchel transformation Ψ∗ : H−1(Ω)m → R of Ψ is obviously given byΨ∗(ξ) = 12∥ξ∥2−1,2. Furthermore, it is readily seen that the energy functional is not Gâteaux differentiable and its effective domain is given by dom(E) = W01,p(Ω)m∩L4(Ω)m. The valuesp, q, r ≥1 are to be chosen such that all assumptions are fulfilled. We can choose, e.g.,
d= 1, p∈(1,+∞), r∈[1,2], q∈[1, p/2 + 1], d= 2, p∈(1,+∞), r∈[1,2], q∈
[1, pd/(p−d))∩[1, p/2 + 1] if p∈(1,2), [1, p/2 + 1] if p≥2,
d≥3, p∈(1,+∞), r∈[1,2], q ∈
[1, q∗) if p∈(1,2), [1, p/2 + 1] if p≥3,
where q∗ = min{d(p+2)2(d−p),3d+4d , p+ 1}. Then, by the Sobolev embedding theorem and theRellich–Kondrachov theorem,U and V are densely, continuously and compactly embedded in W and H, respectively. We will verify for illustration the assumptions for the case d≥3. Since the dissipation potential is state-independent, it is induced by the bilinear form a:V ×V →R,
a(v,w) := 1 2
Z
Ω
∇v· ∇wdx,
and therefore satisfies all conditions. The conditions (5.Eb)-(5.Ed) are obviously fulfilled by the energy functional. In order to verify (5.Ea), we note that every convex and lower semicontinuous functional on a Banach space is weakly lower semicontinuous. Taking the latter into account, we observe that foru∈dom(E), the
energy functional V. The lower semicontinuity of W on V follows immediately from the converse of the dominated convergence theorem (see, e.g., Brézis [35, Theorem 4.9, p. 94]) and Fatou’s lemma. Further, due to the compact embedding of V in H, the concave function is continuous on V with respect to the weak topology. This implies E to be weakly lower semicontinuous onV. In fact, the convex part of the energy is perturbed by the negativeHilbert space norm ofH squared which by the parallelogram law and the embedding V ,→ H leads to the λ-convexity of E with λ := C being the constant of the very same embedding. Now, we show the closedness property (5.Ee).
First, we note that for each u ∈ D(∂E), there holds ξ ∈ ∂UE(u) = ∂UW(u)−u
andΛu =∇u. We note thatΛ is linear and bounded and has as adjoint operator Λ∗ : L∞(Ω)d×m → W−1,p∗(Ω)m,A 7→ −∇ ·A the divergence operator. Let u ∈ dom(∂UW), then by the variational sum rule (Lemma2.2.7) and Lemma 2.2.8, there holds
Second, we express ∂XW2(Λu) with the aid of Lemma 2.3.1 equivalently through the equation
⟨p, Λu⟩X∗×X =W2(Λu) +W2∗(p). (7.1.1) Third, by Ekeland & Temam [69, Proposition 1.2, p. 87], the conjugate W2∗ is given by
W2∗(B) =
Z
Ω
ıB
Rm×d(0,1)(B(x)) dx, with the indicator function ıB
Rm×d(0,1) → {0,+∞} defined by ıB
Rm×d(0,1)(A) =
0 if |A| ≤1 +∞ otherweise.
This implies
W2∗(B) =
0 if |B(x)| ≤1 a.e. in Ω +∞ otherweise.
Inserting the latter expression into the equality (7.1.1), we obtain
Z
Ω
p(x) :∇u(x) dx=
Z
Ω
|p(x)|dx
and |p(x)| ≤ 1 a.e. in Ω. Since p(x) :∇u(x) ≤ |p(x)| by theFenchel–Young (or Cauchy–Schwarz) inequality, we deduce
p(x) :∇u(x) =|p(x)| a.e. in Ω.
Therefore, p(x) ∈ BRm×d(0,1) if |∇u(x)| = 0 and p(x) = |∇u(x)|∇u(x) otherwise. We obtainp(x)∈Sgn(∇u(x)) a.e. inΩ. Now let un ⇀∗ u in L∞(0, T;U)∩H1(0, T;V), un → u in L2(0, T;V), and ξn ⇀∗ ξ in L2(0, T;U∗) as n → ∞ such that ξn(t) ∈
∂E(un(t)) for almost every t∈(0, T), supn∈N,t∈[0,T]E(un(t))≤C2, and lim sup
n→∞
Z T 0
⟨ξn(t),un(t)⟩U∗×Udt ≤
Z T 0
⟨ξ(t),u(t)⟩U∗×Udt. (7.1.2) We note that we can decomposeξ =ζ−u∈V∗ with ζ ∈∂W(u). Then, defining ζn:=ξn+un, there holdsζn∈∂W(un) and ζn⇀ζ :=ξ+uin L2(0, T;U∗). By the Lions–Aubin lemma, we obtain the strong convergence ofun →u in C([0, T];H).
Thus, in view of (7.1.2), we deduce lim sup
n→∞
Z T 0
⟨ζn(t),un(t)⟩U∗×Udt≤
Z T 0
⟨ζ(t),u(t)⟩U∗×Udt.
Since W is convex, by Theorem 2.3.7, there holds ζ(t) ∈ ∂UW(u(t)) in U∗ and W(un(t)) → W(u(t)) as n → ∞ a.e. in (0, T), whence ξ(t) ∈ ∂UE(u(t)) a.e. in (0, T) and E(un(t))→ E(u(t)) as n→ ∞ a.e. in (0, T). We proceed with showing
the control of the subgradient of E, i.e., Condition (5.Eg). Let u ∈ D(∂E) and ξ∈∂UE(u). Then, byHölder’s andYoung’s inequality, the Sobolevembedding theorem, we obtain uniformly bounded, from which (5.Eg) follows. Finally, we verify the assumptions on the perturbationB. The continuity condition (5.Ba) can easily be checked with the dominated convergence theorem.
Ad (5.Bb). Let u ∈ dom(E) and v, w ∈ V. Then, by the Hölder & Young inequalities as well as the Sobolev embedding theorem, there holds
⟨B(u,v),w⟩V∗×V = to (P1) in the sense that
Z T
with p(x, t)∈Sgn(∇u(x, t)) a.e. in ΩT, and the energy-dissipation inequality