You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/232907645

Studies of Newmark method for solving nonlinear systems: (I) basic analysis

Article  in  Journal- Chinese Institute of Engineers · July 2004


DOI: 10.1080/02533839.2004.9670913

CITATIONS READS

30 1,048

1 author:

Shuenn-Yih Chang
National Taipei University of Technology
139 PUBLICATIONS   1,315 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Not available View project

All content following this page was uploaded by Shuenn-Yih Chang on 25 January 2014.

The user has requested enhancement of the downloaded file.


Journal of the Chinese Institute of Engineers, Vol. 27, No. 5, pp. 651-662 (2004) 651

STUDIES OF NEWMARK METHOD FOR SOLVING NONLINEAR


SYSTEMS: (I) BASIC ANALYSIS

Shuenn-Yih Chang

ABSTRACT
The performance of the Newmark method in the solution of linear elastic sys-
tems can be reliably predicted by the currently available evaluation techniques.
However, its performance for solving nonlinear systems is still not clear. Therefore,
its numerical characteristics in the solution of nonlinear systems are explored in this
study by using a newly developed technique. Only a specific implementation of the
Newmark method is studied in this paper although there are several possible imple-
mentations for this integration method. It seems numerical properties of the Newmark
method in the solution of linear elastic and nonlinear systems can be entirely captured
by the newly developed technique. Although this technique is only aimed at a spe-
cific time step, it is still indicative for the whole step-by-step integration procedure
since this procedure consists of each time step.

Key Words: nonlinear system, degree of nonlinearity, evaluation, stability, accuracy.

I. INTRODUCTION by a newly developed technique (Chang, 2003b). In


fact, its numerical characteristics, such as local trun-
Step-by-step integration methods (Bathe and cation error, order of accuracy, overshooting, stabil-
Wilson, 1973; Chang, 1996; 1997; 2001a; 2002; ity and accuracy, in the solution of nonlinear systems
Goudreau and Taylor, 1972; Hilber et al., 1977; are evaluated by this technique. The details of using
Newmark, 1959) are widely used for solving equa- this technique to evaluate the Newmark method are
tions of motion so that dynamic behaviors of the sys- presented. Meanwhile, analytical results are dis-
tem under a specific loading can be obtained. In fact, cussed and correlated to numerical phenomena
it is well recognized that the step-by-step integration experienced.
method is the most powerful technique for nonlinear
dynamic analysis. The performance of integration II. EQUATIONS OF MOTION
methods is very important in practical application and
thus they have been thoroughly studied. However, In dynamic analysis, a structural system is of-
numerical properties of the integration method in the ten idealized as a discrete model and thus the equa-
solution of equations of motion are usually evaluated tions of motion for the discrete structural model are
for linear elastic systems only (Belytschko and expressed as:
Hughes, 1983; Chang, 2001b; 2003a; Goudreau and
Taylor, 1972; Krieg, 1973) and those for nonlinear M u +C u +Ku=f (1)
systems are not assessed (Belytschko and Schoeberle,
1975). The Newmark method is a widely used step- where M, C and K are the mass, viscous damping and
by-step integration method in dynamic response stiffness matrices; u, u and u are the displacement,
analysis and thus is chosen for this study. Its perfor- velocity and acceleration vectors, respectively, and f
mance in the solution of nonlinear systems is assessed is an external force vector. In general, M and C are
constant matrices while the stiffness matrix K is var-
ied with time for a nonlinear system.
S. Y. Chang is with the Department of Civil Engineering, Na-
tional Taipei University of Technology, Taipei, Taiwan 106, R.O.
The Newmark method can be used to obtain the
C. (Tel: 886-2-27712171 ext. 2653; Fax: 886-2-27814518; Email: approximate solutions of Eq. (1) and its general for-
changsy@ntut.edu.tw) mulation is
652 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

Table 1 Four well-known members of Newmark method


Method β γ Stability condition
Newmark explicit method 0 1 0< Ω ≤2
2
Fox-Goodwin method 1 1 0< Ω ≤ 6
12 2
Linear acceleration method 1 1 0< Ω ≤2 3
6 2
Average acceleration method 1 1 0<Ω ≤∞
4 2

Ma i+1+Cv i+1+Kd i+1=f i+1 (i+1)-th time step. Similarly, k i+1 is used to denote
the stiffness at the end of (i+1)-th time step or at the
di+1=d i +( ∆ t)v i+( ∆ t) 2[( 1 − β )a i+ β a i+1] beginning of the (i+2)-th time step. This implies that
2
the stiffness during the (i+1)-th time step is varied
v i+1 =v i+( ∆ t)[(1− γ )a i+ γ a i+1] (2) from the beginning of k i to the end of k i+1.
Immediately following the end of the i-th time
where d i, v i and a i are approximations to u(t i), u (t i) step, computation details for the (i+1)-th time step
and u (t i ), respectively, and β and γ are specified co- are described next. The first step is to find the dis-
efficients which govern the stability, accuracy and nu- placement for the next time step. In fact, using the
merical dissipation of the integration method. To obtain second line of Eq. (3), the acceleration a i+1 can be
the basic properties, it is convenient to consider the rewritten as a function of d i+1. Meanwhile, the ve-
single degree of freedom analogue of Eq. (2) locity vi+1 can be expressed in terms of d i+1 after sub-
stituting the obtained a i+1 into the third line of Eq.
ma i+1+cv i+1+kd i+1=f i+1 (3). As a result, the displacement di+1 can be obtained
after the substitutions of the derived vi+1 and ai+1 into
di+1=d i +( ∆ t)v i+( ∆ t) 2[( 1 − β )a i+ β a i+1] the first line of Eq. (3) and are found to be
2
v i+1 =v i+( ∆ t)[(1− γ )a i+ γ a i+1] (3) [ 1 m + γ c + k ]d
β(∆t) 2 β(∆t) i i+1

