You are on page 1of 38

On the growth and decay of acceleration

waves in random media


By Martin Ostoja-Starzewski1 a n d J e r z y T r e
bicki2
1

Institute of Paper Science and Technology, and


Georgia Institute of Technology, Atlanta, GA 30318-5794, USA
2
Institute of Fundamental Technological Research,
Polish Academy of Sciences, 00-049 Warsaw, Poland
Received 9 October 1997; accepted 8 December 1998

We study the effects of material spatial randomness on the growth to shock or decay
of acceleration waves. In the deterministic formulation, such waves are governed by
a Bernoulli equation d/dx = (x) + (x)2 , in which the material coefficients
and represent the dissipation and elastic nonlinearity, respectively. In the case
of a random microstructure, the wavefront sees the local details: it is a mesoscale
window travelling through a random continuum. Upon a stochastic generalization
of the Bernoulli equation, both coefficients become stationary random processes,
and the critical amplitude c as well as the distance to form a shock x become
random variables. We study the character of these variables, especially as compared
to the deterministic setting, for various cases of the random process: (i) one white
noise; (ii) two independent white noises; (iii) two correlated Gaussian noises; and
(iv) an OrnsteinUhlenbeck process. Situations of fully positively, negatively or zero
correlated noises in and are investigated in detail. Particular attention is given to
the determination of the average critical amplitude hc i, equations for the evolution
of the moments of , the probability of formation of a shock wave within a given
distance x, and the average distance to form a shock wave. Specific comparisons of
these quantities are made with reference to a homogeneous medium defined by the
mean values of the {, }x process.
Keywords: acceleration wave; random media; mesoscale; stochastic mechanics;
wave propagation; Bernoulli equation

1. Introduction
(a) Microscale heterogeneity versus wavefront thickness
Acceleration waves are moving singular surfaces with a jump in particle acceleration
(see, for example, Kosi
nski 1986). The behaviour of acceleration waves is governed
by a Bernoulli equation
d
= (x) + (x)2 ,
(1.1)
dx
where x is position, is the jump in particle acceleration, while the coefficients and
represent the dissipation and elastic nonlinearity, respectively. Classical references
on this subject include Bland (1969), Chen (1973), Coleman & Gurtin (1965), Menon
et al . (1983), McCarthy (1975) and Christensen (1982). The resulting competition
Proc. R. Soc. Lond. A (1999) 455, 25772614
Printed in Great Britain

2577

c 1999 The Royal Society



TEX Paper

2578

M. Ostoja-Starzewski and J. Trebicki


t

t=0

nt

fro

ve
wa

p
x

0
x0 + L

x0

(a)

(b)

(c)

Figure 1. Propagation of a wavefront in the x, t space-time. The wavefront is a zone of finite


thickness L (between x0 and x0 +L at time t = 0) propagating in the direction p (co-oriented with
x) in a microstructure of characteristic grain size d. Three cases are distinguished: (a) L  d,
which shows the trend to a classical (deterministic) continuum limit, in which fluctuations die out
to zero as L/d ; (b) L finite relative to d, where spatial fluctuations render the wavefront
a statistical mesoscale volume element; and (c) L < d, which leads to a piecewise-constant
evolution.

between the two aforementioned effects leads to a possibility of shock formation


in a finite distance x (so-called distance to form a shock ), providing the initial
amplitude 0 exceeds a so-called critical amplitude c . In the case of a homogeneous
medium, these are



(1.2)
,
c = .
x = ln 1

Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2579

For initial amplitudes 0 < c , the dissipation wins over the nonlinearity, and the
wave decays to zero exponentially. While a number of various cases of deterministic
spatial dependence of and had been studied, principally in the 1960s and 1970s
in the rational mechanics literature, the analysis has to be conducted ab initio when
the material displays spatial random fluctuations. Let us begin by considering a
space-time view of a wavefront of finite thickness L propagating in the x-positive
direction (the wavefront in figure 1). The field variable is f (x, t), and so we see
the initial profile f |t=0 . The material has a spatially homogeneous (in terms of its
statistics) ergodic microstructure with a single characteristic grain size d. Here, for
the sake of illustration of basic concepts, we use a two-dimensional Voronoi mosaic
(based on a Poisson point process), which is an excellent generic model of a number
of heterogeneous materials. With reference to figure 1, there are, in general, three
possible cases as follows.
(a) L  d: this is the classical (deterministic) continuum limit, in which fluctuations die out to zero as L/d .
(b) L is finite relative to the grain size: here fluctuations are significant and not
negligible.
(c) L > d: the waves evolution is random and piecewise-constant. Locally, it is governed by classical continuum mechanics, but the random properties of grains it
encounters give it a stochastic character. In fact, one can distinguish two subcases: grain boundaries are either normal to the direction of wave propagation;
or at arbitrary angles and include corners.
We recognize the wavefront of case (a) to be analogous to a representative volume
element (RVE) of deterministic continuum mechanics; the RVE, i.e. L, is so large
relative to the grain size that the wavefront does not see the local material disorder.
However, as L decreases (case (b)), the wavefront becomes a statistical mesoscale
window affected by random continuum type fluctuations of the microstructure. In
case (c), the grains are uniform continua and the wavefront evolves as a random jump
process. This requires consideration of backscattering of a pulse at all the graingrain
boundaries similar to what was done in Ostoja-Starzewski (1991).
In this paper we focus on case (b): the and coefficients are now certain random
processes in x, so that becomes a stochastic process driven by a {, }x vector
process according to a stochastic Bernoulli equation (1.1), and the aforementioned
competition becomes stochastic. This means that c and x become random variables. In other words, there is a finite range of values of the initial amplitude, rather
than just a single deterministic value (1.2)2 , which may result in either a shock or a
decay. Furthermore, the distance to form a shock, given a specific initial amplitude,
is diffused over a certain range. This situation is exemplified in figure 2a, b with the
help of a numerical simulation of a now random differential equation (1.1) driven by
a white noise (see 3).
(b) Outline of the paper
Given the fact that is a stochastic process driven by the {, }x vector process,
our major objective is to statistically characterize c and x . This stochastic problem
Proc. R. Soc. Lond. A (1999)

2580

M. Ostoja-Starzewski and J. Trebicki

Figure 2. Simulation of ten exemplary evolutions of an acceleration wave (a) and its inverse
= 1/ (b) originating from the critical amplitude of a reference homogeneous deterministic
medium c(det) = hihi as functions of distance x in a random medium described by one
white noise (stochastic model of 3). Observe that either a growth to or a decay to 0 occur.
Parameters: hi = 1, hi = 1, S1 = 0.2 and S2 = 0.35.

of c was first studied by Ostoja-Starzewski (1993) in the setting of a white noise


medium; the analysis was carried out in terms of the inverse amplitude = 1/,
which allowed a transformation of the original problem into a linear one. Later on,
Ostoja-Starzewski (1995) considered the non-white-noise character of either one or
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2581

both components of the {, }x process. The present investigation, first signalled in


Ostoja-Starzewski & Trebicki (1999), is more complete in that:
(i) the wavefront is examined as a zone of finite thickness relative to a microstructure in which it propagates;
(ii) it deals with a fully nonlinear problem;
(iii) it assesses both c and x ; and
(iv) it considers a wider range of random media.
In particular, we study the following issues in this paper.
(a) Set-up of the Bernoulli equation in a random continuum due to a finite size of
the wavefront.
(b) Bernoulli equation with one white noise, including:
(1) formulae for the average critical amplitude hc i, which yield conditions
for hc i to be greater or smaller than the critical amplitude c(det) of the
deterministic problem, which is defined by the mean values of and
(see below);
(2) an exact solution for the inverse amplitude;
(3) an approximation of the probability distribution of the inverse amplitude
by a Winterstein method, and a resulting approximation of the amplitude
.
(c) Bernoulli equation with two independent white noises, and analyses similar to
those listed in point (a) above.
(d) Bernoulli equation with one Wiener process, the case that is equivalent to a
perturbation by two correlated Gaussian noises and also includes a discussion
of the probability distributions of and .
(e) Equations for the moments of , and explicit solutions in the special cases.
(f) Numerical comparisons of the first four moments of c and c for various models.
(g) The probability of formation of a shock wave as a function of the mediums
randomness.
(h) Bernoulli equation perturbed by an OrnsteinUhlenbeck process.
Before we proceed with the analysis, we would like to note that the Bernoulli
equation (1.1) describes the behaviour of acceleration waves in a wide variety of
media, both solids and fluids (see numerous references in the aforementioned rational
mechanics literature). Our analysis is, therefore, set in a general context of materials
described by nonlinear functional constitutive relations with the presence of randomness without any special reference to a particular material; we shall return to this
point in the final section of the paper.
Proc. R. Soc. Lond. A (1999)

2582

M. Ostoja-Starzewski and J. Trebicki

The basic class of random media we shall investigate is wide-sense stationary,


ergodic processes (x) and (x), understood as perturbations 0 (x) and 0 (x) imposed upon the constant mean values hi and hi. As noted in point (g) above, the
OrnsteinUhlenbeck process will also allow a study of a non-stationary material randomness. Furthermore, we shall always consider 0 (x) and 0 (x) to be smaller than
the means hi and hi. Indeed, these means define a reference homogeneous medium,
which serves as a deterministic counterpart to all the stochastic problems. Comparisons between the ensemble average solutions for c and x in the latter problem,
and the solutions of the deterministic reference homogeneous medium problem, are
made throughout the paper.

