Title: Splitting methods

Name: Sergio Blanes1 , Fernando Casas2 , Ander Murua3

Affil./Addr. 1: Instituto de Matemática Multidisciplinar

Universitat Politècnica de València

E-46022 València

Spain

Affil./Addr. 2: Departament de Matemàtiques and IMAC

Universitat Jaume I

E-12071 Castellón

Spain

Affil./Addr. 3: Konputazio Zientziak eta A.A. saila

Informatika Fakultatea, EHU/UPV

Donostia/San Sebastián

Spain

Splitting methods
Synonyms

Operator-splitting methods, fractional step methods

Introduction

Splitting methods constitute a general class of numerical integration schemes for dif-

ferential equations whose vector field can be decomposed in such a way that each

sub-problem is simpler to integrate than the original system. For ordinary differen-

tial equations (ODEs) this idea can be formulated as follows. Given the initial value

problem
 2

x′ = f (x), x0 = x(0) ∈ RD (1)

with f : RD −→ RD and solution ϕt (x0 ), assume that f can be expressed as f =

Pm
i=1 f [i] for certain functions f [i] , such that the equations

x′ = f [i] (x), x0 = x(0) ∈ RD , i = 1, . . . , m (2)

[i]
can be integrated exactly, with solutions x(h) = ϕh (x0 ) at t = h, the time step. The

different parts of f may correspond to physically different contributions. Then, by

combining these solutions as

[m] [2] [1]

χh = ϕh ◦ · · · ◦ ϕh ◦ ϕh (3)

and expanding into series in powers of h, one finds that χh (x0 ) = ϕh (x0 ) + O(h2 ),

so that χh provides a first-order approximation to the exact solution. Higher order

approximations can be achieved by introducing more flows with additional coefficients,

[i]
ϕaij h , in composition (3).

Splitting methods involve three steps: (i) choosing the set of functions f [i] such

f [i] ; (ii) solving either exactly or approximately each equation x′ = f [i] (x);
P
that f = i

and (iii) combining these solutions to construct an approximation for (1) up to the

desired order.

The splitting idea can also be applied to partial differential equations (PDEs)

involving time and one or more space dimensions. Thus, if the spatial differential op-

erator contains parts of a different character (such as advection and diffusion), then

different discretization techniques may be applied to each part, as well as for the time

integration.

Splitting methods have a long history and have been applied (sometimes with

different names) in many different fields, ranging from parabolic and reaction-diffusion

PDEs to quantum statistical mechanics, chemical physics and Hamiltonian dynamical

systems [6].
 3

Some of the advantages of splitting methods are the following: they are simple

to implement, are explicit if each sub-problem is solved with an explicit method, and

often preserve qualitative properties the differential equation might possess.

Splitting methods for ODEs

Increasing the order

Very often in applications the function f in the ODE (1) can be split in just two parts,
[b] [a] [a] [b]
f (x) = f [a] (x)+f [b] (x). Then both χh = ϕh ◦ϕh and its adjoint, χ∗h ≡ χ−1
−h = ϕh ◦ϕh ,

are first order integration schemes. These formulae are often called the Lie–Trotter

splitting. On the other hand, the symmetric version

[2] [a] [b] [a]

Sh ≡ ϕh/2 ◦ ϕh ◦ ϕh/2 (4)

provides a second order integrator, known as the Strang–Marchuk splitting, the leapfrog

or the Störmer–Verlet method, depending on the context where it is used [2]. Notice
[2]
that Sh = χ∗h/2 ◦ χh/2 .

More generally, one may consider a composition of the form

[a] [b] [a] [a] [b] [a]

ψh = ϕas+1 h ◦ ϕbs h ◦ ϕas h ◦ · · · ◦ ϕa2 h ◦ ϕb1 h ◦ ϕa1 h , (5)

and try to increase the order of approximation by suitably determining the parameters
[b] [a]
ai , bi . The number s of ϕh (or ϕh ) evaluations in (5) is usually referred to as the

number of stages of the integrator. This is called time-symmetric if ψh = ψh∗ , in which

case one has a left-right palindromic composition. Equivalently, in (5) one has

a1 = as+1 , b1 = bs , a2 = as , b2 = bs−1 , . . . (6)

