Introduction
The general problem
A recurrent question in the theory of dissipative infinite-dimensional dynamical systems is how the long-time behaviour depends on a small parameter in the equation. When the parameter enters through a memory kernel that becomes concentrated in the limit, the perturbed and limiting systems live in different function spaces, and the natural object to compare is the family of global attractors. The present paper addresses this question for a class of memory kernels that concentrate at the origin while preserving their total mass. The mass-preservation is not a technical convenience: it is precisely the property that keeps the effective elastic coefficients bounded away from degeneration, and hence allows estimates that are uniform in the perturbation parameter.
The prototype situation is the following. Let g g g be an integrable memory kernel on R + \mathbb{R}_{+} R + with total mass β \beta β , and for ε > 0 \varepsilon> 0 ε > 0 set
g ε ( s ) : = 1 ε g ( s ε ) . (1) g_{\varepsilon}(s) := \frac{1}{\varepsilon}g\left(\frac{s}{\varepsilon}\right). \tag*{(1)} g ε ( s ) := ε 1 g ( ε s ) . (1)
Then ∫ 0 + ∞ g ε ( s ) d s = β \int_{0}^{+\infty} g_{\varepsilon}(s)\,\mathrm{d}s = \beta ∫ 0 + ∞ g ε ( s ) d s = β for every ε > 0 \varepsilon> 0 ε > 0 , while g ε ⇀ β δ 0 g_{\varepsilon} \rightharpoonup\beta\delta_{0} g ε ⇀ β δ 0 in the distributional sense as ε → 0 + \varepsilon\to0^{+} ε → 0 + . In a viscoelastic equation in which the memory term enters through a convolution against an elastic operator, the coefficient of that operator is effectively shifted by the mass of the kernel. A rescaling that changes the mass would therefore move the elastic coefficient in the limit, possibly through zero, and the limiting problem would be of a different nature. The mass-preserving rescaling (1) avoids this degeneracy and produces a limit in which the memory contribution becomes a local damping term with the same elastic coefficients. This is the structural mechanism that we isolate and exploit.
Our aim is twofold. First, we develop a continuity theory for global attractors under (1) that is not tied to a particular model: the ingredients are uniform dissipativity, uniform exponential stability of the linear semigroup, uniform quasi-stability in the sense of Chueshov–Lasiecka, finite-time convergence of the semigroups, and, for lower semicontinuity, a structural stability hypothesis on the limit attractor. Second, we verify these ingredients for a coupled nonlinear suspension bridge system with past history, which serves as a concrete and physically motivated model in which the general mechanism can be checked in detail. The bridge system is thus an application of the framework rather than its sole object.
The model
Suspension bridges are important civil engineering structures, and their dynamical behaviour has been widely studied. Early work by McKenna and co-authors [11 , 16 , 17 , 24 , 25 , 22 , 23 ] proposed and analysed coupled models for vertical bridge vibrations of the form
m 1 u t t + μ 1 u x x x x + k [ u − v ] + + γ 1 u t = f , x ∈ ( 0 , ℓ ) , t > 0 , m 2 v t t − μ 2 v x x − k [ u − v ] + + γ 2 v t = g , x ∈ ( 0 , ℓ ) , t > 0 , (2) \begin{aligned}
&m_{1}u_{tt}+\mu_{1}u_{xxxx}+k[u-v]^{+}+\gamma_{1}u_{t}=f,\quad x\in(0,\ell),\ t>0,\\
&m_{2}v_{tt}-\mu_{2}v_{xx}-k[u-v]^{+}+\gamma_{2}v_{t}=g,\quad x\in(0,\ell),\ t>0, \tag*{(2)}
\end{aligned} m 1 u tt + μ 1 u xxxx + k [ u − v ] + + γ 1 u t = f , x ∈ ( 0 , ℓ ) , t > 0 , m 2 v tt − μ 2 v xx − k [ u − v ] + + γ 2 v t = g , x ∈ ( 0 , ℓ ) , t > 0 , (2)
where u = u ( x , t ) u=u(x,t) u = u ( x , t ) denotes the vertical displacement of the bridge deck and v = v ( x , t ) v=v(x,t) v = v ( x , t ) the displacement of the main cable. The first equation describes deck oscillations in the vertical plane, the second the motion of the main cable supported by suspender cables. The positive constants m 1 , m 2 m_{1},m_{2} m 1 , m 2 are masses per unit length, k k k is the spring constant of the suspenders, μ 1 , μ 2 \mu_{1},\mu_{2} μ 1 , μ 2 are the flexural rigidity of the deck and the tensile stiffness of the cable, γ 1 , γ 2 \gamma_{1},\gamma_{2} γ 1 , γ 2 are damping coefficients, and f , g f,g f , g are external forces. The nonlinear term k [ u − v ] + k[u-v]^{+} k [ u − v ] + with [ u − v ] + = max { u − v , 0 } [u-v]^{+}=\max\{u-v,0\} [ u − v ] + = max { u − v , 0 } models the restoring force of tensioned suspenders; it vanishes when the deck lies above the cable, reflecting that suspenders can only exert tension. The boundary conditions are
u ( 0 , t ) = u ( ℓ , t ) = u x x ( 0 , t ) = u x x ( ℓ , t ) = 0 , v ( 0 , t ) = v ( ℓ , t ) = 0 , (3) u(0,t)=u(\ell,t)=u_{xx}(0,t)=u_{xx}(\ell,t)=0,\qquad v(0,t)=v(\ell,t)=0, \tag*{(3)} u ( 0 , t ) = u ( ℓ , t ) = u xx ( 0 , t ) = u xx ( ℓ , t ) = 0 , v ( 0 , t ) = v ( ℓ , t ) = 0 , (3)
for a simply supported deck beam and a pinned main cable. For simplicity we set m 1 = m 2 = 1 m_{1}=m_{2}=1 m 1 = m 2 = 1 and consider time-independent external forces f i = f i ( x ) f_i=f_i(x) f i = f i ( x ) , i = 1 , 2 i=1,2 i = 1 , 2 , so that the system is autonomous.
A substantial body of work has addressed well-posedness, asymptotic stability, and global attractors for (2) and its modifications; see, e.g., [1 , 2 , 8 , 13 , 19 , 21 ] and references therein. Ma and Zhong [20 ] considered the coupled system
u t t + μ 1 u x x x x + k [ u − v ] + + γ 1 u t + h 1 ( u ) = f 1 , x ∈ ( 0 , ℓ ) , t > 0 , v t t − μ 2 v x x − k [ u − v ] + + γ 2 v t + h 2 ( v ) = f 2 , x ∈ ( 0 , ℓ ) , t > 0 , (4) \begin{aligned}
&u_{tt}+\mu_{1}u_{xxxx}+k[u-v]^{+}+\gamma_{1}u_{t}+h_{1}(u)=f_{1},\quad x\in(0,\ell),\ t>0,\\
&v_{tt}-\mu_{2}v_{xx}-k[u-v]^{+}+\gamma_{2}v_{t}+h_{2}(v)=f_{2},\quad x\in(0,\ell),\ t>0, \tag*{(4)}
\end{aligned} u tt + μ 1 u xxxx + k [ u − v ] + + γ 1 u t + h 1 ( u ) = f 1 , x ∈ ( 0 , ℓ ) , t > 0 , v tt − μ 2 v xx − k [ u − v ] + + γ 2 v t + h 2 ( v ) = f 2 , x ∈ ( 0 , ℓ ) , t > 0 , (4)
and proved existence of strong solutions and a global attractor under the assumptions h 1 ∈ C 3 ( R , R ) h_{1}\in C^{3}(\mathbb{R},\mathbb{R}) h 1 ∈ C 3 ( R , R ) , h 2 ∈ C 2 ( R , R ) h_{2}\in C^{2}(\mathbb{R},\mathbb{R}) h 2 ∈ C 2 ( R , R ) , and f 1 , f 2 ∈ L 2 ( 0 , ℓ ) f_{1},f_{2}\in L^{2}(0,\ell) f 1 , f 2 ∈ L 2 ( 0 , ℓ ) . Related results for stationary and evolutionary problems may be found in [2 , 19 ] . Dell’Oro, Giorgi, and Pata [8 ] studied asymptotic behaviour for linear coupled bridge models and derived decay estimates under suitable damping and structural assumptions. These works show that the nonlinear coupling k [ u − v ] + k[u-v]^{+} k [ u − v ] + plays an important part in the long-time dynamics, and that attractor analysis requires careful treatment of its non-smooth nature.
If cable motion is neglected and the main cable is replaced by a rigid frame, deck vibration reduces to the scalar equation
u t t + u x x x x + k [ u ] + + γ u t = f . (5) u_{tt}+u_{xxxx}+k[u]^{+}+\gamma u_{t}=f. \tag*{(5)} u tt + u xxxx + k [ u ] + + γ u t = f . (5)
We refer to [1 , 4 , 14 , 18 , 29 , 32 ] and references therein for results concerning (5) . In particular, Park and Kang [29 ] constructed global attractors for bridge equations with nonlinear damping, while Zhong, Ma, and Sun [32 ] established well-posedness of strong solutions and existence of global attractors for the single beam model.
Because deck and cable materials may exhibit viscoelastic responses, there has been growing interest in suspension-bridge models including memory effects. Following Dafermos [7 ] , past-history contributions can be incorporated by introducing relative displacement fields
σ ( x , t , s ) = u ( x , t ) − u ( x , t − s ) , η ( x , t , s ) = v ( x , t ) − v ( x , t − s ) , (6) \sigma(x,t,s)=u(x,t)-u(x,t-s),\qquad\eta(x,t,s)=v(x,t)-v(x,t-s), \tag*{(6)} σ ( x , t , s ) = u ( x , t ) − u ( x , t − s ) , η ( x , t , s ) = v ( x , t ) − v ( x , t − s ) , (6)
which satisfy the transport equations
σ t + σ s − u t = 0 , η t + η s − v t = 0. (7) \sigma_{t}+\sigma_{s}-u_{t}=0,\qquad\eta_{t}+\eta_{s}-v_{t}=0. \tag*{(7)} σ t + σ s − u t = 0 , η t + η s − v t = 0. (7)
This formulation treats infinite past history as an additional state variable in an appropriate Hilbert space. Messaoudi, Bonfoh, Mukiawa, and Enyi [26 ] studied a memory-dependent suspension-bridge system with partially hinged boundaries and obtained a global attractor. For beam-type systems with memory, Mukiawa and Enyi [28 , 9 ] investigated well-posedness and long-time dynamics for Timoshenko and laminated beams subject to Fourier-type laws and infinite memory.
Mukiawa and Messaoudi [27 ] analysed the coupled nonlinear bridge system with past history
{ u t t + γ 1 u t + μ 1 u x x x x + k [ u − v ] − ∫ − ∞ t g 1 ( t − s ) u x x x x ( x , s ) d s + h 1 ( u ) = f 1 ( x ) , x ∈ ( 0 , ℓ ) , t > 0 , v t t + γ 2 v t − μ 2 v x x − k [ u − v ] + ∫ − ∞ t g 2 ( t − s ) v x x ( x , s ) d s + h 2 ( v ) = f 2 ( x ) , x ∈ ( 0 , ℓ ) , t > 0. (8) \begin{cases}
u_{tt}+\gamma_{1}u_t+\mu_{1}u_{xxxx}+k[u-v]-\int_{-\infty}^{t}g_{1}(t-s)u_{xxxx}(x,s)\,\mathrm{d}s+h_{1}(u)=f_{1}(x), & x\in(0,\ell),\ t>0,\\
v_{tt}+\gamma_{2}v_t-\mu_{2}v_{xx}-k[u-v]+\int_{-\infty}^{t}g_{2}(t-s)v_{xx}(x,s)\,\mathrm{d}s+h_{2}(v)=f_{2}(x), & x\in(0,\ell),\ t>0. \tag*{(8)}
\end{cases} { u tt + γ 1 u t + μ 1 u xxxx + k [ u − v ] − ∫ − ∞ t g 1 ( t − s ) u xxxx ( x , s ) d s + h 1 ( u ) = f 1 ( x ) , v tt + γ 2 v t − μ 2 v xx − k [ u − v ] + ∫ − ∞ t g 2 ( t − s ) v xx ( x , s ) d s + h 2 ( v ) = f 2 ( x ) , x ∈ ( 0 , ℓ ) , t > 0 , x ∈ ( 0 , ℓ ) , t > 0. (8)
together with boundary conditions (3) and suitable initial data, proving well-posedness and existence of a compact global attractor on the natural energy space, where the memory kernels g i g_i g i satisfy
∫ 0 + ∞ g i ( s ) d s = β i ∈ ( 0 , μ i ) , g i ′ ( s ) + λ i g i ( s ) ≤ 0 a.e. s ≥ 0 , (9) \int_{0}^{+\infty}g_i(s)\,\mathrm{d}s=\beta_i\in(0,\mu_i),\qquad g_i'(s)+\lambda_i g_i(s)\leq0\quad\text{a.e. }s\geq0, \tag*{(9)} ∫ 0 + ∞ g i ( s ) d s = β i ∈ ( 0 , μ i ) , g i ′ ( s ) + λ i g i ( s ) ≤ 0 a.e. s ≥ 0 , (9)
for some λ i > 0 \lambda_i>0 λ i > 0 , i = 1 , 2 i=1,2 i = 1 , 2 . The nonlinearities h i : R → R h_i:\mathbb{R}\to\mathbb{R} h i : R → R , i = 1 , 2 i=1,2 i = 1 , 2 , are of class C 1 C^1 C 1 , and their primitives
H i ( s ) : = ∫ 0 s h i ( τ ) d τ , i = 1 , 2 , (10) H_i(s):=\int_{0}^{s}h_i(\tau)\,\mathrm{d}\tau,\qquad i=1,2, \tag*{(10)} H i ( s ) := ∫ 0 s h i ( τ ) d τ , i = 1 , 2 , (10)
satisfy the following growth and dissipativity conditions: there exist constants C 1 , C 2 > 0 C_1,C_2>0 C 1 , C 2 > 0 such that for all s ∈ R s\in\mathbb{R} s ∈ R and i = 1 , 2 i=1,2 i = 1 , 2 ,
lim inf ∣ s ∣ → + ∞ H i ( s ) s 2 ≥ 0 , lim inf ∣ s ∣ → + ∞ s h i ( s ) − C 1 H i ( s ) s 2 ≥ 0 , h i ′ ( s ) ≥ − C 2 . (11) \liminf_{|s|\to+\infty}\frac{H_i(s)}{s^2}\geq0,\qquad\liminf_{|s|\to+\infty}\frac{s h_i(s)-C_1H_i(s)}{s^2}\geq0,\qquad h_i'(s)\geq-C_2. \tag*{(11)} ∣ s ∣ → + ∞ lim inf s 2 H i ( s ) ≥ 0 , ∣ s ∣ → + ∞ lim inf s 2 s h i ( s ) − C 1 H i ( s ) ≥ 0 , h i ′ ( s ) ≥ − C 2 . (11)
As observed in [27 , 33 ] , condition (11) implies that for every ϵ > 0 \epsilon>0 ϵ > 0 there exist constants C ϵ , C ϵ ′ > 0 C_\epsilon,C'_\epsilon>0 C ϵ , C ϵ ′ > 0 such that
H i ( s ) + ϵ s 2 ≥ − C ϵ , s h i ( s ) − C 1 H i ( s ) + ϵ s 2 ≥ − C ϵ ′ , ∀ s ∈ R , i = 1 , 2. (12) H_i(s)+\epsilon s^2\geq-C_\epsilon,\qquad s h_i(s)-C_1H_i(s)+\epsilon s^2\geq-C'_\epsilon,\qquad\forall s\in\mathbb{R},\ i=1,2. \tag*{(12)} H i ( s ) + ϵ s 2 ≥ − C ϵ , s h i ( s ) − C 1 H i ( s ) + ϵ s 2 ≥ − C ϵ ′ , ∀ s ∈ R , i = 1 , 2. (12)
A typical example is h i ( s ) = s ∣ s ∣ γ h_i(s)=s|s|^\gamma h i ( s ) = s ∣ s ∣ γ with 0 ≤ γ < + ∞ 0\leq\gamma<+\infty 0 ≤ γ < + ∞ , which satisfies (11) for suitable C 1 , C 2 C_1,C_2 C 1 , C 2 .
Relation to previous work and the role of the correction
The long-time analysis in [27 ] contains one step that, in our view, requires further justification, and we take this opportunity to explain how the present work relates to it. In [27 ] , Proposition 4.1, the authors derive the estimate
∥ U 1 ( t ) − U 2 ( t ) ∥ H 2 ≤ c ∥ U 0 1 − U 0 2 ∥ H 2 + c B ∫ 0 t ( ∥ u 1 ( s ) − u 2 ( s ) ∥ 2 + ∥ v 1 ( s ) − v 2 ( s ) ∥ 2 ) d s , (13) \left\|U^1(t)-U^2(t)\right\|_{\mathcal{H}}^2\leq c\left\|U_0^1-U_0^2\right\|_{\mathcal{H}}^2+c_{\mathcal{B}}\int_{0}^{t}\left(\left\|u^1(s)-u^2(s)\right\|^2+\left\|v^1(s)-v^2(s)\right\|^2\right)\,\mathrm{d}s, \tag*{(13)} U 1 ( t ) − U 2 ( t ) H 2 ≤ c U 0 1 − U 0 2 H 2 + c B ∫ 0 t ( u 1 ( s ) − u 2 ( s ) 2 + v 1 ( s ) − v 2 ( s ) 2 ) d s , (13)
where c B > 0 c_{\mathcal{B}}>0 c B > 0 depends on the absorbing set B \mathcal{B} B , and then assert that for any ϵ > 0 \epsilon>0 ϵ > 0 and sufficiently large T > 0 T>0 T > 0 ,
∥ S ( T ) U 0 1 − S ( T ) U 0 2 ∥ H 2 ≤ ϵ + Υ T ( U 0 1 , U 0 2 ) , (14) \left\|S(T)U_0^1-S(T)U_0^2\right\|_{\mathcal{H}}^2\leq\epsilon+\Upsilon_T(U_0^1,U_0^2), \tag*{(14)} S ( T ) U 0 1 − S ( T ) U 0 2 H 2 ≤ ϵ + Υ T ( U 0 1 , U 0 2 ) , (14)
with Υ T \Upsilon_T Υ T defined in [27 (4.26)] . This step is not immediately clear from (13) : the term c ∥ U 0 1 − U 0 2 ∥ H 2 c\left\|U_0^1-U_0^2\right\|_{\mathcal{H}}^2 c U 0 1 − U 0 2 H 2 is independent of T T T and hence cannot be made arbitrarily small by taking T T T large. For initial data inside a bounded absorbing set, initial differences may remain bounded away from zero, and no decaying factor is available in the above inequality to control this contribution. A rigorous treatment requires exponential decay of the underlying linear semigroup in the form needed here, which was not established in [27 ] . To our knowledge, this issue has not been addressed elsewhere in the literature.
Motivated by this observation, Zhou [33 ] revisited system (8) . Using the Gearhart–Prüss theorem and detailed resolvent estimates, exponential stability was shown for the associated linear semigroup. Combining this stability property with local Lipschitz behaviour of the nonlinear terms, the fact that nonlinearities act only on elastic components, smoothing effects of the elastic operators, and exponential decay of the memory kernels, quasi-stability in the Chueshov–Lasiecka sense follows, which yields asymptotic smoothness. Together with dissipativity results from [27 ] , this gives a compact global attractor on the natural energy space. In the present work we build the singular-limit analysis on this corrected global-attractor result for the fixed-kernel system, rather than directly on the original proof in [27 ] . We stress, however, that the correction is a matter of rigour in an ingredient, not the main contribution of this paper: the main contribution is the continuity theory under the mass-preserving singular limit, for which the fixed-kernel attractor is an input.
The singular perturbation and the continuity questions
We consider a physically motivated singular perturbation of (8) , corresponding to a vanishing viscoelastic relaxation time. This setting relates to the general theory of memory kernels converging in distribution towards a Dirac mass, developed by Conti, Pata, and Squassina [6 ] for evolution equations with memory. Physically, the kernel g i g_i g i encodes the relaxation modulus of the viscoelastic material: a short relaxation time means that the material forgets its past almost immediately, so that the memory term should reduce to a local damping term rather than disappear. We employ the mass-preserving rescaling
g i , ε ( s ) : = 1 ε g i ( s ε ) , ε > 0 , i = 1 , 2 , (15) g_{i,\varepsilon}(s) := \frac{1}{\varepsilon}g_i\left(\frac{s}{\varepsilon}\right), \qquad\varepsilon> 0,\; i=1,2, \tag*{(15)} g i , ε ( s ) := ε 1 g i ( ε s ) , ε > 0 , i = 1 , 2 , (15)
which satisfies three key properties: the total mass ∫ 0 + ∞ g i , ε ( s ) d s = β i \int_{0}^{+\infty} g_{i,\varepsilon}(s)\,\mathrm{d}s = \beta_i ∫ 0 + ∞ g i , ε ( s ) d s = β i remains unchanged for all ε > 0 \varepsilon> 0 ε > 0 ; g i , ε ⇀ β i δ 0 g_{i,\varepsilon} \rightharpoonup\beta_i\delta_0 g i , ε ⇀ β i δ 0 in the distributional sense as ε → 0 + \varepsilon\to0^+ ε → 0 + , so memory terms formally reduce to local damping; and the rescaling describes a short-memory physical regime with vanishing relaxation time. The choice of a mass-preserving rescaling, rather than an arbitrary concentration family, is essential: it is precisely this property that prevents the effective elastic coefficients from degenerating in the limit.
Formally, as ε → 0 + \varepsilon\to0^+ ε → 0 + , the memory-containing system (8) tends to the memoryless damped system
u t t + γ 1 u t + μ 1 u x x x x + k [ u − v ] + + h 1 ( u ) = f 1 ( x ) , x ∈ ( 0 , ℓ ) , t > 0 , v t t + γ 2 v t − μ 2 v x x − k [ u − v ] + + h 2 ( v ) = f 2 ( x ) , x ∈ ( 0 , ℓ ) , t > 0. (16) \begin{aligned}
u_{tt}+\gamma_1u_t+\mu_1u_{xxxx}+k[u-v]^+ + h_1(u) &= f_1(x), \qquad x\in(0,\ell),\ t>0,\\
v_{tt}+\gamma_2v_t-\mu_2v_{xx}-k[u-v]^+ + h_2(v) &= f_2(x), \qquad x\in(0,\ell),\ t>0. \tag*{(16)}
\end{aligned} u tt + γ 1 u t + μ 1 u xxxx + k [ u − v ] + + h 1 ( u ) v tt + γ 2 v t − μ 2 v xx − k [ u − v ] + + h 2 ( v ) = f 1 ( x ) , x ∈ ( 0 , ℓ ) , t > 0 , = f 2 ( x ) , x ∈ ( 0 , ℓ ) , t > 0. (16)
keeping the same elastic coefficients μ i \mu_i μ i . This is the useful feature of (15) : the quantity μ ˉ i , ε = μ i − ∫ 0 + ∞ g i , ε ( s ) d s = μ i − β i \bar{\mu}_{i,\varepsilon}=\mu_i-\int_{0}^{+\infty}g_{i,\varepsilon}(s)\,\mathrm{d}s=\mu_i-\beta_i μ ˉ i , ε = μ i − ∫ 0 + ∞ g i , ε ( s ) d s = μ i − β i is independent of ε \varepsilon ε , so effective elastic coefficients do not degenerate in the singular limit. The limiting system coincides with the classical memoryless coupled suspension-bridge model studied by Ma and Zhong [20 ] . Thus the singular limit provides a rigorous bridge between the memory-dependent model (8) and the classical model (16) , which is the standard justification for neglecting viscoelastic memory on time scales much longer than the relaxation time.
The present paper is concerned with continuity properties of the global attractor under the singular perturbation (15) . We establish upper semicontinuity
lim ε → 0 + dist H ε ( A ε , A 0 ) = 0 , (17) \lim_{\varepsilon\to0^+}\operatorname{dist}_{\mathcal{H}_\varepsilon}(\mathcal{A}_\varepsilon,\mathcal{A}_0)=0, \tag*{(17)} ε → 0 + lim dist H ε ( A ε , A 0 ) = 0 , (17)
and, under a structural stability hypothesis on A 0 A_0 A 0 , lower semicontinuity
lim ε → 0 + dist H ε ( A 0 , A ε ) = 0 , (18) \lim_{\varepsilon\to0^+} \operatorname{dist}_{\mathcal{H}_\varepsilon}(\mathcal{A}_0,\mathcal{A}_\varepsilon)=0, \tag*{(18)} ε → 0 + lim dist H ε ( A 0 , A ε ) = 0 , (18)
where A ε \mathcal{A}_\varepsilon A ε is the global attractor of the perturbed semigroup S ε ( t ) S_\varepsilon(t) S ε ( t ) and A 0 \mathcal{A}_0 A 0 the attractor for the limiting semigroup S 0 ( t ) S_0(t) S 0 ( t ) . Upper semicontinuity says that the perturbed attractors do not exhibit new asymptotic regimes far from the limit attractor; lower semicontinuity says, conversely, that every asymptotic regime of the limit system survives under small perturbations of the memory kernel. Together, they justify using the simpler memoryless model (16) as a faithful approximation of (8) for small relaxation times.
Contributions and organisation
The contributions of this paper are the following.
(i) A general continuity mechanism. We identify the mass-preserving rescaling as the structural condition under which the effective elastic coefficients remain independent of the perturbation parameter, and we formulate the continuity problem in a way that separates the general mechanism (uniform dissipativity, uniform exponential stability, uniform quasi-stability, finite-time convergence, and structural stability of the limit attractor) from the model-specific verification. The mechanism is stated in terms of the abstract semigroups S ε ( t ) S_\varepsilon(t) S ε ( t ) and S 0 ( t ) S_0(t) S 0 ( t ) and is not tied to the bridge model.
(ii) Uniform exponential stability. We extend the exponential stability result for the linear semigroup obtained in [33 ] to a version uniform with respect to ε \varepsilon ε , so that decay rates and prefactors do not deteriorate as ε → 0 + \varepsilon\to0^+ ε → 0 + . This is Theorem 2.2 , and it is the analytic core of the uniform estimates.
(iii) Uniform quasi-stability and exponential attraction. Combining (ii) with the sequential compactness of the integral remainder K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_0) K ε , T ( B 0 ) , we obtain uniform quasi-stability (Proposition 4.1 ) and uniform exponential attraction (Proposition 5.2 ) on a common absorbing ball.
(iv) Finite-time convergence and Hölder continuity. We prove finite-time convergence of the perturbed semigroups to the limit semigroup (Proposition 4.2 ) and uniform Hölder continuity with respect to ε \varepsilon ε on compact subintervals of ( 0 , 1 ] (0,1] ( 0 , 1 ] (Proposition 5.4 ). The former yields upper semicontinuity; the latter is recorded for completeness and for possible quantitative refinements.
(v) Upper and lower semicontinuity. We prove upper semicontinuity of the global attractors (Theorem 4.3 ), and, under a structural stability hypothesis on A 0 A_0 A 0 in the sense of Hale [12 ] , lower semicontinuity (Proposition 5.6 ). A non-vacuous example satisfying the structural hypothesis is given (Example 5.8).
The paper is organised as follows. § 2 introduces the functional setting, formulates the perturbed problem, and establishes the uniform-in-ε \varepsilon ε exponential stability of the linear semigroup. § 3 proves uniform dissipativity and the existence of a common absorbing set. § 4 proves upper semicontinuity for the family of attractors. § 5 derives lower semicontinuity under an additional structural hypothesis. § 6 summarises the findings and points out several open questions.
Functional Setting and the Perturbed Problem
Following Dafermos [7 ] and [27 , 33 ] , we introduce the relative displacements
σ ε ( x , t , s ) = u ε ( x , t ) − u ε ( x , t − s ) , η ε ( x , t , s ) = v ε ( x , t ) − v ε ( x , t − s ) , \sigma_{\varepsilon}(x,t,s)=u_{\varepsilon}(x,t)-u_{\varepsilon}(x,t-s),\qquad\eta_{\varepsilon}(x,t,s)=v_{\varepsilon}(x,t)-v_{\varepsilon}(x,t-s), σ ε ( x , t , s ) = u ε ( x , t ) − u ε ( x , t − s ) , η ε ( x , t , s ) = v ε ( x , t ) − v ε ( x , t − s ) ,
which satisfy the transport equations
∂ t σ ε + ∂ s σ ε − u ε , t = 0 , ∂ t η ε + ∂ s η ε − v ε , t = 0. (19) \partial_t\sigma_{\varepsilon}+\partial_s\sigma_{\varepsilon}-u_{\varepsilon,t}=0,\qquad\partial_t\eta_{\varepsilon}+\partial_s\eta_{\varepsilon}-v_{\varepsilon,t}=0. \tag*{(19)} ∂ t σ ε + ∂ s σ ε − u ε , t = 0 , ∂ t η ε + ∂ s η ε − v ε , t = 0. (19)
Let
H a 2 ( 0 , ℓ ) = { w ∈ H 2 ( 0 , ℓ ) : w ( 0 ) = w ( ℓ ) = 0 } , H_a^2(0,\ell)=\{w\in H^2(0,\ell):w(0)=w(\ell)=0\}, H a 2 ( 0 , ℓ ) = { w ∈ H 2 ( 0 , ℓ ) : w ( 0 ) = w ( ℓ ) = 0 } ,
and define the weighted spaces
M g 1 , ε = L g 1 , ε 2 ( R + ; H a 2 ( 0 , ℓ ) ) , M g 2 , ε = L g 2 , ε 2 ( R + ; H 0 1 ( 0 , ℓ ) ) , \mathcal{M}_{g_1,\varepsilon}=L^2_{g_1,\varepsilon}(\mathbb{R}_+;H_a^2(0,\ell)),\qquad\mathcal{M}_{g_2,\varepsilon}=L^2_{g_2,\varepsilon}(\mathbb{R}_+;H_0^1(0,\ell)), M g 1 , ε = L g 1 , ε 2 ( R + ; H a 2 ( 0 , ℓ )) , M g 2 , ε = L g 2 , ε 2 ( R + ; H 0 1 ( 0 , ℓ )) ,
with inner products
( σ , σ ˉ ) M g 1 , ε = ∫ 0 + ∞ g 1 , ε ( s ) ( σ x x ( s ) , σ ˉ x x ( s ) ) d s , ( η , η ˉ ) M g 2 , ε = ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , η ˉ x ( s ) ) d s . (\sigma,\bar{\sigma})_{\mathcal{M}_{g_1,\varepsilon}}=\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{xx}(s),\bar{\sigma}_{xx}(s))\,\mathrm{d}s,\qquad(\eta,\bar{\eta})_{\mathcal{M}_{g_2,\varepsilon}}=\int_0^{+\infty}g_{2,\varepsilon}(s)(\eta_x(s),\bar{\eta}_x(s))\,\mathrm{d}s. ( σ , σ ˉ ) M g 1 , ε = ∫ 0 + ∞ g 1 , ε ( s ) ( σ xx ( s ) , σ ˉ xx ( s )) d s , ( η , η ˉ ) M g 2 , ε = ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , η ˉ x ( s )) d s .
Set
μ ˉ i , ε = μ i − ∫ 0 + ∞ g i , ε ( s ) d s = μ i − β i > 0 , i = 1 , 2 , (20) \bar{\mu}_{i,\varepsilon}=\mu_i-\int_0^{+\infty}g_{i,\varepsilon}(s)\,\mathrm{d}s=\mu_i-\beta_i>0,\qquad i=1,2, \tag*{(20)} μ ˉ i , ε = μ i − ∫ 0 + ∞ g i , ε ( s ) d s = μ i − β i > 0 , i = 1 , 2 , (20)
where β i = ∫ 0 + ∞ g i ( s ) d s \beta_i=\int_0^{+\infty}g_i(s)\,\mathrm{d}s β i = ∫ 0 + ∞ g i ( s ) d s is the mass appearing in (9) . For each ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] we introduce the energy space
H ε = H a 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × M g 1 , ε × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × M g 2 , ε , (21) \mathcal{H}_{\varepsilon}=H_a^2(0,\ell)\times L^2(0,\ell)\times\mathcal{M}_{g_1,\varepsilon}\times H_0^1(0,\ell)\times L^2(0,\ell)\times\mathcal{M}_{g_2,\varepsilon}, \tag*{(21)} H ε = H a 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × M g 1 , ε × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × M g 2 , ε , (21)
endowed with the norm
∥ ( u , ψ , σ , v , φ , η ) ∥ H ε 2 = μ ˉ 1 , ε ∥ u x x ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + μ ˉ 2 , ε ∥ v x ∥ 2 + ∥ φ ∥ 2 + ∥ η ∥ M g 2 , ε 2 , (22) \|(u,\psi,\sigma,v,\varphi,\eta)\|_{\mathcal{H}_{\varepsilon}}^2=\bar{\mu}_{1,\varepsilon}\|u_{xx}\|^2+\|\psi\|^2+\|\sigma\|_{\mathcal{M}_{g_1,\varepsilon}}^2+\bar{\mu}_{2,\varepsilon}\|v_x\|^2+\|\varphi\|^2+\|\eta\|_{\mathcal{M}_{g_2,\varepsilon}}^2, \tag*{(22)} ∥ ( u , ψ , σ , v , φ , η ) ∥ H ε 2 = μ ˉ 1 , ε ∥ u xx ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + μ ˉ 2 , ε ∥ v x ∥ 2 + ∥ φ ∥ 2 + ∥ η ∥ M g 2 , ε 2 , (22)
with μ ˉ i , ε \bar{\mu}_{i,\varepsilon} μ ˉ i , ε as in (20) . This is a crucial observation: the rescaling (15) preserves the mass, hence the effective elastic coefficients μ ˉ i , ε \bar{\mu}_{i,\varepsilon} μ ˉ i , ε do not degenerate as ε → 0 + \varepsilon\to0^+ ε → 0 + .
Remark 2.1 (Uniform topology). The spaces H ε \mathcal{H}_{\varepsilon} H ε depend on ε \varepsilon ε through the weighted memory spaces M g i , ε \mathcal{M}_{g_i,\varepsilon} M g i , ε . However, all estimates below involve the memory spaces only through the mass β i \beta_i β i , which is independent of ε \varepsilon ε by (20) . More precisely, for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] the norm (22) coincides with the quantity
μ ˉ 1 ∥ u x x ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + μ ˉ 2 ∥ v x ∥ 2 + ∥ φ ∥ 2 + ∥ η ∥ M g 2 , ε 2 , \bar{\mu}_1\|u_{xx}\|^2+\|\psi\|^2+\|\sigma\|_{\mathcal{M}_{g_1,\varepsilon}}^2+\bar{\mu}_2\|v_x\|^2+\|\varphi\|^2+\|\eta\|_{\mathcal{M}_{g_2,\varepsilon}}^2, μ ˉ 1 ∥ u xx ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + μ ˉ 2 ∥ v x ∥ 2 + ∥ φ ∥ 2 + ∥ η ∥ M g 2 , ε 2 ,
whose only ε \varepsilon ε -dependent ingredient is the weight g i , ε g_{i,\varepsilon} g i , ε of total mass β i \beta_i β i . Consequently, all constants appearing in the estimates below are independent of ε \varepsilon ε , and the convergence statements are understood in the following sense: for a sequence ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + and U ε n ∈ H ε n U_{\varepsilon_n}\in\mathcal{H}_{\varepsilon_n} U ε n ∈ H ε n , we write U ε n → U U_{\varepsilon_n}\to U U ε n → U if the elastic components converge in H a 2 × L 2 × H 0 1 × L 2 H_a^2\times L^2\times H_0^1\times L^2 H a 2 × L 2 × H 0 1 × L 2 and the memory components converge in the sense of the weighted norms with weights g i , ε n g_{i,\varepsilon_n} g i , ε n . On compact subintervals [ ε 0 , 1 ] ⊂ ( 0 , 1 ] [\varepsilon_0,1]\subset(0,1] [ ε 0 , 1 ] ⊂ ( 0 , 1 ] , the norms ∥ ⋅ ∥ H ε \|\cdot\|_{\mathcal{H}_{\varepsilon}} ∥ ⋅ ∥ H ε and ∥ ⋅ ∥ H ε ′ \|\cdot\|_{\mathcal{H}_{\varepsilon'}} ∥ ⋅ ∥ H ε ′ are uniformly equivalent, with a constant depending only on ε 0 \varepsilon_0 ε 0 ; this follows from the lower bound on the kernels imposed in Assumption 5.3 (Section 5).
Throughout, we write μ ˉ i : = μ ˉ i , ε = μ i − β i \bar{\mu}_i:=\bar{\mu}_{i,\varepsilon}=\mu_i-\beta_i μ ˉ i := μ ˉ i , ε = μ i − β i (independent of ε \varepsilon ε ) and M g i , ε : = M g i , ε \mathcal{M}_{g_i,\varepsilon}:=\mathcal{M}_{g_i,\varepsilon} M g i , ε := M g i , ε when no confusion can arise. All constants appearing below are independent of ε \varepsilon ε unless explicitly stated. Setting U ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) U_{\varepsilon}=(u_{\varepsilon},u_{\varepsilon,t},\sigma_{\varepsilon},v_{\varepsilon},v_{\varepsilon,t},\eta_{\varepsilon}) U ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) , the perturbed system can be written abstractly as
∂ t U ε + A ε U ε = F ( U ε ) , U ε ( 0 ) = U 0 , ε . (23) \partial_tU_{\varepsilon}+A_{\varepsilon}U_{\varepsilon}=F(U_{\varepsilon}),\qquad U_{\varepsilon}(0)=U_{0,\varepsilon}. \tag*{(23)} ∂ t U ε + A ε U ε = F ( U ε ) , U ε ( 0 ) = U 0 , ε . (23)
where the linear operator A ε A_{\varepsilon} A ε is
A ε U = ( − ψ γ 1 ψ + μ ˉ 1 , ε u x x x x + ∫ 0 + ∞ g 1 , ε ( s ) σ x x x x ( s ) d s σ s − ψ − ϕ γ 2 ϕ − μ ˉ 2 , ε v x x − ∫ 0 + ∞ g 2 , ε ( s ) η x x ( s ) d s η s − ϕ ) , (24) A_{\varepsilon}U =
\begin{pmatrix}
-\psi\\
\gamma_{1}\psi+\bar{\mu}_{1,\varepsilon}u_{xxxx}+\int_{0}^{+\infty}g_{1,\varepsilon}(s)\sigma_{xxxx}(s)\,\mathrm{d}s \\
\sigma_{s}-\psi\\
-\phi\\
\gamma_{2}\phi-\bar{\mu}_{2,\varepsilon}v_{xx}-\int_{0}^{+\infty}g_{2,\varepsilon}(s)\eta_{xx}(s)\,\mathrm{d}s \\
\eta_{s}-\phi
\end{pmatrix}, \tag*{(24)} A ε U = − ψ γ 1 ψ + μ ˉ 1 , ε u xxxx + ∫ 0 + ∞ g 1 , ε ( s ) σ xxxx ( s ) d s σ s − ψ − ϕ γ 2 ϕ − μ ˉ 2 , ε v xx − ∫ 0 + ∞ g 2 , ε ( s ) η xx ( s ) d s η s − ϕ , (24)
and the nonlinearity is
F ( U ) = ( 0 − k [ u − v ] + − h 1 ( u ) + f 1 0 0 k [ u − v ] + − h 2 ( v ) + f 2 0 ) . (25) F(U)=
\begin{pmatrix}
0 \\
-k[u-v]^{+}-h_{1}(u)+f_{1} \\
0 \\
0 \\
k[u-v]^{+}-h_{2}(v)+f_{2} \\
0
\end{pmatrix}. \tag*{(25)} F ( U ) = 0 − k [ u − v ] + − h 1 ( u ) + f 1 0 0 k [ u − v ] + − h 2 ( v ) + f 2 0 . (25)
The domain D ( A ε ) \mathcal{D}(A_{\varepsilon}) D ( A ε ) is defined as in [33 Section 2] , with g i g_i g i replaced by g i , ε g_{i,\varepsilon} g i , ε .
