Anda di halaman 1dari 5

Optics Communications 282 (2009) 1435–1439

Contents lists available at ScienceDirect

Optics Communications
journal homepage: www.elsevier.com/locate/optcom

Adaptive higher-order split-step Fourier algorithm for simulating


lightwave propagation in optical fiber
Xueming Liu
State Key Laboratory of Transient Optics and Photonics, Xi’an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi’an 710119, China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we propose an adaptive step-size control algorithm for solving nonlinear Schrödinger’s
Received 16 August 2008 equations. The proposed algorithm has a fourth-order local accuracy and is a system-independent rule
Received in revised form 8 December 2008 for adjusting the step sizes. This algorithm has two potential advantages, an automatic step adjustment
Accepted 8 December 2008
mechanism and higher-order accuracy. A test example shows that, by comparing to the fixed step-size
method and the local-error method, our method decreases the computational time by about 10 and 70
times, respectively. The performance of the proposed method is validated and compared to commonly
Keywords:
used step-size selection methods by simulating the evolution of a third-order soliton, the collision of a
Nonlinear Schrödinger’s equations
Split-step Fourier method
fundamental soliton pair, and the supercontinuum generation. Numerical simulations show that the pro-
Soliton posed method can increase the computational efficiency by more than one and two orders of magnitude
in comparison with the symmetrized split-step Fourier method. In addition, the computational efficiency
is improved with the increase of the accuracy of solutions.
Ó 2008 Elsevier B.V. All rights reserved.

1. Introduction is based on the fiber parameters such as nonlinearity and/or


dispersion.
Differential equations have been a powerful tool for describing In this paper, an adaptive split-step Fourier algorithm with the
the behavior of a wide variety of physical phenomena. The nonlin- fourth-order local accuracy is proposed to solve the NLSE describ-
ear Schrödinger’s equations (NLSEs) [1–5], which are nonlinear ing lightwave propagation in optical fiber. The numerical results
partial differential equations, have been applied to many nonlinear show that our method increases the computational speed by more
problems such as nonlinear optics, fluid dynamics, plasma physics, than one and two orders of magnitude in comparison with the
and wave propagation in the atmosphere. Especially, the standard symmetrized (S) SSFM.
or generalized (G) NLSEs have extensively been used for modeling
of optical solitons [6], supercontinuum generation [7,8], and opti- 2. Generalized nonlinear Schrödinger’s equations
cal communications. Although analytical solutions may be found
for some simple problems, the complexity of most cases requires A form of the G-NLSE for the wave propagation in optical fiber is
numerical simulations frequently. The split-step Fourier method described as [3,9,17]
(SSFM) costs the least computational time among all compared
numerical schemes for NLSEs because it can be attributed to the @uðz; TÞ
use of fast Fourier-transform (FFT) [1–3]. ¼ ðD þ NÞuðz; TÞ; ð1Þ
@z
Usually, there are two ways to further increase the computa-
!
tional efficiency of SSFM. One way is based on the higher-order a X i
n1
@n
split-step Fourier schemes. For example, the four-order schemes D¼  bn ; ð2Þ
2 nP2
n! @T n
constructed from the Runge–Kutta method, Adams method, pre-
dictor–corrector method etc. had been proposed [2,4,9–11]. These   Z 1 
1 1 @
kinds of methods are usually constant step-size methods although N ¼ ic 1þ ð1  fR Þujuj2 þ fR u hR ðsÞjuðz; T  sÞj2 ds :
they can achieve higher local error. The other way is attributed to u x0 @T 0

the performance of adaptive step-size control [12]. Furthermore, ð3Þ


the suitable step-size selection rule also enhance the computa-
tional efficiency of SSFM [5,13–16], but the selection of step size Here a is the attenuation coefficient, bn are the higher-order disper-
sion coefficients, D is a linear operator including the dispersion and
the loss of fiber, c is the nonlinear parameter, x0 is the carrier fre-
E-mail address: liuxm@opt.ac.cn quency, N is a nonlinear operator, and u(z, T) is a complex-valued

