Mod´elisation Math´ematique et Analyse Num´erique Vol. 35, No3, 2001, pp. 389–405
ON THE CONVERGENCE OF A LINEAR TWO-STEP FINITE ELEMENT METHOD FOR THE NONLINEAR SCHR ¨ODINGER EQUATION∗
Georgios E. Zouraris
1Abstract. We discretize the nonlinear Schr¨odinger equation, with Dirichlet boundary conditions, by a linearly implicit two-step finite element method which conserves theL2 norm. We prove optimal ordera priorierror estimates in theL2 andH1norms, under mild mesh conditions for two and three space dimensions.
Mathematics Subject Classification. 65M12, 65M60.
Received: January 14, 2000.
1. Introduction
1.1. The i.b.v. problemLetd= 1,2 or 3, and Ω⊂Rdbe a bounded domain with smooth boundary∂Ω. Fort∗>0, we setI∗:= [0, t∗] and consider the following initial and boundary value problem for the nonlinear Schr¨odinger equation:
wt= i∆w+ if(|w|2)w in Ω×I∗,
w= 0 on∂Ω×I∗, (1.1)
w(x,0) =w0(x) forx∈Ω,
where w0 : Ω−→C and f ∈ C3([0,+∞);R) are given functions. The nonlinear Schr¨odinger equation, often, appears as a model in mathematical physics (see [1, 7, 8, 16, 22, 26]); for more information on the theory and applications we refer to [5, 6, 10, 14, 15, 17, 19, 21, 25] and the references therein. In the sequel, we will assume that problem (1.1) admits a unique solution which is sufficiently smooth for our purposes.
1.2. Notation and preliminaries
For integer s ∈ N0, we denote by Hs(Ω) the Sobolev space consisting of complex-valued functions which, along with their distributional derivatives of order up tos, are inL2(Ω), and byk · ksthe corresponding norm.
The inner product onL2(Ω) =H0(Ω) is denoted by (·,·) and the associated norm byk · k. H01(Ω) consists of
Keywords and phrases. Nonlinear Schr¨odinger equation, two-step time discretization, linearly implicit method, finite element method,L2andH1error estimates, optimal order of convergence.
∗Work supported by the EU-TMR project HCL # ERBFMRXCT960033 and the EU-HCM scheme Reaction Diffusion Equa- tions # ERBCHRXCT930409.
1Department of Numerical Analysis and Computing Science (NADA), Royal Institute of Technology (KTH), 10044 Stockholm, Sweden.
Present address: Centre de Recherche en Math´ematiques de la D´ecision (CEREMADE), UMR CNRS 7534, Universit´e de Paris IX-Dauphine, Place du Mar´echal de Lattre-de-Tassigny, 75775 Paris Cedex 16, France.
c EDP Sciences, SMAI 2001
the functions ofH1(Ω) that vanish at∂Ω in the sense of trace. In addition, form∈Nand foru,v∈(L2(Ω))m we use (u, v) :=Pm
j=1(uj, vj) and kuk:={(u, u)}1/2. Finally, we setH:=H01(Ω)∩C(Ω), and denote by| · |p the standard norm onLp(Ω) forp≥2 (i.e. | · |2=k · k).
We mention the Poincar´e–Friedrichs inequality
kvk ≤ CΩk∇vk, ∀v∈H01(Ω), (1.2) and the Sobolev–type inequalities
|v|∞≤ C1,∞k∇vk21kvk12, ∀v∈H01(Ω), d= 1, (1.3) and
|v|s≤ C2,sk∇vks−s2kvk2s, ∀s∈[2,+∞), ∀v∈H01(Ω), d= 2. (1.4) For an integerr≥2, let {Sh}h∈(0,1)be a family of finite dimensional subspaces ofHsatisfying
χinf∈Sh
kv−χk+hkv−χk1 ≤C hskvks, ∀v∈Hs(Ω)∩H01(Ω), s= 2, . . . , r, ∀h∈(0,1).
