Abstract
This paper studies the Cauchy problem for first order systems,
Assume that for \(\xi \in \mathbb {R}^d,\) \(\sum A_j\xi _j\) has only real eigenvalues. For coefficients and Cauchy data sufficiently Gevrey regular the Cauchy problem has a unique sufficiently Gevrey regular solution. We prove stability and error estimates for the spectral Crank-Nicholson scheme. Approximate solutions can be computed with accuracy \(\epsilon \) in \(L^\infty ( [0,T]\times \mathbb {R}^d)\) with cost growing at most polynomially in \(\epsilon ^{-1}\). The proofs uses pseudodifferential symmetrizers.
We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.
Avoid common mistakes on your manuscript.
1 Introduction
1.1 Hyperbolic background
Consider the Cauchy problem (0.1). The coefficients \(A_j\) and B are \(m\times m\) complex matrix valued functions that are independent of x for x outside a fixed compact set in \(\mathbb {R}^d\). Denote
The operator is assumed to satisfy the very weak hyperbolicity condition,
This hypothesis is best understood by considering first the case where the coefficients are independent of t, x. In that case, the initial value problem is solved by Fourier transform indicated by a hat,
In the case \(B=0\), the hypothesis implies that for all \(\xi \in \mathbb {R}^d\)
The Cauchy problem is well set in Sobolev spaces with at worst a loss of \(m-1\) derivatives. For general B not zero one has the weaker estimate
This estimate does not allow one to solve the initial value problem for \(g\in C^\infty _0(\mathbb {R}^d)\). However, its subexponential growth shows that the Cauchy problem is well set in Gevrey spaces. Those spaces can be localized by Gevrey partitions of unity so provide a reasonable setting for the initial value problem (0.1). In the constant coefficient case, condition (1.1) is necessary and sufficient for Gevrey well posedness.
The operators hyperbolic in the sense of Petrowsky and Gårding [1] are characterized by the stronger estimate
equivalent to Sobolev solvability with a loss of no more than \(m-1\) derivatives.
Estimate (1.2) corresponds to a sort of instability at high frequency that is stronger than permitted for coefficient problems that are hyperbolic in the sense of Petrovsky and Gårding [1].
The remarkable fact is that provided that the coefficients of L are Gevrey regular, the Cauchy problem for L is well-posed in Gevrey classes if and only if (1.1) holds. The sufficiency is a result of Bronshtein [2]. The necessity is proved in the trio of articles [3,4,5]. In (1.1), no hypothesis is made about the singularities of the characteristic variety of L for t, x fixed, nor on how the geometry of that variety changes as t, x vary. The precise Gevrey regularity required does depend on such structures. Roughly, the more variable are the multiplicities the stronger is the required Gevrey regularity.
The present paper provides additional evidence that the weakly hyperbolic operators characterized by (1.1) deserve the right to be considered hyperbolic. We give an algorithm that computes approximate solutions with reasonable computational cost. The stability analysis of discrete approximations took flight in the the classic paper of Courant Friedrichs and Lewy [6] followed by the work of Von Neumann who showed that the Fourier Transform offered profound insights on the stability of discrete approximations. That pseudodifferential operators offered additional insights was observed by Kreiss, Yamaguti and Nogi and most importantly Lax and Nirenberg who discovered the sharp Gårding inequality for matrix symbol pseudodifferential operators for such an application. This result is crucial for our analysis too. An excellent overview of the classic results is presented in [7]. The proof of stability of our scheme is as difficult as any stability result that we know. The difficulty has two sources. The first is that the stability of the Cauchy problem is itself very difficult. There is no simple multiplier method. Second the stability is very weak so it is reasonable to suspect that it can be destroyed by replacing the problem by a discrete one.
1.2 Algorithm definition
Choose \(\chi (x)\in C_0^{\infty }(\mathbb {R}^d)\) with \(\chi =1\) in \(|x|\le 2\) and \(\chi =0\) for \(|x|\ge 2\sqrt{2}\) such that \(0\le \chi \le 1\). Denote \(\chi _h(D)=\chi (h D)\). Define a family of spectral truncations of G by
The smoothing operators \(G_h\) generate the ordinary differential operators \(\partial _t-G_h\). The resulting ordinary differential equation is then approximated by the Crank-Nicholson scheme.
Definition 1.1
Define for \(n\in \mathbb {Z}\),
The Crank–Nicholson scheme generating a sequence \(\mathbb {N}\ni n \mapsto u_h^n\) intended to approximate \(u_h(nk)\) is
The uniform stability of the Cauchy problems for \(\partial _tu =G_hu\) is proved in [8]. This equation has a symmetrizer \(R_h=R^*_h\) with \(0<c_h<R_h\le 1\). However as the spectral truncation grows the lower bound \(c_h\) tends to zero.
Therefore, the straight forward stability arguments that would work for the Crank-Nicholson step, as in [9, 10] fail. The proof of stability must be at least as hard as the proof of the a priori estimates in [8]. Indeed they are more complicated. The main effort follows the strategy in [8]. We carefully control the additional errors from discretization in time. The Crank-Nicholson scheme is chosen because it is well adapted to estimates using a symmetrizer.
The precise stability result is Theorem 2.4. The proof that the approximations converge to the exact solution is Theorem 2.5.
For the very special case of operators of the form \(u_{tt} = a(t) u_{xx}\) with nonnegative Gevrey a, the spectral Leap-Frog scheme is analysed in [11]. The computational cost estimates of [11] shows that the cost of computing with error \(\epsilon \) grows no faster than polynomially in \(\epsilon ^{-1}\). Virtually identical cost estimates work for our spectral Crank-Nicholson scheme. They are not repeated here.
Constant coefficient problems that are hyperbolic in the sense of Gårding and Petrowsky are more strongly hyperbolic than those studied in this paper. However variable coefficient operators whose frozen problems are hyperbolic in this sense need not inherit the Sobolev well posedness of the constant coefficient problems. Stability of difference approximations to constant coefficient problems hyperbolic in the sense of Gårding and Petrowsky have been studied in a number of works. We refer to [12] for a review of these.
