Introduction
Computational algorithms for simulating the propagation of time-dependent acoustic waves in heterogeneous and unbounded media are fundamental to numerous applications [8 ] . Three main strategies can be found in the literature: finite elements in a truncated region, with the exterior replaced by approximate non-reflecting boundary conditions (NRBCs) [20 , 25 , 29 ] or perfectly matched layers (PMLs) [7 , 9 ] ; boundary elements, which represent the unbounded exterior exactly but are restricted to piecewise homogeneous media, since a variable wave speed cannot be accommodated by boundary potentials alone [10 , 23 ] ; and couplings of finite and boundary elements, which combine both, typically with convolution quadrature (CQ) applied only to the boundary integral operators and a time-stepping scheme in the interior, as discussed below.
In this article we consider a time-domain wave propagation model problem in R m \mathbb{R}^m R m , m = 2 , 3 m=2,3 m = 2 , 3 , involving a transient scalar wave field that propagates through a heterogeneous domain Ω \Omega Ω , where the spatially varying wave speed may be discontinuous, and into the unbounded free-space exterior Ω c = R m ∖ Ω \Omega^c=\mathbb{R}^m\setminus\Omega Ω c = R m ∖ Ω . The inhomogeneity is excited by an incident wave, and the resulting time-dependent wave field can be described, through the Laplace transform, by a family of frequency-domain scattering problems with complex wave numbers. CQ replaces this family by a finite set of such problems and reconstructs the full time history from their solutions, providing a robust and flexible framework for numerical simulation. Each transformed problem is solved by an overlapping coupling of the finite element method (FEM) and the boundary element method (BEM), whose construction and analysis are the subject of this paper.
Let us mention [32 ] , where the model problem solved at each complex frequency is exactly the one considered here; the difference lies in its discretization. There, the transmission problem is recast as a volume Lippmann–Schwinger equation (see [12 Ch. 8] , [39 Ch. 10] ) and solved by trigonometric collocation, so that the unknown is supported throughout the inhomogeneity and the resulting operator is dense. Here, we retain the differential formulation and discretize it by finite elements, using boundary integral operators only to account for the unbounded exterior. Consequently, the analysis in [32 ] requires the sound speed to be globally smooth, whereas our formulation accommodates discontinuities in the sound speed across material interfaces.
Problems of this kind have also been treated numerically using non-overlapping FEM–BEM coupling [3 , 4 , 19 , 28 , 38 ] . In all of these approaches, the boundary integral equation, discretized in time by CQ, is posed on the boundary of the finite element domain, which serves as both the truncation boundary and the coupling interface. The interior problem is advanced in time either by an explicit leapfrog scheme [4 , 19 , 3 ] or by the same A-stable scheme used to generate the CQ. Among the implicit schemes, we mention the trapezoidal rule in [28 ] and a Runge–Kutta method in [38 ] , which is devoted to the time-dependent linear Schrödinger equation.
The method we propose and analyze here differs from these approaches in its coupling strategy and, in particular, in its geometric flexibility. The finite element domain is chosen as a simple polygonal/polyhedral domain containing the heterogeneous domain, while the boundary element problem is posed on an auxiliary curve/surface lying inside the finite element domain and surrounding the heterogeneity, in the region where the medium is homogeneous, so that the two discretizations overlap. Consequently, neither the finite element domain nor the boundary element curve/surface needs to conform to, or depend on, the geometry of the heterogeneity. This separation of the discretization geometries provides considerable flexibility in handling complex or irregular heterogeneous domains. Moreover, the auxiliary boundary element curve/surface can be chosen to be smooth, or even analytic, independently of the geometry of the heterogeneity, which is particularly advantageous for high-order BEM.
In contrast, in a non-overlapping coupling, the boundary integral operators are posed on the polygonal/polyhedral boundary of the finite element domain and therefore inherit its corners. The finite element mesh may of course be chosen to resolve the material interfaces, as we do in our numerical experiments, to preserve the accuracy of the finite elements across jumps in the speed of sound; however, this is a property of the mesh and is not a requirement of the coupling strategy; see § 8 for the alternative. This work can be viewed as a time-domain (TD) counterpart of the frequency-domain framework in [14 , 16 ] .
The advantages of our overlapping TD-FEM–BEM approach are several. First, the all-steps-at-once implementation is naturally parallel. This is an algorithmic choice rather than a choice of discretization: the same discrete solution can also be computed by marching in time, in which case one recovers a scheme of the type used in [28 ] ; see Remark 3.1. In the all-steps-at-once formulation, however, the decoupled frequency-domain problems are independent and uniformly damped, and can therefore be solved fully in parallel.
Second, the exterior is accounted for exactly. Truncating the computational domain by means of NRBCs [29 ] introduces an additional approximation whose accuracy must be controlled independently, whereas the boundary element component provides an exact transparent condition on Σ \Sigma Σ , the boundary of the FEM domain.
The price to be paid is a non-local operator. However, this price is seldom significant: unlike in classical acoustic scattering, the complex frequencies generated by the Laplace transform introduce strong damping, so that the transmitted field is negligible on Σ \Sigma Σ for all but a small fraction of the frequencies. The coupling is therefore activated only within a fixed frequency window, whose size we quantify in this article. Even within this window, the coupled FEM–BEM system is solved by GMRES, which converges in very few iterations, typically one or two.
Finally, the interior problems associated with the transformed formulation are well suited to iterative solvers. They are coercive for every Laplace parameter s s s , with positive real part, with a coercivity constant proportional to the relative damping ℜ ( s ) / ∣ s ∣ \Re(s)/|s| ℜ ( s ) /∣ s ∣ . The complex shift therefore improves the smoothing properties of standard relaxation schemes and mitigates the resonance phenomena that would otherwise arise at purely real frequencies; this is the mechanism exploited by shifted-Laplacian preconditioners for the Helmholtz equation [21 ] , where, in contrast to our setting, the shift is introduced artificially. This makes multigrid a viable option, although, as our experiments show, not over the entire range of frequencies produced by the CQ discretization. This observation is particularly important in three dimensions, where sparse direct solvers quickly become prohibitively expensive.
On the theoretical side, this article provides bounds for the solution operators of the transformed problems that are explicit in the Laplace parameter and in the wave speed (Proposition 4.1 ). These bounds place the problem within the framework of [6 ] and yield convergence of the time discretization for BDF1, BDF2, and the trapezoidal rule, with the order of the underlying multistep method and with constants that grow at most polynomially with the final time (Theorem 2.2). We also prove exponential decay estimates for the transformed solutions away from the inhomogeneity (Proposition 4.5). These estimates quantify the frequency window in which the boundary element correction is needed (Theorem 3.2 ) and show that the coupled operator is a contraction outside this window (Theorem 3.3 ).
The analysis in this article is restricted to the time discretization. The dependence on the spatial discretization of the BEM and FEM components is more delicate, but we believe that such an analysis is attainable in a future work. The damping introduced by the Laplace transform, with ℜ ( s ) ≥ σ 0 > 0 \Re(s) \ge\sigma_0 > 0 ℜ ( s ) ≥ σ 0 > 0 uniformly along the quadrature contour, keeps all transformed problems coercive and avoids the resonances associated with purely real frequencies. The main difficulty is therefore to track the dependence of the relevant constants on s s s , in the spirit of the estimates in § 5 and the semidiscrete analyses in [4 , 28 , 40 ] .
Let us point out, however, that this damping does not by itself prevent the pollution effect [37 ] , since a sizable portion of the quadrature frequencies has small relative damping ℜ ( s ) / ∣ s ∣ \Re(s)/|s| ℜ ( s ) /∣ s ∣ . In our experiments, however, the frequencies that carry non-negligible data are resolved by the mesh, while those that are not carry data only at the level of round-off error; see § 8 .
This paper is structured as follows. In § 2 , we introduce the physical model problem under consideration and the mathematical model used for its numerical approximation. We then describe the linear multistep CQ method employed for the time discretization, together with its all-steps-at-once implementation, and present the convergence results in time, relying on the well-established theory for CQ methods [6 ] ; see also [31 , 35 , 40 ] for multistep CQ based on BDF-type schemes. In § 4 , we explain how the transformed problems are addressed using our overlapping FEM–BEM framework and analyze the resulting coupled equation. A key point is the decay of the solution in the exterior of the inhomogeneity, which determines the form of the transformed problem and the role of the BEM component. This issue is examined in detail in § 5 , where we also establish bounds for the continuity constants with respect to the transformed variable—precisely the type of estimates required in the preceding sections. § 8 contains an extensive set of numerical experiments illustrating the efficiency of the algorithm and providing further support for the theoretical results. Finally, § 9 summarizes the main conclusions and outlines directions for future work.
A model problem and time-discretization algorithm
Transient acoustic wave problem
We consider a model problem for transient acoustic wave propagation in R m \mathbb{R}^{m} R m , m = 2 , 3 m = 2, 3 m = 2 , 3 , consisting of a heterogeneous domain Ω \Omega Ω embedded in an unbounded homogeneous exterior region. The wave speed c ∈ L ∞ ( R m ) c \in L^{\infty}(\mathbb{R}^{m}) c ∈ L ∞ ( R m ) is positive and may vary spatially and be discontinuous within Ω \Omega Ω . In the exterior free space Ω c = R m ∖ Ω ‾ \Omega^{c} = \mathbb{R}^{m} \setminus\overline{\Omega} Ω c = R m ∖ Ω , the wave speed is normalized to the constant c e x t = 1 c_{\mathrm{ext}} = 1 c ext = 1 .
The evolution of the total wave field in R m \mathbb{R}^{m} R m is described by the scalar field ϑ = ϑ ( x , t ) \vartheta= \vartheta(\mathbf{x}, t) ϑ = ϑ ( x , t ) , which satisfies the classical time-domain wave equation:
c − 2 ( x ) ϑ t t ( x , t ) = Δ ϑ ( x , t ) , x ∈ R m , t > 0. c^{-2}(\mathbf{x}) \vartheta_{tt}(\mathbf{x}, t) = \Delta\vartheta(\mathbf{x}, t), \qquad\mathbf{x} \in\mathbb{R}^{m},\ t > 0. c − 2 ( x ) ϑ tt ( x , t ) = Δ ϑ ( x , t ) , x ∈ R m , t > 0.
The total field is induced by a smooth incident field ϑ i n c \vartheta^{\mathrm{inc}} ϑ inc , which originates in the exterior region Ω c \Omega^{c} Ω c , and with support that does not intersect Ω \Omega Ω for t ≤ 0 t \leq0 t ≤ 0 . The incident field satisfies the wave equation
ϑ t t i n c ( x , t ) = Δ ϑ i n c ( x , t ) , x ∈ R m . \vartheta^{\mathrm{inc}}_{tt}(\mathbf{x}, t) = \Delta\vartheta^{\mathrm{inc}}(\mathbf{x}, t), \qquad\mathbf{x} \in\mathbb{R}^{m}. ϑ tt inc ( x , t ) = Δ ϑ inc ( x , t ) , x ∈ R m .
The interaction of this wave with the inhomogeneity generates a scattered and propagating field u = ϑ − ϑ i n c u = \vartheta- \vartheta^{\mathrm{inc}} u = ϑ − ϑ inc , which satisfies homogeneous initial data and the evolution equation
u ∈ H 2 ( [ 0 , T ] , L 2 ( R m ) ) , 1 c 2 u ¨ ( t ) = Δ u ( t ) + f ( t ) , t ∈ [ 0 , T ] , u ( 0 ) = 0 , u ˙ ( 0 ) = 0 , (1) \begin{aligned}
u &\in H^{2}([0,T], L^{2}(\mathbb{R}^{m})),\\
\frac{1}{c^{2}}\ddot{u}(t) &= \Delta u(t) + f(t), \qquad t \in[0,T],\\
u(0) &= 0,\qquad\dot{u}(0) = 0, \tag*{(1)}
\end{aligned} u c 2 1 u ¨ ( t ) u ( 0 ) ∈ H 2 ([ 0 , T ] , L 2 ( R m )) , = Δ u ( t ) + f ( t ) , t ∈ [ 0 , T ] , = 0 , u ˙ ( 0 ) = 0 , (1)
where
f ( x , t ) : = c R ( x ) ϑ t t i n c ( x , t ) , c R ( x ) : = 1 − c − 2 ( x ) f(\mathbf{x},t) := c_{\mathrm{R}}(\mathbf{x})\,\vartheta_{tt}^{\mathrm{inc}}(\mathbf{x},t), \qquad c_{\mathrm{R}}(\mathbf{x}) := 1-c^{-2}(\mathbf{x}) f ( x , t ) := c R ( x ) ϑ tt inc ( x , t ) , c R ( x ) := 1 − c − 2 ( x )
is supported inside Ω \Omega Ω since c ( x ) = 1 c(\mathbf{x})=1 c ( x ) = 1 in the free space (i.e. for x ∈ Ω c \mathbf{x}\in\Omega^{c} x ∈ Ω c ). The function c R c_{\mathrm{R}} c R is the contrast function and the regularity in space of the solution u u u depends on the regularity of c R c_{\mathrm{R}} c R as a function in R m \mathbb{R}^{m} R m .
Convolution Quadrature Method
Let us introduce, for now only in a formal way, the Laplace transform:
F ( s ) : = ( L f ) ( s ) : = ∫ 0 ∞ f ( t ) e − s t d t , s ∈ C + : = { s ∈ C : ℜ s > 0 } . F(s):=(\mathcal{L}f)(s):=\int_{0}^{\infty}f(t)e^{-st}\,\mathrm{d}t,\qquad s\in\mathbb{C}_{+}:=\{s\in\mathbb{C}:\Re s>0\}. F ( s ) := ( L f ) ( s ) := ∫ 0 ∞ f ( t ) e − s t d t , s ∈ C + := { s ∈ C : ℜ s > 0 } .
Then U = L u U=\mathcal{L}u U = L u solves the following problem:
U ( s ) ∈ H 1 ( R m ) , s 2 c 2 U ( s ) − Δ U ( s ) = F ( s ) . (2) \begin{aligned}
U(s)&\in H^{1}(\mathbb{R}^{m}),\\
\frac{s^{2}}{c^{2}}U(s)-\Delta U(s)&=F(s). \tag*{(2)}
\end{aligned} U ( s ) c 2 s 2 U ( s ) − Δ U ( s ) ∈ H 1 ( R m ) , = F ( s ) . (2)
The analysis of problem (2), together with the derivation of suitable bounds and its numerical resolution, will occupy most of this work. For now, we simply state that the problem is well posed and gives rise to a solution operator K ( s ) \mathrm{K}(s) K ( s ) such that
U = K ( s ) F , (3) U=\mathrm{K}(s)F, \tag*{(3)} U = K ( s ) F , (3)
where K ( s ) : L 2 ( Ω ) → H 1 ( R m ) \mathrm{K}(s):L^{2}(\Omega)\to H^{1}(\mathbb{R}^{m}) K ( s ) : L 2 ( Ω ) → H 1 ( R m ) ; its range is in fact contained in H 2 ( R m ) H^{2}(\mathbb{R}^{m}) H 2 ( R m ) . Moreover, K ( s ) \mathrm{K}(s) K ( s ) is analytic for s ∈ C + s\in\mathbb{C}_{+} s ∈ C + .
Fix t > 0 t>0 t > 0 and set f ~ ( τ ) : = f ( τ ) χ [ 0 , t ] ( τ ) \widetilde{f}(\tau):=f(\tau)\chi_{[0,t]}(\tau) f ( τ ) := f ( τ ) χ [ 0 , t ] ( τ ) . By causality of the wave equation (the solution of (1) at time t t t only depends on f ∣ [ 0 , t ] f|_{[0,t]} f ∣ [ 0 , t ] ), u ( t ) u(t) u ( t ) is also the value at t t t of the solution with datum f ~ \widetilde{f} f , so that, formally,
u ( t ) = 1 2 π i ∫ μ − i R μ + i R e ζ t K ( ζ ) F ~ ( ζ ) d ζ = 1 2 π i ∫ μ − i R μ + i R K ( ζ ) [ ∫ 0 ∞ f ~ ( τ ) e ζ ( t − τ ) d τ ] d ζ = 1 2 π i ∫ μ − i R μ + i R K ( ζ ) [ ∫ 0 t f ( τ ) e ζ ( t − τ ) d τ ] ⏟ y ( t ; ζ ) d ζ , μ > 0. (4) \begin{aligned}
u(t)&=\frac{1}{2\pi\mathrm{i}}\int_{\mu-\mathrm{i}\mathbb{R}}^{\mu+\mathrm{i}\mathbb{R}}e^{\zeta t}\mathrm{K}(\zeta)\widetilde{F}(\zeta)\,\mathrm{d}\zeta\\
&=\frac{1}{2\pi\mathrm{i}}\int_{\mu-\mathrm{i}\mathbb{R}}^{\mu+\mathrm{i}\mathbb{R}}\mathrm{K}(\zeta)\left[\int_{0}^{\infty}\widetilde{f}(\tau)e^{\zeta(t-\tau)}\,\mathrm{d}\tau\right]\mathrm{d}\zeta\\
&=\frac{1}{2\pi\mathrm{i}}\int_{\mu-\mathrm{i}\mathbb{R}}^{\mu+\mathrm{i}\mathbb{R}}\mathrm{K}(\zeta)\underbrace{\left[\int_{0}^{t}f(\tau)e^{\zeta(t-\tau)}\,\mathrm{d}\tau\right]}_{y(t;\zeta)}\,\mathrm{d}\zeta,\qquad\mu>0. \tag*{(4)}
\end{aligned} u ( t ) = 2 π i 1 ∫ μ − i R μ + i R e ζ t K ( ζ ) F ( ζ ) d ζ = 2 π i 1 ∫ μ − i R μ + i R K ( ζ ) [ ∫ 0 ∞ f ( τ ) e ζ ( t − τ ) d τ ] d ζ = 2 π i 1 ∫ μ − i R μ + i R K ( ζ ) y ( t ; ζ ) [ ∫ 0 t f ( τ ) e ζ ( t − τ ) d τ ] d ζ , μ > 0. (4)
(The argument can be made rigorous in L 2 ( R m ) L^{2}(\mathbb{R}^{m}) L 2 ( R m ) : by Proposition 4.1 below, and since F ~ ( ζ ) = O ( ∣ ζ ∣ − 1 ) \widetilde{F}(\zeta)=O(|\zeta|^{-1}) F ( ζ ) = O ( ∣ ζ ∣ − 1 ) , the integral exists in the classical sense.) Here, y ( t ; ζ ) y(t;\zeta) y ( t ; ζ ) , with ζ ∈ C + \zeta\in\mathbb{C}_{+} ζ ∈ C + , is the solution of
y ′ − ζ y = f , y ( 0 ) = 0. (5) y'-\zeta y=f,\qquad y(0)=0. \tag*{(5)} y ′ − ζ y = f , y ( 0 ) = 0. (5)
Multistep CQ methods are obtained by discretizing (4): define y κ ( t ; ζ ) y^{\kappa}(t;\zeta) y κ ( t ; ζ ) as the solution of the multistep recursion
1 κ ∑ j = 0 r α j y κ ( t − j κ ; ζ ) = ∑ j = 0 r β j f ( t − j κ ) + ζ ∑ j = 0 r β j y κ ( t − j κ ; ζ ) , y κ ( t ; ζ ) = 0 , for t ≤ 0. \frac{1}{\kappa}\sum_{j=0}^{r}\alpha_{j}y^{\kappa}(t-j\kappa;\zeta)=\sum_{j=0}^{r}\beta_{j}f(t-j\kappa)+\zeta\sum_{j=0}^{r}\beta_{j}y^{\kappa}(t-j\kappa;\zeta),\qquad y^{\kappa}(t;\zeta)=0,\ \text{for }t\leq0. κ 1 j = 0 ∑ r α j y κ ( t − j κ ; ζ ) = j = 0 ∑ r β j f ( t − j κ ) + ζ j = 0 ∑ r β j y κ ( t − j κ ; ζ ) , y κ ( t ; ζ ) = 0 , for t ≤ 0.
Applying the Laplace transform, with Y κ ( s ; ζ ) : = L ( y κ ( ⋅ ; ζ ) ) ( s ) Y^{\kappa}(s;\zeta):=\mathcal{L}(y^{\kappa}(\cdot;\zeta))(s) Y κ ( s ; ζ ) := L ( y κ ( ⋅ ; ζ )) ( s ) we have
Y κ ( s ; ζ ) = 1 δ ( e − κ s ) κ − ζ F ( s ) . Y^{\kappa}(s;\zeta)=\frac{1}{\frac{\delta(e^{-\kappa s})}{\kappa}-\zeta}F(s). Y κ ( s ; ζ ) = κ δ ( e − κ s ) − ζ 1 F ( s ) .
where
δ ( z ) = α r z r + α r − 1 z r − 1 + … + α 0 β r z r + β r − 1 z r − 1 + … + β 0 \delta(z)=\frac{\alpha_r z^r+\alpha_{r-1}z^{r-1}+\ldots+\alpha_0}{\beta_r z^r+\beta_{r-1}z^{r-1}+\ldots+\beta_0} δ ( z ) = β r z r + β r − 1 z r − 1 + … + β 0 α r z r + α r − 1 z r − 1 + … + α 0
is the generating function of the linear multistep method, also known in the CQ literature as the discrete differentiation symbol.
The CQ approximation is then defined by replacing y y y with y κ y^\kappa y κ in (3): with σ = ℜ s > 0 \sigma=\Re s>0 σ = ℜ s > 0 ,
u κ ( t ) = 1 2 π i ∫ μ + i R K ( ζ ) y κ ( t ; ζ ) d ζ = 1 2 π i ∫ μ + i R K ( ζ ) L − 1 ( Y κ ( ⋅ ; ζ ) ) ( t ) d ζ = 1 2 π i ∫ μ + i R K ( ζ ) [ 1 2 π i ∫ σ + i R e t s 1 δ ( e − κ s ) κ − ζ F ( s ) d s ] d ζ = 1 2 π i ∫ σ + i R e t s [ 1 2 π i ∫ μ + i R K ( ζ ) 1 δ ( e − κ s ) κ − ζ d ζ ] F ( s ) d s = 1 2 π i ∫ σ + i R e t s K ( δ ( e − κ s ) κ ) F ( s ) d s . (6) \begin{aligned}
u^\kappa(t)=\frac{1}{2\pi i}\int_{\mu+i\mathbb{R}}K(\zeta)y^\kappa(t;\zeta)\,\mathrm{d}\zeta=\frac{1}{2\pi i}\int_{\mu+i\mathbb{R}}K(\zeta)\mathcal{L}^{-1}(Y^\kappa(\cdot;\zeta))(t)\,\mathrm{d}\zeta\\
&=\frac{1}{2\pi i}\int_{\mu+i\mathbb{R}}K(\zeta)\left[\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{ts}\frac{1}{\frac{\delta(e^{-\kappa s})}{\kappa}-\zeta}F(s)\,\mathrm{d}s\right]\mathrm{d}\zeta\\
&=\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{ts}\left[\frac{1}{2\pi i}\int_{\mu+i\mathbb{R}}K(\zeta)\frac{1}{\frac{\delta(e^{-\kappa s})}{\kappa}-\zeta}\,\mathrm{d}\zeta\right]F(s)\,\mathrm{d}s \\
&=\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{ts}K\left(\frac{\delta(e^{-\kappa s})}{\kappa}\right)F(s)\,\mathrm{d}s. \tag*{(6)}
\end{aligned} u κ ( t ) = 2 π i 1 ∫ μ + i R K ( ζ ) y κ ( t ; ζ ) d ζ = 2 π i 1 ∫ μ + i R K ( ζ ) L − 1 ( Y κ ( ⋅ ; ζ )) ( t ) d ζ = 2 π i 1 ∫ μ + i R K ( ζ ) [ 2 π i 1 ∫ σ + i R e t s κ δ ( e − κ s ) − ζ 1 F ( s ) d s ] d ζ = 2 π i 1 ∫ σ + i R e t s [ 2 π i 1 ∫ μ + i R K ( ζ ) κ δ ( e − κ s ) − ζ 1 d ζ ] F ( s ) d s = 2 π i 1 ∫ σ + i R e t s K ( κ δ ( e − κ s ) ) F ( s ) d s . (6)
Observe that the Cauchy integral formula has been applied above in the last step. This requires, first, that K ( δ ( e − κ s ) / κ ) K(\delta(e^{-\kappa s})/\kappa) K ( δ ( e − κ s ) / κ ) be analytic for ℜ s > 0 \Re s>0 ℜ s > 0 , i.e. ℜ δ ( z ) > 0 \Re\delta(z)>0 ℜ δ ( z ) > 0 for ∣ z ∣ < 1 |z|<1 ∣ z ∣ < 1 , which imposes a severe restriction on the multistep method (A–stability, see Remark 2.1); and, second, that the pole ζ = δ ( e − κ s ) / κ \zeta=\delta(e^{-\kappa s})/\kappa ζ = δ ( e − κ s ) / κ lie to the right of the contour μ + i R \mu+i\mathbb{R} μ + i R for every s ∈ σ + i R s\in\sigma+i\mathbb{R} s ∈ σ + i R , that is, μ < inf s ∈ σ + i R ℜ δ ( e − κ s ) / κ \mu<\inf_{s\in\sigma+i\mathbb{R}}\Re\delta(e^{-\kappa s})/\kappa μ < inf s ∈ σ + i R ℜ δ ( e − κ s ) / κ . For the three methods considered below this infimum equals δ ( e − κ σ ) / κ < σ \delta(e^{-\kappa\sigma})/\kappa<\sigma δ ( e − κσ ) / κ < σ , so that μ \mu μ has to be taken strictly smaller than the abscissa σ \sigma σ of the outer integral.