w h e r e d i, v i a n d a i a r e a p p r o x i m a t i o n s o f
= fi + 1 + [ m 2 ][d i + (∆t)vi + ( 1 – β)(∆t) 2a i]
displacement, velocity and acceleration at the i-th β(∆t) 2
time step; m, c and k are the mass, viscous damping
γ
and stiffness. – c[vi + (1 – γ )(∆t)a i] + c[ ][d + (∆t)vi
β(∆t) i
The coefficients for the four well-known mem-
bers of Newmark method (Newmark, 1959), which + ( 1 – β)(∆t) 2a i] (4)
are the Newmark explicit method, Fox-Goodwin 2
method, linear acceleration method and constant av- In this equation, it is important to note that the stiff-
erage acceleration method, are summarized in Table ness at the end of the i-th time step ki is used to deter-
1 for the subsequent discussions and comparisons. In mine the displacement d i+1 for the (i+1)-th time step.
addition, their stability limits for linear elastic sys- Next, the restoring force r i+1 corresponding to this
tems are also listed for comparison. In this table, the displacement can be computed by an assumed mate-
relationship Ω = ω ( ∆ t) is defined and ω = k/m is the rial model, which mathematically describes a stress-
natural frequency of the system. strain relation or a load-displacement relation. In
general, the load-displacement relation for the (i+1)-
III. AN IMPLEMENTATION OF NEWMARK th time step can be expressed as r i+1=k i+1d i+1 where
METHOD ki+1 is the stiffness at the end of the (i+1)-th time step.
After updating the stiffness k i+1 , the velocity can be
In order to evaluate the numerical properties of computed based on the new determined stiffness k i+1
Newmark method in the solution of a nonlinear system, and are found to be
the step-by-step integration details within a time step
are described and then are realistically reflected in vi + 1 = {[ 1 ]m + c} – 1{ fi + 1 – k i + 1d i + 1
the proposed evaluation procedure. In the following γ (∆t)
derivations, ki is used to represent the stiffness at the
+ [ m ][vi + (1 – γ )(∆t)a i]} (5)
end of the i-th time step or at the beginning of the γ (∆t)
S. Y. Chang: Studies of Newmark Method for Solving Nonlinear Systems: (I) Basic Analysis 653

Finally, the acceleration can be computed by Obviously, there is no stiffness change for µi+1=1 dur-
ing the (i+1)-th time step. Meanwhile, the case of
1 )( f – cv – k d )
ai + 1 = ( m (6) µi+1>1 simulates the stiffness hardening, and the stiff-
i+1 i+1 i+1 i+1
ness softening can be mimicked by 0< µ i+1<1. Thus,
In this computing procedure, the stiffness k i is used µi+1 can be considered as a measure of the degree of
to determine the displacement d i+1 while the velocity nonlinearity for the (i+1)-th time step. In general,
and acceleration are computed by using the stiffness for µ i+1>1 the larger µi+1 is, the higher the degree of
ki+1. This reveals that the use of different stiffnesses hardening nonlinearity is while for 0< µ i+1 <1 the
to compute the displacement, velocity and accelera- smaller µ i+1 is, the higher the degree of softening
tion are the major differences between the linear elas- nonlinearity is.
tic and nonlinear systems. After defining the degree of the nonlinearity µi+1
for the (i+1)-th time step, the integration procedure
IV. ANALYSIS as shown in Eq. (4) to Eq. (6) can be rewritten in a
matrix form (Bathe and Wilson, 1973; Belytschko and
It is often found that for a nonlinear system the Hughes, 1983) as
stiffness is k i at the beginning of the (i+1)-th time
step while it becomes k i+1 at the end of this time step. X i+1=A i+1X i+Li+1f i+1 (8)
Since k i+1 is generally different from k i for a nonlin-
ear system, a new parameter µ i+1 is used to describe where X i =[d i, ( ∆ t)v i , ( ∆ t) 2a i ] T; A i+1 is the amplifica-
the change of the stiffness during the (i+1)-th time tion matrix and L i+1 is the load vector for the (i+1)-th
step. For this purpose, the degree of nonlinearity µi+1 time step. In this study, the viscous damping is as-
is defined as the ratio of k i+1 over k i sumed to be zero in order to simplify the algebraic
manipulations. The explicit expression of the matrix
k i+1 = µ i+1k i (7) Ai+1 is

