2538 MONTHLY WEATHER REVIEW VOLUME 137

A Proposed Modification to the Robert–Asselin Time Filter*

PAUL D. WILLIAMS
Department of Meteorology, University of Reading, Reading, United Kingdom, and Kavli Institute for Theoretical Physics,
University of California, Santa Barbara, Santa Barbara, California

(Manuscript received 15 July 2008, in final form 10 January 2009)

ABSTRACT

The Robert–Asselin time filter is widely used in numerical models of weather and climate. It successfully
suppresses the spurious computational mode associated with the leapfrog time-stepping scheme. Unfortu-
nately, it also weakly suppresses the physical mode and severely degrades the numerical accuracy. These two
concomitant problems are shown to occur because the filter does not conserve the mean state, averaged over
the three time slices on which it operates. The author proposes a simple modification to the Robert–Asselin
filter, which does conserve the three-time-level mean state. When used in conjunction with the leapfrog
scheme, the modification vastly reduces the impacts on the physical mode and increases the numerical
accuracy for amplitude errors by two orders, yielding third-order accuracy. The modified filter could easily be
incorporated into existing general circulation models of the atmosphere and ocean. In principle, it should
deliver more faithful simulations at almost no additional computational expense. Alternatively, it may permit
the use of longer time steps with no loss of accuracy, reducing the computational expense of a given simulation.

1. Introduction This paper presents a possible avenue for progress with

the third of these three challenges, which has received
From a functional perspective, the task of predicting
scant attention compared to the extensive research ef-
future weather and climate may be reduced to the fol-
forts devoted to the first two.
lowing iterative procedure. First, given the state of the
Pfeffer et al. (1992) have assessed the sensitivity of
atmosphere, ocean, and other Earth-system components
an atmospheric general circulation model to the time
at any time (the input), use the governing equations to
stepping, for fixed spatial discretization and physical
compute the state at a slightly later time (the output).
parameterizations. They find that two different meth-
Then, repeat the loop as many times as required, always
ods of time discretization—the leapfrog and Matsuno
using the previous output as the next input.
schemes—result in significant quantitative differences
The above prediction framework presents three main
in the simulated climate. For example, the leapfrog
challenges, each of which potentially degrades the reli-
scheme gives much more precipitation over the western
ability of the forecast. First, Earth observations, which
tropical Pacific Ocean and less precipitation over the
always contain measurement errors, are required to
western North Atlantic Ocean. Therefore, time step-
serve as the initial state. Second, the vast array of active
ping appears to be an important contributor to model
physical processes and interactions is incompletely
error.
known and imperfectly represented in the spatially
Many different time-stepping methods have been pro-
truncated governing equations. Third, the discrete
posed, including the leapfrog scheme (or centered
stepping from one time level to the next is merely an
difference scheme), the Matsuno scheme (e.g., Pfeffer
approximation to the exact time-continuous evolution.
et al. 1992), the Adams–Bashforth family of schemes
(e.g., Durran 1991) and the Runge–Kutta family of
* Kavli Institute for Theoretical Physics Report 08-95. schemes (e.g., Kar 2006). The leapfrog scheme has
emerged as the method of choice in weather and climate
models, despite related disciplines choosing differently
Corresponding author address: Paul D. Williams, Department of
Meteorology, University of Reading, P.O. Box 243, Earley Gate,
(e.g., Runge–Kutta schemes are widely used in com-
Reading RG6 6BB, United Kingdom. putational fluid dynamics but hardly ever used in nu-
E-mail: [email protected] merical weather prediction) and despite evidence that