The same assumption makes z ↦ K ( δ ( z ) / κ ) z\mapsto K(\delta(z)/\kappa) z ↦ K ( δ ( z ) / κ ) analytic in the unit disc, so that it admits the power series expansion
K ( 1 κ δ ( z ) ) = ∑ j = 0 ∞ W j ( κ ) z j , (7) K\left(\frac{1}{\kappa}\delta(z)\right)=\sum_{j=0}^{\infty}W_j(\kappa)z^j, \tag*{(7)} K ( κ 1 δ ( z ) ) = j = 0 ∑ ∞ W j ( κ ) z j , (7)
which converges absolutely on ∣ z ∣ = e − κ σ < 1 |z|=e^{-\kappa\sigma}<1 ∣ z ∣ = e − κσ < 1 . Inserting it in (5) and exchanging sum and integral,
u κ ( t ) = ∑ j = 0 ∞ W j ( κ ) [ 1 2 π i ∫ σ + i R e s ( t − j κ ) F ( s ) d s ] = ∑ j = 0 ∞ W j ( κ ) f ( t − j κ ) , (8) u^\kappa(t)=\sum_{j=0}^{\infty}W_j(\kappa)\left[\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{s(t-j\kappa)}F(s)\,\mathrm{d}s\right]=\sum_{j=0}^{\infty}W_j(\kappa)f(t-j\kappa), \tag*{(8)} u κ ( t ) = j = 0 ∑ ∞ W j ( κ ) [ 2 π i 1 ∫ σ + i R e s ( t − j κ ) F ( s ) d s ] = j = 0 ∑ ∞ W j ( κ ) f ( t − j κ ) , (8)
where f f f is extended by zero to t < 0 t<0 t < 0 , so that by causality the sum is finite.
Remark 2.1 Convolution Quadrature methods have been an active topic of research since the seminal papers of Lubich [33 , 34 ] , originally developed for the solution of both parabolic and hyperbolic problems. The method hinges on the expansion (6) and the calculation, or approximation, of the weights { W j ( κ ) } j = 0 ∞ \{W_j(\kappa)\}_{j=0}^{\infty} { W j ( κ ) } j = 0 ∞ . For such an expansion to exist the linear multistep method has to be A–stable, as mentioned above, and by the second Dahlquist barrier no A–stable linear multistep method has order higher than two. Higher order in time can be achieved with Runge–Kutta based CQ, a case that we will not consider here.
Let us also point out that convolution quadrature is compatible with composition, K − 1 ( ∂ t κ ) K ( ∂ t κ ) = I K^{-1}(\partial_t^\kappa)K(\partial_t^\kappa)=I K − 1 ( ∂ t κ ) K ( ∂ t κ ) = I , so that u κ u^\kappa u κ can equivalently be characterized as the solution of
K − 1 ( δ ( e − κ s ) κ ) U κ ( s ) = F ( s ) K^{-1}\left(\frac{\delta(e^{-\kappa s})}{\kappa}\right)U^\kappa(s)=F(s) K − 1 ( κ δ ( e − κ s ) ) U κ ( s ) = F ( s )
i.e. as the result of applying the multistep method directly to the wave equation. For the trapezoidal rule, since
K − 1 ( δ ( z ) κ ) = 4 1 + 2 z + z 2 [ 1 − 2 z + z 2 κ 2 c 2 − 1 + 2 z + z 2 4 Δ ] , K^{-1}\left(\frac{\delta(z)}{\kappa}\right)=\frac{4}{1+2z+z^2}\left[\frac{1-2z+z^2}{\kappa^2c^2}-\frac{1+2z+z^2}{4}\Delta\right], K − 1 ( κ δ ( z ) ) = 1 + 2 z + z 2 4 [ κ 2 c 2 1 − 2 z + z 2 − 4 1 + 2 z + z 2 Δ ] ,
this is the recursion
1 c 2 u κ ( t ) − 2 u κ ( t − κ ) + u κ ( t − 2 κ ) κ 2 − 1 4 Δ ( u κ ( t ) + 2 u κ ( t − κ ) + u κ ( t − 2 κ ) ) = 1 4 ( f ( t ) + 2 f ( t − κ ) + f ( t − 2 κ ) ) . \frac{1}{c^2}\frac{u^\kappa(t)-2u^\kappa(t-\kappa)+u^\kappa(t-2\kappa)}{\kappa^2}-\frac{1}{4}\Delta\left(u^\kappa(t)+2u^\kappa(t-\kappa)+u^\kappa(t-2\kappa)\right)
=\frac{1}{4}\left(f(t)+2f(t-\kappa)+f(t-2\kappa)\right). c 2 1 κ 2 u κ ( t ) − 2 u κ ( t − κ ) + u κ ( t − 2 κ ) − 4 1 Δ ( u κ ( t ) + 2 u κ ( t − κ ) + u κ ( t − 2 κ ) ) = 4 1 ( f ( t ) + 2 f ( t − κ ) + f ( t − 2 κ ) ) .
that is, the Newmark method with γ = 1 / 2 \gamma=1/2 γ = 1/2 , β = 1 / 4 \beta=1/4 β = 1/4 , unconditionally stable and of order two. The all–steps–at–once implementation of Subsection 2.4 and this marching form compute the same discrete solution; the choice between them is one of algorithm, not of discretization.
Convergence of linear multistep CQ methods
We restrict our attention to the BDF1, BDF2 and trapezoidal–rule CQ methods; see Remark 2.3 below for other linear multistep methods. Following C. Lubich [33 ] , and as in [6 , 40 ] , the CQ solution is written as
K ( ∂ t κ ) f ( t ) : = 1 2 π i ∫ σ + i R e s t K ( s κ ) F ( s ) d s = U κ ( t ) , s κ : = δ ( e − s κ ) κ , K(\partial_t^\kappa)f(t):=\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{st}K(s^\kappa)F(s)\,\mathrm{d}s=U^\kappa(t),\qquad s^\kappa:=\frac{\delta(e^{-s\kappa})}{\kappa}, K ( ∂ t κ ) f ( t ) := 2 π i 1 ∫ σ + i R e s t K ( s κ ) F ( s ) d s = U κ ( t ) , s κ := κ δ ( e − s κ ) ,
as an approximation of the exact solution
K ( ∂ t ) f ( t ) : = 1 2 π i ∫ σ + i R e s t K ( s ) F ( s ) d s = U ( t ) . K(\partial_t)f(t):=\frac{1}{2\pi i}\int_{\sigma+i\mathbb{R}}e^{st}K(s)F(s)\,\mathrm{d}s=U(t). K ( ∂ t ) f ( t ) := 2 π i 1 ∫ σ + i R e s t K ( s ) F ( s ) d s = U ( t ) .
The use of ∂ t \partial_t ∂ t as the argument of K K K emphasizes the role of the Laplace transform (since L ( ∂ t ) = s \mathcal{L}(\partial_t)=s L ( ∂ t ) = s ) in both the definition of the method and the convergence analysis, even though the Laplace transform is never actually applied to the right–hand side of the wave equation.
We will consider the linear multistep CQ method for three representative cases: for BDF1 (implicit Euler), with
δ E ( z ) = 1 − z , \delta_{\mathrm{E}}(z)=1-z, δ E ( z ) = 1 − z ,
of order p = 1 p=1 p = 1 ; for BDF2,
δ B D F 2 ( z ) = 1 2 z 2 − 2 z + 3 2 , \delta_{\mathrm{BDF2}}(z)=\frac{1}{2}z^2-2z+\frac{3}{2}, δ BDF2 ( z ) = 2 1 z 2 − 2 z + 2 3 ,
and for the trapezoidal rule,
δ T R ( z ) = 2 ( 1 − z ) 1 + z , \delta_{\mathrm{TR}}(z)=\frac{2(1-z)}{1+z}, δ TR ( z ) = 1 + z 2 ( 1 − z ) ,
of order p = 2 p=2 p = 2 . The proof of the following theorem is given in Section 4.4.
Theorem 2.2 Let δ \delta δ be the generating function of BDF1 (p = 1 p=1 p = 1 ), BDF2 or the trapezoidal rule (p = 2 p=2 p = 2 ). Let r ∈ [ 1 , 2 ] r \in[1,2] r ∈ [ 1 , 2 ] , let p ~ ∈ N \widetilde{p} \in\mathbb{N} p ∈ N with p ~ > r + p + 1 \widetilde{p} > r+p+1 p > r + p + 1 , and let f ∈ C p ~ ( [ 0 , T ] ; L 2 ( Ω ) ) f \in C^{\widetilde{p}}([0,T];L^{2}(\Omega)) f ∈ C p ([ 0 , T ] ; L 2 ( Ω )) satisfy f ( 0 ) = ⋯ = f ( p ~ − 1 ) ( 0 ) = 0 f(0)=\cdots=f^{(\widetilde{p}-1)}(0)=0 f ( 0 ) = ⋯ = f ( p − 1 ) ( 0 ) = 0 . Then there exist κ 0 > 0 \kappa_{0}>0 κ 0 > 0 and C > 0 C>0 C > 0 , depending only on p p p , r r r and δ \delta δ , such that for all κ ∈ ( 0 , κ 0 ) \kappa\in(0,\kappa_{0}) κ ∈ ( 0 , κ 0 ) , σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) and t ∈ ( 0 , T ] t \in(0,T] t ∈ ( 0 , T ]
∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C c max 2 c min − 2 ( r − 1 ) σ p − 1 − p ~ e σ t κ p ∫ 0 t e − σ τ ∥ f ( p ~ ) ( τ ) ∥ L 2 ( Ω ) d τ . \lVert K(\partial_{t}^{\kappa})f(t)-K(\partial_{t})f(t)\rVert_{H^{r}(\mathbb{R}^{m})}\leq Cc_{\max}^{2}c_{\min}^{-2(r-1)}\sigma^{p-1-\widetilde{p}}e^{\sigma t}\kappa^{p}\int_{0}^{t}e^{-\sigma\tau}\lVert f^{(\widetilde{p})}(\tau)\rVert_{L^{2}(\Omega)}\,\mathrm{d}\tau. ∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C c m a x 2 c m i n − 2 ( r − 1 ) σ p − 1 − p e σ t κ p ∫ 0 t e − σ τ ∥ f ( p ) ( τ ) ∥ L 2 ( Ω ) d τ .
Remark 2.3 Convergence in H r ( R m ) H^{r}(\mathbb{R}^{m}) H r ( R m ) for r ∈ [ 0 , 1 ) r\in[0,1) r ∈ [ 0 , 1 ) needs no separate statement, since the case r = 1 r=1 r = 1 already provides it. For r ∈ [ 1 , 2 ) r\in[1,2) r ∈ [ 1 , 2 ) it suffices to take p ~ = p + 3 \widetilde{p}=p+3 p = p + 3 , and the constant in Theorem 2.2 is then proportional to σ − 4 \sigma^{-4} σ − 4 ; for r = 2 r=2 r = 2 one needs p ~ = p + 4 \widetilde{p}=p+4 p = p + 4 , and the constant is proportional to σ − 5 \sigma^{-5} σ − 5 . The dependence on σ \sigma σ is the same for the three schemes. The cases r = 1 + ε r=1+\varepsilon r = 1 + ε for m = 2 m=2 m = 2 and r = 3 / 2 + ε r=3/2+\varepsilon r = 3/2 + ε for m = 3 m=3 m = 3 yield, via the Sobolev embedding theorem, estimates in the uniform norm with p ~ = p + 3 \widetilde{p}=p+3 p = p + 3 . Choosing σ = 1 / t \sigma=1/t σ = 1/ t for t > 1 t>1 t > 1 (and, say, σ = 1 / 2 \sigma=1/2 σ = 1/2 for t ≤ 1 t\leq1 t ≤ 1 ), the estimate takes the form
∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C max { 1 , t } p ~ + 1 − p κ p ∫ 0 t ∥ f ( p ~ ) ( τ ) ∥ L 2 ( Ω ) d τ , \lVert K(\partial_{t}^{\kappa})f(t)-K(\partial_{t})f(t)\rVert_{H^{r}(\mathbb{R}^{m})}\leq C\max\{1,t\}^{\widetilde{p}+1-p}\kappa^{p}\int_{0}^{t}\lVert f^{(\widetilde{p})}(\tau)\rVert_{L^{2}(\Omega)}\,\mathrm{d}\tau, ∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C max { 1 , t } p + 1 − p κ p ∫ 0 t ∥ f ( p ) ( τ ) ∥ L 2 ( Ω ) d τ ,
so that the constant grows at most like t 4 t^{4} t 4 (r < 2 r<2 r < 2 ) or t 5 t^{5} t 5 (r = 2 r=2 r = 2 ) with the final time, with no exponential factor.
Theorem 2.2 is stated for the three multistep methods used in this paper, but nothing in its proof is specific to them: the derivation of § 2 only uses that ℜ δ ( z ) > 0 \Re\delta(z)>0 ℜ δ ( z ) > 0 for ∣ z ∣ < 1 |z|<1 ∣ z ∣ < 1 , and the convergence results of [6 ] on which the proof rests apply to any A-stable linear multistep method whose generating function satisfies the hypotheses stated there, in which case the theorem holds with p p p , the order of the method. Among the A-stable methods of order two, BDF2 and the trapezoidal rule are the two extreme choices, maximal damping of the high frequencies and none at all, and they lead to markedly different distributions of the quadrature frequencies, as discussed in § 8 . Intermediate schemes, such as the truncated trapezoidal rules mentioned in [3 ] , which maintain a polynomial generating function and thus exhibit bounded quadrature frequencies similar to BDF2, can also be considered. Their error constants are close to those of the trapezoidal rule, but still satisfy the criteria for applying the convergence result from Theorem 2.2.
All–steps–at–once implementation
The discretization in time thus depends on computing sufficiently accurate approximations of
W j ( κ ) = 1 j ! d j d z j K ( 1 κ δ ( z ) ) ∣ z = 0 = 1 2 π i ∮ ∣ z ∣ = r z − j − 1 K ( 1 κ δ ( z ) ) d z = [ 1 r j ∫ 0 1 exp ( − 2 π i j x ) K ( 1 κ δ ( r exp ( 2 π i x ) ) ) d x ] . (9) \begin{aligned}
W_{j}(\kappa)=\left.\frac{1}{j!}\frac{\mathrm{d}^{j}}{\mathrm{d}z^{j}}K\left(\frac{1}{\kappa}\delta(z)\right)\right|_{z=0}
=\frac{1}{2\pi\mathrm{i}}\oint_{|z|=r}z^{-j-1}K\left(\frac{1}{\kappa}\delta(z)\right)\,\mathrm{d}z \\
&=\left[\frac{1}{r^{j}}\int_{0}^{1}\exp(-2\pi\mathrm{i}jx)K\left(\frac{1}{\kappa}\delta\left(r\exp(2\pi\mathrm{i}x)\right)\right)\,\mathrm{d}x\right]. \tag*{(9)}
\end{aligned} W j ( κ ) = j ! 1 d z j d j K ( κ 1 δ ( z ) ) z = 0 = 2 π i 1 ∮ ∣ z ∣ = r z − j − 1 K ( κ 1 δ ( z ) ) d z = [ r j 1 ∫ 0 1 exp ( − 2 π i j x ) K ( κ 1 δ ( r exp ( 2 π i x ) ) ) d x ] . (9)
(by the Cauchy integral formula on the circle ∣ z ∣ = r |z|=r ∣ z ∣ = r , 0 < r < 1 0<r<1 0 < r < 1 , which lies inside the disc of analyticity of z ↦ K ( δ ( z ) / κ ) z\mapsto K(\delta(z)/\kappa) z ↦ K ( δ ( z ) / κ ) ). Take a positive integer n Q n_{Q} n Q and set
e ℓ , n Q : = exp ( 2 π i ℓ n Q + 1 ) , z ℓ , n Q κ = 1 κ δ ( r e ℓ , n Q ) , e_{\ell,n_{Q}}:=\exp\left(\frac{2\pi\mathrm{i}\ell}{n_{Q}+1}\right),\qquad z_{\ell,n_{Q}}^{\kappa}=\frac{1}{\kappa}\delta\left(re_{\ell,n_{Q}}\right), e ℓ , n Q := exp ( n Q + 1 2 π i ℓ ) , z ℓ , n Q κ = κ 1 δ ( r e ℓ , n Q ) ,
@page 8/40 starts_mid=0 ends_mid=1
Theorem 2.2 Let δ \delta δ be the generating function of BDF1 (p = 1 p=1 p = 1 ), BDF2 or the trapezoidal rule (p = 2 p=2 p = 2 ). Let r ∈ [ 1 , 2 ] r \in[1,2] r ∈ [ 1 , 2 ] , let p ~ ∈ N \widetilde{p} \in\mathbb{N} p ∈ N with p ~ > r + p + 1 \widetilde{p} > r+p+1 p > r + p + 1 , and let f ∈ C p ~ ( [ 0 , T ] ; L 2 ( Ω ) ) f \in C^{\widetilde{p}}([0,T];L^{2}(\Omega)) f ∈ C p ([ 0 , T ] ; L 2 ( Ω )) satisfy f ( 0 ) = ⋯ = f ( p ~ − 1 ) ( 0 ) = 0 f(0)=\cdots=f^{(\widetilde{p}-1)}(0)=0 f ( 0 ) = ⋯ = f ( p − 1 ) ( 0 ) = 0 . Then there exist κ 0 > 0 \kappa_{0}>0 κ 0 > 0 and C > 0 C>0 C > 0 , depending only on p p p , r r r and δ \delta δ , such that for all κ ∈ ( 0 , κ 0 ) \kappa\in(0,\kappa_{0}) κ ∈ ( 0 , κ 0 ) , σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) and t ∈ ( 0 , T ] t \in(0,T] t ∈ ( 0 , T ]
∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C c max 2 c min − 2 ( r − 1 ) σ p − 1 − p ~ e σ t κ p ∫ 0 t e − σ τ ∥ f ( p ~ ) ( τ ) ∥ L 2 ( Ω ) d τ . \lVert K(\partial_{t}^{\kappa})f(t)-K(\partial_{t})f(t)\rVert_{H^{r}(\mathbb{R}^{m})}\leq Cc_{\max}^{2}c_{\min}^{-2(r-1)}\sigma^{p-1-\widetilde{p}}e^{\sigma t}\kappa^{p}\int_{0}^{t}e^{-\sigma\tau}\lVert f^{(\widetilde{p})}(\tau)\rVert_{L^{2}(\Omega)}\,\mathrm{d}\tau. ∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C c m a x 2 c m i n − 2 ( r − 1 ) σ p − 1 − p e σ t κ p ∫ 0 t e − σ τ ∥ f ( p ) ( τ ) ∥ L 2 ( Ω ) d τ .
Remark 2.3 Convergence in H r ( R m ) H^{r}(\mathbb{R}^{m}) H r ( R m ) for r ∈ [ 0 , 1 ) r\in[0,1) r ∈ [ 0 , 1 ) needs no separate statement, since the case r = 1 r=1 r = 1 already provides it. For r ∈ [ 1 , 2 ) r\in[1,2) r ∈ [ 1 , 2 ) it suffices to take p ~ = p + 3 \widetilde{p}=p+3 p = p + 3 , and the constant in Theorem 2.2 is then proportional to σ − 4 \sigma^{-4} σ − 4 ; for r = 2 r=2 r = 2 one needs p ~ = p + 4 \widetilde{p}=p+4 p = p + 4 , and the constant is proportional to σ − 5 \sigma^{-5} σ − 5 . The dependence on σ \sigma σ is the same for the three schemes. The cases r = 1 + ε r=1+\varepsilon r = 1 + ε for m = 2 m=2 m = 2 and r = 3 / 2 + ε r=3/2+\varepsilon r = 3/2 + ε for m = 3 m=3 m = 3 yield, via the Sobolev embedding theorem, estimates in the uniform norm with p ~ = p + 3 \widetilde{p}=p+3 p = p + 3 . Choosing σ = 1 / t \sigma=1/t σ = 1/ t for t > 1 t>1 t > 1 (and, say, σ = 1 / 2 \sigma=1/2 σ = 1/2 for t ≤ 1 t\leq1 t ≤ 1 ), the estimate takes the form
∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C max { 1 , t } p ~ + 1 − p κ p ∫ 0 t ∥ f ( p ~ ) ( τ ) ∥ L 2 ( Ω ) d τ , \lVert K(\partial_{t}^{\kappa})f(t)-K(\partial_{t})f(t)\rVert_{H^{r}(\mathbb{R}^{m})}\leq C\max\{1,t\}^{\widetilde{p}+1-p}\kappa^{p}\int_{0}^{t}\lVert f^{(\widetilde{p})}(\tau)\rVert_{L^{2}(\Omega)}\,\mathrm{d}\tau, ∥ K ( ∂ t κ ) f ( t ) − K ( ∂ t ) f ( t ) ∥ H r ( R m ) ≤ C max { 1 , t } p + 1 − p κ p ∫ 0 t ∥ f ( p ) ( τ ) ∥ L 2 ( Ω ) d τ ,
so that the constant grows at most like t 4 t^{4} t 4 (r < 2 r<2 r < 2 ) or t 5 t^{5} t 5 (r = 2 r=2 r = 2 ) with the final time, with no exponential factor.
Theorem 2.2 is stated for the three multistep methods used in this paper, but nothing in its proof is specific to them: the derivation of § 2 only uses that ℜ δ ( z ) > 0 \Re\delta(z)>0 ℜ δ ( z ) > 0 for ∣ z ∣ < 1 |z|<1 ∣ z ∣ < 1 , and the convergence results of [6 ] on which the proof rests apply to any A-stable linear multistep method whose generating function satisfies the hypotheses stated there, in which case the theorem holds with p p p , the order of the method. Among the A-stable methods of order two, BDF2 and the trapezoidal rule are the two extreme choices, maximal damping of the high frequencies and none at all, and they lead to markedly different distributions of the quadrature frequencies, as discussed in § 8 . Intermediate schemes, such as the truncated trapezoidal rules mentioned in [3 ] , which maintain a polynomial generating function and thus exhibit bounded quadrature frequencies similar to BDF2, can also be considered. Their error constants are close to those of the trapezoidal rule, but still satisfy the criteria for applying the convergence result from Theorem 2.2.
All–steps–at–once implementation
The discretization in time thus depends on computing sufficiently accurate approximations of
W j ( κ ) = 1 j ! d j d z j K ( 1 κ δ ( z ) ) ∣ z = 0 = 1 2 π i ∮ ∣ z ∣ = r z − j − 1 K ( 1 κ δ ( z ) ) d z = [ 1 r j ∫ 0 1 exp ( − 2 π i j x ) K ( 1 κ δ ( r exp ( 2 π i x ) ) ) d x ] . (10) \begin{aligned}
W_{j}(\kappa)=\left.\frac{1}{j!}\frac{\mathrm{d}^{j}}{\mathrm{d}z^{j}}K\left(\frac{1}{\kappa}\delta(z)\right)\right|_{z=0}
=\frac{1}{2\pi\mathrm{i}}\oint_{|z|=r}z^{-j-1}K\left(\frac{1}{\kappa}\delta(z)\right)\,\mathrm{d}z \\
&=\left[\frac{1}{r^{j}}\int_{0}^{1}\exp(-2\pi\mathrm{i}jx)K\left(\frac{1}{\kappa}\delta\left(r\exp(2\pi\mathrm{i}x)\right)\right)\,\mathrm{d}x\right]. \tag*{(10)}
\end{aligned} W j ( κ ) = j ! 1 d z j d j K ( κ 1 δ ( z ) ) z = 0 = 2 π i 1 ∮ ∣ z ∣ = r z − j − 1 K ( κ 1 δ ( z ) ) d z = [ r j 1 ∫ 0 1 exp ( − 2 π i j x ) K ( κ 1 δ ( r exp ( 2 π i x ) ) ) d x ] . (10)
(by the Cauchy integral formula on the circle ∣ z ∣ = r |z|=r ∣ z ∣ = r , 0 < r < 1 0<r<1 0 < r < 1 , which lies inside the disc of analyticity of z ↦ K ( δ ( z ) / κ ) z\mapsto K(\delta(z)/\kappa) z ↦ K ( δ ( z ) / κ ) ). Take a positive integer n Q n_{Q} n Q and set
e ℓ , n Q : = exp ( 2 π i ℓ n Q + 1 ) , z ℓ , n Q κ = 1 κ δ ( r e ℓ , n Q ) , (11) e_{\ell,n_{Q}}:=\exp\left(\frac{2\pi\mathrm{i}\ell}{n_{Q}+1}\right),\qquad z_{\ell,n_{Q}}^{\kappa}=\frac{1}{\kappa}\delta\left(re_{\ell,n_{Q}}\right), \tag*{(11)} e ℓ , n Q := exp ( n Q + 1 2 π i ℓ ) , z ℓ , n Q κ = κ 1 δ ( r e ℓ , n Q ) , (11)
and define the following approximations for W j ( κ ) W_j(\kappa) W j ( κ ) :
W j n Q ( κ ) g : = 1 n Q + 1 r − j ∑ ℓ = 0 n Q e − ℓ , n Q K ( z ℓ , n Q κ ) g . (12) W_j^{n_{\mathrm Q}}(\kappa)g := \frac{1}{n_{\mathrm Q}+1}r^{-j}\sum_{\ell=0}^{n_{\mathrm Q}}e_{-\ell,n_{\mathrm Q}}K(z_{\ell,n_{\mathrm Q}}^\kappa)g. \tag*{(12)} W j n Q ( κ ) g := n Q + 1 1 r − j ℓ = 0 ∑ n Q e − ℓ , n Q K ( z ℓ , n Q κ ) g . (12)
That is, the weights are approximated by the trapezoidal rule on a uniform grid with n Q + 1 n_{\mathrm Q}+1 n Q + 1 points, which leads to the following approximation of u κ u^\kappa u κ (cf. (8) )
u n Q κ ( t ) : = ∑ j = 0 ⌊ t / κ ⌋ W j n Q ( κ ) f ( t − j κ ) . u_{n_{\mathrm Q}}^\kappa(t) := \sum_{j=0}^{\lfloor t/\kappa\rfloor}W_j^{n_{\mathrm Q}}(\kappa)f(t-j\kappa). u n Q κ ( t ) := j = 0 ∑ ⌊ t / κ ⌋ W j n Q ( κ ) f ( t − j κ ) .
