Numerical Integration in Structural Dynamics

CEE 541. Structural Dynamics

Department of Civil & Environmental Engineering
Duke University
Henri P. Gavin
Fall 2020

Introduction
A damped structural system subjected to dynamic forces and possibly experiencing
nonlinear material behavior is modeled by

M ẍ(t) + C ẋ(t) + Kx(t) + R(x(t), ẋ(t)) = f ext (t), (1)

where x is a vector of displacements of structural coordinates, M is a positive definite

mass matrix, C is a non-negative definite damping matrix, and K is a non-negative definite
stiffness matrix. The nonlinear restoring forces are given in R(x, ẋ) and f ext (t) is a vector
of external dynamic loads. At any point in time, t = ti+1 = (i + 1)h, we may solve for the
accelerations in terms of the displacements, velocities, and the applied forces.

ẍ(ti+1 ) = −M −1 [C ẋ(ti+1 ) + Kx(ti+1 ) + R(x(ti+1 ), ẋ(ti+1 )) − f ext (ti+1 )]. (2)

Given values for accelerations, velocities, displacements, and applied forces at time ti , if we
can extrapolate the velocities and displacements forward in time by a time step h to time
ti+1 = ti + h, then we may compute the acceleration ẍ(ti+1 ) using equation 2. The various
numerical integration algorithms described in this document1 2 differ primarily in the manner
in which x(ti+1 ) and ẋ(ti+1 ) are computed from x(ti ), ẋ(ti ), ẍ(ti ), f ext (ti ), and f ext (ti+1 ).
There are two general classifications of numerical integration methods: explicit and
implicit. In explicit methods, displacements and velocities at ti+1 can be determined in
closed form from displacements, velocities, accelerations, at ti , and from external forcing at
ti and, potentially, ti+1 . For structural systems with linear elastic stiffness and linear viscous
damping, such discrete-time systems may be written
" # " #
x(ti+1 ) x(ti )
=A + B f ext (ti ) (3)
ẋ(ti+1 ) ẋ(ti )

where A is a 2n × 2n discrete time dynamics matrix which depends upon M , C, K, the

time step, h, and some algorithmic parameters. Implicit methods involve the solution of a
set of nonlinear algebraic equations at each time step. For example, the displacements and
velocities at time ti+1 , [x(ti+1 ), ẋ(ti+1 )], are determined from the roots of a nonlinear equation
in terms of [x(ti+1 ), ẋ(ti+1 )]. Explicit numerical methods are typically more computationally-
efficient than implicit methods.
1
Clough, R. and Penzien, J., Dynamics of Structures, 2nd ed., McGraw-Hill, 1993.
2
Hanson, R.D., CE 611: Structural Dynamics, class notes, The University of Michigan, 1988
2 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

Accuracy and Stability

Numerical methods for integrating equations of motion are assessed and evaluated in
terms of their accuracy and stability. In general, accuracy and stability depend upon the
ratio of the time step, h, to the shortest natural period in the system model. For a system
with many coordinates (n > 103 ), the shortest natural period can be much shorter than
the fundamental natural period, Tn /T1 > 104 . Typically, responses of the highest several
modes of a numerical model are physically meaningless, should be insignificantly small, but
are potentially lightly-damped, and can dominate the errors in numerical integration. The
explicit numerical methods described in these notes can artificially add numerical damping to
suppress instabilities of the higher mode responses. Implicit numerical integration methods
are unconditionally stable.

The Central Difference Method

The central difference approximations for the first and second derivatives are
1
ẋ(ti ) = ẋi ≈ (xi+1 − xi−1 ) , (4)
2h
1
ẍ(ti ) = ẍi ≈ 2 (xi+1 − 2xi + xi−1 ) . (5)
h
Substituting approximations (4) and (5) into equation (1), and re-arranging terms, leads to
1 1 2 1 1
     
M + C xi+1 = 2 M − K xi + − 2 M + C xi−1 + fiext , (6)
h2 2h h h 2h
which may be written A0 xi+1 = A1 xi + A2 xi−1 + fiext . As long as M or C is positive definite,
the matrix A0 is also positive definite and may be factorized prior to the iterative solution
for xi at each time step.
The numerical stability of the central difference method depends on the choice of the
time step interval, h. To obtain a stable solution, h < Ti /π, where Ti is the shortest natural
period of the structural system. If the central difference method is used with a time step
larger than Ti /π the solution will increase exponentially.
A step-by-step procedure for the central difference method may be written as3 :

1. Input the mass, M , damping, C, stiffness, K, matrices and the time step interval h.
h2
2. Initialize x0 , ẋ0 , ẍ0 and compute x1 = x0 − hẋ0 + 2 0
ẍ .
3. Form the matrices A0 , A1 and A2 , and triangularize A0 using LDLT factorization.
4. for i = 1 to N
(a) solve A0 xi+1 = A1 xi + A2 xi−1 + fiext for xi+1 using LDLT back-substitution.
(b) calculate ẋi and ẍi using equations (4) and (5), if required.
(c) write results (ti , xi , ẋi , ẍi ) to a data file.
3
K.J. Bathe, Finite Element Procedures in Engineering Analysis, Prentice-Hall, 1982, pp. 439–449, 499–
506.