2. Wavefront dynamics in random microstructures


(a) Classical continuum mechanics viewpoint
Let us first recall that a jump f (x, t) in the classical case (a) is defined by
[[f ]] = f2 f1 ,

(2.1)

where f1 and f2 are, respectively, the quantities immediately ahead of and behind
the wavefront. The discontinuity surface (propagating in the direction of positive X,
material coordinate) is, here, of zero thickness, i.e. L 0, and located at X = Y (t).
It is well known from continuum mechanics that the convected derivative of the
jump [[f ]] travelling at velocity c is governed by
 
 
f
f
d
[[f ]] =
+c
.
(2.2)
dt
t
x
In the special case, when f is continuous, we obtain the (Hadamard) kinematic
compatibility condition of the first order
 
 
f
f
= c
,
(2.3)
t
x
which must hold, for instance, for the displacement field up to the point of material
fracture.
Next, we consider the dynamic condition of compatibility
[[ij ]]pj = c[[u i ]],

(2.4)

where we recognize the left-hand side to be the jump in tractions across the discontinuity surface. In the case of figure 1a, given the limit L/d , the tractions
and velocities on either side of the jump are practically uniform. This homogenization limit is conventionally taken for granted in continuum mechanics analyses when
substituting the constitutive law into (2.4). However, in the case of figure 1b, the
traction and velocity fields are non-uniform, and, thus, what is implied here is a
volume averaging in a direction perpendicular to p.
Let us scrutinize a constitutive law that has to come into this whole picture. In
the classical case (figure 1a), assuming, for simplicity, an isotropic Hookes law (
and are the Lame constants),
ij = Cijkl kl ,
Proc. R. Soc. Lond. A (1999)

Cijkl = ij kl + (jk il + jl ik ),

(2.5)

Growth and decay of acceleration waves in random media

2583

[[ij ]] = ij [[uk,k ]] + ([[ui,j ]] + [[uj,i ]]).

(2.6)

we have
This is clearly interpreted as a very large window limit relative to d, i.e. the RVE of
continuum mechanics. Ostoja-Starzewski (1998) gives quantitative prescriptions of
the approach of mesoscale moduli to unique macroscopic values as functions of the
window size relative to the grain size for a number of specific microstructures (e.g.
circular inclusions in a matrix, needles in a matrix, fibrous media).
The case (b), L finite relative to d, is a situation below the RVE limit: it corresponds to the wavefront moving as a statistical volume element through a microstructure. It requires, in place of (2.5), a random response law so that
{Cijkl (x, L/d, ); }
is a random stiffness tensor field; is the sample space of all possible realizations.
Thus, we should have
[[ij ]] = Cijkl (x, L/d, )([[uk,l ]] + [[ul,k ]]).

(2.7)

We now proceed in the same fashion for a nonlinear elastic/dissipative law expressed as a nonlinear functional of the entire deformation history

[F (t s), F (t)],
ij (X, t) = s=0

(2.8)

where F = x/X is the deformation gradient with explicit dependence on time. As


shown for example in Christensen (1982), one derives in the classical case (a), the
Bernoulli equation (1.1), in which the coefficients and representing the dissipation
and elastic nonlinearity, respectively, are


dEt
1
Et
0
1

,
=
.
(2.9)
4It + 2Gt (0) + 4Et F
=
4Et
dt
2Et c
We now define several symbols appearing in the above. First,

d
s=0 [F (t s), F (t)],
Et =

dF
(2.10)

d2

[F (t s), F (t)],
Et =
dF 2 s=0
are called the instantaneous tangent modulus and the instantaneous second-order
tangent modulus. Next,
d
[ ]|X=Y (t) ,
(2.11)
It =
dF s=0

where s=0
[ ] is a Frechet derivative given by (r is a real constant)

s=0
[F (t s), F (t) | f (t s)] =

Finally, we have

d
[F (t s) + rf (t s), F (t)] |r=0 . (2.12)
dr s=0

d
F 1
Et |t=0 ,
,
F 1 =
dt
t
and the velocity of the acceleration wave
p
c = Et /R ,
G0t (0) =

Proc. R. Soc. Lond. A (1999)

(2.13)
(2.14)

2584

M. Ostoja-Starzewski and J. Trebicki

where R is the position-independent mass density in the reference configuration.


A consideration of case (b) for the acceleration wavefronts with L finite relative to
d implies randomness in all the parameters present in the and coefficients. This,
in turn, leads to a stochastic differential equation (of form (1.1)) of a prohibitively
high degree of complexity, whose solution would be unwieldy. However, just as is
often done in continuum mechanics, we can consider a simpler specialized case of
(2.9): that of acceleration waves entering a homogeneously deformed medium of
some deformation gradient F0 . In that case, for X > Y (t),
x(X, t) = F0 X + X0 ,

(2.15)

where the second term refers to a possible rigid body displacement. Then, ahead of
the wave, F 1 = 0 and
t = E
0 ,
E

Et = Gt (0) = G0 ,

G0t (0) = G00 .

Next, from (2.14), the velocity of the acceleration wave is constant,


p
c = G0 /R ,

(2.16)
(2.17)

while it follows from (2.15) that F (t s) = 0 for X > Y (t). Finally, the Frechet
derivative (2.12) vanishes, which in turn, by (2.14), implies It = 0. As a result, the
and coefficients become
=

G00
,
2G0

0
E
.
2G0 c

Solution of (1.1) in the deterministic case (a) is given as

(t) =
.
[(/(0)) ]et/ +

(2.18)

(2.19)

Physically realistic conditions are


G0 > 0,

G00 6 0,

0 6= 0,
E

(2.20)

and one considers three cases as follows.


0 and |(0)| < |/|, then we have
0 or if sgn (0) = sgn E
(i) If sgn (0) = sgn E
(t) 0 monotonically as t .
(ii) If (0) = /, then we have (t) = (0), a constant; recall the critical amplitude of (1.2)2 .
0 and |(0)| > |/|, then we have |(t)| monoton(iii) If sgn (0) = sgn E
ically in a finite time given by x of (1.2)1 divided by c.
(b) Stationarity and ergodicity of random media models
0 are conClearly, everything is deterministic in the above because G0 , G00 and E
stants dependent on a given physical situation. However, if they are to be considered
random fields in the sense of case (b), analogous to the field of stiffness in (2.7), the
evolution does not follow equation (2.19) any more, and has to be investigated anew.
We can now consider two basic types of randomness:
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2585

0 and, hence, and are uncorrelated random processes in posi(I) G00 and E
tion X;
0 and, hence, and are correlated random processes in posi(II) G00 and E
tion X.
Consequently, we arrive at a stochastic version of the Bernoulli equation driven by a
vector process {, }x . The resulting competition between the dissipation and elastic
nonlinearity becomes stochastic, and the two quantities of interest, distance x
and critical amplitude 0 , become random variables. Consistent with the discussion
leading to (2.7), we see that both coefficients and are stochastic processes in
cases (b) and (c), and, in the following, as stated in 1, we shall focus on (b).
We assume and to be wide-sense stationary random processes
)
(X) = hi + 0 (X), h0 i= 0,
(2.21)
(X) = hi + 0 (X), h 0 i = 0,
where hi and hi are constant, and 0 and 0 are zero-mean wide-sense stationary
random processes. Without much loss of generality, we can take 0 and 0 as zeromean wide-band Gaussian processes with correlation functions of the type
2
f (a, X),
K(X) = X

(2.22)

2
is the variance of a given process, a is a parameter inverse
where X = |X X 0 |, X
to the correlation radius , and function f satisfies conditions

f (a, 0) = 1,

lim

f (a, X) = 0.

(2.23)

In particular, we may consider an OrnsteinUhlenbeck process for which


2 a|X|
K(X) = X
e
.