Forh∈(0,1), we define the discrete Laplacian ∆h:Sh−→Sh by
(∆hϕ, χ) =−(∇ϕ,∇χ), ∀ϕ, χ∈Sh, an elliptic projection operatorRh:H1(Ω)−→Shby
(∇Rhv,∇χ) = (∇v,∇χ), ∀v∈H1(Ω), ∀χ∈Sh,
and finallyPh will be theL2–projection operator ontoSh. The elliptic projectionRh has the following approx- imation property (cf.,e.g., [23])
kRhv−vk+hkRhv−vk1≤CRhskvks, ∀v∈Hs(Ω)∩H01(Ω), s= 2, . . . , r, ∀h∈(0,1), (1.5) and obviously satisfies
k∇Rhvk ≤ k∇vk, ∀v∈H1(Ω), ∀h∈(0,1). (1.6) We will say thatf has the property (D), if there exists %≥1 such that
|f0(x2)x| ≤CD 1 +x%−1
, ∀x >0. (1.7)
1.3. The numerical method
Leth∈(0,1),N ∈N,k:= tN∗, tn:=nk forn= 0, . . . , N, andt1/2:=t0+k2. Let, also,Wh0∈Sh be a given approximation ofw0. First, we construct an approximationWh1/2∈Shofw(·, t1/2) by
Wh1/2−Wh0 (k/2) = i∆h
Wh1/2+Wh0 2
! + iPh
"
f(|Wh0|2) Wh1/2+Wh0 2
!#
. (1.8a)
Then, form= 1, . . . , N, we define an approximationWhm∈Sh tow(·, tm) recursively by Wh1−Wh0
k = i∆h
Wh1+Wh0 2
+ iPh
f(|Wh1/2|2)
Wh1+Wh0 2
(1.8b)
and
Whn−Whn−2 2k = i∆h
Whn+Whn−2 2
+ iPh
f(|Whn−1|2)
Whn+Whn−2 2
, n= 2, . . . , N. (1.8c) Remark 1.1. By dropping (1.8a) and settingWh1/2 =Wh0 in (1.8b), an alternative method is obtained. For this method, our analysis also applies, butWh1 is a suboptimal order approximation ofw(·, t1) in theH1 norm.
This is the reason that the “fractional” step (1.8a) has been introduced.
Remark 1.2. Leth∈(0,1). For givenϕ∈Sh andλ >0, we defineTh(λ, ϕ;·) :Sh−→Sh by Th(λ, ϕ;χ) :=χ−ikλ∆hχ−iλkPh(f(|ϕ|2)χ) for χ∈Sh.
To ensure that the method (1.8) is well-defined i.e. the existence and uniqueness of Wh1/2 and {Whn}Nn=1
, it is enough to show that Th(λ, φ;·) is invertible. Let ψ ∈ Sh be such that Th(λ, ϕ;ψ) = 0. Then, we have Re(Th(λ, ϕ;ψ), ψ) = 0, or Re
hkψk2+ iλkk∇ψk2−iλk(f(|ϕ|2),|ψ|2) i
=kψk2 = 0, which yieldsψ= 0. Hence, we conclude thatTh(λ, ϕ;·) is one–to–one. SinceTh(λ, ϕ;·) is linear and the spaceSh has finite dimension, the fact that it is one-to-one yields its invertibility.
Remark 1.3. Taking theL2 inner product of (1.8a) withWh1/2+Wh0, of (1.8b) withWh1+Wh0 and of (1.8c) withWhn+Whn−2, and then real parts, we conclude thatkWh1/2k=kWh0kandkWh`k=kWh0kfor `= 0, . . . , N. Thus the method (1.8) isL2 conservative.
1.4. Main results and relations to previous work
The time discretization in (1.8c) in conjuction with a finite difference method for the space discretization is proposed in [9] for the numerical approximation of a nonlinear Schr¨odinger equation in one space dimension and with periodic boundary conditions. An optimal order error bound ofO(k2+h2) in a discreteL2 norm is, also, given, only in the case of a cubic Schr¨odinger equation, wheref(x) =λxandλ∈R. This convergence result is based on the fact that the method conserves a discrete Hamiltonian which ford= 1 yields boundedness of the numerical approximations in the discrete L∞ norm by a constant which is independent of the partition of the time and space intervals (cf. (5) in [9] and Rem. 2.22 in Sect. 2.5). In the case of a general nonlinearity this conservation property fails and thus a different technique is needed to prove convergence.