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 3

The Implicit Linear Acceleration Method

Consider the Taylor series expansions for displacement, velocity, and acceleration:

h2 h3 ... h4 ....
x(ti+1 ) = xi+1 = xi + hẋi + ẍi + xi+ x i + ··· (7)
2! 3! 4!
h2 ... h3 ....
ẋ(ti+1 ) = ẋi+1 = ẋi + hẍi + xi+ x i + ··· (8)
2! 3!
... h2 ....
ẍ(ti+1 ) = ẍi+1 = ẍi + h x i + x i + ··· (9)
2!
Rearranging equation (9),

... h2 ....
h x i = ẍi+1 − ẍi − x i − ··· , (10)
2!
and substituting equation (10) into equations (7) and (8) results in
!
h2 h2 h2 .... h4 ....
xi+1 = xi + hẋi + ẍi + ẍi+1 − ẍi − x i − ··· + x i + ··· , (11)
2 6 2! 4!
!
h h2 .... h3 ....
ẋi+1 = ẋi + hẍi + ẍi+1 − ẍi − x i − ··· + x i + ··· (12)
2 2! 3!

Under the assumption that acceleration ẍ(t) varies linearly in time between times ti and
time ti + h the fourth and higher time-derivatives are zero. The resulting finite difference
relations are precisely
h2
xi+1 = xi + hẋi + (ẍi+1 + 2ẍi ), (13)
6
and
h
ẋi+1 = ẋi + (ẍi+1 + ẍi ). (14)
2
These relationships are implicit because ẍi+1 needs to be determined in order to find xi+1 and
ẋi+1 , but ẍi+1 can not be found without knowing xi+1 and ẋi+1 . Note that the substitutions
above, have eliminated the third time-derivative of x, and that the method is accurate to
....
within h4 x (which is zero if ẍ is linear in time).
This is called the linear acceleration method because the third time derivative of x has
been eliminated. If the rate of change of acceleration within a time-step is truly constant,
then the approximation of truncating the Taylor series at the fourth-order term does not
affect the accuracy of the solution.

cbnd H.P. Gavin January 3, 2025

4 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

The Implicit Linear Acceleration Method,

Made Explicit for Linear Structural Dynamics
Recall that at time ti+1 we can satisfy the equations of motion, by calculating the
acceleration with equation (2). Substituting equations (13) and (14) for the displacements
and velocities at time ti+1 into equation (2),

M ẍi+1 + C{ẋi + (h/2)(ẍi + ẍi+1 )} + K{xi + hẋi + (h2 /6)(ẍi+1 + 2ẍi )} = fi+1
ext
, (15)

where any nonlinearities, R(x, ẋ), are assumed to be negligible. Collecting similar derivatives
of x, " # " #
h h2 ext h h2
M + C + K ẍi+1 = fi+1 − Kxi − [C + hK]ẋi − C + K ẍi . (16)
2 6 2 3
and re-arranging,
" # " #
h h2 h h2 ext
M + C + K ẍi+1 = − C − K ẍi − Kxi − C ẋi − hK ẋi + fi+1 . (17)
2 6 2 3

Recall from the equations of motion, that

M ẍi − fiext = −Kxi − C ẋi . (18)

Substituting equation (18) into (17), we obtain the closed-form linear acceleration recurrence
relations for structural dynamics simulation.
" # " #
h h2 h h2 ext
M + C + K ẍi+1 = M − C − K ẍi − hK ẋi + fi+1 − fiext . (19)
2 6 2 3

h
ẋi+1 = ẋi + [ẍi+1 + ẍi ] . (20)
2
and
h2
xi+1 = xi + hẋi + [ẍi+1 + 2ẍi ] . (21)
6
These relationships are now explicit because ẍi+1 can be determined from the current response
values (xi and ẋi ), the current dynamic load fiext , and the next dynamic load fi+1
ext
. Note
that within each time step, the dynamic equations of equilibrium are satisfied both at time
ti and at time ti+1 .

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 5

The Newmark-β method — incremental formulation

The finite difference approximations for the Newmark-β method are
1
  
xi+1 ≈ xi + hẋi + h2 − β ẍi + β ẍi+1 , (22)
2
and
ẋi+1 ≈ ẋi + h [(1 − γ)ẍi + γ ẍi+1 ] . (23)
If β = 1/4 and γ = 1/2 the Newmark-β method is implicit and unconditionally stable. In
this case the acceleration within the time interval t ∈ [ti ti+1 ) is presumed to be constant. If
β = 1/6 and γ = 1/2 the Newmark-β method is identical to the linear acceleration method.
If β = 0 and γ = 1/2 the Newmark-β method is identical to the central difference method.
For linear structural dynamics, if 2β ≥ γ ≥ 1/2, then the Newmark-β method is stable
regardless of the size of the time-step, h. The Newmark-β method is conditionally stable if
γ < 1/2. For γ = 1/2 the Newmark-β method is at least second-order accurate; it is first
order accurate for all other values of γ.
Formulating the finite difference relationships in terms of the increments of displace-
ment (δxi = xi+1 −xi ), velocity (δ ẋi = ẋi+1 − ẋi ), acceleration (δẍi = ẍi+1 − ẍi ), and nonlinear
restoring force (δRi = R(xi+1 , ẋi+1 ) − R(xi , ẋi ))
1 1 1
δẍi = 2
δxi − ẋi − ẍi , (24)
βh βh 2β
and !
γ γ γ
δ ẋi = δxi − ẋi + h 1 − ẍi . (25)
βh β 2β
Now, satisfying incremental equilibrium over the time step, h,
ext
M δẍi + Cδ ẋi + Kδxi + δRi = fi+1 − fiext = δfiext , (26)
re-grouping terms, and solving for the increment in displacements,
" #
6 3 h 6
   
2
M + C + K δxi = δfiext − δRi + 3M + C ẍi + M + 3C ẋi , (27)
h h 2 h
where β = 1/6 and γ = 1/2. Solving this linear system of equations4 for δxi , the displace-
ments are updated with,
xi+1 = xi + δxi , (28)
the velocities are updated with
h 3
ẋi+1 = −2ẋi − ẍi + δxi , (29)
2 h
and the accelerations satisfy the equations of motion,
ẍi+1 = −M −1 [C ẋi+1 + Kxi+1 + R(xi+1 , ẋi+1 ) − fi+1
ext
]. (30)
Similar relations may be found for other values of β and γ. Note that in this incremental
formulation one needs to obtain the inverse of the mass matrix, M , and the inverse of the
matrix [6M/h2 + 3C/h + K].
4
If the incremental nonlinear restoring forces δRi are not negligible, equation (27) is solved using the
Newton-Raphson method, equation (48).

