Stability of Media and Structures

J R Willis
1 Introduction
Any structure that survives after construction must be stable, in the sense that it has
already demonstrated its ability to withstand a range of loads without undergoing un-
acceptable deection or distortion. It is, however, necessary to consider the question of
how great a load a structure can support, before its performance is compromised that
is, before it will collapse. Examples may be of several types: for instance, an exceptional
fall of snow could result in a shell roof supporting more weight than was envisaged in
its design. As the snow continues to fall, the load increases until the roof collapses.
Conceptually, at the instant before the last snowake landed, the roof was in a state of
unstable equilibrium, so that the small load represented by the last snowake caused
the large deection and the onset of the failure. Excessive loads from other sources are
easy to envisage. It is also possible that a structure may be stable but that, under some
exceptional conditions, one of its resonant vibrations is stimulated. In the absence of
sucient damping, this can result in the build-up of a vibration of large amplitude and
consequent failure. Earthquake damage can (but does not always) fall into this category.
The famous collapse of the Tacoma Narrows suspension bridge involved the development
of an oscillation of large amplitude, induced by wind, by the mechanism of utter.
Evidently, the safe design of a structure must take into account its possible modes
and frequencies of vibration, and must incorporate sucient margins of safety. Stability
(in the sense that a small disturbance induces only a small response in the structure)
is necessary but not sucient: what about a moderate disturbance, for instance?
These questions must be addressed quantitatively, in relation to the types of load that
a structure will experience during service.
This course provides an introductory account of the concepts and methodology re-
quired for the assessment of structural stability. Structural collapse can be a global
event, involving the whole structure, or it can result from a large deformation occurring
locally, because the material from which the structure is built reaches a critical condi-
tion. Some attention is also devoted to this topic, though specialized aspects such as the
development and propagation of cracks are not addressed; any such topic would require
a whole course by itself.
The remainder of this introduction is devoted to a simple example, which requires
no specialized knowledge and yet illustrates many of the features present in the analysis
of the stability of any structure.
1
Fig. 1.1. Model structure.
1.1 An elementary one-dimensional example
The conguration shown in Fig. 1.1 displays most of the features of the buckling of
a strut or a column under compression. A rigid rod OA of zero mass and length L is
pivoted at a point O, and its deection from the vertical (with A above O) is resisted
by a nonlinear spring, which exerts a restoring couple
C = f(); f(0) = 0 (1.1)
when OA makes an angle with the upward vertical. It is assumed that f

() > 0. A
point mass M is attached at A. A force acts vertically downwards at A. This could be
due to gravity acting on the mass M, in which case the force would have magnitude Mg.
However, to preserve generality, we let its magnitude be . The equation of motion of
this system follows from the balance of moment of momentum:
ML
2

= Lsin f(). (1.2)

Any equilibrium position is dened by

=

= 0, and must therefore satisfy
Lsin = f(). (1.3)
The number of equilibria depends on the form of the function f and the value of L.
The vertical conguration = 0 is in equilibrium for any value of L (though it need
not be stable). Consider the length L of the rod to be xed but suppose that the load
2
Fig. 1.2. Equilibrium paths. (a)
1
< 0; (b)
1
= 0,
2
> 0; (c)
1
= 0,
2
< 0.
is open to choice. Equation (1.3) then denes an equilibrium path in the - plane. The
path = 0 will be called the fundamental path. If f() has the expansion
f() = K
1
+ K
2

2
+ K
3

3
+ (K
1
> 0), (1.4)
another path dening a buckled state is connected to the fundamental path = 0 at
the critical mass
c
dened by

c
= K
1
/L = f

(0)/L. (1.5)
The buckled state, in the vicinity of = 0, lies on the path
=
c
+
1
+
2

2
+ , (1.6)
where

1
/
c
= K
2
/K
1
,
2
/
c
= (K
3
/K
1
+
1
6
), (1.7)
The point (0,
c
) is called a point of bifurcation. The path (1.6) denes the initial post-
bifurcation (or post-buckling) response. Figure 1.2 illustrates possible paths. Figure
1.2(a) illustrates an asymmetric bifurcation (the case
1
> 0 just reverses the slope
of the bifurcated path). Figures 1.2(b) and (c) illustrate symmetric bifurcations. The
response of an actual structure depends in part on the form of the post-bifurcation
response. Two aspects are investigated below.
3
Fig. 1.3. Imperfect model structure.
Eect of an imperfection
Real structures are never perfect. The eect of an imperfection may be illustrated by
introducing a small oset into the restoring spring. This is modelled by measuring
from the conguration in which the spring exerts no couple, which occurs when the rod
makes a small angle to the vertical, as shown in Fig. 1.3. The equation of motion now
becomes
ML
2

= Lsin( + ) f(), (1.8)

and any equilibrium conguration must satisfy
Lsin( + ) = f(). (1.9)
Thus, near = 0, and since [[ 1,
= [K
1
+ K
2

2
+ K
3

3
+ ]/[( + )
1
6
( + )
3
+ ]L. (1.10)
The path = 0 for the perfect structure is altered, to lowest order when is suciently
below
c
, to


(
c
/ 1)
. (1.11)
However, when is close to
c
, this approximation breaks down. If K
2
,= 0, a better
approximation is given by the solution of the quadratic equation
(K
1
L) + K
2

2
= L, (1.12)
4
Fig. 1.4. Perturbed equilibrium paths. (a) K
2
> 0; (b) K
2
< 0.
or
(K
2
/K
1
)
2
+ (1 /
c
) = (/
c
). (1.13)
The resulting equilibrium paths are sketched (for > 0) in Fig. 1.4. The branch that
passes through the origin has the equation
=
(1 /
c
) + [(1 /
c
)
2
+ 4(K
2
/K
1
)(/
c
)]
1/2
2(K
2
/K
1
)
. (1.14)
When /
c
is suciently smaller than 1, this reproduces the formula (1.11). The more
interesting of the two cases shown is K
2
< 0. The branch passing through the origin
displays a maximum allowed : increasing from zero would generate the deection
shown, until the maximum is reached. Any further attempt to increase must require
the structure to deform substantially. The deformation would be dynamic (limited by
inertia). Such a situation is termed a snap-through buckle.
Elementary analysis shows that the maximum admitted by (1.14) is given asymp-
totically, for small , by the formula
(1
max
/
c
) 2(K
2
/K
1
)
1/2

1/2
. (1.15)
Thus, the presence of the imperfection has reduced the buckling load by a quantity
proportional to

.
This conclusion can also be reached by introducing another small parameter and
writing
=
1
(1.16)
5
and
/
c
= 1 +
1
+
2

2
+ (1.17)
Substituting these into (1.10) and expanding, but keeping only the term of lowest order
in , gives
(1 +
1
+
2

2
+ )(
1

1
6

3
1

3
+ + )
=
1
+ (K
2
/K
1
)
2
1

2
+ (K
3
/K
1
)
3
1

3
+ (1.18)
Simplifying, therefore,
+
1

2
+ (
2

1
6

3
1
)
3
+
= (K
2
/K
1
)
1

2
+ (K
3
/K
1
)
3
1

3
+ (1.19)
All relevant formulae can now be obtained.
First, if = 0, equating terms of like order in gives

1
= (K
2
/K
1
)
1
,
2
= [(K
3
/K
1
) +
1
6
]
2
1
, (1.20)
exactly consistent with formulae (1.7).
Next, suppose that ,= 0 and K
2
,= 0. Retaining just terms of lowest order gives

1
=
K
2

1
K
1

2
, (1.21)
and then, for consistency, it is necessary to choose =
1/2
.
If K
2
= 0, equation (1.21) contains no interaction between the imperfection and
the parameters dening the spring. This suggests that the procedure as so far given is
unsuitable when K
2
= 0. A balance is obtained at order
3
if we take
1
= 0. Then,

2
=
_
K
3
K
1
+
1
6
_

2
1

3
(1.22)
and for consistency the choice =
1/3
must be made.
The most interesting cases correspond to K
2
< 0 and (K
3
/K
1
) + (1/6) < 0, respec-
tively. It follows then from (1.21) that
(/
c
1)
1
2(K
2
/K
1
)
1/2

1/2
(1.23)
and from (1.22) that

1
= 0,
2

2
3[(K
3
/K
1
) +
1
6
]/4
1/3

2/3
. (1.24)
6
The rst of these results implies (1.15), while the second gives

max
/
c
1 3[(K
3
/K
1
) +
1
6
]
1/3
(/2)
2/3
(K
2
= 0, [(K
3
/K
1
) +
1
6
] < 0). (1.25)
The pattern for cases in which more s vanish should now be apparent.
Dynamics and stability
By denition, the structure under consideration is fully described by the ordinary dier-
ential equation (1.2) (or (1.8) if the imperfect structure were considered). This is simple
enough to allow exact analysis. However, the present purpose is to introduce procedures
that can be applied more generally.
First, considering the perfect structure, = 0 denes an equilibrium conguration for
any value of . Stability of equilibrium is addressed, in the rst instance, via analysis of
the dierential equation, linearized about the equilibrium solution. In the present case,
the linear equation is simply
ML
2

= (L K
1
). (1.26)
Its general solution has the form
1
(t) = Ae
it
+ Be
it
, (1.27)
where

2
= (K
1
L)/ML
2
= [
c
]/ML. (1.28)
Thus, the solution is bounded and therefore remains small if it was initiated by a
small disturbance so long as <
c
. Conversely, if >
c
, becomes imaginary,
the solution grows exponentially (except for very special initial conditions that generate
only the negative exponential). The linearization under which it was derived becomes
invalid and study of what actually happens requires a return to the original nonlinear
dierential equation. In the former situation ( <
c
), the equilibrium conguration
= 0 is described as stable; in the latter it is unstable
2
.
The nonlinear dynamics may be investigated, in the vicinity of the critical point
(0,
c
), by retention of the next terms in the dierential equation (1.2). Thus, now,
ML
2

= L(
3
/6 + ) (K
1
+ K
2

2
+ K
3

3
+ ). (1.29)
1
In the case that is real, the most general real solution is obtained by taking A and
B to be complex conjugates.
2
Stability will be dened formally later
7
If K
2
,= 0, retaining just terms up to order
2
gives

=
_

ML
___

c

_

_
K
2

c
K
1

2
_
. (1.30)
The procedure underlying the derivation of (1.30) can be formalised by writing
/
c
= 1 +
1
, =
1
, and =
1/2
t. (1.31)
The variable
1
is regarded as a function of the slow time variable . Substituting into
(1.2) then gives
ML
2

1
=
c
L[(1 +
1
)(
1

3
1

3
/6 + )
(
1
+ (K
2
/K
1
)
2
1

2
+ (K
3
/K
1
)
3
1

3
+ )], (1.32)
the prime denoting dierentiation with respect to . The terms of order cancel.
Equating those of order
2
gives

1
= A
1
+ B
2
1
, (1.33)
where
A =
1

c
/ML, B = (K
2
/K
1
)
c
/ML. (1.34)
The dierential equation (1.33) admits constant solutions
1
= 0 and
1
= A/B.
The latter corresponds exactly to the initial post-buckling path [c.f. (1.6) with (1.7)].
Linearizing about
1
= 0 shows that this solution is stable so long as A < 0. Linearizing
about
1
= A/B gives the equation

1
= A
1
, (1.35)
having written
1
= A/B+
1
. Thus, the solution
1
= A/B is stable if A > 0. Phase
portraits (plots in a
1
-

1
plane) are sketched in Fig. 1.5. When A < 0, so that
1
= 0
is stable against an innitesimal perturbation, study of (1.33) permits an assessment of
exactly how large a perturbation would be allowed, before the solution would deviate
far from
1
=

1
= 0. The interest of equation (1.33) is that it will be seen to emerge
generically from weakly-nonlinear stability analysis in the vicinity of a non-symmetric
bifurcation.
If K
2
= 0, a dierent parametrization is required. A balance is obtained if
/
c
= 1 +
2

2
, =
1
, = t. (1.36)
8
Fig. 1.5. Phase portraits for the dierential equation (1.33).
9
The equation that results is

1
= A
1
+ B
3
1
, (1.37)
where
A =
2

c
/ML, B = (K
3
/K
1
+
1
6
)
c
/ML. (1.38)
The constant solutions
1
= (A/B)
1/2
(which exist if A/B < 0) correspond to the
initial post-buckling path [c.f. (1.6) and (1.7) with K
2
= 0]. Equation (1.37) will emerge
as a generic feature associated with a symmetric bifurcation.
The eect of an imperfection () can be incorporated by adding its lowest-order
contribution to the governing dierential equation. This has the eect of replacing
(1.33) by

1
= A
1
+ B
2
1
+ C(/
2
), (1.39)
where
C =
c
/ML. (1.40)
[For consistency, it is necessary that /
2
= O(1)]. The equilibrium point of (1.39)
agrees with the approximation given by (1.12). If K
2
= 0, an analogous modication
follows for (1.37).
10
2 Stability of systems: general discussion
This section discusses the stability of systems with any nite number of degrees of
freedom. Although real structures are continua, they are almost always modelled as
discrete for the purpose of stress analysis for example by the use of nite elements
and hence in practice this discussion will apply, at least at a formal level, to virtually
all structures, as well as to other dynamical systems. It is usual to consider a rst-order
system,
u = f(u, t; ), (2.1)
where u : R R
n
is a vector-valued function of time t. The function f has the
arguments shown and takes values in R
n
. The system (1.2) ts this pattern. With the
denitions u
1
= , u
2
=

, it can be written
u
1
= u
2
,
u
2
= [Lsin u
1
f(u
1
)]/(ML
2
). (2.2)
In the general equation (2.1), represents any number m of scalar parameters; that
is, it could be an m-dimensional vector. The most signicant dierence between the
general system (2.1) and the realisation (2.2) is that (2.1) contains time t explicitly: it
is not autonomous. In fact, in the sequel only autonomous systems will be considered.
However, some formal denitions of stability will be given for the system (2.1).
Suppose that u
0
(t) is a particular solution of (2.1). It is called stable if the solution of
the initial-value problem comprising (2.1) for t > t
0
, together with the initial condition
u(t
0
) = u
0
(t
0
) + v
0
(2.3)
has the property
|u(t) u
0
(t)| 0 as 0 (2.4)
uniformly for all t t
0
, for all v
0
with |v
0
| = 1. Here, | | can be taken as the
Euclidean norm. This choice of norm is not important, however, because all norms on
a nite-dimensional vector space are equivalent.
If the solution u satises the stronger requirement that
|u(t) u
0
(t)| 0 as t (2.5)
for all suciently small ([[ < for some > 0), the system (2.1) is called asymptotically
stable.
11
The discussion of the preceding section strongly suggests that the solution u = 0
of the system (2.2) is stable but not asymptotically stable, when <
c
. This can be
veried from its original form (1.2) by noting that

=

d(

)/d and integrating, to

obtain the energy integral
1
2
ML
2

2
+
_

0
f(

)d

L(1 cos ) = E, constant. (2.6)

The condition <
c
ensures that the function on the left side of (2.6) is a convex
function of (

, ) in a neighbourhood of (0, 0) and hence, when the constant E on the

right side, which is xed by the initial conditions, is suciently small, (

, ) remains
close to (0, 0) for all t.
Before proceeding, it is appropriate to express a word of caution in relation to con-
tinuous systems. All norms are not equivalent for such systems, and so not only the
denition of stability, but also its relevance or utility, will depend upon the appropriate
choice of norm. A related concern is the possibility that any chosen nite-dimensional
approximation of a continuous system simply might not contain some instability of the
original system, though such a feature would be likely to show up in practice in the form
of strong sensitivity of predictions made from the discrete system to the precise detail,
such as the nite element mesh. A pragmatic approach is adopted throughout these
lectures: formal methods, such as the perturbation theory already used in Section 1,
will be employed. Such methods do not establish rigorously when instability is reached
but they have the virtue of providing fully explicit indications of the nature of likely
instabilities, and associated estimates for quantities such as critical loads.
Having made these general remarks, we now consider a system of the form
3
M u = F(u, ). (2.7)
Here, u is a vector, F is vector-valued and M is a matrix. is a scalar parameter that
represents the intensity of the loading on the system.
Various aspects of the system (2.7) are now considered, in turn.
Equilibrium
Equilibrium congurations of the system must satisfy the equation
F(u, ) = 0. (2.8)
3
Complications will be considered later.
12
This equation may have several solutions (it will be assumed to have at least one, at
least for some range of values of ). Their number may change with the value of . The
functional dependence upon of any one solution may be investigated by regarding
to be an increasing function of time (or any time-like parameter), while insisting that
the equilibrium condition (2.7) remains satised. Then, dierentiating (2.7) gives
F
u
(u, ) u + F

(u, )

= 0. (2.9)
Just to explain the notation employed here, regard F as a vector with r-component F
r
.
Equation (2.8) is equivalent to
F
r,us
(u, ) u
s
+ F
r,
(u, )

= 0, (2.10)
the comma denoting a partial derivative with respect to the following variable, and
summation over the repeated sux s is implied. Assuming that the matrix F
u
is not
singular, this equation has the unique solution
u = [F
u
]
1
F