2 Main theorems
2.1 Definition of the parameter \(\theta \)
First we formulate an important property which follows from the assumption (1.1). Define
then from [8, Proposition 2.2] (see also [8, (2.3)]) it follows that for any compact set \(K\subset \mathbb {R}^d\) and \(T>0\) there are \(\epsilon _0>0\), \(c>0\) such that
for any \(x\in K\), \(|\xi |\le 1\), \(|(y,\eta )|\le 1\), \(|\epsilon |\le \epsilon _0\), \(|t|\le T\).
Following [8] introduce an integer \(\theta \) defined as follows.
Hypothesis 2.1
The system is \(\theta \) -regular with integer \(0\le \theta \le m-1\) in the sense that for any \(T>0\) and any compact \(K\subset \mathbb {R}^d\) there exist \(C>0\), \(c>0\) and \(\epsilon _0>0\) such that with \(N=\max \{2\theta ,m\}\)
for all \(s\ge 0\), \(0<\epsilon \le \epsilon _0\), \(|\xi | \le 1\), \(|(y,\eta )|\le 1\), \(x\in K\), \(|t|\le T\).
Remark 2.1
This definition of \(\theta \)-regularity is little bit more general than that of [8, Hypothesis 2.8]. Here \({\mathcal H}_r(t,x,\xi ,\xi , 0;\epsilon )\) coincides with \({\mathcal H}_r(t,x,\xi ; \epsilon )\) in [8].
Example 2.1
When (1.1) holds, Hypothesis 2.1 always holds with \(\theta =m-1\). If \(A(t,x,\xi )\) is uniformly diagonalizable then Hypothesis 2.1 holds with \(\theta =0\) (for the proof see [8, Examples 2.9 and 2.10]).
Example 2.2
Suppose (1.1). Assume that there exists \(T=T(t,x,\xi ,y,\eta ;\epsilon )\) with bounds on \(\Vert T\Vert \) and \(\Vert T^{-1}\Vert \) independent of \((t,x,\xi ,y,\eta ;\epsilon )\) such that \(T^{-1}{\mathcal H}_mT\) is a direct sum \(\oplus A_j\) where the size of \(A_j\) is at most \(\mu \). Then Hypothesis 2.1 holds with \(\theta =\mu -1\) (for the proof see [8, Example 2.11]).
2.2 Recall the continuous case
Let
then \(Lu=f\) is written
Denote
where \(\ell \ge 1\) is a positive parameter. When \(\ell =1\) we omit the suffix \(\ell \) and write \(\langle {\xi }\rangle _1=\langle {\xi }\rangle \).
Definition 2.1
If \(1<s<\infty \), the function \(a(x)\in C^{\infty }(\mathbb {R}^d)\) belongs to \(G^{s}(\mathbb {R}^d)\) if there exist \(C>0\), \(A>0\) such that
Recall [8, Proposition 4.4].
Proposition 2.1
Suppose Hypothesis 2.1 is satisfied. Define
For some \(1<s'\le s\) suppose that \(A_j(t,x)\) (resp. B(t, x)) are lipschitzian (resp. continuous) in time uniformly on compact sets with values in \(G^{s'}(\mathbb {R}^d)\). Then there exist \(T>0\), \({\hat{c}}>0\), \(C>0\) and \(\ell _0>0\) such that for all u such that \(e^{(T-{\hat{c}}\,t)\langle {D}\rangle _{\ell }^{\rho }}\partial _{t,x}^{\gamma }u\in L^1([0,T/{\hat{c}}];H^{\nu }(\mathbb {R}^d))\) for \(|\gamma |\le 1\) one has
for \(0\le t\le T/{\hat{c}}\) and \(\ell \ge \ell _0\).
This is a small improvement of [8, Proposition 4.4]. Here is a sketch of the easy proof: As noted in Remark 2.1 we use \({\mathcal H}_r(t,x,\xi ,y,\eta ; \epsilon )\) instead of \({\mathcal H}_r(t,x,\xi ;\epsilon )\) in [8] and make the same choice (3.16) below for \(s, \epsilon , \xi ,y,\eta \) where \(\chi _h\equiv 1\), \(\chi _{2h}\equiv 1\) and \({\bar{\tau }}-\tau =T-at\). Therefore (3.17) below holds for \(0\le T-at\le {\bar{\tau }}\) which gets rid off the constraint \(T-at\ge c\) with some \(c>0\) that we have assumed in [8]. This enables us to take \(T_1=T\) in [8, Proposition 4.4]. In the estimate (2.4) the weight for Lu is improved from \(\langle {D}\rangle _{\ell }^{3\nu }\) to \(\langle {D}\rangle _{\ell }^{-\nu }\). That proof is also easy.
Corollary 2.2
There exist \(T>0\), \({\hat{c}}>0\), \(C>0\) and \(\ell _0>0\) such that for all u satisfying \(\partial _tu=Gu\) one has
for \(0\le t\le T/{\hat{c}}\) and \(\ell \ge \ell _0\).
The proof of [8, Theorem 1.3] gives
Proposition 2.3
Assume the same assumption as in Proposition 2.1 and \(e^{T\langle {D}\rangle ^{\rho }}g\in H^{\nu }(\mathbb {R}^d)\). Then there exists a unique u satisfying
such that \(e^{(T-{\hat{c}}\,t)\langle {D}\rangle ^{\rho }}u\in L^{\infty }([0,T/{\hat{c}}];H^{-\nu }(\mathbb {R}^d))\).