DOI: 10.1175/2009MWR2724.1

! 2009 American Meteorological Society

AUGUST 2009 WILLIAMS 2539

other methods may be superior (e.g., the third-order database, mostly in journals of meteorology and the
Adams–Bashforth scheme is more accurate; Durran atmospheric sciences (around 300 citations) but also in
1991). The leapfrog scheme is used so widely in weather journals of oceanography (around 100 citations) and
and climate models probably because it is easy to im- fluid mechanics (around 50 citations). Examples include
plement, computationally inexpensive, and has low run- the use of the filter in models of regional climate (e.g.,
time storage requirements. Indeed, today’s widespread Caya and Laprise 1999), palaeoclimate (e.g., Fraedrich
use of the leapfrog scheme in general circulation models et al. 2005), ocean circulation (e.g., Griffies et al. 2001),
is perhaps merely a legacy of computer memory having geophysical fluid dynamics (e.g., Ford 1994; Bartello
been such a severe constraint when the models were first 2002), rotating laboratory fluids (e.g., Williams et al.
developed. 2009), and the atmosphere of Mars (e.g., Hartogh et al.
A major problem with the leapfrog scheme is that it 2005). André Robert’s contributions to numerical
admits spurious computational modes (e.g., Mesinger modeling, including the time filter, have been reviewed
and Arakawa 1976; Haltiner and Williams 1980; Durran by Staniforth (1997) following a memorial symposium
1999). In general, a differential equation that is first order held at the University of Québec in 1994.
in time has one degree of freedom, but an n-time-level Currently, the Robert–Asselin filter is used in
numerical approximation to it constitutes an (n 2 1)th-
d operational numerical weather prediction models,
order difference equation with n 2 1 degrees of freedom.
including the Mesoscale Model (MSM) of the Japan
Of these n 2 1 modes, one is the physical mode and the
Meteorological Agency (JMA), the global model of
remaining n 2 2 are computational modes. The leapfrog
the Australian Bureau of Meteorology (BOM) and
is a three-time-level scheme, so one computational mode
Research Centre (BMRC), the global model (GME)
arises in it, in addition to the physical mode, because a
and regional model (COSMO-EU) of Deutscher
second-order difference equation is used to approximate
Wetterdienst (DWD), and the Royal Netherlands
a first-order differential equation. The computational
Meteorological Institute (KNMI) model;
mode (or parasitic mode) is manifest as a spurious oscil-
d atmospheric general circulation models for climate
lation between even and odd time steps, which is referred
simulation, including the ECHAM5 model of the
to as time splitting.
Max-Planck-Institut für Meteorologie (MPI-M) and
One possible solution to time splitting is to periodi-
the Community Atmosphere Model (CAM) of the
cally reinitialize the leapfrog scheme by applying a
National Center for Atmospheric Research (NCAR);
single step of a two-time-level scheme, which does not
d ocean general circulation models, including Océan
admit any computational modes. For example, Pfeffer
Parallélisé (OPA), the Nucleus for European Mod-
et al. (1992) apply a single Matsuno step after every 11
eling of the Ocean (NEMO), the oceans of Met Office
leapfrog steps. This approach does not remove the
Hadley Centre climate models [i.e., the Hadley Cen-
computational mode, but merely resets its amplitude to
tre Coupled Climate Model version 3 (HadCM3),
zero periodically so that it never becomes large enough
the Hadley Centre Global Environmental Model
to be problematic.
(HadGEM), the High Resolution Global Environ-
The far more widely used solution to time splitting is
mental Model (HiGEM), and the Fast Met Office
to apply a time filter during the time-stepping procedure.
U.K. Universities Simulator (FAMOUS)], the Hybrid
Robert (1966) designed such a filter for the leapfrog
Coordinate Ocean Model (HYCOM), and (as an
scheme and Asselin (1972) showed that it selectively
option) the Geophysical Fluid Dynamics Laboratory
suppresses the computational mode but leaves the
(GFDL) Modular Ocean Model (MOM); and
physical mode relatively undamped at low frequencies.
d models of the fluids in rotating annulus laboratory
The filter is now referred to as the Robert filter, the
experiments, including the Quasi-Geostrophic Model
Asselin filter, or the Robert–Asselin filter. The behavior
for Investigating Rotating fluids Experiments
of the filter has been investigated not only for simple
(QUAGMIRE) and the Met Office/Oxford Rotating
equation sets, with no space dependence, but also for
Annulus Laboratory Simulation (MORALS).
the shallow-water equations (Schlesinger et al. 1983)
and the hydrostatic primitive equations (Cordero and Despite its unquestioned success, the Robert–Asselin-
Staniforth 2004). filtered leapfrog scheme suffers from two related prob-
As testament to the filter’s success, Asselin (1972)1 lems. First, in addition to suppressing the computational
has been cited over 450 times according to one citation mode, the scheme also weakly suppresses the physical
mode. Therefore, physical quantities (e.g., energy) that
are conserved by the time-continuous equations are not
1
Robert (1966) predates the standard citation databases. necessarily conserved by the time-discretized equations
2540 MONTHLY WEATHER REVIEW VOLUME 137

FIG. 1. Graphical comparison of the operation of (a) the standard Robert–Asselin filter and (b) the modified
family of filters proposed in this paper. Points at three consecutive time levels are shown (marked with times signs)
and a straight line is drawn between the two outer points (dashed). The standard filter moves the inner
point through a displacement d, defined by (1). The modified filter moves the inner and right outer points through
displacements ad and (a 2 1)d, respectively, where 0 # a # 1. For the configuration of three points shown,
d . 0.