. (2.11)
This is equivalent to the dierential equation
du/d = [F
u
]
1
F

,
which denes the branch of the solution of (2.7) that is being followed. Any such branch
is called an equilibrium path.
Suppose, now, that an equilibrium path is followed, with increasing, until a crit-
ical value
c
is reached, with corresponding equilibrium conguration u
c
, at which F
u
becomes singular. There are two possibilities:
(i) Equation (2.8) has no solution. In this case, cannot be increased beyond
c
for this
branch. Any attempt to increase would have to result in motion, in which the inertia
m u is important.
(ii) Equation (2.8) has non-unique solution. The point (u
c
,
c
) is then a point of bifur-
cation.
In either case, (u
c
,
c
) is a critical point.
Uniqueness
Another perspective on the same phenomenon is gained by considering directly
whether equation (2.7) has a unique solution. Suppose that there are two solution
13
branches, u
1
() and u
2
(), and suppose that for some set of values of they are close
together. It follows that
0 = F(u
1
, ) F(u
2
, ) F
u
(u
2
, )(u
1
u
2
), (2.12)
which implies that F
u
must be singular, and that u
1
and u
2
may dier only by a multiple
of the eigenvector of F
u
. In case (ii) above, the two distinct branches cross at (u
c
,
c
).
In case (i), it may occur that u
1
and u
2
are dierent parts of a single branch that turns
over, as illustrated in Fig. 1.3(b).
4
However, it is conceivable that the branch simply
terminates, the nonlinear terms omitted from (2.12) preventing its continuation. Thus,
by itself, linearized analysis provides an indication of what may happen but does not
predict what will happen.
Inuence of an imperfection
Suppose that an imperfection is present, whose magnitude is described by the parameter
. Recall that, in the example given in Section 1, the imperfection was a small tilt of
the bar away from the vertical, when in equilibrium under zero load. More generally,
an imperfection could be any geometrical feature, or perhaps some variation in stiness,
or perhaps both. In the present general formulation, the presence of the imperfection
is represented by replacing the function F(u, ) in (2.8) by F(u, , ). It is possible, in
fact, to let be a vector of any nite dimension, so that it represents the eect of several
types of imperfections. Equilibrium is now governed by the equation
F(u, , ) = 0, (2.13)
with = 0 dening the structure with no imperfection. The primary solution branch
(that is, the one in which we are interested) is denoted u
0
when = 0. That is,
F(u
0
, , 0) = 0. (2.14)
It is convenient to re-dene variables so that u
0
0. Thus,
F(0, , 0) = 0 (2.15)
for all . Now call the perturbed solution (for ,= 0) u, and assume that u is small and
u 0 as 0. Then,
F
u
(0, , 0)u + F

(0, , 0) 0, (2.16)
4
This gure is for a structure regarded as imperfect but the phenomenon can occur
generally.
14
which implies that
u [F
u
(0, , 0)]
1
F

(0, , 0), (2.17)

except when is close to
c
(where the matrix F
u
is singular).
The initial post-bifurcation path
Before considering the perturbed path when is close to
c
, it is useful to examine
further the response of the unperturbed system in this vicinity. Within the present
framework, the primary solution branch is u = u
0
0, so the critical point (u
c
,
c
)
becomes (0,
c
). Assume that this is a point of (simple) bifurcation. Then, when is
close to
c
, there is one other solution v say, and v 0 as
c
. Therefore, expanding
the equation
F(v, , 0) = 0
about (0,
c
, 0),
F
u
v+
1
2
F
uu
v
2
+F
u
v(
c
)+
1
6
F
uuu
v
3
+
1
2
F
uu
v
2
(
c
)+
1
2
F
u
v(
c
)
2
+ = 0. (2.18)
Here, for example, F
uu
v
2
represents the vector whose i-component is F
i,urus
v
r
v
s
(sum-
mation over r and s implied). All derivatives of F are here evaluated at (0,
c
, 0).
Derivatives with respect to by itself (such as F

) are not included because they are

zero, on account of (2.15).
Now clearly, to lowest order, v is a multiple of the right eigenvector of F
u
(0,
c
, 0),
as found earlier
5
. Suppose, therefore, that u

1
and u
1
are respectively left and right
eigenvectors:
u

1
F
u
= 0 and F
u
u
1
= 0. (2.19)
Now when is close to
c
, v is small and (asymptotically) parallel to u
1
. Therefore,
introduce a small parameter and write
v = v
1
+
2
v
2
+ ,
=
c
+
1
+
2

2
+ . (2.20)
Substituting these into (2.17) gives
F
u
v
1
+
2

1
2
F
uu
v
2
1
+
1
F
u
v
1
+ F
u
v
2
+
3

1
6
F
uuu
v
3
1
+
1
2

1
F
uu
v
2
1
+
1
2

2
1
F
u
v
1
+
2
F
u
v
1
+
1
F
u
v
2
+ F
uu
v
1
v
2
+ F
u
v
3
+ = 0.
(2.21)
5
There is only one linearly independent right eigenvector, on account of the assumption
that the bifurcation point is simple.
15
Therefore, by considering the coecient of ,
F
u
v
1
= 0, (2.22)
implying that v
1
=
1
u
1
for some
1
. Now considering the coecient of
2
,
1
2
F
uu
v
2
1
+
1
F
u
v
1
+F
u
v
2
= 0. (2.23)
This implies necessarily (c.f. (2.19)
1
) that

1
=

1
u

1
F
uu
u
2
1
2u

1
F
u
u
1
, (2.24)
and then (2.23) can be solved for v
2
. The solution is unique only up to a term
2
u
1
.
The coecient of
3
can be treated similarly: it yields

2
=
u

1
6
F
uuu
v
3
1
+
1
2

1
F
uu
v
2
1
+
1
2

2
1
F
u
v
1
+
1
F
u
v
2
+ F
uu
v
1
v
2

1
u

1
F
u
u
1
. (2.25)
The simplest way to x
1
and
2
(and corresponding constants in succeeding terms) is
to dene in terms of v by insisting that = u
T
v for some vector u, such that u
T
u
1
= 1.
Then,
1
= 1 and the requirement that u
T
v
2
= 0 xes
2
. Another possibility would be
to insist that =
c
+
1
exactly, so that
2
=
3
= = 0. Then, equation (2.24)
gives
1
in terms of
1
, and (2.25) xes
2
. In any case, v
3
exists so long as (2.25) is
satised.
If u

1
F
u
u
1
= 0, then (2.24) and (2.25) do not apply. A balance of terms is obtained
if v
1
= 0 and v
2
=
2
u
1
. This rather pathological case is not discussed further.
The imperfect system
Now revert to the imperfect system, with ,= 0. Expanding the equation
F(v, , ) = 0
about the point (0,
c
, 0) gives
F
u
v +
1
2
F
uu
v
2
+ F
u
v(
c
) +
1
6
F
uuu
v
3
+
1
2
F
uu
v
2
(
c
)
+
1
2
F
u
v(
c
)
2
+ + F

= 0, (2.26)
having retained only the leading-order term with respect to . Again, set
v = v
1
+
2
v
2
+ ,
16
and let
=
c
+ (
1
+

1
) + (
2
+

2
)
2
+ , (2.27)
where
r
are as before, so that the terms

r
give the additional perturbation of due
to .
Note rst that an attempt to balance the term proportional to with the term
that is linear in generally cannot succeed, because the condition for consistency of the
equation
F
u
v
1
+ F

= 0
is u

1
F

= 0, which is not usually the case. Therefore, it is necessary to assume that is

of order
k
for some k > 1. Then, as obtained previously, F
u
v
1
= 0 and so v
1
=
1
u
1
for
some
1
.
If is of order
2
, equating terms of order
2
gives
F
u
v
2
+
1
2

2
1
F
uu
u
2
1
+
1
(
1
+

1
)F
u
u
1
+F

/
2
= 0. (2.28)
The condition for consistency is
u

1
2

2
1
F
uu
u
2
1
+
1
(
1
+

1
)F
u
u
1
+ F

/
2
= 0.
Therefore,

1
=
u

1
F

1
u

1
F
u
u
1

2
, (2.29)
since
1
as given by (2.24) cancels the other terms. Thus, to this order,

c

_
u

1
F
uu
u
2
1

1
2u

1
F
u
u
1
+
u

1
F

1
F
u
u
1

2
_
. (2.30)
The perturbation is of the form
(A
1
+ B/
1
),
whose greatest value is 2(AB)
1/2
if A > 0 and B > 0. Thus, in this case,

c

2[(u

1
F
uu
u
2
1
)(u

1
F

)[
1/2
[u

1
F
u
u
1
[
. (2.31)
A snap-through buckle is indicated at a level of an amount of order ||
1/2
lower than

c
.
17
If
1
= 0 (i.e. u

1
F
uu
u
2
1
= 0), a dierent scaling is needed to obtain the desired
balance. It is appropriate, in fact, to take k = 3. Then,

1
= 0 and equating terms of
order
3
gives

2
=
u

1
F

1
u

1
F
u
u
1

3
.
Then,

c
+
2

1
F

1
u

1
F
u
u
1

. (2.32)
This has the form

c
A(
1
)
2
+ B/
1
,
and if A > 0 and B > 0, it follows that

c
3A
1/3
(B/2)
2/3
. (2.33)
The reduction in the critical load is of order ||
2/3
.
Stability
We consider now the stability of the primary equilibrium path u = 0 for the perfect
structure. This requires study of the system of dierential equations
M u = F(u, ), with F(0, ) = 0. (2.34)
First, linearizing gives
M u = F
u
(0, )u, (2.35)
for which a solution may be sought of the form
u(t) = ve
it
.
This generates the eigenvalue problem
[F
u
(0, ) +M]v = 0 ( =
2
). (2.36)
Since the system is real, complex eigenvalues must occur in complex conjugate pairs, so
instability is inevitable unless all eigenvalues are real and positive. It is reasonable to
assume that the system is stable at least for small loads (small ), so assume that all
18
eigenvalues are real and positive for <
c
. Suppose, furthermore, that the smallest
eigenvalue
1
= 0 when =
c
and that it is simple. This gives
F
u
(0,
c
)u
1
= 0,
exactly as discussed already. The solution branch u = 0 will be unstable for some range
of with >
c
, if decreases below zero when increases beyond
c
.
Now some weakly-nonlinear analysis can be developed assuming that is close to

c
. Let
=
c
+
1
, (2.37)
u = v
1
+
2
v
2
+ , (2.38)
scale the time so that
=
1/2
t (2.39)
and regard v
i
as functions of . Substituting into (2.34) then gives
F
u
v
1
+
2

1
F
u
v
1
+
1
2
F
uu
v
2
1
Mv

1
+F
u
v
2
+ = 0, (2.40)
the prime denoting dierentiation with respect to . It follows that v
1
= A()u
1
, and
that the scalar-valued function A() must satisfy the equation
(u

1
Mu
1
)A

=
1
(u

1
F
u
u
1
)A +
1
2
(u

1
F
uu
u
2
1
)A
2
. (2.41)
In the presence of an imperfection, an additional term F

is added to the left side

of (2.40). If can be regarded as of order
2
, equation (2.41) simply becomes altered to
(u

1
Mu
1
)A

=
1
(u

1
F
u
u
1
)A +
1
2
(u

1
F
uu
u
2
1
)A
2
+ u

1
F

/
2
. (2.42)
It should be noted that equation (2.41) has equilibrium solutions corresponding to
A = 0 and A =
2
1
(u

1
F
u
u
1
)
u

1
F
uu
u
2
1
.
This is exactly consistent with the static post-bifurcation analysis performed earlier (c.f.
(2.24)).
The main point about the equation (2.41) is that it arose generically, from a study
of a fairly general system with several degrees of freedom, as was announced in the
Introduction, where an equation of this type arose from study of a simple one-dimensional
19
example. Equation (2.42) can be reduced to the form (2.41) by adding a suitable constant
to A. Equation (2.41) provides immediately an estimate for how large a perturbation can
be, when the system is stable but close to instability. That is, it provides an estimate
for the margin of stability. The phase portraits illustrated in Fig. 1.5 provide this
information in graphical form.
It is left as a relatively simple exercise to analyze the case in which u

1
F
uu
u
2
1
= 0.
Flutter
It is appropriate at least to mention a phenomenon that is intrinsically dynamic in na-
ture. It cannot occur unless the matrix F
u
is non-symmetric. In this case, the possibility
exists that two eigenvalues,
1
and
2
say, coincide at a load lower than
c
(where
c
still
denotes the load at which F
u
(0, ) rst becomes singular), so that they are still positive,
but then, as increases further, they split and become complex conjugate pairs, having,
at least initially, positive real parts. The elementary static bifurcation theory gives no
hint of trouble: there is no equilibrium solution close to the primary one u = 0. How-
ever, even a linearized dynamical analysis predicts instability, because there are values
of (c.f. (2.36)) which will have negative imaginary parts. This type of bifurcation is
called a Hopf bifurcation. Weakly-nonlinear analysis is possible for this situation also:
the motion is basically harmonic, but with an amplitude that evolves slowly in time.
It is more complicated than the analysis presented above, because of the need to deal
simultaneously with two time-scales (the ordinary one which appears in the simple
harmonic motion and the slow time , upon which the amplitude depends), and is not
pursued in detail.
Conservative systems
If, in fact, F(u, ) is derived from a scalar potential , so that F(u, ) =
u
(u, ),
then automatically the matrix F
u
=
uu
is symmetric. Conversely, if F
u
is symmetric,
a potential exists. Then, assuming also that the mass matrix M is symmetric and
independent of u, the equation of motion (2.7) has the following rst integral
6
1
2
u
T
M u + (u, ) = E, constant, (2.43)
which is an expression of conservation of energy. Now let u = u
0
+ v, where u
0
is
an equilibrium solution so that F(u
0
, ) =
u
(u
0
, ) = 0, and v is a small time-
6
Exposure to a course on Lagrangian and Hamiltonian dynamics would permit the
derivation of this result in greater generality.
20
dependent perturbation. Expanding (2.41) about u
0
to second order (which is equivalent
to linearizing (2.7)) gives
1
2
v
T
M v +
1
2

uu
v
2
= E (u
0
, ). (2.44)
The kinetic energy quadratic form is positive-denite. It follows that if, also, the
quadratic form
uu
v
2
is positive-denite, a perturbation that is bounded initially remains
bounded, and therefore that the equilibrium conguration u
0
is stable against a small
perturbation. Conversely, if the form
uu
v
2
is indenite (or even negative-denite), there
will exist disturbances for which linearized analysis would predict exponential growth,
and thus instability. This is the Dirichlet condition for stability: u
0
is stable if it attains
a local minimum for the potential energy function .
21
3 The Euler column
The approach outlined in the preceding section can be applied also when the structure
is a continuum (so having an innite number of degrees of freedom). A strict discussion
would entail the introduction of suitable function spaces and corresponding norms. Such
machinery would be out of place here. It is, nevertheless, still possible to track the
reasoning presented in the preceding section virtually line-by-line, to derive at least a
formal (and physically credible) description of the static and dynamic characteristics of
a structure, close to a critical point. This will now be illustrated by considering the
classical example of the buckling of the Euler column. The column is modelled as a one-
dimensional structure that can support tension or compression, and also has resistance
to bending. Such a structure (which is called an elastica) can resist torsion as well, but
this aspect is not needed for the present example.
3.1 Equations of motion
It is necessary rst to set up equations of motion. For this purpose, with reference to Fig.
3.1, the column (or beam) is modelled as initially straight. Arc length s measured, in the
undeformed conguration from one end (labelled O), is taken as Lagrangian coordinate.
Only deformations in the x, y plane are envisaged. Therefore, the deformed conguration
at time t is specied by the mapping s (x(s, t), y(s, t)). The parts of the beam on
either side of any point P exert a resultant force and couple on each other. The part OP
experiences, at P, a force with components T along the beam and N normal to it, and a
couple of moment M, as illustrated. The complementary part experiences the opposite
force and couple. The unit tangent at P has components (x

, y

)/(x
2
+ y
2
)
1/2
, and the
normal has components (y

, x

)/(x
2
+ y
2
)
1/2
, the prime denoting dierentiation with
respect to s. Let the beam have mass per unit length m. Equating the rate of change
of linear momentum of the section OP to the forces applied to it gives
d
dt
_
s
0
m
_
x
y
_
ds =
__
T
_
x

_
+ N
_
y

__
/(x
2
+ y
2
)
1/2
_
s
0
. (3.1)
Equating the rate of change of moment of momentum to the total applied moment gives
d
dt
_
s
0
m(x y y x)ds =
_
T(xy

yx

) + N(xx

+ yy

)
(x
2
+ y
2
)
1/2
+ M
_
s
0
. (3.2)
It follows by dierentiation of these equations with respect to s that
m x = (Tx

Ny

)/(x
2
+ y
2
)
1/2

, (3.3)
22
Fig. 3.1. Forces and couple on the section OP of the beam.
m y = (Ty

+ Nx

)/(x
2
+ y
2
)
1/2

, (3.4)
0 = N(x
2
+y
2
)
1/2
+ M

, (3.5)
the last equation having been simplied by use of the rst two.
The formulation is completed by appending constitutive equations, which charac-
terise the response of the beam being considered. In general, T is related to the local
stretch (x
2
+y
2
)
1/2
, and M is related to the Lagrangian curvature , where
7
=
y

(x
2
+ y
2
)
. (3.6)
However, we will adopt the idealisation that the beam is inextensible. Then it is subject
to the constraint
(x
2
+ y
2
) = 1 (3.7)
and T becomes an undetermined multiplier. The constitutive relation for M will be
taken to be
M = f() = B
1
+ B
3