2.3 Stability and error estimates
The Crank–Nicholson scheme defined in (1.5) is equivalent to
Note that
where
Assuming \({\bar{C}}\,k\,h^{-1}<1\) one has
and \(u^{n+1}_h\) is given by
Reasoning term by term in (2.7), \(\big ( I - \frac{k}{2}\, G^{n}_h\big )^{-1}\) maps functions with spectrum in \(\mathrm{supp}\,\chi _{2h}(\cdot )\) to themselves. Therefore,
Theorem 2.4
Make the same assumption as in Proposition 2.1. Then there exist \({\bar{\tau }}>0\), \({\bar{\beta }}>0\), \({\bar{a}}>0\), \({\bar{h}}>0\) and \(C>0\) such that the estimate
holds for any \(n\in \mathbb {N}\), \(k>0, h>0\) satisfying \(t_n=nk\le {\bar{\tau }}/{\bar{a}}\), \(kh^{-1}\le {\bar{\beta }}\) and \(0<h\le {\bar{h}}\) where \(\nu =\theta (1-\rho )\).
A more precise estimate of the stability is given in Proposition 3.12.
Theorem 2.5
In addition to the assumption in Proposition 2.1, assume that \(A_j(t,x)\) and B(t, x) are \(C^1\) in time uniformly on compact sets with values in \(G^{s'}(\mathbb {R}^d)\). Then there exist \({\bar{\tau }}>0\), \({\bar{\beta }}>0\), \({\bar{a}}>0\), \({\bar{h}}>0\) and \(C>0\) such that for an exact solution u to (0.1) with Cauchy data g satisfying \(\langle {D}\rangle ^{2+\nu }e^{{\bar{\tau }}\langle {D}\rangle ^{\rho }}g\in L^2\) one has
and
for any \(n\in \mathbb {N}\), \(k>0, h>0\) satisfying \(t_n=nk\le {\bar{\tau }}/{\bar{a}}\), \(kh^{-1}\le {\bar{\beta }}\) and \(0<h\le {\bar{h}}\).
Corollary 2.6
With the same assumptions as in Theorem 2.5 there exist \({\bar{\tau }}>0\), \({\bar{\beta }}>0\), \({\bar{a}}>0\), \({\bar{h}}>0\) and \(C>0\) such that for an exact solution u to (0.1) with Cauchy data g satisfying \(\langle {D}\rangle ^{2+\nu }e^{{\bar{\tau }}\langle {D}\rangle ^{\rho }}g\in L^2\) one has
for any \(n\in \mathbb {N}\), \(k>0, h>0\) satisfying \(t_n=nk\le {\bar{\tau }}/{\bar{a}}\), \(kh^{-1}\le {\bar{\beta }}\) and \(0<h\le {\bar{h}}\).
Remark 2.2
Note that
so that one has \(\rho \ge 3\nu +1/2\) under the assumption of Theorems 2.4 and 2.5.
3 Stability for the spectral Crank–Nicholson scheme
3.1 Spectral truncated weight for Crank–Nicholson scheme
Taking (1.5) into account define spectral truncated weights \(W_h(t,D)\) by
and for \(n\in \mathbb {N}\)
where \(a>0\) is a positive parameter which will be fixed later. In what follows we always assume that the parameters \(a>0, k>0, \ell>0, h>0\) are constrained to satisfy
Since \(a\langle {\xi }\rangle _{\ell }^{\rho }\chi _h\le 3\,a\, h^{-\rho }\) because \(\langle {\xi }\rangle _{\ell }\le 3h^{-1}\) if \(\chi _h(\xi )\ne 0\), it follows that
Here recall [8, Definition 2.3].
Definition 3.1
For \(0<\delta \le \rho \le 1\), the family \(a(x,\xi ;\ell )\in C^{\infty }(\mathbb {R}^d\times \mathbb {R}^d)\) indexed by \(\ell \) belongs to \({\tilde{S}}^m_{\rho ,\delta }\) if for all \(\alpha \), \(\beta \in \mathbb {N}^d\) there is \(C_{\alpha \beta }\) independent of \(\ell \ge 1,x,\xi \) such that
Denote \({\tilde{S}}^m={\tilde{S}}^m_{1,0}\).
Since \(|\partial _{\xi }^{\alpha }\chi _h|\le C_{\alpha }h^{|\alpha |}\) and \(2h^{-1}\le \langle {\xi }\rangle _{\ell }\le 3h^{-1}\) on the support of \(\partial _{\xi }^{\alpha }\chi _h\) for \(|\alpha |\ge 1\) it is clear that \(\chi _h\in {\tilde{S}}^0\).
We examine to what extent \(W^n_h\) satisfies the Crank-Nicholson scheme (1.5).
Lemma 3.1
Assume (3.1) then one can write
where \(\omega _h(\xi )\in {\tilde{S}}^{\rho }\) and
Proof
Denote
then it is clear that
Since
one can define \(\omega _h\) by
Then the first assertion is clear from (3.2). Note that
because of (3.1). Therefore one has \( |\partial _{\xi }^{\alpha }\omega _h|\le C_{\alpha }\, \langle {\xi }\rangle ^{\rho -|\alpha |}_{\ell } \). Using (3.2) this implies the second assertion. \(\square \)
3.2 Crank–Nicholson after conjugation
Note that \(u^n_h\) satisfy
thanks to (2.8). Assume that \(u^n_h\) satisfies
where \(\chi _hf^n=f^n\) is not necessarily assumed.
Consider a weighted energy \((R_h^nW^n_hu_h^n,\,W^n_hu_h^n)\) where \(R^n_h\) is a symmetrizer that is symmetric \((R^n_h)^*=R^n_h\) and will be defined in Sect. 3.3 below. The discrete analog of \(\partial _t (R_h^nW^n_hu_h^n,\,W^n_hu_h^n)\) is the time difference
In what follows we omit the subscript h for ease of reading. Write (3.7) as
with
The term (III) is an error term that will be estimated in Sect. 3.3. The first term is equal to
where
The first line of (3.9) is
Note that
Using (3.3) and \(\omega \chi _h\,W^m=W^m \omega \chi _h\) this becomes
Therefore with \(\Omega ^{n}=W^{n+1}+W^n\) one has, since \((R^n)^*=R^n\)
Thus (I) yields
Since \(\chi _h\,\Omega ^n=\Omega ^n\,\chi _h\) and \(\omega \chi _h=\chi _h \omega \) and using \(\chi _hu^{n+j}=u^{n+j}\) that follows from (3.5) one has
The second line of (3.9) yields, with \(U^n=u^{n+1}+u^n\)
Because of (3.6), this is equal to
Similarly the third line of (3.9) is
Therefore the sum of the second and the third lines of (3.9), denoted by (II), yields
Recalling
we have proved the following proposition.