Under assumption [9 ] , which is inherited by the rescaled kernels g i , ε g_{i,\varepsilon} g i , ε with the same constants λ i \lambda_i λ i and the same mass β i = ∫ 0 + ∞ g i , ε ( s ) d s \beta_i=\int_{0}^{+\infty}g_{i,\varepsilon}(s)\,\mathrm{d}s β i = ∫ 0 + ∞ g i , ε ( s ) d s , the operator A ε A_{\varepsilon} A ε is maximal monotone, and − A ε -A_{\varepsilon} − A ε generates a C 0 C_0 C 0 -semigroup of contractions e − A ε t e^{-A_{\varepsilon}t} e − A ε t on H ε \mathcal{H}_{\varepsilon} H ε .
For a fixed kernel g i g_i g i , the exponential stability of this linear semigroup was established in [33 Theorem 3.1] by means of the Gearhart–Prüss theorem together with a detailed resolvent estimate. In the present work we need a uniform version of that result, in which the decay rate and the prefactor do not degenerate as ε → 0 + \varepsilon\to0^{+} ε → 0 + . The reason is twofold: first, the quasi-stability inequality of Section 4 requires a decay rate that is not allowed to deteriorate as ε → 0 + \varepsilon\to0^{+} ε → 0 + ; second, the compactness argument for the upper semicontinuity of attractors uses the resulting estimates on a sequence ε n → 0 + \varepsilon_n\to0^{+} ε n → 0 + , and any loss of uniformity would destroy the convergence.
Theorem 2.2 (Uniform exponential stability). There exist constants α ≥ 1 \alpha\ge1 α ≥ 1 and β > 0 \beta>0 β > 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that
∥ e − A ε t ∥ L ( H ε ) ≤ α e − β t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (26) \lVert e^{-A_{\varepsilon}t}\rVert_{\mathcal{L}(\mathcal{H}_{\varepsilon})}\le\alpha e^{-\beta t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1]. \tag*{(26)} ∥ e − A ε t ∥ L ( H ε ) ≤ α e − β t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (26)
This uniform exponential stability is the key ingredient for both the upper and the lower semicontinuity analysis below.
Proof of Theorem 2.2. By the same Lax–Milgram argument as in [33 Section 2] (see also [27 Lemma 3.3] , whose proof applies verbatim with g i g_i g i replaced by g i , ε g_{i,\varepsilon} g i , ε ), for each fixed ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] the operator A ε A_{\varepsilon} A ε is maximal monotone on H ε \mathcal{H}_{\varepsilon} H ε . Moreover, the Lax–Milgram coercivity and boundedness constants in that argument are uniform in ε \varepsilon ε , because the quantity ∫ 0 + ∞ g i , ε ( s ) ( 1 − e − s ) d s \int_{0}^{+\infty}g_{i,\varepsilon}(s)(1-e^{-s})\,\mathrm{d}s ∫ 0 + ∞ g i , ε ( s ) ( 1 − e − s ) d s ranges in [ 0 , β i ] [0,\beta_i] [ 0 , β i ] and hence stays bounded uniformly in ε \varepsilon ε . Consequently, by the Lümer–Phillips theorem, − A ε -A_{\varepsilon} − A ε generates a C 0 C_0 C 0 -semigroup of contractions on H ε \mathcal{H}_{\varepsilon} H ε . By the Gearhart–Prüss theorem [33 Theorem 3.2] , it suffices to verify
(C1) i R ⊂ ρ ( A ε ) i\mathbb{R}\subset\rho(A_{\varepsilon}) i R ⊂ ρ ( A ε ) ;
(C2) sup λ ∈ R ∥ ( i λ I − A ε ) − 1 ∥ L ( H ε ) < + ∞ \sup_{\lambda\in\mathbb{R}}\lVert(i\lambda I-A_{\varepsilon})^{-1}\rVert_{\mathcal{L}(\mathcal{H}_{\varepsilon})}<+\infty sup λ ∈ R ∥( iλ I − A ε ) − 1 ∥ L ( H ε ) < + ∞ ,
with constants independent of ε \varepsilon ε . All function spaces are complexified Hilbert spaces; since all data are real-valued, we do not distinguish between the real and complexified pairings, and we write ( f , g ) = ∫ f g (f,g)=\int fg ( f , g ) = ∫ f g for real-valued f , g f,g f , g .
Throughout, we write U = ( u , ψ , σ , v , ϕ , η ) ∈ D ( A ε ) U=(u,\psi,\sigma,v,\phi,\eta)\in D(\mathcal{A}_{\varepsilon}) U = ( u , ψ , σ , v , ϕ , η ) ∈ D ( A ε ) and use the norm (22) . From now on we write μ ˉ i \bar{\mu}_i μ ˉ i for μ ˉ i , ε \bar{\mu}_{i,\varepsilon} μ ˉ i , ε whenever no confusion can arise, and similarly we abbreviate M g i , ε = M g i , ε \mathcal{M}_{g_{i,\varepsilon}}=\mathcal{M}_{g_i,\varepsilon} M g i , ε = M g i , ε . Here and below we use the invariance of the mass,
∫ 0 + ∞ g i , ε ( s ) d s = β i , μ ˉ i , ε = μ i − β i = μ ˉ i , i = 1 , 2 , (27) \int_{0}^{+\infty} g_{i,\varepsilon}(s)\,\mathrm{d}s=\beta_i,\qquad\bar{\mu}_{i,\varepsilon}=\mu_i-\beta_i=\bar{\mu}_i,\qquad i=1,2, \tag*{(27)} ∫ 0 + ∞ g i , ε ( s ) d s = β i , μ ˉ i , ε = μ i − β i = μ ˉ i , i = 1 , 2 , (27)
which holds for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . In particular, the norm ∥ ⋅ ∥ H ε \|\cdot\|_{\mathcal{H}_{\varepsilon}} ∥ ⋅ ∥ H ε is given by the same formula for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] .
Monotonicity. Two integrations by parts using the boundary conditions show that the elastic terms μ ˉ 1 ( − ψ x x , u x x ) + μ ˉ 1 ( u x x x x , ψ ) \bar{\mu}_1(-\psi_{xx},u_{xx})+\bar{\mu}_1(u_{xxxx},\psi) μ ˉ 1 ( − ψ xx , u xx ) + μ ˉ 1 ( u xxxx , ψ ) combine into a purely imaginary expression, and likewise for the v v v -elastic terms; the same cancellation occurs for the off-diagonal memory terms ∫ 0 + ∞ g 1 , ε ( s ) ( σ x x ( s ) , ψ x x ) d s \int_{0}^{+\infty}g_{1,\varepsilon}(s)(\sigma_{xx}(s),\psi_{xx})\,\mathrm{d}s ∫ 0 + ∞ g 1 , ε ( s ) ( σ xx ( s ) , ψ xx ) d s and − ( ψ , σ ) M g 1 , ε -(\psi,\sigma)_{\mathcal{M}_{g_{1,\varepsilon}}} − ( ψ , σ ) M g 1 , ε , which are complex conjugates of each other. Hence
Re ⟨ A ε U , U ⟩ H ε = γ 1 ∥ ψ ∥ 2 + γ 2 ∥ ϕ ∥ 2 + Re ( σ s , σ ) M g 1 , ε + Re ( η s , η ) M g 2 , ε . \operatorname{Re}\langle\mathcal{A}_{\varepsilon}U,U\rangle_{\mathcal{H}_{\varepsilon}}=\gamma_1\|\psi\|^2+\gamma_2\|\phi\|^2+\operatorname{Re}(\sigma_s,\sigma)_{\mathcal{M}_{g_{1,\varepsilon}}}+\operatorname{Re}(\eta_s,\eta)_{\mathcal{M}_{g_{2,\varepsilon}}}. Re ⟨ A ε U , U ⟩ H ε = γ 1 ∥ ψ ∥ 2 + γ 2 ∥ ϕ ∥ 2 + Re ( σ s , σ ) M g 1 , ε + Re ( η s , η ) M g 2 , ε .
Using σ ( x , t , 0 ) = 0 \sigma(x,t,0)=0 σ ( x , t , 0 ) = 0 and integrating by parts in s s s , we obtain
Re ( σ s , σ ) M g 1 , ε = − 1 2 ∫ 0 + ∞ g 1 , ε ′ ( s ) ∥ σ x x ( s ) ∥ 2 d s , \operatorname{Re}(\sigma_s,\sigma)_{\mathcal{M}_{g_{1,\varepsilon}}}=-\frac{1}{2}\int_{0}^{+\infty}g'_{1,\varepsilon}(s)\|\sigma_{xx}(s)\|^2\,\mathrm{d}s, Re ( σ s , σ ) M g 1 , ε = − 2 1 ∫ 0 + ∞ g 1 , ε ′ ( s ) ∥ σ xx ( s ) ∥ 2 d s ,
where the boundary term at s = + ∞ s=+\infty s = + ∞ vanishes because g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) ≤ ε − 1 g 1 ( 0 ) e − λ 1 s / ε → 0 g_{1,\varepsilon}(s)=\varepsilon^{-1}g_1(s/\varepsilon)\leq\varepsilon^{-1}g_1(0)e^{-\lambda_1s/\varepsilon}\to0 g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) ≤ ε − 1 g 1 ( 0 ) e − λ 1 s / ε → 0 as s → + ∞ s\to+\infty s → + ∞ . We verify that the rescaled kernel inherits the same exponential decay rate: differentiating g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) g_{1,\varepsilon}(s)=\varepsilon^{-1}g_1(s/\varepsilon) g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) gives
g 1 , ε ′ ( s ) = ε − 2 g 1 ′ ( s / ε ) ≤ − λ 1 ε − 2 g 1 ( s / ε ) = − λ 1 g 1 , ε ( s ) , g'_{1,\varepsilon}(s)=\varepsilon^{-2}g'_1(s/\varepsilon)\leq-\lambda_1\varepsilon^{-2}g_1(s/\varepsilon)=-\lambda_1g_{1,\varepsilon}(s), g 1 , ε ′ ( s ) = ε − 2 g 1 ′ ( s / ε ) ≤ − λ 1 ε − 2 g 1 ( s / ε ) = − λ 1 g 1 , ε ( s ) ,
so that g 1 , ε ′ + λ 1 g 1 , ε ≤ 0 g'_{1,\varepsilon}+\lambda_1g_{1,\varepsilon}\leq0 g 1 , ε ′ + λ 1 g 1 , ε ≤ 0 for a.e. s ≥ 0 s\geq0 s ≥ 0 with the same λ 1 \lambda_1 λ 1 as for g 1 g_1 g 1 . We thus get Re ( σ s , σ ) M g 1 , ε ≥ λ 1 2 ∥ σ ∥ M g 1 , ε 2 \operatorname{Re}(\sigma_s,\sigma)_{\mathcal{M}_{g_{1,\varepsilon}}}\geq\frac{\lambda_1}{2}\|\sigma\|^2_{\mathcal{M}_{g_{1,\varepsilon}}} Re ( σ s , σ ) M g 1 , ε ≥ 2 λ 1 ∥ σ ∥ M g 1 , ε 2 , and analogously for η \eta η . Thus
Re ⟨ A ε U , U ⟩ H ε ≥ γ 1 ∥ ψ ∥ 2 + λ 1 2 ∥ σ ∥ M g 1 , ε 2 + γ 2 ∥ ϕ ∥ 2 + λ 2 2 ∥ η ∥ M g 2 , ε 2 . (28) \operatorname{Re}\langle\mathcal{A}_{\varepsilon}U,U\rangle_{\mathcal{H}_{\varepsilon}}\geq\gamma_1\|\psi\|^2+\frac{\lambda_1}{2}\|\sigma\|^2_{\mathcal{M}_{g_{1,\varepsilon}}}+\gamma_2\|\phi\|^2+\frac{\lambda_2}{2}\|\eta\|^2_{\mathcal{M}_{g_{2,\varepsilon}}}. \tag*{(28)} Re ⟨ A ε U , U ⟩ H ε ≥ γ 1 ∥ ψ ∥ 2 + 2 λ 1 ∥ σ ∥ M g 1 , ε 2 + γ 2 ∥ ϕ ∥ 2 + 2 λ 2 ∥ η ∥ M g 2 , ε 2 . (28)
Both γ i \gamma_i γ i and λ i \lambda_i λ i are independent of ε \varepsilon ε ; this is the first place where the uniformity in ε \varepsilon ε is manifest. Condition (C1). Suppose, for contradiction, that condition (C1) fails uniformly in ε \varepsilon ε . Then there exist a sequence ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + , frequencies ω n ∈ R \omega_n\in\mathbb{R} ω n ∈ R , and U n = ( u n , ψ n , σ n , v n , ϕ n , η n ) ∈ D ( A ε n ) U_n=(u_n,\psi_n,\sigma_n,v_n,\phi_n,\eta_n)\in D(\mathcal{A}_{\varepsilon_n}) U n = ( u n , ψ n , σ n , v n , ϕ n , η n ) ∈ D ( A ε n ) with ∥ U n ∥ H ε n = 1 \|U_n\|_{\mathcal{H}_{\varepsilon_n}}=1 ∥ U n ∥ H ε n = 1 such that
F n : = i ω n U n − A ε n U n → 0 in H ε n . (29) F_n:=i\omega_nU_n-\mathcal{A}_{\varepsilon_n}U_n\to0\quad\text{in }\mathcal{H}_{\varepsilon_n}. \tag*{(29)} F n := i ω n U n − A ε n U n → 0 in H ε n . (29)
(For each fixed ε \varepsilon ε , condition (C1) is already known to hold; what is assumed here is not a failure of (C1) at some fixed ε \varepsilon ε , but rather that the constant in (C1) becomes unbounded as ε → 0 + \varepsilon\to0^+ ε → 0 + . The argument below shows that such a failure of uniformity cannot occur.) Write F n = ( f n 1 , … , f n 6 ) F_n=(f_n^1,\ldots,f_n^6) F n = ( f n 1 , … , f n 6 ) . Taking the real part of the inner product of (29) with U n U_n U n and using (28) , we obtain
γ 1 ∥ ψ n ∥ 2 + λ 1 2 ∥ σ n ∥ M g 1 , ε n 2 + γ 2 ∥ ϕ n ∥ 2 + λ 2 2 ∥ η n ∥ M g 2 , ε n 2 ≤ ∥ F n ∥ H ε n → 0 , (30) \gamma_1\|\psi_n\|^2+\frac{\lambda_1}{2}\|\sigma_n\|^2_{\mathcal{M}_{g_{1,\varepsilon_n}}}+\gamma_2\|\phi_n\|^2+\frac{\lambda_2}{2}\|\eta_n\|^2_{\mathcal{M}_{g_{2,\varepsilon_n}}}\leq\|F_n\|_{\mathcal{H}_{\varepsilon_n}}\to0, \tag*{(30)} γ 1 ∥ ψ n ∥ 2 + 2 λ 1 ∥ σ n ∥ M g 1 , ε n 2 + γ 2 ∥ ϕ n ∥ 2 + 2 λ 2 ∥ η n ∥ M g 2 , ε n 2 ≤ ∥ F n ∥ H ε n → 0 , (30)
where we used Re ⟨ F n , U n ⟩ H ε n = − Re ⟨ A ε n U n , U n ⟩ H ε n \operatorname{Re}\langle F_n,U_n\rangle_{\mathcal{H}_{\varepsilon_n}}=-\operatorname{Re}\langle\mathcal{A}_{\varepsilon_n}U_n,U_n\rangle_{\mathcal{H}_{\varepsilon_n}} Re ⟨ F n , U n ⟩ H ε n = − Re ⟨ A ε n U n , U n ⟩ H ε n , since the term i ω n ∥ U n ∥ H ε n 2 i\omega_n\|U_n\|^2_{\mathcal{H}_{\varepsilon_n}} i ω n ∥ U n ∥ H ε n 2 is purely imaginary. The rate of convergence in (30) is uniform in ε \varepsilon ε , because γ i , λ i \gamma_i,\lambda_i γ i , λ i do not depend on ε \varepsilon ε . We also record the elementary bounds, valid for every test function φ ∈ H 0 2 ( 0 , ℓ ) \varphi\in H^2_0(0,\ell) φ ∈ H 0 2 ( 0 , ℓ ) ,
∣ ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , x x ( s ) , φ x x ) d s ∣ ≤ β 1 1 / 2 ∥ σ n ∥ M g 1 , ε n ∥ φ x x ∥ , (31) \left|\int_{0}^{+\infty}g_{1,\varepsilon_n}(s)(\sigma_{n,xx}(s),\varphi_{xx})\,\mathrm{d}s\right|\leq\beta_1^{1/2}\|\sigma_n\|_{\mathcal{M}_{g_{1,\varepsilon_n}}}\|\varphi_{xx}\|, \tag*{(31)} ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , xx ( s ) , φ xx ) d s ≤ β 1 1/2 ∥ σ n ∥ M g 1 , ε n ∥ φ xx ∥ , (31)
and, valid for every φ ∈ H 0 1 ( 0 , ℓ ) \varphi\in H_0^1(0,\ell) φ ∈ H 0 1 ( 0 , ℓ ) ,
∣ ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , φ x ) d s ∣ ≤ β 2 1 / 2 ∥ η n ∥ M g 2 , ε ∥ φ x ∥ . (32) \left|\int_0^{+\infty} g_{2,\varepsilon_n}(s)(\eta_{n,x}(s),\varphi_x)\,\mathrm{d}s\right| \le\beta_2^{1/2}\|\eta_n\|_{\mathcal{M}_{g_2,\varepsilon}}\|\varphi_x\|. \tag*{(32)} ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , φ x ) d s ≤ β 2 1/2 ∥ η n ∥ M g 2 , ε ∥ φ x ∥. (32)
Both follow from the Cauchy–Schwarz inequality and the invariance of the mass (27) ; since β i \beta_i β i is independent of ε \varepsilon ε , so are the constants in (31) –(32) .
We distinguish two cases according to the behaviour of ω n \omega_n ω n .
Case A: ω n → 0 \omega_n \to0 ω n → 0 (in particular, ω n = 0 \omega_n=0 ω n = 0 is allowed ). We first note that (30) gives ψ n → 0 \psi_n \to0 ψ n → 0 and ϕ n → 0 \phi_n \to0 ϕ n → 0 in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) . Since ∥ u n ∥ , ∥ v n ∥ ≤ C ∥ U n ∥ H ε n = C \|u_n\|,\|v_n\| \le C\|U_n\|_{\mathcal{H}_{\varepsilon_n}}=C ∥ u n ∥ , ∥ v n ∥ ≤ C ∥ U n ∥ H ε n = C and ω n → 0 \omega_n\to0 ω n → 0 , we also have i ω n u n → 0 i\omega_nu_n\to0 i ω n u n → 0 and i ω n v n → 0 i\omega_nv_n\to0 i ω n v n → 0 in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) . Consequently the first and fourth components of (29) give f n 1 = ψ n − i ω n u n → 0 f_n^1=\psi_n-i\omega_nu_n\to0 f n 1 = ψ n − i ω n u n → 0 and f n 4 = ϕ n − i ω n v n → 0 f_n^4=\phi_n-i\omega_nv_n\to0 f n 4 = ϕ n − i ω n v n → 0 , consistently with F n → 0 F_n\to0 F n → 0 . Taking the L 2 L^2 L 2 -inner products of the second and fifth components of (29) with u n u_n u n and v n v_n v n , respectively, and taking real parts, we obtain
μ ˉ 1 ∥ u n , x x ∥ 2 + μ ˉ 2 ∥ v n , x ∥ 2 = − γ 1 Re ( ψ n , u n ) − γ 2 Re ( ϕ n , v n ) − Re ( f n 2 , u n ) − Re ( f n 5 , v n ) − Re ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , x x ( s ) , u n , x x ) d s − Re ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , v n , x ) d s . (33) \begin{aligned}
\bar{\mu}_1\|u_{n,xx}\|^2+\bar{\mu}_2\|v_{n,x}\|^2
={}&-\gamma_1\operatorname{Re}(\psi_n,u_n)-\gamma_2\operatorname{Re}(\phi_n,v_n)\\
&-\operatorname{Re}(f_n^2,u_n)-\operatorname{Re}(f_n^5,v_n)\\
&-\operatorname{Re}\int_0^{+\infty}g_{1,\varepsilon_n}(s)(\sigma_{n,xx}(s),u_{n,xx})\,\mathrm{d}s\\
&-\operatorname{Re}\int_0^{+\infty}g_{2,\varepsilon_n}(s)(\eta_{n,x}(s),v_{n,x})\,\mathrm{d}s. \tag*{(33)}
\end{aligned} μ ˉ 1 ∥ u n , xx ∥ 2 + μ ˉ 2 ∥ v n , x ∥ 2 = − γ 1 Re ( ψ n , u n ) − γ 2 Re ( ϕ n , v n ) − Re ( f n 2 , u n ) − Re ( f n 5 , v n ) − Re ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , xx ( s ) , u n , xx ) d s − Re ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , v n , x ) d s . (33)
Using the Poincaré inequalities ∥ u n ∥ ≤ c e ∥ u n , x x ∥ \|u_n\|\le\sqrt{c_e}\|u_{n,xx}\| ∥ u n ∥ ≤ c e ∥ u n , xx ∥ and ∥ v n ∥ ≤ c p ∥ v n , x ∥ \|v_n\|\le\sqrt{c_p}\|v_{n,x}\| ∥ v n ∥ ≤ c p ∥ v n , x ∥ , the bounds (31) –(32) , and Young inequality with parameters ϵ 1 , … , ϵ 6 > 0 \epsilon_1,\ldots,\epsilon_6>0 ϵ 1 , … , ϵ 6 > 0 to be chosen below, we estimate the right-hand side of (33) as
γ 1 ∣ Re ( ψ n , u n ) ∣ ≤ ϵ 1 ∥ u n , x x ∥ 2 + C ϵ 1 ∥ ψ n ∥ 2 , \gamma_1|\operatorname{Re}(\psi_n,u_n)|\le\epsilon_1\|u_{n,xx}\|^2+C_{\epsilon_1}\|\psi_n\|^2, γ 1 ∣ Re ( ψ n , u n ) ∣ ≤ ϵ 1 ∥ u n , xx ∥ 2 + C ϵ 1 ∥ ψ n ∥ 2 ,
γ 2 ∣ Re ( ϕ n , v n ) ∣ ≤ ϵ 2 ∥ v n , x ∥ 2 + C ϵ 2 ∥ ϕ n ∥ 2 , \gamma_2|\operatorname{Re}(\phi_n,v_n)|\le\epsilon_2\|v_{n,x}\|^2+C_{\epsilon_2}\|\phi_n\|^2, γ 2 ∣ Re ( ϕ n , v n ) ∣ ≤ ϵ 2 ∥ v n , x ∥ 2 + C ϵ 2 ∥ ϕ n ∥ 2 ,
∣ Re ( f n 2 , u n ) ∣ ≤ ϵ 3 ∥ u n , x x ∥ 2 + C ϵ 3 ∥ f n 2 ∥ 2 , |\operatorname{Re}(f_n^2,u_n)|\le\epsilon_3\|u_{n,xx}\|^2+C_{\epsilon_3}\|f_n^2\|^2, ∣ Re ( f n 2 , u n ) ∣ ≤ ϵ 3 ∥ u n , xx ∥ 2 + C ϵ 3 ∥ f n 2 ∥ 2 ,
∣ Re ( f n 5 , v n ) ∣ ≤ ϵ 4 ∥ v n , x ∥ 2 + C ϵ 4 ∥ f n 5 ∥ 2 , |\operatorname{Re}(f_n^5,v_n)|\le\epsilon_4\|v_{n,x}\|^2+C_{\epsilon_4}\|f_n^5\|^2, ∣ Re ( f n 5 , v n ) ∣ ≤ ϵ 4 ∥ v n , x ∥ 2 + C ϵ 4 ∥ f n 5 ∥ 2 ,
and the two memory terms are bounded by
β 1 1 / 2 ∥ σ n ∥ M g 1 , ε ∥ u n , x x ∥ ≤ ϵ 5 ∥ u n , x x ∥ 2 + C ϵ 5 β 1 ∥ σ n ∥ M g 1 , ε 2 , \beta_1^{1/2}\|\sigma_n\|_{\mathcal{M}_{g_1,\varepsilon}}\|u_{n,xx}\|\le\epsilon_5\|u_{n,xx}\|^2+C_{\epsilon_5}\beta_1\|\sigma_n\|_{\mathcal{M}_{g_1,\varepsilon}}^2, β 1 1/2 ∥ σ n ∥ M g 1 , ε ∥ u n , xx ∥ ≤ ϵ 5 ∥ u n , xx ∥ 2 + C ϵ 5 β 1 ∥ σ n ∥ M g 1 , ε 2 ,
β 2 1 / 2 ∥ η n ∥ M g 2 , ε ∥ v n , x ∥ ≤ ϵ 6 ∥ v n , x ∥ 2 + C ϵ 6 β 2 ∥ η n ∥ M g 2 , ε 2 . \beta_2^{1/2}\|\eta_n\|_{\mathcal{M}_{g_2,\varepsilon}}\|v_{n,x}\|\le\epsilon_6\|v_{n,x}\|^2+C_{\epsilon_6}\beta_2\|\eta_n\|_{\mathcal{M}_{g_2,\varepsilon}}^2. β 2 1/2 ∥ η n ∥ M g 2 , ε ∥ v n , x ∥ ≤ ϵ 6 ∥ v n , x ∥ 2 + C ϵ 6 β 2 ∥ η n ∥ M g 2 , ε 2 .
Choose ϵ 1 , … , ϵ 6 > 0 \epsilon_1,\ldots,\epsilon_6>0 ϵ 1 , … , ϵ 6 > 0 so small that
ϵ 1 + ϵ 3 + ϵ 5 < μ ˉ 1 2 , ϵ 2 + ϵ 4 + ϵ 6 < μ ˉ 2 2 . \epsilon_1+\epsilon_3+\epsilon_5<\frac{\bar{\mu}_1}{2},\qquad\epsilon_2+\epsilon_4+\epsilon_6<\frac{\bar{\mu}_2}{2}. ϵ 1 + ϵ 3 + ϵ 5 < 2 μ ˉ 1 , ϵ 2 + ϵ 4 + ϵ 6 < 2 μ ˉ 2 .
These choices depend only on γ i , β i , μ ˉ i , c e , c p \gamma_i,\beta_i,\bar{\mu}_i,c_e,c_p γ i , β i , μ ˉ i , c e , c p , hence not on n n n or ε \varepsilon ε . Absorbing the ϵ j \epsilon_j ϵ j -terms into the left-hand side of (33) , we obtain
μ ˉ 1 ∥ u n , x x ∥ 2 + μ ˉ 2 ∥ v n , x ∥ 2 ≤ C ( ∥ ψ n ∥ 2 + ∥ ϕ n ∥ 2 + ∥ σ n ∥ M g 1 , ε 2 + ∥ η n ∥ M g 2 , ε 2 + ∥ f n 2 ∥ 2 + ∥ f n 5 ∥ 2 ) , (34) \bar{\mu}_1\|u_{n,xx}\|^2+\bar{\mu}_2\|v_{n,x}\|^2\le C\left(\|\psi_n\|^2+\|\phi_n\|^2+\|\sigma_n\|_{\mathcal{M}_{g_1,\varepsilon}}^2+\|\eta_n\|_{\mathcal{M}_{g_2,\varepsilon}}^2+\|f_n^2\|^2+\|f_n^5\|^2\right), \tag*{(34)} μ ˉ 1 ∥ u n , xx ∥ 2 + μ ˉ 2 ∥ v n , x ∥ 2 ≤ C ( ∥ ψ n ∥ 2 + ∥ ϕ n ∥ 2 + ∥ σ n ∥ M g 1 , ε 2 + ∥ η n ∥ M g 2 , ε 2 + ∥ f n 2 ∥ 2 + ∥ f n 5 ∥ 2 ) , (34)
where C C C depends only on γ i , β i , μ ˉ i , c e , c p \gamma_i,\beta_i,\bar{\mu}_i,c_e,c_p γ i , β i , μ ˉ i , c e , c p and on the chosen ϵ j \epsilon_j ϵ j , hence not on n n n or ε \varepsilon ε . Since ∥ f n 2 ∥ 2 + ∥ f n 5 ∥ 2 ≤ ∥ F n ∥ H ε n 2 \|f_n^2\|^2+\|f_n^5\|^2\le\|F_n\|_{\mathcal{H}_{\varepsilon_n}}^2 ∥ f n 2 ∥ 2 + ∥ f n 5 ∥ 2 ≤ ∥ F n ∥ H ε n 2 , the right-hand side tends to zero uniformly in ε \varepsilon ε by (30) and F n → 0 F_n\to0 F n → 0 . Hence ∥ U n ∥ H ε n → 0 \|U_n\|_{\mathcal{H}_{\varepsilon_n}}\to0 ∥ U n ∥ H ε n → 0 , contradicting ∥ U n ∥ H ε n = 1 \|U_n\|_{\mathcal{H}_{\varepsilon_n}}=1 ∥ U n ∥ H ε n = 1 .
Case B: ω n \omega_n ω n does not tend to 0 0 0 along a subsequence. Passing to a subsequence, we may assume ∣ ω n ∣ ≥ c 0 > 0 |\omega_n|\ge c_0>0 ∣ ω n ∣ ≥ c 0 > 0 for some c 0 > 0 c_0>0 c 0 > 0 . From the first and fourth components of (29) , we have i ω n u n = f n 1 − ψ n i\omega_n u_n=f_n^1-\psi_n i ω n u n = f n 1 − ψ n and i ω n v n = f n 4 − ϕ n i\omega_n v_n=f_n^4-\phi_n i ω n v n = f n 4 − ϕ n . Since f n 1 → 0 f_n^1\to0 f n 1 → 0 in H a 2 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) H_a^2(0,\ell)\hookrightarrow L^2(0,\ell) H a 2 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) , f n 4 → 0 f_n^4\to0 f n 4 → 0 in H 0 1 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) H_0^1(0,\ell)\hookrightarrow L^2(0,\ell) H 0 1 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) , and ψ n , ϕ n → 0 \psi_n,\phi_n\to0 ψ n , ϕ n → 0 in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) by (30) , we deduce that
∥ u n ∥ L 2 ≤ 1 ∣ ω n ∣ ( ∥ f n 1 ∥ + ∥ ψ n ∥ ) → 0 , ∥ v n ∥ L 2 ≤ 1 ∣ ω n ∣ ( ∥ f n 4 ∥ + ∥ ϕ n ∥ ) → 0. \lVert u_n\rVert_{L^2}\leq\frac{1}{\lvert\omega_n\rvert}\left(\lVert f_n^1\rVert+\lVert\psi_n\rVert\right)\to0,\qquad\lVert v_n\rVert_{L^2}\leq\frac{1}{\lvert\omega_n\rvert}\left(\lVert f_n^4\rVert+\lVert\phi_n\rVert\right)\to0. ∥ u n ∥ L 2 ≤ ∣ ω n ∣ 1 ( ∥ f n 1 ∥ + ∥ ψ n ∥ ) → 0 , ∥ v n ∥ L 2 ≤ ∣ ω n ∣ 1 ( ∥ f n 4 ∥ + ∥ ϕ n ∥ ) → 0.
the embedding constants of H a 2 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) H_a^2(0,\ell)\hookrightarrow L^2(0,\ell) H a 2 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) and H 0 1 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) H_0^1(0,\ell)\hookrightarrow L^2(0,\ell) H 0 1 ( 0 , ℓ ) ↪ L 2 ( 0 , ℓ ) are independent of ε \varepsilon ε .
Taking the L 2 L^2 L 2 -inner products of the second component of (29) with u n u_n u n and of the first component with ψ n \psi_n ψ n , and adding the real parts, we obtain
μ ˉ 1 ∥ u n , x x ∥ 2 = ∥ ψ n ∥ 2 − Re ( ψ n , f n 1 ) − γ 1 Re ( ψ n , u n ) \bar{\mu}_1\lVert u_{n,xx}\rVert^2=\lVert\psi_n\rVert^2-\operatorname{Re}(\psi_n,f_n^1)-\gamma_1\operatorname{Re}(\psi_n,u_n) μ ˉ 1 ∥ u n , xx ∥ 2 = ∥ ψ n ∥ 2 − Re ( ψ n , f n 1 ) − γ 1 Re ( ψ n , u n )
− Re ( f n 2 , u n ) − Re ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , x x ( s ) , u n , x x ) d s . (35) {}-\operatorname{Re}(f_n^2,u_n)-\operatorname{Re}\int_0^{+\infty}g_{1,\varepsilon_n}(s)(\sigma_{n,xx}(s),u_{n,xx})\,\mathrm{d}s. \tag*{(35)} − Re ( f n 2 , u n ) − Re ∫ 0 + ∞ g 1 , ε n ( s ) ( σ n , xx ( s ) , u n , xx ) d s . (35)
Indeed, the term i ω n ( ψ n , u n ) i\omega_n(\psi_n,u_n) i ω n ( ψ n , u n ) produced by the second component is rewritten, using the first component to substitute i ω n u n = f n 1 − ψ n i\omega_nu_n=f_n^1-\psi_n i ω n u n = f n 1 − ψ n , as
i ω n ( ψ n , u n ) = ∫ ψ n i ω n u n ‾ = ∫ ψ n f n 1 − ψ n ‾ = ( ψ n , f n 1 ) − ∥ ψ n ∥ 2 , i\omega_n(\psi_n,u_n)=\int\psi_n\overline{i\omega_nu_n}=\int\psi_n\overline{f_n^1-\psi_n}=(\psi_n,f_n^1)-\lVert\psi_n\rVert^2, i ω n ( ψ n , u n ) = ∫ ψ n i ω n u n = ∫ ψ n f n 1 − ψ n = ( ψ n , f n 1 ) − ∥ ψ n ∥ 2 ,
and taking real parts gives Re [ i ω n ( ψ n , u n ) ] = Re ( ψ n , f n 1 ) − ∥ ψ n ∥ 2 \operatorname{Re}[i\omega_n(\psi_n,u_n)]=\operatorname{Re}(\psi_n,f_n^1)-\lVert\psi_n\rVert^2 Re [ i ω n ( ψ n , u n )] = Re ( ψ n , f n 1 ) − ∥ ψ n ∥ 2 . On the other hand, from the first component, Re [ i ω n ( u n , ψ n ) ] = Re ( f n 1 , ψ n ) − ∥ ψ n ∥ 2 = − Re ( ψ n , f n 1 ) + ∥ ψ n ∥ 2 \operatorname{Re}[i\omega_n(u_n,\psi_n)]=\operatorname{Re}(f_n^1,\psi_n)-\lVert\psi_n\rVert^2=-\operatorname{Re}(\psi_n,f_n^1)+\lVert\psi_n\rVert^2 Re [ i ω n ( u n , ψ n )] = Re ( f n 1 , ψ n ) − ∥ ψ n ∥ 2 = − Re ( ψ n , f n 1 ) + ∥ ψ n ∥ 2 , and the two contributions cancel in the sum, yielding (35) . Each of the five terms on the right-hand side of (35) tends to zero as n → + ∞ n\to+\infty n → + ∞ , uniformly in ε \varepsilon ε :
∣ Re ( ψ n , f n 1 ) ∣ ≤ ∥ ψ n ∥ ∥ f n 1 ∥ → 0 \lvert\operatorname{Re}(\psi_n,f_n^1)\rvert\leq\lVert\psi_n\rVert\lVert f_n^1\rVert\to0 ∣ Re ( ψ n , f n 1 )∣ ≤ ∥ ψ n ∥ ∥ f n 1 ∥ → 0 by (30) and ∥ F n ∥ H ε n → 0 \lVert F_n\rVert_{\mathcal{H}_{\varepsilon_n}}\to0 ∥ F n ∥ H ε n → 0 ;
∥ ψ n ∥ 2 → 0 \lVert\psi_n\rVert^2\to0 ∥ ψ n ∥ 2 → 0 by (30) ;
γ 1 ∣ Re ( ψ n , u n ) ∣ ≤ γ 1 ∥ ψ n ∥ ∥ u n ∥ → 0 \gamma_1\lvert\operatorname{Re}(\psi_n,u_n)\rvert\leq\gamma_1\lVert\psi_n\rVert\lVert u_n\rVert\to0 γ 1 ∣ Re ( ψ n , u n )∣ ≤ γ 1 ∥ ψ n ∥ ∥ u n ∥ → 0 , since u n → 0 u_n\to0 u n → 0 in L 2 L^2 L 2 and ∥ ψ n ∥ → 0 \lVert\psi_n\rVert\to0 ∥ ψ n ∥ → 0 ;
∣ Re ( f n 2 , u n ) ∣ ≤ ∥ f n 2 ∥ ∥ u n ∥ → 0 \lvert\operatorname{Re}(f_n^2,u_n)\rvert\leq\lVert f_n^2\rVert\lVert u_n\rVert\to0 ∣ Re ( f n 2 , u n )∣ ≤ ∥ f n 2 ∥ ∥ u n ∥ → 0 , since u n → 0 u_n\to0 u n → 0 in L 2 L^2 L 2 and ∥ f n 2 ∥ → 0 \lVert f_n^2\rVert\to0 ∥ f n 2 ∥ → 0 ;
the memory term is bounded by β 1 1 / 2 ∥ σ n ∥ M g 1 , ε ∥ u n , x x ∥ = o ( 1 ) ∥ u n , x x ∥ \beta_1^{1/2}\lVert\sigma_n\rVert_{\mathcal{M}_{g_{1,\varepsilon}}}\lVert u_{n,xx}\rVert=o(1)\lVert u_{n,xx}\rVert β 1 1/2 ∥ σ n ∥ M g 1 , ε ∥ u n , xx ∥ = o ( 1 ) ∥ u n , xx ∥ via (31) and (30) .