(2.24)

Since we focus on the effect of heterogeneities very small relative to the wavefront
thickness L, we may further consider situations where the correlation radii of both
0 and 0 are much smaller than the correlation radius of the amplitude process
(X). It is reasonable, on the basis of theoretical analyses of random microstructures (Ostoja-Starzewski 1994) and of experimental measurements of mechanical
properties of paper (DiMillo & Ostoja-Starzewski 1998), to take of 0 and of
0 to decrease with L increasing, and to increase with d increasing, i.e. d/L.
Consequently, for L/d  1, we see that
 ,

(2.25)

which allows us to treat (X, ) as a diffusion Markov process driven by two zeromean white noise processes 1 (X, ) and 2 (X, ) that are equivalent to processes
0 (X, ) and 0 (X, ). That is, their correlation functions are of the form
Ki (X) = 2Di (X),

i = 1, 2,

in which the intensities Di are determined by


Z
Z
D1 =
K0 (X) dX,
D2 =
0

Proc. R. Soc. Lond. A (1999)

K 0 (X) dX.

(2.26)

(2.27)

2586

M. Ostoja-Starzewski and J. Trebicki

In our analysis, we assume that the physical processes 0 and 0 satisfy (with
probability one) ergodic properties
Z

1 y

i (X, ) dX = mi = hi (X, )i, i = 1, 2,


lim

y y 0
Z y
(2.28)
1
2

lim
(i (X, )i (X + X, )) dX = Ki (X) + mi .
y y 0
In practice, the left- and right-hand sides of (2.28)1 would be replaced by a spatial (or volume) average from a finite number of sampling points n taken over one
realization
i () =

N
1 X
i (Xn , ),
N n=1

i = 1, 2,

(2.29)

and an ensemble average from a finite number of realizations taken at one sampling
point
hi (X)i =

N
1 X
i (X, n ),
N n=1

i = 1, 2.

(2.30)

The ergodicity of these estimators, i.e. i (X, ) = hi (X, )i, is assured for sufficiently
large N , by the property of the correlation function
lim

|X|

Ki (X) = 0,

i = 1, 2,

(2.31)

which is satisfied by our OrnsteinUhlenbeck as well as the white noise process


models. Indeed, many material microstructures may even possess a stronger property,
that of mixing. This, for instance, is the case with Voronoi mosaics based on Poisson
point processes, both in two and three dimensions (Stoyan & Stoyan 1994); our
figure 1 employs such a mosaic.
Finally, we note that the statistics (mean and correlation function) of mesoscale
properties corresponding to any given L/d ratio, can be readily sampled experimentally. This would involve measuring two coefficients specified in (2.18) on appropriate
sized aggregates of grains, i.e. corresponding to the mesoscale window of figure 1b.
While any specific choice of the ratio L/d may suggest so-called local averaging of
stochastic processes i (X, n ) (i = 1, 2), it is just a superficial analogy: the effective
and on a given scale L are not a result of spatial averaging over L. They could
formally be defined via a stochastic homogenization procedure, but, given the nonlinear/dissipative nature of all the grains, a quantitative solution would be prohibitively
complex.

3. Bernoulli equation with one white noise


(a) Basic stochastic formulation
The evolution of an acceleration wave is governed by a Bernoulli equation
d
= (x) + (x)2 ,
dx
Proc. R. Soc. Lond. A (1999)

x (, > 0).

(3.1)

Growth and decay of acceleration waves in random media

2587

Here we use the position x as an independent variable, but the time may also play this
role; interpretation in terms of x (henceforth representing the Lagrangian coordinate
X) is preferable, because (x) and (x) are material properties. We shall assume
take non-negative values, and prescribe a deterministic initial condition
(x0 ) = 0 .

(3.2)

It is, however, straightforward to generalize the ensuing analysis by admitting randomness in the initial condition.
In equation (3.1), (x) describes the attenuation of the wave, and (x) its amplification due to a materials nonlinear elasticity. Next, we consider both (x) and (x)
to be perturbed by the same standard zero-mean white noise (x, ), that is

(x, ) = hi + S1 (x, ),
(x, ) = hi + S2 (x, ),
(3.3)
S1 + S2 = S,
S1 , S2 > 0,
where hi and hi are mean values of (x) and (x), respectively, and hi  S1 ,
hi  S2 . Also, , where (, F, P ) is the basic probability space. For simplicity,
in further exposition, the parameter will be omitted. Equation (3.3) means that the
white noise (x) appearing at the wavefront affects its dissipation with the intensity
S1 , and its nonlinear amplification with the intensity S2 .
Introducing (3.3) into (3.1), we obtain a stochastic differential equation for :
d
= hi + hi2 + [S2 2 S1 ](x).
dx
We next set up a Stratonovich equation for equation (3.4):
d = [hi + hi2 ] dx + [S2 2 S1 ] dW (S) (x),

(x0 ) = 0 .

(3.4)

(3.5)

In (3.5), dW (S) (x) is a Stratonovich-type differential of the Wiener process. We


choose the Stratonovich interpretation because the stochastic perturbations in equation (3.4) are of a parametric type (see, for example, Schuss 1980).
For the sake of computing various functions of , we will need the It
o formula
related to the It
o equation equivalent to the Stratonovich equation (3.5). To this
end, we define
A() = hi + hi2 ,

B() = S2 2 S1 ,

and proceed according to the general formula




d
1
d = A() + 2 B() B() dx + B() dW (x),
d

(x0 ) = 0 ,

(3.6)

(3.7)

to find the It
o equation, equivalent to (3.5), that is

d = A()
dx + B()
dW (x),
where

(x0 ) = 0 ,

(3.8)

A()
= ( 12 S12 hi) + (hi 32 S1 S2 )2 + S22 3 = A() + 12 B() B(),

B()
= B() = S2 2 S1 .
(3.9)

Proc. R. Soc. Lond. A (1999)

2588

M. Ostoja-Starzewski and J. Trebicki

Let us now consider a process y = g(), which is a certain memoryless transformation of the process (x). The It
o equation for this new process is obtained from
the following It
o formula:


dg
d2 g
dg
1 2

dy = A()
+ 2 B () 2 dx + B()
dW (x).
(3.10)
d
d
d
The first application of the above will concern the moments of (x). To this end,
we introduce a function
y = k ,

k = 1, 2, . . . ,

(3.11)

and applying It
os formula (3.10), we find
k1
k1

2 ()k2 ] dx + k B()

+ 12 k(k 1)B
dW (x).
dk (x) = [k A()

(3.12)

Upon averaging (3.12), we arrive at the following system of equations for the moments
of the process (x)

d k
k1
2
k2
1

h i = khA()
i + 2 k(k 1)hB ()
i,
dx
(3.13)

hk (x0 )i = h0k i, k = 1, 2, . . . ,
whereby we note that the last term in (3.12) vanishes because hdW (x)i = 0 and
(x) is a non-anticipating random process. It is noteworthy that (3.13) is not a
closed system of equations.
The inverse amplitude, defined by
(x) = 1/[(x)],

(3.14)

is one of the primary functions of interest in the ensuing analysis; it transforms the
problem of the blow-up of the wave amplitude to infinity to the one of zero-crossing.
We find from It
os formula (3.10) the following equation for (x):






1
1
2 1

d(x) = A() 2 + B 3 dx + B() 2 dW (x).


(3.15)

Noting (3.13), we get


d = [(S12 12 S12 + hi) (hi 32 S1 S2 + 2S1 S2 )] dx + (S1 S2 ) dW (x), (3.16)
so that, finally,

d = [b1 + b2 ] dx + (S1 S2 ) dW (x),


b1 = hi + 12 S12 ,

b2 = hi 12 S1 S2 .

(3.17)

Note that (3.17) is a linear inhomogeneous equation.


Next of interest are moments of the inverse amplitude. These can also be obtained
from It
os formula with a substitution y = k . Thus,

d k

2
k
k1
1
h i = [kb1 + 2 k(k 1)S1 ]h i + [kb2 + k(k 1)S1 S2 ]h
i

dx
2 k2
1
i, (3.18)
+ 2 k(k 1)S2 h

h k (x0 )i = h0k i, k = 1, 2, . . .
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2589

For the first two moments we find

hi = b1 hi + b2 ,
dx
(3.19)
d 2
2
2
2

h i = [2b1 + S1 ]h i + [2b2 S1 S2 ]hi + S2 ,


dx
with b1 and b2 being given by (3.17). Solving the above linear equations we get



b2
b2 b1 (xx0 )

h(x)i = + 0 +
,
e
b1
b1
(3.20)