Let n C Q n_{\mathrm{CQ}} n CQ denote from now on the number of time steps, so that the final time is T = n C Q κ T=n_{\mathrm{CQ}}\kappa T = n CQ κ . We next apply (9) to approximate W n − j n Q ( κ ) W_{n-j}^{n_{\mathrm Q}}(\kappa) W n − j n Q ( κ ) , even for j > n j>n j > n (notice that W n = 0 W_n=0 W n = 0 for n < 0 n<0 n < 0 ). We then define
u ( n κ ) ≈ u κ , n Q ( n κ ) : = ∑ j = 0 n C Q 1 n Q + 1 r j − n [ ∑ ℓ = 0 n Q e ( j − n ) ℓ , n Q K ( z ℓ , n Q κ ) f ( j κ ) ] . u(n\kappa)\approx u^{\kappa,n_{\mathrm Q}}(n\kappa):=\sum_{j=0}^{n_{\mathrm{CQ}}}\frac{1}{n_{\mathrm Q}+1}r^{j-n}\left[\sum_{\ell=0}^{n_{\mathrm Q}}e^{(j-n)\ell,n_{\mathrm Q}}K(z_{\ell,n_{\mathrm Q}}^\kappa)f(j\kappa)\right]. u ( nκ ) ≈ u κ , n Q ( nκ ) := j = 0 ∑ n CQ n Q + 1 1 r j − n [ ℓ = 0 ∑ n Q e ( j − n ) ℓ , n Q K ( z ℓ , n Q κ ) f ( j κ ) ] .
Interchanging the order of summation,
u κ , n Q ( n κ ) = r − n ∑ ℓ = 0 n Q e − ℓ , n Q K ( z ℓ , n Q κ ) [ 1 n Q + 1 ∑ j = 0 n C Q r j e j ℓ , n Q f ( j κ ) ] . u^{\kappa,n_{\mathrm Q}}(n\kappa)=r^{-n}\sum_{\ell=0}^{n_{\mathrm Q}}e_{-\ell,n_{\mathrm Q}}K(z_{\ell,n_{\mathrm Q}}^\kappa)\left[\frac{1}{n_{\mathrm Q}+1}\sum_{j=0}^{n_{\mathrm{CQ}}}r^je_{j\ell,n_{\mathrm Q}}f(j\kappa)\right]. u κ , n Q ( nκ ) = r − n ℓ = 0 ∑ n Q e − ℓ , n Q K ( z ℓ , n Q κ ) [ n Q + 1 1 j = 0 ∑ n CQ r j e j ℓ , n Q f ( j κ ) ] .
Although the approximation (8)–(9) was already considered in the seminal papers of C. Lubich [33 , 34 ] , this implementation—commonly referred to in the literature as the all-steps-at-once approach—was introduced in [5 ] .
The approximation (9) requires n Q + 1 n_{\mathrm Q}+1 n Q + 1 evaluations of K K K , half of them by symmetry (see (17) below), and does not affect the accuracy of the method for the choice of r r r indicated below [5 ] . Its interest is that it allows the whole computation to be organized around the Fast Fourier Transform, which is the so-called all-steps-at-once implementation:
Algorithm 1 (all-steps-at-once evaluation of u κ u^\kappa u κ with n Q = n C Q n_{\mathrm Q}=n_{\mathrm{CQ}} n Q = n CQ )
Input: f ( j κ ) f(j\kappa) f ( j κ ) , j = 0 , … , n Q j=0,\ldots,n_{\mathrm Q} j = 0 , … , n Q ; radius 0 < r < 1 0<r<1 0 < r < 1 ; generating function δ \delta δ .
Output: u ~ r κ ( n κ ) \widetilde{u}_r^\kappa(n\kappa) u r κ ( nκ ) , n = 1 , … , n Q n=1,\ldots,n_{\mathrm Q} n = 1 , … , n Q , approximating u κ ( n κ ) u^\kappa(n\kappa) u κ ( nκ ) in (7).
Scale the data:
f r κ : = ( r j f ( j κ ) ) j = 0 n Q ∈ [ L 2 ( Ω ) ] n Q + 1 . (13) f_r^\kappa:= \left(r^jf(j\kappa)\right)_{j=0}^{n_{\mathrm Q}}\in\left[L^2(\Omega)\right]^{n_{\mathrm Q}+1}. \tag*{(13)} f r κ := ( r j f ( j κ ) ) j = 0 n Q ∈ [ L 2 ( Ω ) ] n Q + 1 . (13)
Transform (inverse FFT):
F n , r κ : = 1 n Q + 1 ∑ j = 0 n Q e j n , n Q f j , r κ , n = 0 , … , n Q . (14) F_{n,r}^\kappa:=\frac{1}{n_{\mathrm Q}+1}\sum_{j=0}^{n_{\mathrm Q}}e_{jn,n_{\mathrm Q}}f_{j,r}^\kappa,\qquad n=0,\ldots,n_{\mathrm Q}. \tag*{(14)} F n , r κ := n Q + 1 1 j = 0 ∑ n Q e j n , n Q f j , r κ , n = 0 , … , n Q . (14)
Solve the transformed problems, in parallel:
Y ℓ , r κ : = K ( z ℓ , n Q κ ) F ℓ , r κ , ℓ = 0 , … , n Q . (15) Y_{\ell,r}^\kappa:=K(z_{\ell,n_{\mathrm Q}}^\kappa)F_{\ell,r}^\kappa,\qquad\ell=0,\ldots,n_{\mathrm Q}. \tag*{(15)} Y ℓ , r κ := K ( z ℓ , n Q κ ) F ℓ , r κ , ℓ = 0 , … , n Q . (15)
by Algorithm 2 in Section 3; due to (11) only ⌊ n Q / 2 ⌋ + 1 \lfloor n_{\mathrm Q}/2\rfloor+1 ⌊ n Q /2 ⌋ + 1 of them are needed.
Transform back (FFT) and rescale:
u ~ r κ ( n κ ) : = r − n ∑ ℓ = 0 n Q Y ℓ , r κ e − ℓ n , n Q , n = 1 , … , n Q . (16) \widetilde{u}^{\kappa}_{r}(n\kappa) := r^{-n}\sum_{\ell=0}^{n_Q}Y^{\kappa}_{\ell,r}e^{-\ell n,n_Q}, \qquad n=1,\ldots,n_Q. \tag*{(16)} u r κ ( nκ ) := r − n ℓ = 0 ∑ n Q Y ℓ , r κ e − ℓ n , n Q , n = 1 , … , n Q . (16)
Steps 2 and 4 are FFTs of length n Q + 1 n_Q+1 n Q + 1 ; the cost is dominated by Step 3, which is obviously parallelizable, one of the strong points of the algorithm.
Note that
z n Q + 1 − ℓ , n Q κ = z ℓ , n Q κ ‾ , ℓ = 1 , … , n Q , z^{\kappa}_{n_Q+1-\ell,n_Q}=\overline{z^{\kappa}_{\ell,n_Q}}, \qquad\ell=1,\ldots,n_Q, z n Q + 1 − ℓ , n Q κ = z ℓ , n Q κ , ℓ = 1 , … , n Q ,
since e n Q + 1 − ℓ , n Q = e ℓ , n Q ‾ e_{n_Q+1-\ell,n_Q}=\overline{e_{\ell,n_Q}} e n Q + 1 − ℓ , n Q = e ℓ , n Q and δ \delta δ has real coefficients. This implies that, for real-valued data f f f , as in our case,
Y ℓ , r κ = Y n Q + 1 − ℓ , r κ ‾ , ℓ = 1 , … , n Q , (17) Y_{\ell,r}^{\kappa}=\overline{Y^{\kappa}_{n_Q+1-\ell,r}}, \qquad\ell=1,\ldots,n_Q, \tag*{(17)} Y ℓ , r κ = Y n Q + 1 − ℓ , r κ , ℓ = 1 , … , n Q , (17)
so that the computations in Step 3, (10c), the most expensive part, can be halved.
Two parameters remain to be chosen. A typical choice is n Q = n C Q n_Q=n_{CQ} n Q = n C Q , which we adopt in the experiments of Section 5; we refer to [10 ] for an analysis of the dependence on n Q n_Q n Q and for the advantages of overresolving, n Q > n C Q n_Q>n_{CQ} n Q > n C Q . As for the radius, the aliasing error of (9) is O ( r n Q + 1 ) O(r^{n_Q+1}) O ( r n Q + 1 ) while the scaling by r − n r^{-n} r − n in (10d) amplifies round-off errors by up to r − n Q r^{-n_Q} r − n Q ; balancing the two gives the standard choice r = eps 1 / ( 2 n Q ) r=\operatorname{eps}^{1/(2n_Q)} r = eps 1/ ( 2 n Q ) , with eps the machine precision, for which both errors are O ( eps 1 / 2 ) O(\operatorname{eps}^{1/2}) O ( eps 1/2 ) . Better trade-offs are possible by taking n Q > n C Q n_Q>n_{CQ} n Q > n C Q : with n Q = 2 n C Q n_Q=2n_{CQ} n Q = 2 n C Q and r = eps 1 / ( 3 n C Q + 1 ) r=\operatorname{eps}^{1/(3n_{CQ}+1)} r = eps 1/ ( 3 n C Q + 1 ) the error becomes O ( eps 2 / 3 ) O(\operatorname{eps}^{2/3}) O ( eps 2/3 ) , and so on; see [6 Sect. 3.4] and [10 ] . Taking r = 1 r=1 r = 1 , which would avoid the round-off amplification, is not an option here: for purely imaginary s s s problem (2) becomes the heterogeneous Helmholtz equation, which has no solution in H 1 ( R m ) H^1(\mathbb{R}^m) H 1 ( R m ) , and indeed the bound (35a) degenerates as ℜ s → 0 \Re s\to0 ℜ s → 0 ; the quadrature frequencies must therefore be kept at a positive distance from the imaginary axis.
A FEM–BEM spatial-discretization algorithm
Overlapping coupling algorithm for evaluating K ( s ) \mathrm{K}(s) K ( s )
The method relies on the calculation of (10c). So we are led to consider the (numerical) solution of the problem
U = K ( s ) F , where { U ∈ H 1 ( R m ) , ( s / c ) 2 U − Δ U = F . (18) U=K(s)F,\qquad\text{where}\qquad
\begin{cases}
U\in H^1(\mathbb{R}^m),\\
(s/c)^2U-\Delta U=F.
\end{cases} \tag*{(18)} U = K ( s ) F , where { U ∈ H 1 ( R m ) , ( s / c ) 2 U − Δ U = F . (18)
Here we are just assuming that F ∈ L 2 ( R m ) F\in L^2(\mathbb{R}^m) F ∈ L 2 ( R m ) , but in our application supp F ⊂ Ω ‾ \operatorname{supp}F\subset\overline{\Omega} supp F ⊂ Ω , a fact that we emphasize by writing F ∈ L 2 ( Ω ) F\in L^2(\Omega) F ∈ L 2 ( Ω ) . We also recall the notation σ = ℜ s > 0 \sigma=\Re s>0 σ = ℜ s > 0 . Although the problem is naturally formulated for finding U U U in H 1 ( R m ) H^1(\mathbb{R}^m) H 1 ( R m ) , we will see later that the solution is smoother, and in fact U ∈ H 2 ( R m ) U\in H^2(\mathbb{R}^m) U ∈ H 2 ( R m ) .
The numerical solution of this problem is obtained by means of an overlapping FEM–BEM approach. We first introduce the continuous framework on which the discrete procedures will be applied. Let Ω 1 \Omega_1 Ω 1 be a smooth domain and Ω 2 \Omega_2 Ω 2 a polygonal one, such that Ω ‾ ⊂ Ω 1 \overline{\Omega}\subset\Omega_1 Ω ⊂ Ω 1 and Ω ‾ 1 ⊂ Ω 2 \overline{\Omega}_1\subset\Omega_2 Ω 1 ⊂ Ω 2 . Their boundaries are denoted by Γ \Gamma Γ (for Ω 1 \Omega_1 Ω 1 ) and Σ \Sigma Σ (for Ω 2 \Omega_2 Ω 2 ), respectively (see Figure 1 for a schematic representation of the domains involved).
Figure 1. Domains for the space solution algorithm: Ω \Omega Ω (dark gray), with Ω ‾ \overline{\Omega} Ω being, or containing, the support of c R = 1 − c − 2 c_{\mathrm{R}}=1-c^{-2} c R = 1 − c − 2 , the contrast function; Ω 1 \Omega_1 Ω 1 (white), the interior of the smooth curve Γ \Gamma Γ (so that Ω 1 c \Omega_1^{\mathrm{c}} Ω 1 c is the exterior); and Ω 2 \Omega_2 Ω 2 , the interior of the polygonal curve Σ \Sigma Σ . The overlapping domain Ω 12 = Ω 1 c ∩ Ω 2 \Omega_{12}=\Omega_1^{\mathrm{c}}\cap\Omega_2 Ω 12 = Ω 1 c ∩ Ω 2 (light gray) is also depicted in the figure.
Let γ Σ \gamma_\Sigma γ Σ and γ Γ \gamma_\Gamma γ Γ be trace operators that evaluate functions on the artificial boundaries Σ \Sigma Σ and Γ \Gamma Γ , respectively. Let W F t r W_F^{\mathrm{tr}} W F tr be the solution of
( s / c ) 2 W F t r − Δ W F t r = F , in Ω 2 , γ Σ W F t r = 0. (19) \begin{aligned}
(s/c)^2W_F^{\mathrm{tr}}-\Delta W_F^{\mathrm{tr}}&=F,\quad\text{in }\Omega_2,\\
\gamma_\Sigma W_F^{\mathrm{tr}}&=0. \tag*{(19)}
\end{aligned} ( s / c ) 2 W F tr − Δ W F tr γ Σ W F tr = F , in Ω 2 , = 0. (19)
We introduce two unknowns
W t r : = U ∣ Ω 2 − W F t r ∈ H 1 ( Ω 2 ) , W c : = U ∣ Ω 1 c ∈ H 1 ( Ω 1 c ) . W^{\mathrm{tr}}:=U|_{\Omega_2}-W_F^{\mathrm{tr}}\in H^1(\Omega_2),\qquad W^{\mathrm{c}}:=U|_{\Omega_1^{\mathrm{c}}}\in H^1(\Omega_1^{\mathrm{c}}). W tr := U ∣ Ω 2 − W F tr ∈ H 1 ( Ω 2 ) , W c := U ∣ Ω 1 c ∈ H 1 ( Ω 1 c ) .
W c W^{\mathrm{c}} W c is analytic in Ω ‾ 1 c \overline{\Omega}_1^{\mathrm{c}} Ω 1 c , since U U U solves the Helmholtz equation s 2 U − Δ U = 0 s^2U-\Delta U=0 s 2 U − Δ U = 0 in the open set Ω c ⊃ Ω ‾ 1 c \Omega^{\mathrm{c}}\supset\overline{\Omega}_1^{\mathrm{c}} Ω c ⊃ Ω 1 c , and W t r ∈ H 2 ( Ω 2 ) W^{\mathrm{tr}}\in H^2(\Omega_2) W tr ∈ H 2 ( Ω 2 ) whenever Ω 2 \Omega_2 Ω 2 is convex, see Remark 4.3.
We take
g Σ = γ Σ U = γ Σ W t r , g Γ = γ Γ U = γ Γ W c g_\Sigma=\gamma_\Sigma U=\gamma_\Sigma W^{\mathrm{tr}},\qquad g_\Gamma=\gamma_\Gamma U=\gamma_\Gamma W^{\mathrm{c}} g Σ = γ Σ U = γ Σ W tr , g Γ = γ Γ U = γ Γ W c
as new unknowns (for the first identity, note that γ Σ W F t r = 0 \gamma_\Sigma W_F^{\mathrm{tr}}=0 γ Σ W F tr = 0 ). Then
Since, with Ω 1 c = R m ∖ Ω ‾ 1 \Omega_1^{\mathrm{c}}=\mathbb{R}^m\setminus\overline{\Omega}_1 Ω 1 c = R m ∖ Ω 1 , c ∣ Ω 1 c ≡ 1 c|_{\Omega_1^{\mathrm{c}}}\equiv1 c ∣ Ω 1 c ≡ 1 , W c W^{\mathrm{c}} W c solves
W c ∈ H 1 ( Ω 1 c ) , γ Γ W c = g Γ , s 2 W c − Δ W c = 0 , in Ω 1 c . (20) \begin{aligned}
W^{\mathrm{c}}&\in H^1(\Omega_1^{\mathrm{c}}),\quad\gamma_\Gamma W^{\mathrm{c}}=g_\Gamma,\\
s^2W^{\mathrm{c}}-\Delta W^{\mathrm{c}}&=0,\quad\text{in }\Omega_1^{\mathrm{c}}. \tag*{(20)}
\end{aligned} W c s 2 W c − Δ W c ∈ H 1 ( Ω 1 c ) , γ Γ W c = g Γ , = 0 , in Ω 1 c . (20)
In other words, W c W^{\mathrm{c}} W c can be computed using BEM schemes provided the Dirichlet datum g Γ = γ Γ W c = γ Γ U g_\Gamma=\gamma_\Gamma W^{\mathrm{c}}=\gamma_\Gamma U g Γ = γ Γ W c = γ Γ U is known, since problem (20) is uniquely solvable (this holds even for σ = ℜ s = 0 \sigma=\Re s=0 σ = ℜ s = 0 , which is never our case, provided the condition W c ∈ H 1 ( Ω 1 c ) W^{\mathrm{c}}\in H^1(\Omega_1^{\mathrm{c}}) W c ∈ H 1 ( Ω 1 c ) is replaced by a radiation condition at ∞ \infty ∞ ).
For W t r W^{\mathrm{tr}} W tr , we instead have
W t r ∈ H 1 ( Ω 2 ) , γ Σ W t r = g Σ , ( s / c ) 2 W t r − Δ W t r = 0 , in Ω 2 , (21) \begin{aligned}
W^{\mathrm{tr}}&\in H^1(\Omega_2),\quad\gamma_\Sigma W^{\mathrm{tr}}=g_\Sigma,\\
(s/c)^2W^{\mathrm{tr}}-\Delta W^{\mathrm{tr}}&=0,\quad\text{in }\Omega_2, \tag*{(21)}
\end{aligned} W tr ( s / c ) 2 W tr − Δ W tr ∈ H 1 ( Ω 2 ) , γ Σ W tr = g Σ , = 0 , in Ω 2 , (21)
which is well posed since σ = ℜ s > 0 \sigma=\Re s>0 σ = ℜ s > 0 . This problem, as well as problem (19) , is better suited to approximation by FEM schemes, because Ω 2 \Omega_2 Ω 2 is bounded and may contain inhomogeneities arising from the wave-speed field c c c .
Hence we can introduce
K Ω 1 c Γ ( s ) : H 1 / 2 ( Γ ) → H 1 ( Ω 1 c ) , K Ω 2 Σ t r ( s ) : H 1 / 2 ( Σ ) → H 1 ( Ω 2 ) , K_{\Omega^c_1\Gamma}(s): H^{1/2}(\Gamma) \to H^1(\Omega^c_1), \qquad K^{\mathrm{tr}}_{\Omega_2\Sigma}(s): H^{1/2}(\Sigma) \to H^1(\Omega_2), K Ω 1 c Γ ( s ) : H 1/2 ( Γ ) → H 1 ( Ω 1 c ) , K Ω 2 Σ tr ( s ) : H 1/2 ( Σ ) → H 1 ( Ω 2 ) ,
as the solution operators for problems (14) and (15), respectively. That is,
K Ω 1 c Γ ( s ) g Γ : = W c , K Ω 2 Σ t r ( s ) g Σ : = W t r K_{\Omega^c_1\Gamma}(s)g_\Gamma:= W^c, \qquad K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_\Sigma:= W^{\mathrm{tr}} K Ω 1 c Γ ( s ) g Γ := W c , K Ω 2 Σ tr ( s ) g Σ := W tr
so that our unknown U U U can be expressed as
U ∣ Ω 2 = K Ω 2 Σ t r ( s ) g Σ + W F t r , U ∣ Ω 1 c = K Ω 1 c Γ ( s ) g Γ , U|_{\Omega_2} = K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_\Sigma+ W^{\mathrm{tr}}_F, \qquad U|_{\Omega^c_1} = K_{\Omega^c_1\Gamma}(s)g_\Gamma, U ∣ Ω 2 = K Ω 2 Σ tr ( s ) g Σ + W F tr , U ∣ Ω 1 c = K Ω 1 c Γ ( s ) g Γ ,
provided that the Dirichlet data g Σ g_\Sigma g Σ and g Γ g_\Gamma g Γ used to feed the operators K Ω 2 Σ t r ( s ) K^{\mathrm{tr}}_{\Omega_2\Sigma}(s) K Ω 2 Σ tr ( s ) and K Ω 1 c Γ ( s ) K_{\Omega^c_1\Gamma}(s) K Ω 1 c Γ ( s ) are known.
The next step is to determine these unknowns ( g Σ , g Γ ) (g_\Sigma,g_\Gamma) ( g Σ , g Γ ) . We notice that it suffices to impose that both representations agree on the overlap,
( W t r + W F t r ) ∣ Ω 12 = W c ∣ Ω 12 , Ω 12 : = Ω 1 c ∩ Ω 2 , \left.(W^{\mathrm{tr}} + W^{\mathrm{tr}}_F)\right|_{\Omega_{12}} = \left.W^c\right|_{\Omega_{12}}, \qquad\Omega_{12} := \Omega^c_1 \cap\Omega_2, ( W tr + W F tr ) Ω 12 = W c ∣ Ω 12 , Ω 12 := Ω 1 c ∩ Ω 2 ,
since then the two fields glue together to an element of H 1 ( R m ) H^1(\mathbb{R}^m) H 1 ( R m ) which solves (12) (recall that supp F ⊂ Ω ‾ \operatorname{supp} F \subset\overline{\Omega} supp F ⊂ Ω ), and hence coincides with U U U . In the notation of the operators introduced above, this amounts to requiring
V : = ( K Ω 2 Σ t r ( s ) g Σ + W F t r − K Ω 1 c Γ ( s ) g Γ ) ∣ Ω 1 c ∩ Ω 2 = 0. V := \left.(K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_\Sigma+ W^{\mathrm{tr}}_F - K_{\Omega^c_1\Gamma}(s)g_\Gamma)\right|_{\Omega^c_1\cap\Omega_2} = 0. V := ( K Ω 2 Σ tr ( s ) g Σ + W F tr − K Ω 1 c Γ ( s ) g Γ ) Ω 1 c ∩ Ω 2 = 0.
Now, V V V solves the homogeneous equation
− Δ V + s 2 V = 0 , in Ω 12 (22) -\Delta V + s^2 V = 0, \qquad\text{in } \Omega_{12} \tag*{(22)} − Δ V + s 2 V = 0 , in Ω 12 (22)
(the support of F F F does not meet Ω 12 \Omega_{12} Ω 12 , and c ≡ 1 c \equiv1 c ≡ 1 there) and, since γ Σ K Ω 2 Σ t r ( s ) \gamma_\Sigma K^{\mathrm{tr}}_{\Omega_2\Sigma}(s) γ Σ K Ω 2 Σ tr ( s ) and γ Γ K Ω 1 c Γ ( s ) \gamma_\Gamma K_{\Omega^c_1\Gamma}(s) γ Γ K Ω 1 c Γ ( s ) are the identity on H 1 / 2 ( Σ ) H^{1/2}(\Sigma) H 1/2 ( Σ ) and H 1 / 2 ( Γ ) H^{1/2}(\Gamma) H 1/2 ( Γ ) respectively, and γ Σ W F t r = 0 \gamma_\Sigma W^{\mathrm{tr}}_F = 0 γ Σ W F tr = 0 , its traces are
γ Σ V = g Σ − γ Σ K Ω 1 c Γ ( s ) g Γ , γ Γ V = γ Γ K Ω 2 Σ t r ( s ) g Σ + γ Γ W F t r − g Γ . \gamma_\Sigma V = g_\Sigma- \gamma_\Sigma K_{\Omega^c_1\Gamma}(s)g_\Gamma, \qquad\gamma_\Gamma V = \gamma_\Gamma K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_\Sigma+ \gamma_\Gamma W^{\mathrm{tr}}_F - g_\Gamma. γ Σ V = g Σ − γ Σ K Ω 1 c Γ ( s ) g Γ , γ Γ V = γ Γ K Ω 2 Σ tr ( s ) g Σ + γ Γ W F tr − g Γ .
Since the Dirichlet problem for (16) in Ω 12 \Omega_{12} Ω 12 is uniquely solvable for σ = ℜ s > 0 \sigma= \Re s > 0 σ = ℜ s > 0 , we have V = 0 V = 0 V = 0 if and only if both traces vanish, that is, if and only if ( g Σ , g Γ ) (g_\Sigma,g_\Gamma) ( g Σ , g Γ ) solves
g Σ − γ Σ K Ω 1 c Γ ( s ) g Γ = 0 , − γ Γ K Ω 2 Σ t r ( s ) g Σ + g Γ = γ Γ W F t r . (23) \begin{aligned}
g_\Sigma- \gamma_\Sigma K_{\Omega^c_1\Gamma}(s)g_\Gamma&= 0,\\
-\gamma_\Gamma K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_\Sigma+ g_\Gamma&= \gamma_\Gamma W^{\mathrm{tr}}_F. \tag*{(23)}
\end{aligned} g Σ − γ Σ K Ω 1 c Γ ( s ) g Γ − γ Γ K Ω 2 Σ tr ( s ) g Σ + g Γ = 0 , = γ Γ W F tr . (23)
Hence ( g Σ , g Γ ) (g_\Sigma,g_\Gamma) ( g Σ , g Γ ) is the only solution of (17): any solution of (17) reconstructs, through the representation above, the unique solution U U U of (12), and therefore g Σ = γ Σ U g_\Sigma= \gamma_\Sigma U g Σ = γ Σ U and g Γ = γ Γ U g_\Gamma= \gamma_\Gamma U g Γ = γ Γ U .