The paper at hand is devoted to the convergence analysis of the method (1.8). We prove an optimal order error bound of O(k2+hr) in the L2 norm and of O(k2+hr−1) in the H1 norm, whithout conditions when d= 1, and under the following mild mesh conditions
p|ln(h)|hr−1≤C2,a and p
|ln(h)|(k32 +hr)≤C2,b when d= 2, (1.9) or
hr−1≤C3,a
ph and k32 +hr≤C3,b
ph when d= 3; (1.10)
here C2,a, C2,b, C3,a and C3,b are constants which depend only on the solution and the data, and h is the minimum of the diameter of the elements of the partition of Ω over which the finite element space is constructed (see Th. 2.14, Th. 2.15 and Rem. 2.16). Also, if d= 2 andf satisfies (D), then we prove an optimal order error bound O(k2+hr) in theL2 norm and a suboptimal one ofO(k32 +hr−1+k−12hr) in theH1 norm, provided that
k2+hr≤C√
k, (1.11)
where C is a constant which depends only on the solution and the data (see Th. 2.21).
Usually, the analysis of numerical methods for the nonlinear Schr¨odinger equation is based on inverse in- equalities between norms on the finite element spaces for quasiuniform or local quasiuniform partitions of Ω (see, e.g., [2, 3, 11–13, 18, 20, 24]). Here, when d= 1, ord= 2 and f has a polynomial growth satisfying (D), we obtain, for the method (1.8), optimal order of convergence in the L2 norm, avoiding inverse inequalities or assumptions on the finite element spaces (ase.g. theH1-boundedness assumption of theL2-projectionPh used in [12]) besides those of Section 1.2. However, for general f and d = 2 or 3, we need an inverse inequality between the L∞and theH1 norm, and thereforehappears in (1.9) and (1.10).
To arrive at the mesh conditions (1.9), (1.10) or (1.11), we prove convergence estimates in the L2 and H1 norm, for the approximations generated by amodified schemewhich is a nonlinear perturbation of (1.8) at the linearized term (see (Λ) and (Υ) in Sect. 2.3). Then the mesh conditions, exhibited above, are introduced to ensure that the modified approximations are bounded in theL∞ or in theH1 norm, by a constant independent of the discretization parameters. Having this boundness property the modified scheme coincides with (1.8) and hence the convergence estimates for it hold also for (1.8). The analysis here has been inspired from the works [12] and [24], but there the methods under consideration, the techniques used and the results obtained are different.
The analysis and the results of the paper extend, easily, to the method obtained substituting (1.8c) by Whn−Whn−2
2k = i∆h
Whn+Whn−2 2
+ iPh
h
f(|Whn−1|2)Whn−1i
which is a nonconservative implicit-explicit method and, as (1.8), yields only one linear system of algebraic equations at every time level, but the matrix remains unchanged.
An overview of the paper is as follows. Section 2 is divided in five parts. In Section 2.1, we prove some function inequalities often used in the convergence analysis, and in Section 2.2 present a consistency result for the time discretization. Section 2.3 contains the definition of the modified schemes and Section 2.4 the convergence theorems for a general function f. Finally, in Section 2.5 we investigate the special case where d= 2 andf satisfies the property (D).
2. Convergence analysis
2.1. Function inequalitiesWe present here some function inequalities that we will often use later.
Lemma 2.1. Foru1, u2∈C(Ω)andg∈C1([0,+∞);R), we have kg(|u1|2)−g(|u2|2)k ≤ sup
x∈I(u1,u2)|g0(x)|(|u1|∞+|u2|∞)ku1−u2k (2.1) with I(u1, u2) :=
0,max{|u1|2∞,|u2|2∞} .