The order conditions the parameters ai , bi have to satisfy can be obtained by relating

the previous integrator ψh with a formal series Ψh of differential operators [1]: it is

 4

known that, the h-flow ϕh of the original system x′ = f [a] (x) + f [b] (x) satisfies, for each
[a] +F [b] )
g ∈ C ∞ (RD , R), the identity g(ϕh (x)) = eh(F [g](x), where F [a] and F [b] are the

Lie derivatives corresponding to f [a] and f [b] respectively, acting as

D D
[a]
X [a] ∂g [b]
X [b] ∂g
F [g](x) = fj (x) (x), F [g](x) = fj (x) (x). (7)
j=1
∂xj j=1
∂xj

Similarly, the approximation ψh (x) ≈ ϕh (x) given by the splitting method (5) satisfies

the identity g(ψh (x)) = Ψ (h)[g](x), where

[a] [b] [a] [b] [a]

Ψ (h) = ea1 hF eb1 hF · · · eas hF ebs hF eas+1 hF . (8)

Hence, the coefficients ai , bi must be chosen in such a way that the operator Ψ (h) is a
[a] +F [b] )
good approximation of eh(F , or equivalently, h−1 log(Ψ ) ≈ F [a] + F [b] .

Applying repeatedly the Baker–Campbell–Hausdorff (BCH) formula [2] one ar-

rives at

1
log(Ψ (h)) = (va F [a] + vb F [b] ) + hvab F [ab] + h2 (vabb F [abb] + vaba F [aba] )
h
+h3 (vabbb F [abbb] + vabba F [abba] + vabaa F [abaa] ) + O(h4 ), (9)

where

F [ab] =[F [a] , F [b] ], F [abb] = [F [ab] , F [b] ], F [aba] = [F [ab] , F [a] ],

F [abbb] =[F [abb] , F [b] ], F [abba] = [F [abb] , F [a] ], F [abaa] = [F [aba] , F [a] ],

the symbol [·, ·] stands for the Lie bracket, and va , vb , vab , vabb , vaba , vabbb , . . . are poly-

nomials in the parameters ai , bi of the splitting scheme (5). In particular, one gets
Ps+1 Ps 1
Ps Pi
va = i=1 ai , vb = i=1 bi , vab = 2
− i=1 bi j=1 aj . The order conditions then

read va = vb = 1 and vab = vabb = vaba = · · · = 0 up to the order considered. To

Pr
achieve order r = 1, 2, 3, . . . , 10 the number of conditions to be fulfilled is j=1 nj ,

where nj = 2, 1, 2, 3, 6, 9, 18, 30, 56, 99. This number is smaller for r > 3 when dealing

with second order ODEs of the form y ′′ = g(y) when they are rewritten as (1) [1].
 5

For time-symmetric methods, the order conditions at even orders are automat-

ically satisfied, which leads to n1 + n3 + · · · + n2k−1 order conditions to achieve order

r = 2k. For instance, n1 +n3 = 4 conditions need to be fulfilled for a symmetric method

(5)-(6) to be of order four.

Splitting and composition methods

When the original system (1) is split in m > 2 parts, higher order schemes can be

obtained by considering a composition of the basic first order splitting method (3) and
[1] [m−1] [m]
its adjoint χ∗h = ϕh ◦ · · · ◦ ϕh ◦ ϕh . More specifically, compositions of the general

form

ψh = χ∗α2s h ◦ χα2s−1 h ◦ · · · ◦ χ∗α2 h ◦ χα1 h , (10)

can be considered with appropriately chosen coefficients (α1 , . . . , α2s ) ∈ R2s so as to

achieve a prescribed order of approximation.

In the particular case when system (1) is split in m = 2 parts, so that χh =

[b] [a]
ϕh ◦ ϕh , method (10) reduces to (5) with a1 = α1 and

bj = α2j−1 + α2j , aj+1 = α2j + α2j+1 , for j = 1, . . . , s, (11)

where α2s+1 = 0. In that case, the coefficients ai and bi are such that
s+1
X s
X
ai = bi . (12)
i=1 i=1

Conversely, any splitting method (5) satisfying (12) can be written in the form (10)
[b] [a]
with χh = ϕh ◦ ϕh .