None of the constants involved in (i)–(v) depends on ε \varepsilon ε . Consequently, (35) gives μ ˉ 1 ∥ u n , x x ∥ 2 ≤ o ( 1 ) + o ( 1 ) ∥ u n , x x ∥ \bar{\mu}_1\lVert u_{n,xx}\rVert^2\leq o(1)+o(1)\lVert u_{n,xx}\rVert μ ˉ 1 ∥ u n , xx ∥ 2 ≤ o ( 1 ) + o ( 1 ) ∥ u n , xx ∥ , where o ( 1 ) o(1) o ( 1 ) denotes a quantity tending to zero as n → + ∞ n\to+\infty n → + ∞ , uniformly in ε \varepsilon ε . By Young inequality, ∥ u n , x x ∥ 2 ≤ o ( 1 ) \lVert u_{n,xx}\rVert^2\leq o(1) ∥ u n , xx ∥ 2 ≤ o ( 1 ) , so u n , x x → 0 u_{n,xx}\to0 u n , xx → 0 in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) . The same argument applied to the fifth and fourth components, using i ω n v n = f n 4 − ϕ n i\omega_nv_n=f_n^4-\phi_n i ω n v n = f n 4 − ϕ n , gives v n , x → 0 v_{n,x}\to0 v n , x → 0 in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) ; the corresponding identity reads
μ ˉ 2 ∥ v n , x ∥ 2 = ∥ ϕ n ∥ 2 − Re ( ϕ n , f n 4 ) − γ 2 Re ( ϕ n , v n ) \bar{\mu}_2\lVert v_{n,x}\rVert^2=\lVert\phi_n\rVert^2-\operatorname{Re}(\phi_n,f_n^4)-\gamma_2\operatorname{Re}(\phi_n,v_n) μ ˉ 2 ∥ v n , x ∥ 2 = ∥ ϕ n ∥ 2 − Re ( ϕ n , f n 4 ) − γ 2 Re ( ϕ n , v n )
− Re ( f n 5 , v n ) − Re ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , v n , x ) d s . (36) {}-\operatorname{Re}(f_n^5,v_n)-\operatorname{Re}\int_0^{+\infty}g_{2,\varepsilon_n}(s)(\eta_{n,x}(s),v_{n,x})\,\mathrm{d}s. \tag*{(36)} − Re ( f n 5 , v n ) − Re ∫ 0 + ∞ g 2 , ε n ( s ) ( η n , x ( s ) , v n , x ) d s . (36)
and each of its five terms tends to zero uniformly in ε \varepsilon ε by the same reasoning as in (i)–(v). Consequently, all six components of ∥ U n ∥ H ε n 2 \lVert U_n\rVert_{\mathcal{H}_{\varepsilon_n}}^2 ∥ U n ∥ H ε n 2 tend to zero, contradicting ∥ U n ∥ H ε n = 1 \lVert U_n\rVert_{\mathcal{H}_{\varepsilon_n}}=1 ∥ U n ∥ H ε n = 1 .
Both cases being impossible, we conclude that there exist no sequences ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + , ω n ∈ R \omega_n\in\mathbb{R} ω n ∈ R , and U n U_n U n with the above properties. Hence i R ⊂ ρ ( A ε ) i\mathbb{R}\subset\rho(A_\varepsilon) i R ⊂ ρ ( A ε ) for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , uniformly in ε \varepsilon ε .
Condition (C2). Let Υ = ( h 1 , … , h 6 ) ∈ H ε \Upsilon=(h^1,\ldots,h^6)\in\mathcal{H}_\varepsilon Υ = ( h 1 , … , h 6 ) ∈ H ε and let U = ( u , ψ , σ , v , ϕ , η ) ∈ D ( A ε ) U=(u,\psi,\sigma,v,\phi,\eta)\in D(A_\varepsilon) U = ( u , ψ , σ , v , ϕ , η ) ∈ D ( A ε ) solve
i λ U − A ε U = Υ , λ ∈ R . (37) i\lambda U-A_\varepsilon U=\Upsilon,\qquad\lambda\in\mathbb{R}. \tag*{(37)} iλ U − A ε U = Υ , λ ∈ R . (37)
We prove that there exists C > 0 C>0 C > 0 , independent of λ \lambda λ and ε \varepsilon ε , such that
∥ U ∥ H ε ≤ C ∥ Υ ∥ H ε . (38) \lVert U\rVert_{\mathcal{H}_\varepsilon}\leq C\lVert\Upsilon\rVert_{\mathcal{H}_\varepsilon}. \tag*{(38)} ∥ U ∥ H ε ≤ C ∥ Υ ∥ H ε . (38)
Taking the real part of the inner product of (37) with U U U and using (28) ,
γ 1 ∥ ψ ∥ 2 + λ 1 2 ∥ σ ∥ M g 1 , ε 2 + γ 2 ∥ ϕ ∥ 2 + λ 2 2 ∥ η ∥ M g 2 , ε 2 ≤ ∥ U ∥ H ε ∥ Υ ∥ H ε . (39) \gamma_1\|\psi\|^2+\frac{\lambda_1}{2}\|\sigma\|_{M_{g_1,\varepsilon}}^2+\gamma_2\|\phi\|^2+\frac{\lambda_2}{2}\|\eta\|_{M_{g_2,\varepsilon}}^2\le\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}. \tag*{(39)} γ 1 ∥ ψ ∥ 2 + 2 λ 1 ∥ σ ∥ M g 1 , ε 2 + γ 2 ∥ ϕ ∥ 2 + 2 λ 2 ∥ η ∥ M g 2 , ε 2 ≤ ∥ U ∥ H ε ∥Υ ∥ H ε . (39)
Hence
∥ ψ ∥ , ∥ σ ∥ M g 1 , ε , ∥ ϕ ∥ , ∥ η ∥ M g 2 , ε ≤ C ∥ U ∥ H ε ∥ Υ ∥ H ε , (40) \|\psi\|,\ \|\sigma\|_{M_{g_1,\varepsilon}},\ \|\phi\|,\ \|\eta\|_{M_{g_2,\varepsilon}}\le C\sqrt{\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}}, \tag*{(40)} ∥ ψ ∥ , ∥ σ ∥ M g 1 , ε , ∥ ϕ ∥ , ∥ η ∥ M g 2 , ε ≤ C ∥ U ∥ H ε ∥Υ ∥ H ε , (40)
where C C C depends only on γ i , λ i \gamma_i,\lambda_i γ i , λ i ; in particular, C C C does not depend on ε \varepsilon ε .
From the first and fourth components of (37) , i λ u = h 1 − ψ i\lambda u=h^1-\psi iλ u = h 1 − ψ and i λ v = h 4 − ϕ i\lambda v=h^4-\phi iλ v = h 4 − ϕ . Taking the L 2 L^2 L 2 -inner products of the second and fifth components with u u u and v v v , respectively, and adding the real parts, we obtain, after the same cancellation of i λ ( ψ , u ) i\lambda(\psi,u) iλ ( ψ , u ) and i λ ( ϕ , v ) i\lambda(\phi,v) iλ ( ϕ , v ) as above,
μ ˉ 1 ∥ u x x ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 = Re ( ψ , h 1 ) + Re ( ϕ , h 4 ) − ∥ ψ ∥ 2 − ∥ ϕ ∥ 2 − γ 1 Re ( ψ , u ) − γ 2 Re ( ϕ , v ) − Re ( h 2 , u ) − Re ( h 5 , v ) − Re ∫ 0 + ∞ g 1 , ε ( s ) ( σ x x ( s ) , u x x ) d s − Re ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , v x ) d s . (41) \begin{aligned}
\bar{\mu}_1\|u_{xx}\|^2+\bar{\mu}_2\|v_x\|^2
={}&\operatorname{Re}(\psi,h^1)+\operatorname{Re}(\phi,h^4)-\|\psi\|^2-\|\phi\|^2\\
&-\gamma_1\operatorname{Re}(\psi,u)-\gamma_2\operatorname{Re}(\phi,v)-\operatorname{Re}(h^2,u)-\operatorname{Re}(h^5,v)\\
&-\operatorname{Re}\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{xx}(s),u_{xx})\,\mathrm{d}s-\operatorname{Re}\int_0^{+\infty}g_{2,\varepsilon}(s)(\eta_x(s),v_x)\,\mathrm{d}s. \tag*{(41)}
\end{aligned} μ ˉ 1 ∥ u xx ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 = Re ( ψ , h 1 ) + Re ( ϕ , h 4 ) − ∥ ψ ∥ 2 − ∥ ϕ ∥ 2 − γ 1 Re ( ψ , u ) − γ 2 Re ( ϕ , v ) − Re ( h 2 , u ) − Re ( h 5 , v ) − Re ∫ 0 + ∞ g 1 , ε ( s ) ( σ xx ( s ) , u xx ) d s − Re ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , v x ) d s . (41)
We bound the right-hand side term by term. First, by Young’s inequality and ∥ h 1 ∥ L 2 ≤ ∥ h 1 ∥ H A 2 ≤ ∥ Υ ∥ H ε \|h^1\|_{L^2}\le\|h^1\|_{H_A^2}\le\|\Upsilon\|_{\mathcal{H}_\varepsilon} ∥ h 1 ∥ L 2 ≤ ∥ h 1 ∥ H A 2 ≤ ∥Υ ∥ H ε ,
Re ( ψ , h 1 ) − ∥ ψ ∥ 2 ≤ 1 2 ∥ ψ ∥ 2 + 1 2 ∥ h 1 ∥ L 2 2 − ∥ ψ ∥ 2 ≤ 1 2 ∥ Υ ∥ H ε 2 , \operatorname{Re}(\psi,h^1)-\|\psi\|^2\le\frac{1}{2}\|\psi\|^2+\frac{1}{2}\|h^1\|_{L^2}^2-\|\psi\|^2\le\frac{1}{2}\|\Upsilon\|_{\mathcal{H}_\varepsilon}^2, Re ( ψ , h 1 ) − ∥ ψ ∥ 2 ≤ 2 1 ∥ ψ ∥ 2 + 2 1 ∥ h 1 ∥ L 2 2 − ∥ ψ ∥ 2 ≤ 2 1 ∥Υ ∥ H ε 2 ,
and similarly for Re ( ϕ , h 4 ) − ∥ ϕ ∥ 2 \operatorname{Re}(\phi,h^4)-\|\phi\|^2 Re ( ϕ , h 4 ) − ∥ ϕ ∥ 2 . Second, by Poincaré’s inequality and (40) ,
γ 1 ∣ Re ( ψ , u ) ∣ + γ 2 ∣ Re ( ϕ , v ) ∣ ≤ ε ( ∥ u x x ∥ 2 + ∥ v x ∥ 2 ) + C ε ( ∥ ψ ∥ 2 + ∥ ϕ ∥ 2 ) ≤ ε ( ∥ u x x ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥ U ∥ H ε ∥ Υ ∥ H ε . \begin{aligned}
\gamma_1|\operatorname{Re}(\psi,u)|+\gamma_2|\operatorname{Re}(\phi,v)|
&\le\varepsilon(\|u_{xx}\|^2+\|v_x\|^2)+C_\varepsilon(\|\psi\|^2+\|\phi\|^2)\\
&\le\varepsilon(\|u_{xx}\|^2+\|v_x\|^2)+C_\varepsilon\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}.
\end{aligned} γ 1 ∣ Re ( ψ , u ) ∣ + γ 2 ∣ Re ( ϕ , v ) ∣ ≤ ε ( ∥ u xx ∥ 2 + ∥ v x ∥ 2 ) + C ε ( ∥ ψ ∥ 2 + ∥ ϕ ∥ 2 ) ≤ ε ( ∥ u xx ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥ U ∥ H ε ∥Υ ∥ H ε .
Third, again by Poincaré’s inequality and ∥ h 2 ∥ L 2 ≤ ∥ Υ ∥ H ε \|h^2\|_{L^2}\le\|\Upsilon\|_{\mathcal{H}_\varepsilon} ∥ h 2 ∥ L 2 ≤ ∥Υ ∥ H ε ,
∣ Re ( h 2 , u ) ∣ + ∣ Re ( h 5 , v ) ∣ ≤ ε ( ∥ u x x ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥ Υ ∥ H ε 2 . |\operatorname{Re}(h^2,u)|+|\operatorname{Re}(h^5,v)|\le\varepsilon(\|u_{xx}\|^2+\|v_x\|^2)+C_\varepsilon\|\Upsilon\|_{\mathcal{H}_\varepsilon}^2. ∣ Re ( h 2 , u ) ∣ + ∣ Re ( h 5 , v ) ∣ ≤ ε ( ∥ u xx ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥Υ ∥ H ε 2 .
Fourth, by (31) –(32) and (40) ,
∣ Re ∫ 0 + ∞ g 1 , ε ( s ) ( σ x x ( s ) , u x x ) d s ∣ + ∣ Re ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , v x ) d s ∣ ≤ ε ( ∥ u x x ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥ U ∥ H ε ∥ Υ ∥ H ε . \begin{aligned}
\left|\operatorname{Re}\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{xx}(s),u_{xx})\,\mathrm{d}s\right|
+\left|\operatorname{Re}\int_0^{+\infty}g_{2,\varepsilon}(s)(\eta_x(s),v_x)\,\mathrm{d}s\right|\\
\le\varepsilon(\|u_{xx}\|^2+\|v_x\|^2)+C_\varepsilon\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}.
\end{aligned} Re ∫ 0 + ∞ g 1 , ε ( s ) ( σ xx ( s ) , u xx ) d s + Re ∫ 0 + ∞ g 2 , ε ( s ) ( η x ( s ) , v x ) d s ≤ ε ( ∥ u xx ∥ 2 + ∥ v x ∥ 2 ) + C ε ∥ U ∥ H ε ∥Υ ∥ H ε .
Summing all these and choosing ε \varepsilon ε so small that 4 ε < min { μ ˉ 1 , μ ˉ 2 } / 2 4\varepsilon<\min\{\bar{\mu}_1,\bar{\mu}_2\}/2 4 ε < min { μ ˉ 1 , μ ˉ 2 } /2 , we absorb the ε ( ∥ u x x ∥ 2 + ∥ v x ∥ 2 ) \varepsilon(\|u_{xx}\|^2+\|v_x\|^2) ε ( ∥ u xx ∥ 2 + ∥ v x ∥ 2 ) terms into the left-hand side and obtain
μ ˉ 1 ∥ u x x ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 ≤ C ( ∥ U ∥ H ε ∥ Υ ∥ H ε + ∥ Υ ∥ H ε 2 ) , (42) \bar{\mu}_1\|u_{xx}\|^2+\bar{\mu}_2\|v_x\|^2\le C\left(\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}+\|\Upsilon\|_{\mathcal{H}_\varepsilon}^2\right), \tag*{(42)} μ ˉ 1 ∥ u xx ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 ≤ C ( ∥ U ∥ H ε ∥Υ ∥ H ε + ∥Υ ∥ H ε 2 ) , (42)
where C C C depends only on γ i , μ i , β i , c e , c p \gamma_i,\mu_i,\beta_i,c_e,c_p γ i , μ i , β i , c e , c p , but not on λ \lambda λ or ε \varepsilon ε . The only place where the kernel enters is the pair of memory convolutions, controlled by the mass β i \beta_i β i through (31) –(32) ; by (27) , β i \beta_i β i is independent of ε \varepsilon ε .
Combining (39) , (42) and the identity
∥ U ∥ H ε 2 = μ ˉ 1 ∥ u x x ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + ∥ ϕ ∥ 2 + ∥ η ∥ M g 2 , ε 2 , \|U\|_{\mathcal{H}_\varepsilon}^2=\bar{\mu}_1\|u_{xx}\|^2+\bar{\mu}_2\|v_x\|^2+\|\psi\|^2+\|\sigma\|_{M_{g_1,\varepsilon}}^2+\|\phi\|^2+\|\eta\|_{M_{g_2,\varepsilon}}^2, ∥ U ∥ H ε 2 = μ ˉ 1 ∥ u xx ∥ 2 + μ ˉ 2 ∥ v x ∥ 2 + ∥ ψ ∥ 2 + ∥ σ ∥ M g 1 , ε 2 + ∥ ϕ ∥ 2 + ∥ η ∥ M g 2 , ε 2 ,
we obtain
∥ U ∥ H ε 2 ≤ C ( ∥ U ∥ H ε ∥ Υ ∥ H ε + ∥ Υ ∥ H ε 2 ) . (43) \|U\|_{\mathcal{H}_\varepsilon}^2\le C\left(\|U\|_{\mathcal{H}_\varepsilon}\|\Upsilon\|_{\mathcal{H}_\varepsilon}+\|\Upsilon\|_{\mathcal{H}_\varepsilon}^2\right). \tag*{(43)} ∥ U ∥ H ε 2 ≤ C ( ∥ U ∥ H ε ∥Υ ∥ H ε + ∥Υ ∥ H ε 2 ) . (43)
If Υ = 0 \Upsilon=0 Υ = 0 , then (43) gives ∥ U ∥ H ε = 0 \lVert U\rVert_{\mathcal{H}_{\varepsilon}}=0 ∥ U ∥ H ε = 0 , and (38) is trivial. Otherwise, set t = ∥ U ∥ H ε / ∥ Υ ∥ H ε t=\lVert U\rVert_{\mathcal{H}_{\varepsilon}}/\lVert\Upsilon\rVert_{\mathcal{H}_{\varepsilon}} t = ∥ U ∥ H ε / ∥ Υ ∥ H ε . Dividing (43) by ∥ Υ ∥ H ε 2 \lVert\Upsilon\rVert_{\mathcal{H}_{\varepsilon}}^2 ∥ Υ ∥ H ε 2 gives t 2 ≤ C ( t + 1 ) t^2\le C(t+1) t 2 ≤ C ( t + 1 ) , hence
t ≤ C ′ : = C + C 2 + 4 C 2 , t\le C' := \frac{C+\sqrt{C^2+4C}}{2}, t ≤ C ′ := 2 C + C 2 + 4 C ,
a constant depending only on C C C , hence only on γ i , μ i , β i , c e , c p \gamma_i,\mu_i,\beta_i,c_e,c_p γ i , μ i , β i , c e , c p , but not on λ \lambda λ or ε \varepsilon ε . Therefore ∥ U ∥ H ε ≤ C ′ ∥ Υ ∥ H ε \lVert U\rVert_{\mathcal{H}_{\varepsilon}}\le C'\lVert\Upsilon\rVert_{\mathcal{H}_{\varepsilon}} ∥ U ∥ H ε ≤ C ′ ∥ Υ ∥ H ε , which is (38) .
We have verified conditions (C1) and (C2) of the Gearhart–Prüss theorem with constants independent of ε \varepsilon ε . Therefore the linear semigroup e − A ε t e^{-A_{\varepsilon}t} e − A ε t is exponentially stable, uniformly in ε \varepsilon ε : there exist α ≥ 1 \alpha\ge1 α ≥ 1 and β > 0 \beta>0 β > 0 , independent of ε \varepsilon ε , such that
∥ e − A ε t ∥ L ( H ε ) ≤ α e − β t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . \lVert e^{-A_{\varepsilon}t}\rVert_{\mathcal{L}(\mathcal{H}_{\varepsilon})}\le\alpha e^{-\beta t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1]. ∥ e − A ε t ∥ L ( H ε ) ≤ α e − β t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] .
This completes the proof of Theorem 2.2.
Formally passing to the limit ε → 0 + \varepsilon\to0^+ ε → 0 + in (8) with g i g_i g i replaced by g i , ε g_{i,\varepsilon} g i , ε , and using g i , ε ⇀ β i δ 0 g_{i,\varepsilon}\rightharpoonup\beta_i\delta_0 g i , ε ⇀ β i δ 0 , the memory convolution satisfies, in the distributional sense,
− ∫ 0 + ∞ g 1 , ε ( s ) u x x x x ( x , t − s ) d s ⇀ − β 1 u x x x x ( x , t ) . -\int_0^{+\infty}g_{1,\varepsilon}(s)u_{xxxx}(x,t-s)\,ds\rightharpoonup-\beta_1u_{xxxx}(x,t). − ∫ 0 + ∞ g 1 , ε ( s ) u xxxx ( x , t − s ) d s ⇀ − β 1 u xxxx ( x , t ) .
and similarly
+ ∫ 0 + ∞ g 2 , ε ( s ) v x x ( x , t − s ) d s ⇀ + β 2 v x x ( x , t ) . +\int_0^{+\infty}g_{2,\varepsilon}(s)v_{xx}(x,t-s)\,ds\rightharpoonup+\beta_2v_{xx}(x,t). + ∫ 0 + ∞ g 2 , ε ( s ) v xx ( x , t − s ) d s ⇀ + β 2 v xx ( x , t ) .
Since μ ˉ i = μ i − β i \bar{\mu}_i=\mu_i-\beta_i μ ˉ i = μ i − β i , the limiting system becomes
{ u t t + γ 1 u t + μ 1 u x x x x + k [ u − v ] + + h 1 ( u ) = f 1 ( x ) , v t t + γ 2 v t − μ 2 v x x − k [ u − v ] + + h 2 ( v ) = f 2 ( x ) , (44) \begin{cases}
u_{tt}+\gamma_1u_t+\mu_1u_{xxxx}+k[u-v]^++h_1(u)=f_1(x),\\
v_{tt}+\gamma_2v_t-\mu_2v_{xx}-k[u-v]^++h_2(v)=f_2(x), \tag*{(44)}
\end{cases} { u tt + γ 1 u t + μ 1 u xxxx + k [ u − v ] + + h 1 ( u ) = f 1 ( x ) , v tt + γ 2 v t − μ 2 v xx − k [ u − v ] + + h 2 ( v ) = f 2 ( x ) , (44)
where we used μ ˉ i + β i = μ i \bar{\mu}_i+\beta_i=\mu_i μ ˉ i + β i = μ i . Thus the limit system is exactly the classical memoryless coupled suspension bridge system [20 ] , with the same elastic coefficients μ i \mu_i μ i as in the original system. Note that the effective coefficients μ ˉ i , ε = μ i − β i \bar{\mu}_{i,\varepsilon}=\mu_i-\beta_i μ ˉ i , ε = μ i − β i appearing in the rewritten system are independent of ε \varepsilon ε , while the limit equation is written in terms of the original coefficients μ i = μ ˉ i + β i \mu_i=\bar{\mu}_i+\beta_i μ i = μ ˉ i + β i ; the two descriptions are equivalent.
The limit semigroup S 0 ( t ) S_0(t) S 0 ( t ) is generated by the operator A 0 A_0 A 0 obtained by replacing the memory convolutions with the local terms − β 1 u x x x x -\beta_1u_{xxxx} − β 1 u xxxx and + β 2 v x x +\beta_2v_{xx} + β 2 v xx and absorbing them into the elastic coefficients, which amounts to replacing μ ˉ i \bar{\mu}_i μ ˉ i by μ i = μ ˉ i + β i \mu_i=\bar{\mu}_i+\beta_i μ i = μ ˉ i + β i . Its global attractor A 0 \mathcal{A}_0 A 0 exists by [20 ] .
Uniform Dissipativity
Throughout this section we use the abbreviation μ ˉ i = μ ˉ i , ε = μ i − β i \bar{\mu}_i=\bar{\mu}_{i,\varepsilon}=\mu_i-\beta_i μ ˉ i = μ ˉ i , ε = μ i − β i , which is independent of ε \varepsilon ε by [20 ] and [27 ] ; we also write M g i , ε = M g i , ε \mathcal{M}_{g_{i,\varepsilon}}=\mathcal{M}_{g_{i,\varepsilon}} M g i , ε = M g i , ε .
The first step toward any continuity result is a uniform dissipativity estimate : there exists a bounded set B ⊂ H ε \mathcal{B}\subset\mathcal{H}_{\varepsilon} B ⊂ H ε , independent of ε \varepsilon ε , which absorbs all bounded subsets uniformly in ε \varepsilon ε . To formulate this uniformly, we say that a set B \mathcal{B} B is a uniform absorbing set if B ⊂ H ε \mathcal{B}\subset\mathcal{H}_{\varepsilon} B ⊂ H ε for every ε \varepsilon ε (in the sense of Remark 2.1) and if there exists R 0 > 0 R_0>0 R 0 > 0 such that B ⊂ { U : ∥ U ∥ H ε ≤ R 0 } \mathcal{B}\subset\{U:\lVert U\rVert_{\mathcal{H}_{\varepsilon}}\le R_0\} B ⊂ { U : ∥ U ∥ H ε ≤ R 0 } for all ε \varepsilon ε .
Proposition 3.1 (Uniform absorbing set). Let [9 ] and [11 ] hold. Then there exist R 0 > 0 R_0>0 R 0 > 0 and, for every bounded set B ⊂ H ε \mathcal{B}\subset\mathcal{H}_{\varepsilon} B ⊂ H ε , a time t 0 = t 0 ( B ) ≥ 0 t_0=t_0(\mathcal{B})\ge0 t 0 = t 0 ( B ) ≥ 0 such that, uniformly in ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] ,
S ε ( t ) B ⊂ B 0 : = { U ∈ H ε : ∥ U ∥ H ε ≤ R 0 } , ∀ t ≥ t 0 . S_{\varepsilon}(t)\mathcal{B}\subset\mathcal{B}_0:=\{U\in\mathcal{H}_{\varepsilon}:\lVert U\rVert_{\mathcal{H}_{\varepsilon}}\le R_0\},\qquad\forall t\ge t_0. S ε ( t ) B ⊂ B 0 := { U ∈ H ε : ∥ U ∥ H ε ≤ R 0 } , ∀ t ≥ t 0 . Moreover, R 0 R_0 R 0 depends only on γ i , μ i , λ i , β i , c e , c p , ∥ f 1 ∥ , ∥ f 2 ∥ \gamma_i,\mu_i,\lambda_i,\beta_i,c_e,c_p,\lVert f_1\rVert,\lVert f_2\rVert γ i , μ i , λ i , β i , c e , c p , ∥ f 1 ∥ , ∥ f 2 ∥ , and the constants in [11 ] , but not on B \mathcal{B} B or on ε \varepsilon ε .
Proof. Let B ⊂ H ε B \subset\mathcal{H}_{\varepsilon} B ⊂ H ε be a bounded set, and let U 0 , ε ∈ B U_{0,\varepsilon} \in B U 0 , ε ∈ B be arbitrary. We write U ε ( t ) = S ε ( t ) U 0 , ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) U_{\varepsilon}(t) = S_{\varepsilon}(t)U_{0,\varepsilon} = (u_{\varepsilon}, u_{\varepsilon,t}, \sigma_{\varepsilon}, v_{\varepsilon}, v_{\varepsilon,t}, \eta_{\varepsilon}) U ε ( t ) = S ε ( t ) U 0 , ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) for the corresponding solution of the perturbed system. Define the energy functional
E ε ( t ) = 1 2 [ ∥ u ε , t ∥ 2 + μ ˉ 1 ∥ u ε , x x ∥ 2 + ∥ σ ε ∥ M g 1 , ε 2 + ∥ v ε , t ∥ 2 + μ ˉ 2 ∥ v ε , x ∥ 2 + ∥ η ε ∥ M g 2 , ε 2 ] + k 2 ∥ [ u ε − v ε ] + ∥ 2 + ( H 1 ( u ε ) , 1 ) + ( H 2 ( v ε ) , 1 ) − ( f 1 , u ε ) − ( f 2 , v ε ) . (45) E_{\varepsilon}(t) = \frac{1}{2}\left[\lVert u_{\varepsilon,t}\rVert^2 + \bar{\mu}_1\lVert u_{\varepsilon,xx}\rVert^2 + \lVert\sigma_{\varepsilon}\rVert_{\mathcal{M}_{g_{1,\varepsilon}}}^2 + \lVert v_{\varepsilon,t}\rVert^2 + \bar{\mu}_2\lVert v_{\varepsilon,x}\rVert^2 + \lVert\eta_{\varepsilon}\rVert_{\mathcal{M}_{g_{2,\varepsilon}}}^2\right]
+ \frac{k}{2}\lVert[u_{\varepsilon}-v_{\varepsilon}]^+\rVert^2 + (H_1(u_{\varepsilon}),1) + (H_2(v_{\varepsilon}),1) - (f_1,u_{\varepsilon}) - (f_2,v_{\varepsilon}). \tag*{(45)} E ε ( t ) = 2 1 [ ∥ u ε , t ∥ 2 + μ ˉ 1 ∥ u ε , xx ∥ 2 + ∥ σ ε ∥ M g 1 , ε 2 + ∥ v ε , t ∥ 2 + μ ˉ 2 ∥ v ε , x ∥ 2 + ∥ η ε ∥ M g 2 , ε 2 ] + 2 k ∥[ u ε − v ε ] + ∥ 2 + ( H 1 ( u ε ) , 1 ) + ( H 2 ( v ε ) , 1 ) − ( f 1 , u ε ) − ( f 2 , v ε ) . (45)
Here μ ˉ i = μ i − β i \bar{\mu}_i = \mu_i-\beta_i μ ˉ i = μ i − β i is independent of ε \varepsilon ε by the invariance of the mass (27) . Throughout, c e , c p > 0 c_e,c_p>0 c e , c p > 0 denote the Poincaré constants satisfying
∥ w ∥ ≤ c e ∥ w x x ∥ , w ∈ H a 2 ( 0 , ℓ ) , ∥ w ∥ ≤ c p ∥ w x ∥ , w ∈ H 0 1 ( 0 , ℓ ) , (46) \lVert w\rVert\le\sqrt{c_e}\lVert w_{xx}\rVert,\quad w\in H_a^2(0,\ell),\qquad\lVert w\rVert\le\sqrt{c_p}\lVert w_x\rVert,\quad w\in H_0^1(0,\ell), \tag*{(46)} ∥ w ∥ ≤ c e ∥ w xx ∥ , w ∈ H a 2 ( 0 , ℓ ) , ∥ w ∥ ≤ c p ∥ w x ∥ , w ∈ H 0 1 ( 0 , ℓ ) , (46)
and C e > 0 C_e>0 C e > 0 denotes the embedding constant of H a 2 ( 0 , ℓ ) ↪ L ∞ ( 0 , ℓ ) H_a^2(0,\ell)\hookrightarrow L^\infty(0,\ell) H a 2 ( 0 , ℓ ) ↪ L ∞ ( 0 , ℓ ) and H 0 1 ( 0 , ℓ ) ↪ L ∞ ( 0 , ℓ ) H_0^1(0,\ell)\hookrightarrow L^\infty(0,\ell) H 0 1 ( 0 , ℓ ) ↪ L ∞ ( 0 , ℓ ) , so that
∥ w ∥ L ∞ ≤ C e ∥ w x x ∥ , w ∈ H a 2 ( 0 , ℓ ) , ∥ w ∥ L ∞ ≤ C e ∥ w x ∥ , w ∈ H 0 1 ( 0 , ℓ ) . (47) \lVert w\rVert_{L^\infty}\le C_e\lVert w_{xx}\rVert,\quad w\in H_a^2(0,\ell),\qquad\lVert w\rVert_{L^\infty}\le C_e\lVert w_x\rVert,\quad w\in H_0^1(0,\ell). \tag*{(47)} ∥ w ∥ L ∞ ≤ C e ∥ w xx ∥ , w ∈ H a 2 ( 0 , ℓ ) , ∥ w ∥ L ∞ ≤ C e ∥ w x ∥ , w ∈ H 0 1 ( 0 , ℓ ) . (47)
Step 1: Dissipativity of E ε E_{\varepsilon} E ε . Multiplying the first equation of the perturbed system by u ε , t u_{\varepsilon,t} u ε , t in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) and the second by v ε , t v_{\varepsilon,t} v ε , t , and using the transport equations ∂ t σ ε + ∂ s σ ε − u ε , t = 0 \partial_t\sigma_{\varepsilon}+\partial_s\sigma_{\varepsilon}-u_{\varepsilon,t}=0 ∂ t σ ε + ∂ s σ ε − u ε , t = 0 and ∂ t η ε + ∂ s η ε − v ε , t = 0 \partial_t\eta_{\varepsilon}+\partial_s\eta_{\varepsilon}-v_{\varepsilon,t}=0 ∂ t η ε + ∂ s η ε − v ε , t = 0 , we obtain
d d t E ε ( t ) = − γ 1 ∥ u ε , t ∥ 2 − γ 2 ∥ v ε , t ∥ 2 + 1 2 ∫ 0 + ∞ g 1 , ε ′ ( s ) ∥ σ ε , x x ( s ) ∥ 2 d s + 1 2 ∫ 0 + ∞ g 2 , ε ′ ( s ) ∥ η ε , x ( s ) ∥ 2 d s . (48) \frac{\mathrm{d}}{\mathrm{d}t}E_{\varepsilon}(t)=-\gamma_1\lVert u_{\varepsilon,t}\rVert^2-\gamma_2\lVert v_{\varepsilon,t}\rVert^2+\frac{1}{2}\int_0^{+\infty}g'_{1,\varepsilon}(s)\lVert\sigma_{\varepsilon,xx}(s)\rVert^2\,\mathrm{d}s+\frac{1}{2}\int_0^{+\infty}g'_{2,\varepsilon}(s)\lVert\eta_{\varepsilon,x}(s)\rVert^2\,\mathrm{d}s. \tag*{(48)} d t d E ε ( t ) = − γ 1 ∥ u ε , t ∥ 2 − γ 2 ∥ v ε , t ∥ 2 + 2 1 ∫ 0 + ∞ g 1 , ε ′ ( s ) ∥ σ ε , xx ( s ) ∥ 2 d s + 2 1 ∫ 0 + ∞ g 2 , ε ′ ( s ) ∥ η ε , x ( s ) ∥ 2 d s . (48)
Indeed, the contribution of the memory terms to d d t E ε \frac{\mathrm{d}}{\mathrm{d}t}E_{\varepsilon} d t d E ε is obtained by differentiating 1 2 ∥ σ ε ∥ M g 1 , ε 2 = 1 2 ∫ 0 + ∞ g 1 , ε ( s ) ∥ σ ε , x x ( s ) ∥ 2 d s \frac{1}{2}\lVert\sigma_{\varepsilon}\rVert_{\mathcal{M}_{g_{1,\varepsilon}}}^2 = \frac{1}{2}\int_0^{+\infty}g_{1,\varepsilon}(s)\lVert\sigma_{\varepsilon,xx}(s)\rVert^2\,\mathrm{d}s 2 1 ∥ σ ε ∥ M g 1 , ε 2 = 2 1 ∫ 0 + ∞ g 1 , ε ( s ) ∥ σ ε , xx ( s ) ∥ 2 d s with respect to t t t , using the transport equation to replace ∂ t σ ε \partial_t\sigma_{\varepsilon} ∂ t σ ε by u ε , t − ∂ s σ ε u_{\varepsilon,t}-\partial_s\sigma_{\varepsilon} u ε , t − ∂ s σ ε , integrating by parts in s s s , and using σ ε ( x , t , 0 ) = 0 \sigma_{\varepsilon}(x,t,0)=0 σ ε ( x , t , 0 ) = 0 and g 1 , ε ( + ∞ ) = 0 g_{1,\varepsilon}(+\infty)=0 g 1 , ε ( + ∞ ) = 0 ; the resulting term − ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , x x ( s ) , u ε , x x , t ) d s -\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{\varepsilon,xx}(s),u_{\varepsilon,xx,t})\,\mathrm{d}s − ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , xx ( s ) , u ε , xx , t ) d s then cancels against the corresponding term + ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , x x ( s ) , u ε , x x , t ) d s +\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{\varepsilon,xx}(s),u_{\varepsilon,xx,t})\,\mathrm{d}s + ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , xx ( s ) , u ε , xx , t ) d s produced by differentiating the memory energy. The same reasoning applies to η ε \eta_{\varepsilon} η ε . By the differential inequality g i , ε ′ ( s ) + λ i g i , ε ( s ) ≤ 0 g'_{i,\varepsilon}(s)+\lambda_i g_{i,\varepsilon}(s)\le0 g i , ε ′ ( s ) + λ i g i , ε ( s ) ≤ 0 for a.e. s ≥ 0 s\ge0 s ≥ 0 , with the same λ i \lambda_i λ i as for g i g_i g i , we deduce
d d t E ε ( t ) ≤ − γ 1 ∥ u ε , t ∥ 2 − γ 2 ∥ v ε , t ∥ 2 − λ 1 2 ∥ σ ε ∥ M g 1 , ε 2 − λ 2 2 ∥ η ε ∥ M g 2 , ε 2 ≤ 0. (49) \frac{\mathrm{d}}{\mathrm{d}t}E_{\varepsilon}(t)\le-\gamma_1\lVert u_{\varepsilon,t}\rVert^2-\gamma_2\lVert v_{\varepsilon,t}\rVert^2-\frac{\lambda_1}{2}\lVert\sigma_{\varepsilon}\rVert_{\mathcal{M}_{g_{1,\varepsilon}}}^2-\frac{\lambda_2}{2}\lVert\eta_{\varepsilon}\rVert_{\mathcal{M}_{g_{2,\varepsilon}}}^2\le0. \tag*{(49)} d t d E ε ( t ) ≤ − γ 1 ∥ u ε , t ∥ 2 − γ 2 ∥ v ε , t ∥ 2 − 2 λ 1 ∥ σ ε ∥ M g 1 , ε 2 − 2 λ 2 ∥ η ε ∥ M g 2 , ε 2 ≤ 0. (49)
In particular, E ε ( t ) ≤ E ε ( 0 ) E_{\varepsilon}(t)\le E_{\varepsilon}(0) E ε ( t ) ≤ E ε ( 0 ) for all t ≥ 0 t\ge0 t ≥ 0 .