Proposition 3.2
We have
where \(\Omega ^n:=W^{n+1}+W^n\) and \(U^n:=u^{n+1}+u^n\).
3.3 Composition with \(W^{\pm n}_h \) and definition of \(R^n_h\)
First recall Definition 2.4 from [8].
Definition 3.2
For \(1<s\), \(m\in \mathbb {R}\), the family \(a(x,\xi ;\ell )\in C^{\infty }(\mathbb {R}^d\times \mathbb {R}^d ) \) belongs to \({\tilde{S}}_{(s)}^m\) if there exist \(C>0\), \(A>0\) independent of \(\ell \ge 1,x,\xi \) such that for all \(\alpha \), \(\beta \in \mathbb {N}^d\),
We often write \(a(x,\xi )\) for \(a(x,\xi ;\ell )\) dropping the \(\ell \). If \(a(x,\xi )\) is the symbol of a differential operator of order m with coefficients \(a_{\alpha }(x)\in G^{s}(\mathbb {R}^d)\) then \(a(x,\xi )\in {\tilde{S}}^m_{(s)}\) because \(|\partial _{\xi }^{\beta }\xi ^{\alpha }|\le CA^{|\beta |}|\beta |!\langle {\xi }\rangle _{\ell }^{|\alpha |-|\beta |}\) and \(|\partial _x^{\beta }a_{\alpha }(x)|\le C_{\alpha }A_{\alpha }^{|\beta |}|\beta |!^s\) for any \(\beta \in \mathbb {N}^d\).
Proposition 3.3
Suppose \(1/2\le \rho <1\) and \(s=1/\rho \) and let \(A(x,\xi )\) be \(m\times m\) matrix valued with entries in \({\tilde{S}}^1_{(s)}\) and \(\partial _x^{\alpha }A(x,\xi )=0\) outside \(|x|<R\) for some \(R>0\) if \(|\alpha |>0\). Define \( m^*:={\max {\{\rho -k(1-\rho ),-1+\rho \}}}\). Then there is \({\bar{\tau }}>0\), \(\ell _0>0\) such that
is a pseudodifferential operator with symbol given by
with \(R_k\in {\tilde{S}}^{m^*}\) uniformly in \(\tau \), \(\ell \) constrained to satisfy
In particular \({\tilde{A}}(x,\xi )\in {\tilde{S}}^1\) uniformly in such \(\tau \), \(\ell \).
Remark 3.1
This proposition with \(\chi _h\equiv 1\) is [8, Proposition 2.6]. The proof for the case \(\chi _h\equiv 1\) works without any change for the case \(\chi _h\in {\tilde{S}}^0\).
Choosing a smaller \({\bar{\tau }}>0\) if necessary one can assume that
In what follows we choose \(T={\bar{\tau }}\) in the definition of \(W(t,\xi )\) yielding
With \(N=\max \{2\theta ,m\}\) denote
Suppressing the subscript h for ease of reading, Proposition 3.3 shows that
The choice of N guarantees that where \(2\theta (1-\rho )+m^*\le \rho \). Define
Then, the definition of \({\mathcal H}_N\) implies that
In (3.15) choose
Using \(0\le \chi _{2h}\le 1\), it follows from (2.2) that
for \(|t|\le T\), \(\ell \ge \ell _0\) where
Following [8] define
with a positive parameter \(b>0\) that will be fixed later. Since \(\Vert e^{sM^h(t,x,\xi ,\tau )}\Vert =e^{-bs\langle {\xi }\rangle _{\ell }^{\rho }}\Vert e^{isH^h(t,x,\xi ,\tau )}\Vert \), (3.17) implies
with \(\nu =\theta (1-\rho )\) for \(|t|\le T\) and \(b\ge b_0\) with some \(b_0>0\) where \(c_1\), \(c_2\) and \(C>0\) are independent of \(\ell \), h and b.
Introduce the symmetrizer
From [8, Theorem 3.1] it follows that
under the constraint
so that \(b\,\langle {\xi }\rangle _{\ell }^{\rho -1}\le b\,\ell ^{-(1-\rho )}\le 1\). Recall [8, page 230] that
that is
Lemma 3.4
We have
for \(0\le t\le T\) uniformly in a, b, n, k, h under the constraint \( ank\le {\bar{\tau }}\).