1 1 1 –β
2
Ai + 1 = 1 – µ i + 1γΩi2 1 + (β – µ i + 1γ )Ωi2 (1 – γ )(1 + βΩi2) – ( 1 – β)(µ i + 1γΩi2) (9)
1 + βΩi2 2
– µ i + 1Ωi2 – µ i + 1Ωi2 – ( 1 – β)µ i + 1Ωi2
2

where Ω i = ω i ( ∆ t); and ω i = k i / m is the natural fre- method, such as the local truncation error,
quency of the system determined from the structural consistency, stability, accuracy and overshoot, can be
properties at the end of the i-th time step. The three further evaluated for both linear and nonlinear
eigenvalues of the amplification matrix A i+1 can be systems.
found from the characteristic equation (Belytschko
and Hughes, 1983) 1. Local Truncation Error and Order of Accuracy

det (Ai+1− λ I)= λ 3−A 1λ 2+A 2λ −A 3 =0 (10) Since the object of numerical integration is to
determine sufficiently accurate approximations with
where minimum effort, a means to evaluate the efficiency
of various approximation methods is required. Local
2 + (β + µ i + 1β – µ i + 1γ – 1 µ i + 1)Ωi2
A1 = 2 truncation error can be used as a criterion for this
1 + βΩi2 purpose. The local truncation error is defined as the
error committed in each time step by replacing the
1 + µ i + 1(β – γ + 1 )Ωi2 differential equation by the difference equation
A2 = 2
1 + βΩi2 (Belytschko and Hughes, 1983). In general, the lo-
cal truncation error for the (i+1)-th time step can be
A 3=0 (11) expressed as
n
After solving the characteristic equation, numerical
properties of the (i+1)-th time step for the Newmark
Ei + 1 = Σ [H pu (p)(∆t) p – 2] + O[(∆t) p – 1]
p=0
(12)
654 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

where u(p) denotes the p-th derivative of the displace- Newmark method and thus the accuracy significantly
ment u with respect to time and the coefficient H p is deteriorates when compared to that for the solution
found to be of a linear system. However, consistency and non-
zero order of accuracy might be achieved if µ i+1 is
H 0 =1−A 1+A 2 close enough to 1. This is because the first and/or
second terms on the right hand side of Eq. (16) is
1 + ( – 1) p A 2 smaller than the subsequent term.
Hp = p=1, 2, 3, ... (13)
p!
The difference equation with local truncation error 2. Overshooting Effect
E i+1 ( ∆ t) is said to be consistent with the differential
equation if Goudreau and Taylor (1972) discovered a very
peculiar property of the Wilson θ method. Although
lim ( max E i + 1(∆t) ) = 0 (14) this method is unconditionally stable for linear systems,
∆t → 0 1 ≤ i + 1 ≤ N
numerical experiments showed a tendency to over-
shoot exact solutions significantly in the early response.
where N is the total number of integration time steps.
Therefore, the tendency to overshoot is an important
If E i+1=O[(∆t)q] where q>0, then q is named the order
factor, which should be considered in an evaluation
of accuracy (Belytschko and Hughes, 1983). It is ap-
of an implicit method. To assess the tendency of an
parent that consistency implies that Ei+1→0 as ∆t→0,
integration method to overshoot the exact solutions,
and the order of accuracy is the rate at which the lo-
one can consider a single degree of freedom system
cal truncation error approaches zero as the time step
subject to any arbitrary initial conditions di and vi and
tends to zero. Furthermore, the difference equation
then compute d i+1 and v i+1 as a function of Ω i. Since
is convergent with the differential equation if
only convergent integration methods are considered
lim ( max u(t i + 1) – d i + 1 ) = 0 (15) there will be no overshooting effect as Ω i →0.
∆t → 0 1 ≤ i + 1 ≤ N
However, the behavior as Ω i→∞ will provide an in-
dication of the behavior of the high frequency modes
where di+1 is the approximation obtained from the dif-
in a structural system. This is a typical condition when
ference equation and u(t i+1) is the exact solution for
an unconditionally stable implicit method is used to
the differential equation.
solve a large multiple degree of freedom system.
After the substitutions of A1 and A2 into Eq. (12),
The following two equations are directly ob-
the local truncation error for the (i+1)-th time step of
tained from Eq. (8) after using the equation of mo-
the Newmark method is found to be
tion to eliminate the acceleration term a i

µi + 1 – 1 2 β – µ i + 1 (β – γ + 1 ) 1
Ei + 1 = ω u + 2 ωΩu di + 1 = {[1 + (β – 1 )Ωi2]d i + (∆t)vi}
1 + βΩi2
1
1 + βΩi2
i i 1 + βΩi2 2

1 – 1 [β + µ (β – γ + 1 )] vi + 1 = 1 { – {1 – γ + γµ i + 1 + [β(1 – γ )
+ 12 2 i+1 2 ω 2Ω 2u 1 + βΩi2
1 + βΩi2 i i
d
+ (β – 1 )γµ i + 1]Ωi2}Ωi2( i )
+ O[(∆t) ] 3
(16) 2 ∆t

In addition, for µ i+1=1, it reduces to be + [1 + (β – µ i + 1γ )Ωi2]vi} (18)