cbnd H.P. Gavin January 3, 2025

6 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

The HHT-α method

The HHT-α method5 is a generalization of the Newmark-β method and reduces to the
Newmark-β method for α = 0. The HHT-α method adopts the finite difference equations of
the Newmark-β method, (equations (22) and (23)). The equations of motion are modified,
however, using a parameter α, which represents a numerical lag in the damping, stiffness,
nonlinear, and external forces.
ext
M ẍi+1 + (1 − α)C ẋi+1 + αC ẋi + (1 − α)Kxi+1 + αKxi = (1 − α)fi+1 + αfiext . (31)

Equation (31) is analogous to equation (26) with the added condition that the acceleration
is constant within the interval t ∈ [ti ti+1 ). If

0 ≤ α ≤ 1/3 (32)
β = (1 + α)2 /4 , and (33)
γ = 1/2 + α , (34)

the HHT-α method is at least second-order accurate and unconditionally stable. The HHT-α
method is useful in structural dynamics simulations incorporating many degrees of freedom,
and in which it is desirable to numerically attenuate (or dampen-out) the response at high
frequencies. Increasing α decreases the response at frequencies above 1/(2h), provided that
β and γ are defined as above.
Substituting the Newmark-β finite difference relationships into equation (31),

M ẍi+1 + (1 − α)C{ẋi + h[(1 − γ)ẍi + γ ẍi+1 ]} + αC ẋi

+ (1 − α)K{xi + hẋi + h2 [(1/2 − β)ẍi + β ẍi+1 ]} + αKxi
ext
= (1 − α)fi+1 + αfiext , (35)

and grouping terms, we obtain

[M + h(1 − α)γC + h2 (1 − α)βK]ẍi+1 + [h(1 − α)(1 − γ)C + h2 (1 − α)(1/2 − β)K]ẍi

+ [C + h(1 − α)K]ẋi
+ Kxi
ext
= (1 − α)fi+1 + αfiext . (36)

These are a set of linear equations for ẍi+1 in terms of ẍi , ẋi , xi , the external dynamic load,
and the structure’s mass, damping, and stiffness. Equations (22), (23), and (36) provide the
recurrence relationship for the HHT-α method.

5
Hughes, T.J.R., Analysis of Transient Algorithms with Particular Reference to Stability Behavior. in
Computational Methods for Transient Analysis, North-Holland, 1983, pp. 67–155.

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 7

The HHT-α method — incremental formulation

The HHT-α method may be formulated in an incremental form, as was done with
the Newmark-β method in the previous section. Again, formulating the finite difference
relationships in terms of the increments of displacement (δxi = xi+1 − xi ), velocity (δ ẋi =
ẋi+1 − ẋi ), acceleration (δẍi = ẍi+1 − ẍi ), and nonlinear restoring force, and difference between
equation (31) (evaluated at time ti+1 ) and the same expression evaluated at time ti , results
in

M δ ẍi +(1−α)Cδ ẋi +αCδ ẋi−1 +(1−α)Kδxi +αKδxi−1 +(1−α)δRi +αδRi−1 = (1−α)δfiext +αδfi−1
ext
.
(37)
Now, substituting in expressions for δ ẋ and δẍ at times ti and ti−1 , and grouping terms
results in
" # " #
1 γ 1 γ
2
M + (1 − α) C + (1 − α)K δxi = M + (1 − α) C ẋi
βh βh βh β
" ! #
1 γ
+ M − (1 − α)h 1 − C ẍi
2β 2β
" #
γ
− α C + K δxi−1
βh
!
γ γ
+ α C ẋi−1 − αh 1 − C ẍi−1
β 2β
− (1 − α)δRi − αδRi−1
+ (1 − α)δfi + αδfi−1 . (38)

If R(x, ẋ) = 0, these are a set of linear equations for δxi in terms of ẋi , ẍi , δxi−1 , ẋi−1 ,
ẍi−1 , δxi−1 , the external dynamic load, and the structure’s mass, damping, and stiffness.
Equations (22), (23), and (38) provide the recurrence relationship for the HHT-α method
in incremental form. Solving this linear system of equations6 for δxi , the displacements are
updated with,
xi+1 = xi + δxi , (39)
the velocities are updated with
! !
γ γ γ
ẋi+1 = 1− ẋi + δxi + h 1 − ẍi . (40)
β βh 2β

and the accelerations satisfy the HHT-α form of the equations of motion,
h i
ẍi+1 = −M −1 (1 − α)C ẋi+1 + αC ẋi + (1 − α)Kxi+1 + αKxi − (1 − α)fi+1
ext
− αfiext . (41)

This algorithm is explicit regardless of the values of α, β, or γ.

