0 200 400 600 800 1000
−4
−2 0 2
4 x 10
−3Figure 4.2.4: Global error of the six order Magnus method for the Airy equation y′′(t) = −ty(t) with [1,0]⊺ initial conditions, 0 ≤ t ≤ 1000 and step-size h= 14.
Performance of the modified Magnus method is better than that of the classicalMagnusmethod due to a number of reasons. Firstly, the fact that the matrix B is small, B(t) =O(t−tn+1
2),contributes to higher order correction to the solution. Secondly, the order of themodified Magnusmethod increases from p= 2s to p= 3s+ 1, [Ise02b].
4.4 The FM method
In this section we present a novel FM method for solving matrix ODEs, [Kha09a]. The method combines ideas on Filon methods and Lie group methods for solving matrix exponential with nested integral commutators.
Consider theAiry-type equation
y′′(t) +g(t)y(t) = 0, g(t)>0 for t >0, and lim
t→∞g(t) = +∞. Replacing the second order differential equation by the first order system, we
4.4 The FM method
0 200 400 600 800 1000
−8
−4 0 4
8 x 10
−4Figure 4.2.5: Global error of the six order Magnus method for the Airy equation y′′(t) = −ty(t) with [1,0]⊺ initial conditions, 0 ≤ t ≤ 1000 and step-size h= 101.
obtain y′ =Aω(t)y, where
Aω(t) =
0 1
−g(t) 0
.
Due to a large imaginary spectrum of the matrixAω, thisAiry-type equa- tion is rapidly oscillating. We apply the modified Magnus method to solve the system y′ =Aω(t)y,
yn+1 = ehAω(tn+h/2)eΩ˜nyn.
Here the integral commutators in ˜Ωn are computed according to the rules of the Filon quadrature. This results in highly accurate numerical method, the FM method. Taking into account that the entries of the matrixB(t) are likely to be oscillatory, the advantage of the Filon-type method is evident.
In the example below we provide a more detailed evaluation of the FM method applied to solve the Airy-type equation y′ =Aω(t)y.
4.4 The FM method
0 200 400 600 800 1000
−5 0
5 x 10
−9Figure 4.2.6: Global error of the six order Magnus method for the Airy equation y′′(t) = −ty(t) with [1,0]⊺ initial conditions, 0 ≤ t ≤ 1000 and step-size h= 251.
Example 4.4.1 Once we have obtained the representation for Ω˜n, the com- mutator brackets are now formed by the matrix B(t). It is possible to reduce cost of evaluation of the matrix B(t) by simplifying it, [Ise04a]. Denote q =q
g(tn+12h) and v(t) =g(tn+t)−g(tn+12h). Then, B(t) =v(t)
(2q)−1sin 2qt q−2sin2qt
−cos2qt −(2q)−1sin 2qt
=v(t)
q−1sinqt
−cosqt
cosqt q−1sinqt , and for the product
B(t)B(s) =v(t)v(s)sinq(s−t) q
q−1sinqt
−cosqt
cosqs q−1sinqs . It was shown in [Ise04a] that
kB(t)k= cos2ωt+sin2ωt
ω2 and B(t) =O(t−tn+1
2).
4.4 The FM method
We can now estimate integral Zh
0
kB(t)kdt= h
2(1 + 1
ω2) + sin 2ωh
2ω (1− 1
ω2) := q(h, ω).
To meet requirements stated in Theorem 4.2.1 we allow q(h, ω)≤π.
In other words, to satisfy Theorem 4.2.1 we obtain the following condition on the step-size h,
h≤ 2πω3−ω2+ 1 ω(ω2+ 1) .
Given the compact representation of the matrix B(t) with oscillatory entries, we solve Ω(t) with a Filon-type method. We approximate functions in B(t) by a polynomial ˜v(t), for example Hermite polynomial (4.1.4), as in classical Filon-type method. Furthermore, in our approximation we use end points only, although the method is general and more nodes of approximation can be required.
In Figure 4.4.1 we present the global error of the fourth order modified Magnus method with exact integrals for the Airy equation y′′(t) = −ty(t) with [1,0]⊺initial conditions, 0 ≤t≤2000 time interval and step-size h= 15. This can be compared with the global error of the FMmethod applied to the same equation with exactly the same conditions and step-size, Figure 4.4.2.
In Figures 4.4.3 and 4.4.4 we compare the fourth order classical Magnus method with a remarkable performance of the fourth order FM method, applied to the Airy equation with a large step-size equal to h= 12. While in Figures 4.4.5 and 4.4.6 we show the solution for the Airyequation with the fourth order FM method with step-sizes h= 14 and h= 101 respectively.