γ–1 1 – β + 1 (γ – 1 ) Results for the limiting condition of Ω i→∞ for these


2 12 2 2 ω 2Ω 2u
Ei + 1 = 2 ω iΩiu + two equations are found to be
1 + βΩi 1 + βΩi 2 i i

d i + 1 ≈ (1 – 1 )d i
+ O[(∆t) 3] (17) 2β
d γ
This equation clearly shows that for µ i+1=1, the order vi + 1≈ [γ – 1 + ( 1 – 1)γµ i + 1]Ωi2( i ) + (1 – µ i + 1 )vi
2β ∆t β
of accuracy is one if γ ≠ 1 , two if γ = 1 and three if β=
2 2 (19)
1 in addition to the case of γ = 1 . Based on Eq. (14),
12 2
consistency cannot be achieved for any degree of The first line of this equation reveals that there is no
hardening or softening nonlinearity since µ i+1≠1. In overshoot in displacement for either linear or nonlin-
fact, the order of accuracy degenerates to 0 in the ear systems using the Newmark method. However,
solution of a nonlinear system for any member of the it has a tendency to overshoot quadratically in Ω i in
S. Y. Chang: Studies of Newmark Method for Solving Nonlinear Systems: (I) Basic Analysis 655

the velocity equation due to the displacement term It is found that the second inequality of this equation
(di/∆ t) for any degree of nonlinearity. It is worth not- will be satisfied only if the degree of nonlinearity µi+1
ing that the constant average acceleration method is is smaller than or equal to 1 for the case of γ = 1 and
the only member of the Newmark method that shows 2
β≠0 while for γ= 1 and β=0 the first inequality is not
no overshoot in velocity for linear elastic systems. 2
met for any µ i+1>0. Thus, in the sub-family of the
This is because β = 1 , γ = 1 and µ i+1 =1, (d i / ∆ t) in the
4 2 Newmark method with γ= 1 an unconditionally stable
second line of Eq. (19) disappears. However, the 2
overshooting in velocity is still present if it is used to algorithm may exist for solving softening or linear
solve a nonlinear system. systems but not for solving hardening systems.

3. Stability (2) Conditional stability

There is a zero eigenvalue for the Newmark On the other hand, the conditional stability lim-
method since A 3 =0 in Eq. (12). Thus, this integra- its for the Newmark method can be also found from
tion method is spectrally stable (Belytschko and Hughes, Eq. (20) and are:
1983) if the other two eigenvalues of the amplifica-
tion matrix Ai+1 satisfy the following two inequalitie µ i+1 (β − γ )+ β <0, µ i+1( β − γ + 1 )< β and
2
−(A 2+1)≤A 1≤(A 2+1) if −1≤A 2<1
0< Ω i≤ 2 (23)
−2<A 1 <2 if A 2=1 (20) µ i + 1 (γ – β ) – β

In this derivation, the two eigenvalues can be real or where the second inequality is met only if µ i+1<1 for
complex. Thus, the response is in a bounded oscilla- γ = 1 and β ≠0. This implies there is no conditionally
2
tory form or an exponential decay form. Although a stable method in the sub-family of γ = 1 and β ≠0 for
stable computation can be guaranteed by Eq. (20), it 2
the system with µi+1≥1. In fact, the Newmark explicit
is advisable to satisfy more stringent stability
conditions, which restricts the two eigenvalues to be method is the only method in the sub-family of γ = 1
2
complex conjugate, and thus a bounded oscillatory that can have conditional stability for any µi+1 and its
response is obtained. As a result, two complex con- stability limits are found to be 0<Ωi≤2/ µ i or 0<Ωi+1≤2.
jugates and |λ 1,2|≤1 leads to
(ii) Two Complex Conjugate Eigenvalues − Eq. (21)
A 21 −4A 2<0
(1) Unconditional stability
0<A 2 ≤1 (21)
Using the inequality of A 21 −4A2<0, the lower and
If any member of the Newmark method can satisfy
upper bounds of β are found to be:
the stability conditions either defined in Eq. (20) or
(21) for any value of Ω i then it possesses uncondi-
a ≤β if µ i+1=1
tional stability for the (i+1)-th time step. On the other
hand, if the stability conditions are met only for a
a≤ β≤b if µ i+1≠1 (24)
certain interval of Ω i then the integration method can
only have the conditional stability. To explore the
where
stability conditions of the Newmark method, Eqs. (20)
and (21) will be thoroughly studied next.
a = 1 (γ + 1 ) 2
4 2
(i) Two Real or Two Complex Eigenvalues − Eq. (20)
a= 1 {µ i + 1[µ i + 1(γ + 1 ) – γ + 3 ]
(1) Unconditional stability (1 – µ i + 1) 2 2 2

For the Newmark method to have the response – 2µ i + 1 µ i + 1 (γ + 1 ) – γ + 1 }


2 2
in a bounded oscillatory form or an exponential de-
cay form, unconditional stability conditions can be 1
b= {µ i + 1[µ i + 1(γ + 1 ) – γ + 3 ]
easily determined after the substitution of A1 and A2 (1 – µ i + 1) 2 2 2
into Eq. (20) and are found to be:
+ 2µ i + 1 µ i + 1 (γ + 1 ) – γ + 1 } (25)
µ i+1( β − γ )+ β≥0 and − β ≤µ i+1( β− γ + 1 )≤ β (22) 2 2
2
656 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

where a is proved to be greater than or equal to zero


if the result of the square root is a real number. It is
clear that the result of the square root in Eq. (25) is a
real number for µi+1>1 while for the case of 0<µ i+1<1
it must satisfy the following inequality:

1 + µi + 1
γ ≤ 1( ) (26)
2 1 – µi + 1

As a result, for a≥0 and a ≥0 Eq.(24) implies β ≥0.