Moreover, compositions of the form (10) make sense for an arbitrary basic first

order integrator χh (and its adjoint χ∗h ) of the original system (1). Obviously, if the

coefficients αj of a composition method (10) are such that ψh is of order r for arbitrary

basic integrators χh of (1), then the splitting method (5) with (11) is also of order r.

Actually, as shown in [7], the integrator (5) is of order r for ODEs of the form (1) with
 6

f = f [a] + f [b] if and only if the integrator (10) (with coefficients αj obtained from (11))

is of order r for arbitrary first order integrators χh .

This close relationship allows one to establish in an elementary way a defining

feature of splitting methods (5) of order r ≥ 3: at least one ai and one bi are necessarily

negative [1]. In other words, splitting schemes of order r ≥ 3 always involve backward

fractional time steps.

Preserving properties

[i]
Assume that the individual flows ϕh share with the exact flow ϕh some defining prop-

erty which is preserved by composition. Then it is clear that any composition of the

form (5) and (10) with χh given by (3) also possesses this property. Examples of such

features are symplecticity, unitarity, volume preservation, conservation of first integrals,

etc. [6]. In this sense, splitting methods form an important class of geometric numerical

integrators [2]. Repeated application of the BCH formula can be used (see (9)) to show

that there exists a modified (formal) differential equation

x̃′ = fh (x̃) ≡ f (x̃) + hf2 (x̃) + h2 f3 (x̃) + · · · , x̃(0) = x0 , (13)

associated to any splitting method ψh such that the numerical solution xn = ψh (xn−1 )

(n = 1, 2, . . .) satisfies xn = x̃(nh) for the exact solution x̃(t) of (13). An important

observation is that the vector fields fk in (13) belong to the Lie algebra generated

by f [1] , . . . , f [m] . In the particular case of autonomous Hamiltonian systems, if f [i] are

Hamiltonian, then each fk is also Hamiltonian. Then one may study the long-time

behavior of the numerical integrator by analyzing the solutions of (13) viewed as a

small perturbation of the original system (1) and obtain rigorous statements with

techniques of backward error analysis [2].

 7

Further extensions

Several extensions can be considered to reduce the number of stages necessary to

achieve a given order and get more efficient methods. One of them is the use of a

processor or corrector. The idea consists in enhancing an integrator ψh (the kernel)

with a map πh (the processor) as ψ̂h = πh ◦ ψh ◦ πh−1 . Then, after n steps one has

ψ̂hn = πh ◦ ψhn ◦ πh−1 , and so only the cost of ψh is relevant. The simplest example

of a processed integrator is provided by the Störmer–Verlet method (4). In that case

[b] [a] [a]
ψh = χh = ϕh ◦ ϕh and πh = ϕh/2 . The use of processing allows one to get methods

with fewer stages in the kernel and smaller error terms than standard compositions [1].

The second extension uses the flows corresponding to other vector fields in

addition to F [a] and F [b] . For instance, one could consider methods (5) such that,
[a] [b] [abb]
in addition to ϕh and ϕh , use the h-flow ϕh of the vector field F [abb] when its

computation is straightforward. This happens, for instance, for second-order ODEs

y ′′ = g(y) [1; 6].

Splitting is particularly appropriate when kf [a] k ≪ kf [b] k in (1). Introducing a

small parameter ε, we can write x′ = εf [a] (x) + f [b] (x), so that the error of scheme (5)

is O(ε). Moreover, since in many practical applications ε < h, one is mainly interested

in eliminating error terms with small powers of ε instead of satisfying all the order

conditions. In this way it is possible to get more efficient schemes. In addition, the use

of a processor allows one to eliminate the errors of order εhk for all 1 < k < n and all

n [6].

Although only autonomous differential equations have been considered here,

several strategies exist for adapting splitting methods also to non-autonomous systems

without deteriorating their overall efficiency [1].

 8

Some good 4th-order splitting methods

In the next table we collect the coefficients of a few selected fourth-order symmetric

methods of the form (5)-(6). Higher order and more elaborated schemes can be found in

[1; 2; 6] and references therein. They are denoted as Xs 4, where s indicates the number

of stages. S6 4 is a general splitting method, whereas SN6 4 refers to a method tailored

for second-order ODEs of the form y ′′ = g(y) when they are rewritten as a first order

system (1) and the coefficients ai are associated to g(y). Finally, SNI5 4 is a method

especially designed for problems of the form x′ = εf [a] (x) + f [b] (x). With s = 3 stages,

there is only one solution, S3 4, given by a1 = b1 /2, b1 = 2/(2 − 21/3 ). In all cases, the
P P
remaining coefficients are fixed by symmetry and consistency ( i ai = i bi = 1).

