Abstract: In structural dynamics, integration algorithms are often used to obtain the solution of temporally discretized equations of
Downloaded from ascelibrary.org by Purdue University Libraries on 07/24/17. Copyright ASCE. For personal use only; all rights reserved.
motion at selected time steps. Various time integration algorithms have been developed in the time domain using different methods. In
order for an integration algorithm to be reliable it must be stable and accurate. A discrete transfer function is used to study the properties
of integration algorithms. A pole mapping rule from control theory in conjunction with a discrete transfer function is used to develop new
integration algorithms for obtaining solutions to structural dynamics problems. A new explicit integration algorithm, called the CR 共Chen
and Ricles兲 algorithm, is subsequently developed based on the proposed method. The properties of the algorithm are investigated and
compared with other well established algorithms such as the Newmark family of integration algorithms. By assigning proper stable poles
to the discrete transfer function the newly developed CR explicit algorithm is unconditionally stable and has the same accuracy as the
Newmark method with constant acceleration. In addition, the CR algorithm is based on expressions for displacement and velocity that are
both explicit in form, making it an appealing integration algorithm for solving structural dynamics problems.
DOI: 10.1061/共ASCE兲0733-9399共2008兲134:8共676兲
CE Database subject headings: Algorithms; Structural dynamics; Transfer functions; Stability; Discrete elements.
Properties of Transfer Functions where xi+1, ẋi+1, ẍi+1⫽displacement, velocity, and acceleration of
the SDOF structure at the 共i + 1兲th time step and Fi+1⫽value of
For a single-degree-of-freedom 共SDOF兲 linear elastic structure, excitation F共t兲 at the 共i + 1兲th time step.
the differential equation of motion under external excitation can For a discrete system, the Z transform is defined by Ogata
be expressed as 共1995兲 as
⬁
m · ẍ共t兲 + 2mn · ẋ共t兲 + m2n · x共t兲 = F共t兲 共1兲
Z兵F共k兲其 = F共z兲 = 兺
k=0
F共k兲 · z−k 共5兲
In Eq. 共1兲 n, , and m are equal to the natural frequency of the
SDOF structure, the viscous damping ratio of the SDOF structure,
In Eq. 共5兲, F共k兲⫽sample of F共t兲, where k refers to the discrete
and the mass of the SDOF structure, respectively; F共t兲⫽external
sample time. If F共t兲 = 0 for t less than 0, and if F共t兲 has the Z
excitation; and x共t兲, ẋ共t兲, ẍ共t兲⫽displacement, velocity, and accel-
transform F共z兲, then by the real translation theorem 共Ogata 1995兲
eration of the SDOF structure, respectively. Assuming at rest ini-
tial conditions, the continuous form of the transfer function
between the excitation F共t兲 and the displacement x共t兲 can be de- Z兵F共k − 1兲其 = z−1 · F共z兲 共6兲
rived in the Laplace domain and is equal to
Thus, from Eq. 共6兲, multiplication of F共z兲 by z−1 has the effect of
X共s兲 1 delaying the time function F共t兲 by the time step, i.e., Fi共z兲
G共s兲 = = 共2兲
F共s兲 共s2 + 2n · s + 2n兲m = z−1 · Fi+1共z兲 with Fi共z兲 and Fi+1共z兲 equal to F共t兲 at time index i
and i + 1, respectively.
When a difference equation for an integration algorithm is
where s⫽variable in the Laplace domain; and X共s兲 and
known, the corresponding discrete transfer function G共z兲 can be
F共s兲⫽Laplace transform of x共t兲 and F共t兲, respectively.
determined through the use of Eq. 共6兲, which typically has the
The values of s that render the denominator of the transfer
following general form
function to be zero are defined as “poles” of the transfer function
共Franklin et al. 2002兲. The denominator is also called the charac-
teristic equation of the transfer function. The poles of the transfer X共z兲 nrzr + ¯ + n1z + n0
function G共s兲 in Eq. 共2兲 are complex variables and are equal to G共z兲 = = 共7兲
F共z兲 dqzq + ¯ + d1z + d0
p1,2 = − · n ⫾ i · n · 冑共1 − 2兲 共3兲 where X共z兲 and F共z兲⫽Z transform of x共t兲 and F共t兲, respectively.