Lemma 2.2. Foru1,u2,w1,w2∈C(Ω)andg∈C2(R;R), we have kg(u1)−g(u2)−g(w1) +g(w2)k ≤ sup
|x|∈I1(u1,u2,w1,w2)|g00(x)| |w1−w2|∞ ku1−w1k+ku2−w2k
+ sup
|x|∈I2(u1,u2)
|g0(x)| k(u1−u2)−(w1−w2)k, (2.2)
with I1(u1, u2, w1, w2) :=
0,max{|u1|∞+|u2|∞,|w1|∞+|w2|∞}
andI2(u1, u2) :=
0,|u1|∞+|u2|∞ .
Proof. (2.2) follows directly from the following expansion g(u1)−g(u2)−g(w1) +g(w2) = (w1−w2)
Z 1 0
g0((1−τ)u2+τ u1)−g0((1−τ)w2+τ w1) dτ +
(u1−u2)−(w1−w2) Z 1
0
g0((1−τ)u2+τ u1)dτ.
Lemma 2.3. Foru1,u2,w1,w2∈C(Ω), we have
|u1|2− |u2|2− |w1|2+|w2|2 ≤ 2|w1−w2|∞ku2−w2k
+B(u1, u2, w1, w2)k(u1−w1)−(u2−w2)k, (2.3) with B(u1, u2, w1, w2) := max{|u1|∞+|w1|∞,|u2|∞+|w2|∞}.
Proof. We obtain (2.3) observing that
(|u1|2− |w1|2)−(|u2|2− |w2|2) = Ren
(u1−w1)−(u2−w2)
(u1+w1) +(u2−w2)
(u1−w1)−(u2−w2) +2(u2−w2)(w1−w2)
o .
Lemma 2.4. Let F :C2−→Cbe defined byF(z, ω) :=fe(|z|)ω where fe(x) :=f(x2) for x∈R. If d= 2andf satisfies(1.7), then, for u1,u2,w1,w2∈C(Ω)withu1−w1,u2−w2∈H01(Ω), we have
kF(u1, u2)−F(w1, w2)k ≤C ku1−w1k+ku2−w2kh 1 +
X2 j=1
|wj|%∞+k∇(uj−wj)k%i
. (2.4)
Proof. Letu1,u2, w1,w2∈C(Ω) withu1−w1,u2−w2∈H01(Ω). Observing that
|F(z1, z2)−F(ω1, ω2)| ≤C |z1−ω1|+|z2−ω2|h
1 + |ω1|+|ω2|+|z1−ω1|+|z2−ω2|%i forz1,z2, ω1,ω2∈C, and using (1.4), we obtain
kF(u1, u2)−F(w1, w2)k ≤ChZ
Ω
1 + (|w1|+|w2|)2%
|u1−w1|+|u2−w2|2
dx +
Z
Ω
|u1−w1|+|u2−w2|2%+2
dxi1/2
≤C h
1 +|w1|2%∞+|w2|2%∞
ku1−w1k2+ku2−w2k2
+ku1−w1k2k∇(u1−w1)k2%+ku2−w2k2k∇(u2−w2)k2%i1/2
which yields (2.4).
2.2. Consistency
We continue by presenting a consistency result concerning the time discretization.
Forn= 0, . . . , N, we defineσn∈L2(Ω) by w`∗n−wjn∗
λnk = i
2∆(w`∗n+wjn∗) + i
2f(|wi∗n|2)(w`∗n+wj∗n) +σn, (2.5) where w1/2 :=w(·, t1/2), wm := w(·, tm) for m = 0, . . . , N, λ0 := 1/2, λ1 := 1, λm := 2 form = 2, . . . , N,
`∗0 := 1/2,`∗m:=mfor m= 1, . . . , N, i∗0 := 0,i∗1:= 1/2,i∗m:=m−1 form= 2, . . . , N, andj0∗:= 0, j1∗ := 0, jm∗ :=m−2 for m= 2, . . . , N. Then, using Taylor expansions, we arrive at
1≤maxn≤Nkσnk+kkσ0k ≤Ck2 and max
4≤n≤Nkσn−σn−2k ≤Ck3. (2.6) 2.3. Modified Schemes
Modifying properly the linearized term in the numerical method (1.8), we construct two modified schemes, (Λ) and (Υ), which we will use later in the convergence analysis. A modified scheme is connected to a real parameterδ >0 and a given norm ofH, and it is not a numerical method. When the approximations that the scheme furnishes are bounded in that norm byδ, then they coincide with those that (1.8) produces provided, of course, that the initial approximation is the same. Even that the original method (1.8) is linear, the modified scheme will be nonlinear. Hence, we cannot ensure the existence of the modified approximations following the argument of Remark 1.2. For this reason, we shall employ the following Brouwer-type fixed-point lemma, for a proof of which we refer to [3].
