Highlights

An Implicit Second Order Time Integration Scheme with Increased

Accuracy for Applications in Fluid Dynamics
Wulf G. Dettmer, Eman Alhayki

• New versions of the generalised-α method are presented that maintain

the user-controlled high frequency damping but offer improved second
order accuracy. The computer implementation is trivial and the addi-
tional computational cost is negligible.

• New versions of the second order accurate backward difference formula

are presented that offer improved second order accuracy. The com-
puter implementation is trivial and the additional computational cost
is negligible.

• The concept of weighted linear combinations of second, third and fourth

order methods is introduced and exploited for the formulation of new
time integration schemes that maintain unconditional stability but offer
improved accuracy.
 An Implicit Second Order Time Integration Scheme
with Increased Accuracy for Applications in Fluid
Dynamics
Wulf G. Dettmer∗, Eman Alhayki
a
Faculty of Science And Engineering, Swansea University, Bay Campus, Fabian
Way, Swansea, SA1 8EN, Wales, United Kingdom

Abstract
The popular generalised-α method Janssen et al. (2000) is extended to a
family of higher order schemes. It is shown that linear combinations of the
second, third and fourth order schemes result in new unconditionally sta-
ble time integration schemes that offer improved accuracy and less numerical
damping in the low frequency regime than the original generalised-α method,
while maintaining the mechanism for user controlled high frequency damp-
ing via the parameter ρ∞ . For ρ∞ = 0, the new schemes coincide with new
backward difference formulae, similarly to the correspondence between the
generalised-α method and the standard backward difference formula of sec-
ond order. Two numerical examples including the vortex shedding behind a
cylinder immersed in incompressible fluid flow illustrate the performance of
the proposed time integration schemes.
Keywords: implicit numerical time integration, generalised-α method,
backward difference formulae, unconditional stability, accuracy

1. Introduction
Implicit unconditionally stable time integration schemes are a crucial in-
gredient to the computer simulation of any dynamic physical process unless
an explicit strategy is employed. In particular, they are extensively used in

∗
corresponding author
Email addresses: [email protected] (Wulf G. Dettmer),
[email protected] (Eman Alhayki)

Preprint submitted to Computer Methods in Applied Mechanics and EngineeringMay 26, 2023
computational fluid and solid dynamics. This article is focused on initial
boundary value problems of first order and therefore directly applicable to
fluid dynamics.
A variety of implicit time integration schemes have been proposed for
computational fluid dynamics. The most prominent strategies include the
backward Euler method, the Crank Nicolson or trapezoidal rule, the back-
ward difference formulae, see for instance Hughes (1987); Quarteroni et al.
(2000), and, more recently, the generalised-α method Janssen et al. (2000).
Presuming unconditional stability, which is a fundamental requirement for
any viable implicit time integration scheme, the criteria for the assessment
of the performance of a particular scheme are, firstly, accuracy and, sec-
ondly, the numerical damping characteristics. The methods are classified
with respect to their order of accuracy. Due to the Dahlquist theorem, see
for instance Quarteroni et al. (2000), they can be only of first or second order
accuracy. The numerical damping characteristics can be studied in terms of
the variation of the spectral radius with the time step size. The analysis of
accuracy and numerical damping is generally based on the standard linear
scalar model problem and the associated amplification matrices.
An optimal implicit time integration scheme features second order accu-
racy with small error values and a noticeable or even substantial amount of
numerical damping in the high frequency regime. The latter is crucial for
industrially relevant applications since it effectively removes spatially unre-
solved small wave length responses that can cause problems ranging from
noise to failure of the simulation.
While the backward Euler method is known for its excessive numeri-
cal damping and poor accuracy, the Crank Nicolson and trapezoidal rule
schemes are more accurate but do not possess any numerical damping. The
generalised-α method represents an excellent alternative as it offers user con-
trolled high frequency damping combined with second order accuracy. There-
fore, it has been successfully used in many application areas, including most
prominently CFD Janssen et al. (2000); Dettmer and Perić (2003); Kadapa
et al. (2020); Bazilevs et al. (2011); Kees et al. (2011) and fluid-structure
interaction Dettmer and Perić (2006); Joosten et al. (2010); Dettmer and
Perić (2013); Kadapa et al. (2018); Bazilevs et al. (2013); Mao and Jaiman
(2023), but also in phase field modelling Gómez et al. (2008); Ashour et al.
(2023), biomechanics Kim et al. (2023), general solid mechanics Kadapa et al.
(2017), thermomechanics Vila-Cha et al. (2023) and others.
In this work, the generalised-α method from Janssen et al. (2000) is ex-