0030-4018/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.optcom.2008.12.051
1436 X. Liu / Optics Communications 282 (2009) 1435–1439

function of T and z. T = t  b1z and z is the abscissa along the fiber. fR unþ1;1 ¼ expðhFðzn ; un ÞÞ  un ; ð14Þ
denotes the fractional contribution of the delayed Raman’s response
to nonlinear polarization, and hR represents the Raman response unþ1;2 ¼ expf0:5hFðzn þ 0:5h; un  exp½0:5hFðzn ; un ÞÞg
function of a silica fiber.  exp½0:5hFðzn ; un Þ  un ; ð15Þ

3. Novel method unþ1;3 ¼ u2nþ1;2 =unþ1;1 : ð16Þ

Let F = D + N (here we use F to substitute D + N only for a short- By using the S-SSFM [3], the numerical solutions of Eqs. (14) and
ened notation), then Eq. (1) can be rewritten as (15) can further be approximated as
@uðz; TÞ Z !
zþh
¼ D þ N ¼ F: ð4aÞ
uðz; TÞ  @z unþ1;1  exp½hD=2 exp NðfÞdf exp½hD=2un ; ð17Þ
z
Although D, N, and F are operators, they have a physical mean-
ing at each step of the simulation. For example, on/oTn transfers to   Z zþh !  
hD4 hD3
(ix)n for operating the dispersion operator D at each step. By using unþ1;2  exp exp Nð1Þd1 exp
4 zþh=2 4
the relationship ou(z, T)/u(z, T) = o[ln(u(z, T))], Eq. (1) can be written
  Z zþh=2 !  
as hD2 hD1
 exp exp NðnÞdn exp  un : ð18Þ
@½lnðuðz; TÞÞ 4 z 4
¼ F: ð4bÞ
@z
Eq. (18) shows that the lightwave propagation in range of z to z + h
Let us express y(z,T) as is divided into four steps. Each dispersion region h/4 corresponds to
yðz; TÞ ¼ ln½uðz; TÞ: ð4cÞ a dispersion operator Dk (k = 1, 2, 3, 4). If the dispersion along the
length of the fiber is uniform, the operator Dk have the same value,
Now by substituting Eq. (4c) into Eq. (4b), we obtain i.e., D1 = D2 = D3 = D4. The nonlinear operator N is integrated over
@yðz; TÞ=@z ¼ F: ð5Þ two separate intervals, i.e., [z, z + h/2] and [z + h/2, z + h]. According
to the theoretical analysis of S-SSFM, the local errors of Eqs. (17)
We denote y(z) to be the exact solution of Eq. (5) and the initial and (18) are the third-order of the step-size h, i.e., O(h3). As a result,
condition is assumed to be yn = y(zn). The solution after a step-size the local error of Eq. (16) is the fourth-order of h, i.e., O(h4).
h (i.e., zn+1 = zn + h) can be approximated as [18,19] The corresponding er in Eq. (12) is changed as
ynþ1;1 ¼ yn þ h  Fðzn ; yn Þ; ð6Þ er ¼ j lnðjjunþ1;2 jjÞ  lnðjjunþ1;1 jjÞj: ð19Þ
R
ynþ1;2 ¼ yn þ 0:5hFðzn ; yn Þ þ 0:5hFðzn þ 0:5h; yn þ 0:5hFðzn ; yn ÞÞ; Here, the norm ||u|| is defined as ||u|| = ( |u(z)|2dt)1/2. Therefore, on
the basis of Eqs. (13)–(19) an automatic step-size control algorithm
ð7Þ
can be constructed.
ynþ1;3 ¼ 2ynþ1;2  ynþ1;1 : ð8Þ
4. Adaptive step-size control algorithm
Here the operator F is only a shortened notation of D + N. Eqs. (6)–
(8) and the operator F are used for mathematical derivations. We The adaptive step-size control allows the solution of problems
use them to derive Eqs. (14)–(16) and the obtained expressions by prescribing error tolerances rather than specifying a step size.
i.e., Eqs. (14)–(16) have the full physical meaning. After Taylor’s In the step-size control technique, two estimates of a solution
expansion, the corresponding local truncation errors for Eqs. (6)– are employed to evaluate the error. We can use this estimate of
(8) are given by, respectively, the error to either accept or reject the current step-size and esti-
e1 ¼ yðzn þ hÞ  ynþ1;1 ¼ Kh2 þ O1 ðh3 Þ; ð9Þ mate a new step-size for future steps. The realization of adaptive
algorithm is shown in the following.

