This action might not be possible to undo. Are you sure you want to continue?

MONTHLY

WEATHER

REVIEW

VOLUME

119

**The Third-Order Adams-Bashforth Method: An Attractive Alternative to Leapfrog Time Differencing
**

DALE R. DURRAN

**Department of Atmospheric Sciences. University of Washington. Seattle. Washington
**

(Manuscript received 21 February 1990, in final form 21 September 1990) ABSTRACT Th~ third-order Adams-Bashforth method is compared with the leapfrog scheme. Like the leapfrog scheme t~e third-order Ada~s-Bashforth method is an explicit technique that requiresjust one function evaluation pe; time step. Yet the third-order Adams-Bashforth method is not subject to time splitting instability and it is more accurate than the leapfrog scheme. In particular, the O[(ru)4] amplitude error of the third-order AdamsBashforth method can be a marked improvement over the O(ru)2] amplitude error generated by the Asselinfil~eredleapfrog scheme-even when the filter factor is very small. The O[(~t)4] phase-speed errors associated With thud-order Adams-Bashforth time differencing can also be significantly less than the O[(~t)2] errors produced by the leapfrog method. T.he third-order Adams-Bashforth method does use more storage than the leapfrog ~ethod, ~ut Its storage n:'Iulrem~nts are not particularly burdensome. Several numerical examples are provided illustrating the supenonty of third-order Adams-Bashforth time differencing. Other higher-order altern~tlves .to the Adams-Bashfo~ method are al~ surve~ed. A discussion is presented describing the general relationship between the truncation error of an ordinary differential solver and the amplitude and phase-speed errors that develop when the scheme is used to integrate oscillatory systems.

1. Introduction Leapfrog time differencing is widely used to numerica~ly simulate advective processes in low-viscosity fluids. The leapfrog scheme is an efficient method that achieves second-order accuracy with just one function evaluation per time step. Leapfrog time differencing does not artificially damp linear oscillatory motion nor does it produce instability by amplifying the oscillations. These advantages are somewhat diminished by the large pha~-speed error of the leapfrog scheme, and the unsuitability of leapfrog differencing for the representation of diffusive and Rayleigh damping processes. However, the most serious problem associated ~~ the leapfrog scheme is the "time splitting" instability that develops when the method is used to model nonlinear fluid dynamics. This time splitting instability is associated with the leapfrog scheme's undamped computational mode. Time splitting can be controlled th~ough the use of an Asselin-Robert time filter (Assehn 1972; Robert 1966) or by periodically discarding the data at the oldest time level and reinitializing the leapfrog solution through a single integration step with a two-level scheme. Atmospheric modelers commonly follow the first strategy representing time derivatives with the Asselin-filtered leapfrog scheme. Asselin filCo"esfOnti,ing auJho~add(ess: Dr. Dale R. Durran, Dept. of Atmosphenc Sciences, University of Washington, Seattle, WA 98195.

© 1991 American Meteorological Society

tering does, nevertheless, degrade the accuracy of the calculations-reducing the O[ (at)2] truncation error of the unfiltered leapfrog scheme to O( at] . One can obviously avoid the problems with the leapfrog scheme's computational mode by using a different time differencing scheme. The purpose of this paper ~sto examine ~terD:ativesto leapfrog differencing; the pnmary emphasis Willbe on one particular alternative: the third-order Adams-Bashforth method. There is no lack of possible substitutes for leapfrog differencing. Any ordinary differential equation (ODE) s,?lyer is a potential time differencing scheme. In addition to the general ODE solvers, special differencing schemes that produce a beneficial cancellation between the temporal and spatial truncation error can be constructed for particular systems of partial differential equations, e.g., the Lax-Wendroff method for the advection equation. This second approach, in which time and space differencing formulas are developed simultaneously, will not be considered further in this paper because the results are inherently specific to the system of equations for which the scheme was constructed. One class of ODE solvers that might serve as subs~itutes for ~he leapfrog scheme is the family of two time-level differencing methods, which have no computational mode. In order to achieve second-order accuracy, however, a two-level scheme must either be implicit (e.g., the trapezoidal method) or iterative. As a consequence, two-level schemes require more com-

Mesinger and Arakawa 1976) require two function evaluations per time step and generate unstable. The third-order Adams-Bashforth method appears to be an even more attractive alternative. each operating with a 6-ns clock time. most modest-sized atmospheric models are not limited by lack of memory.e. The third-order Adams-Bashforth method is described and analyzed in section 2. and it allows a stable and accurate rep- resentation of diffusive and Rayleigh damping processes. The practical considerations that led to this preference for very low-storage algorithms have. At present. Section 6 contains a discussion of the difficulties that arise if the third-order AdamsBashforth method is substituted for leapfrog differencing in semi-implicit models. Concern about storage seems to be the main reason that higher-order multistep methods have not received more attention. which contrast the performance of Adams-Bashforth and timefiltered leapfrog differencing. The third-order AdamsBashforth method is not subject to time splitting. Like leapfrog differencing. most low-order two-level iterative schemes produce undesirable amplitude errors when applied to oscillatory problems. its maximum stable time step is similar to the maximum stable time step permitted by the Asselin-filtered leapfrog scheme. Section 3 contains a discussion of the relationship between time differencing error and the errors introduced by spatial differencing in wave-propagation problems. The instabilities in these second-order methods are relatively slow growing and can sometimes be tolerated if one uses a sufficiently small time step and integrates for a limited time. . changed over the last several years as the incore memory in many computer systems has increased much more rapidly than CPU power. and they can easily accommodate the storage needs of the third-order Adams-Bashforth method.. the most powerful Cray has 128 million words of memory and eight processors. The only potential drawback of the third-order Adams-Bashforth method is its storage requirement. i. Switching to an efficiently coded third-order Adams-Bashforth method will increase the necessary storage by a factor of 3/2. Total processing capacity has thus increased by a factor of 16 while storage capacity has increased by a factor of 128. amplifying oscillations. A second class of potential replacements for leapfrog differencing are the multistep ordinary differential equation solvers with damped computational modes. the second-order Adams-Bashforth method is a three-level scheme that requires only one function evaluation per time step. the loss of accuracy associated with a 12% reduction in spatial resolution can easily be offset by the improvements associated with third-order time differencing. it is not subject to time-splitting instability and it can be used to model diffusive and Rayleigh damping processes.g. The first-order Matsuno scheme (Matsuno 1966) requires two function evaluations per time step and heavily damps the solution. In addition. Young's method A (Young 1968). which is 2If2times that of the leapfrog scheme (Mesinger and Arakawa 1976). the oscillation amplifies with time but the instability can be tolerated if short time steps are used and the total length of the integration is limited. Unlike the leapfrog scheme. suppose that memory truly is the limiting factor in a particular three-dimensional simulation. Nevertheless. Previous researchers (e. A brief survey of other higher-order time differencing schemes appears in section 5 together with a description of the general relationship between the truncation error of a finite-difference scheme and the amplitude and phase errors that occur when the scheme is applied to oscillatory problems. If the' memory-limited model currently employs leapfrog time differencing and the code is optimized to minimize memory requirements. Unfortunately. The remainder of this paper is organized as follows. the original Cray 1 had one million words of memory and one processer with a 12-ns clock time. Moeng 1984). As an example. Another disadvantage of the second-order Adams-Bashforth method is its large phase-speed error. DURRAN 703 putational effort to achieve the same accuracy as the leapfrog scheme..MARCH 1991 DALE R.g. Second-order Runge-Kutta methods (including the "Huen" method. which is second order. Young 1968) have also considered the stable and highly accurate fourth-order Runge-Kutta method. The only member of this class that seems to have undergone much scrutiny is the second-order AdamsBashforth method (Lilly 1965). Nevertheless. the fourth-order Runge-Kutta scheme requires four function evaluations per time step and considerably more storage than the lower-order two time-level schemes. a minority of atmospheric modelers have preferred the weak amplification of the second-order Adams-Bashforth scheme to the time splitting problems of the leapfrog method and have employed the Adams-Bashforth scheme in their numerical models (e. where its theoretical properties are compared with those of the Asselin-filtered leapfrog scheme. the physical mode is subject to the same instability as the second-order Runge-Kutta method. In some applications. storage must be allocated for the complete set of prognostic variables at two different time levels plus a smaller temporary storage array. the third-order Adams-Bashforth method is an explicit scheme requiring one function evaluation per time step. The conclusions appear in section 7. or equivalently by a 12% reduction in the spatial resolution. As a consequence. The disadvantage of the second-order AdamsBashforth method is that when it is used to model oscillatory phenomena.. Like its second-order cousin. requires three function evaluations per time step and is also unstable. Several numerical tests. are presented in section 4. This extra storage can be created by reducing the number of gridpoints in each of the three spatial dimensions by a factor of(2/3)113. however.