Lemma 2.5. Let (X,(·,·)X) be a finite-dimensional inner product space and k · kX the associated norm. Let µ: X −→X be continuous and assume that there exists α >0such that for every z∈ X with kzkX =αthere holdsRe(µ(z), z)X ≥0. Then, there exists az∗∈ X such thatµ(z∗) = 0 andkz∗kX ≤α.
•Modified Scheme (Λ): Letδ >supt∈I∗|w(·, t)|∞ andgδ be an increasingC2(R;R) function, with bounded derivatives up to second order, satisfying
gδ(x) :=
x, if |x| ≤δ
2|xx|δ, if |x|>2δ and gδ(x)∈
[δ,2δ] if x∈[δ,2δ]
[−2δ,−δ] if x∈[−2δ,−δ] for x∈R. Then, we define a functionγδ :C−→Cbyγδ(z) =gδ(Rez) + igδ(Imz).
Forh∈(0,1) andm= 0,1/2,1, . . . , N, let Λmδ,h∈Shbe specified inductively by
Λ0δ,h=Rhw0 (2.7a)
and
Λ`δ,h∗m −Λjδ,h∗m λmk = i∆h
Λ`δ,h∗m + Λjδ,h∗m 2
! + iPh
"
f(|γδ(Λiδ,h∗m)|2)γδ
Λ`δ,h∗m + Λjδ,hm∗ 2
!#
(2.7b) form= 0, . . . , N.
Remark 2.6. It is easily seen thatγδ(w(·, τ)) =w(·, τ) andγδ(w(·,τ1)+w(2 ·,τ2)) =w(·,τ1)+w(2 ·,τ2) forτ, τ1, τ2∈I∗, wherewis the solution of (1.1). Thus, the consistency argument for the method (1.8) (cf. Sect. 2.2) holds also for (2.7).
Remark 2.7. Assuming that |Λ0δ,h|∞ ≤ δ and |Λ`δ,h∗m|∞ ≤ δ for m = 0, . . . , N, we obtain f(|γδ(Λiδ,h∗m)|2) = f(|Λiδ,h∗m|2) and γδ
Λ`δ,h∗m+Λjδ,h∗m 2
= Λ
`∗ m δ,h+Λjδ,hm∗
2 for m = 0, . . . , N. Hence, if Wh0 = Rhw0, then Λ`δ,h∗m = Wh`∗m for m= 0, . . . , N.
We ensure the existence of a Λ`δ,h∗m ∈Sh which solves the nonlinear system in (2.7b), by an argument based on Lemma 2.5. In particular, letm∈ {0, . . . , N}, h∈(0,1), (X,(·,·)X) = (Sh,(·,·)) and Πmδ,h:X −→X be an operator given by
Πmδ,h(χ) :=χ−Λjδ,hm∗ − i
2λmk∆hχ− i 2λmkPh
f(|γδ(Λiδ,h∗m)|2)γδ(χ)
; Πmδ,h is continuous, sinceγδ is continuous andShhas finite dimension. Then, we obtain
Re(Πmδ,h(χ), χ)≥ kχk kχk − kΛjδ,h∗mk −λmkBδ
, ∀χ∈Sh,
where Bδ := √
2δ|Ω|1/2supx∈[0,8δ2]|f(x)| and|Ω| is the area of Ω. Therefore, we have Re(Πmδ,h(χ), χ)>0 for everyχ ∈ Sh with kχk= 1 +kΛjδ,hm∗k+λmkBδ. Applying Lemma 2.5 with µ = Πmδ,h, we conclude that there exists aχmδ,h∈Sh such that Πmδ,h(χmδ,h) = 0. Thus Λ`δ,h∗m = 2χmδ,h−Λjδ,h∗m is a solution of (2.7b).