4.4 The FM method
0 500 1000 1500 2000
−3
−2
−1 0 1 2
x 10
−10Figure 4.4.1: Global error of the fourth order Modified Magnus method with exact evaluation of integral commutators for the Airy equationy′′(t) =−ty(t) with [1,0]⊺ initial conditions, 0≤t≤2000 and step-size h= 15.
0 500 1000 1500 2000
−3
−2
−1 0 1 2
x 10
−10Figure 4.4.2: Global error of the fourth order FM method with end points only and multiplicities all 2 for the Airy equation y′′(t) =−ty(t) with [1,0]⊺ initial conditions, 0 ≤t≤2000 and step-size h = 15.
4.4 The FM method
0 500 1000 1500 2000
−4
−2 0 2 4
x 10
−7Figure 4.4.3: Global error of the fourth order FM method with end points only and multiplicities all 2 for the Airy equation y′′(t) =−ty(t) with [1,0]⊺ initial conditions, 0 ≤t≤2000 and step-size h = 12.
0 500 1000 1500 2000
−0.015
−0.005 0.005 0.015
Figure 4.4.4: Global error of the fourth order Magnus method for the Airy equation y′′(t) = −ty(t) with [1,0]⊺ initial conditions, 0 ≤ t ≤ 2000 and step-size h= 12.
4.4 The FM method
0 500 1000 1500 2000
−1 0 1
x 10
−9Figure 4.4.5: Global error of the fourth order FM method with end points only and multiplicities all 2, for the Airy equation y′′(t) =−ty(t) with [1,0]⊺ initial conditions, 0 ≤t≤2000 and step-size h = 14.
0 500 1000 1500 2000
−4
−2 0 2
4 x 10
−11Figure 4.4.6: Global error of the fourth order FM method with end points only and multiplicities all 2,for the Airy equation y′′(t) =−ty(t) with [1,0]⊺ initial conditions, 0 ≤t≤2000 and step-size h = 101.
Chapter 5
Highly oscillatory linear systems with variable coefficients
5.1 The asymptotic method
In this chapter we introduce our reader further applications of the FM meth- ods. In particular, we develop the FM method for solving highly oscillatory systems of ODEs with a time-dependent matrix, [Kha09b], [Kha09a]. We solve the matrix exponential in Lie groups applying Magnus method. For discretizing nested commutators we apply Filon quadrature and develop the FM method.
We commence from a system of ordinary differential equations written in a vector form,
y′ =Aωy+f(t), y(0) =y0 ∈Rd, t≥0,
whereAω is a non-singular matrix with large imaginary eigenvalues,σ(Aω)⊂ iR, kAωk ≫ 1,ω ≫1 is a real parameter describing frequency of oscillation and f :R×Rd→Rd is a smooth vector-valued function.
In present chapter we allow matrixAωto be time-dependent non-singular matrix as compared to a constant matrix Aω, discussed previously and pre- sented in [Kha08b]. For more generality one can assume that Aω can be a function from [0,∞) to the set of matrices. However, the scope of this
5.1 The asymptotic method
thesis is to provide answers while Aω is the source of the oscillation occur- ring in the system and the methods introduced in our work are beneficial for large frequencies of oscillation. While for small frequencies or even in absence of oscillation our discretization methods are of the same order as Gauss-Christopher rule with Gaussian point.
The analytic solution of the system is given by the variation of constants formula,
y(t) =Xω(t)y0+ Z t
0
Xω(t−τ)f(τ,y(τ))dτ =Xωy0+I[f]. (5.1.1) The fact that Aω is no longer a constant matrix results in a more complex solution of a fundamental matrix Xω, satisfying matrix linear ODE,
Xω′ =AωXω, Xω = eΩ,
where Ω represents Magnus expansion, an infinite recursive series, Ω(t) =
Z t 0
A(t)dt +1
2 Z t
0
[A(τ), Z τ
0
A(ξ)dξ]dτ +1
4 Z t
0
[A(τ), Z τ
0
[A(ξ), Z ξ
0
A(ζ)dζ]dξ]dτ + 1
12 Z t
0
[[A(τ), Z τ
0
A(ξ)dξ], Z τ
0
A(ζ)dζ]dτ +· · · .