In Eq. 共7兲 nr , . . . n0 and dq , . . . d0⫽coefficients of the numerator
where i⫽designation for 冑−1. and denominator of the discrete transfer function G共z兲, respec-
The poles in Eq. 共3兲 determine the response of the system in tively, with q and r equal to the order of the denominator and
Eq. 共1兲 to initial conditions and an external excitation 共Ogata numerator of G共z兲. A value of z that renders the denominator of
2002兲. Referring to Fig 1共a兲, which is a plot showing poles in the the transfer function in Eq. 共7兲 to be zero is defined as a “pole” of
complex s domain, stable oscillatory behavior occurs in a system the discrete transfer function. The stability of a discrete system is
whose poles are in the left-half plane near, or on the imaginary also determined by its pole positions. When the pole of a discrete
axis 共underdamped, with 0 ⱕ ⬍ 1兲, whereas stable exponential system is located on or inside the unit circle of the complex plane,
decay occurs when the poles are on the negative real axis 共over- the system will be stable; otherwise the system is unstable 共Ogata
damped, with ⱖ 1兲. Unstable behavior occurs in a system when 1995兲. The stable and unstable regions of poles for a discrete
the poles have a positive real part, i.e., are located in the right-half system are illustrated in Fig. 1共b兲.
plane. Poles in the right-half plane are associated with negative For the Newmark family of integration algorithms 共Newmark
damping, i.e., ⬍ 0. 1959兲, displacement, velocity, and acceleration at time step 共i兲
The temporally discretized form of Eq. 共1兲 can be written as and 共i + 1兲 are related by
冋 册
responds to the Newmark method with linear acceleration; and
共 , ␥兲 = 共0 , 1 / 2兲 which corresponds to the Newmark explicit m B11 B12
method. B= 共13b兲
1 + 2␥n⌬t + n⌬t B21 B22
2 2
Ramirez 共1992兲 and Mugan and Hulbert 共2001b兲 showed that
the discrete transfer function corresponding to the Newmark fam-
ily of integration algorithms can be written as ui = 冋 册
Fi
Fi+1
共13c兲
冋册 冋 册
X共z兲 n2z2 + n1z + n0
G共z兲 = = 共9兲 xi xi+1
F共z兲 d2z2 + d1z + d0 Xi = and Xi+1 = 共13d兲
ẋi ẋi+1
The values of the coefficients of G共z兲 are given in Table 1.
For the 共i + 1兲th time step, the discrete transfer function can be The coefficients of matrices A and B are given in Table 2.
written in the following form of a difference equation The eigenvalues of the amplification matrix A in Eq. 共12兲 can
be obtained by solving the eigenvalue problem
− d1 − d0 n2 n1 n0
xi+1 = xi + xi−1 + Fi+1 + Fi + Fi−1 共10兲 兩A − I兩 = 0 共14兲
d2 d2 d2 d2 d2
where I⫽identity matrix of the same dimension of A and
For Eq. 共10兲 to represent an explicit integration algorithm, n2
⫽eigenvalue.