when the filter is activated. The damping and noncon- filtered leapfrog scheme. Section 4 concludes the paper
servation may be benign for sufficiently short integra- with a summary and discussion.
tions, but possibly not for longer ones.
Second, the Robert–Asselin filter severely degrades
2. The Robert–Asselin filter and
the leapfrog scheme’s numerical accuracy, measured as
proposed modification
the rate at which the error tends to zero as the time step
is progressively refined. Specifically, a numerical scheme The standard Robert–Asselin filter, and the modified
is defined to be nth-order accurate if, after a given time filter proposed in this paper, are illustrated graphically
interval, the difference between the numerical solution in Fig. 1. Suppose that the values of a dependent vari-
of the time-discretized equations with time step Dt, and able, x, are given at three successive and equally spaced
the exact solution of the time-continuous equations, times, tn21, tn, and tn11. Then, from a geometrical per-
scales as (Dt)n as Dt / 0. Higher-order schemes are spective, the standard filter (Fig. 1a) operates by moving
generally preferred to lower-order schemes, because the inner point, with coordinates [tn, xn], a fraction n
they may permit the use of longer time steps with no loss toward the midpoint, [tn, (xn21 1 xn11)/2], of the two
of accuracy, reducing the computational expense of a outer points. Therefore, the displacement of the inner
given simulation. The Robert–Asselin-filtered leapfrog point under the influence of the filter is
scheme is only first-order accurate for amplitude errors
(although higher-order contributions may dominate for d 5 (n/2)(xn!1 ! 2xn 1 xn11 ). (1)
very small values of the filter parameter).
The filter parameter, n, is usually chosen to be O(0.01–
Because the Robert–Asselin filter is used so widely,
0.2).
simple-to-implement modifications that deliver more
Two relevant properties of the three points are their
faithful simulations are very attractive. The author
mean,
proposes such a modification in this paper. When used
in conjunction with the leapfrog scheme, the modifica- xn!1 1xn 1xn 1 1
tion vastly reduces the impacts on the physical mode Mn 5 , (2)
3
and increases the numerical accuracy for amplitude
errors by two orders, yielding third-order accuracy. and curvature,
Section 2 motivates the modified filter from a geometrical
perspective. Section 3 derives analytically the amplifi- Cn ;xn!1 !2xn 1xn11 . (3)
cation factor and numerical accuracy for the modified
filtered leapfrog scheme, and compares them with the By displacing xn through the amount d, the standard
corresponding results for the standard Robert–Asselin- Robert–Asselin filter reduces the magnitude of the
AUGUST 2009 WILLIAMS 2541

FIG. 2. Two numerical solutions to (4) and (5) with v 5 1 rad s21, both obtained using the leapfrog
scheme with Dt 5 0.2 s. The computational mode is controlled using either the standard Robert–Asselin
filter (with a 5 1 and n 5 0.2) or the modified filter proposed in this paper (with a 5 1/2 and n 5 0.2). A
single two-time-level forward step is used to initiate the leapfrog scheme. The initial condition is X 5 1,
Y 5 0, for which the exact solution (also plotted) is X 5 cosvt and Y 5 sinvt.