2
b1 (xx0 )
2
(2b1 +S12 )(xx0 )
+ [0 + c1 + c2 ]e
,
h (x)i = c1 c2 e
where
S 2 2(b2 /b1 )(b2 S1 S2 )
2(b2 S1 S2 )(0 + (b2 /b1 ))
c1 = 2
,
c2 =
.
(3.21)
2
2b1 + S1
b1 + S12
Taking x0 = 0 we get from (3.20)


hi + 12 S1 S2
hi + 12 S1 S2 (hi+ 1 S12 )x
2
h(x)i =
+ 0
.
(3.22)
e
hi + 12 S12
hi + 12 S12
Since hi + 12 S12 > 0 and hi + 12 S1 S2 > 0, then, assuming
0 >

hi + 12 S1 S2
b2
=
,
b1
hi + 12 S12

(3.23)

we see that hi as x , and we have no shock formation with respect to


the mean value.
On the other hand, if
hi + 12 S1 S2
b2
,
(3.24)
0 < =
b1
hi + 12 S12

then hi as x , which implies crossing the zero-level and, thereby, an


explosion of the process = 1 . It follows that the average critical inverse amplitude
hc i is given by
hi + 12 S1 S2
.
hi + 12 S12

(3.25)

hi + 12 S12
1
=
.
hc i
hi + 12 S1 S2

(3.27)

hc i =

Considering that the reference deterministic homogeneous medium is specified by hi


and hi, and that the corresponding critical amplitude is c(det) = hi/hi (recall
(3.3)), we establish that the assumption of a white noise perturbation of and
leads to the following relationships between hc i and c(det) :

S2 hi
,
>
hc i > c(det) , for
S1 hi
(3.26)
S2 hi

hc i < c(det) , for


.
<
S1
hi
1
Now, since hi > hi , we find, in view of (3.25), the average critical amplitude
hc i given in terms of the average inverse critical amplitude, that is
hc i =
Proc. R. Soc. Lond. A (1999)

2590

M. Ostoja-Starzewski and J. Trebicki

Figure 3. (a) Five exemplary evolutions of the inverse amplitude = 1/ of an acceleration


wave originating from the same initial value 0 = 0.8hc i, with hc i = 1.0, and demonstrating a
blow-up to a shock wave. (b) The same evolutions in terms of . Parameters: hi = 1, hi = 1,
S1 = 0.35 and S2 = 0.35.

It can now be shown from (3.26) that we can expect hc i to be smaller (greater)
than c(det) = hi/hi accordingly as the ratio S2 /S1 is greater (smaller) than the
ratio hi/hi.
(b) Numerical simulations of trajectories
Evolutions of the deterministic Bernoulli equation are smooth functions tending
either to zero or to infinity. In the case of randomness in the and coefficients,
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2591

Figure 4. The mean and standard deviation of the inverse amplitude as functions of
position corresponding to the case of figure 3; all parameters are the same as in figure 3.

there is, as already pointed out, a stochastic competition between the dissipation
effect due to the (x) and the (x)2 terms. This competition means that
(i) the trajectories of the stochastic system are not smooth, but, rather jagged;
(ii) there is a range of different trajectories, decaying and exploding, that emanate
from the same initial value; and
(iii) there might be a crossover from the behaviour where dominates to the one
where dominates.
In order to numerically simulate the system evolution, we adopt Milshteins method
(Milshtein 1978; Sobczyk 1991).
In view of (3.8), this method leads to the following finite difference scheme for the
x process
i) 1 B
i )W (xi ) + 1 B
2 (i )]h + B(
2 (i )[W (xi )],
(3.28)
i+1 = i + [A(
2

while, considering (3.17), we obtain for the x process


i+1 = i + [(b1 i + b2 ) 12 (S1 i S2 )2 ]h

+ (S1 i S2 )2 W (xi ) + 12 (S1 i S2 )2 [W (xi )],

(3.29)

where h is a step on the x-axis.


Using these schemes, in figure 3 we display five trajectories, generated in a Monte
Carlo sense, in terms of and , respectively. They correspond to the case hi =
hi = 1.0, and S1 = S2 = 0.35, from which, given (3.26), we calculate the critical
amplitude hc i = 1/hc i = 1.0. In particular, in figure 3a, b we treat the case of
0 = 0.8 and 0 = 1/0 = 1.25, respectively, which is the case of blow-up of the
wave. These plots are complemented by the graphs of the average and the standard
Proc. R. Soc. Lond. A (1999)

2592

M. Ostoja-Starzewski and J. Trebicki

Figure 5. (a) Five exemplary evolutions of an acceleration wave = 1/ originating from the
same initial value 0 = 1.25 > hc i, and demonstrating a decay; all other parameters are the
same as in figure 3. (b) The same evolutions in terms of .

deviation of in figure 4, computed from (3.20) and corresponding to this initial


value case. Figure 4 is exemplary of a scenario of explosion of an acceleration wave
to a shock, in which the mean hi goes down to zero with x increasing, while the
associated standard deviation grows.
Next, in figure 5a, b, we present, for the same random medium, the case of decay
of an acceleration wave by choosing 0 = 1.25 and 0 = 1/0 = 0.8, respectively.
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2593

Figure 6. The mean hi and standard deviation of the inverse amplitude as functions of
position corresponding to the case of figure 4; all parameters are the same as in figure 2.

Figure 6 shows that the growth of hi is accompanied by an equally rapid growth


of its standard deviation. This means, as may be confirmed by figure 5b, that the
standard deviation of the process actually vanishes as x grows.
Finally, we mention an obvious feature of these simulations: as S1 and S2 go down
to zero, the jaggedness in the trajectories vanishes and they become deterministic
solutions of (1.1).
(c) Exact solution for the inverse amplitude
Let us return to the linear equation
dx = (Ax + Bx x ) dx + (Cx + Dx x ) dWx .

(3.30)

For completeness, after (Kallianpur 1980), its exact solution has the following form:
Z x

Z x
1 2
(Bs 2 Ds ) ds +
Dx dWs
x = exp
0
 Z0 s


Z x
Z s
1 2
exp
(Bu 2 Du ) du
Du dWu (As Ds Cs ) ds
0 +
0
0
0
 Z s


Z s
Z x
exp
(Bu 12 Du2 ) du
Du dWu Cs dWs ,
(3.31)
+
0

wherein, considering (3.17), we identify


Ax = b2 ,
Cx = S2 ,
Proc. R. Soc. Lond. A (1999)

Bx = b1 ,
Dx = S1 .

(3.32)

2594

M. Ostoja-Starzewski and J. Trebicki

(d ) Approximation of the probability distribution of the inverse amplitude


Let us consider the probability distribution of . It could, in principle, be obtained
from (3.31). Another possibility, however, is to approximate it using the first
moments; the point being that the system of equations for moments (3.18) is closed.
For any fixed x, the process (x) becomes a random variable (). In order to approximate (), the following Hermite polynomial transformation will be considered (cf.
Ditlevsen & Madsen 1996)
() =

N
X

bi Hi (X()),

(3.33)

i=0

where the bi are constant coefficients, and X() is a standardized normal random
variable, i.e. its density is of the form

2
(3.34)
fX (x) = (1/2 2)ex /2 .
Also, Hi (x) = 1, H1 (x) = x, H2 (x) = x 1, H3 (x) = x3 3x, H4 (x) = x4 6x2 + 3,
etc., are the Hermite polynomials. Obviously, the formula (3.31) can be rewritten in
the form
N
X
ai X i ().
(3.35)
() =
i=0

Now, in order to calculate the values of the N +1 parameters ai , the information on


the moments h k i, k = 1, . . . , N + 1 is used. Taking into account the transformation
(3.33), these parameters are determined from the system of nonlinear (with respect
to ai ) equations
X
k 
N
k
i
ai X ()
, k = 1, 2, . . . , N + 1.
(3.36)
h i =
i=0

The above equation can be rewritten in a more explicit form since we have
(
0,
for n odd,
k
hX i =
1 3 (n 1),
for n even.

(3.37)

Next, the system (3.36) is solved numerically, and, finally, the probability density
function of () is determined as
X
f (z) =
|gj1 (z)|fX (gj1 (z)),
(3.38)
j

gj1

where
is the jth branch of the inverse g 1 restricted to an appropriate interval
of monotonicity. We note here that in the special case of N = 4, the transformation
(3.35) is called a Winterstein approximation (Ditlevsen & Madsen 1996).
(e) Approximation of the probability distribution of the amplitude
With the approximate form of the distribution f (y) of the inverse amplitude , we
may derive an approximate form of the distribution of the amplitude itself. This
is done by noting that

 
 
1
1
1
1


= 2 fx
, y > 0.
(3.39)
fx (y) = 2 fx
y
y
y
y
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2595

Furthermore, it should be pointed out here that this formula can be used far enough
from the explosion time. The latter is a random variable x () defined as follows:
x () = inf{x; (x, ) > 0}.
Taking four moments calculated from
Z
Z
y k fx (y) dy =
hxk i =
0

1
y

k
fx

(3.40)
 
1
dy,
y

(3.41)

we can use the Winterstein approximation for the probability density of . In the
particular case when the distribution of can be approximated by a Gaussian, the
distribution of is simply obtained as a nonlinear transformation of the Gaussian
distribution. Finally, note that in the case of S1 = 0, S2 > 0, we find that the inverse
amplitude is governed by the OrnsteinUhlenbeck equation, implying that fx is
Gaussian.