In Remark 2.7, we explained that when the approximations produced by (Λ) are bounded in theL∞norm by δandWh0=Rhw0, then they are the numerical approximations of the method (1.8). Next, we present another modified scheme that has this property for any normν onH, instead of theL∞ one.
•Modified Scheme (Υ): Letν be a norm onH,δ >supt∈I∗ν(w(·, t)) be a given constant, andξδ :R−→Rbe a continuous function defined by
ξδ(x) :=
1, if x≤2δ
−xδ + 3, if x∈[2δ,3δ]
0, if x >3δ
for x∈R. Then, fort∈I∗, we define a mapgν,δ(t;·) :H−→Hby
gν,δ(t;ω) :=ω ξδ(ν(ω−w(·, t))) + w(·, t) (1−ξδ(ν(ω−w(·, t)))) for ω∈ H, wherewis, always, the solution of problem (1.1).
Forh∈(0,1) andm= 0,1/2,1, . . . , N, we specify functions Υmδ,h∈Sh, inductively by
Υ0δ,h=Rhw0 (2.8a)
and
Υ`δ,h∗m −Υjδ,hm∗ λmk = i∆h
Υ`δ,h∗m+ Υjδ,h∗m 2
! + iPh
"
f(|gν,δ(ti∗m; Υiδ,h∗m)|2)gν,δ(t`∗m; Υ`δ,h∗m) +gν,δ(tjm∗; Υjδ,hm∗) 2
#
(2.8b) form= 0, . . . , N.
Remark 2.8. Sincegν,δ(t;w(·, t)) =w(·, t) fort∈I∗, the consistency argument for (2.8) is the same with that for (1.8) (cf.Sect. 2.2).
Remark 2.9. Assuming thatν(Υ0δ,h) ≤δand ν(Υ`δ,h∗m) ≤δ form = 0, . . . , N, we have ν(Υ0δ,h−w(·, t))<2δ and ν(Υ`δ,h∗m −w(·, t))≤2δ form= 0, . . . , N and t∈I∗. Thus, we obtaingν,δ(t; Υ`δ,h∗m) = Υ`δ,h∗m form= 0, . . . , N andt∈I∗. In that case, ifWh0=Rhw0, then a simple induction argument yields Υ`δ,h∗m =Wh`∗mform= 0, . . . , N.
Lemma 2.10. Let ν be a norm on Handδ >supt∈I∗ν(w(·, t)). Then, fort∈I∗ andω∈ H, we have
ν(gν,δ(t;ω))≤4δ. (2.9)
Proof. Lett∈I∗ andω ∈ H. Ifν(ω−w(·, t))≥3δ thenν(gν,δ(t;ω)) =ν(w(·, t))≤δ. We assume, now, that ν(ω−w(·, t))<3δ. Then we haveν(ω)≤4δ, and
ν(gν,δ(t;ω)) ≤ ν(ω)ξδ(ν(ω−w(·, t))) +ν(w(·, t)) (1−ξδ(ν(ω−w(·, t))))
≤ 3δ ξδ(ν(ω−w(·, t))) +δ
≤ 4δ.
Lemma 2.11. Let ν be a norm onHand δ >supt∈I∗ν(w(·, t)). Fort∈I∗, the operatorgν,δ(t;·) :H−→His continuous on (H, ν).
Proof. Fort∈I∗ andω1,ω2∈ H, we have ν gν,δ(t;ω1)−gν,δ(t;ω2)
≤ν(ω1−ω2)+
ν(w(·, t)) +ν(ω2) ξδ(ν(ω1−w(·, t)))−ξδ(ν(ω2−w(·, t))) which, together with the continuity ofξδ, yields the continuity ofgν,δ(t;·) on (H, ν).