3
+ , (3.8)
where now the inextensibility constraint reduces to
= y

. (3.9)
Equation (3.5) simplies correspondingly.
7
This choice is in the spirit of taking material derivatives. It is based on the formula
= dtan
1
(y

/x

)/ds. More generally, M could depend on local stretch and . However,

inextensibility will be assumed in any case.
23
Fig. 3.2. The Euler column.
3.2 The problem
The problem to be studied is illustrated in Fig. 3.2. The column is initially vertical, and
its end O is clamped so that x = 0 and x

= 0 when s = 0. A dead load of magnitude

is applied, vertically downwards, at the upper end of the column, s = l. No moment is
applied at the upper end. Therefore, the boundary conditions are that
x(0, t) = y(0, t) = 0, x

(0, t) = 0,
T(l, t) = y

(l, t), N(l, t) = x

(l, t), M(l, t) = 0. (3.10)

Evidently, one solution of the equations of motion is
x = 0, y = s, and T = . (3.11)
This corresponds to the fundamental equilibrium path u
0
of the preceding section; our
basic objective is to examine its incremental uniqueness as increases, and its stability.
Static analysis
First, equilibrium congurations are studied by considering time-independent solutions.
These satisfy the equations
(Tx

Ny

= 0, (3.12)
(Ty

+ Nx

= 0, (3.13)
N + M

= 0, (3.14)
M = f(), (3.15)
x
2
+ y
2
= 1. (3.16)
24
The rst two of these equations may be integrated and the constants xed from their
known values at s = l. Solving the resulting two equations then gives
N = x

, T = y

. (3.17)
Although this is not essential, it is convenient to satisfy the constraint (3.16) identically
by setting
x

= sin , y

= cos . (3.18)
Then, =

. Substituting all of the relations so far established into (3.14) now yields
(B
1
+ 3B
3

2
+ )

+ sin = 0. (3.19)
The boundary conditions are (0, t) =

(l, t) = 0. As already observed, one solution is

= 0.
Bifurcation
To investigate bifurcation, suppose that there is another solution, close to = 0.
This can be investigated by linearizing (3.19):
B
1

+ = 0. (3.20)
The solution for which (0) = 0 is
= Asin[(/B
1
)
1/2
s], (3.21)
and the smallest value of that satises the condition at s = l (with A ,= 0) is

c
=
B
1

2
4l
2
. (3.22)
The post-bifurcation path can be studied asymptotically by setting
=
1
+
2

2
+ , =
c
+
1
+
2

2
+ . (3.23)
Then,
(B
1
+ 3B
3

2
1
+ )(

1
+
2

2
+ )
+ (
c
+
1
+
2

2
+ )(
1
+
2

2
+
3
(
3

1
6

3
1
) + ) = 0. (3.24)
25
Equating to zero the coecient of gives
B
1

1
+
c

1
= 0, (3.25)
which corresponds exactly to the linearization (3.20), and the associated boundary con-
ditions. Thus,

1
= Asin(s/2l), (3.26)
having taken into account the denition (3.22) of
c
. The terms of order
2
give
B
1

2
+
c

2
+
1

1
= 0. (3.27)
There is no solution
2
that satises the boundary conditions unless
1
= 0. Then with
this condition,
2
has the same form as
1
and nothing is lost if it is specied that

2
= 0. The terms of order
3
now give
B
1

3
+
c

3
+ 3B
3

2
1

1
+
2

1
6

3
1
= 0. (3.28)
The condition for consistency of this equation, with
3
satisfying the required boundary
conditions, is obtained in exactly the same way as in Section 3. The analogue of the left
eigenvector is the function sin(s/2l). Multiplying equation (3.28) by sin(s/2l) and
integrating from 0 to l gives, necessarily,
_
l
0
sin(s/2l)3B
3

2
1

1
+
2

1
6

3
1
ds = 0, (3.29)
since integration by parts and use of the boundary conditions cancels out the terms
involving the still-unknown
3
. Changing the variable of integration to u = s/2l and
substituting explicitly for
1
gives
_
/2
0
sin u
_
A
3
_
3B
1

4
16l
4
cos
2
usin u
1
6

c
sin
3
u
_
+ A
2
sin u
_
du = 0. (3.30)
The required integrals are
_
/2
0
sin
2
udu = /4,
_
/2
0
sin
4
udu = 3/16,
_
/2
0
cos
2
usin
2
udu = /16. (3.31)
Thus,
A
2
A
3
_

c
8
+
3B
3

4
64l
4
_
= 0. (3.32)
26
Fig. 3.3. Imperfect structure.
(a) Beam vertical, load o-vertical, (b) Beam o-vertical, load vertical.
Therefore, if A ,= 0, then A and
2
must be related so that

2
= A
2
_

c
8
+
3B
3

4
64l
4
_
. (3.33)
This equation is the analogue of (2.25) for the problem of the Euler column.
The eect of an imperfection
One obvious possible imperfection is that the beam may not be exactly straight.
Analysis of this would require the development of equations of equilibrium (and also
of motion) for such a beam. This is avoided here by considering an alternative simple
model: the direction of the load is not exactly vertical but instead makes an angle
with the downward vertical, as shown in Fig. 3.3(a). This is equivalent to vertical
loading of a beam whose unloaded conguration is not quite vertical (as depicted in Fig.
3.3(b)), because gravity has already been disregarded in the discussion of the perfect
structure, and will continue to be ignored here.
Equations (3.12) and (3.13) still apply, but now their integration in conjunction with
specifying that the force at the end s = l has horizonal and vertical components sin
and cos respectively gives
T = (x

sin + y

cos ), N = (x

cos + y

sin ). (3.34)
Equivalently, with x

and y

expressed in terms of as in (3.18),

T = cos( + ) cos + sin ,
N = sin( +) sin cos , (3.35)
27
having retained only the perturbation of order . This perturbation generates a term
additional to those displayed in equation (3.19). The perturbed equation is
(B
1
+ 3B
3

2
+ )

+sin + cos 0. (3.36)

The perturbation of the primary solution = 0 due to the perturbation, when is
close to
c
, can be investigated by again postulating the expansions (3.23), and now
substituting into (3.36). It is necessary to decide how to relate to . If it is assumed
that they are of the same order, then the term of order in (3.37) gives
B
1

1
+
c

1
+
c
/ = 0. (3.37)
This equation has no solution
1
satisfying the boundary conditions unless the last term
is zero. Equivalently, is (at least) of order
2
. Once this is assumed, it follows that

1
= sin(s/2l), exactly as before. The equation that results from considering the term
of order
2
, subjected to similar reasoning, leads to the conclusion that should be of
order
3
, if the perturbation expansion is to succeed. Then,
2
can be taken to be zero,
without loss. The coecient of
3
now gives
B
1

3
+
c

3
+ 3B
3

2
1

1
+
2

1
6

3
1
+
c
/
3
= 0. (3.38)
The condition for consistency is obtained by multiplying the equation by sin(s/2l) and
integrating from 0 to l. There is one additional term in comparison with (3.29). This is

3
_
l
0
sin(s/2l) ds =
_
2l

_

c

3
. (3.39)
It follows that
A
2

4
A
3
_

32
+
3B
3

5
256l
4
_
+

c

3
= 0. (3.40)
The equilibrium path for the imperfect structure, close to the critical point, therefore
has the asymptotic form
(A) sin(s/2l),
c

4
c

(A)
+
_

c
8
+
3B
3

4
64l
4
_
(A)
2
. (3.41)
A maximum load is indicated, if the term in curly brackets is negative. In this case, the
maximum load is smaller than
c
by an amount of order
2/3
.
28
3.3 Stability
The Dirichlet condition
The system under discussion is conservative. Therefore, one of the elementary ways to
consider the stability of an equilibrium path is to investigate whether the conguration
realises a local energy minimum (the Dirichlet condition). The energy is
E =
_
l
0
[
1
2
B
1

2
+
1
4
B
3

4
+ ] ds + y(l), (3.42)
where
y(l) =
_
l
0
cos ds. (3.43)
The requirement is to compare the energy evaluated at the solution with the energy
evaluated for a neighbouring conguration. For a point on the primary path = 0, the
energy of a neighbouring conguration is given by (3.42), expanded to lowest non-trivial
order when is small. The energy dierence is, asymptotically,
E = E() E(0)
1
2
_
l
0
[B
1

2
] ds. (3.44)
The only restriction on (s) is that (0) = 0. Now we investigate whether there is
any function (with (0) = 0) for which E is negative. Since E is homogeneous of
degree 2 it suces to restrict further so that
_
l
0

2
ds = 1.
It is necessary then to introduce a Lagrange multiplier, say, which has the eect of
replacing by + in the functional (3.44). Candidate minimizers, subject to this
constraint, can be found by perturbing to + . The functional is stationary if
_
l
0
[B
1

( + )] ds = 0 (3.45)
for all allowed . Integrating the rst term by parts and imposing the boundary
condition gives

(l)(l)
_
l
0
[B
1

+ ( + )]ds = 0 (3.46)
for all allowed . It follows that must satisfy
B
1

+ ( + ) = 0,
29
and

(l) = 0 in addition to (0) = 0. There is no such stationary point (apart from

= 0) unless satises
( + )/B
1
= (2k + 1)
2

2
/4l
2
for some integer k. The corresponding is
(s) = (2/l)
1/2
sin[(2k + 1)s/2l].
The stationary value of E then follows as
1
2
_

2
(2k + 1)
2
B
1
4l
2

_
.
This is positive and so the solution = 0 is stable for 0 <
c
.
Linearized dynamics
The study of dynamics requires a return to the system of equations (3.3)-(3.5) and (3.7)-
(3.9). Linearized about the equilibrium solution x = 0, y = s, T = , N = 0, they
give
( x

N)

= m

x, (3.47)

N B
1
x

= 0, (3.48)
the quantities x, etc. representing the perturbations. There is no equation for y because
the constraint of inextensibility gives y 0. Elimination of

N gives
x

B
1
x

= m

x. (3.49)
The boundary conditions (3.10) imply that
x(0, t) = x

(0, t) = x

(l, t) = x

(l, t) +B
1
x

(l, t) = 0. (3.50)
Normal mode solutions may now be sought by assuming exp(it) time dependence. The
partial dierential equation (3.49) then implies
B
1
x

+ x

m
2
x = 0. (3.51)
The solution of this fourth order ordinary dierential equation, together with the
boundary conditions (3.50), is algebraically complicated. It will not be pursued further,
30
except to remark that the problem so dened is an eigenvalue problem, since the dif-
ferential equation and boundary conditions are homogeneous, and that any eigenvalue
m
2
must be real, because the problem is self-adjoint. To see this, multiply equation
(3.51) by a function u and integrate from 0 to l. This gives, employing integration by
parts,
_
l
0
uL x ds
_
l
0
u[B
1
x

+ x

m
2
x] ds
=
_
l
0
[B
1
x

m
2
xu] ds [B
1
x

( x

+ B
1
x

)u]
l
0
,
(3.52)
having called the dierential operator L. Thus, if both x and u satisfy the boundary
conditions (3.50), the form on the right side of (3.52) is symmetric and it follows that
_
l
0
uL x ds =
_
l
0
xLuds. (3.53)
It is easy to deduce from this symmetry just as for symmetric matrices that eigenval-
ues m
2
must be real. The primary solution is thus stable so long as all eigenvalues are
positive, and this requirement is rst lost when the smallest eigenvalue becomes zero. It
is known already, from the preceding subsection, that this occurs when =
c
.
Weakly-nonlinear dynamics
It is interesting to observe that, even though exact linearized analysis is complicated, asymptotic analysis
close to the critical point is relatively easy, and furthermore nonlinear terms can be retained. The pattern
follows that already established in Section 3. Let
x = x
1
+
2
x
2
+ ,
y = s + y
1
+
2
y
2
+ ,
T =
c
+ T
1
+
2
T
2
+ ,
N = N
1
+
2
N
2
+ ,
=
c
+
2

2
,
= t. (3.54)
Note that here, a decision has been taken from the outset to dene the parameter in terms of the
departure of the load from its critical value. The functions x
r
etc. are regarded as functions of
s and . To save introducing more notation, in the equations to follow a superposed dot will mean
dierentiation with respect to .
31
The governing equations now become

3
m( x
1
+ x
2
+ ) =
_
(
c
+ T
1
+
2
T
2
+ )(x

1
+
2
x

2
+
3
x

3
+ )
(N
1
+
2
N
2
+ )(1 + y

1
+
2
y

2
+ )
_

, (3.55)

3
m( y
1
+ y
2
+ ) =
_
(
c
+ T
1
+
2
T
2
+ )(1 + y

1
+
2
y

2
+ )
+(N
1
+
2
N
2
+ )(x

1
+
2
x

2
+ )
_

, (3.56)
(N
1
+
2
N
2
+
3
N
3
+ ) + B
1
_
(y

1
+
2
y

2
+ )(x

1
+
2
x

2
+ )
(1 + y

1
+
2
y

2
+ )(x

1
+
2
x

2
+ )
_

+ 3B
3

3
(x

1
)
2
(x

1
) = 0. (3.57)
The boundary conditions (3.10) and the remaining equation (3.7) are expanded similarly. The coe-
cients of successive powers of are now set to zero. First, the terms of order give
(
c
x

1
N
1
)

= 0,
(
c
y

1
+ T
1
)

= 0,
N
1
B
1
x

1
= 0. (3.58)
The constraint (3.7) implies that y

1
= 0. Therefore,
T
1
= 0, N
1
=
c
x

1
,
c
x

1
B
1
x

1
= 0. (3.59)
It follows (upon use of the boundary conditions) that
x

1
= A() sin(s/2l), (3.60)
the slowly-varying amplitude A() being so far undetermined. Next, the terms of order
2
give
(
c
x

2
N
2
)

= 0,
(
c
y

2
+ T
2
+ N
1
x

1
)

= 0,
N
2
+ B
1
(x

2
)

= 0. (3.61)
The constraint (3.7) gives
y

2
= x
2
1
/2. (3.62)
It follows that N
2
=
c
x

2
, T
2
=
2
+
c
x
2
1
/2 and then
c
x

2
B
1
x

2
= 0. Thus, x

2
has the same
form as x

1
, and can without loss be set to zero. Now considering terms of order
3
,
m x
1
= (
c
x

3
+ T
2
x

1
N
3
N
2
y

1
N
1
y

2
)

,
m y
1
= (
c
y

3
+ T
3
+ N
2
x

1
+ N
1
x

2
)

,
N
3
+ B
1
(y

2
x

1
x

3
y

2
x

1
)

3B
3
(x

1
)
2
x

1
= 0. (3.63)
32
The second of these equations implies that T
3
= 0, since y
1
= 0, x

2
= 0, so N
2
= 0 and the constraint
gives y

3
= 0. The rst and third equations simplify correspondingly:
m x
1
= (
c
x

2
x

1
N
3
)

,
N
3
B
1
x

3
3B
3
(x

1
)
2
x

1

1
2
B
1
(x
2
1
x

1
)

= 0. (3.64)
Therefore, eliminating N
3
,
m x
1
+
2
x

1
+
_
B
1
(x
2
1
x

1
)

/2 + 3B
3
(x

1
)
2
x

1
_

=
c
x

3
B
1
x

3
. (3.65)
The boundary conditions for x
3
are
x
3
(0, ) = x

3
(0, ) = x

3
(l, ) = 0,
c
x

3
(l, ) + B
1
x

3
(l, ) =
2
x

1
(l, )
1
2
B
1
[x

1
(l, )
2
x

1
(l, )]

.
(3.66)
The consistency condition for the existence of x
3
is obtained by multiplying by the eigenvector x
1
and
integrating from 0 to l. Substituting the expression (3.60) for x

1
and integrating by parts as appropriate,
the consistency condition becomes (with the change of variable u = s/2l)
_
2l

_
3
m

A
_
/2
0
(1 cos u)
2
du +

2
Al
2
+ A
3
_
B
1
4l
_
/2
0
sin u[2 sin ucos
2
u sin
3
u] du 3B
3
_

2l
_
3
_
/2
0
sin
2
ucos
2
udu
_
= 0.
(3.67)
Evaluating the integrals gives, nally,
_
3
8
1
_
m

A +

3
8l
2
A
_

2
A
2
_
B
1

2
32l
2
+
3B
3

4
64l
4
__
= 0. (3.68)
The equilibrium post-buckling relation (3.32) is recovered exactly by setting

A = 0.
In conclusion of this discussion, it is remarked that the asymptotic analysis given above provides
information on slow dynamics near a critical point but does not necessarily provide the complete
picture. Depending on the details of the system, there could be other dynamical solution branches
nearby, and nonlinear terms neglected in the low-order asymptotics could couple these to the motion
calculated and introduce signicant deviations, even where equation (3.68) predicts a periodic solution.
Of course, if the equation predicts an unbounded solution, its validity is in any case restricted to the
regime where A() is of order unity. Equation (3.68) should therefore be interpreted as showing just
how the system may rst respond to a small departure from the primary solution path, close to the
critical point.
33
4 Stability of continua
Considerations of the type described in Sections 2 and 3 apply also to bodies that have
to be modelled as two- or three-dimensional continua. It has been remarked already that
problems for continua are most usually approached by performing a discretization. There
is, nevertheless, some advantage in discussing continua directly, for basic understanding
and also because another phenomenon that of localisation of deformation is possible
in a continuum. When this is likely to occur, it is important that any discretization
should be designed so that it can track the deformation with sucient accuracy. It
is also important to understand when a problem is ill-posed. Further comment will
be made when localisation is discussed. First, however, the same basic sequence of
reasoning as has already been seen in the preceding sections will be followed through.
4.1 Notation
A brief self-contained summary of nonlinear continuum mechanics not, strictly, part
of this course is given in Section 6. This subsection simply records the main notation
that is employed.
Under a deformation, a point initially at position X, with Cartesian components
X