The challenge is to evaluate highly oscillatory integral I[f] in (5.1.1) effi- ciently. We already know that ordinary methods based onGauss–Christoffel quadrature fail to approximate highly oscillatory integrals for increasing ω, [IN05]. Hence, we first focus on developing numerical methods to evaluate highly oscillatory integrals with a matrix-valued kernel and next we use our results in numerical approximation to (5.1.1).
Given a vector-valued integral over a compact interval [a, b]
I[f] = Z b
a
Xω(t)f(t)dt, Xω′ =AωXω, Xω = eΩ,
5.1 The asymptotic method
with an arbitrary non-singular matrix Aω of large eigenvalues, kA−ω1k ≪ 1, σ(Aω)⊂iR, a real parameterωand a smooth vector-valued functionf ∈Rd. We would like to remind our reader that if the matrix Aω was constant, then for example the task of obtaining an asymptotic expansion for the inte- gral I[f] was quite simple, [Kha08b]. Applying rules of integration by parts,
I[f] =A−ω1
Xω(b)f(b)−Xω(a)f(a)
−A−ω1 Z b
a
Xω(t)f′(t)dt
=QA1 −A−ω1I[f′],
we obtain theasymptotic methodQAs, defined ass-partial sum of the asymp- totic expansion for I[f],
QAs[f] =−
s
X
m=1
(−Aω)−mh
Xω(b)f(m−1)(b)−Xω(a)f(m−1)(a)i .
This is not the case when the matrix Aω is arbitrary, since substitution Xω =A−ω1Xω′ inI[f] is no longer useful to apply tools of integration by parts,
I[f] = Z b
a
A−ω1(t)Xω′(t)f(t)dt.
However, it is possible to overcome this issue and obtain the asymptotic expansion for I[f] given an arbitrary matrixAω.
In this section we show how applying some matrix transformations inI[f] leads to its explicit asymptotic expansion.
Given, I[f] =
Z b a
Xω(t)f(t)dt, Xω′ =AωXω, Xω = eΩ, (5.1.2) with a matrix valued kernel Xω depending on a real parameter ω ≫ 1 and satisfying matrix linear differential equation (5.1.2). Assume that Aω is an arbitrary non-singular matrix of pure imaginary spectrum, σ(Aω)⊂iR, kA−ω1k ≪1 and f ∈Rd is a smooth vector-valued function.
We have already mentioned previously in this chapter that the current representation of the integral I[f] in (5.1.2) is not suitable for obtaining its
5.1 The asymptotic method
asymptotic expansion due to a more general choice of the matrix Aω. Once we substitute Xω =A−ω1Xω′ in I[f] in its current occurrence, then
I[f] = Z b
a
A−ω1(t)Xω′(t)f(t)dt.
Transforming integral (5.1.2) into an equivalent integral with appropriate representation for our purposes helps to obtain the asymptotic expansion to (5.1.2).
We first vectorize matrixXω into
Xω = [¯x1,x¯2, ...,x¯d].
Next step we take the transpose of Xω, obtaining a corresponding vector X¯ω,
X¯ω = [x¯1,x¯2, ...,x¯d]⊺.
It is easy to verify that the latter satisfies a linear differential equation X¯′ω =BωX¯ω, with Bω =
d
M
i=1
Aω,
where Bω represents direct d-tuple sum of the original matrixAω.
We also scale an identity matrix I by taking its direct product with the vector-valued function f := f⊺ = [f1, f2, . . . , fd], F = f1×dN
Id×d. For instance, take
f = [f1, f2] and I =
1 0 0 1
. Direct product F =fN
I results in the following matrix, F =
f1 0 f2 0 0 f1 0 f2
. Thus, F =
d
S
i=1
Fi, where each entry matrix is a diagonal matrix with the corresponding element fi on the diagonal,
Fi =
fi 0 . . . 0 0 fi . . . 0
0 0 fi 0
... . .. ... ...
0 . . . 0 fi
.
5.1 The asymptotic method
These transformations result in equalityXωf =FX¯ω,and appear to be suit- able for integration by parts and representation of the asymptotic expansion of the integral (5.1.2).
Example 5.1.1 Let Xω =
x11 x12
x21 x22
, Aω =
a11 a12
a21 a22
and f = f1
f2
. Then
X¯ω =
x11
x21 x12
x22
, Bω =
a11 a12 0 0 a21 a22 0 0 0 0 a11 a12
0 0 a21 a22
and
F =
f1 0 f2 0 0 f1 0 f2
. Hence,
Xωf =FX¯ω =
x11f1+x12f2
x21f1+x22f2
, and X¯′ω =BωX¯ω, once Xω′ =AωXω.