4. Bernoulli equation with two independent white noises


(a) Basic considerations
We now perturb the original Bernoulli equation
d
= (x) + (x)2 ,
dx
by two zero-mean white noises 1 and 2
(x) = hi + 1 (x),

(4.1)

(x) = hi + 2 (x),

having a correlation matrix


Ki j (x1 , x2 ) = gij (x2 x1 ),

g
gij = 11
0


0
.
g22

(4.2)
(4.3)

From this we get


d
= (hi + 1 ) + (hi + 2 )2 ,
dx
which is interpreted in the Stratonovich sense as

(S)
(S)
d = (hi + hi2 ) dx g11 dW1 (x) + g22 2 dW2 (x),
(S)
dW1

(4.4)
(4.5)

(S)
dW2

and
are differentials of two independent Wiener processes in a
where
Stratonovich sense.
A more convenient way of writing the above is in a matrix form
where

d = F () dx + () dW (S) (x),

(4.6)

F () = hi + hi2 ,

() = [11 , 12 ] = [ g11 , g22 2 ],

(S)
(S)
dW (S) = [dW1 , dW2 ]T .

(4.7)

An It
o equation, equivalent to (3.6), is
d = F () dx +
() dW (S) (x),
Proc. R. Soc. Lond. A (1999)

(4.8)

2596

M. Ostoja-Starzewski and J. Trebicki

where

1 d T
,
F () = F () +

2 d

() = (),

dW = [ dW1 , dW2 ]T .

(4.9)

In the above, dW1 and dW2 are the It


o-sense differentials of two independent standard Wiener processes. Taking (4.7) into account, we rewrite equation (4.6) as
 

g11
2
1

dx
d = (hi + hi ) dx + 2 [ g11 , 2 g22 ]
2 g22

g11 dW1 + g22 2 dW2 . (4.10)


Finally, the It
o equation for (4.1) takes the form

1
d = [( 2 g11 hi) + hi2 + g22 3 ] dx g11 dW1 + g22 2 dW2 .

(4.11)

We next set up an It
o equation for the inverse amplitude. Thus, from the It
o
formula in the two-dimensional case (see (7.12))

we find

dh 1 d2 h 2
dh
2
+
[ () + 12
()] + [dW ]T ,
dh() = F ()
d 2 d2 11
d

(4.12)

d = [( 12 g11 hi) + hi2 + g22 3 ](1/2 ) dx

+ (1/3 )[g11 3 + g22 4 ] dx + (1/2 )[ g11 dW1 + g22 2 dW2 ]

(4.13)
= [(hi + g11 ) hi] dx + g11 dW1 g22 dW2 .
This linear equation with a parametric noise dW1 and an additive noise dW2 can be
written in a vector form as



dW1
d = [(hi + g11 ) hi] dx + [ g11 , g22 ]
.
(4.14)
dW2
(b) Moment equations for the inverse amplitude
We apply the It
o formula (4.13) to the function h() = k and then average to
obtain
d k
h i = kh(hi + g11 ) k hi k1 i + 12 k(k 1)hg11 k1 + g22 k2 i.
(4.15)
dx
From this we find equations governing the first two moments

hi = (hi + g11 )hi hi,


h0 i = 1/0 ,
dx
(4.16)
d 2
2
2
2

h i = 2(hi + g11 )h i + (g11 hi)hi + g22 , h0 i = 1/0 ,


dx
where the initial conditions are being interpreted in the same sense as in 2. It follows
that the solutions of (4.16) are



hi
hi

+ h0 i
exp[(hi + g11 )x],
hi =
hi + g11
hi + g11
(4.17)

h 2 i = c1 c2 ehix + [h02 i + (c1 + c2 )] exp[2hi + g22 ]x,


Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media


in which
g22 (g11 /2hi)(g11 hi)
,
c1 =
2hi + 12 g11

c2 =

2597

(g11 hi)(h0 i + (g11 /2hi))


. (4.18)
hi + 12 g11

It is interesting that the equation governing the evolution of hi is similar to the


one of the previous section. However, in the present case, the Wiener process dW2
enters additively into (4.14), which leads to hi in (4.17) being independent of the
intensity g22 of the (x) process. From (4.17)1 , we now find the decay and growth
conditions for the acceleration wave

hi

,
hi , for h0 i >
hi + g11
(4.19)

hi

hi 0,
for h0 i <
.
hi + g11
The same result was obtained by a different method in Ostoja-Starzewski (1993).
The analysis regarding the probability distributions of and is entirely analogous
to that conducted in 3, and is, therefore, not repeated.

5. Bernoulli equation with one Wiener process equivalent to a


perturbation by two correlated Gaussian noises
(a) Basic considerations
Again, let us begin with the stochastic Bernoulli equation
d
= (hi + 1 ) + (hi + 2 )2 ,
(5.1)
dx
where 1 and 2 represent two zero-mean Gaussian noises having a correlation matrix


g11 g12
g11 g22 > g 2 .
(5.2)
,
g12 = g21 = g,
gij =
g21 g22
Equation (5.1) is rewritten as

 
d
2
2 1
= hi + hi + [, ]
,
2
dx

which leads to
d = F () + () dW,
() = [11 , 12 ] = [, 2 ],

(5.3)

F () = hi + hi2 ,

dW1

, hdWi , dWj i = gij = Gij .


dW =

dW2


(5.4)

Now, by interpreting equation (5.4) in the Stratonovich sense, we arrive at the equivalent Ito formula
1 d
d = F () dx + () dW,
F () = F () +
G T .
(5.5)
2 d
Proc. R. Soc. Lond. A (1999)

2598

M. Ostoja-Starzewski and J. Trebicki

The solution of (5.5) is given by a Markov process with drift and diffusion coefficients
(5.6)
A() = F (),
B() = G T .
Each diffusion Markov process having such A() and B() is a solution of a stochastic
differential equation; that is, there exists a certain Wiener process W (x) such that
(x) is a solution of
d = A() + B() dW.

(5.7)

Thus, we can replace equation (5.5) with two correlated Wiener processes by the
equation (5.7) with one Wiener process. Considering the expressions for and G,
we obtain

 
g

g
A() = hi + hi2 + 12 [1, 2] 11
g g22 2
= ( 12 g11 hi) + (hi 32 g)2 + g22 3 ,
and
B() = [, 2 ]

g11
g

g
g22

(5.8)

 

= g11 2 12 g3 + g22 4 .
2

(5.9)

Now, applying the It


o formula for a function = h() to equation (4.7), we obtain
dh 1 p
dh
d2 h p
+ 2 [ B()]2 2 + B()
dW,
d
d
d
which leads to the following equation for the inverse amplitude = 1/:

1
1
1 p
d = dh() = 2 A() + 3 B() 2 B() dW,

d = dh() = A()

A()(1/2 ) = (hi 12 g11 ) + ( 32 g hi)2 ,


B()(1/3 ) = g11 2g.
Thus, we finally arrive at
d = (hi + 12 g11 ) (hi + 12 g)

p
g11 2 2g + g22 dW.

(5.10)

(5.11)

(5.12)

(b) Moment equations for the inverse amplitude


We are now ready to study the moments of the inverse amplitude. To this end, we
use the Ito formula (5.10) with respect to equation (5.12) to obtain
d k
h i = hk k1 [(hi + 12 g11 ) ( 12 g + hi)]i + 12 hk(k 1) k2 [g11 2 2g + g22 ]i
dx
= [k(hi + 12 g11 ) + 12 k(k 1)g11 ]h k i
+ [k( 21 g + hi) + k(k 1)g]h k1 i + 12 k(k 1)g22 h k2 i.

In the case k = 1 we find


d
hi = (hi + 12 g11 )hi ( 12 g + hi),
dx
Proc. R. Soc. Lond. A (1999)

(5.13)

(5.14)

Growth and decay of acceleration waves in random media


which leads to



hi + 12 g
hi + 12 g
+ 0
e(hi+(g11 /2))x .
hi =
hi + 12 g11
hi + 12 g11

2599

(5.15)

From this, we note immediately that the average critical inverse amplitude hc i is
now given by
1
2 S1 S2 + hi
.
1 2
2 S1 + hi

hc i =

(5.16)

Let us note here that for g = 0, the analysis reduces to the one conducted earlier in
4. On the other hand, for the case g S1 S2 , g11 S12 and g22 S22 , the analysis
reduces to that of 1.
It will now be convenient to denote a kth order moment and its initial condition
as follows:
h k i mk (x),

h0k i mk (x0 ).