Thus, the following results can be further obtained
from the constraint of 0<A 2≤1 based on β ≥0

c≤ β≤d and γ ≥ 1 if µ i+1>1


2
Fig. 1 Variations of lower and upper bounds of β with µ i+1 for
stiffness hardening
γ≥ 1 if µ i+1=1
2
β≥γ− 1 if 0< µ i+1<1 (27)
2
the increase of µ i+1 as γ > 1 . Besides, it is also found
2
where that the maximum value of µi+1 increases with the in-
µ crease of γ, which must be larger than 1 . For example,
c= γ − 1 and d = ( i + 1 )(γ – 1 ) (28) 2
2 µi + 1 – 1 2 as β = 1 and γ = 3 , the unconditional stability is
2 4
achieved only if 1< µ i ≤2 while for β = 5 and γ = 3 , it
As a further step, the conditions to have uncon- 4 2
ditional stability for the Newmark method can be sum- extends to be 1< µ i+1 ≤5. However, it will be shown
marized from Eqs. (24) and (27) and are later that a larger value of γ will lead to less accuracy
of numerical solution as γ ≥ 1 .
2
µi + 1 + 1 Conditions to have unconditional stability for
a≤ β ≤d, γ ≥ 1 and γ ≥ if µ i+1>1
2 4 µ i+1=1, as shown in the second line of Eq. (29), are
the same as those which have been derived for linear
a ≤ β and γ ≥ 1 if µ i+1=1 systems (Belytschko and Hughes, 1983). It is inter-
2
esting to study the unconditional stability conditions
1 + µi + 1
a≤ β ≤b and γ ≤ 1 ( ) if 0< µ i+1<1 (29) for stiffness softening, as shown in the third line of
2 1 – µi + 1 Eq. (29). In this line, the second inequality can be
In this equation, the inequality γ ≥(µi+1+1)/4 in the first used to obtain the results of ( γ − 1 )/( γ + 1 )≤ µ i+1<1 if
2 2
case of µ i+1>1 implies that the degree of nonlinearity γ > 1 and 0< µ i+1<1 if γ ≤ 1 . This implies that for any
2 2
µ i+1 must be smaller than or equal to 1 if γ = 1 . member of the Newmark method with γ > 1 , the un-
2 2
Obviously, this is contradictory to the assumption of conditional stability can be obtained only if the de-
µ i+1>1 for this case. Thus, it is concluded that any gree of softening nonlinearity is in the range of ( γ −
member of the Newmark method with γ = 1 cannot 1 )/(γ+ 1 )≤µ <1. However, for the Newmark method
i+1
2 2 2
possess the desired unconditional stability for any with γ ≤ 1 , it can be achieved for any degree of soft-
2
degree of hardening nonlinearity. ening nonlinearity. These results can be seen in Fig.
To see the variations of a and d with respect to 2 where the variations of a and b with respect to µi+1
µi+1 for a given value of γ , the first line of Eq. (29) is
for γ=− 1 , 0, 1 , 1 and 3 are depicted. In addition, for
plotted in Fig. 1. The line with an arrow denotes the 2 2 2
a given value of γ , the possible range of β to have
possible range of β, which is 1 ≤β <∞, to have uncon-
4 unconditional stability for stiffness softening in-
ditional stability for µ i+1 =1 as γ = 1 . In addition, the creases with increasing µ i+1. In general, the higher
2
solid circles represent the lower bounds of β as µi+1=1 the degree of softening nonlinearity is, the smaller
for the cases of γ = 1 , 3 , 1, 5 and 3 ; and their corre- the range of β for a given value of γ is.
2 4 4 2 After the study of the variations of β and γ ver-
sponding values of a are found to be a= 1 , 25 , 9 ,
4 64 16 sus µ i+1 to have unconditional stability, it is natural
49 and 1. For stiffness hardening, the possible range
64 to consider the possible range of µ i+1 to possess the
of β to have unconditional stability is decreased with unconditional stability for given β and γ . Thus,
S. Y. Chang: Studies of Newmark Method for Solving Nonlinear Systems: (I) Basic Analysis 657

Fig. 2 Variations of lower and upper bounds of β with µ i+1 for Fig. 3 Variations of lower and upper bounds of µ i+1 with β
stiffness softening

using the procedure similar to the derivation of Eq. (2) Conditional stability
(29), Eq. (21) will lead to
Conditional stability limits for the Newmark
e≤ µ i+1 ≤f if γ − 1 < β≤ γ − 1 method can be derived from Eq. (21) and are summa-
2 4 rized as follows
e≤ µ i+1 ≤g if β ≥ γ − 1 and β ≠γ + 1
4 2 0<Ω i<min( Ωia , Ωib ) if β <γ − 1 and
2
1 β≤µ if β ≥ γ − 1 and β = γ + 1
4 i+1 4 2
(30) µ i+1<e or f< µi+1