As a result, we have obtained the following equality for the two integrals, I[f] =
Z b a
Xω(t)f(t)dt= Z b
a
F(t)X¯ω(t)dt=I[F], where matrix Xω and vectorX¯ω satisfy linear ODEs
Xω′ =AωXω and X¯′ω =BωX¯ω
respectively. This transformation allows us to approximate I[F] and thus I[f]. Applying integration by parts toI[F] we obtain the asymptotic expan- sion for the integral I[f],
I[f] = Z b
a
F(t)X¯ω(t)dt= Z b
a
F(t)Bω−1(t)X¯′ω(t)dt
=
F(t)Bω−1(t)X¯ω(t)b a−
Z b a
F(t)Bω−1(t)′X¯ω(t)dt.
5.1 The asymptotic method
Hereafter we remind our reader the notation on matrix functions asymp- totics depending on a real parameter ω. For the entirety of our work, all norms are L∞ norms, for vectors, matrices and functions. The norm of a function is taken over the interval [a, b]. We say that f = O( ˜f) for an arbitrary function f and non-negative constant ˜f, which depend on a real parameter ω, if the norm of f and its derivatives are all of order O( ˜f) as ω → ∞, namely kf(m)k = O( ˜f) for m = 0,1, . . .. For arbitrary two n×m matrices A(x) = (aij(x)) and ˜A = (˜aij), ˜aij ≥ 0, depending on a real pa- rameter ω, we can thus posit A(x) = O( ˜A), if aij(x) = O(˜aij) element-wise as ω → ∞. We may also say that f = O(1), if f and its derivatives remain bounded on [a, b], as ω → ∞. Let 1 ={1ij} stand for a matrix with all en- tries one. This allows us to write A(x) =O(1), ifaij(x) =O(1) element-wise as ω → ∞. And finally, if A = O( ˜A) and B = O( ˜B), then the integration and multiplication properties are Rb
a A(x)dx =O( ˜A) and AB =O( ˜AB).˜ Letting
σ0=F,
σk+1= (σkBω−1)′, k= 0,1, . . . we define asymptotic method as
QAs[f] =
s−1
X
k=0
(−1)k σk(b)Bω−1(b)X¯ω(b)−σk(a)B−ω1(a)X¯ω(a) .
Below we Theorem 3.1 from [Olv07] and Lemma 2.1 from [Kha08b]. In Theorem 5.1.3 we generalize these results for a matrix-valued kernel Xω, a vector-valued function f and a time dependant matrix Aω inI[f], (5.1.2) . Theorem 5.1.1 (S. Olver, [Olv07]) Suppose that y satisfies the differential equation,
y′(x) =A(x)y(x),
in the interval [a, b], for some invertible matrix-valued function A such that A−1 =O( ˆA), for ω → ∞. Assume that
I[f] = Z b
a
f⊺(x)y(x)dx,
5.1 The asymptotic method
where f ∈ Rd is a smooth vector-valued function and y ∈ Rd is a smooth, highly oscillatory vector- valued function. Define
QAs[f] =
s−1
X
k=0
(−1)k[σk⊺(b)A−1(b)y(b)−σk⊺(a)A−1(a)y(a)], where
σ0 ≡f, σk+1 = (A−⊺σk)′, k = 0,1, . . . . If f =O( ˜f) and y(x) =O(y)˜ for a≤x≤b, then
I[f]−QAs[f] = (−1)s Z b
a
σs⊺ydx=O(f˜⊺Aˆs+1y),˜ as ω → ∞. Lemma 5.1.2 (MK, [Kha08b]) Let
I[f] = Z b
a
Xω(t)f(t)dt, Xω′ =AωXω,
where the matrix kernel Xω satisfies linear matrix ODE as above, Aω is a constant non-singular matrix, kA−ω1k ≪ 1 and f : R → Rd is a smooth vector-valued function. Then, for ω≫1,
I[f]∼ − X∞
m=1
(−Aω)−m
Xω(b)f(m−1)(b)−Xω(a)f(m−1)(a) .
For ψ = max{kf(s)k,kf(s+1)k},
QAs[f]−I[f]∼O(kA−ωs−1kkXωkψ), as ω→ ∞. If Xω =O( ˆXω) and f =O(f˜), then
QAs[f]−I[f] = O(A−ωs−1Xˆωf˜), as ω → ∞, element wise.