following Mesinger and Arakawa 1976.t] = t/I( n I:l. in which aD = ~. (5) The family of Adams-Bashforth schemes are part of the classical literature. -0.4 cOOt FIG. Magnitude of the amplification factor for the third-order Adams-Bashforth scheme plotted as a function of wt:J. (3) J. and F( t/I) into (2) and choosing the aj to cancel all terms less than order (I:l.t > 0.. Comparative analysis of the third-order AdamsBashforth method and the Asselin-filtered leapfrog scheme The general Nth-order Adams-Bashforth approximation to the ordinary differential equation dt/l/dt = 1. t/I[ (n + 1) I:l.t) (Gear 1971).2 IAI 0. (4) j=O The coefficients aj can be evaluated by requiring (4) to be exact for all polynomials less than order N.t)N.t 12 [23F(4)n) . Solid line is the physical mode. The phase-speed and amplitude errors introduced by time differences in the numerical representation of nondissipative wave phenomena may be examined by analyzing solutions to the "oscillation equation" dt/l/dt = Figure 1 shows I A I . .8 0. and the relative phase error as argA 1 (ZS[A]) R == argAe = wl:l. The third-order Adams. . (11).2 where 4> n is the numerical approximation to t/I( n I:l. the phase-speed error of the physical mode is R phys = 1 iwt/l.8 0. for each of the three roots of(9).t) 6 ]. r<n+l)6t F[t/I(t)]dt = na: where A is the (complex valued) "amplification factor. One computational mode becomes unstable for wl:l.704 MONTHLY WEATHER REVIEW VOLUME 119 / 2.tA 12 3 . p.t ~ 1.t.2_ iwl:l.. (10) In the limit of wl:l. (1) has the form (2) 0. and (17)-( 19) were obtained using the symbolic manipulation capabilities of Mathematica (Wolfram 1988). dashed curves are the two computational modes.4 1.) In the limit of wM ~ 1.724. is the familiar Euler (or forward) difference.and second-order cousins. 109).t (7) F [ t/I( t ) ] dt. Thus.4>n = - The amplification factor for the third-order AdamsBashforth solution to the oscillation equation (6) satisfies the cubic equation A3 - (1 + 23 iWl:l. plotted as a function of wl:l. in which ao = 1. al = -1..6 0. Unlike its first. The first-order Adams-Bashforth method. (9) I:l.6 -0. Although it is a classical scheme.t)4 3 + O[(wM)6].".676. (6) + 289 720 (4 wl:l.4 " " "" " "" F(t/I). . the relative amplitude error in the numerical solution may be defined as I A 1/ I Ae I = I A I ." The amplification factor associated with the exact solution to (6) is Ae = e'"?'.16F(4)n-l) + 5F(4)n-2)].t arctan ~[A] . the third-order Adams-Bashforth scheme dampens the physical mode.t 12 = O. (Note also that the damping rate of the physical mode exceeds the damping rate of one computational mode for wb. The coefficients aj may be determined by substituting the Taylor series expansions for 1/.2 0.t. (8) 2: ajF(4)n-j). The second-order case. both computational modes are strongly damped and the amplification factor for the physical mode is 1 I A IPhyS = 1 -"8 (wl:l. ( 11) Assuming w is a real constant.t)A2 + ~ iwl:l. An alternative derivation of the coefficients can be obtained by writing (1) as the equivalent integral equation --"''" . 1.t) + O[(wl:l. Numerical values for the coefficients of all Adams-Bashforth methods through order 6 are listed in Gear ( 1971.t) + The Adams-Bashforth scheme approximates the preceding integral as r: n6t N-l I:l.Bashforth scheme is 4>n+l . is discussed by Lilly ( 1965) and Mesinger and Arakawa ( 1976). numerical solutions to (6) can be obtained from finite-difference approximations of the form (2) such that I The asymptotic results in (10).. the third-order Adams-Bashforth method does not appear to have been widely employed in numerical models for the solution of partial differential equations.l > 0.

06 is typically used in the NCAR community climate model (Williamson 1983).2¢' = = + (2) 1= O. 2. both modes are neutral and the standard leapfrog result is recovered.."(2) .06 (the typical value in the community climate model) the Asselin-filtered leapfrog method produces more damping than the third-order Adams-Bashforth method whenever the period of the oscillation exceeds 221lt. stability requires will ~ 0. As in the third-order Adams-Bashforth scheme."I) 8 (1 . A prototype equation for the time .2 (the typical value for mesoscale convective models) all oscillations with periods greater than 11III are damped more heavily by the Asselin-filtered leapfrog method. The corresponding amplification factors for the third-order Adams-Bashforth scheme are plotted in Fig. (16) In all cases of practical interest n ~ 1 and "I ~ 1. All values of Ae lie on the unit circle in the complex plane. (14) Is the assumption (A¢n) = A( ¢n)justified? In practice. the third-order AdamsBashforth scheme will follow the phase of low-frequency oscillations much more accurately than the Asselin-filtered leapfrog scheme. IAlcomp = (17) (1 . (A¢o) _ A(¢o) = = ¢' _ ¢' 'Y(¢o . The Adams-Bashforth scheme has two computational modes. which is almost as stringent as the third-order Adams-Bashforth stability condition. it damps low-frequency oscillations much less than the secondorder diffusion in the Asselin-filtered leapfrog method. Asselin-Robert filtering increases the phase-speed error doubling it when "I is 0. Even with "I = 0. Amplification factors for the Asselin-filtered leapfrog scheme are plotted in Fig. ( 13) may be rewritten (A¢n+l) _A(¢n+l) 'Y[(A¢n) 'Yn+l[¢J . therefore (16) implies that A may be factored out of the filtering operation with negligible error. the modulus of the amplification factor for the Asselin-filtered leapfrog scheme may be approximated as I A I phys - _ 1 'Y(wlllf _ 'Y(1 + "I . the amplitude error can be reduced by choosing a small value for "I.) Suppose that ( 12) and (13) are used to obtain an approximate solution to the oscillation equation. -A(¢n)] (15) However.25. Note how the damping introduced into the computational mode by Asselin filtering is relatively independent of the time step. In the preceding. an overbar denotes a time-filtered value. 2c for the case "I = 0.0. DURRAN 705 In agreement with the general proposition stated in section 5. (18) As is the case for the standard leapfrog scheme. the initial condition is not time filtered.25-0. Havingjustified the derivation of ( 14). (1983) recommend a value of "I in the range 0. _ 2¢n (12) (13) + ¢n+l). Once again. 2b. The computational mode that becomes unstable for wilt> 0. This study will now turn from the problem of modeling inviscid oscillatory systems to that of simulating dissipative processes.'Y2)(wlll)4 2 (1 .6"1 + 4"12 + O[(wlll)3]. In the limit of will ~ 1.0.2. the phase speed is second-order accurate.816. '. In the case "I = 0. one obtains Asselin's expression for the amplification factor A = "I In the limit "I . an analysis of the amplitude and phase-speed errors of the Asselin-filtered leapfrog solution may be performed in the usual manner. In the limit of will ~ 1. Let us compare the amplitude and phase errors in ( 10) and ( 11) with the equivalent values for the Asselinfiltered leapfrog scheme: ¢n+l ¢n = = ¢n + ««:' ¢n-l + 2IllF(¢n). .2"1) +2 'Y(Wllt)2 .(Wllt)2] 1/2. one simply defines ¢ ° == ¢ 0• Thus. As might be expected from the structure of (13).MARCH 1991 DALE R. the phase speed of the Asselin-filtered physical mode becomes Rphys _ - 1 + 6( 1 _ 1 + 2"1 "I) (wM) 2 .1. Obviously. both of which are strongly damped when wilt is small. The amplification factors for all physical and computational modes of the Asselin-filtered leapfrog scheme and the third-order Adams-Bashforth method are plotted as a function of will in Fig. the third-order Adams-Bashforth method is more accurate for all oscillations with periods greater than 10III . (Schlesinger et al. Observe that the filtering coefficient "I is equivalent to v/2 in Asselin's (1972) notation. 0. (19) + iwllt ± [(1 . Under the assumption that (A¢n) = A(¢n). However.¢l]. for will = 0. If'Y = 0."I) 3 + O[(wllt)5].9. and solutions are sought in the form (7). the amplitude and phase-speed errors generated by the third-order Adams-Bashforth method are both fourth order.2 are common in mesoscale convective models. Figure 2a shows the amplification factor for the true solution to (6). Asselin-Robert time filtering introduces a second-order error in the amplitude of the physical mode. do not recommend leapfrog differencing for the diffusion terms.3 for general advection-diffusion modeling.724 has an approximate period of 41ll. the period of unstable oscillations is approximately 41ll. If "I is reduced to 0. Values of "I = 0. indeed Schlesinger et al. since the damping term in the third-order Adams-Bashforth method is fourth-order.2.2. A value of "I = 0.