Step 2: Modified energy. Define
E ε , ω ( t ) : = E ε ( t ) + ω ( u ε , u ε , t ) + ω ( v ε , v ε , t ) , (50) E_{\varepsilon,\omega}(t):=E_{\varepsilon}(t)+\omega(u_{\varepsilon},u_{\varepsilon,t})+\omega(v_{\varepsilon},v_{\varepsilon,t}), \tag*{(50)} E ε , ω ( t ) := E ε ( t ) + ω ( u ε , u ε , t ) + ω ( v ε , v ε , t ) , (50)
where ω > 0 \omega>0 ω > 0 is a parameter to be chosen below, independent of ε \varepsilon ε . Multiplying the first equation of the perturbed system by u ε u_{\varepsilon} u ε in L 2 ( 0 , ℓ ) L^2(0,\ell) L 2 ( 0 , ℓ ) and the second by v ε v_{\varepsilon} v ε , and adding the results to (48) , we obtain
d d t E ε , ω ( t ) + ( γ 1 − ω ) ∥ u ε , t ∥ 2 + ( γ 2 − ω ) ∥ v ε , t ∥ 2 + ω μ ˉ 1 ∥ u ε , x x ∥ 2 + ω μ ˉ 2 ∥ v ε , x ∥ 2 + ω γ 1 ( u ε , u ε , t ) + ω γ 2 ( v ε , v ε , t ) + ω k ∥ [ u ε − v ε ] + ∥ 2 + λ 1 2 ∥ σ ε ∥ M g 1 , ε 2 + λ 2 2 ∥ η ε ∥ M g 2 , ε 2 + ω ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , x x ( s ) , u ε , x x ) d s + ω ∫ 0 + ∞ g 2 , ε ( s ) ( η ε , x ( s ) , v ε , x ) d s − ω ( h 1 ( u ε ) , u ε ) − ω ( h 2 ( v ε ) , v ε ) − ω ( f 1 , u ε ) − ω ( f 2 , v ε ) ≤ 0. (51) \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}E_{\varepsilon,\omega}(t)&+(\gamma_1-\omega)\lVert u_{\varepsilon,t}\rVert^2+(\gamma_2-\omega)\lVert v_{\varepsilon,t}\rVert^2+\omega\bar{\mu}_1\lVert u_{\varepsilon,xx}\rVert^2+\omega\bar{\mu}_2\lVert v_{\varepsilon,x}\rVert^2\\
&+\omega\gamma_1(u_{\varepsilon},u_{\varepsilon,t})+\omega\gamma_2(v_{\varepsilon},v_{\varepsilon,t})+\omega k\lVert[u_{\varepsilon}-v_{\varepsilon}]^+\rVert^2\\
&+\frac{\lambda_1}{2}\lVert\sigma_{\varepsilon}\rVert_{\mathcal{M}_{g_{1,\varepsilon}}}^2+\frac{\lambda_2}{2}\lVert\eta_{\varepsilon}\rVert_{\mathcal{M}_{g_{2,\varepsilon}}}^2\\
&+\omega\int_0^{+\infty}g_{1,\varepsilon}(s)(\sigma_{\varepsilon,xx}(s),u_{\varepsilon,xx})\,\mathrm{d}s+\omega\int_0^{+\infty}g_{2,\varepsilon}(s)(\eta_{\varepsilon,x}(s),v_{\varepsilon,x})\,\mathrm{d}s\\
&-\omega(h_1(u_{\varepsilon}),u_{\varepsilon})-\omega(h_2(v_{\varepsilon}),v_{\varepsilon})-\omega(f_1,u_{\varepsilon})-\omega(f_2,v_{\varepsilon})\le0. \tag*{(51)}
\end{aligned} d t d E ε , ω ( t ) + ( γ 1 − ω ) ∥ u ε , t ∥ 2 + ( γ 2 − ω ) ∥ v ε , t ∥ 2 + ω μ ˉ 1 ∥ u ε , xx ∥ 2 + ω μ ˉ 2 ∥ v ε , x ∥ 2 + ω γ 1 ( u ε , u ε , t ) + ω γ 2 ( v ε , v ε , t ) + ω k ∥[ u ε − v ε ] + ∥ 2 + 2 λ 1 ∥ σ ε ∥ M g 1 , ε 2 + 2 λ 2 ∥ η ε ∥ M g 2 , ε 2 + ω ∫ 0 + ∞ g 1 , ε ( s ) ( σ ε , xx ( s ) , u ε , xx ) d s + ω ∫ 0 + ∞ g 2 , ε ( s ) ( η ε , x ( s ) , v ε , x ) d s − ω ( h 1 ( u ε ) , u ε ) − ω ( h 2 ( v ε ) , v ε ) − ω ( f 1 , u ε ) − ω ( f 2 , v ε ) ≤ 0. (51)
Note the minus sign in front of ω ( h i ( ⋅ ) , ⋅ ) \omega(h_i(\cdot),\cdot) ω ( h i ( ⋅ ) , ⋅ ) : it comes from substituting u t t u_{tt} u tt and v t t v_{tt} v tt from the equations into ω ( u , u t t ) \omega(u,u_{tt}) ω ( u , u tt ) and ω ( v , v t t ) \omega(v,v_{tt}) ω ( v , v tt ) .
Using the growth condition (12) on H i H_i H i and h i h_i h i , there exists C 1 > 0 C_1>0 C 1 > 0 (the same constant as in (11) ), independent of ϵ \epsilon ϵ , such that
− ω ( h 1 ( u ϵ ) , u ϵ ) − ω ( h 2 ( v ϵ ) , v ϵ ) ≤ − ω C 1 ( H 1 ( u ϵ ) , 1 ) − ω C 1 ( H 2 ( v ϵ ) , 1 ) + ω ϵ ∥ u ϵ ∥ 2 + ω ϵ ∥ v ϵ ∥ 2 + ω C ϵ ′ , (52) -\omega(h_1(u_\epsilon),u_\epsilon)-\omega(h_2(v_\epsilon),v_\epsilon)\leq-\omega C_1(H_1(u_\epsilon),1)-\omega C_1(H_2(v_\epsilon),1)+\omega\epsilon\lVert u_\epsilon\rVert^2+\omega\epsilon\lVert v_\epsilon\rVert^2+\omega C'_\epsilon, \tag*{(52)} − ω ( h 1 ( u ϵ ) , u ϵ ) − ω ( h 2 ( v ϵ ) , v ϵ ) ≤ − ω C 1 ( H 1 ( u ϵ ) , 1 ) − ω C 1 ( H 2 ( v ϵ ) , 1 ) + ω ϵ ∥ u ϵ ∥ 2 + ω ϵ ∥ v ϵ ∥ 2 + ω C ϵ ′ , (52)
where C ϵ ′ C'_\epsilon C ϵ ′ is the constant appearing in the second inequality of (12) . By Poincaré’s inequality (46) ,
ω ϵ ∥ u ϵ ∥ 2 ≤ ω ϵ c e ∥ u ϵ , x x ∥ 2 , ω ϵ ∥ v ϵ ∥ 2 ≤ ω ϵ c p ∥ v ϵ , x ∥ 2 . \omega\epsilon\lVert u_\epsilon\rVert^2\leq\omega\epsilon c_e\lVert u_{\epsilon,xx}\rVert^2,\qquad\omega\epsilon\lVert v_\epsilon\rVert^2\leq\omega\epsilon c_p\lVert v_{\epsilon,x}\rVert^2. ω ϵ ∥ u ϵ ∥ 2 ≤ ω ϵ c e ∥ u ϵ , xx ∥ 2 , ω ϵ ∥ v ϵ ∥ 2 ≤ ω ϵ c p ∥ v ϵ , x ∥ 2 .
Similarly, by the Cauchy–Schwarz and Young inequalities,
∣ ∫ 0 + ∞ g 1 , ϵ ( s ) ( σ ϵ , x x ( s ) , u ϵ , x x ) d s ∣ ≤ c δ 1 ∥ σ ϵ ∥ M g 1 , ϵ 2 + δ 1 μ ˉ 1 ∥ u ϵ , x x ∥ 2 , (53) \left\lvert\int_0^{+\infty}g_{1,\epsilon}(s)(\sigma_{\epsilon,xx}(s),u_{\epsilon,xx})\,\mathrm{d}s\right\rvert\leq c_{\delta_1}\lVert\sigma_\epsilon\rVert^2_{M_{g_{1,\epsilon}}}+\delta_1\bar{\mu}_1\lVert u_{\epsilon,xx}\rVert^2, \tag*{(53)} ∫ 0 + ∞ g 1 , ϵ ( s ) ( σ ϵ , xx ( s ) , u ϵ , xx ) d s ≤ c δ 1 ∥ σ ϵ ∥ M g 1 , ϵ 2 + δ 1 μ ˉ 1 ∥ u ϵ , xx ∥ 2 , (53)
and the analogous estimate for η ϵ \eta_\epsilon η ϵ and v ϵ v_\epsilon v ϵ reads
∣ ∫ 0 + ∞ g 2 , ϵ ( s ) ( η ϵ , x ( s ) , v ϵ , x ) d s ∣ ≤ c δ 1 ′ ∥ η ϵ ∥ M g 2 , ϵ 2 + δ 1 μ ˉ 2 ∥ v ϵ , x ∥ 2 , (54) \left\lvert\int_0^{+\infty}g_{2,\epsilon}(s)(\eta_{\epsilon,x}(s),v_{\epsilon,x})\,\mathrm{d}s\right\rvert\leq c'_{\delta_1}\lVert\eta_\epsilon\rVert^2_{M_{g_{2,\epsilon}}}+\delta_1\bar{\mu}_2\lVert v_{\epsilon,x}\rVert^2, \tag*{(54)} ∫ 0 + ∞ g 2 , ϵ ( s ) ( η ϵ , x ( s ) , v ϵ , x ) d s ≤ c δ 1 ′ ∥ η ϵ ∥ M g 2 , ϵ 2 + δ 1 μ ˉ 2 ∥ v ϵ , x ∥ 2 , (54)
where
c δ 1 = β 1 4 δ 1 μ ˉ 1 , c δ 1 ′ = β 2 4 δ 1 μ ˉ 2 , (55) c_{\delta_1}=\frac{\beta_1}{4\delta_1\bar{\mu}_1},\qquad c'_{\delta_1}=\frac{\beta_2}{4\delta_1\bar{\mu}_2}, \tag*{(55)} c δ 1 = 4 δ 1 μ ˉ 1 β 1 , c δ 1 ′ = 4 δ 1 μ ˉ 2 β 2 , (55)
and δ 1 > 0 \delta_1>0 δ 1 > 0 is arbitrary. Note that c δ 1 c_{\delta_1} c δ 1 and c δ 1 ′ c'_{\delta_1} c δ 1 ′ depend only on β i \beta_i β i , δ 1 \delta_1 δ 1 , and μ ˉ i \bar{\mu}_i μ ˉ i , and are independent of ϵ \epsilon ϵ .
Step 3: Choice of parameters. We choose the parameters in the following order, all independently of ϵ \epsilon ϵ . First, set δ 1 = 1 / 4 \delta_1=1/4 δ 1 = 1/4 . Second, choose ϵ > 0 \epsilon>0 ϵ > 0 so small that
ϵ ≤ min { μ ˉ 1 8 c e , μ ˉ 2 8 c p } . (56) \epsilon\leq\min\left\{\frac{\bar{\mu}_1}{8c_e},\frac{\bar{\mu}_2}{8c_p}\right\}. \tag*{(56)} ϵ ≤ min { 8 c e μ ˉ 1 , 8 c p μ ˉ 2 } . (56)
Third, choose ω > 0 \omega>0 ω > 0 so small that
ω ≤ min { γ 1 2 , γ 2 2 , λ 1 4 c δ 1 , λ 2 4 c δ 1 ′ , μ ˉ 1 2 , μ ˉ 2 2 , 1 4 c e , 1 4 c p , 1 2 , μ ˉ 1 4 c e , μ ˉ 2 4 c p , 1 2 C 1 } . (57) \omega\leq\min\left\{\frac{\gamma_1}{2},\frac{\gamma_2}{2},\frac{\lambda_1}{4c_{\delta_1}},\frac{\lambda_2}{4c'_{\delta_1}},\frac{\bar{\mu}_1}{2},\frac{\bar{\mu}_2}{2},\frac{1}{4c_e},\frac{1}{4c_p},\frac{1}{2},\frac{\bar{\mu}_1}{4c_e},\frac{\bar{\mu}_2}{4c_p},\frac{1}{2C_1}\right\}. \tag*{(57)} ω ≤ min { 2 γ 1 , 2 γ 2 , 4 c δ 1 λ 1 , 4 c δ 1 ′ λ 2 , 2 μ ˉ 1 , 2 μ ˉ 2 , 4 c e 1 , 4 c p 1 , 2 1 , 4 c e μ ˉ 1 , 4 c p μ ˉ 2 , 2 C 1 1 } . (57)
All the quantities appearing in (56) –(57) —γ i \gamma_i γ i , λ i \lambda_i λ i , μ ˉ i \bar{\mu}_i μ ˉ i , c e c_e c e , c p c_p c p , c δ 1 c_{\delta_1} c δ 1 , c δ 1 ′ c'_{\delta_1} c δ 1 ′ , C 1 C_1 C 1 —are independent of ϵ \epsilon ϵ . In particular, the last bound ω ≤ 1 / ( 2 C 1 ) \omega\leq1/(2C_1) ω ≤ 1/ ( 2 C 1 ) guarantees 1 − ω C 1 ≥ 1 / 2 > 0 1-\omega C_1\geq1/2>0 1 − ω C 1 ≥ 1/2 > 0 , which is needed when the terms − ω C 1 ( H i ( u i ) , 1 ) -\omega C_1(H_i(u_i),1) − ω C 1 ( H i ( u i ) , 1 ) are combined with the terms ( H i ( u i ) , 1 ) (H_i(u_i),1) ( H i ( u i ) , 1 ) present in E ϵ , ω E_{\epsilon,\omega} E ϵ , ω . Substituting the above estimates into (51) and absorbing the ϵ \epsilon ϵ -terms and the memory terms with the help of (56) and (57) , we arrive at
d d t E ϵ , ω ( t ) + c 0 E ϵ , ω ( t ) ≤ c 1 , (58) \frac{\mathrm{d}}{\mathrm{d}t}E_{\epsilon,\omega}(t)+c_0E_{\epsilon,\omega}(t)\leq c_1, \tag*{(58)} d t d E ϵ , ω ( t ) + c 0 E ϵ , ω ( t ) ≤ c 1 , (58)
where c 0 > 0 c_0>0 c 0 > 0 and c 1 ≥ 0 c_1\geq0 c 1 ≥ 0 are constants depending only on γ i \gamma_i γ i , λ i \lambda_i λ i , μ i \mu_i μ i , β i \beta_i β i , c e c_e c e , c p c_p c p , ∥ f 1 ∥ \lVert f_1\rVert ∥ f 1 ∥ , ∥ f 2 ∥ \lVert f_2\rVert ∥ f 2 ∥ , and the constants in (11) , but not on ϵ \epsilon ϵ or on the initial data.
Step 4: Equivalence of E ϵ , ω E_{\epsilon,\omega} E ϵ , ω and ∥ ⋅ ∥ H ϵ 2 \lVert\cdot\rVert^2_{\mathcal{H}_\epsilon} ∥ ⋅ ∥ H ϵ 2 . Let R B > 0 R_B>0 R B > 0 be the bound on the initial data appearing in Step 5, i.e., ∥ U 0 , ϵ ∥ H ϵ ≤ R B \lVert U_{0,\epsilon}\rVert_{\mathcal{H}_\epsilon}\leq R_B ∥ U 0 , ϵ ∥ H ϵ ≤ R B for all U 0 , ϵ ∈ B U_{0,\epsilon}\in B U 0 , ϵ ∈ B . We claim that there exist constants c 2 , c 3 > 0 c_2,c_3>0 c 2 , c 3 > 0 , depending on R B R_B R B but not on ϵ \epsilon ϵ , such that for every t ≥ 0 t\geq0 t ≥ 0 with ∥ U ϵ ( t ) ∥ H ϵ ≤ R B + 1 \lVert U_\epsilon(t)\rVert_{\mathcal{H}_\epsilon}\leq R_B+1 ∥ U ϵ ( t ) ∥ H ϵ ≤ R B + 1 ,
c 2 ( ∥ U ϵ ( t ) ∥ H ϵ 2 − 1 ) ≤ E ϵ , ω ( t ) ≤ c 3 ( ∥ U ϵ ( t ) ∥ H ϵ 2 + 1 ) . (59) c_2\left(\lVert U_\epsilon(t)\rVert^2_{\mathcal{H}_\epsilon}-1\right)\leq E_{\epsilon,\omega}(t)\leq c_3\left(\lVert U_\epsilon(t)\rVert^2_{\mathcal{H}_\epsilon}+1\right). \tag*{(59)} c 2 ( ∥ U ϵ ( t ) ∥ H ϵ 2 − 1 ) ≤ E ϵ , ω ( t ) ≤ c 3 ( ∥ U ϵ ( t ) ∥ H ϵ 2 + 1 ) . (59)
For the lower bound, we estimate each term in E ϵ , ω E_{\epsilon,\omega} E ϵ , ω as follows:
ω ( u ϵ , u ϵ , t ) + ω ( v ϵ , v ϵ , t ) ≥ ω 2 ∥ u ϵ , t ∥ 2 − ω c e 2 ∥ u ϵ , x x ∥ 2 − ω 2 ∥ v ϵ , t ∥ 2 − ω c p 2 ∥ v ϵ , x ∥ 2 . \omega(u_\epsilon,u_{\epsilon,t})+\omega(v_\epsilon,v_{\epsilon,t})\geq\frac{\omega}{2}\lVert u_{\epsilon,t}\rVert^2-\frac{\omega c_e}{2}\lVert u_{\epsilon,xx}\rVert^2-\frac{\omega}{2}\lVert v_{\epsilon,t}\rVert^2-\frac{\omega c_p}{2}\lVert v_{\epsilon,x}\rVert^2. ω ( u ϵ , u ϵ , t ) + ω ( v ϵ , v ϵ , t ) ≥ 2 ω ∥ u ϵ , t ∥ 2 − 2 ω c e ∥ u ϵ , xx ∥ 2 − 2 ω ∥ v ϵ , t ∥ 2 − 2 ω c p ∥ v ϵ , x ∥ 2 .
and, by (12) and Young’s inequality,
( H 1 ( u ε ) , 1 ) + ( H 2 ( v ε ) , 1 ) − ( f 1 , u ε ) − ( f 2 , v ε ) ≥ − ε c e ∥ u ε , x x ∥ 2 − ε c p ∥ v ε , x ∥ 2 − c ε , (60) (H_1(u_\varepsilon),1)+(H_2(v_\varepsilon),1)-(f_1,u_\varepsilon)-(f_2,v_\varepsilon)\ge-\varepsilon c_e\lVert u_{\varepsilon,xx}\rVert^2-\varepsilon c_p\lVert v_{\varepsilon,x}\rVert^2-c_\varepsilon, \tag*{(60)} ( H 1 ( u ε ) , 1 ) + ( H 2 ( v ε ) , 1 ) − ( f 1 , u ε ) − ( f 2 , v ε ) ≥ − ε c e ∥ u ε , xx ∥ 2 − ε c p ∥ v ε , x ∥ 2 − c ε , (60)
where c ε c_\varepsilon c ε depends only on ε \varepsilon ε , ∥ f 1 ∥ \lVert f_1\rVert ∥ f 1 ∥ , ∥ f 2 ∥ \lVert f_2\rVert ∥ f 2 ∥ , and the constants in (12) . The terms produced here are the combined contribution of the H i H_i H i -terms (via (12) ) and of Young’s inequality applied to the f i f_i f i -terms: the former yields − ε ∥ u ε ∥ 2 -\varepsilon\lVert u_\varepsilon\rVert^2 − ε ∥ u ε ∥ 2 , the latter − ε 2 ∥ u ε ∥ 2 -\frac{\varepsilon}{2}\lVert u_\varepsilon\rVert^2 − 2 ε ∥ u ε ∥ 2 , so that together they give − 3 ε c e 2 ∥ u ε , x x ∥ 2 − C ε -\frac{3\varepsilon c_e}{2}\lVert u_{\varepsilon,xx}\rVert^2-C_\varepsilon − 2 3 ε c e ∥ u ε , xx ∥ 2 − C ε (and analogously for v ε v_\varepsilon v ε ). These are distinct from the analogous terms − ω ε c e ∥ u ε , x x ∥ 2 -\omega\varepsilon c_e\lVert u_{\varepsilon,xx}\rVert^2 − ω ε c e ∥ u ε , xx ∥ 2 and − ω ε c p ∥ v ε , x ∥ 2 -\omega\varepsilon c_p\lVert v_{\varepsilon,x}\rVert^2 − ω ε c p ∥ v ε , x ∥ 2 produced in Step 2 by the h i h_i h i -terms.
Combining the above bounds with (45) and (50) , we see that the coefficients of ∥ u ε , t ∥ 2 \lVert u_{\varepsilon,t}\rVert^2 ∥ u ε , t ∥ 2 , ∥ u ε , x x ∥ 2 \lVert u_{\varepsilon,xx}\rVert^2 ∥ u ε , xx ∥ 2 , ∥ v ε , t ∥ 2 \lVert v_{\varepsilon,t}\rVert^2 ∥ v ε , t ∥ 2 , ∥ v ε , x ∥ 2 \lVert v_{\varepsilon,x}\rVert^2 ∥ v ε , x ∥ 2 in the resulting lower bound for E ε , ω ( t ) E_{\varepsilon,\omega}(t) E ε , ω ( t ) are bounded below by
1 2 − ω 2 , μ ˉ 1 2 − ω c e 2 − 3 ε c e 2 − ω ε c e , μ ˉ 2 2 − ω c p 2 − 3 ε c p 2 − ω ε c p . (61) \frac{1}{2}-\frac{\omega}{2},\qquad\frac{\bar{\mu}_1}{2}-\frac{\omega c_e}{2}-\frac{3\varepsilon c_e}{2}-\omega\varepsilon c_e,\qquad\frac{\bar{\mu}_2}{2}-\frac{\omega c_p}{2}-\frac{3\varepsilon c_p}{2}-\omega\varepsilon c_p. \tag*{(61)} 2 1 − 2 ω , 2 μ ˉ 1 − 2 ω c e − 2 3 ε c e − ω ε c e , 2 μ ˉ 2 − 2 ω c p − 2 3 ε c p − ω ε c p . (61)
The choices (56) –(57) already ensure that all three quantities are strictly positive. Indeed, since ω ≤ 1 / 2 \omega\le1/2 ω ≤ 1/2 , the first is at least 1 / 4 1/4 1/4 . For the second, using ω c e ≤ μ ˉ 1 / 4 \omega c_e\le\bar{\mu}_1/4 ω c e ≤ μ ˉ 1 /4 , ε c e ≤ μ ˉ 1 / 8 \varepsilon c_e\le\bar{\mu}_1/8 ε c e ≤ μ ˉ 1 /8 , and ω ε c e ≤ μ ˉ 1 / 16 \omega\varepsilon c_e\le\bar{\mu}_1/16 ω ε c e ≤ μ ˉ 1 /16 (which follows from ω ≤ 1 / 2 \omega\le1/2 ω ≤ 1/2 and ε c e ≤ μ ˉ 1 / 8 \varepsilon c_e\le\bar{\mu}_1/8 ε c e ≤ μ ˉ 1 /8 ), we get
μ ˉ 1 2 − ω c e 2 − 3 ε c e 2 − ω ε c e ≥ μ ˉ 1 2 − μ ˉ 1 8 − 3 μ ˉ 1 16 − μ ˉ 1 16 = μ ˉ 1 8 > 0 , \frac{\bar{\mu}_1}{2}-\frac{\omega c_e}{2}-\frac{3\varepsilon c_e}{2}-\omega\varepsilon c_e\ge\frac{\bar{\mu}_1}{2}-\frac{\bar{\mu}_1}{8}-\frac{3\bar{\mu}_1}{16}-\frac{\bar{\mu}_1}{16}=\frac{\bar{\mu}_1}{8}>0, 2 μ ˉ 1 − 2 ω c e − 2 3 ε c e − ω ε c e ≥ 2 μ ˉ 1 − 8 μ ˉ 1 − 16 3 μ ˉ 1 − 16 μ ˉ 1 = 8 μ ˉ 1 > 0 ,
and similarly for the v v v -coefficient. This yields the lower bound in (59) , the constant term − 1 -1 − 1 on the left-hand side absorbing c ε c_\varepsilon c ε , ω C ε ′ \omega C'_\varepsilon ω C ε ′ , and the ∥ f i ∥ 2 \lVert f_i\rVert^2 ∥ f i ∥ 2 -terms produced by Young’s inequality. All these constants depend only on ε \varepsilon ε , ω \omega ω , ∥ f 1 ∥ \lVert f_1\rVert ∥ f 1 ∥ , ∥ f 2 ∥ \lVert f_2\rVert ∥ f 2 ∥ , and the constants in (12) , but not on ε \varepsilon ε .
For the upper bound, we argue as follows. Fix any t ≥ 0 t\ge0 t ≥ 0 with ∥ U ε ( t ) ∥ H ε ≤ R B + 1 \lVert U_\varepsilon(t)\rVert_{H_\varepsilon}\le R_B+1 ∥ U ε ( t ) ∥ H ε ≤ R B + 1 . By the embedding (47) and the definition (22) of ∥ ⋅ ∥ H ε ‘ , t h e r e e x i s t s a c o n s t a n t \lVert\cdot\rVert_{H_\varepsilon}`, there exists a constant ∥ ⋅ ∥ H ε ‘ , t h er ee x i s t s a co n s t an t C'_e>0$, depending only on C e C_e C e and μ ˉ i \bar{\mu}_i μ ˉ i , such that
∥ u ε ( t ) ∥ L ∞ ≤ C e ′ ∥ u ε , x x ( t ) ∥ ≤ C e ′ ∥ U ε ( t ) ∥ H ε ≤ C e ′ ( R B + 1 ) , \lVert u_\varepsilon(t)\rVert_{L^\infty}\le C'_e\lVert u_{\varepsilon,xx}(t)\rVert\le C'_e\lVert U_\varepsilon(t)\rVert_{H_\varepsilon}\le C'_e(R_B+1), ∥ u ε ( t ) ∥ L ∞ ≤ C e ′ ∥ u ε , xx ( t )∥ ≤ C e ′ ∥ U ε ( t ) ∥ H ε ≤ C e ′ ( R B + 1 ) ,
and similarly for v ε ( t ) v_\varepsilon(t) v ε ( t ) . Since h i h_i h i is continuous and H i H_i H i is its primitive, both functions are bounded on the compact interval [ − C e ′ ( R B + 1 ) , C e ′ ( R B + 1 ) ] [-C'_e(R_B+1),C'_e(R_B+1)] [ − C e ′ ( R B + 1 ) , C e ′ ( R B + 1 )] by a constant C 3 = C 3 ( R B ) C_3=C_3(R_B) C 3 = C 3 ( R B ) depending only on R B R_B R B , C e ′ C'_e C e ′ , and the constants in (11) , but not on ε \varepsilon ε or t t t . Hence
∣ ( H i ( u ε ( t ) ) , 1 ) ∣ ≤ C 3 ( R B ) , ∣ ( h i ( u ε ( t ) ) , u ε ( t ) ) ∣ ≤ C 3 ( R B ) , i = 1 , 2. |(H_i(u_\varepsilon(t)),1)|\le C_3(R_B),\qquad|(h_i(u_\varepsilon(t)),u_\varepsilon(t))|\le C_3(R_B),\qquad i=1,2. ∣ ( H i ( u ε ( t )) , 1 ) ∣ ≤ C 3 ( R B ) , ∣ ( h i ( u ε ( t )) , u ε ( t )) ∣ ≤ C 3 ( R B ) , i = 1 , 2.
Combining this with (45) and (50) , and using the boundedness of u ε u_\varepsilon u ε , u ε , t u_{\varepsilon,t} u ε , t , v ε v_\varepsilon v ε , v ε , t v_{\varepsilon,t} v ε , t , σ ε \sigma_\varepsilon σ ε , η ε \eta_\varepsilon η ε in their respective norms (all controlled by ∥ U ε ( t ) ∥ H ε ≤ R B + 1 \lVert U_\varepsilon(t)\rVert_{H_\varepsilon}\le R_B+1 ∥ U ε ( t ) ∥ H ε ≤ R B + 1 ), we obtain the upper bound in (59) with a constant c 3 c_3 c 3 that depends on R B R_B R B , γ i \gamma_i γ i , μ i \mu_i μ i , β i \beta_i β i , c e c_e c e , c p c_p c p , and the constants in (11) , but not on ε \varepsilon ε .
Step 5: Gronwall and the absorbing ball. Applying Gronwall’s inequality to (58) , we obtain
E ε , ω ( t ) ≤ E ε , ω ( 0 ) e − c 0 t + c 1 c 0 ( 1 − e − c 0 t ) ≤ E ε , ω ( 0 ) e − c 0 t + c 1 c 0 , t ≥ 0. (62) E_{\varepsilon,\omega}(t)\le E_{\varepsilon,\omega}(0)e^{-c_0t}+\frac{c_1}{c_0}(1-e^{-c_0t})\le E_{\varepsilon,\omega}(0)e^{-c_0t}+\frac{c_1}{c_0},\qquad t\ge0. \tag*{(62)} E ε , ω ( t ) ≤ E ε , ω ( 0 ) e − c 0 t + c 0 c 1 ( 1 − e − c 0 t ) ≤ E ε , ω ( 0 ) e − c 0 t + c 0 c 1 , t ≥ 0. (62)
Combining (62) with the lower bound in (59) , we get
c 2 ( ∥ U ε ( t ) ∥ H ε 2 − 1 ) ≤ E ε , ω ( 0 ) e − c 0 t + c 1 c 0 . c_2(\lVert U_\varepsilon(t)\rVert_{H_\varepsilon}^2-1)\le E_{\varepsilon,\omega}(0)e^{-c_0t}+\frac{c_1}{c_0}. c 2 (∥ U ε ( t ) ∥ H ε 2 − 1 ) ≤ E ε , ω ( 0 ) e − c 0 t + c 0 c 1 .
By the upper bound in (59) applied at t = 0 t=0 t = 0 (where ∥ U 0 , ε ∥ H ε ≤ R B < R B + 1 \lVert U_{0,\varepsilon}\rVert_{H_\varepsilon}\le R_B<R_B+1 ∥ U 0 , ε ∥ H ε ≤ R B < R B + 1 ), we have E ε , ω ( 0 ) ≤ c 3 ( R B 2 + 1 ) E_{\varepsilon,\omega}(0)\le c_3(R_B^2+1) E ε , ω ( 0 ) ≤ c 3 ( R B 2 + 1 ) . Hence
∥ U ε ( t ) ∥ H ε 2 ≤ 1 + c 3 ( R B 2 + 1 ) c 2 e − c 0 t + c 1 c 0 c 2 . \lVert U_\varepsilon(t)\rVert_{H_\varepsilon}^2\le1+\frac{c_3(R_B^2+1)}{c_2}e^{-c_0t}+\frac{c_1}{c_0c_2}. ∥ U ε ( t ) ∥ H ε 2 ≤ 1 + c 2 c 3 ( R B 2 + 1 ) e − c 0 t + c 0 c 2 c 1 .
Set
R 0 : = ( 2 + c 1 c 0 c 2 ) 1 / 2 . R_0:=\left(2+\frac{c_1}{c_0c_2}\right)^{1/2}. R 0 := ( 2 + c 0 c 2 c 1 ) 1/2 .
Choose t 0 = t 0 ( B ) ≥ 0 t_0=t_0(B)\ge0 t 0 = t 0 ( B ) ≥ 0 , independent of ε \varepsilon ε , such that
e − c 0 t 0 ≤ 1 c 3 ( R B 2 + 1 ) / c 2 ; e^{-c_0t_0}\le\frac{1}{c_3(R_B^2+1)/c_2}; e − c 0 t 0 ≤ c 3 ( R B 2 + 1 ) / c 2 1 ;
for instance, it suffices to take
t 0 : = max { 0 , 1 c 0 log c 3 ( R B 2 + 1 ) c 2 } . t_0:=\max\left\{0,\frac{1}{c_0}\log\frac{c_3(R_B^2+1)}{c_2}\right\}. t 0 := max { 0 , c 0 1 log c 2 c 3 ( R B 2 + 1 ) } .
Then for every t ≥ t 0 t\ge t_0 t ≥ t 0 , using e − c 0 t ≤ e − c 0 t 0 e^{-c_0t}\le e^{-c_0t_0} e − c 0 t ≤ e − c 0 t 0 , we get
∥ U ε ( t ) ∥ H ε 2 ≤ 1 + c 3 ( R B 2 + 1 ) c 2 e − c 0 t 0 + c 1 c 0 c 2 ≤ 2 + c 1 c 0 c 2 = R 0 2 , \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}^2\le1+\frac{c_3(R_B^2+1)}{c_2}e^{-c_0t_0}+\frac{c_1}{c_0c_2}\le2+\frac{c_1}{c_0c_2}=R_0^2, ∥ U ε ( t ) ∥ H ε 2 ≤ 1 + c 2 c 3 ( R B 2 + 1 ) e − c 0 t 0 + c 0 c 2 c 1 ≤ 2 + c 0 c 2 c 1 = R 0 2 ,
where the second inequality uses the choice of t 0 t_0 t 0 ; if t 0 = 0 t_0=0 t 0 = 0 , then c 3 ( R B 2 + 1 ) c 2 ≤ 1 \frac{c_3(R_B^2+1)}{c_2}\le1 c 2 c 3 ( R B 2 + 1 ) ≤ 1 and the same inequality holds a fortiori. Therefore ∥ U ε ( t ) ∥ H ε ≤ R 0 \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le R_0 ∥ U ε ( t ) ∥ H ε ≤ R 0 for all t ≥ t 0 t\ge t_0 t ≥ t 0 , uniformly in U 0 , ε ∈ B U_{0,\varepsilon}\in B U 0 , ε ∈ B and ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . This proves the proposition.
Lemma 3.2 (Finite-time uniform boundedness). For every T > 0 T>0 T > 0 and every bounded set B ⊂ H ε B\subset\mathcal{H}_\varepsilon B ⊂ H ε , there exists R 1 = R 1 ( T , B ) > 0 R_1=R_1(T,B)>0 R 1 = R 1 ( T , B ) > 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that
sup ε ∈ ( 0 , 1 ] sup U 0 ∈ B sup 0 ≤ t ≤ T ∥ S ε ( t ) U 0 ∥ H ε ≤ R 1 . \sup_{\varepsilon\in(0,1]}\sup_{U_0\in B}\sup_{0\le t\le T}\|S_\varepsilon(t)U_0\|_{\mathcal{H}_\varepsilon}\le R_1. ε ∈ ( 0 , 1 ] sup U 0 ∈ B sup 0 ≤ t ≤ T sup ∥ S ε ( t ) U 0 ∥ H ε ≤ R 1 . Proof. Let U 0 , ε ∈ B U_{0,\varepsilon}\in B U 0 , ε ∈ B with ∥ U 0 , ε ∥ H ε ≤ R B \|U_{0,\varepsilon}\|_{\mathcal{H}_\varepsilon}\le R_B ∥ U 0 , ε ∥ H ε ≤ R B . By the mild formulation and Theorem 2.2,
∥ U ε ( t ) ∥ H ε ≤ α e − β t R B + α ∫ 0 t e − β ( t − s ) ∥ F ( U ε ( s ) ) ∥ H ε d s . \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le\alpha e^{-\beta t}R_B+\alpha\int_0^t e^{-\beta(t-s)}\|F(U_\varepsilon(s))\|_{\mathcal{H}_\varepsilon}\,ds. ∥ U ε ( t ) ∥ H ε ≤ α e − β t R B + α ∫ 0 t e − β ( t − s ) ∥ F ( U ε ( s )) ∥ H ε d s . Since F F F is locally Lipschitz and maps bounded sets into bounded sets, as long as ∥ U ε ( t ) ∥ H ε ≤ R \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le R ∥ U ε ( t ) ∥ H ε ≤ R we have ∥ F ( U ε ( t ) ) ∥ H ε ≤ L R R + ∥ F ( 0 ) ∥ H ε \|F(U_\varepsilon(t))\|_{\mathcal{H}_\varepsilon}\le L_RR+\|F(0)\|_{\mathcal{H}_\varepsilon} ∥ F ( U ε ( t )) ∥ H ε ≤ L R R + ∥ F ( 0 ) ∥ H ε . Multiplying by e β t e^{\beta t} e β t and applying Gronwall’s inequality, we obtain
∥ U ε ( t ) ∥ H ε ≤ ( α R B + α ∥ F ( 0 ) ∥ H ε β ) e α L R t , 0 ≤ t ≤ T . \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le\left(\alpha R_B+\frac{\alpha\|F(0)\|_{\mathcal{H}_\varepsilon}}{\beta}\right)e^{\alpha L_Rt},\qquad0\le t\le T. ∥ U ε ( t ) ∥ H ε ≤ ( α R B + β α ∥ F ( 0 ) ∥ H ε ) e α L R t , 0 ≤ t ≤ T . The right-hand side is independent of ε \varepsilon ε . A standard continuation argument yields a bound on [ 0 , T ] [0,T] [ 0 , T ] : let R > 0 R>0 R > 0 be a candidate radius and let L R L_R L R be the Lipschitz constant of F F F on the ball of radius R R R . As long as ∥ U ε ( t ) ∥ H ε ≤ R \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le R ∥ U ε ( t ) ∥ H ε ≤ R , the preceding estimate gives
∥ U ε ( t ) ∥ H ε ≤ ( α R B + α ∥ F ( 0 ) ∥ H ε β ) e α L R T . \|U_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le\left(\alpha R_B+\frac{\alpha\|F(0)\|_{\mathcal{H}_\varepsilon}}{\beta}\right)e^{\alpha L_RT}. ∥ U ε ( t ) ∥ H ε ≤ ( α R B + β α ∥ F ( 0 ) ∥ H ε ) e α L R T . Since L R L_R L R is nondecreasing in R R R , one can choose R > 0 R>0 R > 0 large enough so that the right-hand side is strictly less than R R R ; then the solution remains in the ball of radius R R R on [ 0 , T ] [0,T] [ 0 , T ] , hence is bounded on [ 0 , T ] [0,T] [ 0 , T ] by a constant R 1 R_1 R 1 depending only on T T T , B B B and the parameters, but not on ε \varepsilon ε . This proves the lemma.
Upper Semicontinuity
We now outline the proof of the upper semicontinuity (17) . The general strategy is standard (see, e.g., [5 Theorem 7.9.6] and [31 , 12 ] ):
(A1) Uniform dissipativity: the semigroups S ε ( t ) S_\varepsilon(t) S ε ( t ) possess a common bounded absorbing set B 0 B_0 B 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] (Proposition 3.1 ). (A2) Uniform asymptotic smoothness of the semigroups S ε ( t ) S_\varepsilon(t) S ε ( t ) , uniformly in ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . In the framework of Chueshov–Lasiecka, this is implied by the uniform quasi-stability inequality on B 0 \mathcal{B}_0 B 0 (Proposition 4.1 ).