5.1 The asymptotic method
Theorem 5.1.3 [Kha09b] Postulate that I[f] =
Z b a
Xω(t)f(t)dt, and Xω′ =Aω(t)Xω,
whereAω is an arbitrary non-singular matrix,σ(Aω)⊂iR,ωis an oscillatory parameter and f ∈Rd is a smooth vector-valued function.
If Fω =O(Fˆ), B−ω1 =O( ˆB) and X¯ω =O(Xˆω), then
I[f]−QAs[f] = (−1)s Z b
a
σs(t)X¯ω(t)dt
= O(FˆBˆs+1X),ˆ as ω → ∞. Proof: We first prove the identity by induction,
I[F] =
F(t)Bω−1(t)X¯ω(t)b a−
Z b a
F(t)Bω−1(t)′X¯ω(t)dt
=
σ0(t)Bω−1(t)X¯ω(t)b a−
Z b a
σ1(t)X¯ω(t)dt.
Suppose the identity holds for some k ∈N, we prove fork+ 1.
Z b a
σk+1(t)X¯ω(t)dt= Z b
a
σk+1(t)Bω−1(t)X¯′ω(t)dt
=
σk+1(t)Bω−1(t)X¯ω(t)b a−
Z b a
σk+2(t)X¯ω(t)dt.
By induction, σk =O(FˆBˆk). Indeed, fork = 0,σ0 =F =O(Fˆ). We assume the equality holds for some σk, then
σk+1=
σkBω−1′
=σ′kBω−1+σkBω−1′
=O(FˆBˆk)O( ˆB) +O(FˆBˆk)O( ˆB) =O(FˆBˆk+1).
And finally, Z b
a
σsX¯ωdt=
σsBω−1X¯ωb a−
Z b a
σs+1X¯ωdt
= O(FˆBˆs+1X) + O(ˆ FˆBˆs+1X) = O(ˆ FˆBˆs+1Xˆ).
5.1 The asymptotic method
Corollary 5.1.4 [Kha09b] If
f(k)(a) =f(k)(b) = 0, for k = 0,1, ..., s−1, then,
I[f] = O(FˆBˆs+1Xˆ).
Example 5.1.2 For practical purposes consider integral
Ih[f] = Z h
0
eΩ(h−τ)f(τ)dτ.
Once we vectorize the system to obtain asymptotic method, I¯stands for a vectorized identity matrix I, andX¯ω stands for a vectorized matrix eΩ. For s = 2 the asymptotic method using information at the end points only, is
QA2 = [F(h)Bω−1(0)−F′(h)Bω−2(0)−F(h)Bω−1′(0)Bω−1(0)]I¯
−[F(0)Bω−1(h)−F′(0)Bω−2(h)−F(0)Bω−1′(h)B−ω1(h)]X¯ω(h).
Employing initial-value integrator, yn+1 = eΩsyn+
Z h 0
eΩs(h−τ)f(tn+τ)dτ,
where Ωs stands for a truncatedMagnus expansion, we can now solve (5.1.1) with the asymptoticmethod,
yn+1 = eΩsyn+QAs.
We solve matrix exponential eΩs with theFMmethod described in [Kha09a].
The method combines the ideas ofLie groupsolvers, namelyMagnusmethod, with Filon quadrature for solving nested integral commutators in Ω.
The following lemma for analysis of highly oscillatory ODEs instead of Taylor method, present a paragraph on it
Lemma 5.1.5 [Kha09a] The approximation with the asymptotic method QAs applied to an integral I[f], is independent of the step-size of integration.
5.2 The FM method
Proof: For∀step-sizehand∀numbernof node pointstn =tn−1+h, the components evaluated at the internal node points of the truncated asymptotic sum QAs appear in pairs with an opposite sign, except the values at the end points
QAs[f](t1)−QAs[f](t0) +QAs[f](t2)−QAs[f](t1) +QAs[f](t3)−QAs[f](t2) +...
+QAs[f](tn−1)−QAs[f](tn−2) +QAs[f](tn)−QAs[f](tn−1)
=QAs[f](tn)−QAs[f](t0).
As a result, no matter how we truncate the interval we obtain the same representation for the asymptotic method QAs evaluated at the end points.
5.2 The FM method
Employing classical Filon-type quadrature proven to have high accuracy for increasing ω, as in [Kha08a], we construct anr-degree polynomial interpola- tion v for the vector-valued function f in (5.1.2),
v(t) =
ν
X
l=1 θl−1
X
j=0
αl,j(t)f(j)(tl),
which agrees with function values and derivatives v(j)(tl) =f(j)(tl) at node points a =t1 < t2 < ... < tν = b, with associated θ1, θ2, ..., θν multiplicities, j = 0,1, ..., θl−1, andl = 1,2, ..., ν, [Kha09b].