(5.17)

From the linear equation (5.13) we determine


Z x
 Z x
 Z t

r1 (t) dt +
r1 ( ) d dt,
r2 (t) exp
mk (x) = 0k exp
x0

x0

where

r1 (t) = k[b1 + 12 (k 1)g11 ],

r2 (t) = 12 k(k 1)g22 mk2 (t) k[b2 + (k 1)g]mk1 (t).

Taking x0 = 0, we find
r1 x

mk (x) = e

0k

Z
+

(5.18)

r2 (t)e

r1 t


dt .

(5.19)

(5.20)

While we have already found m1 (x) in (5.15), the second moment is given explicitly
as
m2 (x) = exp[(2b1 + g11 )x]{2b21 b2 (02 + 1) + b2 g11 (g22 2g0 + 302 b1 2b0 + 02 g11 )
+ b1 b2 (g22 4b2 0 4g0 ) + 2gb21 }
+ exp(b1 y){2b1 b2 (2b2 0 g11 2b1 + 2g0 ) b1 g(2b1 + g11 )

+ 2b2 g11 0 (g + b2 ) + 2b1 g11 (g + g11 ) + b1 b2 (2b1 g22 ) + b2 g11 g22 + 2b21 g,
(5.21)
where
b1 = hi + 12 g,

b2 = hi + 12 g11 .

(5.22)

6. Specification of random events at the explosion of


an acceleration wave
Consideration of the numerical simulations of trajectories in figures 2, 3, 5 and 6
leads us to analyse several issues as follows.
(i) Explosion of an acceleration wave starting from a deterministic initial condition
is a random, rather than a deterministic, phenomenon, and thus, one can only
talk of a probability to form a shock within a given finite distance.
Proc. R. Soc. Lond. A (1999)

2600

M. Ostoja-Starzewski and J. Trebicki

Figure 7. The probability of formation of a shock wave within a distance x, P ((x) < 0), for
various types of correlation when 0 = 0.7: (1) full positive correlation; (2) zero correlation;
(3) full negative correlation; other parameters are hi = hi = 1, g11 = g22 = 0.01.

(ii) The mean distance to such an event (i.e. crossing the = 0 line) is a certain
function of random noises in the material.
(iii) The inverse amplitude is a certain random variable at the mean distance to
form a shock.
(a) Probability of formation of a shock wave
It is clear from the foregoing analysis that when 0 = 1/0 is sufficiently far
from cr = b2 /b1 , then the probability distribution of may be approximated by a
Gaussian distribution N (m (x), (x)). The formula
Z
fz (z; x) dz,
(6.1)
P ((x) < 0) = 1 P ((x) > 0) = 1
0

gives the probability of formation of a shock wave within a distance x. For small
fluctuations of the white noises in and and an appropriate choice of 0 = 1/0 ,
we can, therefore, work with


1
(z m (x))2
.
(6.2)
exp
f (z; x) =
22 (x)
2 (x)
In the following, we plot (6.1), assuming (6.2), for various cases of the initial inverse
amplitude 0 , means hi and hi, noise strengths g11 , g22 , and cross-correlation
g1 between both noises. Thus, in figure 7 we display the probability of formation
of a shock wave within a distance x, P ((x) < 0), for three types of correlation
when 0 = 0.7: (1) full positive correlation; (2) zero correlation; (3) full negative
correlation; other parameters are hi = hi = 1, g11 = g22 = 0.01. It is seen that, as
the correlation between and decreases, the range of distances over which a shock
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2601

Figure 8. The probability of formation of a shock wave within a distance x, P ((x) < 0), for
various types of correlation when 0 = 0.6; the meaning of the three curves and particular
parameter values is the same as in figure 6.

Figure 9. A comparison of the probabilities P ((x) < 0) at full positive correlation, for three
cases: (1) 0 = 0.7; (2) 0 = 0.6; (3) 0 = 0.5; all other parameters are the same as in figure 6.

may form increases. While we used here 0 = 0.7, figure 8 is devoted to a situation
when 0 = 0.6, all other parameters being the same. Clearly, the same trends are
observed here with all the curves being shifted to lower values of x.
Proc. R. Soc. Lond. A (1999)

2602

M. Ostoja-Starzewski and J. Trebicki

Figure 10. A comparison of the probabilities P ((x) < 0) at zero correlation, for three cases:
(1) 0 = 0.7; (2) 0 = 0.6; (3) 0 = 0.5; all other parameters are the same as in figure 6.

Figure 11. A comparison of the probabilities P ((x) < 0) at full negative correlation, for three
cases: (1) 0 = 0.7; (2) 0 = 0.6; (3) 0 = 0.5; all other parameters are the same as in figure 6.

To put the shifting of P ((x) < 0) in perspective, a comparison of the probabilities


P ((x) < 0) at full positive correlation is given in figure 9 for three cases of the
initial amplitude. This comparison is continued in figure 10 at zero correlation, and
in figure 11 at full negative correlation.
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2603

Figure 12. A comparison of the average distance to form a shock, x xc , and the standard
deviation of the inverse amplitude at the point of explosion, (x ), as a function of the initial
amplitude 0 z0 , for three cases: (1) full positive correlation; (2) zero correlation; and (3) full
negative correlation; all other parameters are the same as in figure 6.

(b) Distance to form a shock wave


Using again (6.1), in figure 12 we give a comparison of the average distance to
form a shock, x , as a function of the initial amplitude 0 , for three cases of the
{, }x process: (1) full positive correlation; (2) zero correlation; and (3) full negative
correlation; all other parameters are the same as in figure 7. We do not carry the plots
beyond 0 = 0.9, as the distance x then tends to go to infinity. In this figure we
also display the standard deviation of the inverse amplitude at the point of explosion,
(x ).
Our next three figures continue such a comparison for other sets of parameters at
full positive, zero and negative correlation. In particular, we have here three cases: (1)
g11 = g22 = 0.01; (2) g11 = 0.04 and g22 = 0.01; and (3) g11 = 0.01 and g22 = 0.04.
Figures 13, 14 and 15 give the full positive, zero and negative correlation cases,
respectively. Note that, with correlation between and decreasing, the distance
x tends to go up. It is also interesting to note from these figures that the standard
deviation of the inverse amplitude (x ) tends to grow with 0 as well, and is
strongest when both noises in the material are fully negatively correlated.
(c) Character of the inverse amplitude at the point of shock formation
The final issue we address is that of the character of the probability distribution
of the inverse amplitude at the point of shock formation x . We do it for a range
of 0 values that make the explosion to shock possible. First, in figure 16, we show
the skewness 1 and flatness 2 of the inverse amplitude at the point of blow-up,
Proc. R. Soc. Lond. A (1999)

2604

M. Ostoja-Starzewski and J. Trebicki

Figure 13. A comparison of the average distance to form a shock, x xc , and the standard
deviation of the inverse amplitude at the point of explosion, (x ), as a function of the initial
amplitude 0 z0 , at full positive correlation, for various levels of noise. Notation: (1) for
g11 = g22 = 0.01; (2) for g11 = 0.04 and g22 = 0.01; (3) g11 = 0.01 and g22 = 0.04.

x , as functions of 0 . This is all meant with respect to the mean trajectory, that
is when m1 (x ) = 0. We use here hi = 1, hi = 1 and g11 = g22 = 0.01, and,
as before, study three cases: (1) full positive correlation; (2) zero correlation; and
(3) full negative correlation between and . It is interesting that the skewness
and flatness are practically non-existent in case (1), and they grow with decreasing
cross-correlation. Also in this figure, in the inset we show the density function of the
inverse amplitude at point x for the initial amplitude 0 = 0.9.
Because the inverse amplitude process is fully positive, the boundary = 0
should, from a mathematical viewpoint, be considered as an absorbing boundary
and an appropriate condition for the inverse amplitudes probability density (i.e.
f (, x)|=0 = 0) should be taken into account. Nevertheless, in practice, and especially when the diffusion effects are weak relative to the drift, it is more convenient
to obtain the probability density of the inverse amplitude process in the whole space
and, further, to consider only a part of the probability density for > 0.
In the next three figures we continue this study. First, in figure 17, which is focused
on the full positive correlation, in addition to the above discussed parameter case we
also plot what happens as either g11 or g22 gets larger: cases (2) and (3), respectively;
case (1) is a reference case from above. In case (2) the skewness and flatness increase,
whereas in case (3) only the flatness increases and the skewness decreases. The change
of g11 or g22 also has the effect of broadening the density function of at x .
In figure 18 we analyse the situation of zero correlation between and as either
g11 or g22 gets larger: cases (2) and (3), respectively; and case (1) is a reference case
from figure 16. Here, the magnitude of changes in the skewness and flatness is much
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2605

Figure 14. A comparison of the average distance to form a shock, x xc , and the standard
deviation of the inverse amplitude at the point of explosion, (x ), as a function of the initial
amplitude 0 z0 , at zero correlation, for various levels of noise. The meaning of (1), (2) and
(3) is the same as in figure 12.

greater than before. However, as we go to the case of fully negative correlation in


figure 19, these changes are more powerful by one more order of magnitude. The
three cases investigated in figure 19 are entirely analogous to those analysed in the
two previous figures. Finally, we note from the last insets that increasing either one
of the noises, g11 or g22 , broadens and skews the density of at point x .
These figures indicate that the analysis of a problem of growth to shock based on
the first moment of is insufficient, and should incorporate either the probability
distribution of this quantity or its higher moments. This is done with the help of
figures 1619. They contain a wide range of information on the coefficients of skewness
and flatness of the inverse amplitude. In particular, we observe here a very nonGaussian character of at x , even for very small intensities of perturbations in
and when their correlation is not fully positive.