the most stringent CFL time step restrictions are imposed by fast-moving waves that have little meteorological significance (i. the third-order Adams-Bashforth solution.. L is the wavelength. but less strongly damped than the physical mode for 0.9 are also labeled.2.== ICAtI ilx = ! ck!ilt I k I ilx = ilt/T .545. The third-order Adams-Bashforth method will produce stable nonoscillatory solutions to (20) whenever Kilt < 0. The familiar Courant-Fredricks-Lewy (CFL) stability condition requires ! f. c is the phase speed.C9 REVIEW VOLUME 119 . Points corresponding to a physical (computational) mode and a value of will = 0. scheme and a smaller 3. Kreiss and Oliger 1973). the Courant number associated with the slower-moving. if diffusion is included in models with leapfrog time differencing and if the diffusion terms are evaluated using an explicit scheme. the third-order Adams-Bashforth method permits the explicit integration of diffusion terms with a slightly larger stable time step than that allowed by the time-lagged Euler method. 0. ·. 2. Points obtained with will = 0. All Adams-Bashforth methods may be applied to (20) without encountering the instability generated by the leapfrog scheme. and the Asselin-filtered leapfrog solution with 'Y = 0. . meteorologically important modes can be very small and the effect oftime truncation error on the accuracy of the important modes will be negligible. and T == 27r/ ( kc) is the period of the wave.. ilx / L (22) The stability criteria necessary to guarantee decaying. Durran and Klemp 1983). The leapfrog technique cannot be used to integrate (20) without a transformation of variables because the resulting difference equations are unstable (Lilly 1965. however.g. the stability constraints imposed by the fastest moving waves are eliminated by the use offiItered equations or by numerical techniques like the semi-implicit method (Kwizak and Robert 1971) or the two time-step method (Klemp .) Thus.2iltK(jJn-l. According to (22).5.359 < Kilt < 0.545.2. Amplification factor plotted in the complex plane for the exact solution.1. Individual points associated with all physical and computational modes are plotted for will = 0.1 are labeled "PI" ("el" or "el"). In some atmospheric models. The third-order Adams-Bashforth solution should also be far more accurate than one produced by the first-order lagged Euler method because it would Here k is the wavenumber. the Courant number may be interpreted as the ratio of the temporal resolution to the spatial resolution in a wave propagating at phase speed c. nonoscillatory solutions to (21) is Kilt < 0.P9 • • • • • • • P1 EXACT 3RD·ORDER ADAMS· BASHFORTH ASSELIN· ALTERED LEAPFROG Fio. (21) be obtained using a higher-order time step. the result is (jJn+l = (jJn-l . Thus.e.. evolution of a dissipative process is the so-called "friction equation" (again following the terminology of Arakawa and Mesinger 1976): (20) where K is a nonnegative real constant.9.0. •. If the lagged Euler approach is used to difference ( 20 ).1 1r. If.L I ~ O( 1) for any explicit difference scheme. those terms are generally lagged in time and integrated with a first-order Euler scheme (e.706 MONTHLY WEATHER . The unit circle is graphed as a reference. gravity waves in a global primitive equation model). Then. (One computational mode is stable. The interaction between time-differencing errors in spatial differencing error and The relative importance of temporal and spatial differencing error in the numerical representation of a traveling sinusoidal wave is largely determined by the absolute value of the Courant number.

¢n+1 _ ¢n-I J J k2t:J.t 2t:J. Thus. In order to minimize the phase-speed error for an 8t:J. In this example... 3/8.L = 0. but it will be dominated by the effects of time differencing. Since wt:J. 3b.t should be just 21 % of the maximum stable time step. the initial condition is {64 [ (x .~(¢J+~~:J-2)] = 0. In practice. 3 are for a nondi.527. According to (27). When f. As a result.L 2)] . The optimal value of f. and Asselin time filand the amplitude of the dispersive ripples is evident tering effectivelyintroduces a second-order spatial filter in Fig. 5/8. 3c). when choosing the optimum time step for the widely used leapfrog-time fourth-order space approximation to (23). t:J. 3!h_very poor accuracy is obtained using f..x wave. is one-fourth. { 0. a Courant number just below the stability The amplitude of a traveling wave is also influenced limit.L produce results very similar to Fig.x2 5 . and ¢ is periodic eliminated when f.x .Lkt:J.. ckat < 1.when third-order Adams-Bashforth time differencing is used in conjunction with centered fourth-order spamensional time of 3. = where it is again assumed that ktxx < 1.. and the leading terms in the phase-speed error are C4/f = are exact when f.t)4. where one unit of nondimensional time is the time required for the flow to complete one tial differences. (24) one must strike a compromise between efficiency and accuracy since the largest stable time step does not give the most accurate result.7272. (26) +c _ [4 (¢IJ 3 J+I . DURRAN 707 and Wilhelmson 1978).5 (Fig.3 damps the solution at a rate proportional to (Wt:J. 1. iJV. Some of this scale selectivity differencing.L = 0. at + cal/.t = f. x':.L is independent of k and the fourth-order phase-speed error is the horizontal mesh size t:J.. the total phase-speed error is always circuit around the periodic domain. the numerical error is dominated by time the shortest wavelengths.t)2] time difference no longer dominates the lagging phase.X2 c 1 . the leading phase-speed error of the leapfrog whereas Asselin time filtering introduces an O[ (Wt:J....L = 0. ck. the dissipation oftravspeed error of the spatial difference and the spike trails eling waves generated by temporal filtering is equivalent the true solution. = 0. ax = 0 ' (23) by spatial differencing. For reasonably resolved waves ka» < 1. the amplitude of the spike is nearly correct.L4)] (27) . although the phase speed of the spike is over. Consider. 3a. very small time steps are required to minimize the phase-speed error oflong. employ fourth-order spatial filters to selectively damp In Fig. section 2.L == 0.L phase-speed error is achieved by taking the largest stable time step.:. Many numerical models reductions in f.. 3d. (Fig.and spatial-differencing error is evident in Figs.L can easily be understood by evaluating the discrete dispersion relation satisfied by traveling wave solutions to (24) (Haltiner and Williams 1980).c.MARCH 1991 DALE R.L = 1. the optimal value of f.dissipation. The advection velocity c than the maximum stable time step of 0.:.f.537. Some beneficial cancellation between the temporal.. Some basic techniques for solving the one-dimensional advection equation. the Courant number associated with the features of interest may approach unity and time-truncation errors may become significant. A further increase in the phase lag to damping from a spatial filter. 3. As indicated in dominated by spatial differencing.¢n ) J-l 2t:J. the phase-speed error of well-resolved waves will be small. Computations utilizing these schemes are most accurate when performed with the largest stable time step. Unfortunately. 3b).. the value of f. The dependence of the phase-speed error on f. This value is slightly larger over the interval 0 .1/2) ¢O(x) = 2- In contrast to the more familiar example where leapfrog differencing is combined with centered second-order spatial differences (Haltiner and Williams 1980). otherwise.1.X4 ( 1.. therefore. in Fig. The leading-order terms in the resulting discrete dispersion relation are C4ab3 1/64 ]} 2... if (25) C[ 1. the Lax-Wendroffmethod.x. As previously discussed in estimated. and the leapfrog scheme with centered second-order spatial differencing. The dependence of solutions to (24) on the Courant number is illustrated in Fig. the importance of time-truncation error in simulations performed with Courant numbers near the stability limit.~ 289f.L decreases linearly with decreasing wavenumber. Continued on all propagating features.x = 1/32.L that minimizes the phase-speed error is dependent on the wavenumber.. well-resolved waves.S: < 1.. 3d the dominant error is produced will be lost if the same model uses Asselin time filtering. Examples of this type include upstream differencing. which shows the case f.. The results in Fig. Much better results are produced when f.by the Courant number.~ k4t:J.x':. the third-order Adams-Bashforth method and the dispersive ripples are small. 3d.--[ 6 2 (k --t:J. and the minimum Fig. One can obtain phase-speed errors that are truly fourth-order by combining third-order Adams-Bashforth time differencing with centered fourth-order spatial differences.

708 1.2 0 -0. \. As in section 3. several numerical tests will be presented in order to illustrate the improvements that can accompany the replacement of leapfrog time differences by the third-order Adams-Bashforth method.2 0 -0. \. \. .2 1..J . \. 3. . Effect of leapfrog stepsize on the accuracy of fourth-order centered-difference solution to the advection equation..6 0.4 0 8 16 MONTHLY WEATHER REVIEW VOLUME 119 ---EXACT --LEAPFROG (b) q./ I (d) \. \. 24 32 0 .8 0. (c) 0. ./ q.4 (c) 1. \.1 .(x) \. All results are for a nondimensional time of3.0 0.4 0. since the dissipation acting on the longest waves will be dominated by the second-order time filter. a.6 0.. Linear advection The first numerical comparison will be for simulations of conservative-tracer transport by a constantvelocity flow. \ \ '_" . Shown are the exact and numerical solutions computed using Courant numbers of (a) 0.8 0..0 0. Comparative tests of the third-order Adams-Bashforth method and the Asselin-filtered leapfrog scheme In this section.5.(x) . .2 1. solutions to (23) are computed on the periodic domain 0 ~ x ~ 1 using a wind speed c = 1/4 and the initial distribution given by (25). (b) 0. \ .4 0.. The behavior of third-order Adams-Bashforth .7272. This problem can be avoided with third-order AdamsBashforth time differencing because the dissipation in that scheme is equivalent to a fourth-order spatial filter.4 (a) 1.3. and (d) 0.2 -0.4 1. 4. . \ . .1./ 8 16 24 32 x (grid points) x (grid points) FIG..2 -0.