6
If the incremental nonlinear restoring forces δRi are not negligible, equation (38) is solved using the
Newton-Raphson method, equation (48).

cbnd H.P. Gavin January 3, 2025

8 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

Numerical Example
The simple linear structural model shown below
111
000
000
111
1 2 3 111
000
000
111
000 200 kN/mm
111
000
111 200 kN/mm 200 kN/mm 000
111
000
111
000
111 000
111
000
111
000
111 000
111
000
111
000
111 000
111
000
111
000
111 000
111
000
111
000
111 1 Mt 1 Mt 1 Mt 000
111
000
111
000
111 000
111
000
111
000
111 000
111
000
111 000
111
000
111
000
111 000
111
000
111
000
111 000
111
000
111 0.15 kN/mm/s 111
000
000 0.35 kN/mm/s
111 0.20 kN/mm/s 0.20 kN/mm/s 000
111

is described by the following stiffness, damping, and mass matrices (units kN,mm,s):
     
400 −200 0 1 0 0 0.55 −0.20 0
K =  −200 400 −200  C =  −0.20
M = 0 1 0  0.40 −0.20 
   
 
0 −200 200 0 0 1 0 −0.20 0.35
(42)
This system has natural periods of 1.00 s, 0.36 s, and 0.25 s, with about 1.5% critical damp-
ing in each mode. Free acceleration responses to initial conditions of x(0) = [0, 0, 0]T and
ẋ(0) = [1, 1, 1]T using the HHT-α method with time steps of 0.001 s (black solid) and 0.1 s
(blue dash-dot), and values of α of 0.0 (left) and 0.1 (right) are shown below.
10 10

α = 0.00 α = 0.10

5 5
accel #1, mm/s2

accel #1, mm/s2

0 0

-5 -5

-10 -10
0 2 4 6 8 10 0 2 4 6 8 10
time, s time, s

For the α = 0 cases, the response computed with h = 0.1 s has a magnified short-period
response as compared to the h = 0.001 s case. For α = 0.1 and h = 0.1 s, the response
is more heavily damped especially the response at the shorter periods. Also, the responses
simulated with α = 0.1 and h = 0.1 s has a longer period than the system’s true fundamental
period of 1.0 s. Note that the linear acceleration method (α = 0, β = 1/6, γ = 1/2) would
be numerically unstable for this system with a time step of 0.1 s.
A more formal analysis of numerical damping, stability, and period lengthening effects
in explicit numerical integration methods can be carried out by examining the eigenvalues
of the discrete time dynamics matrix, A, in equation (3).

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 9

Simulation of SDOF Bilinear Hysteretic Behavior

Bilinear hysteretic behavior may be simulated using linear dynamic models. The basic
idea is that each branch of the hysteresis may be described by an equation of the form
R = K(x − d), where the restoring force is zero when x = d and K will take the value of the
pre-yield stiffness or the strain hardening stiffness. The procedure is as follows:

1. Initialize K using the pre-yield stiffness Khi . Initialize x0 , ẋ0 , ẍ0 , f0ext , and d to be zero
at time t0 = 0. The value d is the equilibrium displacement of the bilinear hysteretic
system.
ext
2. At time t = ti+1 , read the value of the external forcing fi+1 , solve equation (27) for
δxi , and calculate xi+1 and ẋi+1 using equations (28) and (29).
3. Compute the next acceleration ẍi+1 from equilibrium
ẍi+1 = −M −1 [C ẋi+1 + K(xi+1 − d) − fi+1
ext
]. (43)
4. Compute the yielding force level from
f yield (xi+1 , ẋi+1 ) = Klo xi+1 + (Khi − Klo ) xy sgn(ẋi+1 ) (44)
5. Check for yielding:
(a) If K = Khi and ẋi+1 > 0 and K(xi+1 − d) > f yield ,
then the system has just exceeded its positive yield force level.
Set K = Klo and d = (1 − Khi /Klo )xy .
(b) If K = Khi and ẋi+1 < 0 and K(xi+1 − d) < f yield ,
then the system has just exceeded it’s negative yield force level.
Set K = Klo and d = (Khi /Klo − 1)xy .
6. Check for load reversal:

If K = Klo and ẋi+1 ẋi < 0,

Set K = Khi and d = xi+1 − (Klo /Khi )(xi+1 − d)
7. If any of the checks in steps 5 or 6 were true, then correct the velocity and acceleration
(ẋi+1 , ẍi+1 ) using equations (27), (29), and (43) with a very small time step (h/104 ),
over which the external forcing is assumed to be constant (δfiext = 0).
8. Record xi+1 , ẋi+1 , ẍi+1 , and fi+1 − M ẍi+1 . Increment the time step. i := i + 1.
9. Go back to step 2 and repeat until the end of the time history.
.
R(x(t),x(t)) yield .
f(x, x > 0)
K hi
K lo xy x y (K hi / K lo− 1)
d d
x y (1 − Khi / Klo) x(t)
yield .
f(x, x < 0)

cbnd H.P. Gavin January 3, 2025

10 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

Newton-Raphson Iterative Solution for Nonlinear Behavior

Equations (27) or (38) may be written K̃ δxi = f˜i . If the nonlinear restoring force
increment δRi is not negligible, then f˜ depends on δx and δ ẋ. For the case of the linear
acceleration method (α = 0, β = 1/6, γ = 1/2),
6 3
 
K̃ = 2 M + C + K , (45)
h h
and
" #
h 6
 
f˜i (δxi ) = δfiext − R(xi + δxi , ẋi + δ ẋi ) + R(xi , ẋi ) + 3M + C ẍi + M + 3C ẋi . (46)
2 h
The Newton-Raphson algorithm7 is an efficient method to solve the nonlinear equations
K̃ δxi = f˜i (δxi ) for the incremental displacement δxi . The algorithm is iterative and proceeds
as follows:
(0)
1. The initial value for δxi is denoted δxi and is arbitrarily set to zero. The corresponding
value for f˜i is
" #
h 6
 
(0)
f˜i (δxi ) = δfiext + 3M + C ẍi + M + 3C ẋi (47)
2 h

2. The Newton-Raphson recurrence relation is simply

(n+1) (n)
δxi = K̃ −1 f˜i (δxi ) (48)
where
(n) (0) (n) (n)
f˜i (δxi ) = f˜i (δxi ) − R(xi + δxi , ẋi + δ ẋi ) + R(xi , ẋi ), (49)
and
(n) 3 (n) h
δ ẋi = δxi − 3ẋi − ẍi . (50)
h 2
3. Equation (48) is iterated upon until the incremental displacement δxi converges, i.e.,
(n+1) (n)
||δxi − δxi || < ϵ.