S6 4 a1 =0.07920369643119565 b1 = 0.209515106613362

a2 = 0.353172906049774 b2 = −0.143851773179818

a3 = −0.04206508035771952

SN6 4 a1 =0.08298440641740515 b1 = 0.245298957184271

a2 = 0.396309801498368 b2 = 0.604872665711080

a3 = −0.3905630492234859

SNI5 4 a1 =0.81186273854451628884 b1 = −0.0075869131187744738

a2 = −0.67748039953216912289 b2 = 0.31721827797316981388

Numerical example: a perturbed Kepler problem

To illustrate the performance of the previous splitting methods, we apply them to the

time integration of the perturbed Kepler problem described by the Hamiltonian

1 1 ε
H = (p21 + p22 ) − − 5 q22 − 2q12 ,


(14)
2 r 2r
p
where r = q12 + q22 . We take ε = 0.001 and integrate the equations of motion

qi′ = pi , p′i = −∂H/∂qi , i = 1, 2, with initial conditions q1 = 4/5, q2 = p1 = 0,

 9

0 S2
RK4
S34
−1 S64
SN64
−2
SNI54
LOG10(ERROR)
−3

−4

−5

−6

−7

−8
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6
LOG10(N. EVALUATIONS)

Fig. 1. Error in the solution (q1 (tf ), q2 (tf )) vs. the number of evaluations for different 4th-order

splitting methods (the extra cost in the method SNI5 4, designed for perturbed problems, is not taken

into account).

p
p1 = 3/2. Splitting methods are used with the partition into kinetic and poten-

tial energy. We measure the 2-norm error in the position at tf = 2000, (q1 , q2 ) =

(0.318965403761932, 1.15731646810481), for different time steps and plot the corre-

sponding error as a function of the number of evaluations for each method in Figure 1.

Notice that although the generic method S6 4 has three more stages than the minimum

given by S3 4, this extra cost is greatly compensated by a higher accuracy. On the

other hand, since this system corresponds to the second order ODE q ′′ = g(q), method

SN6 4 leads to a higher accuracy with the same computational cost. Finally, SNI5 4 takes

profit of the near-integrable character of the Hamiltonian (14) and the two extra stages

to achieve an even higher efficiency. It requires solving the Kepler problem separately

from the perturbation. This requires a more elaborated algorithm with a slightly in-

crease in the computational cost (not reflected in the figure). Results provided by the

leap-frog method S2 and the standard fourth-order Runge-Kutta integrfator RK4 are

also included for reference.

 10

Splitting methods for PDEs

In the numerical treatment of evolutionary PDEs of parabolic or mixed hyperbolic-

parabolic type, splitting time-integration methods are also widely used. In this setting

the overall evolution operator is formally written as a sum of evolution operators,

typically representing different aspects of a given model. Consider an evolutionary

PDE formulated as an abstract Cauchy problem in a certain function space U ⊂ {u :

RD × R → R},

ut = L(u), u(t0 ) = u0 , (15)

where L is a spatial partial differential operator. For instance,

d d
!
∂ X ∂ X ∂
u(x, t) = ci (x) u(x, t) + f (x, u(x, t)), u(x, t0 ) = u0 (x)
∂t j=1
∂xj i=1 ∂xi

or in short, L(x, u) = ∇ · (c∇u) + f (u), corresponds to a diffusion-reaction problem.

In that case it makes sense to split the problem into two sub-equations, corresponding

to the different physical contributions,

ut = La (u) ≡ ∇ · (c∇u), ut = Lb (u) ≡ f (u), (16)

[a] [b]
solve numerically each equation in (16), thus giving u[a] (h) = ϕh (u0 ), u[b] (h) = ϕh (u0 ),
[a] [b]
respectively for a time step h, and then compose the operators ϕh , ϕh to construct an
[b] [a]
approximation to the solution of (15). Thus, u(h) ≈ ϕh (ϕh (u0 )) provides a first-order
[a] [b] [a]
approximation, whereas the Strang splitting u(h) ≈ ϕh/2 (ϕh (ϕh/2 (u0 ))) is formally

second-order accurate for sufficiently smooth solutions. In this way, especially adapted

numerical methods can be used to integrate each sub-problem, even in parallel [3; 4].