As for (Λ), we discuss, now, the existence of a solution for the nonlinear system in (2.8b) following a similar argument. Letm∈ {0, . . . , N},h∈(0,1), (X,(·,·)X) = (Sh,(·,·)) and Φmδ,h:X −→X given by
Φmδ,h(χ) :=χ−Υjδ,hm∗ − i
2λmk∆hχ− i 4λmkPh
h
f |gν,δ(ti∗m; Υiδ,h∗m)|2
gν,δ(t`∗m; 2χ−Υjδ,hm∗) +gν,δ(tjm∗; Υjδ,hm∗)i ,
which is continuous, since gν,δ(t`∗m;·) is continuous (cf. Lem. 2.11) and Sh has finite dimension. Also, in Sh
the norms ν and | · |∞ are equivalent, and hence there exists an h–dependent constant Ch,∞,ν, such that
|χ|∞≤Ch,∞,νν(χ) forχ∈Sh. Using (2.9), we arrive at the following general estimates
|gν,δ(τ;ω)|∞≤Be1,h,δ:= 4δCh,∞,ν, ∀τ∈I∗, ∀ω∈ H, and
f |gν,δ(τ1;ω1)|2 gν,δ(τ2;ω2) +gν,δ(τ3;ω3)≤Be2,h,δ:= 2Be1,h,δ
p|Ω| sup
[0,(Be1,h,δ)2]
|f|,
forτ1, τ2, τ3∈I∗ andω1, ω2, ω3∈ H. Thus, we obtain
Re(Φmδ,h(χ), χ)≥ kχk kχk − kΥjδ,h∗mk − λm4kBe2,h,δ
, ∀χ∈Sh.
Therefore, we have Re(Φmδ,h(χ), χ)>0 for everyχ∈Shwithkχk= 1 +kΥjδ,h∗mk+λm4kBe2,h,δ, and by Lemma 2.5, withµ= Φmδ,h, we conclude the existence of aχmδ,h∈Sh such that Φmδ,h(χmδ,h) = 0. Finally, Υ`δ,h∗m = 2χmδ,h−Υjδ,h∗m is a solution of (2.8b).
2.4. L2- and H1-convergence
Next we will prove the following optimal order error estimates for the method (1.8):
0≤maxm≤NkWhm−wmk ≤ CA(k2+hr) and max
0≤m≤Nk∇(Whm−wm)k ≤ CA(k2+hr−1) (2.10) forh∈(0, h0) andN ≥N0. Hereh0∈(0,1),N0∈NandCA is a constant independent ofhandN.
The basic step in the convergence analysis is the estimation of the differencesϑ1/2Λ,δ,h := Λ1/2δ,h −Rhw1/2 and ϑnΛ,δ,h:= Λnδ,h−Rhwn,n= 0, . . . , N, where due to (2.7a), we haveϑ0Λ,δ,h= 0.
Proposition 2.12. Let d= 1,2or 3, andδ >supt∈I∗|w(·, t)|∞. Then, there existsNδ ∈Nsuch that kϑ1/2Λ,δ,hk ≤CA,δ(k2+khr) and max
1≤m≤NkϑmΛ,δ,hk ≤CA,δ(k2+hr), ∀h∈(0,1), ∀N ≥Nδ. (2.11) The constantCA,δ is independent of handN, but depends onδ, the solution or its derivatives, and the data.
Proof. Leth∈(0,1). Form= 0, . . . , N, we defineσmΛ,δ,h∈Sh by ϑ`Λ,δ,h∗m −ϑjΛ,δ,h∗m
λmk = i∆h
ϑ`Λ,δ,h∗m +ϑjΛ,δ,hm∗ 2
+σmΛ,δ,h. (2.12)
Combining (2.12), (2.5), Remark 2.6, and (1.8), we obtain
σmΛ,δ,h=−Ph(σm+ωmw) + iPhωmΛ,δ, m= 0, . . . , N, (2.13) with
ωmw :=Rh
w`∗m−wjm∗ λmk
−w`∗m−wjm∗ λmk
(2.14) and
ωmΛ,δ:=f(|γδ(Λiδ,h∗m)|2)γδ
Λ`∗m δ,h+Λjδ,hm∗
2
−f(|wi∗m|2)γδ
w`∗m+wj∗m 2
. (2.15)
Taking real parts of theL2inner product of (2.12) withϑ`h∗m+ϑjh∗m and using (2.13), form= 0, . . . , N, we have kϑ`Λ,δ,h∗m k2− kϑjΛ,δ,h∗m k2=−λmkh
Re(σm+ωmw, ϑ`Λ,δ,h∗m +ϑjΛ,δ,hm∗ ) + Im(ωΛ,δm , ϑ`Λ,δ,h∗m +ϑjΛ,δ,hm∗ )i
. (2.16)
By (1.5) and the Taylor formula, we conclude that
0≤maxm≤Nkωwmk ≤Chr. (2.17)
Letm∈ {0, . . . , N}. To estimateωmΛ,δ in theL2 norm, we split it as follows:
ωmΛ,δ=ωΛ,δ,1m +ωmΛ,δ,2
with ωΛ,δ,1m :=
f(|γδ(Λiδ,h∗m)|2)−f(|wi∗m|2) γδ
Λ`∗m δ,h+Λjδ,h∗m
2
and ωΛ,δ,2m :=f(|wi∗m|2) h
γδ
Λ`∗m δ,h+Λjδ,h∗m
2
−γδ
w`∗m+wjm∗
2
i .