The convergence of this form of the Newton-Raphson method depends on the local
smoothness of R(x, ẋ). Convergence can be improved, for a particular time step, by making
h smaller.
Note that K̃ is strongly positive definite; it does not depend on δxi ; and it is inverted
or factorized only once at the beginning of the simulation. If convergence is problematic,
then a set of K̃ matrices may be formed, some with small time steps, and factorized prior
to the simulation.
As an alternative to Newton-Raphson iterations, the approximation δRi ≈ δRi−1 re-
sults in an explicit formulation for the simulation of the response of nonlinear systems, and
can significantly reduce computational time, but with some loss of accuracy.
7
Press, W.H., et al., Numerical Recipes, Cambridge Univ. Press, 1993. http://www.nr.com/

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 11

Systems of First Order Ordinary Differential Equations

The second order ordinary differential equations of motion for a structural system are
M r̈(t) + C ṙ(t) + Kr(t) + R(r(t), ṙ(t)) = f ext (t), (51)
where we have re-named the displacement vector x(t) of equation (1) with r(t). This set
of n second order ordinary differential equations may be written as a set of 2n first order
ordinary differential equations, as follows:
" # " #" # " # " #
d r 0N ×N IN ×N r 0N ×1 0N ×N
= + + f ext (t). (52)
dt ṙ −M K −M −1 C
−1
ṙ −M −1 R(r, ṙ) M −1

The vector [rT ṙT ]T is called the state vector; along with the external forces, it can completely
describe the state of the system at any point in time. State-space models are conventionally
written as
ẋ = Ax + g(x) + Bu, (53)
where x is the state vector, A is the dynamics matrix, g(x) contains any nonlinear terms in
the equations of state, B is the input matrix, and u is the input to the system. For a linear
systems, g(x) = 0 and the eigenvalues of A contain the complex poles of the system. For
under-damped dynamic systems the poles are complex-valued. The imaginary part of the
pole is the damped natural frequency, and the real part of the pole is the negative of the
natural frequency times the damping ratio.
A convenient feature of state-variable formulations is that first order systems can be
easily combined. If components of the structural system have first-order dynamics (for
example, visco-elasticity or Bouc-Wen hysteresis), then the first order dynamics of those
components can be appended to the state vector, and the simulation of the second order
structural system coupled with the first order structural components can proceed in parallel.
The state variable formulation is the most general description of a dynamic system,
and many methods exist for the simulation of state-variable models. A set of these methods
has been implemented in Matlab. See, for example, the Matlab command ode45 .
For large second-order systems (n > 1000), however, the HHT-α method is typically
more efficient in terms of memory and speed.

A Fixed-Step Fourth-Order Runge-Kutta Solver

In the computation of transient response analyses of linear or nonlinear systems, it
is common for the external forcing to be sampled at uniformly-spaced increments in time,
and to desire the response to be computed at these same points in time. If the sample
interval h = ∆t is less than one-tenth of the shortest natural period involved in the system
(over the course of the simulation) then a fixed time-step Runge-Kutta solver can compute a
solution quickly, and with reasonable accuracy.8 The most famous of these solvers is fourth-
order accurate. The goal of the solver is to advance the solution, x, of a general system of
8
Press, W.H., et al., Numerical Recipes, Cambridge Univ. Press, 1993. Section 16.1 http://www.nr.com/

cbnd H.P. Gavin January 3, 2025

12 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

non-homogeneous first-order ordinary differential equations

ẋ(t) = f (t, x(t), u(t)) , x(0) = xo , (54)

from time t to time t+∆t. Given the state vector x(t), we endeavor to find the state x(t+∆t).
The solution scheme is illustrated in Figure 1. Within each time-step, the algorithm evaluates

. .
x2 = f ( t+dt/2, x(t)+x1 (dt/2) , u(t+dt/2) )

. . . .
x(t+dt) = x(t) + ( x1 + 2(x2+x3) + x4) (dt/6)
x(t+dt)

. .
x3 = f ( t+dt/2, x(t)+x2 (dt/2) , u(t+dt/2) )

x(t)
. . .
x1 = f ( t, x(t) , u(t) ) x4 = f (t+dt, x(t)+x3 dt , u(t+dt) )

t t + dt/2 t+dt

Figure 1. Solution Scheme of the Fourth Order Runge-Kutta Method.

ẋ four times.

ẋ1 = f ( t , x(t) , u(t)) (55)

!!
∆t ∆t ∆t
ẋ2 = f t + , x(t) + ẋ1 , u t + (56)
2 2 2
!!
∆t ∆t ∆t
ẋ3 = f t + , x(t) + ẋ2 , u t + (57)
2 2 2
ẋ4 = f ( t + ∆t , x(t) + ẋ3 ∆t , u(t + ∆t)) (58)

The desired solution, x(t + ∆t), is computed from the four state derivatives

∆t
x(t + ∆t) = x(t) + (ẋ1 + 2(ẋ2 + ẋ3 ) + ẋ4 ) , (59)
6
time is incremented, t = t + ∆t, and the solution marches on. The solution scheme is
implemented in the Matlab function ode4u.m9 . As long as ∆t is shorter than 0.45Tn where
Tn is the shortest natural period in the system, this constant time-step Runge-Kutta method
is stable.