The solution of (17) is obtained by a Schur complement approach:
g Γ − γ Γ K Ω 2 Σ t r ( s ) γ Σ K Ω 1 c Γ ( s ) g Γ = γ Γ W F t r , g Σ = γ Σ K Ω 1 c Γ ( s ) g Γ . g_\Gamma- \gamma_\Gamma K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)\gamma_\Sigma K_{\Omega^c_1\Gamma}(s)g_\Gamma= \gamma_\Gamma W^{\mathrm{tr}}_F, \qquad g_\Sigma= \gamma_\Sigma K_{\Omega^c_1\Gamma}(s)g_\Gamma. g Γ − γ Γ K Ω 2 Σ tr ( s ) γ Σ K Ω 1 c Γ ( s ) g Γ = γ Γ W F tr , g Σ = γ Σ K Ω 1 c Γ ( s ) g Γ .
Hence the resulting spatial FEM–BEM algorithm is as follows:
Algorithm 2 (Overlapping FEM–BEM evaluation of K ( s ) F K(s)F K ( s ) F )
Input: s ∈ C + s \in\mathbb{C}_+ s ∈ C + , F ∈ L 2 ( Ω ) F \in L^2(\Omega) F ∈ L 2 ( Ω ) ; domains Ω ⊂ Ω 1 ⊂ Ω 2 \Omega\subset\Omega_1 \subset\Omega_2 Ω ⊂ Ω 1 ⊂ Ω 2 with boundaries Γ \Gamma Γ and Σ \Sigma Σ .
Output: U = K ( s ) F U = K(s)F U = K ( s ) F on Ω 2 \Omega_2 Ω 2 and on Ω 1 c \Omega^c_1 Ω 1 c .
Interior solve (FEM): find W F t r W^{\mathrm{tr}}_{F} W F tr with
( s / c ) 2 W F t r − Δ W F t r = F , in Ω 2 , γ Σ W F t r = 0 , (24) \begin{aligned}
(s/c)^2 W^{\mathrm{tr}}_{F}-\Delta W^{\mathrm{tr}}_{F} &= F,\quad\text{in }\Omega_2,\\
\gamma_{\Sigma}W^{\mathrm{tr}}_{F} &= 0, \tag*{(24)}
\end{aligned} ( s / c ) 2 W F tr − Δ W F tr γ Σ W F tr = F , in Ω 2 , = 0 , (24)
and set
f Γ : = γ Γ W F t r . (25) f_{\Gamma}:=\gamma_{\Gamma}W^{\mathrm{tr}}_{F}. \tag*{(25)} f Γ := γ Γ W F tr . (25)
Coupling (GMRES on Γ \Gamma Γ ): solve the second kind equation
g Γ − γ Γ K Ω 2 Σ t r ( s ) γ Σ K Ω 1 Γ c ( s ) g Γ = f Γ , (26) g_{\Gamma}-\gamma_{\Gamma}K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)\gamma_{\Sigma}K_{\Omega_1\Gamma}^{\mathrm{c}}(s)g_{\Gamma}=f_{\Gamma}, \tag*{(26)} g Γ − γ Γ K Ω 2 Σ tr ( s ) γ Σ K Ω 1 Γ c ( s ) g Γ = f Γ , (26)
each application of the operator requiring one exterior solve (BEM) and one interior solve (FEM), and set g Σ : = γ Σ K Ω 1 Γ c ( s ) g Γ g_{\Sigma}:=\gamma_{\Sigma}K_{\Omega_1\Gamma}^{\mathrm{c}}(s)g_{\Gamma} g Σ := γ Σ K Ω 1 Γ c ( s ) g Γ .
Assemble: with W t r : = K Ω 2 Σ t r ( s ) g Σ W^{\mathrm{tr}}:=K^{\mathrm{tr}}_{\Omega_2\Sigma}(s)g_{\Sigma} W tr := K Ω 2 Σ tr ( s ) g Σ and W c : = K Ω 1 Γ c ( s ) g Γ W^{\mathrm{c}}:=K_{\Omega_1\Gamma}^{\mathrm{c}}(s)g_{\Gamma} W c := K Ω 1 Γ c ( s ) g Γ ,
K ( s ) F ∣ Ω 2 = W t r + W F t r , K ( s ) F ∣ Ω 1 c = W c . (27) K(s)F\big|_{\Omega_2}=W^{\mathrm{tr}}+W^{\mathrm{tr}}_{F},\qquad K(s)F\big|_{\Omega_1}^{\mathrm{c}}=W^{\mathrm{c}}. \tag*{(27)} K ( s ) F Ω 2 = W tr + W F tr , K ( s ) F Ω 1 c = W c . (27)
Note that Algorithm 2 provides a double representation of K ( s ) F K(s)F K ( s ) F on the overlap Ω 12 \Omega_{12} Ω 12 , by W t r + W F t r W^{\mathrm{tr}}+W^{\mathrm{tr}}_{F} W tr + W F tr and by W c W^{\mathrm{c}} W c , which coincide by construction. The coupled equation (26) is analyzed in the next subsection.
Remark 3.1 (Marching form) The equivalence of Remark 2.1 extends to the coupled scheme: applying the multistep method directly to (24) –(27) yields a time-stepping algorithm in which the differential part becomes a Newmark-type recursion on Ω 2 \Omega_2 Ω 2 , with one real symmetric positive definite solve per step, while the boundary element part, being nonlocal in time, becomes a convolution quadrature whose history has to be stored and advanced, and the two are linked at each step through the coupled equation (26) . This is the structure of [28 ] and, for the Schrödinger equation with Runge–Kutta methods, of [38 ] , with the overlapping coupling in place of the non-overlapping one. The history can be handled by the recursive, FFT-based algorithms of [2 , 27 ] .
We favor the all-steps-at-once variant for two reasons. In the scattering problems considered here the right-hand side is known in advance on the whole interval [ 0 , T ] [0,T] [ 0 , T ] , so that early availability of the solution brings no benefit; and the decoupled formulation yields independent, uniformly damped frequency problems, most of which do not even require the coupling (§ 5 ), and which can be solved fully in parallel. The price is that no approximation of the solution u u u of the original time-domain problem (1) is available until all the transformed problems have been solved, and that the corresponding volume solutions must be stored.
Analysis of the coupled equation
The algorithm of the previous subsection rests on two facts: that the boundary element correction U − W F t r U-W^{\mathrm{tr}}_{F} U − W F tr is small for strongly damped frequencies, so that the coupling step can be skipped and, when it cannot, the second kind equation (26) is well conditioned. The two theorems below make these facts quantitative, with constants explicit in the Laplace parameter s s s , the wave speed and the distances between Ω \Omega Ω , Γ \Gamma Γ and Σ \Sigma Σ ; the proofs are given in Section 4.4. Before stating them let us introduce, for any pair of sets A , B ⊂ R m A,B\subset\mathbb{R}^{m} A , B ⊂ R m ,
d ( A , B ) = inf { ∣ x − y ∣ : x ∈ A , y ∈ B } , d(A,B)=\inf\{|\mathbf{x}-\mathbf{y}|:\mathbf{x}\in A,\ \mathbf{y}\in B\}, d ( A , B ) = inf { ∣ x − y ∣ : x ∈ A , y ∈ B } ,
the distance between A A A and B B B .
Theorem 3.2. Let σ 0 > 0 \sigma_0 > 0 σ 0 > 0 and let ε > 0 \varepsilon> 0 ε > 0 be sufficiently small. Set
d : = ( 1 − ε ) d ( Ω , Σ ) d := (1-\varepsilon)d(\Omega,\Sigma) d := ( 1 − ε ) d ( Ω , Σ ) , c max : = ∥ c ∥ L ∞ ( R m ) ≥ 1 c_{\max} := \lVert c\rVert_{L^\infty(\mathbb{R}^m)} \ge1 c m a x := ∥ c ∥ L ∞ ( R m ) ≥ 1 , c 0 , R : = ∥ c R ∥ L ∞ ( Ω 2 ) = ∥ c − 2 − 1 ∥ L ∞ ( Ω 2 ) c_{0,R} := \lVert c_R\rVert_{L^\infty(\Omega_2)} = \lVert c^{-2}-1\rVert_{L^\infty(\Omega_2)} c 0 , R := ∥ c R ∥ L ∞ ( Ω 2 ) = ∥ c − 2 − 1 ∥ L ∞ ( Ω 2 ) .
Then there exists C > 0 C > 0 C > 0 , depending only on σ 0 \sigma_0 σ 0 , ε \varepsilon ε , Ω \Omega Ω , Γ \Gamma Γ and Σ \Sigma Σ (or Ω 1 , Ω 2 \Omega_1,\Omega_2 Ω 1 , Ω 2 ), such that for all s s s with σ = ℜ s ≥ σ 0 \sigma= \Re s \ge\sigma_0 σ = ℜ s ≥ σ 0 , all positive wave-speed fields c c c with c ∣ R m ∖ Ω ≡ 1 c|_{\mathbb{R}^m\setminus\Omega} \equiv1 c ∣ R m ∖ Ω ≡ 1 and all F ∈ L 2 ( Ω ) F \in L^2(\Omega) F ∈ L 2 ( Ω ) the following holds:
∣ s ∣ ∥ W F t r − U ∥ L 2 ( Ω 2 ) + ∥ ∇ ( W F t r − U ) ∥ L 2 ( Ω 2 ) ≤ C c max 2 e − σ d σ − 2 ∣ s ∣ m / 2 + 1 ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ F ∥ L 2 ( Ω 2 ) . (28) |s|\lVert W_F^{\mathrm{tr}}-U\rVert_{L^2(\Omega_2)}+\lVert\nabla(W_F^{\mathrm{tr}}-U)\rVert_{L^2(\Omega_2)}
\le Cc_{\max}^2e^{-\sigma d}\sigma^{-2}|s|^{m/2+1}(1+c_{\max}^2c_{0,R}\sigma^{-1}|s|)\lVert F\rVert_{L^2(\Omega_2)}. \tag*{(28)} ∣ s ∣ ∥ W F tr − U ∥ L 2 ( Ω 2 ) + ∥ ∇ ( W F tr − U ) ∥ L 2 ( Ω 2 ) ≤ C c m a x 2 e − σ d σ − 2 ∣ s ∣ m /2 + 1 ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ F ∥ L 2 ( Ω 2 ) . (28) Besides, if d 1 = ( 1 − ε ) d ( Γ , Σ ) d_1=(1-\varepsilon)d(\Gamma,\Sigma) d 1 = ( 1 − ε ) d ( Γ , Σ ) , d 2 = ( 1 − ε ) d ( Γ , Ω ) d_2=(1-\varepsilon)d(\Gamma,\Omega) d 2 = ( 1 − ε ) d ( Γ , Ω ) and f Γ = γ Γ W F t r f_\Gamma=\gamma_\Gamma W_F^{\mathrm{tr}} f Γ = γ Γ W F tr is as in (19),
∥ f Γ ∥ H 1 / 2 ( Γ ) ≤ C c max 2 e − σ min { d 1 , d 2 } ∣ s ∣ ( m − 1 ) / 2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) . (29) \lVert f_\Gamma\rVert_{H^{1/2}(\Gamma)} \le Cc_{\max}^2e^{-\sigma\min\{d_1,d_2\}}|s|^{(m-1)/2}\sigma^{-1}\lVert F\rVert_{L^2(\Omega_2)}. \tag*{(29)} ∥ f Γ ∥ H 1/2 ( Γ ) ≤ C c m a x 2 e − σ m i n { d 1 , d 2 } ∣ s ∣ ( m − 1 ) /2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) . (29)
Theorem 3.3. Under the assumptions of Theorem 3.2, there exists C > 0 C > 0 C > 0 such that
∥ γ Γ K Ω 2 Σ t r ( s ) ∥ H 1 / 2 ( Σ ) → H 1 / 2 ( Γ ) ≤ C e − σ min { d 1 , d 2 } ∣ s ∣ m / 2 + 1 σ − 1 ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) , (30) \lVert\gamma_\Gamma K_{\Omega_2\Sigma}^{\mathrm{tr}}(s)\rVert_{H^{1/2}(\Sigma)\to H^{1/2}(\Gamma)}
\le Ce^{-\sigma\min\{d_1,d_2\}}|s|^{m/2+1}\sigma^{-1}(1+c_{\max}^2c_{0,R}\sigma^{-1}|s|), \tag*{(30)} ∥ γ Γ K Ω 2 Σ tr ( s ) ∥ H 1/2 ( Σ ) → H 1/2 ( Γ ) ≤ C e − σ m i n { d 1 , d 2 } ∣ s ∣ m /2 + 1 σ − 1 ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) , (30) ∥ γ Σ K Ω 1 Γ c ( s ) ∥ H 1 / 2 ( Γ ) → H 1 / 2 ( Σ ) ≤ C e − σ d 1 ∣ s ∣ m / 2 + 1 σ − 1 . (31) \lVert\gamma_\Sigma K_{\Omega_1\Gamma}^{\mathrm{c}}(s)\rVert_{H^{1/2}(\Gamma)\to H^{1/2}(\Sigma)}
\le Ce^{-\sigma d_1}|s|^{m/2+1}\sigma^{-1}. \tag*{(31)} ∥ γ Σ K Ω 1 Γ c ( s ) ∥ H 1/2 ( Γ ) → H 1/2 ( Σ ) ≤ C e − σ d 1 ∣ s ∣ m /2 + 1 σ − 1 . (31)
eigenvalues decay very fast, which is the setting in which GMRES converges superlinearly. All of this is stated at the continuous level; after discretization it holds up to mesh-dependent norm equivalences.
The polynomial factors in (32) prevent ρ ( s ) < 1 \rho(s) < 1 ρ ( s ) < 1 in the weakly damped window ℜ s ≈ σ 0 \Re s \approx\sigma_0 ℜ s ≈ σ 0 , ∣ s ∣ ≫ σ |s| \gg\sigma ∣ s ∣ ≫ σ , which is where the coupling is mostly activated in the experiments of § 8 . In those experiments GMRES nevertheless converges in very few iterations—typically one or two, never more than four—uniformly in the frequency, which suggests that these factors are an artifact of the proof and that the exponential factor alone governs the convergence.
Operator estimates and proofs of main results
The aim of this section is to derive estimates for the solution operator K ( s ) K(s) K ( s ) and for the operators of the coupled formulation, explicit in the Laplace parameter s s s . The former are the ingredient of the convergence analysis of Theorem 2.2; the latter, together with the decay estimates of Propositions 4.4 and 4.5, yield the proofs of Theorems 3.2 and 3.3 and quantify when the BEM component is truly necessary in our algorithm.
Preliminaries
In this subsection only, let Ω \Omega Ω be an arbitrary bounded Lipschitz domain with boundary Σ \Sigma Σ , and consider the boundary value problem:
W ∈ H 1 ( R m ∖ Σ ) = H 1 ( Ω ∪ Ω c ) , s 2 W − Δ W = 0 , in R m ∖ Σ , γ Σ W = g Σ . (33) \begin{aligned}
W &\in H^1(\mathbb{R}^m \setminus\Sigma)=H^1(\Omega\cup\Omega^c),\\
s^2W-\Delta W &= 0,\qquad\text{in }\mathbb{R}^m\setminus\Sigma,\\
\gamma_\Sigma W &= g_\Sigma. \tag*{(33)}
\end{aligned} W s 2 W − Δ W γ Σ W ∈ H 1 ( R m ∖ Σ ) = H 1 ( Ω ∪ Ω c ) , = 0 , in R m ∖ Σ , = g Σ . (33)
Then
∥ ∇ W ∥ L 2 ( R m ∖ Σ ) + ∣ s ∣ ∥ W ∥ L 2 ( R m ) ≤ C Σ ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ σ − 1 ∥ g Σ ∥ H 1 / 2 ( Σ ) (34) \|\nabla W\|_{L^2(\mathbb{R}^m\setminus\Sigma)}+|s|\|W\|_{L^2(\mathbb{R}^m)}
\le C_\Sigma(1+|s|)^{1/2}|s|^{\sigma-1}\|g_\Sigma\|_{H^{1/2}(\Sigma)} \tag*{(34)} ∥∇ W ∥ L 2 ( R m ∖ Σ ) + ∣ s ∣∥ W ∥ L 2 ( R m ) ≤ C Σ ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ σ − 1 ∥ g Σ ∥ H 1/2 ( Σ ) (34)
see, for example, [40 ] or [6 Lemma 4.13 and Theorem 4.15] . The proof relies on a closely related estimate for s > 0 s>0 s > 0 originally proven in [1 ] . We will use estimate (34) later in this section.
The goal of this section is twofold. First, we aim to derive analogs of estimates (34) for the operator K ( s ) : L 2 ( R m ) → H r ( R m ) K(s):L^2(\mathbb{R}^m)\to H^r(\mathbb{R}^m) K ( s ) : L 2 ( R m ) → H r ( R m ) with s ∈ C + s\in\mathbb{C}_+ s ∈ C + , defined by
U = K ( s ) F , U ∈ H 1 ( R m ) , s 2 c 2 U − Δ U = F . (35) \begin{aligned}
U &= K(s)F,\\
U &\in H^1(\mathbb{R}^m),\\
\frac{s^2}{c^2}U-\Delta U &= F. \tag*{(35)}
\end{aligned} U U c 2 s 2 U − Δ U = K ( s ) F , ∈ H 1 ( R m ) , = F . (35)
Here, c ∈ L ∞ ( R m ) c\in L^\infty(\mathbb{R}^m) c ∈ L ∞ ( R m ) is a positive function such that c ≡ 1 c\equiv1 c ≡ 1 in Ω c \Omega^c Ω c . For any open domain D ⊂ R m D\subset\mathbb{R}^m D ⊂ R m , we denote the inner product in L 2 ( D ) L^2(D) L 2 ( D ) by
( U , V ) D : = ∫ D U V ‾ (36) (U,V)_D:=\int_D U\overline{V} \tag*{(36)} ( U , V ) D := ∫ D U V (36)
and the associated sesquilinear form by
a s , c , D ( U , V ) : = s 2 ( c − 2 U , V ) D + ( ∇ U , ∇ V ) D , (37) a_{s,c,D}(U,V):=s^2(c^{-2}U,V)_D+(\nabla U,\nabla V)_D, \tag*{(37)} a s , c , D ( U , V ) := s 2 ( c − 2 U , V ) D + ( ∇ U , ∇ V ) D , (37)
with the convention that for vector-valued functions, such as ∇ U \nabla U ∇ U , the inner product is understood componentwise.
With this notation, problem (28) is equivalent to the weak formulation
U ∈ H 1 ( R m ) , a s , c , R m ( U , V ) = ( F , V ) R m , ∀ V ∈ H 1 ( R m ) . (38) \begin{aligned}
U &\in H^1(\mathbb{R}^m),\\
a_{s,c,\mathbb{R}^m}(U,V) &= (F,V)_{\mathbb{R}^m}, \qquad\forall V\in H^1(\mathbb{R}^m). \tag*{(38)}
\end{aligned} U a s , c , R m ( U , V ) ∈ H 1 ( R m ) , = ( F , V ) R m , ∀ V ∈ H 1 ( R m ) . (38)
Since Δ U = s 2 c − 2 U − F ∈ L 2 ( R m ) \Delta U=s^2c^{-2}U-F\in L^2(\mathbb{R}^m) Δ U = s 2 c − 2 U − F ∈ L 2 ( R m ) and
∥ W ∥ H 2 ( R m ) 2 ≤ 2 ( ∥ W ∥ L 2 ( R m ) 2 + ∥ Δ W ∥ L 2 ( R m ) 2 ) , (39) \|W\|_{H^2(\mathbb{R}^m)}^2\le2\left(\|W\|_{L^2(\mathbb{R}^m)}^2+\|\Delta W\|_{L^2(\mathbb{R}^m)}^2\right), \tag*{(39)} ∥ W ∥ H 2 ( R m ) 2 ≤ 2 ( ∥ W ∥ L 2 ( R m ) 2 + ∥Δ W ∥ L 2 ( R m ) 2 ) , (39)
which follows from the Fourier characterization of the Sobolev norms on R m \mathbb{R}^m R m [36 Ch. 3] and the elementary inequality ( 1 + t ) 2 ≤ 2 ( 1 + t 2 ) (1+t)^2\le2(1+t^2) ( 1 + t ) 2 ≤ 2 ( 1 + t 2 ) , we conclude that in fact U ∈ H 2 ( R m ) U\in H^2(\mathbb{R}^m) U ∈ H 2 ( R m ) . For smoother F F F , U U U gains regularity.
In the next result, we establish essential continuity constants for the operator K ( s ) \mathrm{K}(s) K ( s ) in different norms, merely assuming that F ∈ L 2 ( R m ) F\in L^2(\mathbb{R}^m) F ∈ L 2 ( R m ) .
Recall, before proceeding, that we have followed the convention
σ = ℜ s > 0 , \sigma=\Re s>0, σ = ℜ s > 0 ,
which yields the elementary bound
1 + ∣ s ∣ α ≤ 2 σ ‾ − α ∣ s ∣ α , α ≥ 0 , (40) 1+|s|^\alpha\le2\underline{\sigma}^{-\alpha}|s|^\alpha,\qquad\alpha\ge0, \tag*{(40)} 1 + ∣ s ∣ α ≤ 2 σ − α ∣ s ∣ α , α ≥ 0 , (40)
with
σ ‾ : = min { 1 , σ } , (41) \underline{\sigma}:=\min\{1,\sigma\}, \tag*{(41)} σ := min { 1 , σ } , (41)
a bound that will be used repeatedly in the sequel.
Estimates for the solution operator K ( s ) \mathrm{K}(s) K ( s ) and for related operators on truncated domains
Proposition 4.1 The following bounds hold:
∥ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ c max 2 σ ‾ − 1 ∣ s ∣ − 1 (42) \|\mathrm{K}(s)\|_{L^2(\mathbb{R}^m)\to L^2(\mathbb{R}^m)}
\le c_{\max}^2\underline{\sigma}^{-1}|s|^{-1} \tag*{(42)} ∥ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ c m a x 2 σ − 1 ∣ s ∣ − 1 (42) ∥ ∇ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ 1 2 c max σ ‾ − 1 (43) \|\nabla\mathrm{K}(s)\|_{L^2(\mathbb{R}^m)\to L^2(\mathbb{R}^m)}
\le\frac{1}{2}c_{\max}\underline{\sigma}^{-1} \tag*{(43)} ∥∇ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ 2 1 c m a x σ − 1 (43) ∥ Δ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ 1 + ( c max c min ) 2 σ ‾ − 1 ∣ s ∣ ≤ 2 ( c max c min ) 2 σ ‾ − 1 ∣ s ∣ (44) \|\Delta\mathrm{K}(s)\|_{L^2(\mathbb{R}^m)\to L^2(\mathbb{R}^m)}
\le1+\left(\frac{c_{\max}}{c_{\min}}\right)^2\underline{\sigma}^{-1}|s|
\le2\left(\frac{c_{\max}}{c_{\min}}\right)^2\underline{\sigma}^{-1}|s| \tag*{(44)} ∥Δ K ( s ) ∥ L 2 ( R m ) → L 2 ( R m ) ≤ 1 + ( c m i n c m a x ) 2 σ − 1 ∣ s ∣ ≤ 2 ( c m i n c m a x ) 2 σ − 1 ∣ s ∣ (44) ∥ K ( s ) ∥ L 2 ( R m ) → H 1 ( R m ) ≤ c max 2 ( 1 + ∣ s 2 c max ∣ 2 ) 1 / 2 σ ‾ − 1 ∣ s ∣ − 1 (45) \|\mathrm{K}(s)\|_{L^2(\mathbb{R}^m)\to H^1(\mathbb{R}^m)}
\le c_{\max}^2\left(1+\left|\frac{s}{2c_{\max}}\right|^2\right)^{1/2}\underline{\sigma}^{-1}|s|^{-1} \tag*{(45)} ∥ K ( s ) ∥ L 2 ( R m ) → H 1 ( R m ) ≤ c m a x 2 ( 1 + 2 c m a x s 2 ) 1/2 σ − 1 ∣ s ∣ − 1 (45) ≤ 2 c max 2 σ ‾ − 1 σ ‾ − 1 (46) \le\sqrt{2}c_{\max}^2\underline{\sigma}^{-1}\underline{\sigma}^{-1} \tag*{(46)} ≤ 2 c m a x 2 σ − 1 σ − 1 (46) with
c max = ∥ c ∥ L ∞ ( R m ) ≥ 1 , c min = ess inf c ≤ 1. c_{\max}=\|c\|_{L^\infty(\mathbb{R}^m)}\ge1,\qquad c_{\min}=\operatorname{ess\,inf}c\le1. c m a x = ∥ c ∥ L ∞ ( R m ) ≥ 1 , c m i n = ess inf c ≤ 1. Furthermore, for r ∈ [ 1 , 2 ] r\in[1,2] r ∈ [ 1 , 2 ] ,
∥ K ( s ) ∥ L 2 ( R m ) → H r ( R m ) ≤ c max 2 [ 2 2 c min − 2 ] r − 1 ( 1 + ∣ s ∣ 2 ) r / 2 σ ‾ − 1 ∣ s ∣ − 1 (47) \|\mathrm{K}(s)\|_{L^2(\mathbb{R}^m)\to H^r(\mathbb{R}^m)}
\le c_{\max}^2\left[2\sqrt{2}c_{\min}^{-2}\right]^{r-1}(1+|s|^2)^{r/2}\underline{\sigma}^{-1}|s|^{-1} \tag*{(47)} ∥ K ( s ) ∥ L 2 ( R m ) → H r ( R m ) ≤ c m a x 2 [ 2 2 c m i n − 2 ] r − 1 ( 1 + ∣ s ∣ 2 ) r /2 σ − 1 ∣ s ∣ − 1 (47) ≤ 2 c max 2 [ 2 2 c min − 2 ] r − 1 σ ‾ − r σ ‾ − 1 ∣ s ∣ r − 1 . \le\sqrt{2}c_{\max}^2\left[2\sqrt{2}c_{\min}^{-2}\right]^{r-1}\underline{\sigma}^{-r}\underline{\sigma}^{-1}|s|^{r-1}. ≤ 2 c m a x 2 [ 2 2 c m i n − 2 ] r − 1 σ − r σ − 1 ∣ s ∣ r − 1 .