where 0<Ω i < Ωia if β = γ − 1 and 0< µ i+1<∞


2
β[β + 2 – (γ + 1 )] – 2β β – (γ – 1 ) 0< Ω i< Ωia if γ− 1 < β< γ − 1 and µ i+1<e or f<µ i+1≤g
e= 2 2 2 4
1
[β – (γ + )] 2
2
0<Ω i< Ωia if γ − 1 ≤β and µ i+1<e (32)
4
β[β + 2 – (γ + 1 )] + 2β β – (γ – 1 )
f= 2 2 where
1
[β – (γ + )] 2
2 µi + 1
Ωia = 2
β [(1 – µ i + 1)β + µ i + 1(γ + 1 )] 2 – 4µ i + 1β
g= (31) 2
β–γ+1
2
Ωib = –1 (33)
The first line of Eq. (30) indicates that there is no µ i + 1 (β – γ + 1 )
2
unconditionally stable integration method if β−γ+ 1 ≤0
2
and it may exist only if β − γ + 1 >0. Details can be The Newmark explicit method is in the second
2 case of Eq. (32); the Fox-Goodwin method and lin-
found in Fig. 3, where Eq. (30) with γ =− 1 , 0, 1 , 1
2 2 ear acceleration method are in the third case; and the
and 3 for the variation of µ i+1 with β is plotted. It is constant average acceleration method is in the fourth
2
found that for γ ≤ 1 and β >0, the integration method case. As a result, conditional stability limits for the
2
can have unconditional stability only for stiffness four integration methods reduce to be
softening while for γ > 1 and β ≥ γ − 1 the method to µi + 1
2 4
have unconditional stability may exist for both stiff- 0 ≤ Ωi ≤ Ωia = 2
[(1 – µ i + 1)β + µ i + 1] 2 – 4µ i + 1β
ness hardening and softening and is limited to some
specific range of µ i+1 . In fact, the interval between (34)
the thin and thick lines, which have the same line type,
is the specific range for given β and γ . for γ = 1 .
2
658 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

Table 2. Possible range of µ i to have unconditional and conditional stability for the four members of the
Newmark method
Method Two real or two complex eigenvalues Two complex conjugate eigenvalues
β γ Unco. Stab. Conditional Stability Unco. Stab. Conditional Stability
0 1
2
none 0<µ i+1<∞ 0< Ω i none 0< µ i+1<∞ 0< Ω i≤ 4
µi + 1
≤ 4
µi + 1

1 1 0< µ i+1 ≤ 1 1 < µ ≤1 0< Ω i 13 – 4 3 ≤µ


i+1 0< µ i+1 < 121
13 – 4 3 or 0< Ω
12 2 5 5 i+1 121 i
144µ i + 1
≤ 24 ≤ 121
13 + 4 3 13 + 4 3 < µ i+1<1 ≤
5µ i + 1 – 1 121 121µ i2+ 1 – 26µ i + 1 + 1

1
6
1
2
0< µ i+1 ≤ 12 1 < µ ≤1
2 i+1
0< Ω i 7 – 2 6 ≤µ
i+1 0< µ i+1< 7 – 2 6 or 0< Ω i
25 25
7 + 2 6 < µ <1 144µ i + 1
≤ 12
2µ i + 1 – 1
≤ 7 + 2 6 ≤
25µ i2+ 1 – 14µ i + 1 + 1
25 25 i+1

1
4
1
2
0< µ i+1 ≤1 none none 1 ≤ µ ≤1
9 i+1
0< µ i+1< 19 0< Ω i≤
64µ i + 1
9µ i2+ 1 – 10µ i + 1 + 1

by a dash-dot arrow, and is conditionally stable for


the rest of the interval of 0<µi+1≤1, as is indicated by
an arrow. These intervals for the three integration
methods will be found and summarized later. For
example, the constant average acceleration method
is unconditionally stable for 1 ≤µ i+1 ≤1 and is condi-
9
tional stable for 0< µ i+1< 1 . This can be also seen in
9
Fig. 4.

(iii) Summary

Since the four members of the Newmark method


are widely used it is useful to summarize the results
Fig. 4 Variations of upper stability limit with µi+1 for well-known of the stability analysis in Table 2. This table is con-
members of Newmark method sidered as eight columns except the first two rows.
The third column is obtained from Eq. (22) and Eq.
(23) determines the fourth and fifth columns. In
Variations of the upper stability limits with µ i+1 addition, the sixth column is computed from Eq. (30)
for the four integration methods are plotted in Fig. 4. and the seventh and eighth columns are calculated
It should be mentioned that the actual range for the from Eq. (32). In this table, it is found that for the
Newmark explicit method is 0< µ i+1 <∞ although the Fox-Goodwin method, linear acceleration method
upper stability limit shown in this figure is only the and constant average acceleration method a stable
range of 0.01<µ i+1 <100. Whereas those of the other computation can be achieved only for stiffness soft-
integration methods are limited to the range of ening and constant stiffness while for any stiffness
0<µ i+1≤1 since they lead to instability for any degree hardening they will lead to instability. Although the
of hardening nonlinearity. It is found that the Newmark explicit method can only possess condi-
Newmark explicit method can only have conditional tional stability, numerical instability will not occur
stability and its upper stability limit is decreased with in the solution of any degree of nonlinearity if the
the increase of µi+1. On the other hand, for the other stability limits 0<Ω i ≤2/ µ i + 1 are met. It should be
integration methods, each method is unconditionally mentioned that the fifth or eighth column reduces to
stable in a certain interval of 0<µi+1≤1, as is indicated be the same as the fourth column of Table 1 if µ i+1 is
S. Y. Chang: Studies of Newmark Method for Solving Nonlinear Systems: (I) Basic Analysis 659

taken to be 1. It becomes clear that an uncondition-


ally stable method for linear systems may lead to in-
stability for nonlinear systems after the stability
analysis.