Using (2.1), the mean value theorem and (1.5), we obtain kωΛ,δ,1m k ≤ C δ(1 +δ) sup
x∈[0,8δ2]|f0(x)| kγδ(Λiδ,h∗m)−γδ(wi∗m)k
≤ Cδ sup
|x|∈[0,2δ]
|gδ0(x)| hr+kϑiΛ,δ,h∗m k ,
and
kωΛ,δ,2m k ≤C sup
|x|∈[0,2δ]|g0δ(x)| hr+kϑ`Λ,δ,h∗m k+kϑjΛ,δ,h∗m k .
Thus, we get the following estimate
kωmΛ,δk ≤Cδ(hr+kϑ`Λ,δ,h∗m k+kϑjΛ,δ,hm∗ k+kϑiΛ,δ,h∗m k). (2.18) From (2.16), (2.17), (2.6) and (2.18), it follows that
kϑ1/2Λ,δ,hk+kϑ1Λ,δ,hk ≤Ceδk(kϑ1/2Λ,δ,hk+kϑ1Λ,δ,hk) +Cδ(k2+khr) (2.19a) and
(1−Ceδk)(kϑmΛ,δ,hk+kϑmΛ,δ,h−1k)≤(1 +Ceδk)(kϑmΛ,δ,h−1k+kϑmΛ,δ,h−2k) +Cδk(k2+hr), m= 2, . . . , N. (2.19b) Finally, assuming thatCeδk≤13 and applying a discrete Gr¨onwall argument on (2.19a-b), we arrive at (2.11).
Proposition 2.13. Let d= 1,2or 3, and δ >supt∈I∗|w(·, t)|∞. Then there exists an integerNeδ ≥Nδ such that
k∇ϑ1/2Λ,δ,hk ≤ CA,δ(k32 +k12hr) and max
1≤m≤Nk∇ϑmΛ,δ,hk ≤ CA,δ(k2+hr), ∀h∈(0,1), ∀N≥Neδ. (2.20) The constantCA,δ is independent ofhandN, but depends onδ, the solution or its derivatives, and the data.
Proof. Let h∈ (0,1) and N ≥ Nδ (cf. Prop. 2.12). For m = 0, . . . , N, we define σΛ,δ,hm , ωwm and ωmΛ,δ as in (2.12), (2.14) and (2.15), respectively.
Taking the L2-inner product of (2.12) byϑ`Λ,δ,h∗m −ϑjΛ,δ,h∗m and then imaginary parts, we obtain
k∇ϑ1/2Λ,δ,hk2= 2Im(σΛ,δ,h0 , ϑ1/2Λ,δ,h), k∇ϑ1Λ,δ,hk2= 2Im(σΛ,δ,h1 , ϑ1Λ,δ,h), (2.21) and
k∇ϑmΛ,δ,hk2+k∇ϑmΛ,δ,h−1k2=k∇ϑmΛ,δ,h−1k2+k∇ϑmΛ,δ,h−2k2+ 2Im(σmΛ,δ,h, ϑmΛ,δ,h−ϑmΛ,δ,h−2), m= 2, . . . , N. (2.22)