7. Bernoulli equation perturbed by an OrnsteinUhlenbeck process


(a) Basic considerations
We now perturb the original Bernoulli equation
d
= (x) + (x)2 ,
dx
by two noises (x) and (x)
(x) = hi + S1 U (x),
Proc. R. Soc. Lond. A (1999)

(x0 ) = 0 ,

(x) = hi + S2 U (x),

S1 , S2 > 0,

(7.1)

(7.2)

2606

M. Ostoja-Starzewski and J. Trebicki

Figure 15. A comparison of the average distance to form a shock, x xc , and the standard
deviation of the inverse amplitude at the point of explosion, (x ), as a function of the initial
amplitude 0 z0 , at full negative correlation, for various levels of noise. The meaning of (1),
(2) and (3) is the same as in figure 12.

Figure 16. Skewness 1 ( ) and flatness 2 () of the inverse amplitude at the point x
(m1 (x ) = 0), as functions of 0 at hi = 1, hi = 1, and g11 = g22 = 0.01, for three cases: full
positive correlation (1); zero correlation (2); and full negative correlation (3) between and .
The inset shows the density function of the inverse amplitude at point x for 0 z0 = 0.9.
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2607

Figure 17. Skewness 1 ( ) and flatness 2 () of the inverse amplitude at the point x
(m2 (xc ) = 0), as functions of 0 at hi = 1, hi = 1, at full positive correlation for three cases:
(1) g11 = g22 = 0.01; (2) g11 = 0.04 and g22 = 0.01; and (3) g11 = 0.01 and g22 = 0.04. The
inset shows the density function of the inverse amplitude at point x for 0 z0 = 0.9.

Figure 18. Skewness 1 ( ) and flatness 2 () of the inverse amplitude at the point
x (m2 (xc ) = 0), as functions of 0 at hi = 1, hi = 1, at zero correlation for three cases:
(1) g11 = g22 = 0.01; (2) g11 = 0.04 and g22 = 0.01; and (3) g11 = 0.01 and g22 = 0.04. The
inset shows the density function of the inverse amplitude at point x for 0 z0 = 0.9.

Proc. R. Soc. Lond. A (1999)

2608

M. Ostoja-Starzewski and J. Trebicki


150

0.3

density function
z 0 = 0.9

skewness and flatness

110

0.2

(1)

(3)

70

(2)

0.1

(3)
(2)

30
0.0

(1)
6

10
0.6

0.7

0.8

0.9

z0
Figure 19. Skewness 1 ( ) and flatness 2 () of the inverse amplitude at the point x
(m2 (xc ) = 0), as functions of 0 at hi = 1, hi = 1, at full negative correlation for three cases:
(1) g11 = g22 = 0.01; (2) g11 = 0.04 and g22 = 0.01; and (3) g11 = 0.01 and g22 = 0.04. The
inset shows the density function of the inverse amplitude at point x for 0 z0 = 0.9.

where, as before, hi and hi are the mean values of stationary (x) and (x) processes, but U (x) is an OrnsteinUhlenbeck process. That is,
dU = aU (x) dx + (x),

U (x0 ) = U0 ,

a > 0,

(7.3)

and (x) is a Gaussian white noise, such that


hi = 0,

K (x, x0 ) = 2D(x x0 ).

The Ito equation for (7.3) is


dU = aU (x) dx +

(7.4)

2D dW (x),

(7.5)

where W (x) is a standard Wiener process. For every x, U (x) has a normal distribution with mean and variance (assuming initial condition U (x0 ) = 0)
hU i = 0,

2
[1 e2ax ],
hhU ii = U

(7.6)

2
where U
= D/a. The correlation function of the process U (x) is
0

2 a|xx |
KU (x, x0 ) = U
e
.

Let us also assume that for every x


p
p
hi  hhS1 U ii = S1 hhU ii,

hi 

hhS2 U ii = S2

p
hhU ii.

(7.7)

We thus arrive at a stochastic differential equation for the vector y = [, U ]T


[y1 , y2 ]T
)
dy1 = [hiy1 + hiy12 ] dx + [S1 y1 + S2 y12 ]y2 dx,

(7.8)
dy2 = ay2 dx + 2D dW (x).
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2609

Figure 20. The influence of the correlation coefficient a of the process U (x) on the mean and
standard deviation of the inverse amplitude in the case when U (x0 ) = 1. Values of other
parameters are: S1 = S2 = 0.15, hi = hi = 1, hU i = 0.

We next introduce an inverse (vector) amplitude via a transformation of coordinates


= [1 , 2 ] [f1 (y1 ), f2 (y2 )] [(1/y 2 ), y2 ] [(1/), U ],
so that (7.8) leads to
d1 = [hi1 hi] dx + [S1 1 S2 ]2 dx,

d2 = a2 dx + 2D dW (x),
or, in a vector form,
d = A() dx + () dW (x),


hi1 hi + (S1 1 S2 )2
= [ai ],
A() =
a2


0 0
() =
= [ij ].
2D
0

(7.9)

i, j = 1, 2,

(7.10)

(7.11)

We treat the above by considering the moments of .


(b) Moment equations for the inverse amplitude
In order to obtain the moment equations for the inverse amplitude in the case of
a Bernoulli equation perturbed by the OrnsteinUhlenbeck process, we begin with
Proc. R. Soc. Lond. A (1999)

2610

M. Ostoja-Starzewski and J. Trebicki

Figure 21. A comparison of the influence of the correlation coefficient, a = 1.5, of the process
U (x) on the mean and standard deviation of the inverse amplitude in the case when U (x0 ) = 1
() vis-`
a-vis U (x0 ) = 0 ( ). Values of other parameters are the same as in figure 19:
S1 = S2 = 0.15, hi = hi = 1.

an It
o formula in an n-dimensional case
dh(x1 , x2 , . . . , xn )

X
n
n
r
n
n
n
h
2h X
1XX
1 X X h
ai +
lk mk +
lm dWm (x) ,
=
2
2
xl
xl xm
xl
i=1
l=1 m=1
k=1
l=1 m=1
(7.12)
and apply it to
h(1 , 2 ) = 1i 2j .

(7.13)

Noting that, as before, x plays the role of the time parameter with the i being the
components of state vector , we get
d i j
h i = hi1i1 2j a1 () + j1i 2j1 a1 () + Dj(j 1)1i 2j2 i,
dx 1 2
i, j = 0, 1, 2, . . . , M,

0 6= i + j 6 M. (7.14)

M is the highest order of a mixed moment to be considered. Upon carrying out the
averaging in (7.12), the last term vanishes, since the average value of a stochastic
integral of a non-anticipating function with respect to a Wiener process equals zero.
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2611

Now, assuming M = 2, and taking into account the vector A(), we find the
following system of equations for the moments:

h1 i = hih1 i S2 h2 i + S1 h1 2 i hi,

dt

2
2

h1 2 i = (hi a)h1 2 i + S1 h1 2 i S2 h2 i,

dt

d 2
2
2
(7.15)
h i = 2[hih1 i S2 h1 2 i + S1 h1 2 i hih1 i],

dt 1

h2 i = ah2 i,

dt

d 2

2
h2 i = 2ah2 i + 2D.
dt
Due to the presence of moments h1 22 i and h12 2 i of order higher than M in (7.15),
this system is not closed. According to the cumulant closure method (see, for example, Cramer 1946; Stratonovich 1963), the indicated moments can be approximated
as
)
h1 22 i = h1 ih22 i,
(7.16)
h12 2 i = h1 ih1 2 i,
so that (7.15) becomes closed and linear.
It is interesting to note that with the OrnsteinUhlenbeck process, we can model
stationary as well as non-stationary material statistics. Thus, in the case when initial
conditions for the process 2 (x) are h2 (x0 )i = 0 and h22 (x0 )i = U , the disturbance
acting on and has a constant variance in the whole material domain. For other
cases of initial values of h22 (x0 )i, the variance of the disturbance increases according
to (7.6). These two cases are analysed numerically in figures 20 and 21.
In figure 20, we illustrate the dependence of the mean and standard deviation of
the inverse amplitude on the correlation coefficient a of U (x), when U (x0 ) = 1,
which is the same as in the stationary state of the OrnsteinUhlenbeck process.
Clearly, as the correlation radius = 1/a increases, so does the standard deviation
of the inverse amplitude process. However, the mean remains practically unchanged.
In figure 21, we demonstrate the difference between wave evolutions in the stationary versus non-stationary media at a = 1.5. This is done again in terms of the
mean and standard deviation of 1 = 1/. The perturbation in the non-stationary
medium starts from zero and grows to the level typical of the stationary case.
Let us end this section with a note on the critical value c of the inverse amplitude.
If the process {1 , 2 } starts from the following set of initial values
)
h12 (x0 )i = m210 ,
h1 (x0 )i = m10 ,
(7.17)
2
h1 (x0 )i = 0,
h22 (x0 )i = U
,
h1 (x0 )(2 (x0 ))i = 0,
and a  0, which describes a stationary random medium, we obtain from (7.15)
d
h1 i = hih1 i hi.
(7.18)
dx
It follows from this that
(7.19)
c = hi/hi,
which is the same as for the reference homogeneous case.
Proc. R. Soc. Lond. A (1999)

2612

M. Ostoja-Starzewski and J. Trebicki

8. Closure
As noted in 1, the beauty of the Bernoulli equation (1.1) lies in its ability to describe
the behaviour of acceleration waves, and, thus, the possibility of shock formation,
in a wide variety of media, both solid and fluid. These wavefronts have conventionally been treated as singular surfaces propagating in deterministic continua. The
presence of a disordered material microstructure, however, has led us to consider
these wavefronts as zones of a thickness finite relative to the microscale of material
microstructure, such as the single grain size. Consequently, the wavefront is recognized as a mesoscale window (analogous to a statistical volume element) travelling
through a sample realization of a random continuum. In the ensemble sense, we thus
need to treat the Bernoulli equation as a stochastic differential equation describing a random evolution. As the wavefront thickness becomes very large relative to
the microscale, the variability of the effective mesoscale response over the wavefront
domain tends to zero: this is the limit of the classical representative volume element (RVE) of continuum mechanics. This issue of mesoscale windows and of the
RVE dilemma has also been studied, among others, in problems of linear elasticity
and stochastic finite elements (Ostoja-Starzewski 1998; Ostoja-Starzewski & Wang
1999), plasticity (Ostoja-Starzewski & Ilies 1996), and damage mechanics (Alzebdeh
et al . 1998).
By considering the spatial correlations of two material coefficients, we arrive at
the {, }x process driving the stochastic Bernoulli equation, modelled as either a
white noise process, or an OrnsteinUhlenbeck process. We then derive a range of
stochastic models that analyse relative effects of strength of noises of two underlying random processes that drive the stochastic competition of dissipation (x ) and
material elastic nonlinearity (x ). The principal conclusions of this investigation are
summarized as follows.
(1) It has been shown that, for small perturbations in (x) and (x), and for
the initial value of 0 = 1/0 sufficiently far away from the critical amplitude
c = 1/c , the probability density of is approximately Gaussian. This allows
one to rapidly compute the probability of an acceleration wave explosion for a
fixed value of x.
(2) The method of Winterstein can be employed for the approximation of as well
as in the cases when the conditions required for the aforementioned Gaussian
fit are not being met.
(3) The degree of correlatedness in the (x) and (x) processes may have a significant effect on the probabilistic characteristics of the inverse amplitude process
(x).
(4) In the case of the random medium being modelled by the OrnsteinUhlenbeck
process, an increase in the correlation radius of this process results in the
growth of the standard deviation of the inverse amplitude process, but leaves
its mean practically unchanged.
(5) Many of our results indicate that the analysis of a problem of growth to shock
based on the first moment of is insufficient and should incorporate either the
probability distribution of this quantity or its higher moments.
Proc. R. Soc. Lond. A (1999)

Growth and decay of acceleration waves in random media

2613

All these results were obtained for microinhomogeneous media described by widesense stationary processes. Thus, given the correlation radius of the (x) and (x)
processes being much smaller than that of the wave process, the medium could
clearly be taken as ergodic, and even mixing. However, in accord with the classical
interpretation of random processes as a family of realizations over the space, our
analyses set the stage for materials not satisfying the ergodic property (e.g. globally
inhomogeneous media), where the averaging over an ensemble is more natural.
In conclusion, the results presented in this paper may become useful not only
in the direct prediction of the behaviour of acceleration waves, but also in inversetype problems, which could allow one to infer the statistical character of microscale
properties from global observations of acceleration wave behaviours.
The constructive comments of a reviewer helped improve this paper. Support of this research
by the National Science Foundation under grants MSS-9202772 and CMS-9713764 is gratefully
acknowledged.

References
Alzebdeh, K., Al-Ostaz, A., Jasiuk, I. & Ostoja-Starzewski, M. 1998 Fracture of random matrixinclusion composites: scale effects and statistics. Int. J. Solids Struct. 35, 25372566.
Bland, D. R. 1969 Nonlinear dynamic elasticity. Waltham, MA: Blaisdell.
Chen, P. J. 1973 Growth and decay of acceleration waves in solids. In Encyclopedia of physics
(ed. S. Fl
ugge & C. Truesdell), vol. VI a/3. Berlin: Springer.
Christensen, R. M. 1982 Theory of viscoelasticity: an introduction. New York: Academic.
Coleman, B. D. & Gurtin, M. E. 1965 Waves in materials with memory. Arch. Ration. Mech.
Analysis 19, 239265.
Cramer, H. 1946 Mathematical methods of statistics. Princeton University Press.
DiMillo, M. & Ostoja-Starzewski, M. 1998 Paper strength: statistics and correlation structure.
Int. J. Fract. 90, L33L38.
Ditlevsen, O. & Madsen, H. O. 1996 Structural reliability methods. Wiley.
Kallianpur, G. 1980 Stochastic filtering theory. New York: Springer.
Kosi
nski, W. 1986 Field singularities and wave analysis in continuum mechanics. Chichester:
Ellis-Horwood, and Warsaw: Polish Scientific.
McCarthy, M. F. 1975 Singular surfaces and waves. In Continuum physics (ed. A. C. Eringen),
vol. 2. New York: Academic.
Menon, V. V., Sharma, V. D. & Jeffrey, A. 1983 On the general behavior of acceleration waves.
Applicable Analysis 16, 101120.
Milshtein, G. N. 1978 A method of second order accuracy integration of stochastic differential
equations. Theory Probab. Appl. 23, 2.
Ostoja-Starzewski, M. 1991 Transient waves in a class of random heterogeneous media. Appl.
Mech. Rev. (Special Issue: Mechanics Pan-America 1991) 44, S199S209.
Ostoja-Starzewski, M. 1993 On the critical amplitudes of acceleration wave to shock wave transition in white noise random media. J. Appl. Math. Phys. (ZAMP) 44, 865879.
Ostoja-Starzewski, M. 1994 Micromechanics as a basis of continuum random fields. Appl. Mech.
Rev. (Special Issue: Micromechanics of Random Media) 47, S221S230.
Ostoja-Starzewski, M. 1995 Transition of acceleration waves into shock waves in random media.
Appl. Mech. Rev. (Special Issue: Nonlinear Waves in Solids) 48, 300308.
Ostoja-Starzewski, M. 1998 Random field models of heterogeneous materials. Int. J. Solids
Struct. 35, 24292455.
Proc. R. Soc. Lond. A (1999)

2614

M. Ostoja-Starzewski and J. Trebicki

Ostoja-Starzewski, M. & Ilies, H. 1996 The Cauchy and characteristic boundary value problems
for weakly random rigid-perfectly plastic media. Int. J. Solids Struct. 33, 11191136.
Ostoja-Starzewski, M. & Trebicki, J. 1999 Acceleration waves in random nonlinear elastic/dissipative media. In 4th Int. Conf. Stochastic Structural Dynamics (ed. B. F. Spencer &
E. A. Johnson), pp. 235240. Rotterdam: Balkema.
Ostoja-Starzewski, M. & Wang, X. 1999 Stochastic finite elements as a bridge between random
material microstructure and global response. Comp. Meth. Appl. Mech. Engng 168, 3549.
Schuss, Z. 1980 Theory and application of stochastic differential equations. Wiley.
Sobczyk, K. 1991 Stochastic differential equations: with applications to physics and engineering.
Dordrecht: Kluwer.
Stoyan, D. & Stoyan, H. 1994 Fractals, random shapes and point fields. Wiley.
Stratonovich, R. L. 1963 Topics in the theory of random noise. New York: Gordon and Breach.

Proc. R. Soc. Lond. A (1999)

You might also like