Asselin filtering with l' = 0..2 0 -0. the value l' = 0.. 'Y = . Significant phase-speed errors and negative concentrations are produced by the leapfrog scheme. Although the unfiltered leapfrog scheme is nominally free from "amplitude" error.2 -0.4 0.2. which implies the time step is 63% of the maximum permitted by the stability requirements of the unfiltered leapfrog scheme.ORDER ADAMS . How do the preceding results change if the accuracy of the spatial differences are reduced to fourth-order? The answer appears in Figs.06 was selected because it is typical of the filtering employed in global spectral models.. and two Asselin-filtered leapfrog solutions are shown at a nondimensional time of 3. 4 faster than the O[(W~t)4] damping produced by the third-order Adams-Bashforth method. 4a were calculated using a Courant number of 0. 4b have been run out to a nondimensional time of 40 (twice the time interval over which the integrations were performed in Fig. the relative performance of the three schemes is unchanged.8 0. creating substantial negative concentrations ahead of the main spike. Nevertheless. The four curves represent the exact solution and results obtained using the third-order Adams-Bashforth method. In the following tests.2. the case l' = 0.08. 5 and 6. third-order AdamsBashforth. and (b) for tnd = 40 and JI. the Adams-Bashforth solution was started by taking a forward time step. The spectral method is essentially an infinite-order finite-difference scheme. The unfiltered leapfrog solution retains more amplitude.. which shows results at a nondimensional time of 20.0 0.MARCH 1991 (a) DALE R. Comparison of an exact solution to the advection equation with results obtained using leapfrog. Figure 4a. 4b are considerably smaller than those in Fig... Figure 4 shows the numerical solutions computed using a 32-wavenumber spectral representation of the spatial derivatives.LEAPFROG ---3RD . but the large. 4a). the dispersion generated by leapfrog differencing reduces the maxima in Fig.4 1. The third-order Adams-Bashforth scheme exhibits negligible phase-speed error. Figure 5 shows computations performed using centered fourth-order differences [see (24)] with a spatial mesh size of 1/32• The exact. the unfiltered leapfrog method.ASSELIN LEAPFROG.4 0 ---EXACT -.2 1. 8 16 24 32 o 8 16 24 32 x (grid points) x (grid points) FIG. followed by a secondorder Adams-Bashforth step. The value l' = 0. = 0.\j.BASHFORTH . and leapfrog time differencing will be compared in two different models: a spectral model and a fourth-order finite-difference model. and Asselinfiltered leapfrog time differencing in a spectral model: (a) for nondimensional time tnd = 20 and Courant number 11 = 0. Figure 4b shows results obtained with the time step reduced so that IL = 0.08. its amplitude error is slightly less than the error in the unfiltered leapfrog solution. indeed the Adams-Bashforth solution looks nearly perfect. The third-order Adams-Bashforth scheme produces the smallest amplitude and phase-speed errors. 4.6 <1>(x) 0. the errors in Fig. Although the simulations in Fig.06 .06 and 0.06. The filtering coefficients are 0. the leapfrog solution was started by taking· a single forward time step.2 is typical of the filtering used in finite-difference me- .06 may be compared with the preceding results. frequency-dependent phase-speed error of the leapfrog scheme disperses the solution. indicates that even weak time filtering strongly damps the solution.2. and the leapfrog method with Asselin-filtering coefficient l' = 0. DURRAN (b) 709 1. 4a. Adams-Bashforth.06 still produces heavy damping. The solutions plotted in Fig..

2 -0. Comparison of an exact solution to the advection equation with results obtained using Adams-Bashforth and Asselin-filtered leapfrog time differencing in a fourth-order finite-difference model at a nondimensional time of 3.20 0 8 16 x (grid points) 24 32 o 8 16 x (grid points) 24 32 FIG.4 0.2 1.06 ---3RD .L = 0. y = . As in Fig.6 <p(x) 0. except that the spatial resolution has been doubled.ORDER ADAMS .5.20 10 20 30 x (grid points) 40 50 60 o 10 20 30 x (grid points) 40 50 60 FIG. y = . 5. for (a) J.4 1.2 -0.2. soscale models. the maximum allowed by the unfiltered leapfrog scheme.8 0. (This time step is also 84% of the maximum allowed by the filtered leapfrog (a) scheme when 'Y = 0.06 3RD .ASSELIN LEAPFROG.0 0.4 0. The calculations shown in Fig.2. 6. implying that the time step was 68%.2 1.y=. the phase-speed error introduced by fourth-order spatial differencing almost (b) 1. .5 and (b) J.8 0.6 <p(x) 0.4 0 ---EXACT -.) As per the discussion in section 4. when the time step is close to the stability limit.4 -------------- EXACT ASSELIN LEAPFROG. 5. and 95% of the maximum allowed by the third-order Adams-Bashforth method.0 0.2 0 -0.710 (a) MONTHLY WEATHER (b) REVIEW VOLUME 119 1.4 1.BASH FORTH ASSELIN LEAPFROG. Sa were performed with a Courant number of 0.L = 0. Y = .BASH FORTH -------ASSELIN LEAPFROG.ORDER ADAMS .2 0 -0.

and as might be expected from a higher-order scheme. lagging phase-speed error produced by fourth-order differencing. 6. The goal is to examine a nonchaotic solution illustrating the susceptibility of leapfrog time differencing to instability. Figure Sb indicates that when the Courant number is reduced to 0. 7. Yo. were integrated using lagged forward differences [as per (21)]. The leading phase-speed error introduced by all three time differencing schemes is reduced as /-t is decreased. and the square brackets are used to indicate the diffusion terms. 6 are identical to those in Fig. the leading phase-speed error of the leapfrog method dominates the smaller. there is no sensitive dependence on the initial conditions. 5. which rapidly generates a 26. The leapfrog scheme. 5.t noise in the leapfrog solution. Equations (28)-(30) were integrated using a fourthorder Runge-Kutta scheme (see Table 1). again following the practice in most mesoscale models. however. (28) rX[Y]. The least accurate solutions are those produced by unfiltered leapfrog differencing. enclosed by the square brackets in (28)-(30). On the other hand. and both leapfrog solutions move too fast. suggesting that the RungeKutta scheme has converged to the exact solution.2. all three numerical solutions lag the exact solution. the larger phase-speed errors introduced by leapfrog time differencing are better able to compensate the error introduced by spatial differencing. a 0.2. The remaining terms in (28)-(30) represent advection and buoyancy forces and. As a consequence. In these simulations the computational mode does not grow without bound. however. Chaotic solutions to (28)-(30) are inappropriate for these tests because they would exaggerate the difference between the various numerical schemes. y= -XZ+ (29) (30) Z=XY-[bZ] where the dot represents differentiation with respect to a nondimensional time. 25).03. 6a and slightly worse in Fig. The phase error in the Adams-Bashforth solution is better than that of the leapfrog method in Fig. the diffusion terms.015 time step was used in Fig. 7a.2. the third-order AdamsBashforth method produces less amplitude error than the Asselin-filtered leapfrog scheme.5. the amplitude error in all solutions is reduced significantly. b = 6. In all cases. Following the practice in most mesoscale convective models. In addition. It is easy to find other values of (u. As in Fig. The simulations plotted in Fig. Sa is very small. except that the horizontal mesh spacing and the time step were halved. Sa. 7b. As expected. this is an uninteresting case-the solution decays with time. the phase error of the Adams-Bashforth solution in Fig. The behavior of this system will be examined for the case a = 12. In contrast to Fig. DURRAN 711 cancels that generated by the third-order AdamsBashforth method. at the relatively short wavelengths that dominate this test problem. This last result is not surprising. -10. the thirdorder Adams-Bashforth scheme. the Asselin time filter damps both leapfrog solutions more rapidly than the AdamsBashforth solution. The degree of improvement may be assessed by comparing Fig. the error in the Adams-Bashforth result decreases more rapidly than that in the Asselinfiltered leapfrog solution as the time step is halved. the finite-difference results are less accurate than those obtained with the spectral method. Thus.MARCH 1991 DALE R. is not suitable for the direct integration of (28)-(30). A considerable increase in the quality of the finite-difference solution can be achieved by doubling the horizontal resolution. 5 with Fig. The results in Fig. b. Lorenz ( 1963) has shown that the equations governing two-dimensional Rayleigh convection can be truncated to three components whose amplitudes are governed by the system of equations X=uY-[uX]. the Runge-Kutta solution provides a reference against which the other methods may be compared. b) and other initial conditions that generate significant 26. From the standpoint of predictability. and the Asselin-filtered leapfrog method. The application of the first two schemes to these equations is straightforward. it remains at a steady amplitude as the low-frequency oscillation in the physical mode decays. Asselin filtering with 'Y = 0. and as a result the lagging phase-speed error generated by the fourth-order spatial difference dominates the total phase error. The third-order Adams-Bashforth scheme continues to produce the least dissipation.t computational mode. and initial conditions (Xo. the Adams-Bashforth solution should show greater improvement because it is a higher-order method. 7a were obtained using a time step of 0. 6b. Examples that produce catastrophic instability are. these were differenced using the leapfrog scheme. The error in the third-order Adams-Bashforth solution is much smaller than the error in the Asselin-filtered leapfrog solution.b. Zo) = (-10. the improvement in the solution accompanying the increase in resolution is significantly greater for the third-order Adams-Bashforth method. The fourth-order Runge-Kutta solution is essentially identical in both Figs. The numerical solution for X is plotted in Fig. The Asselin time filter was set to l' = 0. rare. r. (The solution will "blow up" if a .2 generates heavy damping. r = 12. when /-t = 0. Nonlinear convection Now consider a situation in which the leapfrog solution must be filtered in order to avoid spurious amplification of the computational mode. the damping is particularly severe when l' = 0. The phase error in the Adams-Bashforth solution is slightly worse than that in either leapfrog solution because.