2
tended and reformulated to include an order of accuracy parameter p ≥ 2.
Due to the natural limitation to second order accuracy, the extended family
of schemes is unconditionally stable only for p = 2, where it coincides with
the standard generalise-α method. However, it is shown that linear combina-
tions of the schemes with p = 2, 3, 4 are also unconditionally stable, provided
the weighting of the higher order schemes does not exceed a critical value.
The resulting schemes are assessed in terms of their accuracy and numeri-
cal damping properties. It is shown that, for ρ∞ < 1, they are superior to
the standard generalised-α method. Furthermore, new backward difference
formulae are proposed that are unconditionally stable, more accurate than
the standard formula and equivalent to the proposed extended generalised-α
method, if ρ∞ = 0. The proposed methods allow for improved accuracy and
reduced low frequency damping at negligible addtional computational cost.
Their implementation in existing code is straightforward and only requires
one or two more current state or history arrays.
The remainder of this article is organised as follows: In Section 2, the
extended generalised-α method is introduced. Unconditionally stable linear
combinations of the second, third and fourth order accurate schemes are pro-
posed in Section 3. In Section 4, it is shown that the new schemes give rise
to new backward difference formulae with higher accuracy than the stan-
dard second order formula. In Section 5, the linear scalar oscillator and,
in Section 6, a finite element model of incompressible fluid flow around a
stationary cylinder are used to illustrate the superior performance of the
proposed schemes. Conclusions are summarised in Section 7.

2. A new family of time integration schemes

Consider the scalar linear first order initial value problem

u̇(t) − λ u(t) = 0 with u(0) = u0 , (1)

where λ = −ξ + i ω and u̇ = du/dt. The exact solution can be expressed as

u(t) = u0 eλ t . (2)

Introducing the sequence of discrete time instants tn with n = 0, 1, 2, ... and

the step size ∆t = tn+1 − tn gives rise to the exact amplification factor

ζexact = eλ ∆t with un+1 = ζexact un . (3)

3
The set of equations
u̇n+β − λ un+α = 0 (4)
un+α = α un+1 + (1 − α) un (5)
p−1
0
X
u̇n+β = β0 u̇n+1 + βi u(i)
n ∆t
i−1
(6)
i=1
 
(i)0 0 (i+1)0 0
un+1 = u(i)
n + ∆t γ un+1 + (1 − γ) un(i+1) for i = 0, 1, 2, ..., p − 2 (7)
with p ≥ 2 represents a family of implicit numerical time integration schemes.
(i)0
The term un denotes the i-th derivative of the variable u at the time instant
tn . The associated amplification matrix A of dimensions p × p with
   
 un+1   un 

 
 
 

u̇ ∆t u̇ ∆t
   
n+1 n

 
 
 


 
 
 

 (2) 0   (2) 0 
 un+1 ∆t
 2 
  un ∆t
 2 

(3)0
= A (3)0
(8)
3 3


 un+1 ∆t 




 u n ∆t 



 
 
 



 ... 




 ... 


   
 (p−1)0

p−1

 
 (p−1) 0

p−1 
un+1 ∆t un ∆t

is easily obtained from symbolic mathematical software and given in Ap-

pendix Appendix A. The spectral radius is ρ(A) = max(|ζi |), where ζi with
i = 1, 2, ..., p are the eigenvalues of A. Introducing the spectral radius ρ∞ for
an infinte time step, it is easily shown that
 
1−α 1−γ
ρ∞ = lim ρ(A) = max − , − . (9)
∆t→∞ α γ
Setting
1
α = γ = (10)
1 + ρ∞
ensures that the absolute values of all eigenvalues converge to ρ∞ as ∆t tend
to infinity. By keeping ρ∞ as a free parameter user-controlled high-frequency
damping is enabled.
Next, the amplification factor ζexact is inserted into the characteristic poly-
nomial of A and the resulting expression expanded as a series of powers of
∆t. Equating the first p terms to zero renders a set of equations which can

4
be solved for the coefficients βi with i = 0, 1, 2, ..., p − 1. The results are
summarised in Table 1.
Hence, Equations (4) to (7) represent a family of implicit numerical time
integration schemes of order p ≥ 2 with one user-controlled parameter ρ∞ .
In anticipation of Remark 2.2 below, it is termed generalised-α method of
order p, or briefly, GA-p. The variation of the spectral radius ρ(A) with the
time step size ∆t is visualised in Figure 1.

Remark 2.1: As presented by Equations (4) to (7), the scheme is not

suitable for p = 1. An unconditionally stable first order scheme, namely
the generalised midpoint rule GM, is obtained by setting p = 1 and γ = 1,
see for instance Dettmer and Perić (2003). In this case, the amplification
matrix reduces to a scalar amplification factor and, with ρ∞ = 0 and ρ∞ =
1, respectively, the backward Euler and the trapezoidal rule methods are
recovered. For convenience the generalised midpoint method is summarised
in Box 1.

Remark 2.2: For p = 2, the scheme coincides with the popular generalised-
α method proposed by Janssen et al. Janssen et al. (2000). If α, β0 and β1
are replaced, respectively, by αf , αm and 1 − αm , the notation from Janssen
et al. (2000) is recovered. The scheme GA-2 is summarised in Box 2.

Remark 2.3: For ρ∞ = 0, the proposed family of schemes is identical to

the well-known backward difference formulae of order p. This is further
elaborated in Section 4.