(A3) Convergence of the perturbed semigroups to the limit semigroup on finite time intervals: for every T > 0 T>0 T > 0 and every sequence ε n → 0 + \varepsilon_n \to0^+ ε n → 0 + with U 0 , ε n → U 0 U_{0,\varepsilon_n} \to U_0 U 0 , ε n → U 0 ,
sup t ∈ [ 0 , T ] ∥ S ε n ( t ) U 0 , ε n − S 0 ( t ) U 0 ∥ H ε n → 0. (63) \sup_{t\in[0,T]}\left\|S_{\varepsilon_n}(t)U_{0,\varepsilon_n}-S_0(t)U_0\right\|_{\mathcal{H}_{\varepsilon_n}} \to0. \tag*{(63)} t ∈ [ 0 , T ] sup ∥ S ε n ( t ) U 0 , ε n − S 0 ( t ) U 0 ∥ H ε n → 0. (63)
Given (A1)–(A3), the upper semicontinuity ((17) ) follows by a standard compactness argument: any sequence { U ε n } \{U_{\varepsilon_n}\} { U ε n } with U ε n ∈ A ε n U_{\varepsilon_n}\in\mathcal{A}_{\varepsilon_n} U ε n ∈ A ε n is relatively compact, and any accumulation point belongs to A 0 \mathcal{A}_0 A 0 .
Uniform asymptotic smoothness
The uniform exponential stability ((26) ) of the linear semigroup, combined with the quasi-stability framework of Chueshov–Lasiecka [5 ] as developed in [33 ] , yields the following uniform version.
Proposition 4.1 (Uniform quasi-stability). There exist T > 0 T>0 T > 0 and q ∈ [ 0 , 1 ) q\in[0,1) q ∈ [ 0 , 1 ) , both independent of ε \varepsilon ε , such that for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] the following holds: there exists a map K ε , T : B 0 → H ε K_{\varepsilon,T}:\mathcal{B}_0\to\mathcal{H}_{\varepsilon} K ε , T : B 0 → H ε satisfying
∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , (64) \left\|S_{\varepsilon}(T)U_0-S_{\varepsilon}(T)V_0\right\|_{\mathcal{H}_{\varepsilon}}^2 \le q^2\left\|U_0-V_0\right\|_{\mathcal{H}_{\varepsilon}}^2+\left\|K_{\varepsilon,T}(U_0)-K_{\varepsilon,T}(V_0)\right\|_{\mathcal{H}_{\varepsilon}}^2, \tag*{(64)} ∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , (64) for all U 0 , V 0 ∈ B 0 U_0,V_0\in\mathcal{B}_0 U 0 , V 0 ∈ B 0 , and such that the following sequential compactness property holds: for every sequence ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + and every bounded sequence U n ∈ B 0 U_n\in\mathcal{B}_0 U n ∈ B 0 , there exist a subsequence { ε n k , U n k } \{\varepsilon_{n_k},U_{n_k}\} { ε n k , U n k } and an element K ∗ K_* K ∗ in the sense of Remark 2.1 such that
K ε n k , T ( U n k ) → K ∗ ( k → + ∞ ) (65) K_{\varepsilon_{n_k},T}(U_{n_k})\to K_*
\qquad(k\to+\infty) \tag*{(65)} K ε n k , T ( U n k ) → K ∗ ( k → + ∞ ) (65) in the uniform topology of Remark 2.1.
Proof. Fix ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . Let U 0 , V 0 ∈ B 0 U_0,V_0\in\mathcal{B}_0 U 0 , V 0 ∈ B 0 and set U ε ( t ) = S ε ( t ) U 0 U_\varepsilon(t)=S_\varepsilon(t)U_0 U ε ( t ) = S ε ( t ) U 0 , V ε ( t ) = S ε ( t ) V 0 V_\varepsilon(t)=S_\varepsilon(t)V_0 V ε ( t ) = S ε ( t ) V 0 . Define W ε ( t ) : = U ε ( t ) − V ε ( t ) W_\varepsilon(t):=U_\varepsilon(t)-V_\varepsilon(t) W ε ( t ) := U ε ( t ) − V ε ( t ) and Φ ε ( t ) : = F ( U ε ( t ) ) − F ( V ε ( t ) ) \Phi_\varepsilon(t):=F(U_\varepsilon(t))-F(V_\varepsilon(t)) Φ ε ( t ) := F ( U ε ( t )) − F ( V ε ( t )) . Then W ε W_\varepsilon W ε solves the difference equation
∂ t W ε + A ε W ε = Φ ε ( t ) , W ε ( 0 ) = U 0 − V 0 , (66) \partial_tW_\varepsilon+A_\varepsilon W_\varepsilon=\Phi_\varepsilon(t),\qquad W_\varepsilon(0)=U_0-V_0, \tag*{(66)} ∂ t W ε + A ε W ε = Φ ε ( t ) , W ε ( 0 ) = U 0 − V 0 , (66)
in the mild sense. By the variation-of-constants formula for the linear semigroup e − A ε t e^{-A_\varepsilon t} e − A ε t ,
W ε ( T ) = e − A ε T ( U 0 − V 0 ) + ∫ 0 T e − A ε ( T − s ) Φ ε ( s ) d s . (67) W_\varepsilon(T)=e^{-A_\varepsilon T}(U_0-V_0)+\int_0^T e^{-A_\varepsilon(T-s)}\Phi_\varepsilon(s)\,\mathrm{d}s. \tag*{(67)} W ε ( T ) = e − A ε T ( U 0 − V 0 ) + ∫ 0 T e − A ε ( T − s ) Φ ε ( s ) d s . (67)
Define
K ε , T ( U ) : = ∫ 0 T e − A ε ( T − s ) F ( S ε ( s ) U ) d s , U ∈ B 0 . (68) K_{\varepsilon,T}(U):=\int_0^T e^{-A_\varepsilon(T-s)}F(S_\varepsilon(s)U)\,\mathrm{d}s,\qquad U\in\mathcal{B}_0. \tag*{(68)} K ε , T ( U ) := ∫ 0 T e − A ε ( T − s ) F ( S ε ( s ) U ) d s , U ∈ B 0 . (68)
Then
K ε , T ( U 0 ) − K ε , T ( V 0 ) = ∫ 0 T e − A ε ( T − s ) Φ ε ( s ) d s , K_{\varepsilon,T}(U_0)-K_{\varepsilon,T}(V_0)=\int_0^T e^{-A_\varepsilon(T-s)}\Phi_\varepsilon(s)\,\mathrm{d}s, K ε , T ( U 0 ) − K ε , T ( V 0 ) = ∫ 0 T e − A ε ( T − s ) Φ ε ( s ) d s ,
so that ((67) ) becomes
W ε ( T ) = e − A ε T ( U 0 − V 0 ) + K ε , T ( U 0 ) − K ε , T ( V 0 ) . (69) W_\varepsilon(T)=e^{-A_\varepsilon T}(U_0-V_0)+K_{\varepsilon,T}(U_0)-K_{\varepsilon,T}(V_0). \tag*{(69)} W ε ( T ) = e − A ε T ( U 0 − V 0 ) + K ε , T ( U 0 ) − K ε , T ( V 0 ) . (69)
Step 1: Compression by the linear semigroup. By the uniform exponential stability established in Theorem 2.2 , there exist constants α ≥ 1 \alpha\ge1 α ≥ 1 and β > 0 \beta>0 β > 0 , independent of ε \varepsilon ε , such that ∥ e − A ε t ∥ L ( H ε ) ≤ α e − β t \left\|e^{-A_\varepsilon t}\right\|_{\mathcal{L}(\mathcal{H}_\varepsilon)}\le\alpha e^{-\beta t} e − A ε t L ( H ε ) ≤ α e − β t for all t ≥ 0 t\ge0 t ≥ 0 . Choose T > 0 T>0 T > 0 large enough so that
q : = α e − β T < 1. (70) q:=\alpha e^{-\beta T}<1. \tag*{(70)} q := α e − β T < 1. (70)
Then
∥ e − A ε T ( U 0 − V 0 ) ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 . (71) \left\|e^{-A_{\varepsilon}T}(U_{0}-V_{0})\right\|_{\mathcal{H}_{\varepsilon}}^{2}\leq q^{2}\left\|U_{0}-V_{0}\right\|_{\mathcal{H}_{\varepsilon}}^{2}. \tag*{(71)} e − A ε T ( U 0 − V 0 ) H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 . (71)
Step 2: Sequential compactness of K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_{0}) K ε , T ( B 0 ) . It remains to establish the sequential compactness property (65) . Fix ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] and let U ∈ B 0 U\in\mathcal{B}_{0} U ∈ B 0 . We split
K ε , T ( U ) = ∫ 0 T − σ e − A ε ( T − s ) F ( S ε ( s ) U ) d s ⏟ = : K ε , T ( 1 ) ( U ) + ∫ T − σ T e − A ε ( T − s ) F ( S ε ( s ) U ) d s ⏟ = : K ε , T ( 2 ) ( U ) , (72) K_{\varepsilon,T}(U)=\underbrace{\int_{0}^{T-\sigma}e^{-A_{\varepsilon}(T-s)}F(S_{\varepsilon}(s)U)\,\mathrm{d}s}_{=:K_{\varepsilon,T}^{(1)}(U)}+\underbrace{\int_{T-\sigma}^{T}e^{-A_{\varepsilon}(T-s)}F(S_{\varepsilon}(s)U)\,\mathrm{d}s}_{=:K_{\varepsilon,T}^{(2)}(U)}, \tag*{(72)} K ε , T ( U ) = =: K ε , T ( 1 ) ( U ) ∫ 0 T − σ e − A ε ( T − s ) F ( S ε ( s ) U ) d s + =: K ε , T ( 2 ) ( U ) ∫ T − σ T e − A ε ( T − s ) F ( S ε ( s ) U ) d s , (72)
where σ ∈ ( 0 , T ) \sigma\in(0,T) σ ∈ ( 0 , T ) is a small parameter to be chosen below, independent of ε \varepsilon ε .
(a) Smallness of K ε , T ( 2 ) K_{\varepsilon,T}^{(2)} K ε , T ( 2 ) . Since − A ε -A_{\varepsilon} − A ε generates a C 0 C_{0} C 0 -semigroup of contractions, we have C 0 : = sup ε ∈ ( 0 , 1 ] sup 0 ≤ τ ≤ T ∥ e − A ε τ ∥ L ( H ε ) = 1 C_{0}:=\sup_{\varepsilon\in(0,1]}\sup_{0\leq\tau\leq T}\left\|e^{-A_{\varepsilon}\tau}\right\|_{\mathcal{L}(\mathcal{H}_{\varepsilon})}=1 C 0 := sup ε ∈ ( 0 , 1 ] sup 0 ≤ τ ≤ T e − A ε τ L ( H ε ) = 1 . By Lemma 3.2, the orbits starting from B 0 \mathcal{B}_{0} B 0 remain in a fixed ball of radius R 1 R_{1} R 1 on [ 0 , T ] [0,T] [ 0 , T ] , uniformly in ε \varepsilon ε . Since F F F maps bounded sets into bounded sets, we have
M : = sup ε ∈ ( 0 , 1 ] sup U ∈ B 0 sup 0 ≤ s ≤ T ∥ F ( S ε ( s ) U ) ∥ H ε < + ∞ . M:=\sup_{\varepsilon\in(0,1]}\sup_{U\in\mathcal{B}_{0}}\sup_{0\leq s\leq T}\left\|F(S_{\varepsilon}(s)U)\right\|_{\mathcal{H}_{\varepsilon}}<+\infty. M := ε ∈ ( 0 , 1 ] sup U ∈ B 0 sup 0 ≤ s ≤ T sup ∥ F ( S ε ( s ) U ) ∥ H ε < + ∞.
Therefore
∥ K ε , T ( 2 ) ( U ) ∥ H ε ≤ C 0 M σ = M σ , ∀ U ∈ B 0 , ∀ ε ∈ ( 0 , 1 ] . (73) \left\|K_{\varepsilon,T}^{(2)}(U)\right\|_{\mathcal{H}_{\varepsilon}}\leq C_{0}M\sigma=M\sigma,\qquad\forall U\in\mathcal{B}_{0},\ \forall\varepsilon\in(0,1]. \tag*{(73)} K ε , T ( 2 ) ( U ) H ε ≤ C 0 M σ = M σ , ∀ U ∈ B 0 , ∀ ε ∈ ( 0 , 1 ] . (73)
For every δ > 0 \delta>0 δ > 0 , choose σ > 0 \sigma>0 σ > 0 so small that M σ < δ M\sigma<\delta M σ < δ .
(b) Relative compactness of K ε , T ( 1 ) ( B 0 ) K_{\varepsilon,T}^{(1)}(\mathcal{B}_{0}) K ε , T ( 1 ) ( B 0 ) . Write
K ε , T ( 1 ) ( U ) = e − A ε σ Ψ ε , σ ( U ) , Ψ ε , σ ( U ) : = ∫ 0 T − σ e − A ε ( T − σ − s ) F ( S ε ( s ) U ) d s . (74) K_{\varepsilon,T}^{(1)}(U)=e^{-A_{\varepsilon}\sigma}\Psi_{\varepsilon,\sigma}(U),\qquad\Psi_{\varepsilon,\sigma}(U):=\int_{0}^{T-\sigma}e^{-A_{\varepsilon}(T-\sigma-s)}F(S_{\varepsilon}(s)U)\,\mathrm{d}s. \tag*{(74)} K ε , T ( 1 ) ( U ) = e − A ε σ Ψ ε , σ ( U ) , Ψ ε , σ ( U ) := ∫ 0 T − σ e − A ε ( T − σ − s ) F ( S ε ( s ) U ) d s . (74)
Recall that F ( U ) F(U) F ( U ) has zero memory components, i.e.,
F ( U ) ∈ H e : = H a 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) , F(U)\in\mathcal{H}_{e}:=H_{a}^{2}(0,\ell)\times L^{2}(0,\ell)\times H_{0}^{1}(0,\ell)\times L^{2}(0,\ell), F ( U ) ∈ H e := H a 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) ,
regarded as a subspace of H ε \mathcal{H}_{\varepsilon} H ε with the memory components set to zero. We decompose the state space as H ε = H e ⊕ H m , ε \mathcal{H}_{\varepsilon}=\mathcal{H}_{e}\oplus\mathcal{H}_{m,\varepsilon} H ε = H e ⊕ H m , ε with H m , ε : = M g 1 , ε × M g 2 , ε \mathcal{H}_{m,\varepsilon}:=\mathcal{M}_{g_{1,\varepsilon}}\times\mathcal{M}_{g_{2,\varepsilon}} H m , ε := M g 1 , ε × M g 2 , ε , and correspondingly write Ψ ε , σ ( U ) = ( Ψ ε , σ e ( U ) , Ψ ε , σ m ( U ) ) \Psi_{\varepsilon,\sigma}(U)=(\Psi_{\varepsilon,\sigma}^{e}(U),\Psi_{\varepsilon,\sigma}^{m}(U)) Ψ ε , σ ( U ) = ( Ψ ε , σ e ( U ) , Ψ ε , σ m ( U )) .
(b.1) Regularity of the elastic component. On D ( A ε ) D(A_{\varepsilon}) D ( A ε ) , the operator A ε A_{\varepsilon} A ε admits the block decomposition
A ε = ( A e , ε B ε C ε A m , ε ) . A_{\varepsilon}=
\begin{pmatrix}
A_{e,\varepsilon} & B_{\varepsilon}\\
C_{\varepsilon} & A_{m,\varepsilon}
\end{pmatrix}. A ε = ( A e , ε C ε B ε A m , ε ) .
where A e , ε A_{e,\varepsilon} A e , ε is the elastic operator (with domain D ( A e , ε ) ⊂ H e D(A_{e,\varepsilon})\subset\mathcal{H}_{e} D ( A e , ε ) ⊂ H e ), A m , ε A_{m,\varepsilon} A m , ε is the memory operator generated by the transport equations, B ε : H m , ε → H e B_{\varepsilon}:\mathcal{H}_{m,\varepsilon}\to\mathcal{H}_{e} B ε : H m , ε → H e is the bounded coupling induced by the memory convolutions, and C ε : H e → H m , ε C_{\varepsilon}:\mathcal{H}_{e}\to\mathcal{H}_{m,\varepsilon} C ε : H e → H m , ε is the coupling induced by the transport equations. The operator A e , ε A_{e,\varepsilon} A e , ε has compact resolvent, uniformly in ε \varepsilon ε , because its resolvent is the inverse of an elliptic operator with coefficients μ ˉ 1 , μ ˉ 2 \bar{\mu}_{1},\bar{\mu}_{2} μ ˉ 1 , μ ˉ 2 independent of ε \varepsilon ε .
By the analytic smoothing estimates for the sectorial operator A e , ε A_{e,\varepsilon} A e , ε , for every θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) there exists C θ > 0 C_{\theta}>0 C θ > 0 , independent of ε \varepsilon ε , such that
∥ e − A e , ε t V ∥ D ( A e , ε θ ) ≤ C θ t − θ ∥ V ∥ H e , t > 0 , V ∈ H e . (75) \left\|e^{-A_{e,\varepsilon}t}V\right\|_{D(A_{e,\varepsilon}^{\theta})}\leq C_{\theta}t^{-\theta}\left\|V\right\|_{\mathcal{H}_{e}},\qquad t>0,\ V\in\mathcal{H}_{e}. \tag*{(75)} e − A e , ε t V D ( A e , ε θ ) ≤ C θ t − θ ∥ V ∥ H e , t > 0 , V ∈ H e . (75)
Here C θ C_{\theta} C θ depends only on θ \theta θ , μ ˉ 1 \bar{\mu}_{1} μ ˉ 1 , μ ˉ 2 \bar{\mu}_{2} μ ˉ 2 , and the domain, hence not on ε \varepsilon ε , because the elliptic coefficients μ ˉ i , ε = μ ˉ i \bar{\mu}_{i,\varepsilon}=\bar{\mu}_{i} μ ˉ i , ε = μ ˉ i are independent of ε \varepsilon ε by (27) . More generally, for 0 ≤ θ ′ < θ ≤ 1 0\leq\theta'<\theta\leq1 0 ≤ θ ′ < θ ≤ 1 ,
∥ e − A e , ε t V ∥ D ( A e , ε θ ) ≤ C θ , θ ′ t − ( θ − θ ′ ) ∥ V ∥ D ( A e , ε θ ′ ) , t > 0 , V ∈ D ( A e , ε θ ′ ) , (76) \left\|e^{-A_{e,\varepsilon}t}V\right\|_{D(A_{e,\varepsilon}^{\theta})}\leq C_{\theta,\theta'}t^{-(\theta-\theta')}\left\|V\right\|_{D(A_{e,\varepsilon}^{\theta'})},\qquad t>0,\ V\in D(A_{e,\varepsilon}^{\theta'}), \tag*{(76)} e − A e , ε t V D ( A e , ε θ ) ≤ C θ , θ ′ t − ( θ − θ ′ ) ∥ V ∥ D ( A e , ε θ ′ ) , t > 0 , V ∈ D ( A e , ε θ ′ ) , (76)
with C θ , θ ′ > 0 C_{\theta,\theta'} > 0 C θ , θ ′ > 0 independent of ε \varepsilon ε . We shall use (75) with θ = 1 / 2 \theta= 1/2 θ = 1/2 . For the elastic operator A e , ε A_{e,\varepsilon} A e , ε , the domain D ( A e , ε 1 / 2 ) D(A_{e,\varepsilon}^{1/2}) D ( A e , ε 1/2 ) is continuously embedded into
H a 2 ( 0 , ℓ ) × H a 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) H_a^2(0,\ell) \times H_a^2(0,\ell) \times H_0^1(0,\ell) \times H_0^1(0,\ell) H a 2 ( 0 , ℓ ) × H a 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ )
on the four elastic components ( u , ψ , v , ϕ ) (u,\psi,v,\phi) ( u , ψ , v , ϕ ) ; this is the standard regularity of the square root of the elastic operator, and the same embedding is used in the proof of [33 Lemma 4.3] . In particular, on the velocity block, D ( A e , ε 1 / 2 ) ↪ H a 2 ( 0 , ℓ ) D(A_{e,\varepsilon}^{1/2}) \hookrightarrow H_a^2(0,\ell) D ( A e , ε 1/2 ) ↪ H a 2 ( 0 , ℓ ) for the second component ψ \psi ψ and ↪ H 0 1 ( 0 , ℓ ) \hookrightarrow H_0^1(0,\ell) ↪ H 0 1 ( 0 , ℓ ) for the fourth component ϕ \phi ϕ . This is the natural regularity furnished by the square root of the fourth-order operator μ ˉ 1 u x x x x \bar{\mu}_1u_{xxxx} μ ˉ 1 u xxxx and the second-order operator − μ ˉ 2 v x x -\bar{\mu}_2v_{xx} − μ ˉ 2 v xx , whose coefficients are independent of ε \varepsilon ε . Consequently, for 0 ≤ r < θ ≤ T − σ 0 \le r < \theta\le T-\sigma 0 ≤ r < θ ≤ T − σ ,
( θ − r ) 1 / 2 ∥ [ e − A e , ε ( θ − r ) F e ( S ε ( T − σ − θ ) U ) ] 2 ∥ H a 2 ( 0 , ℓ ) ≤ C 1 / 2 M . (77) (\theta-r)^{1/2}\left\|\left[e^{-A_{e,\varepsilon}(\theta-r)}F_e(S_\varepsilon(T-\sigma-\theta)U)\right]_2\right\|_{H_a^2(0,\ell)} \le C_{1/2}M. \tag*{(77)} ( θ − r ) 1/2 [ e − A e , ε ( θ − r ) F e ( S ε ( T − σ − θ ) U ) ] 2 H a 2 ( 0 , ℓ ) ≤ C 1/2 M . (77)
Since F ( S ε ( s ) U ) F(S_\varepsilon(s)U) F ( S ε ( s ) U ) has zero memory components, we may apply (75) with θ = 1 / 2 \theta= 1/2 θ = 1/2 to obtain
∥ ∫ 0 T − σ e − A e , ε ( T − σ − s ) F e ( S ε ( s ) U ) d s ∥ D ( A e , ε 1 / 2 ) ≤ C 1 / 2 M ∫ 0 T − σ ( T − σ − s ) − 1 / 2 d s ≤ C 1 / 2 M ( 2 T ) 1 / 2 . (78) \left\|\int_0^{T-\sigma}e^{-A_{e,\varepsilon}(T-\sigma-s)}F_e(S_\varepsilon(s)U)\,ds\right\|_{D(A_{e,\varepsilon}^{1/2})}
\le C_{1/2}M\int_0^{T-\sigma}(T-\sigma-s)^{-1/2}\,ds
\le C_{1/2}M(2T)^{1/2}. \tag*{(78)} ∫ 0 T − σ e − A e , ε ( T − σ − s ) F e ( S ε ( s ) U ) d s D ( A e , ε 1/2 ) ≤ C 1/2 M ∫ 0 T − σ ( T − σ − s ) − 1/2 d s ≤ C 1/2 M ( 2 T ) 1/2 . (78)
Applying e − A e , ε σ e^{-A_{e,\varepsilon}\sigma} e − A e , ε σ and using (76) with θ ′ = 1 / 2 \theta' = 1/2 θ ′ = 1/2 and θ = 1 \theta= 1 θ = 1 ,
∥ e − A e , ε σ ∫ 0 T − σ e − A e , ε ( T − σ − s ) F e ( S ε ( s ) U ) d s ∥ D ( A e , ε ) ≤ C 1 , 1 / 2 σ − 1 / 2 C 1 / 2 M ( 2 T ) 1 / 2 , \left\|e^{-A_{e,\varepsilon}\sigma}\int_0^{T-\sigma}e^{-A_{e,\varepsilon}(T-\sigma-s)}F_e(S_\varepsilon(s)U)\,ds\right\|_{D(A_{e,\varepsilon})}
\le C_{1,1/2}\sigma^{-1/2}C_{1/2}M(2T)^{1/2}, e − A e , ε σ ∫ 0 T − σ e − A e , ε ( T − σ − s ) F e ( S ε ( s ) U ) d s D ( A e , ε ) ≤ C 1 , 1/2 σ − 1/2 C 1/2 M ( 2 T ) 1/2 ,
where C 1 , 1 / 2 C_{1,1/2} C 1 , 1/2 denotes the constant C θ , θ ′ C_{\theta,\theta'} C θ , θ ′ from (76) with θ = 1 \theta= 1 θ = 1 and θ ′ = 1 / 2 \theta' = 1/2 θ ′ = 1/2 ; it is independent of ε \varepsilon ε . Since A e , ε A_{e,\varepsilon} A e , ε has compact resolvent, D ( A e , ε ) D(A_{e,\varepsilon}) D ( A e , ε ) is compactly embedded in H e H_e H e , uniformly in ε \varepsilon ε . Therefore { e − A e , ε σ Ψ ε , σ e ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{e^{-A_{e,\varepsilon}\sigma}\Psi^e_{\varepsilon,\sigma}(U): U \in\mathcal{B}_0,\ \varepsilon\in(0,1]\} { e − A e , ε σ Ψ ε , σ e ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} is relatively compact in H e H_e H e .
(b.2) Compactness of the memory component. We show that { Ψ ε , σ m ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi^m_{\varepsilon,\sigma}(U): U \in\mathcal{B}_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ m ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} is relatively compact in H m , ε H_{m,\varepsilon} H m , ε . We treat only the first component Ψ ε , σ σ ( U ) \Psi^\sigma_{\varepsilon,\sigma}(U) Ψ ε , σ σ ( U ) ; the argument for the second is identical.
Since F ( U ) F(U) F ( U ) has zero memory components, the memory component Ψ ε , σ σ ( U ) \Psi^\sigma_{\varepsilon,\sigma}(U) Ψ ε , σ σ ( U ) is generated by the elastic velocity u ε , t u_{\varepsilon,t} u ε , t through the transport equation. Explicitly, for s ≥ 0 s \ge0 s ≥ 0 ,
Ψ ε , σ σ ( U ) ( s ) = − ∫ 0 T − σ ∫ max { 0 , T − σ − τ − s } T − σ − τ ψ ε , σ ( r ; τ ) d r d τ , (79) \Psi^\sigma_{\varepsilon,\sigma}(U)(s)
=
-\int_0^{T-\sigma}\int_{\max\{0,T-\sigma-\tau-s\}}^{T-\sigma-\tau}
\psi_{\varepsilon,\sigma}(r;\tau)\,dr\,d\tau, \tag*{(79)} Ψ ε , σ σ ( U ) ( s ) = − ∫ 0 T − σ ∫ m a x { 0 , T − σ − τ − s } T − σ − τ ψ ε , σ ( r ; τ ) d r d τ , (79)
where ψ ε , σ ( r ; τ ) : = [ e − A e , ε ( r − τ ) F e ( S ε ( τ ) U ) ] 2 \psi_{\varepsilon,\sigma}(r;\tau):=[e^{-A_{e,\varepsilon}(r-\tau)}F_e(S_\varepsilon(\tau)U)]_2 ψ ε , σ ( r ; τ ) := [ e − A e , ε ( r − τ ) F e ( S ε ( τ ) U ) ] 2 for r ≥ τ r \ge\tau r ≥ τ and ψ ε , σ ( r ; τ ) = 0 \psi_{\varepsilon,\sigma}(r;\tau)=0 ψ ε , σ ( r ; τ ) = 0 otherwise, and [ ⋅ ] 2 [\cdot]_2 [ ⋅ ] 2 denotes the second component of an element of H e H_e H e (i.e., the velocity u ε , t u_{\varepsilon,t} u ε , t in the elastic u u u -block). Changing variables θ = T − σ − τ \theta=T-\sigma-\tau θ = T − σ − τ ,
Ψ ε , σ σ ( U ) ( s ) = − ∫ 0 T − σ ∫ max { 0 , θ − s } θ ψ ε , σ ( r ; T − σ − θ ) d r d θ . (80) \Psi^\sigma_{\varepsilon,\sigma}(U)(s)
=
-\int_0^{T-\sigma}\int_{\max\{0,\theta-s\}}^\theta
\psi_{\varepsilon,\sigma}(r;T-\sigma-\theta)\,dr\,d\theta. \tag*{(80)} Ψ ε , σ σ ( U ) ( s ) = − ∫ 0 T − σ ∫ m a x { 0 , θ − s } θ ψ ε , σ ( r ; T − σ − θ ) d r d θ . (80)
We claim that the family { Ψ ε , σ σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi^\sigma_{\varepsilon,\sigma}(U): U \in\mathcal{B}_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} satisfies the hypotheses of the Fréchet–Kolmogorov theorem in the weighted space M g 1 , ε \mathcal{M}_{g_1,\varepsilon} M g 1 , ε , hence is relatively compact. To this end, we verify:
(i) Uniform boundedness in M g 1 , ε \mathcal{M}_{g_1,\varepsilon} M g 1 , ε . By (77) , for 0 ≤ r < θ ≤ T − σ 0 \le r < \theta\le T-\sigma 0 ≤ r < θ ≤ T − σ ,
( θ − r ) 1 / 2 ∥ [ e − A e , ε ( θ − r ) F e ( S ε ( T − σ − θ ) U ) ] 2 ∥ H a 2 ( 0 , ℓ ) ≤ C 1 / 2 M . (81) (\theta-r)^{1/2}\left\|\left[e^{-A_{e,\varepsilon}(\theta-r)}F_e(S_\varepsilon(T-\sigma-\theta)U)\right]_2\right\|_{H_a^2(0,\ell)} \le C_{1/2}M. \tag*{(81)} ( θ − r ) 1/2 [ e − A e , ε ( θ − r ) F e ( S ε ( T − σ − θ ) U ) ] 2 H a 2 ( 0 , ℓ ) ≤ C 1/2 M . (81)
Combining this with (80) and the integrability of ( θ − r ) − 1 / 2 (\theta-r)^{-1/2} ( θ − r ) − 1/2 over r ∈ [ max { 0 , θ − s } , θ ] r \in[\max\{0,\theta-s\},\theta] r ∈ [ max { 0 , θ − s } , θ ] , for every s ≥ 0 s \ge0 s ≥ 0 we obtain
∥ Ψ ε , σ σ ( U ) ( s ) ∥ H a 2 ( 0 , ℓ ) ≤ C , C : = 2 C 1 / 2 M T 1 / 2 , \left\|\Psi^\sigma_{\varepsilon,\sigma}(U)(s)\right\|_{H_a^2(0,\ell)} \le C,\qquad C:=2C_{1/2}MT^{1/2}, Ψ ε , σ σ ( U ) ( s ) H a 2 ( 0 , ℓ ) ≤ C , C := 2 C 1/2 M T 1/2 ,
with C C C independent of U ∈ B 0 U \in B_0 U ∈ B 0 , ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , and s s s . Therefore
∥ Ψ ε , σ σ ( U ) ∥ M g 1 , ε 2 = ∫ 0 + ∞ g 1 , ε ( s ) ∥ Ψ ε , σ σ ( U ) ( s ) ∥ H a 2 ( 0 , ℓ ) 2 d s ≤ C 2 β 1 , \left\|\Psi^{\sigma}_{\varepsilon,\sigma}(U)\right\|_{\mathcal{M}_{g_{1,\varepsilon}}}^{2}=\int_{0}^{+\infty}g_{1,\varepsilon}(s)\left\|\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s)\right\|_{H_a^2(0,\ell)}^{2}\,\mathrm{d}s\le C^2\beta_1, Ψ ε , σ σ ( U ) M g 1 , ε 2 = ∫ 0 + ∞ g 1 , ε ( s ) Ψ ε , σ σ ( U ) ( s ) H a 2 ( 0 , ℓ ) 2 d s ≤ C 2 β 1 ,
uniformly in U U U and ε \varepsilon ε .
(ii) Uniform tail smallness. Let δ > 0 \delta>0 δ > 0 . By (9) , g 1 ( s ) ≤ g 1 ( 0 ) e − λ 1 s g_1(s)\le g_1(0)e^{-\lambda_1s} g 1 ( s ) ≤ g 1 ( 0 ) e − λ 1 s , hence g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) ≤ ε − 1 g 1 ( 0 ) e − λ 1 s / ε g_{1,\varepsilon}(s)=\varepsilon^{-1}g_1(s/\varepsilon)\le\varepsilon^{-1}g_1(0)e^{-\lambda_1s/\varepsilon} g 1 , ε ( s ) = ε − 1 g 1 ( s / ε ) ≤ ε − 1 g 1 ( 0 ) e − λ 1 s / ε . Therefore for every R > 0 R>0 R > 0 ,
∫ R + ∞ g 1 , ε ( s ) d s ≤ g 1 ( 0 ) λ 1 e − λ 1 R / ε ≤ g 1 ( 0 ) λ 1 e − λ 1 R , \int_R^{+\infty}g_{1,\varepsilon}(s)\,\mathrm{d}s\le\frac{g_1(0)}{\lambda_1}e^{-\lambda_1R/\varepsilon}\le\frac{g_1(0)}{\lambda_1}e^{-\lambda_1R}, ∫ R + ∞ g 1 , ε ( s ) d s ≤ λ 1 g 1 ( 0 ) e − λ 1 R / ε ≤ λ 1 g 1 ( 0 ) e − λ 1 R ,
where the last inequality uses ε ≤ 1 \varepsilon\le1 ε ≤ 1 . The right-hand side is independent of ε \varepsilon ε and tends to 0 0 0 as R → + ∞ R\to+\infty R → + ∞ , uniformly in ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . Combining this with the uniform bound on ∥ Ψ ε , σ σ ( U ) ( s ) ∥ H a 2 ( 0 , ℓ ) \left\|\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s)\right\|_{H_a^2(0,\ell)} Ψ ε , σ σ ( U ) ( s ) H a 2 ( 0 , ℓ ) , we conclude that for every δ > 0 \delta>0 δ > 0 there exists R = R ( δ ) > 0 R=R(\delta)>0 R = R ( δ ) > 0 , independent of ε \varepsilon ε , such that
sup U ∈ B 0 sup ε ∈ ( 0 , 1 ] ∫ R + ∞ g 1 , ε ( s ) ∥ Ψ ε , σ σ ( U ) ( s ) ∥ H a 2 ( 0 , ℓ ) 2 d s < δ , \sup_{U\in B_0}\sup_{\varepsilon\in(0,1]}\int_R^{+\infty}g_{1,\varepsilon}(s)\left\|\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s)\right\|_{H_a^2(0,\ell)}^{2}\,\mathrm{d}s<\delta, U ∈ B 0 sup ε ∈ ( 0 , 1 ] sup ∫ R + ∞ g 1 , ε ( s ) Ψ ε , σ σ ( U ) ( s ) H a 2 ( 0 , ℓ ) 2 d s < δ ,
i.e. the tail of the M g 1 , ε \mathcal{M}_{g_{1,\varepsilon}} M g 1 , ε -norm is uniformly small.
(iii) Hölder continuity in s s s . We claim that there exists C 3 > 0 C_3>0 C 3 > 0 , independent of U U U and ε \varepsilon ε , such that for all 0 ≤ s 1 < s 2 0\le s_1<s_2 0 ≤ s 1 < s 2 ,
∥ Ψ ε , σ σ ( U ) ( s 1 ) − Ψ ε , σ σ ( U ) ( s 2 ) ∥ H a 2 ( 0 , ℓ ) ≤ C 3 ∣ s 1 − s 2 ∣ 1 / 2 . (82) \left\|\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s_1)-\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s_2)\right\|_{H_a^2(0,\ell)}\le C_3\lvert s_1-s_2\rvert^{1/2}. \tag*{(82)} Ψ ε , σ σ ( U ) ( s 1 ) − Ψ ε , σ σ ( U ) ( s 2 ) H a 2 ( 0 , ℓ ) ≤ C 3 ∣ s 1 − s 2 ∣ 1/2 . (82)
Indeed, from (80) ,
Ψ ε , σ σ ( U ) ( s 1 ) − Ψ ε , σ σ ( U ) ( s 2 ) = − ∫ 0 T − σ ∫ max { 0 , θ − s 2 } max { 0 , θ − s 1 } ψ ε , σ ( r ; T − σ − θ ) d r d θ . \Psi^{\sigma}_{\varepsilon,\sigma}(U)(s_1)-\Psi^{\sigma}_{\varepsilon,\sigma}(U)(s_2)
=-\int_0^{T-\sigma}\int_{\max\{0,\theta-s_2\}}^{\max\{0,\theta-s_1\}}\psi_{\varepsilon,\sigma}(r;T-\sigma-\theta)\,\mathrm{d}r\,\mathrm{d}\theta. Ψ ε , σ σ ( U ) ( s 1 ) − Ψ ε , σ σ ( U ) ( s 2 ) = − ∫ 0 T − σ ∫ m a x { 0 , θ − s 2 } m a x { 0 , θ − s 1 } ψ ε , σ ( r ; T − σ − θ ) d r d θ .
For fixed θ \theta θ , the inner integral is over an interval of length at most ∣ s 1 − s 2 ∣ \lvert s_1-s_2\rvert ∣ s 1 − s 2 ∣ . To handle the singularity at r = θ r=\theta r = θ , we split the integration region into { θ − r ≥ δ 0 } \{\theta-r\ge\delta_0\} { θ − r ≥ δ 0 } and { θ − r < δ 0 } \{\theta-r<\delta_0\} { θ − r < δ 0 } for a small δ 0 > 0 \delta_0>0 δ 0 > 0 . On the first region, by (81) , ∥ ψ ε , σ ( r ; T − σ − θ ) ∥ H a 2 ≤ C 1 / 2 M δ 0 − 1 / 2 \left\|\psi_{\varepsilon,\sigma}(r;T-\sigma-\theta)\right\|_{H_a^2}\le C_{1/2}M\delta_0^{-1/2} ∥ ψ ε , σ ( r ; T − σ − θ ) ∥ H a 2 ≤ C 1/2 M δ 0 − 1/2 , while on the second region we use the endpoint bound ∥ ψ ε , σ ( θ ; ⋅ ) ∥ H a 2 ≤ M \left\|\psi_{\varepsilon,\sigma}(\theta;\cdot)\right\|_{H_a^2}\le M ∥ ψ ε , σ ( θ ; ⋅ ) ∥ H a 2 ≤ M (for r = θ r=\theta r = θ , the source is F e ( S ε ( T − σ − θ ) U ) F_e(S_\varepsilon(T-\sigma-\theta)U) F e ( S ε ( T − σ − θ ) U ) itself) and the integrability of ( θ − r ) − 1 / 2 (\theta-r)^{-1/2} ( θ − r ) − 1/2 . Combining these two estimates with the elementary inequality ∣ a 1 / 2 − b 1 / 2 ∣ ≤ ∣ a − b ∣ 1 / 2 \lvert a^{1/2}-b^{1/2}\rvert\le\lvert a-b\rvert^{1/2} ∣ a 1/2 − b 1/2 ∣ ≤ ∣ a − b ∣ 1/2 , we obtain
∥ ∫ max { 0 , θ − s 1 } max { 0 , θ − s 2 } ψ ε , σ ( r ; T − σ − θ ) d r ∥ H a 2 ( 0 , ℓ ) ≤ 2 C 1 / 2 M ∣ s 1 − s 2 ∣ 1 / 2 . \left\|\int_{\max\{0,\theta-s_1\}}^{\max\{0,\theta-s_2\}}\psi_{\varepsilon,\sigma}(r;T-\sigma-\theta)\,\mathrm{d}r\right\|_{H_a^2(0,\ell)}\le2C_{1/2}M\lvert s_1-s_2\rvert^{1/2}. ∫ m a x { 0 , θ − s 1 } m a x { 0 , θ − s 2 } ψ ε , σ ( r ; T − σ − θ ) d r H a 2 ( 0 , ℓ ) ≤ 2 C 1/2 M ∣ s 1 − s 2 ∣ 1/2 .