Y= . 5. which shows an integration equivalent to that in Fig.oJS.. As in Fig. and r are changed to 10..712 ( a) 0 ___ MONTHLY WEATHER (b) REVIEW VOLUME 119 LEAPFROG -2 ---ASSELIN LEAPFROG. The variable plotted is X.03 and (b) O. once again. The highly truncated nature of Lorenz's model apparently prevents the additional interactions responsible for the continued amplification of the computational mode. Other larger-amplitude linear solutions also failed to generate a discernable computational mode. the leapfrog solution is. 7a.03. 7a. Zo).. Although there is no time splitting in the linearized problem.. and t:J. The initial development of the computational mode in Fig.ORDER ADAMS . 7 is dependent on nonlinear processes. __ --- 4th-ORDER RUNGE-KUTIA ADAMS-BASH FORTH -2 - . -12 -14 0 10 20 I (lime steps) 30 40 50 0 20 40 60 80 100 I (lime steps) FIG.4TH ORDER RUNGE-KUTT A -4 -6 X (I) -8 -10 / . Numerical approximations to a nonchaotic solution of the Lorenz system.LEAPFROG 3RD ORDER -4 X{t) -6 -12 t (time steps) FIG. The importance of nonlinearity is demonstrated by Fig. . Other time differencing schemes Are there other high-order time differencing schemes that might be even more attractive than the third-order Adams-Bashforth method? Are there second-order schemes that provide a better way to control time splitting than Asselin-Robert time filtering? A partial answer to these questions can be provided by a brief sur- vey of other ordinary-differential-equation solvers.) Practical experience with less highly truncated convective models suggests that time splitting instabilities will inevitably arise in leapfrog integrations. it is helpful to describe the relationship between the overall truncation error of an ODE solver and the amplitude and phase-speed errors introduced by that solver when it is used to model nondissipative oscillatory phenomena. except that the governing equations have been linearized. less accurate than that obtained with the third-order AdamsBashforth method.t = 0. 8. except that the Lorenz equations were linearized about (Xo. the time step is (a) 0.BASHFORTH ____ . 7. In order to organize the survey in a systematic manner. 8.20 _ .. The qualitative character of the linear solution is similar to the nonlinear solution in that both are damped oscillations.3RD . Yo. yet the unfiltered leapfrog solution does not develop a computational mode.

where m~r+ 1. It follows that second-order time differencing produces second-order phase-speed error and fourth-order (or better) amplitude error. the third. ( m= r. (6) is integrated using a linear finite-difference scheme and if the truncation error of the resulting finite-difference approximation to the oscillation equation is order r. the column labeled "efficiency factor" lists the maximum stable time step with which the oscillation equation can be integrated. it does not measure the work required to achieve a given accuracy. If the efficiency factor was all that mattered. then as wl:J. whereas the firstand second-order Adams-Bashforth and Runge-Kutta techniques! are unstable. if r is odd. temporary storage array. consider the third-order Adams-Bashforth method. When integrating partial differential equations. The preceding sections of this paper have demonstrated that the third-order Adams-Bashforth method can be significantly more accurate than the Asselinfiltered leapfrog scheme. The storage factor for implicit methods can vary from problem to problem. at some point in the integration cycle ¢/ will no longer be needed and at that stage it may be overwritten by ¢jn+l. As an example. divided by the number off unction evaluations . (31) and the phase-speed error is 1 + O[(wl:J. if r is even. ( n ~ r + 2. where n = r + 1. 2 per time step. because ¢/ may be required for the computation of ¢Jtl. if r is odd. (32) The proof of this proposition. Therefore. During the interim between the calculation of ¢/+l and the last use of ¢/' ¢/+l may be held in a temporary storage array. which was partially anticipated by Takacs (1985). with the exception of the leapfrog scheme. The column labeled "storage factor" indicates the number of full arrays that must be allocated for each unknown variable in order to implement each scheme. DURRAN 713 Proposition: if the oscillation Eq. one would use backward differencing or the trapezoidal method and a very large time step.t . the (low storage) fourth-order Runge-Kutta scheme might be an even 3 The first-order Adams-Bashforth and Runge-Kutta both equivalent to forward differencing. Switching to a third-order scheme will not improve the fourth-order amplitude error.MARCH 1991 DALE R. schemes are . storage factors are not provided for the implicit methods in Table 1. but will reduce the phase error to fourth order or better. In many applications.and fourth-order versions of these schemes are stable. However. and if a given numerical application will accommodate an iterative method.t)m]. however. computed ¢/+l directly into the storage occupied by ¢/ . select the third-order AdamsBashforth method solely on the basis of its high efficiency factor. appears in the Appendix. A summary of the properties of several elementary methods appears in Table 1. it is possible to utilize less memory than that suggested by the storage factor if newly computed quantities are initially placed in a small. Some improvement relative to Asselin-filtered leapfrog differencing might be obtained from the leapfrog-trapezoidal method (Kurihara 1965) or the Magazenkov scheme because they both damp the computational mode without the loss of O[(l:J. This procedure may be used to reduce the storage requirement of the third-order Adams-Bashforth method from four full arrays to three full arrays and a smaller. Table 1 also indicates that. This high efficiency factor implies that.0 the amplitude error in the numerical solution is 1 + O[(wl:J. but unstable at the third and fourth orders. The leapfrog-trapezoidal scheme is an iterative method however.t)2) accuracy introduced by the Asselin filter. A survey of Table 1 reveals that. but first consider the stable explicit second-order methods listed in Table 1. depending on the algorithm used to solve the implicit system. among the explicit methods. the efficiency factor of the third-order Adams-Bashforth scheme is second only to the leapfrog scheme and its Asselin-filtered variant. the temporary storage array is much smaller than the full array required to hold a complete set of ¢n. if r is even. The other important consideration is accuracy and as defined in this study the efficiency factor only indicates the minimum work required to advance the solution in a stable manner. Conversely. In Table 1. temporary storage array. This assumption is appropriate when integrating all but the simplest partial difference equations. The storage factors given in Table I are upper limits that allow each method to be programmed in a straightforward manner. Note that for each entry the relationship between the total truncation error and the phase speed and amplitude error is consistent with the preceding proposition. In many instances. the third-order Adams-Bashforth method is capable of producing a stable solution with less computational effort than any other explicit scheme listed in Table 1. it is not generally possible to write the newly 2 The definition of the efficiency factor is based on the assumption that the bulk of the computational effort is associated with the evaluation of the function that determines the time derivative.ty]. the implicit Adams-Moulton schemes are stable at the first and second orders (the backward and trapezoidal methods). Most of the following discussion will be devoted to other higher-order methods. One would not.

.~ .. - \'1.~ I ~ + li "'IN s: + "eli ""..5:~C.... II i:<.....' ::::: .'i:<. _1""[::! 0 I ~ 1 ""S:! + 00 o 8 o o 8 o o -e N ~ + II ::::: t-e+ r~ k.' ~~~ II II "e:... ~-$ -<:: .:::Is:! *sII ~ ~ + II + .714 MONTHL Y WEATHER REVIEW VOLUME 119 o 8 o o 8 o o "! o N - c-I~ +I N"" + ... ~ ~ -E o ..." II II $'$'$' II ~-e:~ t.. -e _ 1 ~I ~~ +II ~ _-e: ~ i:<..:...: "'IN I ~ ~ + ~ I !+ 1 II -e- +~ cI -s~ _..: ... .: .. . N _ 00 '" + ..:0 ""10 . -e- .IN ~ .