must be equal to zero. That is, the solution for xi+1 must be based
Substituting the elements of A from Table 2 into Eq. 共14兲 gives
on the equilibrium condition of Eq. 共4兲 associated with the exci-
tation at a time corresponding to the indices i and 共i − 1兲. 共1 + 2␥n⌬t + 2n⌬t2兲2 + 关共− 2 + 0.5 + ␥兲2n⌬t2
On that basis, from Table 1, the Newmark family of integra-
tion algorithms can therefore be written in an explicit form only if + 共2 − 4␥兲n⌬t − 2兴 + 关共 − ␥ + 0.5兲2n⌬t2
 = 0. As noted previously, the well known Newmark explicit + 共2␥ − 2兲n⌬t + 1兴 = 0 共15兲
method with zero algorithmic damping 共Hughes 1987兲 has the
integration parameters for  and ␥ equal to 0 and 1/2, Eq. 共15兲 is similar to the denominator in Eq. 共9兲, which means
respectively. that the eigenvalues of the amplification matrix are exactly the
The poles of the transfer function of Eq. 共9兲 can be obtained by same as the poles of the discrete transfer function. Hence the
solving the following quadratic equation properties of the poles of the discrete transfer function are equiva-
lent to those of the eigenvalues of the amplification matrix for the
d 2z 2 + d 1z + d 0 = 0 共11兲 same integration algorithm. Thus, in the development of a new
integration algorithm, the stability and accuracy of the integration
Eq. 共11兲 is called the characteristic equation of the discrete sys- algorithm will be ensured if proper stable pole values are assigned
tem. As mentioned previously, to ensure the stability of the inte- to its discrete transfer function.
gration algorithm, the poles of the corresponding discrete transfer
function have to be on or inside the unit circle of the discrete z
domain.
Discretization in Control Theory
The Newmark family of integration algorithms can also be
written in following matrix form 共Hughes 1987兲
Two methods are typically used in control theory to obtain the
response of a system at selected time steps. One is to sample the
Xi+1 = A · Xi + B · ui 共12兲
solution of the continuous system at desired time steps; the other
where A⫽amplification matrix; and B is related to the excitation; is to discretize the continuous system and solve the corresponding
ui⫽excitation; and Xi+1 and Xi⫽state vector values at the ith and discrete system. Different discretization methods have been de-
共i + 1兲th time step, respectively. A, B, ui, Xi and Xi+1 are defined veloped in control engineering, such as the zero-order hold
as method and the first-order hold method. The values of the poles of
z = es⌬t 共16兲
where ⌬t⫽sample period and can be taken to be the same as time
step used in the integration algorithm. Based on Eq. 共16兲, a stable
continuous pole located in the left-half plane in the s domain
corresponds to a stable discrete pole located inside the unit circle Fig. 2. Comparison of first-order Pade approximation with exact
in the z domain. An oscillatory pole located on the imaginary axis mapping
in the s domain corresponds to a discrete pole located on the unit
circle in the z domain; an unstable continuous pole located in the
right-half plane of the s domain corresponds to an unstable dis- that the stability of the continuous pole is also conserved during
crete pole located outside the unit circle in the z domain. By using the discretization using Eq. 共17兲.
Eq. 共16兲, the stability of the poles in the s domain is maintained in The error due to the first-order approximation for z can be
the process of discretization. studied by plotting the difference between Eqs. 共16兲 and 共17兲 for
Since the use of an exponential function can make the discreti- negative real values of s · ⌬t, as shown in Fig. 2. It is shown to be
zation difficult, a first-order Pade approximation of the exponen- relatively small when s · ⌬t is larger than −0.5.
tial function is often used instead, where Using Eq. 共17兲, the continuous poles from Eq. 共3兲 can be trans-
formed to the z domain, where
1 + s · ⌬t/2
z⬇
1 − s · ⌬t/2
共17兲 2 + 共− ⫾ i冑1 − 2兲n⌬t
z1,2 = 共18兲
2 − 共− ⫿ i冑1 − 2兲n⌬t
Eq. 共17兲 is also called “Tustin’s method” in control theory 共Fran-
klin et al. 2002兲. For an s value with a nonpositive real part, Eq. Applying the discretization rule of Eq. 共17兲 to the continuous
共17兲 will give a value of z inside or on the unit circle. This means transfer function in Eq. 共2兲 results in
which can be shown to be the same as z1,2 from Eq. 共18兲. Fig. 3
shows the spectral radius of the complex conjugate poles p1 and
p2 from Eq. 共20兲 as a function of n⌬t for various values of the
damping ratio less than 1.0. From Fig. 3 it can be observed that
for increasing values of ⌬t the spectral radius remains less
than one 共i.e., inside the unit circle in the z domain兲 for the cases
when ⱖ 0, which shows that G共z兲 from Eq. 共19兲 is uncondition-
ally stable when ⱖ 0. This is consistent with the statement made
earlier that the stable poles of G共s兲 of Eq. 共2兲 lie in the left-hand
plane. Fig. 3. Spectral radius 共兲 for different damping ratios
Numerator Denominator The diagonal matrices in Eq. 共28兲 are defined as M* = ⌽TM⌽,
C* = ⌽TC⌽, and K* = ⌽TK⌽.