Remark 2.4: It is observed in Figure 1 that the scheme is unconditionally

stable only for p = 2. Clearly it is subject to the Dahlquist Theorem, see for

p β0 β1 β2 β3
−ρ∞ +3 3ρ∞ −1
2 2(ρ∞ +1) 2(ρ∞ +1)

2ρ2∞ −5ρ∞ +11 −2ρ2∞ +11ρ∞ −5 −ρ2∞ +ρ∞ −1

3 6(ρ∞ +1) 6(ρ∞ +1) 3(ρ∞ +1)

−3ρ3∞ +7ρ2∞ −13ρ∞ +25 3ρ3∞ −7ρ2∞ +25ρ∞ −13 3ρ3∞ −7ρ2∞ +7ρ∞ −7 ρ3∞ −ρ2∞ +ρ∞ −1
4 12(ρ∞ +1) 12(ρ∞ +1) 12(ρ∞ +1) 4(ρ∞ +1)2

Table 1: Coefficients βi for the family of time integration schemes GA-p.

5
 1
ρ∞ = 0 ρ∞ = 3

2
ρ∞ = 3
ρ∞ = 1

Figure 1: Spectral radii of GM, GA-2, GA-3 and GA-4 for ξ = 0.

instance Quarteroni et al. (2000).

3. Ensuring Unconditional Stability

The method GA-p described in Section 2 is unconditionally stable only for
p = 2. Yet, in order to improve the accuracy while maintaining unconditional
stability the following variation of Equation (4) is considered
(3) (2)
δ3 u̇n+β + (1 − δ3 ) u̇n+β − λ un+α = 0 , (11)
(j)
where u̇n+β is defined by Equation (6) with p = j. The coefficients βi in
Equation (6) are given in Table 1 and also need to be selected for p = j.

6
 u̇n+1 − λ un+α = 0
un+α = α un+1 + (1 − α) un
un+1 = un + ∆t u̇n+1
1
α =
1 + ρ∞

Box 1: Summary of the generalised midpoint rule method GM.

u̇n+β − λ un+α = 0
un+α = α un+1 + (1 − α) un
u̇n+β = β u̇n+1 + (1 − β) u̇n


un+1 = un + ∆t γ u̇n+1 + (1 − γ) u̇n
1 3 − ρ∞
α = γ = , β =
1 + ρ∞ 2(1 + ρ∞ )

Box 2: Summary of method GA-2 (equivalent to the generalised-α method).

Equation (11) represents the linear combination of the GA-2 and GA-3
schemes: For δ3 = 0 the scheme reduces to GA-2 and is unconditionally
stable, while for δ3 = 1 it recovers GA-3 and is more accurate but only
conditionally stable. Clearly, it is desirable to find the largest possible value
for δ3 ≥ 0 that ensures unconditional stability.
The Jury Table, see for instance Raible (1974), offers a suitable tool
for this investigation. However, the analysis is tedious. It results in long
expressions and is most easily done using symobolic mathematical software.
Some details are described in Appendix Appendix B. The critical value of
δ3 is obtained as a function of ρ∞ , i. e.

(1 − ρ∞ )2
δ3 = . (12)
2 (1 − ρ∞ + ρ2∞ )

Clearly, ρ∞ = 1 renders δ3 = 0 and the contribution of GA-3 vanishes.

The scheme then coincides with the trapezoidal rule. Yet, for ρ∞ = 0, one
obtains δ3 = 21 which corresponds to an equal weighting of the second and

7
third order accurate terms. As ρ∞ is increased and approaches the value of
1, the weighting of GA-3 decreases. In them remainder of this work, the time
integration scheme described by Equations (11), (5), (6), (7), (10) and (12),
together with the relevant expressions from Table 1, is referred to as GA-23.
Linear combinations of further higher order terms can also be investi-
gated. For that purpose, consider
 
(4) (3) (2)
δ4 u̇n+β + (1 − δ4 ) δ3 u̇n+β + (1 − δ3 ) u̇n+β − λ un+α = 0 , (13)

which is the linear combination of GA-23 and GA-4. Again, the Jury Table
is used to find the critical value for δ4 . The investigation renders
(1 − ρ∞ )2
δ4 = . (14)
5 (1 + ρ2∞ )
For ρ∞ = 1, one obtains δ4 = δ3 = 0 and the scheme reduces to the trape-
zoidal rule. For ρ∞ = 0, the weighting factors for GA-2, GA-3 and GA-4 are,
respectively, 52 , 52 and 15 . Hence, the contribution of GA-4 is significant, but
smaller than that of GA-3. The scheme based on Equation (13) is referred
to as GA-234.
Figure 2 shows the spectral radii for GM, GA-2, GA-23 and GA-234 as
functions of the time step size. The eigenvalues of the amplification matrices
are shown in detail Figure 3. The full sets of equations describing GA-23
and GA-234 are summarised in Boxes 3 and 4.