. i:il 0 = e + "e.e 1:: .. it is noteworthy because of its low storage requirements. Magazenkov ( 1980) suggested alternating each leapfrog step with a secondorder Adams-Bashforth step.§ on § ...: s '" -e "0 J.L set to the equal-work value of 116. Note. ~ < E e Mloo I '"I" "':! I 0 '" ~ I "..9 oU ci. however.e ~ + N ~ ~ M ~ ~ V) ~ ~ r-. In order to assess the relative accuracy of the two methods when poorly resolved high-frequency modes are present. I '" "3 E t13- '" 0 ~ ~ . stable..MARCH 1991 DALE M 0 R. OJ . Magazenkov's method is possibly the most attractive second-order alternative to Asselin-filtered leapfrog time differencing.. + "e.<: ~ <o!! <I) ~ ~8 . Although the third-order Runge-Kutta method is less efficient and less accurate than the third-order Adams-Bashforth method. DURRAN 715 '>< e = e.:::1. The remainder of this section will. 9. 5: + * 13- .<: 00 N ~'" ~Io N + r- ~"'~ 1 I 0 M~ N ~'" ~Io 00 I ~Io -~ + '". = ~.. . '".I~ ~I.. Willimson ( 1980) discovered a small class of lowstorage third-order Runge-Kutta methods and rec- ... the amplitude and phase-speed errors of the third-order AdamsBashforth method are much less than those of the Runge-Kutta method. +++ I: r. This accuracy comparison is based solely on the truncation error however. The Courant number of liz is near the maximum stable time step permitted by the third-order Runge. so it only applies to well-resolved solutions. S!. ~ + ".. The two solutions are compared at a nondimensional time of 20 in Fig.a N '" ~ 3 0 r-. + r "e. + I . J..~ '" I ~ ~::::.:: . s ~ ~ 0 t 13- ~~~~ II ~ * 13- ~ ~ .:::1.:::. N oo N 00 "": g <I) "" ~ .!!l .!:. that neither of these second-order schemes achieve quite the same efficiency factor as the third-order Adams-Bashforth method.l o:l ::: M f-o < ~ ~ on + r ~ on + t: ~ I ~ ~ + ~ ~ I "e. When the two schemes are compared on this equal-work basis. therefore. ~ J. "0 '" .:::~ II II II NN ~ . The third-order Runge-Kutta method presented in Table 1 is just one member of a large family of possible Runge-Kutta schemes. II I: r 13- ~ ~~~ -s::.L = liz..:: '" ~ M :!.:: I: -I'" + "e.:::I~ 13II ~~~~ 13-13-13-13- .. i:! . The third-order Runge-Kutta scheme is stable and has less amplitude and phase-speed error than the Adams-Bashforth method.. .M onl'" "'". ..<: U :> ~ '" <I) <0. M + N ~ ~ ~ ~ M + + e- "..: s * ~ ~ better choice since it is higher order and has a similar efficiency factor.. e- <0. since the RungeKutta scheme requires three function evaluations per time step..: "e. However.:::I~ + "e. "0 '" .~ t) i£i<o!! '-IJ N ": ": ~ "1 M 0 "": ~ .~ 00 0 ~ t "E 0 "0 <I) " " . .itf! 1. and the third-order Adams-Bashforth method with J. 4 was repeated using a third-order Runge-Kutta integrator with J. N -1 '" on r0 ~'" I g <I) "0 <I) "'". second-order scheme with a damped computational mode. g5 .. .<: ~ ~ "8 <E a:l 1:: o!. which are actually smaller than those of the Asselin-filtered leapfrog scheme.Kutta scheme. -s 0 e . " M * :Q 0.j...:: + ~ M 13- ~ ~ o-: on V) N . which demonstrates that the AdamsBashforth method produces a significantly better equalwork solution. "'". the spectral-model advection test presented in Fig. All third-order Runge-Kutta schemes produce the same truncation error if they are used to integrate the simple oscillation equation... be devoted to the consideration of other higher-order schemes.>. '" -s <I) c: .. it may best be compared with the AdamsBashforth method by assuming the Adams-Bashforth integration proceeds with one-third the time step of the Runge-Kutta scheme.. Different Runge-Kutta formulae do produce different truncation errors in more complex nonlinear problems and most third-order formulae have a storage factor of 3. The adjective "equal work" will be used to describe this type of comparison in which the time step used with each method is adjusted to hold constant the ratio of the step size to the number off unction evaluations per step. ~Ioo 13~ ~ ~ ~ ~ . the result is a fully explicit.....: a ~ ~o =~ ~~t: Co _g = ~ 0 c: <E a:l . .

9.L~:!-:-.2 o -0. Young's method A (short-dashed line). The preceding discussion applies only to the errors in well-resolved solutions.L.LLL-'::-'-. Otherwise identical to Fig.. the WilliamsonRunge-Kutta method is truly third order.L. if the comparison is made on an equal-work basis (four Adams-Bashforth steps per each Runge-Kutta step). 4 ( 1962) _ In order to save space.L..4~L_. In addition.-.4 0. the phase-speed error of the Runge-Kutta scheme is five times worse that the error in the Adams-Bashforth scheme. the classical fourth-order Runge- Kutta method is listed in Table I.. .:-_ .3RD . 4a. Thus. Turning to the fourth-order methods.2 -0. The Williamson-Runge-Kutta method appears to be distinctly superior to two other time differencing techniques that have appeared in the meteorological literature-Young's method A (Young 1968) and the Lorenz three-cycle method (Lorenz 1971).6 0. The other stable third-order method appearing in Table 1 is the Adams-Bashforth-Moulton predictor corrector. ORDER ADAMS-BASHFORTH .. The Williamson-RungeKutta method is preferred over Young's method A because the latter is unstable.6 <I>(x) OA 0. This improvement in amplitude error may be lost however. I "" 2 2.. with a third-order Adams-Bashforth solution (long-dashed lined) calculated using Il = %./ 1. if one A storage factor of three is achieved using the algorithm of Blum FIG. Magnitude of the amplification factor plotted as a function of w~l. however. both generate the same truncation error. At a fixed value of w~t.3RD .L.5 I / 0... and the exact solution (dash-dotted line). the Adams-Bashforth-Moulton predictor corrector re- 1. lower order. I I _J_ I IAI 0..L_L.. However. 10. 10.ORDER RUNGE-KUTIA 1. . The three-cycle method achieves third-order accuracy only in the special case where the governing differential equation is linear (Lorenz 1971). the lagging phase-speed error of the fourth-order Runge-Kutta scheme will reinforce the lagging phase-speed error introduced by any spatial differencing. when used to integrate the oscillation equation.L..L-'.. 1. it will produce O[(wM)6] amplitude errors.8 0.-:. 10.. The amplification factors for the third-order Runge-Kutta method and Young's method A are compared in Fig. the fourth-order Runge-Kutta scheme (long-dashed line). whereas the three-cycle method is just second order.2 " . . is provided by the plot of amplification factor versus w~t in Fig.4 . for the third-order Runge-Kutta scheme (solid line).5 1. Being fourth order.L_LLLL. As was the case with the third-order RungeKutta scheme. its region of high accuracy is similar to the region in which the third-order RungeKutta scheme is both accurate and stable.8 0.~~ o 8 16 24 32 x (gridpoints) quires the same storage as the third-order AdamsBashforth method.. which will be smaller than the O[ (w~t) 4] amplitude errors produced by the third-order AdamsBashforth method whenever w~t is sufficiently small.2 <.0 0.. Some idea of the quality of a poorly resolved solution.2 ---EXACT .. It is curious that two unstable schemes (second-order Adams-Bashforth and third-order AdamsMoulton) combine to yield a stable predictor-corrector formula.L. if the method is used in combination with the fourth-order spatial filters commonly employed to control short-wavelength noise in atmospheric models. The fourth-order Runge-Kutta method strongly damps the poorly resolved modes. Equal-work comparison of a third-order Runge-Kutta solution (short-dashed line) computed using Il = '/2._L. All three methods are iterative schemes requiring three function evaluations per time step. The Williamson-RungeKutta method is similar to the Lorenz three-cycle method in that both require the same storage and. Moreover./ " .. the Adams-Bashforth-Moulton predictor corrector is inferior to the Adams-Bashforth method in equal-work comparisons. inspection of Table 1 shows that the fourth-order Runge-Kutta method has the lowest storage factor" and the highest efficiency factor of all listed fourth-order schemes. the phase-speed error of the fourth-order Runge-Kutta scheme is less than that of the third-order Adams-Bashforth method.. : .L_L.5 3 FIG. obtained using time steps near the stability limit.716 MONTHLY WEATHER REVIEW VOLUME 119 ommended the particular scheme shown in Table 1.. and seems to require more storage..