curvature of the three points, |Cn|. When used in con- dY

5 1vX, (5)
junction with the leapfrog scheme, this feature of the dt
filter strongly suppresses the computational mode, as
desired. But, crucially, the application of the filter does by alternately applying a leapfrog step and the modified
not conserve the three-time-level mean, Mn. The theo- filter with either (in two separate integrations) a 5 1
retical analysis of section 3 will show that, when used (i.e., the standard Robert–Asselin filter) or a 5 1/2. The
in conjunction with the leapfrog scheme, it is this fea- numerical solutions so obtained are compared with each
ture of the filter that severely degrades the numerical other, and with the exact solution, in Fig. 2, at the pa-
accuracy. rameter values given in the caption. Amplitude errors
In an attempt to include the possibility of conserving are clearly much smaller with the modified filter than with
the three-time-level mean, the modified filter proposed the standard filter. One consequence is that X2 1 Y2,
in this paper (Fig. 1b) acts on the right outer point as which is conserved by the continuous equations and
well as the inner point. Specifically, for any a satisfying corresponds to the energy of the oscillation, decreases
0 # a # 1, the modified filter displaces xn through the by 89% using the standard filter, but is approximately
amount ad and xn11 through the amount (a 2 1)d, conserved using the modified filter, between the be-
where d is given by (1). All members of this family of ginning and end of the integration shown in the figure.
modified filters reduce the magnitude of the curvature
of the three points, jCnj, with a controlling the relative
contributions to the reduction from the displacements 3. Theoretical analysis
of the inner and right outer points. The special case Amplitude and phase errors of time-stepping schemes
a 5 1 yields the standard Robert–Asselin filter dis- are traditionally examined by analyzing solutions to the
cussed above, which displaces the inner point only. The oscillation equation (e.g., Durran 1999), which, for the
special case a 5 0 displaces the right outer point only. complex variable F(t), is
The special case a 5 1/2 will be of particular interest in
this paper, because it displaces the inner and right outer dF
points equally and oppositely, conserving the three- 5 ivF, (6)
dt
time-level mean, Mn.
pffiffiffiffiffiffi
Before embarking upon the theoretical analysis, we where i 5 !1 and v is a given (real) angular fre-
briefly demonstrate the improvement that may be quency. Equation (6) is related to (4) and (5) by F 5
achieved by the proposed modification, when used in X 1 iY. Using the modified filter proposed in section 2
conjunction with the leapfrog scheme. We numerically to control the computational mode, the leapfrog scheme
integrate the equations of simple harmonic motion, for (6), with time step Dt, is
dX
5 ! vY and (4)
dt F(t 1 Dt) 5 F(t ! Dt) 1 2ivDtF(t), (7)
2542 MONTHLY WEATHER REVIEW VOLUME 137

na It follows from (8) and (9) that the unfiltered, singly

F(t) 5 F(t) 1 [F(t ! Dt) ! 2F(t) 1 F(t 1 Dt)], (8)
2 filtered, and doubly filtered values share a common
complex amplification factor A, defined by
and

n(1 ! a) F(t 1 Dt) F(t 1 Dt) F(t 1 Dt)

F(t 1 Dt) 5 F(t 1 Dt) ! [F(t ! Dt) ! 2F(t) A5 5 5 . (10)
2 F(t) F(t) F(t)
1 F(t 1 Dt)]. (9)
Rewriting (7)–(9) with function evaluations at time
In this three-stage method, (7) implements the basic
t only, using (10), yields three equations in the three
leapfrog scheme and (8) and (9) implement the modi-
unknowns: A, F(t)/F(t), and F(t)/F(t). Solving for A
fied filter, with 0 # a # 1. Here F denotes a provisional
gives
value, obtained by applying (7) during the current time
step; F denotes another (singly filtered) provisional $ " # %
n(1 ! a)
value, obtained by applying (9) during the current time A2 ! n 1 2 1 ! ivDt A1 n ! 11 naivDt 5 0,
step; and F denotes the definitive (doubly filtered) value, 2
obtained by applying (8) during the next time step. The (11)
occurrence of filtered values on the right sides of (7)–(9)
makes the scheme recursive: F is overwritten with F as from which the numerical amplification factor is found
soon as it is calculated, and so is F with F. to be

" #
n n(1 ! a)
A6 (a, n, vDt) 5 1 1! ivDt
2 2
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (12)
& " #
n '2 n(1 ! a) 2 & n'
6 1! ! 1! (vDt)2 1 n 1 ! (1 ! a)ivDt.
2 2 2

Assuming 1 2 n/2 . 0, and taking the output of the the shrinking radius for a 5 1 (Fig. 3d), correspond
square root operator to be the branch with nonnegative respectively to an artificial amplification and suppres-
real part, then the positive sign (A1) corresponds to the sion of the physical mode. The numerical amplification
physical mode and the negative sign (A2) to the com- factors for the computational mode (Figs. 3b–d) rotate
putational mode. For the special case a 5 1, (12) re- clockwise in the second quadrant as vDt increases from
duces to the amplification factor derived by Asselin 0 to 1. They each remain inside the unit circle, corre-
(1972) for the standard Robert–Asselin-filtered leap- sponding to a suppression of the computational mode,
frog scheme, as expected. The exact solution to (6) is as desired.
F(t) 5 F(0) exp(ivt), from which the exact amplification As suggested by Fig. 3, the standard Robert–Asselin
factor is found to be filter (a 5 1) behaves qualitatively differently from all
other filters in the modified family (a 6¼ 1). Only the
Aexact (vDt) 5 exp(ivDt), (13) standard filter (Fig. 3d), exhibits a point in the complex
plane at which the amplification factors for the physical
for comparison with (12). and computational modes meet. The singularity occurs
Figure 3 compares the amplification factor for the because a 5 1 is the only case for which the imaginary
numerical solution, in the three cases a 5 0, a 5 ½, and term within the square root of (12) vanishes, allowing
a 5 1, with the amplification factor for the exact solu- A1 5 A2 at vDt 5 1 2 n/2. For all other values of a, the
tion. The exact amplification factor (Fig. 3a) lies on the presence of the imaginary term ensures that there is no
unit circle in the first quadrant and rotates anticlockwise value of vDt for which A1 5 A2, and the singular be-
as vDt increases from 0 to 1. The numerical amplifica- havior of the standard Robert–Asselin filter is avoided.
tion factors for the physical mode (Figs. 3b–d) also ro- Figure 4 shows in more detail how the magnitudes
tate anticlockwise in the first quadrant, but depart of the amplification factors depend upon vDt. The qual-
slightly from the unit circle as vDt increases from 0. The itatively different behavior between the cases a 6¼ 1
growing radii for a 5 0 (Fig. 3b) and a 5 ½ (Fig. 3c), and and a 5 1 is clearly visible. The singularity for the case
AUGUST 2009 WILLIAMS 2543