Proof. The sesquilinear form associated to K ( s ) K(s) K ( s ) is coercive after a rotation in the complex plane: taking V = s / ∣ s ∣ U V=s/|s|\,U V = s /∣ s ∣ U , we have
s ∣ s ∣ ∥ c − 1 U ∥ L 2 ( R m ) 2 + s ‾ ∣ s ∣ ∥ ∇ V ∥ L 2 ( R m ) 2 = a s , c , R m ( U , V ) = ( F , V ) R m s|s|\|c^{-1}U\|_{L^2(\mathbb{R}^m)}^2+\frac{\overline{s}}{|s|}\|\nabla V\|_{L^2(\mathbb{R}^m)}^2=a_{s,c,\mathbb{R}^m}(U,V)=(F,V)_{\mathbb{R}^m} s ∣ s ∣∥ c − 1 U ∥ L 2 ( R m ) 2 + ∣ s ∣ s ∥∇ V ∥ L 2 ( R m ) 2 = a s , c , R m ( U , V ) = ( F , V ) R m
and therefore
∥ F ∥ L 2 ( R m ) ∥ U ∥ L 2 ( R m ) ≥ ∣ ( F , V ) R m ∣ ≥ ℜ ( a s , c , R m ( U , V ) ) ≥ σ ( c max − 2 ∣ s ∣ ∥ U ∥ L 2 ( R m ) 2 + 1 ∣ s ∣ ∥ ∇ V ∥ L 2 ( R m ) 2 ) \begin{aligned}
\|F\|_{L^2(\mathbb{R}^m)}\|U\|_{L^2(\mathbb{R}^m)}\geq|(F,V)_{\mathbb{R}^m}|\geq\Re(a_{s,c,\mathbb{R}^m}(U,V)) \\
&\geq\sigma\left(c_{\max}^{-2}|s|\|U\|_{L^2(\mathbb{R}^m)}^2+\frac{1}{|s|}\|\nabla V\|_{L^2(\mathbb{R}^m)}^2\right)
\end{aligned} ∥ F ∥ L 2 ( R m ) ∥ U ∥ L 2 ( R m ) ≥ ∣ ( F , V ) R m ∣ ≥ ℜ ( a s , c , R m ( U , V )) ≥ σ ( c m a x − 2 ∣ s ∣∥ U ∥ L 2 ( R m ) 2 + ∣ s ∣ 1 ∥∇ V ∥ L 2 ( R m ) 2 )
≥ 2 σ ( c max − 1 ∥ U ∥ L 2 ( R m ) ∥ ∇ V ∥ L 2 ( R m ) ) . \geq2\sigma\left(c_{\max}^{-1}\|U\|_{L^2(\mathbb{R}^m)}\|\nabla V\|_{L^2(\mathbb{R}^m)}\right). ≥ 2 σ ( c m a x − 1 ∥ U ∥ L 2 ( R m ) ∥∇ V ∥ L 2 ( R m ) ) .
(In the last inequality we have used a 2 + b 2 ≥ 2 a b a^2+b^2\geq2ab a 2 + b 2 ≥ 2 ab .) Hence
∥ U ∥ L 2 ( R m ) ≤ c max 2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) , \|U\|_{L^2(\mathbb{R}^m)}\leq c_{\max}^2\sigma^{-1}|s|^{-1}\|F\|_{L^2(\mathbb{R}^m)}, ∥ U ∥ L 2 ( R m ) ≤ c m a x 2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ,
∥ ∇ V ∥ L 2 ( R m ) ≤ 1 2 c max σ − 1 ∥ F ∥ L 2 ( R m ) . \|\nabla V\|_{L^2(\mathbb{R}^m)}\leq\frac{1}{2}c_{\max}\sigma^{-1}\|F\|_{L^2(\mathbb{R}^m)}. ∥∇ V ∥ L 2 ( R m ) ≤ 2 1 c m a x σ − 1 ∥ F ∥ L 2 ( R m ) .
Clearly, (35c) follows from
∥ Δ U ∥ L 2 ( R m ) ≤ ∥ F ∥ L 2 ( R m ) + ∣ s / c min ∣ 2 ∥ U ∥ L 2 ( R m ) . \|\Delta U\|_{L^2(\mathbb{R}^m)}\leq\|F\|_{L^2(\mathbb{R}^m)}+|s/c_{\min}|^2\|U\|_{L^2(\mathbb{R}^m)}. ∥Δ U ∥ L 2 ( R m ) ≤ ∥ F ∥ L 2 ( R m ) + ∣ s / c m i n ∣ 2 ∥ U ∥ L 2 ( R m ) .
On the other hand, we invoke (32) to derive
∥ U ∥ H 2 ( R m ) 2 ≤ 2 [ ∥ U ∥ L 2 ( R m ) 2 + ∥ Δ U ∥ L 2 ( R m ) 2 ] \|U\|_{H^2(\mathbb{R}^m)}^2\leq2\left[\|U\|_{L^2(\mathbb{R}^m)}^2+\|\Delta U\|_{L^2(\mathbb{R}^m)}^2\right] ∥ U ∥ H 2 ( R m ) 2 ≤ 2 [ ∥ U ∥ L 2 ( R m ) 2 + ∥Δ U ∥ L 2 ( R m ) 2 ]
≤ 4 [ ( 1 + ∣ s / c min ∣ 4 ) ∥ U ∥ L 2 ( R m ) 2 + ∥ F ∥ L 2 ( R m ) 2 ] \leq4\left[(1+|s/c_{\min}|^4)\|U\|_{L^2(\mathbb{R}^m)}^2+\|F\|_{L^2(\mathbb{R}^m)}^2\right] ≤ 4 [ ( 1 + ∣ s / c m i n ∣ 4 ) ∥ U ∥ L 2 ( R m ) 2 + ∥ F ∥ L 2 ( R m ) 2 ]
≤ 4 [ ( 1 + ∣ s / c min ∣ 4 ) c max 4 σ − 2 ∣ s ∣ − 2 + 1 ] ∥ F ∥ L 2 ( R m ) 2 \leq4\left[(1+|s/c_{\min}|^4)c_{\max}^4\sigma^{-2}|s|^{-2}+1\right]\|F\|_{L^2(\mathbb{R}^m)}^2 ≤ 4 [ ( 1 + ∣ s / c m i n ∣ 4 ) c m a x 4 σ − 2 ∣ s ∣ − 2 + 1 ] ∥ F ∥ L 2 ( R m ) 2
≤ 8 ( 1 + ∣ s / c min ∣ 4 ) c max 4 σ − 2 ∣ s ∣ − 2 ∥ F ∥ L 2 ( R m ) 2 \leq8(1+|s/c_{\min}|^4)c_{\max}^4\sigma^{-2}|s|^{-2}\|F\|_{L^2(\mathbb{R}^m)}^2 ≤ 8 ( 1 + ∣ s / c m i n ∣ 4 ) c m a x 4 σ − 2 ∣ s ∣ − 2 ∥ F ∥ L 2 ( R m ) 2
≤ 8 ( c max / c min ) 4 ( 1 + ∣ s ∣ 4 ) σ − 2 ∣ s ∣ − 2 ∥ F ∥ L 2 ( R m ) 2 \leq8(c_{\max}/c_{\min})^4(1+|s|^4)\sigma^{-2}|s|^{-2}\|F\|_{L^2(\mathbb{R}^m)}^2 ≤ 8 ( c m a x / c m i n ) 4 ( 1 + ∣ s ∣ 4 ) σ − 2 ∣ s ∣ − 2 ∥ F ∥ L 2 ( R m ) 2
≤ 16 ( c max / c min ) 4 σ − 4 σ − 2 ∣ s ∣ 2 ∥ F ∥ L 2 ( R m ) 2 . \leq16(c_{\max}/c_{\min})^4\sigma^{-4}\sigma^{-2}|s|^2\|F\|_{L^2(\mathbb{R}^m)}^2. ≤ 16 ( c m a x / c m i n ) 4 σ − 4 σ − 2 ∣ s ∣ 2 ∥ F ∥ L 2 ( R m ) 2 .
which proves (36) for r = 2 r=2 r = 2 . (Notice that (33) has been used in the last inequality.) Also, from (35a)–(35b),
∥ U ∥ H 1 ( R m ) = [ ∥ U ∥ L 2 ( R m ) 2 + ∥ ∇ V ∥ L 2 ( R m ) 2 ] 1 / 2 ≤ c max 2 ( 1 + ∣ s 2 c max ∣ 2 ) 1 / 2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ≤ c max 2 ( 1 + ∣ s ∣ 2 ) 1 / 2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ≤ 2 c max 2 σ − 1 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) , \begin{aligned}
\|U\|_{H^1(\mathbb{R}^m)}=\left[\|U\|_{L^2(\mathbb{R}^m)}^2+\|\nabla V\|_{L^2(\mathbb{R}^m)}^2\right]^{1/2}\leq c_{\max}^2\left(1+\left|\frac{s}{2c_{\max}}\right|^2\right)^{1/2}\sigma^{-1}|s|^{-1}\|F\|_{L^2(\mathbb{R}^m)} \\
&\leq c_{\max}^2(1+|s|^2)^{1/2}\sigma^{-1}|s|^{-1}\|F\|_{L^2(\mathbb{R}^m)}\leq\sqrt{2}c_{\max}^2\sigma^{-1}\sigma^{-1}|s|^{-1}\|F\|_{L^2(\mathbb{R}^m)},
\end{aligned} ∥ U ∥ H 1 ( R m ) = [ ∥ U ∥ L 2 ( R m ) 2 + ∥∇ V ∥ L 2 ( R m ) 2 ] 1/2 ≤ c m a x 2 ( 1 + 2 c m a x s 2 ) 1/2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ≤ c m a x 2 ( 1 + ∣ s ∣ 2 ) 1/2 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ≤ 2 c m a x 2 σ − 1 σ − 1 ∣ s ∣ − 1 ∥ F ∥ L 2 ( R m ) ,
which proves (35d)–(35e) and (36) for r = 1 r=1 r = 1 . The result for r ∈ ( 1 , 2 ) r\in(1,2) r ∈ ( 1 , 2 ) follows by interpolation. □
In order to simplify the exposition we will use in this section the usual s s s –dependent norm, equivalent to the H 1 ( D ) H^1(D) H 1 ( D ) norm,
∥ U ∥ s , D 2 = ∣ s ∣ 2 ∥ U ∥ L 2 ( D ) 2 + ∥ ∇ V ∥ L 2 ( D ) 2 . \|U\|_{s,D}^2=|s|^2\|U\|_{L^2(D)}^2+\|\nabla V\|_{L^2(D)}^2. ∥ U ∥ s , D 2 = ∣ s ∣ 2 ∥ U ∥ L 2 ( D ) 2 + ∥∇ V ∥ L 2 ( D ) 2 .
Clearly,
K ( s ) F ∣ Ω 2 = K t r ( s ) ( F , γ Σ K ( s ) F ) , K(s)F|_{\Omega_2}=K^{\mathrm{tr}}(s)(F,\gamma_{\Sigma}K(s)F), K ( s ) F ∣ Ω 2 = K tr ( s ) ( F , γ Σ K ( s ) F ) ,
where
K t r ( s ) ( F , g Σ ) : = U , { U ∈ H 1 ( Ω 2 ) , γ Σ U = g Σ , s 2 c 2 U − Δ U = F , in Ω 2 . (48) K^{\mathrm{tr}}(s)(F,g_{\Sigma}):=U,\qquad
\begin{cases}
U\in H^1(\Omega_2),\quad\gamma_{\Sigma}U=g_{\Sigma},\\
\frac{s^2}{c^2}U-\Delta U=F,\quad\text{in }\Omega_2.
\end{cases} \tag*{(48)} K tr ( s ) ( F , g Σ ) := U , { U ∈ H 1 ( Ω 2 ) , γ Σ U = g Σ , c 2 s 2 U − Δ U = F , in Ω 2 . (48)
Proposition 4.2. For any s s s with σ = ℜ s > 0 \sigma=\Re s>0 σ = ℜ s > 0 the following holds:
∥ K t r ( s ) ( F , g Σ ) ∥ s , Ω 2 ≤ C σ − 1 c max 2 ∥ F ∥ L 2 ( Ω 2 ) + C σ − 1 ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1 / 2 ( Σ ) . (49) \left\|K^{\mathrm{tr}}(s)(F,g_{\Sigma})\right\|_{s,\Omega_2}
\le C\sigma^{-1}c_{\max}^{2}\left\|F\right\|_{L^2(\Omega_2)}
+C\sigma^{-1}(1+|s|)^{1/2}|s|\left(1+c_{\max}^{2}c_{0,R}\sigma^{-1}|s|\right)\left\|g_{\Sigma}\right\|_{H^{1/2}(\Sigma)}. \tag*{(49)} K tr ( s ) ( F , g Σ ) s , Ω 2 ≤ C σ − 1 c m a x 2 ∥ F ∥ L 2 ( Ω 2 ) + C σ − 1 ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1/2 ( Σ ) . (49) where C C C is a constant depending only on the domain Ω 2 \Omega_2 Ω 2 , and
c max = ∥ c ∥ L ∞ ( R m ) ≥ 1 , c 0 , R = ∥ c R ∥ L ∞ ( Ω 2 ) = ∥ c − 2 − 1 ∥ L ∞ ( Ω 2 ) . c_{\max}=\|c\|_{L^\infty(\mathbb{R}^m)}\ge1,\qquad
c_{0,R}=\|c_R\|_{L^\infty(\Omega_2)}=\|c^{-2}-1\|_{L^\infty(\Omega_2)}. c m a x = ∥ c ∥ L ∞ ( R m ) ≥ 1 , c 0 , R = ∥ c R ∥ L ∞ ( Ω 2 ) = ∥ c − 2 − 1 ∥ L ∞ ( Ω 2 ) .
Proof. Throughout the proof, C C C denotes a constant depending only on Ω 2 \Omega_2 Ω 2 , which may change from line to line.
Let U 1 U_1 U 1 and U 2 U_2 U 2 be the respective solutions of the problems
{ U 1 ∈ H 1 ( Ω 2 ) , − Δ U 1 + s 2 U 1 = 0 , γ Σ U 1 = g Σ , { U 2 ∈ H 0 1 ( Ω 2 ) , − Δ U 2 + ( s / c ) 2 U 2 = F 2 , γ Σ U 2 = 0 , \begin{cases}
U_1\in H^1(\Omega_2),\\
-\Delta U_1+s^2U_1=0,\\
\gamma_{\Sigma}U_1=g_{\Sigma},
\end{cases}
\qquad
\begin{cases}
U_2\in H_0^1(\Omega_2),\\
-\Delta U_2+(s/c)^2U_2=F_2,\\
\gamma_{\Sigma}U_2=0,
\end{cases} ⎩ ⎨ ⎧ U 1 ∈ H 1 ( Ω 2 ) , − Δ U 1 + s 2 U 1 = 0 , γ Σ U 1 = g Σ , ⎩ ⎨ ⎧ U 2 ∈ H 0 1 ( Ω 2 ) , − Δ U 2 + ( s / c ) 2 U 2 = F 2 , γ Σ U 2 = 0 ,
with
F 2 : = F − [ c − 2 − 1 ] s 2 U 1 . (50) F_2:=F-[c^{-2}-1]s^2U_1. \tag*{(50)} F 2 := F − [ c − 2 − 1 ] s 2 U 1 . (50)
Notice that U 1 + U 2 = K t r ( s ) ( F , g Σ ) U_1+U_2=K^{\mathrm{tr}}(s)(F,g_{\Sigma}) U 1 + U 2 = K tr ( s ) ( F , g Σ ) , so it suffices to bound these two functions to prove the result.
For U 1 U_1 U 1 , we can use (27) (see also [6 Lemma 4.13] ) to conclude
∣ s ∣ ∥ U 1 ∥ L 2 ( Ω 2 ) + ∥ ∇ U 1 ∥ L 2 ( Ω 2 ) ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ ∥ g Σ ∥ H 1 / 2 ( Σ ) . (51) |s|\|U_1\|_{L^2(\Omega_2)}+\|\nabla U_1\|_{L^2(\Omega_2)}
\le C\sigma^{-1}(1+|s|)^{1/2}|s|\|g_{\Sigma}\|_{H^{1/2}(\Sigma)}. \tag*{(51)} ∣ s ∣∥ U 1 ∥ L 2 ( Ω 2 ) + ∥∇ U 1 ∥ L 2 ( Ω 2 ) ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣∥ g Σ ∥ H 1/2 ( Σ ) . (51)
As a consequence,
∥ F 2 ∥ L 2 ( Ω 2 ) ≤ ∥ F ∥ L 2 ( Ω 2 ) + C c 0 , R σ − 1 ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ 2 ∥ g Σ ∥ H 1 / 2 ( Σ ) . (52) \|F_2\|_{L^2(\Omega_2)}
\le\|F\|_{L^2(\Omega_2)}
+Cc_{0,R}\sigma^{-1}(1+|s|)^{1/2}|s|^2\|g_{\Sigma}\|_{H^{1/2}(\Sigma)}. \tag*{(52)} ∥ F 2 ∥ L 2 ( Ω 2 ) ≤ ∥ F ∥ L 2 ( Ω 2 ) + C c 0 , R σ − 1 ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ 2 ∥ g Σ ∥ H 1/2 ( Σ ) . (52)
On the other hand, and similarly to (31), we have that U 2 U_2 U 2 solves
{ U 2 ∈ H 0 1 ( Ω 2 ) , a s , c , Ω 2 ( U 2 , V ) = ( F 2 , V ) Ω 2 , ∀ V ∈ H 0 1 ( Ω 2 ) . (53) \begin{cases}
U_2\in H_0^1(\Omega_2),\\
a_{s,c,\Omega_2}(U_2,V)=(F_2,V)_{\Omega_2},\quad\forall V\in H_0^1(\Omega_2). \tag*{(53)}
\end{cases} { U 2 ∈ H 0 1 ( Ω 2 ) , a s , c , Ω 2 ( U 2 , V ) = ( F 2 , V ) Ω 2 , ∀ V ∈ H 0 1 ( Ω 2 ) . (53)
Proceeding as in the proof of Proposition 4.1,
∥ U 2 ∥ L 2 ( Ω 2 ) ≤ c max 2 σ − 1 ∣ s ∣ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) (54) \|U_2\|_{L^2(\Omega_2)}
\le c_{\max}^2\sigma^{-1}|s|^{-1}\|F_2\|_{L^2(\Omega_2)} \tag*{(54)} ∥ U 2 ∥ L 2 ( Ω 2 ) ≤ c m a x 2 σ − 1 ∣ s ∣ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) (54)
∥ ∇ U 2 ∥ L 2 ( Ω 2 ) ≤ 1 2 c max σ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) , (55) \|\nabla U_2\|_{L^2(\Omega_2)}
\le\frac{1}{2}c_{\max}\sigma^{-1}\|F_2\|_{L^2(\Omega_2)}, \tag*{(55)} ∥∇ U 2 ∥ L 2 ( Ω 2 ) ≤ 2 1 c m a x σ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) , (55)
From (43)–(44), we derive
∥ U 2 ∥ s , Ω 2 ≤ ( c max + 1 2 ) 2 σ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) . (56) \|U_2\|_{s,\Omega_2}
\le\left(c_{\max}+\frac{1}{2}\right)^2\sigma^{-1}\|F_2\|_{L^2(\Omega_2)}. \tag*{(56)} ∥ U 2 ∥ s , Ω 2 ≤ ( c m a x + 2 1 ) 2 σ − 1 ∥ F 2 ∥ L 2 ( Ω 2 ) . (56)
Combining (39)–(45), the result follows. □
Exponential decay away from the inhomogeneity
Proposition 4.4 Let D ⊂ R m D\subset\mathbb{R}^m D ⊂ R m , m = 2 , 3 m=2,3 m = 2 , 3 , be open, let E ⊂ D E\subset D E ⊂ D be bounded, and let 0 < d 0 < d : = d ( E , ∂ D ) 0<d_0<d:=d(E,\partial D) 0 < d 0 < d := d ( E , ∂ D ) and σ 0 > 0 \sigma_0>0 σ 0 > 0 . Set
d ′ = 1 2 ( d 0 + d ) , D d 0 , E : = { x ∈ D : d 0 < d ( x , E ) < d ′ } . d'=\frac{1}{2}(d_0+d),\qquad D_{d_0,E}:=\{\mathbf{x}\in D:d_0<d(\mathbf{x},E)<d'\}. d ′ = 2 1 ( d 0 + d ) , D d 0 , E := { x ∈ D : d 0 < d ( x , E ) < d ′ } . Then for every multi-index α ∈ N 0 m \boldsymbol{\alpha}\in\mathbb{N}_0^m α ∈ N 0 m there exists C > 0 C>0 C > 0 , depending only on m m m , ∣ α ∣ |\boldsymbol{\alpha}| ∣ α ∣ , σ 0 \sigma_0 σ 0 , d 0 d_0 d 0 , d − d 0 d-d_0 d − d 0 and ∣ D d 0 , E ∣ |D_{d_0,E}| ∣ D d 0 , E ∣ , such that for every s s s with σ = ℜ s ≥ σ 0 \sigma=\Re s\ge\sigma_0 σ = ℜ s ≥ σ 0 and every W ∈ H 1 ( D ) W\in H^1(D) W ∈ H 1 ( D ) satisfying
s 2 W − Δ W = 0 in D , s^2W-\Delta W=0\quad\text{in }D, s 2 W − Δ W = 0 in D , the following holds:
∥ ∂ α W ∥ L ∞ ( E ) ≤ C e − σ d 0 ∣ s ∣ ∣ α ∣ + ( m − 1 ) / 2 ∥ W ∥ L 2 ( D d 0 , E ) . (61) \|\partial_{\boldsymbol{\alpha}}W\|_{L^\infty(E)}
\le C e^{-\sigma d_0}|s|^{|\boldsymbol{\alpha}|+(m-1)/2}\|W\|_{L^2(D_{d_0,E})}. \tag*{(61)} ∥ ∂ α W ∥ L ∞ ( E ) ≤ C e − σ d 0 ∣ s ∣ ∣ α ∣ + ( m − 1 ) /2 ∥ W ∥ L 2 ( D d 0 , E ) . (61)
Proof. Let η ∈ C ∞ ( R m ) \eta\in C^\infty(\mathbb{R}^m) η ∈ C ∞ ( R m ) be a cut-off function with 0 ≤ η ≤ 1 0\leq\eta\leq1 0 ≤ η ≤ 1 , η ≡ 1 \eta\equiv1 η ≡ 1 on E d 0 E_{d_0} E d 0 and η ≡ 0 \eta\equiv0 η ≡ 0 outside E d ′ E_{d'} E d ′ , with
E ε : = { x ∈ R m : d ( x , E ) < ε } E_\varepsilon:=\{x\in\mathbb{R}^m:d(x,E)<\varepsilon\} E ε := { x ∈ R m : d ( x , E ) < ε }
so that supp ∇ η ⊂ D ‾ d 0 , E \operatorname{supp}\nabla\eta\subset\overline{D}_{d_0,E} supp ∇ η ⊂ D d 0 , E . Constructing η \eta η by mollification of the characteristic function of E ( d 0 + d ′ ) / 2 E_{(d_0+d')/2} E ( d 0 + d ′ ) /2 with a kernel of radius 1 2 ( d ′ − d 0 ) \frac{1}{2}(d'-d_0) 2 1 ( d ′ − d 0 ) , we have in addition
∥ ∂ β η ∥ L ∞ ( R m ) ≤ C β ( d ′ − d 0 ) − ∣ β ∣ , (62) \|\partial^\beta\eta\|_{L^\infty(\mathbb{R}^m)}\leq C_\beta(d'-d_0)^{-|\beta|}, \tag*{(62)} ∥ ∂ β η ∥ L ∞ ( R m ) ≤ C β ( d ′ − d 0 ) − ∣ β ∣ , (62)
with C β C_\beta C β depending only on m m m and β \beta β , in particular neither on E E E nor on D D D (see Figure 2 for 1D sketch of this construction).
Figure 2. One-dimensional sketch of the construction in the proof of Proposition 4.4 : the cut-off function η \eta η equals 1 on E d 0 E_{d_0} E d 0 and 0 outside E d ′ E_{d'} E d ′ , so that ∇ η \nabla\eta ∇ η is supported in the shell D d 0 , E = { d 0 < d ( x , E ) < d ′ } D_{d_0,E}=\{d_0<d(x,E)<d'\} D d 0 , E = { d 0 < d ( x , E ) < d ′ } (shaded), which lies at distance at least d 0 d_0 d 0 from E E E and inside D D D .
Let V : = η W V:=\eta W V := η W , extended by zero to R m \mathbb{R}^m R m . Then V ∈ H 1 ( R m ) V\in H^1(\mathbb{R}^m) V ∈ H 1 ( R m ) , supp V ⊂ E ‾ d ′ ⊂ D \operatorname{supp}V\subset\overline{E}_{d'}\subset D supp V ⊂ E d ′ ⊂ D , V ∣ E = W ∣ E V|_E=W|_E V ∣ E = W ∣ E , and
s 2 V − Δ V = f : = − 2 ∇ η ⋅ ∇ W − ( Δ η ) W ∈ L 2 ( R m ) , supp f ⊂ D ‾ d 0 , E . s^2V-\Delta V=f:=-2\nabla\eta\cdot\nabla W-(\Delta\eta)W\in L^2(\mathbb{R}^m),\qquad\operatorname{supp}f\subset\overline{D}_{d_0,E}. s 2 V − Δ V = f := − 2∇ η ⋅ ∇ W − ( Δ η ) W ∈ L 2 ( R m ) , supp f ⊂ D d 0 , E .