By definition, Filon-type method is QFs[f] =
Z b a
Xω(t)v(t)dt=
ν
X
l=1 θl−1
X
j=0
βl,jf(j)(tl),
where βl,j =Rb
a Xω(t)αl,j(t)dt.
Theorem 5.2.1 [Kha09b] Postulate that
I[f] = Z b
a
Xω(t)f(t)dt, and Xω′ =Aω(t)Xω,
5.2 The FM method
Aω is a non-singular matrix, σ(Aω) ⊂ iR, ω is an oscillatory parameter and f ∈Rd is a smooth vector-valued function.
If Fω =O(Fˆ), B−ω1 =O( ˆB) and X¯ω = O(Xˆω), then
I[f]−QFs[f] = (−1)s Z b
a
σs(t)X¯ω(t)dt
= O(FˆBˆs+1X),ˆ as ω → ∞.
Proof: Observing that I[f −v] = I[f]−QFs[f], the proof follows by replacing f in the asymptotic method (5.1.3) by f −v and the fact that [f −v](j)(a) =[f −v](j)(b) = 0, for j = 0,1, ..., s−1.
Corollary 5.2.2 [Kha09b] If Xω =O(1) and f =O(1), then QFs[f]−I[f] = O(A−ωs−11).
Theorem 5.2.3 [Kha09b] Let θ1, θν ≥ s, r = Pν
l=1θl −1. Then r is the numerical order of the Filon-type method applied to the linear systems,
y(tn)−yn =O(hr+1).
Numerical experiments withQF Ms method, a combination ofFilonquadra- ture and Magnus method, are available in Figures 5.2.1, 5.2.2, 5.2.3, 5.2.4, 5.2.5, 5.2.6 and 5.2.7.
By definition, FM method is QF Ms [f] =
Z b a
eΩs(t)v(t)dt=
ν
X
l=1 θl−1
X
j=0
βl,jf(j)(tl),
where βl,j = Rb
a eΩs(t)αl,j(t)dt. In our case it is numerical discretization of the integral
I[f] = Z h
0
eΩsf(tn+τ)dτ,
5.3 The modified FM method
applied to solve
yn+1 = eΩsyn+QF Ms [f].
We present numerical results forFM method QF M2 with end points only and multiplicities all 2, for the equation y′′(t) =−ty(t)−cos(t), 0≤t≤100, with [1,0]⊺ initial conditions and various step-sizes in Figures 5.2.1-5.2.7.
Initially, we show that employing the FM method one can afford large step- sizes, such ash= 1 andh= 12,Figures 5.2.1. Decreasing step-size up toh= 14 and h= 101 we obtain accuracy up to 10−8, Figure 5.2.2. In Figure 5.2.4 we show that for a fixed large step-size the accuracy of approximation improves with an increase in frequency,h= 14, ω= 102and 103.For a smaller step-sizes we demonstrate high accuracy in approximation for ω = 104, Figure 5.2.5 and ω = 105, Figure 5.2.6. While in Figure 5.2.7 we show that for ω = 106 and hω = 104 the method introduces highly desirable accuracy, the error is 10−14. This can be compared with Figure 5.2.3 for the same step-size size.
5.3 The modified FM method
In our further discussion we develop a modified FM method QF MS employ- ing ideas from the modified Magnus method, [Ise02c], [Kha09a]. By solving the equation for variable constants locally with a constant matrix we intro- duce local change of variables. Applications of the modified Magnus method include systems of highly oscillatory linear systems of ODEs, [Kha09a]
y′(t) =Aω(t)y(t) +f(t), y(t0) =y0, with analytic solution
y(t) =Xω(t)y0+ Z t
0
Xω(t−τ)f(τ)dτ =Xωy0+I[f].
We assume that the spectrum of the matrixAω(t) has large imaginary values.
Introducing local change of variables at each mesh point, we write the solution in the form
y(t) = e(t−tn)A(tn+12h)x(t−tn), t ≥tn.
5.3 The modified FM method
0 20 40 60 80 100
−0.01
−0.005 0 0.005 0.01
0 20 40 60 80 100
−2
−1 0 1 2 x 10
−4Figure 5.2.1: Global error of the FM methodQF M2 with end points only and multiplicities all 2, for the equation y′′(t) = −ty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions and step-size h= 1 and h= 12 (top two figures from left to right)
We apply modifiedMagnustechniques to the system and solve the system locally with a constant matrix Aω, while x(t) satisfies a highly oscillatory equation
x′(t) =Bω(t)Xω(t) + e−tAω(tn)f(t), t≥0, with the same matrix Bω as for linear matrix ODEs,
Bω(t) = e−tAω(tn)[Aω(t)−Aω(tn)]etAω(tn).