FIG. 3. Trajectories through the complex plane traced out by various amplification factors for the
oscillation equation, (6), as vDt increases from 0 to 1. The plots compare (a) the exact amplification
factor, given by (13), with (b)–(d) the numerical amplification factors for the modified filtered leapfrog
scheme, given by (12). The filter parameters are (b)a 5 0 and n 5 0.2, (c) a 5 1/2 and n 5 0.2, and (d) a 5 1 and
n 5 0.2. The case a 5 1, shown in (d), corresponds to the standard Robert–Asselin filter. In (b)–(d), solid
lines denote the physical mode (A1) and dashed lines denote the computational mode (A2). The unit
circle, centered at the origin, is drawn in gray for reference.

a 5 1 (Fig. 4c), which renders the curves non- case a 5 1 compared to the cases a 6¼ 1, the suppression is
differentiable at vDt 5 1 2 n/2, is replaced for the cases much less uniform as vDt increases from 0 to 1.
a 5 0 (Fig. 4a) and a 5 ½ (Fig. 4b) with a smooth tran- Figure 5 shows an enlarged view of how, as vDt / 0,
sition from the small vDt regime to the large vDt regime. the magnitudes of the numerical amplification factors
The consequence for the physical mode (A1) is that, for for the physical mode approach the magnitude of the
the case a 5 1 only, jA1j 2 1 changes sign as vDt in- exact amplification factor. The limiting value, unity,
creases from 0 to 1, corresponding to a transition from appears to be approached much more rapidly for the
artificial suppression to artificial amplification. The con- case a 5 ½ than for the cases a 5 0 and a 5 1, sug-
sequence for the computational mode (A2) is that, for the gesting a higher numerical accuracy. To confirm this

FIG. 4. Magnitudes of various amplification factors for the oscillation equation, (6), plotted as functions of vDt. The plots compare the
magnitude of the exact amplification factor, given by (13) to be unity, with the magnitudes of the numerical amplification factors for the
modified filtered leapfrog scheme, given by (12). The filter parameters are (a) a 5 0, (b) a 5 ½, and (c) a 5 1, with various values of n in
each case. The case a 5 1, shown in (c), corresponds to the standard Robert–Asselin filter. Solid lines denote the physical mode (A1) and
dashed lines denote the computational mode (A2).
2544 MONTHLY WEATHER REVIEW VOLUME 137