Since σ = ℜ s > 0 \sigma=\Re s>0 σ = ℜ s > 0 , V V V is the unique H 1 ( R m ) H^1(\mathbb{R}^m) H 1 ( R m ) solution of this equation, and therefore
V ( x ) = ∫ R m ϕ ( x − y ) f ( y ) d y , where ϕ ( z ; s ) = { i 4 H 0 ( 1 ) ( i s ∣ z ∣ ) , m = 2 , e − s ∣ z ∣ 4 π ∣ z ∣ , m = 3 , V(x)=\int_{\mathbb{R}^m}\phi(x-y)f(y)\,\mathrm{d}y,
\qquad\text{where}\qquad
\phi(z;s)=
\begin{cases}
\frac{\mathrm{i}}{4}H_0^{(1)}(\mathrm{i}s|z|), & m=2,\\
\frac{e^{-s|z|}}{4\pi|z|}, & m=3,
\end{cases} V ( x ) = ∫ R m ϕ ( x − y ) f ( y ) d y , where ϕ ( z ; s ) = { 4 i H 0 ( 1 ) ( i s ∣ z ∣ ) , 4 π ∣ z ∣ e − s ∣ z ∣ , m = 2 , m = 3 ,
is the fundamental solution of s 2 − Δ s^2-\Delta s 2 − Δ . For x ∈ E x\in E x ∈ E and y ∈ D d 0 , E y\in D_{d_0,E} y ∈ D d 0 , E we have ∣ x − y ∣ ≥ d 0 |x-y|\geq d_0 ∣ x − y ∣ ≥ d 0 , so that ϕ ( x − ⋅ ; s ) \phi(x-\cdot\,;s) ϕ ( x − ⋅ ; s ) is smooth on D ‾ d 0 , E \overline{D}_{d_0,E} D d 0 , E and, since ϕ ( x − ⋅ ; s ) ∇ η \phi(x-\cdot\,;s)\nabla\eta ϕ ( x − ⋅ ; s ) ∇ η has compact support in D D D , we may integrate by parts to obtain
W ( x ) = V ( x ) = ∫ D d 0 , E [ 2 ∇ y ϕ ( x − y ; s ) ⋅ ∇ η ( y ) + ϕ ( x − y ; s ) Δ η ( y ) ] W ( y ) d y , x ∈ E . (63) W(x)=V(x)=\int_{D_{d_0,E}}\left[2\nabla_y\phi(x-y;s)\cdot\nabla\eta(y)+\phi(x-y;s)\Delta\eta(y)\right]W(y)\,\mathrm{d}y,\qquad x\in E. \tag*{(63)} W ( x ) = V ( x ) = ∫ D d 0 , E [ 2 ∇ y ϕ ( x − y ; s ) ⋅ ∇ η ( y ) + ϕ ( x − y ; s ) Δ η ( y ) ] W ( y ) d y , x ∈ E . (63)
Next, for ∣ z ∣ ≥ d 0 |z|\geq d_0 ∣ z ∣ ≥ d 0 , ℜ s ≥ σ 0 \Re s\geq\sigma_0 ℜ s ≥ σ 0 and any multi-index β \beta β ,
∣ ∂ β ϕ ( z ; s ) ∣ ≤ C e − σ ∣ z ∣ ∣ s ∣ ∣ β ∣ + ( m − 3 ) / 2 , (64) |\partial^\beta\phi(z;s)|\leq C e^{-\sigma|z|}|s|^{|\beta|+(m-3)/2}, \tag*{(64)} ∣ ∂ β ϕ ( z ; s ) ∣ ≤ C e − σ ∣ z ∣ ∣ s ∣ ∣ β ∣ + ( m − 3 ) /2 , (64)
with C C C depending only on m m m , ∣ β ∣ |\beta| ∣ β ∣ , d 0 d_0 d 0 and σ 0 \sigma_0 σ 0 . For m = 3 m=3 m = 3 this follows by direct differentiation, each derivative producing a factor ∣ s ∣ |s| ∣ s ∣ or ∣ z ∣ − 1 ≤ d 0 − 1 ≤ ∣ s ∣ / ( σ 0 d 0 ) |z|^{-1}\leq d_0^{-1}\leq|s|/(\sigma_0d_0) ∣ z ∣ − 1 ≤ d 0 − 1 ≤ ∣ s ∣/ ( σ 0 d 0 ) . For m = 2 m=2 m = 2 it follows from ∣ H 1 ( 1 ) ( w ) ∣ ≤ C ν ∣ w ∣ − 1 / 2 e − ℑ w |H_1^{(1)}(w)| \le C_\nu|w|^{-1/2}e^{-\Im w} ∣ H 1 ( 1 ) ( w ) ∣ ≤ C ν ∣ w ∣ − 1/2 e − ℑ w , valid for ℑ w ≥ 0 \Im w \ge0 ℑ w ≥ 0 and ∣ w ∣ ≥ σ 0 d 0 |w| \ge\sigma_0d_0 ∣ w ∣ ≥ σ 0 d 0 , applied with w = i s ∣ z ∣ w=is|z| w = i s ∣ z ∣ , so that ℑ w = σ ∣ z ∣ \Im w=\sigma|z| ℑ w = σ ∣ z ∣ , together with ( H 0 ( 1 ) ) ′ = − H 1 ( 1 ) (H_0^{(1)})'=-H_1^{(1)} ( H 0 ( 1 ) ) ′ = − H 1 ( 1 ) and the recurrences for higher derivatives.
Differentiating (51) under the integral sign, applying (52) with ∣ β ∣ ≤ ∣ α ∣ + 1 |\beta| \le|\alpha|+1 ∣ β ∣ ≤ ∣ α ∣ + 1 and the Cauchy–Schwarz inequality, we obtain for x ∈ E \mathbf{x} \in E x ∈ E
∣ ∂ α W ( x ) ∣ ≤ C e − σ d 0 ∣ s ∣ ∣ α ∣ + ( m − 3 ) / 2 ( ∣ s ∣ ∥ ∇ η ∥ L 2 ( D d 0 , E ) + ∥ Δ η ∥ L 2 ( D d 0 , E ) ) ∥ W ∥ L 2 ( D d 0 , E ) . |\partial^\alpha W(\mathbf{x})| \le C e^{-\sigma d_0}|s|^{|\alpha|+(m-3)/2}\left(|s|\|\nabla\eta\|_{L^2(D_{d_0,E})}+\|\Delta\eta\|_{L^2(D_{d_0,E})}\right)\|W\|_{L^2(D_{d_0,E})}. ∣ ∂ α W ( x ) ∣ ≤ C e − σ d 0 ∣ s ∣ ∣ α ∣ + ( m − 3 ) /2 ( ∣ s ∣∥∇ η ∥ L 2 ( D d 0 , E ) + ∥Δ η ∥ L 2 ( D d 0 , E ) ) ∥ W ∥ L 2 ( D d 0 , E ) .
By (50),
∥ ∇ η ∥ L 2 ( D d 0 , E ) ≤ C ( d − d 0 ) − 1 ∣ D d 0 , E ∣ 1 / 2 \|\nabla\eta\|_{L^2(D_{d_0,E})} \le C(d-d_0)^{-1}|D_{d_0,E}|^{1/2} ∥∇ η ∥ L 2 ( D d 0 , E ) ≤ C ( d − d 0 ) − 1 ∣ D d 0 , E ∣ 1/2
∥ Δ η ∥ L 2 ( D d 0 , E ) ≤ C ( d − d 0 ) − 2 ∣ D d 0 , E ∣ 1 / 2 ≤ C ( d − d 0 ) − 2 σ 0 − 1 ∣ s ∣ ∣ D d 0 , E ∣ 1 / 2 , \|\Delta\eta\|_{L^2(D_{d_0,E})} \le C(d-d_0)^{-2}|D_{d_0,E}|^{1/2} \le C(d-d_0)^{-2}\sigma_0^{-1}|s||D_{d_0,E}|^{1/2}, ∥Δ η ∥ L 2 ( D d 0 , E ) ≤ C ( d − d 0 ) − 2 ∣ D d 0 , E ∣ 1/2 ≤ C ( d − d 0 ) − 2 σ 0 − 1 ∣ s ∣∣ D d 0 , E ∣ 1/2 ,
which proves (49). □
The next result analyzes the decay of K ( s ) F K(s)F K ( s ) F far away from Ω \Omega Ω , the support of F F F .
Proposition 4.5. Let σ 0 > 0 \sigma_0>0 σ 0 > 0 , let E ⊂ Ω c E\subset\Omega^c E ⊂ Ω c be bounded, let 0 < d < d 0 : = d ( E , Ω ) 0<d<d_0:=d(E,\Omega) 0 < d < d 0 := d ( E , Ω ) and α ∈ N 0 m \alpha\in\mathbb{N}_0^m α ∈ N 0 m . Then there exists C > 0 C>0 C > 0 , depending only on m m m , ∣ α ∣ |\alpha| ∣ α ∣ , σ 0 \sigma_0 σ 0 , d d d , d 0 d_0 d 0 and E E E , such that for any s s s with σ = ℜ s ≥ σ 0 \sigma=\Re s\ge\sigma_0 σ = ℜ s ≥ σ 0
∥ ∂ α K ( s ) ∥ L 2 ( Ω ) → L ∞ ( E ) ≤ C c max 2 σ − 1 e − σ d ∣ s ∣ ( m − 3 ) / 2 + ∣ α ∣ , c max = ∥ c ∥ L ∞ ( R m ) ≥ 1. (65) \|\partial^\alpha K(s)\|_{L^2(\Omega)\to L^\infty(E)} \le Cc_{\max}^2\sigma^{-1}e^{-\sigma d}|s|^{(m-3)/2+|\alpha|}, \qquad c_{\max}=\|c\|_{L^\infty(\mathbb{R}^m)}\ge1. \tag*{(65)} ∥ ∂ α K ( s ) ∥ L 2 ( Ω ) → L ∞ ( E ) ≤ C c m a x 2 σ − 1 e − σ d ∣ s ∣ ( m − 3 ) /2 + ∣ α ∣ , c m a x = ∥ c ∥ L ∞ ( R m ) ≥ 1. (65) Proof. For F ∈ L 2 ( Ω ) F\in L^2(\Omega) F ∈ L 2 ( Ω ) , W 1 : = K ( s ) F ∣ Ω c ∈ H 1 ( Ω c ) W_1:=K(s)F|_{\Omega^c}\in H^1(\Omega^c) W 1 := K ( s ) F ∣ Ω c ∈ H 1 ( Ω c ) satisfies s 2 W 1 − Δ W 1 = 0 s^2W_1-\Delta W_1=0 s 2 W 1 − Δ W 1 = 0 in Ω c \Omega^c Ω c , since c ≡ 1 c\equiv1 c ≡ 1 and F ≡ 0 F\equiv0 F ≡ 0 there. Proposition 4.4, applied with D = Ω c D=\Omega^c D = Ω c and d 0 d_0 d 0 replaced by d < d ( E , ∂ D ) d<d(E,\partial D) d < d ( E , ∂ D ) , together with (35a), gives
∥ ∂ α W 1 ∥ L ∞ ( E ) ≤ C e − σ d ∣ s ∣ ∣ α ∣ + ( m − 3 ) / 2 ∣ s ∣ ∥ K ( s ) F ∥ L 2 ( R m ) \|\partial^\alpha W_1\|_{L^\infty(E)} \le Ce^{-\sigma d}|s|^{|\alpha|+(m-3)/2}|s|\|K(s)F\|_{L^2(\mathbb{R}^m)} ∥ ∂ α W 1 ∥ L ∞ ( E ) ≤ C e − σ d ∣ s ∣ ∣ α ∣ + ( m − 3 ) /2 ∣ s ∣∥ K ( s ) F ∥ L 2 ( R m ) ≤ C c max 2 σ − 1 e − σ d ∣ s ∣ ∣ α ∣ + ( m − 3 ) / 2 ∥ F ∥ L 2 ( Ω ) . \le Cc_{\max}^2\sigma^{-1}e^{-\sigma d}|s|^{|\alpha|+(m-3)/2}\|F\|_{L^2(\Omega)}. ≤ C c m a x 2 σ − 1 e − σ d ∣ s ∣ ∣ α ∣ + ( m − 3 ) /2 ∥ F ∥ L 2 ( Ω ) . □
Proof of Theorems 2.2, 3.2 and 3.3 Proof of Theorem 2.2 Following [6 ] , let A ( μ ; B ( X , Y ) ) \mathcal{A}(\mu;B(X,Y)) A ( μ ; B ( X , Y )) denote the class of analytic functions F : C + → B ( X , Y ) F:\mathbb{C}_+\to B(X,Y) F : C + → B ( X , Y ) such that
∥ F ( s ) ∥ ≤ C F ( σ ) ∣ s ∣ μ , s ∈ C + \|F(s)\|\le C_F(\sigma)|s|^\mu,\qquad s\in\mathbb{C}_+ ∥ F ( s ) ∥ ≤ C F ( σ ) ∣ s ∣ μ , s ∈ C + with C F : ( 0 , ∞ ) → ( 0 , ∞ ) C_F:(0,\infty)\to(0,\infty) C F : ( 0 , ∞ ) → ( 0 , ∞ ) non-increasing and polynomially bounded as σ → 0 \sigma\to0 σ → 0 . By Proposition 4.1, estimate (36),
K ∈ A ( r − 1 ; B ( L 2 ( Ω ) , H r ( R m ) ) ) K\in\mathcal{A}(r-1;B(L^2(\Omega),H^r(\mathbb{R}^m))) K ∈ A ( r − 1 ; B ( L 2 ( Ω ) , H r ( R m ))) where
C K ( σ ) = C r σ ‾ − r σ − 1 , C r : = 2 c max 2 ( 2 2 ) r − 1 c min − 2 ( r − 1 ) , C_K(\sigma)=C_r\underline{\sigma}^{-r}\sigma^{-1},\qquad C_r:=\sqrt{2}\,c_{\max}^2(2\sqrt{2})^{r-1}c_{\min}^{-2(r-1)}, C K ( σ ) = C r σ − r σ − 1 , C r := 2 c m a x 2 ( 2 2 ) r − 1 c m i n − 2 ( r − 1 ) , with
σ ‾ : = min { 1 , σ } . \underline{\sigma}:=\min\{1,\sigma\}. σ := min { 1 , σ } .
Since the left-hand side at time t t t only depends on f ∣ [ 0 , t ] f|_{[0,t]} f ∣ [ 0 , t ] , we may extend f f f by a C p ~ C^{\widetilde p} C p function vanishing for t ≥ T + 1 t \ge T+1 t ≥ T + 1 , so that f ∈ W 0 p ~ ( L 2 ( Ω ) ) f \in W^{\widetilde p}_0(L^2(\Omega)) f ∈ W 0 p ( L 2 ( Ω )) in the notation of [6 ] . The result is then [6 Thm. 2.21] for BDF1 and BDF2, and [6 Thm. 2.23] for the trapezoidal rule, applied with μ = r − 1 ∈ [ 0 , 1 ] \mu=r-1\in[0,1] μ = r − 1 ∈ [ 0 , 1 ] . In both cases the regularity requirement reduces to p ~ > μ + p + 2 = r + p + 1 \widetilde p>\mu+p+2=r+p+1 p > μ + p + 2 = r + p + 1 (for the trapezoidal rule, max { 2 μ + 3 , μ + 4 } = μ + 4 \max\{2\mu+3,\mu+4\}=\mu+4 max { 2 μ + 3 , μ + 4 } = μ + 4 because μ ≤ 1 \mu\le1 μ ≤ 1 ), and the constant is proportional to σ p + r − p ~ C K ( σ ~ ) \sigma^{p+r-\widetilde p}C_K(\widetilde\sigma) σ p + r − p C K ( σ ) with
σ ~ = 1 2 σ min { 1 , C 2 } < 1 \widetilde\sigma=\frac{1}{2}\sigma\min\{1,C_2\}<1 σ = 2 1 σ min { 1 , C 2 } < 1
where C 2 C_2 C 2 depends only on δ \delta δ and κ 0 \kappa_0 κ 0 . Hence, C K ( σ ~ ) = C r σ ~ − ( r + 1 ) C_K(\widetilde\sigma)=C_r\widetilde\sigma^{-(r+1)} C K ( σ ) = C r σ − ( r + 1 ) , which yields the stated dependence on σ \sigma σ , c max c_{\max} c m a x and c min c_{\min} c m i n .
Proof of Theorem 3.2
Notice that by the trace theorem, if Ξ \Xi Ξ is a bounded Lipschitz curve/surface contained in a bounded Lipschitz domain E E E
∥ γ Ξ V ∥ H 1 / 2 ( Ξ ) ≤ C ( Ξ , E ) ∥ V ∥ H 1 ( E ) ≤ C ( Ξ , E ) ∣ E ∣ 1 / 2 ( ∥ V ∥ L ∞ ( E ) + ∥ ∇ V ∥ L ∞ ( E ) ) ≤ C ( Ξ , E ) max { 1 , σ 0 − 1 } ∣ E ∣ 1 / 2 ⏟ C ′ = C ′ ( σ 0 , Ξ , E ) ( ∣ s ∣ ∥ V ∥ L ∞ ( E ) + ∥ ∇ V ∥ L ∞ ( E ) ) (66) \begin{aligned}
\|\gamma_{\Xi}V\|_{H^{1/2}(\Xi)} &\le C(\Xi,E)\|V\|_{H^1(E)} \le C(\Xi,E)|E|^{1/2}\left(\|V\|_{L^\infty(E)}+\|\nabla V\|_{L^\infty(E)}\right)\\
&\le\underbrace{C(\Xi,E)\max\{1,\sigma_0^{-1}\}|E|^{1/2}}_{C'=C'(\sigma_0,\Xi,E)}\left(|s|\|V\|_{L^\infty(E)}+\|\nabla V\|_{L^\infty(E)}\right) \tag*{(66)}
\end{aligned} ∥ γ Ξ V ∥ H 1/2 ( Ξ ) ≤ C ( Ξ , E ) ∥ V ∥ H 1 ( E ) ≤ C ( Ξ , E ) ∣ E ∣ 1/2 ( ∥ V ∥ L ∞ ( E ) + ∥∇ V ∥ L ∞ ( E ) ) ≤ C ′ = C ′ ( σ 0 , Ξ , E ) C ( Ξ , E ) max { 1 , σ 0 − 1 } ∣ E ∣ 1/2 ( ∣ s ∣∥ V ∥ L ∞ ( E ) + ∥∇ V ∥ L ∞ ( E ) ) (66)
for all s s s with σ ≥ σ 0 \sigma\ge\sigma_0 σ ≥ σ 0 .
First, with the notation of this section
U ∣ Ω 2 = K t r ( s ) ( F , γ Σ K ( s ) F ) , W F t r = K t r ( s ) ( F , 0 ) . U|_{\Omega_2}=K^{\mathrm{tr}}(s)(F,\gamma_{\Sigma}K(s)F),\qquad W_F^{\mathrm{tr}}=K^{\mathrm{tr}}(s)(F,0). U ∣ Ω 2 = K tr ( s ) ( F , γ Σ K ( s ) F ) , W F tr = K tr ( s ) ( F , 0 ) .
By Proposition 4.2,
∥ W F t r − U ∥ s , Ω 2 ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ γ Σ K ( s ) F ∥ H 1 / 2 ( Σ ) . \|W_F^{\mathrm{tr}}-U\|_{s,\Omega_2}\le C\sigma^{-1}(1+|s|)^{1/2}|s|\left(1+c_{\max}^2c_{0,R}\sigma^{-1}|s|\right)\|\gamma_{\Sigma}K(s)F\|_{H^{1/2}(\Sigma)}. ∥ W F tr − U ∥ s , Ω 2 ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ γ Σ K ( s ) F ∥ H 1/2 ( Σ ) .
Take now E E E a bounded open neighborhood of Σ \Sigma Σ with d < d ( E , Ω ) d<d(E,\Omega) d < d ( E , Ω ) . Proposition 4.5 with ∣ α ∣ ≤ 1 |\alpha|\le1 ∣ α ∣ ≤ 1 and (54) with Ξ = Σ \Xi=\Sigma Ξ = Σ give
∥ γ Σ K ( s ) F ∥ H 1 / 2 ( Σ ) ≤ C c max 2 σ − 1 e − σ d ∣ s ∣ ( m − 1 ) / 2 ∥ F ∥ L 2 ( Ω ) , \|\gamma_{\Sigma}K(s)F\|_{H^{1/2}(\Sigma)}\le Cc_{\max}^2\sigma^{-1}e^{-\sigma d}|s|^{(m-1)/2}\|F\|_{L^2(\Omega)}, ∥ γ Σ K ( s ) F ∥ H 1/2 ( Σ ) ≤ C c m a x 2 σ − 1 e − σ d ∣ s ∣ ( m − 1 ) /2 ∥ F ∥ L 2 ( Ω ) ,
and (28) follows after absorbing the factor ( 1 + ∣ s ∣ ) 1 / 2 ≤ ( 1 + σ 0 − 1 ) 1 / 2 ∣ s ∣ 1 / 2 (1+|s|)^{1/2}\le(1+\sigma_0^{-1})^{1/2}|s|^{1/2} ( 1 + ∣ s ∣ ) 1/2 ≤ ( 1 + σ 0 − 1 ) 1/2 ∣ s ∣ 1/2 into the constant.
The proof of (23) is very similar, noticing first that
W F t r = K t r ( s ) ( F , 0 ) W_F^{\mathrm{tr}}=K^{\mathrm{tr}}(s)(F,0) W F tr = K tr ( s ) ( F , 0 )
which yields a first bound by Proposition 4.2:
∥ W F t r ∥ s , Ω 2 ≤ C c max 2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) . \|W_F^{\mathrm{tr}}\|_{s,\Omega_2}\le Cc_{\max}^2\sigma^{-1}\|F\|_{L^2(\Omega_2)}. ∥ W F tr ∥ s , Ω 2 ≤ C c m a x 2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) .
Take now Ξ = Γ \Xi=\Gamma Ξ = Γ , E E E a bounded open neighborhood of Γ \Gamma Γ with d ~ = min { d 1 , d 2 } < d ( E , Ω ∪ Ω 2 c ) \widetilde d=\min\{d_1,d_2\}<d(E,\Omega\cup\Omega_2^c) d = min { d 1 , d 2 } < d ( E , Ω ∪ Ω 2 c ) , D = Ω 2 ∩ Ω 2 c D=\Omega_2\cap\Omega_2^c D = Ω 2 ∩ Ω 2 c and notice that W F t r W_F^{\mathrm{tr}} W F tr solves the constant-coefficient Helmholtz equation with complex wave number i s \mathrm{i}s i s in D D D . Then, by (54) and Proposition 4.4 (with d 0 = d ~ d_0=\widetilde d d 0 = d ) we conclude
∥ γ Γ W F t r ∥ H 1 / 2 ( Γ ) ≤ C ( ∣ s ∣ ∥ W F t r ∥ L ∞ ( E ) + ∥ ∇ W F t r ∥ L ∞ ( E ) ) ≤ C ′ e − σ d ~ ∣ s ∣ ( m − 1 ) / 2 ∥ W F t r ∥ s , D ≤ C ′ ′ c max 2 e − σ d ~ ∣ s ∣ ( m − 1 ) / 2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) , \begin{aligned}
\|\gamma_{\Gamma}W_F^{\mathrm{tr}}\|_{H^{1/2}(\Gamma)}
&\le C\left(|s|\|W_F^{\mathrm{tr}}\|_{L^\infty(E)}+\|\nabla W_F^{\mathrm{tr}}\|_{L^\infty(E)}\right)\\
&\le C' e^{-\sigma\widetilde d}|s|^{(m-1)/2}\|W_F^{\mathrm{tr}}\|_{s,D}\\
&\le C''c_{\max}^2e^{-\sigma\widetilde d}|s|^{(m-1)/2}\sigma^{-1}\|F\|_{L^2(\Omega_2)},
\end{aligned} ∥ γ Γ W F tr ∥ H 1/2 ( Γ ) ≤ C ( ∣ s ∣∥ W F tr ∥ L ∞ ( E ) + ∥∇ W F tr ∥ L ∞ ( E ) ) ≤ C ′ e − σ d ∣ s ∣ ( m − 1 ) /2 ∥ W F tr ∥ s , D ≤ C ′′ c m a x 2 e − σ d ∣ s ∣ ( m − 1 ) /2 σ − 1 ∥ F ∥ L 2 ( Ω 2 ) ,
which is (23).
Proof of Theorem 3.3
Bound (30) follows by the argument used to prove (29) , with W F t r W_F^{\mathrm{tr}} W F tr replaced by
V : = K Ω 2 Σ t r ( s ) g Σ = K t r ( s ) ( 0 , g Σ ) , V := K_{\Omega_2\Sigma}^{\mathrm{tr}}(s)g_\Sigma= K^{\mathrm{tr}}(s)(0,g_\Sigma), V := K Ω 2 Σ tr ( s ) g Σ = K tr ( s ) ( 0 , g Σ ) ,
which also solves the constant–coefficient Helmholtz equation in D = Ω 2 ∩ Ω c D = \Omega_2 \cap\Omega^c D = Ω 2 ∩ Ω c and satisfies
∥ V ∥ s , D ≤ ∥ V ∥ s , Ω 2 ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1 / 2 ( Σ ) \|V\|_{s,D} \le\|V\|_{s,\Omega_2} \le C\sigma^{-1}(1+|s|)^{1/2}|s|(1+c_{\max}^2c_{0,R}\sigma^{-1}|s|)\|g_\Sigma\|_{H^{1/2}(\Sigma)} ∥ V ∥ s , D ≤ ∥ V ∥ s , Ω 2 ≤ C σ − 1 ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1/2 ( Σ )
as a consequence of Proposition 4.2. Now, with E E E a bounded open neighborhood of Γ \Gamma Γ with d ~ = min { d 1 , d 2 } < d ( E , Ω ∪ Ω 2 c ) \widetilde d=\min\{d_1,d_2\}<d(E,\Omega\cup\Omega_2^c) d = min { d 1 , d 2 } < d ( E , Ω ∪ Ω 2 c ) we get after applying Proposition 4.4 (with d 0 = d ~ d_0=\widetilde d d 0 = d ) and (66) (with Ξ = Γ \Xi=\Gamma Ξ = Γ )
∥ γ Γ V ∥ H 1 / 2 ( Γ ) ≤ C ( ∣ s ∣ ∥ V ∥ L ∞ ( E ) + ∥ ∇ V ∥ L ∞ ( E ) ) ≤ C ′ e − σ d ~ ∣ s ∣ ( m − 1 ) / 2 ∥ V ∥ s , Ω 2 ≤ C ′ ′ e − σ d ~ ( 1 + ∣ s ∣ ) 1 / 2 ∣ s ∣ ( m + 1 ) / 2 σ − 1 ( 1 + c max 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1 / 2 ( Σ ) . \begin{aligned}
\|\gamma_\Gamma V\|_{H^{1/2}(\Gamma)} \le C\left(|s|\|V\|_{L^\infty(E)}+\|\nabla V\|_{L^\infty(E)}\right) \\
&\le C'e^{-\sigma\widetilde d}|s|^{(m-1)/2}\|V\|_{s,\Omega_2} \\
&\le C''e^{-\sigma\widetilde d}(1+|s|)^{1/2}|s|^{(m+1)/2}\sigma^{-1}(1+c_{\max}^2c_{0,R}\sigma^{-1}|s|)\|g_\Sigma\|_{H^{1/2}(\Sigma)}.