e2 ¼ yðzn þ hÞ  ynþ1;2 ¼ 0:5Kh2 þ O2 ðh3 Þ; ð10Þ INITIALIZE: Given an initial tolerance e0, minimum and
maximum step-size hmin and hmax.
e3 ¼ yðzn þ hÞ  ynþ1;3 ¼ 2O2 ðh3 Þ  O1 ðh3 Þ ¼ O3 ðh3 Þ; ð11Þ Step 0: 1z = z0, h = hmin, FLAG = 1.
 Step 1: While (FLAG = 1), do STEP 2–7.
2 
where K ¼ 12 d dzyðzÞ
2  . Step 2: Compute Eqs. (16)–(19), and then achieve the solution
z¼z0

We can observe from Eqs. (9)–(11) that Eq. (8) can offer higher-
un+1,3 and the errorpestimate
ffiffiffiffiffiffiffiffiffiffiffi er.
Step 3: Compute d ¼ 0:8  e0 =er .
order solution than that of Eq. (6) or Eq. (7). The difference be-
Step 4: If er 6 e0 , un+1, 3 is accepted and z = z + h.
tween Eq. (9) and Eq. (10) is
Step 5: If d 6 0.1, then h = 0.1 h.
er ¼ e1  e2 ¼ ynþ1;2  ynþ1;1  0:5Kh2 : ð12Þ
Else if d P 4, then h = 4h.
If er is less than a desired error e0 (i.e., er < e0), we can accept yn+1,3 as
an approximate value of y(zn + h). In contrast, if er > e0, we reject
Else, set h = d  h.
yn+1,3 and repeat the current step with a new trial step-size h0 . Be-
Step 6: If h > hmax, then h = hmax.
cause K approximates a constant, the step-size h is proportional
pffiffiffiffi Step 7: If z P L, then FLAG = 0.
to er [Eq. (12)]. Therefore, the new and old step sizes have the fol-
lowing relationship among them
pffiffiffiffiffiffiffiffiffiffiffi Else if z + h > L, then h = L  z.
0
h ¼ e0 =er h: ð13Þ
Else if h < hmin, then terminate the program.
From Eq. (4c), let u(z, T) = exp(y(z, T)). Then, Eqs. (6)–(8) can be
Step 8: Stop the program.
changed as
X. Liu / Optics Communications 282 (2009) 1435–1439 1437

In the adaptive control procedure, we should avoid taking too Fig. 1 also shows that the globally relative errors for the LEM
small step-size h or too large h. If the step-size h is smaller than hmin, and the proposed method decrease in range of x = 2–3.6 and
the computation should be stopped (STEP 7). If the step-size h is 2.6–6, respectively. This decrease can be explained as follows.
larger than hmax, the computation will continue with h = hmax (STEP We can see from Eq. (11) that the local truncation error e3 of the
3 3 
6). In STEP 3 and STEP 5, we usepthe new step-size d  h to replace
ffiffiffiffiffiffiffiffiffiffiffi proposed method is 2O2(h3)  O1(h3), where O2 ðh Þ ¼ 16 d dzyðzÞ3  ,
0  z¼n
the old step size i.e., h 0:8  e0 =er h. This is a modified result 3 1 d yðzÞ
3
O1 ðh Þ ¼ 6 dz3  , and both n, f 2 [zn,zn+1]. Therefore, depending
compared to Eq. (13). The modification comes from the fact that z¼f