Proof
We show that
with \(C_{\alpha \beta }\) independent of b, h and \(0\le \tau \le {\bar{\tau }}\). If (3.21) is proved then writing
the assertion follows immediately. To prove the estimate (3.21) we apply the same arguments in the proof of [8, Theorem 3.1]. First consider \(\partial _{\tau }H(t,x,\xi ,\tau )\). Since
it follows that
Denote
Since
then (3.22) and Duhamel’s representation yields
where \(E(s)=\langle {\xi }\rangle _{\ell }^{\nu }e^{-cb s\langle {\xi }\rangle _{\ell }^{\rho }}\). Repeating the same arguments in the proof of [8, Theorem3.1] one can prove
from which we obtain (3.21) by exactly the same way as in the proof of [8, Theorem 3.1]. \(\square \)
Lemma 3.5
With \( R_h^n(x,\xi ) \, :=\, R_h(nk,x,\xi ,ank) \), one has
for \(0\le (n+1)k\le T\) uniformly in a, b, n, k, h under the constraint
Proof
Write
Express
Using \(b\,\partial _tR_h(t,x,\xi ,\tau )\in {\tilde{S}}^{3\nu +1-\rho }_{\rho -\nu ,1-\rho +\nu }\), one obtains
where \(3\nu +1-\rho \le -2\nu +\rho \) in view of (2.2). For the term \(R_h((n+1)k,x,\xi ,a(n+1)k)-R_h((n+1)k,x,\xi ,ank)\) we apply Lemma 3.4 to get
Here note that (2.9) implies \( 1/2>3\nu \) because \(1>\rho \ge 3\nu +1/2\) and hence
Then noting that \(a\langle {\xi }\rangle _{\ell }^{3\nu }\le a\langle {\xi }\rangle ^{-\rho /6}_{\ell }\langle {\xi }\rangle ^{\rho -2\nu }_{\ell }\le a\,\ell ^{-\rho /6}\langle {\xi }\rangle ^{\rho -2\nu }_{\ell }\) one has
under the constraint \(a\,\ell ^{-\rho /6}\le 1\). Thus the proof is complete. \(\square \)
Definition 3.3
For a \(m\times m\) complex matrix \(\mathcal {M}=\mathcal {M}^*\), the notation \(\mathcal {M}\gg 0\) means that for all \(v\in \mathbb {C}^m\) one has \((\mathcal {M}v,\,v)_{\mathbb {C}^m}\ge 0\). For two such matrices, \(\mathcal {M}_1\gg \mathcal {M}_2\) means that \(\mathcal {M}_1-\mathcal {M}_2 \gg 0\).
Equation (3.17) yields for any \(v\in \mathbb {C}^m\)
This is an important pointwise lower bound for the symbol
where c is independent of b, a, n, k, h constrained to satisfy (3.23) and (3.19).
3.4 Estimate of (I)
Suppressing the suffix h again, denote
so that \(W^{n+1}=W^nW\) where \(1/2\le W\le 1\) and \(W^{\pm 1}\in {\tilde{S}}^0\) which follows from (3.2) and (3.4). Consider \({\mathsf {Re}}(\omega \,R^n\,\Omega ^n\,w,\, \Omega ^n\,w)\). Write \(\Omega ^n=W^{n+1}+W^n=(1+W)\,W^n\) and hence
Note that \(R^n\in {\tilde{S}}^{2\nu }_{\rho -\nu ,1-\rho +\nu }\) uniformly for parameters satisfying the constraints (3.23) and (3.19).
Lemma 3.6
One can write
with \((R_1^n)^*=R_1^n\) and \(R_2^n\in {\tilde{S}}^{4\nu -\rho }_{\rho -\nu ,1-\rho +\nu }\).
Proof
Denote \(f(\xi )=(1+W(\xi ))\omega (\xi )\) and \(g(\xi )=1+W(\xi )\). Since one has \(f(\xi )\in {\tilde{S}}^{\rho }\subset {\tilde{S}}^{\rho }_{\rho -\nu ,1-\rho +\nu }\) and \(g(\xi )\in {\tilde{S}}^0\subset {\tilde{S}}^{0}_{\rho -\nu ,1-\rho +\nu }\) applying [13, Theorem 18.5.4] one can write
where \(R_N\in {\tilde{S}}^{2\nu +\rho -N(2\rho -1-2\nu )}_{\rho -\nu ,1-\rho +\nu }\).
Choose N so large that \(2\nu +\rho -N(2\rho -1-2\nu )\le 4\nu -\rho \). The second term on the right-hand side, denoted by \(iR_1^n\), clearly satisfies \((R_1^n)^*=R_1^n\) because \(f(\xi )\) and \(g(\xi )\) are real scalar symbols. Since \(\partial _{\xi }^{\alpha }f\in {\tilde{S}}^{\rho -|\alpha |}_{\rho -\nu ,1-\rho +\nu }\) and \(\partial _{\xi }^{\beta }g\in {\tilde{S}}^{-|\beta |}_{\rho -\nu ,1-\rho +\nu }\) it is clear that the third term on the right-hand side is in \({\tilde{S}}^{4\nu -\rho }_{\rho -\nu ,1-\rho +\nu }\). \(\square \)
Thanks to Lemma 3.6 we have
It follows from Lemma 3.1 and (3.25) that
with \(c>0\) uniformly in the constrained parameters k, n, h, a, b. Note that
since for any \(k(\xi ) \in {\tilde{S}}^{q}\) one has
which is bounded by \(C_{\alpha }b^{-|\alpha |}\langle {\xi }\rangle _{\ell }^{q-|\alpha |(\rho -\nu )}\) because of (3.19). Repeating the arguments proving [8, (4.6)] it follows from the sharp Gårding inequality [13, Theorem 18.6.7] that there is \(\ell _0>0\) such that
for \(\ell \ge \ell _0\). Since \(-\nu +\rho /2\ge -\rho /2+1/2+2\nu \) by (2.9), choosing another \(b_0\) if necessary one obtains
with \(c'>0\) uniform for \(\ell \ge \ell _0\) and \(b\ge b_0\). From (2.9) again one sees that
and one concludes that choosing another \(\ell _0\) if necessary
with \(c>0\) uniform for \(\ell \ge \ell _0\) and \(b\ge b_0\). Repeating the same arguments one obtains
Summarizing we have
Lemma 3.7
There are \(c>0\), \(\ell _0\) and \(b_0\) such that for \(\ell \ge \ell _0\) and \(b\ge b_0\) one has
Lemma 3.7 together with (3.11) prove the following
Proposition 3.8
There exist \(c>0\), \(\ell _0>0\) and \(b_0\) such that for \(\ell \ge \ell _0\) and \(b\ge b_0\) one has
From (3.25) one has \(R^n\gg c\langle {\xi }\rangle ^{-2\nu }I\) and \(\langle {\xi }\rangle ^{\rho }_{\ell }R^n\gg c\langle {\xi }\rangle ^{\rho -2\nu }I\) with some \(c>0\) then repeating the same arguments proving (3.26) above there is \(c>0\) such that
for \(b\ge b_0\). In particular \(\mathrm{Op}(\langle {\xi }\rangle ^{\rho }_{\ell }R^n)\) is nonnegative hence
for any \(\delta >0\).