It is possible to reduce cost of evaluation of the matrixB(t) by simplifying it, [Ise04a]. Denoteq=q
g(tn+12h) andv(t) = g(tn+t)−g(tn+12h). Then, B(t) =v(t)
(2q)−1sin 2qt q−2sin2qt
−cos2qt −(2q)−1sin 2qt
=v(t)
q−1sinqt
−cosqt
cosqt q−1sinqt ,
5.3 The modified FM method
0 20 40 60 80 100
−4
−2 0 2 4 x 10
−60 20 40 60 80 100
−4
−2 0 2 4 x 10
−8Figure 5.2.2: Global error of the FM methodQF M2 with end points only and multiplicities all 2, for the equation y′′(t) = −ty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions and step-size h = 14 and h = 101 (from left to right).
and for the product
B(t)B(s) =v(t)v(s)sinq(s−t) q
q−1sinqt
−cosqt
cosqs q−1sinqs . It was shown in [Ise04a] that
kB(t)k= cos2ωt+sin2ωt
ω2 and B(t) =O(t−tn+1
2).
This approximation results in the following algorithm, yn+1= ehA(tn+h/2)xn
xn= eΩ˜nyn.
Given the compact representation of the matrix B(t) with oscillatory entries, we solve Ω(t) with a Filon-type method, approximating function
5.3 The modified FM method
0 20 40 60 80 100
−4
−2 0 2
4 x 10
−13Figure 5.2.3: Global error of the FM methodQF M2 with end points only and multiplicities all 2, for the equation y′′(t) = −ty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions and step-size h= 1001 .
v(t) by a polynomial ˜v(t), for example Hermite polynomial, as in classical Filon-type method. Furthermore, in our approximation we use end points only, although the method is general and more nodes of approximation can be required.
As a result we can now solve (5.1.2) we Modified Magnus method in combination with Filon quadrature, which we call the modified FM method, and this proves to be more accurate approximation than solving if we solved the commutators with Gaussian quadrature. Performance of the modified Magnus method improves as compared to classical Magnus method due to a number of reasons. The fact due to the fact that the matrix B is small, B(t) = O(t −tn+1
2), contributes to higher order correction to the solution, [Ise02b], [Kha09a]. The order of the modified Magnus method increase from p= 2s top= 3s+ 1.
5.3 The modified FM method
0 20 40 60 80 100
−1.5
−1
−0.5 0 0.5 1 1.5x 10−6
0 20 40 60 80 100
−2
−1 0 1 2x 10−7
Figure 5.2.4: Global error of the QF M2 FMmethod with end points only and multiplicities all 2, for the equation y′′(t) = −ωty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions and step-size h= 14 and ω= 102 ω = 103 (from left to right).
0 20 40 60 80 100
−1.5
−1
−0.5 0 0.5 1 1.5x 10−9
0 20 40 60 80 100
−4
−2 0 2 4x 10−11
Figure 5.2.5: Global error of the QF M2 FMmethod with end points only and multiplicities all 2, for the equation y′′(t) = −ωty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions, ω = 104, step-sizes h = 101 and h = 251 (from left to right),
5.3 The modified FM method
0 20 40 60 80 100
−3
−2
−1 0 1 2 3x 10−12
0 20 40 60 80 100
−6
−4
−2 0 2 4 6x 10−14
Figure 5.2.6: Global error of the QF M2 FMmethod with end points only and multiplicities all 2, for the equation y′′(t) = −ωty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions and step-size h = 1001 with ω = 25×102 and ω = 105 (from left to right.)
0 20 40 60 80 100
−2
−1 0 1
2 x 10
−14Figure 5.2.7: Global error of the QF M2 FMmethod with end points only and multiplicities all 2, for the equation y′′(t) = −ωty(t)−cos(t), 0 ≤ t ≤ 100, with [1,0]⊺ initial conditions, step-size h= 1001 and ω = 106.