quadratic term in (20) is negative, yielding jA1j 2 1

; 2(vDt)2 as vDt / 0, in agreement with the
leading-order behavior shown in Fig. 5. Hence, the am-
plitude error per time step varies as (Dt)2 and the am-
plitude error per unit time varies as Dt. Therefore, the
numerical scheme with a 5 1 is first-order accurate for
amplitude errors and unconditionally stable. Alterna-
tively, for the special case a 5 0, the coefficient of
the quadratic term in (20) is positive, yielding jA1j 2 1
; 1(vDt)2 as vDt / 0, also in agreement with the
leading-order behavior shown in Fig. 5. Therefore, the
numerical scheme with a 5 0 is first-order accurate for
amplitude errors and unconditionally unstable.
In contrast, for the special case a 5 ½, the coefficient
of the quadratic term in (20) vanishes and the coeffi-
cient of the quartic term is positive, yielding jA1j 2 1
FIG. 5. Magnitudes of various amplification factors for the os-
cillation equation, (6), plotted as functions of vDt. The curves show ; 1(vDt)4 as vDt / 0, in agreement with the
how, as vDt / 0, the magnitudes of the numerical amplification leading-order behavior shown in Fig. 5. Hence, the am-
factors for the physical mode of the modified filtered leapfrog plitude error per time step varies as (Dt)4 and the am-
scheme, given by (12), approach the magnitude of the exact plitude error per unit time varies as (Dt)3. Therefore, the
amplification factor, given by (13) to be unity. The filter parame-
numerical scheme with a 5 ½ is third-order accurate for
ters are a 5 0, a 5 ½, a 5 0.53, and a 5 1, with n 5 0.2 in each case.
The case a 5 1 corresponds to the standard Robert–Asselin filter. amplitude errors and unconditionally unstable. In sum-
mary, the filter that conserves the three-time-level mean
suggestion, we Taylor expand the square root in (12) to gives a numerical scheme that is two orders more accurate
obtain, for the physical mode, for amplitude errors than the standard Robert–Asselin
filter. The increased accuracy may be exploited, by using it
‘

! cn(a, n)(ivDt)n.
either to decrease the error for a given time step, or to
A1(a, n, vDt) 5 (14) increase the time step without increasing the error.
n50
The unconditional instability of the case a 5 ½ may
The coefficients, cn(a, n), of the terms in the series ex- be avoided by instead choosing a U ½, which almost
pansion are real. The first few are given by conserves the three-time-level mean. The resulting filter
c0 5 1, (15) is effectively a weighted blend of the third-order filter
with a 5 ½ and the first-order filter with a 5 1, the
c1 5 1, (16) weighting of the latter being comparatively tiny. For this
case, the coefficient of the quadratic term in (20) is
1 ! n(1 ! a) negative, but very small, and the coefficient of the
c2 5 , (17)
2(1 ! n/2) quartic term is positive. The negative quadratic term
dominates for small vDt and the positive quartic term
n(1 ! a)[1 ! n(1 ! a)]
c3 5 ! , and (18) dominates for larger vDt, in agreement with the be-
4(1 ! n/2)2 havior shown in Fig. 5 for the case a 5 0.53. Therefore,
the numerical scheme with a U ½ is conditionally stable.
1 ! 2n(1 ! a) 1 n3 (1 ! a)3
c4 5 ! . (19) The finite stable range, for which jA1j # 1, may be es-
8(1 ! n/2)3 timated from (20) by approximating the quartic term (but
not the quadratic term) by its value when a 5 ½, to give
It follows from (13) and (14) that
sffiffiffi)
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi*
ffiffiffi)
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi*
ffiffiffi
( ( ( ( ( ( 1 1 n
(A ( ! (A ( ( ( 2
1 exact 5 A1 ! 1 5 2 (1 ! 2c2 )(vDt) 0 # vDt # 8 a ! 1! , (21)
2 2
1
1 (!1 1 4c2 ! 8c3 1 8c4 )(vDt)4 1 # # #:
8 or, for n $ 1,
(20)
sffiffiffi)
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi*
ffiffiffi
1
For the special case a 5 1, which corresponds to the 0 # vDt # 8 a ! . (22)
standard Robert–Asselin filter, the coefficient of the 2
AUGUST 2009 WILLIAMS 2545