Systems of hyperbolic conservation laws, such as

ut + f (u)x + g(u)x = 0, u(x, t0 ) = u0 (x)

can also be treated with splitting methods, in this case by fixing a step size h and apply-

ing a especially tailored numerical scheme to each scalar conservation law ut +f (u)x = 0
 11

and ut +g(u)x = 0. This is a particular example of dimensional splitting where the orig-

inal problem is approximated by solving one space direction at a time. Early examples

of dimensional splitting are the so-called locally one-dimensional (LOD) methods (such

as LOD-backward Euler and LOD Crank–Nicolson schemes) and alternating direction

implicit (ADI) methods (e.g., the Peaceman–Rachford algorithm) [4].

Although the formal analysis of splitting methods in this setting can also be

carried out by power series expansions, several fundamental difficulties arise, however.

First, nonlinear PDEs in general possess solutions that exhibit complex behavior in

small regions of space and time, such as sharp transitions and discontinuities. Second,

even if the exact solution of the original problem is smooth, it might happen that

the composition defining the splitting method provides non-smooth approximations.

Therefore, it is necessary to develop sophisticated tools to analyze whether the numer-

ical solution constructed with a splitting method leads to the correct solution of the

original problem or not [3].

On the other hand, even if the solution is sufficiently smooth, applying splitting

methods of order higher than two is not possible for certain problems. This happens,

in particular, when there is a diffusion term in the equation, since then the presence

of negative coefficients in the method leads to an ill-posed problem. When c = 1 in

(16), this order-barrier has been circumvented, however, with the use of complex-valued
[a]
coefficients with positive real parts: the operator ϕzh corresponding to the Laplacian La

is still well defined in a reasonable distribution set for z ∈ C, provided that ℜ(z) ≥ 0.

There exist also relevant problems where high order splitting methods can be

safely used as is in the integration of the time dependent Schrödinger equation iut =
1
− 2m ∆u + V (x)u split into kinetic T = −(2m)−1 ∆ and potential V energy operators

and with periodic boundary conditions. In this case, the combination of the Strang

splitting in time and the Fourier collocation in space is quite popular in chemical
 12

physics (with the name of split-step Fourier method). These schemes have appealing

structure-preserving properties, such as unitarity, symplecticity and time-symmetry [5].

Moreover, it has been shown that for a method (5) of order r with the splitting into

kinetic and potential energy and under relatively mild assumptions on T and V , one

has a rth-order error bound kψhn u0 − u(nh)k ≤ Cnhr+1 max0≤s≤nh ku(s)kr in terms of

the rth-order Sobolev norm [5].

Cross-references

Composition methods; Symplectic methods; One-step methods, order, convergence;

Stoermer-Verlet; Symmetric methods;

References

1. Blanes S, Casas F, and Murua A (2008) Splitting and composition methods in the numerical

integration of differential equations. Bol. Soc. Esp. Mat. Apl. 45: 89-145

2. Hairer E, Lubich C, and Wanner G (2006) Geometric Numerical Integration. Structure-Preserving

Algorithms for Ordinary Differential Equations, 2nd Ed, Springer (Berlin)

3. Holden H, Karlsen K H, Lie K A, and Risebro N H (2010) Splitting Methods for Partial Differential

Equations with Rough Solutions, European Mathematical Society (Zürich)

4. Hundsdorfer W, Verwer J G (2003) Numerical Solution of Time-Dependent Advection-Diffusion-

Reaction Equations, Springer (Berlin)

5. Lubich C (2008) From Quantum to Classical Molecular Dynamics: Reduced Models and Numerical

Analysis, European Mathematical Society (Zürich)

6. McLachlan R I, Quispel R (2002) Splitting methods, Acta Numer. 11: 341–434

7. McLachlan R I (1995), On the numerical integration of ordinary differential equations by sym-

metric composition methods, SIAM J. Numer. Anal., 16: 151–168