n2 0 d2 m
The poles of a MDOF structure for each mode of vibration are
n1 ␣2⌬t2 d1 共␣22n⌬t2 + 2␣1n⌬t − 2兲m
determined from the following expression in the modal coordi-
n0 共␣1 − ␣2兲⌬t2 d0 共共␣1 − ␣2兲2n⌬t2 − 2␣1n⌬t + 1兲m nate system
miis2i + ciisi + kii = 0 共29兲
4
␣1 = ␣2 = 共25兲 where mii, cii, and kii⫽ith diagonal term of the diagonalized mass
4 + 4n⌬t + 2n⌬t2 matrix M*, damping C*, and stiffness matrix K*, respectively,
The accuracy of the proposed CR algorithm is discussed later. and si⫽pole of the ith mode.
Downloaded from ascelibrary.org by Purdue University Libraries on 07/24/17. Copyright ASCE. For personal use only; all rights reserved.
The proposed CR algorithm remains unconditionally stable when For each mode of vibration of a MDOF structure, two complex
applied to nonlinear structures with a softening behavior, for the conjugate poles are obtained from the solution to Eq. 共29兲. The
poles for the algorithm can be shown to remain on the unit circle solution of the poles of Eq. 共29兲 in the modal coordinate system
in the discrete z domain as the structural natural frequency n can be written in the following matrix form, where
goes to zero due to a softening of the structure 共Chen 2007兲. If a M * · S *2 + C * · S * + K * = 0 共30兲
different form of the variation in either the velocity or displace-
ment over a time step were selected than that appearing in Eqs. In Eq. 共30兲, S*⫽diagonal matrix of dimension n ⫻ n, with each
共24a兲 and 共24b兲, but where Eqs. 共24a兲 and 共24b兲 both remain diagonal term representing poles of a mode in the modal coordi-
explicit expressions 共i.e., xi+1 and ẋi+1 are a function of the previ- nate system. The solution of S* can be written as
ous time step兲, the integration algorithm would remain uncondi- 1
tionally stable by assigning the poles from Eq. 共18兲. However, the S* = 2 · 共M*兲−1 · 关− C* ⫾ i · 共4 · M* · K* − C*2兲1/2兴 共31兲
integration ␣1 and ␣2 would be different than that given by Eq. Premultiplying and postmultiplying Eq. 共30兲 by 共⌽ 兲 and 共⌽兲−1,T −1
共25兲. respectively, leads to the characteristic equation for the MDOF
structure in the original coordinate system
Accuracy Analysis
Acknowledgments
References
Fig. 6. Period elongation for Chang’s algorithm and the proposed CR
algorithm, and comparison with Newmark family of integration
Chang, S. Y. 共1999兲. “An unconditionally stable explicit algorithm in time
algorithms
history analysis.” Rep. No. NCREE-99-001, National Center for Re-
search on Earthquake Engineering, National Taiwan Univ., Taipei,
eration, as well as the proposed CR algorithm and Chang’s Taiwan.