TABLE 1. Summary of the conservation, stability, and accuracy for parameterizing unresolved physical processes. These
properties of the modified filter proposed in this paper, when used advances have helped to improve the fidelity of weather
in conjunction with the leapfrog scheme. The case a 5 1 corre-
and climate simulations. However, many general circu-
sponds to the standard Robert–Asselin filter.
lation models still use the same Robert–Asselin-filtered
Conserves Order of accuracy leapfrog time-stepping scheme as when they were first
three-time-level Amplitude Phase developed, despite evidence that time stepping appears to
a mean? Stability be an important contributor to model error.
0 No Unconditionally 1 2 This paper proposes a simple modification to the
unstable Robert–Asselin filter. The modified filter displaces the
½ Yes Unconditionally 3 2 state at the future time slice as well as the current time
unstable
slice. The modification yields a generalized family of
U½ Almost Conditionally stable ’3 2
1 No Unconditionally 1 2 filters, with a parameter controlling the relative sizes of
stable the two displacements. The standard Robert–Asselin
filter is a special case.
The behavior of the family of modified filters, when
These approximate formulas for the finite stable range
used in conjunction with the leapfrog scheme, is ana-
work well. For example, for the case a 5 0.53 and n 5 0.2,
lyzed. The standard Robert–Asselin filter is shown to
(21) gives 0 # vDt # 0.46, which is in good agreement
behave qualitatively differently from all other filters in
with Fig. 5.
the family. Each filter reduces the magnitude of the
The numerical scheme with a U ½ is strictly only
curvature at the three time slices operated upon, sup-
first-order accurate for amplitude errors. However, it is
pressing the computational mode. For the physical
clear from Fig. 5 that the error for the case a 5 0.53 is
mode, each filter yields second-order accuracy for phase
much smaller than the error of the first-order scheme
errors. But only the filter that conserves the three-time-
with a 5 1. Indeed, the error for the case a 2 ½ $ 1 is
level mean yields third-order accuracy for amplitude
comparable in magnitude to the error of the third-order
errors, whereas all other filters in the family (including
scheme with a 5 ½ across much of the finite stable
the standard Robert–Asselin filter) yield only first-order
range. Therefore, for practical purposes, the numerical
accuracy. The filter that conserves the three-time-level
scheme with a U ½ is as good as third-order accurate for
mean yields an unconditionally unstable scheme, but
amplitude errors.
conditional stability is recovered by adding a tiny amount
Turning finally to consider phase errors, it follows
of the standard Robert–Asselin filter, to yield a scheme
from (13) and (14) that
that is as good as third-order accurate.
arg(A1 ) ! arg(Aexact ) 5 arg(A1 ) ! vDt The modified filter proposed in this paper could im-
) * prove weather and climate models. For example, it may
1
5 c2 ! c3 ! (vDt)3 1 # # #: permit the use of longer time steps with no loss of ac-
3
curacy, reducing the computational expense of a given
(23)
simulation. Alternatively, if the time step cannot be
For all cases of a and n in the practical range, the co- lengthened because it is constrained more strongly by
efficient of the cubic term in (23) is positive, yielding other conditions (e.g., the CFL criterion) than by ac-
arg(A1) 2 vDt ; 1(vDt)3 as vDt / 0. Hence, the phase curacy requirements, then the modified filter may per-
error per time step varies as (Dt)3 and the phase error mit an increase in accuracy at almost no additional
per unit time varies as (Dt)2. Therefore, all numerical computational expense. The modified filter would be
schemes in the modified family are second-order accu- extremely easy to implement in an existing computer
rate for phase errors. model: the Robert–Asselin-filtered leapfrog routine
Table 1 summarizes the conservation, stability, and could be upgraded by changing only a few lines of code.
accuracy properties of the modified filter, when used in There may be a slight increase in the computational
conjunction with the leapfrog scheme, for various values expense—the standard scheme is a two-stage method
of a. and the modified scheme is a three-stage method—but
no extra function evaluations are required.
4. Summary and discussion
There are alternative methods for controlling the
In the decades that have elapsed since the first general computational mode of the leapfrog scheme, which do
circulation models were developed, there have been not involve the application of a time filter. Kurihara
major advances in the Earth observation systems from (1965) proposed the leapfrog–trapezoidal method, which
which initial conditions are derived, and in techniques consists of obtaining a provisional value by applying a
2546 MONTHLY WEATHER REVIEW VOLUME 137