4. Accuracy

For the Newmark method, if there exist two com-


plex conjugate eigenvalues in addition to the zero
eigenvalue, these eigenvalues λ1,2 can be further ex-
pressed as:

A 1 ± j 4A 2 – A 21 – ξi+1 Ωi+1 ± j Ωi+1


λ 1, 2 = =e (35)
2

where Ω i + 1 = ω i + 1 (∆t) and j= – 1 . According to this


equation, the phase Ω i + 1 and the numerical damp-
ing ratio ξ i + 1 are found to be

4A 2 ln(A 2)
Ω i + 1 = tan – 1( – 1 ) , ξi +1 = –
A1 2 Ωi+1

(36)

In addition, the relative period error is proposed to


measure period distortion and is defined as
Fig. 5 Numerical characteristics for the linear acceleration method
T – Ti+1 Ω
Pish+ 1 = i+1 = 1 – i+1 if T i+1 > T i + 1
Ti + 1 Ωi+1

T i + 1 – Ti + 1 Ωi + 1 damping ratio, and spectral radius, five values of µ i+1


Piel+ 1 = = –1 if T i+1 < T i + 1
Ti + 1 Ωi+1 are thoroughly considered for the Newmark explicit
(37) method and the constant average acceleration method.
However, numerical characteristics for the linear ac-
where T i+1 =(2 π )/ ω i+1 and Ti+1 =(2 π )/ ω i + 1 . The celeration method and Fox-Goodwin method are also
symbols Pish+ 1 and Piel+ 1 are the relative period errors provided in Figs. 5 and 6, respectively.
corresponding to the cases of period shrinkage and
elongation during the (i+1)-th time step, respectively. (i) Newmark Explicit Method
The spectral radius is generally defined as
ρ i+1 =max| λ j| for j=1, 2, 3 and is equal to ρ i+1 = A 2 = It is evident from Eq. (38) that the spectral ra-
e – ξ i + 1 Ω i + 1 if two eigenvalues are complex conjugate. dius for the Newmark explicit method is 1 for any
The explicit expression of the spectral radius in the degree of nonlinearity. Thus, this method possesses
case of two complex eigenvalues for the Newmark no numerical dissipation for either linear elastic or
method is found to be nonlinear systems since ξ i + 1 =0. However, its rela-
tive period error highly depends on the degree of
nonlinearity µ i+1 and the details are shown in Fig. 7.
1 + µ i + 1(β – γ + 1 )Ωi2 It should be mentioned that the first line of Eq. (37)
ρi + 1 = 2 (38) is used to compute the relative period error since the
1 + βΩi2
Newmark explicit method always shows a period
shrinkage for any degree of nonlinearity. For each
It is clear that ρ i+1 >1 will lead to instability since it
curve, the relative period error is gradually increased
may induce exponential growth. Thus, for γ = 1 , Eq. with the increase of ∆ t/T i while as ∆ t/T i tends to the
2
(38) implies a stable computation may be obtained only stability limit it grows very rapidly. This figure also
if 0<µi+1≤1 while instability will occur if µ i+1>1. shows that the relative period error is increased with
To illustrate the influence of degree of the increasing value of µ i+1 for a specific value of
nonlinearity µ i+1 to relative period error, numerical ∆ t/T i .
660 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

Fig. 8 Spectral radius for the constant average acceleration method

Fig. 6 Numerical characteristics for the Fox-Goodwin method

Fig. 9 Relative period error for the constant average accelera-


tion method

Fig. 7 Relative period error for the Newmark explicit method

(ii) Constant Average Acceleration Method

Numerical properties for the constant average


Fig. 10 Numerical dissipation for the constant average accelera-
acceleration method are plotted in Figs. 8 to 10. The tion method
variation of spectral radius with ∆ t/T i is plotted in
Fig. 8 where a solid circle is the point at which com-
plex eigenvalues bifurcate into real and distinct degree of softening nonlinearity and greater than 1
eigenvalues. It can be seen that for the stiffness in- for any degree of hardening nonlinearity. This im-
variant the spectral radius is always equal to 1 for plies that numerical solutions will be eventually
any value of ∆ t/T i. However, it is less than 1 for any blown up for any degree of hardening nonlinearity
S. Y. Chang: Studies of Newmark Method for Solving Nonlinear Systems: (I) Basic Analysis 661