Remark 3.1: It is likely that more schemes, based on the inclusion of fur-
ther terms of higher order accuracy, can be formulated. Yet, it is expected
that the associated weighting factors are very small. Furthermore, the appli-
cation of the Jury Table becomes more tedious and eventually impractical.
Therefore, this investigation is ommitted.

Remark 3.2: The diagrams in Figure 2 confirm that GA-23 and GA-234
are unconditionally stable. Considering the range of methods, from GM to
GA-234, the onset of numerical damping is clearly shifted to the regime of
larger time steps as more higher order terms are included.

Remark 3.3: It can be observed in Figure 2 that the shift of the numerical
damping from GA-2 to GA-23 is larger than from GA-23 to GA-234. This
is consistent with the relatively small weighting of GA-4 in GA-234 and sup-
ports the decision made in Remark 4.1 not to investigate linear combinations

8
 1
ρ∞ = 0 ρ∞ = 3

2
ρ∞ = 3
ρ∞ = 1

Figure 2: Spectral radii of GM, GA-2, GA-23 and GA-234 for ξ = 0.

involving further higher order contributions.

Remark 3.4: It is also observed in Figure 2 that the shift of the numerical
damping from GA-2 to GA-23 or to GA-234 is similarly significant as the
one from GM to GA-2, i. e. the contribution of the terms of higher order
accuracy significantly changes the scheme, even though, overall, the scheme
retains second order accuracy.

4. Backward difference formulae

For ρ∞ = 0, the time integration schemes GA-p described by Equations
(4) to (7) are equivalent to the well-known backward difference formulae
BDF-p, see for instance Quarteroni et al. (2000). In order to show this,

9
 GM GA-2 GA-23 GA-234
ρ∞ = 0
1
3
ρ∞ =
2
3
ρ∞ =
ρ∞ = 1

Figure 3: Eigenvalues of GM, GA-2, GA-23 and GA-234 for ξ = 0, displayed in the
complex plain for a range of time step sizes ∆t.

Equation (7) is rewritten as

(i)0 (i)0
(i+1)0 u − un
un+1 = n+1 , (15)
∆t
where γ = 1 due to ρ∞ = 0. This formula can be applied recursively to each
term on its right-hand-side, thereby reducing the order of the derivatives and
involving terms from earlier time instants. Eventually all derivatives on the

10
 u̇n+β − λ un+α = 0
un+α = α un+1 + (1 − α) un
u̇n+β = β0 u̇n+1 + β1 u̇n + β2 ün ∆t
(i)0 0 (i+1)0 0

un+1 = u(i)
n + ∆t γ un+1 + (1 − γ) un(i+1) for i = 0, 1
1 10 − 5ρ∞ + ρ2∞
α=γ= , β0 =
1 + ρ∞ 6(1 + ρ∞ )
(1 − ρ∞ )2
β1 = 1 − β0 , β2 = −
6(1 + ρ∞ )

Box 3: Summary of method GA-23. The coefficients βi have been redefined for brevity of
presentation.

p β0 β1 β2 β3 β4

1 1 −1
3 1
2 2
−2 2
11 3
3 6
−3 2
− 13
25
4 12
−4 3 − 43 1
4

Table 2: Coefficients βi for backward difference formulae BDF-p.

right-hand-side of Equation (6) can be expressed in terms of historic values

of the primary variable u and the backward difference formula is obtained.
The BDF-p schemes can be expressed as

u̇n+1 − λ un+1 = 0 (16)

p
1 X
u̇n+1 = βi un+1−i , (17)
∆t i=0
where the coefficients βi are given in Table 2. The same procedure can be
applied to the methods GA-23 and GA-234, resulting in linear combinations
of BDF-2, BDF-3 and BDF-4. The new methods are denoted as BDF-23
and BDF-234 and are summarised in Boxes 5 and 6, respectively. For lin-
ear problems and given the correct initial conditions, they render numerical

11
 u̇n+β − λ un+α = 0
un+α = α un+1 + (1 − α) un
...
u̇n+β = β0 u̇n+1 + β1 u̇n + β2 ün ∆t + β3 u n ∆t2
(i)0 0 (i+1)0 0

un+1 = u(i)
n + ∆t γ un+1 + (1 − γ) un(i+1) for i = 0, 1, 2
1 35 − 21ρ∞ + 7ρ2∞ − ρ3∞
α=γ= , β0 =
1 + ρ∞ 20(1 + ρ∞ )
(1 − ρ∞ )2 (5 − ρ∞ ) (1 − ρ∞ )3
β1 = 1 − β0 , β2 = − , β3 = −
20(1 + ρ∞ ) 20(1 + ρ∞ )2

Box 4: Summary of method GA-234. The coefficients βi have been redefined for brevity
of presentation.

u̇n+1 − λ un+1 = 0
10 un+1 − 15 un + 6 un−1 − un−2
u̇n+1 =
6 ∆t

Box 5: Summary of method BDF-23.

results that are exactly equal to those obtained from GA-23 and GA-234
when ρ∞ = 0. Therefore, BDF-23 and BDF-234 are not considered directly
in the numerical examples in Sections 5 and 6, but are represented by their
counterparts GA-23 and GA-234 with ρ∞ = 0.
As opposed to GA-23 and GA-234, the schemes BDF-23 and BDF-234
do not allow for varying time step sizes or for user-controlled high-frequency
damping. However, their implementation in any existing computer code
based on BDF-2 is trivial and renders an immediate benefit in terms of
increased accuracy.