leapfrog predictor and then improving it by recursively semi-Lagrangian discretizations. Mon. Wea. Rev., 132, 600–
applying a corrector. Being a predictor–corrector 610.
Déqué, M., and D. Cariolle, 1986: Some destabilizing properties of
method, however, this scheme is iterative and poten-
the Asselin time filter. Mon. Wea. Rev., 114, 880–884.
tially computationally demanding. Magazenkov (1980) Durran, D. R., 1991: The third-order Adams–Bashforth method:
proposed the alternate application, from one time step An attractive alternative to leapfrog time differencing. Mon.
to the next, of a leapfrog step and a second-order Wea. Rev., 119, 702–720.
Adams–Bashforth step. The need to execute a different ——, 1999: Numerical Methods for Wave Equations in Geophysical
Fluid Dynamics. Springer-Verlag, 482 pp.
algorithm at even- and odd-numbered time steps is
Ford, R., 1994: Gravity-wave radiation from vortex trains in ro-
cumbersome, however. In contrast, the modified filtered tating shallow-water. J. Fluid Mech., 281, 81–118.
leapfrog scheme proposed in this paper is noniterative Fraedrich, K., H. Jansen, E. Kirk, U. Luksch, and F. Lunkeit, 2005:
and nonalternating, yet still suppresses the computa- The Planet Simulator: Towards a user friendly model. Meteor.
tional mode and achieves third-order numerical accur- Z., 14, 299–304.
Griffies, S. M., R. C. Pacanowski, M. Schmidt, and V. Balaji, 2001:
acy for amplitude errors.
Tracer conservation with an explicit free surface method for
Déqué and Cariolle (1986) have shown that, despite z-coordinate ocean models. Mon. Wea. Rev., 129, 1081–1098.
the demonstrated ability of the standard Robert–Asselin Haltiner, G. J., and R. T. Williams, 1980: Numerical Prediction and
filter to stabilize numerical solutions to atmospheric Dynamic Meteorology. 2nd ed. Wiley, 496 pp.
motion equations for certain combinations of temporal Hartogh, P., A. Medvedev, T. Kuroda, R. Saito, G. Villanueva,
A. Feofilov, A. Kutepov, and U. Berger, 2005: Description
differencing and physical processes, in some other cases
and climatology of a new general circulation model of the
even a very weak filter may lead to an instability that Martian atmosphere. J. Geophys. Res., 110, E11008, doi:10.1029/
can only be suppressed by a severe reduction of the time 2005JE002498.
step. It remains to be seen whether the modified filter Kar, S. K., 2006: A semi-implicit Runge–Kutta time-difference
proposed in this paper also exhibits this unexpected scheme for the two-dimensional shallow-water equations.
Mon. Wea. Rev., 134, 2916–2926.
behavior. Finally, it is possible that the modified filter
Kurihara, Y., 1965: On the use of implicit and iterative methods for
could also improve the Robert–Asselin-filtered Adams– the time integration of the wave equation. Mon. Wea. Rev., 93,
Bashforth schemes (e.g., Tandon 1987), but the explo- 33–46.
ration of this possibility is left for future work. Magazenkov, L. N., 1980: Time integration schemes for fluid
dynamics equations, effectively damping the high frequency
components (in Russian). Tr. Gl. Geofiz. Obs., 410, 120–
Acknowledgments. The author is funded through a
129.
Fellowship from the U.K. Natural Environment Re- Mesinger, F., and A. Arakawa, 1976: Numerical Methods Used in
search Council (NE/D009138/1). This study was sup- Atmospheric Models. Global Atmospheric Research Program
ported in part by the National Science Foundation (GARP) Publications Series 17, Vol. I, GARP, 64 pp.
(PHY05-51164) and by a travel grant from the Royal Pfeffer, R., I. Navon, and X. Zou, 1992: A comparison of the im-
pact of two time-differencing schemes on the NASA GLAS
Astronomical Society. The suggestions of two anony-
climate model. Mon. Wea. Rev., 120, 1381–1393.
mous reviewers are gratefully acknowledged. The car- Robert, A. J., 1966: The integration of a low order spectral form of
bon footprint of this study—due to international air the primitive meteorological equations. J. Meteor. Soc. Japan,
travel and office power consumption—is estimated by 44, 237–245.
the author to be 3100 kg of CO2. Schlesinger, R., L. Uccellini, and D. Johnson, 1983: The effects of
the Asselin time filter on numerical solutions to the linear-
ized shallow-water wave equations. Mon. Wea. Rev., 111,
REFERENCES 455–467.
Staniforth, A., 1997: André Robert (1929–1993): His pioneering
Asselin, R., 1972: Frequency filter for time integrations. Mon. Wea. contributions to numerical modeling. Numerical Methods in
Rev., 100, 487–490. Atmospheric and Oceanic Modelling: The André J. Robert
Bartello, P., 2002: A comparison of time discretization schemes for Memorial Volume, C. A. Lin, R. Laprise, and H. Ritchie, Eds.,
two-timescale problems in geophysical fluid dynamics. NRC Research Press, 25–54.
J. Comput. Phys., 179, 268–285. Tandon, M., 1987: Robert’s recursive frequency filter: A re-
Caya, D., and R. Laprise, 1999: A semi-implicit semi-Lagrangian examination. Meteor. Atmos. Phys., 37, 48–59.
regional climate model: The Canadian RCM. Mon. Wea. Rev., Williams, P. D., T. W. N. Haine, P. L. Read, S. R. Lewis, and
127, 341–362. Y. H. Yamazaki, 2009: QUAGMIRE v1.3: A quasi-geostrophic
Cordero, E., and A. Staniforth, 2004: A problem with the model for investigating rotating fluids experiments. Geosci.
Robert–Asselin time filter for three-time-level semi-implicit Model Dev., 2, 13–32.