, moves to x, with Cartesian components x

i
. The deformation gradient matrix
A has components
A
i
=
x
i
X

.
Principal stretches
r
, r = 1, 2, 3, are dened so that the symmetric matrix A
T
A has
eigenvalues
2
r
.
A strain measure e
f
is dened, relative to a function f, to have the same principal
axes as A
T
A, and eigenvalues f(
r
). The function f is monotone increasing, f(1) = 0
and f

(1) = 1. Green strain corresponds to

f() =
1
2
(
2
1).
The corresponding strain measure, denoted by E, is
E =
1
2
(A
T
AI),
where I denotes the identity.
The stress T
f
, conjugate to the strain e
f
, is dened so that T
f

e
f

is the rate
of working of the stress per unit initial volume, during the deformation. For an elastic
34
medium, with energy density function per unit initial volume W, expressed as a function
of e
f
, it follows that
T
f

=
W
e
f

.
The stress T
(2)
that is conjugate to E is the second PiolaKirchho stress tensor. It
is also convenient to introduce S, with components S
i
, as the nominal stress tensor. Its
transpose is also called the rst PiolaKirchho stress tensor, or the Boussinesq tensor.
It has the property that the rate of working of the stress, per unit initial volume, is
S
i

A
i
, and it follows that
S
i
=
W
A
i
.
4.2 Equilibrium
A three-dimensional body, in equilibrium under some system of loading, adopts a con-
guration that satises the equations of equilibrium
S
i,
+
0
b
i
= 0, X B
0
. (4.1)
This is simply the time-independent version of the equations of motion (c.f. (6.12)).
The loading comprises the body-force b, together with boundary conditions. At each
point of B
0
, three conditions must be given. For instance, all three components x
i
of
x may be prescribed, or all three components N

S
i
of surface traction may be given
as functions of X, or some mixture, such as the normal component of traction and the
tangential components of x. In addition, it may be that a component of traction is
specied as a function not only of X, but also of x and A. It is possible, also, that
the body-force b could depend on the current position x of the material point if, for
example, it were applied via a non-uniform magnetic eld.
In any case, it will be assumed here that body-force, and the given combination
of surface displacements and surface tractions depend on a parameter , so that = 0
corresponds to no loading and the loading increases in some sense with . As in previous
sections, there may be more than one solution branch, but for the branch being followed,
rst the question of uniqueness of the increment of solution associated with an increment
of will be addressed. As previously, it is convenient to discuss rates of change in place of
increments, while disregarding inertia. These must conform to the equilibrium condition

S
i,
+
0

b
i
= 0, X B
0
, (4.2)
35
which is the rate equation corresponding to (4.1), together with rate forms of the bound-
ary conditions. In the most general conguration-dependent case (c.f. (6.73)), these
would take the form
N

S
i
= f
i
+ k
ij
x
j
+ C
ij
x
j,
(4.3)
wherever N

S
i
is given. Here, f
i
=
i
/t, k
ij
=
i
/x
j
and C
ij
=
i
/A
j
. In
addition,

b
i
could contain a term (b
i
/x
j
) x
j
.
It is also necessary to specify the constitutive relation of the body, in rate form.
Elastic body, simple boundary conditions
First, consider an elastic body, subjected to a combination of dead loading and given
boundary displacements, as considered in Section 2.3. Thus, in (4.3), k
ij
= C
ij
= 0.
The body force b is similarly assumed to be of dead-loading type. The elastic constitutive
relation (2.41), in rate form, gives

S
i
= c
ij
x
j,
, (4.4)
where
c
ij
=

2
W
A
i
A
j
(4.5)
(c.f. (6.47)). We wish now to examine the uniqueness of the solution of the equilibrium
equations (4.2) (in which

b is given as a function of X), together with the constitutive
relation (4.4) and boundary conditions.
The usual way to discuss uniqueness is to assume that there are two dierent so-
lutions. Then their dierence, denoted with the prex , satises the corresponding
system of homogeneous equations. Thus,

S
i,
= 0, X B
0
, (4.6)
where

S
i
= c
ij
x
j,
, (4.7)
together with homogeneous boundary conditions.
Now multiply equation (4.6) by x
i
, sum over i and integrate over B
0
. This gives
0 =
_
B
0
[ x
i
c
ij
x
j,
]
,
dX
_
B
0
x
i,
c
ij
x
j,
dX
=
_
B
0
x
i
N

c
ij
x
j,
dS
0

_
B
0
x
i,
c
ij
x
j,
dX
=
_
B
0
x
i,
c
ij
x
j,
dX, (4.8)
36
having employed the divergence theorem and made use of the fact that the boundary
conditions are homogeneous.
Uniqueness of x
i,
is guaranteed if
_
B
0
x
i,
c
ij
x
j,
dX > 0 (4.9)
for all x
i
not identically zero, that are consistent with the boundary conditions. That is,
x
i
must be zero wherever on the boundary x
i
is prescribed. Furthermore, x
i
is unique
provided it is prescribed at some point of the boundary. (Negative-deniteness of the
quadratic form would do equally well, but it will be seen later that positive-deniteness
is needed for stability).
A sucient condition for (4.9) to hold is that
a
i
c
ij
a
j
0, X B
0
, (4.10)
with equality only if a
i
= 0. It is also necessary for (4.9) in some cases. All-round
dead loading, generating uniform stress and deformation, is an example. It is, however,
possible that (4.9) may hold for all x
i,
allowed by other boundary conditions, even
if (4.10) does not hold. Suppose, conversely, that the minimum value of the quadratic
form (4.9) is zero, and that it is attained for some x
i
not identically zero. Then x
i
satises the equilibrium equations (4.6), and hence the solution is not unique. Such a
eld is called an eigenmode.
Relation to work-conjugate variables
Insensitivity of the energy function W to rigid rotations implies that W can only depend
on the deformation gradient A through some measure of strain, such as e
f
. Perhaps the
simplest of these is the Green strain, e
(2)
E =
1
2
(A
T
AI). It follows from the chain
rule for partial dierentiation that
S
i
= A
i
W
E

. (4.11)
Therefore,

S
i
= A
i

2
W
E

A
j

A
j
+
ij
W
E

A
i
. (4.12)
Hence,
c
ij
= A
i
L

A
j
+
ij
S
k
B
k
, (4.13)
37
where
L

=

2
W
E

(4.14)
and B
T
is the inverse of A, so that B
k
=

/x
k
, A
j
B
k
=
kl
. The formula (4.13)
is derived in Section 6 in the wider context of inelastic deformations.
It is plausible, for example, that the quadratic form e

may be positive-
denite with respect to symmetric e

. The form (4.13) can be expressed

a
i
c
ij
a
j
= b
ik
A
i
A
k
L

A
j
A
l
+ det(A)
ij
T
kl
b
jl
, (4.15)
where a
i
= b
ik
A
k
or, equivalently, b
ik
= a
i
B
k
. T is Cauchy stress. Evidently, the
quadratic form (4.10), or (4.15), cannot be positive-denite if any of the principal Cauchy
stresses are negative (this can be demonstrated by choosing b
ik
to be skew-symmetric).
Tensile stress, on the other hand, enhances the positive-deniteness. Therefore, in the
presence of tensile stress, bifurcation from a uniform state of deformation, maintained by
all-round dead loading, would need to be associated with the quadratic form generated
from L

, becoming indenite by a suciently large amount. This form is a property

of the energy function W. It is remarked, however, that there is nothing special about
the choice of the Green strain tensor E, except that it made the calculations simple. If
some other strain measure were employed, a formula of the same general form would
result, but L

would be dierent, and the dierence would be accounted for by an

additional term involving the current stress. The formula (6.35) provides this additional
term.
General boundary conditions
If the boundary conditions are of the general conguration-dependent form discussed
above, it is still possible to write down the system of linear partial dierential equations
and boundary conditions that govern any possible x
i
. Since the system is homogeneous,
an eigenvalue problem for the loading parameter is dened, and any solution is again
an eigenmode. No further discussion is given here.
Post-bifurcation behaviour
It is possible to study the initial post-bifurcation path, by following the pattern already
established in Sections 2 and 3. An elementary example will be presented in the context
of weakly-nonlinear dynamics.
38
4.3 Localisation
In contrast to the type of bifurcation which was envisaged above, we introduce now
the notion that material may become locally unstable, in the sense of admitting the
development of a discontinuity in the velocity eld. If such a discontinuity survives
for any nite time, a discontinuity in displacement ensues if the surface across which
the discontinuity exists remains stationary; otherwise, it moves through the material,
forming a shock. The formation of a stationary discontinuity is called localisation. In
practice, discontinuities are not realised exactly. There will be some ne structure. This,
however, is not captured by the simple constitutive description adopted so far. Although
much is already known, understanding is still far from complete for inelastic solids. Here,
we conne attention to the possible onset of localisation by identifying conditions under
which the rate equations of equilibrium (4.2) permit the development of discontinuities
in rates of stress and deformation gradient.
Suppose, therefore, that velocity is continuous but that stress-rate is discontinuous
across a surface S
0
in the reference conguration, which maps onto the surface S in the
current conguration. If velocity is continuous across S
0
, the tangential components of
its gradient must be continuous. Therefore, at most,
[ x
i,
] = a
i
N

(4.16)
for some vector a
8
. Here, N is the normal to S
0
and the square brackets denote the
jump across S
0
of the quantity enclosed. The equations of equilibrium (4.2) cannot hold
at S
0
but equilibrium still requires that
[N

S
i
] = 0. (4.17)
This condition can be interpreted as an enforcement of (4.2) in the weak sense. Now
substituting the constitutive relation (4.4) gives
c
ij
N

a
j
= 0. (4.18)
This is the condition for localisation: equation (4.18) should have non-trivial solution a
for some direction N.
Note that if the quadratic form given in (4.10) is positive-denite, then
c
ij
(a
i
N

)(a
j
N

) > 0 (4.19)
8
In fact, from (4.16), a is the jump in the normal derivative of x, a
i
= [N

x
i,
].
39
for all a and N, and localisation is impossible. The condition (4.19) is the condition
for strong ellipticity of the system of partial dierential equations (4.2) with (4.4). It
is weaker than the condition (4.10) and therefore it is possible that the onset of
bifurcation may occur before the onset of localisation. There is, however, a result called
Van Hoves theorem, that states that if displacements are prescribed over the whole of
the boundary, then the solution of the rate problem is unique if (4.19) is satised. For
this particular problem, therefore, bifurcation does not precede localisation.
Recall, again, that the constants c
ij
that appear in (4.19) depend explicitly as well
as implicitly on the current level of stress.
4.4 Linearized dynamics
Now suppose that the fundamental equilibrium solution, denoted with a superscript
zero, is perturbed dynamically, with the loading held xed. Denote the perturbed stress
and perturbed position
S
i
= S
0
i
+ s
i
, x
i
= x
0
i
+u
i
. (4.20)
Then, the linearized equations of motion give
s
i,
=
0
u
i
, (4.21)
together with homogeneous boundary conditions, and (still considering an elastic body)
the constitutive relations
s
i
= c
ij
u
j,
. (4.22)
Normal mode solutions have the time-dependence exp(it) and satisfy the system of
equations
(c
ij
u
j,
)
,
+
0

2
u
i
= 0. (4.23)
For the simple types of boundary conditions considered above, these equations are self-
adjoint and all eigenvalues
2
are real
9
. For suciently small loads, it is reasonable to
expect that the equilibrium conguration is stable, and therefore that the eigenvalues
2
are all positive. Instability rst becomes possible when the smallest eigenvalue is zero.
The usual argument, involving multiplication of the equation by u
i
and integrating over
B
0
, gives

2
_
B
0

0
u
i
u
i
dX =
_
B
0
u
i,
c
ij
u
j,
dX. (4.24)
9
The proof is very similar to that given in detail for the Euler column.
40
Thus, the smallest eigenvalue becomes zero at the value of for which
min
_
B
0
u
i,
c
ij
u
j,
dX = 0, (4.25)
the minimum being taken over elds u
i
that are compatible with any given displacements
on the boundary, and for which
_
B
0

0
u
i
u
i
dX = 1. Equation (4.25) denes u as an
eigenmode, as introduced in the context of static bifurcation.
It may be noted that the potential energy of the system is given by (6.51). Therefore,
the dierence in energy, between the congurations x and x
0
, is
E =
_
B
0
_
W(x
0
i,
+ u
i,
) W(x
0
i,
)
0
b
i
u
i
_
dX
_
B
0
N

S
0
i
u
i
dS
0
. (4.26)
Expanded to second order in u
i,
, this gives
E
_
B
0
1
2
u
i,
c
ij
u
j,
dX. (4.27)
The term linear in u
i
and u
i,
vanishes because the energy is stationary at x
0
. The
conguration x
0
is stable (all eigenvalues
2
are positive) if E, as given by (4.27), is
positive-denite. This is the Dirichlet condition for stability; it coincides exactly with
the condition (4.9).
Wave propagation
Consider now an innitesimal plane wave disturbance, propagating through uniform
material, uniformly pre-deformed to the level dened by S
0
and x
0
. The general plane
wave has the form
u
i
= a
i
f(t N

/c), (4.28)
where a is the amplitude of the wave and c is its speed. Substituting this form into the
equations of motion (4.21) gives
[c
ij
N

0
c
2

ij
]a
j
f

(t N X/c) = 0. (4.29)
This is satised, for any wave-form f, if
[c
ij
N

0
c
2

ij
]a
j
= 0. (4.30)
The matrix c
ij
N

is symmetric and therefore has real eigenvalues. The corre-

sponding wave speeds are real so long as the eigenvalues are positive. This is precisely
41
the condition for strong ellipticity of the static equations, which precludes localisation.
There is an obvious sense in which the material can be regarded as locally stable: there
should exist three real wave speeds
10
. Conversely, localisation of deformation occurs
when some disturbance cannot propagate, and therefore has no alternative but to build
up. c
ij
N

is also called the acoustic tensor, or the Christoel tensor. If it has

three positive eigenvalues, the equations of motion (4.21) are called totally hyperbolic.
It should be noted that, when the equations of motion are not totally hyperbolic, the
usual problem in which initial values of u and u are prescribed, becomes ill-posed.
Conversely, when the condition (4.30) for localisation is met, the corresponding prob-
lems for equilibrium fail to be elliptic, and problems with the usual kinds of boundary
conditions become ill-posed. The correct resolution must be to admit into the physical
model features so far neglected: viscosity, dependence of stress on higher-order gradients
of deformation, etc. This should be reected in any nite-element representation. If it is
not, the discretized problem will have a solution but it is unavoidably mesh-dependent.
Arbitrary choice of any particular mesh is equivalent to the injection of some additional
physics. If this is not identied explicitly, there is no reason to suppose that the nite-
element model will reect physical reality. It has to be remarked that this expedient is
nevertheless frequently adopted by practitioners!
4.5 Weakly-nonlinear dynamics
The dynamics of the system will now be investigated, at a load close to that which produces bifurcation.
Towards this end, let the primary solution be x
0
, and let the given tractions and body-forces be t
0
i
and
b
0
i
. These all depend on the parameter . The bifurcation level is given the superscript c in place of 0.
Now modify the applied loading, so that
b
i
= b
c
i
+ b
(1)
i
+
2
b
(2)
i
, (4.31)
and any given components of displacement or traction on the boundary have the forms
x
i
= x
c
i
+ x
(1)
i
+
2
x
(2)
i
, t
i
= t
c
i
+ t
(1)
i
+
2
t
(2)
i
, X B
0
. (4.32)
It is important that the quantities with superscript (1) should be directed tangentially to the original
loading path (that is, b
(1)
i
is proportional to db
0
i
/d, etc.), but the quantities with superscript (2) are
unrestricted. Let the perturbation u have the expansion
u = u
(1)
+
2
u
(2)
+ . (4.33)
10
Counting multiplicity: it does not matter if two wave speeds coincide, as in the case
of an unstressed isotropic material.
42
The perturbed stress satises
s
i
= c
ij
u
j,
+ d
ijk
u
j,
u
k,
+ , (4.34)
where
d
ijk
=
1
2

3
W
A
i
A
j
A
k
(A
0
). (4.35)
The equation of motion governing the perturbation is
s
i,
+
0
(b
(1)
i
+
2
b
(2)
i
) =
0
u
i
, (4.36)
exactly. Assume that the elds with the superscripts (1) or (2) depend on time only through the
slow variable
=
1/2
t.
Then the equations of motion give, upon substituting the series,
[c
ij
(u
(1)
j,
+
2
u
(2)
j,
]
,
+
2
[d
ijk
u
(1)
j
u
(1)
k,
]
,
+
0
(b
(1)
i
+
2
b
(2)
i
) =
0

2
u
(1)
i
+ O(
3
), (4.37)
where the prime denotes dierentiation with respect to . The boundary conditions are, to order
2
,
that
either u
(1)
i
+
2
u
(2)
i
= x
(1)
i
+
2
x
(2)
i
or N