\end{aligned} ∥ γ Γ V ∥ H 1/2 ( Γ ) ≤ C ( ∣ s ∣∥ V ∥ L ∞ ( E ) + ∥∇ V ∥ L ∞ ( E ) ) ≤ C ′ e − σ d ∣ s ∣ ( m − 1 ) /2 ∥ V ∥ s , Ω 2 ≤ C ′′ e − σ d ( 1 + ∣ s ∣ ) 1/2 ∣ s ∣ ( m + 1 ) /2 σ − 1 ( 1 + c m a x 2 c 0 , R σ − 1 ∣ s ∣ ) ∥ g Σ ∥ H 1/2 ( Σ ) .
Absorbing again the factor ( 1 + ∣ s ∣ ) 1 / 2 (1+|s|)^{1/2} ( 1 + ∣ s ∣ ) 1/2 into ∣ s ∣ 1 / 2 |s|^{1/2} ∣ s ∣ 1/2 , this proves (30) .
Finally, to prove (31) we simply take Ξ = Σ ⊂ E ⊂ D : = Ω 1 c \Xi=\Sigma\subset E\subset D:=\Omega_1^c Ξ = Σ ⊂ E ⊂ D := Ω 1 c in (66) , E E E being a bounded neighborhood of Σ \Sigma Σ with d ( E , Γ ) > d 1 d(E,\Gamma)>d_1 d ( E , Γ ) > d 1 so that for
W = K Ω 1 Γ ( s ) g Γ W=K_{\Omega_1\Gamma}(s)g_\Gamma W = K Ω 1 Γ ( s ) g Γ
we can conclude
∥ γ Σ W ∥ H 1 / 2 ( Σ ) ≤ C ′ e − σ d 1 ∣ s ∣ ( m − 1 ) / 2 ∥ W ∥ s , Ω 1 c ≤ C ′ ′ e − σ d 1 ∣ s ∣ ( m + 1 ) / 2 ( 1 + ∣ s ∣ ) 1 / 2 σ − 1 ∥ g Γ ∥ H 1 / 2 ( Γ ) , \begin{aligned}
\|\gamma_\Sigma W\|_{H^{1/2}(\Sigma)} \le C'e^{-\sigma d_1}|s|^{(m-1)/2}\|W\|_{s,\Omega_1^c} \\
&\le C''e^{-\sigma d_1}|s|^{(m+1)/2}(1+|s|)^{1/2}\sigma^{-1}\|g_\Gamma\|_{H^{1/2}(\Gamma)},
\end{aligned} ∥ γ Σ W ∥ H 1/2 ( Σ ) ≤ C ′ e − σ d 1 ∣ s ∣ ( m − 1 ) /2 ∥ W ∥ s , Ω 1 c ≤ C ′′ e − σ d 1 ∣ s ∣ ( m + 1 ) /2 ( 1 + ∣ s ∣ ) 1/2 σ − 1 ∥ g Γ ∥ H 1/2 ( Γ ) ,
where we have applied estimate (34) in the last step. Since ( 1 + ∣ s ∣ ) 1 / 2 ≤ ( 1 + σ 0 − 1 ) 1 / 2 ∣ s ∣ 1 / 2 (1+|s|)^{1/2}\le(1+\sigma_0^{-1})^{1/2}|s|^{1/2} ( 1 + ∣ s ∣ ) 1/2 ≤ ( 1 + σ 0 − 1 ) 1/2 ∣ s ∣ 1/2 , this proves (31) .
Numerical experiments
All computational results in this section were obtained with fully discrete counterparts of Algorithms 1 and 2: the all-steps-at-once Algorithm 1 of Subsection 2.4 organizes the computation, and at each quadrature frequency s ℓ s_\ell s ℓ (by (17) , only half of them need to be computed) the evaluation of K ( s ℓ ) F K(s_\ell)F K ( s ℓ ) F , i.e., the solution of problem (18) , is carried out independently and in parallel by the FEM–BEM coupling of Algorithm 2 of Section 3. We next briefly describe how these components are discretized. (A rigorous convergence analysis of the spatial discretization is beyond the scope of this article, and is left for future work.)
FEM–BEM discretization and coupling
Finite elements. For the heterogeneous bounded domain problem (24) on Ω 2 \Omega_2 Ω 2 we use standard conforming P 2 \mathbb{P}_2 P 2 Lagrange elements. This simple choice is adequate here: for frequencies with ℜ s \Re s ℜ s proportional to ∣ s ∣ |s| ∣ s ∣ , the sesquilinear form is coercive with constant ℜ s / ∣ s ∣ \Re s/|s| ℜ s /∣ s ∣ and hence the Galerkin solution is quasi-optimal with constants independent of s s s .
Since f f f is smooth in time, the weakly damped frequencies case with ∣ s ∣ |s| ∣ s ∣ large, which the mesh cannot resolve, carry a right-hand side F ℓ , r F_{\ell,r} F ℓ , r at round-off level. The mesh resolves ∂ Ω \partial\Omega ∂ Ω and the interfaces between the pieces of Ω \Omega Ω , which is why Ω \Omega Ω is chosen polygonal; c c c is then smooth on each element and c − 2 c^{-2} c − 2 is integrated by element-wise Gaussian quadrature of its analytic expression. The Dirichlet datum on Σ \Sigma Σ is imposed by nodal interpolation at the vertices and edge midpoints lying on Σ \Sigma Σ . The resulting matrices, s 2 M c + S s^2M_c+S s 2 M c + S with M c M_c M c the c − 2 c^{-2} c − 2 -weighted mass matrix and S S S the stiffness matrix, are complex symmetric but not Hermitian, and are solved either with a sparse direct solver or with the multigrid iteration described at the end of this section; in both cases the factorization (or the hierarchy) is computed once per frequency and reused in all the solves at that frequency.
Boundary elements. The exterior problem defining K Ω 1 c Γ ( s ) K_{\Omega^c_1\Gamma}(s) K Ω 1 c Γ ( s ) is solved by an indirect single layer potential representation, W c = SL ( i s ) λ W^c=\operatorname{SL}(\mathrm{i}s)\lambda W c = SL ( i s ) λ , with λ \lambda λ solving the first-kind equation V ( i s ) λ = g Γ V(\mathrm{i}s)\lambda=g_\Gamma V ( i s ) λ = g Γ on Γ \Gamma Γ , where i s \mathrm{i}s i s is the wave number of the associated Helmholtz equation, see [30 , 36 ] ; the strong damping makes more elaborate (combined) formulations unnecessary, and we observed no difference when using them. For numerically solving the first-kind equation, we use the deltaBEM method [17 , 18 ] , a Nyström-type fully discrete Calderón calculus of order three on a uniform grid of n n n points in the parameter domain of Γ \Gamma Γ , which is local in nature, very simple to implement and requires no splitting of the kernel into regular and singular parts. (For the complex frequencies arising in our CQ formulation, such a splitting causes both terms to grow exponentially with ∣ ℜ s ∣ ∣ z ∣ |\Re s||z| ∣ℜ s ∣∣ z ∣ , even though their sum decays exponentially, which leads to severe cancellation.) An alternative in the same spirit is a Galerkin discretization with piecewise polynomials. In either case it is worth noticing that, since the fundamental solution kernels decay like e − σ ∣ x − y ∣ e^{-\sigma|x-y|} e − σ ∣ x − y ∣ , the matrices become numerically banded — entries corresponding to well-separated points fall below any prescribed tolerance — precisely for strongly damped frequencies, that is, in the regime where, by Theorem 3.2 , the correction is least needed: the cost of the boundary term fades at the same rate as its influence on the solution. Conversely, for the few weakly damped frequencies the matrices are dense, but there the correction is essential. Finally, the trace γ Σ W c \gamma_\Sigma W^c γ Σ W c is obtained by evaluating the discrete potential at the Dirichlet nodes of the finite element mesh on Σ \Sigma Σ ; since these lie at distance at least d 1 d_1 d 1 from Γ \Gamma Γ , this is a plain quadrature with a smooth kernel, encoded in a dense matrix assembled once per frequency.
Traces and coupling. The trace γ Γ \gamma_\Gamma γ Γ of a finite element function, needed for f Γ = γ Γ W F t r f_\Gamma=\gamma_\Gamma W_F^{\mathrm{tr}} f Γ = γ Γ W F tr and for the operator K Γ ( s ) K_\Gamma(s) K Γ ( s ) of Remark 3.4 , is computed by evaluating the P 2 \mathbb{P}_2 P 2 function at the n n n deltaBEM points of Γ \Gamma Γ , which lie in the interior of the elements. With these two trace operators, the Schur complement equation (20) becomes a dense n × n n\times n n × n system, which is never assembled: it is solved by GMRES, each iteration requiring one finite element solve, one boundary element solve and the two trace evaluations. The GMRES tolerance is 10 − 8 10^{-8} 1 0 − 8 on the relative residual, and the whole coupling step is skipped whenever ∥ γ Σ K Ω 1 c Γ ( s ) f Γ ∥ ∞ ≤ 10 − 7 ∥ W F t r ∥ ∞ \|\gamma_\Sigma K_{\Omega^c_1\Gamma}(s)f_\Gamma\|_\infty\leq10^{-7}\|W_F^{\mathrm{tr}}\|_\infty ∥ γ Σ K Ω 1 c Γ ( s ) f Γ ∥ ∞ ≤ 1 0 − 7 ∥ W F tr ∥ ∞ , in which case U ∣ Ω 2 U|_{\Omega_2} U ∣ Ω 2 is taken to be W F t r W_F^{\mathrm{tr}} W F tr (see (27) in Algorithm 2).
Numerical example and simulation
In this section we consider the example configuration depicted in Figure 3 , based on the heterogeneous region Ω \Omega Ω . The FEM domain is the rectangle Ω 2 = ( − 5 , 5 ) × ( − 4 , 4 ) \Omega_2=(-5,5)\times(-4,4) Ω 2 = ( − 5 , 5 ) × ( − 4 , 4 ) , so that Σ = ∂ Ω 2 \Sigma=\partial\Omega_2 Σ = ∂ Ω 2 , and the BEM curve Γ \Gamma Γ is the ellipse with semi-axes 2.5 2.5 2.5 and 2 2 2 , rotated by π / 6 \pi/6 π /6 and centered at approximately ( 0.90 , − 0.06 ) (0.90,-0.06) ( 0.90 , − 0.06 ) , whose parameterization is given in (57).
Figure 3. Configuration of the experiment. Left: the wave speed c c c in Ω ‾ 2 = ( − 5 , 5 ) × ( − 4 , 4 ) \overline{\Omega}_2 = (-5,5) \times(-4,4) Ω 2 = ( − 5 , 5 ) × ( − 4 , 4 ) , with the chain-link inhomogeneity Ω \Omega Ω made up of three polygonal pieces. Right: initial finite element mesh, the ellipse Γ \Gamma Γ (red) supporting the boundary element unknowns, and the domains Ω ‾ ⊂ Ω 1 \overline{\Omega} \subset\Omega_1 Ω ⊂ Ω 1 and Ω ‾ 1 ⊂ Ω 2 \overline{\Omega}_1 \subset\Omega_2 Ω 1 ⊂ Ω 2 .
The heterogeneous domain Ω \Omega Ω is the chain-link shaped region shown in Figure 3 : a non-convex domain with re-entrant corners and holes, made up of three polygonal pieces Ω ( 1 ) \Omega^{(1)} Ω ( 1 ) (left link), Ω ( 2 ) \Omega^{(2)} Ω ( 2 ) (central overlap) and Ω ( 3 ) \Omega^{(3)} Ω ( 3 ) (right link). The wave speed is c ≡ 1 c \equiv1 c ≡ 1 outside Ω ‾ \overline{\Omega} Ω , whereas inside each piece it is a smooth, oscillatory perturbation of the background,
c ( x , y ) = 1 + ∑ j = 1 3 f j ( x , y ) χ Ω ( j ) ( x , y ) , f 1 ( x , y ) = 1 2 ( 2 + cos ( 4 2 ( x + y ) ) ) , f 2 ( x , y ) = 1 2 ( 1 + cos ( 12 y ) ) , f 3 ( x , y ) = 1 2 ( 1 + cos ( 4 x ) ) , (67) \begin{aligned}
c(x,y) &= 1 + \sum_{j=1}^{3} f_j(x,y)\,\chi_{\Omega^{(j)}}(x,y), \\
f_1(x,y) &= \frac{1}{2}\left(2+\cos\left(4\sqrt{2}(x+y)\right)\right), \\
f_2(x,y) &= \frac{1}{2}\left(1+\cos(12y)\right), \\
f_3(x,y) &= \frac{1}{2}\left(1+\cos(4x)\right), \tag*{(67)}
\end{aligned} c ( x , y ) f 1 ( x , y ) f 2 ( x , y ) f 3 ( x , y ) = 1 + j = 1 ∑ 3 f j ( x , y ) χ Ω ( j ) ( x , y ) , = 2 1 ( 2 + cos ( 4 2 ( x + y ) ) ) , = 2 1 ( 1 + cos ( 12 y ) ) , = 2 1 ( 1 + cos ( 4 x ) ) , (67)
where χ Ω ( j ) \chi_{\Omega^{(j)}} χ Ω ( j ) denotes the characteristic function of Ω ( j ) \Omega^{(j)} Ω ( j ) . Hence c c c takes values in [ 1 , 2 ] [1,2] [ 1 , 2 ] on Ω ( 1 ) \Omega^{(1)} Ω ( 1 ) and Ω ( 3 ) \Omega^{(3)} Ω ( 3 ) , and in [ 3 / 2 , 5 / 2 ] [3/2,5/2] [ 3/2 , 5/2 ] on Ω ( 2 ) \Omega^{(2)} Ω ( 2 ) , so that c min = 1 c_{\min}=1 c m i n = 1 and c max = 5 / 2 c_{\max}=5/2 c m a x = 5/2 . Notice that c c c is discontinuous across ∂ Ω \partial\Omega ∂ Ω (except at isolated points) and across the interfaces between the three pieces, which produces a right-hand side F F F with jumps, as anticipated in § 5 . The three oscillation directions ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 1 , 1 ) / 2 (1,1)/\sqrt{2} ( 1 , 1 ) / 2 and ( 0 , 1 ) (0,1) ( 0 , 1 ) and the increasing frequencies 4 4 4 , 4 2 4\sqrt{2} 4 2 and 12 12 12 are chosen to avoid any accidental alignment with the mesh.
The problem is driven by an incident plane pulse propagating in the background medium, where c ≡ 1 c \equiv1 c ≡ 1 , with unit speed:
u i n c ( x , t ) = exp ( − ( t − ( x − x 0 ) ⋅ d ) 2 4 ς 2 ) cos ( ω ( t − ( x − x 0 ) ⋅ d ) ) , (68) u^{\mathrm{inc}}(x,t)=\exp\left(-\frac{\left(t-(x-x_0)\cdot\mathbf{d}\right)^2}{4\varsigma^2}\right)\cos\left(\omega\left(t-(x-x_0)\cdot\mathbf{d}\right)\right), \tag*{(68)} u inc ( x , t ) = exp ( − 4 ς 2 ( t − ( x − x 0 ) ⋅ d ) 2 ) cos ( ω ( t − ( x − x 0 ) ⋅ d ) ) , (68)
x 0 = ( − 3 , 0 ) , d = ( 1 , − 0.2 ) 1.04 , ς = 0.2 , ω = π , x_0=(-3,0),\qquad\mathbf{d}=\frac{(1,-0.2)}{\sqrt{1.04}},\qquad\varsigma=0.2,\qquad\omega=\pi, x 0 = ( − 3 , 0 ) , d = 1.04 ( 1 , − 0.2 ) , ς = 0.2 , ω = π ,
that is, a Gaussian-modulated wave front of width of order ς \varsigma ς and carrier frequency ω \omega ω , traveling in the direction d \mathbf{d} d . At t = 0 t=0 t = 0 the pulse is centered on the line through x 0 x_0 x 0 orthogonal to d \mathbf{d} d , at distance about 2 from Ω ‾ \overline{\Omega} Ω , so that its trace on the inhomogeneity is of order e − 25 e^{-25} e − 25 and the homogeneous initial conditions in (1) hold up to round-off. We choose the final time for our simulation to be T = 8 T=8 T = 8 .
The initial mesh, generated with Gmsh [24 ] , consists of 1146 triangles and 2365 nodes. Finer meshes are obtained by uniform refinements, so that the meshes form a nested hierarchy; this is what makes multigrid solvers applicable, a point to which we return below. This initial mesh is fitted to the material interfaces, which comes at no cost here since the interfaces are polygonal, and it preserves the full order of the finite element approximation under refinement. For curved interfaces some accuracy is lost, less so if the mesh follows the shape of the inhomogeneity with straight-sided elements, as in [15 ] , than if the interface is ignored by the mesh. The BEM unknowns live on the ellipse with parameterization p \mathbf{p} p :
Γ : p ( x ) = ( cos π 6 − sin π 6 sin π 6 cos π 6 ) ( 0.75 + 2.5 cos x − 0.5 + 2 sin x ) , x ∈ [ 0 , 2 π ) , (69) \Gamma:\quad\mathbf{p}(x)=
\begin{pmatrix}
\cos\frac{\pi}{6} & -\sin\frac{\pi}{6}\\
\sin\frac{\pi}{6} & \cos\frac{\pi}{6}
\end{pmatrix}
\begin{pmatrix}
0.75+2.5\cos x\\
-0.5+2\sin x
\end{pmatrix},\qquad x\in[0,2\pi), \tag*{(69)} Γ : p ( x ) = ( cos 6 π sin 6 π − sin 6 π cos 6 π ) ( 0.75 + 2.5 cos x − 0.5 + 2 sin x ) , x ∈ [ 0 , 2 π ) , (69)
which is discretized with n = 300 n=300 n = 300 points.
Error measurement metrics. Since no closed-form solution of the above example is available, the role of the exact solution for error measurements is played by a reference solution u r e f u^{\mathrm{ref}} u ref computed with the same scheme and a finer discretization of the parameter under study, the others being kept fixed: the time step κ = 8 / 3072 \kappa=8/3072 κ = 8/3072 (n C Q = 3072 n_{\mathrm{CQ}}=3072 n CQ = 3072 ), half that of the finest discretization considered, n C Q = 192 , 384 , 768 , 1536 n_{\mathrm{CQ}}=192, 384, 768, 1536 n CQ = 192 , 384 , 768 , 1536 , when the convergence in time is examined; and one further refinement of the finite element mesh or of the boundary element discretization when the convergence in space is studied. For each time level t j = j κ t_j=j\kappa t j = j κ of the numerical solution under study we compute the relative error
( E κ ( t j ) ) 2 : = ∑ ℓ ∣ u κ ( x ℓ , t j ) − u r e f ( x ℓ , t j ) ∣ 2 ∑ ℓ ∣ u r e f ( x ℓ , t j ) ∣ 2 ≈ ∫ Ω 2 ∣ u κ ( x , t j ) − u r e f ( x , t j ) ∣ 2 d x ∫ Ω 2 ∣ u r e f ( x , t j ) ∣ 2 d x , \left(E^\kappa(t_j)\right)^2:=\frac{\sum_{\ell}\left|u^\kappa(\mathbf{x}_{\ell},t_j)-u^{\mathrm{ref}}(\mathbf{x}_{\ell},t_j)\right|^2}{\sum_{\ell}\left|u^{\mathrm{ref}}(\mathbf{x}_{\ell},t_j)\right|^2}\approx\frac{\int_{\Omega_2}\left|u^\kappa(\mathbf{x},t_j)-u^{\mathrm{ref}}(\mathbf{x},t_j)\right|^2\,\mathrm{d}x}{\int_{\Omega_2}\left|u^{\mathrm{ref}}(\mathbf{x},t_j)\right|^2\,\mathrm{d}x}, ( E κ ( t j ) ) 2 := ∑ ℓ ∣ u ref ( x ℓ , t j ) ∣ 2 ∑ ℓ u κ ( x ℓ , t j ) − u ref ( x ℓ , t j ) 2 ≈ ∫ Ω 2 ∣ u ref ( x , t j ) ∣ 2 d x ∫ Ω 2 u κ ( x , t j ) − u ref ( x , t j ) 2 d x ,
where the sums run over the vertices and edge midpoints of the mesh of u r e f u^{\mathrm{ref}} u ref , on which the coarser finite element solution, being piecewise polynomial, is evaluated exactly; this is a discrete relative L 2 ( Ω 2 ) L^2(\Omega_2) L 2 ( Ω 2 ) error at time t j t_j t j . Here u u u denotes the total wave; note that the error itself is the same for the total and for the scattered field, since the incident wave is known exactly, and the total wave is used only for normalization purposes because its norm stays away from zero for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
In addition to E κ ( t ) E^\kappa(t) E κ ( t ) we report its maximum in time
E ∞ κ : = max j E κ ( t j ) . (70) E^\kappa_\infty:=\max_j E^\kappa(t_j). \tag*{(70)} E ∞ κ := j max E κ ( t j ) . (70)
Results. We have run the experiment with the BDF2- and the trapezoidal-rule-based CQ schemes for n C Q = 192 , 384 , 768 , 1536 , n_{\mathrm{CQ}}=192, 384, 768, 1536, n CQ = 192 , 384 , 768 , 1536 , 1 i.e. κ = 1 / 24 , 1 / 48 , 1 / 96 , 1 / 192 \kappa=1/24,1/48,1/96,1/192 κ = 1/24 , 1/48 , 1/96 , 1/192 , with continuous P 2 P_2 P 2 finite elements on the second and fourth uniform refinements of the initial mesh (36 961 and 587 905 degrees of freedom, respectively) and 300 boundary element unknowns on Γ \Gamma Γ . Figure 4 shows the total field computed with the trapezoidal rule and n C Q = 3072 n_{\mathrm{CQ}}=3072 n CQ = 3072 , the discretization used below as reference solution, at four instants, and Figure 5 the corresponding scattered field.
Figure 4. Total field u = u i n c + u s c u = u^{\mathrm{inc}} + u^{\mathrm{sc}} u = u inc + u sc at t = 1.92 t = 1.92 t = 1.92 , 3.29 3.29 3.29 , 5.79 5.79 5.79 and 8.00 8.00 8.00 , computed on the FEM domain Ω 2 = ( − 5 , 5 ) × ( − 4 , 4 ) \Omega_2 = (-5,5) \times(-4,4) Ω 2 = ( − 5 , 5 ) × ( − 4 , 4 ) , whose boundary Σ \Sigma Σ is the frame of each panel; the boundary of the inhomogeneity Ω \Omega Ω is drawn in gray. The color scale is the same in all panels and is saturated at ± 1.7 \pm1.7 ± 1.7 . Since c > 1 c > 1 c > 1 inside Ω \Omega Ω , the pulse travels faster through the inhomogeneity than in the background, and the wave front is visibly advanced after crossing it (t = 5.79 t = 5.79 t = 5.79 ). Note the absence of spurious reflections at Σ \Sigma Σ , where the exact exterior condition is enforced through the BEM.
Figure 5. Scattered field u s c u^{\mathrm{sc}} u sc at the same times as in Figure 4. The color scale is again common to all panels, now saturated at ± 0.6 \pm0.6 ± 0.6 (the amplitude of u s c u^{\mathrm{sc}} u sc is considerably smaller than that of the incident pulse); saturated regions mark the concentration of the scattered energy inside Ω \Omega Ω and at its right tip. The waves generated by the multiple internal reflections in the chain-link geometry leave the computational domain through Σ \Sigma Σ without any visible artifact.
Figure 6 shows E κ ( t ) E^\kappa(t) E κ ( t ) for both schemes, with the same BEM discretization (single layer formulation with 300 degrees of freedom) and two levels of refinement of the finite element space. The curves are practically indistinguishable, which indicates that the time error does not depend on the spatial resolution. For illustrative purposes we also depict the error at the final time T = 8 T = 8 T = 8 in Figure 7 .
Figure 6. Relative error E κ ( t ) E^{\kappa}(t) E κ ( t ) at the time steps t j = j κ t_j=j\kappa t j = j κ for the BDF2- (left) and trapezoidal-rule-based (right) CQ schemes, n C Q = 192 , 384 , 768 , 1536 n_{\mathrm{CQ}}=192,\,384,\,768,\,1536 n CQ = 192 , 384 , 768 , 1536 ; the reference solution uses the same scheme with n C Q = 3072 n_{\mathrm{CQ}}=3072 n CQ = 3072 . Top: a continuous P 2 \mathbb{P}_2 P 2 finite element space on the second uniform refinement of the initial mesh, 36 961 nodes (vertices and edge midpoints) on 18 336 triangles. Bottom: two further uniform refinements, 587 905 nodes on 293 376 triangles.
Figure 7. (a) BDF2. (b) Trapezoidal rule. Error u κ − u r e f u^{\kappa}-u^{\mathrm{ref}} u κ − u ref at t = 8 t=8 t = 8 for n C Q = 192 n_{\mathrm{CQ}}=192 n CQ = 192 , 384, 768, 1536 (left to right), each against a reference solution of the same scheme with n C Q = 3072 n_{\mathrm{CQ}}=3072 n CQ = 3072 . Note the different color scales: the error is roughly divided by four at each refinement while its spatial pattern is unchanged. It concentrates at the incident front, which is leaving Ω 2 \Omega_2 Ω 2 through Σ \Sigma Σ at this time.