5. Example 1: Linear Model Problem

The standard scalar linear initial value problem described by Equation (1)
is used to compare the performance and accuracy of the methods presented
in Section 2. The exact solution is given in Equation (2). The frequency is
set to ω = 1. Hence, the oscillation period is T = 2π/ω = 2π. The initial

12
 u̇n+1 − λ un+1 = 0
35 un+1 − 56 un + 28 un−1 − 8 un−2 + un−3
u̇n+1 =
20 ∆t

Box 6: Summary of method BDF-234.

condition considered is u(0) = 1 + 0i. The results obtained with GM, GA-2,
GA-23 and GA-234 from different time step sizes, different values of ρ∞ and
ξ are displayed in Figures 4 and 5. The figures show the real part of the
numerical solutions. The evolution of the imaginary part is phase shifted
but qualitatively and quantitatively similar and therefore not presented for
the sake of brevity.
Figure 6 shows the convergence of the numerical solution as ∆t decreases
for GM, GA-2, GA-23, and GA-234 for different values of ρ∞ . The error
is computed by comparing the exact solution ũ from Equation (2) to the
numerical solution ū obtained from each method, i. e.
v
u
u1 X N
ε= t |ūn − ũn |2 (18)
N n=1

where N is the number of time steps used in the respective computation to

reach tN = 35.
The following observations are made from Figures 4 and 5:
1. As ∆t is reduced, all methods converge to the exact solution.
2. The solutions deviate less for ρ∞ = 1/3 than for ρ∞ = 0. This agrees
with the fact that all schemes considered recover the trapezoidal rule
for ρ∞ = 1.
3. For large time steps, the methods GA-23 and GA-234 are noticably
more accurate than the generalised-α method GA-2. GA-23 and GA-
234 exhibit less numerical damping and a smaller frequency error. GA-
234 is marginally more accurate than GA-23.
The following observations are made from Figure 6:
4. The slope of the graphs in the diagram confirms that methods GA-2,
GA-23 and GA-234 are second-order accurate.

13
 Exact GA-2 GA-23 GA-234

T 1 T
(a) ρ∞ = 0 and ∆t = 32 (b) ρ∞ = 3 and ∆t = 32

T 1 T
(c) ρ∞ = 0 and ∆t = 8 (d) ρ∞ = 3 and ∆t = 8

Figure 4: Linear model problem, response obtained for ξ = 0.0.

5. The difference between the graphs for GA-2, GA-23 and GA-234 re-
duces as ρ∞ approaches the value of one. The graphs coincide for
ρ∞ = 1 where all schemes reduce to the trapezoidal rule.
6. The accuracy gain from GA-2 to GA-23 is more significant than that
from GA-23 to GA-234. This agrees with the small change in the
weighting of the higher order contributions from GA-23 to GA-234.
7. In Figure 6(a), for ρ∞ = 0 and ∆t ≈ T /32, the improvement in ac-
curacy from GA-2 to GA-234 compares to that from the first order
accurate method GM to the second order accurate scheme GA-2. This
is also reflected in Figures 4 and 5 where, for ρ∞ = 0, GA-234 renders
significantly more accurate results than GA-2.
Importantly, Observation 7 suggests that, in the range of time steps T /50 <
∆t < T /25 which, in the context of implicit time integration, is most rele-
vant for industrial applications, the scheme GA-234 offers a significant per-
formance gain over the standard generalised-α method GA-2. Due to the

14
 Exact GA-2 GA-23 GA-234

T 1 T
(a) ρ∞ = 0 and ∆t = 32 (b) ρ∞ = 3 and ∆t = 32

T 1 T
(c) ρ∞ = 0 and ∆t = 8 (d) ρ∞ = 3 and ∆t = 8

Figure 5: Linear model problem, response obtained for ξ = 0.1.

equivalence between BDF-2 and GA-2 and between BDF-234 and GA-234,
for ρ∞ = 0, this also suggests the superiority of BDF-234 over the standard
method BDF-2.

6. Example 2: Flow Around Cylinder

This benchmark example examines the flow of an incompressible fluid
around a stationary circular cylinder. Figure 7 shows the geometry, bound-
ary conditions and the employed finite element mesh consisting of 6,205 nodes
and 12,258 elements. The spatial discretisation is based on linear equal-order
velocity-pressure elements. The convection and incompressibility effects are
stabilised by standard SUPG/PSPG stabilisation terms as described, for in-
stance, in Dettmer and Perić (2003) and originally proposed in Tezduyar
et al. (1992).
The vertical velocity component v is set to zero at the upper and lower
boundaries, while the horizontal velocity component u remains unconstrained.