2 o -0. advantage of the large stable time step permitted by In summary.8. DURRAN 717 wishes to accurately represent the highest frequencies. On balance...2. An additional method can produce a stable solution to the oscillation assessment of the relative accuracy of the fourth-order equation with fewer function evaluations than any of Runge-Kutta and third-order Adams-Bashforth the other third. with a thirdorder Adams-Bashforth solution (long-dashed line) calculated using 11 = 0. Being fourth order..O.L.. Fig. mentioned three-cycle scheme.L-s ~\ . In atmospheric applications. in this equal-work test.. but exhibits larger dispersion curacy of the Runge-Kutta method manifests itself as and phase-speed errors than the third-order Adams.8.fourth-order method appears to be the (low storage) order Runge-Kutta solution was computed using IL Runge-Kutta scheme.2--'--' x (grid points) FiG.-.L.J..4 o'--'--'-'-J. When ap= 0. .2 -0.. higher-order Runge-Kutta forlations and in those circumstances. (33) Here W = WI + W2._ . use of the semiimplicit scheme allows such models to be efficiently integrated with a large time step that would violate the CFL condition for gravity waves ( .0 ¢n+1 _ ¢n-I o. The most attractive model to a nondimensional time of 20. scheme is second-order...L..4 2!1t . When ap.3. The phase-speed error Bashforth result. Use of the Adams-Bashforth method in semi-implicit schemes plied to the oscillation equation. one could take full mulae and the third-order Adams-Bashforth method.. < 1) and lead to instability in a fully explicit scheme.L.. because of its relatively lower Nevertheless.. (¢n+1 + ¢n-I) 2 .. The third-order Adams-Bashforth method is also shown in Fig. the third-order Adams-Bashforth the fourth-order Runge-Kutta method. Yet when applied implicit time differencing.. 4 and 9..3RD • ORDER ADAMS· BASHFORTH 1. The advantage of the fourthe maximum step selected for a fourth-order Runge..a 0. W2 is identified with the frequencies characteristic of gravity wave propagation... The basic ---EXACT properties of this technique may be illustrated by the .2 4TH .. larger than that of the third-order Adams-Bashforth One other scheme worthy of discussion is the Lorenz method when the two are compared on an equal-work four-cycle method (Lorenz 1971).L. (6): 1. in which the Asselin-filtered to nonlinear differential equations.. 11. The fourth-order Runge-Kutta solution retains plied to the oscillation equation. Kwizak and Robert ( 1971) have shown that (33) will be stable whenever (34) A sufficient condition for the satisfaction of (34) is IWI I~ Iw21.. 11. and.MARCH 1991 DALE R.a reduction in amplitude error. 0. Like Figs.. and the condition I WI I ~'W2' is almost always satisfied.L. Otherwise identical to Figs..J..that the low-storage.. The fourth.L.cycle scheme is that its storage factor is only two. on the linear test problem.of the fourth-order Runge-Kutta method is actually Bashforth solution looks the best.. the four-cycle scheme was designed to achieve high accuracy for linear differential equations while conserving storage. " I' . W2At.ORDER RUNGE-KUTIA following semi-implicit approximation to the oscillation Eq._. the third-order Adams.or fourth-order methods listed in Table methods is provided by the equal-work advection test 1. AlKutta integration should be similar to the maximum though the four-cycle scheme will perform very well step used in the third-order Runge-Kutta scheme.. the four-cycle scheme produces truncation errors identical to those of a Many large meteorological models employ semifourth-order Runge-Kutta method. = lWI¢ n + lW2 .... in some situations one might welcome order it is unlikely to be a better general purpose scheme the opportunity to damp the highest-frequency oscil.2.6 $(X) 0. the four-cycle leapfrog scheme is used to represent the advection terms while the pressure gradient and divergence terms are integrated using the trapezoidal method.. WI is identified with those frequencies characteristic of advection..LJ-L---'-c'16~'--'-'--'--'-'-2L4. Like the previously basis. 11 shows more accurate than the other third-order methods when results obtained by integrating a spectral advection compared on an equal work basis. this scheme = 0. the fourth-order acslightly more amplitude.. the third-order will exhibit less overall truncation error than the thirdAdams-Bashforth solution was calculated using IL order Adams-Bashforth method as !1t . Equal-work comparison of a fourth-order Runge-Kutta solution (short-dashed line) computed using 11 = 0.. and the Asselin filter is neglected. 4a and 9.6. Since the accurate simulation of high-speed gravity waves is thought to be unnecessary in large-scale atmospheric models.

6 fourth-order Runge-Kutta methods. the third-order AdamsBashforth method is more accurate than the leapfrog method: phase-speed errors are smaller and amplitude errors are considerably less than those generated by Asselin time filtering. Stable der accuracy of the overall integration. it has been demonstrated that if the truncation error of a difference scheme is order r and the scheme is applied 1. noniterative method that damps .4 0.and 0. provided that r is odd. As an alternative to Asselin time filtering. OCr). together with a detailed need to be engineered for each application. dashed curves are computational modes. In connection with this survey. Not surprinsingly..-2Asselin filtering. Among the more 0. Runge-Kutta method. Both Asselin time filtering and the grated at Courant numbers (W2~t) much greater than periodic reinitialization of the leapfrog solution involve arbitrary parameters: the filtering coefficient and the unity. Unfortunately. one might the accuracy of the W2 component by halving the stepattempt to control leapfrog time splitting by periodisize of the trapezoidal integration. for4iwI~t 5iwI~t + -3-A . Conclusions can be rather obscure. or better. Use of the third-order AdamsBashforth method eliminates the computational mode An analysis of the third-order Adams-Bashforth without the introduction of arbitrary parameters that method has been presented. 5 If r is even.0. number of time steps between restarts.8 attractive schemes revealed by the survey are the Magazenkov method and the low-storage third. and the IAI phase-speed error will be order r. such as the second-order in Fig.A ( . the robust stability of the original semi-implicit scheme is cally discarding data from the oldest time level. The third-order Adams-Bash. Thus.t for the case W2t. In addition. but mentation of such a scheme is more complex. even though the local amplitude and phase-speed errors of an odd-order scheme might both be O( r + 1). the scheme itself will be accurate only to FIG.2 One might expect to obtain a more accurate semiimplicit scheme by replacing the leapfrog difference in (33) with a third-order Adams-Bashforth formula. the amplitude error will be order r + I and the phase-speed error will be order r 1. Magnitude of the amplification factor for the AdamsBashforth trapezoidal semi-implicit scheme (36).2 0. the amplitude error will be order r + 2.-0. .-the computational mode while retaining second-order 0. but the impleintegrations can be achieved by reducing w2~t.___----'-_ = -I (23cjJn .4 to the oscillation equation.restarting the integration using a two time-level mula (35). The influence of these parameter values on the numerical solution 7. efficient way to avoid leapfrog time-splitin the WI component of the solution.4 -----------------method is an explicit. Restarting The magnitude of the three roots of (36) are plotted with a second-order scheme.718 MONTHLY WEATHER REVIEW VOLUME 119 scheme.6 0. This scheme is easy to implement. would preserve the second-orNote that the physical mode is always unstable. it degrades the second-order accuracy of the unadulterated leapfrog method. as a function of Wlt.1+ 12 + -2.16cjJn-1 + 5cjJn-2) forth method is also superior to the leapfrog scheme ~t 12 in that it can be easily used to integrate diffusive and Rayleigh damping processes.8 5 Note that the local truncation error of an O( r) approximation to a differential equation is O( r + 1) as per (A2). comparison of the third-order Adams-Bashforth Other accurate time-integration techniques have also method with the widely used Asselin-filtered leapfrog been surveyed. and most there is little point in using a semi-implicit scheme explicit second-order schemes are also subject to slowly unless the high-frequency modes can be stably inte. The simple Euler or forward-difference method is commonly used to reinitialize leapfrog in1 iW2~t)A3 ( 23iwI~t iW2~t) 2 tegrations.0. 12as a function of WI~t for the case w2~t = 2. but like . The Magazenkov 0. the greatest advantage of the thirdorder Adams-Bashforth method may simply be that This modification not only reduces the truncation error it is a simple.-12. (36) ward differencing is unstable and tends to amplify the high-frequency components of the solution.growing instability. The amplification factor for (35) satisfies method.2 + 1. and not shared by the Adams-Bashforth trapezoidal for. it also improves ting instability. Solid line is the physical mode.t = 2.= O. 12. or better. The preceding comments notwithstanding. in which case cjJn+1 cjJn iw - .

which is both higher order and requires less storage + O[(Wt. that the third.(1971 ). The author has benefited from a order schemes lies in the position ofthe imaginary facreview by Piotr Smolarkiewicz.n+l = eiwAtl/.has the form quires the fewest function evaluations to stably integrate the solution over a given time interval. Substitution of1/.t). Kay Dewar drafted the figures.t)'+I]. Harry Edmon assisted with the computer A = e'r" + CI(Wt. for odd-order schemes 0.extended to O[(wt.t). one A = e':" + e. and from discussions with Rashid Akmaev and Mark Borges. into (AI). it follows that. (A4) per step is held constant by adjusting the time step j=r+1 used with each method).t)'+1 + O[(wM)'+2]. the reduces to an expression of the form (A3) with er+1 same algorithm is executed every time step. in the case of method has a number of practical advantages that make a two-level scheme. (A2) where the right-hand side of(A2) is the local truncation error and r is the order of accuracy of the numerical scheme.t). The proceeds as follows. (A4) leapfrog scheme. In addition. obtained by third-order Adams-Bashforth scheme is.t).n+1 L(1/. it is frog scheme.. more efficient than any of these schemes in that it re.. will not be repeated here. Using these relations. (A3) than the Magazenkov and Asselin-filtered leapfrog schemes. which must be order Adams-Bashforth method is not a suitable sub. Grace Gudmundson edited the manuscript + O[(Wt. The derivation technique examined in this survey would appear to be of (A3 ) is particularly easy for single-step methods and the low-storage fourth-order Runge-Kutta scheme.t).. substituting Taylor series expansions for I/. (A7) (A8) where ¢n is the numerical approximation to 1/.) + = O[(t. Milne's method (Gear 1971).n) by the definition it particularly suitable as a direct substitute for the of the amplification factor.+2 + O[(wt. whereas for even-order schemes = 1 + (CI + C2)(wt.n L ciiwt..+3]. The most accurate and most efficient iterative where er+1 and er+2 are real coefficients. the trapezoidal method. When it is compared on an equal-work basis (in which the ratio of the step size to the number of function evaluations = I/.. AI/. The details of Gear's proof.+I and e +2 = -C +2.+2 graphics.. there is no need to iterate.the form (7). The truncation error...+2 + O[(wt. The storage requirements of the Magazenkov time step. Unlike the Magazenkov method. method.n. the third-order Adams-Bashforth equation 1/. and L( ¢) is a linear function of ¢ at previous times. and the derivation follows Gear applications.+1 + iC2(wt..L(I/. the third-order AdamsBashforth method is more accurate than all the pre. If the scheme is a multistep r r Runge-Kutta schemes.. and Young's method A (Young 1968) are examples of even-order .useful to first establish that for an rth order scheme modate an iterative method and storage is critical. (A6) where CI and C2 are real constants.+3]. In order to prove the proposition. It is assumed that w is constant in (6) and that the coefficients of L are therefore independent of the This proves the first part of the proposition.t). except the fourth-order Runge-Kutta mula.+2 might utilize the Williamson-Runge-Kutta scheme.+1(iWt.t).t). Since 1/..+3]. Furthermore. The amplification factor for odd-order schemes can be expressed in the form and Chris Bretherton introduced the author to Mathematica. (A3) applies to the "principle root" (or the which can simplify the coding in complex modeling physical mode).t).is the exact solution to the oscillation method. Unlike the = -C.MARCH 1991 DALE R. DURRAN 719 accuracy.t). can be written 00 APPENDIX A = eiwAt + iCI(wt.+1 + er+2(iwt. however. The leapfrog scheme. in which case solutions may be sought in method are similar to those of the Asselin-filtered leap.and oddAcknowledgments. It does appear. If the numerical application will accom. Note that the amplitude error associated with odd-order methods must always be order r + 1.(nt. Evaluating (AA *) 1/2... are somewhat tedious and stitute for leapfrog differencing in semi-implicit models. Joe Barsugli tor i in the truncation error..into (AI) yields an expression of the form 1/. however.+3]. The important difference between even. but that cancellation can reduce the amplitude error of even-order schemes beyond order r + 2.t)'+1 Proof of Proposition Suppose that approximate solutions to the oscillation Eq. (6) are computed by the numerical scheme ¢n+1 + L(¢) = + C2(wt. (AI) IAI IAI = I + CI(Wt..t)'+3]. This research was supported by NSF Grants ATM-8796281 and the amplification factor for even-order methods and ATM-89l4852.t)j.t).where the Cjare determined by the finite-difference forceding schemes. (A5) and typed Table 1.n = ..