while stable solutions can be obtained for the case of results are still indicative for nonlinear dynamic
stiffness invariant and any degree of softening analysis. Numerical properties for the solution of
nonlinearity. This figure also shows that the spectral nonlinear systems highly depend on the degree of
radius is only a little greater than 1 for µ i+1 >1 as nonlinearity in each time step. This explains why the
∆ t/T i is small enough. This strongly indicates that performance of a step-by-step integration method in
reliable solutions for any degree of hardening the solution of nonlinear systems cannot be reliably
nonlinearity may still be achieved if the integration predicted by the evaluation methods used for linear
time step is small enough. In fact, it will be illus- elastic systems. It seems that the theoretical evalua-
trated by means of numerical examples in the accom- tion results for the Newmark method are consistent
panying paper (Chang, 2004) that the way to yield with the phenomena experienced in the step-by-step
reliable solutions for any degree of hardening solution of nonlinear systems. However, numerical
nonlinearity is essentially to rely upon the use of a verifications are needed. In fact, detailed verifica-
small time step to keep the spectral radius smaller tions and guidelines to have accurate solutions for
than 1 and thus to delay the numerical instability un- nonlinear dynamic analysis are studied in the accom-
til accurate results are obtained. This is because that panying paper.
as ∆ t/T i is small enough the solution is still in an os-
cillatory form and the spectral radius is less than 1 ACKNOWLEDGMENTS
since there are two complex conjugate eigenvalues
with the norm below 1. The relative period error for The author is grateful to acknowledge that this
the constant average acceleration method is shown study is financially supported by the National Sci-
in Fig. 9. For each curve, the relative period error ence Council, Taiwan, R.O.C., under Grant No. NSC-
increases gradually with the increasing value of 91-2218-E-027-010.
∆ t/T i while as ∆ t/T i tends to the stability limit it be-
comes steep except that the curves for µ i+1 =10 and REFERENCES
100 show irregularity as their stability limits are
approached. It can also be found that for a given value Bathe, K. J., and Wilson, E. L., 1973, “Stability and
of ∆t/T i , constant stiffness exhibits the least relative Accuracy Analysis of Direct Integration Me-
period error while for stiffness hardening the rela- thods,” Earthquake Engineering and Structural
tive period error is increased with the increasing value Dynamics, Vol. 1, pp. 283-291.
of µ i+1 and for stiffness softening it increases with Belytschko, T., and Schoeberle, D. F., 1975, “On the
the decrease of µ i+1 . Thus, either for stiffness hard- Unconditional Stability of an Implicit Algorithm
ening or softening, a smaller value of ∆ t/T i is in gen- for Nonlinear Structural Dynamics,” Journal of
eral required to yield about the same accuracy in a Applied Mechanics, Vol. 17, pp. 865-869.
given period when compared to that for constant Belytschko, T., and Hughes, T. J. R., 1983, Compu-
stiffness. Fig. 10 depicts the variation of the numeri- tational Methods for Transient Analysis, Elsevier
cal damping ratio ξ i + 1 with ∆t/T i. As expected, the Science Publishers B.V., North-Holland.
constant average acceleration method is non-dissipa- Chang, S. Y., 1996, “A Series of Engergy Conserv-
tive for the case of stiffness invariant. However, it is ing Algorithms for Structural Dynamics,” Jour-
interesting to find that a positive damping ratio is gen- nal of the Chinese Institute of Engineers, Vol. 19,
erated for the stiffness softening while a negative No. 2, pp. 219-230.
damping ratio is introduced for stiffness hardening. Chang, S. Y., 1997, “Improved Numerical Dissipa-
Besides, the numerical damping ratio increases with tion for Explicit Methods in Pseudodynamic
the increasing value of the degree of hardening Tests,” Earthquake Engineering and Structural
nonlinearity while the absolute value of the numeri- Dynamics, Vol. 26, No. 9, pp. 917-929.
cal damping ratio increases with the decreasing value Chang, S. Y., 2001a, “Analytical Study of the Supe-
of the degree of softening nonlinearity. It also can riority of the Momentum Equations of Motion for
be found that the absolute value of the numerical Impulsive Loads,” Computers and Structures,
damping ratio is increased with the increase of ∆ t/T i Vol. 79, No.15, pp.1377-1394.
for a given value of the degree of nonlinearity. Chang, S. Y., 2001b, “Application of the Momentum
Equations of Motion to Pseudodynamic Test-
V. CONCLUSIONS ing,” Philosophical Transactions of the Royal So-
ciety, Series A, Vol. 359, No. 1786, pp.1801-1827.
Numerical characteristics of the Newmark Chang, S. Y., 2002, “Explicit Pseudodynamic Algo-
method for both linear and nonlinear systems are rithm with Unconditional Stability,” Journal of
evaluated by using the new evaluation technique. Engineering Mechanics, ASCE, Vol. 128, No.9,
Although only a specific time step is evaluated the pp. 935-947.
662 Journal of the Chinese Institute of Engineers, Vol. 27, No. 5 (2004)

Chang, S. Y., 2003a, “Accuracy of Time History Hilber, H. M., Hughes, T. J. R., and Taylor, R. L.,
Analysis of Impulse,” Journal of Structural En- 1977, “Improved Numerical Dissipation for Time
gineering, ASCE, Vol. 129, No.3, pp. 357-372. Integration Algorithms in Structural Dynamics,”
Chang, S. Y., 2003b, “Nonlinear Error Propagation Earthquake Engineering and Structural Dynam-
Analysis for Explicit Pseudodynamic Algo- ics, Vol. 5, pp. 283-292.
rithm,” Journal of Engineering Mechanics, ASCE, Krieg, R. D., 1973, “Unconditional Stability in Nu-
Vol. 129, No.8, pp. 841-850. merical Time Integration Methods,” Journal of
Chang, S. Y., 2004, “Studies of Newmark Method for Applied Mechanics, Vol. 40, No. 3, pp. 417-421.
Solving Nonlinear Systems: (II) Verification and Newmark, N. M., 1959, “A Method of Computation
Guideline,” Journal of the Chinese Institute of for Structural Dynamics,” Journal of Engineering
Engineers, Vol. 27, No. 5, pp. 663-675. Mechanics Division, ASCE, Vol. 85, pp. 67-94.
Goudreau, G. L., and Taylor, R. L., 1972, “Evalua-
tion of Numerical Integration Methods in Elasto- Manuscript Received: May 02, 2003
dynamics,” Computer Methods in Applied Mechan- Revision Received: Dec. 10, 2003
ics and Engineering, Vol. 2, No. 1, pp. 69-97. and Accepted: Jan. 20, 2004

View publication stats

You might also like