Chapter 6
Highly oscillatory non-linear systems
6.1 Waveform relaxation methods
In this chapter we present iteration techniques, namely, waveform relaxation algorithms for solving non-linear ordinary differential equations with associ- ated initial conditions:
d
dtz(t) =f(t,z), z(0) =z0, t∈[0, T], (6.1.1) where T ≥0 f : [0, T]×Rn→Rn, z0 = [z1,0, z2,0, ..., zn,0]∈Rn is the vector of initial values, andz(t) = [z1(t), z2(t), ..., zn(t)]∈Rn is the solution vector.
In other words the system is written as follows, d
dtz1(t) =f1(t, z1, z2, ..., zn), z1(0) =z1,0
d
dtz2(t) =f2(t, z1, z2, ..., zn), z2(0) =z2,0
... ...
d
dtzn(t) =fn(t, z1, z2, ..., zn), zn(0) =zn,0.
6.1 Waveform relaxation methods
A simple example of waveform relaxation scheme, which maps the previous iterate z(ν−1) into new iterate z(ν) is,
d
dtz1(ν)(t) =f1(t, z(ν)1 , z2(ν−1), z3(ν−1), ..., zn(ν−1)), z1(ν)(0) =z1,0 d
dtz2(ν)(t) =f2(t, z(ν)1 , z2(ν), z(ν3 −1), ..., zn(ν−1)), z2(ν)(0) =z2,0
... ...
d
dtzn(ν)(t) =fn(t, z1(ν), z2(ν), z3(ν), ..., zn(ν)), z(ν)n (0) =zn,0. The above relaxation scheme is calledGauss-Seidelwaveform relaxation, similarly toGauss-Seideliterative method to solve linear and non-linear sys- tems of algebraic equations. The iterative methods simplify the problem of solving large systems of differential equations of n variables to a sequence of differential equations of a single variable. Another example of iteration scheme is the Jacobi waveform relaxation method,
d
dtz1(ν)(t) =f1(t, z1(ν), z2(ν−1), z(ν3 −1), ..., zn(ν−1)), z1(ν)(0) =z1,0
d
dtz2(ν)(t) =f2(t, z1(ν−1), z2(ν), z(ν3 −1), ..., zn(ν−1)), z2(ν)(0) =z2,0
... ...
d
dtzn(ν)(t) =fn(t, z(ν1 −1), z(ν2 −1), z3(ν−1), ..., zn(ν)), zn(ν)(0) =zn,0.
In both cases the iterations scheme picks up initial conditions as an initial approximation z(0)(t)
z(0)i (t) =zi,0, t ∈[0, T], i= 1, ..., n,
which is defined along the whole interval of integration. Easy to notice that unlikeJacobialgorithm, theGauss–Seidelmethod is sequential, i.e. one have to solve equations one after another, while the Jacobi scheme can be solved in parallel, simultaneously.
Let us definewaveform relaxation operator F as z(ν)=F(z(ν−1)).
6.1 Waveform relaxation methods
Then a one-step iteration method can be written as d
dtz(ν)(t) =F(t,z(ν−1),z(ν)), z(ν)(0) =z0, where F(u, v) is defined so, that F(t, v, v) =f(t, v).
The well known Picard iteration, as well as Jacobi and Gauss–Seidel schemes can be written in the formulae,
Fi(t, u, v) =Fi(t, u1, u2, ..., ui, ..., un), (Picard), Fi(t, u, v) =Fi(t, u1, u2, ..., ui−1, vi, ui+1, ..., un), (Jacobi), Fi(t, u, v) =Fi(t, v1, v2, ..., vi−1, vi, ui+1, ..., un), (Gauss-Seidel).
Convergence analysis ofwaveform relaxationmethods is presented [Van93], [Nev89a], [Nev89b] and [MN87]. The convergence of sequences {z(ν)}∞ν=0 is analysed in the context of Banach spaces.
We consider continuous vector-valued functions defined on [0, T], i.e.
C([0, T];Rn), with maximum of the norm, defined as:
kzkT = max
t∈[0,T]kz(t)k, with vector norm k · k inRn.
Definition 6.1.1 Let (X,k · kX) be a normed linear space and the associated norm. An operator U:X→Xis called a contraction if there exists a γ, with 0≤γ <1, such that for all x, y ∈X it is true that
kU(x)−U(y)k ≤γkx−ykX.
It is shown in [Van93] that thewaveform relaxationis a contraction map- ping.
Theorem 6.1.1 [Van93] Let X be a Banach space and U a contraction.
Then there is a unique x⋆ ∈ X, such that U(x⋆) = x⋆. Moreover, if x(0) is any point in X, and we define the sequence {x(ν)}∞ν=0 by x(1) = U(x(1)), then x(ν) →x⋆ as ν→ ∞.