[c
ij
(u
(1)
j,
+
2
u
(2)
j,
) +
2
d
ijk
u
(1)
j,
u
(1)
k,
] = t
(1)
i
+
2
t
(2)
i
, X B
0
. (4.38)
Equating terms of order gives
[c
ij
u
(1)
j,
]
,
+ b
(1)
i
= 0, (4.39)
together with the corresponding boundary conditions. Now by hypothesis, this system does have a
solution, which continues the primary branch. Call this u
(1)
i
. It is not unique, however. Therefore,
u
(1)
i
= u
(1)
i
+ A()v
(1)
i
(4.40)
for some function A(), where here the eigenmode has been designated v
(1)
. Now equating terms of
order
2
,
[c
ij
u
(2)
j,
+ d
ijk
( u
(1)
j,
+ Av
(1)
j,
)( u
(1)
k,
+ Av
(1)
k,
)]
,
+ b
(2)
i
=
0
A

v
(1)
i
. (4.41)
Here, it has been assumed that the terms with superscript (1) are independent of . Dierentiation
with respect to is indicated by a prime. The corresponding boundary conditions are
either u
(2)
i
= x
(2)
i
or N

[c
ij
u
(2)
j,
+ d
ijk
( u
(1)
j,
+ Av
(1)
j,
)( u
(1)
k,
+ Av
(1)
k,
)] = t
(2)
i
, X B
0
. (4.42)
The condition of consistency, for a solution u
(2)
to exist, can be found by multiplying equation (4.41)
by v
(1)
i
and integrating over B
0
. This gives, upon use of the divergence theorem,
_
B0
_
v
(1)
i
(b
(2)
i

0
A

v
(1)
i
) v
(1)
i,
d
ijk
( u
(1)
j,
+ Av
(1)
j,
)( u
(1)
k,
+ Av
(1)
k,
)
_
dX
+
_
B0
[v
(1)
i
t
(2)
i
v
(1)
i,
c
ij
N

x
(2)
j
]dS
0
= 0. (4.43)
43
The rst term in the surface integral only involves the given components t
(2)
i
because v
(1)
i
satises
homogeneous boundary conditions and therefore is zero wherever t
(2)
i
is not given. The symmetry
c
ij
= c
ji
of the elastic constants ensures that the second term similarly involves only the prescribed components
x
(2)
j
. This symmetry also ensured that [v
(1)
i,
c
ij
]
,
= 0, which was also exploited in the derivation.
Equation (4.43) is written more tidily as
A

_
B0

0
v
(1)
i
v
(1)
i
dX+ 2A
_
B0
v
(1)
i,
d
ijk
u
(1)
j,
v
(1)
k,
dX+ A
2
_
B0
v
(1)
i,
d
ijk
v
(1)
j,
v
(1)
k,
dX
=
_
B0
(v
(1)
i
b
(2)
i
v
(1)
i,
d
ijk
u
(1)
j,
u
(1)
k,
) dX+
_
B0
(v
(1)
i
t
(2)
i
N

c
ij
v
(1)
j,
x
(2)
i
) dS
0
. (4.44)
Equilibrium points are found by setting A

= 0 in (4.44). The resulting algebraic equation embodies

the static post-bifurcation response. It is slightly dierent in form from equation (3.20) which was
derived for a system with a nite number of degrees of freedom for two reasons. One is that the
fundamental solution is here not identied as zero. The other is that the extra loading, described by
b
(2)
etc., is somewhat more general, in that it may dene some deviation from the loading path of the
primary solution.
If the term containing A
2
should vanish (as it could if the system and its loading path had some
symmetry), then a dierent parameterisation would be needed: the preceding Section which dealt with
the Euler column provides a template.
The health warning given at the end of the corresponding discussion for the Euler column is
repeated here: the weakly nonlinear dynamics developed here are interesting and relevant but, de-
pending on the details of the problem, it is possible that there could be other solution branches nearby
whose analysis would require more sophisticated methods.
4.6 Inelastic media
This Section is concluded with a very brief discussion of inelastic media. The discus-
sion will be conned to equilibrium problems. The equation for continuing equilibrium
remains (4.2). Now, however, a dierent constitutive relation is adopted. The relation
(4.4) becomes nonlinear because the tangent moduli c
ij
are homogeneous functions of
degree zero in

A. The tangent moduli are, in fact, often taken to be piecewise constant
functions: for example, the plastic response of a single crystal is usually viewed as re-
sulting from slip on a denite set of slip systems, and the tangent moduli take constant
values which depend on which slip systems are activated. If, in addition, the moduli
have the symmetry
c
ij
= c
ji
,
44
the relation between stress-rate and deformation-rate can be given in the form

S
i
= U/

A
i
, (4.45)
where
U(

A) =
1
2

A
i
c
ij

A
j
. (4.46)
so long as certain other conditions are also met, to ensure that the potential (4.46) is
continuous and dierentiable.
The constitutive response can be modelled for many materials by the relation (4.45),
where U is any homogeneous function of degree 2 in

A. This model will be assumed in the
discussion to follow. It is justied at least for the usual model of single crystal response,
and, in fact, then follows (from micromechanical considerations) for any polycrystalline
material whose individual crystals respond in this way.
Incremental uniqueness and bifurcation
Suppose that the rate equations of equilibrium admit two dierent solutions, with su-
perscripts 1 and 2. Then

S
k
i,
+
0

b
i
= 0, (k = 1, 2), (4.47)
where

S
k
i
=
U

A
i
(

A
k
). (4.48)
The elds must also satisfy boundary conditions, which will be taken as a mixture of
prescribed displacements or dead-load tractions, as before. It follows that
_
B
0
(

A
1
i

A
2
i
)
_
U

A
i
(

A
1
)
U
A
i
(

A
2
)
_
dX = 0. (4.49)
This result follows from the use of the divergence theorem and the equations of equilib-
rium, coupled with the fact that both elds satisfy the same boundary conditions.
Suppose rst that U is a strictly convex function of

A: this is a generalisation of the
condition (4.10). It is a general property of a strictly convex function (of

A) that, for
any

A
1
and

A
2
,
(

A
1
i

A
2
i
)
_
U

A
i
(

A
1
)
U
A
i
(

A
2
)
_
0, (4.50)
with equality only if

A
1
=

A
2
. It follows that the solution of the rate problem is unique
if U is strictly convex.
45
Now introduce a comparison potential
U
0
(

A) =
1
2

A
i
c
0
ij

A
j
, (4.51)
where the c
0
ij
are constants, with the symmetry specied above. Suppose now that the
function U U
0
is convex. It follows that
(

A
1
i

A
2
i
)
_
(U U
0
)

A
i
(

A
1
)
(U U
0
)

A
i
(

A
2
)
_
0, (4.52)
and hence that
_
B
0
(

A
1
i

A
2
i
)
_
U

A
i
(

A
1
)
U
A
i
(

A
2
)
_
dX
_
B
0
(

A
1
i

A
2

)c
0
ij
(

A
1
j

A
2
j
) dX.
(4.53)
Uniqueness is assured if the quadratic form on the right side of (4.53) is positive-denite
for all

A (

A
1

A
2
) consistent with the boundary conditions. It is not necessary
that the function U
0
should be convex. The advantage of the use of a comparison
potential is that the form on the right side of (4.53) can be investigated by studying
its eigenmodes, using standard methods of linear analysis. Positive-deniteness of
the form is sucient for uniqueness of the rate problem for the actual material. The
prediction of bifurcation for the comparison medium does not, however, automatically
imply bifurcation for the actual medium. Whether or not it does depends (at least) on
the choice of comparison medium, and needs to be considered case by case. The device
of introducing a comparison potential was introduced by R. Hill, and the uniqueness
result that follows from (4.53) is referred to as Hills comparison theorem.
Localisation
Now consider again the possibility that rates of stress and deformation gradient may be
discontinuous across the surface which maps back to S
0
in the reference conguration.
The relations (4.16) and (4.17) now imply that
N

_
U

A
i
(

A
2
)
U

A
i
(

A
1
)
_
= 0, (4.54)
where

A
2
i
=

A
1
i
+ a
i
N

. (4.55)
46
Multiplying (4.53) by a
i
and summing over i then yields
(

A
2
i

A
1
i
)
_
U

A
i
(

A
2
)
U

A
i
(

A
1
)
_
= 0. (4.56)
It follows immediately that localisation cannot occur if the potential U is convex, for
then the convexity condition (4.50) together with (4.56) implies that a = 0. It may be
noted further that localisation is similarly excluded if (4.50) holds for all

A
1
and

A
2
that
satisfy (4.55). This states that

A
1
and

A
2
dier by a matrix of rank one (a matrix with
only one linearly independent row or column). The associated restriction on U is that
U is rank-one convex. Of course if a comparison medium is introduced, and localisation
cannot occur in that medium, then it cannot occur in the actual medium. However, there
is no advantage in introducing a comparison potential U
0
just for discussing localisation,
since the criterion for localisation is purely algebraic.
Media not satisfying normality
If the tangent moduli of a material do not possess the symmetry c
ij
= c
ji
, it is pos-
sible for two real eigenvalues of the acoustic tensor c
ij
N

to be real up to some level

of deformation at which they coalesce and then, for increased deformation, move into
the complex plane as complex conjugate pairs. It should be noted that the equations of
continuing equilibrium remain elliptic, because the determinant of the acoustic tensor
remains non-zero. Therefore, depending on the boundary conditions, it is possible that
the equilibrium path suers no bifurcation. However, if any dynamic disturbance is con-
sidered, the corresponding equations of motion must be employed, and these cease to
be totally hyperbolic. The natural initial value problems become ill-posed. By analogy
with utter, this particular type of material condition is called the utter ill-posedness.
Exactly what to do about it is so far undecided. Analysis to date has been conned
almost exclusively to identifying when the condition might occur. Clearly features of
material response that are usually unimportant and so are neglected in the models dis-
cussed above need to be recognised and allowed for. These will include rate-dependence
(for which systematic descriptions exist) and non-local response, upon which there is as
yet no universal agreement.
47
4.7 Illustrative example
The formulae derived above will now be developed more explicitly, for the constitutive
model (6.67). Thus, during plastic loading,
e = /

T+PQ
T

T/h. (4.57)
Here, T and e are conjugate measures of stress and strain; a particular choice will be
made below.
The rst task is to invert the relation (4.57). Elementary algebra yields

T = L e
LPQ
T
L e
Q
T
LP+ h
. (4.58)
Thus, during loading,
L = L
LPQ
T
L
Q
T
LP+ h
. (4.59)
Now to be denite, take the conjugate stress and strain pair to be the second Piola-
Kirchho stress T
(2)
and the Green strain e
(2)
E. Then, the corresponding moduli
relating rates of nominal stress and deformation gradient are given by (6.71). If, in
addition, the current state is chosen as reference conguration, so that A = I at the
present instant, it follows that
c
kilj
= L
kilj
+
ij
T
kl

L
kipq
P
qp
Q
rs
L
srlj
P
pq
L
qpsr
Q
rs
+ h
. (4.60)
Here, since the reference conguration is the current one, Greek suxes have been
dispensed with. Also, since at this instant all measures of stress coincide, T can be
viewed equally as second Piola-Kirchho stress, or nominal stress, or Cauchy stress. It
is important to remember, though, that the moduli c
kilj
relate the nominal stress-rate,
given by (6.38) with (6.36), to the rate of deformation.
Now specialise further, by taking
P = Q = T

/|T

|, (4.61)
where T

is the deviatoric stress

T

ij
= T
ij

1
3

ij
T
kk
(4.62)
48
and |T

| = (T

ij
T

ji
)
1/2
. It is consistent with the implied isotropy to take
L
kilj
=
ki

lj
+ (
kl

ij
+
kj

li
). (4.63)
The relation (4.61) means that plastic strain-rate and deviatoric stress are parallel to
one another; volume change and hydrostatic stress are related elastically. The constitu-
tive relation so developed is a generalisation to nite deformations of the conventional
J
2
, or von Mises theory, which states that plastic deformation occurs when the mag-
nitude of the shear stress, |T

|, reaches a critical value. The coecient h denes the

amount of hardening: during simple shear, it relates the rate of increase of the shear
stress to the rate of increase of the plastic part of the strain-rate
11
. It should be em-
phasised that the constitutive relation that has been developed depends on the choice
of conjugate stress and strain measures. Any other choice would provide a dierent
generalisation of the small-deformation theory. It should be noted, however, that this
or any other choice of conjugate variables yields a tangent modulus tensor that displays
the symmetry c
kilj
= c
ljki
exactly
12
.
The corresponding potential U now takes the form
U(

A) =
1
2

A
ik
_
L
kilj
+
ij
T
kl

L
kipq
P
qp
P
rs
L
srlj
P
pq
L
qpsr
P
rs
+ h
H(P
ab
L
badc

A
cd
)
_

A
jl
, (4.64)
where H denotes the Heaviside step function. Evidently, the potential U is rank-one
convex if (4.19) is satised, with c
kijl
taking their inelastic values. Substituting into
this condition, with a orthogonal to N (because it is already known that this model of
plasticity only allows for inelastic shear deformations), gives
a
i
N
k
c
kilj
a
j
N
l
=
4
2
(a
i
T
ik
N
k
)(N
l
T
lj
a
j
)
(2 +h)T

rs
T

sr
+ N
k
T
kl
N
l
> 0. (4.65)
Now suppose that T has the form
_
_
_
p T
12
0
T
12
p 0
0 0 p
_
_
_.
11
This description is accurate when the strains are innitesimal. It is given only for the
purpose of motivation.
12
It can be noted that, in contrast, the common assumption that the co-rotational or
Jaumann derivative of Kirchho stress is related to strain-rate through a set of moduli
of the form (4.59) introduces some asymmetry; it is also not compatible exactly with the
existence of an energy function, in the case of no plastic deformation.
49
Take a
i
=
i1
and N
k
=
k2
. The condition for avoidance of localisation then reduces to
h > 0. However, for more complicated patterns of stress, it is possible that localisation
might not occur even for (some) negative values of h. It is equally possible that the
onset of localisation could occur at some positive value of h.
Finally, an example of a simple model of dilatant plasticity is presented:
P
ki
= T

ki
/|T

| +
ki
, Q
ki
= T

ki
/|T

| +
ki
. (4.66)
In this case, plastic yielding depends on the hydrostatic part of the stress as well as
on the shear, and the inelastic deformation likewise has a dilatant component. Soils
display this type of behaviour. There is, however, no universal agreement on the relative
values of the parameters and , except that they are most unlikely to be equal, so
that utter ill-posedness is a potential problem.
50
Fig. 5.1. (a) Sketch of the cross-section of a buckled pipe. (b) The one degree-of-freedom
model, with four hinges.
5 Propagating Instabilities
Considering small small departures from some equilibrium path as in previous sections
serves to identify bifurcation points, critical loads, and static or dynamic response close
to such points. Achievement of instability is most likely to lead to a large departure
from the conguration from which the instability commenced. Furthermore, there are
situations in which a local large perturbation can act as a trigger for an instability
that aects the whole structure, even though the loading that is applied is insucient
to generate instability in the absence of the local perturbation. An example is the
propagation of a buckle along an undersea pipeline. The pipeline is modelled as a long
circular cylinder, subjected to external pressure loading. The essence of the phenomenon
is captured by a simple model, with just one degree of freedom. After buckling, the
cross-section of the pipe usually has the form depicted in Fig. 5.1(a). This motivates
considering the simple model shown in Fig. 5.1(b). Four quarter-circle segments of
radius a are joined by nonlinear hinges, each of which resists bending through an angle
by exerting a resisting couple of moment
M = f(). (5.1)
51
The hinges may be elastoplastic but, since only monotonically increasing loads will be
considered, the moment M can be considered to be a single-valued function of . The
conguration of the cross-section is dened by the angle shown in Fig. 5.1(b). If
the pipe is initially circular, then = 0 denes its initial conguration. An initial
imperfection is modelled by taking =
0
,= 0 prior to deformation.
5.1 The critical pressure
The response modelled by (5.1) can be treated as though it is elastic, even if it is not;
the dierence would emerge only if unloading were considered. Therefore, it is possible
to dene a potential energy per unit length of pipe
U() =
_

0
f(q)dq. (5.2)
The external pressure has xed magnitude p. The potential associated with this (per
unit length of pipe) is p times the area of cross-section of the pipe. This area, A() say,
is 4 times the area enclosed by a curved segment, plus the area of the rhombus inside,
whose sides are of length

2a. Thus,
A() = (a
2
2a
2
) + 2a
2
cos(2), (5.3)
since the acute angle of the rhombus is /2 2. The total energy (per unit length) of
the system is therefore
E() = pA() + 2U(2(
0
)) + 2U(2(
0
)), (5.4)
since two hinges undergo the deection = 2(
0
) while the other two undergo the
deection = 2(
0
). Equilibrium requires that E() should be stationary. Thus,
by dierentiating (5.4) with respect to ,
4f(2(
0
)) 4f(2(
0
)) 4pa
2
sin(2) = 0. (5.5)
For the perfect structure,
0
= 0. Therefore, since f(0) = 0, the symmetric congu-
ration = 0 is in equilibrium. It is stable so long as d
2
E()/d
2
> 0 when = 0. That
is,
p < p
c
, (5.6)
where
p
c
= 2f