3.5 Estimate of (II)
Consider the term \({\mathsf {Re}}\,(R^nW^{n}\,U^n,\,W^{n}\,G^n\,U^n)\). Recall \(G^n=\chi _{2h}(iA(nk,x,D)+B(nk,x))\chi _{2h}\) and with \(t_n=nk\)
where
Then thanks to Proposition 3.3,
Since \(\chi _{2h}\in {\tilde{S}}^0\) and \(H(t_n,x,\xi ,at_n)\in {\tilde{S}}^1\), one sees that
where \({\tilde{R}}_n\in {\tilde{S}}^0\) uniformly in all parameters satisfying \(at_n=ank\le {\bar{\tau }}\). Define \(K^n:={\tilde{R}}_n+\chi _{2h}\#R(t_n,at_n)\#\chi _{2h}+\chi _{2h}\#W^n\#B(t_n)\#W^{-n}\#\chi _{2h}\). Then
so,
In addition,
Note that \(2\nu +{\bar{m}}\le \rho \) since \(2\nu +{{m}^*}\le \rho \). Recall
and \(R^n_h=R_h(t_n,x,\xi ,at_n)\) so that from (3.20) it follows that
In view of (3.29), denoting \(H(t_n)=H(t_n,x,\xi ,at_n)\), one has
It follows from (3.30) that
where \(b\, L^n\in {\tilde{S}}^{1-\rho +3\nu }_{\rho -\nu ,1-\rho +\nu }\) and \({\tilde{L}}^n\in {\tilde{S}}^{2\nu +{\bar{m}}}_{\rho -\nu ,1-\rho +\nu }\subset {\tilde{S}}^{\rho }_{\rho -\nu ,1-\rho +\nu }\). Since \(\rho \ge 1-\rho +3\nu \) taking another \(b_0\) if necessary one concludes
for \(b\ge b_0\). Thanks to (3.28) one has
Combining these estimates one obtains for \(b\ge b_0\),
Next study \({\mathsf {Re}}\,(R^nW^{n+1}\,G^n\,U^n,\,W^{n+1}\,U^n)\).
Lemma 3.9
One has
Proof
Write \(W^{n+1}\#G^n\#W^{-(n+1)}=W\#\big (W^n\#G^n\#W^{-n})\#W^{-1}\) so that
Since \(W^{\pm 1}\in {\tilde{S}}^{0}\) and \(H(t_n)\in {\tilde{S}}^1\) it is clear that
This proves the lemma. \(\square \)
Lemma 3.9 implies that
with
Since \(R^n\#T^n+(T^n)^*\#R^n\in {\tilde{S}}^{2\nu }_{\rho -\nu ,1-\rho +\nu }\) and \(\rho \ge 4\nu \) by (3.24) repeating the same arguments proving (3.31) one obtains for \(b\ge b_0\)
Equations (3.31) and (3.32) yield the following lemma.
Lemma 3.10
There exist \(b_0>0\) and \(\ell _0>0\) such that for \(b\ge b_0\) and \(\ell \ge \ell _0\) one has
Next estimate \(\sum _{i=0}^1{\mathsf {Re}}\,(W^{n+i}R^nW^{n+i}f^n,\,U^n)\). Since \(R^n\in {\tilde{S}}^{2\nu }_{\rho -\nu ,1-\rho +\nu }\), it follows that
Equation (3.24) implies that \(-\nu >2\nu -\rho /2\) so
Lemma 3.10 together with (3.12) and (3.33) yield the following proposition.
Proposition 3.11
There exist \(C>0\), \(b_0>0\) and \(\ell _0>0\) such that for \(b\ge b_0\) and \(\ell \ge \ell _0\) one has
3.6 Proof of Theorem 2.4
First choose \(b={\bar{b}}\) and \(\ell _1\) such that Propositions 3.8 and 3.11 and (3.27) hold with \(b={\bar{b}}\) and \(\ell \ge \ell _1\). Next choose \(a={\bar{a}}\) such that \(c\,{\bar{a}}\ge {\bar{b}}/2\) then taking (3.27) into account it follows from Propositions 3.8 and 3.11 that
Finally we estimate (III). Thanks to Lemma 3.5 one has
Increase \({\bar{a}}\) if necessary so that \(c\,{\bar{a}}\ge 2C'\,{\bar{b}}^{-1}\), in view of (3.34) and (3.35), recalling (3.13), we conclude that
Noting (3.23) and (3.19) we set
In what follows we assume \(\ell \ge \ell _2\). Taking (3.1) into account define
Note that \(\Vert \langle {D}\rangle _{\ell }^{-\nu }\,W^{n+1}\,f^n\Vert \le \Vert \langle {D}\rangle _{\ell }^{-\nu }\,W^{n}\,f^n\Vert \) thanks to (3.2). Summing (3.36) from 0 to \(n-1\) yields
Since \(W^p=e^{({\bar{\tau }}-{\bar{a}}t_p)\langle {D}\rangle _{\ell }^{\rho }\chi _h}\) with \(\chi _h=1\) on \(\mathrm{supp}\,\chi _{2h}\), and recalling (2.8), it follows from (3.27) and (3.37) that
Equation (3.24) implies that \(\rho /2-\nu >2\nu \) yielding the following proposition.
Proposition 3.12
There exist \({\bar{\tau }}>0\), \({\bar{a}}>0\), \({\bar{\beta }}>0\), \(C>0\) and \({\bar{\ell }}\,(\ge \ell _2 )\) such that one has
for any \(n\in \mathbb {N}\), \(k>0, \ell>0, h>0\) satisfying \(nk\le {\bar{\tau }}/{\bar{a}}\), \(kh^{-1}\le {\bar{\beta }}\) and \(h^{-1}\ge \ell \ge {\bar{\ell }}\).