9
http://www.duke.edu/∼hpgavin/ode4u.m

cbnd H.P. Gavin January 3, 2025

 Numerical Integration for Structural Dynamics 13
1 function [ time , x_sol , x_drv , y_sol ] = ode4u ( dxdt , time , x0 , u , cnstnt )
2 % [ time , x s o l , x d r v , y s o l ] = ode4u ( d x d t , time , x0 , u , c n s t n t )
3 %
4 % S o l v e a sy s te m o f nonhomogeneous o r d i n a r y d i f f e r e n t i a l e q u a t i o n s u s i n g t h e
5 % 4 t h o r d e r Runge−Kutta method .
6 %
7 % Input Variable Description
8 % −−−−−−−−−−−−−− −−−−−−−−−−−
9 % dxdt : a f u n c t i o n o f t h e form [ x d o t , y ] = d x d t ( t , x , u , c n s t n t )
10 % which p r o v i d e s t h e s t a t e d e r i v a t i v e g i v e n t h e time , t ,
11 % t h e s t a t e , x , e x t e r n a l f o r c i n g , u , and c o n s t a n t s
12 % a v e c t o r sy s te m o u t p u t s , y , may a l s o be r e t u r n e d
13 % time : 1−by−p v e c t o r o f time v a l u e s , u n i f o r m l y i n c r e a s i n g
14 % ( or d e c r e a s i n g ) , a t which t h e s o l u t i o n i s computed
15 % x0 : n−by−1 v e c t o r o f i n i t i a l s t a t e v a l u e s , a t time ( 1 )
16 % u : m−by−p m a t r i x o f s y st e m f o r c i n g d a t a f o r t h e ode
17 % sampled a t p o i n t s i n t h e 1−by−p v e c t o r time
18 % cnstnt : optional set of constants for the function dxdt
19 %
20 % Output V a r i a b l e Description
21 % −−−−−−−−−−−−−− −−−−−−−−−−−
22 % time : i s r e t u r n e d , un−changed .
23 % x sol : the solution to the d i f f e r e n t i a l equation at
24 % t i m e s g i v e n i n t h e 1−by−p v e c t o r time .
25 % x drv : the d e r i v i t i v e of the solution
26 % y sol : a d d i t i o n a l s y st e m o u t p u t s
27 %
28 % Henri Gavin , C i v i l and Environmental E n g i n e e r i n g , Duke U n i v s e r s i t y
29 % WH Press , e t a l , Numerical R e c i p e s i n C, 1992 , S e c t i o n 1 6 . 1
30
31 points = length ( time );
32 i f nargin < 5 , cnstnt = 0; end
33 i f nargin < 4 , u = zeros (1 , points ); end
34
35 [ dxdt1 , y1 ] = f e v a l ( dxdt , 0 , x0 , u (: ,1) , cnstnt ); % compute i n i t i a l o u t p u t
36
37 n = length ( x0 (:));
38 m = length ( y1 (:)); % number o f o u t p u t s
39
40 x_sol = NaN(n , points ); % memory a l l o c a t i o n f o r s t a t e h i s t o r y
41 x_drv = NaN(n , points ); % memory a l l o c a t i o n f o r s t a t e d e r i v a t i v e h i s t o r y
42 y_sol = NaN(m , points ); % memory a l l o c a t i o n f o r o u t p u t h i s t o r y
43
44 x_sol (: ,1) = x0 (:); % current state
45 x_drv (: ,1) = dxdt1 (:); % current rate
46 y_sol (: ,1) = y1 (:); % current output
47
48 [ ru , cu ] = s i z e ( u );
49 i f ( cu < points ) , u (: , cu +1: points ) = 0; end % pad u w i t h z e r o s t o l e n g t h o f t
50
51 f o r p = 1: points -1 % advance t h e s o l u t i o n from p t o p+1
52
53 t = time ( p );
54 dt = time ( p +1) - t ; % t h e time s t e p f o r t h i s i n t e r v a l
55 dt2 = dt /2; % h a l f o f t h e time s t e p
56
57 [ dxdt2 , y2 ] = f e v a l ( dxdt , t + dt2 , x0 + dxdt1 * dt2 , ( u (: , p )+ u (: , p +1))/2.0 , cnstnt );
58 [ dxdt3 , y3 ] = f e v a l ( dxdt , t + dt2 , x0 + dxdt2 * dt2 , ( u (: , p )+ u (: , p +1))/2.0 , cnstnt );
59 [ dxdt4 , y4 ] = f e v a l ( dxdt , t + dt , x0 + dxdt3 * dt , u (: , p +1) , cnstnt );
60
61 x0 = x0 + ( dxdt1 + 2*( dxdt2 + dxdt3 ) + dxdt4 ) * dt /6.0; % u p d a t e x0
62
63 [ dxdt1 , y1 ] = f e v a l ( dxdt , time ( p +1) , x0 , u (: , p +1) , cnstnt );
64
65 x_sol (: , p +1) = x0 (:); % current state
66 x_drv (: , p +1) = dxdt1 (:); % current rate
67 y_sol (: , p +1) = y1 (:); % current output
68
69 % i f max( a b s ( x0 ) ) > 1 e10 , b r e a k ; end % c l o s e enough t o a f l o a t i n g p o i n t e r r o r
70 i f ˜ a l l ( isfinite ( x0 )) , break; end % a f l o a t i n g point error

cbnd H.P. Gavin January 3, 2025

 14 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

71 end
72 % −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− ODE4U
73 % 2014 , 2022−04−29, 2023−01−21

cbnd H.P. Gavin January 3, 2025

Numerical Integration for Structural Dynamics 15

A Variable-Step Fourth-Fifth-Order Runge-Kutta Solver

For problems in which the time increments are not much shorter than the shortest
natural periods involved in the system, and for problems involving hard nonlinearities, such
as problems involving impact or friction, fixed-step Runge-Kutta solvers may be numerically
unstable. In such cases one may attempt a re-analysis using a shorter time step for every
time step, or one may use a solver in which the large fixed time-steps are automatically sub-
divided only when necessary. A feature of such variable time-step solvers is that a desired
level of accuracy may be enforced throughout the simulation, by comparing a fourth-order
accurate solution with a fifth-order accurate solution.
The implementation of the fourth-fifth order solver discussed here was proposed by
Cash and Karp in 1990.10 To advance the solution from a time t to a time t + ∆t, six state
derivatives are computed