Chang, S. Y. 共2002兲. “Explicit pseudodynamic algorithm with uncondi-
algorithm, have no algorithmic damping and therefore no ampli-
tional stability.” J. Eng. Mech., 128共9兲, 935–947.
tude decay. The period elongation shown in Fig. 6 that occurs in
Chen, C. 共2007兲. “Development and numerical simulation of a hybrid
the CR algorithm is the same as that in Chang’s algorithm and the effective force testing method.” Ph.D. dissertation, Lehigh Univ.,
Newmark method with constant acceleration, and larger than that Bethlehem, Pa.
which occurs in the Newmark methods in both explicit form and Chopra, A. K. 共2001兲. Dynamics of structures: Theory and applications
with linear acceleration. to earthquake engineering, 2nd Ed., Prentice-Hall, Upper Saddle
One can reduce the PE that occurs in the CR algorithm by River, N.J.
selecting a different set of poles other than those from Eq. 共18兲. Franklin, G. F., Powell, J. D., and Naeini, A. E. 共2002兲. Feedback control
However, the unconditional stability property of the algorithm of dynamic system, 4th Ed., Prentice-Hall, Upper Saddle River, N.J.
may be impaired. For example, if the poles from the discrete Hilber, H. M., Hughes, T. J. R., and Taylor, R. L. 共1977兲. “Improved
transfer function for the Newmark explicit method had been used numerical dissipation for time integration algorithms in structural me-
to develop the CR algorithm, the result would be ␣1 = 1 and ␣2 chanics.” Earthquake Eng. Struct. Dyn., 5共3兲, 283–292.
= 1 / 2. In effect, the CR algorithm in this case would be identical Hughes, T. J. R. 共1987兲. The finite-element method: Linear static and
to the Newmark explicit method, which is conditionally stable. dynamic finite-element analysis, Prentice-Hall, Englewood Cliffs, N.J.
Mugan, A. 共2003兲. “Discrete equivalent time integration methods for
transient analysis.” Int. J. Numer. Methods Eng., 57共14兲, 2043–2075.
Mugan, A., and Hulbert, G. M. 共2001a兲. “Frequency domain analysis of
Summary and Conclusions time integration methods for semidiscrete finite-element equations,
Part I. Parabolic problems.” Int. J. Numer. Methods Eng., 51共3兲,
A discrete transfer function approach is used to develop explicit 333–350.
unconditionally stable integration algorithms for linear structural Mugan, A., and Hulbert, G. M. 共2001b兲. “Frequency domain analysis of
dynamics problems. The poles of the discrete transfer function are time integration methods for semidiscrete finite-element equations.
shown to be the same as the eigenvalues of the amplification Part II: Hyperbolic and parabolic-hyperbolic problems.” Int. J. Numer.
matrix for the integration algorithm. By applying the mapping Methods Eng., 51共3兲, 351–376.
rule of poles using the first-order Pade approximation, a new ex- Newmark, N. M. 共1959兲. “A method of computation for structural dy-
plicit integration algorithm called the CR algorithm is developed. namics.” J. Engrg. Mech. Div., 85共3兲, 67–94.
Both linear elastic SDOF and MDOF cases are discussed. Ogata, K. 共1995兲. Discrete-time control systems, 2nd Ed., Prentice-Hall,
The proposed method offers a viable means for developing Englewood Cliffs, N.J.
Ogata, K. 共2002兲. Modern control engineering, 4th Ed., Prentice-Hall,
integration algorithms with desirable characteristics for solving
Upper Saddle River, N.J.
structural dynamics problems. By selecting stable poles that are Ramirez, M. R. 共1992兲. “The numerical transfer function for time inte-
based on the continuous transfer function for a SDOF system with gration analysis.” Proc., New Methods in Transient Analysis, ASME,
a positive viscous damping ratio, the explicit CR algorithm devel- New York, PVP-Vol. 246/AMD-Vol. 143, 79–85.
oped in this paper is unconditionally stable and exhibits properties Wilson, E. L. 共1968兲. “A computer program for the dynamic stress analy-
of zero algorithmic damping. The period elongation is identical to sis of underground structures.” SESM Rep. No. 68-1, Division of
that in the Newmark method with constant average acceleration. Structural Engineering and Structural Mechanics, Univ. of California,
The CR algorithm, besides being unconditionally stable, has the Berkeley, Calif.
advantage of being based on expressions for displacement and Wood, W. L. 共1990兲. Practical time-stepping schemes, Clarendon, Ox-
velocity that are both explicit in form. These features make the ford, U.K.