the stability of algorithm is further enhanced after the step size is on the variable, e3 can be positive or negative. The global error is
shortened by a factor of 0.8. the sum of the local truncation errors and therefore, the relative
error of the proposed method may decrease at some range of the
5. Test and comparison variable. Similar explanations can explain why the global error of
LEM reaches zero at x  3.6. But the local truncation error for the
A typical solution commonly contains both slow variational fixed step-size method is O1(h3) which is always positive and con-
segments, where larger steps are acceptable, and rapid variational sequently, its relative error monotonously increases.
segments, where smaller steps are required. The key to implement-
ing adaptive step-size control is to develop an estimate of the local
6. Numerical solution of NLSE describing solitons and
truncation error. The nonlinear phase-rotation method (NPRM),
supercontinuum generation
the logarithmic step-size method, and the walk-off method as
shown in Refs. [12,15] can adjust their step-size according to the
Numerical solutions of the G-NLSE are often time-consuming.
fiber parameters, such as nonlinearity and/or dispersion. But they
The simulation of G-NLSE describing the evolution of a high-order
are lacking an automatic step adjustment mechanism and are un-
soliton, the collision of a soliton pair, and the supercontinuum gen-
able to construct the really adaptive step-size control algorithm. To
eration are very sensitive to numerical errors [9,12], and hence
author’s best knowledge, the local-error method (LEM) proposed
they are used to evaluate the performance of the proposed method.
by Sinkin et al. is the best one of the automatic step adjustment
Since the S-SSFM and its improved method, NPRM, are widely used
methods [12], and is more excellent than the NPRM, the logarith-
in optical fiber transmission simulators, both are used to compare
mic method, and the walk-off method etc.
to our method in this section.
Because G-NLSE is the complex partial differential equation, it
Fig. 2a–c show the relationship between the CPU time and the
does not have the exact solution. Therefore, it is hard to accurately
global relative error e, where the symbols of star, up-triangle and
know the local truncation and the global errors of numerical solu-
circle denote that they are calculated from the S-SSFM, the NPRM,
tion along the variable. In order to analyze the detailed variation of
and the proposed method, respectively. Here, the global relative er-
the local and global errors along the variable and to show the
ror e is defined by e = ||un  ue||/||ue||, and un and ue are the numer-
advantage of our method, an ordinary differential equation is cal-
ical solution and the exact value, respectively. Fig. 2d illustrates the
culated by the proposed method, the LEM, and the fixed step-size
evolution of step size along the fiber at our method, where the red,
method, respectively
blue and black curves correspond to Fig. 2a–c at e  105, respec-
dy x 1 tively. ‘a.u.’ in Fig. 2 denotes the arbitrary unit. In simulations,
¼y þ ; and yð0Þ ¼ 1: ð20Þ we assumed that: (1) b2 = 1 ps2/km, c = 9 W1km1, the fiber
dx 2 2
length L = 1.57 km (a soliton period), the initial peak power
The exact solution of Eq. (20) is given by y(x) = exp (x) + x/2. Fig. 1
P0 = 1 W, and the third-order soliton width T0 = 1 ps for Fig. 2a
shows the relative error between the numerical solution yn and
[3]; (2) b2 = 1 ps2/km, c = 1 W1 km1, L = 50 km (a collision
the exact value y(xn), i.e., |(y(xn)  yn)/y(xn)|. We can observe from
length), P0 = 1 W, T0 = 1 ps, and the initial separation of the funda-
Fig. 1 that to satisfy the globally relative error of <1.15  103 at
mental soliton pair q0 = ln(100/p) for Fig. 2b; (3) b2 = 1.276 
x = 10, (1) only 36 steps are used for the proposed method to solve
102 ps2/m, b3 = 8.119  105 ps3/m, b4 = 1.321107 ps4/m,
Eq. (20); (2) the LEM employs 354 steps to finish this task; (3) there
b5 = 3.032  1010 ps5/m, b6 = 4.196  1013 ps6/m, b7 = 2.57 
are 2502 steps for the fixed step-size method; and (4) the proposed
1016 ps7/m, c = 45 W1 km1, L = 0.1 m, P0 = 10 kW, and T0 = 28.4
method shortens the CPU time by about 10 times and 70 times in
fs for Fig. 2c [17]. Because the CPU time depends on the source
comparison with the LEM and the fixed step-size method, respec-
codes of program (e.g., C source codes are much faster than MAT-
tively. Therefore the proposed method in this paper is better than
LAB source codes) and the computer configurations (e.g., memory,
the LEM.
CPU frequency, etc.), we use the normalized CPU time for

-3
1.2x10 Fixed step-size method: N=2502 steps

-3
1.0x10
Local-error method: N=354 steps
Relative error

-4
8.0x10

-4
6.0x10

-4
4.0x10

2.0x10
-4 The proposed method: N=36 steps

0.0
0 1 2 3 4 5 6 7 8 9 10
x

Fig. 1. Relationship of relative error with variable x. The up-triangular, down-triangular, and square symbols in figure denote the points that are calculated during the
computing procedure by the proposed method, the LEM, and the fixed step-size method, respectively.
1438 X. Liu / Optics Communications 282 (2009) 1435–1439

a S-SSFM b S-SSFM
NPRM NPRM
2
10
Our method 2
Our method
10
CPU time (a.u.)