455-467. and J. Wea.. 99. Rev. Comput. J. Note that reductions in the phase-speed error of odd-order schemes are possible through cancellation. R. but all even rth-order schemes have order-r phase-speed errors. CO 80307... 111. NCAR/TN-21O + STR. 107 pp. 10. 64 pp.] Williamson. N.. 88 pp... 1988: Mathematica: A System/or Doing Mathematics by Computer. 1968:Comparative properties of some time differencing schemes for linear and nonlinear oscillations. John Wiley and Sons.. for r odd.C1)(wAt)'+1 + O[(WAt)'+2]. 477 pp. Wea.. R. 1. W. (AI3) This concludes the proof. P. Japan. J. Gear. Mon. M. 1983: A compressible model for the simulation of moist mountain waves. Mesinger. NCAR Technical Note. Robert. Rev. Sci.720 MONTHL Y WEATHER REVIEW VOLUME119 schemes with amplitude errors smaller than order r The phase-speed error may be obtained by observing that. Magazenkov. 100. 1971: An N-cycle time differencing scheme for stepwise numerical integration. 11-26. 1984: A large-eddy simulation model for the study of planetary boundary-layer turbulence. and R. 1976: Numerical Methods Used in Atmospheric Models. GARP Publ. Mon. 1966: Numerical integrations of the primitive equations by a simulated backward difference method. H. one may show that for rodd R=1 + (C2 . J. R.. 1070-1096. Y. 1980: Numerical Prediction and Dynamic Meteorology. 44. 237-244.. H. 1050-1065. B. K. A. 93. Kreiss. Mon. Mon. Lilly.. Meteor.. the author is not aware of an odd rth-order scheme that achieves better than order r + 1 phasespeed error. Rev.. 1. C... 1980: Tr. Atmos. and for r even. Phys. 1963: Deterministic nonperiodic flow. Car < 1. Sci.Cd(wAt). World Meteorological Organization. Addison-Wesley. Ser.. 33-46. ~[A] 9t[A] = tanwAt +2.487-490. D. 17. 1978: The simulation of threedimensional convective storm dynamics.. 41.. Inc. 113. Gl. + (C2 . 1962: A modification of the Runge. 120-129. Wea. Soc. B. c. J. E. Wea. Williamson. 357-364. Japan. Boulder. Durran. Vol. 48-56.. 111. Atmos. 176-187. Wea. 35. 1983: Description of NCAR Community Climate Model (CCMOB). Haltiner. T. and A. Wea. Moeng. Prentice-Hall. S. K. Mon. Rev. 410. 93. and A. These results notwithstanding. J.. Mon. Wea. Klemp. J.. World Meteorological Organization. D. 1965: On the use of implicit and iterative methods for the time integration of the wave equation. 1965: On the computational stability of numerical solutions of time-dependent nonlinear geophysical fluid dynamics problems.O. T. REFERENCES Asselin. F. Ser. 96. 99. --. 749 pp.. 1.. 644-648. 32-36. 20522062. 1.. Mon.. 1985: A two-step scheme for the advection equation with minimized dissipation and dispersion errors. W. Wilhelmson.. Wolfram. 130-141. Oliger. 1. A.Kutta fourth-order method... Williams.. Arakawa. 1966: The integration of a low-order spectral form of the primitive meteorological equations. 44. 1983: The effects of the Asselin time filter on numerical solutions to the linearized shallow-water wave equations. E.. 1971: Numerical Initial Value Problems in Ordinary Differential Equations. Atmos. Rev.. E. 253 pp. [Available from Information Services. ~[AJ 9t[AJ = tanwAt (A9) + C1(wAt)'+1 + O[(WAt)'+2]. 1. N. Mon. Rev. Wea. 20. J. Klemp.. J. and R... Rev. 2. Observ. NCAR. Young. Kurihara. 76-84... Meteor. Robert. Rev. Math.+2 + O[(wAt). G.+3]. . Wea. J. J. 2341-2361. Soc. Mon. Ser. 35. Rev. 1972: Frequency filter for time integrations. (All) which holds whenever a. 1973: Methods/or the approximate solution of time-dependent problems. Box 3000. GARP Pub!. Geojiz. (AlO) Using the relation tana + Car = tan ( a + Car) + O[ ar+2]. 1980: Low-storage Runge-Kutta schemes. B. Johnson. 1971: A semi-implicit scheme for gridpoint atmospheric models of the primitive equations. Sci. 1. Kwizak. Schlesinger. 16. Blum. D.. Comput. Uccellini and D. Lorenz. A. Takacs..-H. Matsuno. (AI2) whereas for r even R=1 + C1(wAtY + O[(WAty+lJ. and J. R.

- Unphysical Solutions of the Kadanoff-Baym Equations in Linear Response
- Galactic Dynamics
- E. M. Cioran Cahiers 1957-1972 (French Edition).pdf
- Lehmann representation of the nonequilibrium self-energy
- Psychedelic Medicine Ibogaine
- Development Countries Warsaw Pact
- The Lysenko Effect and the autonomy of science
- The Lysenko Effect
- The Lysenko Effect
- In Memoriam Nicolaas Gotfried Van Kampen (1921-2013)
- Nonequilibrium mean-field theory
- The Threefold Way - Dyson (1961)
- The Cephalopoda
- Inglourious Basterds Script
- Gnu Plot
- h6
- Quantum Theory of Condensed Matter
- p406_1
- Dynamic Response Functions of Hubbard Model in Gutzwiller Approximation
- Death of Lonesome George the Tortoise Gives Extinction a Face - NYTimes
- Second Order
- EM Genius and Mystery
- Dziga Vertov | Senses of Cinema
- Statistical Physics of Social Dynamics Rev Mod Phys 81 591
- p882_1

- Unphysical Solutions of the Kadanoff-Baym Equations in Linear Response
- Galactic Dynamics
- E. M. Cioran Cahiers 1957-1972 (French Edition).pdf
- Lehmann representation of the nonequilibrium self-energy
- Psychedelic Medicine Ibogaine
- Development Countries Warsaw Pact
- The Lysenko Effect and the autonomy of science
- The Lysenko Effect
- The Lysenko Effect
- In Memoriam Nicolaas Gotfried Van Kampen (1921-2013)
- Nonequilibrium mean-field theory
- The Threefold Way - Dyson (1961)
- The Cephalopoda
- Inglourious Basterds Script
- Gnu Plot
- h6
- Quantum Theory of Condensed Matter
- p406_1
- Dynamic Response Functions of Hubbard Model in Gutzwiller Approximation
- Death of Lonesome George the Tortoise Gives Extinction a Face - NYTimes
- Second Order
- EM Genius and Mystery
- Dziga Vertov | Senses of Cinema
- Statistical Physics of Social Dynamics Rev Mod Phys 81 591
- p882_1
- Quantum Theory of Optical and Electronic Properties of Semiconductors 4th Ed - H. Haug, S. Koch (World, 2006) WW
- Merzbacher Quantum Mechanics
- 1. Charles Darwin - On the Origin of Species (1859)

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd