Structural Safety, 3 (1986) 167-194 167

Elsevier Science Publishers B.V., Amsterdam - Printed in The Netherlands

METHODS OF STOCHASTIC STRUCTURAL DYNAMICS

Y.K. Lin .i, F. Kozin 2, Y.K. Wen a, F. Casciati 4, G.I. SchuOller s, A. Der Kiureghian e,
O. Ditlevsen 7 and E.H. Vanmarcke a

1 Florida Atlantic University (U.S.A.), 2 Polytechnic Institute of New York (U.S.A.),

University of Illinois, Urbana (U.S.A.), 4 University of Pavia (Italy), 5 University of Innsbruck (Austria),
6 University of California, Berkeley (U.S.A.), 7 Technical University of Denmark, Lyngby (Denmark),
8 Princeton University (U.S.A.)

ABSTRACT

A concise review is given of the analytical non-parametric excitations appear as inho-

methods of stochastic structural dynamics which mogeneous terms on the right hand side. Physi-
deals with structural systems under time-vary- cally, random parametric excitations represent
ing random excitation. Included in the review the variation of structural properties with time,"
are both linear and nonlinear structures and therefore, they can affect the stability of struct-
both parametric and non-parametric random ural response. Approximate methods are de-
excitations. Mathematically, parametric exci- scribed for those cases for which exact solutions
tations appear in the coefficients for the un- are presently not available.
knowns in the equations of motion, whereas

2.1. INTRODUCTION bles have been discretized using a suitable

scheme, for example, the Galerkin or a finite
Stochastic structural dynamics is con- element scheme. Equations (1.1) are first order
cerned with the solution of equations of the differential equations; however, this is not a
following type restriction since a higher order differential
equation or a differential-integral equation
d Xj(t)=f/(X, t ) + g + k ( X , t)Fk(t )" can be converted equivalently to set of first
dt
order differential equations.
j = 1, 2 . . . . . N; k = 1, 2 , . . . , M (2.1.1)
Clearly, the functional form of each g j+, in
where t represents time, Fk(t ) are M random Eqns. (2.1.1) provides a clue as how the ran-
excitations, Xj(t) are components of and N- d o m force F~ interacts with the system. For
dimensional response state vector X(t), and a ease of discussion, let us re-write this last
repeated index indicates a summation. It is term as follows:
assumed that the spatial independent varia-
g+k(x, t)r (t) =
* Chairman of Subcommittee on Methods of Stochastic
Structural Dynamics. +gjt,2(t)Ft(t) (2.1.2)
168

Fk(t) gj k,2 It) Fk(t) 2.2. LINEAR TIME-INVARIANT STRUC-

TURES UNDER NON-PARAMETRIC RAN-
DOM EXCITATIONS

Many random vibration problems are

(a)
solved using the elementary spectral analysis
Fig. 1. Linear time-invariant structure under (a) sta-
of one form or another. This method is an
tionary random excitation, (b) uniformly modulated extension of the well-known operational
random excitation. calculus, applicable only to linear time-in-
variant systems. A time-invariant system can
indicating that two types of gjk function are be represented by differential equations with
possible: the first type, gjk.l(X), is autono- constant coefficients; therefore, the method is
mous (dependent on the response X but not not applicable to the case of parametric exci-
explicitly on t), whereas the second type tation. Further restriction to linearity is due
gj/.2(t) depends only on t. The roles played to the fact that the principle of superposition
by the random excitations, Fk(t ) and Et(t ), is used in the analysis.
are f u n d a m e n t a l l y d i f f e r e n t . T h r o u g h Referring to (2.1.1), functions fj must now
gjk.~(X), random excitations Fk(t ) appear in be linear and functions gjk must not be func-
the coefficients of the unknowns in the tions of the unknown response. More specifi-
governing equations; thus, they are capable of cally, the governing equations may be written
changing the characteristics of the dynamic as follows:
system randomly with time. In particular, they d
can make an originally stable system unstable - ~ Xj = ajk X k + gjk.2( t ) Fa ( t )"
or vice versa. On the other hand, random
excitations F~(t), being associated with j = 1, 2 . . . . . N; k = 1, 2 . . . . . M (2.2.1)
gjt,2(t), appear only in the inhomogeneous in which ajk a r e now constants. Two cases
terms, and they cannot affect the systems will be considered. In one case, the gjk.2 func-
stability. Due to their distinctive roles, the tions are constants (all are equal to unity),
random excitations in the first group, Fk(t ), and Fk(t ) are stationary random processes.
are called the parametric excitations, and In another, the gjk,2 functions do vary with t,
those in the second group, F~(t), the non- but Fk(t) remain to be stationary processes.
parametric excitations, although they are also The general idea may be outlined with the
known by other names in different areas of help of Figs. 2.2.1a and b. We may assume,
applications. Intuitively, the parametric exci- without loss of generality, that the ensemble
tations are more "intimately" related to the average of every random excitation is zero.
structural response than non-parametric exci- Then the ensemble average of every response
tations. variable, Xj(t), of a linear system is also zero.
In the present paper, a state-of-art review The effect of non-zero means of the excita-
will be given of the general field of stochastic tion processes, if they exist, can be evaluated
structural dynamics. We shall begin with the separately in a deterministic analysis, and the
simplest case of linear time-invariant systems results superimposed on the zero-mean solu-
under purely non-parametric excitations, for tions.
which the elementary spectral analysis or its In Fig. 2.2.1a, an arbitrary linear structure
generalized version is applicable. The case of is excited by a number of stationary random
nonlinear structures will be discussed next, processes, typically Fk(t ) at location k. The
followed by parametrically excited systems. immediate objective is to obtain certain star-
 169

istical properties of a response quantity Xj(t) The cross-spectral density Oxx of the re-
(treated as a component of a suitable re- sponse variables Xj(t) and X,,(t~ can be used
sponse vector) when the response also reaches to compute the cross-correlation function:
statistical stationarity. Since Fk(t) and Xj(t)
are stationary random processes, they have
the Fourier-Stieljes integral representations:
OG
Fk(t ) = f]oc
e i'°t d/g'k(*0)
--OO

X (I)F~F~(.0) ei'°" d*0 (2.2.6)

oo
X,(t) = f£ e i~' dXj(*0) (2.2.2)
oo which is a function of the time-difference ~-.
in which Fk(*0) and A'j(*0) are random When j - - m , a cross-correlation becomes an
processes with orthogonal increments, having auto-correlation whose value at the zero
the property time-difference, ~-= 0, is the mean-square
value of the response quantity in question.
E [dFk(*0,)d F/* (.02)] The interpretation of a typical excitation
0, .01 4:.0= Fk(t ) and a typical response variable Xj(t)
x d*0, *0, = ,02 = .0
can, of course, be generalized. In particular,
when classical normal modes of vibration of a
E [d2j(*0,) dX*(*0=)] linear time-invariant system exist and can
easily be found, it is usually more convenient
0, *0, • *02 = .0 (2.2.3) to formulate the problem such that Xj(t)
= ~ff~g,Xm(~,l)) d*0, *0, = *o2 = *0 represents the generalized displacement in the
j th mode, and Fk(t) the generalized force in
in which E[. ] denotes an ensemble average, the k th mode. Even when classical normal
an asterick indicates the complex conjugate, modes do not exist, a set of assumed modes
and • represents a spectral density or a may, sometimes, be selected, in which case
cross-spectral density of the random processes the results obtained will be mathematically
identified in the subscripts. approximate in the sense of Galerkin.
Equation (2.2.2) shows that F k and Xj are The case of non-constant gjk,2 functions is
composed of infinitely many infinitesmal shown in Fig. 2.2.1b. The products gjk,Z(t)
sinusoidal components. Then from the princi- Fk(t) play the role of excitations. These prod-
ple of linear superposition, ucts are non-stationary random processes,
d)~'j(*0) = Hjk(*0 ) d•(*0) (2.2.4) even though Fk(t ) are stationary. These prod-
ucts are known as uniformly modulated ran-
in which each Hjk (.0) has the physical mean- dom processes, a class of non-stationary ran-
ing of the complex amplitude of a response dom processes which has found important
variable at location j if the excitation is a practical applications. Under non-stationary
unit sinusoid at location k alone. The Hjk excitations, the response quantities are also
functions are called the frequency response non-stationary, and the cross-correlation of
functions; their determination often con- two response quantities is a function of two
sumes the major portion of time and effort in different times, instead of the time difference.
an analysis. Without giving the details, it suffices to note
It follows from the orthogonal increment that eqn. (2.2.6) must now be replaced by
property, eqn. (2.2.3), that
d~x,x,,(w ) = H)k(*0)H*I(*0)dPF~F,(*0) (2.2.5) (OXjxm(ll, 12)=E [ X j ( t l ) X r n ( t 2 ) ]
170

= f* M:.(t,, responding higher order statistical properties

- DC
are required, and they can be expressed most
×e conveniently in terms of the impulse response
functions. For example, if the structure is
(2.2.7)
initially quiescent, the third moment relation-
in which ship is given by

a4,k(t, _
u[x,(,,)xk(,2)x,(,,)]
=£
×
fo'gjk,l(r) e -i(v-~)t d r
×(t2-
(no summation on j nor k) (2.2.8)
×E[Fp(r,)Fq('r2)F,('r3) ] d% (2.2.10)
When the modulation functions gs~,2 are of where each impulse response function, say
the exponential, sinusoidal or polynomial hjp(U), is related to the corresponding
type, or their combinations, eqn. (2.2.8) may frequency response function H/p(o:) as fol-
sometimes be evaluated in closed form; fre- lows:
quently, it must be computed numerically.
It is of interest to note that Eqns. (2.2.7) h j e ( u ) = - ~ _1 f~ Hse(~0) e~'~"do: (2.2.11)
and (2.2.8) can also be used to compute corre-
lation functions for transient non-stationary Other higher moments of the response
response for a system which is initially processes can be similarly formulated. If the
quiescent and is exposed to stationary excita- excitation processes are jointly stationary and
tions at t =0. This is recognized to be a if moments of response processes int he sta-
special case in which tionary state are required, then the lower limits
in the multi-fold integration should be re-
1, t> 0 (2.2.9) placed by - m, and the result is independent
gJk'2(t) = O, t<0
of the choice of time origin for the t 's. Higher
This formulation is especially useful in order spectra (e.g. bi-spectrum, tri-spectrum)
modeling the 'zero-start' conditions char- may be defined as Fourier transforms of the
acterizing earthquake excitation. Nonsta- higher (3rd, 4th) moments of the stationary
tionary statistics of the response to suddenly response, and they are related to the respec-
applied and modulated random excitation tive spectra of the excitation analogously to
may be found e.g. in [95] through [100]. eqn. (2.2.5). The probability density of the
Thus far, our discussion has been restricted response may then be approximated by series
to the second order statistical properties expansions utilizing these higher statistical
(either spectral densities or correlations) of moments, e.g. Ref. [80].
the excitation and response processes, the The literature pertaining to the subject of
mean functions of these processes being as- linear time-invariant systems under random
sumed to be zero. If the excitation processes non-parametric excitation is enormously rich.
are Gaussian distributed, then the response Yet, for this elementary and well-established
processes of a linear system are also Gaussian portion of stochastic structural dynamics, it
distributed, and they are completely char- appears adequate to list only several text
acterized by the first and second order statis- books as references, for example, [1-9]. Al-
tical properties. For non-Gaussian excitations though spectral representation is often the
and responses, the relationship between cor- same basic tool for the analysis of such sys-
 171

terns, it is refreshing to find occasionally novel Under Gaussian white noise excitation, the
applications such as [10] and [11]. In the response state vector Y(t) = (Yl(t), (Y2(t)} is
particular area of earthquake engineering, a Markov vector. The joint probability den-
proposals to modify the commonly used re- sity of Yl(t) and Y2(t) conditional on Yl(to)
sponse spectrum approach incorporating con- =Ylo and Y2(to)=Y20 is governed by a
cepts of stochastic structural dynamics were parabolic partial differential equation, called
made in [12] and [101]. the Fokker-Planck equation which, for the
present case, is given by
8p Op 8 ,s 8 2p
2.3. NON-LINEAR STRUCTURES UNDER
NON-PARAMETRIC RANDOM EXCITA-
8--7 +Y2 8y2 8y2[Pg(Yl,
Y2)]- 'rr/x~-~- = 0
8Y2
TIONS (2.3.4)

2.3.1 Statement of the Problem where K is the spectral density of the excita-
tion white noise. The unknown p is called the
Consider the following equations of mo- transition probability density of the Markov
tion: vector Y(t). Equation (2.3.4) is solved with an
mijX j + gi(X, X ) = F/(t); initial condition

i, j --- 1, 2 .... , N (2.3.1) P(Yl, Y2, t0lYl0, Y20, to)

where rnij are constants and g,;(X, X) are = 8(y, - y,0) 8(y2 - Y20)- (2.3.5)
nonlinear functions of X and X. Equations A general solution for eqn. (2.3.4) is not
(2.3.1) may represent a structural model for a known for second order nonlinear systems.
building, an offshore platform, an aircraft or However, some exact solutions are known for
a ship, and the nonlinearity may arise from the following reduced equation,
large deflection or inherent nonlinear material
8p* 8 8 2 -*
property, or both. Although the system is Y2 8y 1 8yz[P*g(Ya, Y2]-~rK 8Y2 - 0
nonlinear, its dynamic properties are still
time-invariant and the excitations are non- (2.3.6)
parametric. The objective is to solve for the
response processes X i and k i given the prob- obtained by setting the time-derivative term
abilistic or statistical information about the in eqn. (2.3.4) to zero. The unknown p* in
random excitations F, (t). eqn. (2.3.6) represents the unconditional joint
probability density for Yl(t) and Y2(t) when
2.3.2. Some exact solutions t - t o is sufficiently long so that the effect of
the initial condition is negligible, and these
When the excitation processes can be mod- response processes have reached statistical
eled as Gaussian white noises, exact solutions stationarity. In particular [13,5] if
are sometimes possible. In the case of single
degree of freedom, we have g(Yl, Y2)=Y2gz(H)+g1(Yl) (2.3.7)
X+ g(X, J()= W(t) (2.3.2) where

Let Y I = X and Y2=X, eqn. (2.3.2) is re-

placed by two first order equations H= ½y~ + foV'ga(u) du (2.3.8)
J', = II2 (2.3.3a)
and g2 satisfies rather general conditions to
J~2 = -g(Y1, Y2)+ W(t) (2.3.3b) ensure the existence of stationary response,
172

then above solution is that Yg2 be a positive func-

tion.
p* = C exp - ~ J 0 g2(v/) d-q (2.3.9) In closing the discussion on exact solu-
tions, it should be pointed out that white
The coefficient C is determined from the noise processes, which are delta correlated or
condition that the total probability is unity. have constant spectral densities, are idealiza-
For practical purposes the conditions re- tions. Therefore, the solutions presented in
quired of function g2 are satisfied if it is a this section are mathematically exact, but they
positive function. are physically approximate in the sence that
Equation (2.3.7) includes an important spe- physically realistic random excitations have
cial case been replaced by ideal white noises. The re-
g(Yl, Y2) = flY2 + gl(Yl) (2.3.10) placement is justified provided that the corre-
lation times of the actual excitation processes
for which
are short compared with the relaxation time
of the dynamic system. In such a case the net
area under a correlation curve, be it an auto-
(2.3.11) correlation or a cross-correlation, may be
lumped at the origin, effectively changing the
This special case restricts nonlinearity in the physical processes by white noises. If para-
displacement alone, in which case the velocity metric excitations are also present, then the
II2(t) is a Gaussian random variable at every replacement cannot be so simply accom-
t and is independent of the non-Gaussian plished, but it can still be done using a well-
displacement Y](t) at the same t. Clearly, established procedure, known as the stochas-
Yz(t) is not a Gaussian random process. tic averaging method of Stratonovich [t5] and
The stationary solution, eqn. (2.3.9), has Khasminskii [16].
been generalized [13] for multi-degree-of-free-
dom systems, governed by 2.3.3 Approximate Methods
X, + ~,g2(H) 2, + 3U(X)/3X,= W,(t);
Since exact solutions are not always ob-
no sum in i; i = 1, 2 . . . . . N (2.3.12)
tainable for nonlinear systems under random
in which U is a potential function, and Wi(t ) excitations, even when the excitations are
are uncorrelated Gaussian white noises of idealized white noises, a number of approxi-
spectral densities Kj. If the following ratios mation techniques have been developed [14].
p,/Kj =V are all equal, independent of j, Some of these techniques will be briefly dis-
then the solution to the reduced Fokker- cussed below.
Planck equation for the stationary joint prob-
ability density is given by I. Iterative Solution for the Fokker-Planck
equation [13]
p* = C e x p l - ;y f0/4g2(T/)d~] (4.3.13) If the nonlinearity is small, then the transi-
tion probability density is expected to be
where close to that of the linear case. Consider, for
example, the following equation of motion.
N
H = ½• Xi2 + U ( X ) (4.3.14) X+ flJ(+cooX+eg(X, 2 ) = W(t) (2.3.15)
i=l
Where I ~ l i s a small constant. An approxi-
A sufficient condition of the existance of the mate solution for the Fokker-Planck equa-
 173

tion associated with this nonlinear system may OjlOk,= 2rrKrsgjr(X , t)gks(X, t) (2.3.21)
be obtained as the sum of the solution for a
where K,s is the cross-spectral density of
linear system and a correction attributable to
Wr(t ) and Ws(t). When r = s, a cross-spectral
the nonlinear term ~g(X, X) in the equation
density becomes a spectral density. Without
of motion; namely
parametric excitations, the second term on
p ( y , t l Y0) =P0(Y, t l Y0) the right hand side of (2.3.20) will be zero.
The conversions rules, (2.3.20) and (2.3.21)
-,fo'dtfg( )p(L yo) are general; they are applicable even when
parametric excitations are present.
× ~~Po(Y, t-~l~) d~ Equation (2.3.18) can be used to obtain an
It#) equation for an arbitrary scalar function
(2.3.16) ~ ( X ) of X [20]:
where P0 is the transition probability density
if e = 0. eqn. (2.3.16) may be solved itera- dq~ = --~ + mJ-~ii + ~oJ,Ok, OXjOXk dt
tively. Convergence can be expected when the
nonlinearity is small. • a~
+ qgk~-~j dBk(t) (2.3.22)
II. Cumulant Closure
If we are interested only in some statistical The equation for E[~(X)] follows from en-
moments of structural response to Gaussian semble averaging:
white noise excitations, then a solution to the d
associated Fokker-Planck equation may not
be required. The equations for moments can
be derived from the Fokker-Planck equation a~ 1 02~ ]
=E m-~j + ~OjlOkla~j~ Xk
(e.g. [17]), but they can also be derived more
conveniently from the It8 type stochastic dif-
J
ferential equations [18] for the system. (2.3.23)
Let the physical equations and the corre-
sponding It8 equations be, respectively, where ~ is known as the backward diffusion
operator of the Markov process. Equations
d xj(~)-~-fj(X, t)-[-gjk(X, t)Wk(t ) for the statistical moments are obtained by
dt
(2.3.17) letting q~(X) be Xj, XjXk, XjXkXt, etc.
Unfortunately, the equations for the statis-
and
tical moments for the response of a nonlinear
dXj(t) = rnj(X, t)dt + Ojk(X , t) dBk(t ) system form an infinite hierarchy, in the sense
(2.3.18) that an equation derived for an n th moment
where Bk(t) are independent unite Wiener generally contains moments of both higher
processes, satisfying order and lower order than n. Therefore, a
suitable closure scheme must be employed to
E[Bk(t,)B,(t2) ] =8k, min(t,, t2) (2.3.19) obtain approximate solutions. One such
It can be shown [19] that scheme is cumulant closure which has the
a advantage of being applicable to large nonlin-
rag(X, t) = ~ ( X , t) +qTKkr-~rgjk(X , l) ear systems, including those under both para-
metric and non-parametric excitations.
X g,, ( X, t ) (2.3.20) The statistical moments are related to
174

cumulants as follows [5,15]: necessary. The perturbation method, e.g. [22],

is one such approach.
E[x,] Consider, for example, a single-degree-of-
E[ x xk] : xj, xk] + freedom system,

E[ X, XkX,] Xj, Xk, X,] LX+cg(X, )() = F ( t ) (2.3.25)

where L is a linear operator and F(t) is a
+ 3{ ~,[ xj]~2[ x~., x,] }, weakly stationary excitation. It is assumed
that ¢ is sufficiently small and the solution
for eqn. (2.3.25) can be expanded as a power
E[xjx x,x,.] series,
x(t) = Xo(t) + ¢x,(t) + ~2x2(t) +...
(2.3.26)
÷ 3( K2[ Xj, Xk]K2[ XI, Xml } s
Substituting (2.3.26) into (2.3.25) and equating
+4( ~[ xjl~[ x~, x,. x.,] }, like powers of ~, we obtain
+ 6 ( Ka[ Xj] K, [ Xk]K2[ X~, X,, I }, LXo=r(t )
LX, = - g ( X 0 , X0) (2.3.27)
(...) (2.3.24) (...)
where Kj [ ] denotes a j t h cumulant, and ( }s Equations (2.3.27) are linear and can be solved
indicates a symmetrizing operation with re- using standard methods given in Section 2.2.
sl~ect to all its arguments; namely, taking the In particular, if the system is time-invariant,
arithmetic mean of different permuted terms the spectral density of the stationary re-
similar to the one within the braces. sponse, calculated to order ¢, is given by
The simplest cumulant closure is the
cumulant-neglect closure [21] in which all the O x x ( ~O) = O XoXo( ~o) + 2c Re[ qbXoXl(¢0)]
cumulants higher than a given order N are set (2.3.28)
to zero. The statistical moments of an order
where Re denotes the real part, and Ox,,x,(¢o)
higher than N can then be expressed in terms
is the cross-spectral density of Xo(t) and
of lower moments using relations (2.3.24).
Very accurate results have been obtained using
Xl(t).
this scheme [21]. It is of interest to note that IV. Equivalent linearization technique
for a r a n d o m process defined on ( - o o , ~ ) , This particular approximation method is
neglecting K, for n > 2 is equivalent to the quite versatile when applied to practical
so-called Gaussian closure; namely, ap- multi-degree-of-freedom systems. It was first
proximating this r a n d o m process by a Oaus- developed for r a n d o m vibration analysis by
sian process. Booton [23] and Caughey [24], and has since
been extended for more complex systems with
III. P e r t u r b a t i o n m e t h o d varied nonlinear restoring forces, including
The approximation methods outlined in I hysteretic forces. The basic idea is to replace
an II are applicable when excitation processes a nonlinear system by an equivalent linear
are white noises or can be replaced by white one, such that the mean square difference
noises. When the white noise assumption is between the two systems (comparing the
not appropriate, more general approaches are equations of motion) is minimized. Recent
 175

works by Iwan and Spanos [25,26] have laid linear system may be used to start the itera-
further theoretical foundation for this method tion. Alternatively, a set of ordinary differen-
for stationary as well as nonstationary cases. tial equations of the response variances and
The coefficients so obtained for the lin- covariances may be obtained, which can be
earized system are generally functions of sta- solved iteratively for a stationary solution and
tistics of the response variables. They may be integrated numerically for a nonstationary
in implicit forms which could hamper the solution. The extension of the method to the
solution procedure, especially for multi-de- case involving excitations with non-zero
gree-or-freedom systems. means can be found in [27] where a separate
Let the equation of motion of a general set of differential equations are obtained for
multi-degree-of-freedom nonlinear system be the mean responses.
given by The kinds of nonlinear systems that are
g(~', ,i', X)=F(t) (2.3.29) amenable to equivalent linearization are re-
ally quite diverse. I n most applications, the
in which g is a vector function representing restoring forces may be modeled as algebraic
the total internal force acting on the system, functions of displacement and velocity, but
each component being a single valued and nonlinearity associated with the inertial forces
odd function of its arguments, and F is zero- as well as self-oscillatory systems have also
mean Gaussian vector process. Equation been considered [28,29]. The restriction that
(2.3.29) is replaced by a linear system with the restoring forces are anti-symmetric (odd)
the following equation of motion functions of their arguments has also been
Mr( + CX + KX= F(t) (2.3.30) relaxed [30].
A special class of nonlinear systems, of
in which the elements in the mass, damping particular interest in the study of the perfor-
and stiffness matrices are determined by the mance of a structure under severe loads, is
requirements that the mean square error that of inelastic systems. Most structures be-
E[~-T~.] is minimized, where come inelastic before damage occurs. How-
¢ = g - Mr( - CX - KX (2.3.31) ever, a major difficulty in treating such sys-
tems lies in the hereditary nature of the re-
The conditions that the partial derivative of
storing force, i.e., it depends on the response
E[iT~] with respect to each undetermined
time history. Invoking the Krylov-Bogoliu-
coefficient vanishes and that response
bov (K-B) assumption of slowly varying
processes of a Gaussian excited linear system
parameters, Caughey [31] obtained the lin-
are Gaussian lead to the following results:
earized system for a hysteresis system, with
coefficients expressed in integral form. How-
mi/ [8XjJ ever, the K-B assumption is tantamount to
that of the system being narrow banded, which
[ ag, 1 is valid only for slightly to moderately inelas-
(2.3.32)
tic behavior [32].
A recent approach to modeling hysteretic
For most nonlinear systems these coefficients behavior by a nonlinear differential equation
can be expressed explicitly in terms of the appears to be very versatile. It is capable of
response statistics. Since M, C and K are modeling system deterioration and biaxial in-
functions of the response statistics, a iterative teraction, and is adaptable to the equivalent
solution procedure is generally required. In linearization method in that the coefficients
this regard, the solution to the corresponding for the equivalent linear system can be ob-
176

tained in closed form without resort to the governed by a linear differential equation
K-B assumption [33,34]. This method of mod- while the response is within a given domain D
eling and solution procedure have been suc- in the response space, and by another dif-
cessfully applied to response and damage pre- ferential equation, linear or nonlinear, while
diction of a variety of structural systems un- the response is in the complement of D. Then
der seismic excitation [35,36]. given that the excitation is stationary, ergodic
Comparison of results from equivalent lin- and Gaussian and that the excursions of the
earization, from exact solutions if available response process out of D are rare and of
(e.g., from the Fokker-Planck equation) and short duration, it may be possible to obtain
from Monte Carlo simulations indicates that an approximate probabilistic description of
the accuracy of the equivalent linearization the samples of these response excursions in
method is generally very good. An error in the vicinities of the out-crossing times by
the range of 0-20% is representative. Further- using an almost deterministic representation
more, unlike some other approximation meth- for the excitation. This description which is
ods, the accuracy of this method is quite valid for smooth stationary and ergodic
independent of the level of nonlinearity; be it Gaussian response processes, is given as the
of a geometric origin or material origin. How- linear regression of the excitation on the value
ever, experience seems to indicate that the of the response and its derivative at the out-
errors tend to be on the unconservative side; crossing point plus a nonstationary residual
e.g., the statistical moments of the response Gaussian process. Modeling the derivative of
are usually underestimated. It must also be the response process at an out-crossing point
noted that since linearization is based on as a random variable with a distribution which
minimization of the mean square error, the can be obtained by sampling various realiza-
prediction of the response statistics other than tions at out-crossing points, the linear regres-
the second moments, such as auto correlation sion plus the residual process is known as a
function, extreme value, etc. may not be as Slepian process which is now the model for
reliable. For example, by comparing the exact the excitation before and after an arbitrary
probability density of the response of a Duff- out-crossing of the response.
ing oscillator under Gaussian white noise ex- If the previous excursion is far back in
citation with the approximate result obtained time, the response behaves like a stationary
from equivalent linearization [81], the tail re- and ergodic Gaussian process prior to the
gions of the two probability densities are next excursion. Thus the Slepian model for
found to be greatly different. The two com- the excitation may be used in the differential
puted probabilities for x > 30 may differ by a equation which governs the system behavior
factor as large as 250. Thus the prediction of in the complement of D. For certain types of
crossing statistics by equivalent linearization problems the residual process in the Slepian
may be seriously wrong for high threshold model starts up with a vanishing or small
levels. It has also been shown that for nonlin- variance as compared to both the variance of
early damped systems, the first excursion the original stationary excitation and the vari-
probability obtained from the equivalent lin- ance of the linear regression part of the model.
ear system may be in error by several orders Thus an essential part of the response may be
of magnitude [82,83]. obtained as the solution corresponding to this
linear regression part of the excitation alone.
V. Slepian process approximation When the response leaves the complement of
In the application of this approximation, D after a very short time duration, the vari-
we shall assume that a nonlinear system is ance of the residual portion of the Slepian
 177

process may not have grown significantly dur- the structure out of the elastic range, thereby
ing that time. U p o n re-entering to domain D, changing the system parameters, finally there
the response becomes a Gaussian process may be end loadings such as on a beam or a
conditional on the random entrance values at plate, that lead to randomly varying coeffi-
the boundary of D. Since the excursions out cients.
of D are rare, the response process will under For the general linear structure, we may
suitable conditions have time to become write the system equations as
nearly a stationary and ergodic process again, d X ( t ) / d t = A ( t ) X ( t ) = ( A o + Aa(t))X(t)
making the Slepian model applicable once
(2.4.1)
more at the next out-crossing.
Several papers using this Slepian model exclusive of non-homogeneous forcing terms,
technique have been published by Lindgren. where the matrix A 0 denotes fixed nominal
Applications to random vibration problems parameters, and the matrix Al(t ) contains
are demonstrated in [84] and [85]. elements which are random processes.
A very similar method of analysis has been There are two classes of questions that are
used to estimate response statistics for of concern to us relative to these simple linear
elasto-plastic and bilinear hysteretic systems random systems. We are concerned, quantita-
in Refs. [102] and [103]; use is made of the tively, with the statistical properties of the
work of Karnopp and Scharton [104] who response vector X(t). That is, we want to
obtained an estimate for the average rate of know the joint densities, moments, correla-
energy dissipation due to isolated excursions tions, first passage time statistics, etc., etc. We
above high yield levels. The inelastic system are also concerned with qualitative properties
response is approximated as the sum of an of the response, in particular the stability of
inelastic displacement (i.e., the effect of an the equilibrium solution X - O.
accumulation of plastic excursions) and an It is an interesting fact that we can say
oscillatory component. The latter is the re- more about the qualitative properties than we
sponse of an associated linear system with the can say about the quantitative properties. In a
same initial hysteretic system. Further exten- sense, this reflects the same situation as in the
sions permit inclusion of the effect of gravity case of deterministic linear systems, where we
on a yielding structure [106] and prediction of may not be able to solve (2.4.1) for a given
the random response of multi-degree elasto- time varying matrix A(t), but we can in
plastic shear beam structures [107]. m a n y cases obtain the stability boundaries for
various classes of functions aij(t ), which are
the elements of A(t).
2.4 RANDOM PARAMETRIC EXCITA- The basic reason that makes it difficult to
TIONS determine, for example, the moments of the
response vector X(t), is that we cannot in
As we have seen in eqn. (2.1.1) the general general separate the statistics of the A(t)
equation for structural dynamics includes the matrix process from the X-process. Thus,
possibility of randomly fluctuating parame- upon simply taking the expected values of
ters. Typically such terms arise when the (2.4.1), we have,
properties of a structure are varying due to
dE(X(t)}
internal or external causes. There may be -E{A(t)X(t)}, (2.4.2)
ageing in the system, which would lead to dt
changes in the structure, there may be ther- where the right hand side of (2.4.2), cannot be
mal gradients, or large excitations that take obtained directly, in general. We note, how-
178

ever, that if we can write For the linear system, there are only two
general classes of stochastic systems for which
E(A(t)X(t)}=E{.4(t)}E{X(t)), (2.4.3)
one can determine the moment properties ex-
then it follows that (2.4.2) simply remains as plicitly.
a linear system with coefficient matrix These are the cases for which the system
E{.4(t)} (2.4.1) may be written as
If, further, the .4(t) matrix contains ele-
ments which are stationary processes, if fol- dX(t)dt - [ A ° + i ~ l ~ A i f i ( t ) l X ( t ) (2.4.7a)
lows that E ( A ( t ) ) is a constant matrix which
allows exact solution of (2.4.2), given that or
(2.4.3) holds. Higher moments can also be
obtained. However, a problem still exists in dX(t)=AoX(t) dt + ~-'.AiX(t) dBi(t )
the random coefficient case, even in the sim- i=1
plest examples. It may happen that the mo- (2.4.7b)
ments do not exist, as the following simple
example shows. In the case of (2.4.7a), if the constant matrices
(.4, . 4 1 , " ' , .4m) generate what is known as a
solvable Lie algebra, then for the vector of
Example 4-1
Gaussian processes
Consider the scalar equation [L(t), f2(t) ....

d X ( t ) _ aZX(t) ' X(0) = 1 (2.4.4) the p t h m o m e n t of the components of the

dt solution vector X(t) can be obtained ex-
where a is a gaussian random variable with plicitly. A comprehensive discussion of this
zero man, and unit variance. The sample solu- class of system can be found in references
tions for this process are the family of in- [52-551.
creasing exponentials Unfortunately, the linear system of equa-
tions that are of most interest to structural
X ( t ) = e a2', (2.4.5) engineers do not generate solvable Lie alge-
bras. Therefore, the available results in this
which are very well behaved.
case do not appear at this time to be of
We cannot find the moments from (2.4.4),
practical significance for structural vibrations
but since we have the sample solution form
problems.
(2.4.5), we know that
The linear system (2.4.7b) is the general
E{Xk(t)}=E{ek~'} linear homogeneous Ito equation, in which
each dBi(t ) denotes the differential of a
OC
__ 1 f e (kt-+ '/2)a2 da Brownian motion. This is, of course, a special
VI2q7 -o¢ case of eqn. (2.3.18) when the system is linear
(2.4.6) and the random excitations are purely para-
metric.
It is obvious that he integral is divergent for In the linear case, indeed one can obtain a
t > 1/2k. Hence, all moments cease to exist closed set of equations for the p t h moments,
for t > 1/2.
Therefore, we see that even though this E { X P , Xp2 . . . Xpm } ,
stochastic linear equation is very simple, with
well behaved sample solutions, the statistical where
properties may be quite singular. P, + P2 + ... + p . , = p
 179

Recently, the general equations for the ar- that yields

bitrary pth order moments have been pre- ytpl = A [PlxtP] (2.4.10)
sented. These results are due mainly to Broc-
kett and his co-workers [70,71]. Motivated by Finally, for linear systems
the algebraic theory of linear differential X(t) = A ( t ) X ( t ) (2.4.11)
equations, one can define for a given n-vector
the differential equality
X and a given positive integer p, the associ-
ated vector X [p] whose components are X(t + h)= (I + hA(t))X(t) + O(h z)
[(pP)(P ;Pa)-.. holds, yielding from the definition (2.4.10)
X[Pl(t + h ) = (I + hA(t))IP]xtpI(t) + O(h 2)
(2.4.12)
x pp
Upon defining the limit
X~IX;2... X•, (2.4.8) 1
lim -~((I + hA(t)) [p]- l[pl)) = Atp](t),
for h J,0
n
we determine the associated differential equa-
Pi = P , P, > 0. tion for X [°l,
i=1
The components of X [pl a r e ordered lexico- XtP]( t ) = AEp]( t )XtP]( t ) (2.4.13)
graphically. For example, for n = p = 3, the
components of X [3], for In order to relate these ideas back to the
original eqn. (2.4.7b) one can apply the It6
calculus directly or study the Stratonovich
X= X2 equivalent to obtain the It6 differential equa-
X3 tion for X[ pl as
are, in order, m
+ ) - - 2Ar[pl
]
d~'tP](t) = A - ½ A
x¢ [Pl r=l

x?x2 × x [ P l ( t ) dt
rn
+ Y'~ ArI,IXIpI(t) dBr(t)
r=l
~/-6XIX2X3
(2.4.14)
X [3]= vl3Xl X2 (2.4.9)

x2 where the Br-processes are independent

~t3X2X3 standard Brownian motions.
Applying expectations to eqn. (2.4.14) im-
mediately yields the linear differential equa-
tion for the pth order moments as
The vector X [p] satisfies [[ X [pl [[ = [[ X [[P,
where [[ X[[ z = (X, X) and more generally,
( X , Y ) P = ( X [pl, Y[P]). Furthermore, this
concept extends to matrices through the defi- = A - 1 A + A r[pl
nition Y = A X , defining A[p], o r that matrix r=l [p]
180

xE{ are available, and certain problems that re-

main. We hope that the interested reader will
= KpE{ XIeJ(t) } (2.4.15)
look further into these questions, and, per-
Although, in principle, the equations for the haps, provide the needed answers.
p t h moments for linear homogeneous It6 dif-
ferential equations have been known for m a n y 2.5.2 Stability properties for stochastic
years, the general form (2.4.15) can be quite equations with physical noise coefficients
useful. These ideas extend quite simply to the
non-homogeneous case as well. The general linear homogeneous stochastic
Except for special cases, in general, all differential equation with physical noise coef-
other stochastic systems with randomly vary- ficients may be written as
ing coefficients require approximations such
as perturbation expansions, closure tech- d X ( t ) _ [(A + F ( t ) ] X ( t ) , (2.5.1)
dt
niques, etc., to determine the moments, corre-
lations and approximations to the desired where X(t) is an n-vector, A is a constant
probability densities. n X n matrix and F ( t ) is an n X n matrix
For qualitative properties, such as stability, whose non-identically zero elements C / ( t ) are
zero-mean stochastic processes.
an extensive literature as well as many results
do exist as we see in the next section. The problem of stability has generally been
concerned with the asymptotic properties of
the equilibrium solution X ( t ) - 0.
2.5. STOCHASTIC STABILITY Although stability of the solutions of sto-
chastic differential equations may be defined
2.5.1 Introductory remarks in a n u m b e r of modes corresponding to the
various forms of convergence of probability
We shall discuss a few problems related to theory, the most often studied stability con-
r a n d o m parametric excitation that have oc- cepts are stability of the mean, in particular,
curred to many in the field for which results asymptotic stability of the second moments,
are available, but complete answers have not and almost sure sample asymptotic stability.
as yet been forthcoming. As pointed out in For asymptotic stability of the second mo-
the Introduction, parametric excitation affect ments, one requires
the system stability. Surveys of the topic of lim E(Ir X(t)]12 ) = 0. (2.5.2)
stability for stochastic systems can be found tt~
in [37,38 and 39]. If the second m o m e n t s approach zero ex-
We concentrate mainly on linear systems. ponentially, then the second m o m e n t s are said
Although certain results are available for the to possess exponential stability.
non-linear stochastic systems, the results do For almost sure sample asymptotic stabil-
not contain as large a body of stability condi- ity, one requires
tions as are known for the linear case.
lim II X(t)II = 0 (2.5.3)
Non-linear systems have been studied by t~'~
many researchers, for example, Kliemann [40],
with probability one (almost surely).
as well as Sunahara and his co-workers [41].
We should note that (2.5.3) implies stabil-
However, more remains to be developed for
ity (Lyapunov) for linear systems. For non-
the stability properties of the non-linear sto-
linear systems, we also require
chastic system.
Our objective is to state certain results that lim sup ]t X ( t : t o , Xo)]l = 0 (2.5.4)
]IX'oil {0 t > t o
 181

with probability one, where x 0 denotes the be obtained, thereby allowing moment stabil-
initial value of X(t) at t = t 0. ity properties to be studied in the wideband
In general, conditions for stability of the case.
second moments cannot be obtained exactly For the small perturbation problem, early
for systems subjected to physical noise excita- works of Stratonovich [15], Weidenhammer
tions. The basic reason for this is that the [48], and more recently by Ariaratnam [49]
exact moments cannot, in general, be ob- and Wedig [50] find mean square stability
tained. conditions in terms of the values of the spec-
The fact that for the system (2.5.1) the tral density at twice the undamped natural
vector X(t) is correlated with the stochastic frequency of the linear oscillator when para-
coefficient F(t) does not allow us to obtain metric random excitations occur in the spring
an exact equation satisfied by the moments of terms. For the slowly varying case, Samuels
X(t). That is, we cannot write [51] obtained early results.
There is, however, another relatively recent
E{F(t)X(t)}=E(F(t)}E(X(t)) (2.5.5)
direction for studying the m o m e n t (as well as
for the linear system (2.5.1), where the matrix the sample) properties of linear systems with
F(t) contains components which are physical arbitrary physical noise coefficients. This ap-
random processes. proach involves the ideas of Lie Algebras,
The major approaches that are available first motivated by Brockett [52], and studied
presently to study the moment behavior of by Willems [53], Willsky et al [54], as well as
systems with physical random coefficients are Kistner [55].
approximate methods. These are based upon Let us now turn to sample properties and
the fact that the coefficient process is a small conditions that will guarantee asymptotic sta-
perturbation, the coefficient process is nar- bility almost surely. Consider the system
rowband or, finally, that the process is wide- (2.5:1) written as the general linear homoge-
band so that the Markov diffusion approxi- neous differential equation
mations may be applied via ideas of Stra- d X ( t ) / d t = A(t) X(t). (2.5.6)
tonovich [42], Wong-Zakai [43] or Khaz-
minskii [44], in which the original physical It can easily be shown that for the norm
coefficient equation may be substituted by an II.g lip = x T p x , ~v = p, p positive definite
associated white noise coefficient equation,
whose moments can be studied exactly in the the equality
linear case, but still require approximations in log II S'(t) lip - log II .g(0) lip
the general non-linear case.
t
In particular, if the random excitations are
wideband physical noises, the conversion 1 as
rules, (2.3.20) and (2.3.21), can be used to =TJo
obtain the equivalent It6 differential equa- (2.5.7)
tions.
These ideas have been recently applied by holds, where " T " denotes transpose.
Lin [45] and Ariaratnam [46]. In [47] the If the quotient on the left hand side of
approach was used for the first time to study (2.5.7) remains negative as t approaches infin-
the stability of 2nd order linear systems. Ex- ity, it must follow that
act results were obtained there. lim II x(t)lip = 0.
t]'oo
For the linear case, the associated It6 equa-
tion is also linear, and thus all moments can Therefore, the system will be asymptotically
182

stable. Thus, the time average of the quotient he found for w = 1,

within the integral if it exists, must determine
the asymptotic properties of X(t).
o~_=l£{F2(t)} < 4 ~ 2 (2.5.14)
We notice that the transformation as the sufficient condition extending the re-
sults in [57] and [58]. These results are based
X(t) = X(t)/ll X(t)lip (2.5.8)
only upon knowledge of the second m o m e n t
transforms the trajectories onto the surface of properties of the F-process.
an ellipse, since P is positive definite. Later, in 1972 Kozin and Wu [59], based
From the properties of pencils of quadratic u p o n the Infante approach, presented an ex-
forms, it is known that in terms of X(t) the tension to allow use of the properties of the
quotient in the integral of (2.5.7) satisfies probability density for the F-process, again
,o(s)< XT(s)[AT(s)e+ eA(s)]X(s) applied to second order oscillators. Indeed, if
we define Pt = Prob ( F ( t ) > l }, one can show
(2.5.9) that the sufficient condition becomes.
where o~(s), ~2(s) are, the smallest and largest
characteristic values, respectively, of the ma- [/(1- 2P,)+ 2f~f dP(f)]D < 2~,
trix (2.5.15)
[AX(s)e + PA(s)]P -1. (2.5.10) where D = (l + I - ~.,2)1/2.
Clearly, these are real functions since the As in Infante's procedure, l must be cho-
matrix is symmetric. sen to maximize the sufficiency region.
Thus, it will follow that if the elements Numerical procedures must be applied to ob-
aq(t) of the A-matrix are stationary, ergodic tain the region. However, the region of suf-
r a n d o m processes, then we can write from ficiency is greatly increased.
(2.5.1), with probability one, A similar result holds for the randomly
d a m p e d second order oscillator
E{~o(.)} = lim (s) ds
t'~ )((t) + [ 2 ~ + g ( t ) ] )((t) + o.}2X(/) = 0

log 1[X ( t ) l i p - log II ~(0) lip (2.5.16)

< lim
tl"~ t and the sufficient condition obtained again
leads to a greatly increased region of suf-
< lim
tl"oo
fo'a(s)ds= E{ f~ (s)}. ficiency, for almost sure sample asymptotic
(2.5.11) stability.
In an interesting paper [60], Blankenship
Thus, we obtain a sufficient condition for introduced a more general norm into this
almost sure sample stability, as problem. If the norm of a matrix is denoted
E{~2(.)} < 0 (2.5.12) by "ll II", then the measure of a matrix A,
denoted as/~A) is defined as
The condition in this form for physical noise
processes was first obtained by Infante [56]. I,(A)=lim[ III+OAII-1 ] (2.5.17)
0+0 0
I n particular, for the classical second order
oscillator with zero mean, stationary ergodic Thus for example, in the simple case,
physical noise coefficient, F(t),
l l x , = ( E &- 2''j2
) , II A II max •i(ATA),
X(t) = 0
=

X ( t ) + 2~o0J((t) + [o~2 + F ( t ) ] i

(2.5.13) where ~ i ( M ) denotes the ith eigenvalue of

 183

the matrix M, one finds that boundary was achieved by simulation. Our
past experience has always led to numerical
~(A) = ½max X,(A +AT).
i difficulties for determining sample behavior
The basic result is the following lemma, near the stability boundary.
LEMMA (Coppel [61]). Of course, we know that the limit in (2.5.19)
If ~ ' ( t ) = A ( t ) X ( t ) , then is much smoother than the sample behavior.
However, further clarification of the result in
exp( [63] would certainly be desirable.
It has been known at least since 1971 [64]
that if the random spring parameter is Gaus-
<llxollexPftS.[A(s)] ds (2.5.18) sian white noise, then the undamped second
order oscillator is unstable with probability
Blankenship shows that the Kozin-Wu ap- one.
proach is an application of Copper's lemma. In a recent paper Kotani has shown that if
Thus, it may be possible through this general the randomly varying spring term is sta-
measure /~, to extend the sufficiency condi- tionary, ergodic and bounded (in general the
tion of [59] further for systems of the form range of values belong to a compact state
(2.5.6). space), then the solution is unstable almost
Almost sure sample stability problems re- surely. Again, in this problem the unbounded
main to be studied further for applications in (for example, Gaussian) case is still of inter-
the physical noise parameter case. est.
For exact stability boundaries in general, if Although much of the theory for sample
we denote the integrand in the right hand stability is for arbitrary n th order coupled
integral of (2.5.7) as Q(s), then if suitable linear systems, the sample stability regions
ergodic properties hold, it follows from the have not been easily obtained for higher order
discussion related to (2.5.7) that we will have systems. In a certain sense this has been due
to the complexity of obtaining useful ap-
lira lfo'Q(s ) ds{ < 0, stability (2.5.19) proximations and numerical studies of the
t t o~ > 0, instability
sufficient conditions.
Therefore, evaluation of this integral is the However, in [65] an extension of the results
key factor in determining the stability proper- of [59] was obtained and applied to coupled
ties for linear systems with random coeffi- systems of the form
cients. This approach was taken in [62] to
attempt to establish it as a viable approach 21 "1-~21 "[-(C1 --]-FI(t))X 1
for studying stability properties statistically +(c2 + F2(t))X2=O
and numerically. However, many statistical 22 + 822 + (c2 + 6 ( t ) ) x , (2.5.20)
questions remain relative to the convergence
rate and asymptotic variance of the limit in
+(c3 + F,(t))X2=O
(2.5.19). where Fl(t), F2(t) are independent Gaussian
This numerical approach was studied later processes.
by Arnold and his co-workers. In particular A sufficient condition of the form
in [63], rather smooth boundaries are pre-
sented for the case of the second order oscil- -fl + [( K11)a/2+ ( K22)a/2]
lator (2.5.13) where F(t) is the Ornstein- [(l, + 12) + 2 [ W , ( E , - ll)
Uhlenbeck Gaussian process. We have no
details, however, on how such a smooth × w 2 ( e 2 - t2)]] < - ¢
(2.5.21)
184

was obtained, where example,

KI1, K22 are functions of (c 1, c 2, O, fi, 11, 12) lim E { X V ( t ) X ( t ) }
Wi = Prob{ F , ( t ) > / i}, i= 1,2 i or, generally, (2.5.24)
E,=E{F,(t)IF,(t)>li}, i 1, 2 f lim E { I I X ( t ) H pp}
tT~
(2.5.22)
where
and the region of sufficiency is maximized by t1
choosing the constants 11, 12 optimally. U x(t)Ill = xf(t), p > o.
The problem was motivated by study of i--I
the stability of moving elastic strips in the Of course for p = 2, this is the second mo-
presence of in-plane loads. Another approach ment stability problem, studied for many years
to such problems with physical noise coeffi- by [67], [68], [69], [17] as well as many others
cients can be found in the dissertation by Ly from many countries.
[66], written under the direction of Professor The important difference between the
Ariaratnam. physical noise case discussed in Section II
and the white noise case is that for integer p,
2.5.3 Stability properties for stochastic the p th moments of the process generated by
differential equations with Gaussian white the linear equation (2.5.23) can be obtained
noise coefficients--It6 equations explicitly. This is due to the fact that the
following relation holds,
The linear It6 equation is certainly the d
most widely studied and understood, due d-~E [4~(X)] = E[5'~x(~(X)] (2.5.25)
basically to the m a n y analytical tools that are
available for determining the properties of its where ~ is the backward diffusion operator
solutions. which has been defined previously in eqn.
The general linear homogeneous It6 equa- (2.3.23). It is a second order operator on the
tion may be written as, components of the X-vector. Indeed for in-
/
tegral powers of the components of the vector
X, the right hand side will also yield the same
d X ( t ) = A X ( t ) d t + y" C X ( t ) dB~(t),
r=l
integral powers which lead to a closed set of
(2.5.23) equations for solution of the moments. This
idea was applied at least as early as 1962 in
where X is an n-vector, A, Cr, r = 1 . . . . . l, [17], and has been used many times since
are constant n × n matrices, and the Br(t), then.
r = 1 . . . . . l are independent Wiener processes. In particular for the second order oscillator
That is, they are zero mean stationary, inde- with white noise spring term,
pendent increment processes with E{ B 2(t) } X(t)+2~J(+(w2+oB)X=O, (2.5.26)
=ozt. The increments dBr(t) are rigorous
interpretations of the so-called ideal Gaussian (Here B denotes a unit ideal Gaussian white
white noise. The solution processes of (2.5.23) noise), we can write it in It6 linear system
are Markov diffusion processes and thus there form as
is quite a large body of analyses available to d Xl = X2 dt
study them. As in the case of the physical
noise coefficient system we are concerned with d X 2 = - (2~X2 + oa2Xl) dt+Xlo dB
asymptotic properties of the moments, for (2.5.27)
 185

In this case the backward diffusion operator Khasminskii's approach very simply by look-
(also referred to as the generator of the solu- ing at ordinary differential equations.
tion process) is of the form We have already seen from equations (2.5.7)
and (2.5.8) that for the Q function defined in
~ x = X2 ~_~l _ (Z;X2 + w2X1 ) a (2.5.19) the exact stability boundaries may be
- aX2
obtained in theory. The point is, however,
02 22 that the limit in (2.5.19) must be evaluated.
+ -~- X 2 3X 2 (2.5.28) For the general physical noise parameter case
this is, as yet, not possible analytically. How-
Thus, for qffX) given respectively by )(12, ever, for the ideal Gaussian white noise (It6)
X1X2, X 2 we find the second moment equa- case, this is possible as first noted by Khaz-
tions as minskii.
[ E(X 2 } 0 2 0 Let us consider the linear It6 system. With
A = (aij), C, = (ci'j), and E(Bi(tl)Bj(t2) } =
d El{XIX2}I = -w 2 -2; 1
dt I 3,j min(t 1, t2), we can write the backward
/ E{ X 2 } o2 --2ta 2 -4; operator as
O O2
e(x}} £ P = a i ; X J - ~ i + ½cq(X) 3XiOXj , (2.5.31)
X E{ XlX2} (2.5.29)
E(x } where
= c;kc;,.XkX,.;
and whose solution can easily be seen to be
asymptotically stable (that is, mean square r = 1 , 2 .... ,/; k = 1 , 2 . . . . . n; m = l , 2 ..... n
stability) if [17] and we define C( X) = ( cij( X)).
4~2; > 02 (2.5.30) Making the change in variables

These ideas extend to higher order mo- X(t) = X(t)/[I X(t)[I, (2.5.32)
ments as well, [70,71]. In [53], for example, where ]l" II denotes Euclidean norm, Khaz-
Willems notes that if the eigenvalues of the minskii recognized that the Markov process
matrices C, are not all imaginary, then there generated by (2.5.32) is mapped onto the
is no value of the noise intensity that would surface of the n-dimensional unit sphere and
yield exponentially stable p t h moments, for is a Markov process, satisfying an It6 dif-
all p. ferential equation. Furthermore, if we define
Indeed, if C, has an eigenvalue with non- p = l o g tlXII, the It6 differential formula
zero real part, then there are unstable p-mo- yields
ments, for p large enough. He considers other
examples as well.
do = Q(]k ) dt + ]kfc,]k dBr, (4.5.33)
The explicit determination of the asymp- where
totic behavior of the sample solutions of
Q(k) = ~.TA ]K + ½trC(k) - •Tc(k),
(2.5.23) in the It6 case was studied by Khas-
minskii [72], who derived the fundamental (4.5.34)
result which can yield a necessary and suffi-
where tr denotes trace.
cient condition for sample stability. This al-
Indicating the dependence of the k ( t ) pro-
lows, in principle, the determination of the
cess on its initial condition X 0, by
exact boundary of the region of stability for
the linear It6 equation. We can motivate xx°(t) = x °(t)/ll II,
186

where hold. Upon setting

Xo = x0/ll x0 II, )`1 = cos 4,, )`2 = sin 4',
we have
the backward operator for the 4' process is of
Theorem (Khasminskii [72]).
the form
If there exist n linearly independent vec-
tors X, . . . . . k , in E, such that with prob- d2 d
ability one ½a2(4') -d4'
- +2m ( 4 ' ) d4' (2.5.38)

limt-lfotO(kX'(r))dr<O (2.5.35)
l ---~ oO
From diffusion process theory, [73], we have
for i = 1. . . . . n, then the system (2.5.1) is al- Definition
most surely asymptotically sample stable. If, A singular point of the ~process is any
however, for any )% in E, point satisfying 02(4')= 0.
If a2(4')= 0, m(4')~: 0, then 4' is called a
lim t 1
I - - * OO
f'o
"0
( XXo( • )) d r > 0 (2.5.36) shunt.
If a2(4')--0, m ( 4 ' ) = 0, then 4' is called a
then the system (2.5.1) is unstable in the sense trap.
P { l i m t ~ [I X*°(t) 11 = oo) = 1.
In general, (2.5.35) and (2.5.36) are quite At a shunt, the process can only pass
difficult to apply without an ergodic property through in one direction. At a trap the pro-
to enable their evaluation. However, as can be cess will remain there forever.
shown by examples, they are very useful in For the second order linear equation, it can
specific instances. It should be pointed out be shown that there are at most four singular
that (2.5.35) provides a method for comput- points.
ing the sample stability boundaries for linear The ergodic properties of the @process are
stochastic differential equations of order completely determined by the properties of
higher than two. One procedure would be to these singular points. For example, for the
generate solutions X(t) on a digital computer second order oscillator
for a linearly independent set of k i and con-
struct the time averages indicated in (2.5.35)
dXl(t)= X2(t) dt
for each ki, then statistically test the signs of dX2(t ) = - [2~'wX2(t ) + ~2Xl(t)] dt
these time averages for large t. This was ini-
-Xl(t ) dB(t) (2.5.39)
tially done in [62].
These results have been applied in [47] and the only singular points are shunts located at
[64] to determine the exact sample stability __+~'/2 [47,64].
boundaries for the general second order linear The 4'-process moves in a clockwise direc-
It6 differential equation, tion around the unit circle and is an ergodic
{ d X~ = X2 dt (recurrent) process on the unit circle. Hence,
the entire circle is essentially the only ergodic
d X 2 = -[G02 d t + a l d B l ( t ) ] X 1 (2.5.37)
component for the 4'-process. For o~ = 1, the
[2~o~ at + 02 dB2(t)] X 2 oscillator (2.5.39) yields

For the second order equations (2.5.37) the E{ Q()`(4'))}

),-process is a one dimensional diffusion de-
fined on the boundary of the unit circle. = E { TCOS24'
°2 COS 24' -- 2~" sin24' }
Indeed for any second order system this will
 187

The interesting point is that the It6

= 2f~;zQ(X(¢))
equivalent equations are of exactly the same
form as the original physical noise equations
x I s ( - 2 ) - S(~p)] dM(q~) (2.5.40)
for systems (2.5.42). That is, the correction
terms are zero. Thus, the stability boundaries
where
for the wide band Gaussian coefficient pro-
dS(~) 1 2 cess in (2.5.42) will be the same as for the
exp ~5oztan ¢
d~ cos2q~ Gaussian white noise (It6) case. In fact it can
be shown that there are non-zero corrections
× (3 + 3~'tan q~+ tan2¢)] only when the random noise coefficient multi-
plies terms of order one less than the order of
dM(q~) 2 [2 the differential equation.
exp - -~02 tan q~ What about higher order linear It6 equa-
dq5 o 2 cos2q5
tions. In theory the technique that we have
X (3 + 3~'tan ~ + tan2~)] described can be used to obtain exact stabil-
ity boundaries for the higher order case.
(2.5.41) Unfortunately, there exist basic problems
The functions S, M are the so-called speed in applying this approach. Again noting that
and scale measures of the diffusion process. the X-process defined by (2.5.32) is a unit
Upon evaluating the integral (2.5.40) where norm vector defined on the surface of the
S, M are given by (2.5.41) the exact sample n - 1 dimensional unit sphere, we must study
stability region for (2.5.40)can be obtained. the expected value of Q(X) on the sphere.
For the undamped oscillator, However, for higher dimensional spherical
surfaces, we cannot classify the singularities
dXl(t)= X2(t) dt sufficiently well to determine the associated
dX2(t ) = -W2Xl(t) d t - Xl(t ) d B ( t ) ergodic properties of the diffusion process.
(2.5.42) The singularities on the n-dimensional spheri-
cal surface are whole curves or boundaries
we find that the evaluation of (2.5.40) for instead of simple points. Also the diffusion
(2.5.41) with ~"= 0 yields E{ Q(X(¢))) > 0 for can behave in many different ways near these
any o 2 > 0. Hence, the undamped oscillator boundaries.
with white noise coefficients is always unsta- More study will have to be made of the
ble. singularities and ergodic properties of diffu-
Many examples with the exact stability sions defined on general manifolds, in order
boundaries are contained in [47]. A study of for us to be able to apply the above tech-
the general second order system case can be niques.
found in [74]. A recent study for random Of course if the diffusion has no singulari-
telegraphic coefficient processes using this ap- ties on the surface of the unit n-sphere, then
proach appears in [75]. we know that the process is ergodic. How-
One point of importance is related to the ever, most physical systems of interest gener-
results in [47]. The idea was to study the ate diffusion processes that have singularities.
stability of second order linear oscillators with Therefore, the question of exact stability re-
wide band excitations in the coefficients. gions for higher order linear It6 equations
Hence, the Wong-Zakai [19], or Stratonovich remains an important open question for ap-
[15] correction terms were included in the plications.
study of the It6 equivalent equations (2.3.18). The connection between the asymptotic
188

properties of moments and of samples of the n th moment. This is not necessary for sample
solution processes to stochastic differential stability as given by the region (2.5.45). As n
equations has been of interest for many years. increases, the stability region decreases and
A partial answer to this problem was estab- lies below the line 02 = 2 a / ( n - l), (a > 0) in
lished quite recently [76] for linear It6 sto- the (a, o 2) plane.
chastic differential equations. This illustrates the known facts about sta-
We shall motivate the results by studying bility for the first order It6 differential equa-
the simple first order stochastic differential tion. The regions of stability for higher in-
equation with a white noise coefficient. teger moments are included in the regions of
We consider the first order It6 equation stability of lower integer moments, and all
dX(t)+adt+oX(t)dB(t)=O, (2.5.43) integer m o m e n t stability regions are included
in the region of sample stability.
where a and o are constants and B(t) repre- Going back to (2.5.46) we notice that the
sents a unit Brownian Motion process with equality hold for all p > 0. Therefore, let us
zero mean and variance t. break away from traditional studies of integer
It is well known that the solution process moments and considered the vase 0 < p < l.
of (2.5.43) is For this range of values of p, the stability
X( t ) = X o e x p ( - ( a + o2/2)t - oB( t ) ). region obtained from (2.5.46) is
(2.5.44) o2>2a/(p-1),(O<p<l). (2.5.48)

Again, it is well known that since B(t) grows A rather curious and interesting result fol-
like ~t log log t with probability one, then the lows from (2.5.48). We see that as p ap-
stability properties of the sample solutions proaches zero, the half line that bounds the
are determined by the deterministic term in stability region for the p t h m o m e n t (the other
the exponent of (2.5.44) Hence, the region of boundary is the positive a-axis) rotates coun-
sample stability is clearly given by a + o 2 / 2 terclockwise and approaches the half line that
> 0 , i.e. bounds the region of sample stability. It ap-
pears, at least for the simple stochastic equa-
O2 > -2a. (2.5.45) tion (2.5.43), that the region of sample stabil-
We now consider the m o m e n t s of the solu- ity can be characterized as the union of all
tion process (2.5.44) The absolute p t h mo- p t h m o m e n t stability regions for p > 0. The
ment is simply found to be next obvious question is, does this property
hold more generally?
E { I X ( t ) I p} = Ix0l p e x p ( - p ( a + o Z / z ) t ) For the linear It6 system, the answer is yes,
×E{exp(-poB(t))} as shown in [76].
Very recently, Arnold [77], has extended
=Ix01 p exp 7 [ ( p - 1)o 2 - 2a] the results in [76] to include linear systems of
the form
(2.5.46)
X( t) = A ( t ) X ( t ) , (2.5.49)
where X(0) = x o with probability one.
The region of stability for the integer mo- where the stochastic elements of the matrix
ment p = n( > 1) as given by (2.5.46) is A(t) are stationary ergodic processes.
Unfortunately, this explicit connection does
o 2 < 2 a / ( n - 1). (2.5.47) not lead us directly to the sample stability
F r o m (2.5.27) it is obvious that the constant a boundary, since the limiting m o m e n t stability
must be greater than zero for stability of the boundaries are not directly obtainable as p $ 0
 189

for higher order systems. In any case the Then [5]

exact connection between moment stability
and sample stability does exist and may J_ooZPXrz(4, O, z, t) dz
ultimately lead to a means of determining the
exact stability boundaries or, at least, good
p(4, t)=ff oo dx f_o p yz(x,O, z, t)
approximations to the exact boundaries. These (2.6.2)
questions are still open for study.
For the case of a narrow band process, eqn.
(2.6.1) may be approximated by
d oO
2.6 STRUCTURAL FAILURE UNDER RAN-
DOM EXCITATIONS d4 I YPx, r(4, Y) dy
P(4, t ) - - -o N+(4, t) (2.6.3)

2.6.1 Types of failure so that only the joint probability density of X

and Y is required.
A structure may fail when its response to
random excitation crosses out, for the first 2.6.3 Envelope processes, excursion
time, a safe domain or its strength continues probability
to deteriorate to eventual fracture. These two
modes of failure are known as first passage Several definitions have been proposed for
and fatigue, respectively. Analyses related to the envelope of a stationary random process
structural failures include the statistics of level [86,108,109]. The concept has also been ex-
crossings [86], the distribution of peak magni- tended to nonstationary processes [110,111].
tude given the occurrence of a peak [5] etc. Of special interest to the first excursion prob-
The deterioration of material strength has lem is the discretized form of the envelope, in
been quantified as cumulative damage [93], which the realizations of the envelope process
and more recently as fatigue crack propa- a i are sampled at the peaks of the process
gation. X(t). The excursion probability P E of a nar-
row band process x may then be approxi-
mated by the probability that at least one
2.6.2 Level crossings and peak statistics
single realization of the discrete envelope pro-
cess exceeds the threshold level. Thus [87]
If the joint probability density
Px, r (x, y, t) of the response X(t) and its PE = 1 -- P(a a < 4 ) P ( a 2 < 4[a I < 4 ) . . -
time derivative Y(t) is known, the expected × . P ( a , < 4 l a l < 4 A ... A a , _ l < 4 )
number of crossings over a threshold 4 from (2.6.4)
below, N+(4, t) is found to be [86]
in which n is the expected number of zero
N+(4, t ) = L y ' P x ' r ( 4 ' Y, t) d x d y
crossings. The conditional probabilities in eqn.
(2.6.1) (2.6.4) are usually approximated. The sim-
plest approximation is the assumption of in-
The distribution of peak height p.-- (4, t) dependence between successive crossings,
conditional on the occurrence of a peak re- [87,88]. In that case
quires the knowledge of the joint probability
density of the response X(t) and its first and PE = 1 -- P(al < 4)P(a2 < 4)... P ( a , < 4)
second derivatives Y(t), Z(t) respectively. (2.6.5)
190

in which P ( a i < ~ ) may easily be derived better results [94]. For more detailed analysis
from eqn. (2.6.3). In general this assumption of the physical background to cumulative
leads to conservative results, especially for damage we refer to the report of Subcommit-
lightly damped systems. Since successive tee Nr. 4. The mean and variance of the
crossings are not expected to be independent, damage D in one cycle may be obtained from
an improved approximation is to assume the probability density function of the peaks,
Markovian ("one-step-memory") property. Pz (~) utilizing
Then D = a~ k (2.6.7)
PE = 1 - P ( a , < ~ ) P ( a 2 < ~ ] a 1 < ~ ) . . .
in which a, k are material parameters. Tak-
×P(a, < ~ I a,,_, < ~) (2.6.6) ing into account the fact that failure is in
general caused by a large number of load
The conditional probabilities in eqn (2.6.6)
cycles, the central limit theorem may be in-
may be found for both linearly and nonlin-
voked to obtain the PDF of total damage in
early damped systems [87,89]. Further im-
lifetime. Thus the probability of fatigue is
provement may be obtained by considering a
obtained by a cumulative Gaussian distribu-
"two-step-memory" process, as shown for lin-
early and nonlinearly damped systems [83]. tion [88,83]. This probability is, of course,
conditional on particular realizations of the
The dependence between subsequent peaks
may also be taken into account in terms of material parameters a, k (which are in gen-
eral random variables) and on the fact that
the average clump size, i.e. the average num-
ber of successive peaks in each group above excursion failure does not occur.
threshold [5,90]. Since the excursion probabil-
it,y satisfies the Kolmogorov backward equa-
tion, several methods utilizing this fact are 2.7 CONCLUDING REMARKS
available [112]. In general the moments of
first passage time can be evaluated [89] and The review of the general field of stochastic
the probability density of first passage time structural dynamics presented in the fore-
ap- going ranges from the very elementary to the
proximated (e.g. using maximum entropy very advanced, and in several occasions it
criteria) [113]. It should also be mentioned touches upon some very recent research topics.
that numerical (Finite Element) solutions of Due to the space and resource limitations, not
the backward equation are available for linear only must the discussions be brief on the
and nonlinear single-degree-of-freedom sys- subjects we do cover, but omissions entirely
tems in stationary and non-stationary cases of other subjects are deemed unavoidable. We
[91,92]. Also, approaches to the first passage have, attempted at a compromise in the
problem for non-stationary response processes coverage so that in conjunction with compa-
are treated in [114] and are reviewed in [105]. nion reviews from other sub-committees, the
entire package can be useful to as large a
2.6.4 Cumulative damage readership as possible. It should be emphati-
cally pointed out that many subjects not
Applying the well-known Palmgren-Miner covered in this review are by no means unim-
hypothesis [93] cumulative damage for varia- portant from the viewpoint of stochastic
ble amplitude loading may be assessed. Al- structural dynamics.
though this hypothesis does not agree very One conspicious omission is the publica-
well with experimental data, more com- tion on simulation techniques, known collec-
plicated theories do not provide consistently tively as the Monte Carlo method. For non-
 191

linear p r o b l e m s w h i c h are too c o m p l i c a t e d to 12 A. Der Kiureghian, A. Response spectrum method

a t t e m p t a n a n a l y t i c a l solution or w h e n the for random vibration analy0is of MDF systems,
Earthquake Eng. Struct. Dynamics, 9 (1981)
a c c u r a c y of a n a p p r o x i m a t e solution is b e i n g
419-435.
q u e s t i o n e d , s i m u l a t e d s a m p l e results m i g h t b e 13 T.K. Caughey, Nonlinear theory of random vibra-
the o n l y a v a i l a b l e recourse. A r e p r e s e n t a t i v e tions, In: Advances in Applied Mechanics, 11.
r e f e r e n c e in this a r e a is a p a p e r b y Vaicaitis, Academic Press, 1971, pp. 209-253.
S h i n o z u k a a n d T a k e n o [78]. In o r d e r to o b - 14 Y.K. Wen, Approximate method for non-linear
tain m e a n i n g f u l averages b y simulation, a random vibrations. J. of Eng. Mech. Div., ASCE,
101 (1975) 389-401.
large n u m b e r of s a m p l e s m u s t b e g e n e r a t e d ;
15 R.L. Stratonovich, Topics in the Theory of Ran-
thus, the c o m p u t a t i o n a l cost c a n b e quite dom Noise, 1. Translation by R.A. Silverman,
high. This cost c a n be r e d u c e d using a p r o c e - Gordon and Breach, 1963.
dure, called the r e s p o n s e surface m e t h o d , as 16 R.Z. Khasminskii, A limit theorem for the solution
s h o w n b y V e n e z i a n o , Casciati a n d Faravelli of differential equations with random right hand
sides. Theory of Probability and Applications, 2
[79], in a p a p e r on seismic p e r f o r m a n c e
(1966) 390-405.
evaluation. 17 J.L. bogdanoff and F. Kozin, Moments of the
output of linear random systems. J. Acoustical Soc.
Am., 34 (1962) 1063-1066.
REFERENCES 18 K. Ito, On stochastic differential equations, Mem-
1 G. Augusti, A. Baratta and F. Casciati, Probabilis- oirs of the American Mathematical Society, 4 (1951)
tic Methods in Structural Engineering, Chapman 51.
and Hall Ltd., London, New York, 1984. 19 E. Wong and M. Zakai, On the relation between
2 V.V. Bolotin, Statistical Methods in Structural ordinary and stochastic equations. Int. J. Eng. Sci.,
Mechanics, Holden-Day, San Francisco, 1969. 2 (1965) 213-229.
3 S.H. Crandall and W.D. Mark, Random Vibration 20 K. Ito, On a formula concerning stochastic dif-
of Mechanical Systems, Academic Press, New York, ferentials. Nagoya Math. J. Jap., 3 (1951) 55.
N.Y., 1963. 21 W.F. Wu and Y.K. Lin, Cumulant-neglect closure
4 I. Elishakoff, Probabilistic Methods in the Theory for non-linear oscillators under random parametric
of Structures. John Wiley & Sons, Inc., New York, and external excitations. Int. J. Non-linear Mech.,
N.Y., 1983. 19 (1984) 349-362.
5 Y.K. Lin, Probabilistic Theory of Structural Dy- 22 S.H. Crandall, Perturbation techniques for random
namics. Mc Graw-Hill Book Co., Inc., New Yrok, vibrations of nonlinear systems. J. Acoustical Soc.
N.Y., 1967. (Reprint by Robert E. Krieger Publish- Am., 35 (1963) 1700-1705.
ing Co., Melbourne, FL, 1976). 23 R.C. Booton, The analysis of nonlinear central
6 D.E. Newland, An Introduction to Random Vibra- systems with random inputs, IRE Trans. Circuit
tion and Spectral Analysis. Longmans, London, Theory, 1 (1954) 32-34.
1975. 24 T.K. Caughey, Equivalent linearization technique.
7 N.C. Nigam, Introduction to Random Vibrations. J. Acoustical Soc. Am., 35 (1963) 1706-1711.
MIT Press, Cambridge Mass., 1983. 25 W.O. Iwan, A generalization of the concept of
8 J.D. Robson, An Introduction to Random. Vibra- equivalent linearization. Int. J. Nonlinear mech., 8
tion, Edinburgh University Press, 1963. (1973) 279-287.
9 G.I. SchuEller, Einffahrung in die Sicherheit und 26 P.D. Spanos and W.D. Iwan, On the existance and
Zuverl~tssigkeit von Tragwerken, Wilhelm Ernst & uniqueness of solution generated by equivalent lin-
Sohn-Verlag, Berlin, Mi~nchen, 1981. earization. Int. J. Nonlinear Mech., 13 (1978) 71-78.
10 Y.K. Lin, S. Maekawa, H. Nijim and L. Maestrello, 27 P.T.D. Spanos, Formulation of stochastic lineariza-
Response of periodic beam to supersonic bounda- tion for symmetric or assymetric M.D.O.F. nonlin-
ry-layer turbulence. In: B.L. Clarkson (Ed.), Sto- ear systems. J. Appl. Mech., Trans. ASME, 47
chastic Problems in Dynamics. Pitman, London, (1980) 209-211.
1977, pp. 468-486. 28 I.E. Kazakov, Statistical analysis of systems with
11 Y.K. Lin and W.F. Wu, Along-wind motion of multi-dimensional non-linearities, Automation and
building on compliant soil. J. Eng. Mech., 110 (1) Remote Control, 26 (1965) 458-464.
(1984) 1-19. 29 T.K. Caughey, Response of Van Der Pors oscilla-
192

tor to random excitation. J. Appl. Mech. Trans. it) of coupled linear systems under colnbincd
ASME, 26 (1959) 345 348. harmonic and stochastic excitations. In: B.L.
30 L.E. Borgman, Spectral analysis of ocean wave Clarkson (Ed.), Stochastic Problems in l)vnanlics.
forces on piling. J. Waterways Harbours Div. Proc. Pitman, London, 1977. pp. 90 105.
ASCE, 93 (1967) 126-156. 47 R.R. Mitchell and F. Kozin, Sample stability of
31 T.K. Caughey, Random excitation of a system with second order linear differential equations with wide
bilinear hysteresis, J. Appl. Mech. Trans. ASME, band noise coefficients, SIAM J. Appl. Math.. 27
27 (1960) 649 652. (1974) 571 605.
32 W.O. Ivan and L.O. Lutes, Response of the bilin- 48 F. Weidenhammer, Stabilit~itsbedingungen fiSr
ear hysteretic system to stationary random exci- schwinger mit zuf',~lligen parametererregungen, ln-
tations. J. Acoustical Soc. Am., 43 (1968) 545-552. genieur Archiv, 33 (1964) 404--415.
33 Y.K. Wen, Equivalent linearizalion for hysteretic 49 S.T. Ariaratnam, Dynamic stability of a column
systems under random excitation, J. Appl. Mech. under random loading. In: Proc. Int. Conf. on
Trans. ASME, 47 (19800 150-154. Dynamic Stability of Structures, Pergamon, New
34 T.T. Baber and Y.K. Wen, Random vibration of York, 1967, pp. 267.
hysteretic degrading systems. J. Eng. Mech. Div. 50 W. Wedig, Regions of Instability for a Linear
ASCE, 107 (EM6) (1981) 1069-1087. System with random parametric excitation. Lecture
35 Y.K. Wen, Stochastic response and damage analy- Notes in Math., 294, Springer Verlag, New York,
sis of inelastic structures. Int. J. Probabil. Eng. 1972, pp. 160-172.
Mech. 1 (1) (1986) 49 57. 51 J.C. Samuels, On the mean square stability of
36 F. Casciati and L. Faravelli, Reliability assessment random systems. Trans. IRE PGIT, 5 (1959) 248.
for non-linear random frames, In: Proc. of Weibull 52 R.W. Brockett, Lie theory and control systems
1UTAM Symposium, Stockholm, Sweden, June, defined on spheres. SIAM J. Appl. Math., 25 (1973)
1984. 213 225.
37 F. Kozin, A survey of stability of stochastic sys- 53 J.L. Willems, Moment stability of linear white noise
tems, Automatica, 5 (1969) 95 112. and colored noise systems. In: B.L. Clarkson (Ed.),
38 F. Kozin, Stability of the linear stochastic system. Stochastic Problems in Dynamics. Pitman, London,
Lecture Notes in Math., Springer Verlag, 294, 1972 1977, pp. 67 89.
pp. 86-229. 54 A.S. Willsky, S.I. Marcus and D.N. Martin, On the
39 L. Arnold and W. Kliemann, Qualitative theory of stochastic stability of linear systems containing col-
stochastic systems. In: A.H. Bharucha-Reid (Ed.), ored multiplicative noise, IEEE Trans. AC. 20
Prob. Analysis and Related Topics, Academic Press, (1975) 711-713.
3, 1981. 55 A. Kistner, On the moments of linear systems
40 W. Kliemann, Qualitative Theorie Nichtlinearen excited by a colored noise process. In: B.L. Clark-
Stochastischer Systems. Report No. 32, Dept. of son (Ed.), Stochastic Problems in Dynamics. Pit-
Math, Un. of Bremen, Aug. 1980. mann Process, 1977, pp. 36-53.
41 Y. Sunahara et al. Noise stabilization of non-linear 56 E.R. Infante, On the stability of some linear non-
dynamical systems. Trans. Japan Soc. Mech. Eng., autonomous random systems. ASME J. App.
40 (1974) 338. Mech., 35 (1968) 7 12.
42 R.L. Stratonovich, A new representation for sto- 57 F. Kozin, On almost sure stability of linear systems
chastic integrals and equations. SIAM Journal with random coefficients. J. Math. and Phys., 42
Control, 4 (1966) 362-371. (1963) 59-67.
43 E. Wong and M. Zakai, On the convergence of 58 T.K. Caughey and A.H. Gray, Jr., On the almost
ordinary integrals to stochastic integrals. Ann. sure stability of linear dynamic systems with sto-
Math, Statist., 36 (1965) 1560-1564. chastic coefficients. ASME J. Appl. Mech., 32
44 R.Z. Khazminskii, On stochastic processes defined (1965) 365.
by differential equations with a small parameter. 59 F. Kozin and C.M. Wu, On the stability of linear
Theory Prob. and Appls., 11 (1966) 211-228. stochastic differential equations. J. Appl. Mech., 40
45 Y.K. Lin, Structural response under turbulent flow (1973) 87-92.
excitations. In: H. Parkus (Ed.), Random Excita- 60 G. Blankenship, Stability of linear differential
tion of Structures by Earthquake and Atmospheric equations with random coefficients, IEEE Trans.
Turbulence, CISM Lectures, No. 225, Springer- AC, 22 (1977) 834-838.
Verlag, Wien, New York, 1977. 61 W.A. Coppel, Stability and Asymptotic Behavior of
46 S.T. Ariaratnam, and D.S.F. Tam, Moment stabil- Differential Equations. Heath, Boston, 1965.
 193

62 F. Kozin and S. Sugimoto, Decision criteria for 77 L. Arnold, A. Formula Connecting Sample and
stability of stochastic systems from observed data. Moment Stability of Linear Stochastic Systems,
In: B.L. Clarkson (Ed.), Stochastic Problems in Report No. 92, Dept. of Math., Un. of Bremen,
Dynamics. Pitman, London, 1977, pp. 8-35. May 1983.
63 W. Kliemann and W. Rumelin, On the Growth of 78 R. Vaicaitis, M. Shinozuka and M. Takeno, Re-
Linear System Parametrically Disturbed by a Dif- sponse analysis of tall building to wind loading, J.
fusion Process, Report No. 27, Department of Struct. Div. ASCE, 101 (ST3) (March, 1975)
Mathematics, Un. of Bremen, 8 July 1981. 585-600.
64 F. Fozin and S. Prodromou, Necessary and suffi- 79 D. Veneziano, F. Casciati and L. Faravelli, Meth-
cient conditions for almost sure sample stability of ods of seismic fragility for complicated systems. In:
linear Ito-equations. SIAM J. Appl. Math., 21 Proc. of. 2nd CSNI Specialist Meeting on Prob.
(1971) 413-424. Methods in Seismic Assess. for Nuclear Power
65 R.M. Milstead and F. Kozin, The stability of a Plants, Livermore, 1983, pp. 67-88.
moving elastic strip subjected to random paramet- 80 H.J. Melzer and G.I. Schu~ller, On the reliability of
ric excitation. ASME J. Appl. Mech., 46 (1979) flexible structures under non-normal loading
404-410. processes, In: K. Henning (Ed.), Proc. of IUTAM-
66 B.L. Ly, Topics in the Stability of Stochastic Sys- Symp. on Random Vibrations and Reliability.
tems. Ph.D. Dissertation, Dept. of Civil Eng., Un. Akademie Verlag, Berlin, 1983, pp. 73-83.
of Waterloo, Ontario, Canada, 1974. 81 N.C. Hampl, Non-Gaussian stochastic analysis of
67 J.C. Samuels, On the stability of random systems nonlinear systems, In: F. Casciati, L. Faravelli
and the stabilization of deterministic systems with (Ed.), Proc. 2nd Int. Workshop on Stochastic
random noise. J. Acoust. Soc. of America, 32 (1960) Methods in Structural mechanics, Univ. Pavia,
594-601, 1356. 1986, pp. 243-254.
68 W. Wedig, Stabilit~it stochastischer Systeme. 82 C.G. Bucher, The effect of damping on the relia-
ZAMM, 55 (1975) 185-187. bility of structures under dynamic loading. In: I.
69 M.B. Nevelson, Some remarks ont he stability of a Konishi, A.H.-S. Ang and M. Shinozuka (Eds.),
linear stochastic system, Jour. Appl. Math. and Proc. 4th Int. Conference on Structural Safety and
Mech., 30 (1966) 1332-1335. Reliability, (ICOSSAR' 85), Kobe, Japan. IAS-
70 R.W. Brockett, Parametrically stochastic linear dif- SAR, New York, 1, 1985, pp. 607-614.
ferential equations. Math. Programming Study, 5 83 G.I. Schu~ller, C.G. Bucher, Nonlinear damping
(1976) 8-21. and its effect on the reliability estimates of struc-
71 R.W. Brockett, Lie algebras and Lie grou0s in tures, In: I. Elishakoff and R.H. Lyon (Ed.), Ran-
control theory, In: Brockett and Mayne (Eds.), dom Vibration--Status and Recent Developments,
Geometric Methods in System Theory. Redidel. The Stephen Harry Crandall Festschrift. Elsevier
Boston, 1973. Science Publ. Co., Amsterdam, 1986 pp. 389-402.
72 R.Z. Khasminskii, Necessary and sufficient condi- 84 G. Lindgren, Jumps and bumps on random roads.
tions for the asymptotic stability of linear stochas- J. Sound Vibration, 78 (1981) 383-395.
tic systems, Theor. Prob. and Appls., 12 (1967) 85 O. Ditlevsen, Elasto-plastic oscillator with Gaus-
144-147. sian excitation. DCAMM Report 291, Technical
73 K. Ito and H.P. McKean, Diffusion Processes and University of Denmark, Lyngby, Denmark, Aug.
Their Sample Paths. Springer-Verlag, New York, 1984.
1965. 86 S.O. Rice, Mathematical analysis of random noise.
74 K. Nishioka, On the stability of two-dimensional Bell System Tech. J., 23 ((1944) 282-332, 24 (1945)
linear stochastic systems. Kodai Math. Sem. Rep., 46-156. Repr. in N. Wax (Ed.), Selectred Papers
27 (1976) 211-230. on Noise and Stochastic Processes'. Dover, New
75 K.A. Loparo and G.L. Blankenship, Almost sure York, 1954.
instability of a class of linear stochastic systems 87 J.N. Yang and M. Shinozuka, On the first excur-
with jump process coefficients. Proc. Workshop in sion probability in stationary narrow-band random
Liapunov Exponents. Springer-Verlag Lecture vibration, II. J. Appl. Mech. 39 (3) (1972) 733-738.
Notes in Math, Nov. 1984, University of Bremen. 88 W.-Q Zhu, A method for estimating reliability of
76 F. Kozin and S. Sugimoto, Relations between sam- structures subject to quasi-stationary a n d / o r
ple and moment stability for linear stochastic dif- quasi-homogeneous Gaussian random excitations.
ferential equations. In: D. Mason (Ed.), Proc. Conf. Structural Safety and Reliability, Vol. I, Proc.
Stochastic Diff. Eq. Academic Press, 1977. pp. ICOSSAR'85, Kobe, May 27-29, Japan, pp 1-
145-162. 219-228.
194

89 J.B. Roberts, First passage probability for non-lin- quakes, In: C. Lomnitz and E. Rosenblueth (Eds.),
ear oscillators. J. Mech. Div. Proc. ASCE, 102, Seismic Design and Engineering Decisions, Elsevier
(EM5) (Oct. 1976) 851 866. Publ., Amsterdam, 1976, Chap. 8.
90 E.H. Vanmarcke, On the distribution of the first- 102 E.H. Vanmarcke, P.I. Yanev and M.B. De Estrada,
passage time for normal stationary random Response of Simple Hysteretic Systems to Random
processes. ASME, J. Appl. Mech., 42 (March 1975) Excitation. Department of Civil Engineering, M.I.T.
215-220. Research Report No. R70-66, 1977.
91 L.A. Bergman and B.F. Spencer, Solution of the 103 N.R. lyengar and J.K. lyengar, Stochastic analysis
First Passage Problem for Simple Linear and Non- of yielding system, J. Eng. Mech. Div. ASCE, 104
linear Oscillators by the finite Element Method. (1978) 383-398.
T.A.M. Report Nr. 461, Univ. Illinois, Urbana- 104 D. Karnopp and T.D. Scharton, Plastic deforma-
Champaign, I l l . 1983. tion in random vibration. J. Acoustical Soc. Am.,
92 L.A. Bergman and B.F. Spencer, First passage of a 39 (1966) 1154-1161.
rigid structure on a sliding foundation. Struct. 105 E.H. Vanmarcke, Some recent developments in
Safety and Reliability, I. Konishi, A.H.-S. Ang and random vibration, App. Mech. Rev., 32 (10) (1979)
M. Shinozuka (Eds.) Proc. ICOSSAR'85, Kyoto, 1197-1202.
1985. 106 E.H. Vanmarcke and D. Veneziano, Probabilistic
93 A. Palmgren, Die Lebensdauer von Kugellagern. seismic response of simple inelastic systems, Proc.
VDI, 68 (14) (1924) 339-341. 5th World Conference on Earthquake Engineering,
94 Committee on Fatigue and Fracture Reliability of Rome, 1973.
the Committee on Structural Safety and Reliability 107 G.C. Gazetas, Random Vibration Analysis of In-
of the Structural Division, Fatigue reliability: vari- elastic Multi-Degree-of-Freedom Systems Sub-
able amplitude loading. ASCE J. Struct. Div., 108 jected to Earthquake Ground Motions, M.I.T. De-
(ST1) (Jan. 1982) 47-69. partment of Civil Engineering Research Report
95 T.K. Caughey and H.J. Stumpf, Transient response R76-39, August 1976.
of a dynamic system under random excitation, J. 108 H. Cramer and M.R. Leadbetter, Stationary and
Appl. Mech. Trans. ASME, 28 (1961) 562-566. Related Stochastic Processes, Wiley, New York,
96 R.B. Corotis and E.H. Vanmarcke, Time-depen- 1967.
dent spectral content of system response, J. Eng. 109 S.H. Crandall, The envelope of a lightly damped
Mech. Division, ASCE, (101) (1975) 623-637. nonlinear oscillator, Zagadnienia drgan nielinio-
97 J.K. Hammond, On the response of single and wych, 5 (1964) 120-139.
multi-degree of freedom systems to non-stationary 110 J.N. Yang, Nonstationary envelope process and
random excitations, J. Sound Vibration, 7 (1968) first-excursion probability, J. Struct. Mech., 1 (1972)
393-416. 231-245.
98 M. Chakravorty, Transient Spectral Analysis of 111 J.N. Yang, Simulation of random envelope
Linear Elastic Structures and Equipment Under processes, J. Sound Vibration, 21 (1972) 73-85.
Random Excitation J. Department of Civil En- 112 J.N. Yang and M. Shinozuka, First-passage time
gineering, M.I.T. Research Report No. R72-18, problem, J. Acoust. Soc. Amer., 43(2) (1970)
1972. 393-394.
99 J.B. Roberts, The covariance response of linear 113 Y.K. Lin, On first-excusion failure of randomly
systems to non-stationary random excitation. J. excited structures, I and II, AIAA J., (1970)
Sound and Vibration, 14 (1971) 385-400. 720-725 and 1888-1890.
100 D.A. Gasparini, Response of M D O F systems to 114 J.N. Yang, First-excursion probability in nonsta-
nonstationary random excitation, J. Eng. Mech. tionary random vibration, J. Sound Vibration, 27(2)
Div. ASCE, 105 (1979) 13-27. (1973) 165-182.
101 E.H. Vanmarcke, Structural response to earth-