ẋ1 = f (t + a1 ∆t, x(t), u(t + a1 ∆t)) (60)

ẋ2 = f (t + a2 ∆t, x(t) + ẋ1 b21 ∆t, u(t + a2 ∆t)) (61)
ẋ3 = f (t + a3 ∆t, x(t) + (ẋ1 b31 + ẋ2 b32 )∆t, u(t + a3 ∆t)) (62)
ẋ4 = f (t + a4 ∆t, x(t) + (ẋ1 b41 + ẋ2 b42 + ẋ3 b43 )∆t, u(t + a4 ∆t)) (63)
ẋ5 = f (t + a5 ∆t, x(t) + (ẋ1 b51 + ẋ2 b52 + ẋ3 b53 + ẋ4 b54 )∆t, u(t + a5 ∆t)) (64)
ẋ6 = f (t + a6 ∆t, x(t) + (ẋ1 b61 + ẋ2 b62 + ẋ3 b63 + ẋ4 b64 + ẋ5 b65 )∆t, u(t + a6 ∆t)) (65)

The fourth order predictor at time t + ∆t, is x4 (t + ∆t), and the fifth order predictor at time
t + ∆t, is x5 (t + ∆t).

x4 (t + ∆t) = x(t) + (ẋ1 c41 + ẋ3 c43 + ẋ4 c44 + ẋ5 c45 + ẋ6 c46 )∆t (66)
x5 (t + ∆t) = x(t) + (ẋ1 c51 + ẋ3 c53 + ẋ4 c54 + ẋ6 c56 )∆t (67)

The coefficients ai , bij , and cij are provided in Cash and Karp’s paper. If every term of the
truncation error vector, ϵ = |x4 − x5 |/|x5 | is less than the required tolerance, ϵtol , then the
solution x5 is considered sufficiently accurate. In this case, x(t + ∆t) is assigned to x5 , t is
assigned to t + ∆t, and the solution marches on. Otherwise, the time step is divided into
Nss sub-steps,11 where
l h   im
Nss = Nss max max (ϵ/ϵtol )1/4 , 1.1 , (68)

and Nss is initialized to 1, and ⌈·⌉ is the round-up operator.

These concepts are implemented in the Matlab function ode45u.m12

10
Cash, J.R. and Karp, A.H., “A Variable Order Runge-Kutta Method for Initial Value Problems with
Rapidly Varying Right-Hand Sides,” ACM Transactions on Mathematical Software, vol 16, no 3, pp 201-222
(1990). http://www.duke.edu/∼hpgavin/cee541/Cash+Karp-ACMTMS-1990.pdf
11
Press, W.H., et al., Numerical Recipes, Cambridge Univ. Press, 1993. Section 16.2 http://www.nr.com/
12
http://www.duke.edu/∼hpgavin/ode45u.m

cbnd H.P. Gavin January 3, 2025

 16 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

The usage of ode45u.m is very similar to the usage of ode4u.m

ode45u.m
1 % [ time , x , x d o t , y ] = ode45u ( d x d t , time , x0 , u , params , t o l e r a n c e , d i s p l a y )
2 %
3 % S o l v e a sy s te m o f nonhomogeneous o r d i n a r y d i f f e r e n t i a l e q u a t i o n s
4 % u s i n g t h e embedded Runge−Kutta f o r m u l a s o f J .R. Cash and A.H. Karp , and
5 % u s i n g l i n e a r i n t e r p o l a t i o n o f e x t e r n a l f o r c i n g w i t h i n time−s t e p s .
6 %
7 % Input Variable Description
8 % −−−−−−−−−−−−−− −−−−−−−−−−−
9 % dxdt : a f u n c t i o n o f t h e form [ x d o t , y ] = d x d t ( t , x , u , params )
10 % which p r o v i d e s t h e s t a t e d e r i v a t i v e g i v e n t h e time , t ,
11 % t h e s t a t e , x , e x t e r n a l f o r c i n g , u , and p a r a m e t e r s
12 % a v e c t o r s y st e m o u t p u t s , y , may a l s o be r e t u r n e d .
13 % time : 1−by−p v e c t o r o f time v a l u e s , u n i f o r m l y i n c r e a s i n g
14 % ( or d e c r e a s i n g ) , a t which t h e s o l u t i o n i s computed .
15 % x0 : n−by−1 v e c t o r o f i n i t i a l s t a t e v a l u e s , a t time ( 1 ) .
16 % u : m−by−p m a t r i x o f s y st e m f o r c i n g d a t a f o r t h e ode ,
17 % sampled a t p o i n t s i n t h e 1−by−p v e c t o r time .
18 % params : o p t i o n a l p a r a m e t e r s used i n t h e f u n c t i o n d x d t
19 % tolerance : the desired tolerance constant ( d e f a u l t 0.001)
20 % may be a v e c t o r o f t h e l e n g t h o f x0 , or a s c a l a r
21 % display : 3= d i s p l a y l o t s o f r e s u l t s ; 2= d i s p l a y l e s s ; 1= even l e s s
22 %
23 % Output V a r i a b l e Description
24 % −−−−−−−−−−−−−−− −−−−−−−−−−−
25 % time : i s r e t u r n e d , un−changed
26 % x : the s o l u t i o n to the d i f f e r e n t i a l equation at times
27 % x dot : the d e r i v a t i v e to the d i f f e r e n t i a l equation at times
28 % y : the output of the d i f f e r e n t i a l equation at times
29 %