The maximum in time of the relative error E ∞ κ E_\infty^\kappa E ∞ κ (see (70) ) decays from 0.31 0.31 0.31 to 4.4 × 10 − 3 4.4 \times10^{-3} 4.4 × 1 0 − 3 for BDF2 and from 0.1 0.1 0.1 to 1.2 × 10 − 3 1.2 \times10^{-3} 1.2 × 1 0 − 3 for the trapezoidal rule, with estimated orders of convergence log 2 ( E ∞ κ / E ∞ κ / 2 ) \log_2(E_\infty^\kappa/E_\infty^{\kappa/2}) log 2 ( E ∞ κ / E ∞ κ /2 ) of 1.6 1.6 1.6 , 2.3 2.3 2.3 , 2.1 2.1 2.1 and 2.0 2.0 2.0 , 2.1 2.1 2.1 , 2.3 2.3 2.3 , respectively. The last value of each sequence is slightly biased upwards, since the reference solution has only half the time step of the finest discretization.
Both schemes exhibit the second order rate of convergence predicted by Theorem 2.2, but they do so in different ways. The trapezoidal rule is in the asymptotic regime from the coarsest step onwards, whereas BDF2 shows a reduced order between the two coarsest levels (EOC ≈ 1.6 \approx1.6 ≈ 1.6 ) and reaches second order only for κ ≤ 1 / 48 \kappa\le1/48 κ ≤ 1/48 . The latter is a pre-asymptotic effect: a relative error of 0.31 0.31 0.31 is close to saturation, and the error can no longer scale with κ 2 \kappa^2 κ 2 at that level.
The comparison between the two schemes is otherwise governed by the phase error, which for a mode of frequency ω \omega ω accumulates at the rate κ 2 ω 3 / 12 \kappa^2\omega^3/12 κ 2 ω 3 /12 per unit time for the trapezoidal rule and κ 2 ω 3 / 3 \kappa^2\omega^3/3 κ 2 ω 3 /3 for BDF2, a factor of four [6 Sect. 4.13] ; accordingly, the errors of the trapezoidal rule are 3–4 times smaller than those of BDF2 at equal κ \kappa κ throughout the whole range, or, equivalently, the trapezoidal rule with step κ \kappa κ matches
BDF2 with step κ / 2 \kappa/2 κ /2 , that is, with twice as many frequency problems. This is consistent with the recommendation of [6 Sect. 4.13] in favor of the trapezoidal rule on accuracy grounds.
We also consider the influence of the spatial discretization, keeping n C Q = 768 n_{\mathrm{CQ}}=768 n CQ = 768 fixed. Figure 8 shows E κ ( t ) E^{\kappa}(t) E κ ( t ) for an increasing number of degrees of freedom on Γ \Gamma Γ , the finite element space being fixed. Up to t ≈ 4 t\approx4 t ≈ 4 the scattered field has not yet reached Γ \Gamma Γ , the coupling step is skipped at almost all frequencies, and the curves merely reflect the round-off level of the scaling by r r r ; the boundary element discretization only comes into play afterward. From then on the error at the final time decreases rapidly, by one to two orders of magnitude each time the number of degrees of freedom is doubled, from about 0.5 with 10 to about 10 − 4 10^{-4} 1 0 − 4 with 80, and then levels off around 10 − 5 10^{-5} 1 0 − 5 for 160 and 320. This plateau is three orders of magnitude below the time discretization error at this n C Q n_{\mathrm{CQ}} n CQ and may be caused by the tolerances of the coupling step rather than by the boundary element resolution. With 320 degrees of freedom the boundary element contribution therefore plays no role in the results reported above.
Figure 8. Relative error E κ ( t ) E^{\kappa}(t) E κ ( t ) at the time steps t j = j κ t_j = j\kappa t j = j κ for the BDF2- (left) and trapezoidal-rule-based (right) CQ schemes with n C Q = 768 n_{\mathrm{CQ}} = 768 n CQ = 768 and a P 2 \mathrm{P}_2 P 2 finite element space with 36 961 nodes on 18 336 triangles (second uniform refinement). The deltaBEM solver uses 10, 20, 40, 80, 160 and 320 degrees of freedom, the reference solution being computed with 640.
Keeping instead the boundary discretization fixed and refining the finite element space (Figure 9 ), the estimated orders of convergence are about 3. This is the L 2 L^2 L 2 rate of P 2 \mathrm{P}_2 P 2 elements for piecewise H 3 H^3 H 3 solutions, and it is the rate one expects here: the mesh resolves ∂ Ω \partial\Omega ∂ Ω and the interfaces between the pieces of Ω \Omega Ω , so that the jumps of c c c and F F F across them do not affect the elementwise interpolation error, while the corners of these curves, where
Figure 9. Relative error E κ ( t ) E^{\kappa}(t) E κ ( t ) at the time steps t j = j κ t_j = j\kappa t j = j κ for the BDF2- (left) and trapezoidal-rule-based (right) CQ schemes with n C Q = 768 n_{\mathrm{CQ}} = 768 n CQ = 768 and 320 degrees of freedom in the deltaBEM solver. The error is computed from the initial mesh (2365 nodes on 1146 triangles) up to its third uniform refinement, the reference solution being computed with the same discretization parameters on the next refinement (587 905 nodes on 293 376 triangles).
Δ U \Delta U Δ U is discontinuous, and the right-angle corners of Σ \Sigma Σ [26 Section 5.1] only reduce the regularity to H 3 − ε ( Ω 2 ) H^{3-\varepsilon}(\Omega_2) H 3 − ε ( Ω 2 ) , piecewise, for every ε > 0 \varepsilon>0 ε > 0 .
Iterative solvers for the Schur complement equation. The number of GMRES iterations for the Schur complement equation (26) is usually one or two, as Figure 10 shows for two representative runs, and we never observed more than four in any of our experiments; at most frequencies the coupling step is skipped altogether. This is better than what Theorem 3.3 guarantees, since the contraction estimate (32) does not hold in the weakly damped window where the coupling is active, see Remark 3.4 . As mentioned above, the coupling is activated whenever γ Σ K Ω c 1 Γ ( s ) f Γ \gamma_\Sigma K_{\Omega_c1\Gamma}(s)f_\Gamma γ Σ K Ω c 1Γ ( s ) f Γ , the first iterate of (26) , exceeds 10 − 7 ∥ W t r r ∥ ∞ 10^{-7}\lVert W_{\mathrm{tr}}^r\rVert_\infty 1 0 − 7 ∥ W tr r ∥ ∞ . By (29) and (31) this quantity decays like e − σ ( d 1 + min { d 1 , d 2 } ) e^{-\sigma(d_1+\min\{d_1,d_2\})} e − σ ( d 1 + m i n { d 1 , d 2 }) up to polynomial factors, so that the coupling is needed only for ℜ s ≤ σ ⋆ \Re s\leq\sigma^\star ℜ s ≤ σ ⋆ , with σ ⋆ ≈ 3.3 \sigma^\star\approx3.3 σ ⋆ ≈ 3.3 fixed by the geometry and the tolerance. On the other hand, the scaling radius r = ε 1 / ( 2 n C Q ) r=\varepsilon^{1/(2n_{\mathrm{CQ}})} r = ε 1/ ( 2 n CQ ) keeps every frequency at a distance from the imaginary axis that is independent of κ \kappa κ . Indeed, for the three methods considered ℜ δ ( r e i θ ) \Re\delta(re^{\mathrm{i}\theta}) ℜ δ ( r e i θ ) attains its minimum over θ \theta θ at θ = 0 \theta=0 θ = 0 , so that ℜ s ℓ ≥ δ ( r ) / κ = : σ 0 \Re s_\ell\geq\delta(r)/\kappa=:\sigma_0 ℜ s ℓ ≥ δ ( r ) / κ =: σ 0 for all ℓ \ell ℓ , and since δ ( e − x ) = x + O ( x 2 ) \delta(e^{-x})=x+O(x^2) δ ( e − x ) = x + O ( x 2 ) ,
Figure 10. Frequencies s ℓ = δ ( r e ℓ , n C Q ) / κ s_{\ell}=\delta(r e_{\ell,n_{\mathrm{CQ}}})/\kappa s ℓ = δ ( r e ℓ , n CQ ) / κ used by the all-steps-at-once solver for n C Q = 384 n_{\mathrm{CQ}}=384 n CQ = 384 and n C Q = 1536 n_{\mathrm{CQ}}=1536 n CQ = 1536 , marked according to the solver applied: pure FEM (squares) or FEM–BEM coupling with one or two GMRES iterations (crosses). Right: detail of the region σ 0 ≤ ℜ s ≤ σ ∗ \sigma_0\leq\Re s\leq\sigma^* σ 0 ≤ ℜ s ≤ σ ∗ , where the coupling is active. Fourth uniform refinement of the initial mesh, quadratic continuous finite elements with 587 905 nodes (vertices and midpoints) on 293 376 triangles, and 300 dof on the BEM solver set on Γ \Gamma Γ .
σ 0 = δ ( r ) κ ≈ ∣ ln ε ∣ 2 n C Q ⋅ n C Q T = ∣ ln ε ∣ 2 T ≈ 2.25. \sigma_0=\frac{\delta(r)}{\kappa}\approx\frac{|\ln\varepsilon|}{2n_{\mathrm{CQ}}}\cdot\frac{n_{\mathrm{CQ}}}{T}=\frac{|\ln\varepsilon|}{2T}\approx2.25. σ 0 = κ δ ( r ) ≈ 2 n CQ ∣ ln ε ∣ ⋅ T n CQ = 2 T ∣ ln ε ∣ ≈ 2.25.
The BEM is therefore needed only in the fixed window σ 0 ≤ ℜ s ≤ σ ⋆ \sigma_0\leq\Re s\leq\sigma^\star σ 0 ≤ ℜ s ≤ σ ⋆ , marked by the dashed lines in the right panels of Figure 10 , and the proportion of pure FEM problems is 80%–88% for BDF2 and, much more steadily, 57%–60% for the trapezoidal rule.
The frequencies that actually contribute to the solution can be identified from the transformed data. Writing the quadrature node as r e ℓ , n C Q = r e i θ ℓ re_{\ell,n_{\mathrm{CQ}}}=re^{i\theta_\ell} r e ℓ , n CQ = r e i θ ℓ , θ ℓ = 2 π ℓ / ( n C Q + 1 ) \theta_\ell=2\pi\ell/(n_{\mathrm{CQ}}+1) θ ℓ = 2 π ℓ / ( n CQ + 1 ) , and using r ≈ e − κ σ 0 r\approx e^{-\kappa\sigma_0} r ≈ e − κ σ 0 , step 2 of Algorithm 1 reads
F ℓ , r κ = 1 n C Q + 1 ∑ j = 0 n C Q f ( j κ ) e − ( σ 0 − i τ ℓ ) j κ ≈ 1 T ∫ 0 T f ( t ) e − σ 0 t e i τ ℓ t d t , τ ℓ : = θ ℓ κ = 2 π ℓ T + κ ≈ 2 π ℓ T . (71) \begin{aligned}
F_{\ell,r}^{\kappa}=\frac{1}{n_{\mathrm{CQ}}+1}\sum_{j=0}^{n_{\mathrm{CQ}}}f(j\kappa)e^{-(\sigma_0-i\tau_\ell)j\kappa} \\
&\approx\frac{1}{T}\int_0^T f(t)e^{-\sigma_0t}e^{i\tau_\ell t}\,\mathrm{d}t,\qquad\tau_\ell:=\frac{\theta_\ell}{\kappa}=\frac{2\pi\ell}{T+\kappa}\approx\frac{2\pi\ell}{T}. \tag*{(71)}
\end{aligned} F ℓ , r κ = n CQ + 1 1 j = 0 ∑ n CQ f ( j κ ) e − ( σ 0 − i τ ℓ ) j κ ≈ T 1 ∫ 0 T f ( t ) e − σ 0 t e i τ ℓ t d t , τ ℓ := κ θ ℓ = T + κ 2 π ℓ ≈ T 2 π ℓ . (71)
that is, heuristically and up to the factor 1 / T 1/T 1/ T , the Fourier transform of the damped datum at the frequencies τ ℓ \tau_\ell τ ℓ of a Fourier series on [ 0 , T ] [0,T] [ 0 , T ] , whatever the multistep method. For the pulse (56) this transform is, up to smooth factors, the sum of two Gaussians e − ς 2 ( τ ∓ ω ) 2 e^{-\varsigma^2(\tau\mp\omega)^2} e − ς 2 ( τ ∓ ω ) 2 of width 1 / ς = 5 1/\varsigma=5 1/ ς = 5 centered at ± ω \pm\omega ± ω , which falls below 10 − 13 10^{-13} 1 0 − 13 of its maximum for ∣ τ ∣ ≳ 30 |\tau|\gtrsim30 ∣ τ ∣ ≳ 30 . Hence only the nodes with τ ℓ ≲ 30 \tau_\ell\lesssim30 τ ℓ ≲ 30 , about 30 T / π ≈ 75 30T/\pi\approx75 30 T / π ≈ 75 of them regardless of n C Q n_{\mathrm{CQ}} n CQ , carry a nonnegligible right-hand side; refining the time step only adds nodes with data at round-off level. For these nodes δ ( r e ℓ , n C Q ) ≈ κ σ 0 − i θ ℓ \delta(re_{\ell,n_{\mathrm{CQ}}})\approx\kappa\sigma_0-i\theta_\ell δ ( r e ℓ , n CQ ) ≈ κ σ 0 − i θ ℓ , so that s ℓ ≈ σ 0 − i τ ℓ s_\ell\approx\sigma_0-i\tau_\ell s ℓ ≈ σ 0 − i τ ℓ (their conjugates give the other half of the contour): they are the least damped frequencies of the contour, the ones for which the coupling is needed and, as discussed next, the ones for which the interior solver is hardest. All the others, however many, are cheap.
Iterative solvers for the FEM problems. Most of the FEM problems are in fact solved for completeness of the FFT step only: for the pulse (56), F ℓ , r κ F_{\ell,r}^{\kappa} F ℓ , r κ is smaller than 10 − 13 10^{-13} 1 0 − 13 relative to its maximum for ∣ ℑ s ∣ ≳ 30 |\Im s|\gtrsim30 ∣ℑ s ∣ ≳ 30 , as discussed in (59), and these frequencies could be skipped altogether by setting Y ℓ , r κ = 0 Y_{\ell,r}^{\kappa}=0 Y ℓ , r κ = 0 . Moreover, for such strongly damped frequencies the solution is close to c 2 F / s 2 c^2F/s^2 c 2 F / s 2 , a rescaled copy of the datum, up to boundary layers of width O ( 1 / ∣ s ∣ ) O(1/|s|) O ( 1/∣ s ∣ ) along the interfaces where F F F jumps; the mesh does not resolve these layers, but they carry a negligible part of an already negligible solution and do not propagate.
Regarding the FEM solver, and motivated by the damping present in the complex wave number, we have implemented a multigrid solver based on V-cycles with Gauss–Seidel smoothing: one pre-smoothing step, no post-smoothing, undamped coarse-grid correction. The depth of the V-cycle is therefore limited by the number of levels available in each mesh hierarchy. Let us stress that designing a multigrid solver for this problem is not among the aims of this paper: a naive implementation suffices for our purposes, and what follows is merely an exploration of the terrain, intended to show that iterative solvers are a realistic option here and to identify the regimes in which they can be expected to work. Its interest is not speed — on the present two-dimensional problem it is comparable to a sparse direct solver — but memory: no fill-in is stored, only the matrices of the mesh hierarchy. In three dimensions, where sparse direct solvers cease to be affordable, some such alternative is not merely convenient but unavoidable. It should also be emphasized that we are in a far more comfortable situation than that of the genuine Helmholtz equation: the wave number k = i s k=is k = i s has a positive imaginary part bounded away from zero, uniformly in κ \kappa κ , and it is this damping that makes a plain V-cycle work at all for a sizable part of the frequencies.
Figure 11 shows where the V-cycle converged and with how many levels. Two features are worth pointing out. First, the number of levels that can be used decreases as ∣ s ∣ |s| ∣ s ∣ grows, one level being lost roughly each time ∣ s ∣ |s| ∣ s ∣ doubles, which is consistent with the coarsest mesh of the hierarchy having to resolve the length scale 1 / ∣ s ∣ 1/|s| 1/∣ s ∣ ; the thresholds are the same for both schemes, as one would expect from a mechanism that depends on the spatial mesh alone. Second, for sufficiently large ∣ s ∣ |s| ∣ s ∣ and insufficient relative damping the smoother fails, and a sparse direct solver is used instead. This is consistent with the analysis of the shifted Helmholtz operator introduced in [21 ] : a complex shift of size O ( k 2 ) O(k^2) O ( k 2 ) , that is, of relative size O ( 1 ) O(1) O ( 1 ) , is required for this operator to be invertible by multigrid [11 ] , whereas a shift of size at most O ( k ) O(k) O ( k ) is needed for it to be a good preconditioner for the unshifted problem [22 ] . In our setting the second requirement plays no role, since the shifted problem is the one to be solved. In terms of the Laplace parameter, k 2 = − s 2 k^2=-s^2 k 2 = − s 2 and the relative size of the shift is ∣ ℑ k 2 ∣ / ∣ ℜ k 2 ∣ ≈ 2 ℜ s / ∣ s ∣ |\Im k^2|/|\Re k^2| \approx2\Re s/|s| ∣ℑ k 2 ∣/∣ℜ k 2 ∣ ≈ 2ℜ s /∣ s ∣ , so that the threshold observed here corresponds to a shift of the order predicted there.
The proportion of frequencies for which the smoother fails grows as the time step decreases — from 0% to 48% for BDF2 between n C Q = 384 n_{\mathrm{CQ}}=384 n CQ = 384 and 1536 — and is consistently larger for the trapezoidal rule, 63% and 78% for the same two values, whose contour accumulates a large number of poorly damped frequencies. The two contours differ where it matters for the smoother. The relative damping σ / ∣ s ∣ = ℜ δ ( z ) / ∣ δ ( z ) ∣ \sigma/|s|=\Re\delta(z)/|\delta(z)| σ /∣ s ∣ = ℜ δ ( z ) /∣ δ ( z ) ∣ depends on the quadrature node z = r e i θ z=re^{\mathrm{i}\theta} z = r e i θ alone, θ = 2 π ℓ / ( n Q + 1 ) \theta=2\pi\ell/(n_{\mathrm{Q}}+1) θ = 2 π ℓ / ( n Q + 1 ) , and near θ = 0 \theta=0 θ = 0 it is ≈ ( 1 − r ) / θ \approx(1-r)/\theta ≈ ( 1 − r ) / θ for both schemes, with 1 − r ≈ κ σ 0 1-r\approx\kappa\sigma_0 1 − r ≈ κ σ 0 forced by round-off: in the band of frequencies that carry the data the two schemes are equally hard, and both resort there to the direct solver or to few levels. Beyond that band they part ways. For BDF2,
σ ∣ s ∣ ≈ 1 θ ( ( 1 − r ) + 1 4 θ 4 ) , \frac{\sigma}{|s|}\approx\frac{1}{\theta}\left((1-r)+\frac{1}{4}\theta^4\right), ∣ s ∣ σ ≈ θ 1 ( ( 1 − r ) + 4 1 θ 4 ) ,
which reaches a minimum of order ( 1 − r ) 3 / 4 (1-r)^{3/4} ( 1 − r ) 3/4 and then grows to O ( 1 ) O(1) O ( 1 ) : the weakly damped frequencies form a narrow band, and the remaining failures, at large ∣ s ∣ |s| ∣ s ∣ , are due to the coarsest meshes not resolving the length scale 1 / ∣ s ∣ 1/|s| 1/∣ s ∣ , an inverse CFL condition h ≲ κ h\lesssim\kappa h ≲ κ that constrains the robustness of the iterative solver but not the stability of the scheme. For the trapezoidal rule, instead,
σ ∣ s ∣ ≈ 1 − r sin θ , \frac{\sigma}{|s|}\approx\frac{1-r}{\sin\theta}, ∣ s ∣ σ ≈ sin θ 1 − r ,
which stays below 2 ( 1 − r ) 2(1-r) 2 ( 1 − r ) on two thirds of the nodes. A small relative damping can still be compensated by resolution, since the coarse-grid correction of the near-resonant modes requires the coarsest mesh to resolve the wavelength to a degree that grows as the damping decreases [11 ] , but for the trapezoidal rule this would have to hold on most of the contour, where ∣ s ∣ |s| ∣ s ∣ is already of order κ − 1 \kappa^{-1} κ − 1 while the relative damping is only of order κ \kappa κ , which is not a practical option; see Figure 11 (b). An iterative interior solver for this scheme would rather require smoothers adapted to the weakly damped Helmholtz regime. This is the situation of the classical Helmholtz equation and remains a challenge.
Figure 11. Frequencies s ℓ = δ ( r e ℓ , n C Q ) / κ s_\ell= \delta(r e_\ell,n_{\mathrm{CQ}})/\kappa s ℓ = δ ( r e ℓ , n CQ ) / κ used by the all-steps-at-once solver for n C Q = 384 n_{\mathrm{CQ}} = 384 n CQ = 384 and n C Q = 1536 n_{\mathrm{CQ}} = 1536 n CQ = 1536 , marked according to the number of levels of the V-cycle that converged, the cross denoting the frequencies for which the multigrid solver failed and was replaced by a sparse direct solver. Right: detail of the region of small σ = ℜ s \sigma= \Re s σ = ℜ s . Same discretization as in Figure 10 .
Table 1 gives a detailed account of the frequency problems behind the time-convergence experiment of Figure 6 . The second column is the number of independent frequencies generated by the quadrature once conjugate symmetry is taken into account, n C Q / 2 + 1 n_{\mathrm{CQ}}/2+1 n CQ /2 + 1 , which grows linearly with the number of time steps. The third column shows that only 40 of them were actually solved, in all runs and for both schemes: these are the frequencies with the largest data, all with ∣ ℑ s s ∣ ≤ 30 |\Im s_s|\le30 ∣ℑ s s ∣ ≤ 30 , and the remaining ones, whose data lie below 10 − 11 10^{-11} 1 0 − 11 of the maximum, were set to zero. Since the convergence rates of Figure 6 are unaffected, the cost of the solver phase does not depend on n C Q n_{\mathrm{CQ}} n CQ ; what grows with the number of time steps is only the FFT and the bookkeeping of the discarded frequencies. The fourth column counts the problems in which the BEM correction was activated.
Table 1. Frequency problems in the time-convergence experiment, for each scheme and time step: independent frequencies after conjugate symmetry (n C Q / 2 + 1 n_{\mathrm{CQ}}/2 + 1 n CQ /2 + 1 ); problems actually solved, the remaining ones carrying data below 10 − 11 10^{-11} 1 0 − 11 of the maximum and being set to zero; problems in which the BEM correction was activated; problems in which the multigrid V-cycle converged with four, three, two or one coarse levels (Gauss–Seidel smoothing, LU on the coarsest level after a first failure at full depth); and problems that required a direct solver on the finest mesh.
For the trapezoidal rule this happens in all 40 problems, which lie within the window σ 0 ≤ ℜ s ≤ σ ⋆ \sigma_0 \le\Re s \le\sigma^\star σ 0 ≤ ℜ s ≤ σ ⋆ of Figure 10 ; for BDF2 and the two coarsest time steps some of them fall outside it, since for large κ \kappa κ the nodes carrying the data spread over a wider arc of the contour, where the BDF2 damping is larger, and are solved as pure FEM problems. In every case the coupled system was solved with two GMRES iterations. The last columns describe the interior solver. All 40 problems were solved by the multigrid V-cycle, at full depth or after reducing the number of coarse levels, and the direct solver on the finest mesh was never needed. The distribution of depths is the same for all time steps except the coarsest one, for the reason discussed below.
Conclusions
We have proposed and analyzed a fully CQ method for efficient simulation of transient acoustic waves in a heterogeneous and unbounded medium. Multistep CQ applied to the solution operator reduces the problem to independent Helmholtz–type problems with complex frequencies, each of them solved by an overlapping FEM–BEM coupling in which the finite element domain and the boundary element curve are chosen independently of the geometry of the inhomogeneity. The analysis of the time discretization rests on bounds for the solution operators of the transformed problems that are explicit in the Laplace parameter and in the wave speed; the exponential decay of the transformed solutions away from the inhomogeneity shows, in turn, that the boundary element correction is needed only in a fixed window of weakly damped frequencies, independent of the time step, and that the coupled operator is a contraction outside it.
The numerical experiments confirm the second order in time of the BDF2 and trapezoidal schemes, with the errors in time and in space decoupled from each other, and show that the coupled equation is solved by GMRES in very few iterations, usually one or two, at every frequency where the coupling is active, a behavior that our contraction estimate does not yet explain. They also show that a plain multigrid V–cycle solves the interior problems for a large part of the frequencies, failing only where the relative damping is small, in agreement with the known theory for the shifted Helmholtz operator.
Several questions remain open. The analysis of the spatial discretization, with constants explicit in the Laplace parameter, is the natural next step; the regularity of the transformed solutions on polygonal domains and the s s s –dependence of the lifting operators on Σ \Sigma Σ are the main difficulties. Higher order in time calls for Runge–Kutta convolution quadrature, which fits the same all–steps–at–once framework. In three dimensions, where the method is most attractive, the interior solver becomes the bottleneck, and the design of smoothers adapted to the weakly damped frequencies deserves further study. Finally, the extension to the acoustic model with variable density, where the interior operator takes divergence form while the coupling is unaffected, is left for future work.
Declaration of generative AI and AI-assisted technologies in the writing process
During the preparation of this work the first author (vD) used Claude (Anthropic), mainly in the debugging and documentation of the MATLAB code used in the numerical experiments, for checking some of the computations and, to a lesser extent, revising parts of the text. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.
Acknowledgements
The first author (vD) is partially supported by the project “Adquisición de conocimiento y minería de datos, funciones especiales y métodos numéricos avanzados” of Universidad Pública de Navarra. The second author (MG) is supported by the Simons Foundation through Grant No. 518882. vD is deeply grateful to Francisco J. Gaspar for several enlightening discussions on multigrid methods for wave problems, which greatly improved his understanding of this iterative method in this context and led to a sound implementation in the present setting, and for his generous and selfless help.