Integrating over θ ∈ [ 0 , T − σ ] \theta\in[0,T-\sigma] θ ∈ [ 0 , T − σ ] yields (82) with C 3 : = 2 C 1 / 2 M ( T − σ ) C_3:=2C_{1/2}M(T-\sigma) C 3 := 2 C 1/2 M ( T − σ ) .
(iv) Compactness on compact intervals. Combining (i)–(iii), the family { Ψ ε , σ σ ( U ) ∣ [ 0 , R ] : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi^{\sigma}_{\varepsilon,\sigma}(U)|_{[0,R]}:U\in B_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ σ ( U ) ∣ [ 0 , R ] : U ∈ B 0 , ε ∈ ( 0 , 1 ]} is uniformly bounded in L ∞ ( 0 , R ; H a 2 ( 0 , ℓ ) ) L^\infty(0,R;H_a^2(0,\ell)) L ∞ ( 0 , R ; H a 2 ( 0 , ℓ )) and uniformly Hölder continuous of order 1 / 2 1/2 1/2 . By the Arzelà–Ascoli theorem and the compact embedding H 1 / 2 ( 0 , R ; H a 2 ( 0 , ℓ ) ) ↪ L 2 ( 0 , R ; H a 2 ( 0 , ℓ ) ) H^{1/2}(0,R;H_a^2(0,\ell))\hookrightarrow L^2(0,R;H_a^2(0,\ell)) H 1/2 ( 0 , R ; H a 2 ( 0 , ℓ )) ↪ L 2 ( 0 , R ; H a 2 ( 0 , ℓ )) , this family is relatively compact in L 2 ( 0 , R ; H a 2 ( 0 , ℓ ) ) L^2(0,R;H_a^2(0,\ell)) L 2 ( 0 , R ; H a 2 ( 0 , ℓ )) .
(v) Conclusion for the memory component. Combining (ii) and (iv) with a standard diagonal argument in the weighted space M g 1 , ε \mathcal{M}_{g_{1,\varepsilon}} M g 1 , ε yields the relative compactness of { Ψ ε , σ σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi^{\sigma}_{\varepsilon,\sigma}(U):U\in B_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} in M g 1 , ε \mathcal{M}_{g_{1,\varepsilon}} M g 1 , ε . The same argument applies to the second memory component in M g 2 , ε \mathcal{M}_{g_{2,\varepsilon}} M g 2 , ε . Therefore { Ψ ε , σ m ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi^m_{\varepsilon,\sigma}(U):U\in B_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ m ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} is relatively compact in H m , ε \mathcal{H}_{m,\varepsilon} H m , ε . Moreover, the compact set furnished by the Fréchet–Kolmogorov criterion can be chosen independently of ε \varepsilon ε , because the constants in (i)–(iii) and the tail estimate in (ii) are uniform in ε \varepsilon ε .
(c) Relative compactness of K ε , T ( 1 ) ( B 0 ) K^{(1)}_{\varepsilon,\mathcal{T}}(B_0) K ε , T ( 1 ) ( B 0 ) . Combining (b.1) and (b.2), the set { Ψ ε , σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{\Psi_{\varepsilon,\sigma}(U):U\in B_0,\ \varepsilon\in(0,1]\} { Ψ ε , σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} is relatively compact in the uniform topology of Remark 2.1. Since e − A ε σ e^{-\mathcal{A}_{\varepsilon}\sigma} e − A ε σ is a bounded linear operator on H ε \mathcal{H}_{\varepsilon} H ε (uniformly in ε \varepsilon ε ), its image of a relatively compact set is relatively compact. Therefore
{ K ε , T ( 1 ) ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } = e − A ε σ { Ψ ε , σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ] } \{K_{\varepsilon,T}^{(1)}(U): U \in\mathcal{B}_0,\ \varepsilon\in(0,1]\}=e^{-A_{\varepsilon}\sigma}\{\Psi_{\varepsilon,\sigma}(U): U \in\mathcal{B}_0,\ \varepsilon\in(0,1]\} { K ε , T ( 1 ) ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]} = e − A ε σ { Ψ ε , σ ( U ) : U ∈ B 0 , ε ∈ ( 0 , 1 ]}
is relatively compact.
(d) Sequential compactness. Combining (a) and (c), for every δ > 0 \delta>0 δ > 0 we have exhibited a relatively compact set K δ K_{\delta} K δ (independent of ε \varepsilon ε ) such that dist ( K ε , T ( U ) , K δ ) < δ \operatorname{dist}(K_{\varepsilon,T}(U),K_{\delta})<\delta dist ( K ε , T ( U ) , K δ ) < δ for all U ∈ B 0 U\in\mathcal{B}_0 U ∈ B 0 and ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , where the distance is measured in the uniform topology of Remark 2.1 (on a common reference space). Now let ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + and U n ∈ B 0 U_n\in\mathcal{B}_0 U n ∈ B 0 be arbitrary. For each j ∈ N j\in\mathbb{N} j ∈ N , choose δ = 1 / j \delta=1/j δ = 1/ j and the corresponding relatively compact set K 1 / j K_{1/j} K 1/ j . A nested diagonal extraction over j j j yields a single subsequence { ( ε n k , U n k ) } \{(\varepsilon_{n_k},U_{n_k})\} {( ε n k , U n k )} along which K ε n k , T ( U n k ) K_{\varepsilon_{n_k},T}(U_{n_k}) K ε n k , T ( U n k ) is Cauchy, hence converges in the uniform topology of Remark 2.1. This proves the sequential compactness property (65).
Step 3: Conclusion. Combining (69), (71), and the sequential compactness property of K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_0) K ε , T ( B 0 ) established in Step 2(d), we obtain (64). All the constants α , β , T , q , C 1 / 2 , C 0 , M , C 3 , β 1 \alpha,\beta,T,q,C_{1/2},C_0,M,C_3,\beta_1 α , β , T , q , C 1/2 , C 0 , M , C 3 , β 1 depend only on γ i , μ i , λ i , β i , c e , c p \gamma_i,\mu_i,\lambda_i,\beta_i,c_e,c_p γ i , μ i , λ i , β i , c e , c p , the constants in (11), and the radius of B 0 \mathcal{B}_0 B 0 , but not on ε \varepsilon ε . This completes the proof.
Convergence of semigroups
The next ingredient is the finite-time convergence (63).
Proposition 4.2 (Finite-time convergence). Let B ⊂ H ε B\subset\mathcal{H}_{\varepsilon} B ⊂ H ε be bounded and let U 0 , ε ∈ B U_{0,\varepsilon}\in B U 0 , ε ∈ B with U 0 , ε → U 0 U_{0,\varepsilon}\to U_0 U 0 , ε → U 0 as ε → 0 + \varepsilon\to0^+ ε → 0 + in the sense of Remark 2.1. Then for every T > 0 T>0 T > 0 ,
sup t ∈ [ 0 , T ] ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0. \sup_{t\in[0,T]}\|S_{\varepsilon}(t)U_{0,\varepsilon}-S_0(t)U_0\|_{\mathcal{H}_{\varepsilon}}\to0. t ∈ [ 0 , T ] sup ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0. Proof. Throughout, let
H e : = H 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) , H m , ε : = M g 1 , ε × M g 2 , ε , \mathcal{H}_e:=H^2(0,\ell)\times L^2(0,\ell)\times H_0^1(0,\ell)\times L^2(0,\ell),\qquad\mathcal{H}_{m,\varepsilon}:=\mathcal{M}_{g_1,\varepsilon}\times\mathcal{M}_{g_2,\varepsilon}, H e := H 2 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) × H 0 1 ( 0 , ℓ ) × L 2 ( 0 , ℓ ) , H m , ε := M g 1 , ε × M g 2 , ε , so that H ε = H e ⊕ H m , ε \mathcal{H}_{\varepsilon}=\mathcal{H}_e\oplus\mathcal{H}_{m,\varepsilon} H ε = H e ⊕ H m , ε . Let A e , 0 A_{e,0} A e , 0 denote the elastic part of the limit operator, acting on H e \mathcal{H}_e H e by
A e , 0 ( u ψ v ϕ ) = ( − ψ γ 1 ψ + μ 1 u x x x x − ϕ γ 2 ϕ − μ 2 v x x ) , (83) A_{e,0}
\begin{pmatrix}
u\\
\psi\\
v\\
\phi
\end{pmatrix}
=
\begin{pmatrix}
-\psi\\
\gamma_1\psi+\mu_1u_{xxxx}\\
-\phi\\
\gamma_2\phi-\mu_2v_{xx}
\end{pmatrix}, \tag*{(83)} A e , 0 u ψ v ϕ = − ψ γ 1 ψ + μ 1 u xxxx − ϕ γ 2 ϕ − μ 2 v xx , (83) where μ i = μ ˉ i + β i \mu_i=\bar{\mu}_i+\beta_i μ i = μ ˉ i + β i are the original elastic coefficients. The limit operator A 0 A_0 A 0 is defined on H e ⊕ H m , ε \mathcal{H}_e\oplus\mathcal{H}_{m,\varepsilon} H e ⊕ H m , ε as A 0 = A e , 0 ⊕ 0 A_0=A_{e,0}\oplus0 A 0 = A e , 0 ⊕ 0 , that is, A e , 0 A_{e,0} A e , 0 on the elastic components and zero on the memory components; its action is independent of the particular weighted space H m , ε \mathcal{H}_{m,\varepsilon} H m , ε because the limit system is memoryless. Since the memory components of U 0 ( s ) = S 0 ( s ) U 0 U_0(s)=S_0(s)U_0 U 0 ( s ) = S 0 ( s ) U 0 vanish for all s ≥ 0 s\geq0 s ≥ 0 , we may regard U 0 ( s ) U_0(s) U 0 ( s ) as an element of H ε \mathcal{H}_{\varepsilon} H ε via the natural embedding H e ↪ H ε \mathcal{H}_e\hookrightarrow\mathcal{H}_{\varepsilon} H e ↪ H ε (memory components set to zero), and we have F 0 ( U 0 ( s ) ) = F ( U 0 ( s ) ) F_0(U_0(s))=F(U_0(s)) F 0 ( U 0 ( s )) = F ( U 0 ( s )) , where F 0 F_0 F 0 is the nonlinearity of the limit system.
Let U ε ( t ) = S ε ( t ) U 0 , ε U_{\varepsilon}(t)=S_{\varepsilon}(t)U_{0,\varepsilon} U ε ( t ) = S ε ( t ) U 0 , ε and U 0 ( t ) = S 0 ( t ) U 0 U_0(t)=S_0(t)U_0 U 0 ( t ) = S 0 ( t ) U 0 denote the solutions of the perturbed and limit systems, respectively, and set W ε ( t ) : = U ε ( t ) − U 0 ( t ) W_{\varepsilon}(t):=U_{\varepsilon}(t)-U_0(t) W ε ( t ) := U ε ( t ) − U 0 ( t ) . Both solutions satisfy the mild formulation
U ε ( t ) = e − A ε t U 0 , ε + ∫ 0 t e − A ε ( t − s ) F ( U ε ( s ) ) d s , (84) U_{\varepsilon}(t)=e^{-A_{\varepsilon}t}U_{0,\varepsilon}+\int_0^t e^{-A_{\varepsilon}(t-s)}F(U_{\varepsilon}(s))\,\mathrm{d}s, \tag*{(84)} U ε ( t ) = e − A ε t U 0 , ε + ∫ 0 t e − A ε ( t − s ) F ( U ε ( s )) d s , (84) U 0 ( t ) = e − A 0 t U 0 + ∫ 0 t e − A 0 ( t − s ) F 0 ( U 0 ( s ) ) d s . (85) U_0(t)=e^{-A_0t}U_0+\int_0^t e^{-A_0(t-s)}F_0(U_0(s))\,\mathrm{d}s. \tag*{(85)} U 0 ( t ) = e − A 0 t U 0 + ∫ 0 t e − A 0 ( t − s ) F 0 ( U 0 ( s )) d s . (85)
Subtracting (85) from (84) , and using F 0 ( U 0 ( s ) ) = F ( U 0 ( s ) ) F_0(U_0(s)) = F(U_0(s)) F 0 ( U 0 ( s )) = F ( U 0 ( s )) , we obtain
W ε ( t ) = e − A ε t ( U 0 , ε − U 0 ) + ∫ 0 t e − A ε ( t − s ) [ F ( U ε ( s ) ) − F ( U 0 ( s ) ) ] d s + R ε ( t ) , (86) W_\varepsilon(t) = e^{-A_\varepsilon t}(U_{0,\varepsilon}-U_0) + \int_0^t e^{-A_\varepsilon(t-s)}[F(U_\varepsilon(s))-F(U_0(s))]\,\mathrm{d}s + R_\varepsilon(t), \tag*{(86)} W ε ( t ) = e − A ε t ( U 0 , ε − U 0 ) + ∫ 0 t e − A ε ( t − s ) [ F ( U ε ( s )) − F ( U 0 ( s ))] d s + R ε ( t ) , (86)
where the remainder is
R ε ( t ) : = ∫ 0 t [ e − A ε ( t − s ) − e − A 0 ( t − s ) ] F ( U 0 ( s ) ) d s . (87) R_\varepsilon(t) := \int_0^t [e^{-A_\varepsilon(t-s)}-e^{-A_0(t-s)}]F(U_0(s))\,\mathrm{d}s. \tag*{(87)} R ε ( t ) := ∫ 0 t [ e − A ε ( t − s ) − e − A 0 ( t − s ) ] F ( U 0 ( s )) d s . (87)
Step 1: Uniform boundedness on [ 0 , T ] [0,T] [ 0 , T ] . By Lemma 3.2, the solutions U ε ( t ) U_\varepsilon(t) U ε ( t ) remain in a fixed ball B 1 : = { U : ∥ U ∥ H ε ≤ R 1 } B_1 := \{U : \|U\|_{\mathcal{H}_\varepsilon} \le R_1\} B 1 := { U : ∥ U ∥ H ε ≤ R 1 } for t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , uniformly in ε \varepsilon ε . Consequently, there exists L 1 > 0 L_1 > 0 L 1 > 0 , independent of ε \varepsilon ε , such that
∥ F ( U ) − F ( V ) ∥ H ε ≤ L 1 ∥ U − V ∥ H ε , U , V ∈ B 1 , (88) \|F(U)-F(V)\|_{\mathcal{H}_\varepsilon} \le L_1\|U-V\|_{\mathcal{H}_\varepsilon}, \qquad U,V \in B_1, \tag*{(88)} ∥ F ( U ) − F ( V ) ∥ H ε ≤ L 1 ∥ U − V ∥ H ε , U , V ∈ B 1 , (88)
and M 1 : = sup U ∈ B 1 ∥ F ( U ) ∥ H ε < + ∞ M_1 := \sup_{U\in B_1}\|F(U)\|_{\mathcal{H}_\varepsilon} < +\infty M 1 := sup U ∈ B 1 ∥ F ( U ) ∥ H ε < + ∞ .
Step 2: Vanishing of the remainder R ε R_\varepsilon R ε . We claim that
sup t ∈ [ 0 , T ] ∥ R ε ( t ) ∥ H ε → 0 , ε → 0 + . (89) \sup_{t\in[0,T]}\|R_\varepsilon(t)\|_{\mathcal{H}_\varepsilon} \to0, \qquad\varepsilon\to0^+. \tag*{(89)} t ∈ [ 0 , T ] sup ∥ R ε ( t ) ∥ H ε → 0 , ε → 0 + . (89)
We split R ε = R ε e + R ε m R_\varepsilon= R_\varepsilon^e + R_\varepsilon^m R ε = R ε e + R ε m according to the decomposition H ε = H e ⊕ H m , ε \mathcal{H}_\varepsilon= \mathcal{H}_e \oplus\mathcal{H}_{m,\varepsilon} H ε = H e ⊕ H m , ε , where R ε e R_\varepsilon^e R ε e and R ε m R_\varepsilon^m R ε m denote the elastic and memory components of R ε R_\varepsilon R ε , respectively.
( a ) (a) ( a ) Elastic component. We first show the strong convergence
sup τ ∈ [ 0 , T ] ∥ [ e − A e , ε τ − e − A e , 0 τ ] V ∥ H e → 0 , ε → 0 + , ∀ V ∈ H e . (90) \sup_{\tau\in[0,T]}\|[e^{-A_{e,\varepsilon}\tau}-e^{-A_{e,0}\tau}]V\|_{\mathcal{H}_e} \to0, \qquad\varepsilon\to0^+, \quad\forall V \in\mathcal{H}_e. \tag*{(90)} τ ∈ [ 0 , T ] sup ∥ [ e − A e , ε τ − e − A e , 0 τ ] V ∥ H e → 0 , ε → 0 + , ∀ V ∈ H e . (90)
By the Trotter–Kato theorem (see, e.g., [15 , 30 ] ), it suffices to verify that
( μ I + A e , ε ) − 1 V → ( μ I + A e , 0 ) − 1 V in H e , ε → 0 + , (91) (\mu I+A_{e,\varepsilon})^{-1}V \to(\mu I+A_{e,0})^{-1}V \quad\text{in } \mathcal{H}_e,\qquad\varepsilon\to0^+, \tag*{(91)} ( μ I + A e , ε ) − 1 V → ( μ I + A e , 0 ) − 1 V in H e , ε → 0 + , (91)
for some μ > 0 \mu>0 μ > 0 and every V ∈ H e V\in\mathcal{H}_e V ∈ H e . Fix μ > 0 \mu>0 μ > 0 and V = ( h 1 , h 2 , h 4 , h 5 ) ∈ H e V=(h^1,h^2,h^4,h^5)\in\mathcal{H}_e V = ( h 1 , h 2 , h 4 , h 5 ) ∈ H e , where the four components are ordered as ( u , ψ , v , ϕ ) (u,\psi,v,\phi) ( u , ψ , v , ϕ ) ; equivalently, V = ( h 1 , h 2 , 0 , h 4 , 0 , h 5 ) ∈ H ε V=(h^1,h^2,0,h^4,0,h^5)\in\mathcal{H}_\varepsilon V = ( h 1 , h 2 , 0 , h 4 , 0 , h 5 ) ∈ H ε when V V V is regarded as an element of H ε = H e ⊕ H m , ε \mathcal{H}_\varepsilon=\mathcal{H}_e\oplus\mathcal{H}_{m,\varepsilon} H ε = H e ⊕ H m , ε , so that its memory components h 3 h^3 h 3 and h 6 h^6 h 6 vanish identically. The resolvent equation ( μ I + A e , ε ) − 1 V = W ε = ( u ε , ψ ε , v ε , ϕ ε ) (\mu I+A_{e,\varepsilon})^{-1}V=W_\varepsilon=(u_\varepsilon,\psi_\varepsilon,v_\varepsilon,\phi_\varepsilon) ( μ I + A e , ε ) − 1 V = W ε = ( u ε , ψ ε , v ε , ϕ ε ) reads
{ μ u ε − ψ ε = h 1 , μ ψ ε + γ 1 ψ ε + μ ˉ 1 u ε , x x x x + ∫ 0 + ∞ g 1 , ε ( s ) σ ε , x x x x ( s ) d s = h 2 , μ σ ε + ∂ s σ ε − ψ ε = h 3 , μ v ε − ϕ ε = h 4 , μ ϕ ε + γ 2 ϕ ε − μ ˉ 2 v ε , x x − ∫ 0 + ∞ g 2 , ε ( s ) η ε , x x ( s ) d s = h 5 , μ η ε + ∂ s η ε − ϕ ε = h 6 , (92) \begin{cases}
\mu u_\varepsilon-\psi_\varepsilon=h^1,\\
\mu\psi_\varepsilon+\gamma_1\psi_\varepsilon+\bar{\mu}_1u_{\varepsilon,xxxx}+\int_0^{+\infty}g_{1,\varepsilon}(s)\sigma_{\varepsilon,xxxx}(s)\,\mathrm{d}s=h^2,\\
\mu\sigma_\varepsilon+\partial_s\sigma_\varepsilon-\psi_\varepsilon=h^3,\\
\mu v_\varepsilon-\phi_\varepsilon=h^4,\\
\mu\phi_\varepsilon+\gamma_2\phi_\varepsilon-\bar{\mu}_2v_{\varepsilon,xx}-\int_0^{+\infty}g_{2,\varepsilon}(s)\eta_{\varepsilon,xx}(s)\,\mathrm{d}s=h^5,\\
\mu\eta_\varepsilon+\partial_s\eta_\varepsilon-\phi_\varepsilon=h^6, \tag*{(92)}
\end{cases} ⎩ ⎨ ⎧ μ u ε − ψ ε = h 1 , μ ψ ε + γ 1 ψ ε + μ ˉ 1 u ε , xxxx + ∫ 0 + ∞ g 1 , ε ( s ) σ ε , xxxx ( s ) d s = h 2 , μ σ ε + ∂ s σ ε − ψ ε = h 3 , μ v ε − ϕ ε = h 4 , μ ϕ ε + γ 2 ϕ ε − μ ˉ 2 v ε , xx − ∫ 0 + ∞ g 2 , ε ( s ) η ε , xx ( s ) d s = h 5 , μ η ε + ∂ s η ε − ϕ ε = h 6 , (92)
with the boundary conditions σ ε ( x , 0 ) = η ε ( x , 0 ) = 0 \sigma_\varepsilon(x,0)=\eta_\varepsilon(x,0)=0 σ ε ( x , 0 ) = η ε ( x , 0 ) = 0 . Here ( h 1 , … , h 6 ) (h^1,\ldots,h^6) ( h 1 , … , h 6 ) are the components of V V V in H ε \mathcal{H}_\varepsilon H ε , ordered as ( u , ψ , σ , v , ϕ , η ) (u,\psi,\sigma,v,\phi,\eta) ( u , ψ , σ , v , ϕ , η ) ; since V ∈ H e V\in\mathcal{H}_e V ∈ H e , its memory components vanish, i.e. h 3 = h 6 = 0 h^3=h^6=0 h 3 = h 6 = 0 . The solutions of the transport equations for σ ε , η ε \sigma_\varepsilon,\eta_\varepsilon σ ε , η ε can be written as
σ ε ( x , s ) = ∫ 0 s e − μ ( s − τ ) ( ψ ε ( x ) + h 3 ( x , τ ) ) d τ , η ε ( x , s ) = ∫ 0 s e − μ ( s − τ ) ( ϕ ε ( x ) + h 6 ( x , τ ) ) d τ , (93) \sigma_\varepsilon(x,s)=\int_0^s e^{-\mu(s-\tau)}(\psi_\varepsilon(x)+h^3(x,\tau))\,\mathrm{d}\tau,\qquad
\eta_\varepsilon(x,s)=\int_0^s e^{-\mu(s-\tau)}(\phi_\varepsilon(x)+h^6(x,\tau))\,\mathrm{d}\tau, \tag*{(93)} σ ε ( x , s ) = ∫ 0 s e − μ ( s − τ ) ( ψ ε ( x ) + h 3 ( x , τ )) d τ , η ε ( x , s ) = ∫ 0 s e − μ ( s − τ ) ( ϕ ε ( x ) + h 6 ( x , τ )) d τ , (93)
for s ≥ 0 s \ge0 s ≥ 0 . Substituting (93) into (92) , we obtain, for every φ ∈ H a 2 ( 0 , ℓ ) \varphi\in H_a^2(0,\ell) φ ∈ H a 2 ( 0 , ℓ ) , the identity
⟨ ∫ 0 + ∞ g 1 , ε ( s ) σ ε ( s ) d s , φ x x ⟩ = ∫ 0 + ∞ g 1 , ε ( s ) ∫ 0 s e − μ ( s − τ ) ⟨ ψ ε + h 3 ( τ ) , φ x x ⟩ d τ d s . (94) \left\langle\int_0^{+\infty} g_{1,\varepsilon}(s)\sigma_\varepsilon(s)\,\mathrm{d}s,\varphi_{xx}\right\rangle= \int_0^{+\infty} g_{1,\varepsilon}(s)\int_0^s e^{-\mu(s-\tau)}\left\langle\psi_\varepsilon+h^3(\tau),\varphi_{xx}\right\rangle\mathrm{d}\tau\mathrm{d}s. \tag*{(94)} ⟨ ∫ 0 + ∞ g 1 , ε ( s ) σ ε ( s ) d s , φ xx ⟩ = ∫ 0 + ∞ g 1 , ε ( s ) ∫ 0 s e − μ ( s − τ ) ⟨ ψ ε + h 3 ( τ ) , φ xx ⟩ d τ d s . (94)
and, analogously, for every φ ∈ H 0 1 ( 0 , ℓ ) \varphi\in H_0^1(0,\ell) φ ∈ H 0 1 ( 0 , ℓ ) ,
⟨ ∫ 0 + ∞ g 2 , ε ( s ) η ε ( s ) d s , φ x ⟩ = ∫ 0 + ∞ g 2 , ε ( s ) ∫ 0 s e − μ ( s − τ ) ⟨ ϕ ε + h 6 ( τ ) , φ x ⟩ d τ d s . (95) \left\langle\int_0^{+\infty} g_{2,\varepsilon}(s)\eta_\varepsilon(s)\,\mathrm{d}s,\varphi_x\right\rangle= \int_0^{+\infty} g_{2,\varepsilon}(s)\int_0^s e^{-\mu(s-\tau)}\left\langle\phi_\varepsilon+h^6(\tau),\varphi_x\right\rangle\mathrm{d}\tau\mathrm{d}s. \tag*{(95)} ⟨ ∫ 0 + ∞ g 2 , ε ( s ) η ε ( s ) d s , φ x ⟩ = ∫ 0 + ∞ g 2 , ε ( s ) ∫ 0 s e − μ ( s − τ ) ⟨ ϕ ε + h 6 ( τ ) , φ x ⟩ d τ d s . (95)
As ε → 0 + \varepsilon\to0^+ ε → 0 + , using the invariance of the mass (27) and the fact that g 1 , ε g_{1,\varepsilon} g 1 , ε concentrates near s = 0 s=0 s = 0 , the right-hand side of (94) tends to β 1 ⟨ ψ ε , φ x x ⟩ \beta_1\langle\psi_\varepsilon,\varphi_{xx}\rangle β 1 ⟨ ψ ε , φ xx ⟩ ; the contribution of h 3 h^3 h 3 vanishes in the limit, because the inner integral satisfies
∥ ∫ 0 s e − μ ( s − τ ) h 3 ( τ ) d τ ∥ H a 2 ≤ 1 − e − μ s μ ess sup τ ∈ ( 0 , s ) ∥ h 3 ( τ ) ∥ H a 2 = O ( s ) \left\|\int_0^s e^{-\mu(s-\tau)}h^3(\tau)\,\mathrm{d}\tau\right\|_{H_a^2} \le\frac{1-e^{-\mu s}}{\mu}\mathop{\operatorname{ess\,sup}}_{\tau\in(0,s)}\|h^3(\tau)\|_{H_a^2}=O(s) ∫ 0 s e − μ ( s − τ ) h 3 ( τ ) d τ H a 2 ≤ μ 1 − e − μ s ess sup τ ∈ ( 0 , s ) ∥ h 3 ( τ ) ∥ H a 2 = O ( s )
as s → 0 + s\to0^+ s → 0 + , where the essential supremum is finite for a.e. s > 0 s>0 s > 0 because h 3 ∈ L g 1 , ε 2 ( R + ; H a 2 ) h^3\in L^2_{g_{1,\varepsilon}}(\mathbb{R}_+;H_a^2) h 3 ∈ L g 1 , ε 2 ( R + ; H a 2 ) ; and g 1 , ε g_{1,\varepsilon} g 1 , ε concentrates at s = 0 s=0 s = 0 with total mass β 1 \beta_1 β 1 . Analogously, the right-hand side of (95) tends to β 2 ⟨ ϕ ε , φ x ⟩ \beta_2\langle\phi_\varepsilon,\varphi_x\rangle β 2 ⟨ ϕ ε , φ x ⟩ ; the contribution of h 6 h^6 h 6 vanishes in the limit, because
∥ ∫ 0 s e − μ ( s − τ ) h 6 ( τ ) d τ ∥ H 0 1 ≤ 1 − e − μ s μ ess sup τ ∈ ( 0 , s ) ∥ h 6 ( τ ) ∥ H 0 1 = O ( s ) \left\|\int_0^s e^{-\mu(s-\tau)}h^6(\tau)\,\mathrm{d}\tau\right\|_{H_0^1} \le\frac{1-e^{-\mu s}}{\mu}\mathop{\operatorname{ess\,sup}}_{\tau\in(0,s)}\|h^6(\tau)\|_{H_0^1}=O(s) ∫ 0 s e − μ ( s − τ ) h 6 ( τ ) d τ H 0 1 ≤ μ 1 − e − μ s ess sup τ ∈ ( 0 , s ) ∥ h 6 ( τ ) ∥ H 0 1 = O ( s )
as s → 0 + s\to0^+ s → 0 + , where the essential supremum is finite for a.e. s > 0 s>0 s > 0 because h 6 ∈ L g 2 , ε 2 ( R + ; H 0 1 ) h^6\in L^2_{g_{2,\varepsilon}}(\mathbb{R}_+;H_0^1) h 6 ∈ L g 2 , ε 2 ( R + ; H 0 1 ) ; and g 2 , ε g_{2,\varepsilon} g 2 , ε concentrates at s = 0 s=0 s = 0 with total mass β 2 \beta_2 β 2 . Consequently, passing to the limit in (92) in the weak sense, we obtain
{ μ u − ψ = h 1 , μ ψ + γ 1 ψ + μ 1 u x x x x = h 2 , μ v − ϕ = h 4 , μ ϕ + γ 2 ϕ − μ 2 v x x = h 5 . (96) \begin{cases}
\mu u-\psi=h^1,\\
\mu\psi+\gamma_1\psi+\mu_1u_{xxxx}=h^2,\\
\mu v-\phi=h^4,\\
\mu\phi+\gamma_2\phi-\mu_2v_{xx}=h^5. \tag*{(96)}
\end{cases} ⎩ ⎨ ⎧ μu − ψ = h 1 , μ ψ + γ 1 ψ + μ 1 u xxxx = h 2 , μv − ϕ = h 4 , μ ϕ + γ 2 ϕ − μ 2 v xx = h 5 . (96)
which is precisely the resolvent equation ( μ I + A e , 0 ) − 1 V = W = ( u , ψ , v , ϕ ) (\mu I+A_{e,0})^{-1}V=W=(u,\psi,v,\phi) ( μ I + A e , 0 ) − 1 V = W = ( u , ψ , v , ϕ ) . Here we used that V ∈ H e V\in H_e V ∈ H e has zero memory components, and that the contributions of the memory convolutions vanish in the limit, so that the remaining equations are exactly those of A e , 0 A_{e,0} A e , 0 on the elastic components ( u , ψ , v , ϕ ) (u,\psi,v,\phi) ( u , ψ , v , ϕ ) . The convergence W ε → W W_\varepsilon\to W W ε → W in H e H_e H e follows from the elliptic regularity theory for the fourth-order operator u ↦ μ ‾ 1 u x x x x u\mapsto\overline{\mu}_1u_{xxxx} u ↦ μ 1 u xxxx and the second-order operator v ↦ − μ ‾ 2 v x x v\mapsto-\overline{\mu}_2v_{xx} v ↦ − μ 2 v xx , both with ε \varepsilon ε -independent coefficients. This proves (91) and hence (90) .
Since s ↦ F ( U 0 ( s ) ) s\mapsto F(U_0(s)) s ↦ F ( U 0 ( s )) is continuous from [ 0 , T ] [0,T] [ 0 , T ] into H e H_e H e and [ 0 , T ] [0,T] [ 0 , T ] is compact, the family { F ( U 0 ( s ) ) : s ∈ [ 0 , T ] } \{F(U_0(s)):s\in[0,T]\} { F ( U 0 ( s )) : s ∈ [ 0 , T ]} is relatively compact in H e H_e H e . By (90) and a standard compactness argument, the convergence is uniform over this family:
sup s ∈ [ 0 , T ] sup τ ∈ [ 0 , T ] ∥ [ e − A e , ε τ − e − A e , 0 τ ] F ( U 0 ( s ) ) ∥ H e → 0 , ε → 0 + . \sup_{s\in[0,T]}\sup_{\tau\in[0,T]}\left\|\left[e^{-A_{e,\varepsilon}\tau}-e^{-A_{e,0}\tau}\right]F(U_0(s))\right\|_{H_e}\to0,\qquad\varepsilon\to0^+. s ∈ [ 0 , T ] sup τ ∈ [ 0 , T ] sup [ e − A e , ε τ − e − A e , 0 τ ] F ( U 0 ( s )) H e → 0 , ε → 0 + .
Consequently,
sup t ∈ [ 0 , T ] ∥ R e σ ( t ) ∥ H e ≤ ∫ 0 T sup s ∈ [ 0 , T ] sup τ ∈ [ 0 , T ] ∥ [ e − A e , ε τ − e − A e , 0 τ ] F ( U 0 ( s ) ) ∥ H e d s → 0. \sup_{t\in[0,T]}\|R_e^\sigma(t)\|_{H_e}\le\int_0^T\sup_{s\in[0,T]}\sup_{\tau\in[0,T]}\left\|\left[e^{-A_{e,\varepsilon}\tau}-e^{-A_{e,0}\tau}\right]F(U_0(s))\right\|_{H_e}\,\mathrm{d}s\to0. t ∈ [ 0 , T ] sup ∥ R e σ ( t ) ∥ H e ≤ ∫ 0 T s ∈ [ 0 , T ] sup τ ∈ [ 0 , T ] sup [ e − A e , ε τ − e − A e , 0 τ ] F ( U 0 ( s )) H e d s → 0.