Remark 3.2
To obtain Proposition 3.12 the spectral condition \(\chi _{h}u^n=u^n\) is assumed while for \(f^n\) no spectral condition is assumed.
Proof of Theorem 2.4
Fix \(\ell ={\bar{\ell }}\) in Proposition 3.12. Since
the proof is immediate. \(\square \)
4 Error estimates for the spectral Crank–Nicholson scheme
4.1 Continuous case revisited
Start by extending estimates (2.5) in Corollary 2.2 to \(\partial _t^ju\) for \(j=1,2\). It is clear that one can assume \({\bar{\tau }}\le T\) and \({\bar{a}}\ge {\hat{c}}\). Then it is easy to examine that Corollary 2.2 holds with \(T={\bar{\tau }}\) and \({\hat{c}}={\bar{a}}\). Suppose \(\partial _tu=Gu\). Write
so \(\langle {D}\rangle _{\ell }^{\mu }u\) satisfies \(\partial _t(\langle {D}\rangle _{\ell }^{\mu }u)=(G+B_{\mu })\langle {D}\rangle _{\ell }^{\mu }u\). The \(B_\mu \) satisfy the following bounds.
Lemma 4.1
There is \(A>0\) such that for any \(\alpha , \beta \in \mathbb {N}^d\) one has
Proof
Up to a multiplicative constant \(B_{\mu }\) is given by
Therefore \(\partial _{\xi }^{\alpha }\partial _x^{\beta }B_{\mu }\) is, after change of variables \(x+\theta y\mapsto y\), \(\theta ^{-1}\eta \mapsto \eta \), a sum of terms
with \(\alpha '+\alpha ''=\alpha \). Recall that
with some \(c>0\) (see [8, Lemma 6.2]). In addition, it is easy to see that
Using \(\langle {x}\rangle ^{2d}e^{ix\eta }=\langle {D_{\eta }}\rangle ^{2d}e^{ix\eta }\), an integration by parts in \(\eta \) proves the assertion. \(\square \)
Thanks to Lemma 4.1 it follows from the proof of Proposition 3.3 that
Apply Corollary 2.2 to \(\partial _tv=(G+B_{\mu })v\) with \(v=\langle {D}\rangle _{\ell }^{\mu }u\) to find that choosing a smaller \({\bar{\tau }}>0\) and larger \({\bar{a}}>0\) and \(\ell _0\) if necessary,
for \(0\le t\le {\bar{\tau }}/{\bar{a}}\) and \(\ell \ge \ell _0\). Indeed in the proof of Proposition 2.1 the term B satisfies \(e^{(T-{\hat{c}}t)\langle {\xi }\rangle _{\ell }^{\rho }}\#B\# e^{(T-{\hat{c}}t)\langle {\xi }\rangle _{\ell }^{\rho }}\in {\tilde{S}}^0\) so choosing \({\hat{c}}\) large, it is irrelevant. Write
Proposition 3.3 and Lemma 4.1 imply that \(e^{({\bar{\tau }}-{\bar{a}}t)\langle {\xi }\rangle _{\ell }^{\rho }}\#(G+B_{\mu })\#e^{-({\bar{\tau }}-{\bar{a}}t)\langle {\xi }\rangle _{\ell }^{\rho }}\in {\tilde{S}}^1\). It follows that
from (4.1). Next assume that \(A_j(t,x)\) and B(t, x) are \(C^1\) in time uniformly on compact sets with values in \(G^{s'}(\mathbb {R}^d)\). Since \(\partial _t^2u=(\partial _tG)u+G\partial _tu\) repeating the same arguments one has
Choosing \(\mu =-\nu +i\), \(i=0,1,2\) one obtains the following lemma.
Lemma 4.2
Assume that \(A_j(t,x)\) and B(t, x) are \(C^1\) in time uniformly on compact sets with values in \(G^{s'}(\mathbb {R}^d)\) and that \(\partial _tu=Gu\). Then there exist \(C>0\), \(\ell _0>0\) such that
for \(0\le t\le {\bar{\tau }}/{\bar{a}}\), \(\ell \ge \ell _0\) and \(0\le i,j\le 2\).
4.2 Error estimate for the spectral Crank–Nicholson scheme
Suppose that u(t, x) satisfies
where \(G(t,x,D)=iA(t,x,D)+B(t,x)\). Denote \({\tilde{u}}=\chi _{2h} u\) so that \(\chi _{h}{\tilde{u}}={\tilde{u}}\). Thus
Next estimate to what extent \({\tilde{u}}(t_n,x)\) satisfies the difference scheme. The error, denoted by \(g(n)=g(n,\cdot )\), is given by
where \(G^n=\chi _{2h}(iA(nk,x,D)+B(nk,x))\chi _{2h}\). Note that
The approximate solution \(u^n=u^n_h\) satisfies
At \(t=0\) the approximate solution is equal to the spectral truncation of the exact solution, \(u^0=\chi _{2h}g={\tilde{u}}(0)\).
Noting \(\mathrm{supp}{\mathcal F}\big ({\tilde{u}}(t_n)-u^n\big )\subset \mathrm{supp}\,\chi _{2h}\) and hence \(\chi _h({\tilde{u}}(t_n)-u^n)={\tilde{u}}(t_n)-u^n\), Proposition 3.12 implies
for any \(t_n=kn\le {\bar{\tau }}/{\bar{a}}\).
Lemma 4.3
There is \(C>0\) so that
for \(0\le j\le n-1\) and \(0\le t_n\le {\bar{\tau }}/{\bar{a}}\).