Numerical Example
In this example ode4u.m is used to solve the transient response of an inelastic system
forced by ground accelerations ẅ(t), as described by the equations:
r̈(t) + 2ζωn ṙ(t) + µz(t) = −ẅ(t) (69)
 
3
ż(t) = 1 − z sgn(ṙ(t)) ṙ(t)/dy . (70)

with initial conditions r(0) = 0, ṙ(0) = 0, z(0) = 0. The paramter µ is a plastic force
level divided by the mass (µ = fp /m), and the paramter dy is a yield displacement. The
external forcing is ground acceleration, −ẅ(t), given by a single full-period sinusoidal pulse
of amplitude Ap and period Tp ,
(
−Ap sin(2πt/Tp ) 0 ≤ t ≤ Tp
ẅ(t) = (71)
0 otherwise
In state space, these equations may be expressed as ẋ = f (t, x, u, p)
   
r(t) ṙ(t)
d 
ṙ(t) 
= 
−2ζω n ṙ(t) − µz(t) − ẅ(t)  (72)
dt
   
z(t) (1 − z 3 sgn(ṙ(t))) ṙ(t)/dy .
These ordinary differential equations are computed in the short .m-file inelasSys.m

cbnd H.P. Gavin January 3, 2025

 Numerical Integration for Structural Dynamics 17

1 function [ dxdt , fi ] = inelasSys (t ,x , a_g , p )

2 % [ dxdt , f i ] = i n e l a s S y s ( t , x , a g , p )
3 % dynamics o f an SDOF i n e l a s t i c s ys t e m
4
5 % state vector ...
6 r = x (1); % displacement response , m
7 r_dot = x (2); % v e l o c i t y r e s p o n s e , m/ s
8 z = x (3); % hysteretic variable
9
10 % parameters . . .
11 wn = p (1); % n a t u r a l f r e q u e n c y , rad / s
12 zeta = p (2); % v i s c o u s damping r a t i o
13 mu = p (3); % y i e l d f o r c e d i v i d e d by mass , m/ s ˆ2
14 dy = p (4); % y i e l d displacement , m
15
16 fi = mu * z ; % inelastic force
17
18 % state equations
19 dxdt = [ r_dot ; % v e l o c i t y response
20 -2* zeta * wn * r_dot - fi - a_g ; % acceleration response
21 (1 - z ˆ3 * sign ( r_dot )) * r_dot / dy ]; % inelastic force derivitive

And the simulation may be executed using the m-file inelasSim.m

1 % i n e l a s S i m .m − s i m u l a t e t h e r e s p o n s e o f a SDOF i n e l a s t i c s y st e m u s i n g ode4u .m
2 % parameters . . .
3 g = 9.81; % g r a v i t a t i o n a l a c c e l e r a t i o n , m/ s ˆ2
4 k = 5000; % s t i f f n e s s , N/m
5 m = 100; % mass , kg
6 zeta = 0.02; % v i s c o u s damping r a t i o
7 fp = 95.0; % plastic force level , N
8 Tp = 1.0; % pulse period , s
9 Ap = 1.0; % p u l s e a m p l i t u d e , m/ s ˆ2
10
11 mu = fp / m ; % y i e l d f o r c e d i v i d e d by mass , m/ s ˆ2
12 wn = sqrt ( k / m ); % n a t u r a l f r e q u e n c y , rad / s
13 dy = fp / k ; % y i e l d displacement
14
15 p = [ wn , zeta , mu , dy ]; % v e c t o r o f s y st e m p a r a m e t e r s
16
17 % time a x i s d a t a . . .
18 T = 5; % duration of simulation , sec
19 dt = 0.010; % time s t e p , s
20 N = f l o o r ( T / dt ); % number o f d a t a p o i n t s
21 t = [0: N -1]* dt ; % time , s
22
23 % ground a c c e l e r a t i o n i n p u t d i s t u r b a n c e , m/ s ˆ2 . . .
24 a_g = - Ap * s i n (2* pi * t / Tp ); a_g ( f l o o r ( Tp / dt ):end ) = 0;
25 x0 = [ 0; 0; 0]; % i n i t i a l states : [ displ ; veloc ; inelastic force ]
26
27 [t ,x , dxdt , fi ] = ode4u ( ’ inelasSys ’ , t , x0 , a_g , p ); % s i m u l a t e r e s p o n s e s
28
29 % P l o t s −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
30 f i g u r e (1)
31 clf
32 hold on
33 plot ( t ,100* x (1 ,:) ); % d i s p l a c e m e n t r e l a t i v e t o t h e ground r ( t )
34 plot ( t , x (2 ,:) ); % velocity r e l a t i v e t o t h e ground
35 plot ( t , dxdt (2 ,:)+ a_g ); % t o t a l a c c e l e r a t i o n , r e l a t i v e t o an i n e r t i a l frame
36 hold off
37 f i g u r e (2)
38 clf
39 hold on
40 plot (0 ,0 , ’ ok ’)
41 plot ( 100* x (1 ,:) , fi );
42 hold off

cbnd H.P. Gavin January 3, 2025

18 CEE 541. Structural Dynamics – Duke University – Fall 2020 – H.P. Gavin

6
relative displacement response, cm
relative velocity response, m/s
4 total acceleration response, m/s/s

2
responses

-2

-4

-6
0 1 2 3 4 5
time, t, s

1
2
inelastic force / mass, µ z(t) / m, m/s

0.5

0
(r(0), µz(0))

-0.5

-1
-6 -4 -2 0 2 4 6
relative displacement, r(t), cm

cbnd H.P. Gavin January 3, 2025