15
 GM GA-2 GA-23 GA-234

1 2
(a) ρ∞ = 0 (b) ρ∞ = 3 (c) ρ∞ = 3 (d) ρ∞ = 1

Figure 6: Linear model problem, convergence rates.

In addition, the pressure at the outlet boundary is set to zero and at the inlet
boundary u is prescribed uniformly as u∞ = 1, consistent with the premise
that the flow far from the cylinder is uniform. The cylinder diameter is
d = 1, while the fluid density and viscosity are given by ρ = 1 and µ = 0.01,
respectively. Hence, the Reynolds number is Re = ρ u∞ d/µ = 100. The non-
dimensional lift and drag coefficients CL and CD and the Strouhal number
St are defined as
2FL 2FD fd
CL = , CD = , St = (19)
ρu2∞ d ρu2∞ d u∞
where FL and FD are the lift and drag forces respectively, and f is the
frequency of the vortex shedding. The flow is simulated with the schemes
GM, GA-2, GA-23, and GA-234.
Figure 8 illustrates the evolution of the drag and lift coefficients over time.
Figure 9 shows the convergence of the Strouhal number and lift coefficient
as the time step size ∆t is reduced. Clearly, as ∆t is reduced, all methods
converge to CL = 0.3558 and St = 0.1728. These values depend on the
Reynolds number, the size of the domain and the mesh density employed.
They agree well with the results reported, for instance, in Dettmer and Perić
(2003); Kadapa et al. (2020) and references therein.

16
 v=0

u = u∞
15 d=1

p=0
15

v=0

10 20

(a)

(b) (c)

Figure 7: Flow around cylinder; geometry and boundary conditions (a); finite element
mesh (b), mesh refinement around the cylinder (c).

The convergence pattern for the Strouhal number shown in Figures 9(a)
to 9(c) is identical to that observed for the linear model problem in Section
5. In brief, the schemes GA-23 and GA-234 perform better than GA-2 and
this is most evident in Figure 9(a), i. e. for values of ρ∞ close to zero and
in the range of physically relevant time step sizes.

17
 For the lift coefficient, the convergence pattern is highly nonlinear. The
influence of the parameter ρ∞ and the performance comparison are less clear.
This is attributed to the fact that the lift coefficient is not computed from
the primary variable, namely the fluid velocity, which is subject to the time
integration scheme, but derived from the boundary traction forces.

Figure 8: Flow around cylinder; typical evolution of the lift and drag coefficients CL and
CD .

7. Conclusions
A generalisation of the well-known generalised-α method first proposed in
Janssen et al. (2000) to higher order accuracy with the same user controlled
high frequency damping mechanism has been introduced. It has been shown
that by using appropriate weighting factors, the linear combinations of the
second, third and fourth order accurate schemes render unconditionally stable
time integration schemes, namely the methods GA-23 and GA-234, that
are noticably more accurate than the original generalised-α method. The
associated additional computational cost is negligible. The good performance
of the methods GA-23 and GA-234 has been illustrated in two numerical
examples, including the finite element based simulation of incompressible
fluid flow around a cylinder.
Furthermore, it has been shown that GA-23 and GA-234 give rise to the
backward difference formulae BDF-23 and BDF-234 which are more accurate
than the standard BDF-2 method and require, respectively, only one or two
additional history terms.

18
 GM GA-2 GA-23 GA-234

1 2
(a) ρ∞ = 0 (b) ρ∞ = 3 (c) ρ∞ = 3

Figure 9: Flow around cylinder; CL amplitude and St convergence.

Appendix A. Amplification matrix

The amplification matrix A(4) for the time integration scheme GA-4 as
defined by Equation (8) for p = 4 is obtained as
 
δ + γλ∆t β0 η − β1 γ −β2 γ −β3 γ
1  λ∆t αηλ∆t − β1 −β2 −β3
A(4)

=  
δ  λ∆t/γ (αλ∆t − β0 − β1 )/γ −(δη + β2 )/γ −β3 /γ 
λ∆t/γ 2 2 2
(αλ∆t − β0 − β1 )/γ −(δ + β2 )/γ −(γδη + β3 )/γ 2

(A.1)

where δ = β0 − α γ λ∆t and η = 1 − γ. The amplification matrices A(2)

19
and A(3) for the schemes GA-2 and GA-3 are equal to the two and three
dimensional top left submatrices of A(4) , respectively.