Proof
Use (4.3) to write
The triangle inequality yields
Write
and note that
Since \(0\le s'-t_j\le k\) if \(t_j\le s'\le t_{j+1}\) it follows from (3.2) that
thanks to Lemma 4.2. Therefore one has
Turn to the second term on the right-hand side of (4.6). Use
to write
Proposition 3.3 implies that \(\langle {\xi }\rangle _{\ell }^{-\nu }\#W^j\#G^j\#W^{-j}\in {\tilde{S}}^{1-\nu }\). In addition, \(W^je^{-({\bar{\tau }}-{\bar{a}}s')\langle {D}\rangle _{\ell }^{\rho }\chi _h}=e^{{\bar{a}}(s'-t_j)\langle {D}\rangle _{\ell }^{\rho }\chi _h}\) when \(0\le s'-t_j\le k\). Repeat the same arguments as above to find
Then
Next study the third and fourth term on the right-hand side of (4.6).
Lemma 4.4
Let \(\alpha \ge 0\). There is \(C>0\) such that
Proof
Since \(1-\chi _{2h}(\xi )=0\) unless \(|\xi |\ge h^{-1}\) one has
which proves the assertion. \(\square \)
Since \(G^j-G(t_j)=\chi _{2h}G(t_j)(\chi _{2h}-I)+(\chi _{2h}-I)G(t_j)\) one can write
Using \(\langle {\xi }\rangle _{\ell }^{-\nu }\#W^j\#G(t_j)\#W^{-j}\in {\tilde{S}}^{1-\nu }\) together with Lemma 4.4 one finds
Therefore by Lemma 4.2,
Turn to \(f(j):=[\chi _{2h},G(t_j)]u(t_j)\). Since
repeating the same arguments as above one obtains that
This finishes the proof of Lemma 4.3. \(\square \)
4.3 Proof of Theorem 2.5
Noting that \(\mathrm{supp}{\mathcal {F}}\big ({\tilde{u}}(t_n)-u^n\big )\subset \mathrm{supp}\,\chi _{2h}\) and \(\chi _h=1\) on the support of \(\chi _{2h}\) it follows from (4.5) and Lemma 4.3 that
Since \(\langle {\xi }\rangle _{\ell }\le \sqrt{3}h^{-1}\) on the support of \(\chi _{2h}\), (4.7) implies that
Finally estimate \(\Vert \langle {D}\rangle _{\ell }^{-\nu }W^n(u(t_n)-{\tilde{u}}(t_n))\Vert \). Since \(u(t_n)-{\tilde{u}}(t_n)=(1-\chi _{2h})u(t_n)\) the same arguments as above prove that
Similarly one has
Combining (4.7), (4.8) and (4.9), (4.10) yields the following proposition.
Proposition 4.5
There exist \({\bar{\tau }}>0, {\bar{a}}>0, {\bar{\beta } }>0, C>0\) and \({\bar{\ell }}>0\) such that for any exact solution u to (4.2) with Cauchy data u(0) such that \(\langle {D}\rangle ^{2+\nu }_{\ell }e^{{\bar{\tau }}\langle {D}\rangle _{\ell }^{\rho }}u(0)\in L^2\) one has
and
for any \(0\le t_n=nk\le {\bar{\tau }}/{\bar{a}}\), \(kh^{-1}\le {\bar{\beta }}\) and \(h^{-1}\ge \ell \ge {\bar{\ell }}\).
Remark 4.1
In order for a difference approximation to be accurate, the time discretization must be taken sufficiently fine [6]. Here Proposition 4.5 shows that one could constrain k to satisfy a CFL type condition \(kh^{-1}\le {\bar{\beta }}\). More precisely, the proof shows that it suffices to constrain k to satisfy
Proof of Theorem 2.5
Taking (3.39) into account it is enough to choose \(\ell ={\bar{\ell }}\) in Proposition 4.5. \(\square \)
References
Gårding, L.: Linear hyperbolic partial differential equations with constant coefficients. Acta Math. 85, 1–62 (1951)
Bronshtein, M.D.: The Cauchy problem for hyperbolic operators with characteristics of variable multiplicity. Trudy Moskov. Mat. Obsc. 41, 83–99 (1980)
Lax, P.D.: Asymptotic solutions of oscillatory initial value problems. Duke Math. J. 24, 627–646 (1957)
Mizohata, S.: Some remarks on the Cauchy problem. J. Math. Kyoto Univ. 1, 109–127 (1961/1962)
Nishitani, T.: On the Lax-Mizohata theorem in the analytic and Gevrey classes. J. Math. Kyoto Univ. 18, 509–521 (1978)
Courant, R., Friedrichs, K.O., Lewy, H.: Über die partiellen differenzengleichungen der mathematischen physik. Math. Ann. 100, 32–74 (1928)
Richtmyer, R., Morton, K.: Difference Methods for Initial-Value Problems, 2nd edn. Wiley, New York (1967)
Colombini, F., Nishitani, T., Rauch, J.: Weakly hyperbolic systems by symmetrization. Ann. Scuola Norm. Sup. Pisa. 19, 217–251 (2019)
Lax, P.D.: Differential equations, difference equations and matrix theory. Commun. Pure Appl. Math. 11, 175–194 (1958)
Yamaguti, M., Nogi, T.: An algebra of pseudo difference schemes and its applications. Publ. Res. Inst. Math. Sci. 3, 151–166 (1967)
Colombini, F., Rauch, J.: Numerical analysis of very weakly well-posed hyperbolic Cauchy problems. IMA J. Numer. Anal. 35, 989–1010 (2015)
Petit-Bergez, S.: Problèmes faiblement bien posés: discrétisations et applications. Thèse Docteur de Mathématiques Université de Paris 13 (2006)
Hörmander, L.: The Analysis of Linear Partial Differential Operators. III. Pseudo-Differential Operators. Springer, Berlin (1994)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Colombini, F., Nishitani, T. & Rauch, J. A discrete algorithm for general weakly hyperbolic systems. J. Pseudo-Differ. Oper. Appl. 13, 20 (2022). https://doi.org/10.1007/s11868-022-00452-z
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11868-022-00452-z
Keywords
- Spectral method
- Crank–Nicholson
- Stability
- Weak hyperbolicity
- Pseudodifferential symmetrizer
- Gevrey regularity
- Error estimates