CPU time (a.u.)


1
10

1
10

0
10

0
10
-9 -8 -7 -6 -5 -4 -3 -9 -8 -7 -6 -5 -4 -3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Global relative error ε Global relative error ε

c d Supercontinuum generation
S-SSFM
3
10 5 A soliton pair
NPRM
Our method
4
Step size h (a.u.)
CPU time (a.u.)

10 2
3

10 1 2

10 0 Third-order soliton
0
-5 -4 -3 -2 0.0 0.2 0.4 0.6 0.8 1.0
10 10 10 10
Global relative error ε z

Fig. 2. Relationships of the CPU time versus the global relative error e and the step size versus fiber distance z. (a) third-order soliton, (b) a soliton pair, (c) supercontinuum
generation, (d) the step-size h versus normalized distance z.

comparing the computational efficiency of above mentioned the step-size h decreases to the minimum value at z  0.23 and
methods. 0.76 where soliton evolves to have peak intensity, and h decreases
We can find from Fig. 2a–c that the curve slope of the proposed to the second minimum value at z = 0.5 where soliton splits into
method is less than those of the S-SSFM and NPRM. It originates two distinct pulses [3]; (2) the step size is greatly reduced to accu-
from the fact that both S-SSFM and NPRM are locally third-order rately resolve the soliton collision when the collision occurs at
accurate (corresponding to globally second-order accurate) while z = 0.5 (the red curve); (3) h decreases with the parabolic approxi-
our method is locally fourth-order accurate. Our method is much mation at z < 0.8 where the spectrum of pulse broadens for the
more efficient when the global relative error e is less than 104. supercontinuum generation, while it slowly decreases at other part
For example, Fig. 2a shows that the computational time of the pro- (the black curve). The first step sizes are rapidly increased because
posed method is one and two orders of magnitude less than that of the initial guess h is set to be a much small value, and the last step-
the S-SSFM when e < 5  106 and 109, respectively. Fig. 2b exhib- size h is smaller than its previous step-size because the remaining
its that, comparing to the standard S-SSFM, our method enhances section of the fiber is shorter than the previous h.
the computational efficiency by a factor of 10 at e = 105 and in- We find from Fig. 2 that the NPRM works better than the
creases the efficiency more and more with the increase of accuracy. S-SSFM for simulating the evolution of soliton because it is deter-
The performance of our method is comparable in the range of low mined by the peak power in the time domain and the nonlinear
accuracy values, e.g., our method and the other two methods per- interactions are critical in the propagation. But the NPRM per-
form equally well for e > 103 in Fig. 2b. This reason comes from forms as well as the S-SSFM for simulating the supercontinuum
the fact that, although our method has a higher-order solution with generation.
the adaptive step size, it requires an additional step i.e., Eq. (18). As
a consequence, in comparison to the standard S-SSFM method, the 7. Discussion
proposed method can significantly enhance the computational effi-
ciency at the high accuracy of solutions. From Fig. 2d, we can see The computational efficiency of algorithm depends on the type
that (1) during a period of third-order soliton (the blue curve), of given problem, the variable region of problem, etc. For instance,
X. Liu / Optics Communications 282 (2009) 1435–1439 1439

8. Conclusions

6000
We have proposed an adaptive step-size control algorithm
using local fourth-order accuracy to solve the G-NLSE. Besides
Step size h (m)