(0)/a
2
. (5.7)
52
Fig. 5.2. Qualitative plots of pressure p versus deformation as measured by
0
, for the
perfect structure (
0
= 0) and an imperfect structure (
0
> 0). The pressure can take any
value when = /4, corresponding to contact of opposing faces.
Now to be more explicit, suppose that the hinges are elastic-perfectly plastic, so that
f() =
_
k, [[
y
k
y
, [[
y
.
(5.8)
Then,
U() =
_
1
2
k
2
, 0
y
1
2
k
2
y
+ k
y
(
y
),
y
.
(5.9)
Also, U() = U(). The equilibrium condition (5.5) now gives
p
p
c
=
_
2(
0
)/ sin(2), 0 (
0
)
y
/2

y
/ sin(2), (
0
)
y
/2.
(5.10)
Only the case
0
0 is considered. The form of the relation (5.10) is sketched in Fig.
5.2, for the perfect structure (
0
= 0) and an imperfect structure (
0
> 0). It suggests
that collapse will occur once the maximum pressure, dened by putting =
0
+
y
/2 in
(5.10), is attained. The maximum pressure is reduced by the presence of an imperfection.
There is, in fact, an upper limit to , given by = /4. This corresponds to opposing
53
Fig. 5.3. Schematic picture of a pipe buckled over part of its length. The pipe buckles
over an additional length d in going from conguration (a) to conguration (b). The
transition region between buckled and unbuckled parts retains its form but is translated
though distance d.
faces of the cylinder coming into contact, after which p may become arbitrarily large
without inducing further deformation in this model. This is what we mean by collapse.
Equivalently, the tube is said to have buckled.
5.2 Propagation of a buckle
If, for some reason, the pipe buckles (essentially as described above) over some limited
portion of its whole length, the deformation must be three-dimensional, since there must
be one or two regions of transition between the buckled and unbuckled congurations.
There is the possibility that the buckle may spread along the pipe, even if the pressure
is less than the critical pressure p
c
that is required to initiate a buckle in a perfect
section. Fortunately, the pressure required to propagate a buckle can be estimated,
at least approximately, without recourse to three-dimensional analysis. If the buckled
section is already long, the congurations shown in Fig. 5.3 may be considered. Figure
5.3(a) shows the buckle in a certain position and Fig. 5.3(b) depicts the conguration
after the buckle has propagated through a distance d. The transition region retains its
54
form but has translated through a distance d. The work done by the pressure between
congurations (a) and (b) is just d times the dierence in the energy per unit length
for the undeformed and the buckled sections. Thus, from (5.3), with = 0 for the
undeformed section, the work done is
W = 2a
2
pd, (5.11)
since = /4 in the buckled conguration.
It is impossible for the buckle to have propagated, unless the work W is at least as
great as the work required to extend the buckle through the distance d. If the material
of the pipe is (nonlinearly) elastic (so that the work done within it is independent of
the loading path), this latter work is exactly d times the work done (per unit length)
against the hinges, in taking the tube from its initial state to its completely buckled
state, since the region over which the deformation is three-dimensional has translated
but the pattern of deformation within it is unchanged. Thus, for an elastic pipe, the
work done within the tube is exactly
2[U(/2) +U(/2)]d. (5.12)
It is plausible, though no proof is oered, that, if the pipe is made of inelastic material,
the work done against the stresses must be at least this value. Thus, a plausible lower
bound for the pressure required to propagate the buckle is given by
p
p
= [U(/2) +U(/2)]/a
2
. (5.13)
When U() has the form given in equation (5.9),
p
p
= k
y
[
y
]/a
2
. (5.14)
If
0
denotes yield stress of the pipe material in tension or compression, the most ele-
mentary estimate for the yield moment M
y
= k
y
of a hinge is
M
y
= 2
_
t/2
0

0
z dz =
0
t
2
/4, (5.15)
where t denotes the thickness of the tube. Specializing also to rigid-perfectly plastic
response, so that
y
0, the lower-bound critical pressure for propagation p
p
becomes
p
p
=
0
(t/2a)
2
. (5.16)
This is very much smaller than the pressure p
c
required to initiate a buckle in a perfect
section of pipe, which formula (5.7) gives as twice the elastic bending stiness, divided
by a
2
and so is innite in the rigid-perfectly plastic limit!
55
5.3 Other propagating instabilities
There are other problems in which an instability, once started, can propagate along a
structure under a load much lower than that required to initiate it. One example, in
common experience, is provided by blowing up a long thin balloon: signicant pressure
is required to achieve a large expansion over some portion of the balloon, but thereafter
it becomes relatively easy to blow up, at moderate pressure, by the mechanism of length-
ening of the inated section. The equations are not presented, but again the key is to
recognise that the pressure versus radius relation has a form similar to that given for the
imperfect pipe in Fig. 5.2, with radius replacing the parameter
0
. The vertical
part of the response shown in Fig. 5.2 is not vertical in the case of the balloon, but it
does rise steeply as radius increases beyond a certain value.
Materials that can undergo phase transformation can display a related phenomenon.
There is an impressive body of theory for phase-transforming materials that can be
modelled as (nonlinearly) elastic
13
. Problems involving more than one spatial dimension
are very far from trivial, and completely beyond the scope of this course. It is possible,
however, to give a brief outline for the one-dimensional case, realised by the tension or
compression of a bar. The stress-strain response is dened by the relation
= W/e, (5.17)
where denotes nominal tensile stress (load divided by cross-sectional area prior to
deformation) and e is the tensile strain, e = u/x, where u is displacement and x
is the Lagrangian coordinate along the bar. The energy function W has two or more
minima. The case of exactly two minima will be discussed, as depicted in Fig. 5.4(a).
At a minimum, W/e = 0, and so there is zero stress. Dierent minima correspond
to dierent phases of the material. If e is measured relative to the unstressed state in
phase 0, then e = 0 denes the minimum in the energy associated with phase 0, and
some other strain, e = e
1
, denes the corresponding minimum for phase 1. The stress
is zero at e = 0 and e = e
1
, and each of these congurations is stable. There must
be, however, a maximum for W in between, at e = e

say, at which the stress is also

zero but the associated conguration is unstable. The stress-strain relation thus has the
general character depicted in Fig. 5.4(b). The stress-strain curve passes through the
origin, rises to a maximum value
max
at some strain between 0 and e

, passes through
zero at e = e

, then falls to a minimum value

min
at a strain between e

and e
1
, and
13
There are, of course, also materials that undergo plastic deformation in addition to
phase transformation, for which theory is much less advanced.
56
Fig. 5.4. (a) Energy function of a material that can exist in two phases, 0 and 1. (b) The
corresponding stress-strain relation.
thereafter rises. Suppose that the bar is subjected to end displacements that produce
nominal stress , and that the bar is initially in phase 0, with zero strain. The relative
displacement of the ends of the bar is e times its length, where e is the mean tensile
strain. As e increases from zero, the strain is uniform along the bar and has value e,
up to the level at which =
max
. Thereafter, the stress lies between
max
and
min
,
but the material has a choice: it can either adopt the uniform strain e along its entire
length, corresponding to the point C in Fig. 5.4(b), or it can adopt the strain e
C
over a
fraction f of its length, and the strain e
A
over the remaining fraction, (1 f), so that
(1 f)e
A
+ fe
C
= e. (5.18)
If it adopts a conguration of this latter type, then at least each part of the bar corre-
sponds to a state of stress and strain that is stable. The material at strain e
A
is in phase
0 and the material at strain e
C
is in phase 1. There are innitely many congurations
of this type. Even the level of the line ABC is not xed, except that must lie between

max
and
min
. Exactly what happens will depend upon how deformation past the point
corresponding to
max
may be triggered.
Suppose now that, for some reason, the bar (which occupies the region 0 < x < L
before deformation) adopts the strain e
C
for 0 < x < fL and the strain e
A
for fL <
57
x < L: the phase transformation has spread from the end x = 0. If the mean strain
is increased by an amount e, the transformation will spread further along the bar, a
distance x = fL, say, and e
A
and e
C
will undergo changes e
A
, e
C
. These changes
are related so that they are consistent with (5.18):
e = f(e
C
e
A
) + (1 f)e
A
+ fe
C
. (5.19)
Also, since the stress has to be constant along the bar, for equilibrium, = W

(e
A
) =
W

(e
C
).
The work done on the bar during this process must at least suce to provide the
additional strain energy in the bar. The former is Le, while the latter is
U = Lf[W(e
C
) W(e
A
)] + (1 f)W

(e
A
)e
A
+ fW

(e
C
)e
C

= Lf[W(e
C
) W(e
A
)] + [e f(e
C
e
A
)] , (5.20)
having used (5.19). Thus, propagation is not possible unless
= W

(e
A
) = W

(e
C
)
W(e
C
) W(e
A
)
e
C
e
A
=:
M
. (5.21)
The level at which equality is achieved in (5.21) denes the Maxwell stress
M
. It
corresponds to equality of the areas shown hatched in Fig. 5.4(b). The inequality states
that the stress required to propagate the phase transformation must equal or exceed
the Maxwell stress. This is exactly analogous to the lower bound p
p
for the pressure
required to propagate a buckle along a pipe. Estimation of the precise value of the stress
at which a phase transformation front may propagate would require a detailed model of
the kinetics of the phase transformation process. This is analogous to the need to model
the three-dimensional deformation in the transition region in the buckle propagation
problem but no microscopic model of the phase transformation process has yet found
wide acceptance.
58
Fig. 6.1. Sketch of a body in undeformed and deformed congurations.
6 Review of Nonlinear Solid Mechanics
This section presents a quick overview of continuum mechanics, as applied to solids. It
is not strictly part of the course but serves to make these notes self-contained.
6.1 Deformation and Stress
An outline of the basic notions of deformation and stress, relevant to all materials,
is presented here. A wide variety of notations is in current use; here, the arbitrary
choice has been made to follow that used in the book by R.W. Ogden (Nonlinear Elastic
Deformations, Ellis Horwood, Chichester 1984).
Deformation
The deformation of a body is depicted in Fig. 6.1. It occupies a domain B
0
in its
reference conguration and B
1
currently; if the deformation varies with time t, then
B
1
depends on t. A generic point of the body has position vector X B
0
initially,
and x B
1
at time t, relative to origins O and o respectively. Relative to Cartesian
bases E

for the initial conguration and e

i
currently, the vectors X and x have
coordinate representations
X = X

and x = x
i
e
i
, (6.1)
with implied summation over the values 1,2,3 for the repeated suxes.
59
The deformation is dened by an invertible map from B
0
to B
1
. In terms of X and
x,
x = (X, t) (6.2a)
or, in components,
x
i
=
i
(X, t). (6.2b)
The deformation gradient A is then dened as
A = A
i
e
i
E

, (6.3)
where
A
i
=
x
i
X

. (6.4)
Then, an innitesimal line segment dX deforms into the segment dx, where
dx = AdX, dx
i
= A
i
dX

. (6.5)
Strain tensors relate lengths and angles before and after deformation. If innitesimal
line segments dX and dY transform respectively into dx and dy, then
dx.dy = dx
T
dy = dX
T
A
T
AdY. (6.6)
All information on length and angle changes is thus contained in A
T
A. Perhaps the
simplest strain measure the Green strain is then
E =
1
2
_
A
T
AI
_
, or E

=
1
2
(A
i
A
i

) . (6.7)
A general class of strain measures is obtained by rst dening the eigenvalues and
normalized eigenvectors of A
T
A as
2
i
and u
(i)
, so that
A
T
A =
3

i=1

2
i
u
(i)
u
(i)
. (6.8)
Then, if f is any monotone increasing function for which f(1) = 0 and f

(1) = 1, a
strain tensor e is dened as
e =
3

i=1
f(
i
)u
(i)
u
(i)
. (6.9)
60
The strain tensor (6.7) ts this pattern, with f() =
1
2
(
2
1).
The polar decomposition theorem
A result needed later, related to (6.8), is the polar decomposition theorem. Dene
U =
3

i=1

i
u
(i)
u
(i)
and then set
R = AU
1
= A
3

i=1

i
1
u
(i)
u
(i)
.
It follows that R represents a rotation (so that R
T
R = I), while U is symmetric. The
representation
A = RU
is the required result.
Stress and equations of motion
Suppose the body is acted upon by surface and body forces, which may vary with
time t. These forces can be represented either as functions of x and t, relative to the
current conguration, or as functions of X and t, relative to the initial conguration, to
which the current conguration is related by (6.2). Since the mapping is usually not
known in advance of solving the problem, we choose to employ the latter representation.
Thus, with the mass density of the body given as
0
per unit volume in the reference
conguration, the body force b per unit mass can equivalently be expressed as
0
b
per unit initial volume. Since the force is actually applied to the body in its current
conguration, it is usual to express b, in components, as
b = b
i
e
i
.
The forces applied to the surface are dened similarly, per unit of surface area in the
reference conguration. Thus, if an element of surface dS
0
is mapped by (6.2) into a
surface element ds, the force applied to ds is
df = tdS
0
,
where t is the force per unit reference area and dS
0
denotes the magnitude of dS
0
.
61
Fig. 6.2. Element of volume employed for deriving (6.11).
Balance of linear momentum then requires that
d
dt
_
B
0

0
v dX =
_
B
0

0
bdX+
_
B
0
t dS
0
, (6.10)
where v = dx/dt is the particle velocity and B
0
denotes the surface of B
0
. A similar
relation must apply to any part of B
0
. In particular, by taking a small volume element
of the type shown in Fig. 6.2, it follows that t has the representation
t = S
T
N, or t
i
= S
i
N

, (6.11)
where N denotes the unit normal to the surface element dS
0
14
. If the representation
(6.11) is now substituted into (6.10) (applied to any part of B
0
), the divergence theorem
then yields the equation of motion
S
i,
+
0
b
i
=
0
dv
i
/dt, X B
0
. (6.12)
It may be noted that (6.12) is exactly like the more usual equation involving Cauchy
stress, but
0
and B
0
are known; the equation is thus linear, even though it is exact.
Balance of moment of momentum requires that
d
dt
_
B
0

ijk
x
j
(
0
v
k
)dX =
_
B
0

ijk
x
j
(
0
b
k
)dX+
_
B
0

ijk
x
j
(S
k
N

)dS
0
.
14
The derivation of (6.11) can, if preferred, be by-passed by treating (6.11) as a
postulate.
62
Transformation of the surface integral to one over B
0
by the divergence theorem, followed
by use of (6.12) and the recognition that B
0
may be chosen arbitrarily gives the result

ijk
A
j
S
k
= 0. Equivalently,
A
i
S
j
= A
j
S
i
, or AS = (AS)
T
. (6.13)
In view of its reference to initial area, S is called the tensor of nominal stress; its
transpose is called the rst Piola-Kirchho stress tensor. It can be shown that Cauchy
stress T is related to S by
T = AS/det(A).
Work-conjugate stresses and strains
The rate of working of the forces applied to the body is obtained by multiplying forces
by velocities; thus, the total rate of working is
_
B
0

0
b
i
v
i
dX+
_
B
0
N

S
i
v
i
dS
0
= w, say.
Application of the divergence theorem and use of the equation of motion (6.12) trans-
forms this to
w =
_
B
0
__
d
dt
_
1
2

0
v
i
v
i
+ S
i
v
i,
_
dX. (6.14)
The last term in the integrand represents the rate of working, per unit reference volume,
of the stresses. It can also be written
S
i
v
i,
= S
i

A
i
. (6.15)
(The superposed dot means d/dt.)
Now consider the strain tensor obtained by taking f() = 1, so that e = UI.
Employing in (6.15) the polar decomposition of A gives
S
i

A
i
= S
i

R
i
U

+ S
i
R
i

U

= S
i
R
i

U

.
The term involving

R is zero. This can be seen by noting that
S
i

R
i
U

= S
i

R
i
R
j
R
j
U

= A
j
S
i

R
i
R
j
,
63
and invoking the symmetry (6.13) together with the antisymmetry

RR
T
= R

R
T
.
Hence, exploiting the symmetry of U,
S
i

A
i
= T
(1)

, (6.16)
where
T
(1)

=
1
2
[S
i
R
i
+ S
i
R
i
] . (6.17)
The tensor T
(1)
is the stress which is conjugate to the strain e.
The same idea applies to other measures of strain. Another simple example is pro-
vided by the Green strain (6.7). The associated conjugate stress is T
(2)
, where
T
(2)

= S
i
B
i
, (6.18)
with the notation
B
T
= A
1
. (6.19)
The stress tensor T
(2)
is the second Piola-Kirchho stress tensor; it is symmetric, since
(6.13) gives
SB = B
T
ASB = B
T
S
T
A
T
B = B
T
S
T
.
Convected coordinates
Suppose a coordinate net is scribed into the initial conguration. It is rectilinear initially
but, after deformation, an innitesimal segment E

dX

is transformed into a segment

e
i
A
i
dX

= e

dX

, say. The vectors e

form a basis, but this is not orthonormal;

it is associated instead with curvilinear coordinates X

in the current conguration.