(b) Memory component. Since F ( U 0 ( s ) ) F(U_0(s)) F ( U 0 ( s )) has zero memory components, e − A 0 τ F ( U 0 ( s ) ) e^{-A_0\tau}F(U_0(s)) e − A 0 τ F ( U 0 ( s )) has zero memory components for every τ ≥ 0 \tau\ge0 τ ≥ 0 . Hence
R ε σ ( t ) = ∫ 0 t [ e − A ε ( t − s ) F ( U 0 ( s ) ) ] σ d s . (97) R_\varepsilon^\sigma(t)=\int_0^t\left[e^{-A_\varepsilon(t-s)}F(U_0(s))\right]^\sigma\,\mathrm{d}s. \tag*{(97)} R ε σ ( t ) = ∫ 0 t [ e − A ε ( t − s ) F ( U 0 ( s )) ] σ d s . (97)
where [ ⋅ ] σ [\cdot]^\sigma [ ⋅ ] σ denotes the M g 1 , ε \mathcal{M}_{g_1,\varepsilon} M g 1 , ε -component. For Φ \Phi Φ with zero memory components, the memory component of e − A ε τ Φ e^{-A_\varepsilon\tau}\Phi e − A ε τ Φ is given by the transport equation: for ξ ≥ 0 \xi\ge0 ξ ≥ 0 ,
[ e − A ε τ Φ ] σ ( x , ξ ) = ∫ max { 0 , τ − ξ } τ [ e − A e , ε θ Φ ] ψ ( x ) d θ , (98) \left[e^{-A_\varepsilon\tau}\Phi\right]^\sigma(x,\xi)=\int_{\max\{0,\tau-\xi\}}^\tau\left[e^{-A_{e,\varepsilon}\theta}\Phi\right]^\psi(x)\,\mathrm{d}\theta, \tag*{(98)} [ e − A ε τ Φ ] σ ( x , ξ ) = ∫ m a x { 0 , τ − ξ } τ [ e − A e , ε θ Φ ] ψ ( x ) d θ , (98)
where [ ⋅ ] ψ [\cdot]^\psi [ ⋅ ] ψ denotes the elastic velocity component (the second component of H e \mathcal{H}_e H e ). By the analyticity of the semigroup e − A e , ε θ e^{-A_{e,\varepsilon}\theta} e − A e , ε θ and the elliptic regularity of A e , ε A_{e,\varepsilon} A e , ε (whose coefficients μ ˉ 1 , μ ˉ 2 \bar{\mu}_1,\bar{\mu}_2 μ ˉ 1 , μ ˉ 2 are independent of ε \varepsilon ε ), there exists C 1 > 0 C_1>0 C 1 > 0 , independent of ε \varepsilon ε , such that for every Φ ∈ H e \Phi\in\mathcal{H}_e Φ ∈ H e and θ > 0 \theta>0 θ > 0 ,
∥ [ e − A e , ε θ Φ ] ψ ∥ H a 2 ( 0 , ℓ ) ≤ C 1 θ − 1 / 2 ∥ Φ ∥ H e . (99) \left\|\left[e^{-A_{e,\varepsilon}\theta}\Phi\right]^\psi\right\|_{H_a^2(0,\ell)}\le C_1\theta^{-1/2}\|\Phi\|_{\mathcal{H}_e}. \tag*{(99)} [ e − A e , ε θ Φ ] ψ H a 2 ( 0 , ℓ ) ≤ C 1 θ − 1/2 ∥Φ ∥ H e . (99)
(Here we used the embedding D ( A e , ε 1 / 2 ) ↪ H a 2 ( 0 , ℓ ) D(A_{e,\varepsilon}^{1/2})\hookrightarrow H_a^2(0,\ell) D ( A e , ε 1/2 ) ↪ H a 2 ( 0 , ℓ ) for the second component, as in Step 2 of Proposition 4.1.) Combining (98) –(99) and using the Cauchy–Schwarz inequality,
∥ [ e − A ε τ Φ ] σ ∥ M g 1 , ε 2 = ∫ 0 + ∞ g 1 , ε ( ξ ) ∥ ∫ max { 0 , τ − ξ } τ [ e − A e , ε θ Φ ] ψ d θ ∥ H a 2 ( 0 , ℓ ) 2 d ξ ≤ C 1 2 ∥ Φ ∥ H e 2 ∫ 0 + ∞ g 1 , ε ( ξ ) ( ∫ max { 0 , τ − ξ } τ θ − 1 / 2 d θ ) 2 d ξ ≤ 4 C 1 2 ∥ Φ ∥ H e 2 ∫ 0 + ∞ g 1 , ε ( ξ ) min { τ , ξ } d ξ . (100) \begin{aligned}
\left\|\left[e^{-A_\varepsilon\tau}\Phi\right]^\sigma\right\|_{\mathcal{M}_{g_1,\varepsilon}}^2
&=\int_0^{+\infty}g_{1,\varepsilon}(\xi)\left\|\int_{\max\{0,\tau-\xi\}}^\tau\left[e^{-A_{e,\varepsilon}\theta}\Phi\right]^\psi\mathrm{d}\theta\right\|_{H_a^2(0,\ell)}^2\mathrm{d}\xi\\
&\le C_1^2\|\Phi\|_{\mathcal{H}_e}^2\int_0^{+\infty}g_{1,\varepsilon}(\xi)\left(\int_{\max\{0,\tau-\xi\}}^\tau\theta^{-1/2}\mathrm{d}\theta\right)^2\mathrm{d}\xi\\
&\le4C_1^2\|\Phi\|_{\mathcal{H}_e}^2\int_0^{+\infty}g_{1,\varepsilon}(\xi)\min\{\tau,\xi\}\mathrm{d}\xi. \tag*{(100)}
\end{aligned} [ e − A ε τ Φ ] σ M g 1 , ε 2 = ∫ 0 + ∞ g 1 , ε ( ξ ) ∫ m a x { 0 , τ − ξ } τ [ e − A e , ε θ Φ ] ψ d θ H a 2 ( 0 , ℓ ) 2 d ξ ≤ C 1 2 ∥Φ ∥ H e 2 ∫ 0 + ∞ g 1 , ε ( ξ ) ( ∫ m a x { 0 , τ − ξ } τ θ − 1/2 d θ ) 2 d ξ ≤ 4 C 1 2 ∥Φ ∥ H e 2 ∫ 0 + ∞ g 1 , ε ( ξ ) min { τ , ξ } d ξ . (100)
Using the change of variables ξ = ε ζ \xi=\varepsilon\zeta ξ = εζ and the exponential decay of g 1 g_1 g 1 in (9) ,
∫ 0 + ∞ g 1 , ε ( ξ ) min { τ , ξ } d ξ ≤ ∫ 0 + ∞ g 1 , ε ( ξ ) ξ d ξ = ε ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ , \int_0^{+\infty}g_{1,\varepsilon}(\xi)\min\{\tau,\xi\}\mathrm{d}\xi\le\int_0^{+\infty}g_{1,\varepsilon}(\xi)\xi\mathrm{d}\xi=\varepsilon\int_0^{+\infty}g_1(\zeta)\zeta\mathrm{d}\zeta, ∫ 0 + ∞ g 1 , ε ( ξ ) min { τ , ξ } d ξ ≤ ∫ 0 + ∞ g 1 , ε ( ξ ) ξ d ξ = ε ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ,
where ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ≤ g 1 ( 0 ) / λ 1 2 < + ∞ \int_0^{+\infty}g_1(\zeta)\zeta\mathrm{d}\zeta\le g_1(0)/\lambda_1^2<+\infty ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ≤ g 1 ( 0 ) / λ 1 2 < + ∞ . Therefore
∥ [ e − A ε τ Φ ] σ ∥ M g 1 , ε ≤ 2 C 1 ∥ Φ ∥ H e ⋅ ε 1 / 2 ⋅ ( ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ) 1 / 2 = O ( ε 1 / 2 ) , (101) \left\|\left[e^{-A_\varepsilon\tau}\Phi\right]^\sigma\right\|_{\mathcal{M}_{g_1,\varepsilon}}\le2C_1\|\Phi\|_{\mathcal{H}_e}\cdot\varepsilon^{1/2}\cdot\left(\int_0^{+\infty}g_1(\zeta)\zeta\mathrm{d}\zeta\right)^{1/2}=O(\varepsilon^{1/2}), \tag*{(101)} [ e − A ε τ Φ ] σ M g 1 , ε ≤ 2 C 1 ∥Φ ∥ H e ⋅ ε 1/2 ⋅ ( ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ) 1/2 = O ( ε 1/2 ) , (101)
uniformly in τ ≥ 0 \tau\ge0 τ ≥ 0 . Combining (97) and (101) ,
∥ R ε σ ( t ) ∥ M g 1 , ε ≤ 2 C 1 M 1 ε 1 1 / 2 T ( ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ) 1 / 2 , \left\|R_\varepsilon^\sigma(t)\right\|_{\mathcal{M}_{g_1,\varepsilon}}\le2C_1M_1\varepsilon_1^{1/2}T\left(\int_0^{+\infty}g_1(\zeta)\zeta\mathrm{d}\zeta\right)^{1/2}, ∥ R ε σ ( t ) ∥ M g 1 , ε ≤ 2 C 1 M 1 ε 1 1/2 T ( ∫ 0 + ∞ g 1 ( ζ ) ζ d ζ ) 1/2 ,
which tends to 0 as ε → 0 + \varepsilon\to0^+ ε → 0 + , uniformly in t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . The same argument gives the vanishing of R ε η R_\varepsilon^\eta R ε η in M g 2 , ε \mathcal{M}_{g_2,\varepsilon} M g 2 , ε . Combining (a) and (b), we obtain (89) .
Step 3: Gronwall argument. Taking the norm of (86) and using the uniform exponential stability (26) , the Lipschitz bound (88) , and (89) , we obtain
∥ W ε ( t ) ∥ H ε ≤ α e − β t ∥ U 0 , ε − U 0 ∥ H ε + α L 1 ∫ 0 t e − β ( t − s ) ∥ W ε ( s ) ∥ H ε d s + ∥ R ε ( t ) ∥ H ε . \|W_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le\alpha e^{-\beta t}\|U_{0,\varepsilon}-U_0\|_{\mathcal{H}_\varepsilon}+\alpha L_1\int_0^t e^{-\beta(t-s)}\|W_\varepsilon(s)\|_{\mathcal{H}_\varepsilon}\,\mathrm{d}s+\|R_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}. ∥ W ε ( t ) ∥ H ε ≤ α e − β t ∥ U 0 , ε − U 0 ∥ H ε + α L 1 ∫ 0 t e − β ( t − s ) ∥ W ε ( s ) ∥ H ε d s + ∥ R ε ( t ) ∥ H ε .
Multiplying by e β t e^{\beta t} e β t and applying Gronwall’s inequality,
∥ W ε ( t ) ∥ H ε ≤ ( α ∥ U 0 , ε − U 0 ∥ H ε + sup s ∈ [ 0 , T ] ∥ R ε ( s ) ∥ H ε ) e α L 1 T , t ∈ [ 0 , T ] . \|W_\varepsilon(t)\|_{\mathcal{H}_\varepsilon}\le\left(\alpha\|U_{0,\varepsilon}-U_0\|_{\mathcal{H}_\varepsilon}+\sup_{s\in[0,T]}\|R_\varepsilon(s)\|_{\mathcal{H}_\varepsilon}\right)e^{\alpha L_1T},\qquad t\in[0,T]. ∥ W ε ( t ) ∥ H ε ≤ ( α ∥ U 0 , ε − U 0 ∥ H ε + s ∈ [ 0 , T ] sup ∥ R ε ( s ) ∥ H ε ) e α L 1 T , t ∈ [ 0 , T ] .
Since ∥ U 0 , ε − U 0 ∥ H ε → 0 \|U_{0,\varepsilon}-U_0\|_{\mathcal{H}_\varepsilon}\to0 ∥ U 0 , ε − U 0 ∥ H ε → 0 and sup s ∈ [ 0 , T ] ∥ R ε ( s ) ∥ H ε → 0 \sup_{s\in[0,T]}\|R_\varepsilon(s)\|_{\mathcal{H}_\varepsilon}\to0 sup s ∈ [ 0 , T ] ∥ R ε ( s ) ∥ H ε → 0 as ε → 0 + \varepsilon\to0^+ ε → 0 + , we conclude that
sup t ∈ [ 0 , T ] ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0 , \sup_{t\in[0,T]}\|S_\varepsilon(t)U_{0,\varepsilon}-S_0(t)U_0\|_{\mathcal{H}_\varepsilon}\to0, t ∈ [ 0 , T ] sup ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0 ,
which is the desired convergence.
Conclusion of upper semicontinuity
Combining Propositions 3.1, 4.1, and 4.2, we can apply the standard abstract theorem on upper semicontinuity of attractors (see, e.g., [12 ] or [5 Chapter 7] ).
Theorem 4.3 (Upper semicontinuity). Let (11) and (9) hold. Then, for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , the semigroup S ε ( t ) S_{\varepsilon}(t) S ε ( t ) possesses a global attractor A ε \mathcal{A}_{\varepsilon} A ε , the limit semigroup S 0 ( t ) S_{0}(t) S 0 ( t ) possesses a global attractor A 0 \mathcal{A}_{0} A 0 , and the family { A ε } ε ∈ ( 0 , 1 ] \{\mathcal{A}_{\varepsilon}\}_{\varepsilon\in(0,1]} { A ε } ε ∈ ( 0 , 1 ] satisfies
lim ε → 0 + dist H ε ( A ε , A 0 ) = 0. \lim_{\varepsilon\to0^{+}}\operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_{\varepsilon},\mathcal{A}_{0})=0. ε → 0 + lim dist H ε ( A ε , A 0 ) = 0.
Proof. The proof is a direct application of the abstract upper semicontinuity theorem [5 Chapter 7] . We first record that the semigroups S ε ( t ) S_{\varepsilon}(t) S ε ( t ) , ε ∈ [ 0 , 1 ] \varepsilon\in[0,1] ε ∈ [ 0 , 1 ] , are continuous on their respective spaces H ε \mathcal{H}_{\varepsilon} H ε for ε > 0 \varepsilon> 0 ε > 0 and H e \mathcal{H}_{e} H e for ε = 0 \varepsilon= 0 ε = 0 : for ε > 0 \varepsilon> 0 ε > 0 this follows from Theorem 2.2 (the operator − A ε -A_{\varepsilon} − A ε generates a C 0 C_{0} C 0 -semigroup of contractions on H ε \mathcal{H}_{\varepsilon} H ε ), and for ε = 0 \varepsilon= 0 ε = 0 from the standard theory for Lipschitz perturbations of C 0 C_{0} C 0 -semigroups applied to the limit operator A 0 A_{0} A 0 on H e \mathcal{H}_{e} H e . We then verify the three hypotheses of the theorem.
( H 1 ) (H1) ( H 1 ) Uniform dissipativity. By Proposition 3.1 , there exists a bounded absorbing set B 0 \mathcal{B}_{0} B 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that S ε ( t ) B ⊂ B 0 S_{\varepsilon}(t)B \subset\mathcal{B}_{0} S ε ( t ) B ⊂ B 0 for every bounded B B B and every t ≥ t 0 ( B ) t \ge t_{0}(B) t ≥ t 0 ( B ) . Moreover, the limit semigroup S 0 ( t ) S_{0}(t) S 0 ( t ) also possesses a bounded absorbing set: this follows either from the same energy estimates applied to the limit system (which is obtained by replacing the memory convolutions with the local terms − β 1 u x x x x -\beta_{1}u_{xxxx} − β 1 u xxxx and + β 2 v x x +\beta_{2}v_{xx} + β 2 v xx , equivalently by letting g i , ε ⇀ β i δ 0 g_{i,\varepsilon} \rightharpoonup\beta_{i}\delta_{0} g i , ε ⇀ β i δ 0 ), or by passing to the limit ε → 0 + \varepsilon\to0^{+} ε → 0 + in the uniform absorbing estimate.
( H 2 ) (H2) ( H 2 ) Uniform asymptotic smoothness. By Proposition 4.1 , the semigroups S ε ( t ) S_{\varepsilon}(t) S ε ( t ) , ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , satisfy the uniform quasi-stability inequality
∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , U 0 , V 0 ∈ B 0 , \left\|S_{\varepsilon}(T)U_{0}-S_{\varepsilon}(T)V_{0}\right\|_{\mathcal{H}_{\varepsilon}}^{2}\le q^{2}\left\|U_{0}-V_{0}\right\|_{\mathcal{H}_{\varepsilon}}^{2}+\left\|K_{\varepsilon,T}(U_{0})-K_{\varepsilon,T}(V_{0})\right\|_{\mathcal{H}_{\varepsilon}}^{2},\qquad U_{0},V_{0}\in\mathcal{B}_{0}, ∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , U 0 , V 0 ∈ B 0 ,
where q < 1 q<1 q < 1 and T > 0 T>0 T > 0 , and K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_{0}) K ε , T ( B 0 ) satisfies the sequential compactness property of Proposition 4.1 , all independent of ε \varepsilon ε . By [5 Chapter 7, Theorem 7.9.4] , the quasi-stability inequality, together with the uniform dissipativity (H1), implies that the semigroups S ε ( t ) S_{\varepsilon}(t) S ε ( t ) are asymptotically smooth. The proof of [5 Chapter 7, Theorem 7.9.4] uses only the constants appearing in the quasi-stability inequality and the dissipativity estimate, both of which are independent of ε \varepsilon ε by Proposition 4.1 and Proposition 3.1 ; hence the resulting asymptotic smoothness is uniform in ε \varepsilon ε . For the limit semigroup S 0 ( t ) S_{0}(t) S 0 ( t ) , the existence of a compact global attractor A 0 \mathcal{A}_{0} A 0 , invariant under S 0 ( t ) S_{0}(t) S 0 ( t ) and attracting every bounded subset, is established in [20 ] ; hence no separate asymptotic-smoothness argument is needed for S 0 S_{0} S 0 .
( H 3 ) (H3) ( H 3 ) Finite-time convergence. By Proposition 4.2 , for every T > 0 T>0 T > 0 , every bounded B B B , and every sequence U 0 , ε ∈ B U_{0,\varepsilon}\in B U 0 , ε ∈ B with U 0 , ε → U 0 U_{0,\varepsilon}\to U_{0} U 0 , ε → U 0 ,
sup t ∈ [ 0 , T ] ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0. \sup_{t\in[0,T]}\left\|S_{\varepsilon}(t)U_{0,\varepsilon}-S_{0}(t)U_{0}\right\|_{\mathcal{H}_{\varepsilon}}\to0. t ∈ [ 0 , T ] sup ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0.
By [5 Chapter 7] , hypotheses (H1)–(H3) imply that for every sequence ε n → 0 + \varepsilon_{n}\to0^{+} ε n → 0 + ,
dist H ε n ( A ε n , A 0 ) → 0. \operatorname{dist}_{\mathcal{H}_{\varepsilon_{n}}}(\mathcal{A}_{\varepsilon_{n}},\mathcal{A}_{0})\to0. dist H ε n ( A ε n , A 0 ) → 0.
Since ε n → 0 + \varepsilon_{n}\to0^{+} ε n → 0 + is arbitrary, we obtain the upper semicontinuity
lim ε → 0 + dist H ε ( A ε , A 0 ) = 0. \lim_{\varepsilon\to0^{+}}\operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_{\varepsilon},\mathcal{A}_{0})=0. ε → 0 + lim dist H ε ( A ε , A 0 ) = 0.
Lower Semicontinuity
Lower semicontinuity is more delicate. It requires that every point of A 0 \mathcal{A}_0 A 0 can be approximated by points of A ε \mathcal{A}_\varepsilon A ε . A standard approach is to combine the uniform exponential attraction property with a uniform Hölder continuity of the semigroup with respect to the parameter ε \varepsilon ε . Throughout this section we work on compact subintervals [ ε 0 , 1 ] ⊂ ( 0 , 1 ] [\varepsilon_0,1]\subset(0,1] [ ε 0 , 1 ] ⊂ ( 0 , 1 ] , on which the weighted norms ∥ ⋅ ∥ M g i , ε \|\cdot\|_{\mathcal{M}_{g_i,\varepsilon}} ∥ ⋅ ∥ M g i , ε are uniformly equivalent (see Assumption 5.3 and Proposition 5.4 ); the passage to ε 0 → 0 + \varepsilon_0\to0^+ ε 0 → 0 + is handled by a diagonal argument.
Exponential attraction
We begin with the notion of exponential attraction.
Definition 5.1. Let X X X be a Hilbert space, let A ⊂ X \mathcal{A}\subset X A ⊂ X be a compact global attractor for a continuous semigroup S ( t ) S(t) S ( t ) on X X X , and let B ⊂ X B\subset X B ⊂ X be bounded. We say that A \mathcal{A} A exponentially attracts B B B if there exist constants M B ≥ 1 M_B\ge1 M B ≥ 1 and ω > 0 \omega>0 ω > 0 such that
dist X ( S ( t ) B , A ) ≤ M B e − ω t , ∀ t ≥ 0. (102) \operatorname{dist}_{X}(S(t)B,\mathcal{A})\le M_B e^{-\omega t},\qquad\forall t\ge0. \tag*{(102)} dist X ( S ( t ) B , A ) ≤ M B e − ω t , ∀ t ≥ 0. (102)
For the perturbed system, we have the following uniform version.
Proposition 5.2 (Uniform exponential attraction ). There exist constants M ≥ 1 M\ge1 M ≥ 1 and ω > 0 \omega>0 ω > 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that the following holds: for every bounded set B ⊂ H ε B\subset\mathcal{H}_\varepsilon B ⊂ H ε there exists M B ≥ 1 M_B\ge1 M B ≥ 1 , depending on B B B but not on ε \varepsilon ε , such that
dist H ε ( S ε ( t ) B , A ε ) ≤ M B e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (103) \operatorname{dist}_{\mathcal{H}_\varepsilon}(S_\varepsilon(t)B,\mathcal{A}_\varepsilon)\le M_B e^{-\omega t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1]. \tag*{(103)} dist H ε ( S ε ( t ) B , A ε ) ≤ M B e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (103) In particular, if B ⊂ B 0 B\subset\mathcal{B}_0 B ⊂ B 0 (the absorbing ball of Proposition 3.1 ), then one can take M B = M M_B=M M B = M , so that
dist H ε ( S ε ( t ) B , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (104) \operatorname{dist}_{\mathcal{H}_\varepsilon}(S_\varepsilon(t)B,\mathcal{A}_\varepsilon)\le M e^{-\omega t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1]. \tag*{(104)} dist H ε ( S ε ( t ) B , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . (104)
Proof. The proof is a direct application of the Chueshov–Lasiecka theory on exponential attraction for quasi-stable semigroups [5 Chapter 7] . We proceed in two steps.
Step 1: Exponential attraction on the absorbing ball. By Proposition 4.1 , the semigroups S ε ( t ) S_\varepsilon(t) S ε ( t ) , ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , satisfy the uniform quasi-stability inequality (64) , namely
∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , U 0 , V 0 ∈ B 0 , \|S_\varepsilon(T)U_0-S_\varepsilon(T)V_0\|_{\mathcal{H}_\varepsilon}^{2}\le q^{2}\|U_0-V_0\|_{\mathcal{H}_\varepsilon}^{2}+\|K_{\varepsilon,T}(U_0)-K_{\varepsilon,T}(V_0)\|_{\mathcal{H}_\varepsilon}^{2},\qquad U_0,V_0\in\mathcal{B}_0, ∥ S ε ( T ) U 0 − S ε ( T ) V 0 ∥ H ε 2 ≤ q 2 ∥ U 0 − V 0 ∥ H ε 2 + ∥ K ε , T ( U 0 ) − K ε , T ( V 0 ) ∥ H ε 2 , U 0 , V 0 ∈ B 0 ,
where q ∈ [ 0 , 1 ) q\in[0,1) q ∈ [ 0 , 1 ) and T > 0 T>0 T > 0 , and K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_0) K ε , T ( B 0 ) satisfies the sequential compactness property of Proposition 4.1 , all independent of ε \varepsilon ε .
Moreover, by the Chueshov–Lasiecka existence theorem [5 Chapter 7] , the uniform quasi-stability inequality (64) together with the uniform dissipativity of Proposition 3.1 implies that for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] the semigroup S ε ( t ) S_\varepsilon(t) S ε ( t ) possesses a compact global attractor A ε ⊂ B 0 \mathcal{A}_\varepsilon\subset\mathcal{B}_0 A ε ⊂ B 0 , which is invariant under S ε ( t ) S_\varepsilon(t) S ε ( t ) and attracts every bounded subset.
We claim that the quasi-stability inequality (64) together with the compactness of A ε \mathcal{A}_\varepsilon A ε implies the exponential attraction estimate
dist H ε ( S ε ( t ) B 0 , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] , (105) \operatorname{dist}_{\mathcal{H}_\varepsilon}(S_\varepsilon(t)\mathcal{B}_0,\mathcal{A}_\varepsilon)\le M e^{-\omega t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1], \tag*{(105)} dist H ε ( S ε ( t ) B 0 , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] , (105)
with constants M ≥ 1 M\ge1 M ≥ 1 and ω > 0 \omega>0 ω > 0 independent of ε \varepsilon ε . This is precisely the content of [5 Chapter 7] , applied to the semigroup S ε ( t ) S_\varepsilon(t) S ε ( t ) with the quasi-stability inequality (64) and the attractor A ε \mathcal{A}_\varepsilon A ε . We verify that the hypotheses of that theorem are satisfied uniformly in ε \varepsilon ε :
Quasi-stability. The inequality (64) holds with q q q and T T T independent of ε \varepsilon ε , and K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_0) K ε , T ( B 0 ) satisfies the sequential compactness property of Proposition 4.1 ;
(ii) Existence of the attractor. The attractor A ε \mathcal{A}_{\varepsilon} A ε exists, is compact, and is invariant under S ε ( t ) S_{\varepsilon}(t) S ε ( t ) [5 Chapter 7] ;
(iii) Attraction of the absorbing ball. The attractor A ε \mathcal{A}_{\varepsilon} A ε attracts the absorbing ball B 0 \mathcal{B}_{0} B 0 : that is, dist H ε ( S ε ( t ) B 0 , A ε ) → 0 \operatorname{dist}_{\mathcal{H}_{\varepsilon}}(S_{\varepsilon}(t)\mathcal{B}_{0},\mathcal{A}_{\varepsilon}) \to0 dist H ε ( S ε ( t ) B 0 , A ε ) → 0 as t → + ∞ t \to+\infty t → + ∞ .
Since all the constants appearing in [5 Chapter 7] depend only on q q q , T T T , the sequential compactness property of K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_{0}) K ε , T ( B 0 ) , the diameter of B 0 \mathcal{B}_{0} B 0 , and the attracting properties of A ε \mathcal{A}_{\varepsilon} A ε , and since all of these quantities are independent of ε \varepsilon ε by Propositions 3.1 and 4.1, we conclude that the constants M M M and ω \omega ω in (105) can be chosen independently of ε \varepsilon ε .
Step 2: Extension to general bounded sets. Let B ⊂ H ε B \subset\mathcal{H}_{\varepsilon} B ⊂ H ε be an arbitrary bounded set. By Proposition 3.1, there exists t 0 = t 0 ( B ) ≥ 0 t_{0}=t_{0}(B)\geq0 t 0 = t 0 ( B ) ≥ 0 , independent of ε \varepsilon ε , such that S ε ( t 0 ) B ⊂ B 0 S_{\varepsilon}(t_{0})B\subset\mathcal{B}_{0} S ε ( t 0 ) B ⊂ B 0 for all ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . For t ≥ t 0 t\geq t_{0} t ≥ t 0 , using the semigroup property,
S ε ( t ) B = S ε ( t − t 0 ) ( S ε ( t 0 ) B ) ⊂ S ε ( t − t 0 ) B 0 , S_{\varepsilon}(t)B=S_{\varepsilon}(t-t_{0})(S_{\varepsilon}(t_{0})B)\subset S_{\varepsilon}(t-t_{0})\mathcal{B}_{0}, S ε ( t ) B = S ε ( t − t 0 ) ( S ε ( t 0 ) B ) ⊂ S ε ( t − t 0 ) B 0 ,
so that, by (105) ,
dist H ε ( S ε ( t ) B , A ε ) ≤ dist H ε ( S ε ( t − t 0 ) B 0 , A ε ) ≤ M e − ω ( t − t 0 ) = M e ω t 0 e − ω t . \operatorname{dist}_{\mathcal{H}_{\varepsilon}}(S_{\varepsilon}(t)B,\mathcal{A}_{\varepsilon})\leq\operatorname{dist}_{\mathcal{H}_{\varepsilon}}(S_{\varepsilon}(t-t_{0})\mathcal{B}_{0},\mathcal{A}_{\varepsilon})\leq Me^{-\omega(t-t_{0})}=Me^{\omega t_{0}}e^{-\omega t}. dist H ε ( S ε ( t ) B , A ε ) ≤ dist H ε ( S ε ( t − t 0 ) B 0 , A ε ) ≤ M e − ω ( t − t 0 ) = M e ω t 0 e − ω t .
For 0 ≤ t < t 0 0\leq t<t_{0} 0 ≤ t < t 0 , the set S ε ( t ) B S_{\varepsilon}(t)B S ε ( t ) B is bounded uniformly in ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] by Lemma 3.2. Consequently the distance is bounded by
D B : = sup 0 ≤ t ≤ t 0 diam ( S ε ( t ) B ∪ B 0 ) , D_{B}:=\sup_{0\leq t\leq t_{0}}\operatorname{diam}(S_{\varepsilon}(t)B\cup\mathcal{B}_{0}), D B := 0 ≤ t ≤ t 0 sup diam ( S ε ( t ) B ∪ B 0 ) ,
which is finite and independent of ε \varepsilon ε . Setting
M B : = e ω t 0 max { M , D B } , M_{B}:=e^{\omega t_{0}}\max\{M,D_{B}\}, M B := e ω t 0 max { M , D B } ,
we obtain
dist H ε ( S ε ( t ) B , A ε ) ≤ M B e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . \operatorname{dist}_{\mathcal{H}_{\varepsilon}}(S_{\varepsilon}(t)B,\mathcal{A}_{\varepsilon})\leq M_{B}e^{-\omega t},\qquad\forall t\geq0,\ \forall\varepsilon\in(0,1]. dist H ε ( S ε ( t ) B , A ε ) ≤ M B e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] .
Indeed, for t ≥ t 0 t\geq t_{0} t ≥ t 0 we have M e ω t 0 e − ω t ≤ M B e − ω t Me^{\omega t_{0}}e^{-\omega t}\leq M_{B}e^{-\omega t} M e ω t 0 e − ω t ≤ M B e − ω t , while for 0 ≤ t < t 0 0\leq t<t_{0} 0 ≤ t < t 0 we have D B ≤ M B e − ω t 0 ≤ M B e − ω t D_{B}\leq M_{B}e^{-\omega t_{0}}\leq M_{B}e^{-\omega t} D B ≤ M B e − ω t 0 ≤ M B e − ω t . This proves (103) . The constant M B M_{B} M B depends on B B B (through the diameter of S ε ( [ 0 , t 0 ] ) B ∪ B 0 S_{\varepsilon}([0,t_{0}])B\cup\mathcal{B}_{0} S ε ([ 0 , t 0 ]) B ∪ B 0 and the time t 0 ( B ) t_{0}(B) t 0 ( B ) ) but not on ε \varepsilon ε , while ω \omega ω and M M M are independent of both B B B and ε \varepsilon ε .
The second ingredient is a continuity estimate of the form: for every T > 0 T>0 T > 0 , every bounded set B B B , and every ε 0 ∈ ( 0 , 1 ] \varepsilon_{0}\in(0,1] ε 0 ∈ ( 0 , 1 ] , there exists L T ( B , ε 0 ) > 0 L_{T}(B,\varepsilon_{0})>0 L T ( B , ε 0 ) > 0 such that
∥ S ε ( t ) U 0 − S ε ′ ( t ) U 0 ∥ H ε ≤ L T ( B , ε 0 ) ∣ ε − ε ′ ∣ 1 / 2 , (106) \lVert S_{\varepsilon}(t)U_{0}-S_{\varepsilon'}(t)U_{0}\rVert_{\mathcal{H}_{\varepsilon}}\leq L_{T}(B,\varepsilon_{0})\lvert\varepsilon-\varepsilon'\rvert^{1/2}, \tag*{(106)} ∥ S ε ( t ) U 0 − S ε ′ ( t ) U 0 ∥ H ε ≤ L T ( B , ε 0 ) ∣ ε − ε ′ ∣ 1/2 , (106)
for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , U 0 ∈ B U_{0}\in B U 0 ∈ B , and ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon'\in[\varepsilon_{0},1] ε , ε ′ ∈ [ ε 0 , 1 ] .
The key observation is twofold. On the one hand, on compact subintervals of ( 0 , 1 ] (0,1] ( 0 , 1 ] , the rescaled kernels g i , ε g_{i,\varepsilon} g i , ε depend continuously on ε \varepsilon ε in the L 1 L^{1} L 1 -norm. Precisely, we assume that for every ε 0 ∈ ( 0 , 1 ] \varepsilon_{0}\in(0,1] ε 0 ∈ ( 0 , 1 ] there exists C i ( ε 0 ) > 0 C_{i}(\varepsilon_{0})>0 C i ( ε 0 ) > 0 such that
∥ g i , ε − g i , ε ′ ∥ L 1 ( R + ) ≤ C i ( ε 0 ) ∣ ε − ε ′ ∣ , ε , ε ′ ∈ [ ε 0 , 1 ] , i = 1 , 2. (107) \lVert g_{i,\varepsilon}-g_{i,\varepsilon'}\rVert_{L^{1}(\mathbb{R}_{+})}\leq C_{i}(\varepsilon_{0})\lvert\varepsilon-\varepsilon'\rvert,\qquad\varepsilon,\varepsilon'\in[\varepsilon_{0},1],\ i=1,2. \tag*{(107)} ∥ g i , ε − g i , ε ′ ∥ L 1 ( R + ) ≤ C i ( ε 0 ) ∣ ε − ε ′ ∣ , ε , ε ′ ∈ [ ε 0 , 1 ] , i = 1 , 2. (107)
This assumption is understood on compact subintervals [ ε 0 , 1 ] ⊂ ( 0 , 1 ] [\varepsilon_{0},1]\subset(0,1] [ ε 0 , 1 ] ⊂ ( 0 , 1 ] . It is satisfied, for instance, when g i ∈ W 1 , 1 ( R + ) g_{i}\in W^{1,1}(\mathbb{R}_{+}) g i ∈ W 1 , 1 ( R + ) with ∫ 0 + ∞ ( 1 + s ) ∣ g i ′ ( s ) ∣ d s < + ∞ \int_{0}^{+\infty}(1+s)\lvert g_{i}'(s)\rvert\,\mathrm{d}s<+\infty ∫ 0 + ∞ ( 1 + s ) ∣ g i ′ ( s )∣ d s < + ∞ : indeed, by the change of variables s = ε ζ s=\varepsilon\zeta s = εζ ,
∥ ∂ ε g i , ε ∥ L 1 ( R + ) ≤ ε − 1 ∥ g i ∥ L 1 ( R + ) + ε − 2 ∫ 0 + ∞ ζ ∣ g i ′ ( ζ ) ∣ d ζ , \lVert\partial_{\varepsilon}g_{i,\varepsilon}\rVert_{L^{1}(\mathbb{R}_{+})}\leq\varepsilon^{-1}\lVert g_{i}\rVert_{L^{1}(\mathbb{R}_{+})}+\varepsilon^{-2}\int_{0}^{+\infty}\zeta\lvert g_{i}'(\zeta)\rvert\,\mathrm{d}\zeta, ∥ ∂ ε g i , ε ∥ L 1 ( R + ) ≤ ε − 1 ∥ g i ∥ L 1 ( R + ) + ε − 2 ∫ 0 + ∞ ζ ∣ g i ′ ( ζ )∣ d ζ ,
which is bounded uniformly for ε ∈ [ ε 0 , 1 ] \varepsilon\in[\varepsilon_0,1] ε ∈ [ ε 0 , 1 ] . In particular, (107) holds for the exponential kernel g i ( s ) = b i e − a i s g_i(s)=b_i e^{-a_i s} g i ( s ) = b i e − a i s on every [ ε 0 , 1 ] [\varepsilon_0,1] [ ε 0 , 1 ] , with C i ( ε 0 ) C_i(\varepsilon_0) C i ( ε 0 ) blowing up as ε 0 → 0 + \varepsilon_0 \to0^+ ε 0 → 0 + . On the other hand, the weighted norms M g i , ε \mathcal{M}_{g_i,\varepsilon} M g i , ε corresponding to different values of ε ∈ [ ε 0 , 1 ] \varepsilon\in[\varepsilon_0,1] ε ∈ [ ε 0 , 1 ] must be uniformly equivalent, so that the memory component can be controlled in a single reference norm.
Assumption 5.3. There exist constants c 0 > 0 c_0>0 c 0 > 0 and λ 0 > 0 \lambda_0>0 λ 0 > 0 such that the kernels g i g_i g i satisfy
g i ( s ) ≥ c 0 e − λ 0 s , a.e. s ≥ 0 , i = 1 , 2. (108) g_i(s) \ge c_0 e^{-\lambda_0 s}, \qquad\text{a.e. }s \ge0,\ i=1,2. \tag*{(108)} g i ( s ) ≥ c 0 e − λ 0 s , a.e. s ≥ 0 , i = 1 , 2. (108)
This assumption is compatible with (9) provided c 0 ≤ g i ( 0 ) c_0 \le g_i(0) c 0 ≤ g i ( 0 ) and λ i ≤ λ 0 \lambda_i \le\lambda_0 λ i ≤ λ 0 : it holds, for instance, for any kernel of the form g i ( s ) = b i e − a i s g_i(s)=b_i e^{-a_i s} g i ( s ) = b i e − a i s with b i ≥ c 0 b_i \ge c_0 b i ≥ c 0 and λ i ≤ a i ≤ λ 0 \lambda_i \le a_i \le\lambda_0 λ i ≤ a i ≤ λ 0 . It ensures that, for every pair ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon' \in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] with ε 0 > 0 \varepsilon_0>0 ε 0 > 0 , the weighted norms ∥ ⋅ ∥ M g i , ε \|\cdot\|_{\mathcal{M}_{g_i,\varepsilon}} ∥ ⋅ ∥ M g i , ε and ∥ ⋅ ∥ M g i , ε ′ \|\cdot\|_{\mathcal{M}_{g_i,\varepsilon'}} ∥ ⋅ ∥ M g i , ε ′ are uniformly equivalent, with a constant depending only on ε 0 \varepsilon_0 ε 0 , c 0 c_0 c 0 , λ 0 \lambda_0 λ 0 , λ i \lambda_i λ i , g i ( 0 ) g_i(0) g i ( 0 ) , and ℓ \ell ℓ .
This assumption is not a consequence of the standing hypotheses (9) ; it is an extra hypothesis used only in the lower semicontinuity analysis.
Proposition 5.4 (Hölder continuity in ε \varepsilon ε ). Assume (9) , (107) , and Assumption 5.3 . Let B B B be bounded and let ε 0 ∈ ( 0 , 1 ] \varepsilon_0 \in(0,1] ε 0 ∈ ( 0 , 1 ] . For every T > 0 T>0 T > 0 there exists L T ( B , ε 0 ) > 0 L_T(B,\varepsilon_0)>0 L T ( B , ε 0 ) > 0 such that (106) holds for all t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , U 0 ∈ B U_0 \in B U 0 ∈ B , and ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon' \in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] .
Proof. Let ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon' \in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] and U 0 ∈ B U_0 \in B U 0 ∈ B be fixed. Set U ε ( t ) : = S ε ( t ) U 0 U_\varepsilon(t):=S_\varepsilon(t)U_0 U ε ( t ) := S ε ( t ) U 0 , U ε ′ ( t ) : = S ε ′ ( t ) U 0 U_{\varepsilon'}(t):=S_{\varepsilon'}(t)U_0 U ε ′ ( t ) := S ε ′ ( t ) U 0 , and W ( t ) : = U ε ( t ) − U ε ′ ( t ) W(t):=U_\varepsilon(t)-U_{\varepsilon'}(t) W ( t ) := U ε ( t ) − U ε ′ ( t ) . Write U ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) U_\varepsilon=(u_\varepsilon,u_{\varepsilon,t},\sigma_\varepsilon,v_\varepsilon,v_{\varepsilon,t},\eta_\varepsilon) U ε = ( u ε , u ε , t , σ ε , v ε , v ε , t , η ε ) and similarly for U ε ′ U_{\varepsilon'} U ε ′ . By the mild formulation of both solutions, we have
U ε ( t ) = e − A ε t U 0 + ∫ 0 t e − A ε ( t − s ) F ( U ε ( s ) ) d s , U_\varepsilon(t)=e^{-A_\varepsilon t}U_0+\int_0^t e^{-A_\varepsilon(t-s)}F(U_\varepsilon(s))\,\mathrm{d}s, U ε ( t ) = e − A ε t U 0 + ∫ 0 t e − A ε ( t − s ) F ( U ε ( s )) d s ,
U ε ′ ( t ) = e − A ε ′ t U 0 + ∫ 0 t e − A ε ′ ( t − s ) F ( U ε ′ ( s ) ) d s . U_{\varepsilon'}(t)=e^{-A_{\varepsilon'}t}U_0+\int_0^t e^{-A_{\varepsilon'}(t-s)}F(U_{\varepsilon'}(s))\,\mathrm{d}s. U ε ′ ( t ) = e − A ε ′ t U 0 + ∫ 0 t e − A ε ′ ( t − s ) F ( U ε ′ ( s )) d s .