4500 the automatic step adjustment mechanism, the proposed method


is globally third-order accurate rather than globally second-order
3000
accurate for the standard S-SSFM. A test example shows that our
method decreases the computational time by a factor of about 10
and 70 in comparison with the fixed step-size method and the
1500 LEM, respectively. On the contribution of higher accuracy and
adaptive step control mechanism, our method consumes less com-
0 putational time to solve G-NLSE describing the lightwave propaga-
0 100 200 300 400
tion in optical fiber. Numerical simulations show that in case of
solving NLSEs which describe the evolution of soliton, the compu-
Distance (km)
tational efficiency of the proposed methods is one and two orders
Fig. 3. Step-size h as a function of distance. of magnitude more than that of S-SSFM. Additionally, the compu-
tational efficiency of our method is enhanced with the increase
of the solution accuracy.
the relative error reaches zero at variable x  3.6 by using the LEM
method to simulate Eq. (20). On the other hand, it is impossible for Acknowledgments
our method to present the same performance (Fig. 1). However, the
proposed method uses only 36 steps for simulating Eq. (20) to sat- This work was supported by the ‘‘Hundreds of Talents Pro-
isfy the globally relative error of <1.15  103 at x = 10, while the grams” of the Chinese Academy of Sciences and by the National
LEM method requires 354 steps. Numerical results show that (1) Natural Science Foundation of China under Grant 10874239,
the proposed method does not offer advantage for simulating the 10604066, and 60537060. The author would like to thank Dr. Asi-
higher-order solitons in comparison to the LEM; (2) our method of- fullah Khan for editing and improving the English of this paper.
fers advantage for simulating the collision of a fundamental soliton
pair in comparison to the LEM; (3) the LEM is less effective to sim- References
ulate the supercontinuum generation as it depends on the initial
[1] T.R. Taha, M.J. Ablowitz, J. Comput. Phys. 55 (1984) 203.
step size, while our method is highly effective for this task.
[2] B. Fornberg, T.A. Driscoll, J. Comput. Phys. 155 (1999) 456.
An example for the comparison of the proposed method and the [3] G.P. Agrawal, Nonlinear Fiber Optics, 4th ed., Academic Press, Boston, 2007.
local-error method is shown in Fig. 3. In the simulation, all of the [4] K.J. Blow, D. Wood, IEEE J. Quantum Electron. 25 (1989) 2665.
[5] Q. Zhang, M.I. Hayee, J. Lightwave Technol. 26 (2008) 302.
parameters are the same as Fig. 3 in Ref. [12]. We can see from
[6] H. Hasegawa, Y. Kodama, Solitons in Optical Communications, Oxford
Fig. 3 that (1) the step-size h decreases to the minimum value at University Press, New York, 1995.
the distance z of 200 km, (2) h increase to the maximum value at [7] W.H. Reeves, D.V. Skyabin, F. Biancalana, J.C. Knight, et al., Nature 424 (2003)
z  180 and 220 km, and (3) h approximately holds on the same 511.
[8] S. Martin-Lopez, L. Abrardi, P. Corredera, et al., Opt. Express 16 (2008) 6745.
value at the range of 0 to 100 km and 300 to 400 km. The reasons [9] J. Hult, J. Lightwave Technol. 25 (2007) 3770.
originate from the fact that (1) the pulse collision occurs at [10] X. Liu, B. Lee, IEEE Photon. Technol. Lett. 15 (2003) 1549.
z = 200 km and the peak power of soliton is maximum so that h [11] G.M. Muslu, H.A. Erbay, Math. Comput. Simul. 67 (2005) 581.
[12] V. Sinkin, R. Holzlöhner, J. Zweck, C.R. Menyuk, J. Lightw. Technol. 21 (2003)
is minimum in order to accurately resolve the collision, (2) the 61.
peak power is minimum at z  180 and 220 km so that h is [13] S.G. Li et al., Chin. Phys. 15 (2006) 437.
maximum, and (3) the soliton approximately keeps the same peak [14] Rieznik, T. Tolisano, F.A. Callegari, D. Grosz, H. Fragnito, Opt. Express 13 (2005)
3822.
power at z = 0100 km and z = 300–400 km so that h is also [15] G. Bosco, A. Carena, V. Curri, et al., IEEE Photon. Technol. Lett. 12 (2000) 489.
nonvariant. By comparing Fig. 3 of this paper to that of Fig. 3 in [16] T. Kremp, W. Freude, J. Lightw. Technol. 23 (2005) 1491.
Ref. [12], it can be easily observed that our method offers better [17] J.M. Dudley, S. Coen, IEEE J. Sel. Top. Quantum Electron. 8 (2002) 651.
[18] U.M. Ascher, L.R. Petzold, Computer Methods for Ordinary Differential
performance in this given problem. At the same time, the numeri-
Equations and Differential-Algebraic Equations, SIAM, Philadelphia, 1998.
cal results show that the initial step size has little influence on [19] Quarteroni, R. Sacco, F. Saleri, Numerical Mathematics, Springer Verlag, New
simulating Fig. 3. York, 2000.

Anda mungkin juga menyukai