Now we can associate with T
(2)
, for example, a tensor

T
(2)
= T
(2)

=

T
(2)
ij
e
i
e
j
, (6.20)
where

T
(2)
ij
= A
i
A
j
T
(2)

. (6.21)
The components T
(2)

can be viewed as the contravariant components of the tensor

T
(2)
,
relative to the basis e

. It can be checked by calculation that

T
(2)
ij
= det(A)T
ij
(T
denoting Cauchy stress). The tensor

T
(2)
is called the Kirchho stress.
Similar constructions could be based upon the components of other stress tensors;
there is no particular advantage to this, but the discussion presented perhaps explains
the source of the wide variety of possible descriptions of stress.
It should be noted, nally, that all of the measures of stress that have been discussed
coincide, when the reference conguration is chosen as the current conguration at the
instant of interest.
64
Stress rates
The subject of stress rates will only be touched upon during the lectures. The account
to follow, though itself only a sketch, is included so that the exposition in these notes is
in a sense complete.
First we establish some notation. The strain measure (6.9), constructed with the
function f, will be denoted e
f
:
e
f
=
3

r=1
f(
r
)u
(r)
u
(r)
or in sux notation, e
f

=
3

r=1
f(
r
)u
(r)

u
(r)

. (6.22)
The conjugate stress is now denoted T
f
, with components T
f

. The corresponding stress

rate is nothing other than

T
f
, with components

T
f

. As was discussed above for the

second Piola-Kirchho stress T
(2)
, another tensor,

T
f
, can be formed by taking T
f

to
be its contravariant components on the basis e

, so that its Cartesian components are

given by a formula exactly like (6.21). Then, the components

T
f

can be employed to
form a certain rate for the tensor

T
f
, say

T
f
t
=

T
f

, or


T
f
ij
t
= A
i
A
j

T
f

. (6.23)
It is of interest to express the stress-rate

T
f
/t in terms of the components

T
f
ij
and
their derivatives, particularly in the case when the current conguration is taken as the
reference conguration, so that A = I. Relative to this conguration, it has already
been remarked (and will be proved below) that all stress measures become identical. So
also, do all strain-rates, but the same is not true of the stress-rates.
To begin, the strain-rate corresponding to e
f
has components
e
f

=
3

r=1
[f

(
r
)

r
u
(r)

u
(r)

+ f(
r
)(d/dt)(u
(r)

u
(r)

)]. (6.24)
(Here and elsewhere, d/dt has the same meaning as a superposed dot, either representing the time
derivative at a xed material point, so that X is kept xed.) The stress components T
f

are related to
those of the nominal stress through
S
i

A
i
= T
f

e
f

= T
f

e
f

A
i

A
i
. (6.25)
Thus,
S
i
= T
f

e
f

A
i
. (6.26)
65
An expression for e
f

/A
i
is therefore required. A calculation summarised in Appendix 2A gives
e
f

A
i
=
3

r=1
_
f

(
r
)

r
u
(r)

u
(r)

u
(r)

A
i
u
(r)

s=r
f(
r
)

2
r

2
s
(u
(r)

u
(s)

+ u
(s)

u
(r)

)(u
(r)

A
i
u
(s)

+ u
(s)

A
i
u
(r)

)
_
. (6.27)
This formula is somewhat inconvenient for investigating the case when A = I, because it requires a
limiting operation. Appendix 2A derives this limit directly. It involves the Eulerian strain-rate
ij
,
which is the symmetric part of the Eulerian deformation-rate
ij
= x
i
/x
j
=

A
i
B
j
, which reduces
to

A
i

j
when A is the identity, A
i
=
i
. The time derivative of
r
, and the eigenvector u
(r)
, satisfy

ij

j
u
(r)

i
u
(r)

. (6.28)
Thus,

ij
=
3

r=1

j
u
(r)

u
(r)

. (6.29)
This demonstrates explicitly, by comparison with (6.24) when
r
= 1, that all stress-rates coincide with
the Eulerian strain-rate when the current conguration is chosen as reference. Expressed dierently,
when A = I,
e
f

=
e
f

A
i

A
i
=
j

jk
=
1
2
(
j

A
j
+
j

A
j
). (6.30)
It follows that
e
f

A
i
=
1
2
(
i

+
i

), (6.31)
and hence that
S
i
=
i
T
f

. (6.32)
Thus, relative to the current conguration, all measures of stress coincide, and may be identied with
a single stress T

which in fact is also coincident with the Cauchy stress.

Relations between dierent stress-rates are now investigated by dierentiating (6.26) to give

S
i
=

T
f

e
f

A
i
+ T
f

2
e
f

A
i
A
j

A
j
. (6.33)
It is shown in Appendix 2A that, when A = I,

2
e
f

A
i
A
j
=
1
2
_
1
4
(f

(1) 1)(

ij

ij

j
+
j

i
+
j

+
j

+
i

+
j

i
)
+

r
u
(r)

u
(r)

u
(r)

u
(r)

ij
(

)
+
1
2

s=r
(u
(r)

u
(s)

+ u
(s)

u
(r)

)
ij
(

)
_
. (6.34)
66
The desired relation between

S
i
and

T
f

now follows by substituting (6.31) and (6.34) into (6.33). This

is not written out explicitly, but it is noted that the rate associated with T
f
depends on the function
f only through the term multiplying the coecient (f

(1) 1). It is now not dicult to conclude, for

T
f
and

T
g
(the latter being dened with the function g replacing f), that

T
f
ij
t
=

T
g
ij
t

1
2
[f

(1) g

(1)](
ik

T
kj
+

T
ik

kj
), (6.35)
where

T denotes either of

T
f
or

T
g
, since these coincide when the current conguration is taken as
reference.
All that remains now is to evaluate one particular stress-rate. This is easy to do for
the tensor

T
(2)
. Dierentiating (6.21) and then setting A = I gives
d

T
(2)
ij
dt
=
i

j

T
(2)

+ (

A
i

j
+
i

A
j
)T
(2)

, (6.36)
whence it follows that

T
(2)
ij
t
=
d

T
ij
dt

ik

T
kj

jk

T
ki
. (6.37)
It is also of interest, for later use, to derive an expression for the nominal stress-rate.
The simplest course is to dierentiate the relation (6.18). This gives (when A = I)

T
(2)

=

S
i

i
S
i

i

A
k

k
.
This delivers the nominal stress-rate

j

S
i
=


T
(2)
ji
t
+
ik

T
kj
=
d

T
ji
dt

jk

T
ki
. (6.38)
It is remarked nally that the Cauchy stress tensor, previously introduced as T
with components T
ij
, is not conjugate to any strain measure. However, the relation

T
(2)
= det(A)T between Kirchho stress and Cauchy stress implies, when A = I, that
d

T
ij
dt
=
dT
ij
dt
+
kk
T
ij
. (6.39)
6.2 Elastic constitutive equations
This section records the constitutive relations of nite-deformation elasticity and then
their specialization, rst to incremental deformation and then, further, to classical
linear elasticity (which may be viewed as an increment of deformation from an unstressed
reference conguration).
67
The general elastic constitutive relation
An elastic body is taken as one which stores any energy that is put into it. If thermal
eects are disregarded, any mechanical work done on the body must either generate
kinetic energy or else be stored within the body as strain energy. The energy that is
stored, per unit reference volume, is W, which is a function, W(A), of the local deforma-
tion gradient. If the deformation is suciently slow for isothermal conditions to prevail,
then W is the free energy function, with the temperature xed at its ambient value.
More generally, energy accounts must be performed, making allowance for temperature
(and entropy) variations and the input of heat, but this is not discussed further.
The total rate of input of mechanical work has already been given, as w, in equation
(6.14). The statement just made translates into the equation
w =
_
d
dt
_
_
B
0
[
1
2

0
v
i
v
i
+ W(A)] dX. (6.40)
Comparison of (6.14) and (6.40) (both of which also apply when B
0
is replaced by any
part of B
0
) shows that
S
i

A
i
=
dW(A)
dt
=
W
A
i

A
i
.
Since this equation remains true for any motion
15
, it follows that
S
i
=
W
A
i
. (6.41)
Equation (6.41) is the constitutive equation for nite-deformation elasticity.
There is, however, a restriction on the form of the function W: it has to be objective.
It is common experience that the stored energy of an elastic body undergoes no change if
it is subjected to a rigid motion. Hence, W can depend on A only in some combination
that recognises only length and angle changes that is, A must appear only through
some measure of strain. All are equivalent, since all can be constructed from (6.8) or
from U, but not all are equally convenient. If, however, W is expressed as a function of
15
It is a simple exercise to construct deformations which are independent of X, for which
A and

A take chosen values at some specic time.
68
some particular strain e
f
, it follows immediately from the denition of work-conjugacy
that the associated stress, T
f
, is given by
T
f

=
W
e
f

. (6.42)
Incremental deformations
This discussion is restricted to the case of a small, possibly dynamic, perturbation of a
static nite deformation, dened by a deformation gradient A
0
, say, and corresponding
nominal stress S
0
, so that
S
0
i
=
W
A
i
(A
0
).
This initial deformation is maintained by body force b
0
, so that, from the equation of
motion (6.12) in the case of no time-dependence,
S
0
i,
+
0
b
0
i
= 0. (6.43)
Some boundary condition must also be specied; for simplicity we assume that x =

0
(X) is prescribed for X B
0
.
Now change the body force to b
0
+ f and displace the boundary to
0
+ w, where
f and w are small, but possibly depend on time t. The body undergoes an increment
of displacement u and the total deformation gradient and nominal stress now have
components
A
i
= A
0
i
+ u
i,
and S
i
= S
0
i
+ s
i
, say.
The equation of motion (6.12) now implies, on taking account of (6.43),
s
i,
+
0
f
i
=
0
d
2
u
i
dt
2
(6.44)
and the boundary condition gives
u = w, X B
0
. (6.45)
The system (6.44), (6.45) is completed by appending an incremental version of the
constitutive equation (6.41). This is
s
i
= c
ij
u
j,
, (6.46)
69
where
c
ij
=

2
W
A
i
A
j
(A
0
). (6.47)
It should be noted that, although the tensor c does not have all of the usual symmetries
assigned to elastic moduli, it does have the crucial symmetry
c
ij
= c
ji
, (6.48)
which renders the equations self-adjoint. These equations form the basis for analy-
ses of stability and, when their solution is unique, most of the standard techniques of
linear elasticity can be deployed for their solution. Computational schemes for static
nite-deformation problems are also usually approached by an incremental formulation,
equivalent to the one given here.
Linear elasticity
Classical linear elasticity is a special case of small deformations superposed on a nite
deformation: it is necessary only to choose the special values b
0
= 0, A
0
= I, S
0
= 0.
The coordinate bases E

and e
i
can be taken to coincide and it is usual to take the
coordinates to be x
i
and avoid the use of Greek suxes. All stress tensors coincide
and increments too, in the absence of pre-strain and can be regarded as Cauchy stress,
with components T
ij
. All strain tensors likewise reduce to the innitesimal strain tensor
called e with components
e
ij
=
1
2
(u
i,j
+u
j,i
)
and, since the strain energy function W must depend upon e, the elastic moduli are
given by
c
ijkl
=

2
W
e
ij
e
kl
and have the symmetries
c
ijkl
= c
jikl
= c
klij
.
(This is consistent with the symmetry (6.13) of the stress tensor S, which follows when
A = I.)
The energy function at zero strain is taken to be zero, and zero strain is taken to
correspond to zero stress (unless pre-stress was a feature to be modelled). Hence, for
small strains, the energy function becomes
W(e) =
1
2
c
ijkl
e
ij
e
kl
. (6.49)
70
Since the energy at any non-zero level of strain should be positive, a restriction on the
quadratic form (6.49) is that it should be positive-denite.
In the special case of isotropy, the tensor of elastic moduli c takes the form
c
ijkl
=
ij

kl
+ (
ik

jl
+
il

jk
) . (6.50)
The energy function (6.49) can be written
W(e) =
1
2
e
ii
e
kk
+ e
ij
e
ij
=
1
2
e
ii
e
kk
+ e

ij
e

ij
,
where
= +
2
3

and
e

ij
= e
ij

1
3
e
kk
.
Positive-deniteness of (6.49) is then equivalent to the inequalities
> 0, > 0.
6.3 Some Energy Considerations
This section deals with the classical energy principles. Also, just to nish o, a discussion
of energy ux is included; this can equally well be done for a general continuum and
specialized to elasticity afterwards, so this sequence is followed. The topics have some
fundamental signicance. In particular, they have signicant bearing on the theory of
forces on defects.
The minimum energy principle
It is a reasonable physical postulate that an elastic body, when in equilibrium, adopts
the conguration that minimizes its total energy, allowing for the constraints to which
it is subjected
16
. This statement can be put into mathematical form, only once these
constraints have been made explicit. Although there are other possibilities, it will be
assumed here that the body is subjected to body force b(X) per unit mass and that, at
each point of its boundary B
0
, one of the pair x
i
, t
i
is prescribed, for each i, where t
i
16
Of course this can apply only to constraints which can be associated with a potential
energy.
71
denotes a prescribed value for N

S
i
. Thus, any loads that are prescribed are of dead
loading type; conguration-dependent loads (that depend on x) are more complicated.
The energy of the system comprising the body and its loading mechanism is now
E(x(X)) =
_
B
0
[W(A)
0
b
i
x
i
] dX
_
B
0
t
i
x
i
dS
0
, (6.51)
where the curly bracket implies evaluating the sum at any point X only over those values
of i for which t
i
is prescribed. The physical postulate that the body selects for itself the
function x(X) that minimizes E, subject to its components x
i
taking any values that may
be prescribed for X B
0
, is hard to verify mathematically (and, when W has several
minima corresponding to phase transformations, is a subject of active research) but it
is easy to verify that the equations of equilibrium are satised when E is stationary. To
see this, let x(X) be the solution and let u be any admissible variation. The statement
that, to rst order,
E(x +u) = E(x)
implies
_
B
0
_
W
A
i
u
i,

0
b
i
u
i
_
dX
_
B
0
t
i
u
i
dS
0
= 0. (6.52)
An application of the divergence theorem gives
_
B
0
[N

S
i
u
i
t
i
u
i
] dS
0

_
B
0
[S
i,
+
0
b
i
] u
i
dX = 0. (6.53)
The requirement that (6.53) should hold for any u(X) for which u
i
(X) = 0 whenever
x
i
is prescribed, generates the equilibrium equation (the time-independent version of
(6.12)) and the traction boundary conditions.
In the case of linear elasticity, W is a convex function of e and it is easy to prove
that the stationary point is a minimum. This follows from the calculation
E( x) E(x) =
_
B
0
[W(e) W(e) b
i
( x
i
x
i
)] dX
_
B
0
t
i
( x
i
x
i
)dS
0
,
where x is any admissible eld. It follows since W is convex that
W(e) W(e) ( e
ij
e
ij
)
W
e
ij
(e)
= ( e
ij
e
ij
)T
ij
.
72
This inequality, coupled with an application of the divergence theorem, gives the desired
result, that
E( x) E(x).
Notice that the proof, in this form, requires W to be convex but not necessarily quadratic,
and so applies also to physically nonlinear problems, under the assumption of small
deformations (the deformation theory of plasticity falls within this class).
The complementary energy principle
Although there is a stationary principle of complementary energy for nite deformations,
a minimum principle has only been established in the case of small deformations. Then,
W is a convex function of e and there is no diculty in dening a complementary energy
density function
W

(T) = sup
e
[T
ij
e
ij
W(e)] . (6.54)
The supremum is attained when
T
ij
=
W
e
ij
;
the equality
e
ij
=
W

T
ij
is satised simultaneously. The complementary energy principle states that
F(T) =
_
B
0
W

(T)dX
_
B
0
N
j
T
ji
x
i
dS
0
is minimized by the actual stress eld T, amongst stress elds that satisfy the equilibrium
equations and any prescribed traction boundary conditions (N
j
T
ji
= t
i
). The double
curly bracket in (6.54) implies summation only over those values of i for which x
i
is
prescribed. The function W

is convex and the proof follows that outlined for the

minimum energy principle.
Energy ux considerations
This section is valid for any continuum and even thermal eects are admitted; the
formulae can easily be specialized to elasticity. Suppose that (part of) a body occupies a
domain B
0
in the reference conguration, as discussed earlier. It is subjected to surface
forces t
i
= N

S
i
. In addition, there is a ux of heat out of B
0
across the surface
73
B
0
, which is expressed as N

per unit area in the reference conguration, so that q

represents a nominal heat ux vector. For simplicity, it is assumed that there is no
body force or direct input of heat from an external source, except through the boundary.
Since energy is conserved, the rate of energy input into B
0
must equal the rate of increase
of energy within B
0
. Thus,
_
B
0
[N

S
i
v
i
N

]dS
0
=
_
B
0
d
dt
[U +
1
2

0
v
i
v
i
] dX, (6.55)
where U denotes the internal energy per unit reference volume (this will, in general, be
a function of the current state, as specied by the deformation gradient A, the entropy
and some set of internal variables). Application of the divergence theorem to (6.55),
coupled with the equation of motion (6.12) with b = 0, now implies the local energy
balance equation
S
i
v
i,
q
,
=

U, (6.56)
since (6.55) must apply for any domain B
0
.
So far, the domain B
0
, once chosen, is xed. It transforms, during the motion of
the body, to a domain B
1
, which depends upon t but always consists of the same set of
material points. Now consider, however, a more general case, where a domain B
1
(t) is
chosen a priori, restricted only so that it varies smoothly with t. Such a domain maps
back onto the reference conguration, through the inverse of the mapping (6.2), to a
domain B
0
(t), which now depends on t. Of course, B
0
(t) could be chosen rst, to induce
a corresponding B
1
(t). The rate of change of energy within B
1
(t) becomes
_
d
dt
_
_
B
0
(t)
[U +
1
2