Appendix B. Jury Test

The root distribution of the characteristic polynomial of a system with
respect to the unit circle is a critical aspect of evaluating the stability of
discrete-time systems. Several procedures in control theory, such as the
Schur-Cohn stability test, the Jury stability test, and bilinear transforma-
tion coupled with the Routh-Hurwitz stability criterion, can be directly ap-
plied to the characteristic equation to indicate the existence of any unstable
roots (Ogata, 1995).
The Jury test involves a set of conditions that the system must satisfy
in order to be stable. In this work, the modified Jury procedure described
in (Raible, 1974) is employed as it allows for easier mathematical manipula-
tion and direct evaluation of root dispersion. The Jury Table is formed as
illustrated in Table B.3 for a given characteristic polynomial D(z),

D(z) = an z n + an−1 z n−1 + · · · + a1 z + a0 . (B.1)

20
 an an−1 an−2 an−3 ··· a1 a0 ka
bn−1 bn−2 bn−3 ··· b1 b0 kb
cn−2 cn−3 ··· c1 c0 kc
.. .. .. ..
. . . .
v1 v0 kv
w0

with
a0 b0 c0 v0
ka = , kb = , kc = , · · · , kv =
an bn−1 cn−2 v1
bn−i = an−i+1 − ka ai−1 , for i = 1, 2, · · · , n
cn−i = bn−i+1 − kb bi−2 , for i = 2, 3, · · · , n
..
.
w0 = v1 − kv v0
Table B.3: Jury’s table for a real polynomial. If an > 0, then D(z) is stable only if
bn−1 > 0, cn−2 > 0, · · · w0 > 0.

The application of the procedure to a polynomial with complex coeffi-

cients requires the modification described in (Choo and Kim, 2009). Specif-
ically, let
a0 b0 v0
ka = , kb = , · · · , kv = ,
an bn−1 v1
where the bar denotes the complex conjugation. Stability requires

Re(an ) > 0, Re(bn−1 ) > 0, · · · , Re(w0 ) > 0 .

The characteristic polynomial of GA-23 obtained from the amplification ma-

trix that results from Equation (11) reads

a 2 b c
p(z) = z 3 + z + z + (B.2)
d d d

21
where

a = −6 δ3 (ρ3∞ + 1) − 9 ρ∞ (2 ∆t λ − 1) − 3 ρ2∞ (ρ∞ − 6) − 12

b = 6 δ3 (ρ3∞ + 1) − 9 ρ2∞ (2 ∆t λ + 1) + 6 ρ∞ (2 ρ2∞ − 3) + 3
c = −2 δ3 (ρ3∞ + 1) − 3 ρ3∞ (2 ∆t λ + 3) − 3 ρ∞ (2 ρ∞ − 1)
d = 3 ρ∞ (2 − ρ∞ ) + 2 δ3 (1 + ρ3∞ ) − 6 ∆t λ + 9 .

By employing symbolic mathematical software, the following four conditions

are derived from the Jury Table B.3,

(C1) : 1 > 0 (B.3)

(C2) : 1 − s2 > 0 (B.4)
(C3) : s1 > 0 (B.5)

1  (s7 − s3 s6 s9 )(s9 − s4 s5 s7 ) 
(C4) : s1 − s10 + − s3 s6 s8
s1 s2 − 1
!
 (s8 − s4 s5 s10 )(s10 − s3 s6 s8 ) 
× s9 + − s4 s5 s7 >0 (B.6)
s2 − 1

where
(s7 − s3 s6 s9 ) (s8 − s4 s5 s10 )
s1 = − s2 + 1
s2 − 1

(−6 ∆t ω ρ3∞ + s12 ) (6 ∆t ω ρ3∞ + s12 )

s2 =
(s11 − 6 ∆t ω) (s11 + 6 ∆t ω)

1 1
s3 = s4 =
(s11 − 6 ∆t ω) (s11 + 6 ∆t ω)

s5 = s12 − 6 ∆t ω ρ3∞ s6 = s12 + 6 ∆t ω ρ3∞

s7 = s3 (s13 − 18 ∆t ω ρ2∞ ) s8 = s4 (s13 + 18 ∆t ω ρ2∞ )

s9 = s4 (s14 − 18 ∆t ω ρ∞ ) s10 = s3 (s14 + 18 ∆t ω ρ∞ )

22
 s11 = 2δ3 + 6ρ∞ + 2δ3 ρ3∞ − 3ρ2∞ + 6∆tξ + 9 i

s12 = 2δ3 − 3ρ∞ + 2δ3 ρ3∞ + 6ρ2∞ + 9ρ3∞ − 6∆tρ3∞ ξ i

s13 = 3 2δ3 − 6ρ∞ + 2δ3 ρ3∞ − 3ρ2∞ + 4ρ3∞ + 6∆tρ2∞ ξ + 1 i

s14 = 3 2δ3 − 3ρ∞ + 2δ3 ρ3∞ − 6ρ2∞ + ρ3∞ − 6∆tρ∞ ξ + 4 i

By applying the constraints ∆t ≥ 0, ξ ≥ 0, ω ≥ 0 and 0 ≤ ρ∞ ≤ 1 and

solving the Inequalities (B.4) to (B.6) for δ3 , the following requirements for
stability are obtained:
3 (3 − 2ρ∞ + 3 ρ2∞ )
(C2) : δ3 ≥ − (B.7)
4 (1 − ρ∞ + ρ2∞ )
3 (1 − ρ∞ )2
(C3) : δ3 ≥ − (B.8)
2 (1 − ρ∞ + ρ2∞ )
(1 − ρ∞ )2
(C4) : δ3 ≥ (B.9)
2 (1 − ρ∞ + ρ2∞ )
The critical term arises from Condition C4. Therefore, setting
(1 − ρ∞ )2
δ3 = (B.10)
2 (1 − ρ∞ + ρ2∞ )
ensures that all requirements are met and the scheme GA-23 is uncondition-
ally stable.
The same methodology is employed to derive the expression for δ4 given
in Equation (14). However, the analysis is tedious, and the expressions are
lengthy and complex. They are not presented here for the sake of brevity.