Subtracting and using the variation-of-constants formula, we obtain
W ( t ) = [ e − A ε t − e − A ε ′ t ] U 0 + ∫ 0 t e − A ε ( t − s ) [ F ( U ε ( s ) ) − F ( U ε ′ ( s ) ) ] d s + ∫ 0 t [ e − A ε ( t − s ) − e − A ε ′ ( t − s ) ] F ( U ε ′ ( s ) ) d s . (109) \begin{aligned}
W(t)={}&\left[e^{-A_\varepsilon t}-e^{-A_{\varepsilon'}t}\right]U_0\\
&+\int_0^t e^{-A_\varepsilon(t-s)}\left[F(U_\varepsilon(s))-F(U_{\varepsilon'}(s))\right]\,\mathrm{d}s\\
&+\int_0^t\left[e^{-A_\varepsilon(t-s)}-e^{-A_{\varepsilon'}(t-s)}\right]F(U_{\varepsilon'}(s))\,\mathrm{d}s. \tag*{(109)}
\end{aligned} W ( t ) = [ e − A ε t − e − A ε ′ t ] U 0 + ∫ 0 t e − A ε ( t − s ) [ F ( U ε ( s )) − F ( U ε ′ ( s )) ] d s + ∫ 0 t [ e − A ε ( t − s ) − e − A ε ′ ( t − s ) ] F ( U ε ′ ( s )) d s . (109)
Step 1: Uniform boundedness on [ 0 , T ] [0,T] [ 0 , T ] . By Lemma 3.2 , the solutions U ε ( t ) U_\varepsilon(t) U ε ( t ) and U ε ′ ( t ) U_{\varepsilon'}(t) U ε ′ ( t ) remain in a fixed ball B 1 : = { U : ∥ U ∥ H ε ≤ R 1 } B_1:=\{U:\|U\|_{\mathcal{H}_\varepsilon}\le R_1\} B 1 := { U : ∥ U ∥ H ε ≤ R 1 } for all t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , where R 1 > 0 R_1>0 R 1 > 0 depends on T T T , B B B , and the parameters, but not on ε , ε ′ \varepsilon,\varepsilon' ε , ε ′ . Consequently, there exists L 1 > 0 L_1>0 L 1 > 0 , independent of ε , ε ′ \varepsilon,\varepsilon' ε , ε ′ , such that
∥ F ( U ) − F ( V ) ∥ H ε ≤ L 1 ∥ U − V ∥ H ε , U , V ∈ B 1 , (110) \|F(U)-F(V)\|_{\mathcal{H}_\varepsilon}\le L_1\|U-V\|_{\mathcal{H}_\varepsilon},\qquad U,V\in B_1, \tag*{(110)} ∥ F ( U ) − F ( V ) ∥ H ε ≤ L 1 ∥ U − V ∥ H ε , U , V ∈ B 1 , (110)
and M 1 : = sup U ∈ B 1 ∥ F ( U ) ∥ H ε < + ∞ M_1:=\sup_{U\in B_1}\|F(U)\|_{\mathcal{H}_\varepsilon}<+\infty M 1 := sup U ∈ B 1 ∥ F ( U ) ∥ H ε < + ∞ .
Step 2: Estimating the difference of the linear semigroups. We claim that
∥ [ e − A ε t − e − A ε ′ t ] V ∥ H ε ≤ C 2 ( ε 0 , T ) ∣ ε − ε ′ ∣ 1 / 2 ∥ V ∥ H ε , t ∈ [ 0 , T ] , V ∈ H ε , (111) \left\|\left[e^{-A_\varepsilon t}-e^{-A_{\varepsilon'}t}\right]V\right\|_{\mathcal{H}_\varepsilon}\le C_2(\varepsilon_0,T)|\varepsilon-\varepsilon'|^{1/2}\|V\|_{\mathcal{H}_\varepsilon},\qquad t\in[0,T],\ V\in\mathcal{H}_\varepsilon, \tag*{(111)} [ e − A ε t − e − A ε ′ t ] V H ε ≤ C 2 ( ε 0 , T ) ∣ ε − ε ′ ∣ 1/2 ∥ V ∥ H ε , t ∈ [ 0 , T ] , V ∈ H ε , (111)
with a constant C 2 ( ε 0 , T ) > 0 C_2(\varepsilon_0,T)>0 C 2 ( ε 0 , T ) > 0 independent of ε , ε ′ \varepsilon,\varepsilon' ε , ε ′ and V V V . To prove (111) , we use the integral representation
e − A ε t − e − A ε ′ t = ∫ 0 t e − A ε ( t − s ) ( A ε ′ − A ε ) e − A ε ′ s d s . e^{-A_\varepsilon t}-e^{-A_{\varepsilon'}t}=\int_0^t e^{-A_\varepsilon(t-s)}(A_{\varepsilon'}-A_\varepsilon)e^{-A_{\varepsilon'}s}\,\mathrm{d}s. e − A ε t − e − A ε ′ t = ∫ 0 t e − A ε ( t − s ) ( A ε ′ − A ε ) e − A ε ′ s d s .
which follows from the identity ∂ s ( e − A ε ( t − s ) e − A ε ′ s ) = e − A ε ( t − s ) ( A ε ′ − A ε ) e − A ε ′ s \partial_s\left(e^{-A_\varepsilon(t-s)}e^{-A_{\varepsilon'}s}\right)=e^{-A_\varepsilon(t-s)}(A_{\varepsilon'}-A_\varepsilon)e^{-A_{\varepsilon'}s} ∂ s ( e − A ε ( t − s ) e − A ε ′ s ) = e − A ε ( t − s ) ( A ε ′ − A ε ) e − A ε ′ s . For V ∈ D ( A ε ′ ) V\in\mathcal{D}(A_{\varepsilon'}) V ∈ D ( A ε ′ ) and t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , the integrand is well defined because e − A ε ′ s V ∈ D ( A ε ′ ) e^{-A_{\varepsilon'}s}V\in\mathcal{D}(A_{\varepsilon'}) e − A ε ′ s V ∈ D ( A ε ′ ) for s ≥ 0 s\ge0 s ≥ 0 . For general V ∈ H ε V\in\mathcal{H}_\varepsilon V ∈ H ε , we first obtain the estimate for V ∈ D ( A ε ′ ) V\in\mathcal{D}(A_{\varepsilon'}) V ∈ D ( A ε ′ ) and then extend it to all V ∈ H ε V\in\mathcal{H}_\varepsilon V ∈ H ε by density. Thus, for V ∈ D ( A ε ′ ) V\in\mathcal{D}(A_{\varepsilon'}) V ∈ D ( A ε ′ ) , every φ 2 ∈ H d 2 ( 0 , ℓ ) \varphi_2\in H^2_d(0,\ell) φ 2 ∈ H d 2 ( 0 , ℓ ) , and every φ 4 ∈ H 0 1 ( 0 , ℓ ) \varphi_4\in H^1_0(0,\ell) φ 4 ∈ H 0 1 ( 0 , ℓ ) , testing ( A ε ′ − A ε ) U (A_{\varepsilon'}-A_\varepsilon)U ( A ε ′ − A ε ) U against the element ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) (0,\varphi_2,0,0,\varphi_4,0) ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) yields
∣ ⟨ ( A ε ′ − A ε ) U , ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) ⟩ H ε ∣ ≤ ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ x x ( s ) ∥ ∥ φ 2 , x x ∥ d s + ∫ 0 + ∞ ∣ g 2 , ε ′ − g 2 , ε ∣ ( s ) ∥ η x ( s ) ∥ ∥ φ 4 , x ∥ d s . (112) \left|\left\langle(A_{\varepsilon'}-A_\varepsilon)U,(0,\varphi_2,0,0,\varphi_4,0)\right\rangle_{\mathcal{H}_\varepsilon}\right|
\le\int_0^{+\infty}|g_{1,\varepsilon'}-g_{1,\varepsilon}| (s)\|\sigma_{xx}(s)\|\|\varphi_{2,xx}\|\,\mathrm{d}s+\int_0^{+\infty}|g_{2,\varepsilon'}-g_{2,\varepsilon}| (s)\|\eta_x(s)\|\|\varphi_{4,x}\|\,\mathrm{d}s. \tag*{(112)} ⟨ ( A ε ′ − A ε ) U , ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) ⟩ H ε ≤ ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ xx ( s ) ∥∥ φ 2 , xx ∥ d s + ∫ 0 + ∞ ∣ g 2 , ε ′ − g 2 , ε ∣ ( s ) ∥ η x ( s ) ∥∥ φ 4 , x ∥ d s . (112)
where σ \sigma σ and η \eta η are the memory components of U U U . Using the Cauchy–Schwarz inequality in s s s and the elementary bound ∣ g 1 , ε ′ − g 1 , ε ∣ ≤ g 1 , ε ′ + g 1 , ε |g_{1,\varepsilon'}-g_{1,\varepsilon}|\le g_{1,\varepsilon'}+g_{1,\varepsilon} ∣ g 1 , ε ′ − g 1 , ε ∣ ≤ g 1 , ε ′ + g 1 , ε ,
∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ x x ( s ) ∥ ∥ φ 2 , x x ∥ d s ≤ ( ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ x x ( s ) ∥ 2 d s ) 1 / 2 ( ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) d s ) 1 / 2 ∥ φ 2 , x x ∥ ≤ 2 ∥ σ ∥ M g 1 , ε ∥ g 1 , ε ′ − g 1 , ε ∥ L 1 ( R + ) 1 / 2 ∥ φ 2 , x x ∥ , \begin{aligned}
\int_0^{+\infty}|g_{1,\varepsilon'}-g_{1,\varepsilon}| (s)\|\sigma_{xx}(s)\|\|\varphi_{2,xx}\|\,\mathrm{d}s
&\le\left(\int_0^{+\infty}|g_{1,\varepsilon'}-g_{1,\varepsilon}| (s)\|\sigma_{xx}(s)\|^2\,\mathrm{d}s\right)^{1/2}\left(\int_0^{+\infty}|g_{1,\varepsilon'}-g_{1,\varepsilon}| (s)\,\mathrm{d}s\right)^{1/2}\|\varphi_{2,xx}\|\\
&\le\sqrt{2}\|\sigma\|_{\mathcal{M}_{g_{1,\varepsilon}}}\|g_{1,\varepsilon'}-g_{1,\varepsilon}\|_{L^1(\mathbb{R}_+)}^{1/2}\|\varphi_{2,xx}\|,
\end{aligned} ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ xx ( s ) ∥∥ φ 2 , xx ∥ d s ≤ ( ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) ∥ σ xx ( s ) ∥ 2 d s ) 1/2 ( ∫ 0 + ∞ ∣ g 1 , ε ′ − g 1 , ε ∣ ( s ) d s ) 1/2 ∥ φ 2 , xx ∥ ≤ 2 ∥ σ ∥ M g 1 , ε ∥ g 1 , ε ′ − g 1 , ε ∥ L 1 ( R + ) 1/2 ∥ φ 2 , xx ∥ ,
where we used ∥ σ ∥ M g 1 , ε ≤ C ( ε 0 ) ∥ σ ∥ M g 1 , ε ′ \|\sigma\|_{\mathcal{M}_{g_{1,\varepsilon}}}\le C(\varepsilon_0)\|\sigma\|_{\mathcal{M}_{g_{1,\varepsilon'}}} ∥ σ ∥ M g 1 , ε ≤ C ( ε 0 ) ∥ σ ∥ M g 1 , ε ′ , with a constant C ( ε 0 ) C(\varepsilon_0) C ( ε 0 ) independent of ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon'\in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] . This equivalence follows from Assumption 5.3 : for ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon'\in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] ,
g 1 , ε ( s ) g 1 , ε ′ ( s ) = ε ′ ε ⋅ g 1 ( s / ε ) g 1 ( s / ε ′ ) ≤ ε ′ ε ⋅ g 1 ( 0 ) e − λ 1 s / ε c 0 e − λ 0 s / ε ′ . \frac{g_{1,\varepsilon}(s)}{g_{1,\varepsilon'}(s)}=\frac{\varepsilon'}{\varepsilon}\cdot\frac{g_1(s/\varepsilon)}{g_1(s/\varepsilon')}\le\frac{\varepsilon'}{\varepsilon}\cdot\frac{g_1(0)e^{-\lambda_1s/\varepsilon}}{c_0e^{-\lambda_0s/\varepsilon'}}. g 1 , ε ′ ( s ) g 1 , ε ( s ) = ε ε ′ ⋅ g 1 ( s / ε ′ ) g 1 ( s / ε ) ≤ ε ε ′ ⋅ c 0 e − λ 0 s / ε ′ g 1 ( 0 ) e − λ 1 s / ε .
and the right-hand side is bounded uniformly in s ≥ 0 s\ge0 s ≥ 0 and in ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon'\in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] by a constant depending only on ε 0 \varepsilon_0 ε 0 , c 0 c_0 c 0 , λ 0 \lambda_0 λ 0 , λ 1 \lambda_1 λ 1 , and g 1 ( 0 ) g_1(0) g 1 ( 0 ) . The same estimate applies to the second integral. Combining with (107) and the uniform bound ∥ σ ∥ M g 1 , ε ′ , ∥ η ∥ M g 2 , ε ′ ≤ ∥ U ∥ H ε \|\sigma\|_{\mathcal{M}_{g_{1,\varepsilon'}}},\|\eta\|_{\mathcal{M}_{g_{2,\varepsilon'}}}\le\|U\|_{\mathcal{H}_\varepsilon} ∥ σ ∥ M g 1 , ε ′ , ∥ η ∥ M g 2 , ε ′ ≤ ∥ U ∥ H ε , we obtain
∥ ( A ε ′ − A ε ) U ∥ D ( A ε ′ ) ∗ ≤ C ( ε 0 ) ∣ ε − ε ′ ∣ 1 / 2 ∥ U ∥ H ε , U ∈ D ( A ε ′ ) , (113) \|(A_{\varepsilon'}-A_\varepsilon)U\|_{\mathcal{D}(A_{\varepsilon'})^*}\le C(\varepsilon_0)|\varepsilon-\varepsilon'|^{1/2}\|U\|_{\mathcal{H}_\varepsilon},\qquad U\in\mathcal{D}(A_{\varepsilon'}), \tag*{(113)} ∥ ( A ε ′ − A ε ) U ∥ D ( A ε ′ ) ∗ ≤ C ( ε 0 ) ∣ ε − ε ′ ∣ 1/2 ∥ U ∥ H ε , U ∈ D ( A ε ′ ) , (113)
with C ( ε 0 ) > 0 C(\varepsilon_0)>0 C ( ε 0 ) > 0 independent of ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon'\in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] ; here D ( A ε ′ ) ∗ \mathcal{D}(A_{\varepsilon'})^* D ( A ε ′ ) ∗ denotes the dual of D ( A ε ′ ) \mathcal{D}(A_{\varepsilon'}) D ( A ε ′ ) with respect to the H ε \mathcal{H}_\varepsilon H ε -inner product. The constant C ( ε 0 ) C(\varepsilon_0) C ( ε 0 ) depends only on ε 0 \varepsilon_0 ε 0 , c 0 c_0 c 0 , λ 0 \lambda_0 λ 0 , λ 1 \lambda_1 λ 1 , g 1 ( 0 ) g_1(0) g 1 ( 0 ) , and the analogous quantities for g 2 g_2 g 2 , and blows up as ε 0 → 0 + \varepsilon_0\to0^+ ε 0 → 0 + . Since ( A ε ′ − A ε ) U (A_{\varepsilon'}-A_\varepsilon)U ( A ε ′ − A ε ) U has nonzero components only in the elastic equations and depends on U U U only through its memory components, it suffices to test against elements of the form ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) (0,\varphi_2,0,0,\varphi_4,0) ( 0 , φ 2 , 0 , 0 , φ 4 , 0 ) with φ 2 ∈ H d 2 ( 0 , ℓ ) \varphi_2\in H^2_d(0,\ell) φ 2 ∈ H d 2 ( 0 , ℓ ) and φ 4 ∈ H 0 1 ( 0 , ℓ ) \varphi_4\in H^1_0(0,\ell) φ 4 ∈ H 0 1 ( 0 , ℓ ) ; these are dense in the elastic subspace, and the memory components of the test function do not contribute. Combining (113) with the uniform boundedness of e − A ε ( t − s ) e^{-A_\varepsilon(t-s)} e − A ε ( t − s ) and e − A ε ′ s e^{-A_{\varepsilon'}s} e − A ε ′ s , we obtain (111) .
Step 3: Estimating the third term in (109). By (111) , the third term in (109) is bounded by
∥ ∫ 0 t [ e − A ε ( t − s ) − e − A ε ′ ( t − s ) ] F ( U ε ′ ( s ) ) d s ∥ H ε ≤ C 2 ( ε 0 , T ) M 1 T ∣ ε − ε ′ ∣ 1 / 2 . \left\|\int_0^t\left[e^{-A_\varepsilon(t-s)}-e^{-A_{\varepsilon'}(t-s)}\right]F(U_{\varepsilon'}(s))\,\mathrm{d}s\right\|_{\mathcal{H}_\varepsilon}\le C_2(\varepsilon_0,T)M_1T|\varepsilon-\varepsilon'|^{1/2}. ∫ 0 t [ e − A ε ( t − s ) − e − A ε ′ ( t − s ) ] F ( U ε ′ ( s )) d s H ε ≤ C 2 ( ε 0 , T ) M 1 T ∣ ε − ε ′ ∣ 1/2 .
Step 4: Gronwall argument. Taking the norm of (109) , estimating the first term by (111) with V = U 0 V=U_0 V = U 0 , and using the estimates from Steps 2 and 3 together with the Lipschitz bound (110) , we obtain
∥ W ( t ) ∥ H ε ≤ C 2 ( ε 0 , T ) ( 1 + M 1 T ) ∣ ε − ε ′ ∣ 1 / 2 + α L 1 ∫ 0 t e − β ( t − s ) ∥ W ( s ) ∥ H ε d s . \|W(t)\|_{\mathcal{H}_\varepsilon}\le C_2(\varepsilon_0,T)(1+M_1T)|\varepsilon-\varepsilon'|^{1/2}+\alpha L_1\int_0^t e^{-\beta(t-s)}\|W(s)\|_{\mathcal{H}_\varepsilon}\,\mathrm{d}s. ∥ W ( t ) ∥ H ε ≤ C 2 ( ε 0 , T ) ( 1 + M 1 T ) ∣ ε − ε ′ ∣ 1/2 + α L 1 ∫ 0 t e − β ( t − s ) ∥ W ( s ) ∥ H ε d s .
Applying Gronwall’s inequality,
∥ W ( t ) ∥ H ε ≤ C 2 ( ε 0 , T ) ( 1 + M 1 T ) ∣ ε − ε ′ ∣ 1 / 2 e α L 1 T , t ∈ [ 0 , T ] . \|W(t)\|_{\mathcal{H}_\varepsilon}\le C_2(\varepsilon_0,T)(1+M_1T)|\varepsilon-\varepsilon'|^{1/2}e^{\alpha L_1T},\qquad t\in[0,T]. ∥ W ( t ) ∥ H ε ≤ C 2 ( ε 0 , T ) ( 1 + M 1 T ) ∣ ε − ε ′ ∣ 1/2 e α L 1 T , t ∈ [ 0 , T ] .
Taking the supremum over t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] and U 0 ∈ B U_0 \in B U 0 ∈ B , and setting
L T ( B , ε 0 ) : = C 2 ( ε 0 , T ) ( 1 + M 1 T ) e α L 1 T , L_T(B,\varepsilon_0) := C_2(\varepsilon_0,T)(1+M_1T)e^{\alpha L_1T}, L T ( B , ε 0 ) := C 2 ( ε 0 , T ) ( 1 + M 1 T ) e α L 1 T ,
we obtain (106) . Note that C 2 ( ε 0 , T ) C_2(\varepsilon_0,T) C 2 ( ε 0 , T ) , M 1 M_1 M 1 , L 1 L_1 L 1 depend only on T T T , B B B , ε 0 \varepsilon_0 ε 0 , and the parameters, but not on ε , ε ′ ∈ [ ε 0 , 1 ] \varepsilon,\varepsilon' \in[\varepsilon_0,1] ε , ε ′ ∈ [ ε 0 , 1 ] . This completes the proof.
Remark 5.5. The exponent 1 / 2 1/2 1/2 in (106) is a natural consequence of the L 1 L^1 L 1 -estimate (107) together with the Cauchy–Schwarz inequality in the memory variable s s s : only the square root ∥ g i , ε − g i , ε ′ ∥ L 1 1 / 2 \lVert g_{i,\varepsilon}-g_{i,\varepsilon'}\rVert_{L^1}^{1/2} ∥ g i , ε − g i , ε ′ ∥ L 1 1/2 can be extracted from the weighted inner product, while the full L 1 L^1 L 1 -norm remains inside the M g i , ε M_{g_{i,\varepsilon}} M g i , ε -norm of the memory variable. To obtain Lipschitz continuity in ε \varepsilon ε , it would suffice, together with the norm equivalence of Assumption 5.3, to have the stronger estimate
∫ 0 + ∞ ∣ g i , ε ( s ) − g i , ε ′ ( s ) ∣ 2 g i , ε ( s ) d s ≤ C ( ε 0 ) ∣ ε − ε ′ ∣ 2 , ε , ε ′ ∈ [ ε 0 , 1 ] , \int_0^{+\infty}\frac{\lvert g_{i,\varepsilon}(s)-g_{i,\varepsilon'}(s)\rvert^2}{g_{i,\varepsilon}(s)}\,\mathrm{d}s \le C(\varepsilon_0)\lvert\varepsilon-\varepsilon'\rvert^2,\qquad\varepsilon,\varepsilon' \in[\varepsilon_0,1], ∫ 0 + ∞ g i , ε ( s ) ∣ g i , ε ( s ) − g i , ε ′ ( s ) ∣ 2 d s ≤ C ( ε 0 ) ∣ ε − ε ′ ∣ 2 , ε , ε ′ ∈ [ ε 0 , 1 ] ,
with C ( ε 0 ) > 0 C(\varepsilon_0)>0 C ( ε 0 ) > 0 ; for the exponential kernel g i ( s ) = b i e − a i s g_i(s)=b_i e^{-a_i s} g i ( s ) = b i e − a i s this holds on every compact subset of ( 0 , 1 ] (0,1] ( 0 , 1 ] with C ( ε 0 ) = O ( ε 0 − 3 ) C(\varepsilon_0)=O(\varepsilon_0^{-3}) C ( ε 0 ) = O ( ε 0 − 3 ) . Since lower semicontinuity only requires continuity (not Lipschitz continuity) of the map ε ↦ S ε ( t ) U 0 \varepsilon\mapsto S_\varepsilon(t)U_0 ε ↦ S ε ( t ) U 0 , the Hölder estimate (106) is sufficient for our purposes.
Conclusion of lower semicontinuity
With Propositions 5.2 and 5.4 at hand, the lower semicontinuity follows from a standard argument.
Proposition 5.6 (Lower semicontinuity under a structural condition). Assume the hypotheses of Theorem 4.3 and Assumption 5.3. Suppose in addition that the limit attractor A 0 A_0 A 0 satisfies the following structural stability condition:
(S) A 0 A_0 A 0 is the union of finitely many hyperbolic equilibria of the limit semigroup S 0 ( t ) S_0(t) S 0 ( t ) and their unstable manifolds.
Then
lim ε → 0 + dist H ε ( A 0 , A ε ) = 0. (114) \lim_{\varepsilon\to0^+}\operatorname{dist}_{\mathcal H_\varepsilon}(A_0,A_\varepsilon)=0. \tag*{(114)} ε → 0 + lim dist H ε ( A 0 , A ε ) = 0. (114)
Proof. The proof is a direct application of the abstract lower semicontinuity theorem [5 Chapter 7] (see also [12 ] ). We verify the four hypotheses of that theorem.
( L 1 ) (L1) ( L 1 ) Existence of attractors. By Theorem 4.3 and its proof, for every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] the semigroup S ε ( t ) S_\varepsilon(t) S ε ( t ) possesses a compact global attractor A ε ⊂ B 0 \mathcal A_\varepsilon\subset\mathcal B_0 A ε ⊂ B 0 , and the limit semigroup S 0 ( t ) S_0(t) S 0 ( t ) possesses a compact global attractor A 0 ⊂ B 0 \mathcal A_0 \subset\mathcal B_0 A 0 ⊂ B 0 . All the attractors are contained in the common absorbing ball B 0 \mathcal B_0 B 0 of Proposition 3.1.
( L 2 ) (L2) ( L 2 ) Finite-time convergence. By Proposition 4.2, for every T > 0 T>0 T > 0 and every sequence U 0 , ε → U 0 U_{0,\varepsilon}\to U_0 U 0 , ε → U 0 ,
sup t ∈ [ 0 , T ] ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0 , ε → 0 + . \sup_{t\in[0,T]}\lVert S_\varepsilon(t)U_{0,\varepsilon}-S_0(t)U_0\rVert_{\mathcal H_\varepsilon}\to0,\qquad\varepsilon\to0^+. t ∈ [ 0 , T ] sup ∥ S ε ( t ) U 0 , ε − S 0 ( t ) U 0 ∥ H ε → 0 , ε → 0 + .
In particular, taking U 0 , ε = U 0 U_{0,\varepsilon}=U_0 U 0 , ε = U 0 , we get S ε ( t ) U 0 → S 0 ( t ) U 0 S_\varepsilon(t)U_0\to S_0(t)U_0 S ε ( t ) U 0 → S 0 ( t ) U 0 as ε → 0 + \varepsilon\to0^+ ε → 0 + , uniformly in t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
( L 3 ) (L3) ( L 3 ) Uniform exponential attraction. By Proposition 5.2, there exist constants M ≥ 1 M\ge1 M ≥ 1 and ω > 0 \omega>0 ω > 0 , independent of ε \varepsilon ε , such that
dist H ε ( S ε ( t ) B 0 , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] . \operatorname{dist}_{\mathcal H_\varepsilon}(S_\varepsilon(t)\mathcal B_0,\mathcal A_\varepsilon)\le Me^{-\omega t},\qquad\forall t\ge0,\ \forall\varepsilon\in(0,1]. dist H ε ( S ε ( t ) B 0 , A ε ) ≤ M e − ω t , ∀ t ≥ 0 , ∀ ε ∈ ( 0 , 1 ] .
In particular, the attractors A ε \mathcal A_\varepsilon A ε are uniformly exponentially attracting on $\mathcal B_0.атәи
(L4) Uniform continuity in ε \varepsilon ε (optional). By Proposition 5.4 , for every ε 0 ∈ ( 0 , 1 ] \varepsilon_0 \in(0,1] ε 0 ∈ ( 0 , 1 ] and every bounded set B B B there exists L T ( B , ε 0 ) > 0 L_T(B,\varepsilon_0)>0 L T ( B , ε 0 ) > 0 such that
∥ S ε ( t ) U 0 − S ε ′ ( t ) U 0 ∥ H ε ≤ L T ( B , ε 0 ) ∣ ε − ε ′ ∣ 1 / 2 , t ∈ [ 0 , T ] , U 0 ∈ B , ε , ε ′ ∈ [ ε 0 , 1 ] . \|S_{\varepsilon}(t)U_0-S_{\varepsilon'}(t)U_0\|_{\mathcal{H}_{\varepsilon}}\leq L_T(B,\varepsilon_0)|\varepsilon-\varepsilon'|^{1/2},\qquad t\in[0,T],\ U_0\in B,\ \varepsilon,\varepsilon'\in[\varepsilon_0,1]. ∥ S ε ( t ) U 0 − S ε ′ ( t ) U 0 ∥ H ε ≤ L T ( B , ε 0 ) ∣ ε − ε ′ ∣ 1/2 , t ∈ [ 0 , T ] , U 0 ∈ B , ε , ε ′ ∈ [ ε 0 , 1 ] .
This estimate is not needed for the lower semicontinuity conclusion below: hypotheses (L1)–(L3) and (L5) already suffice in the Chueshov–Lasiecka framework. We record it here only for completeness and for possible quantitative refinements. (If one wishes to invoke it, the dependence of L T ( B , ε 0 ) L_T(B,\varepsilon_0) L T ( B , ε 0 ) on ε 0 \varepsilon_0 ε 0 is harmless: for any sequence ε n → 0 + \varepsilon_n\to0^+ ε n → 0 + , one first chooses ε 0 > 0 \varepsilon_0>0 ε 0 > 0 , applies the estimate on [ ε 0 , 1 ] [\varepsilon_0,1] [ ε 0 , 1 ] , and uses a diagonal argument for the finitely many remaining indices.)
(L5) Structural stability of A 0 \mathcal{A}_0 A 0 . By hypothesis (S), A 0 \mathcal{A}_0 A 0 is the union of finitely many hyperbolic equilibria of S 0 ( t ) S_0(t) S 0 ( t ) and their unstable manifolds. This is the structural stability condition required in [5 Chapter 7] and [12 ] .
By [5 Chapter 7] , hypotheses (L1)–(L5) imply the lower semicontinuity
lim ε → 0 + dist H ε ( A 0 , A ε ) = 0 , (115) \lim_{\varepsilon\to0^+}\operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_0,\mathcal{A}_{\varepsilon})=0, \tag*{(115)} ε → 0 + lim dist H ε ( A 0 , A ε ) = 0 , (115)
which is (114) .
Remark 5.7. Condition (S) is a standard sufficient condition for lower semicontinuity in the Hale–Chueshov–Lasiecka framework; see [5 Chapter 7] and [12 ] . We do not verify (S) for the specific nonlinearities h i h_i h i and k [ ⋅ ] + k[\cdot]^+ k [ ⋅ ] + here: it is an additional hypothesis under which Proposition 5.6 holds. For gradient systems, weaker structural conditions may suffice; see, e.g., [3 ] for related continuity results. A non-vacuous instance of (S) is provided in Example 5.8.
Example 5.8 (A non-vacuous instance of condition (S)). Let ( 0 , ℓ ) = ( 0 , π ) (0,\ell)=(0,\pi) ( 0 , ℓ ) = ( 0 , π ) and take
μ 1 = μ 2 = 5 , γ 1 = γ 2 = 1 , k = 1 , f 1 = f 2 = 0 , h 1 ( u ) = u , h 2 ( v ) = v , \mu_1=\mu_2=5,\qquad\gamma_1=\gamma_2=1,\qquad k=1,\qquad f_1=f_2=0,\qquad h_1(u)=u,\qquad h_2(v)=v, μ 1 = μ 2 = 5 , γ 1 = γ 2 = 1 , k = 1 , f 1 = f 2 = 0 , h 1 ( u ) = u , h 2 ( v ) = v ,
together with the exponential kernels g 1 ( s ) = g 2 ( s ) = 2 e − s g_1(s)=g_2(s)=2e^{-s} g 1 ( s ) = g 2 ( s ) = 2 e − s . Then (9) holds with β i = 2 ∈ ( 0 , 5 ) = μ i \beta_i=2\in(0,5)=\mu_i β i = 2 ∈ ( 0 , 5 ) = μ i and λ i = 1 \lambda_i=1 λ i = 1 ; the rescaled kernels g i , ε ( s ) = ε − 1 g i ( s / ε ) = 2 ε − 1 e − s / ε g_{i,\varepsilon}(s)=\varepsilon^{-1}g_i(s/\varepsilon)=2\varepsilon^{-1}e^{-s/\varepsilon} g i , ε ( s ) = ε − 1 g i ( s / ε ) = 2 ε − 1 e − s / ε preserve the mass β i = 2 \beta_i=2 β i = 2 , so that μ ˉ i , ε = μ i − β i = 3 > 0 \bar{\mu}_{i,\varepsilon}=\mu_i-\beta_i=3>0 μ ˉ i , ε = μ i − β i = 3 > 0 for all ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . Assumption 5.3 holds with c 0 = 1 c_0=1 c 0 = 1 , λ 0 = 1 \lambda_0=1 λ 0 = 1 , and (107) follows from the explicit computation for exponential kernels. The nonlinearities satisfy (11) with H i ( s ) = 1 2 s 2 H_i(s)=\frac{1}{2}s^2 H i ( s ) = 2 1 s 2 , C 1 = 2 C_1=2 C 1 = 2 , C 2 = 1 C_2=1 C 2 = 1 , since
lim inf ∣ s ∣ → ∞ H i ( s ) s 2 = 1 2 ≥ 0 , s h i ( s ) − 2 H i ( s ) = 0 , h i ′ ( s ) = 1 ≥ − 1. \liminf_{|s|\to\infty}\frac{H_i(s)}{s^2}=\frac{1}{2}\geq0,\qquad sh_i(s)-2H_i(s)=0,\qquad h_i'(s)=1\geq-1. ∣ s ∣ → ∞ lim inf s 2 H i ( s ) = 2 1 ≥ 0 , s h i ( s ) − 2 H i ( s ) = 0 , h i ′ ( s ) = 1 ≥ − 1.
The limit system is
{ u t t + u t + 5 u x x x x + [ u − v ] + + u = 0 , v t t + v t − 5 v x x − [ u − v ] + + v = 0 , \begin{cases}
u_{tt}+u_t+5u_{xxxx}+[u-v]^++u=0,\\
v_{tt}+v_t-5v_{xx}-[u-v]^++v=0,
\end{cases} { u tt + u t + 5 u xxxx + [ u − v ] + + u = 0 , v tt + v t − 5 v xx − [ u − v ] + + v = 0 ,
with u = u x x = v = 0 u=u_{xx}=v=0 u = u xx = v = 0 at x = 0 , π x=0,\pi x = 0 , π . Let ( u , v ) (u,v) ( u , v ) be an equilibrium. Multiplying the first equation by u u u and the second by v v v , integrating over ( 0 , π ) (0,\pi) ( 0 , π ) , and adding the two identities gives
5 ∥ u x x ∥ 2 + 5 ∥ v x ∥ 2 + ∥ u ∥ 2 + ∥ v ∥ 2 + ∫ 0 π [ u − v ] + ( u − v ) d x = 0. 5\|u_{xx}\|^2+5\|v_x\|^2+\|u\|^2+\|v\|^2+\int_0^\pi[u-v]^+(u-v)\,dx=0. 5∥ u xx ∥ 2 + 5∥ v x ∥ 2 + ∥ u ∥ 2 + ∥ v ∥ 2 + ∫ 0 π [ u − v ] + ( u − v ) d x = 0.
Since every term is nonnegative and the last one equals ∫ 0 π ( [ u − v ] + ) 2 d x \int_0^\pi([u-v]^+)^2dx ∫ 0 π ([ u − v ] + ) 2 d x , we deduce u = v = 0 u=v=0 u = v = 0 . Hence ( 0 , 0 ) (0,0) ( 0 , 0 ) is the only equilibrium. At ( 0 , 0 ) (0,0) ( 0 , 0 ) the nonlinearity [ ⋅ ] + [\cdot]^+ [ ⋅ ] + is inactive on the non-activation side u − v ≤ 0 u-v\leq0 u − v ≤ 0 , and the linearisation there decouples into
5 u x x x x + u = 0 , − 5 v x x + v = 0 , 5u_{xxxx}+u=0,\qquad-5v_{xx}+v=0, 5 u xxxx + u = 0 , − 5 v xx + v = 0 ,
whose eigenvalues are 5 n 4 + 1 > 0 5n^{4}+1>0 5 n 4 + 1 > 0 and 5 n 2 + 1 > 0 5n^{2}+1>0 5 n 2 + 1 > 0 , n ≥ 1 n\geq1 n ≥ 1 . Hence ( 0 , 0 ) (0,0) ( 0 , 0 ) is a hyperbolic equilibrium with trivial unstable manifold, and A 0 = { ( 0 , 0 ) } \mathcal{A}_0=\{(0,0)\} A 0 = {( 0 , 0 )} is the union of finitely many hyperbolic equilibria and their unstable manifolds. Thus condition (S) of Proposition 5.6 is satisfied, and the lower semicontinuity lim ε → 0 + dist H ε ( A 0 , A ε ) = 0 \lim_{\varepsilon\to0^{+}}\operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_0,\mathcal{A}_{\varepsilon})=0 lim ε → 0 + dist H ε ( A 0 , A ε ) = 0 holds for this example. This shows that hypothesis (S) is non-vacuous.
Summary and Open Problems
We have developed a continuity theory for the global attractor of a dissipative evolution equation with memory under a mass-preserving singular limit of the memory kernel, and we have verified its ingredients for a coupled suspension bridge system with past history. The key elements are:
The mass-preserving rescaling (15), which keeps the effective elastic coefficients μ ˉ i \bar{\mu}_i μ ˉ i fixed and is the key structural feature behind the uniform dissipativity estimates. This is the feature that distinguishes the present limit from arbitrary concentration families and that makes the continuity analysis tractable.
The uniform exponential stability (26) of the linear semigroup, established in [33 ] for fixed ε \varepsilon ε and extended to a uniform-in-ε \varepsilon ε version in Theorem 2.2.
The uniform quasi-stability (Proposition 4.1), obtained from the uniform exponential stability and the sequential compactness of the integral remainder K ε , T ( B 0 ) K_{\varepsilon,T}(\mathcal{B}_0) K ε , T ( B 0 ) , and the uniform exponential attraction (Proposition 5.2), which follows from the uniform quasi-stability together with the common absorbing ball.
The finite-time convergence (63), which together with (1)–(3) yields the upper semicontinuity; the uniform Hölder continuity (106) in ε \varepsilon ε is established as well, although the lower semicontinuity proof does not rely on it.
Several open problems remain:
(a) Rate of convergence. Can one obtain a quantitative rate for the upper and lower semicontinuity? For instance, is it true that dist H ε ( A ε , A 0 ) ≤ C ε θ \operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_{\varepsilon},\mathcal{A}_0)\leq C\varepsilon^{\theta} dist H ε ( A ε , A 0 ) ≤ C ε θ and dist H ε ( A 0 , A ε ) ≤ C ′ ε θ ′ \operatorname{dist}_{\mathcal{H}_{\varepsilon}}(\mathcal{A}_0,\mathcal{A}_{\varepsilon})\leq C'\varepsilon^{\theta'} dist H ε ( A 0 , A ε ) ≤ C ′ ε θ ′ for some θ , θ ′ > 0 \theta,\theta'>0 θ , θ ′ > 0 ?
(b) Non-mass-preserving perturbations. What happens if the rescaling changes the mass, i.e., ∫ 0 + ∞ g i , ε ( s ) d s = β i , ε → β i , 0 ≠ β i \int_{0}^{+\infty}g_{i,\varepsilon}(s)\,\mathrm{d}s=\beta_{i,\varepsilon}\to\beta_{i,0}\ne\beta_i ∫ 0 + ∞ g i , ε ( s ) d s = β i , ε → β i , 0 = β i ? In this case, the effective elastic coefficients μ ˉ i , ε = μ i − β i , ε \bar{\mu}_{i,\varepsilon}=\mu_i-\beta_{i,\varepsilon} μ ˉ i , ε = μ i − β i , ε may degenerate (or even change sign), and the analysis becomes significantly more difficult. The mass-preserving hypothesis is thus not merely convenient but structurally necessary for the present method.
(c) Other singular limits. Can one treat the limit λ i → + ∞ \lambda_i\to+\infty λ i → + ∞ (infinitely fast memory decay) or λ i → 0 + \lambda_i\to0^{+} λ i → 0 + (infinitely slow memory decay) within the same framework? Note that λ i → + ∞ \lambda_i\to+\infty λ i → + ∞ is a singular limit of a related nature to ε → 0 + \varepsilon\to0^{+} ε → 0 + , in which the kernel concentrates near the origin.
(d) Lower semicontinuity without structural assumptions. Is the lower semicontinuity true for the full range of nonlinearities h i h_i h i satisfying (11), without additional structural assumptions on A 0 \mathcal{A}_0 A 0 ? Example 5.8 shows that condition (S) is non-vacuous; whether it can be removed in general remains open.
Declarations
Conflict of interest. The author has no relevant financial or non-financial interests to disclose.