0
v
i
v
i
] dX
=
_
B
0
(t)
d
dt
[U +
1
2

0
v
i
v
i
] dX+
_
B
0
(t)
[U +
1
2

0
v
i
v
i
] (N

)dS
0
, (6.57)
where the components of the velocity of B
0
are C

(these may depend on position

X B
0
). Use of the divergence theorem, the local energy balance (6.56) and the
equation of motion shows that the volume integral on the right side of (6.57) satises
(6.56), even though B
0
depends on t, and hence
_
d
dt
_
_
B
0
(t)
[U +
1
2

0
v
i
v
i
] dX =
_
B
0
(t)
N

[S
i
v
i
q

+ C

(U +
1
2

0
v
i
v
i
)] dS
0
. (6.58)
The right side of (6.58) may thus be interpreted as the ux of energy across the moving
surface B
1
(t), whose image in the reference conguration is B
0
(t).
74
6.4 Inelastic constitutive equations
The study of the response of solids which are not elastic of course is extremely broad.
This section just presents the briefest outline, sucient for the purposes of these notes.
A rather general framework for discussing the response of solids is to postulate that
stress at some material point X and some time t depends on the entire history of the
strain at the point X, at all times before t, and including the time t. (Non-local
continua, in which stress depends on the strain in a neighbourhood of X are excluded for
the present.) Conversely, the strain at X at time t depends on the history of the stress.
We shall assume here that stress and strain are some chosen work-conjugate pair,
as discussed earlier.
No progress can be made unless some further structure is assumed. Here, we suppose
that the rate of strain at time t is expressible in the form
e
f
= F(T
f
,

T
f
, H), or e
f

= F

(T
f
,

T
f
, H), (6.59)
where F is a (tensor-valued) function and H symbolically denotes the entire history of
the deformation at times prior to t, perhaps encapsulated in some set of parameters,
called state variables. A possible response to an imposed stress is displayed in Fig.
6.3. This depicts just a uniaxial stress S and the corresponding component e of strain
17
.
The stress increases from zero at O, up to a maximum at B, and then decreases down to
C. The noteworthy feature is that the response is linear from O to A. Then, it deviates
signicantly from linear behaviour. Of course the entire curve that is shown will depend
not only upon the stress but also on its rate of change. That is, the gure actually
shows the curve obtained by eliminating the time t between the stress and the strain,
both obtained as functions of t.
It is common experience that the linear portion OA that is depicted really does exist
for the majority of solid materials, and that this portion is independent of the rate of
change of the stress. Thus, at low stress or strain, the material behaves according to
linear elasticity. The point A is not so easy to dene rigorously. It is true that a rapid
deviation from linear elastic behaviour occurs in the vicinity of A. The phenomenon is
called plastic yielding. Exactly where it is considered to occur depends on the sensi-
tivity of the experiment. It is probably true to assert that there is some contamination
of the linear elastic response, even at very small stress or strain, but this is insignicant
until the point A is reached. Thus, the point A will be dened in practice as that point
beyond which the linear elastic approximation is no longer acceptable.
17
S can be regarded as the 11 component of nominal stress and e as the proportional
elongation,
1
1; these are natural observables in a simple tension test.
75
Fig. 6.3. Representation of elastoplastic response.
Now consider the point B, at which the stress is reduced: the incremental response
again becomes linear, but if the stress were reduced to zero there would be some residual
strain. Again, exactly what happens near B will depend on the rate of change of the
stress, and some deviation from linearity will occur, though this can almost invariably
be ignored in practice.
Unless the rate of stress (or, equivalently, the strain-rate) is very large, it is found for
metals at moderate temperatures (less than one-third the melting point, for instance)
that the dependence on stress- or strain-rate is negligible. An example stress-strain
relation, in one dimension, that illustrates this is
e =

S/E + e
p
, (6.60)
where the plastic strain-rate e
p
is given by
e
p
= A([S[/S
0
)
N
S/[S[, (6.61)
where E is an elastic modulus, S
0
depends on the history H in some rate-independent
fashion (such as H =
_
t
0
[ e
p
[dt) and N 1. Since N is large, the stress will remain
close to S
0
so long as S S
0
and is increasing with time, since otherwise the strain-rate
would be large. Conversely, if [S[ < S
0
, the plastic strain-rate is negligible. Thus, to
lowest approximation, the response (6.60) during continued loading can be expressed in
the form
e =
_
1
E
+
1
h
_

S, (6.62)
76
where h = S

0
(H) if H =
_
t
0
[ e
p
[ dt. This approximation will not be uniform near the
point A, which requires the more complete relation (6.60, 2.61) for resolution.
However, for many practical purposes, the rate-independent idealisation is sucient.
This is now developed more systematically. Some generic instant, at which the stress
and strain tensors have the values T
f
, e
f
, is considered. There is an elastic domain in
stress space (and a corresponding domain in strain space). The stress lies either within
the elastic domain or on its boundary. The boundary is called the yield surface. If
the stress lies within the elastic domain, then increments of stress and strain are related
elastically: say
e
f

= /

T
f

. (6.63)
The compliance tensor /

may depend on whatever inelastic deformation has oc-

curred previously that is, upon H. Its inverse, L

, is the tensor of elastic moduli.

If the stress is currently at the boundary of the elastic domain that is, on the yield
surface the relation (6.63) still applies if the stress-rate is directed into the interior of
the elastic domain. This is assumed to be convex, so the appropriate restriction when
the stress is on the yield surface is that the stress-rate lies within a cone / in stress-rate
space, which locally denes the interior of the elastic domain. In the simplest case, the
yield surface is smooth in a neighbourhood of T
f
; then, / will have the form
/ =

T
f
: Q

T
f

0, (6.64)
where Qis symmetric and depends upon the local state of stress and strain (and history).
However, it is also possible that the yield surface might have a vertex. If T
f
is on the
yield surface and its rate

T
f
lies outside the cone /, then T
f
remains on the yield
surface which therefore, in general, moves as the stress changes. The strain-rate has a
plastic contribution as well, so that
e
f

= /

T
f

+F

(

T
f
, H), (6.65)
where F

are homogeneous functions of degree one in

T
f
, and H represents history, as
before. Most models of plasticity, in fact, can be expressed, when the current stress is
on the yield surface,
e
f

= /

T
f

if

T
f
/
= M

T
f

otherwise, (6.66)
where the tensor M is a homogeneous function of degree zero in

T
f
. The case described
as otherwise is referred to as during plastic loading, or during plastic deformation.
The elastic relation also applies when the stress is within the elastic domain.
77
In the case that the cone takes the form (6.64) at a smooth point on the yield
surface the relation (6.66) is often given in the form
e
f

= /

T
f

if Q

T
f

0
= /

T
f

+ P

(Q

T
f

)/h otherwise. (6.67)

Thus, in this case,
M

= /

+ P

/h. (6.68)
It is usual to normalise P and Q so that P

= 1 (and similarly for Q).

Another very common assumption (good for metals but not applicable to granular
media, for instance) is that of normality: during loading, the increment of plastic
strain must lie within the cone of normals to the yield surface. In the simple case
represented by (6.68), this means simply that P = Q.
In summary, therefore, our theory of plasticity is as follows: the constitutive equation
will be taken to be (6.66). Usually, the instantaneous compliances M

will have the

symmetry M

= M

. The stress-rate versus strain-rate relation will be assumed

to be invertible. Thus, when the stress is on the yield surface,

T
f

= L

e
f

if

T /,
= L

e
f

if

T / /. (6.69)
The tensor L is homogeneous of degree zero in

T
f
. Equivalently, L may be considered
to be a homogeneous function of degree zero of e
f
.
Three remarks will be made, in conclusion. The rst is that nonlinear elastic response
can be described in the form

T
f

= L

e
f

, for any increment of stress. It is necessary

only to take
L

=

2
W
e
f

e
f

(e
f
),
which automatically has the symmetry given above. The second is that materials other
than those normally considered as elastic-plastic may have local constitutive response
with the form (6.66). The last is that, if the relation (6.66) is known for any conjugate
stress-strain pair, then the corresponding relation can be deduced for any other. Further-
more, a similar relation can be deduced for the relation between nominal stress and defor-
mation gradient, by use of (6.32), (6.33) and (6.34). In the particular case that T
f
is the
second Piola-Kirchho stress, S
i
= A
i
T
(2)

from (6.18), and e

(2)

=
1
2
(

A
k
A
k
+A
k

A
k
).
Therefore,

S
i
= c
ij

A
j
, (6.70)
78
where
c
ij
= A
i
L

A
j
+
ij
S
k
B
k
if

S /

,
= A
i
L

A
j
+
ij
S
k
B
k
if

S / /

, (6.71)
where /

is dened so that

S /

implies

T
f
/, and conversely.
6.5 Virtual work; relation to nite element computation
It has been seen in Section 2.3 that an elastic body subjected to loads of a certain fairly
general type is in equilibrium when it attains a minimum-energy conguration. In fact,
equation (6.53) was shown to generate the eld equations and the boundary conditions.
Now consider the opposite reasoning, for a somewhat more general problem. Assume
the equations of motion (6.12), together with initial conditions
x = x
0
(X), x = x
0
(X), X B
0
, t = 0 (6.72)
and boundary conditions of the type considered in Section 2.3: one of the pair (x
i
, N

S
i
)
is given at each point of the boundary, for each i, except that the traction component
N

S
i
, where given, may be conguration-dependent,
N

S
i
=
i
(x, A, X, t). (6.73)
The functions
i
are restricted so that they only involve derivatives of x in directions
tangent to the surface, so that they can be expressed in terms of surface values of x
alone.
Now multiply the equation of motion (6.12) by a virtual displacement w
i
(X, t),
sum over i and integrate over B
0
. Use of the divergence theorem then gives
_
B
0
N

S
i
w
i
dS
0
=
_
B
0
[w
i,
S
i

0
b
i
w
i
+
0
v
i
w
i
]dX. (6.74)
This is true for any eld w
i
. Now restrict w
i
so that w
i
= 0 wherever x
i
is prescribed.
Then, the left side of (6.74) reduces to
_
B
0

i
w
i
dS
0
=
_
B
0
[w
i,
S
i

0
b
i
w
i
+
0
v
i
w
i
]dX, (6.75)
where the curly bracket implies summation over only those values of i for which
i
is
given.
79
The statement (6.75) permits the construction of nite-element approximations. The
eld x
i
is approximated as
x
i
=

K
U
K
i
(t)
K
(X), (6.76)
where the functions
K
take the value 1 at the node labelled K and are zero at all
other nodes. Thus, U
K
i
provides an approximation for u
i
, evaluated at the node K. Now
w
i
is given similarly:
w
i
=

K
W
K
i
(t)
K
(X). (6.77)
The requirement that (6.75) should be satised for all w
i
of the form (6.77) generates a
system of nonlinear equations for the nodal values U
K
i
. Formally, the system has the
structure
F
Ki
(U) =

L
M
KL

U
L
i
, or F(U) = M

U, (6.78)
together with initial conditions. The function (or functional) F cannot easily be given
explicitly in the general case. It requires that the nominal stress be expressed, through
the constitutive equation, in terms of the displacement eld, in the approximation rep-
resented by (6.76). The boundary term
i
has to be similarly approximated, in terms
of the values of the displacements at the boundary nodes. The mass matrix has the
components
M
KL
=
_
B
0

L
dX. (6.79)
The practical implementation of the nite element method is beyond the scope of
this set of notes.
Appendix 6A: Calculations relating to stress- and strain-rates
The problem that is addressed is to calculate the rates of change of the principal stretches
r
and the
corresponding principal directions u
(r)
(r = 1, 2, 3). They satisfy the equations
A
T
Au
(r)
=
2
r
u
(r)
, or A
k
A
k
u
(r)

=
2
r
u
(r)

. (6A.1)
Dierentiating with respect to time gives
(

A
k
A
k
+ A
k

A
k
)u
(r)

+ A
k
A
k
u
(r)

= 2
r

r
u
(r)

+
2
r
u
(r)

. (6A.2)
Now multiply (6A.2) by u
(r)

and sum over . This gives

2
r

r
= u
(r)

(

A
k
A
k
+ A
k

A
k
)u
(r)

, (6A.3)
80
since (6A.1) holds, and u
(r)
is a unit vector. Substituting the now-known value of

r
back into (6A.2)
gives
(A
k
A
k

2
r

) u
(r)

= u
(r)

[u
(r)

(

A
k
A
k
+ A
k

A
k
)u
(r)

] (

A
k
A
k
+ A
k

A
k
)u
(r)

. (6A.4)
The matrix on the left side of this equation is singular, but a solution exists because the right side
is orthogonal to the eigenvector u
(r)
. It is rendered unique by the requirement that u
(r)
must be
orthogonal to u
(r)
, since the latter must remain a unit vector. The explicit solution is obtained by
writing (6A.4) in the spectral representation

s=r
(
2
s

2
r
)u
(s)

u
(s)

) u
(r)

= u
(r)

[u
(r)

(

A
k
A
k
+ A
k

A
k
)u
(r)

] (

A
k
A
k
+ A
k

A
k
)u
(r)

. (6A.5)
Inversion is immediate. The matrix on the left, regarded as an operator on the subspace spanned by
the eigenvectors u
(s)
with s ,= r, has inverse with component

s=r
1

2
s

2
r
u
(s)

u
(s)

.
Thus,
u
(r)

s=r
1

2
r

2
s
u
(s)

[u
(s)

(

A
k
A
k
+ A
k

A
k
)u
(r)

]. (6A.6)
These values may now be substituted into (6.24) to give
e
f

r
_
f

(
r
)
2
r
u
(r)

u
(r)

u
(r)

(

A
k
A
k
+ A
k

A
k
)u
(r)

s=r
f(
r
)

2
r

2
s
(u
(r)

u
(s)

+ u
(s)

u
(r)

)[u
(r)

(

A
k
A
k
+ A
k

A
k
)u
(s)

]
_
. (6A.7)
Since
e
f

=
e
f

A
i

A
i
,
a formal dierentiation of (6A.7) with respect to

A
i
yields the result (6.26).
The case A = I is degenerate, and it is easier to start again than to take the necessary limits.
Equation (6A.1) is satised identically. The eigenvectors u
(r)
are arbitrary at the precise instant that
A = I, but it is sensible to require them to be continuous functions of time. The relation (6A.2) reduces
to
1
2
(

A
k

k
+
k

A
k
)u
(r)

r
u
(r)

. (6A.8)
Equivalently,

ij

j
u
(r)

i
u
(r)

.
Thus,

r
(r = 1, 2, 3) are the eigenvalues of the Eulerian strain-rate tensor, and u
(r)
are the
corresponding eigenvectors, resulting in the spectral representation (6.29).
81
Now dierentiate (6A.2) with respect to time, and afterwards set A = I (and
r
= 1). This gives
(

A
k

k
+
k

A
k
+ 2

A
k

A
k
)u
(r)

+ (

A
k

k
+
k

A
k
) u
(r)

= 2(

r
+

2
r
)u
(r)

+ 4

r
u
(r)

. (6A.9)
Multiplying by u
(r)

and summing over gives

u
(r)

(

A
k

k
+
k

A
k
+ 2

A
k

A
k
)u
(r)

= 2(

r
+

2
r
). (6A.10)
Then, substituting back into (6A.9) gives
(

A
k

k
+
k

A
k
2

) u
(r)

= [u
(r)

(

A
k

k
+
k

A
k
+ 2

A
k

A
k
)u
(r)

]
k
(

A
k

k
+
k

A
k
+ 2

A
k

A
k
)u
(r)

(6A.11)
This equation can be treated in the same way that (6A.5) was. It has solution
u
(r)

=
1
2

s=r
1

s
u
(s)

u
(s)

[
1
2
(

A
k

k
+
k

A
k
) +

A
k

A
k
]u
(r)

. (6A.12)
Now dierentiating (6.24) with respect to time, and then setting A = I, gives
e
f

r
(f

(1)

2
r
+

r
)u
(r)

u
(r)

+ 2

r
( u
(r)

u
(r)

+ u
(r)

u
(r)

). (6A.13)
Also,
e
f

=
e
f

A
i

A
i
+

2
e
f

A
i
A
j

A
i

A
j
. (6A.14)
Therefore, the expressions (6A.12) for u
(r)

and (6A.10) for

r
can be substituted into (6A.13), and
comparison of the resulting expression with (6A.14) will give
2
e
f

/A
i
A
j
. It suces for this
purpose to consider the special case

A = 0. It is noted, too, that

2
r
u
(r)

u
(r)

can be given explicitly,

by exploiting the spectral representation of the operator in (6A.8): the sum is just that operator,
squared.
Implementing this plan gives the result

2
e
f

A
i
A
j

A
i

A
j
=
_
1
4
(f

(1) 1)(

A
k

k
+

A
k

k
)(

A
l

l
+

A
l

k
)
+

r
u
(r)

u
(r)

u
(r)

A
k

A
k
u
(r)

+
1
2

s=r
(u
(s)

A
k

A
k
u
(r)

)(u
(s)

u
(r)

+u
(r)

u
(s)

)
_
. (6A.15)
Formal dierentiation of this result with respect to

A
i
and

A
j
gives twice
2
e
f

/A
i
A
j
. The
result (6.34) follows directly.
82