References
Ashour, M., Valizadeh, N., Rabczuk, T., 2023. Phase-field navier–stokes
model for vesicle doublets hydrodynamics in incompressible fluid flow.
Computer Methods in Applied Mechanics and Engineering 412.
Bazilevs, Y., Calo, V., Cottrell, J., Hughes, T., Reali, A., Scovazzi, G.,
2011. Variational multiscale residual-based turbulence modeling for large
eddy simulation of incompressible flows. Computer Methods in Applied
Mechanics and Engineering 197, 173–201.

23
Bazilevs, Y., Takizawa, K., Tezduyar, T.E., 2013. Computational Fluid-
Structure Interaction. Wiley.

Choo, Y., Kim, D., 2009. Schur stability of complex polynomials. Journal of
Institute of Control, Robotics and Systems 15, 671–674.

Dettmer, W., Perić, D., 2003. An analysis of the time integration algorithms
for the finite element solutions of incompressible navier–stokes equations
based on a stabilised formulation. Computer Methods in Applied Mechan-
ics and Engineering 192, 1177–1226.

Dettmer, W., Perić, D., 2006. A computational framework for fluid–structure

interaction: finite element formulation and applications. Computer Meth-
ods in Applied Mechanics and Engineering 195, 5754–5779.

Dettmer, W., Perić, D., 2013. A new staggered scheme for fluid–structure
interaction. International Journal for Numerical Methods in Engineering
93, 1–22.

Gómez, H., Calo, V., Bazilevs, Y., Hughes, T., 2008. Isogeometric analysis
of the cahn–hilliard phase-field model. Computer Methods in Applied
Mechanics and Engineering 197, 4333–4352.

Hughes, T., 1987. The finite element method: Linear static and dynamic
finite element analysis. Prentice-Hall.

Janssen, K., Whiting, C., Hulbert, G., 2000. A generalized-α method for
integrating the filtered navier–stokes equations with a stabilized finite ele-
ment method. Computer Methods in Applied Mechanics and Engineering
190, 305–319.

Joosten, M., Dettmer, W., Perić, D., 2010. On the temporal stability and
accuracy of coupled problems with reference to fluid–structure interaction.
International Journal for Numerical Methods in Fluids 64, 1363–1378.

Kadapa, C., Dettmer, W., Perić, D., 2017. On the advantages of using
the first-order generalised-alpha scheme for structural dynamic problems.
Computers and Structures 193, 226–238.

Kadapa, C., Dettmer, W., Perić, D., 2018. A stabilised immersed frame-
work on hierarchical b-spline grids for fluid-flexible structure interaction

24
 with solid–solid contact. Computer Methods in Applied Mechanics and
Engineering 335, 472–489.

Kadapa, C., Dettmer, W., Perić, D., 2020. Accurate iteration-free mixed-
stabilised formulation for laminar incompressible navier-stokes: Applica-
tions to fluid-structure interaction. Journal of Fluids and Structures 97.

Kees, C., Akkerman, I., Farthing, M., Bazilevs, Y., 2011. A conservative level
set method suitable for variable-order approximations and unstructured
meshes. Journal of Computational Physics 230, 4536–4558.

Kim, H.J., Rundfeldt, H.C., Lee, I., Lee, S., 2023. Tissue-growth-based syn-
thetic tree generation and perfusion simulation. Biomechanics and Mod-
eling in Mechanobiology 22, 1095–1112.

Mao, X., Jaiman, R., 2023. An interface and geometry preserving phase-field
method for fully eulerian fluid-structure interaction. Journal of Computa-
tional Physics 476.

Ogata, K., 1995. Discrete-time control systems. Prentice-Hall, Inc.

Quarteroni, A., Sacco, R., Saleri, F., 2000. Numerical Mathematics. Springer.

Raible, R., 1974. A simplification of jury’s tabular form. IEEE Transactions

on Automatic Control 19, 248–250.

Tezduyar, T., Mittal, S., Ray, S., Shih, R., 1992. Incompressible flow
computations with stabilized bilinear and linear equal-order-interpolation
velocity-pressure elements. Computer Methods in Applied Mechanics and
Engineering 95, 221–242.

Vila-Cha, J., Carneiro, A.C., Ferreira, B., Pires, F.A., 2023. A numerical as-
sessment of partitioned implicit methods for thermomechanical problems.
Computers and Structures 277-278.

25